University of Southern California Dissertations and Theses

Essays on environmental economics: education, employment, and experiments

(USC Thesis Other)

Essays on environmental economics: education, employment, and experiments

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ESSAYS ON ENVIRONMENTAL ECONOMICS: EDUCATION, EMPLOYMENT, AND EXPERIMENTS by Taraq Khan A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (ECONOMICS) May 2023 Copyright 2023 Taraq Khan Dedicated to my beloved father, Md Shahanur Ahmed Khan Who instilled in me a love of learning in my youth And supported me in fulfilling this whimsical, dunya dream. ii Contents Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix Chapter 1: Heat and Long Term Earnings: Evidence from Schooling in Brazil . . 1 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 The Brazilian Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 A Regions and Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 B Institutional Details:RAIS/Job Market . . . . . . . . . . . . . . . . . . 7 C Institutional Details: SAEB/Quality of Education Indicators . . . . . 9 1.3 Data and Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 A Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 B Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 C Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.4 Heat and Long Term Impact on Earnings . . . . . . . . . . . . . . . . . . . . 16 A Wages by Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 B Mechanisms: Summer and Weekend Temperature Effects . . . . . . . 20 C Comparison to Middle and High-School Temperature Shocks . . . . 22 D Decomposing the Effect by Year of Elementary School . . . . . . . . . 23 E Further Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 E.1 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . 25 E.2 Non-Linearities . . . . . . . . . . . . . . . . . . . . . . . . . 26 E.3 The Effect of Extra Education . . . . . . . . . . . . . . . . . . 27 1.5 Heterogeneity in the South . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 A Wages by Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 B Wages by Education Level . . . . . . . . . . . . . . . . . . . . . . . . . 31 B.1 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . 31 1.6 Heat and Schooling Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 A The Region Divide, Revisited . . . . . . . . . . . . . . . . . . . . . . . 32 B Test Scores by Socioeconomic Level . . . . . . . . . . . . . . . . . . . 32 iii 1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 Chapter 2: Last Hired, First Fired: Droughts and Labor Market Inequality in Brazil - with Antonio Bento and Edson Severnini . . . . . . . . . . . . . . . . . . 35 2.1 Introduction and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.2 Research Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 A Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 B Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.3 Preliminary Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 Chapter 3: Religious Figures and their Impact on Non-Religious Decisions, for the Religious and ”A”-religious: An Experiment on Environmental Donations 44 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 A Study Motivation & Contributions . . . . . . . . . . . . . . . . . . . . 47 3.2 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 A Sample Population and Randomisation . . . . . . . . . . . . . . . . . 49 B Treatment Arms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 B.1 Control (Generic Leader) . . . . . . . . . . . . . . . . . . . . 50 B.2 Pope Francis . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 B.3 Shia Arm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 B.4 Pope Benedict . . . . . . . . . . . . . . . . . . . . . . . . . . 51 C Balance and Stratification . . . . . . . . . . . . . . . . . . . . . . . . . 51 D Regression Specifications . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.3 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 A Full Sample Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 B Heterogeneous Treatment Effects by Religiosity . . . . . . . . . . . . 56 C Alternative Heterogeneity Regressions . . . . . . . . . . . . . . . . . 61 3.4 Robustness Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 A Indicator for Donations Above the Mean . . . . . . . . . . . . . . . . 65 B Survey Durations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 A Heat and Long Term Earnings: Evidence from Schooling in Brazil . . . . . . 74 A.1 Figures and Summary Statistics . . . . . . . . . . . . . . . . . . . . . . 74 A.2 Regressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 iv A.3 Regressions - Main Specification Robustness . . . . . . . . . . . . . . 83 A.4 Regressions - South and Southeast Robustness . . . . . . . . . . . . . 91 A.5 Regressions - SAEB Scores . . . . . . . . . . . . . . . . . . . . . . . . . 94 B Religious Figures and their Impact on Non-Religious Decisions, for the Re- ligious and ”A”-religious: An Experiment on Environmental Donations . . . 97 B.1 Robustness to Survey Durations . . . . . . . . . . . . . . . . . . . . . 97 B.2 Robustness to Donation Indicator . . . . . . . . . . . . . . . . . . . . 101 v List of Tables 1.1 Summary Statistics, Brazil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.2 Elementary Schooling Days Temperature on Wage by Region . . . . . . . . . 17 1.3 Elementary Schooling Heat Coefficients scaled, for South . . . . . . . . . . . 19 1.4 Elementary Days on Wage: School Days vs Holidays and Weekends . . . . . 21 1.5 School Days Temperature on Wage by School Exposure Period - Completed Only . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 1.6 Elementary Schooling Days Temperature on Wage by Region, School Year Split . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 1.7 Elementary Schooling Days Temperature on Wage by Region and Gender . 30 2.1 Droughts by Individual Sectors . . . . . . . . . . . . . . . . . . . . . . . . . . 41 2.2 Droughts by Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 2.3 Droughts by Sector - High/Low . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3.1 Balance Table on Covariates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 3.2 Number of Participants by Strata and Religiosity . . . . . . . . . . . . . . . . 