University of Southern California Dissertations and Theses

Climate change, air pollution and health: how policy can induce adaptation

(USC Thesis Other)

Climate change, air pollution and health: how policy can induce adaptation

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content CLIMATE CHANGE, AIR POLLUTION AND HEALTH: HOW POLICY CAN INDUCE ADAPTATION by Noah Miller A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (PUBLIC POLICY AND MANAGEMENT) August 2024 Copyright 2024 Noah Miller Dedication To Bao and Qi, your love and laughter makes everything better. ii Acknowledgements I couldn’t have made it to this point without the help of many people in my life. I’d like to take this opportunity to thank them for their support. First, my sincere thanks to my dissertation committee. Dr. Adam Rose, who got me started down the path of environmental economics and has been a source of constant support. Dr. Edson Severnini, whose advice, insights, and humor have helped me during tough times. And of course, my wonderful advisor Dr. Antonio Bento, who has pushed me to not just succeed, but constantly improve, and who has always been quick to offer his knowledge, experience, and wisdom. I’d also like to thank Julie Kim, for helping me navigate the vagaries of academic bureaucracy. Next, I’d like to thank my partner, Yadanar. You light up my life and are a source of constant joy and inspiration. I couldn’t have done this without you. Last but not least, I’d like to acknowledge the immense support provided by my family and friends over these past years. Thank you. iii TABLE OF CONTENTS Dedication. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi Chapter 1: A Unifying Approach to Measuring Climate Change Impacts and Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 I Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 1 II Prior Methods and Our Unifying Approach to Measuring Climate Change Impacts and Adaptation . . . . . . . . . . . . . . 6 A Prior Methods . . . . . . . . . . . . . . . . . . . . . . . 6 B Our Unifying Approach . . . . . . . . . . . . . . . . . . . 8 C Decomposition of Meteorological Variables: Climate Norms vs. Weather Shocks . . . . . . . . . . . . . . . . . . . . . . . 10 III Empirical Application: Climate Impacts on Ambient Ozone . . . . . 14 A Conceptual Framework . . . . . . . . . . . . . . . . . . . 15 B Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 C Empirical Strategy . . . . . . . . . . . . . . . . . . . . . 22 IV Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 A Impacts of Temperature on Ambient Ozone Concentration . . . . 27 B Measuring Adaptation to Climate Change . . . . . . . . . . . 30 C Robustness Checks . . . . . . . . . . . . . . . . . . . . . 31 D Estimating Nonlinear Effects of Temperature . . . . . . . . . . 36 E Exploring Heterogeneity . . . . . . . . . . . . . . . . . . . 39 V Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . 40 iv Chapter 2: Incidental Adaptation: The Role of Non-Climate Regulations. . . . . . 42 I Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 42 II Analytical Framework . . . . . . . . . . . . . . . . . . . . . 47 A The Nature of Existing Regulations Influencing Adaptation. . . . 48 B A Schematic Representation of the Framework for Ambient Ozone and NAAQS . . . . . . . . . . . . . . . . . . . . . . 52 III Data and Data Descriptions . . . . . . . . . . . . . . . . . . 56 A NAAQS, Ozone Pollution, and Climate: Background and Data . . 56 B Basic Trends in Pollution, Attainment Status, and Weather: Implications for the Importance of Regulations . . . . . . . . . . 58 IV Empirical Framework . . . . . . . . . . . . . . . . . . . . . 62 V Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 A The Role of Regulations for Inducing Adaptation to Climate Change . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 B Robustness Checks . . . . . . . . . . . . . . . . . . . . . 74 C Heterogeneity in Regulation-Induced Adaptation . . . . . . . . 85 D Climate Adaptation Co-Benefits from Existing Regulations: Some Calculations . . . . . . . . . . . . . . . . . . . . . . . . . 88 VI Concluding Remarks . . . . . . . . . . . . . . . . . . . . . 91 Chapter 3: Policy Induced Defensive Investment: The Role of Information in Reducing the Healthcare Burden of Air Pollution . . . . . . . . . . . . 94 I Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 94 II Ozone Formation, Seasonality, and Wind Direction . . . . . . . . . 100 III Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 A Health Data . . . . . . . . . . . . . . . . . . . . . . . . 103 B Air Pollution Data . . . . . . . . . . . . . . . . . . . . . 104 C Wind & Other Meteorological Data . . . . . . . . . . . . . . 106 IV Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . 107 A Health Impacts of Acute (Unexpected) and Long-run (Expected) Air Pollution . . . . . . . . . . . . . . . . . . . . . . . . . 107 B Decomposing Ozone . . . . . . . . . . . . . . . . . . . . 113 C Instrumenting with Wind Direction . . . . . . . . . . . . . . 115 V Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 A Impacts of Ambient Ozone Concentration on Health. . . . . . . 118 B Robustness Checks . . . . . . . . . . . . . . . . . . . . . 123 VI Concluding Remarks . . . . . . . . . . . . . . . . . . . . . 133 v Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 Appendix A: A Unifying Approach to Measuring Climate Change Impacts and Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 A.1 Additional Data Discussion . . . . . . . . . . . . . . . . . . 149 A.1.a Background Details on Ozone . . . . . . . . . . . . . . . 150 A.1.b Further Details on the Construction of the Data . . . . . . . 151 A.2 Further Robustness Checks and Heterogeneity . . . . . . . . . . 169 A.2.a Further Robustness Checks . . . . . . . . . . . . . . . . 170 A.2.b Heterogeneity. . . . . . . . . . . . . . . . . . . . . . 176 A.3 Sources of Variation to Identify Climate Impacts . . . . . . . . . 193 Appendix B: Incidental Adaptation: The Role of Non-Climate Regulations . . . . . 195 B.1 The National Air Quality Standards, Ozone Formation, and Additional Data Discussion . . . . . . . . . . . . . . . . . . . . . 196 B.1.a Background Details on the National Ambient Air Quality Standards . . . . . . . . . . . . . . . . . . . . . . . . . . 197 B.1.b Background Details on Ozone . . . . . . . . . . . . . . . 199 B.1.c Further Details on the Construction of the Data . . . . . . . 200 B.2 Further Robustness Checks and Heterogeneity . . . . . . . . . . 221 B.2.a Further Robustness Checks . . . . . . . . . . . . . . . . 222 B.2.b Heterogeneity. . . . . . . . . . . . . . . . . . . . . . 229 B.3 Formal Derivations of Analytical Results . . . . . . . . . . . . 248 B.3.a Model Derivations . . . . . . . . . . . . . . . . . . . . 248 B.3.b Model Extensions . . . . . . . . . . . . . . . . . . . . 252 vi List of Tables Table 1.1. Climate Impacts and Adaptation – Our Unifying Approach vs. Prior Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 Table 1.2. Key Robustness Checks . . . . . . . . . . . . . . . . . . . . . . 33 Table 2.1. Climate Impacts on Ambient Ozone and Adaptation . . . . . . . . . . 72 Table 2.2. Parallel Trends & Alternative Outcomes . . . . . . . . . . . . . . . 76 Table 2.3. Accounting for Competing Input Regulations Aimed at Ambient Ozone Reductions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 Table 2.4. Results by Distance of Ozone Concentrations to NAAQS Threshold . . . 83 Table 2.5. Implied Impacts of Ambient Ozone Climate Penalty . . . . . . . . . . 90 Table 3.1. Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . 104 Table 3.2. OLS and IV Estimates of the Effects of Acute Ozone Shocks and Expected Ozone Norms on Elderly Hospitalization for Cardiovascular or Respiratory Disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 Table 3.3. Parallel Trends . . . . . . . . . . . . . . . . . . . . . . . . . 125 Table 3.4. Moving Average Norm Alternative Lengths . . . . . . . . . . . . . 127 Table 3.5. Controlling for Other Air Pollutants . . . . . . . . . . . . . . . . 129 Table 3.6. Further Robustness Checks & Alternative Health Outcomes . . . . . . 131 Table A.1.1.Yearly Summary Statistics for Daily Maximum Temperature . . . . . . 165 Table A.1.2.Yearly Summary Statistics for Ozone Monitoring Network . . . . . . . 166 Table A.1.3.Ozone Monitoring Season by State . . . . . . . . . . . . . . . . . 167 Table A.1.4.County Summary Statistics by Belief in Climate Change . . . . . . . 168 Table A.2.1.Alternative Criteria for Selection of Weather Stations . . . . . . . . . 181 Table A.2.2.Comparison to Alternative Estimation Methods (Semi-Balanced Panel) . 182 Table A.2.3.Excluding Areas with Regional Air Pollution Policies . . . . . . . . . 183 Table A.2.4.Further Robustness Checks . . . . . . . . . . . . . . . . . . . . 184 Table A.2.5.Bootstrapped Standard Errors . . . . . . . . . . . . . . . . . . . 185 Table A.2.6.Non-Linear Effects of Temperature . . . . . . . . . . . . . . . . . 186 Table A.2.7.Comparison of Adaptation Under Nonlinear Specifications . . . . . . . 187 Table A.2.8.Results by Decade . . . . . . . . . . . . . . . . . . . . . . . . 188 Table A.2.9.Adaptation by Belief in Climate Change . . . . . . . . . . . . . . 189 Table A.2.10.Adaptation by Belief in Climate Change Regulation . . . . . . . . . 190 Table A.2.11.Adaptation by Political Leaning . . . . . . . . . . . . . . . . . 191 Table A.2.12.Adaptation by VOC- or NOx-limited Atmosphere . . . . . . . . . . 192 vii Table B.1.1.History of Ambient Ozone NAAQS . . . . . . . . . . . . . . . . . 217 Table B.1.2.Ozone Monitoring Season by State . . . . . . . . . . . . . . . . . 218 Table B.1.3.Yearly Summary Statistics for Ozone Monitoring Network . . . . . . . 219 Table B.1.4.Yearly Summary Statistics for Temperature and Decomposition . . . . 220 Table B.2.1.Alternative Lag Lengths of Nonattainment Indicator . . . . . . . . . 235 Table B.2.2.Alternative Lengths of Climate Norms . . . . . . . . . . . . . . . 236 Table B.2.3.Adaptation Responses . . . . . . . . . . . . . . . . . . . . . . 237 Table B.2.4.Alternative Specifications and Sample Restrictions . . . . . . . . . . 238 Table B.2.5.Alternative Criteria for Selection of Weather Stations . . . . . . . . . 239 Table B.2.6.Bootstrapped Standard Errors . . . . . . . . . . . . . . . . . . . 240 Table B.2.7.Results by Decade . . . . . . . . . . . . . . . . . . . . . . . . 241 Table B8a. Non-Linear Effects of Temperature . . . . . . . . . . . . . . . . . 242 Table B8b. Non-Linear Effects of Temperature (continued) . . . . . . . . . . . . 243 Table B9. Adaptation by Local Beliefs in Climate Change . . . . . . . . . . . . 244 Table B10. Beliefs in Climate Change: Summary Stats . . . . . . . . . . . . . 245 Table B11. Placebo: Preferences for Single Parenting . . . . . . . . . . . . . . 246 Table B12. Adaptation by VOC- or NOx-limited Atmosphere . . . . . . . . . . . 247 viii List of Figures Figure 1.1. Theoretical Relationship Between Marginal Cost of Dirty Production and Temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 Figure 1.2. Climate Norms and Shocks . . . . . . . . . . . . . . . . . . . . . 21 Figure 1.3. Decomposition of Temperature Norms & Shocks (Los Angeles, 2013) . . . 24 Figure 1.4. Comparing Linear, Binned, and Nonlinear Specifications . . . . . . . . 38 Figure 2.1. Conceptual Framework on Regulation-Induced Adaptation . . . . . . . 54 Figure 2.2. Evolution of Maximum Ozone Concentration and Counties in Nonattainment . . . . . . . . . . . . . . . . . . . . . . . . . . 59 Figure 2.3. Climate Norms and Shocks Over the Period of Analysis (1980-2013) . . . 61 Figure 3.1. Relationship Between Wind Direction and Ozone Concentration Shocks and Norms for Counties in and Around the Greater Boston Area, MA . . 102 Figure 3.2. Evolution of Ambient Ozone Air Pollution and Monitoring (1980 – 2017) . 105 Figure 3.3. Distributions of Daily Ozone Concentrations in Counties with or without Alert Day Programs . . . . . . . . . . . . . . . . . . . . . . . 112 Figure 3.4. Decomposition of Ambient Ozone Concentration into Daily Shocks and Monthly Norms . . . . . . . . . . . . . . . . . . . . . . . . . 115 Figure A.1.1.Temperature Relative to Baseline (1950-1979) . . . . . . . . . . . . 156 Figure A.1.2.Comprehensive Location of Weather Monitors . . . . . . . . . . . 157 Figure A.1.3.Climate Norms and Shocks (semi-balanced sample) . . . . . . . . . 158 Figure A.1.4.Climate Norms and Shocks (main model sample) . . . . . . . . . . 159 Figure A.1.5.Evolution of Maximum Ambient Ozone Concentration . . . . . . . . 160 Figure A.1.6.Ozone Monitor Location by Decade of First Appearance . . . . . . . 161 Figure A.1.7.Ozone Monitors and their Matched Weather Monitors . . . . . . . . 162 Figure A.1.8.Relationship between Ozone and Decomposed Temperature . . . . . . 163 Figure A.1.9.Decomposition of Temperature Norms & Shocks (Los Angeles, All Years) . . . . . . . . . . . . . . . . . . . . . . 164 Figure A.2.1.Climate Impacts and Adaptation Over Time . . . . . . . . . . . . 180 Figure B.1.1.Ozone Monitor Location by Decade of First Appearance . . . . . . . 207 Figure B.1.2.Temperature Relative to Baseline (1950-1979) . . . . . . . . . . . . 208 Figure B.1.3.Ozone Monitors and Matched Weather Monitors . . . . . . . . . . 209 Figure B.1.4.Evolution of 4th Highest Ozone Concentration . . . . . . . . . . . 210 Figure B.1.5.Evolution of Nonattainment Designation in Monitored Counties . . . . 211 Figure B.1.6.Decomposition of Climate Norms and Shocks – Estimating Sample . . . 212 Figure B.1.7.Relationship between Ambient Ozone and Temperature . . . . . . . 213 ix Figure B.1.8.Decomposition of Temperature Norms and Shocks (Los Angeles, 2013) . . . . . . . . . . . . . . . . . . . . . . . . 214 Figure B.1.9.Decomposition of Temperature Norms and Shocks (Los Angeles, All Years) . . . . . . . . . . . . . . . . . . . . . . 215 Figure B.1.10.Evolution of Ozone Concentration by Belief in Climate Change . . . . 216 x Abstract Climate change is unequivocal, with global temperatures expected to rise from 1.5 to 4.5C over the 21st century. Therefore, it is crucial to develop methods to measure climate impacts and adaptation. This dissertation first develops a unifying approach to measure both climate impacts and adaptation in the same estimating model, applying the method to examine the “climate penalty” on ambient ozone air pollution. The second chapter then examines how policies interact with climate change and preexisting market-failures to induce or inhibit adaptation. Specifically, this chapter develops a tractable analytical framework of a corrective regulation where the market failure interacts with climate, highlighting the mechanism through which the regulation may also incidentally induce climate adaptation. Combined with the unifying approach from chapter one, the analysis finds that the Clean Air Act incidentally induces statistically and economically significant climate adaptation co-benefits with respect to controlling ozone pollution. The final chapter extends the prior methods to allow for the examination of whether policy-induced adaptation complements or crowds out adaptation that economic agents would otherwise intrinsically engage in. Specifically, this chapter examines the negative health impacts of ozone exposure, and whether economic agents intrinsically adapt to expected changes in ambient ozone – often referred to as “defensive investments”. Taking advantage of the staggered implementation of air quality alert programs across the US between 2004 and 2017, the analysis finds no evidence that air quality alert programs crowd-out intrinsic defensive investments, with suggestive – though statistically insignificant – evidence that they in fact increase intrinsic defensive-investments. xi Chapter 1: A Unifying Approach to Measuring Climate Change Impacts and Adaptation I. Introduction Failure to achieve climate mitigation goals puts increasing pressure on climate adaptation strategies.1 Therefore, it is crucial to develop methods to measure climate impacts and adaptation. Inspired by the macroeconomic literature on the effects of unanticipated versus anticipated shocks on the economy (e.g., Lucas, 1972), the labor literature on the importance of distinguishing transitory versus permanent income shocks (e.g., Solon, 1992), and the properties of the Frisch-Waugh-Lovell theorem (Frisch and Waugh, 1933; Lovell, 1963), we develop a unifying approach to measuring climate impacts and adaptation. The proposed approach is then applied to examine the impact of climate change on ambient “bad” ozone concentration in U.S. counties over the period 1980-2013. Ozone is not emitted directly into the air, but rather formed by a Leontief-like production function of Nitrogen Oxides (NOx) and Volatile Organic Compounds (VOCs) in the presence of sunlight and warm temperatures; hence, affected by climate change (e.g., Jacob and Winner, 2009). Our unifying approach overcomes key challenges of the literature by decomposing meteorological conditions into climatic variation and weather shocks, and estimating climate and weather effects in the same panel fixed-effects equation. The pioneer cross-sectional approach to estimate the impact of climate change on economic outcomes (Mendelsohn, 1According to the Sixth Assessment Report from the Intergovernmental Panel on Climate Change (IPCC, 2022), the warming of the climate system is unequivocal, and global temperatures are likely to rise from 1.5 to 4.5 degree Celsius over the 21st century, depending on the emissions scenario. 1 Nordhaus and Shaw, 1994) has relied on permanent, anticipated components behind meteorological conditions, but may suffer from omitted variable bias. In contrast, the panel fixed-effects approach (Deschenes and Greenstone, 2007) exploits transitory, unanticipated weather shocks, and deals with that bias, but identification of climate effects using weather variation is not trivial. Current hybrid approaches combining cross-sectional and panel data variation also face challenges (see a recent review by Kolstad and Moore, 2020). The partitioning variation approach also decomposes meteorological conditions and estimates climate and weather effects jointly, but typically does not include spatially-disaggregated fixed effects leaving it susceptible to omitted variable bias (e.g., Kelly, Kolstad and Mitchell, 2005; Moore and Lobell, 2014; Merel and Gammans, 2021).2 Our unifying approach combines the strengths of the prior methods while addressing their shortcomings by relying on the properties of the Frisch-Waugh-Lovell theorem. Influential studies have proposed measuring adaptation as the difference between the estimates of impacts in fixed-effects and cross-sectional approaches (Dell, Jones and Olken, 2012, 2014). Estimates of climate impacts based on cross-sectional analysis are usually inclusive of adaptation, whereas those from fixed-effects are typically not. Our unifying approach estimates the short- and long-run impacts in the same equation. As a result, our approach enables a straightforward test for the statistical significance of the measure of adaptation and addresses two other shortcomings from existing approaches. First, it recovers a measure of adaptation directly from the jointly estimated impacts of weather and climate. In contrast, a common approach in the literature tackles adaptation indirectly, by flexibly estimating economic damages due to weather shocks, then assessing climate damages by using shifts in the future weather distribution predicted by climate models (e.g., Deschenes and Greenstone, 2011). Second, and analogous to the Lucas Critique (Lucas, 1976), our approach overcomes the challenges of identifying adaptation by comparing the profiles of weather 2The long differences approach is a special case of partitioning variation which leverages panel data variation in weather over a range of timescales (e.g., annual, decadal, and multi-decadal) to identify climate impacts, but does not estimate climate and weather effects jointly (e.g., Dell, Jones and Olken, 2012; Moore and Lobell, 2015; Burke and Emerick, 2016). 2 responses across time and space, under the assumption that preferences are constant across those dimensions (Barreca et al., 2016; Heutel, Miller and Molitor, 2021; Carleton et al., 2022).3 Instead, we identify adaptation by comparing how economic agents in the same season and location respond to weather shocks – which, by definition, limit opportunities to adapt – with their own response to climatic changes, which should incorporate adaptive behavior. We apply our unifying approach to the context of daily temperature and ambient ozone concentration across the continental United States. In our analysis, we merge location-byday ozone concentration data with temperature data across the United States for the period 1980-2013. In a typical climate impact setting, the outcome of interest is (i) affected by temperature, (ii) something of value to the agent, and (iii) responsive to adaptive behavior that dampens the temperature effect. By definition, adaptation involves adjusting to or coping with climatic change with the goal of reducing vulnerability to its harmful effects.4 In our setting, for agents to be adapting to rising temperatures in a way that changes atmospheric ozone levels, one needs all of the following: (i) agents must be worried about ozone’s detrimental impacts, (ii) agents have some knowledge of the process of ozone formation such that they are aware not only of temperature’s role but also the impact of an agent’s emissions, and (iii) agents believe their actions can sufficiently alter ozone concentrations. There is evidence that on high ozone days, individuals may avoid outdoor exposure (e.g., Neidell, 2009a) and buy medicines to remediate exposure (e.g., Deschenes, Greenstone and Shapiro, 2017). Also, they may drive less and use public transit in smog alert days (e.g., Cutter and Neidell, 2009). Indeed, the alerts educate the public on the impact of temperature and the agents’ actions on ambient ozone levels. Hence, it not unreasonable to assume that our 3One way to address this issue is to use experimental or quasi-experimental variation in those attributes in order to causally capture the extent to which they offset weather effects. One example is Garg, McCord and Montfort (2020), who leverage quasi-experimental variation in eligibility to a cash transfer program in Mexico to identify how income may mitigate the temperature-homicide relationship. 4The IPCC defines adaptation as “the process of adjustment to actual or expected climate and its effects in order to moderate harm or take advantage of beneficial opportunities,” and further states that “[a]daptation plays a key role in reducing exposure and vulnerability to climate change. (...) In human systems, adaptation can be anticipatory or reactive, as well as incremental and/or transformational.” (IPCC, 2022). 3 research setting satisfies the three conditions for adaptation enumerated above. Our approach has two key features. The first is the decomposition of meteorological variables into “climate” and “weather.” The second is identifying responses to weather shocks and longer-term climatic changes in the same estimating equation. As noted, the difference between those short- and long-run responses is what the literature refers to as adaptation.5 Indeed, ozone, as with most climate-related outcomes of interest, responds to realized temperature – regardless of how that temperature may be decomposed into “weather” or “climate.” It is only agents, by virtue of being able to adjust to long-run climate, that may affect the ozone response to climatic changes. In the absence of any adaptive behavior, the ozone response to equivalent changes in weather or climate would be the same. For the first feature of our approach, the daily temperature variable is used to construct two variables. The first, T empC, is operationalized as a 30-year moving average of monthspecific average temperatures (e.g., take the average of June daily temperatures for each year and location and then apply a 30-year moving average). This is what we interpret as “climate.”6 The second temperature variable, T empW , is daily temperature with T empC subtracted, interpreted as “weather.” For the second feature of our approach, both variables enter in our estimating equation along with a set of location-by-season-by-year fixed effects, ϕis (e.g., Chicago-Spring 1990, Chicago-Summer 1990, etc.). Because we create the variable “weather” as a first step, the Frisch-Waugh-Lovell theorem guarantees we do not need to include granular time fixed effects to identify weather effects (Lovell, 1963, Theorem 4.1, p.1001).7 On the other hand, the inclusion of ϕis allows us to leverage two sources of climatic variation to identify climate 5Although we focus on adaptive behavior, we are agnostic about the true impacts. There may be adaptation or intensification effects (Dell, Jones and Olken, 2014). 6Climate normals are, by definition, 30-year averages of weather variables such as temperature (WMO, 2017). The monthly frequency for the moving averages in our empirical decomposition is without loss of generality. All we need is a time frame that economic agents can easily remember information from the past. Our robustness checks using daily moving averages provide nearly identical results. 7 Intuitively, by decomposing observed temperature into its moving average “climate” component and daily “weather” component as the difference from this average, the Frisch-Waugh-Lovell theorem shows that the “weather” variable is already de-meaned as if we had included location-by-month-by-year fixed-effects in the final estimating equation. 4 impacts. Conditional on location-by-season fixed effects, the first source of variation comes from adding the most recent year’s monthly weather information and dropping the oldest portion from the 30-year moving average. The underlying idea is similar to filtering different frequencies of temperature, as has been done in the time series literature (e.g., Baxter and King, 1999; Christiano and Fitzgerald, 2003). In other words, we identify the agents’ response to their new climate expectation. The second source of variation arises from demeaning T empC from a location-specific season-by-year fixed effect.8 Take, for example, days in April, May, and June in Chicago, all within the spring season of the same year. After demeaning from a spring fixed effect, the average April moving-average measure of climate will likely be a negative value and the average June climate a positive value. Our methodology contributes to the estimation of climate damage functions and the costs of climate change (e.g., Auffhammer, 2018b; Tol, 2018). Our unifying approach to uncover climate impacts and adaptation should be of interest to a broad set of applications due to its simplicity. Our novel application to the impact of climate change on ambient ozone adds an overlooked force behind determinants of ozone pollution (e.g., Salvo and Wang, 2017). This paper proceeds as follows: Section II provides an overview of the previous methodological approaches used to identify climate impacts and proposes our unifying approach and the resulting measure of adaptation. Section III provides a conceptual framework of an agent’s adaptation decision-making, describes our data, and presents our empirical strategy. Section IV reports our main findings, examines the robustness of our estimates, generalizes our approach to nonlinear settings, and explores heterogeneity in adaptive responses. Finally, Section V concludes. 8Note that we use “location” here in the general sense as the spatial unit of analysis. For example, in our empirical setting location is taken as an individual ozone monitor. 5 II. Prior Methods and Our Unifying Approach to Measuring Climate Change Impacts and Adaptation A. Prior Methods Prior literature on estimating climate impacts and adaptation has usually relied on two approaches. The first is the cross-sectional approach (e.g., Mendelsohn, Nordhaus and Shaw, 1994; Schlenker, Hanemann and Fisher, 2005), which exploits permanent, anticipated components behind meteorological conditions, leveraging climate variation across locations to estimate climate impacts inclusive of adaptation, but may suffer from omitted variable bias. The other is the panel fixed-effects approach (e.g., Deschenes and Greenstone, 2007; Schlenker and Roberts, 2009), which deals with that bias but identifies the effect of transitory, unanticipated weather shocks, most likely exclusive of adaptation, making the transition to estimated climate effects nontrivial.9 By using either the short- or long-run variation behind meteorological conditions to identifying climate impacts, those research designs trade off key assumptions. More recent literature (e.g., Dell, Jones and Olken, 2012, 2014) has proposed various hybrid approaches for combining these two strands of the literature, but face issues of their own (Kolstad and Moore, 2020). The cross-sectional (CS) approach estimates the following equation: yi = α + βCSxi + (µi + νi) = α + βCSxi + ei , (1.1) where yi is an outcome variable measured at location i, and is affected by the climatological variable of interest, xi – typically taken as temperature. µi represents the vector of all timeinvariant unobserved covariates that may be correlated to xi , while νi reflects the standard idiosyncratic error term. Thus, if µi is non-empty and cov(xi , µi) ̸= 0, βˆ CS suffers from omitted variable bias (OVB). The panel fixed-effects (FE) approach instead estimates the following equation: 9Only in certain conditions does weather variation exactly identify the effects of climate (e.g., Hsiang, 2016; Lemoine, 2020). 6 yit = α + βF Exit + µi + λt + νit, (1.2) where the outcome variable, yit, and climatic variable of interest, xit, are now additionally measured at some recurring time interval t. By averaging each variable in Equation (1.2) for each unit i over time, we obtain: y¯i = α + βF Ex¯i + µi + ¯νi , (1.3) where ¯yi ≡ 1/T PT t=1 yit, and the other variables are defined similarly.10 Subtracting Equation (1.3) from Equation (1.2), we highlight the source of variation in the identification of βF E: (yit − y¯i) = βF E(xit − x¯i) + λt + (νit − ν¯i). (1.4) Because (xit − x¯i) is the deviation of observed temperature from its local long-run value, βF E is clearly identified from temperature shocks. Thus, in this approach, although most OVB problems are resolved by the µi term, βˆ F E now identifies the impact of meteorological, rather than climatological, phenomena. Recently, focus has expanded from simply estimating climate impacts to estimating adaptation to climate change. Some authors have noted that βCS identifies climate impacts inclusive of any adaptation, while βF E, by its nature, identifies meteorological impacts which can be taken as an approximation of climate impacts exclusive of any adaptation (e.g., Dell, Jones and Olken, 2012, 2014). Thus, they propose measuring adaptation as the difference between βˆ F E and βˆ CS. Although this principle to recovering a measure of adaptation is accurate, the approach faces two empirical challenges. First, to the extent that OVB may impact βˆ CS in the cross-sectional model, this will translate directly into bias in the estimate of climate adaptation. Second, even if an unbiased estimate of βCS could be obtained, βˆ CS 10Note that via the inclusion of the intercept, the λt and µi fixed effects are both relative to the same baseline, α, and thus the λt term drops out when averaging over time by the restriction that P t λt = 0. 7 and βˆ F E arise from two different estimating equations. While OLS, equation by equation, allows us to easily test hypotheses about the coefficients within an equation, it does not provide a convenient way for testing hypotheses involving coefficients from different equations. Thus, in practice, one must resort to seemingly unrelated regression (SUR) models to explicitly test whether the measure of adaptation is statistically distinguishable from zero.11 Aside from SUR, it would be possible to statistically test the difference between coefficients recovered via the CS and FE models using re-sampling methods – i.e., block bootstrap or Bayesian bootstrap with random weights assigned at the block-level. However, while these methods may solve the hypothesis testing issue for inferring the significance of adaptation, they would not address the issue of potential bias in the underlying estimating equations, making it difficult to interpret the magnitude of adaptation. B. Our Unifying Approach Our unifying approach nests both of those strands of the climate-economy literature in the same estimating equation. It simultaneously identifies long-run climatological impacts and short-run effects of meteorological shocks, and thus allows for an explicitly testable measure of adaptation in the spirit of prior comparisons between short- and long-run effects (e.g., Dell, Jones and Olken, 2012, 2014). Specifically, we begin by posing the ideal estimating equation, although infeasible: yit = α + βW (xit − x¯i) + βCx¯i + µi + λt + νit. (1.5) If this infeasible equation were estimable, βW – the effect of weather shocks – would exactly identify βF E by the Frisch-Waugh-Lovell theorem. On the other hand, βC – the effect of 11As is well known, a SUR system is a generalization of a linear regression model that consists of several regression equations – each having its own dependent variable and potentially different sets of exogenous explanatory variables – that has cross-equation error correlation, that is, the error terms in the regression equations are correlated. Also recall that all equations in a SUR system are estimated jointly, but that such estimation usually requires feasible generalized least squares with a specific assumption on the form of the variance-covariance matrix regarding the structure of the correlation among the error terms. Hence, further structural assumptions are needed for statistical inference of the measure of adaptation. 8 changes in climate – would identify βCS minus OVB due to the inclusion of fixed effects. Unfortunately, βC cannot be identified because ¯xi is perfectly collinear with µi . Notice that emerging hybrid approaches have also relied on such “partitioning variation” (e.g., Kelly, Kolstad and Mitchell, 2005; Moore and Lobell, 2014; Merel and Gammans, 2021). They have attempted to address this collinearity issue by dropping the unit fixedeffect, µi , instead including a set of location controls, ci , in their estimating equation, taking the general form of yit = f(xit − x¯i) + g(¯xi) + ciγ + ϵit, where f(.) and g(.) can take flexible functional forms. While this approach can include spatially-aggregate and time fixed-effects, identification would still ultimately rely on cross-sectional variation within the spatiallyaggregate region, and thus may suffer from similar OVB concerns as the CS model. We therefore propose the following feasible approximation of the ideal Equation (1.5), which allows for the inclusion of unit fixed-effects by letting the measure of climate vary across time within the sample:12 yit = α + βW (xit − x¯ip¯) + βCx¯ip¯ + µi + λs + νit. (1.6) As time can be aggregated into multiple subset levels – day, month, season, year, decade, etc. – we first define a time period, p, as a weakly larger aggregation of t. Agents, however, may observe and react to the slow evolution of climate. Thus, we define ¯p to incorporate data from the same time period p in the past. Furthermore, agents may need time to adjust, so we additionally restrict ¯p to exclude contemporaneous data. We also replace λt with λs – where s is a one-level higher aggregation in time than p – in order to retain relevant variation in ¯xip¯. 13 Depending on the study context, µi and λs may be interacted to flexibly control for unit-level effects that may vary over time. Defined in this way, variation in ¯xip¯ comes from two separate sources. First, although 12Observe that for simplicity, and to keep the comparison with the prior CS and FE strands of the literature as clear as possible, our unifying approach uses a linear specification, which should also capture the firstorder effects of potentially nonlinear responses. Later, in Section IV.D, we show how this approach can be easily extended to include higher order nonlinear effects. 13Note that just as t, by convention, represents a specific time-step of the sample, e.g. day-of-the-sample, we take s as similarly representing a more aggregate time-step of the sample, e.g. season-of-the-sample. 9 more aggregate than t, ¯p still varies across time within the the higher level time period s. Second, ¯p is defined to include historical data, and thus “updates” its value from year to year. Following the same steps as with the fixed-effects model and averaging each variable in Equation (1.6) for each cross-sectional unit i over time, we obtain: y¯i = α + βW (¯xi − x¯i) + βCx¯i + µi + ¯νi = α + βCx¯i + µi + ¯νi , (1.7) where, once again, ¯yi ≡ 1/T PT t=1 yit, and the other variables are defined similarly.14 Subtracting Equation (1.7) from Equation (1.6), we highlight the source of variation that allows for the identification of both βW and βC: (yit − y¯i) = βW (xit − x¯ip¯) + βC(¯xip¯ − x¯i) + λs + (νit − ν¯i). (1.8) In Equation (1.8) we can observe that βˆW is identified from temperature shocks, therefore approximately equivalent to βˆ F E, whereas βˆ C is identified from climatic changes, approximately equivalent to βˆ CS, though now critically free from a number of OVB concerns. We thus naturally define adaptation as the difference βˆW − βˆ C. Because both coefficients of interest are estimated in a single equation, statistical inference on the measure of adaptation is straightforward. Furthermore, observe that while our method does require the researcher to take a stance on the temporal granularity of the climate variable, ¯xip¯, and time fixedeffects, λs, the recovered measure of adaptation leverages the behavioral responses of the same economic agents to both weather shocks and climatic changes via the inclusion of unit fixed effects, µi . C. Decomposition of Meteorological Variables: Climate Norms vs. Weather Shocks As mentioned above and seen in Equation (1.6), implementing our approach requires that we first decompose xit into its long-run component, ¯xip¯, and its short-run deviation from 14Note that in Equation (1.7) the ¯xi derived from the ¯xip¯ term would rely on a longer time-series of information than the ¯xi derived from the xit term. Still, they are approximately equivalent, with correlation between these two terms above 0.95 in our empirical application. 10 this value, (xit − x¯ip¯). Econometrically, from the Frisch-Waugh-Lovell theorem, we can decompose xit into its longer term seasonal component and a contemporaneous de-seasonalized component. For example, as weather varies day-to-day, t, and local climate varies both seasonally (e.g., month-to-month within a year) and over time (e.g., year-to-year), we could take “month-of-the-sample,” my, as representing the seasonal component and pose the following first-stage regression: xit = γimy + ϵit, (1.9) such that temperature in location i on day t (of month m in year y) is regressed on a set of location-by-month-by-year fixed effects. In this case, the matrix of coefficients ˆγimy would constitute the matrix of monthly average temperature values ¯ximy, while the estimated residuals (xit − x¯imy) (≡ ϵˆit) would reflect the de-seasonalized daily local deviations of temperature. Because this regression simply de-means xit over the my period in the time-series dimension for each individual location i, we could instead recover the xit − x¯imy values in Equation (1.9) arithmetically via the following: T emp | {z } xit = T empC | {z } x¯imy + T empW | {z } (xit−x¯imy) , (1.10) such that T empC (≡ x¯imy) represents climate patterns, and T empW (≡ xit−x¯imy) deviations from those longer-run patterns. Notice that although the above example uses daily temperatures, de-seasonlized at the monthly level, the choice of timing can be selected to match the study context. To use the example of agriculture, a common focus in the climate literature, it may be that a year, or the growing seasons within a year, would be better suited to the analysis than the months of the year example illustrated in equations (1.9) and (1.10). Economically, however, this presents a potential problem. As mentioned in the previous section, agents may need time to adapt, and prior information sets likely inform agents’ beliefs. Thus, ¯ximy is not strictly equivalent to ¯xip¯ as defined in Equation (1.6). To address 11 this, we propose, as a first step, replacing ¯ximy with a lagged function of its historical values: x¯ip¯ ≡ 1 J X J<y j=1 ωjx¯imj ≈ x¯imy, (1.11) where ωj represents a scalar weighting of ¯ximj , such that the function defining ¯xip¯ can be generalized to fit various contexts.15 Returning to the agriculture example, it’s possible that farmers need more than a single year to adjust production processes or change crop choice, in which case the (ωy−k, ..., ωy−1) weighting scalars of Equation (1.11) could all simply be set to zero, with k > 1. Furthermore, the functional form of Equation (1.11) itself can be chosen to best suit the application by changing the specific values of ωj . Myopic and Bounded agents may simply assume that contemporaneous monthly temperature will be equal to what it was in the previous year, that is, ωj simply evaluates to zero for all j ∈ {1, ..., y − 2}. Other agents may flexibly fit values of ωj to the historical data in an attempt to predict x¯ip¯ through statistical means. A similar idea has been used in macroeconomics to measure business cycles,16 and in the literature of intergenerational mobility following Solon’s (1992) seminal work.17 Note that ¯xip¯ can be calculated from a longer time-series of x to take into account historical information beyond the sample period of the outcome variable. We then return to Equation (1.10), substituting ¯xip¯ for ¯ximy in representing T empC, and recovering xit − x¯ip¯ (≈ xit − x¯imy) for T empW , giving us all the components necessary for estimating Equation (1.6).18 Notice that by the properties of the Frisch-Waugh-Lovell theorem (specifically, point 4 of Lovell (1963, Theorem 4.1, p.1001)) it is unnecessary to 15These weights, ωj , can be defined by values derived from other literatures, such as climatology, which defines a climate normal as the average temperature over the last 30 years: “The 30 year interval was selected by international agreement, based on the recommendations of the International Meteorological Conference in Warsaw in 1933. The 30 year interval is sufficiently long to filter out many of the short-term interannual fluctuations and anomalies, but sufficiently short so as to be used to reflect longer term climatic trends” (Climatology Office, 2003). Alternative filtering techniques could also be implemented (e.g., Baxter and King, 1999; Christiano and Fitzgerald, 2003), and would implicitly follow from this expression by varying the values of ωj . 16See, for example, Baxter and King (1999), Christiano and Fitzgerald (2003) and Hsiang (2016). 17In Solon’s context, observed income is noisy: it includes a permanent and a transitory component. To establish a relationship between permanent income of sons and fathers, Solon proposes averaging fathers’ income for a number of years to reduce the errors-in-variables bias. 18In our preferred decomposition detailed in the following section, Cor(¯xip¯, x¯imy) > 0.95 and Cor((xit − x¯ip¯),(xit − x¯imy)) > 0.90. 12 de-seasonalize the outcome variable yit in the same way as (xit − x¯ip¯), which allows us to estimate both effects of interest in the same equation.19 This decomposition highlights the two sources of variation that have been used in the climate-economy literature. T empC and T empW in the decomposition above are associated with different sets of information. On the one hand, T empC includes climate patterns that economic agents can only gather by experiencing weather realizations over a long period of time, and can be thought of as the “climate normal” temperature. On the other hand, T empW represents weather shocks, which by definition are revealed to economic agents virtually at the time of the weather realization. Usually one adjusts to something they happen to know by experience. Therefore, adaptation can be measured as the difference between responses to changes in T empC relative to effects of weather shocks T empW . This is analogous to Lucas’ powerful insight that economic agents respond differently depending on the set of information that is available to them. Lucas (1977), for instance, provides an example of a producer that makes no changes in production or works less hard when facing a permanent increase in the output price, but works harder when the price increase is transitory. 20 It is also important to emphasize that this decomposition does not make any assumption on how individuals and firms process and use the information from the past. Rational agents would respond optimally to all information at hand when deciding the degree of adaptation, while myopic and inattentive agents (e.g., Gabaix and Laibson, 2006; Reis, 2006), on the other hand, may find it costly to absorb and process all the information at all times, and may respond only to partial information or only sporadically. Our measure of adaptation is agnostic to either type of behavior; the goal of our approach is to empirically assess the 19“Theorem 4.1: Consider the following alternative regression equations, where the subscript α indicates that the data have been adjusted by the least squares procedure with D as the matrix of explanatory variables: 1. Y = Xb1 +Dα1 +e1 2. Yα = Xαb2 +e2 3. Y = Xb3 +e3 4. Y = Xαb4 +e4 ... The identity b2 = b4 reveals that it is immaterial whether the dependent variable is adjusted or not, provided the explanatory variables have been seasonally corrected” (Lovell, 1963). 20Notably, in our context the behavior would be reversed. Due to the contemporaneous nature of transitory weather shocks, little to no change in production is possible, while the producer would be able to change behavior in response to permanent changes in climate. 13 economic and statistical significance of adaptation, regardless of how economic agents make decisions on whether to adapt, or the extent of adaptation. Finally, notice that this decomposition represents a first-order Taylor approximation of a potentially nonlinear relationship between climate and realized temperature. Two types of variation are often associated with a changing climate: changes in averages, and changes in the frequency of extreme weather events (IPCC, 2022). For simplicity, and to keep the comparison with prior approaches as clear as possible, our temperature decomposition focuses on increases in averages, not on variability. In fact, in the following section we show that our weather data, comprised of the comprehensive set of national weather monitors, suggests a gradual increase in average temperature, but that the magnitude of temperature shocks, defined as deviations from the 30-year moving averages, are relatively stable over time, and narrowly bounded. Therefore, in our approach, dispersion shows up only implicitly in the sense that long-run norms take into account the frequency and intensity of daily temperature extremes. III. Empirical Application: Climate Impacts on Ambient Ozone We apply our unifying approach to measure climate impacts on ambient ozone concentration, and adaptation to climate change in this context, and examine the heterogeneity in adaptive behavior. This application is ideal for three reasons. First, ozone is not emitted directly into the air, but rather rapidly formed by Leontief-like chemical reactions between nitrogen oxides (NOx) and volatile organic compounds (VOCs) in the presence of sunlight and warm temperatures.21 Hence, meteorological conditions do matter in determining surface ozone levels, and climate change may increase ozone concentration in the near future (e.g., Jacob and Winner, 2009). Furthermore, ozone is rapidly destroyed during the night; thus, correlation between ambient concentrations across two consecutive days is limited. Second, nationwide high-frequency data on ambient ozone and meteorological conditions are publicly available 21See Appendix A.1.a for further details. 14 for a long period of time in the United States: we use daily measurements for the typical ozone season from 1980-2013.22 Third, this is a highly policy-relevant issue. The so-called “climate penalty” on ozone means that climate change might deteriorate air quality in the near future, with important implications for public health and labor productivity.23 A. Conceptual Framework In the context of ozone, economic agents could be polluting firms, households engaging in consumption that produces precursor pollutants, or local regulators concerned with pollution and public health. For example, households may respond to an ozone alert day by mowing their lawns or refueling their cars earlier or later in the day – or on a different day altogether – to avoid VOC emissions, taking public transit, carpooling, or working from home to reduce emissions altogether, or purchasing hybrid or electric vehicles to reduce local emissions. On the other hand, firms may (i) reshuffle their production activities within the day to avoid VOC emissions in peak hours, such as painting in construction sites, or even between different months, increasing emissions during colder months in order to reduce emissions during hotter months; (ii) install pollution abatement technologies, or otherwise change their production function, for instance by electrifying emissions-intensive production processes such as switching from oil or gas furnaces to electric. Additionally, local regulators may provide ground-level ozone information to at-risk populations to avoid intense ozone exposure on hot days, e.g., by issuing an ozone alert when a heat wave is forecasted, and coordinating local actions with households and firms to reduce permanently or shift emissions-intensive activities within the day or across days, weeks, or months. Importantly, these agents could be reacting to either the realized or anticipated outcome of climate change, and could be undertaking small or large actions – adjusting behaviors within a day might be a small action 22The ozone season varies by state and usually consists of only six months (typically April-September), but concerns are mounting that longer spring and fall would expand the ozone season in some states (e.g., Zhang and Wang, 2016). 23Exposure to ambient ozone has been causally linked to asthma hospitalization, pharmaceutical expenditures, mortality, and labor productivity (e.g., Neidell, 2009a; Moretti and Neidell, 2011; Graff Zivin and Neidell, 2012; Deschenes, Greenstone and Shapiro, 2017). 15 that adds up across many agents, for example, while the switch to alternative commuting or production methods may be more transformational.24 For simplicity of exposition, consider the case of a polluting firm. The agent minimizes cost by selecting the optimal production schedule for the given input costs, climate, and other local factors faced by the agent. But, ambient ozone itself can impose an additional shadow price on the agent’s chosen production schedule, implied by, e.g., public or regulatory pressures. Specifically, for the agent engaging in dirty production, the emission of ozone precursor pollutants (VOCs and NOx) are de facto “inputs” into the agent’s production schedule.25 Any shadow price on ozone faced by the agent would thus translate into an implicit shadow price on the emission of either of these precursors as inputs in their production process, conditional on local climate and atmospheric composition.26 Ceteris paribus, the agent would thus minimize costs taking into account the implicit shadow prices on these precursors.27 In practice, the optimizing decisions are often over changes in input mix or timing of production (Henderson, 1996). In other words, the agent is implicitly considering ozone levels whenever they choose the cost-minimizing inputs for production of goods and services.28 To better understand why agents may adapt to climatic changes in ways that reduce 24Observe that some local regulators are making a direct case for reducing precursor pollutants to control climate change driven increases in ozone (e.g., BAAQMD, 2017), and that the EPA also acknowledges the role of climate change in worsening ozone concentrations, stating that “[i]n addition to being affected by changing emissions, future O3 concentrations may also be affected by climate change” (USEPA, 2015a). 25That is, they are emitted in proportion to the choice, and quantity used, of actual production inputs. 26Naturally, there may also be regulatory pressures for the precursors themselves, therefore explicitly defining (shadow) prices for them as well (Auffhammer and Kellogg, 2011; Deschenes, Greenstone and Shapiro, 2017). In the robustness checks, however, we provide evidence that these regulations do not seem to play an important role in agents’ adaptation measures regarding climatic changes. This is not surprising, given that it is ozone formation, not the precursors, that primarily depends on climate. 27Unlike in the agriculture setting, a common focus of prior studies, where markets exist for most inputs, in our context markets for ozone precursors (de facto inputs in production) existed only in some areas and in specific periods of time. Notwithstanding, the implicit shadow prices – reflecting social valuation of ambient ozone reductions – may provide incentives for producers similar to those provided by market prices. 28Of course there are other factors that may affect ambient ozone concentrations, climate being the obvious one, but precursor emissions are the only source that is controllable by the agent. While this could lead to measurement error in the direct relationship between agents’ decisions and ozone concentration, ozone – in this context – is the outcome variable, and any measurement error in ozone would simply be absorbed by the error term in a reduced form model. 16 ambient ozone, compare the ozone context to a standard agricultural setting. As has been shown in that context (e.g., Mendelsohn, Nordhaus and Shaw, 1994; Schlenker, Hanemann and Fisher, 2005; Deschenes and Greenstone, 2007; Schlenker and Roberts, 2009), the agent maximizes profit by optimizing over their choice of crop and other inputs such as irrigation, conditional on anticipated or realized climate, controlling for other local factors such as soil quality. Restated, the agent minimizes cost by selecting the optimal production schedule for the given set of input costs, climate, and other local factors faced by the agent. Figure 1.1 illustrates this “cost-minimizing” optimization decision agents face with respect to ozone and its precursors, depicting the envelope of minimum-cost production schedules, conditional on realized climate, in the spirit of Deschenes and Greenstone (2007). Cost of production is on the left y-axis, associated ozone concentration is on the right y-axis, and temperature is on the x-axis.29 For simplicity in illustration, we assume that factors such as precipitation and other exogenous determinants have been adjusted for. The production schedule 1 and 2 cost functions reveal the relationship between cost and temperature, as well as ozone and temperature, when these production schedules are chosen. It is evident that schedule specific costs, and associated ozone concentrations, vary with temperature. Further, the cost-minimizing production schedule varies with temperature. For example, production schedule 1 minimizes cost between T1 and T2; the agent would be indifferent between the two at T2 where the cost functions cross (i.e., point B); and production schedule 2 minimizes cost between T2 and T3. The long-run equilibrium is denoted by the dashed gray line and represents the long-run optimum when the agent can freely adjust their production schedule in response to changes in temperature. Consider first an agent that is initially faced with a climate normal temperature of T1. Their optimal choice would thus be to minimize cost under production schedule 1, at point A. Now consider two alternative scenarios: one in which the agent is faced with a transitory 29Notice that from the cost minimization problem, we observe a derived demand function for VOCs and NOx, conditional on the agent’s chosen level of output. In turn, that demand for precursors maps into resultant ambient ozone levels, conditional on the temperature. 17 Figure 1.1: Theoretical Relationship Between Marginal Cost of Dirty Production and Temperature Production Schedule 1 Cost Function Production Schedule 2 Cost Function Long-run Equilibrium (Inclusive of Adaptation) A B C C' Ambient Ozone Concentration Cost of Production T1 T2 T3 Temperature Notes: This figure illustrates a stylized example of how changes in temperature could affect the cost of production through the shadow price on ozone, and thus the implicit shadow prices on VOCs or NOx that are emitted under the chosen production schedule. The profit-maximizing firm minimizes cost – the amounts inputs used in production multiplied by their respective prices, as well as the quantity of VOCs and NOx produced under the chosen production schedule multiplied by the shadow prices of these ozone precursor pollutants implied by the local shadow price on ozone and conditions of the local atmosphere. While in many cases firms may not face an observable market price for their emissions of VOCs or NOx, they may face a shadow price for doing so based on, for example, public or regulatory pressures. As depicted, at a temperature of T1, production schedule one dominates schedule two, and the firm minimizes cost at point A, with associated daily maximum ozone concentration. At a temperature of T2 the firm is indifferent between either production schedule one or two at point B. At a temperature of T3, however, production schedule two now dominates schedule one, and the firm minimizes cost at point C. A firm may not, however, be capable of adjusting their production schedule on a day-to-day basis. Thus, a firm facing a climate normal temperature of T1 may opt to produce at point A, but end up producing at point C ′ , and a much higher ozone concentration, when faced with a temperature shock of T3. A firm that experiences many such shocks would thus update their beliefs about the underlying climate norm and shift their production schedule towards schedule two. temperature shock of T3, and a second in which the agent is faced with a permanent change to a new climate normal temperature of T3. Under the first scenario, the agent would be 18 unable, or unwilling,30 to adapt to the temperature shock and would temporarily produce at point C ′ , with higher associated ozone concentration and higher cost of production. Under the second scenario, the agent would adjust to this permanent change in the climate normal temperature and change to production schedule 2, now producing at point C rather than C ′ . Notice, however, that while point C is lower cost than point C ′ , it still implies a higher cost of production and associated ozone concentration than point A. This is to be expected. Adaptation is typically not costless (e.g., Kelly, Kolstad and Mitchell, 2005; Carleton et al., 2022) – as production schedule 1 was cost-minimizing under the original climate norm of T1, this implies that schedule 2 must be (weakly) more costly to implement in the absence of any climatic changes. Finally, notice that our unifying approach estimates simultaneously both of these reduced form relationships between ambient ozone concentration and temperature, accounting for agents’ differential responses to temperature shocks versus changes in the climate norm. The recovered estimate for temperature shocks – βW in Equation (1.6) – reflects the difference between the ozone concentrations associated with points C ′ and A, while the recovered estimate for changes in the climate norm – βC in Equation (1.6) – reflects the difference between points C and A, and thus adaptation can be clearly taken as the difference between C ′ and C. B. Data Weather Data — For meteorological data, we use daily measurements of maximum temperature as well as total precipitation from the National Oceanic and Atmospheric Administration’s Global Historical Climatology Network database (NOAA, 2014). This data-set provides detailed weather measurements at over 20,000 weather stations across the country for the period 1950-2013. Figure A.1.1, in Appendix A.1, presents the yearly temperature 30From a purely mechanical standpoint, the agent may be technologically unable to adjust their production schedule on such short notice – i.e., daily. From an economic standpoint, even if such adjustments were technologically feasible, they may not be economically sound, as such adjustments would likely incur greater costs than could be saved by avoiding the additional cost associated with transitory sub-optimal production. 19 fluctuations and overall climate trend in the US as measured by these weather stations, relative to a 1950-1979 baseline average temperature, while Figure A.1.2 illustrates the geographical location of the complete sample of weather stations from 1950-2013. Figure 1.2, by comparison, depicts the variation and trend of our decomposed temperature variables, T empC and T empW , between 1980 and 2013 for the comprehensive set of national weather stations, indicating that while average temperature has been gradually increasing, temperature variability has remained relatively stable.31 These weather stations are typically not located adjacent to the ozone monitors. Hence, we develop an algorithm to obtain a weather observation at each ozone monitor in our sample.32 Our preferred matching algorithm uses information from the two closest weather stations within 30 km of each ozone monitor, as these stations are likely to better reflect the local environment than stations that are further away. The final sample under this matching algorithm includes 97.25% of all daily ozone observations (97.91% of all ozone monitors). However, we also expand the matching algorithm to include the closest five weather stations within 80 km, for a final sample that includes over 99.99% of all daily ozone observations (100% of all ozone monitors). Table A.1.1, in Appendix A.1, reports the summary statistics for daily temperature and our decomposed variables, for each year in our sample from 1980-2013. Ozone Data — For ground-level ozone concentrations, we use daily readings from the nationwide network of the EPA’s air quality monitoring stations. In our preferred specification we use an unbalanced panel of ozone monitors.33 Appendix A.1 Figure A.1.5 illustrates the evolution of ambient ozone concentrations over our sample period for both the full unbalanced panel of monitors, as well as a smaller balanced panel. Figure A.1.6 depicts the evolution of our sample of ozone monitors over the three decades in our data, and illustrates 31Figures A.1.3 and A.1.4 in Appendix A.1 present similar patterns using a semi-balanced sample of weather stations, and our final sample of weather stations once matched to ozone monitors. 32We detail the steps taken in Appendix A.1.b as well as conduct robustness checks on the sensitivity of our results to changes in the algorithm in Appendix A.2.a. 33We discuss the reasoning for this approach as well as our results using a balanced panel in Appendix A.2.a. 20 Figure 1.2: Climate Norms and Shocks 25.95 26.00 26.05 26.10 26.15 26.20 30-year Moving Average (Temperature C°) 1980 1990 2000 2010 Panel A. Average Climate Norm Over Time -1 -0.5 0.0 0.5 1.0 1.5 Deviation from Moving Average (Temperature C°) 1980 1990 2000 2010 Panel B. Average Temperature Shock Over Time Notes: This figure depicts US temperature over the years in our sample (1980-2013), decomposed into their climate norm and temperature shock components. The climate norm (Panel A) and temperature shocks (Panel B) are constructed from a complete, unbalanced panel of weather stations across the US from 1950 to 2013, restricting the months over which measurements were gathered to specifically match the ozone season of April–September (see Appendix A.1 Table A.1.3 for a complete list of ozone seasons by state). The solid line in Panel A smooths out the annual averages of the 30-year moving averages, and the horizontal dashed lines in Panel B highlights that temperature shocks are bounded in our period of analysis. Appendix A.1 Figure A.1.3 depicts these same norms and shocks when restricting the dataset to include only a semi-balanced panel of weather stations, while Appendix A.1 Figure A.1.4 depicts these when the dataset is restricted to only those weather stations that are matched to an ambient ozone monitor for our main estimation sample. 21 the expansion of the network over time. Table A.1.2 describes some features of the sample of ozone monitors used in our analysis, for every year between 1980 and 2013. Consolidating information from the above sources, we reach our final unbalanced sample of ozone monitors over the period 1980-2013.34 Figure A.1.7 illustrates the proximity of our final sample of ozone monitors to the matched weather stations. We carry out the analysis focusing on the effect of daily maximum temperature on daily maximum ozone concentration since 1980. We choose this relationship because increases in temperature are expected to be the principal factor driving increases in ambient ozone concentrations (Jacob and Winner, 2009). Indeed, data on ozone and temperature from our sample, plotted in Appendix A.1 Figure A.1.8, highlights the close correlation between these two variables. Interestingly, we see that not only does contemporaneous temperature have an effect on ambient ozone, but the long-term climate normal temperature also seems to be affecting it, although perhaps to a lesser extent. We leverage both relationships in the empirical framework we now describe. C. Empirical Strategy Decomposition of Meteorological Variables: An Empirical Counterpart — Focusing on temperature (T emp), our primary variable of interest, we express it around ozone monitor i in day t of month m and year y, and decompose it into T empC (≡ x¯ip¯) and T empW (≡ xit−x¯ip¯) as in Section II. For our application, we define: x¯ip¯ = 1 30 X y−1 j=y−30 x¯imj , (1.12) Implicitly defining ωj as equal one for all j ∈ {y − 30, ..., y − 1} – where y denotes the contemporaneous year – and zero otherwise, such that T empC (≡ x¯ip¯) is equal to the 30- year monthly moving average (MA) of past temperatures.35 34For further details regarding the construction of the final dataset for our analysis, see Appendix A.1.b. 35Our decomposition of meteorological variables into a 30-year moving average (norms) and deviations 22 We choose a one-year lag to make this variable part of the information set held by economic agents at the time that the outcome of interest is measured. At the same time, we average temperature over 30 years because it is how climatologists usually define climate normals, and because we wanted individuals and firms to be able to observe climate patterns for a long period of time, enough to potentially make adjustments.36 For example, the 30- year MA associated with May 1982 is the average of May temperatures for all years in the period 1952-1981. Therefore, economic agents should have had at least one year to respond to unexpected changes in climate normals at the time ambient ozone is measured. We use monthly MAs, rather than daily or seasonal, because it is likely that individuals recall climate patterns by month, not by day of the year. Indeed, meteorologists on TV and social media often talk about how a month has been the coldest or warmest in the past 10, 20, or 30 years, but not how a particular day of the year has deviated from the norm for that specific day.37 Taking this approach, T empW represents weather shocks and is defined as the deviation of the daily temperature from the lagged 30-year monthly MA. By definition, these shocks are revealed to economic agents only at the time ambient ozone is being measured. Thus, in this case agents may have had only a few hours to adjust, limiting their ability to respond to unexpected temperatures.38 Figure 1.3 provides from it (shocks), as discussed in Section II, is a data filtering technique to separate the “signal” from the “noise.” This should not be confused with (a special case of) an autoregressive integrated moving average (ARIMA) model of climate change. 36It is possible, however, that agents form beliefs regarding expected climate over much shorter and more recent time windows (e.g., Kaufmann et al., 2017), or that organizational inertia slows the rate at which firms adapt to a changing climate. In our robustness checks we provide similar estimates using 3-, 5-, 10-, and 20-year moving averages, as well as longer lag lengths between the contemporaneous weather shock and the defined climate normal. 37There may be a concern that because temperature can have a within-month trend, defining temperature as a monthly average (climate norm) with daily (weather) shocks could mechanically lead to a stronger relationship between ozone and weather than between ozone and climate. As another robustness check, we redefine ¯xip¯ in Equation (1.12) to the special case in which p = t, using daily instead of monthly moving averages, discussed further in the following subsection. Economic agents, however, may still associate a day with its corresponding month when making adjustment decisions. 38Because precise weather forecasts are made available only a few hours before its realization, economic agents may have limited time to adjust prior to the ozone measurement. This might be true even during Ozone Action Days (OAD). An OAD is declared when weather conditions are likely to combine with pollution emissions to form high levels of ozone near the ground that may cause harmful health effects. Individuals and firms are urged to take action to reduce emissions of ozone-causing pollutants, but usually only a day in advance or in the same day. Unlike what happens in a few developing countries, however, neither production 23 an illustrative example of our preferred decomposition in Panel A, compared to a traditional fixed-effects decomposition in Panel B, using data for Los Angeles in 2013.39 Figure 1.3: Decomposition of Temperature Norms & Shocks (Los Angeles, 2013) Panel A. Our Preferred Decomposition 0 20 40 Temperature (C°) Apr Jun Aug Oct Temp Shock Climate Norm Daily Temp Panel B. Fixed-Effect Decomposition 0 20 40 Temperature (C°) Apr Jun Aug Oct Deviation from Avg Average Temp Daily Temp Notes: This figure compares our preferred decomposition method with a standard fixed-effects approach using data from the 2013 Los Angeles ozone season. Panel A depicts the daily measure of temperature, decomposed into climate norm and temperature shock. By contrast, Panel B depicts the same daily measure of temperature, but decomposed into a typical fixed-effect average temperature and the deviations, after controlling for monthly fixed effects. The dashed lines indicate observed daily maximum temperature while the black solid lines represent long-run norms. The gray solid lines indicate temperature shocks which are nearly identical in both panels, as would be expected from the Frisch-Waugh-Lovell theorem, illustrating the variation used for identifying βW and βF E. Panel A additionally highlights the variation in climate used to identify βC in our proposed approach, while Panel B lacks any such variation in the measure of climate. nor driving is forced to stop in those days, limiting the impact of short-run adjustments. In the robustness checks, we find no evidence of any additional adaptation occurring due to OAD announcements. That is, short-run adjustments, if any, do not seem large enough to be comparable to what happens in the long run. 39Figure A.1.9, in Appendix A.1, illustrates this same concept but over the entire 34-year sample period. 24 Econometric Model — Given the decomposition of meteorological variables into two sources of variation, our parsimonious econometric specification to estimate the impact of temperature on ambient ozone is the following: Ozoneit = βW T empW it + βCT empC ip¯ + X ′ itδ + ϕis + ϵit, (1.13) where i represents an ozone monitor, t stands for day, and s for season-of-the-sample (Spring or Summer, in each year). As mentioned in the prior section, our analysis focuses on the most common ozone season in the U.S. – April to September – in the period 1980-2013.40 The dependent variable Ozone captures daily maximum ambient ozone concentration. T emp’s represent the two components of the decomposition proposed for meteorological variables.41 The matrix of additional control covariates Xit contains a similar decomposition of precipitation.42 Finally, we replace the monitor fixed effects, µi , and time fixed effects, λs, from the generalized model presented in Equation (1.6) with ϕis – fixed effects for monitor-by-seasonby-year, and include ϵit, an idiosyncratic term.43 From a theoretical standpoint this change is not necessary – and in fact the empirical results are qualitatively similar in our context when implemented using µi and λs as separate fixed effects. We nevertheless combine them to more flexibly control for local factors that may have changed across seasons and years, allowing us to more closely approximate the ideal experiment.44 Analogous to Isen, Rossin-Slater and Walker (2017), notice that by including fixed effects for monitor-by-season-by-year, it is as if we regressed our main specification monitor by 40Table A.1.3 in Appendix A.1 lists the official ozone season by state. 41We further explore the nonlinear effects of temperature on ozone in Section IV.D, providing two alternative approaches for extending the linear model to allow for nonlinearities in the response function of ozone to weather shocks and climate norms. 42Although Dawson, Adams and Pandisa (2007) find it to be less important than temperature, Jacob and Winner (2009) point out that higher water vapor in the future climate may decrease ground-level ozone concentration. Our estimates are in line with those authors’ assessment, and are available upon request. 43Appendix A.3 details how both sources of monitor-level variation in ¯xip¯, within-season and across-year, are still leveraged within this monitor-by-season-by-year fixed-effects structure. 44One may be concerned that we do not include fixed effects for “predictable” within-season variation such as the “ozone weekend effect.” As a robustness check we re-estimated Equation (1.13) after further extending our monitor-by-season-by-year fixed effects, ϕis, to monitor-by-season-by-year-by-weekday/end. Our results were quantitatively unchanged to the third decimal digit. 25 monitor, individually, for each season of the sample, and then took the weighted average of all recovered coefficients. Conceptually, consider the following thought experiment that we observe in our data many thousands of times for both daily temperature shocks and monthly climate norms: Take two days (months) in the same location, same season, and same year. Now, suppose that one of the days (months) experiences a larger temperature shock (hotter climate norm) than the other. Our estimation strategy quantifies the extent to which this difference in temperature shock (climate norm) affected the ozone concentration observed on that day (month). Therefore, this approach controls for a number of potential time-invariant and time-varying confounding factors that one may be concerned with, such as the composition of the local atmosphere, regulatory burden, and technological progress. Measuring Adaptation — Once we credibly estimate the impact of the two components of temperature – daily shocks and within-season changes in climate normals – on ambient ozone concentration, we uncover our measure of adaptation. The average adaptation across all monitored locations in our sample is the difference between the coefficients βˆW and βˆ C estimated in Equation (1.13). If economic agents engaged in full adaptive behavior, βˆ C would be zero, and the magnitude of the average adaptation would be equal to the size of the weather shock effect on ambient ozone concentration.45 As previously discussed, agents would react to “permanent” increases in temperature by reducing ozone precursor emissions to offset potential increases in ozone concentration. In our preferred econometric specification, behavioral responses are allowed to occur only in the year after the change in temperature norm is observed. Those adjustments, however, might be related to innovations in temperature happening both in the previous year and 30 years before. Indeed, the “moving” feature of the 30-year MA is, by definition, associated with the removal of the earliest observation included in the average – 31 years before, and the inclusion of the most recent observation – one year before. Nevertheless, in the robustness 45This outcome is unlikely because, as noted previously, adaptation is typically not costless and thus the costs of engaging in ‘full adaptive behavior’ likely outweigh the benefits (Kelly, Kolstad and Mitchell, 2005; Carleton et al., 2022). 26 checks we consider cases where economic agents can take a decade or two to adjust. IV. Results A. Impacts of Temperature on Ambient Ozone Concentration Column (1) of Table 1.1 presents the effects on ambient ozone of the two components of observed temperature: climate norm, represented by the lagged 30-year monthly MA, and temperature shock, represented by the deviation from that long-run norm.46 Although the effects are uncovered by estimating Equation (1.13), columns (2) and (3), respectively, benchmark them against effects that would have been found if one had exploited either only the panel (e.g., Deschenes and Greenstone, 2007; Schlenker and Roberts, 2009) or only the crosssectional (e.g., Mendelsohn, Nordhaus and Shaw, 1994; Schlenker, Hanemann and Fisher, 2005) structure of the data. Column (2) reports the effect of temperature on ozone identified by exploiting withinmonitor daily variation in maximum temperature after controlling for monitor-by-month-byyear fixed effects. The coefficient indicates that a 1◦C increase in maximum temperature leads to a 1.66 parts per billion (ppb) increase in maximum ambient ozone concentration. Column (3) reports results from a cross-sectional estimation of daily maximum ozone concentration on daily maximum temperature around each monitor, averaged over the entire period of analysis 1980-2013. These variables capture information for all the years in our sample and are good proxies for the average pollution and climate around each monitor. The estimate suggests that a 1◦C increase in average maximum temperature is associated with an increase of 1.17 ppb in ozone concentration, approximately. When we decompose daily maximum temperature into our two components in column (1), as expected the estimated effect of temperature shocks on ambient ozone is statistically the same as the fixed-effects approach 46As mentioned before, even though we use monthly moving averages in our main analysis, as a robustness check we also estimate our preferred specifications using daily moving averages. The results are virtually identical, and are reported in Appendix A.2.a Table A.2.4. 27 Table 1.1: Climate Impacts and Adaptation – Our Unifying Approach vs. Prior Approaches Daily Max Ozone Levels (ppb) Unifying Fixed-Effects Cross-Section (1) (2) (3) Temperature Shock 1.678*** (0.063) Climate Norm 1.164*** (0.051) Max Temperature 1.659*** (0.063) Average Max Temperature 1.166*** (0.106) Implied Adaptation 0.514*** 0.493** (0.041) (0.225) Fixed Effects: Monitor-by-Season-by-Year Yes Monitor-by-Month-by-Year Yes State Yes Precipitation Controls Yes Yes Yes Latitude & Longitude Yes Non-Attainment Control Yes Observations 5,139,523 5,139,523 2,712 R2 0.481 0.542 0.352 Notes: This table reports the weather and climate impacts on ambient ozone concentrations, estimated by different methodologies. Column (1) reports the estimates of our unifying approach, in which we decompose daily maximum temperature into climate norms and weather shocks. Column (2) reports the effect of daily maximum temperature on ambient ozone from the panel fixed-effects approach, exploiting day-to-day variation in temperature, capturing the effect of a change in weather. Column (3) reports cross-sectional estimates using average maximum temperature and ambient ozone concentrations over all the years from 1980-2013 for each ozone monitor in the sample, capturing the effect of a change in climate. Note that while estimates in column (3) must additionally control for whether a county is in violation of the CAA ozone standards, this is implicitly controlled for via the fixed-effects in columns (1) and (2). Combining our estimates in column (1) with climate projections from the U.S. Fourth National Climate Assessment (Vose et al., 2017) under the business-as-usual scenario (RCP 8.5) – 1.6◦C temperature increase by 2050, and 4.8◦C by 2100 – ambient ozone concentrations would rise by 1.9 and 5.6 ppb, respectively. This should be the so-called “climate penalty” – the response of economic agents to longer-term climatic changes, which is inclusive of adaptation. Wrongly using the response to temperature shocks as the penalty, which is exclusive of adaptation, those numbers would be larger: 2.7 and 8 ppb, respectively. Standard errors are clustered at the county level in columns (1) and (2), while column (3) uses standard heteroskedastic robust errors. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 28 in column (2). Coincidentally, the effect for the lagged 30-year MA climate norm is also statistically the same as its counterpart in column (3). Specifically, a 1◦C temperature shock increases ozone concentration by 1.68 ppb, and a 1◦C change in climate norm increases ozone concentration by 1.16 ppb. To be clear, this does not imply that the cross-sectional approach is free of omitted variable bias concerns. In this specific context there may simply be both upward and downward bias simultaneously affecting the estimate as in Griliches (1977).47 In fact, when we re-estimate our model on a more balanced sample of monitors as a robustness check the bias in the cross-sectional approach becomes much more evident, leading to an over-estimation of the implied measure of adaptation by more than 100 percent.48 It is widely recognized that the cross-sectional approach is plagued with omitted variable bias. In our context, if more informed/concerned local monitoring agencies inspect heavy emitters of ozone precursors more often when average temperature rises, and more intense enforcement of environmental regulations induces reductions in ozone concentration, then this unobserved behavior might lead to underestimation of the long-run impact of temperature. On the other hand, as emphasized in the conceptual framework, estimates from the standard panel data fixed-effects methodology and our approach should be statistically the same due to the properties of the Frisch-Waugh-Lovell theorem. The deseasonalization embedded in the fixed-effects model is approximately equivalent to the use of deviations from 30-year norms in our regression model. Our estimates imply a so-called “climate penalty” on ozone on the lower end of the ranges found in the literature. Indeed, Jacob and Winner (2009), in their review of the effects of climate change on air quality, find that climate change alone may lead to a rise in summertime surface ozone concentrations by 1-10 ppb – a wide interval partly driven 47More generally, in contexts where one is able to control for all key covariates (e.g., in an agricultural setting with measurements of soil quality, storage, irrigation, and other relevant variables), then the fixed effects would be capturing much of the same content. But there are likely to be many contexts where the crucial controls are omitted – e.g., in analyses of mortality and health outcomes, several confounding factors may be unobserved such as genetic traits, defensive investments, and lifestyle choices such as smoking, drinking, and exercising. 48See estimates in Appendix A.2.a Table A.2.2. 29 by the different regional focuses of the studies they review. The U.S. EPA, in its 2009 Interim Assessment, claims that “the amount of increase in summertime average ... O3 concentrations across all the modeling studies tends to fall in the range 2-8 ppb” (USEPA, 2009, p.25). Combining our estimates in column (1) with climate projections from the U.S. Fourth National Climate Assessment (Vose et al., 2017) under a business-as-usual scenario (RCP 8.5), one would predict an increase in ambient ozone concentrations by the mid and end of the century in the range of 1.9-5.6 ppb, approximately.49 To be clear, “climate penalty” in our setting is the response of economic agents to longer-term climatic changes, which is inclusive of adaptation, as it will be discussed below. If one would wrongly use the response to temperature shocks as the penalty, which is exclusive of adaptation, the range would be 2.7-8 ppb, a nontrivial shift to the right. In fact, this may be one of the reasons why our estimate of the penalty is on the lower ranges of the values produced by simulation studies (again, for a review, see Jacob and Winner, 2009); they usually do not take into account behavioral responses. To put those values in perspective, each of the last few times EPA revised the air quality standards for ambient ozone, they decreased it by 5 ppb. B. Measuring Adaptation to Climate Change Our results indicate that short-run temperature shocks have a larger impact on ozone levels compared to long-run temperature norms. The comparison between the short- and longrun effects of temperature may provide a measure of adaptive responses by economic agents (Dell, Jones and Olken, 2012, 2014). Our measure of adaptation – also a comparison between the impact of changes in the long-run climate normal temperature (lagged 30-year MA) and the effect of the temperature shock (deviation from the MA) – is 0.51 ppb, suggesting that 49To be clear, while our estimate of adaptation does not rely on extrapolation, any prediction of the future “climate penalty” must do so by construction. In that sense, the “climate penalty” implied by our estimates may still be an upper bound. As we will discuss later in Subsection E, although our measure of adaptation has remained relatively constant over time, the impact of the climate norm on ozone has decreased. This could imply that long-run changes in the economic or regulatory landscape, driven, e.g., by technological advancement or shifting preferences, could lead to further decreases in this impact in the future. At the same time, we also find non-linear and increasing effects of temperature on ozone formation, indicating that there may be counter-acting intensification effects. 