53 3.3 Average donations conditional on winning the$100 prize draw . . . . . . . 53 3.4 Treatment Effect: Donation of Potential $100 Lottery . . . . . . . . . . . . . 57 3.5 Heterogeneous Treatment Effects: Donation of Potential $100 Lottery . . . . 60 3.6 Heterogeneous Treatment Effects: Donation of Potential $100 Lottery by Politics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 3.7 Heterogeneous Treatment Effects: Donation of Potential $100 Lottery by Religion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 3.8 Heterogeneous Treatment Effects: Donations Above the Mean of 21 . . . . . 65 A.1 Means by Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 A.2 Correlation Across Birth, Elementary, and Work Year Temperatures . . . . . 75 A.3 School Days Temperature on Wage by School Exposure Period . . . . . . . . 77 A.4 Elementary Schooling Days Temperature on Wage by Region - Non-Linear . 78 A.5 Elementary Schooling Days Temperature on Wage by School Completion . . 79 A.6 Elementary Schooling Days Temperature on Wage by Region and High School Completion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 A.7 High School Completion by Region . . . . . . . . . . . . . . . . . . . . . . . . 81 vi A.8 Elementary Schooling Days Temperature on Wage - Robustness to Controls, Full Sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84 A.9 Elementary Schooling Days Temperature on Wage with Other School Tem- perature Bin Controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 A.10 Elementary Schooling Days Temperature on Wage by Region - 3C Bins . . . 86 A.11 Elementary Schooling Days Temperature on Wage by Region - 1C Bins . . . 87 A.12 Elementary Schooling Days Temperature on Wage - Robustness to Age . . . 88 A.13 Elementary Schooling Days Temperature on Wage - Robustness to Controls in the South . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 A.14 Elementary Schooling Days Temperature on Wage - Robustness to Age in the South . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 A.15 Elementary Schooling Days Temperature on Mathematics Score by Region . 94 A.16 Elementary Schooling Days Temperature on Mathematics Score by Region and SEC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 vii List of Figures 1.1 Climate of Brazil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.2 Temperature across Brazil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1.3 Temperature Bins by Region . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.4 Increasing Temepratures in Brazil . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.5 Main Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3.1 Coefficients and 95% confidence intervals as the upper and lower bounds vary for Benedict treatment arm, very and areligious. . . . . . . . . . . . . . 66 A.1 Regions of Brazil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 A.2 Population Density, Brazil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 A.3 High School Completion by Region . . . . . . . . . . . . . . . . . . . . . . . . 82 A.4 Elementary Svhooling Days Temperature on Wage by Region - 1C and 3C Bins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 A.5 Brazil Earnings Split by Age . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 A.6 South-Southeast by Age . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 A.7 SAEB Mathematics Test Scores by SEC . . . . . . . . . . . . . . . . . . . . . . 96 B.1 Coefficients and 95% confidence intervals for the very religious: Upper and Lower Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 B.2 Coefficients and 95% confidence intervals for the somewhat and non-religious: Upper and Lower Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 B.3 Coefficients and 95% confidence intervals for agnostic and atheists: Upper and Lower Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 B.4 Coefficients and 95% confidence intervals for dependent indicator variable cutoff: for the very religious, somewhat, and non-religious. . . . . . . . . . . 102 B.5 Coefficients and 95% confidence intervals for dependent indicator variables cutoff: for Agnostics and Atheists . . . . . . . . . . . . . . . . . . . . . . . . . 103 viii Abstract This dissertation consists of three chapters, with different approaches and insights into topics related to Environmental Economics. The first chapter provides causal evidence of a negative impact of heat exposure dur- ing school age on long term earnings in Brazil. Using natural variation in temperature exposure over the first five years of elementary education, I find that a one standard devi- ation increase in the number of elementary schooling days between 30-32C reduces long term earnings by 0.94% in the South and Southeast. Although ex-ante these impacts could be operating through a number of mechanisms, the heterogeneity in the effects of shocks across school and non-school days supports a learning channel: there are no long-term earnings effects of heat exposure during summer and weekends. No effects are found for middle or high-school exposure, but heat shocks remain significant in all but the final year of elementary education. Consistent with some degree of adaptation, effects are concen- trated in climates unaccustomed to continuous heat, and are most prominent amongst the least educated. A secondary set of data and results examining schooling outcomes further complements the evidence that the effect occurs through disrupted education and falls disproportionately on the most disadvantaged socioeconomic class. These results suggest the learning inhibited by heat is sizeable enough to be captured in later life earnings, and that public policy could potentially mitigate these impacts by supporting human capital accumulation during childhood. The second chapter, joint work with Antonio Bento and Edson Severnini, proposes a paper in which we ask and attempt to answer the question of whether labor market dispar- ities seen during economic downturns are also evidenced from the result of environment- related declines in activity. Using a fixed-effects approach with variation in droughts both temporally and spatially we intend to estimate the impact on the hiring and firing of work- ers by race and gender. This chapter documents promising preliminary results. The third chapter studies the heterogeneous effect of religious figures on preferences for the environment proxied by donations from a$100 prize draw. Quotes from religious leaders are provided and juxtaposed either with a religious leader or a generic ”world leader”. In the preferred specification’s LATE, I find a backlash effect among those that ix identify as Atheists or Agnostics who reduce their donations by $8.53-$9.62 across treat- ment arms compared to the control group. There is no statistically significant effect for the Non-Religious and Somewhat Religious. The Very Religious select to increase dona- tions by $9.83-$11.28 when less controversial religious figures are involved. Robustness checks to specification and subsampling are discussed, as is suggestive evidence that it is indeed the heterogeneity in religiosity rather than a correlated variable driving the results. I finish with a discussion on why I have found such an effect that does not appear very prominently in related literature. x Chapter 1 Heat and Long Term Earnings: Evidence from Schooling in Brazil 1.1 Introduction Extreme heat continues to become more and more frequent across the globe, with as of yet unforetold impacts. In recent years areas of the world reached over 40C, surpassing temperatures not expected in those regions for decades ([1]). Research has also pointed to an increase in temperature variation within a locality ([2, 3]) with both the rising av- erage and variation expected to have detrimental economic impacts ([4–6]). Brazil is no exception with the highest temperature recorded at 44.8C in 2020. The literature details evidence of effects during childhood and adolescences, such as parental loss or unemployment, and its impact on later life relationships and socioeco- nomic status ([7, 8]). I show that climate shocks such as excess heat can also be associated with negative effects into the distant future. In many long-run studies pinning down the mechanism is difficult, however I provide evidence consistent with some channels while also ruling out others. This paper focuses on the inefficient accumulation of human capital during the forma- tive years of early elementary schooling in Brazil. I find the effects of heat most damaging in areas unaccustomed to high temperatures. My main contributions are to show such damage exists, to provide evidence of a learning channel that is only measurable during 1 early schooling years, to decompose the schooling effect, and to examine the heterogeneity therein. Heterogeneity in the impacts of climate shocks is not unique and has been observed in many other papers, including for floods and wildfire smoke exposure ([9, 10]). There have been many studies detailing the detrimental impacts of excessive heat on school out- comes (e.g. [11, 12]), with parallel effects found due to other climate anomalies such as air pollution ([13, 14]). Excessive heat makes learning uncomfortable, and much research has studied the im- pact on exam results, some more conclusively ([15, 16]), than others ([17]). A lapse of concentration is also found in certain non-school situations within controlled environ- ments ([18–20]). My paper speaks to the debate on whether the causal connection be- tween heat and schooling benefits operates primarily through learning, or through per- formance during tests. I show results that are consistent with the former explanation. The prolonged lack of concentration over several years may eventually lead to a reduction in earnings once students enter the job market and demonstrate lower levels of human capital. To the best of my knowledge, this paper is the first the connection between early education and long-term impacts on earnings, and to provide evidence for the mechanism of these impacts. Sub optimal weather may induce poorer health and this too can have long term impacts ([21]). An individual who is unwell and suffers poorer long term health outcomes may find themselves able to work less efficiently. They may also miss school when rainfall is high, which increases the opportunity cost of work 1 , thereby reducing the scope of their human capital accumulation as they shift their time away from school ([22, 23]). In India, there is evidence of effects mediated through agriculture ([24]), and high stakes has also been show to mitigate some effects ([25]) but not in other situations ([26]). I find effects 1 Agriculture is more productive when it rains compared to drought periods. 2 for schooling days but not weekends and the holidays, which suggests that it is through an education related channel that my results propagate 2 . Following in the footsteps of papers looking at the long term effects of exposure to heat around the time of birth ([27–29]), I modify the empirical strategy for a schooling frame- work, estimating heat exposure during the entire schooling year as opposed to just during the exam period. Using the natural variation in temperature exposure across students in different cohorts to provide a source of exogeneity in temperature exposure over the first five years of elementary education, I estimate the impact on long term earnings at the age of 30, subdividing by region, gender, and education level. Municipality fixed-effects con- trol for municipality level time-invariant characteristics. There are over 11 million workers in the final sample spread across over 5,700 municipalities. I also present the effects of heat exposure during middle and highschool. I show suggestive evidence that the early years are where learning is affected the most, that it is males and the least educated that are worst affected, and support these results with a secondary set of analysis on schooling outcomes. To supplement these findings I look into more recent educational data, further supporting the hypothesis that it is indeed the education channel I am identifying. My research suggests that the detrimental effects are strongest for the youngest stu- dents in elementary. The effects are significant and economically meaningful when scaled by standard deviation changes. I do not detect any effects from middle or highschool expo- sure to heat. No effects are found in the final year of elementary school, which determines progress into middle school. The grade one achieves during elementary does not tend to be presented as a signal in the job market, which further suggests that it is the learning route I am identifying. There is no evidence of a ”building blocks” mechanism at play, in which the first year of elementary school is most pivotal. Indeed, other than the final year, heat during each year of elementary remains damaging to students’ long term earnings. 