30 economic agents may be adapting to climate change. In the case of polluting firms, for example, they might be making adjustments to their production processes so that whenever average temperature rises, the emissions of ozone precursors reduce to keep ambient ozone at controllable levels. Such adjustments might be driven by public and regulatory pressures and/or technological innovation. If we ignored such adaptive responses by economic agents, then we would be overestimating the “climate penalty” on ozone by more than 44 percent. Again, we would be making the mistake of taking the effect of weather shocks as the penalty, when we should be looking at the impact of climatic changes, which incorporates adaptive responses by economic agents. Using the climate projections from the U.S. Fourth National Climate Assessment under the business-as-usual scenario (RCP 8.5), we would overestimate the climate penalty by 0.82 ppb by mid-century, and 2.47 ppb by the end of the century. C. Robustness Checks Measurement Error & Agents’ Expectations — A concern regarding our decomposition of meteorological variables in Equation (1.10) might be measurement error. Because both components are intrinsically unobserved, we define the long-run climate norm as the 30- year MA, and weather shocks as deviations from that moving average. If there is classical measurement error, the estimates of the coefficients of interest in Equation (1.13) will suffer from attenuation bias. Moreover, the bias will be magnified in fixed-effect regressions. To investigate the robustness of our results to measurement error, we carry out analyses using moving averages of different length. We start by using a 3-year MA, then 5-, 10-, and 20-year MAs, relative to our preferred specification using 30 years. As argued seminally by Solon (1992), as we increase the time window of a moving average, the permanent component of a variable that also includes a transitory component will be less mismeasured. If this is the case, we should observe the coefficients of interest increasing as longer windows are used for the moving averages. Our estimates in columns (1) through (4) of Table 1.2 remain 31 remarkably stable over the different lengths of the moving averages, but if anything they get slightly larger until the 20-year moving average. As pointed out by Blanc and Schlenker (2017), a fixed-effects regression with variables under classical measurement error is plagued by larger attenuation bias. The identifying variation in a standard panel analysis comes from deviations from the cross-sectional averages in the panel structure. Once the variables of interest are demeaned, the share of measurement error variation is magnified, and the coefficients of interest will be even more attenuated. Again, our estimates in Table 1.2 remain largely unchanged over the different lengths of the moving averages, with a slight attenuation of the coefficient of the moving average when we move from the 20- to the 30-year moving average. This latter result suggests that the widely used climate normals are close to the “optimal” long-run norms. The improvements from reducing measurement error might be offset by the panel-driven attenuation bias between 20- and 30-year time windows. At the same time, it is possible that agents form climate expectations in a way that exhibits recency weighting (e.g., Kaufmann et al., 2017). This presents a second trade-off. Longer, 20- to 30-year MAs, guided by climatology, appear “optimal” in our setting for navigating the first trade-off between potential measurement error and fixed effect induced attenuation bias for the purposes of estimating a long-run climate impact. Shorter, 3- to 5-year MAs, however, may better reflect agents’ internalized information set with regards to forming expectations over the current climate conditions and thus better capture mediumrun adaptive behavior (Moore et al., 2019). It is plausible, therefore, that the observed increases, however slight, in the coefficient on climate norm as we move from a 3- to a 20- year MA are, at least in part, due to agents’ stronger adaptive response to recent events than to longer-run trends in the climate norm. 32 Table 1.2: Key Robustness Checks Alternative Lengths of Climate Norm Adaptation Responses 3-year 5-year 10-year 20-year Long Run Long Run Short-Run MA MA MA MA 10-year Lag 20-year Lag 2004-2013 only (1) (2) (3) (4) (5) (6) (7) Temperature Shock 1.669*** 1.670*** 1.670*** 1.673*** 1.681*** 1.685*** 1.179*** (0.063) (0.062) (0.062) (0.062) (0.063) (0.063) (0.029) Climate Norm 1.158*** 1.166*** 1.176*** 1.175*** 1.155*** 1.143*** 0.581*** (0.049) (0.050) (0.051) (0.051) (0.050) (0.049) (0.034) Implied Adaptation 0.511*** 0.504*** 0.495*** 0.499*** 0.527*** 0.542*** 0.597*** (0.040) (0.040) (0.041) (0.041) (0.041) (0.041) (0.029) Shock x Action Day 0.068 (0.188) All Controls Yes Yes Yes Yes Yes Yes Yes Observations 5,139,523 5,139,523 5,139,523 5,139,523 5,131,943 5,127,886 1,879,041 R2 0.481 0.481 0.481 0.481 0.481 0.481 0.444 Notes: This table reports the results for key robustness checks investigating sensitivity to alternative definitions for the climate norm, as well as allowing more or less time for economic agents to engage in adaptive behavior. Columns (1) through (4) report estimates when we adjust the length of the constructed climate norm (moving averages of temperature) using different time windows. The estimates in columns (5) and (6) are obtained by Equation (1.13), but using 10- and 20-year lags between the moving average and contemporaneous temperature, rather than 1-year lag. Column (7) continues using the 1-year lag of the main specification, but adds an additional interaction term on temperature shock using clean air action day announcements (days in which the relevant air quality authority observes, or expects to observe, unhealthy levels of pollution on the Air Quality Index and releases a public service announcement to this effect) at the county-level to estimate short-run adaptive behavior. Note that although action day policies first began in the 1990’s, EPA data only begins from 2004 onwards, leading to a restricted overall sample (approximatley 35% of our full sample). The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 33 Lagged & Short-run Adaptive Responses — Another potential concern with our preferred specification might be the fact that we have used the 1-year lagged 30-year moving average to capture the long-term climate norm, implying that agents adapt within one year. Hence, we check the sensitivity of our results when agents have 10 or 20 years to adapt, instead of just one. In columns (5) and (6) of Table 1.2, we provide estimates from our preferred specification but using respectively 20-year moving averages of temperature lagged by 10 years, and 10-year moving averages lagged by 20 years. By doing so, we are providing agents more time to potentially adjust to climate change. Even though we would expect that the effects of the weather shocks to be similar, we anticipate the effects of the climate norm to be slightly smaller than before, as agents should now be able to adapt more than before. This is what we find from our estimates reported in Table 1.2, although the magnitude of the coefficients is remarkably close to that of our main results. Alternatively, one might be concerned that agents are in fact able to respond rapidly and adapt to weather shocks, in which case the coefficient on temperature deviations would be inclusive of any such adaptive responses, and thus our estimate of adaptation would be biased downwards. In column (7) we make use of a widespread policy of “Ozone Action Day” (OAD) alerts, where a local air pollution authority would issue an alert, usually a day in advance, that meteorological conditions are expected to be more conducive to a high concentration of ambient ozone in the following day. If agents are adapting to contemporaneous weather shocks, these “action days” would be the days we would be most likely to observe an adaptive response. Indeed, individuals are urged to take voluntary action to reduce emissions of ozone precursors such as working from home, carpooling to work, or using public transportation; combining auto trips while running errands; and reducing home landscaping projects. Firms are also urged to provide work schedule flexibility, reduce refueling of the corporate fleet during daytime, and save AC-related energy usage by adjusting indoor temperature (USEPA, 1997b). Interacting an indicator variable for days in which OAD alerts were issued for a given county with our other covariates, we find that such alerts have a negligible and statistically 34 insignificant impact on the effect of a 1◦C change in the contemporaneous temperature shock.50 Although previous studies have provided evidence of some decline in driving and increases in the use of public transportation in a few locations (e.g., Cutter and Neidell, 2009; Sexton, 2012), we find little indication that agents engage in meaningful short-run adaptive responses across the country. Further Robustness Checks — We conduct additional robustness checks regarding features in the construction of the data, selection of the estimating sample, and alternative econometric specifications in Appendix A.2.a Tables A.2.1, A.2.2, A.2.3, and A.2.4. Specifically, Table A.2.1 examines the sensitivity of our results to our algorithm for matching ozone and temperature monitoring stations. Table A.2.2 restricts our sample of ozone monitors to a semi-balanced panel, including only monitors with data for every year of our sample; however, as pointed out by Muller and Ruud (2018a), our preferred unbalanced panel is likely more nationally representative. Table A.2.3 examines the sensitivity of our results to the exclusion of regions that had implemented policies aimed at reducing ambient ozone concentrations by specifically targetting ozone precursors. Table A.2.4 contains four additional robustness checks: (i) implementing a daily MA rather than monthly; (ii) purposefully aggregating our data to the monthly level to simulate our methodology with lower frequency data; (iii) controlling for wind speed and sunlight with the subset of data for which that information is available; and (iv) examining the sensitivity of our results to inter-regional NOx transport by restricting the estimating sample to exclude, or conversely, only include, the states designated by the EPA as part of the “ozone transport region” (OTR). Across all of these models results remain qualitatively similar to our central findings. Finally, Appendix A.2.a Table A.2.5 provides bootstrapped standard errors for our main estimates, finding little difference relative to the standard errors clustered at the county 50Although the recovered coefficients of temperature shock, climate norm, and implied adaptation are quantitatively different for column (7) than columns (5) and (6), this is due to a difference in the underlying sample. EPA data on “action day” alerts were only provided from 2004 onwards, leading to a restricted overall sample (approximately 36% of our full sample). 35 level. In addition, that table presents standard errors clustered at the state level. Although they double in magnitude, they do not change the statistical significance of the results. D. Estimating Nonlinear Effects of Temperature In many empirical settings there has been a focus in the economics literature on allowing for nonlinear effects of temperature or climate on the outcome of interest. Thus, while our central model adopts a linear specification for simplicity in interpretation and comparison with prior methods, we note that our proposed approach is easily extendable to any nonlinear setting with nth order polynomial effects by simply including higher-order polynomial terms for both the weather shock, (xit − x¯ip¯), and climate norm, ¯xip¯. The following equation presents the quadratic model: yit = α + βW (xit − x¯ip¯) + βCx¯ip¯ + βW2(xit − x¯ip¯) 2 + βC2x¯ 2 ip¯ + µi + λs + νit, (1.14) while a cubic model would add the terms βW3(xit −x¯ip¯) 3 and βC3x¯ 3 ip¯ . Adaptation could then be inferred for the quadratic model as: Adaptation = (βW − βC) + 2(βW2(xit − x¯ip¯) − βC2(¯xip¯)), (1.15) while adaptation for a cubic model would add the term 3(βW3(xit−x¯ip¯) 2−βC3(¯xip¯) 2 ). Notably, for a marginal deviation of the daily temperature from the climate norm, i.e., xit − x¯ip¯ ≈ 0, Equation (1.15) simplifies to: Adaptation = βW − βC − 2βC2(¯xip¯), (1.16) with marginal adaptation in the cubic model additionally including the term −3βC3(¯xip¯) 2 . Note that estimating the impacts of climate and weather in a setting with nonlinear effects will also inherently include the interaction of these two channels of temperature response, as 36 discussed by Mendelsohn (2016), because the marginal impact of weather will vary with the underlying climate norm from which it is deviating.51 As an alternative to including nonlinear terms, one could construct a set of indicator variables denoting whether realized temperature at location i on day t fell within a certain temperature bin. By interacting these indicators with the shock, norm, and control variables in a linear model, the response function of the outcome variable to both weather and climate would be allowed to flexibly adjust across the temperature distribution in a piece-wise linear fashion.52 Allowing for a more flexible response function may be especially desirable in settings where the underlying functional form is unknown. Furthermore, by estimating a (locally) linear relationship within each bin, the specification allows for intuitive and easily interpretable measures of weather and climate impacts and implied measure of adaptation. The exact functional form of the ozone-temperature relationship is unknown because ozone formation may be intensified with higher temperatures, but also exhibits a shorter half-life (McClurkin, Maier and Ileleji, 2013). We thus examine the nonlinear effects of weather shocks and climate norms on ambient ozone concentrations across the temperature distribution using quadratic and cubic versions of Equation (1.13) by including the additional terms outlined in Equation (1.14). We also estimate a “binned” specification as described above by creating indicator variables denoting whether the contemporaneous daily maximum temperature at a given ozone monitor falls within a certain 5◦C temperature bin.53 Figure 1.4 depicts the ozone relationship and marginal response to climate and weather, as well as marginal adaptation, across the temperature distribution for the linear, quadratic, cubic, and binned specifications.54 The linear specification appears to provide an adequate 51To see this mathematically, one need only expand the higher order weather terms to see that they include the interaction effects. For example, the expansion of (xit − x¯ip¯) 2 includes the term −2xitx¯ip¯. 52In this way, the marginal effect of a 1◦C change in either component of temperature is constrained to be constant within its respective temperature bin, but is allowed to vary across each bin. 53The lowest bin is below 20◦C (just over the 10th percentile of our temperature distribution), and the highest bin is above 35◦C (90th percentile of our temperature distribution), with the middle bin, 25–30◦C, approximately centered around the temperature distribution median (27.8◦C) and mean (27.1◦C). 54Recall that the effects of climate and weather under higher order models depend on the level of the other variable. For graphing the climate norm effects we assume a weather shock of zero – approximately the sample average as the shocks are constructed as deviations from the norm. For graphing the weather 37 Figure 1.4: Comparing Linear, Binned, and Nonlinear Specifications Panel A. Relationship Between Ozone Concentration and Climate Norm 40 60 80 100 120 140 Ozone Concentration (ppb) 10 15 20 25 30 35 40 Climate Norm Temperature (C°) Panel B. Relationship Between Ozone Concentration and Weather (at Sample Mean Climate Norm) 40 60 80 100 120 140 Ozone Concentration (ppb) 10 15 20 25 30 35 40 Daily Max Temperature (C°) Panel C. Marginal Effect on Ozone Concentration of Climate Norm -1.0 0.0 1.0 2.0 3.0 Ozone Concentration (ppb) 10 15 20 25 30 35 40 Climate Norm Temperature (C°) Panel D. Marginal Effect on Ozone Concentration of Weather (at Sample Mean Climate Norm) -1.0 0.0 1.0 2.0 3.0 Ozone Concentration (ppb) 10 15 20 25 30 35 40 Daily Max Temperature (C°) Panel E. Implied Marginal Adaptation 0.0 0.5 1.0 1.5 2.0 Ozone Concentration (ppb) 10 15 20 25 30 35 40 Climate Norm Temperature (C°) Linear Binned Quadratic Cubic Specification Notes: This figure compares our central linear specification with a 5◦C binned linear specification, as well as quadratic and cubic specifications following Equation (1.14). For clarity in the figures, we trim the top and bottom one-percent of the temperature distribution. Panels A and B depict the relationship between ozone and climate or weather, respectively, across the temperature distribution. Panels C and D depict the marginal impacts of climate and weather on ozone concentration, with both the flexible binned specification and the cubic reflecting an “inverted u” shape, suggesting that while ozone increases with temperature, above a certain temperature it increases at a decreasing rate. Finally, Panel E shows marginal adaptation, wherein both the binned and cubic specifications exhibit a “normal u” shape, suggesting that adaptation is larger when temperature is higher. shock effects we assume the sample average climate norm of approximately 27.5◦C. 38 first-order approximation of the nonlinearities captured by the cubic and binned specifications, while the quadratic model appears to mis-specify the ozone-weather relationship compared to the other models. Although both the cubic and binned specifications find similar ozone and adaptation responses across the majority of the temperature distribution, due to the functional form restrictions of the cubic it also implies a large, and rather unintuitive, level of adaptation at lower temperatures. With this in mind, our preferred approach for capturing potential nonlinearities in our empirical context is the binned specification. Table A.2.6, column (1), in Appendix A.2 presents the results of our preferred specification when interacting each of the independent variables with the 5◦C temperature bin indicators. The implied measure of adaptation is then presented in column (2).55 Similar to Figure 1.4, we find that the ozone/temperature response is increasing at an increasing rate at lower temperature ranges, but increases at a decreasing rate at higher temperatures, particularly for increases in the climate norm. These results suggest that agents may be making extra effort to reduce ozone precursor emissions when temperatures are the highest and could otherwise lead to greater ozone formation. E. Exploring Heterogeneity Earlier studies have inferred adaptation indirectly, by flexibly estimating economic damages due to weather shocks, then assessing climate damages through shifts in the future weather distribution. We have pointed out the shortcomings of that time/space extrapolation approach in the spirit of the Lucas Critique (Lucas, 1976). Importantly, once we have recovered a measure of adaptation from responses to weather shocks and longer-term climatic changes by the same economic agents, then we are able to explore the heterogeneity in their degree of adaptation. In Appendix A.2.b we report results of heterogeneity analyses examining adaptive behavior over time in Figure A.2.1 and Table A.2.8, across varying measures of belief in climate change in Tables A.2.9 – A.2.11. Table A.2.12 then examines how the effect 55Table A.2.7 additionally compares the implied level of adaptation under the linear, binned, quadratic, and cubic specifications. 39 of temperature on ozone may be attenuated if the local atmosphere is limited in one ozone precursor (NOx or VOCs) relative to the other. V. Concluding Remarks We developed a unifying approach to measuring climate change impacts and adaptation that considers both responses to weather shocks and longer-term climatic changes in the same estimating equation. By bridging the two earlier strands of the climate-economy literature – cross-sectional studies that relied on permanent, anticipated components behind meteorological conditions (e.g., Mendelsohn, Nordhaus and Shaw, 1994; Schlenker, Hanemann and Fisher, 2005), and panel fixed effects that exploit transitory, unanticipated weather shocks (e.g., Deschenes and Greenstone, 2007; Schlenker and Roberts, 2009) – we have overcome identification concerns from earlier cross-sectional studies, improved on the measurement of adaptation, and provided a test for the statistical significance of this measure. Our approach rests on two rather simple but powerful ideas. First, the decomposition of meteorological variables into long-run climate norms and contemporaneous weather shocks. Second, the properties of the Frisch-Waugh-Lovell theorem, which enables the simultaneous identification of these two short- and long-run impacts. In the spirit of Dell, Jones and Olken (2012, 2014), we recovered a measure of adaptation defined as the difference between those short- and long-run responses. Unlike previous studies, however, this measure was derived directly from coefficients estimated in the same fixed-effects model; hence, less susceptible to omitted variable biases from cross-sectional estimates. In addition, it compares the responses of the same economic agents, overcoming the challenges of identifying adaptation by comparing the profiles of weather responses across time and space (e.g., Deschenes and Greenstone, 2011; Barreca et al., 2016; Heutel, Miller and Molitor, 2021), which requires that preferences be constant across those dimensions. We applied our unifying approach to study the impact of climate change on ambient ozone in U.S. counties over the period 1980-2013. Others have relied on atmospheric-sciences sim40 ulation models to study the so-called “climate penalty” on ozone (see a review in Jacob and Winner, 2009). By ignoring the adaptive behavior of economic agents, they may have substantially overestimated the magnitude of this penalty – in our study setting, disregarding adaptation overestimates the climate penalty by approximately 44 percent. When considering the impacts of climate change on air pollution, the application of our unifying methodology led to four main findings. First, a changing climate appears to be affecting ambient ozone concentrations in two ways. A 1◦C shock in temperature increases ozone levels by 1.68 parts per billion (ppb) on average, which is expectedly what would have been found in the standard fixed-effects approach. A change of similar magnitude in the 30-year moving average increases ozone concentration by 1.16 ppb. Second, we found strong evidence of adaptive behavior. For a 1 ◦C change in temperature, our measure of adaptation in terms of ozone concentration is 0.51 ppb, which is statistically and economically significant. Third, by extending our central model to flexibly recover estimates accounting for the nonlinear relationship between ozone and temperature, we found that adaptation is highest on the hottest days, which would ex ante be likely to lead to higher levels of ambient ozone. Finally, we provided evidence of nontrivial heterogeneity in the degree of temperature response and adaptation across time and space, which highlights the potential biases of existing approaches in assigning weather responses or adaptation from one period and/or location to other periods and locations, consistent with insights by Olmstead and Rhode (2011) and Bleakley and Hong (2017). Notably, although we made use of high frequency data in this study, our unifying framework is generalizable to any empirical setting where one can obtain short-term variation in weather associated with limited opportunities to adapt, and long-term climatological variation allowing for adaptation. Settings in which opportunities to adapt are limited at the daily level, but may exist at the monthly or seasonal level are reliant on temporally disaggregated data, while those in which such opportunities are limited even at the monthly or seasonal level may be able to use more aggregate data. 41 Chapter 2: Incidental Adaptation: The Role of Non-Climate Regulations I. Introduction Many government institutions, policies, and regulations have been established to help smooth out private shocks in varied contexts such as employment, health, or housing.1 Due to the nature of climate change, however, some of these shocks are likely to become more frequent and/or severe (IPCC, 2021), and it is unclear, a priori, whether these existing institutions, policies, and regulations may induce or constrain private climate adaptation.2 On the one hand, existing government policies may distort private decisions, inhibiting agents’ adaptation; on the other hand, they may correct market failures, inducing adaptation. Given the political gridlock surrounding climate change mitigation efforts, understanding whether existing policies can induce climate adaptation is of particular importance (Nordhaus, 2019; Aldy and Zeckhauser, 2020; Goulder, 2020), though notably such policies should be viewed as complements, rather than substitutes, of first-best climate policy.3 This study conceptualizes and demonstrates the possibility for regulation-induced adaptation, examining the context of 1For example, unemployment insurance helps households smooth out the income effects of a labor shock, leading to more efficient labor outcomes (Acemoglu and Shimer, 1999) and helping to avoid home foreclosures (Hsu, Matsa and Melzer, 2018). Similarly, medical insurance – whether directly provided via Medicare/Medicaid, or facilitated through, e.g., a local Affordable Care Act health exchange – can smooth out negative health shocks by increasing access to care (e.g., Doyle Jr, 2005), while the National Flood Insurance Program covers over a trillion dollars’ worth of housing and related assets (Michel-Kerjan, 2010). 2The IPCC defines adaptation as “the process of adjustment to actual or expected climate and its effects in order to moderate harm or take advantage of beneficial opportunities,” and further states that “[a]daptation plays a key role in reducing exposure and vulnerability to climate change. (...) In human systems, adaptation can be anticipatory or reactive, as well as incremental and/or transformational.” (IPCC, 2022). 3Furthermore, although market forces can lead to adaptation, that alone may not be enough to adequately adapt to climatic changes and may lead to devastating distributional impacts. 42 the existing National Ambient Air Quality Standards (NAAQS) for ozone pollution.4 We develop a tractable analytical framework to highlight how pre-existing regulatory incentives can affect behavioral responses to climate change and thus may incidentally induce or inhibit climate adaptation. We then econometrically recover key parameters to calculate the welfare effects of the regulation-induced adaptation co-benefit of the ozone NAAQS. An advantage of our econometric approach is that it recovers a measure of adaptation arising from the behavior of the same economic agents. This allows us to compare the relative magnitudes of adaptation between counties in or out of attainment with the NAAQS regulation, in the same estimating equation, to empirically recover a measure of regulation-induced adaptation without making further assumptions over preferences across time or place. We estimate adaptation in both attainment and nonattainment counties as the difference between the ozone response to increases in temperature due to transitory weather shocks and shifts in the climate normal temperature. Weather shocks, by their nature, are observed simultaneously with their impact on ozone concentrations, affecting ozone formation directly – conditional on the level of ozone precursor emissions – such that agents have few if any avenues to adjust their behavior in response to weather shocks. On the other hand, shifts in the expected climate norm are observable by agents by, for example, looking at the average temperature of previous years, and thus may affect the level of precursor emissions if agents adapt to a shifting climate by changing their emissions profile. Therefore, while an increase in temperature would typically increase ozone concentrations, counties in violation of the ozone air quality standard – those designated as out of attainment, or in “nonattainment” with the NAAQS – would be pressured to take action to bring those levels down, adapting 4Ozone is a local pollution externality formed by a production function that exhibits Leontief-like properties in its inputs (Auffhammer and Kellogg, 2011) – “precursor” emissions of Nitrogen Oxides (NOx) and Volatile Organic Compounds (VOCs) – in the presence of sunlight and warm temperatures; hence, affected by climate change. Exposure to ambient ozone has important economic implications because it leads to increases in hospitalization, medication expenditure, and mortality (e.g., Neidell, 2009a; Moretti and Neidell, 2011; Deschenes, Greenstone and Shapiro, 2017). Beginning in 1980, the Environmental Protection Agency (EPA) has monitored and regulated ambient ozone concentrations via the ozone NAAQS to protect human health. The NAAQS themselves set a pollution concentration threshold that counties cannot exceed, reducing the frequency and magnitude with which individuals face pollution exposure shocks. 43 to expected climate normal temperatures and thus attenuating the climate impact. In other words, climate adaptation induced by the NAAQS. We account for any “baseline” level of adaptation or other confounding effects by differencing out the measure of adaptation in attainment counties from nonattainment counties, recovering an estimate of regulationinduced adaptation (RIA) akin to a difference-in-differences estimator. Ultimately, we embed our estimates into our analytical framework, combined with an estimate of the marginal damages of ozone from the literature, allowing us to calculate a back-of-the-envelope measure of the welfare effects of the additional adaptation induced by the NAAQS regulation. While our empirical analysis focuses on a negative production externality – ambient ozone – regulation-induced adaptation may occur in any context where (i) the corrective policy reduces the market failure of interest, by directly targeting the relevant outcome, and (ii) climate change would otherwise exacerbate the market failure. Among many possible examples, consider existing programs and policies intended to correct the under-provision of vaccines to individuals, which can provide potentially large external benefits (White, 2021). Climate change may increase the incidence or severity of disease outbreaks.5 Individuals may respond to this increase by taking advantage of existing vaccine provision programs. That is, the existence of the vaccine provision program allows (induces) these individuals to engage in adaptive behavior, incidentally attenuating the impact of climate change. Similarly, consider institutions to correct the under-provision of public safety, such as governmentmaintained law enforcement agencies. Increasing temperatures may increase the probability of violence or unlawful activity (e.g., Ranson, 2014; Mukherjee and Sanders, 2021; Hsiang, Burke and Miguel, 2013), but individuals may respond to this increase by calling on existing law enforcement to deter or reduce the severity of incidents, attenuating the climate impact.6 To understand the mechanism behind regulation-induced adaptation in our setting, con5For example, warmer winters are associated with milder influenza seasons, but often lead to a more severe influenza season the following winter (Towers et al., 2013). 6Notice that defined in this way, RIA is not the regulator responding directly to climate change by, e.g., amending existing regulation or policy, but individual economic agents who are taking advantage of existing regulation or policy to adapt to the effects of a changing climate. 44 sider a location where emissions of ozone precursor pollutants — Nitrogen Oxides (NOx) and Volatile Organic Compounds (VOCs) — are under control in the baseline. If a rise in temperature leads to more intense ozone formation and the violation of the NAAQS, economic agents will be designated as in nonattainment and pressured by the U.S. Environmental Protection Agency (EPA) to adopt pollution abatement strategies to reduce emissions of NOx and VOCs, and ultimately ambient ozone concentration. Since those actions would have to be taken not because of higher ozone precursor emissions but rather higher temperatures, we refer to the resulting decline in ozone levels as adaptation to climate change induced by the ozone NAAQS.7 At the end of the day, in addition to smoothing out “status quo” pollution shocks, this existing Clean Air Act (CAA) regulation may encourage behavioral adjustments that also attenuate the pollution shocks triggered by climate change. Our results demonstrate that existing policies unrelated to climate change can indeed facilitate adaptation, and the magnitude of the effect is of economic significance. In the absence of adaptation, a 1◦C increase in temperature would increase the ambient ozone concentration in nonattainment counties by 1.99 parts per billion (ppb), on average. Adaptation reduces this impact by 0.64 ppb, with 0.33 ppb due to regulation-induced adaptation (RIA). In other words, adaptation reduces the climate impact on ozone by about one third in nonattainment counties, with over half of the effect attributable to RIA. To put this effect in perspective, a 1.5◦C temperature increase – the midpoint of the representative concentration pathway (RCP) 4.5 and 8.5 warming scenarios for mid-century – would increase ozone by approximately 3 ppb in the absence of adaptation, but only 2 ppb once accounting for adaptation, with 0.5 ppb of this decrease due to RIA. Combined with an estimate of the social costs of ozone increases from the literature (Deschenes, Greenstone and Shapiro, 2017), our 7By definition, climate adaptation involves adjusting to or coping with climatic change with the goal of reducing our vulnerability to its harmful effects. So, this is not a new use of the term climate adaptation. In the context of responses to natural disasters, for example, Kousky (2012) explains that “[t]he negative impacts of disasters can be blunted by the adoption of risk reduction activities. (...) [T]he hazards literature (...) refers to these actions as mitigation, whereas in the climate literature, mitigation refers to reductions in greenhouse gas emissions. The already established mitigation measures for natural disasters can be seen as adaptation tools for adjusting to changes in the frequency, magnitude, timing, or duration of extreme events with climate change.” (p.37, our highlights). 45 estimates would translate to between $794-908 million (2015 USD) per year in total adaptation welfare benefits by mid-century depending on the warming scenario (i.e., RCP 4.5 or 8.5), with $412-471 million attributable to the regulation-induced adaptation co-benefit of the NAAQS. For comparison, the cost of reducing the current NAAQS threshold by 1 ppb is $296 million per year (USEPA, 2015b). Importantly, corresponding RIA measures for other key outcomes of local economic activity – employment and wages – are precise zeros, suggesting that our RIA measure captures differential responses to regulation rather than differences in other county-level drivers of emissions. Additionally, sample restrictions based on persistent vs. changing attainment status provide supportive evidence that our results are not driven by a sub-set of counties, and that the parallel trends assumption is satisfied. Our findings are robust to a wide variety of sample restrictions and specification checks, such as: accounting for competing regulations on ozone precursors, allowing for differential responses based on counties’ proximity to the NAAQS nonattainment threshold, employing alternative climate measurements, allowing agents to have longer periods of adjustment to climatic changes, allowing for instantaneous adaptation from ozone alert days, among others. We also find suggestive evidence that regulation-induced adaptation is greater on days when temperature is higher (and higher ozone concentrations would thus be more likely, ceteris paribus), when local beliefs in the existence of climate change are stronger, and when the chemical composition of the local atmosphere is “limited” in either of the two ozone precursor pollutants.8 This study makes three main contributions to the literature. First, it provides an analytical framework and credible empirical evidence that non-climate policies correcting existing market failures can be used as a buffer to climate shocks while also inducing climate adaptation. When the outcome of interest arises from market failures, and climate change would exacerbate those failures (e.g., Goulder and Parry, 2008; Bento et al., 2014), existing non8To simplify our analytical framework, we follow Auffhammer and Kellogg (2011) and represent ozone formation by a Leontief-like production function. However, we recognize the complexity of ozone formation and run heterogeneity analysis by the composition of the local atmosphere. Additional reductions in the limiting precursor pollutant will typically lead to a larger overall reduction in ambient ozone concentrations. 46 climate policies may be able to smooth out the climate-exacerbated impacts and induce adaptation.9 In contrast, prior research had highlighted perverse adaptation incentives generated by existing non-climate policies due to distortion of private behavior – e.g., Annan and Schlenker (2015) show that farmers may not engage in the optimal protection against extreme heat when crop losses are covered by the federal crop insurance program. Second, it demonstrates that existing government policy can also provide a catalyst for adaptation. Previous work had examined the role of market forces or private responses in adapting to climatic changes (e.g., Barreca et al., 2016), but private incentives may be limited in scope or distribution. Third, it points out a nontrivial incidental co-benefit of the Clean Air Act (CAA) – climate adaptation. Prior literature had analyzed the impacts of the CAA on air quality itself (e.g., Henderson, 1996; Auffhammer and Kellogg, 2011; Deschenes, Greenstone and Shapiro, 2017), and a variety of other economic outcomes (see a recent review by Aldy et al., 2020), including unintended consequences (e.g., Becker and Henderson, 2000; Gibson, 2019), but interactions between existing CAA regulations and climate change had been overlooked. The paper proceeds as follows. Section II presents our analytical framework to understand how existing government regulations and policy may affect adaptation to climate change. Section III provides a background on the NAAQS for ambient ozone, ozone formation, and the data used in our empirical analysis. Section IV introduces the empirical strategy; Section V reports and discusses the results; and Section VI concludes. II. Analytical Framework The creation of new regulations can often prove politically or technologically infeasible, but existing regulations may mimic key incentives of a new regulation. In the context of climate change, several global climate policy architectures – basically new regulations – have been 9 In the same spirit, Mullins and White (2020) find that the improved access to primary care services provided by the publicly-funded Community Health Centers rolled out across U.S. counties in the 1960s and 1970s moderated the heat-mortality relationship by 14.2 percent. 47 proposed over the years (e.g., Nordhaus, 2019; Aldy and Zeckhauser, 2020). Nevertheless, because of free-riding concerns, political polarization, and disagreement over the distribution of the costs of climate change mitigation, it has proven difficult to convince countries to join into an international agreement with significant emission reductions, or to enact federal legislation addressing climate change. Recognizing the difficulty of implementing first-best climate policy, and the urgency in tackling the challenges of climate change, Goulder (2020) advocates for considerations of political feasibility and costs of delayed implementation in the choice of climate policy. Second-best policies may be socially inefficient, but if they are politically feasible for nearterm implementation, they might move up in the ordering of the policies considered by the federal government, akin to the discussion by Goulder (2020) in the context of climate policies.10 In this study, we demonstrate that under certain conditions existing government regulations are already providing incentives for producers and consumers to adapt to climate change – much like a second-best policy – and argue that policymakers should take these co-benefits into consideration when enforcing or revising them. A. The Nature of Existing Regulations Influencing Adaptation To understand how existing “smoothing” institutions, policies, and regulations may induce climate adaptation, consider a simple formalization using a static analytical framework with a representative agent in the spirit of, e.g., Bovenberg and Goulder (1996) and Goulder et al. (1999). Assume that this agent enjoys utility from both a consumption good, Y , and an emissions-producing consumption good, X, with E the economy-wide emissions concentration from producing X. 11 The agent’s utility function is given by: 10Many other second-best policies have been implemented around the world. The economic rationale has been laid out many decades ago (Lipsey and Lancaster, 1956). In the context of climate change, a prominent example in the United States is the corporate average fuel economy (CAFE) standards. A first-best policy would be taxing tailpipe emissions directly. 11Note that E refers to the emissions of, e.g., a local pollutant and not greenhouse gas emissions. As this is a representative agent model, all production, consumption, and emissions can be considered local to the agent; in our empirical context this would be interpreted as local production causing local emissions, 48 U = u(X, Y ) − ϕ(E), (2.1) where u(.) is utility from non-external goods and is quasi-concave, ϕ(.) is disutility from the economy-wide concentration of emissions and is weakly convex, and we assume additive separability of terms.12 The representative agent competitively produces both X and Y using a fixed (exogenous) endowment of labor, L, as the only factor of production. Furthermore, assume that the marginal product of labor is constant in each industry, and normalize output such that marginal products – and thus wage rate – is unity, implying that the unit cost of producing X or Y is also unity. Additionally, assume that the emissions produced per unit of X is e, such that the economy-wide level of emissions, E, is equal to eX. 13 The agent’s budget constraint is thus: pXX + Y = L + G, (2.2) where pX is the demand price of X (equal to unity in the absence of any smoothing policy), L is the exogenous endowment of labor, and G is a lump-sum government transfer to the agent equal to any revenue raised through the chosen smoothing policy (zero in the absence of any such policy). The representative agent chooses X and Y to maximize utility subject to this budget constraint, taking external damages as given. First, consider a scenario in which the chosen smoothing policy is a corrective “quota”- based regulation imposed on emissions above a certain concentration, as with the NAAQS in our empirical setting.14 That is, the government defines some threshold, E¯, above which i.e., within the same county. In other contexts, this could be, e.g., a metropolitan area, state, or transport region. Furthermore, the use of local pollution emissions is without loss of generality, as the same framework would apply to any external output produced in proportion to X – positive or negative – where its creation or economy-wide level is somehow impacted by climate change. 12For the sake of simplicity in exposition we will focus on a negative externality, as in our empirical context, though this is without loss of generality, as all concepts and insights would similarly apply to the context of a positive externality by simply reversing the sign on ϕ. 13This implies that reductions in emissions can only be achieved by reducing production of X or through abatement activities which incur costs equivalent to reducing production of X. 14Again, the assumption here of a specific policy type is without loss of generality — any corrective policy, such as a standard Pigouvian-style tax or Coasian permit-based policy would yield similar overall results. 49 they impose a (virtual) tax of tE on each additional unit of E. Thus, G =    tE(E − E¯), if E > E. ¯ 0, otherwise. (2.3) and similarly, profit per unit of X would be:    pX − {1 + tEe}, for each unit of X > X. ¯ pX − {1}, otherwise. (2.4) where X¯ = E¯ e is the implicit “production threshold” arising from the regulation, and profits in equilibrium are equal to zero. The corrective regulation thus raises the marginal cost of X – for any production at or above X¯ – inducing output substitution towards the now comparatively more profitable production of Y . Now, consider this same scenario but additionally assume that climate interacts with the economy-wide level of emissions by allowing E to be conditional on climate, c, that is, E ≡ E(c) = e(c)X. Thus, for the same level of existing regulation, tE, under an increasing climate the representative agent would potentially face a more stringent constraint on their production of X, and would re-optimize to maximize profits such that Xc ≤ X0. Notably, this re-optimization would only occur if the agent were constrained by the regulation’s threshold; if E0 ≤ Ec ≤ E¯, the regulation would remain non-binding and the agent would observe a “silent” increase in the level of economy-wide emissions. In other words, if the pre-existing corrective policy is binding (or becomes binding in the presence of climate change) it would induce behavioral adjustments, i.e., additional reductions in X, in response to an increasing climate – that is, regulation-induced adaptation. Let us compare a scenario with no corrective smoothing regulation, denoted with superscript N, against a scenario with a corrective smoothing regulation, with superscript R. For a marginal change in climate, dc, the general equilibrium welfare effect of regulation-induced adaptation (RIA) consists of two key components (see Appendix B.3.a for a proof): 50 1 λ dV dc ∆R = − ϕ ′ λ | {z } Marginal Damages dE dc R − dE dc N | {z } RIA , (2.5) where 1 λ dV dc is the change in welfare due to an incremental change in climate, ∆R denotes the discrete change from a context without a binding regulation on E to a context with one, and importantly: (i) ϕ ′ λ is the monetized marginal damages of the emissions concentration, while (ii) dE dc R and dE dc N reflect the change in the economy-wide level of emissions due to a changing climate with and without regulation, respectively. Thus, with an estimate of ϕ ′ λ from the literature (e.g., Deschenes, Greenstone and Shapiro, 2017, provide an estimate of the WTP to avoid a marginal increase in ambient ozone), and our own econometrically recovered estimates of dE dc R and dE dc N , we can calculate the welfare “co-benefit” of the preexisting NAAQS regulation as the monetized value of regulation-induced adaptation. In the following subsection, Figure 2.1 presents a schematic representation of regulation-induced adaptation in nonattainment counties, relative to any actions taken by attainment counties. Notably, there may be systematic deviations between regulated and unregulated regions (or even for the same region across different time periods) that could lead to level differences in “off the shelf” estimates of dE dc R and dE dc N which would in turn contaminate any welfare calculation. For example, in our empirical setting, counties are only constrained by the NAAQS if their ozone concentration levels are above the set threshold – thus, by definition, regulated counties will have inherently higher levels of baseline emissions and any increases in climate will in turn lead to comparatively higher levels of new ozone formation, before accounting for adaptation. At the same time, there may be other, exogenous, drivers of adaptation that could affect both regulated and unregulated counties. In order to account for these and other possible issues, we first econometrically estimate overall adaptation for both regulated and unregulated counties – in the same estimating equation – and use our estimates of adaptation in regulated and unregulated counties as the welfare parameters dE dc R and dE dc N respectively, de-facto “differencing out” any level differences as well as any 51 adaptation that is exogenous to the regulation of interest. Not every pre-existing policy or regulation with climate interactions will induce adaptation, however, as has been documented in prior literature (e.g., Annan and Schlenker, 2015). Thus, it is useful to examine under what conditions we can expect such regulations to induce or inhibit adaptation. As shown above, policies which correct a pre-existing market failure will incidentally induce adaptation if that market-failure has climate interactions. In Appendix B.3.b, we extend our analytical framework to show how and why the opposite also holds true – policies or regulations that distort private behavior will inhibit adaptation if the distortion has climate interactions. We additionally extend the original framework to examine input, rather than output, regulations on emissions or other externalities – showing that even when the output has climate interactions, if the input lacks any climate interaction, then the regulation will fail to induce adaptation. Specifically, while such input regulations may reduce climate impacts – for example, by reducing the baseline level of precursor emissions – if the input(s) lack clear climate interactions then the regulation would not create any incentive for agents to adjust their behavior in response to climate change. That is, the regulation would not induce any climate adaptation. B. A Schematic Representation of the Framework for Ambient Ozone and NAAQS We apply the analytical framework in an empirical setting, focusing on the existing Clean Air Act (CAA) regulation — specifically, the National Ambient Air Quality Standard (NAAQS) for ozone. With the CAA Amendments of 1970, the EPA was authorized to set up and enforce a NAAQS for ambient ozone.15 Since then, a nationwide network of air pollution monitors has allowed EPA to track ozone concentrations, and a threshold is used to determine whether pollution levels are sufficiently dangerous to warrant regulatory action.16 Counties with ozone levels exceeding the NAAQS threshold are designated as in “nonattainment” and 15For further details of the ozone NAAQS see Appendix B.1.a. 16Exposure to ambient ozone has been causally linked to increases in asthma hospitalization, medication expenditures, and mortality, and decreases in labor productivity (e.g., Neidell, 2009a; Moretti and Neidell, 2011; Graff Zivin and Neidell, 2012; McGrath et al., 2015; Deschenes, Greenstone and Shapiro, 2017). 52 the corresponding state is required to submit a state implementation plan (SIP) outlining its strategy for the nonattainment county to reduce air pollution levels in order to reach compliance.17 Depending on the severity of the exceedance, counties are given between 3- and 20-years to reach compliance, but in all cases, counties must show active progress within the first three years (USEPA, 2004).18 In cases of persistent nonattainment, the CAA mostly mandates command-and-control regulations, requiring that plants use the “lowest achievable emissions rate” technology (LAER) in their production processes. Furthermore, if pollution levels continue to exceed the standards or if a county fails to abide by the approved plan, sanctions may be imposed on the county in violation, such as retention of funding for transportation infrastructure. To make the concept of regulation-induced adaptation as clear as possible in the context we are studying, we use the schematic representation depicted in Figure 2.1. In this representation, we follow Auffhammer and Kellogg (2011) and use a simplified characterization of the process of ozone formation as a Leontief-like production function using two inputs – NOx and VOCs.19 In Panel A, the y-axis represents the regulated output – ozone formation – and the x-axis represents a composite index I(.) of those two inputs, whose levels move along the production function F(I(NOx,VOCs), Climate) represented by the upward-sloping black line. F(I(NOx,VOCs), Climate) is equivalent to E(c) = e(c)X in the formalization above. The blue horizontal line represents the maximum ambient ozone concentration, E¯, a county may reach while still complying with the NAAQS for ozone. Above that threshold, a county would be deemed out of compliance with the standards, or in nonattainment. Panel 17Appendix Table B.1.1 details the current and historical thresholds used to determine nonattainment status under the prevailing NAAQS. 18In later robustness checks we examine the sensitivity of our estimates to shortening or lengthening the time counties are given to reach compliance, finding no statistical or economically meaningful difference. We thus opt to follow the EPA’s regulatory schedule by using a 3-year lag of nonattainment status. 19Naturally, ozone production is much more complicated. Notably the relationship varies significantly with the composition of the atmosphere and physical forcings. The exact relationship is often times proxied by the ratio of VOCs to NOx, but mixing ratios and the reactivity of available VOCs add a lot of uncertainty to the actual production (e.g., Sillman and He, 2002). To account for these nuances in ozone formation, in our empirical analysis we explore the heterogeneity of our main effects to the composition of the local atmosphere – VOC-limited vs. NOx-limited. 53 Figure 2.1: Conceptual Framework on Regulation-Induced Adaptation Panel A. Ozone, Ozone Precursors, and Climate Panel B. Ozone Precursors (Inputs) Notes: This figure provides a schematic representation of the conceptual framework used in our analysis. We follow Auffhammer and Kellogg (2011) and use a simplified characterization of ozone formation as a Leontief-like production function using two inputs – NOx and VOCs. In reality, ozone formation is much more complicated, as discussed in the text. In Panel A, the y-axis represents ozone formation and the x-axis represents a composite index I(.) of NOx and VOCs, whose levels move along the linear production function F(I(NOx,VOCs), Climate) represented by the upward-sloping black curve. The blue horizontal line represents the maximum ambient ozone concentration a county may reach while still complying with the NAAQS. In point A, a county is complying with the standards. When average temperature rises, the chemical production function shifts upward and to the left, represented by the red upward-sloping curve. For the same level of I(NOx,VOCs), ozone concentration increases to point B. Because the county is now out of attainment, they are required to make adjustments to comply with the NAAQS. As they take steps to reduce emissions of ozone precursors to reach attainment – moving along the new chemical production function curve until point C – those economic agents are in fact adjusting to a changing climate. Indeed, as Panel B shows, agents must reduce the production of ozone precursors in order to reach point C. B illustrates the Leontief-like production function of ozone with respect to its precursors, VOCs and NOx, on the x- and y-axis, respectively, and resulting ozone “isoquant” curves 54 increasing up and to the right.20 Assume that an ozone monitor is sited in a county that is initially complying with the standards, as in point A. Moreover, suppose for simplicity that emissions of ozone precursors are such that ozone levels are initially under control, but then temperature rises. Because Panel A depicts a bidimensional diagram representing ozone as a function of I(NOx,VOCs) – taking climate as given – an increase in temperature shifts the production function upward and to the left. This new production function under climate change is represented by the red upward-sloping line. Because we assumed emissions of ozone precursors were initially under control, an increase in average temperature raises ozone concentration for the same level of the index I(NOx,VOCs), reaching point B. Since the ozone concentration is now above the NAAQS threshold, the county is designated as out of attainment, and firms are pressured to make adjustments in their production process to comply with the air quality standards in the near future, usually three years after a county receives the nonattainment designation. Notice that firms need to respond to the regulation not because they were careless in controlling emissions in the baseline, but rather because climate has changed. As they take steps to reduce emissions to reach attainment, moving along the new production function until point C as shown in both Panel A and B, those economic agents are in fact adjusting to a changing climate. This new production function (technology) may have a cost advantage in the abatement of ozone precursors in the state of the world with climate change. Thus, the agents are adapting to climate change because of the ozone NAAQS regulation, that is, they are engaging in regulation-induced adaptation.21 20In reality, the ozone isoquants might bend inward, especially on the vertical (NOx) axis. There is a fairly large region in which NOx decreases actually increases ozone. 21Ambient ozone concentration is a negative externality. For completeness, public policy can also induce adaptation to climate change in addressing positive externalities. Besides the social desirability of increasing the level of those outcomes, such policies can create a co-benefit of adjusting to a changing climate. One example is the Medicaid-covered influenza vaccination. Severe influenza seasons are likely to emerge with global warming (Towers et al., 2013), but publicly-funded annual vaccination allows Medicaid beneficiaries to cope with climatic changes. This is in addition to the herd-immunity impact of influenza vaccination (White, 2021). Thus, the concept of policy-induced adaptation is quite broad, and incentives affecting adaptive behavior are already in place in a variety of policies implemented around the world. 55 III. Data and Data Descriptions Ambient ozone is one of the six criteria pollutants regulated under the existing Clean Air Act. However, unlike other pollutants, it is not emitted directly into the air. Rather, it is formed by Leontief-like chemical reactions between nitrogen oxides (NOx) and volatile organic compounds (VOCs), under sunlight and warm temperatures. Because ambient ozone is affected by both climate and regulations, and high-frequency data are available since 1980, this is an ideal setting to study regulation-induced adaptation. In Appendix B.1, we provide further details regarding the ozone standards, ozone formation and the data. A. NAAQS, Ozone Pollution, and Climate: Background and Data NAAQS data. For data on the Clean Air Act nonattainment designations associated with exceeding the NAAQS for ambient ozone, we use the EPA Green Book of Nonattainment Areas for Criteria Pollutants, which provides an indicator of nonattainment status for each county-year in our sample. In our empirical analysis, we use the nonattainment status lagged by three years because EPA gives counties with heavy-emitters at least three years to comply with NAAQS for ambient ozone (USEPA, 2004, p.23954).22 Specifically, with regards to nonattainment status, if any monitor within a county exceeds the NAAQS, EPA designates the county to be out of attainment (USEPA, 1979, 1997a, 2004, 2008, 2015a). While the structure of enforcement is dictated by the CAA and the EPA, much of the actual enforcement activity is carried out by regional- and state-level environmental protection agencies, with local agencies having discretion over enforcement as long as they are within attainment for the NAAQS. Regional EPA offices do, however, conduct inspections to confirm attainment status and/or issue sanctions when a state’s enforcement is below 22EPA allows nonattainment counties with polluting firms between 3 to 20 years to adjust their production processes. Nonattainment counties are “classified as marginal, moderate, serious, severe or extreme (...) at the time of designation” (USEPA, 2004, p.23954). They must reach attainment in: “Marginal – 3 years, Moderate – 6 years, Serious – 9 years, Severe – 15 or 17 years, Extreme – 20 years” and show active progress within the first three years (USEPA, 2004, p.23954). 56 required levels, and assist states with major cases. Thus, while there may be heterogeneity in local enforcement for nonattainment counties, we would expect that those counties achieve at least the minimum level of increased regulation mandated by the EPA. Ozone data. For ambient ozone concentrations, we use daily readings from the nationwide network of the EPA’s air quality monitoring stations. Following Auffhammer and Kellogg (2011) and the regulatory design implemented by the Clean Air Act for designating a county as out of attainment, in our preferred specification we use an unbalanced panel of ozone monitors and make only two restrictions to construct our analysis sample. First, we include only monitors with valid daily information. According to EPA, daily measurements are valid for regulation purposes only if (i) 8-hour averages are available for at least 75 percent of the possible hours of the day, or (ii) daily maximum concentration is higher than the standard. Second, as a minimum data completeness requirement, for each ozone monitor we include only years for which at least 75 percent of the days in the typical ozone monitoring season (April-September) are valid; years having concentrations above the standard are included even if they have incomplete data.23 Our final sample consists of valid ozone measurements for a total of 5,139,129 monitor-days.24 Weather data. For climatological data, we use daily measurements of maximum temperature as well as total precipitation from the National Oceanic and Atmospheric Administrations’s Global Historical Climatology Network databse (NOAA, 2014). This dataset provides detailed weather measurements at over 20,000 weather stations across the country. We use information from 1950-2013, because we need 30 years of data prior to the period of analysis 23The typical ozone monitoring season around the country is April-September, but in fact it varies across states. Appendix Table B.1.2 reports the season for each state. In our empirical analysis we use only the common ozone season across all states, which includes the six months from April through September. 24Appendix Figure B.1.1 depicts the evolution of ambient ozone monitors over the three decades in our data, and illustrates the expansion of the network over time. Appendix Table B.1.3 provides annual summary statistics on the ozone monitoring network. The number of monitors increased from 1,361 in the 1980s to 1,851 in the 2000s. The number of monitored counties also grew from 585 in the 1980s to 840 in the 2000s. While Muller and Ruud (2018a) find that compliance with the NAAQS for ambient ozone is not consistently associated with network composition, Grainger, Schreiber and Chang (2019) provide evidence that local regulators do avoid pollution hotspots when siting new ozone monitors. Later, as a robustness check, we show qualitatively similar results for a semi-balanced panel of ozone monitors. 57 to construct a moving average measure of climate.25 The weather stations are typically not located adjacent to the ozone monitors. Hence, we match ozone monitors to nearby weather stations using a straightforward procedure.26 B. Basic Trends in Pollution, Attainment Status, and Weather: Implications for the Importance of Regulations To give a sense of the data, Figure 2.2 illustrates the evolution of ozone concentrations and the proportion of counties in nonattainment over our sample period, while Figure 2.3 does the same for our two components of daily temperature – climate norms and weather shocks. Ozone concentrations and nonattainment designations. Figure 2.2, Panel A, depicts the annual average of the highest daily maximum ambient ozone concentration recorded at each monitor from 1980-2013 in the United States. The sample is split according to whether counties were in or out of attainment with the NAAQS for ambient ozone. Counties out of compliance with the NAAQS experienced, on average, a steeper reduction in the daily maximum ozone levels than counties in compliance.27 Figure 2.2, Panel B, shows that as ambient ozone concentrations fell, the number of counties out of attainment also declined. Notice that when the 1997 NAAQS revisions were implemented in 2004 after litigation, the share of counties out of attainment increased more than 50 percent. Such a jump is not observed in the implementation of the 2008 revision, however. In the latter case, the share of counties in nonattainment remained stable around 30 percent. Appendix Figure B.1.5 shows that most counties out of attainment were 25Appendix Figure B.1.2 presents the yearly temperature fluctuations and overall trend in climate for the contiguous US as measured by these monitors, relative to a 1950-1979 baseline average temperature. 26Using information on the geographical location of ozone monitors and weather stations, we calculate the distance between each pair of ozone monitor and weather station using the Haversine formula. Then, for every ozone monitor we exclude weather stations that lie beyond a 30-km radius. Moreover, for every ozone monitor we use weather information from only the closest two weather stations within the 30-km radius. Appendix Figure B.1.3 illustrates the proximity of our final sample of ozone monitors to these matched weather stations. Once we apply this procedure, we exclude ozone monitors that do not have any weather stations within 30km. As will be discussed later, our results do not seem sensitive to these choices. 27Appendix Figure B.1.4 further compares similar trends in ozone levels with the updated 1997, 2008, and 2015 NAAQS levels which, while much lower, are based instead on the observed 4th highest 8-hour average ambient ozone concentration. 58 first designated in nonattainment in the 1980’s. The map displays concentrations of those counties in California, the Midwest, and in the Northeast. Nevertheless, a nontrivial number of counties went out of attainment for the first time in the 1990’s and 2000’s. Figure 2.2: Evolution of Maximum Ozone Concentration and Counties in Nonattainment 1979 NAAQS 80 100 120 140 160 Daily Maximum Ozone Concentration (ppb) 1980 1990 2000 2010 Attainment Counties Nonattainment Counties Panel A. Maxiumum Ozone Concentration by Attainment Status 1979 NAAQS Implemented 1997 NAAQS Implemented 2008 NAAQS Implemented 0 .2 .4 .6 .8 1 Share of Monitored Counties in Nonattainment 1980 1990 2000 2010 National Average Panel B. Proportion of Monitored Counties in Nonattainment Notes: This figure displays the evolution of maximum ambient ozone concentrations in the United States over the period 1980-2013 and the evolution of the proportion of counties violating the ambient ozone standards among the counties with ozone monitors. Panel (A) depicts daily maximum 1-hour ambient ozone concentrations over time (annual average), split by counties designated as in- or out- of attainment under the National Ambient Air Quality Standards (NAAQS). The 1979 NAAQS for designating a county’s attainment status was based on an observed 1-hour maximum ambient ozone concentration of 120 parts per billion (ppb) or higher. Here we contrast this attainment status cutoff with the maximum yearly ozone concentrations of attainment and nonattainment counties. Appendix Figure B.1.4 further compares these heterogeneous trends in ozone levels with the updated 1997 (implemented in 2004 due to lawsuits), 2008, and 2015 NAAQS levels. Panel (B) depicts the share of monitored counties that were out of attainment with the NAAQS for ozone during each year of our sample period. As can be clearly seen, this proportion has declined over time as the NAAQS regulations took effect. Also, observe that the policy change in 2004 resulted in many additional counties falling out of attainment, indicating that there was a nontrivial number of counties with ozone levels at the margin of nonattainment. Decomposing temperature into long-run climate norms and short-run weather shocks. In order to disentangle variation in weather versus climate, we decompose average temperature into a climate norm – a 30-year monthly moving average (MA) following (WMO, 2017), and a weather shock – the daily deviation from the norm.28 Figure 2.3, Panel A, plots the annual 28Our decomposition of meteorological variables into a 30-year moving average (norms) and deviations from it (shocks) is a data filtering technique to separate the “signal” from the “noise.” This should not be 59 average of the 30-year MA in the dotted line, as well as a smoothed version of it in the solid line; note that due to the nature of the MA, this takes into account information since 1950. Panel B plots the annual average of the shocks. Notice that the average deviations from the 30-year MA are bounded around zero, with bounds relatively stable over time, suggesting little changes in the variance of the climate distribution.29 Using our final sample, not surprisingly Appendix Figure B.1.7 shows that ambient ozone is closely related to both components of temperature, which we examine more formally in the empirical analysis. confused with a moving-average model of climate change. We average temperature over 30 years because it is how climatologists usually define climate normals, though other filtering techniques could be used. Interestingly, when we run robustness checks regarding a potential measurement error of the temperature norm related to the window of the moving average, we find suggestive evidence that 30 years is approximately where the error is minimum. In further robustness checks, we examine the sensitivity of our results to using a daily rather than monthly moving-average. 29Figure 2.3 is constructed using the comprehensive sample of NOAA weather stations in order to provide a sense of the climate norms and weather shocks that is nationally representative. Appendix Figure B.1.6 presents a similar illustration to Figure 2.3 using our final sample of weather monitors once matched to ozone monitors. Appendix Table B.1.4 reports the summary statistics for daily temperature and our decomposed variables, for each year in our sample from 1980-2013. 60 Figure 2.3: Climate Norms and Shocks Over the Period of Analysis (1980-2013) 25.95 26.00 26.05 26.10 26.15 26.20 30-year Moving Average (Temperature C°) 1980 1990 2000 2010 Panel A. Average Climate Norm Over Time -1 -0.5 0.0 0.5 1.0 1.5 Deviation from Moving Average (Temperature C°) 1980 1990 2000 2010 Panel B. Average Temperature Shock Over Time Notes: This figure depicts US temperature over the years in our sample (1980-2013), decomposed into their climate norm and temperature shock components. The climate norm (Panel A) and temperature shocks (Panel B) are constructed from a complete, unbalanced panel of weather stations across the US from 1950 to 2013, restricting the months over which measurements were gathered to specifically match the ozone season of April–September, the typical ozone season in the US (see Appendix Table B.1.2 for a complete list of ozone seasons by state). Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between daily observed maximum temperature and the climate norm. The solid line in Panel (A) smooths out the annual averages of the 30-year moving averages, and the horizontal dashed lines in Panel (B) highlights that temperature shocks are bounded in our period of analysis. 61 IV. Empirical Framework In the empirical analysis, we focus on estimating the extent to which ozone concentration is affected by climate change under the NAAQS regulation, relative to a benchmark without (or lower levels of) regulation. The goal is to recover dE dc R − dE dc N in Equation (2.5), the measure of regulation-induced adaptation. Thus, with an estimate of ϕ ′ λ , the marginal damage of ozone pollution, from the literature (e.g., Deschenes, Greenstone and Shapiro, 2017), we are able to provide some back-of-the-envelope calculations regarding welfare changes. We build upon a unifying approach to estimating climate impacts (Bento et al., 2020) which bridges the two leading approaches of the climate-economy literature identifying both weather and climate impacts in the same equation. Moreover, because our approach critically identifies adaptation by comparing how the same economic agents respond to both weather and climate variation, we are able to recover our measure of regulation-induced adaptation by comparing heterogeneous adaptation from counties in and out of attainment with the NAAQS for ozone without needing to make assumptions over preferences.30 In contrast, previous studies have inferred adaptation indirectly, by flexibly estimating economic damages due to weather shocks – sometimes for different time periods and locations – then assessing climate damages by using shifts in the future weather distribution predicted by climate models (e.g., Deschenes and Greenstone, 2011; Barreca et al., 2016; Auffhammer, 2018a; Carleton et al., 2019; Heutel, Miller and Molitor, forthcoming). That implies an extrapolation of weather responses over time and space, which requires preferences to be constant across those dimensions, an assumption that can be challenging for reasons similar to the Lucas Critique (Lucas, 1976). As a first step to implement our approach, we decompose the observed daily maximum temperature into a climate norm and a daily weather shock. The norm is operationalized by the 30-year monthly moving average (MA), akin to the concept of climate normals used 30In our context, adaptation could be driven by, for example, individuals responding to pollution information, firms adjusting to environmental regulation, and local regulators implementing federal laws. The estimation strategy should capture the sum of all responses together, without separating them out. 62 in climatology.31 The shock is merely the deviation of the observed daily temperature from that norm. Because ozone formation is directly tied to temperature, as discussed in Section III, the impact of temperature on ambient ozone is the focus of our analysis. Given that decomposition, we estimate the following equation: Ozoneit = β W N (T empW it × Nonattainc,y−3) + β C N (T empC im × Nonattainc,y−3) + β W A (T empW it × Attainc,y−3) + β C A (T empC im × Attainc,y−3) + Xitγ + ηis + ϕrsy + ϵit, (2.6) where i represents an ozone monitor located in county c of NOAA climate region r, observed on day t, month m, season s (Spring or Summer), and calendar year y. Our analysis focuses on the most common ozone season in the U.S. – April to September, as mentioned in the background section – over the period 1980-2013. Ozone represents daily maximum ambient ozone concentration, T empW represents the weather shock, and T empC the climate norm. Hence, the response of ambient ozone to the temperature shock β W represents the short-run effect of weather, and the response to the climate norm β C reflects the long-run impact of climate. Nonattaincy denotes nonattainment designation, which is a binary variable equals to one if a county c is not complying with the NAAQS for ambient ozone in year y. Given the structure of fixed effects described below, the identifying variation regarding attainment status is essentially “within-county variation.”32 This variable is lagged by three calendar 31To make this variable part of the information set held by economic agents at the time ambient ozone is measured, we lag it by one year. For example, the 30-year MA associated with May 1982 is the average of May temperatures for all years in the period 1952-1981. Therefore, economic agents should have had at least one year to respond to unexpected changes in climate normals at the time ozone is measured. Later, we discuss almost identical results for longer lags. Also, we use monthly MAs because it is likely that individuals recall climate patterns by month, not by day of the year. Indeed, broadcast meteorologists often talk about how a month has been the coldest or warmest in the past 10, 20, or 30 years, but not how a particular day of the year has deviated from the norm for that specific day. Later, we discuss qualitatively similar results when we use daily instead of monthly moving averages. 32Because there is variation in the timing of nonattainment designations, but we have a never treated group (the persistent attainment counties), identification can rely on the weakest parallel trends assumption considered by Marcus and Sant’Anna (2021), which does not impose any restriction on pretreatment trends across groups. In fact, when there is a “reasonably large” number of never treated units – as is the case in our setting – that assumption can identify policy-relevant parameters even “if researchers are not comfortable 63 years because EPA allows counties with heavy polluters at least three years to comply with the ozone NAAQS, as discussed in the background section. X represents time-varying control variables such as precipitation – similarly decomposed into a norm and shock. Although less important than temperature, Jacob and Winner (2009) point out that higher water vapor in the future climate may decrease ambient ozone concentration.33 η represents monitor-byseason fixed effects, ϕ climate-region-by-season-by-year fixed effects, and ϵ an idiosyncratic term.34 Standard errors are clustered at the county level.35 This approach has two key elements. The first is the decomposition of meteorological variables into two components: long-run climate norms and transitory weather shocks, the latter defined as deviations from those norms. This decomposition is meant to have economic content. It is likely that individuals and firms respond to information on climatic variation they have observed and processed over the years. In contrast, economic agents may be constrained in their ability to respond to weather shocks, by definition. As mentioned above, our measure of adaptation is the difference between those two responses by the same economic agents. In practice, we decompose temperature into a monthly moving average incorporating information from the past three decades, often referred to as climate normal, and a daily deviation from that 30-year average. This moving average is purposely lagged in the empirical analysis to reflect all the information available to individuals and firms up to, and including, the year prior to the measurement of the outcome variables.36 The second key element of our approach is identifying responses to weather shocks and with a priori ruling out nonparallel pretrends” (Marcus and Sant’Anna, 2021, p.251). 33Although temperature is the primary meteorological factor affecting tropospheric ozone concentrations, other factors such as wind and sunlight have also been noted as potential contributors. Later, we discuss qualitatively similar results for a subsample with information on wind speed and sunlight. 34In unreported analyses we examine specifications with alternative fixed effects structures, such as including latitude and longitude interacted with season-by-year, or replacing region-by-season-by-year with state-by-season-by-year. Estimates from our preferred, more parsimonious specification are similar in magnitude and significance to each of these alternatives. 35In later robustness checks we assess the sensitivity of our results to changes in our estimation of the standard errors: increasing the spatial dimension of the clustering to the state-level, adding a temporaldimension via two-way clustering by both county and week, and estimation via county-level block bootstrap. All coefficients remain statistically significant at the 1% level regardless of the choice of cluster or bootstrap. 36A graphical representation of our decomposition is illustrated for Los Angeles county in 2013 in Appendix B.1.c Figure B.1.8, and over the entire sample period of 1980-2013 in Figure B.1.9. 64 longer-term climatic changes in the same estimating equation. We are able to leverage both sources of variation in the same estimating equation because of the properties of the Frisch-Waugh-Lovell theorem (Frisch and Waugh, 1933; Lovell, 1963). The deseasonalization embedded in the standard fixed-effects approach is approximately equivalent to the construction of weather shocks as deviations from long-run norms as a first step. Furthermore, there is no need to deseasonalize the outcome variable to identify the impact of those shocks (Lovell, 1963, Theorem 4.1, p.1001).37 As a result, we do not need to saturate the econometric model with highly disaggregated time fixed effects; thus, we are able to also exploit variation that evolves slowly over time to identify the impacts of longer-term climatic changes. We exploit plausibly random, daily variation in weather, and monthly variation in climate normals to simultaneously identify the impact of weather shocks and climate change on ambient ozone concentration. Identification of the weather effect is similar to the standard fixed effect approach (e.g., Deschenes and Greenstone, 2007; Schlenker and Roberts, 2009), with the exception that because we isolate the temperature shock as a first step, we do not need to include highly disaggregated time fixed effects (Frisch and Waugh, 1933; Lovell, 1963). Identification of the climate effect relies on plausibly random, within-season monitorlevel monthly variation in lagged 30-year MAs of temperature after flexibly controlling for regional shocks at the season-by-year level.38 To better understand the identification of climate impacts, consider the following thought experiment that we observe in our data many thousands of times: take two months in 37“Theorem 4.1: Consider the following alternative regression equations, where the subscript α indicates that the data have been adjusted by the least squares procedure with D as the matrix of explanatory variables: 1. Y = Xb1 + Da1 + e1 2. Yα = Xαb2 + e2 ... 4. Y = Xαb4 + e4 ... The identity b1 = b2 reveals that inclusion of the matrix of seasonal dummy variables in the regression analysis is equivalent to working with least squares adjusted time series. The identity b2 = b4 reveals that it is immaterial whether the dependent variable is adjusted or not, provided the explanatory variables have been seasonally corrected” (Lovell, 1963). 38Because the climate norm is a constructed variable, there may be a concern that measurement error in this variable could lead to attenuation of the estimated coefficient. In later robustness checks we examine this concern by implementing alternative lengths of the moving average, finding results that are statistically and economically similar across all MA lengths. Furthermore, as noted by Solon (1992) in his examination of the effect of parents’ income on that of their child, using a longer moving-average should reduce any measurement error in this variable. 65 the same location and season (Spring or Summer). Now, suppose that one of the months experiences a hotter climate norm than the other, after accounting for any time-varying fluctuations in, e.g., atmospheric or economic conditions that affected the overarching climate region at the season-by-year level. Our estimation strategy quantifies the extent to which this difference in the climate norm affected the ozone concentrations observed on that month. Therefore, this approach controls for a number of potential time-invariant and time-varying confounding factors that one may be concerned with, such as the composition of the local and regional atmosphere, or technological progress. Furthermore, note that because the monthly climate norm is operationalized as a 30-year moving average for that month, the climate norm is “updated” from year to year as the temperature from 31 years ago drops out, and the temperature from last year enters into the moving average. This updating feature of the MA also mimics the ideal “climate experiment” by, for example, making the April climate norm in one year appear more like the May climate norm. For instance, if the average temperature in April 31-years ago was particularly cold, while the average temperature in April of last year was particularly warm, the 30-year moving average climate norm in this year’s April may be meaningfully warmer than last year’s April climate norm. In other words, we identify agents’ response to their new climate expectation using both within-season variation across months and year-to-year variation for the same month. Our ultimate goal, however, is not just to identify adaptation via estimates of climate impacts vis-`a-vis weather shocks, but to identify whether there is a different level of adaptation in nonattainment versus attainment counties. As the EPA was given substantial enforcement powers to ensure that the goals of the Clean Air Act were met, policy variation itself is plausibly exogenous conditional on observables and the unobserved heterogeneity embedded in the fixed effects structure considered in our analysis (see, e.g., Greenstone, 2002; Chay and Greenstone, 2005). In order to reach compliance, some states initiated their own inspection programs and frequently fined non-compliers. However, for states that failed to adequately enforce the standards, EPA was required to impose its own procedures for attaining com66 pliance. The inclusion of monitor-by-season fixed effects allows us to control for the strong positive association observed in cross-sections among location of polluting activity, high concentration readings, and nonattainment designations while preserving inter-annual variation in attainment status for each individual monitor. Thus, the variation used in our analysis comes from both cross-sectional differences in attainment status between counties and from changes in status within the same county over time, as previously shown in Figure 2.2: from attainment to nonattainment, or vice versa. Measuring regulation-induced adaptation. Once we credibly estimate the impact of the two components of temperature interacted with county attainment status, we recover a measure of regulation-induced adaptation. The average adaptation in nonattainment counties is the difference between the coefficients β W N and β C N in Equation (2.6). If economic agents engaged in full adaptive behavior, β C N would be zero, and the magnitude of the average adaptation in those counties would be equal to the size of the weather effect on ambient ozone concentration (for a review of the concept of climate adaptation, see Dell, Jones and Olken, 2014). Indeed, under full adaptive behavior, any unexpected increase in the climate norm would lead economic agents to pursue reductions in ozone precursor emissions to avoid an increase in ambient ozone concentration of identical magnitude to the weather effect in the same month of the following year.39 In other words, agents would respond to “permanent” changes in temperature by adjusting their production processes to offset that increase in the climate norm. Unlike weather shocks, which influence ozone formation by triggering chemical reactions conditional on a level of ozone precursor emissions, changes in the 30-year MA should affect the level of emissions. We can measure adaptation in attainment counties in the same way: (β W A − β C A ). This adaptation could arise from technological innovations, market forces, or regulations other 39Again, later we consider cases where economic agents can take a decade or two to adjust. Because EPA may give counties with heavy emitters up to two decades to comply with the ozone NAAQS, as discussed in the background section, adaptive responses many years after agents observe changes in climate norms may be plausible. Interestingly, we will find almost identical results. 67 than the NAAQS for ambient ozone.40 Sources of this type of adaptation would be, for example, the adoption of solar electricity generation, which reaches maximum potential by mid-day, when ozone formation is also at high speed, or other existing policies and regulations that have interactions with both ozone and climate, such as incentives to adopt low or zero emissions vehicles, which may reduce precursor emissions during rush-hours when ozone formation is typically at its highest.41 Once we have measured adaptation in both attainment and nonattainment counties, we can express adaptation induced by the NAAQS for ambient ozone matching Equation (2.5) as the difference: RIA ≡ (β W − β C ) | {z } dE/dc × (1N − 1A) | {z } ∆R = (β W N − β C N ) − (β W A − β C A ). (2.7) Because our RIA measure is analogous to a difference-in-differences parameter, it must satisfy a parallel trends assumption on the estimates of adaptation for nonattainment and attainment counties. To provide suggestive evidence supporting that assumption, we re-run equation 2.6 for sub-samples based on attainment status to examine pre-trends, as well as other outcomes that capture key dimensions of local economic activity – employment and wages.42 We will show that the corresponding RIAs for these alternative outcomes are precise zeros. An important advantage of this approach is to have all those coefficients estimated in the same equation. Hence, we can straightforwardly run a test of this linear combination to obtain a coefficient and standard error for the measure of regulation-induced adaptation 40Indeed, EPA mandates “best available control technology” (BACT) to curb emissions of local pollutants from large point sources even in attainment counties. As mentioned earlier, EPA mandates the more stringent “lowest achievable emissions rate” (LAER) technology in nonattainment counties. Abatement costs are considered in formulating BACT standards, but not LAER standards. 41For regulatory purposes, all the EPA considers is the observed measurement of ozone concentration by the pollution monitor, irrespective of weather conditions. It is important to mention, however, that EPA considers weather conditions when determining trends in ozone concentrations. In fact, EPA uses statistical models to adjust for the variability in seasonal ozone concentrations due to weather to provide a more accurate assessment of the underlying trend in ozone caused by emissions (see https://www.epa.gov/airtrends/trends-ozone-adjusted-weather-conditions). 42One could also think of other pollutants such as particulate matter (PM), but as emphasized by Jacob and Winner (2009), temperature seems to play a minor role in ambient PM concentration. 68 (RIA), and proceed with statistical inference. Note that while in our study context we exploit daily variation in weather and monthly variation in climate norms, the empirical strategy is general and can be applied to any study context that meets the following conditions: First, the weather shock should be at a temporal frequency in which agents have limited opportunities to adapt, ideally at the same temporal frequency as the outcome of interest. Second, the climate norm should be at the temporal frequency that agents would think about climatic changes triggering adjustments that would affect the outcome of interest, and needs to be weakly longer than the weather shock. The climate norm should be lagged such that agents have time to internalize any climatic shifts and make corresponding adjustments. Recall that while the contemporaneous weather shock may affect the outcome variable through a number of potential channels, prior years’ climate normal temperature can only impact the current time period’s outcome variable through permanent changes, which include adaptation. Third, the temporal frequency of the fixedeffects must be longer than the climate norm in order to maintain variation in the norm. Finally, the policy or regulation of interest must have heterogeneity in its implementation across time and/or space, i.e., turning on or off across different regions or at different times periods. Among many possible applications in, e.g., agriculture, wildfire management, or even tourism, consider the two examples of vaccine provision and law enforcement that we posed previously. For law enforcement, the outcome might be the number of dispatch calls, measured daily, or even hourly, depending on available data. The temperature shock could thus reflect the observed temperature at the same frequency, where both individuals and law enforcement may otherwise be limited in their ability to respond to temperature shocks. Meanwhile, the climate norm may reflect the norm for the respective month (lagged by, e.g., one year), corresponding to the temporal frequency at which agents may remember climate normal temperatures. The respective temporal granularity of the fixed-effects structure could thus be at the seasonal level. Finally, the policy could be, e.g., some exogenous shift 69 in law enforcement budget, or change in legal landscape, that might affect law enforcement agencies’ ability to respond to reported crimes. Alternatively, in the context of influenza vaccine provision, the outcome may be the number of vaccines administered weekly (or monthly), while the temperature shock would reflect the average weekly (or monthly) temperature, and the norm may reflect this same, or somewhat longer, temporal frequency – lagged by 1-year. Intuitively, large decreases in temperature may trigger agents to get their yearly flu shot, and moreover agents may internalize historical seasonality in when this shift occurs, e.g., associating it with the first week of October, middle of November, or whenever would happen to correspond to their local region’s climate norms. As there is typically only one flu season per year, in the winter, the fixed-effects structure might then encompass the 12-month period from July through June of the following year. Finally, the policy may be some exogenous shift in the level of vaccine provision – e.g., increasing the level of outreach, the number of individuals who are eligible, or decreasing the cost of receiving the vaccine. V. Results As discussed, our ultimate goal is to use Equation (2.6) to recover empirical estimates of the coefficients β W N , β C N , β W A , and β C A in Equation (2.7), which we can then incorporate into Equation (2.5) to recover back-of-the-envelope calculations of the welfare impacts of regulationinduced adaptation under various climate scenarios. Thus, we begin by presenting our main econometric findings on the impacts of temperature on ambient ozone concentration, average adaptation, and adaptation induced by the existing NAAQS regulation under the Clean Air Act. We then discuss the robustness of our results when accounting for coinciding input regulations on ozone precursors, as well as considering the distance of ozone concentrations from the NAAQS threshold. Following this, we discuss a number of additional robustness checks regarding the measurement of climate, alternative timings for economic agents to process changes in climate and engage in adaptive behavior, and further specification checks and 70 sample restrictions. Then, we examine heterogeneity in our recovered measure of adaptive response over time and across the temperature distribution, as well as by local (county-level) factors such as belief in climate change or precursor-limited ambient atmosphere. Finally, we map our econometric results into the analytical framework developed in Section II to estimate the welfare effects of regulation-induced adaptation due to the ozone NAAQS. A. The Role of Regulations for Inducing Adaptation to Climate Change Table 2.1 reports our main findings on the role of existing government regulations and policy in inducing climate adaptation. Before discussing the ozone NAAQS regulation-induced adaptation, we present the average climate impacts and adaptation across all counties in our sample. For this purpose, we run a simplified version of Equation (2.6), where the temperature shock and norm are not interacted with attainment status. Column (1) shows that a 1◦C temperature shock increases average daily maximum ozone concentration by about 1.65ppb. This can be seen as a benchmark for the ozone response to temperature because of the limited opportunities to adapt in the short run.43 A 1◦C-increase in the 30-year MA, lagged by one year and thus revealed in the year before ozone levels are observed, increases daily maximum ozone concentration by about 1.16ppb, an impact that is significantly lower than the response to a 1◦C temperature shock, indicating adaptive behavior by economic agents. Indeed, column (3) presents the measure of adaptation – 0.49ppb – which is economically and statistically significant. If adaptation was not taken into consideration, the impact of temperature on ambient ozone would be overestimated by roughly 42 percent. The estimates above represent average treatment effects. Because we are interested in the role of regulations in potentially affecting adaptive behavior, we estimate heterogeneous treatment effects by attainment status, as specified in Equation (2.6). Table 2.1, column 43We see it as a benchmark because we assume that economic agents are not able to respond to weather shocks. In reality, there might be some opportunities to make short-run adjustments in the context of ambient ozone. Although developed countries have usually not taken drastic measures to attenuate unhealthy levels of ambient ozone because concentrations are generally low, developing countries have often constrained operation of industrial plants and driving in days of extremely high levels of ozone. 71 Table 2.1: Climate Impacts on Ambient Ozone and Adaptation Daily Max Ozone Levels (ppb) Implied Adaptation (1) (2) (3) (4) Temperature Shock 1.648*** (0.058) Climate Norm 1.161*** 0.487*** (0.049) (0.036) Nonattainment x Shock 1.990*** (0.079) Nonattainment x Norm 1.351*** 0.639*** (0.067) (0.054) Attainment x Shock 1.263*** (0.027) Attainment x Norm 0.956*** 0.308*** (0.035) (0.029) Regulation Induced 0.332*** (0.056) Nonattainment Control Yes Yes Precipitation Controls Yes Yes Fixed Effects: Monitor-by-Season Yes Yes Region-by-Season-by-Year Yes Yes Observations 5,139,529 5,139,529 R2 0.428 0.434 Notes: This table reports our main findings regarding the climate impacts on ambient ozone concentrations (in parts per billion – ppb) over the period 1980-2013, as well as the implied estimates of adaptation, in particular regulation-induced adaptation. Column (1) reports climate impact estimates (national average), with daily temperature decomposed into climate norms and temperature shocks. In column (2) we interact the climate norm and temperature shock with indicators for whether counties have been designated as inor out- of attainment under the National Ambient Air Quality Standards (NAAQS) for ambient ozone, to estimate heterogeneous effects across attainment and nonattainment counties, as specified in Equation (2.6). The attainment status is lagged by 3 years, because EPA allows at least this time period for counties to return to attainment levels. The last two columns report our adaptation estimates. By comparing the impacts of climate norm and temperature shock from column (1), we obtain our estimate of overall adaptation in column (3). Similarly, in column (4) we report the adaptation in attainment and nonattainment counties separately, which we obtain by comparing the impacts of climate norm and temperature shock reported in column (2). As defined in Equation (2.7), the difference between adaptation in nonattainment and attainment counties is our measure of regulation-induced adaptation. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 72 (2), reports the estimates disaggregated by whether the ozone monitors are located in attainment or nonattainment counties. Given that attainment counties have cleaner air by definition, on average the ozone response to temperature changes in these counties is significantly lower than for nonattainment counties. However, as shown in column (4), adaptation in nonattainment counties is over 107 percent larger than in attainment counties. Specifically, adaptation in nonattainment counties reduces the impact of a 1◦C increase in temperature on ambient ozone concentration by 0.64 parts per billion (ppb), or about one-third of the total impact. As defined in Equation (2.7), the difference between adaptation estimates in nonattainment and attainment counties – 0.33ppb – is our measure of regulation-induced adaptation, shown at the bottom of column (4), which represents just over half of the total adaptation in nonattainment counties. Therefore, a regulation put in place to correct an externality – the NAAQS for ambient ozone – generates a co-benefit in terms of adaptation to climate change, on top of the documented direct impact on ambient ozone concentrations (Henderson, 1996). Recall that for tractability, our analytical framework focuses mainly on climate adaptation that may be induced by the existing regulation of interest, and is agnostic about the real-world magnitude of dE dc N – any adaptation that is plausibly exogenous to the regulation. That is, while the framework shows that we should expect induced adaptation in attainment counties to be zero, that does not mean that the total level of adaptation in those counties is zero. Thus, recovering a baseline measure of “non-induced” adaptation – that which occurs in attainment counties – is a key feature of our econometric approach, allowing us to “difference-out” the adaptation in nonattainment counties that is plausibly exogenous to the NAAQS regulation.44 Specifically, the second estimate in column (4) – 0.31 ppb – indicates that adaptive behavior is in fact present in attainment counties. The underlying reasons 44One may worry that attainment and nonattainment counties could be systematically different in ways that are not fully controlled for by the included set of fixed-effects. Later, as a robustness check, we examine the results of our main specification estimated on two alternative sub-samples: one in which we include only those counties that were consistently in or out of attainment throughout the entire sample period, and another in which we instead include only those counties that switched attainment status at least once during the sample period. In both cases the results are similar to our full-sample estimates. 73 might be technological innovation and market forces, as highlighted in previous studies (e.g., Barreca et al., 2016), other regulations affecting both attainment and nonattainment counties (e.g., Auffhammer and Kellogg, 2011; Deschenes, Greenstone and Shapiro, 2017), or even preventive responses in counties with ozone readings near the threshold of the NAAQS for ambient ozone, as examined in our robustness checks below. An example of adaptation triggered by innovation, market forces, and other regulations in the context of ambient ozone arises from the adoption of solar panels for electricity generation. Higher temperatures lead to more ozone formation, but they also constrain the operations of coal-fired power plants. Regulations under the Clean Water Act restrict the use of river waters to cool the boilers when water temperature rises (e.g., McCall, Macknick and Hillman, 2016). Because coal plants are important contributors of VOC and NOx emissions, those constraints lead to a reduction in the concentration of ozone precursors. At the same time, solar panels are more suitable for electricity generation in hotter areas, with higher incidence of sunlight; thus, more extensively used in those places. Now, higher temperatures combined with lower levels of ozone precursors – enabled by the adoption of solar panels – may lead to lower levels of ambient ozone. Hence, adaptation driven by innovation, market forces, and regulations other than the ozone NAAQS. B. Robustness Checks Parallel-trends and estimates of firm responses to climatic changes. The measure of regulationinduced adaptation (RIA) recovered by our main specification is analogous to a differencein-differences parameter, as the difference between adaptation, which is itself the difference between the weather and climate responses, in counties designated either in attainment or nonattainment. Thus, an important condition for identifying RIA is parallel pre-trends prior to counties’ nonattainment designations. We investigate this assumption via two different approaches. First, by re-estimating equation (2.6) with three alternative sample restrictions: (i) including only counties with a persistent NAAQS designation across the entire sample 74 period – i.e., always either in attainment or nonattainment; (ii) including only counties that had their NAAQS designation switched at least one time – i.e., from attainment to nonattainment, or vice-versa; and (iii) including counties that were persistently in attainment, as well as only the periods of attainment for counties that were ever in nonattainment. Second, by re-estimating equation (2.6) for other county-level outcomes that capture key dimensions of local economic activity – monthly employment and quarterly wages.45 Results reported in Table 2.2 correspond to the first three sample restrictions in columns (1) through (3), and the two alternative outcomes in columns (4) and (5). Across both sub-samples reported in columns (1) and (2), the estimate of RIA is statistically indistinguishable from our full-sample estimate, suggesting that our central result is not driven by a differential response in a sub-sample of counties. Results reported in column (3) correspond to a more explicit test of pre-trends. While the ozone response to weather and climate does appear to have a level difference between the persistent attainment counties and the attainment periods of “ever nonattainment” counties, the estimate of regulation-induced adaptation is small in magnitude and statistically indistinguishable from zero, indicating similar pre-trends between both sets of counties. 45Note that while our main specification makes use of daily, monitor-level, observations, because these alternative outcomes are measured at the county level, and at a longer temporal frequency, we first construct average values of each independent variable at the county level and corresponding temporal frequency to each outcome variable of interest. 75 Table 2.2: Parallel Trends & Alternative Outcomes Only Counties Only Counties Persistent Attainment Counties Alternative Outcomes With Persistent That Switched vs Attainment Periods of Employment Wages NAAQS Status NAAQS Status Counties Ever in Nonattainment (Log) (Log) (1) (2) (3) (4) (5) Nonattainment x Shock 1.948*** 1.996*** 1.434*** −0.002 0.004* (0.115) (0.083) (0.041) (0.001) (0.002) Nonattainment x Norm 1.270*** 1.404*** 1.025*** 0.002*** −0.002 (0.137) (0.071) (0.044) (0.000) (0.001) Attainment x Shock 0.970*** 1.444*** 0.973*** 0.000 −0.001 (0.027) (0.042) (0.027) (0.001) (0.001) Attainment x Norm 0.489*** 1.168*** 0.490*** 0.001*** 0.000 (0.034) (0.046) (0.034) (0.000) (0.001) Implied Adaptation Nonattainment 0.678*** 0.593*** 0.483*** 0.000 0.000 (0.098) (0.056) (0.034) (0.001) (0.002) Attainment 0.480*** 0.276*** 0.409*** −0.001 −0.001 (0.034) (0.037) (0.039) (0.001) (0.002) Regulation Induced 0.198* 0.317*** 0.074 0.001 0.001 (0.104) (0.058) (0.051) (0.001) (0.001) Observations 1,122,101 4,017,428 2,455,854 84,423 28,390 R2 0.352 0.445 0.394 0.996 0.972 Notes: This table reports the results of alternative sample restrictions in columns (1) through (3) and alternative outcome variables in columns (4) and (5) to examine the parallel trends assumption. Column (1) includes only counties with a persistent NAAQS status across the entire sample period – 497 attainment and 51 nonattainment counties. Column (2) only includes counties that switched their NAAQS status at least once during the sample period – 458 counties. Column (3) compares the 497 counties persistently in attainment with the periods of attainment for the 458 counties that ever switched attainment status. Finally, columns (4) and (5) report the effects of temperature shocks and changes in the climate norm on monthly log employment and quarterly log wages at the county level for all estimating sample counties, 1990-2013. The lack of response implies that the main channel for RIA, and adaptation in general, is likely “in-place” behavioral or production adjustments, rather than, e.g., shifts in production location. The full list of controls are the same as in the main model, depicted in column (2) of Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 76 Finally, results reported in columns (4) and (5) reveal differences between attainment and nonattainment counties with respect to both employment and wages that are precise zeros, further suggesting that the two groups of counties satisfy the parallel trends assumption. In other words, because employment and wages are not responding to the interactions of attainment status with weather and climate in the same way as ozone, the coefficients in our central specification can be reasonably interpreted as causal moderators – how attainment status may affect the marginal impact of weather and climate on ozone formation.46 Furthermore, although Henderson (1996) and Becker and Henderson (2000) have shown that manufacturing plants may relocate in response to an ozone nonattainment designation, our results in Table 2.2 show that county-level employment and wages do not respond differentially to changes in climate across attainment and nonattainment counties, implying that our central estimate of RIA is driven by “in-place” behavioral or production adjustments, rather than permanent or transitory shifts in production location. Estimates considering input regulations for ozone precursors. During our period of analysis (1980-2013), three other policies aiming at reducing ambient ozone concentrations were implemented in the United States: (i) regulations restricting the chemical composition of gasoline, intended to reduce VOC emissions from mobile sources (Auffhammer and Kellogg, 2011), (ii) the NOx Budget Trading Program (Deschenes, Greenstone and Shapiro, 2017), (iii) the Regional Clean Air Incentives Market (RECLAIM) NOx and SOx emissions trading program (Fowlie, Holland and Mansur, 2012). Notably, as these were all input regulations on ozone precursor emissions, which lack explicit climate interactions themselves, our theoretical framework suggests that they should have no impact on adaptation (see Appendix B.3.b for further discussion and a proof of this extension). However, because our goal is to econometrically recover an empirical estimate of climate adaptation induced specifically by 46Conversely, if employment or wages were responding to the interactions of attainment status with weather and climate, we would be uncovering effect moderators, where the coefficients would be capturing both the effect of attainment/nonattainment status and any other factors that could be correlated with this status while also moderating the ozone-temperature relationship. 77 the NAAQS for ambient ozone, it is imperative to examine the sensitivity of our estimates of regulation-induced adaptation when taking into account these input regulations targeted at ozone precursors. Auffhammer and Kellogg (2011) demonstrate that the 1980s and 1990s federal regulations restricting the chemical composition of gasoline, intended to curb VOC emissions, were ineffective in reducing ambient ozone concentration. Since there was flexibility regarding which VOC component to reduce, to meet federal standards refiners chose to remove compounds that were cheapest, yet not so reactive in ozone formation. Beginning in March 1996, California Air Resources Board (CARB) approved gasoline was required throughout the entire state of California. CARB gasoline targeted VOC emissions more stringently than the federal regulations. These precisely targeted, inflexible regulations requiring the removal of particularly harmful compounds from gasoline significantly improved air quality in California (Auffhammer and Kellogg, 2011). Therefore, we re-estimate our analysis removing the state of California from 1996 onwards. The results reported in Table 2.3 reveal that the estimate for regulation-induced adaptation in column (2), derived from column (1) estimates of the impact of temperature shocks and norms on ambient ozone concentration, is remarkably close to our overall estimate of regulation-induced adaptation. Hence, it appears that VOC regulations in California do not drive our estimate of climate adaptation induced by the NAAQS for ozone, in line with our theoretical framework’s predictions regarding such input regulations. 78 Table 2.3: Accounting for Competing Input Regulations Aimed at Ambient Ozone Reductions VOC Regulations NOx Regulations VOCs and NOx (Exclude California) (Exclude NBP States) (Exclude RECLAIM) (Exclude OTR States) Ozone (ppb) Adaptation Ozone (ppb) Adaptation Ozone (ppb) Adaptation Ozone (ppb) Adaptation (1) (2) (3) (4) (5) (6) (7) (8) Nonattainment x Shock 2.032*** 2.050*** 1.987*** 1.988*** (0.092) (0.090) (0.082) (0.098) Nonattainment x Norm 1.370*** 0.662*** 1.430*** 0.620*** 1.320*** 0.667*** 1.359*** 0.629*** (0.061) (0.064) (0.080) (0.062) (0.055) (0.061) (0.083) (0.066) Attainment x Shock 1.275*** 1.267*** 1.263*** 1.260*** (0.028) (0.031) (0.027) (0.029) Attainment x Norm 0.970*** 0.305*** 0.978*** 0.290*** 0.946*** 0.317*** 0.968*** 0.292*** (0.034) (0.028) (0.041) (0.034) (0.033) (0.029) (0.039) (0.032) Regulation Induced 0.358*** 0.331*** 0.349*** 0.337*** (0.065) (0.063) (0.062) (0.067) Observations 4,631,413 4,338,183 5,008,323 4,437,345 R2 0.432 0.443 0.439 0.438 Notes: This table reports results from our main specification when excluding locations with competing input regulations aimed at reducing ambient ozone concentrations via reductions in ozone precursors (VOCs and NOx). Column (1) excludes California from 1996 onwards, when stringent VOC regulations were in place. Column (3) excludes the states participating in the NBP from 2003 onwards, when the program was in effect. Column (5) excludes the four California counties within the South Coast Air Quality Management District from 1994 onwards, when the RECLAIM was in operation. Column (7) excludes all states within the OTR from 1993 onwards, when it went into effect. The implied adaptation estimates presented in columns (2), (4), (6), and (8) are derived from the estimates reported in columns (1), (3), (5), and (7) respectively. The full list of controls are the same as in the main model, depicted in column (2) of Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 79 Deschenes, Greenstone and Shapiro (2017) and Fowlie, Holland and Mansur (2012) both find a substantial decline in air pollution emissions and ambient ozone concentrations from the introduction of an emissions market for nitrogen oxides (NOx), another ozone precursor. The NOx Budget Trading Program (NBP) examined by Deschenes, Greenstone and Shapiro (2017) operated a cap-and-trade system for over 2,500 electricity generating units and industrial boilers in the eastern and midwestern United States between 2003 and 2008. Thus, we re-estimate our analysis excluding the states participating in the NBP, from 2003 onwards.47 The RECLAIM NOx and SOx trading program examined by Fowlie, Holland and Mansur (2012) similarly operated a cap-and-trade system at 350 stationary sources of NOx for the four California counties within the South Coast Air Quality Management District (SCAQMD) starting in 1994. Thus, we again re-estimate our analysis, excluding the SCAQMD counties from 1994 onwards.48 Table 2.3 reports the results excluding NBP states in columns (3) and (4), and excluding RECLAIM counties in columns (5) and (6). The estimate for regulation-induced adaptation in columns (4) and (6) are quite similar to our overall estimate of regulation-induced adaptation. Despite being effective in reducing NOx and ozone concentrations, the NBP and RECLAIM programs do not seem to affect climate adaptation induced by the NAAQS for ozone. Again, this is in line with our theoretical framework’s predictions regarding such input regulations and thus not surprising. In addition to these three policies, the CAA amendments of 1990 designated many states in the northeastern United States as part of an Ozone Transport Region (OTR). Within this region, even attainment counties were required to act to reduce emissions of NOx and VOCs (USCFR, 2013). Similar to the three cases above, we re-estimate our analysis excluding the states that were designated as part of the OTR starting from 1993 – when the affected 47NBP participating states include: Alabama, Connecticut, Delaware, Illinois, Indiana, Kentucky, Maryland, Massachusetts, Michigan, Missouri, New Jersey, New York, North Carolina, Ohio, Pennsylvania, Rhode Island, South Carolina, Tennessee, Virginia, and West Virginia, and Washington, DC. The NBP operated only in northeastern states on May 1 of 2003, and expanded to the other states on May 31 of 2004 (Deschenes, Greenstone and Shapiro, 2017). 48Participating counties include: Los Angeles, Riverside, San Bernardino, and Orange. 80 states’ implementation plans (SIP) had been revised to include all areas in the OTR.49 Table 2.3 reports the results excluding OTR states in columns (7) and (8). The estimate for regulation-induced adaptation in column (8) are once again quite similar to our overall estimate of regulation-induced adaptation amd in line with our theoretical framework’s predictions regarding such input regulations. These estimates have the added benefit of addressing a separate potential concern: cross-county adaptation spillovers. Theoretically, adaptation efforts in a nonattainment county could reduce the pollution in a neighboring attainment county. This would imply a higher level of adaptation in the attainment county than occurred, leading to a downward bias in the estimate of RIA. This potential concern would be most pronounced in areas where pollution is likely to transport across county boundaries – for example, in the OTR. As we find no statistically significant difference between our central estimate of RIA and the estimate when excluding OTR states, this suggests that cross-county spillover effects, should they exist, are not of meaningful magnitude. Estimates by distance of ozone concentrations to NAAQS threshold. One may ponder that the ideal setting to identify regulation-induced adaptation would be to randomly assign regulation, and compare the impact of climatic changes in regulated versus unregulated jurisdictions. Nevertheless, this would work only if the regulation was unanticipated and imposed only once. If regulations are anticipated, and can be assigned multiple times, in multiple rounds, such as the Clean Air Act nonattainment designations, economic agents may respond more similarly to the threat of regulation, even when it is randomly assigned. They might be indifferent between making adjustments before or after being affected by the regulation if more rounds of regulatory action are on the horizon. The intuition for these results is similar to the outcomes of finitely versus infinitely repeated games (or games that are being repeated an unknown number of times). Consider the prisoner’s dilemma game. 49Affected states include: Connecticut, Delaware, Maine, Maryland, Massachusetts, New Hampshire, New Jersey, New York, Pennsylvania, Rhode Island, Vermont, and the Consolidated Metropolitan Statistical Area that includes the District of Columbia. For the latter, we include both DC and the entire state of Virginia, as the amended SIP may have affected statewide policy. 81 If played a finite number of times, defection may yield higher payoffs, following familiar backward-induction arguments. But if played an infinite (or an unknown) number of times, cooperation may emerge as a preferable outcome. In the case of the Clean Air Act, EPA designates counties out of compliance with NAAQS if their pollution concentrations are above a known threshold. Such designations may change over time depending on the adjustments made by economic agents in those jurisdictions. For counties whose pollution concentration is around the threshold, economic agents may have incentives to make efforts to comply with NAAQS no matter whether those counties are just above or just below the threshold. If counties are even a little above the standards, EPA mandates them to adopt emissions control technologies and practices to reduce pollution, which is costly. If counties are a little under the standards, they may want to keep it that way to avoid regulatory oversight. As a result, they may end up making efforts to maintain the area under attainment. This somewhat similar adaptive behavior around the ozone standards may reduce the estimates for regulation-induced adaptation near the NAAQS threshold.50 Table 2.4 reports estimates recovered by interacting our main specification with monitorlevel indicators for whether the daily ozone concentration fell within 20 percent, above or below, the NAAQS threshold in Panel A, between 20-40 percent away from the threshold in Panel B, and over 40 percent away from the threshold in Panel C.51 The observations within 50It is important to mention that before the 1990 CAA amendments, EPA used a “too close to call” nonattainment category with minimal requirements for areas just violating the NAAQS. Areas in this category (with ozone levels up to 138ppb, hence above the threshold of 120ppb) were not subject to full SIP requirements, but rather watched closely to see if their air quality was getting worse (Krupnick and Farrell, 1996). This malleability in enforcement may also reduce the estimate for regulation-induced adaptation near the NAAQS threshold. 51Recall that the EPA changed the criteria for designating a county out of attainment in 1997 (implemented in 2004 after litigation) and again in 2008 to use the 4th highest 8-hour concentration level – 80 ppb and 75 ppb respectively – rather than the 1st highest 1-hour concentration level of 120 ppb. In our analysis we compare the 1st highest 1-hour concentration level, our outcome of interest, against the prevailing NAAQS threshold for constructing these daily indicators. As noted by the EPA, the 4th highest 8-hour concentration of 80 ppb should approximate the 1st highest 1-hour concentration level of 120 ppb. Thus in unreported analyses we also consider estimates where we use only the 1-hour 120 ppb threshold in constructing the indicator variables, and estimates where we use only the observations prior to the NAAQS change in 2004. In both cases results are qualitatively similar to our preferred specification reported in Table 2.4. 82 Table 2.4: Results by Distance of Ozone Concentrations to NAAQS Threshold Panel A. Ozone (ppb) Within 20% of NAAQS Threshold Nonattainment Attainment Induced Ozone (ppb) Adaptation Ozone (ppb) Adaptation Adaptation (1) (2) (3) (4) (5) Temperature Shock 0.610*** 0.382*** (0.024) (0.014) Climate Norm 0.539*** 0.071** 0.395*** −0.013 0.084*** (0.033) (0.034) (0.017) (0.014) (0.029) Sub-sample Obs. 676,068 Panel B. Ozone (ppb) Within 20% - 40% of NAAQS Threshold Temperature Shock 0.758*** 0.300*** (0.077) (0.011) Climate Norm 0.484*** 0.274*** 0.264*** 0.036** 0.238*** (0.061) (0.036) (0.025) (0.018) (0.043) Sub-sample Obs. 1,300,386 Panel C. Ozone (ppb) Over 40% away from NAAQS Threshold Temperature Shock 1.225*** 0.772*** (0.123) (0.024) Climate Norm 0.673*** 0.552*** 0.479*** 0.293*** 0.259*** (0.063) (0.076) (0.038) (0.028) (0.089) Sub-sample Obs. 3,162,755 Observations 5,139,209 R2 0.709 Notes: This table reports results from our main specification when including interactions with indicator variables for ozone monitor readings over the period 1980-2013 with concentrations falling within 20 percent of the NAAQS threshold in Panel A, within 20-40 percent of the threshold in Panel B, and over 40 percent away from the threshold in Panel C. Note that all reported estimates for nonattainment and attainment counties reported in Columns (1) and (3) come from a single estimating equation. Columns (2) and (4) represent the implied measures of adaptation, while Column (5) reports the resulting measure of regulationinduced adaptation as their difference. The full list of controls are the same as in the main model, depicted in column (2) of Table 2.1. For reference, the 1979 NAAQS for designating a county’s attainment status was based on an observed 1-hour maximum ambient ozone concentration of 120ppb or higher, while the 1997 amendment (implemented in 2004) changed this to an observed maximum 8-hour average ambient ozone concentration of 80ppb or higher, and the 2008 update further reduced this to 75ppb. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 83 20 percent of the NAAQS threshold comprise about 13 percent of the overall sample. As expected, the empirical evidence we provide for this subset indicates limited differential adaptation across attainment and nonattainment counties, but still of nontrivial magnitude. The estimate for regulation-induced adaptation, which is the difference between the adaptation estimates in columns (2) and (4), is still economically and statistically significant. For the observations of ambient ozone concentration within 20-40 percent of the NAAQS threshold (25 percent of the overall sample), and over 40 percent away from the threshold (62 percent of the overall sample), we cannot rule out that the estimates of regulation-induced adaptation reported in column (5) are similar to our main estimate. Given that together these observations make up 87 percent of the overall sample, it is fair to say that most of the regulation-induced adaptation arises from monitors with ozone readings relatively far from the NAAQS threshold. Other robustness checks and sample restrictions. We further examine the sensitivity of our results to a host of additional robustness checks in Appendix B.2. Table B.2.1 examines the choice of a 3-year lag on counties’ nonattainment status, as the EPA may give some counties a longer deadline to reach compliance. Conversely, a 3-year lag implicitly assumes that counties which had re-entered attainment status would continue act as if they were in nonattainment for the first few years. We re-estimate equation (2.6) using a 1-year and a 6-year lag on the nonattainment indicator, finding results that are economically and statistically similar to our central results, suggesting that the choice of the 3-year lag does not meaningfully impact our estimates. Table B.2.2 varies our moving average measure of climate to investigate whether measurement error may be of concern, potentially arising from our decomposition of meteorological variables using a 30-year MA. Alternatively, there may be concern with our choice of a 1-year lagged 30-year MA in our preferred specification, implying that agents adapt within one year – or the assumption that agents are constrained to adapt in the short-run. To investigate the first concern we repeat our analysis using a 10-year and 20-year lag in place of the 1- 84 year lag, with results presented in columns (1) and (2) of Table B.2.3.52 To address the second concern we make use of a widespread “Ozone Action Day” alert policy, whereby the local air pollution authority would release a public alert, typically a day or two in advance, that meteorological conditions are expected to be especially conducive to ozone formation. To the extent that agents are adapting to contemporaneous weather shocks, we would be most likely to observe an adaptive response on these high impact days, especially considering the prior warning. Table B.2.4 explores further specification checks – using a daily rather than monthly MA, or including other meteorological controls, and sample restrictions – constraining the estimating sample to a semi-balanced panel. Furthermore, we provide results using a variety of alternative matching rules between ozone monitors and weather stations in Table B.2.5: varying the distance cut-off, the number of monitors in the matching, and the averaging procedure. Estimates in all of the above analyses are relatively stable across the alternative approaches. Lastly, recall that our standard errors are clustered at the county level. Since the 30-year MAs and temperature shocks could be considered generated regressors, we also provide standard errors block bootstrapped at the county level for our main estimates in Appendix Table B.2.6. Bootstrapped standard errors are all within 6% of those estimated via clustering at the county level. Because the changes were usually relatively minor, for simplicity we use clustered standard errors at the county level in the remainder of the analyses.53 C. Heterogeneity in Regulation-Induced Adaptation Once we have recovered a measure of regulation-induced adaptation from the differential responses to weather shocks and longer-term climatic changes in nonattainment and attainment counties, we are then able to explore heterogeneity in the degree of adaptation across 52Note that NOAA weather data only has nationwide coverage available from approximately 1950 onwards. Thus, when using a 10-year lag the MA is comprised of only 20 years, while with the 20-year lag the MA consists of only 10 years. 53Appendix Table B.2.6 also reports standard errors clustered at the state level and two-way clustered by county and week. The estimated standard error for RIA increases by up to 39%, but the coefficient remains statistically significant at the 1% level. 85 other dimensions. Specifically, we examine heterogeneity along four dimensions: across time and the temperature distribution, as well as by local belief in climate change and local atmospheric composition. Adaptation across time and temperature. So far we have demonstrated that existing government regulations and policy can be effective in inducing climate adaptation. Now, we examine these estimates by decade. As reported in Appendix Table B.2.7, the magnitude of regulation-induced adaptation in the 1980’s is marginally larger, declining somewhat in the 1990’s, and further still in the 2000’s – for all three decades, however, estimates of regulationinduced adaptation are not statistically different from our central result. Looking at the recovered coefficients for β W and β C specifically, however, reveals an interesting trend. The ozone-temperature gradient itself declines meaningfully over time in both attainment and nonattainment counties, in line with what one might expect from previous studies suggesting that the CAA may induce innovation and diffusion of pollution abatement technologies (e.g., Popp, 2003, 2006). To that extent, our results – which focus on the static adaptation induced by the NAAQS – may present a lower-bound of the total adaptation induced by the CAA which may also have dynamic elements. Examining the estimates across the temperature distribution in Tables B8a and B8b, RIA ranges between 0.182 ppb to 0.268 ppb for the three temperature bins below 30◦C, approximately doubling to 0.452 ppb in the 30-35◦C bin, and almost tripling to 0.689 ppb when above 35◦C – in line with the idea that nonattainment counties may especially focus adaptive efforts on months with the hottest days, when they would otherwise have been most likely to exceed the NAAQS threshold. Adaptation by local climate beliefs and local atmospheric composition. While the above analyses examine heterogeneity in adaptive response across time and the temperature distribution, one may wonder how adaptation varies across other dimensions, i.e., spatially, such as between areas with different climate beliefs or different underlying atmospheric conditions. 86 In the absence of direct climate policy at the national and international stage, action driven by local culture may help address the challenge of climate change (Stavins et al., 2014). At the same time, the underlying composition of precursor emissions in the local atmosphere may also play an important role. Table B9 in Appendix B.2.b examines this first point, using the results of a relatively recent county-level survey regarding residents’ beliefs in climate change (Howe et al., 2015).54 We create county-level indicators for terciles of high, medium, and low belief, and interact the indicators for high- and low-belief counties with our temperature and control variables, taking the median-belief tercile of counties as the baseline.55 Our results suggest that climate beliefs may significantly affect the level and channel of regulation-induced adaptation: high-belief counties are associated with approximately 45% higher adaptation when in nonattainment, but are no different from baseline counties when in attainment; conversely, low-belief counties are associated with approximately 44% lower adaptation when in attainment, but maintain a similar level of adaptation as baseline counties when in nonattainment.56 This could be due to, e.g., low-belief counties only engaging in adaptive behavior when forced to do so, i.e., when designated as in nonattainment, while conversely, high-belief counties may take a nonattainment designation as a call to action, engaging in greater levels of adaptation than may otherwise be necessary if simply trying to meet the NAAQS requirements for ozone concentration levels. Similarly, Table B12 in Appendix B.2.b examines the second point. Due to the Leontief54Specifically, Howe et al. (2015) develop a modelling technique to estimate local climate beliefs at a high degree of granularity using less granular survey results in combination with demographic characteristics. Their model results are externally validated against independently conducted surveys and are found to have an average margin of error of +/-8% at the county level using bootstrap and a 95% confidence interval. In either case, as we only have cross-sectional variation in beliefs, which may be correlated with other demographic and local characteristics, we interpret these results as suggestive of an effect moderator, rather than a causal moderator, on the magnitude of RIA caused by the ozone NAAQS. 55Appendix Figure B.1.10 depicts the evolution of ozone concentration for these three sets of counties from 1980-2013. While the pattern for low- and median-belief counties track quite similarly, high-belief counties began with higher ozone concentrations, on average, but have now mostly converged with the other counties. Additionally, Table B10 provides summary statistics of basic demographic characteristics across these three county groupings using data from the 2006-2010 5-year American Community Survey. 56As a placebo check on these findings, we also examine the heterogeneity in our results when separating counties into low- median- and high-belief regarding “preferences” for single-parenthood in Table B11. 87 like production function of ozone, counties may find themselves with an atmospheric composition that is “limited” in either precursor component – VOCs or NOx. We create county-level indicators, at 5-year intervals, for whether a county is, in general, VOC- or NOx-limited and interact these indicators with our temperature and control variables, taking the counties with non-limited atmosphere as the baseline. Our results suggest that while counties without a precursor-limited atmosphere still observe regulation-induced adaptation, the effect is almost quadrupled in VOC-limited counties and doubled in NOx-limited counties, though the latter is statistically imprecise. This result is perhaps unsurprising. Areas which have a local atmosphere that is already limited with regards one of the precursors may be able to focus their efforts on continuing to reduce the limiting pollutant, which would likely have a larger impact on ozone formation than similar efforts in areas where neither NOx nor VOCs are a limiting factor. D. Climate Adaptation Co-Benefits from Existing Regulations: Some Calculations Having presented our main findings, we now provide some back-of-the-envelope calculations on the co-benefits of the existing Clean Air Act associated with climate adaptation induced by the NAAQS for ambient ozone. Following the sufficient statistic approach (Harberger, 1964; Chetty, 2009; Kleven, forthcoming) as outlined in Section II, these calculations combine our main estimates from Table 2.1 with climate projections from the U.S. Fourth National Climate Assessment (Vose et al., 2017), and the social benefits of ozone reductions from Deschenes, Greenstone and Shapiro (2017). As detailed in Equation (2.7), all of these elements can be mapped directly into the components of Equation (2.5), allowing us to interpret the resulting values as welfare changes. Additionally, we also discuss how these co-benefits are affected by the projected changes in climate over the 21st century. Formally, we map each of these three “sufficient statistics” to the components of Equation (2.5), summing across every county n in the set of counties ever designated as nonattainment (NA) within our sample period: 88 1 λ ∆V ∆c ≈ −X n∈NA ϕ ′ λ |{z} DGS dE dc n |{z} T able2.1 ∆cn |{z} V ose , (2.8) where ϕ ′ λ is treated as a fixed value, approximately equal to $1.75 million (2015 US) per county per year, following Deschenes, Greenstone and Shapiro (2017). The value of ∆c varies depending on the chosen climate projection from Vose et al. (2017), while dE dc varies depending on whether, and which type, of adaptation is being calculated, following directly from our central results in columns (2) and (4) of Table 2.1. Table 2.5 presents the costs of climate change, the savings from overall adaptation, and particularly the savings from regulation-induced adaptation – the co-benefit of the ozone NAAQS. We focus on the 509 counties most affected by the NAAQS for ambient ozone (nonattainment counties), representing about two thirds of the U.S. population. The row labeled costs “without adaptation” uses the estimated effects of temperature shocks on ambient ozone – β W N – and the one labeled “with adaptation” uses the estimated impacts of changes in climate norms (lagged 30-year MAs) – β C N . These are the main results reported in Table 2.1 – the estimated coefficients for nonattainment counties from column (2). In addition, the row labeled savings “from adaptation” report the difference between the costs with and without adaptation – β W N − β C N – and the row labeled “regulation-induced adaptation” displays the portion of the adaptation due to the NAAQS for ambient ozone – RIA as in Equation (2.7). Table 2.5, column (1), reports the costs associated with increased ambient ozone, and potential savings from adaptation, from a 1◦C increase in temperature – i.e., ∆c = 1. The costs arising from additional ambient ozone amount to approximately $1.77 billion (2015 USD) per year when we use the benchmark effect of temperature shocks that do not take into account adaptation. They reduce to approximately $1.2 billion using the impact of changes in climate norms, which does incorporate adaptive behavior. The difference of $567 million per year is the total potential savings from adaptation, 52 percent of which is 8 Table 2.5: Implied Impacts of Ambient Ozone Climate Penalty Nonattainment Counties 1 ◦C Increase RCP 4.5 Scenario RCP 8.5 Scenario 2050 2100 2050 2100 (1) (2) (3) (4) (5) Costs (Millions 2015 USD/year) Without Adaptation 1,766 2,473 4,946 2,826 8,479 With Adaptation 1,199 1,679 3,357 1,918 5,755 Savings (Millions 2015 USD/year) From Adaptation 567 794 1,589 908 2,723 Regulation Induced Adaptation 294 412 824 471 1,412 Net RIA Welfare Co-Benefit 196 275 549 314 940 Notes: This table reports some back-of-the-envelope calculations on a class of co-benefits of the existing Clean Air Act regulations – climate adaptation induced by the NAAQS for ambient ozone. The calculations are derived from the main estimates in Table 2.1 and the costs associated with those climate penalties on ambient ozone in the United States, for all 509 counties ever in nonattainment in our sample, under a variety of climate scenarios. The social costs of ozone increases are inferred from the estimated willingness to pay (WTP) for a 1 ppb decrease in the mean 8-hour summer ozone concentration in the states participating in the U.S. NOx Budget Program – about $1.7 million (2015 USD) per county per year (Deschenes, Greenstone and Shapiro, 2017, p.2985, Table 6, Panel D, Column 5). Column (1) reports the impacts of a 1◦C increase in temperature, while columns (2) and (3) report the impacts under the Representative Concentration Pathway (RCP) 4.5 climate scenario at mid- and late- century. Similarly, columns (4) and (5) report the effects for mid- and late- century under the RCP 8.5 climate scenario. Increases for mid-century relative to 1976-2005 are projected to be 1.4◦C for RCP4.5 and 1.6◦C for RCP8.5. By late-century, the RCPs diverge significantly, leading to different rates of warming: approximately 2.8◦C for RCP4.5, and 4.8◦C for RCP8.5 (Vose et al., 2017, p.195). In this table, the first row reports the expected effect of the relevant temperature increase by using the estimate of temperature shock from column (2) of Table 2.1. The second row then reports what these impacts would be after including adaptation by instead using the estimate of climate norm from the same column of Table 2.1. Row three displays the implied savings, the difference between the first two rows. Further, by taking the difference between the measures of adaptation in nonattainment and attainment counties from Table 2.1, column (4), row four reports the component of these savings that can be attributed to adaptation induced by the NAAQS for ambient ozone. Finally, row 5 accounts for the fact that adaptation is unlikely to be costless. Using the EPA’s estimate of the cost of reducing the ozone NAAQS by 1ppb – $296 million per year (USEPA, 2015a) – multiplied by the 0.332ppb value of RIA, implies a cost of RIA of $98 million per year per 1◦C. Subtracting this value, scaled corresponding to each respective column, from the gross welfare benefits reported in row 4 gives an approximate estimate of the net welfare co-benefits of adaptation induced by the NAAQS regulation. 90 induced by the NAAQS for ambient ozone. The portion induced by the NAAQS represents the co-benefits of the Clean Air Act in terms of climate adaptation, and can be interpreted as additional societal welfare gains from that existing regulation, as informed by Equation (2.5). In the next four columns, all estimates are scaled up with the temperature projections from Vose et al. (2017) – e.g., ∆c = 1.4 in column (2). Regulation-induced adaptation, in particular, reaches the range of $412-471 million per year by mid-century, and $824-1,412 million by the end of the century. Adaptation, however, is typically not costless and the above estimates in Table 2.5 reflect the gross annual co-benefits of (regulation-induced) adaptation. Using the EPA’s estimated cost of strengthening the current NAAQS by 1ppb, $296 million per year USEPA (2015b), multiplied by 1/3 to reflect the 0.332 ppb reduction per 1◦C arising from RIA, suggests that RIA is associated with an approximate annual cost of $98 million. Thus, the net adaptation co-benefits of the ozone NAAQS are approximately $196 million per year per 1 ◦C, or approximately $275-314 annually by mid-century, depending on warming scenario. These are nontrivial additional welfare gains brought about by the air quality standards regarding ambient ozone. VI. Concluding Remarks Understanding whether and how we can adapt to a changing climate is essential for individuals and policymakers seeking to develop efficient climate policies. Faced with the political challenges of creating new, first-best climate policies, the urgency to address climate change, and the often slow pace and distributional implications of market-based adaptation, it may be relatively easier in the short-run to adjust existing policy to maximize adaptation cobenefits while working towards comprehensive climate policy. This study develops an analytical framework and presents the first credible estimates of regulation-induced adaptation. We develop an analytical framework to examine the interactions between climate change and existing corrective policy or regulations established 91 for reasons unrelated to climate change, revealing that when climate change would exacerbate a market failure the existing regulation or policy would trigger an adaptive response, reducing climate impacts. We then demonstrate this induced adaptation effect empirically, examining the impact of temperature changes on ambient ozone concentration in the United States from 1980-2013. Comparing the adaptive response to long-run climatic changes in temperature between counties in or out of attainment with the Clean Air Act’s National Ambient Air Quality Standard for ambient ozone reveals an adaptive response that is more than twice as large in nonattainment counties. This regulation-induced adaptation in nonattainment counties has non-trivial welfare effects, implying an additional co-benefit of the ozone NAAQS of up to $412-471 million per year by mid-century – or approximately 5-10% of the EPA’s estimated range of direct health benefits of the ozone NAAQS (USEPA, 2015b). The NAAQS for ozone is an ideal setting for examining regulation-induced adaptation, both because of its direct policy relevance and because climate change is expected to increase ozone concentrations in the near future. Thus, by highlighting an additional benefit of the NAAQS that had previously been unaccounted for, our findings may contribute to the design or revision of pollution control policy as well. However, while this analysis focuses on ozone as one instance of regulation-induced adaptation, the proposed concept and methodological approach are general and could be applied to examine a broad class of existing non-climate corrective policies with potential climate interactions.57 For example, law enforcement and the military may additionally act as buffers against climate-related crime and conflict; vaccination campaigns, such as with influenza, may additionally help to cope with more severe influenza seasons that are likely to emerge with global warming (Towers et al., 2013). Importantly, while our empirical analysis makes use of daily data and compares within-season monthly climate normals to identify the climate impact, this is not a necessary requirement of the underlying method. For example, in the context of agriculture or vaccine provision, the 57Notably, these potential adaptation co-benefits are in addition to the intended direct effects of the government policies, regulations, or provision of public goods. For example, Dell, Jones and Olken (2014) note that snowfalls that occasionally disrupts Southern U.S. states have negligible effects in the Northeast, in part because of policy-induced investments in snow removal. 92 appropriate temporal frequency of the climate norm may be longer, such as the within-year growing season or flu season. The method is flexible, allowing the researcher to adjust the parameters of the estimating equation to their specific study context’s outcome of interest and period of analysis. 93 Chapter 3: Policy Induced Defensive Investment: The Role of Information in Reducing the Healthcare Burden of Air Pollution I. Introduction Exposure to ambient air pollution can have significant impacts on physical health, particularly cardiovascular (Brook et al., 2009) and respiratory disease (Bentayeb et al., 2012), leading to increases in morbidity and mortality (Bell et al., 2004; Dominici et al., 2006; Di et al., 2017; Cohen et al., 2017; Deryugina et al., 2019). It can also affect mental health (Nguyen, Malig and Basu, 2021), cognitive function (Zhang, Chen and Zhang, 2018; Krebs and Luechinger, 2024), and worker productivity (Graff Zivin and Neidell, 2012). Many governments have enacted various regulations on pollution emissions to protect public health such as the National Ambient Air Quality Standards in the US (USEPA, 1979), as well as information-based programs alerting individuals to expected high concentration days in an effort to reduce the health impacts of these potentially high exposure days (Neidell, 2009b; Graff Zivin and Neidell, 2009). At the same time, individuals may also take it upon themselves to reduce their own exposure – or the health impacts of unavoidable exposure – by, for example, purchasing and wearing respirators (such as n95 face masks), using air purifiers, or taking pharmaceuticals that reduce negative health responses from exposure to air pollution.1 These actions are often termed avoidance behaviors, defensive investments, or 1While these reflect day-to-day marginal defensive investment options, Chay and Greenstone (2005), for example, find that reductions in local air pollution increase housing values – implying that individuals also engage in discrete defensive investments, i.e., the choice of housing location, when selecting their optimal defensive investment strategy. 94 defensive expenditures in the literature, where prior studies have typically examined distinct channels of defensive investment to infer willingness to pay (WTP) to reduce air pollution (Deschenes, Greenstone and Shapiro, 2017; Zhang and Mu, 2018; Ito and Zhang, 2020; Barwick et al., 2018) or the value of information-based air pollution programs (Barwick et al., 2024; Wang and Zhang, 2023).2 Combining the universe of Medicare beneficiary data from 2004 to 2017 with daily air pollution data from 1980 to 2017 and “air quality alert” announcements from 2004 to 2017, this paper examines the effect on health outcomes from expected and unexpected exposure to ambient ozone in the same estimating equation, allowing for straightforward inference of the effectiveness of defensive investments as the difference between the two. By simultaneously estimating the effect of air quality alerts, any reductions in health impacts caused by these alerts can be compared with this measure of defensive investments. This paper builds on two recent methodological advancements in the literature, one in climate economics regarding the analysis of climate adaptation (Bento et al., 2023), and the other in the air pollution and health literature (Deryugina et al., 2019), which proposes an instrumental variables (IV) approach using wind direction to address many of the omitted variable bias and measurement error concerns in studies of the health effects of air pollution exposure. Bento et al. (2023) show that by relying on the properties of the Frisch-Waugh-Lovell theorem (Frisch and Waugh, 1933; Lovell, 1963), the effects of short-run weather shocks and long-run climate norms can be simultaneously estimated in the same equation – with the difference reflecting the level of climate adaptation that economic agents undertake in response to expected climatic norms.3 In this same spirit, the following analysis adapts this 2Some studies have formalized theoretical bounds on the cost of non-marginal pollution exposure (Bartik, 1988) when abstracting away from specific channels of defensive investments and assuming a continuous defensive investment function that, in real world applications might consist of the combination of multiple defensive investment technologies or behavioral adjustments. Other studies have estimated bounds on the total cost of defensive investments by comparing results from separate OLS and IV regressions (Moretti and Neidell, 2011). 3Specifically, by decomposing climatic variables into a long-run “norm” component, e.g., as a multi-year moving average at the monthly level, and short-run “shock” component as the daily deviation from this monthly average, the Frisch-Waugh-Lovell theorem shows that the “shock” component is already de-meaned as if the estimating equation had included location-by-month-by-year fixed effects in the final estimating 95 approach to simultaneously estimate the effects of short-run (acute) ozone exposure shocks, and long-run (expected) ozone norms, taking the difference as the implied total effectiveness of all defensive investments taken by economic agents to reduce negative health outcomes from anticipatable ozone exposure.4 Mechanically, the ozone norm is represented by a longrun moving average at the monthly level – e.g., the ozone norm for June, 2010 would be constructed as the mean ozone concentration for June in the years 2005 to 2009 – while the shock is represented by the daily deviation from this norm. Intuitively, consider two months of the year for the same location, one in which ozone concentrations are known to be high in expectation, e.g., by agents experiencing high levels of ozone for that month in prior years, and one in which ozone concentrations are known to be low in expectation. In the high expectation month, agents may be more likely to take steps to avoid exposure, i.e., undertake defensive investments, while in the low expectation month they may be less likely to do so. Thus, an unexpected high ozone concentration day, i.e., acute exposure shock, in the low expectation month may have a larger health impact than a day with the same ozone concentration in the month where the ozone norm was expected to be high, as agents may have engaged in defensive investments in the latter case but may be constrained in their ability to respond to the unexpected exposure shock in the first case. Throughout this paper, the difference between these two effects is referred to as agents’ intrinsic defensive investment (IDI).5 At the same time, “air quality alert” programs are designed to inform residents of their respective region if a poor air quality day is expected.6 By informing agents ahead of time for potential high air pollution days, such programs may induce agents to engage in more equation. 4Similar to seasonal variation in climatic conditions, ambient ozone has predictable seasonal variation, making it an ideal air pollutant for examining this concept in the context of air pollution. 5Such intrinsic defensive investments may be of particular importance in the future, as studies have shown that climate change is likely to cause an increase in ambient ozone (Jacob and Winner, 2009; Bento et al., 2023). If agents can adapt to this increase, however, adjusting their expectations to the new ozone norm, IDI could help reduce the climate change impact on health from increased ambient ozone. 6Such announcements are often made day of, or a few days in advance, and may be based on the general air quality index (AQI) measure, or on a specific pollutant, such as ozone or PM2.5. 96 defensive investment than they would have otherwise, particularly in the case where the agent did not a priori expect a high ozone concentration day. By including an indicator for such alert days in the analysis described above, the effectiveness of these programs for reducing negative health outcomes can be examined, and this channel of policy-induced defensive investment (PDI) can be compared with what agents may intrinsically undertake themselves. There is the possibility, however, that agents may become reliant on such alerts, forgoing their own intrinsic defensive investment activities in favor of only engaging in such actions when induced by the alerts; that is, PDI may crowd-out IDI. Conversely, if such information disclosures increase general awareness of ozone, e.g., if announcements help agents to develop a better understanding of the seasonality of ozone norms, alert programs could instead lead to higher levels of IDI. 7 In this case, rather than PDI crowding-out IDI, not only would the alert programs be complementary to IDI, but they may provide a “double-dividend” benefit – both directly benefiting agents through PDI and indirectly through increases in IDI. By interacting the ozone shock and ozone norm variables described above with an indicator variable for whether a county maintained an alert program, the differential level of IDI in these counties can be examined to determine whether PDI crowds-out or complements IDI. Notably, standard OLS estimates of the health impacts of air pollution are likely to suffer from measurement error driven attenuation bias among other possible concerns. Thus, this study adapts the wind direction IV proposed by Deryugina et al. (2019) to work within the context of simultaneously estimating both decomposed acute and seasonal ozone exposure. As with Deryugina et al. (2019), this paper uses daily wind direction as the instrument for the daily ozone shock – i.e., whether the average wind direction for that day was blowing from the Northeast, Northwest, Southeast, or Southwest. For the ozone norm, which is operationalized as the long-run monthly moving average ozone concentration, a commensu7Barwick et al. (2024), for example, find this to be the case in with the roll-out of China’s real-time air quality information program, although this program was much more comprehensive than a simple alert program. 97 rate IV is constructed as the long-run monthly moving average of wind direction – i.e., the proportion of the same long-run period when the wind was blowing from each of the four directions. The identifying assumption of the IV approach is thus similar to that of Deryugina et al. (2019); after flexibly controlling for highly granular fixed effects and meteorological variables, changes in a county’s daily wind direction are unrelated to changes in the county’s hospitalization rates or associated healthcare spending, except through their influence on the contemporaneous ozone shock. Similarly, changes in the proportional share of different wind directions for the same month in prior years are unrelated to current hospitalizations and spending except through their influence on the expected ozone norm. Overall, the results indicate that a 10 parts-per-billion (ppb) acute ozone exposure shock – approximately the mean magnitude of daily ozone shocks – leads to a sample average 7.3 percent increase in hospital admissions for cardiovascular or respiratory illness over the fourday window that spans the day of the shock and the following three days. A similar 10 ppb increase in the norm leads to a 3.9 percent increase, implying that IDI reduces admissions by almost half.8 By comparison, air quality alerts reduce four-day hospitalizations by 98 per million Medicare beneficiaries, implying that for the same 10 ppb ozone increase PDI reduces hospitalizations by just over 1.9 percent, or about half of the effect of IDI in counties with alert programs. The analysis finds no evidence that PDI crowds out IDI, with the sign of the effect suggesting that, if anything, there may be a double-dividend benefit, although the results are statistically insignificant. In line with the estimated effects on hospitalizations, the impacts on healthcare costs are not insignificant, with a 10 ppb ozone shock increasing the four-day average spending by between $1 to $1.7 million per million beneficiaries – a 7.7 to 12.3 percent increase, although IDI reduces the cost by 40-45 percent with PDI again about half of the effect of IDI. This study makes three main contributions to the literature. First, it examines the role of agents’ behavior in reducing the negative impact of ozone exposure on health outcomes. 8As with Deryugina et al. (2019), the IV estimates are significantly larger than the OLS estimates. 98 In contrast, prior studies on estimating the health impacts of acute ozone and air pollution exposure had predominantly examined the relationship without considering human behavioral responses (Bell et al., 2004; Dominici et al., 2006; Di et al., 2017; Deryugina et al., 2019), which may overstate the true health effects of air pollution levels if individuals have available channels of adjustment to avoid expected exposure. Second, it examines the total effectiveness of agents’ defensive investments for reducing the health impacts of air pollution exposure. Previous studies have often used a single channel of adjustment, such as mask purchases (Zhang and Mu, 2018; Wang and Zhang, 2023), air purifier sales (Ito and Zhang, 2020) or pharmaceutical sales (Deschenes, Greenstone and Shapiro, 2017) to estimate the cost of air pollution via agents’ WTP to avoid exposure. By instead examining the effects of ozone shocks and norms on hospitalizations and healthcare spending directly, the difference, intrinsic defensive investments (IDI), reflects the avoided hospitalizations and associated healthcare costs attributable to the sum of all channels of defensive investments. Third, it is able to compare this implied level of IDI with a measure of policy-induced defensive investment (PDI) driven by the information-based air quality alert programs, and test whether these two channels of adjustment are complementary or lead to agents substituting from one channel to the other. Prior literature on the role of information-based policy instruments for reducing the health impacts of air pollution have shown the benefits of PDI, but have often been regional in scope (Neidell, 2009b; Graff Zivin and Neidell, 2009) or, due to the nature of the informational program, have not examined their effects separately from individuals’ intrinsic defensive investment activities (Barwick et al., 2024). The paper proceeds as follows. Section II presents background information on ambient ozone formation and how this gives rise to predictable seasonality which can be affected by wind direction. Section III describes the data used in the analysis and Section IV presents the empirical strategy. Section V reports the main findings, examines the robustness of the estimates, and explores alternative health outcomes. Finally, Section VI concludes. 99 II. Ozone Formation, Seasonality, and Wind Direction Unlike most other air pollutants, ambient ozone is not emitted directly into the atmosphere. Rather, it is formed by the complex chemical reactions of two “precursor” pollutants, oxides of nitrogen (NOx) and volatile organic compounds (VOCs), following a Leontief-like production function in the presence of sunlight and warm temperatures. As both sunlight and warm temperatures are important catalysts in its formation, ambient ozone concentrations follow a predictable seasonality, typically increasing throughout Spring and into Summer, before declining into Fall and Winter. In fact, in many states the EPA does not mandate the monitoring of ambient ozone in Fall and Winter months.9 This predictability makes ambient ozone an ideal setting for examining whether agents intrinsically engage in defensive investments to protect themselves against expected exposure but does not preclude similar IDI behavior for other pollutants. As noted by Deryugina et al.’s (2019) paper proposing the use of wind direction as an instrumental variable for PM2.5 exposure, so long as wind direction is sufficiently correlated with ozone, and not otherwise correlated with the health outcomes of interest, it would be a valid instrument. Unlike PM2.5 however, ozone is highly reactive, with its half-life time in atmosphere decreasing with both humidity and airflow (McClurkin, Maier and Ileleji, 2013), thus it may be less likely that a significant amount of local ozone is due to transportation of ozone itself from outside the county via wind.10 NOx, however, has been shown to transport long distances, and this transported NOx can often have significant impacts on local ozone (Mauzerall et al., 2005, e.g.,). In fact, the EPA has specifically designated an Ozone Transport Region (OTR) consisting of 12 northeast states and the District of Columbia, where special emissions regulations are enforced due to the prevalence of inter-state NOx transport leading to increased ozone concentrations potentially hundreds of miles away from 9See Appendix Table A.1.3 for a list of required ozone monitoring months by state. 10This will, of course, depend on the size of the counties in question. Geographically small and dense counties may certainly transport some ozone between each other, especially if humidity and wind speed are low. 100 the initial emissions source (USCFR, 2013).11 By exploiting variation in local ozone pollution due to the transport of externally produced Ozone or NOx into a county, rather than within-county transport, the empirical analysis reduces the potential for measurement error in residents’ ozone exposure, as externally produced and transported pollution is likely to affect all residents equally. In the empirical analysis, described in detail in Section IV, ozone is decomposed into a long-run monthly norm and a daily shock as the daily deviation from this norm. Similarly, prior years’ monthly wind direction norms and contemporaneous daily wind direction are used as the respective instruments for these two channels of ozone exposure.12 Figure 3.1 illustrates the variation in wind direction used to estimate the causal health impacts of daily ozone shocks and monthly ozone norms. Specifically, Panel A depicts the relationship between the average daily wind direction at pollution monitors, in 10-degree bins, and daily ozone concentration shocks at these monitors in the Northeast region surrounding Greater Boston, MA.13 Panel B depicts a similar relationship between the prior years’ monthly wind direction and monthly ozone norms. All estimates are relative to the 260-270 degree bin, where 270 degrees corresponds to wind blowing from the West. The figure displays results from regressions that control for county-by-season and state-by-year fixed effects, as well as flexible controls for the interactions of maximum and minimum temperatures, precipitation, wind speed. As shown, wind direction is a particularly strong predictor of local ozone shocks, while certain wind directions are also strong predictors of local ozone norms.14 In both cases the 11The OTR includes all or part of: Connecticut, Delaware, Maine, Maryland, Massachusetts, New Hampshire, New Jersey, New York, Pennsylvania, Rhode Island, Vermont, Verginia, and the District of Columbia. 12As noted by Deryugina et al. (2019), prevailing wind direction may affect the placement of pollution monitors or cause residents to sort into upwind or downwind locations. However, the long-run monthly wind direction instruments are constructed from daily wind direction measurements, such that they are allowed to vary both month-to-month within the year as well as year-to-year within the sample, alleviating this concern. 13Specifically, all monitors in counties assigned to the monitor group that includes Boston, MA. This sub-sample is made up of approximately 17.5% observations from CT, 9% observations from ME, 32.5% observations from MA, 22% observations from NH, 5% observations from NY, 9% observations from RI, and 6% observations from VT. 14Note that while the pattern is clearer for shocks than norms, the magnitude of the effect is much larger for norms, by almost an order of magnitude for some wind directions. 101 Figure 3.1: Relationship Between Wind Direction and Ozone Concentration Shocks and Norms for Counties in and Around the Greater Boston Area, MA Dirty air from the SW -10 -8 -6 -4 -2 0 2 4 6 8 10 N NE E SE S SW W NW N Panel A. Ozone Shock & Daily Wind Dir. Clean air from the Atlantic -100 -80 -60 -40 -20 0 20 40 60 80 100 N NE E SE S SW W NW N Panel B. Ozone Norm & Historical Monthly Wind Dir. Daily Ambient Ozone Concentration (ppb) Notes: Panel A shows regression estimates of Equation (3.6) when using 10-degree angle rather than 90- degree angle bins of daily wind direction and when constructing the proportional 5-year wind direction “norms”, restricting the estimation sample to the single monitor group that includes the greater Boston area. The dependent variable is the daily ozone concentration shock, and the key independent variables are the wind direction instruments. Controls include county-by-season and state-by-year fixed effects, as well as a flexible function of maximum and minimum temperatures, precipitation, wind speed, and the interactions between them. The dashed lines represent 95 percent confidence intervals based on robust standard errors. Panel B shows the results of the same regression when replacing the dependent variable with the monthly ozone norm. patterns are consistent with expectations for the region. There is a clear pattern of lower ozone concentration shocks and norms when the wind is blowing from the Northeast and East, i.e., bringing in clean air from over the Atlantic Ocean. Similarly, both shocks and norms show an increase when the wind is blowing from the West-Southwest, bringing in dirty air from the rest of the Ozone Transport Region (OTR). 102 III. Data A. Health Data Data on hospital admissions and medical costs come from Medicare administrative data. The sample includes all beneficiaries between the ages of 65 and 100, accounting for over 97 percent of US elderly. Health care usage and associated costs are derived from the Medicare Provider Analysis and Review (MedPAR) file, which includes data on hospitalizations for any beneficiary enrolled in fee-for-service (FFS) Medicare. Each MedPAR observation is derived from a healthcare facility service claim corresponding to a single hospitalization event and includes the date of admission and total cost billed by the provider.15 Because ozone has been shown to primarily affect the cardiovascular and respiratory systems, admissions and associated costs are restricted to only those where the primary diagnosis was in either of these two categories.16 Beneficiary level hospital admissions and associated costs are aggregated to the county-day level using the date of admission and the beneficiary’s county of residence. Table 3.1 presents summary statistics for the main estimation sample, consisting of 1,762,556 county-day observations. As shown in Panel A, on average there are 25,851 FFS Medicare beneficiaries in each county, with an average four-day hospitalization rate for cardiovascular or respiratory diagnosis of 1,141 per million beneficiaries, and an associated average cost of care of approximately $12.66 million. 15Total cost includes any payments made by Medicare, as well as any additional payments made by the beneficiary themself or another party on their behalf. 16Using the primary ICD9/ICD10 diagnosis code assigned to the MedPAR observation, the observation is marked as a respiratory admission, cardiovascular admission, or excluded if an unrelated diagnosis is assigned. 103 Table 3.1: Summary Statistics Mean Std. Dev. Panel A. Medicare data Number of FFS beneficiaries 25,851 42,359 Four-day admissions rate (respiratory or cardiovascular) 1,141 4,582 Total cost of care ($ millions) 12,662,598 54,156,050 Panel B. Ozone data Ozone norm (ppb) 46.90 7.09 Ozone shock (ppb) -1.11 11.76 Ozone shock (ppb, absolute magnitude) 9.24 7.36 Panel C. Air quality alert data Total counties 876 Counties with alert program 251 Total alerts issued 14,588 Notes: Table reports unweighted statistics for the estimation sample. Unit of observation is county-day. Hospitalization rate and cost of care are per million fee-for-service (FFS) Medicare beneficiaries. Ozone norms and shocks are constructed following the approach detailed in Section IV, using the daily highest 8-hour average concentration, averaged across all monitors within a county. B. Air Pollution Data Air pollution concentration data is obtained from the EPA’s Air Quality System database, which provides monitor level hourly concentrations for all Clean Air Act regulated pollutants collected from the nationwide network of EPA air quality monitors.17 Comprehensive data for the daily highest 8-hour average ambient ozone concentration are collected for every active monitor for the years 1980 to 2017 during the Spring and Summer months (April through September).18 Ozone is a seasonal pollutant affected by the local climate, thus the EPA designates an “ozone season” for each state, which can range from all twelve months of the year to only the six months of April through September; thus the sample is constrained to this six month period which is common across all states. Data on three other criteria pollutants are similarly obtained for April through September 2004 to 2017, including fine particulate 17The EPA publishes official daily averages and maximums for each pollutant at the monitor-level via the AirData portal accessible at: https://aqs.epa.gov/aqsweb/airdata/download files.html#Daily. 18Pollution data from years prior to the measured health outcomes are necessary for the implementation of the method as will be described in detail in the following section. 104 matter (PM2.5), carbon monoxide (CO) and sulfur dioxide (SO2) as prior literature has linked these criteria pollutants to similar health effects (Currie and Neidell, 2005; Schlenker and Walker, 2016; Deryugina et al., 2019, e.g.,). Monitor level daily concentrations for each pollutant are aggregated to the county-day level by averaging all available pollution concentrations on each day across all monitors within the county. Figure 3.2 Panel A depicts the yearly evolution of ambient ozone, which shows a steady decrease in the maximum observed concentration but a relatively constant mean concentration. Panel B depicts the number of active ozone monitors in each year and number of monitored counties, which both steadily rose from the mid-1980’s until around 2010 but have levelled off since then. Figure 3.2: Evolution of Ambient Ozone Air Pollution and Monitoring (1980 – 2017) 30 40 50 60 70 80 90 100 110 120 Ozone (ppb) 1980 1985 1990 1995 2000 2005 2010 2015 Annual county mean Annual county max Panel A. Ozone concentration (ppb) 400 600 800 1000 1200 1400 Total 1980 1985 1990 1995 2000 2005 2010 2015 Annual total monitors Annual total counties Panel B. Ozone monitors & monitored counties Notes: Panel A depicts the annual county means (solid line) and county maximums (dashed line) for the daily highest 8-hour average ozone concentration. As seen, while the maximum concentrations have steadily fallen since the implementation of the Clean Air Act in 1980, the mean concentrations have remained relatively constant near 40 parts per billion. Panel B depicts the nationwide number of ozone monitors, and monitored counties, which both steadily grew from the mid-1980’s through about 2010 but have leveled off since then. Data on air pollution alert declarations are similarly collected from the EPA, which maintains a database of all local air pollution control agencies’ alert programs, available 105 starting from 2004.19 Alert program data includes: forecast date; whether an alert was declared for the forecast date; and the latitude/longitude of the forecast region. Alert day declarations are matched to their respective county-day by latitude/longitude and date and assigned an indicator value of one, with all other county-days assigned a value of zero.20 The earliest occurrence of an alert day forecast for each county is used to create a second indicator variable denoting the starting year from which an alert program was active versus any prior years where no such alert program was in place. As reported in Panel B of Table 3.1, the average ozone norm in the final estimating sample is 46.90 ppb with a standard deviation of 7.09 ppb; the average ozone shock is 9.24 ppb in absolute magnitude, with a standard deviation of 7.36.21 Of the 876 counties in the final sample, 251 counties maintained an alert program in at least one year of the sample, with 14,588 alerts issued during this period. C. Wind & Other Meteorological Data Wind direction and speed data are obtained from the North American Regional Reanalysis (NARR) daily data for the years 1999 through 2017.22 NARR incorporates raw data from a variety of sources including weather stations, satellites, aircraft, radiosondes and dropsondes (weather instruments either lifted by hot air balloons or dropped from aircraft), and Pibals (specialized balloons used to measure wind at different altitudes) to produce reanalyzed data for dozens of weather parameters, including wind direction and speed, at a 32-by-32 kilometer grid in 3-hour increments. Wind speed is reported as an east-west component (u-vector) and north-south component (v-vector), which are linearly interpolated to each 19Some alert day programs are tied to specific pollutants, such as ozone or PM2.5, while others are based on the overall Air Quality Index (AQI) level. For the purposes of this paper all alert days are treated as the same, with the effectiveness of alternative alert channels left for future research. 20In later robustness checks alert day declarations are instead matched to every county within the respective CBSA associated with the latitude/longitude. 21Due to being constructed as the daily deviation from the ozone norm, the ozone shock is approximately centered around zero, with a mean of -1.11 ppb and a standard error of 11.76 ppb. 22NCEP North American Regional Reanalysis (NARR) data provided by the NOAA PSL, Boulder, Colorado, USA, from their website at https://psl.noaa.gov. 106 ozone pollution monitor from the original grid points and averaged to a daily value for each wind vector at each monitor. Wind speed at each pollution monitor is calculated as the vector-sum of the u- and v- vectors, while wind direction is calculated to reflect the direction the wind is blowing from following Deryugina et al. (2019).23 Both wind direction and wind speed are averaged to the county-day level using all values from all monitors in the county. Finally, daily temperature and precipitation data for the years 2004 through 2017 are obtained from PRISM, which provides daily minimum and maximum temperature and total precipitation for a 4-by-4 kilometer grid covering the contiguous United States.24 The daily measures across all points within a county are then averaged to create a county-day measure. IV. Empirical Strategy A. Health Impacts of Acute (Unexpected) and Long-run (Expected) Air Pollution Local air pollution that follows a seasonal (i.e., predictable) trend, such as ambient ozone, can be decomposed into two constituents: (i) its expected long-run norm, and (ii) its daily, unexpected, deviation from this norm. The norm – for example, the monthly average ozone level in prior years – is experienced by individuals, allowing them to make future adjustments to avoid exposure during periods of expected high concentrations. The daily deviation from this norm reflects a potentially unexpected exposure shock, which individuals may be constrained in their ability to predict or adjust to.25 The difference between the impact of ozone norms and ozone shocks on health outcomes can thus be used to infer the sum effectiveness of all adjustments that agents undertook to avoid expected ozone exposure. To examine this empirically, the analysis builds on, and combines, two separate methods – one for jointly estimating the effects of short-run weather shocks and long-run climate 23Specifically, wind speed equals √ u 2 + v 2, while wind direction is calculated by first calculating θ = 180 π arctan(|v| |u| ), and then translating θ into a 0-360 degree scale depending on the signs of u and v. 24PRISM Climate Group, Oregon State University, https://prism.oregonstate.edu, data created 28 October 2019, accessed 10 April 2022. 25In recent years the availability of local air quality information has increased, and some individuals may check local air quality daily. Even then, they may be constrained in their ability to make day-of adjustments. 107 impacts to infer climate adaptation (Bento et al., 2023), and the other an instrumental variables (IV) approach for addressing measurement error and other concerns in estimating the causal impact of air pollution exposure on health and mortality (Deryugina et al., 2019). The objective is to jointly estimate the separate effects on health outcomes of acute ozone shocks vis-`a-vis expected ozone norms, controlling for potentially confounding factors. Consider the following model of the shock and norm relationships with hospital admissions for cardiovascular or respiratory illness:26 Admitit = β SOzoneS it + β N OzoneN it + X ′ itγ + αis + τky + ϵit, (3.1) where i denotes county in state k, and t denotes day in season s (Spring or Summer) of year y. 27 The parameters of interest are β S and β N , the coefficients on the ozone shock and norm, respectively. As noted in prior studies of the health effects of ozone and other local air pollutants, there could be short-run displacement effects and/or delays between exposure and the onset of symptoms (Bell et al., 2004; Bell, Peng and Dominici, 2006; USEPA, 2006; Deryugina et al., 2019). For example, ozone exposure may cause an individual that was already suffering from respiratory distress to be admitted to the hospital a few days earlier than they would have otherwise. Conversely, symptoms from pollution exposure may not manifest until the following day(s). To account for both possibilities, the dependent variable, Admitit, reflects a four-day total based on day t and the following three days.28 X′ it includes highly granular and flexible weather controls – interacting bins of daily maximum and minimum temperature, precipitation, and wind speed.29 To ensure that β S and β N are not capturing 26For now, consider hospital admissions with either of these two diagnoses as the main outcome of interest. Later analyses will consider other outcomes such as associated healthcare cost. 27The ozone season varies by region, and thus not all counties measure ozone concentration throughout the year, see Table A.1.3 in Appendix A.1 for further details, thus the estimating sample is restricted to Spring and Summer (April – September), the common ozone season across all counties. 