2 A health channel would imply that I find effects thought the year and not just during school. However once schooling days, weekends and holidays are all included, only schooling days remains significant at standard significant levels. 3 I also find a larger impact on the male population relative to females and in the cooler south. This is consistent with the findings in Italy ([30]), except here I also find effects for lower economic status students. Several explanations are provided, from actual exposure due to location ([31–33]), exposure through higher manual labor, and metabolic differ- ences, to the use of AC and conventional fans. I do however note that even by 2020, there was still a low prevalence of air conditioning across Brazil 3 Perhaps the most policy relevant finding I have is that heat is most damaging in the relatively cooler south of Brazil. The reasons for this are discussed in my paper, however the conclusion remains the same; policymakers should think twice before committing re- sources solely to hotter areas when investing in heat mitigation such as air conditioning. I contribute to multiple strands in the literature. A growing series of papers have used the Brazilian RAIS dataset to document earnings and inequality (e.g. [34]), but none insofar have attempted, to the best of my knowledge, to link it to longer term effects of educational attainment. Thus the key novelty of my paper is that I directly measure the effect of heat during the formative years of education on long term earnings, in a similar vein to the methodology used for in-utero temperature effects ([28]). Another key differentiator is my evidence on the early formative schooling years that are likely to have a lasting impact through the learning channel, rather than distorting human capital signals that are derived from test scores later on in the schooling timeline. Indeed, the lack of effects found during middle and highschool suggests either that tra- jectories are determined much sooner in the learning cycle, or that the early schooling years are more sensitive to the effects of excess heat. The insignificant effects in the fifth and final year of elementary education also suggests a learning channel. 4 Prior research has shown that learning is important during higher education, where wages fall due to a reduction in workload ([35]) for students who signal through attending college. Clark 3 This is based on Brazilian Census data, as schools did not report AC details during the period individuals were in school. 4 During the fifth year it is determined whether students advance to middle school or retake the year. 4 and Martorell also argue for the learning channel, finding little evidence of a highschool diploma’s signaling effect ([36]). I do not find significant effects for summer or weekend temperatures after schooling days are included, casting doubt on other mechanisms that should continue to function during this period such as health effects or economic shocks. My examination by days in each temperature bin scaled by standard deviations also allows me to show effects in a more disaggregated way: extremely hot days are more damaging, but also much less common. As such, it is perfectly possible that the damages from extra hot days at 30C reduces wages by more than the less common number of extra hot days at 34C. If extreme heat is damaging but rarely occurs, then one may find one’s attention focused on mitigating the effects of relatively lower levels of ”high” heat a better policy decision. I also examine non-linearities within bins, finding that the more hot days an individual experiences, the smaller the marginal effect on wages lost becomes. Another advantage I possess is in my use of climatological data at the 9km grid level, much more precise than previous datasets at the hourly level at 32km. This is especially important for accuracy of the data in coastal areas ([37]), and with a large number of cities in Brazil facing the South Atlantic Ocean this becomes of critical importance. Without this, previous researchers attempting this may have found their results to be non-conclusive and effects insignificant. My research also ties in with the Environmental Justice literature. I find heterogene- ity in impacts that disfavour males, the least educated, and the least well off, as well as a regional bias. The effects appear to diminish as the level of educational attainment in- creases, but whether this is through selection where those who make it to highschool are less perturbed by heat while they learn, or affected students catch up through extra ed- ucation, is not clear. Nonetheless, I present an abundance of evidence that the burden of heat falls disproportionately on some more than others, and can affect young students long into their working lives. 5 1.2 The Brazilian Context A Regions and Climate The size of Brazil lends itself to multiple climates across the country (Figure 1.1). The north of the country tends to be hotter on average but with low seasonal variation, while the south is cooler with greater monthly fluctuations 5 A large proportion of the population of Brazil is also located in areas cooled by either their altitude or the sea breeze. There are five Macroregions in Brazil, and they lend themselves to the climate differ- ences discussed. The South and Southeast regions have a much higher population with both Rio de Janeiro and S˜ ao Paulo situated therein. There are significant large cities such as Brasilia in the Central region and Salvador in the Northeast. Manaus is the largest city in the North. The income gap between the north and the south is well documented. For example, Maranh˜ ao, a state in the Northeast has only 27% of the GDP of S˜ ao Paulo ([38]). Cou- pled with the higher average temperatures in the North, one may expect that any effect of heat on the poor to be found there. And yet in this paper I find effects solely in the south, and strongest for the least educated (and likely poorest) therein. Indeed, there is some evidence that high variability in temperature is more damaging than hotter aver- age temperatures for the human physiology ([39]). While it is highly unlikely that air conditioning was the main source of mitigation of heat due to the time period studied, traditional fans may have played their part. There are two key temperature metrics that differ between the north and south: aver- age temperature and seasonal variation. The average daily maximum temperature across municipalities in the north and northeastern regions is 30C during schooling days, while only 25C in the south. That is, the North is more accustomed to continuous hot weather 5 Some cities, such as Rio de Janeiro despite being in the south, can get very hot. Central Brazil follows a similar low variation in temperature across the seasons as the North. 6 Figure 1.1: Climate of Brazil (Figure 1.2a). Furthermore, there is higher seasonal variation in the south, with standard deviations at 3C and 4C respectively. This suggests that there is an element of adaptation involved; individuals living in areas with consistently hot weather are less perturbed by temperature. Of course, thermoregulation may break down in extreme, humid heat, but this is as of yet an uncommon occurrence in Brazil. It should be noted that while the average temperature(Figure 1.2a) shares similarities in distribution across location to the schooling days’ bins(Figure 1.2b), it is not identical. This allows me to distinguish effects during school from summer days and weekends. B Institutional Details:RAIS/Job Market Brazil experienced a period of high inflation until 1994, and I follow many other papers in excluding those years from my analysis. While the RAIS dataset only covers the informal 7 (a) Average Daily Maximum Temperature During Sample Period (b) Average Number of Schooling Days per Year over 30C During Sample Period Figure 1.2: Temperature across Brazil sector, there is evidence that the distribution of wages has a high overlap with the informal sector ([40]). To quote Engborn, ”RAIS covers nearly the entire universe of workers in tax-registered firms. It excludes informal workers and firms, firm owners and shareholders unless they are (self-)employed, and certain less-populated regions in the years before 1994.”. Fur- thermore, information on all workers on the payroll in previous years is legally required. There is high compliance due to large penalties for late, incomplete, or inaccurate data. It is used for social programs and federal wage bonuses, so there is an incentive for accurate reporting from both parties involved ([41]). Unemployment insurance is only involved in the formal sector, and the informal sector makes for a weak substitute. This allays some but not all concerns about the unobserved labour market. 8 C Institutional Details: SAEB/Quality of Education Indicators Brazil has experienced recent increases in enrollment and school attainment. Enrollment rates were starkly lower in the past, well below 50% in the 1940s and finally peaking at near full net enrollment by 2000 ([42]. This and policy relevance are the primary reason for restricting my analysis to the latest decade of earnings data I have; individuals in the earnings dataset would have been expected to be at elementary school between the years 1986-1999. Striving to improve the quality of education, policymakers found themselves with a dearth of information. Hence in the 1990s SAEB (National Assessment of Basic Education) was created to biennially assess students in critical fields of knowledge such as language and mathematics. While the primary purpose was to track progress in schools rather than individual students, I nonetheless am able to use this data in my empirical strategy. I do, however, lack the precision that student level fixed effects would afford me as in other related papers due to the nature of my estimates and data; there are no resits. I can and do control for school level variation through school fixed effects. The assessment system has gone through many revisions over its time. In 1995 the methodology was changed to allow year-on-year analysis. In 2005, SAEB was once again restructured into two assessments: National Assessment of Basic Education (ANEB), which retained the sampling structure of SAEB, and Prova Brasil, a census-style examination. I focus on the most recent years of data, from 2013 to 2019 to ensure consistency and allow for a breakdown by Socioeconomic Class (SEC). The census is applied to all public schools in Brazil with at least 30 students enrolled, while the sample includes private schools and had the cutoff at 10 students ([43]). 9 1.3 Data and Empirical Strategy A Data Three major datasets are combined in this paper to measure: earnings, test scores and temperature. Earnings are from the RAIS dataset, which covers the entirety of the formal sector in Brazil. Brazil has had a large shift in its educational system over time, with the retaking of years becoming less and less common. I do not have data on students’ retaking years, so to mitigate this concern my my analysis is based on the subset of 9 years between 2010- 2018. This also increases the external validity of my estimates for policymakers, as any structural shifts in schooling or the labour market over a long period of time are likely to occur. My primary outcome of interest is individual level monthly wages at age 30, which I average across the three years of 29-31 to reduce noise 6 I also examine earnings separately by year verify that the results are not driven by any particular age of earnings that may be correlated with my plausibly exogenous variation in temperature. I cannot directly observe an individual’s years of education in this dataset and thus I assign their elementary school heat exposure period through their age. As for the municipality of the school (and hence temperature exposure), I proxy with the location they are first seen in the dataset between ages 18-21 7 . Brazil has had a large number of its citizens migrating across the country and this is one further reason for only using the latest years of data. The only predetermined characteristic that is available here 6 As I am interested in earnings for the individual rather than by sector I opt to include those who are not currently employed in the data. This may underestimate their earnings (if they have shifted to the informal sector), but leaving them out also induces bias (if those impacted by heat are less stable in their formal sector employment. 