28In later robustness checks, models are examined for totals between one- to seven-days, with the four-day model capturing similar effects as models with longer time windows while reducing computation time. 29Specifically, 17 bins of temperature are defined, ranging from below -15C◦ to above 30C◦ in 3C◦ increments. Additionally, bins for deciles of windspeed and precipitation are created. All possible interactions of these temperature, precipitation, and wind speed bins are used to generate a set of indicators for inclusion in the 108 the effects of weather conditions during this four-day period, three leads of the weather variables are also included in X′ it. Similarly, the OLS estimates of Equation (1) include three leads and lags of both OzoneS it and OzoneN it , while the IV estimates include three leads and lags of the instruments.30 Fixed effects for county-by-season, αis, are included to flexibly control for county-level effects that may vary across seasons, such as geographic and seasonal differences in health or pollution. Similarly, fixed effects for state-by-year, τky, are included to flexibly control for state-level effects that may vary across years, such as changes in health- or environmental-policies.31 Standard errors are clustered at the county level, and estimates are population-weighted using the daily number of FFS Medicare beneficiaries in each county.32 Measuring Intrinsic & Induced Defensive Investments — By estimating both effects of acute (unexpected) and long-run (expected) exposure in the same equation, the coefficients can be directly compared, allowing for straightforward statistical inference. Any difference can then be interpreted as the sample average effectiveness of the sum of all intrinsic defensive investments (IDI) taken by individuals to protect themselves against expected exposure. That is: IDI = β S − β N . At the same time, many local air pollution authorities maintain air pollution alert day programs, which aim to notify residents if a high pollution day is expected, typically a day or two in advance, but in some cases as many as five days prior.33 By alerting individuals to an expected high level of local air pollution on a specific day, individuals may be able estimating equation. 30Specifically, three leads and lags of the IV for OzoneS it are included. As will be discussed later regarding the IV implementation, leads and lags of the IV for OzoneN it cannot be included. However, as the norm for any given day within a month is the same for the following day – with the rare exception of the last day of the month – this is less of an issue than it might seem. By redefining OzoneN it as the sum of the norms for the four days that correspond to the four-day outcome window the estimation can account for the “double counting” that would otherwise be attributed to the norm. 31In later robustness checks alternative fixed effects are examined. 32In later robustness checks alternative clusters are examined. Furthermore, as the ozone shock and norm could be considered generated regressors, bootstrapped standard errors are estimated following the Bayesian bootstrap approach (Rubin, 1981). 33Notification channels differ by region, but may include local news, website, smartphone app, radio, and so on. 109 to make last minute defensive investments. That is, alerts may induce additional defensive investment on relevant days.34 To estimate the effectiveness of such policy-induced defensive investments (PDI), an indicator variable for alert days is added to Equation (3.1), giving the following model: Admitit = β SOzoneS it + β N OzoneN it + β AAlertit + X ′ itγ + αis + τky + ϵit, (3.2) where Alertit is an indicator variable equal to 1 if county i observed an air quality alert for day t, and zero otherwise, and all other variables are as defined previously. The additional coefficient of interest, β A, captures the average change in the admissions rate for cardiovascular or respiratory illness on alert days. Notably, while IDI = β S − β N reflects changes in hospital admissions in terms of a one part per billion (ppb) increase in ozone exposure, β A reflects changes hospital admissions, holding constant the effects of ozone shocks and norms. Thus, to make IDI and PDI comparable, β A is divided by the sample average daily ozone concentration to calculate an effect in the same “per ppb” units as IDI. That is, PDI is defined as β A/Ozoneit. Finally, individuals in regions with alert day programs may become accustomed to receiving alerts on bad air quality days and begin relying on alerts rather than historical norms to inform their defensive investment decisions. If this were the case, PDI would become a substitute for IDI, rather than complementing it.35 To examine this, an additional interaction term for the existence of an alert program is added to Equation (3.2), giving the following 34As many alert programs may also urge individuals to take action to reduce their own impact on emissions, such as by switching from driving to taking public transit, there may be some concern that alerts affect health outcomes through changes in air pollution itself, rather than through changes in defensive investments. For example, Cutter and Neidell (2009) find that “Spare the Air” alerts in the San Francisco Bay Area do indeed induce a shift from driving to public transit use, although it is unclear whether this shift has a meaningful impact on ambient ozone in the region. By comparison, Sexton (2012) finds that programs with a public transit ridership subsidy increase ridership without meaningfully reducing car trips. Further, Bento et al. (2023) find no impact of ozone alert days on the ozone/temperature relationship. In either case, within the estimating sample the average ozone concentration remains higher on alert days (∼65 ppb) than non-alert days (∼46 ppb), and the proposed empirical approach recovers an estimate of alert days holding constant the effects of ozone shocks and norms. 35In this case, while the health benefits of the alert day policy may be lower than expected, there may still be benefits, such as reduced cognitive burden on agents who follow alerts in lieu of actively thinking about ozone exposure. 110 equation: Admitit = β SOzoneS it + β N OzoneN it + β S P (OzoneS it × P rogiy) + β N P (OzoneN it × P rogiy) + β AAlertit + X ′ itγ + αis + τky + ϵit, (3.3) where P rogiy is an indicator variable equal to 1 if county i maintained an air quality alert program in year y and is zero otherwise, with all other variables as previously defined.36 In this case, β S and β N capture the effects of ozone shocks and norms in counties without an alert day program, while (β S +β S P ) captures the effect of a shock in counties with an alert program and (β N + β N P ) does so for the effect of the ozone norm. In this way, (β S P − β N P ) can be interpreted akin to a difference-in-differences estimate of the effect of implementing an air quality alert program on agents’ IDI. If (β S P − β N P ) < 0, this would imply that agents substitute away from IDI in favor of relying on alerts to inform their level of defensive investment, PDI. If, however, (β S P − β N P ) > 0, this would imply that – as suggested by prior literature on air quality information campaigns (Barwick et al., 2024, e.g.,) – the alert programs in fact increase overall awareness of air pollution, such as ozone, and thus agents engage in more IDI than they would have previously.37 Notably, the decision to implement an alert day program is typically under the purview of the local air control authority, and thus may be endogenous with local air quality as, for example, counties with worse air pollution may be more likely to implement alert programs. Figure 3.3 examines potential differences in the ozone distribution between these two sets of counties, plotting the empirical distributions of daily ambient ozone concentrations for counties without an alert program, and for counties with an alert program on days without an alert, days with an alert, and all days combined.Panel A depicts the full distributions of 36The one exception is that X′ it now additionally includes the un-interacted indicator variable P rogiy to control for any systematic level differences between counties with and without alert programs. 37An increase in IDI across alert program counties would reflect a “double dividend” benefit of alert programs, with the direct benefit measured by PDI and this secondary benefit measured by the increase in IDI, (β S P − β N P ). 111 each, which appear quite similar except for the alert day distribution, which is shifted to the right as would be expected.38 Panel B focuses solely on days above 70 ppb, the right-tail of the distribution consisting of days where the ozone level exceeded the allowable threshold under the EPA’s current National Ambient Air Quality Standards (USEPA, 2015a). As shown, counties with alert programs on average have a slightly fatter right-tail than counties without alert programs, which becomes a slightly thinner tail if removing the alert days from the distribution. Figure 3.3: Distributions of Daily Ozone Concentrations in Counties with or without Alert Day Programs Notes: This figure depicts the distributions of daily highest 8-hour average ozone concentrations in counties with and without an alert program, as well as on days where an alert was issued. Panel A includes the entirety of each distribution, with the vertical line at 70 ppb for reference as the current EPA threshold for allowable ozone concentration under the National Ambient Air Quality Standards (NAAQS) for ozone. Panel B includes only the “right-tails” of the distributions, above this 70 ppb threshold, to more clearly illustrate any differences across counties for high ozone days. The distribution on alert days is excluded from Panel B to allow for differences in the other distributions to be more visible. 38Importantly, although the alert day distribution is shifted to the right, it shares a common support with the other distributions, indicating that not all high ozone days are necessarily associated with an alert, and some alert days end up realizing low ozone concentrations. 112 Aside from having relatively similar ozone distributions, any systematic differences in average ozone concentrations or health profiles across counties would be controlled for in the estimating equation via included fixed effects. However, there may be differences in the slope of the ozone-health relationship between the two sets of counties, driven by potential selection, which would not be accounted for. That said, the key metric of interest in Equation (3.3) is not any potential level differences between counties with or without alert programs, but the relative difference in IDI captured by the difference-in-differences analog, (β S P −β N P ), which highlights whether PDI crowds out or compliments IDI. 39 B. Decomposing Ozone Estimating the above equations requires, as a first step, decomposing daily ambient ozone concentrations into an expected norm and a daily shock. This paper adapts the approach of Bento et al. (2023) who decompose daily temperature to jointly estimate the effects of climate and weather by relying on the properties of the Frisch-Waugh-Lovell theorem (Frisch and Waugh, 1933; Lovell, 1963). In the context of Bento et al. (2023), the 30-year moving average of monthly temperature is used as the climate norm following the climate science literature. In the context of air pollution however, there is no pre-existing literature that defines a “pollution norm.” Operationalizing the concept of an ozone pollution norm must therefore consider both the within-year periodicity at which agents may recall norms, as well as the across-year frequency at which old information is forgotten (or deemed unimportant) and new information is added. First, as months are a well-defined unit of time that individuals likely associate 39This potential selection concern is examined in later robustness checks, finding that when re-estimating Equation (3.3) on alternative restricted sub-samples based on alert program status the estimated effect of alert programs on IDI remains relatively unchanged. Because there is both variation in the timing of when counties implemented an alert program and there is a “never treated group” (the counties that never implemented an alert program), identification can rely on the weakest parallel trends assumption considered by Marcus and Sant’Anna (2021), which does not impose any restriction on the pre-treatment trends across groups. As the authors note, if there is a “reasonably large” number of never treated units – as is the case here – that assumption can identify policy-relevant parameters even “if researchers are not comfortable with a priori ruling out nonparallel pretrends” (Marcus and Sant’Anna, 2021, p.251). 113 with environmental phenomena – whether it be, for example, temperature, precipitation, or air pollution – this is taken as the baseline period for defining norms within-year. Second, while agents exhibiting recency bias (Tversky and Kahneman, 1973, e.g.,) may only recall pollution norms from the previous (few) year(s), other agents may recall norms from many years prior. Therefore, the central analysis adopts a middle-ground approach, using a 5-year moving average, as wind direction data are only available beginning five years prior to the beginning of the sample period.40 Note that the norm variable is lagged by one year to give agents time to observe any shifts in monthly ozone norms and make adjustments in response. Thus, in the central analysis the ozone norm for any given county and day is given by: OzoneN it = 1 5 X y−1 j=y−5 Ozoneimj (3.4) where m reflects month of the year, and other notation is as defined previously. The ozone shock follows directly as the difference of the daily ozone concentration from this norm: OzoneS it = Ozoneit − OzoneN it (3.5) Figure 3.4 depicts this decomposition for Los Angeles County, highlighting the two sources of identifying variation. Panel A includes the full sample period, 2004 through 2017, and Panel B includes only data from 2017 to more clearly illustrate variation in the ozone shock and norm vis-`a-vis observed daily ozone concentrations. 40In later robustness checks, shorter 1- and 3-year as well as longer 10- and 20-year moving averages are also considered. Note that due to the aforementioned data limitations, specifications with 10- and 20-year moving averages are still only able to take advantage of five years of preceding wind direction data for the IV. 114 Figure 3.4: Decomposition of Ambient Ozone Concentration into Daily Shocks and Monthly Norms -50 0 50 100 Apr. 2004 Apr. 2008 Apr. 2012 Apr. 2016 Panel A. Los Angeles - All Years -20 0 20 40 60 80 Apr. May Jun. Jul. Aug. Sep. Oct. Panel B. Los Angeles - 2017 Ozone Concentration (ppb) Daily 8hr Ozone Ozone Shock Ozone Norm Notes: This figure illustrates the two sources of variation used in the analysis to simultaneously recover estimates of the health effects of acute ozone shocks and expected ozone norms. Panel A depicts daily data for the entire sample period for Los Angeles county, (April-September, 2004-2017), while Panel B restricts the graph to only 2017 for clarity. Notice that the variation in the ozone shock is nearly identical to the variation in the observed daily 8-hour ozone, the only significant difference is in levels, as the shock has been de-meaned from the norm and is thus approximately centered around zero. At the same time, there is both month-to-month variation in the ozone norm within-year, as well as variation across years. C. Instrumenting with Wind Direction OLS estimates of Equations (3.1), (3.2), and (3.3) are prone to bias because exposure to ozone may not be randomly assigned and is likely to be measured with error.41 Thus, the empirical strategy employed here adapts the IV approach proposed by Deryugina et al. (2019), using wind direction as an instrument for air pollution, to work in combination with the above 41For example, Grainger, Schreiber and Chang (2019) find evidence that local regulators avoid pollution hot-spots when deciding where to site new ozone monitors, and Muller and Ruud (2018b) find that prior period ozone readings are associated with whether to add or drop a monitor in the following period. 115 decomposition of ozone. The daily wind direction remains a viable IV for the daily ozone shock, as shown in Section II, but it is unclear what relationship, if any, contemporaneous wind direction would have with the ozone norm, which is constructed from historical ozone measurements. The ozone norm is, however, likely to be related to historical wind direction over the same measurement period and a “wind direction norm” instrument can thus be constructed using a similar approach to the construction of the ozone norm. Specifically, by counting the number of days in each month that the wind was blowing from each of the four directions, and normalizing such that all directions together sum to unity, four monthly wind direction variables can be constructed with each one representing the proportion of the month that the daily wind was, on average, blowing from the respective direction. For example, if in June for a given area the wind blew from the Northeast 3 days, the Southeast 6 days, the Southwest 9 days and the Northwest 12 days, this would be reflected as values of 0.1, 0.2, 0.3, and 0.4 for each of the four monthly wind direction variables respectively. Next, by taking the 5-year moving average of these proportional monthly wind directions, similarly lagged by one year, a set of wind direction norm instruments can be constructed to match the same frequency and period as the ozone norm variable.42 Thus, following a similar procedure to Equation (3.4), a set of four wind direction norm variables can be defined as {wdirnorm0 imy, wdirnorm90 imy, wdirnorm180 imy, wdirnorm270 imy}. For example, the variable wdirnorm0 imy would reflect the proportion of month m, over the previous 5-years, where the daily average wind direction in county i fell within the [0,90) degree interval. Combining both instruments, the specification for the first-stage regressions is: 42In the case of robustness checks where the ozone norm is constructed using only the previous 1- or 3-years of data, the wind direction “norm” instruments are similarly constructed using only the prior 1- and 3-years of data. However, in the case of the 10- and 20-year ozone norms, the wind direction “norm” instruments are still constrained to only include the prior 5-years of data, as this data is only available from 1999 onwards. 116 Yit = X g∈G X 2 d=0 (σ g d [Gi = g] × wdir90d it + ν g d [Gi = g] × wdirnorm90d imy) + Z ′ itδ + αis + τky + ϵit, (3.6) where Yit represents OzoneS it and OzoneN it and the excluded instruments are P g∈G P2 d=0(ρ g d [Gi = g] × wdir90d it + v g d [Gi = g] × wdirnorm90d imy).43 Each variable in the set {wdir90d it } is equal to one if the daily average wind direction in county i fell within the 90-degree interval [90d,90d+90) and zero otherwise, while the variables in the set {wdirnorm90d imy} are as described above. In both cases, the interval [270,360) is the omitted category. Fifty spatial monitor groups obtained via the k-means cluster algorithm based on location are represented by G, with the variable 1[Gi = g] an indicator for county i being classified into monitor group g from the set of groups G.Intuitively, neighboring monitors are more likely to face the same environmental conditions, and be assigned to the same group, than distant monitors. The coefficients σ g d and ν g d are thus allowed to vary across geographic regions. Included fixed effects are as defined in Equation (3.1), while Z ′ it contains all variables included in X′ it when instrumenting for Equation (3.1), and additionally includes Alertit when instrumenting for Equation (3.2) and P rogiy when instrumenting for Equation (3.3). Furthermore, when instrumenting for Equation (3.3), interactions of P rogiy and P g∈G P2 d=0(ρ g d [Gi = g] × wdir90d it + v g d [Gi = g] × wdirnorm90d imy) are included. Equation (3.6) restricts the effect of wind direction to be constant within each of the four wdir bins and scale linearly with the wdirnorm proportional variables but allows the effects to vary between geographic groupings of monitors (g ∈ G). Thus, despite this restriction it includes hundreds of instruments each for ozone shocks and norms, as well as 43Note that wdirnorm90d imy takes the same value for consecutive days within the same month, and thus including leads or lags of this variable in Equations (3.1), (3.2), or (3.3) would de facto be as if the excluded instruments were included in the second stage regression. For this reason, the four-day sum of ozone norms is used for OzoneN it in the central analysis rather than explicitly including leads or lags of this instrument in the second stage. 117 tens of thousands of control variables and fixed effects and is estimated on over 1.7 million observations.44 Furthermore, weak instrument bias is not a concern in this setting. As demonstrated in Figure 3.1, wind direction is a strong predictor of ambient ozone shocks and norms – which is confirmed by the large first-stage F-statistics shown in the results tables.45 V. Results A. Impacts of Ambient Ozone Concentration on Health As discussed in Section IV, the ultimate goal is to recover empirical estimates of β S , β N , β A, β S P , and β N P from Equation (3.3) which can then be used infer values of IDI, PDI, and whether there are any effects on IDI when alert day programs are implemented. To that end, Table 3.2 presents estimates of the impact on hospital admissions with cardiovascular or respiratory diagnoses from ozone shocks, ozone norms, and air pollution alerts, as well as the implied levels of IDI and PDI recovered via Equations (3.1), (3.2) and (3.3).46 Column (1) reports the coefficients recovered via estimating Equation (3.1) via OLS, while column (2) reports the corresponding IV estimates. When estimated via OLS, a 10 ppb increase in the ozone norm is associated with 9.67 additional hospitalizations per million beneficiaries over the four-day window, or a 0.86 percent increase relative to the average admissions rate, though notably this effect is not statistically significant. In the case of the ozone shock, the effect is statistically significant, but small in magnitude and oppositely signed from what would be expected, with every 10 ppb increase in the daily shock associated 44As noted by Deryugina et al. (2019), increasing the number of instruments is computationally burdensome, without meaningfully changing the estimated effects. The included fixed effects and control variables, however, are partialed out using the algorithm developed by Correia (2017), which automatically performs the necessary degrees-of-freedom adjustments, greatly decreasing the computational burden of including high dimensions of fixed effects and controls. 45Tables of results for IV estimations include the first-stage F-statistic computed assuming errors are homoscedastic, and thus can be compared to the Stock and Yogo (2005) critical values which are valid only under homoscedasticity. As there are two first-stage equations, one for ozone shocks and one for ozone norms, the smaller of the two F-statistics is presented. 46Recall that standard errors are clustered at the county level, with later robustness checks examining alternative specifications. OLS estimates report the model R2 , while IV estimates report the first-stage F-statistic. 118 with 7.32 fewer hospitalizations per million beneficiaries, or a 0.65 percent decrease from the average admissions rate. Taken together, this would imply that on average the sum effect of all intrinsic defensive investments somehow increases the average admissions rate by 17 beneficiaries per million over the four-day window for every 10 ppb increase in ambient ozone, or a 1.51 percent increase over the average admissions rate. Table 3.2: OLS and IV Estimates of the Effects of Acute Ozone Shocks and Expected Ozone Norms on Elderly Hospitalization for Cardiovascular or Respiratory Disease OLS IV IV IV (Alert Day) (Alert Program) (1) (2) (3) (4) Panel A. Effect of ozone (ppb) on hospital admissions Ozone Shock (ppb) −0.732*** 7.917*** 8.247*** 6.661*** (0.099) (1.297) (1.306) (1.296) Ozone Norm (ppb) 0.967 4.331*** 4.409*** 3.262*** (0.653) (0.520) (0.516) (0.530) Shock x Alert Program 10.593*** (1.569) Norm x Alert Program 5.873*** (0.820) Alert Day −63.103*** −98.535*** (11.984) (17.044) Implied IDI −1.699** 3.586*** 3.838*** 3.399*** (0.664) (1.250) (1.264) (1.304) Implied PDI 1.388*** 2.167*** (0.264) (0.375) Implied IDI x Alert Counties 4.720*** (1.632) Program Effect on IDI 1.321 (1.534) Observations 1,777,710 1,762,556 1,762,556 1,762,556 R2/F-statistic 0.925 291 331 176 Continued on next page. 119 Table 3.2: Continued OLS IV IV IV (Alert Day) (Alert Program) (1) (2) (3) (4) Panel B. Percent Change in hospital admissions for a 10-ppb ozone increase Four-day Admissions (sample mean) 1,124 1,129 1,129 1,129 Shock -0.651 7.012 7.304 5.899 Norm 0.861 3.836 3.905 2.889 Shock x Program 9.382 Norm x Program 5.201 Implied IDI -1.512 3.176 3.399 3.010 Implied PDI 1.229 1.919 Implied IDI x Alert Cnty 4.181 Program Effect on IDI 1.170 Notes: Table reports OLS and IV estimates of Equation (3.1) and IV estimates of Equations (3.2) and (3.3). Dependent variable is the four-day hospitalization rate with a cardiovascular or respiratory diagnosis per million beneficiaries. All regressions include county-by-season and state-by-year fixed effects; flexible controls for interactions of temperature, precipitation, and wind speed; and three leads of these weather controls. Three leads and lags of the ozone variables or the instruments are also included or accounted for as discussed in Section IV. Estimates are weighted by the number of beneficiaries in each county. Recall that the ozone norm reflects the 5-year moving average of ozone at the monthly level, while the ozone shock reflects the daily deviation from this norm. Reported F-statistic is from the first-stage regression when instrumenting for the ozone shock, which is the smaller of the first-stage regression F-statistics. Standard errors are clustered at the county level and reported in parentheses. ***, **, and * represent significance at 1%, 5% and 10%, respectively. By comparison, the corresponding IV estimates reported in column (2) are 4–10 times larger in magnitude, and both of the expected sign, suggesting that the OLS estimates indeed suffer from significant bias. Specifically, the IV estimates indicate a causal effect of a 10 ppb increase in the ozone norm on hospitalizations of 43.31 per million beneficiaries, or a 3.84 percent increase in the average admissions rate. Similarly, a 10 ppb increase in the ozone shock causes an increase in hospitalizations of 79.17 per million beneficiaries, or a 7.01 percent increase. Taken together, this implies that IDI reduces the hospitalizations that otherwise would have occurred due to a 10 ppb increase in ozone by 35.86 per million beneficiaries, or 3.18 percent of the average admissions rate. In other words, for the same increase in ambient ozone concentration, when that increase is expected IDI reduces associated hospitalizations by approximately 45 percent. 120 Column (3) reports the IV estimates of Equation (3.2), adding the alert day indicator to simultaneously recover an estimate of β A which can be used to infer PDI. Both coefficients on the ozone norm and shock are of approximately similar magnitude and are statistically indistinguishable from those reported in column (2).47 Specifically, a 10 ppb ozone shock is estimated to cause an increase in admissions of 7.3 percent while a similar 10 ppb increase in the ozone norm causes an increase of 3.91 percent, implying that IDI decreases the admissions impact of ozone exposure by 46.5 percent. By comparison, hospitalizations on alert days are estimated to decrease by 63.1 per million beneficiaries on average, which implies that PDI reduces hospitalizations by 13.88 per million beneficiaries for a 10 ppb increase in ambient ozone, or 1.2 percent of the average admissions rate. In other words, PDI has about 36 percent of the effect of IDI for reducing hospitalizations caused by ambient ozone. Finally, column (4) reports the IV estimates of Equation (3.3), adding the interaction terms for counties with an alert day program to estimate any change in IDI caused by the implementation of such programs, highlighting four points. First, note that the shock β S and norm β N coefficients in counties without an alert program are somewhat lower than the sample average coefficients reported in column (3), while in counties with an alert program the full shock (β S + β S P ) and norm (β N + β N P ) effects are somewhat higher.48 Specifically, a 10 ppb ozone shock increases admissions by 66.61 per million beneficiaries in counties without an alert program and by 105.93 in counties with and alert program, corresponding to an increase in the average admissions rate of 5.9 and 9.38 percent respectively. A 10 ppb increase in the ozone norm, by comparison, increases admissions by 32.62 or 58.73 per million beneficiaries in counties without or with an alert program, respectively, corresponding to an increased average admissions rate of 2.89 and 5.2 47Estimates of both β S and β N increase slightly, as would be expected now that the estimation explicitly includes the alert day indicator, which previously would have caused a (mild) negative omitted variable bias on both coefficients. 48While the effects of the ozone shock and norm in counties with and without an alert program are statistically distinguishable from each other, ozone shocks in either type of county are statistically indistinguishable from the sample average ozone shock reported in column (3). The effects of the ozone norm do appear to be statistically distinguishable from the sample average norm reported in column (3), although this is not explicitly tested as the estimates are recovered via separate regressions. 121 percent. Second, following a similar pattern to the estimated impacts of ozone shocks and norms, the implied effect of IDI for counties without an alert program is slightly smaller than the full sample average, while being slightly larger in counties with an alert program. In counties without an alert program IDI reduces hospitalizations from a 10 ppb ozone increase by 33.99 per million beneficiaries, compared to 47.2 in counties with an alert program – though notably it cannot be statistically ruled out that the effect of IDI is the same across both types of counties. Third, air quality alerts are estimated to reduce hospitalizations by 98.54 per million beneficiaries, implying a PDI effect of 21.67 per million beneficiaries, or a reduction in average hospitalizations of 1.92 percent – approximately 46 percent of the effect of IDI in counties with an alert program. Lastly, as discussed in Section IV, a key metric of interest in column (4) is the relative difference in IDI between the two sets of counties: (β S P − β N P ), which is analogous to a difference-in-differences estimate of the effect of implementing an alert program on residents’ IDI. For a 10 ppb increase in ozone, IDI reduces hospitalizations by an additional 13.21 beneficiaries per million in counties with an alert day program relative to counties without a program, or 1.17 percent of the average admissions rate, indicating a potential “double dividend” benefit of alert programs, though this effect is statistically insignificant. This finding suggests that PDI does not, on average, crowd out IDI as would be the case if agents substituted away from relying on their own expectations of ozone norms formed by experience and instead began relying more on receiving alerts. Rather, alerts may increase agents’ awareness of ozone norms. For example, consider a county in which ozone is, on average, higher in June than in May, and in which there exist two types of agents: attentive and fully rational agents, and inattentive or boundedly rational agents. Attentive and fully rational agents are more likely to form correct expectations that ozone norms are worse in June than May and intrinsically engage in an optimal level of defensive investment in each 122 month. Comparatively, inattentive or boundedly rational agents may not form accurate expectations (or expectations at all) and thus may not engage in an optimal level of defensive investment. Alerting agents to high ozone days, thus has two effects. First, both sets of agents would now potentially be able to respond to contemporaneous ozone shocks and engage in PDI as measured by Equations (3.2) and (3.3). Second, because alerts would likely occur more often in June than May, these discrete alert events may be easier to recognize and recall, helping inattentive or boundedly rational agents to form clearer expectations that ozone norms are worse in June than in May, and thus update their IDI strategy to better match this reality. B. Robustness Checks Selection into Alert Programs & Parallel Trends — As discussed in Section IV, one potential concern with Equation (3.3) is that the “assignment” of an alert program is not necessarily exogenous to the county. Thus, the larger magnitude effects of ozone shocks and norms in alert program counties should not be interpreted as the causal effect of implementing an alert program, but as the heterogeneous causal effects of ozone shocks and norms across the two different sets of counties. The effect of implementing an alert program on IDI, (β S P −β N P ), can however, be interpreted as causal if it satisfies the same parallel trends identifying assumption as a difference-in-differences approach. To that end, Table 3.3 reports the results of three alternative sample restrictions used to examine the parallel trends assumption. First, one subset of alert program counties had implemented their program before the sample period began, thus there are no “pre” or “post” periods for these counties as there would be in a traditional difference-in-differences setting. Column (1) thus reports results of estimating Equation (3.3) when restricting the sample to include only (i) the “nevertreated” counties which never implemented an alert program (625 counties) and (ii) the “always-treated” counties which had a program in place throughout the entire sample (102 counties). Results are of similar magnitude and statistically indistinguishable from the full 123 sample estimates. Second, the other subset of alert program counties implemented their alert program at some point within the sample period, such that there are distinct pre- and postperiods for these counties. Column (2) reports results when restricting the sample to include only (i) the same 625 “never-treated” counties, and (ii) the “staggered-treatment” counties which implemented an alert program at some point within the sample period (149 counties). The level differences between the effects of ozone shocks and norms in alert program counties versus counties without alert programs is more pronounced, but as in column (1), none of the coefficients are statistically indistinguishable from the full sample results. Similarly, while the alert day effect is somewhat larger, the difference is not statistically significant. As in column (1), the program effect on IDI remains positive and statistically insignificant. Finally, by restricting the sample to include the “never-treated” counties and only the “pre-period” for “staggered-treatment” counties – i.e., only the observations from before the alert program was implemented – column (3) reports the results of an analysis which mimics a “pre-trend” analysis from traditional difference-in-differences approaches. Note that due to this sample restriction, the specification changes slightly from Equation (3.3), as there are now no observations where an alert program was active. Thus, the indicator variable used to interact with the ozone shock and norm variable is redefined to be equal to one if the county ever implemented an alert program during the full sample period and is zero otherwise. The coefficients on the ozone norm and shock in counties which never implemented an alert program remain statistically indistinguishable from the full sample estimates, although interestingly both the shock and norm coefficients in the “staggeredtreatment, pre-period” counties are much smaller in magnitude and statistically different from the full sample estimates.49 The estimated difference between IDI in the two sets of counties remains positive and statistically insignificant, suggesting that there may be some inherent underlying difference in IDI behaviors between counties which never implemented an alert day program and counties which eventually did. This inherent difference in IDI 49Note that by the construction of this sample, there are zero observations that include an alert day, and thus an estimate of β A cannot be recovered. 124 Table 3.3: Parallel Trends Alert Program Alert Program Alert Program Persistent Status Switching Status Pre-trend Only (1) (2) (3) Ozone Shock (ppb) 7.448*** 4.881*** 4.348*** (1.581) (1.298) (1.570) Ozone Norm (ppb) 4.130*** 3.085*** 3.716*** (0.618) (0.562) (0.631) Shock x Alert Program 8.968*** 10.355*** 2.761** (1.663) (1.989) (1.290) Norm x Alert Program 4.237*** 6.575*** 0.782* (1.159) (0.935) (0.460) Alert Day −73.140*** −110.277*** (18.707) (25.703) Implied IDI 3.318** 1.796 0.632 (1.557) (1.251) (1.458) Implied PDI 1.596*** 2.437*** (0.408) (0.568) Implied IDI x Alert Counties 4.731** 3.780** 1.980* (2.107) (1.836) (1.168) Program Effect on IDI 1.413 1.983 1.348 (2.108) (1.812) (1.216) Observations 1,424,893 1,532,826 1,316,839 F-statistic 172 140 194 Notes: Table reports IV estimates of Equation (3.3) when restricting the estimating sample based on alert program implementation. Column (1) includes only counties that were either “never treated” or “always treated” with an alert program. Column (2) includes only counties that were either “never treated” or that implemented an alert program at some point during the sample period. Column (3) includes counties that were “never treated” and the “pre-treatment” periods only for counties that eventually implemented an alert program during the sample period. The dependent variable, variables of interest, full list of included controls, and county weights are the same as in Table 3.2 column (4). Standard errors are clustered at the county level and reported in parentheses. ***, **, and * represent significance at the 1%, 5% and 10% level, respectively. 125 during the “pre-trend” period is approximately similar in magnitude to the full sample estimated effect of implementing an alert program on IDI, which would appear to suggest that there may not in fact be any double-dividend benefit from such programs, with the full benefit of alert day programs captured by PDI. 50 Importantly, regardless of chosen sample restriction the estimated impact of alert programs on IDI never implies that PDI crowds out IDI, and thus, at worst, PDI is strictly complimentary to IDI even if it doesn’t induce a double dividend benefit on IDI which has been shown to be the case with more comprehensive air quality information campaigns (Barwick et al., 2024). Agents’ Expectations & Defining Ozone Norms — Unlike in the climatology literature, which Bento et al. (2023) defer to when proposing their decomposition of climate norms and weather shocks, there does not appear to exist any working definitions in the atmospheric science, epidemiology, or behavioral economics literature of what might constitute a “pollution norm.” Particularly, it is unclear at what frequency agents may recall average pollution levels (such as monthly) or the length of time and periodicity at which agents may update their prior beliefs about these pollution norms (such as the most recent 5 years, updating every year). If agents suffer from recency bias (Tversky and Kahneman, 1973, e.g.,), then they may only use the most recent year(s) of information in forming their expectations of ozone norms. Alternatively, if local air pollution has been rapidly improving or worsening in recent years, even rational and informed agents may view recent year(s) as more informative than prior years when considering ozone norms. At the same time, agents may form their expectations over much longer periods, internalizing information from the previous decade(s). To examine whether the results are sensitive to how the ozone norm is defined, Table 3.4 reports results of specifications which estimate Equation (3.3) when replacing the 5-year MA ozone norm with a 1-, 3-, 10-, and 20-year MA in columns (1), (2), (3), and (4) respectively. 50Specifically, if taking the difference between the full sample estimate of the program effect on IDI, and the differential IDI captured by this pre-trend subsample, in the spirit of a triple-difference approach, the effect of implementing an alert program on IDI would be approximately zero. 126 Table 3.4: Moving Average Norm Alternative Lengths 1-year MA 3-year MA 10-year MA 20-year MA (1) (2) (3) (4) Ozone Shock (ppb) 7.195*** 6.311*** 6.066*** 6.126*** (1.211) (1.267) (1.433) (1.548) Ozone Norm (ppb) 2.647*** 3.364*** 3.366*** 3.255*** (0.473) (0.537) (0.511) (0.499) Shock x Alert Program 11.142*** 10.243*** 10.380*** 10.246*** (1.661) (1.622) (1.676) (1.651) Norm x Alert Program 5.188*** 6.065*** 5.714*** 5.696*** (0.693) (0.842) (0.744) (0.698) Alert Day −100.942*** −95.794*** −98.750*** −96.908*** (17.781) (17.409) (17.404) (16.983) Implied IDI 4.548*** 2.947** 2.700* 2.871* (1.087) (1.240) (1.467) (1.607) Implied PDI 2.222*** 2.108*** 2.174*** 2.133*** (0.391) (0.383) (0.383) (0.374) Implied IDI x Alert Counties 5.953*** 4.178** 4.666*** 4.551*** (1.536) (1.670) (1.611) (1.527) Program Effect on IDI 1.405 1.231 1.966 1.680 (1.457) (1.573) (1.547) (1.544) Observations 1,765,834 1,775,605 1,778,730 1,778,861 F-statistic 208 173 183 193 Notes: Table reports IV estimates of Equation (3.3) when using a 1-, 3-, 10-, or 20-year moving average for constructing the monthly ozone norm rather than the 5-year moving average used in the central analysis, with results reported in columns (1), (2), (3), and (4) respectively. Recall that due to data limitations for wind direction, the 10- and 20-year moving average models in columns (3) and (4) continue to use a 5-year moving average when constructing the proportional wind direction instruments for the ozone norm. The dependent variable, variables of interest, full list of included controls, and county weights are the same as in Table 3.2 column (4). Standard errors are clustered at the county level and reported in parentheses. ***, **, and * represent significance at the 1%, 5% and 10% level, respectively. 127 In columns (1) and (2) the historical wind direction from the corresponding 1- and 3-years of prior data is used when constructing the IV for the ozone norm. Because wind direction data is not available prior to 1999, however, the 10- and 20-year ozone norms continue to use the 5-year wind direction norm IV. Across all four models, the estimated effects of ozone shocks, ozone norms, and alert day effects remain of similar magnitude and are statistically indistinguishable from the central estimate reported in column (4) of Table 3.2. Likewise, IDI, PDI and the effect of implementing an alert program on IDI all remain of qualitatively similar magnitude and statistical significance as the central results. In both counties with and without an alert program, IDI appears to be slightly larger when using the 1-year norm, suggesting that, if anything, agents may tend towards using the most recent prior year’s observed ozone levels when forming expectations over the current year’s ozone concentrations. Controlling for Other Air Pollutants — Other air pollutants can also cause health impacts, including carbon monoxide (CO), sulfur-dioxide (SO2), nitrogen dioxide (NO2), and fine particulate matter (PM2.5), and may be co-transported with ozone or ozone precursors.51 In particular, NO2 is an ozone precursor. Recall, however, that the estimating equation considers the effects of ozone shocks and ozone norms on four-day health outcomes. Because NO2 will react with VOCs and sunlight to form ozone over the four-day period, it is not possible to disentangle its independent effects from those of ozone.52 At the same time, NOx can also convert to particulate nitrate, a component of PM2.5, at several percent per hour (Lin and Cheng, 2007), such that NOx transported into a region could lead to increases in both PM2.5 and ozone. Thus, PM2.5 and ozone may also be correlated, but through a mediating relationship with NOx rather than direct effects on each other. 51These pollutants may, however, be produced in different locations or carried at different rates or distances by wind, allowing each to be separately instrumented using wind direction in the same estimating equation. 52For example, NO2 can convert to NO + O in the presence of sunshine, wherein the singular O molecule can combine with O2 and energy from VOCs to form O3 (ozone). The exact rate at which NO2 reacts and forms ozone is dependent on the characteristics of the local atmospheric chemistry, such as prevalence and ratios of VOCs, NOx, peroxides, ozone, and meteorological conditions such as sunlight, temperature, and humidity (Sillman and He, 2002; Leighton, 2012). 128 Table 3.5: Controlling for Other Air Pollutants SO2 CO PM2.5 All Three (1) (2) (3) (4) Ozone Shock (ppb) 6.429** 5.795** 3.679 4.502 (3.167) (2.536) (3.570) (3.818) Ozone Norm (ppb) 3.805*** 3.627*** 3.400*** 3.311*** (0.889) (0.902) (0.915) (0.955) Shock x Alert Program 12.104*** 11.730*** 9.485*** 10.682*** (3.031) (2.842) (3.335) (3.637) Norm x Alert Program 6.184*** 6.177*** 5.829*** 5.902*** (1.205) (1.114) (1.343) (1.300) Alert Day −109.450*** −108.887*** −108.008*** −116.538*** (27.062) (26.943) (24.421) (27.134) Implied IDI 2.624 2.168 0.278 1.191 (3.114) (2.705) (3.394) (3.701) Implied PDI 2.384*** 2.372*** 2.353*** 2.539*** (0.590) (0.587) (0.532) (0.591) Implied IDI x Alert Counties 5.919* 5.552* 3.657 4.780 (3.088) (3.116) (3.126) (3.461) Program Effect on IDI 3.295 3.384 3.378 3.588 (2.798) (2.816) (2.774) (2.897) SO2 (ppm) −21.145 −30.063 (21.830) (22.604) CO (ppm) −195.156 −228.647 (218.616) (226.559) PM2.5 (µg/m3 ) 4.589 8.861 (5.691) (5.996) Observations 235,968 235,968 235,968 235,968 F-statistic 34 34 34 34 Notes: Table reports IV estimates of Equation (3.3) with the addition of the endogenous variables SO2, CO, and PM2.5, individually in columns (1), (2) and (3), and altogether in column (4), which are similarly instrumented using wind direction. The dependent variable, variables of interest, full list of included controls, and county weights are the same as in Table 3.2 column (4). Standard errors are clustered at the county level and reported in parentheses. ***, **, and * represent significance at the 1%, 5% and 10% level, respectively. 129 Restricting the sample to include only county-days where readings for ozone, SO2, CO, and PM2.5 are simultaneously available, Table 3.5 reports the results of specifications that include potentially endogenous variables SO2, CO, PM2.5, and all three into Equation (3.3).53 Estimated effects on ozone shocks, norms, and alert days remain similar to the full sample estimates when including SO2 or CO, reported in columns (1) and (2) respectively. IDI also remains of similar magnitude, though now is statistically insignificant due to much larger standard errors, while PDI is both statistically significant and of similar magnitude. When controlling for PM2.5, reported in column (3), the effect of an ozone shock in counties without an alert program becomes smaller in magnitude, and statistically insignificant, although the effect of the ozone norm remains similar in magnitude and statistical significance to the full sample estimates. The effects of ozone shocks and norms in alert program counties are also of similar magnitude to the full sample estimates and are statistically significant. Finally, column (4) reports estimates when including all three control pollutants, with results similar to those reported in column (3). Further Robustness Checks & Examining Healthcare Costs — In addition to the above robustness checks, Table 3.6 reports the results of additional analyses examining the sensitivity of the central estimates to alternative specifications for the standard errors in columns (1) and (2) and geographically expanding the defined alert day announcements and program areas to encompass all counties within the respective CBSA in column (3). Additionally, column (4) reports results when replacing the outcome variable with the total cost of healthcare attributable to the four-day hospitalization rate for cardiovascular and respiratory diagnosis. First, as both the ozone norm and ozone shock could be considered generated regressors, column (1) reports results when estimating the standard errors via the Bayesian Bootstrap method (Rubin, 1981).54 Second, while the main outcome variable and regressors of interest 53For simplicity, the control pollutants are included as simply their daily averages. While the estimating model could feasibly include a similar decomposition of the control pollutants into shocks and norms, this would not only be computationally burdensome, requiring twelve additional first-stage regressions, but it would further reduce the estimating sample, as PM2.5 only began being widely monitored in the mid-2000’s restricting the number of observations for which 5-year moving average PM2.5 norms could be calculated. 54Specifically, the estimation approach approximates the block-bootstrap method by weighting all obser130 Table 3.6: Further Robustness Checks & Alternative Health Outcomes Bootstrap Monitor Grp. CBSA-level Associated Std. Errors Std. Errors Alert Program Cost ($1,000’s) (1) (2) (3) (4) Ozone Shock (ppb) 6.661*** 6.661*** 5.730*** 103.050*** (1.243) (1.992) (1.533) (24.358) Ozone Norm (ppb) 3.262*** 3.262*** 2.386*** 58.571*** (0.631) (0.835) (0.658) (9.104) Shock x Alert Program 10.593*** 10.593*** 9.575*** 166.316*** (1.361) (2.857) (1.274) (28.442) Norm x Alert Program 5.873*** 5.873*** 5.108*** 97.218*** (0.944) (1.170) (0.613) (14.311) Alert Day −98.535*** −98.535*** −80.684*** −1630.503*** (15.771) (31.536) (11.560) (299.958) Implied IDI 3.399*** 3.399 3.344** 44.479* (1.266) (2.120) (1.656) (24.112) Implied PDI 2.167*** 2.167*** 1.774*** 35.855*** (0.347) (0.693) (0.254) (6.596) Implied IDI x Alert Cnty 4.720*** 4.720 4.467*** 69.098** (1.615) (3.050) (1.268) (28.492) Program Effect on IDI 1.321 1.321 1.123 24.618 (1.646) (2.383) (1.495) (23.387) Observations 1,762,556 1,762,556 1,762,556 1,762,556 F-statistic 176 176 198 176 Notes: Table reports IV estimates of Equation (3.3) under various alternative specifications. Columns (1) and (2) present results when standard errors are estimated via Bayesian Bootstrap (Rubin, 1981) or clustered at the monitor-group level. Column (3) reports estimates when assigning the alert day and alert program indicator variables to the entire CBSA associated with the latitude and longitude, rather than only the singular county. Column (4) replaces the hospitalizations dependent variable with the total cost of care associated with those hospitalizations, presented in $1,000 per million beneficiaries. Other than the specific listed changes, the dependent variable, variables of interest, full list of included controls, and county weights are the same as in Table 3.2 column (4). Standard errors are block-bootstrapped using the Bayesian approach in column (1), clustered at the monitor-group in column (2), clustered at the CBSA level in column (3), clustered at the county level in column (4) and are reported in parentheses. ***, **, and * represent significance at the 1%, 5% and 10% level, respectively. 131 vary at the county level, in the first-stage regression the effect of wind direction is constrained to have the same effect across an entire monitor group. Thus, column (2) reports standard errors when clustering at the monitor-group level rather than the county level. Third, column (3) reports results of the main specification when redefining the Alertit and P rogiy indicators to include all counties within the CBSA associated with the alert and alert program. While in some areas alert programs may be county-specific, in other areas they may span multiple counties all under the jurisdiction of the same air control authority.55 Standard errors estimated via the Bayesian bootstrap are similar in magnitude with the central results, but increase by between 50-80 percent when clustering at the monitor-group level such that although the main coefficients remain strongly significant the implied effects of IDI become statistically insignificant. When expanding alerts and alert programs to the CBSA level, coefficients are slightly smaller in magnitude but remain statistically significant and indistinguishable from the central estimates. Finally, column (4) reports results of the impact of ozone shocks and norms on healthcare costs. Qualitatively, the results mirror trends in the central estimate results examining hospitalizations. Specifically, a 10 ppb ozone shock leads to an increase in the four-day total cost of care of $1.03 and $1.66 million, per million beneficiaries, in non-alert and alert program counties respectively – a 7.65 and 12.34 percent increase from the sample average four-day cost. By comparison, a 10 ppb increase in the ozone norm leads to an additional cost of $0.58 and $0.97 million, per million beneficiaries, in non-alert and alert program counties respectively – a 4.35 to 7.21 percent increase from the average four-day cost. Taken together, this implies that IDI reduces healthcare costs from ozone by 40-45 percent. At the same time, alert days are associated with a $1.63 million decrease in four-day total costs, per million beneficiaries, implying that for a 10 ppb increase in ozone, PDI reduces costs by $0.36 million per million beneficiaries, or about half of the effect of IDI in alert-program vations within each distinct county with a random draw from a multinomial distribution where each county has a uniform chance of being selected. 55In this specification standard errors are clustered at the CBSA level rather than county level to reflect this change. 132 counties. Once again, the results indicate no evidence that PDI crowds out IDI. VI. Concluding Remarks Understanding how air pollution effects health care use and associated costs is essential for designing optimal environmental policies. Similarly, understanding how individuals adjust to air pollution to avoid exposure or reduce the health impacts of exposure is another essential input. Prior literature examining both of these aspects has highlighted the large healthcare burden of air pollution, as well as many of the existing defensive investment strategies that individuals pursue in order to protect themselves. By adapting and combining two recently proposed methodological approaches in the economics literature (Deryugina et al., 2019; Bento et al., 2023) this study simultaneously investigates both the causal health impacts of ambient ozone pollution and the effectiveness of two distinct channels of defensive investment in the same estimating equation, allowing for direct statistical comparisons between the magnitudes of each effect. First, by decomposing daily ambient ozone into acute shocks and expected norms, the analysis is able to examine the potentially different impacts of ozone exposure on health when that pollution is or isn’t expected by individuals. Specifically, exploiting both shortrun (daily) and long-run (monthly, multi-year moving averages) variation in ambient ozone pollution caused by changes in wind direction, the employed empirical approach simultaneously estimates the effects of both acute shocks and expected norms, allowing for a direct comparison of both channels. The results indicate, perhaps unsurprisingly, that when pollution is expected it has lower health impacts, likely because agents are better able to protect themselves against exposure through either avoiding it entirely or engaging in defensive investments such as wearing a mask, using an air purifier, or taking medications. This paper refers to these types of behavioral adjustments as “intrinsic defensive investments” (IDI) – with the empirical analysis inferring the sum effectiveness of all channels of IDI as the difference between the estimated effect of the daily shock and the monthly norm. IDI is 133 found to have statistically significant and economically meaningful effects. For the same 10 ppb increase in daily ambient ozone concentration, if that increase is expected IDI reduces the hospitalizations attributable to this increase by 46.5%, on average. At the same time, in many US counties air quality alert programs provide individuals with information regarding predicted acute pollution shocks with up to few days advance notice, allowing them to potentially engage in defensive investment activities. This paper refers to these activities as “policy-induced defensive investments” (PDI), finding that PDI, while smaller in magnitude than IDI, is still of economically and statistically significant magnitude. On average, alert days decrease hospitalizations in counties with these programs by 8.7%, holding constant the effects of ozone shocks, norms, and included controls. Putting this in similar terms as IDI, for a 10 ppb increase in daily ozone, if an alert was issued the ozone increase will cause 20.5% fewer hospitalizations. Importantly, while there may be concern that individuals could become over-reliant on alerts, such that PDI would crowd out IDI, the analysis finds no evidence of this. In fact, the direction of the effect, though statistically insignificant, is suggestive of a potential double-dividend benefit of alert programs, with higher levels of IDI in counties which had implemented an alert program compared to counties without such a program. Thus, at worst, PDI caused by alert programs is complimentary to individuals’ own IDI, and may have additional benefits. Finally, this study focuses primarily on ambient ozone because the clear seasonality in local ozone concentrations makes it an ideal setting for examining the concept of IDI and drawing clear comparisons with PDI. This does not, however, preclude the method from being applied to the study of other air pollutants, so long as their local concentration levels follow a predictable pattern that individuals might use to form expectations of pollution norms. Furthermore, the decomposition of ozone employed in this analysis was itself adapted from the climate economics literature, illustrating the generalizability of this method for analyzing behavioral adjustments in a variety of empirical settings where individuals may be subject to both long-run (expected) norms and short-run (unexpected) shocks. 134 Bibliography Acemoglu, Daron, and Robert Shimer. 1999. “Efficient unemployment insurance.” Journal of political Economy, 107(5): 893–928. Aldy, Joseph E., and Richard J. Zeckhauser. 2020. “Three Prongs for Prudent Climate Policy.” NBER Working Paper #26991. Aldy, Joseph E., Maximilian Auffhammer, Maureen L. Cropper, Arthur G. Fraas, and Richard Morgenstern. 2020. “Looking Back at Fifty Years of the Clean Air Act.” NBER Working Paper #26687. Annan, Francis, and Wolfram Schlenker. 2015. “Federal Crop Insurance and the Disincentive to Adapt to Extreme Heat.” American Economic Review: Papers & Proceedings, 105(5): 262–66. Auffhammer, Maximilian. 2018a. “Climate Adaptive Response Estimation: Short and Long Run Impacts of Climate Change on Residential Electricity and Natural Gas Consumption Using Big Data.” NBER Working Paper #24397. Auffhammer, Maximilian. 2018b. “Quantifying Economic Damages from Climate Change.” Journal of Economic Perspectives, 32(4): 33–52. Auffhammer, Maximilian, and Ryan Kellogg. 2011. “Clearing the Air? The Effects of Gasoline Content Regulation on Air Quality.” American Economic Review, 101(6): 2687– 2722. BAAQMD, Bay Area Air Quaility Management District. 2017. “2017 Clean Air Plan, Volume 1.” BAAQMD, Bay Area Air Quaility Management District, San Francisco, CA. Barreca, Alan, Karen Clay, Olivier Deschenes, Michael Greenstone, and Joseph S. Shapiro. 2016. “Adapting to Climate Change: The Remarkable Decline in the US Temperature-Mortality Relationship over the Twentieth Century.” Journal of Political Economy, 124(1): 105–159. Bartik, Timothy J. 1988. “Evaluating the benefits of non-marginal reductions in pollution using information on defensive expenditures.” Journal of environmental economics and management, 15(1): 111–127. 135 Barwick, Panle Jia, Shanjun Li, Deyu Rao, and Nahim Bin Zahur. 2018. “The morbidity cost of air pollution: evidence from consumer spending in China.” Available at SSRN 2999068. Barwick, Panle Jia, Shanjun Li, Liguo Lin, and Eric Yongchen Zou. 2024. “From fog to smog: The value of pollution information.” American Economic Review, 114(5): 1338– 1381. Baxter, Marianne, and Robert G. King. 1999. “Measuring Business Cycles: Approximate Band-Pass Filters for Economic Time Series.” Review of Economics and Statistics, 81(4): 575–93. Becker, Randy, and Vernon Henderson. 2000. “Effects of Air Quality Regulations on Polluting Industries.” Journal of Political Economy, 108(2): 379–421. Bell, Michelle L, Aidan McDermott, Scott L Zeger, Jonathan M Samet, and Francesca Dominici. 2004. “Ozone and short-term mortality in 95 US urban communities, 1987-2000.” Jama, 292(19): 2372–2378. Bell, Michelle L, Roger D Peng, and Francesca Dominici. 2006. “The exposure– response curve for ozone and risk of mortality and the adequacy of current ozone regulations.” Environmental health perspectives, 114(4): 532–536. Bentayeb, Malek, Marzia Simoni, Nour Baiz, Dan Norback, Sandra Baldacci, Sara Maio, Giovanni Viegi, Isabella Annesi-Maesano, Geriatric Study in Europe on Health Effects of Air Quality in Nursing Homes (GERIE) Group, et al. 2012. “Adverse respiratory effects of outdoor air pollution in the elderly.” The international journal of tuberculosis and lung disease, 16(9): 1149–1161. Bento, Antonio, Daniel Kaffine, Kevin Roth, and Matthew Zaragoza-Watkins. 2014. “The Effects of Regulation in the Presence of Multiple Unpriced Externalities: Evidence from the Transportation Sector.” American Economic Journal: Economic Policy, 6(3): 1–29. Bento, Antonio, Mehreen Mookerjee, and Edson Severnini. 2020. “RegulationInduced Adaptation: Evidence from Ambient Ozone and the Clean Air Act.” Mimeo. Bento, Antonio M, Noah Miller, Mehreen Mookerjee, and Edson Severnini. 2023. “A unifying approach to measuring climate change impacts and adaptation.” Journal of Environmental Economics and Management, 121: 102843. Bento, Antonio, Noah Miller, Mehreen Mookerjee, and Edson Severnini. 2020. “A Unifying Approach to Measuring Climate Change Impacts and Adaptation.” NBER Working Paper #27247. Blanc, Elodie, and Wolfram Schlenker. 2017. “The Use of Panel Models in Assessments of Climate Impacts on Agriculture.” Review of Environmental Economics and Policy, 11(2): 258–279. 136 Bleakley, Hoyt, and Sok Chul Hong. 2017. “Adapting to the Weather: Lessons from U.S. History.” Journal of Economic History, 77(3): 756–795. Bovenberg, A Lans, and Lawrence H Goulder. 1996. “Optimal environmental taxation in the presence of other taxes: general-equilibrium analyses.” The American Economic Review, 86(4): 985–1000. Brook, Robert D, Bruce Urch, J Timothy Dvonch, Robert L Bard, Mary Speck, Gerald Keeler, Masako Morishita, Frank J Marsik, Ali S Kamal, Niko Kaciroti, et al. 2009. “Insights into the mechanisms and mediators of the effects of air pollution exposure on blood pressure and vascular function in healthy humans.” Hypertension, 54(3): 659–667. Burke, Marshall, and Kyle Emerick. 2016. “Adaptation to Climate Change: Evidence from US Agriculture.” American Economic Journal: Economic Policy, 8(3): 106–40. Carleton, Tamma, Amir Jina, Michael Delgado, Michael Greenstone, Trevor Houser, Solomon Hsiang, Andrew Hultgren, Robert E. Kopp, Kelly E. McCusker, Ishan Nath, James Rising, Ashwin Rode, Hee Kwon Seo, Arvid Viaene, Jiacan Yuan, and Alice Tianbo Zhang. 2022. “Valuing the Global Mortality Consequences of Climate Change Accounting for Adaptation Costs and Benefits.” Quarterly Journal of Economics, 137(4): 2037–2105. Carleton, Tamma, Michael Delgado, Michael Greenstone, Trevor Houser, Solomon Hsiang, Andrew Hultgren, Amir Jina, Robert Kopp, Kelly McCusker, Ishan Nath, James Rising, Ashwin Rode, Hee Kwon Seo, Justin Simcock, Arvid Viaene, Jiacan Yuan, and Alice Zhang. 2019. “Valuing the Global Mortality Consequences of Climate Change Accounting for Adaptation Costs and Benefits.” University of Chicago Becker Friedman Institute Working Paper No. 2018-51. Chay, Kenneth Y., and Michael Greenstone. 2005. “Does Air Quality Matter? Evidence from the Housing Market.” Journal of Political Economy, 113(2): 376–424. Chetty, Raj. 2009. “Sufficient Statistics for Welfare Analysis: A Bridge Between Structural and Reduced-Form Methods.” Annual Review of Economics, 1: 451–488. Chetty, Raj, John N. Friedman, Nathaniel Hendren, Maggie R. Jones, and Sonya R. Porter. 2018. “The Opportunity Atlas: Mapping the Childhood Roots of Social Mobility.” NBER Working Paper #25147. Christiano, Lawrence J., and Terry J. Fitzgerald. 2003. “The Band Pass Filter.” International Economic Review, 44(2): 435–465. Climatology Office, Wisconsin State Climatology Office. 2003. “Climatic Normals.” http: // aos. wisc. edu/ ~ sco/ normals. html . 137 Cohen, Aaron J, Michael Brauer, Richard Burnett, H Ross Anderson, Joseph Frostad, Kara Estep, Kalpana Balakrishnan, Bert Brunekreef, Lalit Dandona, Rakhi Dandona, et al. 2017. “Estimates and 25-year trends of the global burden of disease attributable to ambient air pollution: an analysis of data from the Global Burden of Diseases Study 2015.” The lancet, 389(10082): 1907–1918. Correia, Sergio. 2017. “Linear models with high-dimensional fixed effects: An efficient and feasible estimator.” Unpublished manuscript, http://scorreia. com/research/hdfe. pdf (last accessed 25 October 2019), 4(2). Currie, Janet, and Matthew Neidell. 2005. “Air Pollution and Infant Health: What Can We Learn From California’s Recent Experience?” Quarterly Journal of Economics, 120(3): 1003–1030. Cutter, W. Bowman, and Matthew Neidell. 2009. “Voluntary Information Programs and Environmental Regulation: Evidence from ’Spare the Air’.” Journal of Environmental Economics and Management, 58(3): 253–265. Dawson, John P., Peter J. Adams, and Spyros N. Pandisa. 2007. “Sensitivity of Ozone to Summertime Climate in the Eastern USA: A Modeling Case Study.” Atmospheric Environment, 41(7): 1494–1511. Dell, Melissa, Benjamin F. Jones, and Benjamin A. Olken. 2012. “Temperature Shocks and Economic Growth: Evidence from the Last Half Century.” American Economic Journal: Macroeconomics, 4(3): 66–95. Dell, Melissa, Benjamin F. Jones, and Benjamin A. Olken. 2014. “What Do We Learn from the Weather? The New Climate-Economy Literature.” Journal of Economic Literature, 52(3): 740–798. Deryugina, Tatyana, Garth Heutel, Nolan H Miller, David Molitor, and Julian Reif. 2019. “The mortality and medical costs of air pollution: Evidence from changes in wind direction.” American Economic Review, 109(12): 4178–4219. Deschenes, Olivier, and Michael Greenstone. 2007. “The Economic Impacts of Climate Change: Evidence from Agricultural Output and Random Fluctuations in Weather.” American Economic Review, 97(1): 354–85. Deschenes, Olivier, and Michael Greenstone. 2011. “Climate Change, Mortality, and Adaptation: Evidence from Annual Fluctuations in Weather in the US.” American Economic Journal: Applied Economics, 3(4): 152–85. Deschenes, Olivier, Michael Greenstone, and Joseph S. Shapiro. 2017. “Defensive Investments and the Demand for Air Quality: Evidence from the NOx Budget Program.” American Economic Review, 107(10): 2958–89. 138 Di, Qian, Lingzhen Dai, Yun Wang, Antonella Zanobetti, Christine Choirat, Joel D Schwartz, and Francesca Dominici. 2017. “Association of short-term exposure to air pollution with mortality in older adults.” Jama, 318(24): 2446–2456. Dominici, Francesca, Roger D Peng, Michelle L Bell, Luu Pham, Aidan McDermott, Scott L Zeger, and Jonathan M Samet. 2006. “Fine particulate air pollution and hospital admission for cardiovascular and respiratory diseases.” Jama, 295(10): 1127– 1134. Doyle Jr, Joseph J. 2005. “Health insurance, treatment and outcomes: using auto accidents as health shocks.” Review of Economics and Statistics, 87(2): 256–270. Flegal, Jane A., Anna-Maria Hubert, David R. Morrow, and Juan B. MorenoCruz. 2019. “Solar Geoengineering: Social Science, Legal, Ethical, and Economic Frameworks.” Annual Review of Environment and Resources, 44: 399–423. Fowlie, Meredith, Stephen P Holland, and Erin T Mansur. 2012. “What do emissions markets deliver and to whom? Evidence from Southern California’s NOx trading program.” American Economic Review, 102(2). Frisch, Ragnar, and Frederick V. Waugh. 1933. “Partial Time Regressions as Compared with Individual Trends.” Econometrica, 1(4): 387–401. Gabaix, Xavier, and David Laibson. 2006. “Shrouded Attributes, Consumer Myopia, and Information Suppression in Competitive Markets.” Quarterly Journal of Economics, 121(2): 505–540. Garg, Teevrat, Gordon McCord, and Aleister Montfort. 2020. “Can Social Protection Reduce Environmental Damages?” Mimeo. Gibson, Matthew. 2019. “Regulation-induced pollution substitution.” Review of Economics and Statistics, 101(5): 827–840. Goodman-Bacon, Andrew. 2018. “Difference-in-Differences with Variation in Treatment Timing.” NBER Working Paper #25018. Goulder, Lawrence H. 2020. “Timing Is Everything: How Economists Can Better Address the Urgency of Stronger Climate Policy.” Review of Environmental Economics and Policy, 14(1): 143–156. Goulder, Lawrence H., and Ian W.H. Parry. 2008. “Instrument Choice in Environmental Policy.” Review of Environmental Economics and Policy, 2(2): 152–174. Goulder, Lawrence H., Ian W.H. Parry, Roberton C. III Williams, and Dallas Burtraw. 1999. “The Cost-effectiveness of Alternative Instruments for Environmental Protection in a Second-best Setting.” Journal of Public Economics, 72(3): 329–360. 139 Graff Zivin, Joshua, and Matthew Neidell. 2009. “Days of haze: Environmental information disclosure and intertemporal avoidance behavior.” Journal of Environmental Economics and Management, 58(2): 119–128. Graff Zivin, Joshua, and Matthew Neidell. 2012. “The Impact of Pollution on Worker Productivity.” American Economic Review, 102(7): 3652–3673. Grainger, Corbett, Andrew Schreiber, and Wonjun Chang. 2019. “Do Regulators Strategically Avoid Pollution Hotspots when Siting Monitors? Evidence from Remote Sensing of Air Pollution.” Mimeo. Greenstone, Michael. 2002. “The Impacts of Environmental Regulations on Industrial Activity: Evidence from the 1970 and 1977 Clean Air Act Amendments and the Census of Manufactures.” Journal of Political Economy, 110(6): 1175–1219. Griliches, Zvi. 1977. “Estimating the Returns to Schooling: Some Econometric Problems.” Econometrica, 45(1): 1–22. Harberger, Arnold C. 1964. “The Measurement of Waste.” American Economic Review Papers & Proceedings, 54(3): 58–76. Henderson, J. Vernon. 1996. “Effects of Air Quality Regulation.” American Economic Review, 86(4): 789–813. Heutel, Garth, Juan Moreno-Cruz, and Katharine Ricke. 2016. “Climate Engineering Economics.” Annual Review of Resource Economics, 8: 99–118. Heutel, Garth, Nolan H. Miller, and David Molitor. 2021. “Adaptation and the Mortality Effects of Temperature Across U.S. Climate Regions.” Review of Economics and Statistics, 103(4): 740–753. Heutel, Garth, Nolan H. Miller, and David Molitor. forthcoming. “Adaptation and the Mortality Effects of Temperature Across U.S. Climate Regions.” Review of Economics and Statistics. Howe, Peter D., Matto Mildenberger, Jennifer R. Marlon, and Anthony Leiserowitz. 2015. “Geographic variation in opinions on climate change at state and local scales in the USA.” Nature Climate Change, 5(6): 596–603. Hsiang, Solomon M. 2016. “Climate Econometrics.” Annual Review of Resource Economics, 8: 43–75. Hsiang, Solomon M., Marshall Burke, and Edward Miguel. 2013. “Quantifying the Influence of Climate on Human Conflict.” Science, 341(6151). Hsu, Joanne W, David A Matsa, and Brian T Melzer. 2018. “Unemployment insurance as a housing market stabilizer.” American Economic Review, 108(1): 49–81. 140 IPCC, Intergovernmental Panel on Climate Change. 2021. “Summary for Policymakers.” In Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. , ed. V. Masson Delmotte, P. Zhai, A. Pirani, S.L. Connors, C. P´ean, S. Berger, N. Caud, Y. Chen, L. Goldfarb, M.I. Gomis, M. Huang, K. Leitzell, E. Lonnoy, J.B.R. Matthews, T.K. Maycock, T. Waterfield, O. Yelek¸ci, R. Yu and B. Zhou. Cambridge, United Kingdom and New York, NY:Cambridge University Press. IPCC, Intergovernmental Panel on Climate Change. 2022. “Summary for Policymakers.” In Climate Change 2022: Impacts, Adaptation and Vulnerability. Contribution of Working Group II to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. , ed. H.-O. P¨ortner, D.C. Roberts, M. Tignor, E.S. Poloczanska, Mintenbeck K., A. Alegr´ıa, M. Craig, S. Langsdorf, S. L¨oshke, V. M¨oller, A. Okem and B. Rama. Cambridge, United Kingdom and New York, NY:Cambridge University Press. Isen, Adam, Maya Rossin-Slater, and Reed Walker. 2017. “Relationship between season of birth, temperature exposure, and later life wellbeing.” Proceedings of the National Academy of Sciences, 114(51): 13447–13452. Ito, Koichiro, and Shuang Zhang. 2020. “Willingness to pay for clean air: Evidence from air purifier markets in China.” Journal of Political Economy, 128(5): 1627–1672. Jacob, Daniel J., and Darrel A. Winner. 2009. “Effect of Climate Change on Air Quality.” Atmospheric Environment, 43(1): 51–63. Kaufmann, Robert K., Michael L. Mann, Sucharita Gopal, Jackie A. Liederman, Peter D. Howe, Felix Pretis, Xiaojing Tang, and Michelle Gilmore. 2017. “Spatial heterogeneity of climate change as an experiential basis for skepticism.” Proceedings of the National Academy of Sciences, 114(1): 67–71. Kelly, David L., Charles D. Kolstad, and Glenn T. Mitchell. 2005. “Adjustment Costs from Environmental Change.” Journal of Environmental Economics and Management, 50(3): 468–495. Kleven, Henrik J. forthcoming. “Sufficient Statistics Revisited.” Annual Review of Economics. Kolstad, Charles D., and Frances C. Moore. 2020. “Estimating the Economic Impacts of Climate Change Using Weather Observations.” Review of Environmental Economics and Policy, 14(1): 1–24. Kousky, Carolyn. 2012. “Informing Climate Adaptation: A Review of the Economic Costs of Natural Disasters, Their Determinants, and Risk Reduction Options.” Resources for the Future Discussion Paper 12-28. Krebs, Benjamin, and Simon Luechinger. 2024. “Air pollution, cognitive performance, and the role of task proficiency.” Journal of the Association of Environmental and Resource Economists, 11(4): 921–958. 141 Krupnick, Alan J., and Deirdre Farrell. 1996. “Six Steps to a Healthier Ambient Ozone Policy.” Resources for the Future Discussion Paper 96-13. Leighton, Philip. 2012. Photochemistry of air pollution. Elsevier. Lemoine, Derek. 2020. “Estimating the Consequences of Climate Change from Variation in Weather.” NBER Working Paper #25008. Lin, YC, and MT Cheng. 2007. “Evaluation of formation rates of NO2 to gaseous and particulate nitrate in the urban atmosphere.” Atmospheric Environment, 41(9): 1903–1910. Lipsey, R. G., and Kelvin Lancaster. 1956. “The General Theory of Second Best.” Review of Economic Studies, 24(1): 11–32. Lovell, Michael C. 1963. “Seasonal Adjustment of Economic Time Series and Multiple Regression Analysis.” Journal of the American Statistical Association, 58(304): 993–1010. Lucas, Robert E., Jr. 1972. “Expectations and the Neutrality of Money.” Journal of Economic Theory, 4: 103–124. Lucas, Robert E., Jr. 1976. “Econometric Policy Evaluation: A Critique.” CarnegieRochester Conference Series on Public Policy, 1: 19–46. Lucas, Robert E., Jr. 1977. “Understanding Business Cycles.” Carnegie-Rochester Conference Series on Public Policy, 5: 7–29. Marcus, Michelle, and Pedro H. C. Sant’Anna. 2021. “The Role of Parallel Trends in Event Study Settings: An Application to Environmental Economics.” Journal of the Association of Environmental and Resource Economists, 8(2): 235–275. Mauzerall, Denise L, Babar Sultan, Namsoug Kim, and David F Bradford. 2005. “NOx emissions from large point sources: variability in ozone production, resulting health damages and economic costs.” Atmospheric Environment, 39(16): 2851–2866. McCall, James, Jordan Macknick, and Daniel Hillman. 2016. “Water-Related Power Plant Curtailments: An Overview of Incidents and Contributing Factors.” National Renewable Energy Laboratory (NREL) Technical Report NREL/TP-6A20-67084. McClurkin, Janie D, Dirk E Maier, and Klein E Ileleji. 2013. “Half-life time of ozone as a function of air movement and conditions in a sealed container.” Journal of Stored Products Research, 55: 41–47. McGrath, Justin M., Amy M. Betzelberger, Shaowen Wang, Eric Shook, XinGuang Zhu, Stephen P. Long, and Elizabeth A. Ainsworth. 2015. “An Analysis of Ozone Damage to Historical Maize and Soybean Yields in the United States.” Proceedings of the National Academy of Sciences, 112(46): 14390–14395. Mendelsohn, Robert. 2016. “Measuring weather impacts using panel data.” Mimeo. 142 Mendelsohn, Robert, William D. Nordhaus, and Daigee Shaw. 1994. “The Impact of Global Warming on Agriculture: A Ricardian Analysis.” American Economic Review, 84(4): 753–71. Merel, Pierre, and Matthew Gammans. 2021. “Climate Econometrics: Can the Panel Approach Account for Long-Run Adaptation?” American Journal of Agricultural Economics, 103(4): 1207–1238. Michel-Kerjan, Erwann O. 2010. “Catastrophe economics: the national flood insurance program.” Journal of economic perspectives, 24(4): 165–86. MIT, Election Data & Science Lab. 2018. “County Presidential Election Returns 2000- 2016.” https://doi.org/10.7910/DVN/VOQCHQ, accessed on June 3, 2019. Moore, Frances C., and David B. Lobell. 2014. “Adaptation potential of European agriculture in response to climate change.” Nature Climate Change, 4: 610–614. Moore, Frances C., and David B. Lobell. 2015. “The fingerprint of climate trends on European crop yields.” PNAS, 112(9): 2670–2675. Moore, Frances C., Nick Obradovich, Flavio Lehner, and Patrick Baylis. 2019. “Rapidly Declining Remarkability of Temperature Anomalies May Obscure Public Perception of Climate Change.” Proceedings of the National Academy of Sciences, 116(11): 4905– 4910. Moretti, Enrico, and Matthew Neidell. 2011. “Pollution, Health, and Avoidance Behavior: Evidence from the Ports of Los Angeles.” Journal of Human Resources, 46(1): 154–75. Mukherjee, Anita, and Nicholas J Sanders. 2021. “The causal effect of heat on violence: Social implications of unmitigated heat among the incarcerated.” National Bureau of Economic Research. Muller, Nicholas Z., and Paul A. Ruud. 2018a. “What Forces Dictate the Design of Pollution Monitoring Networks?” Environmental Modeling & Assessment, 23(1): 1–14. Muller, Nicholas Z, and Paul A Ruud. 2018b. “What forces dictate the design of pollution monitoring networks?” Environmental Modeling & Assessment, 23: 1–14. Mullins, Jamie T., and Corey White. 2020. “Can access to health care mitigate the effects of temperature on mortality?” Journal of Public Economics, 191: 104259. Neidell, Matthew. 2009a. “Information, Avoidance Behavior, and Health: The Effect of Ozone on Asthma Hospitalizations.” Journal of Human Resources, 44(2): 450–78. Neidell, Matthew. 2009b. “Information, avoidance behavior, and health: the effect of ozone on asthma hospitalizations.” Journal of Human resources, 44(2): 450–478. 143 Nguyen, Angela-Maithy, Brian J Malig, and Rupa Basu. 2021. “The association between ozone and fine particles and mental health-related emergency department visits in California, 2005–2013.” PloS one, 16(4): e0249675. NOAA, National Oceanic & Atmospheric Administration. 2014. “National Oceanic and Atmospheric Administration (NOAA), Global Historical Climatology Network.” ftp://ftp.ncdc.noaa.gov/pub/data/ghcn/daily/by year/, accessed on November 30, 2014. Nordhaus, William. 2019. “Climate Change: The Ultimate Challenge for Economics.” American Economic Review, 109(6): 1991–2014. Olmstead, Alan L., and Paul W. Rhode. 2011. “Responding to Climatic Challenges: Lessons from U.S. Agricultural Development.” In The Economics of Climate Change: Adaptations Past and Present. , ed. Gary D. Libecap and Richard H. Steckel, Chapter 6, pp. 169–194. Chicago, IL:University of Chicago Press. Popp, David. 2003. “Pollution control innovations and the Clean Air Act of 1990.” Journal of Policy Analysis and Management, 22(4): 641–660. Popp, David. 2006. “International innovation and diffusion of air pollution control technologies: the effects of NOX and SO2 regulation in the US, Japan, and Germany.” Journal of Environmental Economics and Management, 51(1): 46–71. Ranson, Matthew. 2014. “Crime, Weather, and Climate Change.” Journal of Environmental Economics and Management, 67(3): 274–302. Reis, Ricardo. 2006. “Inattentive Producers.” Review of Economic Studies, 73(3): 793–821. Rubin, Donald B. 1981. “The bayesian bootstrap.” The annals of statistics, 130–134. Salvo, Alberto, and Yi Wang. 2017. “Ethanol-Blended Gasoline Policy and Ozone Pollution in Sao Paulo.” Journal of the Association of Environmental and Resource Economists, 4(3): 731–794. Schlenker, Wolfram, and Michael J. Roberts. 2009. “Nonlinear Temperature Effects Indicate Severe Damages to U.S. Crop Yields under Climate Change.” Proceedings of the National Academy of Sciences, 106(37): 15594–98. Schlenker, Wolfram, and W. Reed Walker. 2016. “Airports, Air Pollution, and Contemporaneous Health.” Review of Economic Studies, 83(2): 768–809. Schlenker, Wolfram, W. Michael Hanemann, and Anthony C. Fisher. 2005. “Will U.S. Agriculture Really Benefit from Global Warming? Accounting for Irrigation in the Hedonic Approach.” American Economic Review, 95(1): 395–406. Sexton, Steven E. 2012. “Paying for Pollution? How General Equilibrium Effects Undermine the ’Spare the Air’ Program.” Environmental and Resource Economics, 53: 553–575. 144 Sillman, Sanford, and Dongyang He. 2002. “Some theoretical results concerning O3- NOx-VOC chemistry and NOx-VOC indicators.” Journal of Geophysical Research – Atmospheres, 107(D22): 26–1–26–15. Solon, Gary. 1992. “Intergenerational Income Mobility in the United States.” American Economic Review, 82(3): 393–408. Stavins, R., J. Zou, T. Brewer, M. Conte Grand, M. den Elzen, M. Finus, J. Gupta, N. Hohne, M.-K. Lee, A. Michaelowa, M. Paterson, K. Ramakrishna, G. Wen, J. Wiener, and H. Winkler. 2014. “International Cooperation: Agreements and Instruments.” In Climate Change 2014: Mitigation of Climate Change. Contribution of Working Group III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change. , ed. O. Edenhofer, R. Pichs-Madruga, Y. Sokona, E. Farahani, S. Kadner, K. Seyboth, A. Adler, I. Baum, S. Brunner, P. Eickemeier, B. Kriemann, J. Savolainen, S. Schl¨omer, C. von Stechow, T. Zwickel and J.C. Minx. Cambridge, United Kingdom and New York, NY, USA: Cambridge University Press, Chapter 13, pp. 1001–1082. Stock, James H., and Motohiro Yogo. 2005. “Testing for weak instruments in Linear Iv regression.” Identification and Inference for Econometric Models, 80–108. United Kingdom:Cambridge University Press. Publisher Copyright: © Cambridge University Press 2005. Tol, Richard S. J. 2018. “The Economic Impacts of Climate Change.” Review of Environmental Economics and Policy, 12(1): 4–25. Towers, Sherry, Gerardo Chowell, Rasheed Hameed, Matthew Jastrebski, Maryam Khan, Jonathan Meeks, Anuj Mubayi, and George Harris. 2013. “Climate Change and Influenza: The Likelihood of Early and Severe Influenza Seasons Following Warmer Than Average Winters.” PLoS Currents Influenza, 5: ecurrents.flu.3679b56a3a5313dc7c043fb944c6f138. Tversky, Amos, and Daniel Kahneman. 1973. “Availability: A heuristic for judging frequency and probability.” Cognitive Psychology, 5(2): 207–232. USCFR, U.S. Code of Federal Regulations. 2005. “National Archives and Records Administration, Federal Register, Title 42 Public Health, Chapter 85, Subchapter 1, Part D, Subpart 1, Par. 7509.” USCFR, U.S. Code of Federal Regulations. 2013. “National Archives and Records Administration, Federal Register, Title 42 Public Health, Chapter 85, Subchapter 1, Part D, Subpart 2, Sec. 7511c.” USCFR, U.S. Code of Federal Regulations. 2022. “National Archives and Records Administration, Federal Register, Title 40 Protection of Environment, Chapter 1, Subchapter C, Part 81, Subpart E, Par. 81.457.” 145 USEPA, U.S. Environmental Protection Agency. 1979. “Revisions to the National Ambient Air Quality Standards for Photochemical Oxidants: Final Rulemaking.” Federal Register, 44(28): 8202–8237. February 8, 1979. USEPA, U.S. Environmental Protection Agency. 1997a. “National Ambient Air Quality Standards for Ozone: Final Rule.” Federal Register, 62(138): 38856–38896. July 18, 1997. USEPA, U.S. Environmental Protection Agency. 1997b. “Survey and Review of Episodic Control Programs in the United States.” Washington, DC: U.S. Environmental Protection Agency, Office of Air and Radiation, EPA 420-R-97-003. USEPA, U.S. Environmental Protection Agency. 2004. “Final Rule To Implement the 8-Hour Ozone National Ambient Air Quality Standard – Phase 1.” Federal Register, 69(84): 23951–24000. April 30, 2004. USEPA, U.S. Environmental Protection Agency. 2006. “Air Quality Criteria for Ozone and Related Photochemical Oxidants — Volume II.” Available at epa.gov/ttn/naaqs/aqmguide/collection/cp2/20060228 ord epa600 r05- 004bf ozone criteria document vol2.pdf, accessed on July 23, 2017. USEPA, U.S. Environmental Protection Agency. 2008. “National Ambient Air Quality Standards for Ozone: Final Rule.” Federal Register, 73(60): 16435–16514. March 27, 2008. USEPA, U.S. Environmental Protection Agency. 2009. “Assessment of the Impacts of Global Change on Regional U.S. Air Quality: A Synthesis of Climate Change Impacts on Ground-Level Ozone – An Interim Report of the U.S. EPA Global Change Research Program.” Washington, DC: U.S. Environmental Protection Agency, Office of Research and Development, National Center for Environmental Assessment, EPA/600/R-07/094F. USEPA, U.S. Environmental Protection Agency. 2015a. “National Ambient Air Quality Standards for Ozone: Final Rule.” Federal Register, 80(206): 65291–65468. October 26, 2015. USEPA, U.S. Environmental Protection Agency. 2015b. “Regulatory Impact Analysis of the Final Revisions to the National Ambient Air Quality Standards for Ground-Level Ozone.” Washington, DC: U.S. Environmental Protection Agency, Office of Air and Radiation and Office of Air Quality Planning and Standards, EPA-452/R-15-007. USEPA, U.S. Environmental Protection Agency. 2019. “Integrated Review Plan for the Review of the Ozone National Ambient Air Quality Standards.” Washington, DC: U.S. Environmental Protection Agency, Office of Air Quality Planning and Standards, EPA-452/R-19-002. 146 Vose, R.S., D.R. Easterling, K.E. Kunkel, A.N. LeGrande, and M.F. Wehner. 2017. “Temperature Changes in the United States.” In Climate Science Special Report: Fourth National Climate Assessment, Volume I. , ed. D.J. Wuebbles, D.W. Fahey, K.A. Hibbard, D.J. Dokken, B.C. Stewart, and T.K. Maycock, Chapter 6, pp. 185–206. Washington, DC:U.S. Global Change Research Program. Wang, Zhenxuan, and Junjie Zhang. 2023. “The value of information disclosure: Evidence from mask consumption in China.” Journal of Environmental Economics and Management, 122: 102865. White, Corey. 2021. “Measuring Social and Externality Benefits of Influenza Vaccination.” Journal of Human Resources, 56(3): 749–785. WMO. 2017. WMO guidelines on the calculation of climate normals. World Meteorological Organization Geneva, Switzerland. Zhang, Junjie, and Quan Mu. 2018. “Air pollution and defensive expenditures: Evidence from particulate-filtering facemasks.” Journal of Environmental Economics and Management, 92: 517–536. Zhang, Xin, Xi Chen, and Xiaobo Zhang. 2018. “The impact of exposure to air pollution on cognitive performance.” Proceedings of the National Academy of Sciences, 115(37): 9193–9197. Zhang, Yuzhong, and Yuhang Wang. 2016. “Climate-driven Ground-level Ozone Extreme in the Fall Over the Southeast United States.” Proceedings of the National Academy of Sciences, 113(36): 10025–10030. 