7 Individuals who are not in the dataset in this period are excluded. 10 is their gender. Finally, I examine heterogeneous impacts through the worker’s gender and education level 8 . The second piece of data is the SAEB schooling dataset, built around a biennial test to measure learning outcomes of all schools in Brazil. A sample of students in each school take the test 9 , in the fifth and ninth grades, as well as the third year of high school. I examine the results at the fifth grade 10 for Mathematics and Portuguese in the years be- tween 2013-2019. I directly observe the municipality of the school, as well as the assigned Socioeconomic Level 11 . I use ERA5-Land data to calculate temperature and precipitation. This is a reanalysis dataset that integrates weather stations and satellite data into a climate model to project climate variables at an hourly level, on a 9km grid (0.1 global grid) from 1950 to the near- present. I average this across the ∼ 5570 municipalities in Brazil. The resolution is an improvement over previous papers in the literature which rely on a 32km grid (0.25 global grid). This is particularly important given the small size of many municipalities. At the 25km grid level, only around 3300 municipalities are matched to a gridpoint within their boundaries, while switching to 9km results in around 5300 of the∼ 5570 municipalities having such points 12 . Furthermore, high resolution is important for coastal regions, and with the large pop- ulation situated across the Eastern frontier I expect further dividends in precision ([37]). After averaging across municipality I take the hourly maximum per day. This gives me 8 Education level is a categorical variable indicating whether the worker has completed elementary (5th), middle (9th), high school, or higher education (with Masters and PhDs considered separately.) 9 The test is required if there are at least 20 students enrolled in the grade. 10 This is consistent with my main specification looking at the effect on wages. 11 Socioeconomic level ranges from 1(lowest) to 7(highest), and is provided for the school rather than individual students. 12 For municipalities that intersect with no gridpoints, I assign the nearest neighbour instead. While some municipalities are very large, these are also very sparsely populated and measurement error, defined here as the temperature felt by individuals differing from the temperature of the municipality as a whole, should not greatly bias my results. The smallest unit of geography for population data is at the municipality level and so I am unable to weigh heat by population to measure this more accurately. 11 the daily maximum temperature per municipality from which my independent variables are derived. Brazil mandates 200 days of schooling per year, and my temperature variable encap- sulates this. I do not count weekends or days during the holiday months. I sum the total number of days within each degree centigrade over an individual’s five year elementary schooling period. This is further summed into 2C bins for the main regression specifica- tions. B Summary Statistics Table 1.1 provides summary statistics for the South and Southeast of Brazil, which is the main focus of my results and the most populated area. 63% of the RAIS sample is male, and 76% have completed high school education. Less than 10% of individuals experienced rainfall that was above the 90’th percentile of rain in the region, which is used as a control. Of the approximately 1000 schooling days, the mean number of days in a bin over 28C was 212, with a lower number of days at higher temperature bins. Table 1.1: Summary Statistics, Brazil Mean SD Median Min Max N Male 0.63 0.48 1 0 1 11,019,010 Asinh Wage 30 7.59 2.12 8.00 0 12.43 11,019,010 Highschool 0.76 0.43 1 0 1 11,019,010 Middle 0.91 0.28 1 0 1 11,019,010 Elementary 0.98 0.14 1 0 1 11,019,010 High Rain 0.08 0.28 0 0 1 11,018,997 28-30C 166.69 114.60 151 0 674 11,019,010 30-32C 87.19 86.68 52 0 549 11,019,010 32-34C 30.47 50.07 8 0 446 11,019,010 34C+ 11.77 39.62 0 0 594 11,019,010 As for comparisons to other regions, gender and wages do not differ by much. As expected, on average wages are highest across the south and southeast. The central, north and northeast regions have many more hot days than the south, which is most clearly seen 12 in Figure 1.3. It should be noted that for the South, some areas may not reach the highest temperature bin with common occurrence, if at all, potentially introducing noise and even bias in estimates. Figure 1.3: Temperature Bins by Region There has been a marked increase in average temperatures, and in the number of days over 30C, relative to the exposure of individuals of students in my dataset (See Figure 1.4a, Figure 1.4b). This increase in hotter temperature-days will be one of the multipliers I use to evaluate the economic magnitude of my coefficients for the South and Southeast. 13 (a) 5Y rolling mean Average Daily Maximum Temperature (b) Average Number of Schooling Days per 5 Year over 30C During Sample Period Figure 1.4: Increasing Temepratures in Brazil C Empirical Strategy Taking advantage of an individual’s age, which by large dictates their schooling years, as a source of exogenous variation for heat exposure during their early educational pe- riod, I present regressions of long term earnings and schooling outcomes on temperature exposure. The wage regression is as follows: Wage imt = ∑ b (θ b Temp mtb )+g i +rain mt +γ m +λ t +ε imt Wage is inverse hyperbolic sine of the monthly earnings 13 for individual i, who was first seen in the dataset at age 18-21 in municipality m (to proxy for schooling location), and is aged 30 in yeart. γ are municipality fixed effects, with g a control for predetermined characteristics of gender and rain an indicator of high rainfall. Identification is hence off of within municipality variation across time in earnings. Temp is the number of school- ing days during elementary school in a specific temperature bin b. The interpretation of the coefficient θ is then the percentage change in earnings of an extra schooling day in a 13 The inverse hyperbolic sine results are largely consistent with using thelog(1+x) transformation. 