147 Appendices Appendix A: A Unifying Approach to Measuring Climate Change Impacts and Adaptation This appendix provides details on the construction of the data, descriptive figures, and the tabular results of robustness tests and explorations of heterogeneity using alternate specifications. In Appendix A.1, we provide relevant background information on ozone as a local air pollutant in Section A.1.a, and further details on the sources of our data and construction of final variables in Section A.1.b. We additionally include maps of both weather and ozone monitoring station locations, illustrative figures of our decomposition of temperature and its relationship with ozone concentration, and tables of summary statistics. Appendix A.2 includes additional discussion of alternate specifications, split between those investigating robustness in Section A.2.a, and those examining heterogeneity in Section A.2.b. Appendix A.3 provides further details on the two sources of variation used to empirically identify the effect of changes in the climate norm on ozone concentration. 148 A.1. Additional Data Discussion This appendix section provides background information on ozone pollution in Section A.1.a, while Section A.1.b provides further details on the data sets discussed in Section III, as well as auxiliary data sets used in alternative specifications. It then includes relevant Figures and Tables as outlined below. Figure A.1.1.Temperature Relative to Baseline (1950-1979) Figure A.1.2.Comprehensive Location of Weather Monitors Figure A.1.3.Climate Norms and Shocks (semi-balanced sample) Figure A.1.4.Climate Norms and Shocks (main model sample) Figure A.1.5.Evolution of Maximum Ambient Ozone Concentration Figure A.1.6.Ozone Monitor Location by Decade of First Appearance Figure A.1.7.Ozone Monitors and their Matched Weather Monitors Figure A.1.8.Relationship between Ozone and Decomposed Temperature Figure A.1.9.Decomposition of Temperature Norms and Shocks (Los Angeles, All Years) Table A.1.1.Yearly Summary Statistics for Daily Maximum Temperature Table A.1.2.Yearly Summary Statistics for Ozone Monitoring Network Table A.1.3.Ozone Monitoring Season by State Table A.1.4.County Summary Statistics by Belief in Climate Change 149 A.1.a. Background Details on Ozone Background on Ozone — The ozone the U.S. EPA regulates as an air pollutant is mainly produced close to the ground (tropospheric ozone).56 It results from complex chemical reactions between pollutants directly emitted from vehicles, factories and other industrial sources, fossil fuel combustion, consumer products, evaporation of paints, and many other sources. These highly nonlinear Leontief-like reactions involve volatile organic compounds (VOCs) and oxides of nitrogen (NOx) in the presence of sunlight. In “VOC-limited” locations, the VOC/NOx ratio in the ambient air is low (NOx is plentiful relative to VOC), and NOx tends to inhibit ozone accumulation. In “NOx-limited” locations, the VOC/NOx ratio is high (VOC is plentiful relative to NOx), and NOx tends to generate ozone. As a photochemical pollutant, ozone is formed only during daylight hours, but is destroyed throughout the day and night. It is formed in greater quantities on hot, sunny, calm days. Indeed, major episodes of high ozone concentrations are associated with slow moving, high pressure systems, which are associated with the sinking of air, and result in warm, generally cloudless skies, with light winds. Light winds minimize the dispersal of pollutants emitted in urban areas, allowing their concentrations to build up. Photochemical activity involving these precursors is enhanced because of higher temperatures and the availability of sunlight. Modeling studies point to temperature as the most important weather variable affecting ozone concentrations.57 Ambient ozone concentrations increase during the day when formation rates exceed destruction rates, and decline at night when formation processes are inactive.58 Ozone concen56It is not the stratospheric ozone of the ozone layer, which is high up in the atmosphere, and reduces the amount of ultraviolet light entering the earth’s atmosphere. 57Dawson, Adams and Pandisa (2007), for instance, examine how concentrations of ozone respond to changes in climate over the eastern U.S. The sensitivities of average ozone concentrations to temperature, wind speed, absolute humidity, mixing height, cloud liquid water content and optical depth, cloudy area, precipitation rate, and precipitating area extent were investigated individually. The meteorological factor that had the largest impact on ozone metrics was temperature. Absolute humidity had a smaller but appreciable effect. Responses to changes in wind speed, mixing height, cloud liquid water content, and optical depth were rather small. 58In urban areas, peak ozone concentrations typically occur in the early afternoon, shortly after solar noon when the sun’s rays are most intense, but persist into the later afternoon. 150 trations also vary seasonally. They tend to be highest during the late spring, summer and early fall months.59 The EPA has established “ozone seasons” for the required monitoring of ambient ozone concentrations for different locations within the U.S.60 Recently, there is growing concern that the ozone season may prolong with climate change (e.g., Zhang and Wang, 2016). A.1.b. Further Details on the Construction of the Data Weather Data — Meteorological data was obtained from the National Oceanic and Atmospheric Administration’s Global Historical Climatology Network database (NOAA, 2014). This data set provides detailed weather measurements at over 20,000 weather stations across the country, for which we use the period April-September, 1950-2013, for the contiguous 48 states. In constructing our complete, unbalanced panel of weather stations we make only one restriction: for each weather station in each year, we include only those stations for which valid measurements of maximum and minimum temperature, as well as precipitation, exist for at least 75 percent of the days in the ozone monitoring season (April-September). Figure A.1.2 illustrates the geographical location of the weather stations that we have used from 1950-2013, while Table A.1.1 reports summary statistics for maximum temperature and our decomposed measures of climate norm and temperature shock, averaged across our entire sample for each year 1980-2013. Figure A.1.3 illustrates the variation we have in both components of the maximum temperature, namely, the temperature shocks and the climate norms, using a semi-balanced panel of the comprehensive set of weather stations61 while Figure A.1.4 depicts similar variation, but using only the temperature assigned to each ozone monitor in our final sample. Notice that there seems to be more variation in the 30-year MA in the latter figure because it includes cross-sectional variation as well. Also, the 30-year 59In areas where the coastal marine layer (cool, moist air) is prevalent during summer, the peak ozone season tends to be in the early fall. 60Appendix Table A.1.3 shows the ozone season for each state during which continuous, hourly averaged ozone concentrations must be monitored. 61To create this semi-balanced panel, we impose an additional restriction on our complete, unbalanced sample: for each weather station, we include only those stations with valid readings in every year 1950-2013. 151 MA trends down towards the end of the period of our study due to changes in ozone monitor location over time, as shown in Figure A.1.6. These weather stations are typically not located adjacent to the ozone monitors. Hence, we develop an algorithm to obtain a weather observation at each ozone monitor in our sample. Using information on the geographical location of ozone monitors and weather stations, we calculate the distance between each pair of ozone monitor and weather station using the Haversine formula. Then, for every ozone monitor we exclude weather stations that lie beyond a 30 km radius of that monitor. Moreover, for every ozone monitor we use weather information from only the closest two weather stations within the 30 km radius. Once we apply this algorithm, we exclude ozone monitors that do not have any weather stations within 30km. We calculate weather at each ozone monitor location as the weighted average of these two weather stations using the inverse of the squared distance between them. Figure A.1.7 illustrates the proximity of our final sample of ozone monitors to these matched weather stations. We additionally assess the robustness of our results to changes in this algorithm by increasing the radius to 80 km and using the 5 closest weather stations, and by varying the weights used – unweighted arithmetic mean and simple inverse distance weighting – in calculating the approximate daily weather at each ozone monitoring location. The results of our model under these alternative specifications is discussed further in Appendix A.2.a. Ozone Data — Ambient ozone concentration data was obtained from the Environmental Protection Agency’s Air Quality System (AQS) AirData database, which provides daily readings from the nationwide network of the EPA’s air quality monitoring stations. The data was made available by a Freedom of Information Act (FOIA) request. In our preferred specification we use an unbalanced panel of ozone monitors. We make only two restrictions to construct our final sample. First, we include only monitors with valid daily information. According to EPA, daily measurements are valid for regulation purposes only if (i) 8-hour averages are available for at least 75 percent of the possible hours of the day, or (ii) daily maximum 8-hour average concentration is higher than the standard. Second, as a minimum 152 data completeness requirement, for each ozone monitor we include only years for which at least 75 percent of the days in the ozone monitoring season (April-September) are valid; years having concentrations above the standard are included even if they have incomplete data. We have valid ozone measurements for a total of 5,284,615 monitor-days covering 1980- 2013 and the conterminus United States.62 The number of total valid monitors increased from 1,361 in the 1980s to 1,851 in the 2000s, indicating a growth of 16.6 percent of the ozone monitoring network per decade.63 The number of monitored counties in our main estimating sample also grew from 585 in the 1980s to 840 in the 2000s. Figure A.1.6 depicts the evolution of our sample monitors over the three decades in our data, and illustrates the expansion of the network over time. Table A.1.2 provides some summary statistics regarding the increase in the number of monitors over time.64 Auxiliary Data — In some of our robustness checks and examinations of heterogeneity we incorporate additional datasets. Sources and any necessary data construction steps are described below. In Table A.2.12 we use measures of whether a county is “VOC-limited” or “NOx-limited.” These measures were constructed using data collected by the EPA’s network of respective monitoring stations. Note, however, that these are often separate pollution monitors from our main sample of ozone monitors. Additionally, data – especially for VOCs – is relatively sparse compared to ozone data. Due to these data constraints, we construct measures of whether a county is VOC-limited or NOx-limited for each 5-year period in our sample, e.g. 1980-1984, which we then match with our sample of ozone monitors at the county level. To 62Note that this value refers to all valid ozone measurements, the final samples used in estimation will be smaller due to, e.g., instances where an ozone monitor is not paired with any weather stations under our matching algorithm. For instance, our main estimating sample contains 5,139,523 valid monitor-day observations. 63For our main estimating sample, these are 1,285 and 1,701, respectively. 64Note that not all monitored counties were monitored in every year, and not all monitoring stations were active in every year. Some monitors were phased in to replace others, while others were simply added to the network over time as needed – thus individual years will generally have less unique monitors and monitored counties than existed across an entire decade or the sample period. 153 construct these measures we first combine the EPA’s VOC and NOx data at the county-day level and generate a daily ratio of VOCs to NOx for each county, where possible. Following the scientific literature, observations with a ratio less than or equal to 4 are coded as VOClimited, while those greater than 15 are coded NOx-limited, and the remainder are coded as non-limited. We then sum these three measures by county across each 5-year interval and denote a county as VOC-limited, NOx-limited, or non-limited for that interval based on whichever measure was the most prevalent. For example, a county with 50 VOC-limited day, 20 NOx-limited days, and 30 non-limited days would be marked as VOC-limited for this 5-year window. Admittedly, this creates a somewhat coarse measure of whether a county is VOC- or NOx-limited. Given the available data, however, this appears to be the furthest this question can be pursued at this time, and, if anything, should be expected to bias the observed effect from this heterogeneity towards zero. In Table A.2.4 we include average daily windspeed and total daily sunlight as additional regressors within our main specification. These data, although recorded less frequently, are collected at the same weather monitoring stations as our main temperature and precipitation variables. Due to the sparseness of these data we do not decompose them into a long-run climate norm and transitory weather shock as we do with temperature and precipitation. In Tables A.2.9 and A.2.10 we examine heterogeneity in our results when separating counties into low- median- and high-levels of belief regarding the existence of climate change and the use of regulation to reduce carbon emissions. These measures were constructed using county level survey data collected by Howe et al. (2015) in 2013 which estimate the percentage of each county’s respective population that hold such beliefs. Notably, we do not rely on the explicitly stated aggregate level of belief, but rather the relative level of belief compared to the rest of our sample. Specifically, we separate counties into low- median- or high-belief terciles based on their stated level of belief in the existence of climate change – and separately by their belief in the use of regulations to reduce carbon emissions. In this way we arrive at three approximately equal sized groups for which we are able to examine 154 heterogeneity in climate impacts and adaptive response. For reference, Table A.1.4 provides summary statistics of basic demographic characteristics across these three county groupings using data from the 2006-2010 5-year American Community Survey. In Table A.2.11 we approach the question of hetergeneous beliefs from a different angle, using county-level voting results from the 2008 general presidential election obtained from MIT’s Election Data and Science Lab (2018). We construct a simple indicator variable for whether Barack Obama or John McCain won the popular vote in that county and denote a county as “Democrat” if the former is true. 155 Figure A.1.1: Temperature Relative to Baseline (1950-1979) -1 -.5 0 .5 1 1.5 Deviation from 1950-1979 Baseline Temperature (C°) 1950 1960 1970 1980 1990 2000 2010 Avg. Temp Max. Temp Min. Temp Notes: This figure depicts annual temperature fluctuations and the overall climatic trend for the ozone season in the US relative to a 1950-1979 baseline average. The baseline and the yearly deviations from it are constructed from the comprehensive sample of weather stations across the US from 1950 to 2013 following the data construction steps detailed in Appendix A.1.b. The 1950-1979 baseline represents, generally speaking, the pre-climate change awareness era. The average temperature, relative to this baseline, has been slowly but steadily increasing since the early- to mid-1970’s, with an increase in the average temperature of approximately 0.5 degree Celsius (◦C) by 2010. For clarity, the thin solid line, the short-dashed line, and long-dashed line refer to annual averages for daily average, maximum, and minimum temperature, respectively, as coded in the legend. The thick solid line smooths out the annual observations for average temperature over the period covered in the graph. 156 Figure A.1.2: Comprehensive Location of all Weather Monitors Notes: This figure maps the location of all weather stations across the continental U.S. contained in our complete dataset. 157 Figure A.1.3: Climate Norms and Shocks (semi-balanced sample) 26.05 26.10 26.15 26.20 26.25 30-year Moving Average (Temperature C°) 1980 1990 2000 2010 Panel A. Average Climate Norm Over Time -1 -0.5 0.0 0.5 1.0 1.5 Deviation from Moving Average (Temperature C°) 1980 1990 2000 2010 Panel B. Average Temperature Shock Over Time Notes: This figure depicts US temperature over the years in our sample (1980-2013), decomposed into their climate norm and temperature shock components. The climate norm (Panel A) and temperature shocks (Panel B) are constructed from a semi-balanced panel of weather stations across the US from 1950 to 2013, restricting the months over which measurements were gathered to specifically match the ozone season of April–September, the typical ozone season in the US (see Appendix Table A.1.3 for a complete list of ozone seasons by state). Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. The solid line in Panel A smooths out the annual averages of the 30-year moving averages, and the horizontal dashed lines in Panel B highlights that temperature shocks are bounded in our period of analysis. As described in our data construction in Appendix A.1.b, our full sample of weather stations includes only weather stations with valid measurements for at least 75% of the days in the ozone season. Here we further restrict this to a semi-balanced sample, including only those stations with valid readings in every year of our sample. 158 Figure A.1.4: Climate Norms and Shocks (main model sample) 26.4 26.6 26.8 27.0 27.2 27.4 30-year Moving Average Temperature (C°) 1980 1990 2000 2010 Panel A. Average Climate Norm Over Time -0.5 0.0 0.5 1.0 Deviation from Moving Average Temperature (C°) 1980 1990 2000 2010 Panel B. Average Temperature Shock Over Time Notes: This figure depicts US temperature over the years in our sample (1980-2013), decomposed into their climate norm and temperature shock components. The climate norm (Panel A) and temperature shocks (Panel B) are constructed from the panel of weather stations included in our main model sample across the US from 1950 to 2013, restricting the months over which measurements were gathered to specifically match the ozone season of April–September, the typical ozone season in the US (see Appendix Table A.1.3 for a complete list of ozone seasons by state). The unbalanced feature of our main sample, with ambient ozone monitors moving north over time (see Figure A.1.6), is the likely driving force behind the downward pattern of the average climate norm at the end of our sample period in Panel A. The horizontal dashed lines in Panel B highlights that temperature shocks are bounded in our period of analysis. 159 Figure A.1.5: Evolution of Maximum Ambient Ozone Concentration 1979 NAAQS 80 100 120 140 160 Daily Max Ozone Levels (ppb) 1980 1990 2000 2010 Unbalanced Sample Semi-Balanced Sample Notes: This figure depicts the evolution of the daily maximum 1-hour ambient ozone concentrations over time in the US for both our complete (unbalanced) sample and our restricted (semi-balanced) sample. For reference the horizontal line depicts the 1979 National Ambient Air Quality Standard for Ozone, which was based on an observed 1-hour maximum ambient ozone concentration of 120 ppb or higher. 160 Figure A.1.6: Ozone Monitor Location by Decade of First Appearance Notes: This figure maps the location of each ozone monitor in our final sample, by decade of first appearance. 161 Figure A.1.7: Ozone Monitors and their Matched Weather Monitors Notes: This figure maps the location of each ozone monitor in our final sample, and their matched weather stations. For each ozone monitor, the closest 2 stations within a 30 km radius have been used in the matching. 162 Figure A.1.8: Relationship between Ozone and Decomposed Temperature -4 -2 0 2 4 Detrended Ambient Ozone (ppb) -0.6 -0.3 0.0 0.3 0.6 Detrended 30-year Moving Average (Temperature C°) 1980 1990 2000 2010 Climate Norm Max Ozone Level Panel A. Relationship Between Ozone and Climate Norm -4 -2 0 2 4 Detrended Ambient Ozone (ppb) -1.0 -0.5 0.0 0.5 1.0 Detrended Deviation from Moving Average (Temperature C°) 1980 1990 2000 2010 Temperature Shock Max Ozone Level Panel B. Relationship Between Ozone and Temperature Shock Notes: This figure depicts the general relationship between daily maximum ozone concentrations and temperature over the years in our sample (1980-2013) after decomposing temperature into our measure of climate norm and temperature shock and de-trending the data. Both the climate norm (Panel A) and the temperature shock (Panel B) appear to have a close correlation with ozone concentrations, although the relationship in Panel A appears weaker than that in Panel B, providing suggestive evidence of adaptative behavior. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. 163 Figure A.1.9: Decomposition of Temp. Norms & Shocks (Los Angeles, All Years) Panel A. Our Preferred Decomposition -20 0 20 40 Temperature (C°) 1980 1990 2000 2010 Temp Shock Climate Norm Panel B. Fixed-Effect Decomposition -20 0 20 40 Temperature (C°) 1980 1990 2000 2010 Deviation from Avg Average Temp Notes: his figure compares our preferred decomposition method with a standard fixed-effects approach using data from the Los Angeles ozone season across all years in our sample. Panel A depicts the daily measure of temperature, decomposed into climate norm and temperature shock. By contrast, Panel B depicts the same daily measure of temperature, but decomposed into a typical fixed-effect average temperature and the deviations, after controlling for monthly fixed effects. The dashed lines indicate observed daily maximum temperature while the black solid lines represent long-run norms. The gray solid lines indicate temperature shocks which are nearly identical in both panels, as would be expected from the Frisch-Waugh-Lovell theorem, illustrating the variation used for identifying βW and βF E. Panel A additionally highlights the variation in climate used to identify βC in our proposed approach, while Panel B lacks any such variation in the measure of climate. 164 Table A.1.1: Yearly Summary Statistics for Daily Maximum Temperature Year Max Temp Climate Trend Temp Shock (1) (2) (3) (4) 1980 27.0 26.5 0.5 1981 26.9 26.6 0.4 1982 26.1 26.7 -0.6 1983 26.8 26.8 0.0 1984 26.7 26.8 -0.1 1985 27.0 26.6 0.3 1986 26.7 26.4 0.3 1987 27.3 26.6 0.7 1988 27.4 26.6 0.7 1989 26.4 26.7 -0.3 1990 26.7 26.6 0.1 1991 27.1 26.6 0.5 1992 26.1 26.7 -0.5 1993 26.6 26.6 0.0 1994 26.9 26.6 0.2 1995 26.8 26.7 0.0 1996 26.5 26.7 -0.2 1997 26.4 26.8 -0.4 1998 27.3 27.0 0.4 1999 27.2 27.0 0.2 2000 27.1 27.1 0.0 2001 27.4 27.2 0.3 2002 27.8 27.2 0.6 2003 26.9 27.3 -0.4 2004 27.0 27.2 -0.2 2005 27.6 27.3 0.3 2006 27.7 27.3 0.4 2007 27.7 27.3 0.4 2008 27.3 27.3 0.0 2009 26.9 27.3 -0.3 2010 27.8 27.2 0.6 2011 27.4 27.1 0.3 2012 28.0 27.1 0.9 2013 26.4 26.6 -0.3 Notes: This table outlines the evolution of maximum temperature in our sample from the years 1980-2013 in Column (2). Columns (3) and (4) decompose this into our respective measures of climate norm and temperature shock. 165 Table A.1.2: Yearly Summary Statistics for Ozone Monitoring Network Year # Observations # Counties # Ozone Monitors (1) (2) (3) (4) 1980 88426 361 609 1981 100459 399 659 1982 102111 402 661 1983 102429 408 653 1984 103828 390 649 1985 105457 388 648 1986 103820 375 634 1987 110366 392 668 1988 113232 409 686 1989 119938 425 725 1990 126535 443 757 1991 132046 466 792 1992 137754 482 821 1993 146023 511 863 1994 149400 520 876 1995 154230 528 902 1996 153019 530 894 1997 160024 550 931 1998 164491 568 960 1999 168901 585 982 2000 172686 592 999 2001 180872 616 1047 2002 186261 630 1071 2003 188462 641 1082 2004 189868 653 1087 2005 187709 649 1082 2006 188298 650 1075 2007 190824 661 1092 2008 190682 660 1099 2009 194184 678 1116 2010 196439 688 1130 2011 199948 716 1159 2012 199723 703 1148 2013 148306 658 1039 Notes: This table outlines the summary statistics of our main data sample. The construction of our main sample follows EPA guidelines by including all monitor-days for which 8-hour averages were recorded for at least 18 hours of the day and monitor-years for which valid monitor-days were recorded for at least 75% of days between April 1st and September 30th. 166 Table A.1.3: Ozone Monitoring Season by State State Start Month - End State Start Month - End Alabama March - October Nevada January - December Alaska April - October New Hampshire April - September Arizona January - December New Jersey April - October Arkansas March - November New Mexico January - December California January - December New York April - October Colorado March - September North Carolina April - October Connecticut April - September North Dakota May - September Delaware April - October Ohio April - October D.C. April - October Oklahoma March - November Florida March - October Oregon May - September Georgia March - October Pennsylvania April - October Hawaii January - December Puerto Rico January - December Idaho April - October Rhode Island April - September Illinois April - October South Carolina April - October Indiana April - September South Dakota June - September Iowa April - October Tennessee March - October Kansas April - October Texas1 January - December Kentucky March - October Texas1 March - October Louisiana January - December Utah May - September Maine April - September Vermont April - September Maryland April - October Virginia April - October Massachusetts April - September Washington May - September Michigan April - September West Virginia April - October Minnesota April - October Wisconsin April 15 - October 15 Mississippi March - October Wyoming April - October Missouri April - October American Samoa January - December Montana June - September Guam January - December Nebraska April - October Virgin Islands January - December Notes: This table shows, for each state, the season when ambient ozone concentration is required to be measured and reported to the U.S. EPA. 1The ozone season is defined differently in different parts of Texas. Source: USEPA (2006, p.AX3-3). 167 Table A.1.4: County Summary Statistics by Belief in Climate Change Panel A. Low Belief Counties Count Mean Std. Dev. Minimum Maximum Population (1000’s) 334 80.6 106.9 0.8 837.5 Average Education (Years) 334 12.7 0.6 11.0 14.3 Median Income ($1000/year) 334 48.6 10.4 21.9 83.3 Average Income ($1000/year) 334 61.7 11.3 36.9 111.9 Panel B. Median Belief Counties Population (1000’s) 335 174.7 297.4 1.9 3,951.0 Average Education (Years) 335 13.2 0.6 11.8 15.1 Median Income ($1000/year) 335 53.8 12.4 26.3 109.8 Average Income ($1000/year) 335 68.2 14.6 39.2 142.2 Panel C. High Belief Counties Population (1000’s) 336 466.7 780.8 1.3 9,758.3 Average Education (Years) 336 13.6 0.7 11.5 16.1 Median Income ($1000/year) 336 60.5 16.8 30.4 125.7 Average Income ($1000/year) 336 79.6 21.3 41.1 146.0 Notes: This table reports summary statistics of underlying demographics for each of the terciles of counties used in Table B9. Demographic data were obtained from the 2006-2010 5-year American Community Survey, with income reported in 2015 dollars, and average years of education based on a population weighted average of educational attainment status for the county population over 25 years of age. 168 A.2. Further Robustness Checks and Heterogeneity Section A.2.a of this appendix provides further elaboration of the alternative specifications used for robustness checks as well the results of our nonlinear specifications, as discussed in Section IV, while Section A.2.b does so for our heterogeneity analyses. It then includes relevant Figures and Tables as outlined below. Figure A.2.1.Climate Impacts and Adaptation Over Time Table A.2.1.Alternative Criteria for Selection of Weather Stations Table A.2.2.Comparison to Alternative Estimation Methods (Semi-Balanced Panel) Table A.2.3.Excluding Areas with Regional Air Pollution Policies Table A.2.4.Further Robustness Checks Table A.2.5.Bootstrapped Standard Errors Table A.2.6.Non-Linear Effects of Temperature Table A.2.7.Comparison of Adaptation Under Nonlinear Specifications Table A.2.8.Results by Decade Table A.2.9.Adaptation by Belief in Climate Change Table A.2.10.Adaptation by Belief in Climate Change Regulation Table A.2.11.Adaptation by Political Leaning Table A.2.12.Adaptation by VOC- or NOx-limited Atmosphere 169 A.2.a. Further Robustness Checks Alternative Criteria for Selection of Weather Stations — While our robustness checks presented in Table 1.2 have addressed potential concerns with the manner in which we construct our regressors by decomposing temperature, a possible additional concern arises from the fact that temperature monitors are not necessarily sited next to ozone monitors. Because of this, we do not have an exact measure of temperature at the same geographic point as our measure of ozone. As discussed in our data section and detailed in Appendix A.1.b, we define temperature at an ozone monitoring station as the mean of the reported daily maximum temperatures at the two closest weather stations within 30 kilometers, weighted by the inverse squared distance to the ozone monitor. In so doing, we are likely to approximate a good measure of the daily maximum temperature for the local region as a whole, while also maintaining a close geographic boundary around the ozone monitoring station so as not to influence this approximation with temperature readings from a weather station further away that may be subject to a different set of meteorological conditions. It’s possible, however, that a less strongly distance-weighted mean would provide a more accurate measure of temperature for the overall local region – although likely less accurate at the ozone monitoring station itself – or that the 2-station and 30-kilometer cutoffs are too restrictive. We investigate the effects of lessening the distance weighting in the calculation of expected temperature at the ozone monitoring station, as well as relaxing the constraints on both the number of included weather stations and distance from the ozone monitor in Table A.2.1. Specifically, columns (1) and (2) report results of our main specification when we maintain the 2-station/30-kilometer restriction, but decrease the weighting scheme to either the simple arithmetic mean in column (1), or a non-squared inverse distance weight in column (2). Columns (3) and (4) use the same weighting schemes as in (1) and (2), but now include temperature readings from the 5 closest weather monitoring stations within 80 kilometers. Results in all four columns are relatively stable and consistent with our main specification. 170 Non-Random Siting of Ozone Monitors — In recent work, Muller and Ruud (2018a) argue that the location of pollution monitors is not necessarily random. The U.S. EPA maintains a dense network of pollution monitors in the country for two major reasons: (i) to provide useful data for the analysis of important questions linking pollution to its varied impacts, and (ii) to check and enforce regulations on criteria pollutants. These are conflicting interests: while monitors should be placed in regions having different levels of pollution to provide representative data, they might be placed in areas where pollution levels are the highest to maintain oversight. Not surprisingly, the authors find that most of the monitors tend to be in areas where pollution levels have been high, and compliance with the regulation is a question. Following those authors’ results, we can expect that ozone monitors that have consistently been in our sample across all years may be located in areas having very high pollution levels, thus commanding constant monitoring and regulation by the EPA. To check if this claim is accurate, we re-run our main analysis using a balanced sample of ozone monitors. Starting from our original sample, and using only monitors that have been in the data for every year from 1980-2013, we are left with 89 pollution monitors. The results are reported in Table A.2.2. We find that a 1◦C temperature shock leads to a rise in ozone concentrations by 2.03 ppb, while a 1◦C increase in the climate norm leads to a rise of 1.49 ppb, implying an adaptation level of 0.54 ppb. As expected, the temperature effects obtained from the balanced panel are larger than those in our main results, although the level of adaptation remains largely unchanged. The balanced panel leads to an overestimation of the climate penalty. Therefore, our preferred, unbalanced sample of monitors includes areas with different levels of air pollution, and our estimates should be more representative of the entire country. Accounting for Policies Targeting Ozone Precursors — During our period of analysis (1980- 2013) there were two other major policies aimed at reducing ambient ozone concentrations implemented in the United States: (i) regulations restricting the chemical composition of 171 gasoline, intended to reduce VOC emissions from mobile sources (Auffhammer and Kellogg, 2011), and (ii) the NOx Budget Trading Program (Deschenes, Greenstone and Shapiro, 2017). There may be concern that these input regulations targeted at ozone precursors could be influencing our results. Table A.2.3 examines the sensitivity of our results to the exclusion of the regions and periods affected by these regulations from our estimating sample. Column (1) reports the results of our main specification re-estimated on a sample excluding all observations from California starting in 1996, when new state-wide regulations went into effect – aimed at reducing VOC emissions between April and September by requiring a more stringently regulated type of reformulated gasoline (RFG) be sold.65 Column (2) reports the results of re-estimating our main specification after excluding all states that participated in the NOx Budget Trading Program (NBP) starting in 2003, when the program went into effect. Finally, column (3) re-estimates our model on a sample excluding both subsets of observations. In all three cases the recovered estimates of temperature shock, climate norm, and implied adaptation are statistically indistinguishable from our full-sample estimates. This is not too surprising, because predominantly it is ozone formation, rather than precursors, that depend on climate. Thus, while these policies may have affected precursor levels, they would not necessarily have affected how agents respond to changes in climate. Additional Robustness Checks — In addition to all prior robustness checks, we conduct four final checks in Table A.2.4. First, it may be a concern that our climate norm variable structures the long-run climate normal temperature as the 30-year monthly moving average, despite the fact that seasonal – or within-season – shifts in temperature are unlikely to exactly follow the calendar at a monthly level. We examine the sensitivity of our results to this decision by alternatively constructing this variable as a 30-year daily moving average, allowing it to vary arbitrarily within each month. Results of our main specification, 65We exclude only California in this exercise because Auffhammer and Kellogg (2011) only found effects of gasoline standards on air quality in California. They found no effects for the federal gasoline standards. 172 substituting daily moving averages for the standard monthly ones, are presented in column (1). Both coefficients of interest are nearly identical to our original findings. Ultimately, we prefer the monthly moving average because it is likely that individuals recall climate patterns by month, not by day of the year, making the interpretation of adaptation more intuitive. Indeed, broadcast meteorologists often talk about how a month has been the coldest or warmest in the past 10, 20, or 30 years, but not how a particular day of the year has deviated from a daily norm. Second, it may be a concern that our proposed methodology is heavily reliant on highfrequency data in order to successfully decompose temperature into its climatological and meteorological components. While this concern does not pose a threat to identification in our context per se, if valid it would reduce the generalizability of our method to other contexts with less temporally rich data. We examine this concept by aggregating our data to the monitor-month level, taking the arithmetic mean of all variables for each ozone monitor, by month, for each year of our sample and running our preferred specification on this aggregate sample. As the climate norm variable is already identified from variation in monthly moving averages, we would not expect this coefficient to change other than due to the aggregation of our dependent variable and the temperature deviations, which both would otherwise vary daily. It is less clear, however, how this “smoothing” of daily ozone and temperature deviations might affect the coefficient on temperature shocks. Although our sample size is greatly reduced, now consisting of 178,175 observations compared to the previous 5,139,523 we find qualitatively similar results, reported in column (2). As expected, the coefficient on climate norm is nearly identical, while the coefficient on temperature shock is slightly larger in magnitude than in our full sample model, though statistically very similar. Third, although temperature is the primary meteorological factor affecting tropospheric (ambient) ozone concentrations, other factors such as wind speed and sunlight have also been noted as potential contributors. High wind speed may prevent the build-up of ozone precursors locally, and dilute ozone concentrations. Ultraviolet solar radiation should trigger 173 chemical reactions leading to the formation of more ground-level ozone. To test whether our main estimates are capturing part of the effects of wind speed and sunlight, we control for these variables in an alternative specification using a smaller sample containing those variables. Column (3) presents our main results from estimating Equation (1.13) plus controls for average daily wind speed (meters/second) and total daily sunlight (minutes). As expected, higher wind speeds lead to lower ozone concentrations, and more sunlight leads to higher concentrations. From column (3), we find that a 1 meter/second increase in average daily wind speed would decrease ozone concentrations by 2.3 ppb, whereas a 1 minute increase in daily sunlight leads to 0.02 ppb increase in ozone concentrations. Controlling for these two effects, we find that a shock in daily maximum temperature of 1◦C leads to a 1.75 ppb increase in daily maximum ozone whereas a 1◦C increase in the climate norm leads to a 1.13 ppb increase in ozone, implying a measure of overall adaptation of 0.62 ppb – all statistically similar to our main results. Our primary estimates of the impact of temperature on ozone concentrations, and hence our measures of adaptation, do not seem to rely crucially on other potentially important meteorological factors. Finally, one may be concerned that inter-regional pollution transport may be affecting our results. If, for example, pollution was transported into a region, this may affect the estimated level of adaptation in that region as the local agent would not have full control over the pollution outcome. We note two key points relating to this concern. First, neither ozone itself nor VOCs, an ozone precursor, are likely to be transported long distances due to their high reactivity, thus it is primarily NOx transport that would pose a concern. Second, as with any real-world setting, local agents may need to take exogenous factors into account when making decisions, such as in a region subject to increased baseline levels of NOx due to transport from other regions. Agents in such a region may, for example, opt to prioritize reductions in VOCs, a precursor they have more direct control over, in an effort to curb ozone formation. Furthermore, a collection of regions which suffer from such transport concerns may collectively work together to reduce NOx – this may be of their 174 own volition, or imposed by a higher regulatory body, such as the EPA. In fact, the EPA designates 12 states, in whole or part, in the Northeast as part of the “ozone transport region” (OTR).66 As these states represent the region wherein inter-regional NOx transport poses the greatest concern, we examine the effect of such transport on our central estimates by both explicitly excluding all OTR states from our estimating sample, and, conversely, using only these states as the estimating sample. Columns (4) and (5) present these results, finding that while the coefficients on weather and climate are somewhat higher in the OTRonly sample, as might be expected of a region where ozone is a particular concern, adaptation is statistically indistinguishable from our central estimate. Nonlinear Effects of Temperature on Ambient Ozone Concentration — Table A.2.6, column (1), presents the results of our preferred specification when interacting each of the independent variables with the 5◦C temperature bin indicators. The implied measure of adaptation is then presented in column (2).67 Similar to Figure 1.4, we find that the ozone/temperature response is increasing at an increasing rate at lower temperature ranges, but increases at a decreasing rate at higher temperatures, particularly for increases in the climate norm. Specifically, below 20◦C, a 1◦C temperature shock would raise ozone levels by an additional 0.69 ppb, while a similar increase in the climate norm would raise ozone concentrations by 0.14 ppb. Above 20◦C, however, both effects drastically increase, with temperature shocks increasing ozone by 1.69 ppb between 20-25◦C, and by over 2 ppb above 25◦C. While the effect of a 1◦C increase in the climate norm is increasing with temperature up to 30◦C – at 1.28 ppb and 1.83 ppb for 20-25 and 25-30◦C, respectively – the magnitude of the impact is decreasing with temperature above 30◦C – at 1.50 ppb and 0.90 ppb for 30-35 and above 66“Ozone Transport Region boundary. As of March 14, 2022, the boundary for the Ozone Transport Region consists of the entire States of Connecticut, Delaware, Maryland, Massachusetts, New Hampshire, New Jersey, New York, Pennsylvania, Rhode Island, and Vermont; portions of Maine identified in this section under Table 1; and the Consolidated Metropolitan Statistical Area that includes the District of Columbia and the following counties and cities in Virginia: Arlington County, Fairfax County, Loudoun County, Prince William County, Strafford County, Alexandria City, Fairfax City, Falls Church City, Manassas City, and Manassas Park City.” (USCFR, 2022). 67Table A.2.7 additionally compares the implied level of adaptation under the linear, binned, quadratic, and cubic specifications. 175 35◦C, respectively. This would imply a more than doubling of our full-sample measure of adaptation above 35◦C, at 1.15 ppb. These results suggest that agents may be making extra effort to reduce ozone precursor emissions when temperatures are the highest and could otherwise lead to greater ozone formation. This relatively high level of adaptation above 35◦C can be plausibly explained by at least two reasons. First, regions having temperatures above 35◦C might have higher incidence of sunlight which might lead to more extensive use of solar panels to generate electricity. Since the U.S. as a whole is predominantly NOx limited, we would expect that changes in electricity usage drastically affect ozone concentrations.68 Higher temperatures might be creating an environment that is more suited to shifts away from conventional and dirtier sources of power generation, thus leading to higher levels of adaptation. Second, absent any adaptation, days that are exceptionally hot are more likely to cause exceptionally high levels of ozone, which could trigger additional regulatory oversight. In order to avoid this, firms would be most likely to concentrate adaptation efforts on days where the climate normal temperature is itself the hottest. A.2.b. Heterogeneity Results by Decade — To examine temporal heterogeneity, Figure A.2.1 Panel A illustrates the evolution of temperature’s impacts on ozone formation across our sample period in 5-year increments, while Panel B reports the resulting level of adaptation. As seen in Panel A, the effects of both temperature shocks and the climate norms on ambient ozone concentration are decreasing over time, likely due – at least in part – to regulations (see, for example, our companion paper Bento, Mookerjee and Severnini, 2020). The early 1980’s, which marked the initial phases of ozone monitoring and awareness, and when the average pollution levels 68Electricity generation is a major source of NOx, and, since ozone formation has a Leontief-like production function in terms of NOx and VOCs, changes in electricity use in a NOx limited region would imply large changes in ozone formation. 176 were also higher, exhibit the largest impacts of climate on ambient ozone. Notice in Panel A that responses to temperature shocks a decade ahead approximately mirror responses to longer-term climatic changes a decade before. Nevertheless, the difference between those responses at any single point in time since the 1980’s has been relatively stable, as illustrated by Panel B. This suggests that there may be limits to adaptation unless new technologies are able to affect atmospheric composition, such as in the case of geoengineering (e.g., Heutel, Moreno-Cruz and Ricke, 2016; Flegal et al., 2019). It also highlights the risks of extrapolating flexibly-estimated weather responses over time to estimate adaptation (Olmstead and Rhode, 2011; Bleakley and Hong, 2017), analogous to the Lucas Critique (Lucas, 1976). Table A.2.8 mirrors Figure A.2.1 and reports our results by decade. We split our sample into three “decades” – 1980-90, 1991-2001, and 2002-2013 – so that we have roughly the same number of years in each. We find that the effects of contemporaneous daily maximum temperature, and its two components of our decomposition, are decreasing over time, as shown in column (1). Nevertheless, looking at column (2), we find evidence that adaptation by economic agents reduced slightly from the 1980’s to the 1990’s, but stabilized back at its original levels in the 2000’s. The average adaptation across all counties in our sample drops from 0.54 ppb in the 1980’s to 0.43 ppb in the 1990’s, but increases again to 0.54 ppb in the 2000’s. Adaptation by Beliefs in Climate Change Across Counties — Using the results of a relatively recent county-level survey regarding residents’ beliefs in climate change (Howe et al., 2015), we split the set of counties in our sample into terciles of high, median, and low beliefs. Table A.2.9 presents the results of our preferred specification when interacting indicator variables for high- and low-belief counties with our temperature and control variables in column (1). The implied measure of adaptation is reported in column (2). We find that lowbelief counties, on average, observe a smaller ozone response to a 1◦C temperature shock, relative to the median set of counties, but that this difference is statistically insignificant 177 with regards to changes in the climate norm. High-belief counties, by comparison, observe approximately 31-35 percent larger and statistically significant ozone responses to a 1◦C increase in both components of temperature. As might be expected of counties at opposite ends of the spectrum regarding beliefs that climate is changing, we find that adaptation is roughly 42 percent lower in low-belief counties than median ones, while this effect is similar in magnitude but of opposite sign for high-belief counties.69 This evidence suggests that greater caution is called for when extrapolating flexibly-estimated weather responses over space when dealing with adaptation to climate change. Economic agents might respond heterogeneously according to unobserved preferences, beliefs, and the experience with the local climate. Heterogeneity by Precursor “Limited” Atmospheric Conditions — As detailed in Appendix Section A.1.a, ozone is formed from precursor pollutants – volatile organic compounds (VOCs) and oxides of nitrogen (NOx) – in the presence of sunlight and heat. Specifically, ozone formation appears to follow a Leontief-like production function, implying that regions where the ambient supply of one of the two precursor pollutants, VOCs or NOx, are limited might be less susceptible to increased ozone formation when faced with increased temperatures. To examine potential heterogeneity in the temperature/ozone relationship and adaptation along this channel we collected all available data on VOC and NOx emissions for each county in our sample as reported by the EPA. Due to the sparseness of these data, we construct aggregate indicators of whether a county is VOC-limited, NOx-limited, or neither for each 5-year interval of our overall sample.70 69Table A.2.10 in Appendix A.2.b conducts a similar analysis, separating counties by their belief in the use of regulation on carbon emissions, while Table A.2.11 in Appendix A.2.b instead splits the sample into two groups based on whether they leaned Republican or Democrat in the 2008 presidential election using data from MIT (2018). Results in Table A.2.10 are qualitatively similar to Table A.2.9, while the results in Table A.2.11 paint a similar picture under the assumption that belief or dis-belief in climate change approximately maps to Democratic or Republican political affiliation. Table A.1.4 in Appendix A.1 provides summary statistics of basic characteristics for the three sets of counties used in Table A.2.9. High-belief counties tend to be more populous, better educated, and richer than low-belief ones. 70Because ozone formation follows a Leontief-like production function, a county is “VOC-limited” if the 178 Column (1) of Table A.2.12 presents the results of our main specification when using this restricted sample – approximately 20% of our full sample – finding results that are qualitatively similar, albeit larger in magnitude, to our full sample results for the effects of temperature shock, climate norm, and the resulting measure of adaptation – shown in column (2). The magnitude is likely larger because VOCs may be monitored in places with potentially high concentrations. In column (3) we interact the indicators for VOC- and NOx-limited counties with our other regressors to recover a coarse estimate of the effect that being limited in either precursor has on the relationship between our two measures of temperature and ozone. Both main coefficients, and the resulting measure of adaptation – shown in column (4), remain statistically unchanged for non-limited counties. While the difference from these values is statistically indistinguishable from zero in NOx-limited counties. In VOC-limited counties the effects of temperature shock and climate norm are approximately 31 and 79 percent lower and significant, respectively, although the resulting level of adaptation is not precise enough to conclude that it is statistically different from other counties. This finding appears to corroborate the Leontief-like production function of ozone (e.g., Auffhammer and Kellogg, 2011; Deschenes, Greenstone and Shapiro, 2017); when departing from the balanced mix of ozone precursors, the estimated effects of temperature on ambient ozone concentration decline. ratio of VOC to NOx is too low, while it would be “NOx-limited” if the ratio is too high, and a middle set of counties would not be limited as they face levels of both precursor emissions closer to the “optimal” mix. Further details on this data can be found in Appendix A.1.b. 179 Figure A.2.1: Climate Impacts and Adaptation Over Time 3210 Increase in Daily Max Ozone (ppb) per Degree Increase in Temperature (C°) 1980 1990 2000 2010 Temperature Shock Climate Norm Panel A. Temperature Impact on Ozone Over Time 1 .8 .6 .4 .20 Reduction in Daily Max Ozone Increase (ppb) Due to Adaptation 1980 1990 2000 2010 Adaptation Panel B. Adaptation Over Time Notes: This figure displays the impacts of temperature increases on ambient ozone concentrations over time in the US in Panel A, and the implied measures of adaptation in Panel B. Splitting the main sample into 5-year periods (e.g., 1980-1984, 1985-1989, etc.), Panel A depicts the estimated coefficients on the climate norm and temperature shock variables for each of these periods. All coefficients were estimated by Equation (1.13), modified to include interactions between both components of temperature and indicators for each of the 5-year periods. Panel B depicts the respective measures of adaptation as the differences between the estimated coefficients associated with shocks and norms. The solid lines in Panel A smooth out each set of estimated coefficients plotted in the graph, and the dashed line in Panel B smooths out the implied measures of adaptation. Appendix A.2.b Table A.2.8 examines these same patterns by decade in tabular form. All point estimates included in the figure are statistically significant at the 1% level. 180 Table A.2.1: Alternative Criteria for Selection of Weather Stations Daily Max Ozone Levels (ppb) (1) (2) (3) (4) Temperature Shock 1.721*** 1.700*** 1.773*** 1.764*** (0.063) (0.063) (0.067) (0.066) Climate Norm 1.165*** 1.165*** 1.156*** 1.156*** (0.051) (0.051) (0.050) (0.050) Implied Adaptation 0.557*** 0.535*** 0.617*** 0.608*** (0.041) (0.042) (0.044) (0.043) Distance Cut-off 30 km 30 km 80 km 80 km Stations Included 2 2 5 5 Weighting Scheme Simple Avg 1/Dist Simple Avg 1/Dist All Controls Yes Yes Yes Yes Observations 5,139,523 5,139,523 5,284,420 5,284,420 R2 0.484 0.483 0.484 0.485 Notes: This table reports estimates from models using alternative criteria to match weather stations to ozone monitors. These estimates are obtained by our main specification, Equation (1.13), but using different distance radii, number of weather stations, and weights when matching ozone monitors to weather stations. In our main analysis we use a radius of 30 km, the 2 closest stations, and the inverse squared distance as the weight. In the above columns, we give the same weight to both stations (simple average), or use the inverse distance as an alternative weight. In columns (3) and (4) we also increase the radius to 80 km and use the information from the closest 5 weather stations. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 181 Table A.2.2: Comparison to Alternative Estimation Methods (Semi-Balanced Panel) Daily Max Ozone Levels (ppb) Unifying Fixed-Effects Cross-Section (1) (2) (3) Temperature Shock 2.028*** (0.109) Climate Norm 1.492*** (0.084) Max Temperature 2.009*** (0.109) Average Max Temperature 0.904 (0.950) Implied Adaptation 0.536*** 1.105 (0.082) (0.773) Fixed Effects: Monitor-by-Season-by-Year Yes Monitor-by-Month-by-Year Yes State Yes Precipitation Controls Yes Yes Yes Latitude & Longitude Yes Non-Attainment Control Yes Observations 520,670 520,670 89 R2 0.475 0.534 0.545 Notes: This table reports our main climate impact results using a semi-balanced panel including only those monitors that exist in every year of our data. Column (1) reports the estimates of our unifying approach, in which we decompose daily maximum temperature into climate norms and weather shocks, and exploit variation in both components in the same estimating equation – our Equation (1.13). Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year to allow for economic agents to potentially adapt, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. Column (2) reports the effect of daily maximum temperature on ambient ozone from the panel fixed-effects approach, exploiting day-to-day variation in temperature, hence capturing the effect of a change in weather. Column (3) reports cross-sectional estimates using average maximum temperature and ambient ozone concentrations for each ozone monitor in the sample. Having averaged the variables over all the years from 1980-2013, this estimate captures the effect of a change in climate. Note that while estimates in column (3) must additionally control for whether a county is in violation of the CAA ozone standards, this is implicitly controlled for via the fixed-effects in columns (1) and (2). Standard errors are clustered at the county level in columns (1) and (2), while column (3) uses standard heteroskedastic robust errors. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 182 Table A.2.3: Excluding Areas with Regional Air Pollution Policies Daily Max Ozone Levels (ppb) Gasoline Policy (RFG) NOx Budget Program Both (1) (2) (3) Temperature Shock 1.672*** 1.723*** 1.722*** (0.060) (0.073) (0.073) Climate Norm 1.175*** 1.218*** 1.234*** (0.045) (0.060) (0.054) Implied Adaptation 0.498*** 0.506*** 0.488*** (0.040) (0.049) (0.048) All Controls Yes Yes Yes Observations 4,631,407 4,338,178 3,830,062 R2 0.463 0.491 0.473 Notes: This table reports results from our main specification in Equation (1.13) but excluding locations with input regulations aimed at reducing ozone precursors (VOCs and NOx). Two major regulations were implemented in the United States over our sample period 1980-2013: (i) regulations restricting the chemical composition of gasoline, intended to reduce VOC emissions from mobile sources (Auffhammer and Kellogg, 2011), and (ii) the NOx Budget Trading Program (Deschenes, Greenstone and Shapiro, 2017). Here we examine the sensitivity of our estimates when taking into account these input regulations. Column (1) excludes California from 1996 onwards, when stringent VOC regulations were in place. Column (2) excludes the states participating in the NBP from 2003 onwards, when the program was in effect. Column (3) excludes both subsets of observations. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 183 Table A.2.4: Further Robustness Checks Daily Max Ozone Levels (ppb) Daily Monthly Meteorological Excluding Including Only Moving Average Aggregation Controls OTR States OTR States (1) (2) (3) (4) (5) Temperature Shock 1.684*** 1.806*** 1.749*** 1.558*** 2.077*** (0.064) (0.062) (0.078) (0.082) (0.056) Climate Norm 1.207*** 1.171*** 1.126*** 1.052*** 1.476*** (0.050) (0.050) (0.070) (0.066) (0.059) Average Wind Speed −2.325*** (0.309) Total Daily Sunlight 0.015*** (0.001) Implied Adaptation 0.477*** 0.636*** 0.624*** 0.506*** 0.601*** (0.040) (0.041) (0.064) (0.055) (0.038) Observations 5,139,454 178,175 453,829 4,116,365 1,023,158 R2 0.480 0.859 0.479 0.497 0.426 Notes: This table reports estimates, obtained by Equation (1.13), from models that replace our monthly moving average with a daily one in column (1), aggregate our high-frequency daily data to monthly averages in column (2), and include additional meteorological controls in column (3). Specifically, for column (1) we first decompose contemporaneous maximum temperature into an alternative climate norm, represented by the 30-year daily moving average, and the respective temperature shock, represented by the difference between this value and the contemporaneous maximum temperature. For column (2), we first aggregate our final sample to the monthly level for each ozone monitor before estimating Equation (1.13) in order to simulate the application of our model to contexts with less granular data. This reduces our sample from 5,139,523 observations to 178,175. Despite this reduction, our results remain qualitatively similar. In column (3) we augment our main specification by including further meteorological controls, daily average windspeed and total daily sunlight, in our matrix of additional regressors. While both coefficients are strongly significant, they do not meaningfully affect our coefficients of interest, but drastically restrict our total sample size. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 184 Table A.2.5: Alternative Clustering and Bootstrapped Standard Errors Daily Max Ozone Levels (ppb) (1) Temperature Shock 1.678*** (County Cluster) (0.063) (State Cluster) (0.134) (Bootstrapped) (0.063) Climate Norm 1.164*** (County Cluster) (0.051) (State Cluster) (0.091) (Bootstrapped) (0.051) Implied Adaptation 0.514*** (County Cluster) (0.041) (State Cluster) (0.106) (Bootstrapped) (0.042) All Controls Yes R2 0.481 Observations 5,139,523 Notes: This table compares the standard errors of our main estimates with ones obtained by clustering at the state- rather than county-level, and by bootstrap (block method clustered at the county level, 1000 iterations). The latter addresses the potential concern that because our temperature shocks and norm are constructed, they could be seen as generated regressors. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 185 Table A.2.6: Non-Linear Effects of Temperature Panel A. Below 20◦C Daily Max Ozone Levels (ppb) Adaptation (1) (2) Temperature Shock 0.691*** (0.017) Climate Norm 0.142*** 0.550*** (0.034) (0.030) Panel B. 20-25◦C Temperature Shock 1.694*** (0.072) Climate Norm 1.278*** 0.417*** (0.069) (0.031) Panel C. 25-30◦C Temperature Shock 2.017*** (0.087) Climate Norm 1.826*** 0.191*** (0.092) (0.041) Panel D. 30-35◦C Temperature Shock 2.196*** (0.096) Climate Norm 1.496*** 0.700*** (0.128) (0.070) Panel E. Above 35◦C Temperature Shock 2.049*** (0.135) Climate Norm 0.901*** 1.148*** (0.180) (0.136) Observations 5,139,523 R2 0.494 Notes: This table reports the average marginal effect of a 1◦Celsius increase in the temperature shock and climate norm on the daily maximum ambient ozone concentration (ppb) for days in which the daily maximum temperature fell within different temperature bins. We categorize temperature into 5 bins from < 20◦C to > 35◦C with 5◦C intervals in between. Estimates in column (1) correspond to Equation (1.13), interacting all variables with indicators for each temperature bin, while estimates in column (2) report the implied measure of adaptation. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1, plus the un-interacted indicators for each temperature bin to allow the intercept to vary across bins. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 186 Table A.2.7: Adaptation Under Linear, Binned, and Nonlinear Specifications Panel A. Below 20◦C Linear Binned Quadratic Cubic (1) (2) (3) (4) Implied Adaptation 0.514*** 0.550*** 0.708*** 1.135*** (0.041) (0.030) (0.074) (0.073) Panel B. 20-25◦C Implied Adaptation 0.514*** 0.417*** 0.541*** 0.587*** (0.041) (0.031) (0.040) (0.049) Panel C. 25-30◦C Implied Adaptation 0.514*** 0.191*** 0.375*** 0.420*** (0.041) (0.041) (0.061) (0.048) Panel D. 30-35◦C Implied Adaptation 0.514*** 0.700*** 0.209* 0.633*** (0.041) (0.070) (0.109) (0.101) Panel E. Above 35◦C Implied Adaptation 0.514*** 1.148*** 0.043 1.227*** (0.041) (0.136) (0.161) (0.210) All Controls Yes Yes Yes Yes Observations 5,139,523 5,139,523 5,139,523 5,139,523 R2 0.481 0.494 0.483 0.486 Notes: This table reports implied adaptation estimates across the temperature distribution recovered via four alternative specifications. Our central linear specification is shown in column (1), while column (2) depicts the binned specification shown in Table A.2.6, column (3) and (4) show the results of quadratic and cubic specifications, respectively, following Equation (1.14). Recall that for the quadratic and cubic models, calculating marginal adaptation requires choosing a value for the underlying climate norm. Thus, calculations in columns (3) and (4) use the values of 17.5, 22.5, 27.5, 32.5, and 37.5◦Celsius to correspond to the “mid-points” of the respective temperature bins used in column (2). Additionally, recall that implied adaptation reflects the difference between the climate norm, represented by the 30-year monthly moving average of maximum temperature, lagged by one year, and the temperature shock, represented by the difference between this value and the contemporaneous maximum temperature. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1, plus the un-interacted indicators for each temperature bin in column (2). Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 187 Table A.2.8: Results by Decade Panel A. 1980’s Daily Max Ozone Levels (ppb) Adaptation (1) (2) Temperature Shock 2.264*** (0.142) Climate Norm 1.726*** 0.539*** (0.086) (0.088) Panel B. 1990’s Temperature Shock 1.768*** (0.051) Climate Norm 1.339*** 0.428*** (0.049) (0.037) Panel C. 2000’s Temperature Shock 1.280*** (0.030) Climate Norm 0.743*** 0.537*** (0.034) (0.030) All Controls Yes Observations 5,139,523 R2 0.490 Notes: This table reports our main estimates disaggregated by the three “decades” in our sample: 1980-1990; 1991-2001 and 2002-2013. Estimates in column (1) correspond to Equation (1.13), while estimates in column (2) report the implied measure of adaptation. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 188 Table A.2.9: Adaptation by Belief in Climate Change Daily Max Ozone Levels (ppb) Adaptation (1) (2) Temperature Shock 1.442*** (0.040) x Low Belief −0.141** (0.061) x High Belief 0.503*** (0.114) Climate Norm 0.998*** 0.445*** (0.054) (0.051) x Low Belief 0.047 −0.188*** (0.071) (0.063) x High Belief 0.310*** 0.193** (0.102) (0.085) All Controls Yes Observations 5,139,523 R2 0.484 Notes: This table reports estimates of temperature shock and climate norm interacted with an indicator of whether the residents of the county generally believed in climate change or not. Specifically, all counties in the sample were split into terciles based on the results of a survey conducted on climate change beliefs (Howe et al., 2015). In column (1) the main effect reflects the result for the median tercile of counties, while the interacted effects reflect the difference from this value observed in the lower and higher tercile counties. Column (2) reports the implied measure of adaptation for the median counties along with the differential effect in the low and high belief counties. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 189 Table A.2.10: Adaptation by Belief in Climate Change Regulation Daily Max Ozone Levels (ppb) Adaptation (1) (2) Temperature Shock 1.507*** (0.046) x Low Belief −0.397*** (0.063) x High Belief 0.483*** (0.118) Climate Norm 1.115*** 0.392*** (0.061) (0.047) x Low Belief −0.344*** −0.053 (0.085) (0.068) x High Belief 0.210** 0.273*** (0.104) (0.084) All Controls Yes Observations 5,139,523 R2 0.486 Notes: This table reports estimates of temperature shock and climate norm interacted with an indicator of whether the residents of the county generally believed in the use of regulations on carbon emissions to combat climate change or not. Specifically, all counties in the sample were split into terciles based on the results of a survey conducted on climate change beliefs (Howe et al., 2015). In column (1) the main effect reflects the result for the median tercile of counties, while the interacted effects reflect the difference from this value observed in the lower and higher tercile counties. Column (2) reports the implied measure of adaptation for the median counties along with the differential effect in the low and high belief counties. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 190 Table A.2.11: Adaptation by Political Leaning Daily Max Ozone Levels (ppb) Adaptation (1) (2) Temperature Shock 1.325*** (0.047) x Democrat 0.558*** (0.100) Climate Norm 0.975*** 0.349*** (0.043) (0.042) x Democrat 0.302*** 0.256*** (0.085) (0.071) All Controls Yes Observations 5,139,523 R2 0.484 Notes: This table reports estimate of temperature shock and climate norm interacted with an indicator of whether the county voted Democrat in the 2008 presidential election. Column (1) follows Equation (1.13), with an additional interaction term for Democrat political preference depicting the differential effect of shocks and norms in these counties compared to baseline Republican voting counties. Similarly, column (2) reports the implied measure of adaptation for Republican leaning counties, with the differential effect in Democrat leaning counties noted by the interaction effect. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 191 Table A.2.12: Adaptation by VOC- or NOx-limited Atmosphere Daily Max Ozone Levels (ppb) Main Specification VOC/NOx-Limited Restricted Sample Adaptation Restricted Sample Adaptation (1) (2) (3) (4) Temperature Shock 2.135*** 2.185*** (0.165) (0.206) x VOC-limited −0.674** (0.281) x NOx-limited −0.082 (0.115) Climate Norm 1.378*** 0.757*** 1.347*** 0.838*** (0.140) (0.119) (0.139) (0.151) x VOC-limited −1.070*** 0.397 (0.335) (0.284) x NOx-limited 0.176 −0.258* (0.108) (0.139) All Controls Yes Yes Observations 1,007,563 1,007,563 R2 0.505 0.506 Notes: This table reports estimates of temperature shock and climate norm interacted with an indicator of whether the county was, on average, more VOC-limited, NOx-limited, or non-limited. Using 5-year bins (1980-1984, 1985-1989, etc.) a county is designated as VOC-limited, NOx-limited, or neither for each bin based on whichever of these three categories the county observed the most days of during the 5-year period. We restrict our sample to only those counties for which data on these precursor pollutants is available (approximately 20% of our full sample), and depict the results of our main specification under this restricted sample in column (1) for comparison, with the implied measure of adaption in column (2). In column (3), the main effect reflects the result for non-limited counties, while the interaction term depicts the relative difference in the effect of shocks and norms in precursor limited counties. Column (4) reports the implied measure of adaptation in non-limited counties, and the differential effect in limited ones. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. The full list of controls are the same as in the main model, depicted in column (1) of Table 1.1. Standard errors are clustered at the county level. ***, ** and * represent significance at the 1%, 5% and 10%, respectively. 192 A.3. Sources of Variation to Identify Climate Impacts This appendix provides further elaboration of the two sources of variation used to identify βC, the coefficient on ¯xip¯ in Equation (1.13) in our empirical application. That estimating equation employs monitor-by-season-by-year fixed effects, a more flexible way to control for a number of unobserved time-varying factors in our ambient ozone setting. Many least-squares estimators weight heterogeneity with factors that depend on group sizes and the within and between variances of explanatory variables. A univariate regression coefficient, for example, equals an average of coefficients in mutually exclusive (and demeaned) subsamples weighted by size and subsample x-variance (see Goodman-Bacon, 2018, footnote 11): βˆ = P i (yi − y¯)(xi − x¯) P i (xi − x¯) 2 = P A (yi − y¯)(xi − x¯) + P B (yi − y¯)(xi − x¯) P i (xi − x¯) 2 = nAs A xy + nBs B xy s 2 xx = nAs 2,A xx s 2 xx βˆ A + nBs 2,B xx s 2 xx βˆ B, (3.7) where A and B are mutually-exclusive subsamples, and as usual, βˆ j = s j xy s 2,j xx , j = A, B. (3.8) A simpler version of the estimating equation in our empirical application – Equation (1.13) – focusing on the time-varying temperature norm, ¯xip¯, is: yit = βCx¯ip¯ + ϕis + ϵit, (3.9) where y represents ambient ozone concentrations, i monitor, t day, and s season-of-the-sample, and ¯p refers to the aggregation of time – in our case, month – used to construct the climate 193 normals from past observations – the 30-year monthly moving averages of temperature. Applying a fixed-effects transformation in Equation (3.9), we obtain (yit − y¯j ) = βC(¯xip¯ − x¯j ) + (ϵit − ϵ¯j ), (3.10) where j represents subsamples defined by the trio monitor-by-season-by-year is. Using an alternative notation, we can express the transformed equation as y˜it = βCx˜ip¯ + ˜ϵit. (3.11) By running OLS on the transformed equation, analogous to the decomposition in Equation (3.7), we can express βˆ C as: βˆ C = X j njs 2,j x˜x˜ s 2 x˜x˜ βˆ C,j . (3.12) Equation (3.12) helps us understand how we are leveraging both variation across months within the subsample defined by the monitor-by-season-by-year trio, and variation over time to identify βC. Indeed, βˆ C incorporates variation in temperature norms within each subsample j, s 2,j x˜x˜ , and variation in how economic agents located around the same monitor respond to temperature norms in different points in time, which is captured by βˆ C,j . 194 Appendix B: Incidental Adaptation: The Role of Non-Climate Regulations This appendix provides details on the construction of the data, descriptive figures, and the tabular results of robustness tests and explorations of heterogeneity using alternate specifications, as well as proofs and extensions of our analytical model. In Appendix B.1, we provide background information on the National Ambient Air Quality Standards for ozone, ozone formation, and further details on the sources of our data and construction of final variables. We additionally include maps of both weather and ozone monitoring station locations, illustrative figures of our decomposition of temperature and its relationship with ozone concentration, and tables of summary statistics. Appendix B.2 includes additional discussion of alternate specifications, split between those investigating robustness in Subsection B.2.a, and those examining heterogeneity in Subsection B.2.b. Appendix B.3 provides a formal derivation of our analytical model and extensions to examine alternative regulatory settings as discussed in Section II. 195 B.1. The National Air Quality Standards, Ozone Formation, and Additional Data Discussion This appendix section provides background information on the National Ambient Air Quality Standards in Section B.1.a as well as background information on ozone pollution in Section B.1.b. Section B.1.c then provides further details on the data sets discussed in Section III, as well as auxiliary data sets used in alternative specifications. It then includes relevant Figures and Tables as outlined below. Figure B.1.1.Ozone Monitor Location by Decade of First Appearance Figure B.1.2.Temperature Relative to Baseline (1950-1979) Figure B.1.3.Ozone Monitors and Matched Weather Monitors Figure B.1.4.Evolution of 4th Highest Ozone Concentration Figure B.1.5.Evolution of Nonattainment Designation in Monitored Counties Figure B.1.6.Decomposition of Climate Norms and Shocks - Estimating Sample Figure B.1.7.Relationship between Ambient Ozone and Temperature Figure B.1.8.Decomposition of Temperature Norms and Shocks (Los Angeles, 2013) Figure B.1.9.Decomposition of Temperature Norms and Shocks (Los Angeles, All Years) Figure B.1.10.Evolution of Ozone Concentration by Belief in Climate Change Table B.1.1.History of Ambient Ozone NAAQS Table B.1.2.Ozone Monitoring Season by State Table B.1.3.Yearly Summary Statistics for Ozone Monitoring Network Table B.1.4.Yearly Summary Statistics for Temperature and Decomposition 196 B.1.a. Background Details on the National Ambient Air Quality Standards Ambient ozone is an important component of smog that is capable of damaging living cells, such as those present in the linings of the human lungs. With the Clean Air Act Amendments of 1970, EPA was authorized to set up and enforce a National Ambient Air Quality Standard (NAAQS) for ambient ozone. Since then, a nationwide network of air pollution monitors has allowed EPA to track ozone concentration, and a threshold is used to determine whether pollution levels are sufficiently dangerous to warrant regulatory action. Exposure to ambient ozone has been causally linked to increases in asthma hospitalization, medication expenditures, and mortality, and decreases in labor productivity (e.g., Neidell, 2009a; Moretti and Neidell, 2011; Graff Zivin and Neidell, 2012; Deschenes, Greenstone and Shapiro, 2017). If any monitor within a county exceeds the NAAQS, EPA designates the county to be out of compliance or in “nonattainment” (USEPA, 1979, 1997a, 2004, 2008, 2015a). The corresponding state is required to submit a state implementation plan (SIP) outlining its strategy for the nonattainment county to reduce air pollution levels in order to comply with NAAQS.71 Figure B.1.5 depicts all counties monitored under the NAAQS for ozone during the period 1980-2013, noting the decade in which they were first designated as in “nonattainment,” if ever. While the structure of enforcement is dictated by the CAA and the EPA, much of the actual enforcement activity is carried out by regional- and statelevel environmental protection agencies. In particular, EPA divides the country into 10 geographic regions, and significant portions of the EPA’s operations are conducted through these regional offices. For instance, regional EPA offices conduct inspections and/or issue 71In more details, the Clean Air Act defines air quality control regions (AQCRs) so that air quality is managed in a more localized manner (Section 107 of the CAA as codified in 40 CFR Part 81, Subpart B). Boundaries of AQCRs are usually based upon county lines or other political divisions, but it is important to highlight that each AQCR is a contiguous area where air quality is relatively uniform; where topography is a factor in air movement, AQCRs often correspond with airsheds. AQCRs may consist of two or more cities, counties or other governmental entities, and each region is required to adopt consistent pollution control measures across the political jurisdictions involved. Each AQCR is treated as a unit for the purposes of pollution reduction and achieving the NAAQS. They are designated on a pollutant-by-pollutant basis. For example, for nitrogen dioxide and sulfur dioxide, the AQCR for Nebraska is the entire state. For particulate matter, the state is divided into several AQCRs. 197 sanctions when a state’s enforcement is below required levels, and assist states with major cases. EPA allows counties with polluting firms from 3 to 20 years to adjust their production processes.72 Specifically, the CAA mostly mandates command-and-control regulations, requiring that plants use the “lowest achievable emissions rate” (LAER) technology in their production processes in nonattainment counties.73 However, if pollution levels continue to exceed the standards or if a county fails to abide by the approved plan, sanctions may be imposed on the county in violation. These sanctions may include the withholding of federal highway funds and the imposition of technological “emission offset requirements” on new or modified sources of emissions within the county (USCFR, 2005). The first NAAQS for ambient ozone was established in 1979, when 120 parts per billion (ppb) was defined as the maximum 1-hour concentration that could not be violated more than once a year for a county to be designated as in attainment (USEPA, 1979).74 The CAA requires periodic review and, if appropriate, revision of existing air quality criteria to reflect advances in scientific knowledge on the effects of the pollutant on public health and welfare. So, in 1997, the standards were strengthened to 80ppb, but with a different form for the threshold: annual fourth-highest daily maximum 8-hour concentration averaged over 3 years (USEPA, 1997a).75 The 1997 NAAQS were challenged in court, and not enforced until 2004 (USEPA, 2004). In 2008, the standards were revised downward to 75ppb (USEPA, 2008). The latest revision happened in 2015, and the current 8-hour threshold is 70ppb 72“Nonattainment” counties are “classified as marginal, moderate, serious, severe or extreme (...) at the time of designation” (USEPA, 2004, p.23954). The maximum period to reach attainment is: “Marginal – 3 years, Moderate – 6 years, Serious – 9 years, Severe – 15 or 17 years, Extreme – 20 years” (USEPA, 2004, p.23954). 73EPA also mandates “best available control technology” (BACT) to curb emissions of local pollutants from large point sources even in attainment counties. Abatement costs are considered in formulating BACT standards, but not LAER standards. 74As Appendix Table B.1.1 shows, the standard put in place in 1971 was not focusing on ambient ozone, but rather all photochemical oxidants. 75EPA justified the new form as equivalent to the empirical 1-hour maximum to not be exceeded more than once a year. “The 1-expected-exceedance form essentially requires the fourth-highest air quality value in 3 years, based on adjustments for missing data, to be less than or equal to the level of the standard for the standard to be met at an air quality monitoring site” (USEPA, 1997a, p.38868). 198 (USEPA, 2015a). The EPA is currently conducting a review of the air quality criteria and the NAAQS for photochemical oxidants including ozone (USEPA, 2019).76 In accordance with the prevailing regulatory standard for the majority of our sample period – 1980-2004 – we use the 1-hour maximum ozone concentration level (ppb) for our empirical analysis. B.1.b. Background Details on Ozone Background on Ozone — The ozone the U.S. EPA regulates as an air pollutant is mainly produced close to the ground (tropospheric ozone).77 It results from complex chemical reactions between pollutants directly emitted from vehicles, factories and other industrial sources, fossil fuel combustion, consumer products, evaporation of paints, and many other sources. These highly nonlinear Leontief-like reactions involve volatile organic compounds (VOCs) and oxides of nitrogen (NOx) in the presence of sunlight. In “VOC-limited” locations, the VOC/NOx ratio in the ambient air is low (NOx is plentiful relative to VOC), and NOx tends to inhibit ozone accumulation. In “NOx-limited” locations, the VOC/NOx ratio is high (VOC is plentiful relative to NOx), and NOx tends to generate ozone. As a photochemical pollutant, ozone is formed only during daylight hours, but is destroyed throughout the day and night. It is formed in greater quantities on hot, sunny, calm days. Indeed, major episodes of high ozone concentrations are associated with slow moving, high pressure systems, which are associated with the sinking of air, and result in warm, generally cloudless skies, with light winds. Light winds minimize the dispersal of pollutants emitted in urban areas, allowing their concentrations to build up. Photochemical activity 76A summary of the changes in the form and levels of the NAAQS for ambient ozone is provided in Appendix Table B.1.1. Additionally, during our period of analysis (1980-2013), nitrogen dioxide (NO2) also had its own NAAQS, but there were no changes from 1971 to 2010. Furthermore, from 2003 to 2008, there was a cap-and-trade program created to reduce the regional transport of NOx emissions from power plants and other large combustion sources in the eastern United States – the NOx Budget Trading Program (NBP), which was shown to be effective in reducing ozone concentrations (Deschenes, Greenstone and Shapiro, 2017). There were also regulations targeting VOCs: restrictions on the chemical composition of gasoline that are primarily intended to reduce VOC emissions from mobile sources. Apart from the more stringent regulations in California, these regulations have been shown to be ineffective in reducing ambient ozone concentrations (Auffhammer and Kellogg, 2011). 77It is not the stratospheric ozone of the ozone layer, which is high up in the atmosphere, and reduces the amount of ultraviolet light entering the earth’s atmosphere. 199 involving these precursors is enhanced because of higher temperatures and the availability of sunlight. Modeling studies point to temperature as the most important weather variable affecting ozone concentrations.78 Ambient ozone concentrations increase during the day when formation rates exceed destruction rates, and decline at night when formation processes are inactive.79 Ozone concentrations also vary seasonally. They tend to be highest during the late spring, summer and early fall months.80 The EPA has established “ozone seasons” for the required monitoring of ambient ozone concentrations for different locations within the U.S.81 Recently, there is growing concern that the ozone season may prolong with climate change (e.g., Zhang and Wang, 2016). B.1.c. Further Details on the Construction of the Data Weather Data — Meteorological data was obtained from the National Oceanic and Atmospheric Administration’s Global Historical Climatology Network database (NOAA, 2014). This dataset provides detailed weather measurements at over 20,000 weather stations across the country, for which we use the period April-September, 1950-2013, for the contiguous 48 states. In constructing our complete, unbalanced panel of weather stations we make only one restriction: for each weather station in each year, we include only those stations for which valid measurements of maximum and minimum temperature, as well as precipitation, exist for at least 75 percent of the days in the ozone monitoring season (April-September). Figure 78Dawson, Adams and Pandisa (2007), for instance, examine how concentrations of ozone respond to changes in climate over the eastern U.S. The sensitivities of average ozone concentrations to temperature, wind speed, absolute humidity, mixing height, cloud liquid water content and optical depth, cloudy area, precipitation rate, and precipitating area extent were investigated individually. The meteorological factor that had the largest impact on ozone metrics was temperature. Absolute humidity had a smaller but appreciable effect. Responses to changes in wind speed, mixing height, cloud liquid water content, and optical depth were rather small. 79In urban areas, peak ozone concentrations typically occur in the early afternoon, shortly after solar noon when the sun’s rays are most intense, but persist into the later afternoon. 80In areas where the coastal marine layer (cool, moist air) is prevalent during summer, the peak ozone season tends to be in the early fall. 81Appendix Table B.1.2 shows the ozone season for each state during which continuous, hourly averaged ozone concentrations must be monitored. 200 B.1.2 plots annual deviations of temperature from the 1950-1979 baseline average. These are the thin solid, dotted, and dashed lines, representing average, maximum, and minimum temperature, respectively. The baseline represents both the pre-ozone regulation era as well as, generally speaking, the pre-climate change awareness era. The climate trend relative to this baseline – the smoothed thick solid line in the figure – has been slowly but steadily increasing since the mid-1970s, with an increase in the average temperature of approximately 0.5 degrees Celsius by 2010. This is consistent with findings from the U.S. Fourth National Climate Assessment, which indicate an increase in average temperature of 0.7 degrees Celsius for the period 1986-2016 relative to 1901-1960 (Vose et al., 2017). We decompose average temperature into a climate norm (30 year monthly moving average, lagged by 1 year) and a temperature shock (deviation of daily temperature from the climate norm). Figure B.1.6 depicts similar variation in both the climate norm and temperature shock as Figure 2.3, but using only the temperature assigned to each ozone monitor in our final sample. Notice that there seems to be more variation in the 30-year MA in the latter figure because it includes cross-sectional variation as well. Also, the 30-year MA trends down towards the end of the period of our study due to changes in ozone monitor location over time, as shown in Figure B.1.1. Table B.1.4 reports summary statistics for maximum temperature and our decomposed measures of climate norm and temperature shock, averaged across our entire sample for each year 1980-2013. Figures B.1.8 and B.1.9 provide illustrative examples of this decomposition for Los Angeles county for a single year – 2013 – and for the entire period 1980-2013, respectively. Ozone Data — Ambient ozone concentration data was obtained from the Environmental Protection Agency’s Air Quality System (AQS) AirData database, which provides daily readings from the nationwide network of the EPA’s air quality monitoring stations. The data was made available by a Freedom of Information Act (FOIA) request. In our preferred specification we use an unbalanced panel of ozone monitors. We make only two restrictions to construct our final sample. First, we include only monitors with valid daily information. 201 According to EPA, daily measurements are valid for regulation purposes only if (i) 8-hour averages are available for at least 75 percent of the possible hours of the day, or (ii) daily maximum 8-hour average concentration is higher than the standard. Second, as a minimum data completeness requirement, for each ozone monitor we include only years for which least 75 percent of the days in the ozone monitoring season (April-September) are valid; years having concentrations above the standard are included even if they have incomplete data. We have valid ozone measurements for a total of 5,638,273 monitor-days.82 The number of total valid monitors increased from 1,361 in the 1980s to 1,851 in the 2000s, indicating a growth of 16.6 percent of the ozone monitoring network per decade.83 The number of monitored counties in our main estimating sample also grew from 585 in the 1980s to 840 in the 2000s. Figure B.1.1 depicts the evolution of our sample monitors over the three decades in our data, and illustrates the expansion of the network over time. Table B.1.3 provides some summary statistics regarding the increase in the number of monitors over time.84 Figure B.1.4 depicts the daily maximum 1-hour ambient ozone concentrations from 1980- 2013, split by counties in and out of attainment of the ozone NAAQS. In this figure we compare the trends in ozone concentrations with the updated 1997, 2008 and 2015 NAAQS. These standards were based on the observed 4th Highest 8-hour average ambient ozone concentration of 80, 75 or 70 ppb respectively. Figure B.1.4 contrasts these attainment cut-offs with the maximum yearly ozone concentrations in attainment and nonattainment counties. Table B.1.1 clearly illustrates the evolution of the National Ambient Air Quality Standards for ozone over the years. Alternatively, Figure B.1.10 compares the trends in ozone concentrations from 1980-2013 for counties with low- median- and high-belief in climate 82Note that this value refers to all valid ozone measurements, the final samples used in estimation will be smaller due to, e.g., instances where an ozone monitor is not paired with any weather stations under our matching algorithm. For instance, our main estimating sample contains 5,139,529 valid monitor-day observations. 83For our main estimating sample, these are 1,285 and 1,701, respectively. 84Note that not all monitored counties were monitored in every year, and not all monitoring stations were active in every year. Some monitors were phased in to replace others, while others were simply added to the network over time as needed – thus individual years will generally have less unique monitors and monitored counties than existed across an entire decade or the sample period. 202 change. Notably, the concentrations appear to be converging over time – high-belief counties started out with higher baseline ozone levels, but over time reduced them to almost be in-line with low- and median-belief counties. Matching Ozone and Weather Data — These weather stations are typically not located adjacent to the ozone monitors. Hence, we develop an algorithm to obtain a weather observation at each ozone monitor in our sample. Using information on the geographical location of ozone monitors and weather stations, we calculate the distance between each pair of ozone monitor and weather station using the Haversine formula. Then, for every ozone monitor we exclude weather stations that lie beyond a 30 km radius of that monitor. Moreover, for every ozone monitor we use weather information from only the closest two weather stations within the 30 km radius. Once we apply this algorithm, we exclude ozone monitors that do not have any weather stations within 30km. We calculate weather at each ozone monitor location as the weighted average of these two weather stations using the inverse of the squared distance between them. Figure B.1.3 illustrates the proximity of our final sample of ozone monitors to these matched weather stations. We additionally assess the robustness of our results to changes in this algorithm by increasing the radius to 80 km and using the 5 closest weather stations, and by varying the weights used – unweighted arithmetic mean and simple inverse distance weighting – in calculating the approximate daily weather at each ozone monitoring location. The results of our model under these alternative specifications is discussed further in Appendix B.2.a. After matching ozone monitors with weather stations, we have valid ozone and temperature measurements for a total of 5,139,529 monitor-days. Figure B.1.7 illustrates the close association between ambient ozone concentrations and both components of temperature. Notice that the relationship between ozone and the climate norm, depicted in Panel A of Figure B.1.7 appears to be weaker than that with the temperature shock, in Panel B. This suggests that economic agents undertake adaptive behavior, after having observed the historical climate norm. 203 Auxiliary Data — In some of our robustness checks and examination of heterogeneity we incorporate additional data sets. Sources and any necessary data construction steps are described below. In column (3) of Table B.2.3 we include a monitor-day level interaction term for whether the local air quality authority had issued an ozone “action day” alert for the respective county. These “action day” alerts are often made day-of, or a few days in advance of, days in which the relevant air quality authority observes, or expects to observe, unhealthy levels of pollution on the Air Quality Index and releases a public service announcement to this effect. Individuals and firms are urged to take voluntary action to reduce the emissions of pollutants that are conducive to ozone formation. Note that although action day policies first began in the 1990’s, EPA only provides data beginning in 2004, leading to a restricted overall sample (approximately 36% of our full sample). In Table B.2.4 we include average daily windspeed and total daily sunlight as additional regressors within our main specification. These data, although recorded less frequently, are collected at the same weather monitoring stations as our main temperature and precipitation variables. Due to the sparseness of these data we do not decompose them into a long-run climate component and transitory weather shock as we do with temperature and precipitation. Additionally, it has been shown that, e.g., manufacturing plants have relocated in response to ozone nonattainment designations (Henderson, 1996; Becker and Henderson, 2000). In Table 2.2 we replace our daily ozone dependent variable with measures of (logged) monthly employment or quarterly wages at the county level obtained from the Quarterly Census of Employment and Wages. In Table B9 we examine heterogeneity in our results when separating counties into lowmedian- and high-levels of belief regarding the existence of climate change. These measures were constructed using county level survey data collected by Howe et al. (2015) in 2013 which estimate the percentage of each county’s respective population that hold such beliefs. 204 Notably, we do not rely on the explicitly stated aggregate level of belief, but rather the relative level of belief compared to the rest of our sample. Specifically, we separate counties into low- median- or high-belief terciles based on their stated level of belief in the existence of climate change. In this way we arrive at three equally sized groups for which we are able examine heterogeneity in climate impacts and adaptive response. For reference, Table B10 provides summary statistics of basic demographic characteristics across these three county groupings using data from the 2006-2010 5-year American Community Survey. As a placebo check we also examine the heterogeneity in our results when separating counties into low- median- and high-belief regarding “preferences” for single-parenthood in Table B11. Similar to our construction of “climate beliefs,” we begin with a measure of the fraction of single-parent households at the county level from the Opportunity Atlas (Chetty et al., 2018). We then again separate counties into low- median- or high-belief terciles based on their relative level of “preference” for single-parenthood. In this way we arrive at three equally sized groups for which we are able examine heterogeneity in climate impacts and adaptive response. In Table B12 we use measures of whether a county is “VOC-limited” or “NOx-limited.” These measures were constructed using data collected by the EPA’s network of respective monitoring stations. Note, however, that these are often separate pollution monitors from our main sample of ozone monitors. Additionally, data – especially for VOCs – is relatively sparse compared to ozone data. Due to these data constraints, we construct measures of whether a county is, in general, VOC-limited or NOx-limited for each 5-year period in our sample, e.g. 1980-1984, which we then match with our sample of ozone monitors at the county level. To construct these measures we first combine the EPA’s VOC and NOx data at the county-day level and generate a daily ratio of VOCs to NOx for each county. Following the scientific literature, observations with a ratio less than or equal to 4 are coded as VOClimited, while those greater than 15 are coded NOx-limited, and the remainder are coded as non-limited. We then sum these three measures by county across each 5-year interval 205 and denote a county as VOC-limited, NOx-limited, or non-limited for that interval based on whichever measure was the most prevalent. For example, a county with 50 VOC-limited day, 20 NOx-limited days, and 30 non-limited days would be marked as VOC-limited for this 5-year window. Admittedly, this creates a somewhat coarse measure of whether a county is VOC- or NOx-limited. Given the available data, however, this appears to be the furthest this question can be investigated, and, if anything, should be expected to bias the observed effect from this heterogeneity towards zero. 206 Figure B.1.1: Ozone Monitor Location by Decade of First Appearance Notes: This figure depicts the evolution of ozone monitors in our sample over three decades and illustrates the expansion of the monitoring network. We use an unbalanced panel of ozone monitors, after making the following two restrictions. Firstly, we only include monitors if 8-hour averages are available for at least 75 percent of the possible hours of the day, or (ii) daily maximum concentration is higher than the standard. Secondly, as a minimum data completeness requirement, for each ozone monitor we include only years for which least 75 percent of the days in the typical ozone monitoring season (April-September) are valid; years having concentrations above the standard are included even if they have incomplete data. We have valid ozone measurements for a total of 5,139,529 monitor-days after matching monitors with weather stations. The number of unique valid monitors increased from 1,285 in 1980 to over 1,850 in the 2000’s. 207 Figure B.1.2: Temperature Relative to Baseline (1950-1979) -1 -.5 0 .5 1 1.5 Deviation from 1950-1979 Baseline Temperature (C°) 1950 1960 1970 1980 1990 2000 2010 Avg. Temp Max. Temp Min. Temp Notes: This figure depicts annual temperature fluctuations and the overall climatic trend for the ozone season in the US relative to a 1950-1979 baseline average. The baseline and the yearly deviations from it are constructed from the comprehensive sample of weather stations across the US from 1950 to 2013 following the data construction steps detailed in Appendix B.1.c. The 1950-1979 baseline represents, generally speaking, the pre-climate change awareness era. The average temperature, relative to this baseline, has been slowly but steadily increasing since 1980, with an increase in the average temperature of approximately 0.5 degree Celsius (◦C) by 2010. For clarity, the thin solid line, the short-dashed line, and long-dashed line refer to annual averages for daily average, maximum, and minimum temperature, respectively, as coded in the legend. The thick solid line smooths out the annual observations for average temperature over the period covered in the graph. 208 Figure B.1.3: Ozone Monitors and their Matched Weather Monitors Notes: This figure illustrates the proximity of our final sample of ambient ozone monitors to the matched weather stations. Using information on the geographical location of pollution monitors and weather stations we calculate the Haversine distance between each pair of ozone monitor and weather station. Then every ozone monitor is matched to the closest two weather stations within a 30 km radius of the monitor. We exclude ozone monitors that do not have any weather station within a 30 km radius. Once the monitors are matched to weather stations, we generate the approximate weather realizations at the ozone monitor by averaging the meteorological variables at the matched weather stations, weighted by their inverse squared distance from the monitor. 209 Figure B.1.4: Evolution of the 4th Highest Ambient Ozone Concentration 1997 NAAQS 2008 NAAQS 2015 NAAQS 60 80 100 120 140 Daily Max Ozone Concentration (ppb) 1980 1990 2000 2010 Attainment Counties Nonattainment Counties Notes: This figure depicts the national average of the annual 4th highest daily maximum 1-hour ambient ozone concentration over time in the US, split by counties designated as in- or out- of attainment under the National Ambient Air Quality Standards (NAAQS). The 1997, 2008, and 2015 NAAQS updates for designating a county’s attainment status were based on the observed 4th highest 8-hour average ambient ozone concentration of 80, 75, and 70 ppb or higher, respectively. Here we contrast these attainment status cutoffs with the yearly ozone concentrations in Attainment and Nonattainment counties. 210 Figure B.1.5: Map of Monitored Counties - by First Decade Designated in Nonattainment Notes: This figure illustrates all counties monitored under the NAAQS for ozone during the period 1980-2013, noting the decade in which they were first designated as in “nonattainment,” if ever. While the structure of enforcement is dictated by the CAA and the EPA, much of the actual enforcement activity is carried out by regional- and state-level environmental protection agencies. Most counties out of attainment were first designated in nonattainment in the 1980s. The map displays concentrations of those counties in California, the Midwest, and in the Northeast. Nevertheless, a nontrivial number of counties went out of attainment for the first time in the 1990s and 2000s. 211 Figure B.1.6: Climate Norms and Shocks (Final Sample) 26.4 26.6 26.8 27.0 27.2 27.4 30-year Moving Average (Temperature C°) 1980 1990 2000 2010 Panel A. Average Climate Norm Over Time -0.5 0.0 0.5 1.0 Deviation from Moving Average (Temperature C°) 1980 1990 2000 2010 Panel B. Average Temperature Shock Over Time Notes: This figure depicts US temperature over the years in our sample (1980-2013), decomposed into their climate norm and temperature shock components. The climate norm (Panel A) and temperature shocks (Panel B) are constructed from the unbalanced panel of weather stations included in our main model sample, restricting the months over which measurements were gathered to specifically match the ozone season of April–September, the typical ozone season in the US (see Appendix Table B.1.2 for a complete list of ozone seasons by state). The unbalanced feature of our main sample, with ambient ozone monitors moving north over time (see Figure B.1.1), is the likely driving force behind the downward pattern of the average climate norm at the end of our sample period in Panel A. The horizontal dashed lines in Panel B highlights that temperature shocks are bounded in our period of analysis. 212 Figure B.1.7: Relationship between Ambient Ozone and Temperature -4 -2 0 2 4 Detrended Ambient Ozone (ppb) -0.6 -0.3 0.0 0.3 0.6 Detrended 30-year Moving Average (Temperature C°) 1980 1990 2000 2010 Climate Norm Max Ozone Level Panel A. Relationship Between Ozone and Climate Norm -4 -2 0 2 4 Detrended Ambient Ozone (ppb) -1.0 -0.5 0.0 0.5 1.0 Detrended Deviation from Moving Average (Temperature C°) 1980 1990 2000 2010 Temperature Shock Max Ozone Level Panel B. Relationship Between Ozone and Temperature Shock Notes: This figure depicts the general relationship between daily maximum ozone concentrations and temperature over the years in our sample (1980-2013) after decomposing temperature into our measure of climate norm and temperature shock and de-trending the data. Both the climate norm (Panel A) and the temperature shock (Panel B) appear to have a close correlation with ozone concentrations, although the relationship in Panel (A) appears weaker than that in Panel (B), providing suggestive evidence of adaptative behavior. Recall that the climate norm represents the 30-year monthly moving average of the maximum temperature, lagged by one year, while the temperature shock represents the difference between this value and the contemporaneous maximum temperature. 213 Figure B.1.8: Decomposition of Temp. Norms & Shocks (Los Angeles, 2013) Panel A. Our Preferred Decomposition 0 20 40 Temperature (C°) Apr Jun Aug Oct Temp Shock Climate Norm Daily Temp Panel B. Fixed-Effect Decomposition 0 20 40 Temperature (C°) Apr Jun Aug Oct Deviation from Avg Average Temp Daily Temp Notes: This figure compares our preferred decomposition method with a standard fixed-effects approach using data from the 2013 Los Angeles ozone season. Panel A depicts the daily measure of temperature, decomposed into climate norm and temperature shock. By contrast, Panel B depicts the same daily measure of temperature, but decomposed into a typical fixed-effect average temperature and the deviations, after controlling for monthly fixed effects. The dashed lines indicate observed daily maximum temperature while the black solid lines represent long-run norms. The gray solid lines indicate temperature shocks which are nearly identical in both panels, as would be expected from the Frisch-Waugh-Lovell theorem, illustrating the variation used for identifying βW and βF E. Panel A additionally highlights the variation in climate used to identify βC in our proposed approach, while Panel B lacks any such variation in the measure of climate. 214 Figure B.1.9: Decomposition of Temp. Norms & Shocks (Los Angeles, All Years) Panel A. Our Preferred Decomposition -20 0 20 40 Temperature (C°) 1980 1990 2000 2010 Temp Shock Climate Norm Panel B. Fixed-Effect Decomposition -20 0 20 40 Temperature (C°) 1980 1990 2000 2010 Deviation from Avg Average Temp Notes: This figure illustrates the same comparison as in Figure B.1.8 for Los Angeles, but now using the entire sample period, 1980-2013. Specifically, Panel A depicts the daily measure of temperature, decomposed into climate norm and temperature shock. By contrast, Panel B depicts the same daily measure of temperature, but decomposed into a typical fixed-effect average temperature and the deviations, after controlling for monthly fixed effects. The dashed lines indicate observed daily maximum temperature while the black solid lines represent long-run norms. The gray solid lines indicate temperature shocks which are nearly identical in both panels, as would be expected from the Frisch-Waugh-Lovell theorem, illustrating the variation used for identifying βW and βF E. Panel A additionally highlights the variation in climate used to identify βC in our proposed approach, while Panel B lacks any such variation in the measure of climate. 215 Figure B.1.10: Evolution of Ozone Concentration by Belief in Climate Change 1979 NAAQS 80 100 120 140 160 Daily Maximum Ozone Concentration (ppb) 1980 1990 2000 2010 Low Belief Median Belief High Belief Maxiumum Ozone Concentration by Level of Belief Notes: This figure depicts the national average of the highest daily maximum 1-hour ambient ozone concentration over time in the US, split by counties with low- median- and high-belief in climate change. Notably, the concentrations appear to be converging over time – high-belief counties started out with higher baseline ozone levels, but over time reduced them to almost be in-line with low- and median-belief counties. Here we contrast these concentrations with the 1980’s attainment status cutoff of 120ppb threshold. 216 Table B.1.1: History of Ambient Ozone NAAQS Enacted Primary/ Indicator Averaging Level Form Secondary Time 1971 Primary and Total photo- 1-hour 80 ppb Hourly concentration not to be Secondary chemical oxidants exceeded more than one hour per year 1979 Primary and Ozone 1-hour 120 ppb Hourly concentration not to be Secondary exceeded more than one day per year 1997† Primary and Ozone 8-hour 80 ppb Annual fourth-highest daily maximum 8-hr Secondary concentration, averaged over 3 years 2008 Primary and Ozone 8-hour 75 ppb Annual fourth-highest daily maximum 8-hr Secondary concentration, averaged over 3 years 2015 Primary and Ozone 8-hour 70 ppb Annual fourth-highest daily maximum 8-hr Secondary concentration, averaged over 3 years Notes: This table shows the history of ambient ozone regulations in the U.S. The first standard was put in place in 1971, but targeted all photochemical oxidants. The first National Ambient Air Quality Standards (NAAQS) for ambient ozone was established in 1979, when 120ppb was defined as the maximum 1-hour concentration that could not be violated more than once a year for a county to be designed as in attainment. In 1997, the standards were strengthened to 80ppb, but with a different form for the threshold: annual fourth-highest daily maximum 8-hour concentration averaged over 3 years. With the 2008 and 2015 revisions, the current 8-hour threshold is now 70ppb. EPA justified the new form in 1997 as equivalent to the empirical 1-hour maximum to not be exceeded more than once a year. “The 1-expected-exceedance form essentially requires the fourth-highest air quality value in 3 years, based on adjustments for missing data, to be less than or equal to the level of the standard for the standard to be met at an air quality monitoring site” (USEPA, 1997a, p.38868). Lastly, as the EPA (2005) states, “primary standards set limits to protect public health, including the health of ‘sensitive’ populations such as asthmatics, children, and the elderly. Secondary standards set limits to protect public welfare, including protection against decreased visibility, damage to animals, crops, vegetation, and buildings.” Source: epa.gov/ozone-pollution/table-historical-ozone-national-ambient-air-quality-standards-naaqs. † The 1997 NAAQS was challenged in courts, and not implemented until 2004. 217 Table B.1.2: Ozone Monitoring Season by State State Start Month - End State Start Month - End Alabama March - October Nevada January - December Alaska April - October New Hampshire April - September Arizona January - December New Jersey April - October Arkansas March - November New Mexico January - December California January - December New York April - October Colorado March - September North Carolina April - October Connecticut April - September North Dakota May - September Delaware April - October Ohio April - October D.C. April - October Oklahoma March - November Florida March - October Oregon May - September Georgia March - October Pennsylvania April - October Hawaii January - December Puerto Rico January - December Idaho April - October Rhode Island April - September Illinois April - October South Carolina April - October Indiana April - September South Dakota June - September Iowa April - October Tennessee March - October Kansas April - October Texas1 January - December Kentucky March - October Texas1 March - October Louisiana January - December Utah May - September Maine April - September Vermont April - September Maryland April - October Virginia April - October Massachusetts April - September Washington May - September Michigan April - September West Virginia April - October Minnesota April - October Wisconsin April 15 - October 15 Mississippi March - October Wyoming April - October Missouri April - October American Samoa January - December Montana June - September Guam January - December Nebraska April - October Virgin Islands January - December Notes: This table shows, for each state, the season when ambient ozone concentration is required to be measured and reported to the U.S. EPA. The ozone season is defined differently in different parts of Texas. Source: USEPA (2006, p.AX3-3). 218 Table B.1.3: Yearly Summary Statistics for Ozone Monitoring Network Year # Observations # Counties # Ozone Monitors (1) (2) (3) (4) 1980 88426 361 609 1981 100459 399 659 1982 102111 402 661 1983 102429 408 653 1984 103828 390 649 1985 105457 388 648 1986 103820 375 634 1987 110366 392 668 1988 113232 409 686 1989 119938 425 725 1990 126535 443 757 1991 132046 466 792 1992 137754 482 821 1993 146023 511 863 1994 149400 520 876 1995 154230 528 902 1996 153019 530 894 1997 160024 550 931 1998 164491 568 960 1999 168901 585 982 2000 172686 592 999 2001 180872 616 1047 2002 186261 630 1071 2003 188462 641 1082 2004 189868 653 1087 2005 187709 649 1082 2006 188298 650 1075 2007 190824 661 1092 2008 190682 660 1099 2009 194184 678 1116 2010 196439 688 1130 2011 199948 716 1159 2012 199723 703 1148 2013 148306 658 1039 Notes: This table outlines the summary statistics of our main data sample. The construction of our main sample follows EPA guidelines by including all monitor-days for which 8-hour averages were recorded for at least 18 hours of the day and monitor-years for which valid monitor-days were recorded for at least 75% of days between April 1st and September 30th. 219 Table B.1.4: Yearly Summary Statistics for Daily Maximum Temperature Year Max Temp Climate Trend Temp Shock (1) (2) (3) (4) 1980 27.0 26.5 0.5 1981 26.9 26.6 0.4 1982 26.1 26.7 -0.6 1983 26.8 26.8 0.0 1984 26.7 26.8 -0.1 1985 27.0 26.6 0.3 1986 26.7 26.4 0.3 1987 27.3 26.6 0.7 1988 27.4 26.6 0.7 1989 26.4 26.7 -0.3 1990 26.7 26.6 0.1 1991 27.1 26.6 0.5 1992 26.1 26.7 -0.5 1993 26.6 26.6 0.0 1994 26.9 26.6 0.2 1995 26.8 26.7 0.0 1996 26.5 26.7 -0.2 1997 26.4 26.8 -0.4 1998 27.3 27.0 0.4 1999 27.2 27.0 0.2 2000 27.1 27.1 0.0 2001 27.4 27.2 0.3 2002 27.8 27.2 0.6 2003 26.9 27.3 -0.4 2004 27.0 27.2 -0.2 2005 27.6 27.3 0.3 2006 27.7 27.3 0.4 2007 27.7 27.3 0.4 2008 27.3 27.3 0.0 2009 26.9 27.3 -0.3 2010 27.8 27.2 0.6 2011 27.4 27.1 0.3 2012 28.0 27.1 0.9 2013 26.4 26.6 -0.3 Notes: This table outlines the evolution of maximum temperature in our sample from the years 1980–2013 in column (2). Columns (3) and (4) decompose this into our respective measures of climate norm and temperature shock. 220 B.2. Further Robustness Checks and Heterogeneity This appendix provides further elaboration of the alternative specifications used for robustness checks and examinations of heterogeneity as discussed in Section V. It then includes relevant Tables as outlined below. Table B.2.1.Alternative Lag Lengths of Nonattainment Indicator Table B.2.2.Alternative Lengths of Climate Norms Table B.2.3.Adaptation Responses Table B.2.4.Alternative Specifications and Sample Restrictions Table B.2.5.Alternative Criteria for Selection of Weather Stations Table B.2.6.Bootstrapped Standard Errors Table B.2.7.Results by Decade Tables B8a & B8b. Non-Linear Effects of Temperature Table B9. Adaptation by Local Beliefs in Climate Change Table B10. Beliefs in Climate Change: Summary Stats Table B11. Placebo: Preferences for Single Parenting Table B12. Adaptation by VOC- or NOx-limited Atmosphere 221 B.2.a. Further Robustness Checks Alternative Lag Lengths of Nonattainment Indicator — In our preferred specification we use a 3-year lag from when a county is designated as in nonattainment with the ozone NAAQS, corresponding with the EPA’s minimum timeline for the county to re-enter compliance with the NAAQS. When selecting this lag length on the indicator of a county’s nonattainment designation, however, two potential concerns arise. First, the EPA may choose to allow heavily polluted counties up to 6, 9, 15-17, or even 20 years to re-enter attainment (USEPA, 2004). With this extra time, there may be concern that such counties do not begin engaging in adaptive behavior immediately, delaying efforts until closer to the deadline set by the EPA. If this were the case, it would lead to a downward bias in our estimate of RIA, as these nonattainment counties would be behaving more in line with attainment counties during the early years of the nonattainment “treatment.” Notably, however, even when allowing for a longer deadline the EPA does require that counties show active progress towards re-entering attainment in the first three years of their nonattainment designation (USEPA, 2004). Second, because nonattainment is lagged by three years, the estimation implicitly assumes that any adaptation effects of being designated in nonattainment continue for three years after a county re-enters attainment. If counties that re-enter attainment immediately reduce their adaptation efforts, the 3-year lag could lead to attenuation of the estimated coefficient on adaptation in nonattainment counties and thus the estimate of RIA. Conversely, depending on the types of adaptive actions taken, they may not be easy to quickly undo and may naturally persist for some number of years after a county re-enters attainment, and it is unclear, a priori, which of these two effects might dominate. We investigate both concerns in Table B.2.1, shortening the lag to a single year and extending it to 6-years, the next longest deadline the EPA may assign a nonattainment county. With regards to the first concern, the estimate of adaptation in nonattainment counties is in fact smaller in magnitude, although statistically indistinguishable, when using a 6-year lag rather than our preferred 3-year lag. This suggests that our preferred 3-year 222 lag does not suffer from attenuation bias arising from heavily polluted counties potentially delaying adaptation efforts. This is perhaps unsurprising given the EPA’s requirement of active progress within the first three years. With regards to the second concern, while we find slight suggestive evidence that the estimate of RIA is indeed decreasing with increases in the lag length, the difference is small in magnitude, and we are unable to rule out that RIA is statistically identical across the 1-year, 3-year, and 6-year lagged specifications. Alternative Lengths of Climate Norms — A potential concern with our primary estimates reported in Table 2.1 might be the way in which we define our climate norm. Recall that we define the climate norm as the 30-year monthly moving average of temperature, lagged by one year. Although this is the usual definition of climate used in the literature by climatologists, in Table B.2.2, we address any possible concerns about measurement error impacting our results. In this table, we vary the length of time that we use in constructing the climate norms. In going from column (1) to (4), we report estimates using a 3-year, 5-year, 10-year and 20-year moving average as our climate norm. If we observe the coefficients of the climate norm, we see a slight increase in the magnitude as we move to longer-run averages. However, if we compare effect of the climate norm in column (4) of Table B.2.2 (20-year average) to column (2) of Table 2.1 (30-year average), we see a decline in magnitude. This latter result suggests that the widely used climate normals are close to the “optimal” long-run norms. The improvements from reducing measurement error might be offset by the panel-driven attenuation bias between 20- and 30-year time windows. Adaptation Responses — Given that in this paper, we speak at length about adaptation to climate change, and in particular, regulation-induced adaptation, another major concern might be the time given to economic agents to adapt. Recall that in our preferred specification, we define climate norm as the 30-year monthly moving average of temperature, lagged by one year (e.g., the 30-year moving average associated with May 1982 is the average May temperatures over the years 1952-1981). Thus, economic agents will have had at least one 223 year to respond and adapt to unexpected changes in the climate normal temperature. One might wonder whether one year is enough time for agents to adapt and adjust their behavior. To alleviate such concerns, we check the sensitivity of our results when agents have 10 or 20 years to adapt, instead of just one. In Table B.2.2 column (1), we define climate norm as a 20-year monthly moving average of temperature, lagged by 10 years such that economic agents have a decade to make adjustments in response to unexpected changes in the climate norm (e.g., the climate norm associated with May 1982 would now instead be the average of May temperatures over the years 1952-1971). Similarly, in column (2), we report estimates using a 10-year moving average as our climate norm, lagged by 20 years, giving even more time to economic agents to adapt. The estimated impacts are remarkably similar to our main findings, suggesting that economic agents react as soon as information becomes available to them and that those effects are persistent. In column (3) we turn to possibility of agents responding rapidly to weather shocks. Were this to be the case, such short-run adaptive behaviors would affect our benchmark impacts of temperature shocks and hence bias our estimates of regulation-induced adaptation downwards. To investigate this possibility, we make use of a widespread policy of “Ozone Action Day” (OAD) alerts, where a local air pollution authority would issue an alert, usually a day in advance, that meteorological conditions are expected to be more conducive to forming potentially hazardous levels of ambient ozone in the following day. As a result, individuals and firms are urged to voluntarily take actions that would reduce emissions of ozone precursors. Thus, if agents are adapting to contemporaneous weather shocks, these “action days” would be the days we would be most likely to observe an adaptive response. Interacting an indicator variable for days in which OAD alerts were issued for a given county with our other covariates, we find that such alerts have a negligible and statistically insignificant impact on the effect of a 1◦C change in the contemporaneous temperature shock in both attainment and nonattainment counties – signifying limited opportunities, or willingness, to adapt in the short term.85 85Although the recovered coefficients of temperature shocks, climate norms, and implied adaptation levels are quantitatively different for column (3) than columns (1) and (2), this is likely due to a difference in the 224 Alternative Specifications and Sample Restrictions — In Table B.2.4 we further explore the sensitivity of our results to changes in the primary econometric specification and additional sample restrictions. First, it may be a concern that our climate norm variable structures the long-run climate normal temperature as the 30-year monthly moving average, despite the fact that seasonal – or within-season – shifts in temperature are unlikely to exactly follow the calendar at a monthly level. We examine the sensitivity of our results to this decision by alternatively constructing this variable as a 30-year daily moving average, allowing it to vary arbitrarily within each month. Results of our main specification, substituting daily moving averages for the standard monthly ones, are presented in column (1). The impacts of both components of temperature in attainment as well as nonattainment counties are nearly identical to our original findings. Ultimately, we prefer the monthly moving average because it is likely that individuals recall climate patterns by month, not by day of the year, making the interpretation of adaptation more intuitive. Indeed, as mentioned before, broadcast meteorologists often talk about how a month has been the coldest or warmest in the past 10, 20, or 30 years, but not how a particular day of the year has deviated from the norm. Second, Muller and Ruud (2018a) argue that the location of pollution monitors is not necessarily random. The U.S. EPA maintains a dense network of pollution monitors in the country for two major reasons: (i) to provide useful data for the analysis of important questions linking pollution to its varied impacts, and (ii) to check and enforce regulations on criteria pollutants. These are conflicting interests: while monitors should be placed in regions having different levels of pollution to provide representative data, they might be placed in areas where pollution levels are the highest to maintain oversight. Not surprisingly, the authors find out that most of the monitors tend to be in areas where pollution levels have been high, and compliance with the regulation is a question. Following those authors’ results, we can expect that ozone monitors that have consistently underlying sample. EPA data on “action day” alerts were only provided from 2004 onwards, leading to a restricted overall sample (approximately 36% of our full sample). 225 been in our sample across all years must be located in areas having very high pollution levels, thus commanding constant monitoring and regulation by the EPA. To check if this claim is accurate, we run our analysis using a balanced sample of ozone monitors. Starting from our original sample, and using only monitors that have been in the data for every year from 1980-2013, we are left with 92 pollution monitors. The results are reported in column (2) of Table B.2.4. As expected, the temperature effects obtained from the balanced panel are larger than those in our main results, although the level of adaptation remains largely unchanged. Our preferred, unbalanced sample of monitors includes areas with different levels of air pollution, and thus estimates should be more representative of the entire country. Lastly, although temperature is the primary meteorological factor affecting tropospheric ozone concentrations, other factors such as wind speed and sunlight have also been noted as potential contributors. High wind speed may prevent the build-up of ozone precursors locally, and dilute ozone concentrations. Ultraviolet solar radiation should trigger chemical reactions leading to the formation of ground-level ozone. To test whether our main estimates are capturing part of the effects of wind speed and sunlight, we control for these variables in an alternative specification using a smaller sample containing those variables. Column (3) of Table B.2.4 presents our main results from estimating Equation (2.6) plus controls for average daily wind speed (meters/second) and total daily sunlight (minutes). As expected, higher wind speeds lead to lower ozone concentrations, and more sunlight leads to higher concentrations. We find that a 1 meter/second increase in average daily wind speed would decrease ozone concentrations by 2.2 ppb, whereas a 1 minute increase in daily sunlight leads to 0.01 ppb increase in ozone concentrations. Including these additional variables does not significantly change our primary estimates of interest, however, which remain statistically indistinguishable from our preferred model. Alternative Criteria for Selection of Weather Stations — An additional concern arises from the fact that weather stations are not necessarily sited next to ozone monitors. Because of this, we do not have an exact measure of temperature at the same geographic point as 226 our measure of ozone. As discussed in our data section, we define temperature at an ozone monitoring station as the mean of the reported daily maximum temperatures at the two closest weather stations within 30 kilometers, weighted by the inverse squared distance to the ozone monitor. In doing so, we are likely to approximate a good measure of the daily maximum temperature for the local region as a whole, while also maintaining a close geographic boundary around the ozone monitoring station so as not to influence this approximation with temperature readings from a weather station further away that may be subject to a different set of meteorological conditions. It’s possible, however, that a less strongly distance weighted mean would provide a more accurate measure of temperature for the overall local region – although likely less accurate at the ozone monitoring station itself – or that the 2-station and 30-kilometer cutoffs are too restrictive. We investigate the effects of lessening the distance weighting in the calculation of expected temperature at the ozone monitoring station, as well as relaxing the constraints on both the number of included weather stations and distance from the ozone monitor in Table B.2.5. Specifically, columns (1) and (2) report results of our main specification when we maintain the 2-station/30-kilometer restriction, but decrease the weighting scheme to either the simple arithmetic mean in column (1), or a non-squared inverse distance weight in column (2). Columns (3) and (4) use the same weighting schemes as in columns (1) and (2), but now include temperature readings from the 5 closest weather monitoring stations within 80 kilometers. Results in all four columns are relatively stable and consistent with our main specification. Alternative Clustering and Bootstrapped Standard Errors — In our preferred specification we estimate standard errors clustered at the county level – corresponding to the spatial level of the attainment/nonattainment designations. This approach to estimating the standard errors raises three potential concerns. First, the weather shock and climate norm may be considered generated regressors. To address this first concern we estimate standard errors via block-bootstrap at the county-level using 250 iterations. Second, one may be concerned that the choice of cluster may have the incorrect spatial 227 granularity. For example, while attainment status is defined by the EPA at the county level, actual enforcement is often carried out at the regional (air quality control region, AQCR) or state level via the required State Implementation Plans (SIPs). We thus additionally consider standard errors clustered at the state level – note that state clusters would nest both AQCR-level and county-level clusters, and thus be the most conservative of the three. Finally, because the climate norm variable is constructed as a monthly average, by definition its values are correlated within each month, suggesting that two-way clustering by both county and month may be ideal. This, however, presents an empirical challenge: our sample includes only the six months of April through September (the common ozone season across the US). Standard errors estimated with monthly clusters would thus likely suffer from small sample bias due to only having six clusters. To overcome this challenge we instead use weeks as the temporal clustering period, resulting in 24 clusters, estimating standard errors two-way clustered by county and week.86 Clustering by week in this way allows for arbitrary correlation between observations within a specific week of the year, across all years of the sample, (for each county, respectively), but not the full month associated with the climate norm. Alternatively, we could cluster temporally by month-of-the-sample. While this approach would have the benefit of allowing for arbitrary correlation within each month, addressing fully the concern with monthly defined climate norms, and would allow for enough clusters to avoid small sample bias concerns, it would not account for arbitrary correlation between the same month in different years. In unreported analyses, we find that the standard errors estimated via the month-of-sample two-way clustering approach are less conservative than those estimated via two-way clustering by county and week. Table B.2.6 reports the results of these alternative clustering methods as well as the bootstrapped standard errors. Bootstrapped standard errors are all within 6% of those estimated via clustering at the county level. State clustered and two-way clustered standard 86Because calendar weeks will often split across two different months, defeating the purpose of clustering in our context, we instead define “weeks” to split each month into four approximately equal periods. The first, second, and third weeks all consist of eight days, with the fourth week consisting of all days remaining in the month, either six or seven days depending on whether the month has 30 or 31 days. 228 errors are larger than the county clustered standard errors in our preferred specification, as might be expected, but all estimated coefficients remain statistically significant at the 1% level. B.2.b. Heterogeneity Results by Decade — To examine temporal heterogeneity, Table B.2.7 reports our results by decade. We split our sample into three “decades” – 1980-90, 1991-2001, and 2002-2013 – so that we have roughly the same number of years in each. We find that the effects of both the climate norm and temperature shock in attainment as well as nonattainment counties, are decreasing over time, as shown in column (1). In column (2), we report the implied measures of adaptation in nonattainment and attainment counties, for each of the three decades. By comparing these differential magnitudes of adaptation in nonattainment vs attainment counties, we can get our regulation-induced adaptation measures in each decade. The estimates suggest that regulation-induced adaptation was 0.39 ppb in the 1980’s, 0.28 ppb in the 1990’s, and 0.22 ppb in the 2000’s. While seeming to decrease over time, potentially driven by technological innovation and market forces in attainment counties, we cannot rule out that they are statistically indifferent from our primary estimates in Table 1. Looking at the recovered coefficients of β W and β C specifically, however, reveals an interesting trend. The ozone-temperature gradient itself declines meaningfully over time in both attainment and nonattainment counties, in line with what one might expect from previous studies suggesting that the CAA may induce innovation and diffusion of pollution abatement technologies (e.g., Popp, 2003, 2006). To that extent, our results – which focus on the static adaptation induced by the NAAQS – may present a lower-bound of the total adaptation induced by the CAA which may also have dynamic elements. Nonlinear Effects of Temperature — Because ozone formation may be intensified with higher temperatures, we also examine the heterogeneous nonlinear effects of daily maximum tem229 perature on ambient ozone concentrations. Similar to our previous investigations we start by creating indicator variables denoting whether the contemporaneous daily maximum temperature at a given ozone monitor falls within a certain 5◦C temperature bin. In this way, the marginal effect of a 1◦C change in either component of temperature is allowed to vary across each 5◦C temperature bin. As expected, we find that higher temperatures generally lead to higher ozone concentrations. The lowest bin is below 20◦C (just over the 10th percentile of our temperature distribution), and the highest bin is above 35◦C (90th percentile of our temperature distribution). Tables B8a and B8b present the results of our preferred specification when interacting each of these temperature bin indicators with our other covariates in column (1). The implied measures of adaptation for both nonattainment and attainment counties are presented in column (2). By comparing the adaptation estimates for nonattainment vs attainment counties we arrive at our measure of regulation-induced adaptation for each temperature bin. Below 20◦C, temperature impacts are much lower, as we would expect, although adaptation estimates are in line with our main specification. Between 20-25◦C and 25-30◦C, temperature impacts steadily increase, while adaptation estimates are lower and statistically distinguishable from our main specification. Once the temperature increases above 30◦C, however, the impact of the climate norm begins to attenuate – especially in nonattainment counties – and the estimate of regulation-induced adaptation increases substantially. Between 30-35◦C, the magnitude of regulation-induced adaptation is 50% larger than our main specification, and above 35◦C it is more than double, although we cannot rule out that they are statistically indifferent from our main specification. Notably, in nonattainment counties, adaptation reduces the effect of a 1◦C increase in temperature by over 60 percent when temperatures are above 35◦C, which is all the more relevant given the prospects of ever increasing temperatures in the coming decades. This relatively high level of adaptation above 35◦C – especially in nonattainment counties – can be plausibly explained by at least two reasons. First, regions having temperatures 230 above 35◦C might have higher incidence of sunlight which might lead to more extensive use of solar panels to generate electricity. Higher temperatures might be creating an environment that is more suited to shifts away from conventional and dirtier sources of power generation, thus leading to higher levels of adaptation. Second, and more specific to regulation-induced adaptation, days that are exceptionally hot are more likely to cause exceptionally high levels of ozone, which could trigger additional regulatory oversight. In order to avoid this, firms would be most likely to concentrate adaptation efforts on days where the “climate normal” temperature is itself the hottest. Results by Local Climate Beliefs — Here we examine whether climate change beliefs may alter the effectiveness of existing government regulations and policy in inducing climate adaptation. On the one hand, the enormous heterogeneity in economic and environmental preferences/beliefs across local jurisdictions (e.g., Howe et al., 2015) makes the enactment of comprehensive climate policy difficult (Goulder, 2020). On the other hand, the same heterogeneity in local beliefs may be able to be leveraged to push forward local actions supporting climate adaptation. Using the results of a relatively recent county-level survey regarding residents’ beliefs in climate change (Howe et al., 2015), we split the set of counties in our sample into terciles of high, median, and low belief, and interact indicators for high- and low-belief counties with our temperature and control variables.87 Appendix Table B10 shows that low-belief counties are, on average, less populous, poorer, and more politically conservative than mid-belief counties, while high-belief counties skew more towards the political left, are richer and more populous. Table B9 reports the results. The main temperature effects represent the mid-belief tercile, whose interactions are omitted, and the coefficients of the interactions with low- and high-belief terciles are relative to the omitted category. In column (1) we can see that the ozone response to temperature is consistently larger in high-belief counties relative to the 87Appendix Figure B.1.10 depicts the evolution of ozone concentration for these three sets of counties from 1980-2013. While the pattern for low- and median-belief counties track quite similarly, high-belief counties began with higher ozone concentrations, on average, but have now mostly converged with the other counties. 231 middle tercile, while for low-belief counties the evidence is mixed. This pattern is consistent with more economic activity in the more urban and richer high-belief counties. In column (2), the adaptation estimates for the mid-belief tercile are qualitatively similar to our main estimates for nonattainment and attainment counties, although the implied level of adaptation is somewhat muted for nonattainment counties and somewhat larger for attainment counties. Comparatively, adaptation in low-belief counties is statistically indistinguishable from the middle-tercile when in nonattainment, but 44 percent lower when in attainment. This pattern is reversed for high-belief counties, with statistically indifferent adaptation relative to the middle-tercile when in attainment, but 45 percent higher when out of attainment. These results translate into positive measures of regulation-induced adaptation across all three sets of counties, as seen in column (3) – although critically arising from different channels. Low-belief counties, bound by the NAAQS when in nonattainment, are constrained to meet at least the minimum level of ozone reduction, inducing adaptation levels similar to the middle-tercile. When in attainment, however, low-belief counties make much less effort than other counties to adapt – this is reasonable because they do not face stringent regulation, are generally poorer, and do not believe in climate change. In this case, the NAAQS induces adaptation by enforcing a required level of action. Conversely, high-belief counties engage in normal levels of adaptation when in attainment, but increase their adaptive behavior when in nonattainment. This, too, seems reasonable, as this set of counties is probably the most affected by the NAAQS, are generally richer – thus more able to afford the adjustments implied by the NAAQS – and are more believing in climate change – thus more willing to adjust behaviors or make investments in response to a changing climate. Because local beliefs in climate change are closely related to income, education, and political affiliation, one may wonder whether the heterogeneity in the response to environmental policy is not driven by other local unobserved factors. To provide evidence corroborating the role of environmentally-related local preferences, we investigate whether local views on 232 single parenthood, as proxied by the county fraction of children growing up in single-parent families in 2012-16 (Chetty et al., 2018), affect climate adaptation induced by the NAAQS for ambient ozone. For ease of comparison, we once again split counties into low- medianand high-“belief” counties based on this measure and interact the indicators for low- and high-belief with our other variables, taking the median as the baseline. Table B11 reports the results in the same format as Table B9. In column (1), the interactions of temperature shocks and norms in nonattainment and attainment counties are by and large not statistically significant. The implied adaptation estimates presented in column (2) show no meaningful changes for counties in the low- or high-belief terciles. More importantly, the estimates for regulation-induced adaptation displayed in column (3) are statistically indistinguishable across all terciles of local preferences for single parenthood. Thus, the local unobserved factors that may shape responses to environmental policy seem to be the ones related to local preferences for environmental amenities, as we have hypothesized. Ozone Formation in VOC- and NOx-limited Areas: Implications for Local Adaptation — As shown above, local climate change beliefs may effect the level of adaptation induced by the CAA. At the same time, the underlying composition of precursor emissions in the local atmosphere may also play an important role. Due to the Leontief-like production function of ozone, counties may find themselves with a baseline atmospheric composition that is “limited” in one precursor component – VOC or NOx. Urban areas are more prone to being VOC-limited, due to high levels of NOx pollution from production facilities and transportation, while rural areas are more prone to being NOx-limited due to the lack of such facilities and proximity to more VOC-rich undeveloped land. Counties with such a “limited” atmosphere may find it easier to adapt to climate change because even a small reduction in the limiting precursor’s emissions could lead to meaningful reductions in ozone. Nonattainment counties in particular may exploit this option in an attempt to bring themselves back into attainment, amplifying regulation-induced adaptation in precursor-limited areas. We explore this important feature of the production function of ozone in Table B12 by interacting our 233 main specification with indicators for whether a county is, in general, VOC- or NOx- limited – taking counties with non-limited atmosphere as the baseline. Unfortunately, data on VOC and NOx emissions are less available than for ozone,88 and thus our estimating sample is restricted to approximately 20 percent of our main sample. For reference, we thus first estimate our main specification on this reduced sample, finding results strikingly similar to Table 2.1, reported here in column (1), and in column (2) for the implied measures of adaptation. Columns (3) and (4) report estimated impacts and implied adaptation, respectively, once interacting our measures of VOC- and NOx-limited atmosphere. Our results suggest that while counties without a precursor-limited atmosphere still observe regulation-induced adaptation, the effect is almost quadrupled in VOC-limited counties. NOx-limited counties similarly see a large increase, approximately doubling the effect in non-limited counties, but the estimate is statistically imprecise – likely due to the smaller number counties that fall into this sub-group.89 88See Appendix B.1.c for further details of this data and our construction of the “limited” indicator variables. 89Specifically, observations in non-limited counties account for just under 60 percent of the estimating sample, while just over 36 percent are VOC-limited observations and the remainder, approximately 4 percent, are NOx-limited. 234 Table B.2.1: Alternative Lag Lengths of Nonattainment Indicator 1-year Lag 3-year Lag 6-year Lag (1) (2) (3) Nonattainment x Shock 2.028*** 1.990*** 1.935*** (0.085) (0.079) (0.068) Nonattainment x Norm 1.377*** 1.351*** 1.318*** (0.069) (0.067) (0.062) Attainment x Shock 1.221*** 1.263*** 1.241*** (0.024) (0.027) (0.027) Attainment x Norm 0.925*** 0.956*** 0.923*** (0.033) (0.035) (0.035) Implied Adaptation: Nonattainment Counties 0.651*** 0.639*** 0.617*** (0.055) (0.054) (0.050) Attainment Counties 0.296*** 0.308*** 0.318*** (0.031) (0.029) (0.032) Regulation Induced 0.355*** 0.332*** 0.299*** (0.061) (0.056) (0.053) All Controls Yes Yes Yes Observations 5,139,529 5,139,529 4,855,961 R2 0.435 0.433 0.426 Notes: This table reports the results of our central specification when shortening or increasing the length of the lag on counties’ nonattainment designation. Column (1) uses a shorter 1-year lag, column (2) uses our preferred 3-year lag corresponding with the minimum length of time the EPA affords nonattainment counties to re-enter compliance, and column (3) uses a 6-year lag corresponding to the next longest length of time the EPA allows nonattainment counties which may need more time to to re-enter compliance. Notably, even when counties are given a longer deadline by the EPA they are still expected to show progress towards re-entering attainment within the first three years. The number of observations vary slightly across the three specifications due to the selected length of the lag, although estimates to not meaningfully differ when using a restricted sample which includes only the intersection of observations included in each of the respective full samples. The full list of controls are the same as in the main model, depicted in column (2) of Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 235 Table B.2.2: Alternative Lengths of Climate Norm Daily Max Ozone Levels (ppb) 3-yr MA 5-yr MA 10-yr MA 20-yr MA (1) (2) (3) (4) Nonattainment x Shock 1.992*** 1.991*** 1.986*** 1.987*** (0.082) (0.081) (0.080) (0.079) Nonattainment x Norm 1.346*** 1.350*** 1.362*** 1.360*** (0.064) (0.065) (0.067) (0.067) Attainment x Shock 1.266*** 1.262*** 1.260*** 1.261*** (0.027) (0.027) (0.027) (0.027) Attainment x Norm 0.922*** 0.938*** 0.956*** 0.961*** (0.033) (0.033) (0.034) (0.035) Implied Adaptation: Nonattainment Counties 0.646*** 0.641*** 0.625*** 0.627*** (0.055) (0.056) (0.056) (0.055) Attainment Counties 0.344*** 0.323*** 0.304*** 0.300*** (0.028) (0.028) (0.028) (0.029) Regulation Induced 0.302*** 0.317*** 0.321*** 0.328*** (0.056) (0.056) (0.056) (0.056) All Controls Yes Yes Yes Yes Observations 5,139,529 5,139,529 5,139,529 5,139,529 R2 0.434 0.434 0.434 0.434 Notes: This table addresses the potential concerns with the measurement of the climate norm as a 30-year monthly moving average of temperature, lagged by 1 year. To explore whether measurement error is a cause of concern in our analysis, we estimate Equation (2.6) using alternative definitions for the climate norm. From column (1) through column (4), we report the estimates using a 3-, 5-, 10- and 20-year moving average of temperature as the climate norm. Recall that all moving averages are lagged by one year to allow for the potential adaptation responses by individuals and firms. As argued seminally by Solon (1992), as we increase the time window of a moving average, the permanent component of a variable that also includes a transitory component will be less mismeasured. Our estimates remain remarkably stable over the different lengths of the moving averages, but if anything, they get slightly larger until the 20-year moving average. There is a slight decline in the coefficient of the climate norm as we move from the 20-year to 30-year moving average (as reported in Table 1), which suggests that the widely used three-decade averages of meteorological variables including temperature are close to the long-run norms. The full list of controls are the same as in the main model, depicted in Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 236 Table B.2.3: Adaptation Responses Daily Max Ozone Levels (ppb) Long-Run Long-Run Short-Run 10-year Lag 20-year Lag 2004-2013 only (1) (2) (3) Nonattainment x Shock 1.987*** 1.987*** 1.406*** (0.078) (0.078) (0.047) Nonattainment x Norm 1.353*** 1.351*** 0.715*** (0.067) (0.067) (0.056) Shock x Action Day −0.147 (0.224) Attainment x Shock 1.265*** 1.267*** 0.995*** (0.028) (0.028) (0.020) Attainment x Norm 0.947*** 0.935*** 0.484*** (0.035) (0.034) (0.028) Shock x Action Day −0.056 (0.150) Implied Adaptation: Nonattainment Counties 0.634*** 0.636*** 0.691*** (0.052) (0.050) (0.044) Attainment Counties 0.318*** 0.333*** 0.511*** (0.029) (0.030) (0.029) Regulation Induced 0.316*** 0.303*** 0.179*** (0.054) (0.053) (0.041) Induced x Action Day −0.091 (0.256) Observations 5,131,949 5,127,892 1,879,044 R2 0.434 0.434 0.422 Notes: This table reports estimates when allowing more or less time for economic agents to engage in adaptive behavior. The estimates in columns (1) and (2) are obtained by Equation (2.6), but using 10- and 20-year lags between the moving average and contemporaneous temperature, rather than the usual 1-year lag. By doing so, agents are provided with more time to potentially adjust to climate change. Column (3) continues using the 1-year lag of the main specification, but adds an interaction term for “ozone action day” announcements at the county-level to estimate short-run adaptive behavior. These are days in which the relevant air quality authority expects to observe unhealthy levels of pollution. Individuals and firms are urged to take voluntary action to reduce precursor emissions. The estimate for the interaction between temperature shocks and action days is economically and statistically insignificant, pointing to limited opportunities for economic agents to adjust in the short run. Note that although action day policies first began in the 1990’s, EPA only provided data beginning in 2004, leading to a restricted overall sample (approximately 36% of our full sample). The full list of controls are the same as in the main model, depicted in Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 237 Table B.2.4: Further Robustness Checks Daily Max Ozone Levels (ppb) Daily Semi-Balanced Meteorological Moving Average Panel Controls (1) (2) (3) Nonattainment x Shock 1.997*** 2.177*** 2.056*** (0.080) (0.107) (0.082) Nonattainment x Norm 1.419*** 1.582*** 1.351*** (0.068) (0.085) (0.065) Attainment x Shock 1.265*** 1.562*** 1.228*** (0.028) (0.084) (0.083) Attainment x Norm 0.973*** 1.286*** 0.775*** (0.032) (0.102) (0.089) Average Wind Speed −2.204*** (0.284) Total Daily Sunlight 0.015 (0.015) Implied Adaptation: Nonattainment Counties 0.578*** 0.595*** 0.705*** (0.053) (0.088) (0.086) Attainment Counties 0.292*** 0.276*** 0.453*** (0.028) (0.076) (0.074) Regulation Induced 0.286*** 0.319*** 0.251** (0.054) (0.093) (0.108) Observations 5,139,460 520,670 453,859 R2 0.433 0.408 0.441 Notes: This table checks the sensitivity of our main results in Table 2.1 to changes in the primary econometric specification given by Equation (2.6) and sample restrictions. Column (1) replaces the monthly moving average with a daily moving average of temperature as the climate norm. Column (2) reports estimates from a semi-balanced panel of 92 ozone monitors that form around 11% of our complete sample. Column (3) provides estimates based on the reduced sample for which we have information on additional meteorological variables- average wind speed and total daily sunlight. High wind speeds prevent the build-up of ozone precursors and ultra-violet solar radiation triggers chemical reactions leading to the formation of groundlevel ozone. Having controlled for these additional parameters as well, which have statistically significant impacts on ozone, our primary estimates remain indistinguishable from our results in Table 2.1. The full list of controls are the same as in the main model, depicted in Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 238 Table B.2.5: Alternative Criteria for Selection of Weather Stations Daily Max Ozone Levels (ppb) (1) (2) (3) (4) Nonattainment x Shock 2.043*** 2.019*** 2.149*** 2.135*** (0.080) (0.080) (0.094) (0.091) Nonattainment x Norm 1.353*** 1.352*** 1.344*** 1.343*** (0.067) (0.067) (0.066) (0.065) Attainment x Shock 1.298*** 1.281*** 1.345*** 1.334*** (0.027) (0.027) (0.028) (0.028) Attainment x Norm 0.957*** 0.957*** 0.946*** 0.946*** (0.036) (0.035) (0.036) (0.035) Implied Adaptation: Nonattainment Counties 0.690*** 0.667*** 0.805*** 0.792*** (0.052) (0.053) (0.064) (0.063) Attainment Counties 0.341*** 0.325*** 0.399*** 0.388*** (0.030) (0.029) (0.029) (0.029) Regulation Induced 0.348*** 0.342*** 0.406*** 0.404*** (0.055) (0.056) (0.066) (0.064) Distance Cut-off 30 km 30 km 80 km 80 km Stations Included 2 2 5 5 Weighting Scheme Simple Avg 1/Dist Simple Avg 1/Dist Observations 5,139,529 5,139,529 5,284,426 5,284,426 R2 0.437 0.436 0.439 0.440 Notes: This table reports estimates from models using alternative criteria to match weather stations to ozone monitors. These estimates are from Equation (2.6), but we have varied the distance cut-off, the number of monitors in the matching as well as the averaging strategy to match the weather stations with the ozone monitors. Recall that in our main estimates in Table 2.1, we arrive at our sample by matching each ozone monitor to the closest two weather stations within a 30 km radius and we get the weather realization at each ozone monitor by averaging our weather variables over these closest two weather stations, weighted by their inverse squared distance from the monitor. In columns (1) and (2), we continue to use the closest two weather stations whereas in columns (3) and (4) we use the closest 5 weather stations within a 80 km radius of the ozone monitor. We also vary the weighting scheme: in columns (1) and (3), instead of a weighted average we just use a simple average across all matched weather stations; whereas in columns (2) and (4) we average the weather variables weighted by the inverse of the distance from the monitor. The full list of controls are the same as in the main model, depicted in Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 239 Table B.2.6: Alternative Clustering and Bootstrapped Standard Errors Daily Max Ozone Levels (ppb) Implied Adaptation (1) (2) Nonattainment x Shock 1.990*** (County Cluster) (0.079) (State Cluster) (0.126) (County x Week Cluster) (0.174) (Bootstrapped) (0.081) Nonattainment x Norm 1.351*** 0.639*** (County Cluster) (0.067) (0.054) (State Cluster) (0.103) (0.104) (County x Week Cluster) (0.147) (0.129) (Bootstrapped) (0.065) (0.055) Attainment x Shock 1.263*** (County Cluster) (0.027) (State Cluster) (0.060) (County x Week Cluster) (0.120) (Bootstrapped) (0.028) Attainment x Norm 0.956*** 0.308*** (County Cluster) (0.035) (0.029) (State Cluster) (0.076) (0.058) (County x Week Cluster) (0.118) (0.107) (Bootstrapped) (0.037) (0.029) Regulation Induced 0.332*** (County Cluster) (0.056) (State Cluster) (0.078) (County x Week Cluster) (0.073) (Bootstrapped) (0.056) Observations 5,139,529 R2 0.434 Notes: This table compares the standard errors of our main estimates with ones obtained by clustering at the state- rather than county-level, two-way clustering by county and week, and by bootstrap (block method clustered at the county level, 250 iterations). The latter addresses the potential concern that because our temperature shocks and norm are constructed, they could be seen as generated regressors. Bootstrapped standard errors are all within 6% of those estimated via clustering at the county level, and across all three alternative estimation methods recovered coefficients remain statistically significant at the 1% level. The full list of controls are the same as in the main model, depicted in Table 2.1. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 240 Table B.2.7: Results by Decade Panel A. 1980’s Max Ozone (ppb) Implied Adaptation Induced Adaptation (1) (2) (3) Nonattainment x Shock 2.496*** (0.165) Nonattainment x Norm 1.746*** 0.750*** (0.115) (0.119) Attainment x Shock 1.715*** (0.078) Attainment x Norm 1.356*** 0.359*** 0.391*** (0.064) (0.052) (0.106) Panel B. 1990’s Nonattainment x Shock 2.042*** (0.068) Nonattainment x Norm 1.470*** 0.571*** (0.057) (0.056) Attainment x Shock 1.360*** (0.034) Attainment x Norm 1.068*** 0.292*** 0.279*** (0.037) (0.039) (0.064) Panel C. 2000’s Nonattainment x Shock 1.506*** (0.042) Nonattainment x Norm 0.959*** 0.547*** (0.061) (0.061) Attainment x Shock 1.054*** (0.022) Attainment x Norm 0.729*** 0.324*** 0.223*** (0.034) (0.033) (0.054) Observations 5,139,529 R2 0.441 Notes: This table reports our main estimates disaggregated by the three “decades” in our sample: 1980- 1990; 1991-2001 and 2002-2013. Estimates in column (1) correspond to Equation (2.6), while estimates in column (2) report the implied measure of adaptation. The effects of the climate norm and temperature shock are decreasing over time in both attainment and nonattainment counties. Similarly, the measure of regulation-induced adaptation, column (3), appears to be somewhat decreasing across the three decades, although still statistically indistinguishable from our full sample results in Table 2.1. The full list of controls are the same as in the main model, depicted in Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 241 Table B8a: Nonlinear Effects of Temperature Panel A. Below 20◦C Max Ozone (ppb) Implied Adaptation Induced Adaptation (1) (2) (3) Nonattainment x Shock 0.795*** (0.023) Nonattainment x Norm 0.124*** 0.670*** (0.039) (0.036) Attainment x Shock 0.594*** (0.021) Attainment x Norm 0.192*** 0.403*** 0.268*** (0.036) (0.032) (0.047) Panel B. 20-25◦C Nonattainment x Shock 1.900*** (0.120) Nonattainment x Norm 1.438*** 0.462*** (0.114) (0.040) Attainment x Shock 1.361*** (0.042) Attainment x Norm 1.081*** 0.280*** 0.182*** (0.053) (0.031) (0.048) Observations 5,139,529 R2 0.447 Notes: This table explores the non-linear effects of the climate norm and temperature shock on ambient ozone concentrations. Specifically, we consider five bins of daily temperature: below 20◦C, 20-25◦C, 25-30◦C, 30-35◦C and above 35◦C. Estimates in column (1) correspond to Equation (2.6) after interacting indicator variables for each of these temperature bins, while estimates in column (2) report the implied measure of adaptation. Although regulation-induced adaptation on days between 20-25◦C and 25-30◦C appears to be lower than in our full-sample model, above 35◦C the magnitude of regulation-induced adaptation more than doubles, which is encouraging, given the prospects of ever increasing temperatures over the next decades. Recall that the 30-yr MA is lagged by 1 year, and the Clean Air Act attainment/nonattainment county designation is lagged by 3 years. The full list of controls are the same as in the main model, depicted in Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 242 Table B8b: Nonlinear Effects of Temperature Panel C. 25-30◦C Max Ozone (ppb) Implied Adaptation Induced Adaptation (1) (2) (3) Nonattainment x Shock 2.488*** (0.118) Nonattainment x Norm 2.241*** 0.246*** (0.131) (0.053) Attainment x Shock 1.407*** (0.049) Attainment x Norm 1.365*** 0.042 0.204*** (0.060) (0.033) (0.051) Panel D. 30-35◦C Nonattainment x Shock 2.509*** (0.132) Nonattainment x Norm 1.678*** 0.831*** (0.193) (0.104) Attainment x Shock 1.772*** (0.079) Attainment x Norm 1.394*** 0.379*** 0.452*** (0.099) (0.055) (0.092) Panel E. Above 35◦C Nonattainment x Shock 2.134*** (0.148) Nonattainment x Norm 0.809*** 1.325*** (0.206) (0.185) Attainment x Shock 1.642*** (0.114) Attainment x Norm 1.007*** 0.635*** 0.689*** (0.150) (0.153) (0.225) Observations 5,139,529 R2 0.447 Notes: This table continues the results from Table B8a for the temperature bins 25-30◦C, 30-35◦C and above 35◦C in panels (C), (D), and (E), respectively. The full list of controls are the same as in the main model, depicted in Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 243 Table B9: Adaptation by Local Beliefs in Climate Change Max Ozone (ppb) Implied Adaptation Induced Adaptation (1) (2) (3) Nonattainment x Shock 1.698*** (0.060) x Low Belief 0.020 (0.087) x High Belief 0.388*** (0.108) Nonattainment x Norm 1.171*** 0.527*** (0.085) (0.087) x Low Belief −0.040 0.060 (0.086) (0.094) x High Belief 0.152 0.236** (0.103) (0.107) Attainment x Shock 1.268*** (0.033) x Low Belief −0.093* (0.049) x High Belief 0.057 (0.069) Attainment x Norm 0.874*** 0.394*** 0.133* (0.043) (0.037) (0.074) x Low Belief 0.081 −0.173*** 0.234** (0.062) (0.051) (0.107) x High Belief 0.139* −0.082 0.318** (0.081) (0.071) (0.144) Observations 5,139,529 R2 0.435 Notes: This table reports estimates according to local beliefs on the existence of climate change. All counties in the sample were split into terciles based on the results of a survey conducted on climate change beliefs (Howe et al., 2015). Tercile indicators were interacted with the main variables in Equation (2.6). In column (1), the main impacts of the climate norm and temperature shock represent the effects in counties having beliefs in the middle tercile (for which the interactions have been omitted). The coefficients on the interaction terms reveal the differential effects of the climate norm and temperature shock in low- and high-belief terciles. Column (2) reports our implied measures of adaptation. By comparing the main estimates of the climate norm and shock in column (1), we obtain adaptation in mid-belief counties. Using the coefficients on the interaction terms, we obtain the differential adaptation in low- and high-belief counties in comparison to the mid-belief counties. Column (3) displays the measure of regulation-induced adaptation for the mid-belief tercile, followed by the differential RIA in low- and high-belief terciles. The full list of controls are the same as in the main model, depicted in column (2) of Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 244 Table B10: County Summary Statistics by Belief in Climate Change Panel A. Low Belief Counties Count Mean Std. Dev. Minimum Maximum Population (1000’s) 334 80.8 107.3 0.8 837.5 Average Education (Years) 334 12.7 0.6 11.0 14.3 Median Income ($1000/year) 334 48.5 10.4 21.9 83.3 Average Income ($1000/year) 334 61.5 11.3 36.9 111.9 Voted Democrat in 2008 (%) 334 37.2 10.4 6.6 64.8 Panel B. Median Belief Counties Population (1000’s) 335 162.7 213.3 1.9 1,870.4 Average Education (Years) 335 13.2 0.6 11.8 15.1 Median Income ($1000/year) 335 53.9 12.4 26.3 109.8 Average Income ($1000/year) 335 68.3 14.6 39.2 142.2 Voted Democrat in 2008 (%) 335 45.6 10.7 17.0 74.9 Panel C. High Belief Counties Population (1000’s) 336 478.5 803.3 1.3 9,758.3 Average Education (Years) 336 13.6 0.7 11.5 16.1 Median Income ($1000/year) 336 60.5 16.8 30.4 125.7 Average Income ($1000/year) 336 79.5 21.3 41.1 146.0 Voted Democrat in 2008 (%) 336 56.8 11.6 16.0 92.5 Notes: This table reports summary statistics of underlying demographics for each of the terciles of counties used in Table B9. Demographic data were obtained from the 2006-2010 5-year American Community Survey, with income reported in 2015 dollars, and average years of education based on a population weighted average of educational attainment status for the county population over 25 years of age. Voting data is obtained at the county level from the MIT Election Lab, and refers specifically to votes cast in the 2008 presidential election. 245 Table B11: Adaptation by Local ‘Preferences’ for Single Parenting Max Ozone (ppb) Implied Adaptation Induced Adaptation (1) (2) (3) Nonattainment x Shock 2.147*** (0.167) x Low Tercile −0.159 (0.170) x High Tercile −0.216 (0.164) Nonattainment x Norm 1.431*** 0.716*** (0.142) (0.100) x Low Tercile 0.039 −0.198 (0.115) (0.127) x High Tercile −0.068 −0.147 (0.117) (0.123) Attainment x Shock 1.311*** (0.049) x Low Tercile −0.123** (0.062) x High Tercile −0.021 (0.068) Attainment x Norm 1.009*** 0.302*** 0.414*** (0.068) (0.044) (0.102) x Low Tercile −0.096 −0.027 −0.170 (0.077) (0.062) (0.151) x High Tercile −0.082 0.061 −0.209 (0.089) (0.069) (0.158) Observations 5,139,529 R2 0.435 Notes: This table reports differential estimates according to local beliefs unrelated to environmental amenities – the ‘preference’ for single parenting. All counties in the sample were split into terciles based on the fraction of single-parent households from the Opportunity Atlas (Chetty et al., 2018), and those terciles were then interacted with the main variables in Equation (2.6). In column (1), the main impacts of the climate norm and temperature shock represent the effects in counties classified in the middle tercile (for which the interactions have been omitted). The coefficients on the interaction terms reveal the differential effects of the climate norm and temperature shock in low- and high-fraction terciles. Column (2) reports our implied measures of adaptation. By comparing the main estimates of the climate norm and shock in column (1), we obtain adaptation in mid-fraction counties. Using the coefficients on the interaction terms, we obtain the differential adaptation in low- and high-fraction counties in comparison to the mid-fraction counties. Column (3) displays the measure of regulation-induced adaptation for the mid-fraction tercile, followed by the differential RIA in low- and high-fraction terciles. The full list of controls are the same as in the main model, depicted in Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 246 Table B12: Adaptation by VOC- or NOx-limited Atmosphere Main Specification VOC/NOx-Limited Ozone(ppb) Adaptation Ozone(ppb) Adaptation (1) (2) (3) (4) Nonattainment x Shock 2.097*** 2.139*** (0.136) (0.176) x VOC-limited 0.439* (0.225) x NOx-limited −0.134 (0.273) Nonattainment x Norm 1.398*** 0.699*** 1.406*** 0.733*** (0.149) (0.107) (0.159) (0.118) x VOC-limited 0.126 0.313* (0.142) (0.176) x NOx-limited −0.235 0.101 (0.239) (0.328) Attainment x Shock 1.707*** 1.872*** (0.182) (0.245) x VOC-limited −0.513* (0.262) x NOx-limited −0.421 (0.342) Attainment x Norm 1.326*** 0.381*** 1.385*** 0.487*** (0.133) (0.112) (0.159) (0.135) x VOC-limited −0.106 −0.407** (0.125) (0.182) x NOx-limited −0.288** −0.133 (0.125) (0.307) Regulation Induced 0.318*** 0.246** (0.104) (0.117) x VOC-limited 0.720** (0.346) x NOx-limited 0.233 (0.592) Observations 1,007,563 1,007,563 R2 0.459 0.460 Notes: This table reports estimates when interacted with an indicator of whether a county was VOC- or NOx-limited. Columns (1) and (2) depict the results of our main specification under the restricted sample for which precursor pollutant data is available. In columns (3) and (4) results reflect the effect for non-limited counties, while the interaction terms depict the differential effect in precursor limited counties. The full list of controls are the same as in the main model, depicted in Table 2.1. Standard errors are clustered at the county level. ***, **, and * represent significance at 1%, 5% and 10%, respectively. 247 B.3. Formal Derivations of Analytical Results This appendix provides further elaboration of the conceptual framework and formalized model of regulation-induced adaptation as discussed in Section II. B.3.a. Model Derivations Here we provide a formal derivation of regulation-induced adaptation (RIA) and its associated welfare effects, as discussed in Section II, for a corrective smoothing regulation on an externality with climate interactions. For simplicity we start by constructing a straightforward model in which we focus on incremental changes to the regulation, assuming that it is binding at any level. Afterwards we relax this assumption to compare two scenarios – one in which the regulation is binding, and one in which it is not – to derive the channel and welfare effects of RIA. In Appendix C.2 we extend the model to account for additional complexities such as: (i) adaptation and welfare effects when the smoothing regulation is distortionary rather than corrective, and (ii) input regulations rather than output regulations on the externality of interest. All variables are as defined in the text unless otherwise stated. Suppose the representative agent’s utility is: U = u(X, Y ) − ϕ(E), (B.3.1) and they face the budget constraint: pXX + Y = L + G, (B.3.2) where the price of X depends directly on the stringency of the corrective smoothing regulation on emissions concentration, tE, and indirectly on climate, c, through its effect on E in the presence of tE > 0.90 The agent maximizes (B.3.1) subject to (B.3.2), taking E as given. 90More specifically, the economy-wide emissions concentration, E, is a function of the per-unit emissions, e, multiplied by total production of X, conditional on climate, c, that is: E ≡ E(c) = e(c)X. 248 The solution to which is the (uncompensated) demand functions: X(tE, e(c), G), Y (tE, e(c), G). (B.3.3) Derived “demand” for economy-wide emissions can also be expressed as a function of the exogenous variables: E(tE, e(c), G), (B.3.4) and thus, the agent’s indirect utility is: V = v(tE, e(c), G) − ϕ(E(tE, e(c), G)). (B.3.5) From Roy’s identity: ∂v ∂tE = −λE, ∂v ∂c = 0, ∂v ∂G = λ, (B.3.6) where λ is the marginal utility of income. Differentiating (B.3.5) with respect to tE, holding c constant, and using (B.3.6) gives: 1 λ dV dtE = −E + dG dtE − ϕ ′ λ dE dtE , (B.3.7) where: dE dtE = ∂E ∂tE + ∂E ∂G dG dtE . (B.3.8) The government budget constraint is: G = tEE, (B.3.9) and differentiating (B.3.9) with respect to tE, holding c constant, gives: dG dtE = E + tE dE dtE . (B.3.10) From (B.3.7) and (B.3.10), 1 λ dV dtE = dE dtE tE − ϕ ′ λ , (B.3.11) 249 and the optimal corrective smoothing regulation in the absence of any climate interaction is clearly the Pigouvian tax (or its equivalent), t ⋆ E = ϕ ′ λ . Now, if climate is also changing, differentiating (B.3.11) a second time, with respect to c, simply gives: 1 λ d 2V dtEdc = d 2E dtEdc tE − ϕ ′ λ . (B.3.12) The optimal corrective smoothing regulation is in fact unchanged as it “internalizes” the climate interaction already. But notice that the implied tax on X will change. This implied tax, t ∼ X, is equal to tEe(c), such that: dt∼ X dc = tEe ′ . (B.3.13) Thus, in situations where it is infeasible to directly regulate E, a proxy regulation on X would need to be more (less) stringent in the presence of an increasing climate if e ′ is larger (less) than one. Using the above model, we now provide a formal derivation of regulation-induced adaptation and its associated welfare effects. Consider first a scenario in which climate is changing and the pre-existing smoothing regulation is held constant, but the agent is unconstrained by the regulation.91 Differentiating (B.3.5) with respect to c, holding t¯E = 0 constant, gives: 1 λ dV dc N = ϕ ′ λ dE dc N , (B.3.14) where: dE dc N = e ′X + e dX dc , (B.3.15) and, using (B.3.3) and (B.3.13): dX dc = ∂X ∂t∼ X dt∼ X dc + ∂X ∂G dG dc = ∂X ∂t∼ X tEe ′ + ∂X ∂G tEe ′X + tEe dX dc . (B.3.16) 91In our context, for example, this would apply to counties that are designated as in attainment with the NAAQS but could apply to any policy scenario in which the existing policy is either non-binding or non-existent. 250 As t¯E = 0 in this scenario, dX dc likewise evaluates to zero, and (B.3.16) simplifies to: dE dc N = e ′X. (B.3.17) Now, consider a second scenario in which climate is changing and the pre-existing smoothing policy is still fixed, but now the agent is constrained by the regulation.92 Differentiating (B.3.5) with respect to c, holding t¯E > 0 constant, gives: 1 λ dV dc R = ϕ ′ λ dE dc R , (B.3.18) where: dE dc R = e ′X + e dX dc , (B.3.19) though notably now dX dc does not evaluate to zero, as t¯E > 0. Taking the difference of (B.3.14) from (B.3.18) reflects the welfare benefits of regulation-induced adaptation (RIA) from the discrete implementation of a corrective smoothing regulation: 1 λ dV dc R − dV dc N = − ϕ ′ λ | {z } Marginal Damages dE dc R − dE dc N | {z } RIA . (B.3.20) Substituting (B.3.17) and (B.3.19) into the second term on the right-hand side of (B.3.20) shows that the channel of regulation-induced adaptation operates through induced adjustments in the production of X in response to a change in climate: RIA = e dX dc (B.3.21) And thus, from (B.3.20), the welfare effects of regulation-induced adaptation are clearly equal to the monetized marginal damages of emissions, multiplied by the avoided increase in concentration that would have occurred due to climate change in the absence of the regulation. 92In our context, for example, this would apply to counties that have been designated as in nonattainment with the NAAQS but could apply to any policy scenario in which the existing policy is binding in some form. 251 B.3.b. Model Extensions When Smoothing Policies Inhibit Adaptation. In Appendix B.3.a we show that corrective smoothing regulations with climate interactions can induce adaptation. On the other hand, if the smoothing regulation or policy is not a corrective measure imposed on an externality, but a distortionary policy intended to attenuate a production shock, and that shock is a function of climate, the policy would in fact inhibit adaptation. Consider a similar analytical model, but with agent’s utility given by: U = u(X, Y ), (B.3.22) and the agent faces the budget constraint: pXX + Y = L, (B.3.23) where now X and Y are both consumption goods without external outputs. However, the production of X is subject to some shock, s, that increases the per-unit cost of producing X by some monetary value given by δ ≡ δ(c) which is an increasing function of climate. The effect of this production shock can be attenuated by allocating labor towards a, a behavior or technology that would attenuate the impact of the shock, at a per-unit cost of z(a), such that the profit per unit of X is given by: pX − {1 + z(a) + δ(c)(s − a)}. (B.3.24) The agent thus chooses a to maximize profit, which gives the first-order condition: z ′ (a) = δ(c), (B.3.25) which states that attenuation activity occurs until the marginal cost per unit of X equal the cost of the production shock. Further differentiating (B.3.25) with respect to climate, c gives: z ′ dc = δ ′ . (B.3.26) 252 In other words, the optimal level of attenuation activity – as a function of its marginal cost – is increasing with climate at a rate commensurate with the increase in the marginal cost of the shock due to increasing climate. That is, there are private incentives to adapt to climate change. For example, looking at the context of agriculture as with, e.g., Annan and Schlenker (2015), if climate change increases the severity of heatwaves and thus associated crop-loss, the marginal cost of this production shock has increased, but farmers may invest in additional irrigation or heat-resistant crops as ways of adapting to this climate induced change in the damage function of the shock. And, from (B.3.26) it’s clear that in the absence of a market failure, private incentives would induce optimal adaptation. Now, however, suppose that the government implements a smoothing policy to ameliorate the impacts of the production shock by imposing a labor tax, tL, which it uses to in turn subsidize the production costs (or market price) of X. Returning to the agriculture example, this could resemble the Federal Crop Insurance Program, for example, which heavily subsidizes the premiums farmers can pay in order to insure their crops against loss from production shocks such as heat-waves. The profit per unit of X would now be: pX − {1 + z(a) + δ(c)(s − a) − γδ(c)(s − a)}, (B.3.27) with associated government budget constraint: G = tLL − γδ(c)(s − a)X, (B.3.28) where γ is a scalar between zero and one, reflecting the proportional amount of the shock’s production costs that are subsidized by the smoothing policy. Continuing the crop insurance example, if the premiums faced by farmers are subsidized by 60 percent, γ would be taken as equal to 0.6, as in the absence of the subsidy the full cost of the premium would simply reflect the expected value of crop-loss. From the first-order condition of (B.3.27), we now have: z ′ (a) = (1 − γ)δ(c), (B.3.29) 253 and differentiating (B.3.29) with respect to climate, c, gives: z ′ dc = (1 − γ)δ ′ . (B.3.30) In other words, the smoothing policy inhibits adaptive behavior by decreasing the profit incentive that would otherwise have existed to increase shock attenuating activities in response to a changing climate. Or, conversely, γδ′ represents the contraction in adaptation due to the pre-existing distortionary policy, or – regulation-inhibited adaptation. Input Regulations vs. Output Regulations. Extending the original model to examine indirect “input” regulations on emissions – for example, on NOx or VOC precursor emissions in our empirical setting – rather than direct “output” regulations on the outcome of interest – i.e., Ozone – is rather straightforward. First, redefine the economy-wide emissions function as proportional to the level of its inputs: E = e(c)I, (B.3.31) where inputs, I, are in turn a function of production, but are independent of climate: I = iX. (B.3.32) A regulation on inputs, tI , could thus in theory approximate a direct regulation on the emissions, tE, but since I(.) is not a function of climate, tI would not induce any climate adaptation. Returning to (B.3.19), and substituting (B.3.31) and (B.3.32) for E gives: dE dc R = e ′ iX + e dX dc , (B.3.33) and, using (B.3.31) and (B.3.32) to also redefine t ∼ X = tI i, re-deriving (B.3.16) gives: dX dc = tI ∂X ∂dt∼ X di dc + tIX ∂X ∂G di dc / 1 − tI i ∂X ∂G . (B.3.34) since di/dc = 0, the numerator of (B.3.34) evaluates to zero, and we have that dX/dc = 0. In fact, this is rather intuitive – since the regulation on inputs does not internalize the climate 254 interaction, it does not induce any change in behavior in response to a change in climate. 255