14 particular temperature bin as with respect to the excluded bin. 14 Standard errors are clus- tered at the municipality-year level, in line with the level of variation in my treatment. To better understand the magnitude of the effects I also report the results of a one standard deviation increase in hot days 15 , and the effect in the counterfactual scenario where each individual is assigned the heat exposure of their municipality over the hotter 2016-2020 period 16 . The schooling regression follows a similar strategy, with some minor alterations: Score ismt = ∑ b (θ b Temp mtb )+rain mt +γ s +λ t +ε ismt As I have the exact school the student attended I can take fixed effects at the school level s and furthermore, Temp no longer requires to be proxied by age or early work lo- cation, and is merely computed using the previous five years of temperature data for the municipality the school is located in. The regression is weighed by school enrollment. 17 Score is the standardized SAEB test score for students in their fifth grade in Mathematics and Portuguese. 18 θ can be interpreted in a similar manner to the wages regression: the change test scores measured in standard deviations of an extra schooling day in a hot bin as opposed to the excluded. Once again I present the effect for a one standard deviation increase and under the more recent real, hotter counterfactual set of temperatures. 14 2C temp bins are used in the main specification, and 24-26C is the excluded bin to avoid perfect multi- collinearity. 15 This amounts to 10.23 days over the schooling period for the 30-32C bin. 16 This is 23.8 days for the 30-32C bin 17 The dataset is on an individual level but a students are sampled by school. Smaller schools will likely be oversampled and hence the need to weigh the regression. 18 The exams are completed near the end of the schooling year, so this includes the year that they take the exam. 15 1.4 Heat and Long Term Impact on Earnings A Wages by Climate I begin by examining the impact of elementary school heat shocks on long term earnings at age 30. My full sample consists of around 11 million workers. Table 1.2 column (1) shows magnitudes of between 0.05-0.085 for Brazil as a whole. Taken at face value, this suggests that an extra school day in a higher temp bin (over 30C) reduces worker wages at 30 by 0.05-0.085%. This is smaller in magnitude than the effects found in the literature focusing on heat during early childhood and inter-utero ([28]) 19 , however the period in question I study lasts much longer and so excess heat has the potential to be much more damaging. Columns (2), (3) and (4) split the sample by the regions of Brazil 20 ; the results here are notable. While other papers ([11]) have indeed found smaller impacts in areas used to hot weather, I find that the effects are being purely driven by the cooler south, with the null effects in other regions masking the true magnitude therein. Figure 1.5b plots columns (2) - (4). There is no statistically significant effect in the north or central regions at higher temperatures, while the coefficients rise to 0.09-0.18 % for a one day increase in the number of hot days. While more extreme temperatures days are more damaging, they area also less com- mon. Furthermore, a single day of hot weather is difficult to size in the grand scheme of years of education. To get a more meaningful number, I multiply the coefficients in Ta- ble 1.2 column (4) by Table 1.3a. This second table gives the standard deviations for the number of days in each temperature bin for each municipality, and municipality weighed by workers in the dataset, and finally the difference compared to the 2016-2020, for the 19 The effect found is 0.1% for an extra day over 32C. 20 Most of the population is located near the coast on the east side, so while North-Northeast/South- Southeast is the more accurate split I refer to the areas as North/South for brevity throughout this paper. 16 Table 1.2: Elementary Schooling Days Temperature on Wage by Region Dependent Variable: Wage percent Region Brazil Central North-Northeast South-Southeast Model: (1) (2) (3) (4) Variables 26-28 0.0034 0.0616 -0.0069 -0.0089 (0.0213) (0.0592) (0.0405) (0.0245) 28-30 -0.0353 0.0108 -0.0269 -0.0346 (0.0240) (0.0513) (0.0369) (0.0291) 30-32 -0.0504 ∗∗ 0.0108 0.0003 -0.0917 ∗∗∗ (0.0240) (0.0617) (0.0381) (0.0351) 32-34 -0.0857 ∗∗∗ 0.0900 -0.0320 -0.1866 ∗∗∗ (0.0255) (0.0813) (0.0403) (0.0454) >34 -0.0556 0.2073 ∗∗ 0.0002 -0.1594 ∗∗ (0.0350) (0.0837) (0.0512) (0.0743) Fit statistics Controls Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Observations 11,018,997 926,901 1,748,088 8,344,008 R 2 0.01960 0.03610 0.02368 0.02085 Within R 2 0.00019 0.00165 0.00099 9.92× 10 − 5 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. Controls include Low temperature bins, rain 90th percentlie indicator, and gender. Fixed effects include municipality and year 17 (a) Brazil (b) Regions Notes: Figure 1.5: Main Specification South and Southeast. The standardized coefficients are in Table 1.3b. A 1 standard devi- ation increase in the number of days of temperature over 30C reduces earnings by 0.61- 1.36%. Larger still, had the temperature in 2016-2020 been reflected in the true weather during schooling years, one would find a decrease in wages of between 2.18-3.13%. While this is large, caution must be taken in interpreting the coefficients - the effect is likely to be smaller once general equilibrium effects are taken into account, and modern day heat mitigation technology may reduce the impacts further. I interpret these magnitudes as absolute upper bounds on my estimates. There are a few potential reasons for why it is only in the relatively cooler South that I find effects. Firstly, individuals who expect hotter weather may have already adapted through the adoption of air conditioning. While there is a compelling case in the US, adoption of air conditioning is much lower in Brazil, even if more prominent in the north, and so AC is unlikely to be the whole story here. Cheaper to run fans may still play a part. Secondly, it is possible that those living in the north and central regions are already used to 18 Table 1.3: Elementary Schooling Heat Coefficients scaled, for South (a) Factor Scaling by Bin in South 28-30C 30-32C 32-34C 34C+ Average s.d. across municipalities 17.43 13.69 7.30 3.81 Average s.d. across individuals 11.99 10.23 5.62 4.00 Number of extra days in 2015-2020 23.07 23.80 16.78 13.39 (b) Scaled Coefficients by Bin in South 28-30C 30-32C 32-34C 34C+ By Average s.d. across municipalities -0.60 -1.25 -1.36 -0.61 By Average s.d. across individuals -0.42 -0.94 -1.05 -0.64 By Number of extra days in 2015-2020 -0.80 -2.18 -3.13 -2.13 hot weather, and thus feel less of the psychological impact and discomfort associated with it. Indeed, equatorial weather, while hotter than the south, is much more consistent across the year. Finally, the lower variation in temperature, coupled with the long term nature of the effect against the identified cause may simply mean I do not have the necessary statistical power to uncover smaller effects that may come about from adaptation, and individuals becoming ”used to” hot weather in other regions as mentioned in previous points. The literature suggests a few main mechanisms through which heat may be affecting long term earnings. The first is a story of human capital accumulation. This hypothe- sis dictates that students fail to acquire as much human capital during their critical early schooling period, leaving them permanently behind in the future as captured by lower wages in the job market. A second channel would be long term health, also included in the human capital channel. A third mechanism is not one of human capital accumulation (specifically learning), but one of signaling. Students may have accumulated knowledge and skills, but upon entering the job market find the signal of their skills distorted by poor examination results relative to their non-heat stricken peers. One final mechanism is moving from education into productive work when rainfall is high due to the higher opportunity cost of work ([22]). I control for rain so my results are indicative of a channel outside of this. ) 19 There is also a question on which specific period is critical for students in their edu- cation; whether elementary the only sensitive period, or if heat is damaging later in their schooling years. Is the first or final year of elementary driving the results, due to a ”build- ing blocks” or ”test signal” story respectively? I detail mechanisms as well as these issues in the remainder of Section 1.4. B Mechanisms: Summer and Weekend Temperature Effects Including precipitation as a control suggests that I am not misattributing the effects of another potentially correlated climate indicator in these results. The inclusion of munici- pality fixed effects also allows me to rule out biases from any time-invariant unobservables at the municipality level such as baseline economic conditions. There are still questions of mechanisms left to be answered, and while this tends to be difficult when looking at long run effects I present suggestive evidence in favor of a learning channel. I only consider the effects during schooling days, not weekends or holidays, however these are likely to be correlated since hot days come in waves rather than idiosyncratically each day. Is the effect also being driven by heat exposure during the summer or weekends? For example, students may work on weekends, and become more exhausted than other- wise when the day is hot, not fully recovering by the time they return to school. Summer heat may also affect their longer-term health, and poorer health outcomes may translate into poor future outcomes. I have been hypothesizing a school-related channel, and Table 1.4 leads credence to my claim. In column (1), I find a significant detrimental effects of an extra hot school day 0.055%. Similarly if one examines holidays and weekends, we also see significance as in column (2). However, once hot schooling days are included the coefficients on week- end and holiday hot days are no longer significant the 5% level, instead replaced in by the newly added hot schooling day coefficients at 0.059%, as in column (3), statistically indistinguishable from the estimate without non-schooling days. 20 Table 1.4: Elementary Days on Wage: School Days vs Holidays and Weekends Dependent Variable: Wage Percent Model: (1) (2) (3) Variables School Days>30C -0.0545 ∗∗∗ -0.0586 ∗∗ (0.0117) (0.0267) Holidays>30C -0.0369 ∗∗ -0.0343 ∗ (0.0182) (0.0186) Weekends>30C -0.0847 ∗∗∗ 0.0357 (0.0307) (0.0606) Fit statistics Controls Yes Yes Yes Fixed Effects Yes Yes Yes Observations 11,019,010 11,019,010 11,019,010 R 2 0.01958 0.01958 0.01958 Within R 2 0.00017 0.00016 0.00017 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is < 30C. Coefficients multi- plied by 100 to allow for direct percentage interpretation. Control includes gender. Fixed effects include municipal- ity and year 21 Were the effect health-related, I should find independent effects for both summer and schooling days, which I do not. Furthermore, the lack of weekend effects strongly suggests that my effects are indeed attributable to school directly. Heat-related economic shocks can also be ruled out. Tourism, agriculture, construction and other outdoors economic activity would have affects all days of the year. Indoor activity, given the time period in study (with little air conditioning) would also likely have suffered throughout the year. C Comparison to Middle and High-School Temperature Shocks Table 1.5: School Days Temperature on Wage by School Exposure Period - Completed Only Dependent Variable: Wage Percent from Exposure During: Exposure Period Elementary Middle Highschool Model: (1) (2) (3) Variables 26-28 0.0064 -0.0025 -0.0171 (0.0210) (0.0255) (0.0309) 28-30 -0.0422 ∗ 0.0100 0.0129 (0.0239) (0.0211) (0.0302) 30-32 -0.0599 ∗∗ 0.0310 0.0204 (0.0235) (0.0231) (0.0273) 32-34 -0.1016 ∗∗∗ 0.0335 -0.0542 (0.0251) (0.0318) (0.0491) >34 -0.0808 ∗∗ 0.0170 0.0593 ∗ (0.0332) (0.0277) (0.0352) Fit statistics Controls Yes Yes Yes Fixed Effects Yes Yes Yes Observations 10,799,544 10,041,859 8,334,835 R 2 0.01776 0.01432 0.01312 