Abstract (if available)

Abstract Climate change is unequivocal, with global temperatures expected to rise from 1.5 to 4.5C over the 21st century. Therefore, it is crucial to develop methods to measure climate impacts and adaptation. This dissertation first develops a unifying approach to measure both climate impacts and adaptation in the same estimating model, applying the method to examine the “climate penalty” on ambient ozone air pollution. The second chapter then examines how policies interact with climate change and pre-existing market-failures to induce or inhibit adaptation. Specifically, this chapter develops a tractable analytical framework of a corrective regulation where the market failure interacts with climate, highlighting the mechanism through which the regulation may also incidentally induce climate adaptation. Combined with the unifying approach from chapter one, the analysis finds that the Clean Air Act incidentally induces statistically and economically significant climate adaptation co-benefits with respect to controlling ozone pollution. The final chapter extends the prior methods to allow for the examination of whether policy-induced adaptation complements or crowds out adaptation that economic agents would otherwise intrinsically engage in. Specifically, this chapter examines the negative health impacts of ozone exposure, and whether economic agents intrinsically adapt to expected changes in ambient ozone – often referred to as “defensive investments”. Taking advantage of the staggered implementation of air quality alert programs across the US between 2004 and 2017, the analysis finds no evidence that air quality alert programs crowd-out intrinsic defensive investments, with suggestive – though statistically insignificant – evidence that they in fact increase intrinsic defensive-investments.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Impacts of heat mitigation strategies and pollutant transport on climate and air quality from urban to global scales

PDF

Essays on narrative economics, Climate macrofinance and migration

PDF

Prenatal environmental exposures and fetal growth in the MADRES cohort

Asset Metadata

Creator Miller, Noah (author)

Core Title Climate change, air pollution and health: how policy can induce adaptation

School School of Policy, Planning and Development

Degree Doctor of Philosophy

Degree Program Public Policy and Management

Degree Conferral Date 2024-08

Publication Date 08/13/2024

Defense Date 02/26/2024

Publisher Los Angeles, California (original), University of Southern California (original), University of Southern California. Libraries (digital)

Tag “climate penalty” on ozone,adaptation,ambient ozone concentration,Clean Air Act,climate change estimation methods,climate impacts,government regulations and policy,health behavior,local air pollution,public health,regulation-induced adaptation,valuation of environmental effects

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Bento, Antonio (committee chair), Rose, Adam (committee member), Severnini, Edson (committee member)

Creator Email noah.miller.012@gmail.com,nsmiller@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113998TFM

Unique identifier UC113998TFM

Identifier etd-MillerNoah-13382.pdf (filename)

Legacy Identifier etd-MillerNoah-13382

Document Type Dissertation

Format theses (aat)

Rights Miller, Noah

Internet Media Type application/pdf

Type texts

Source 20240813-usctheses-batch-1197 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu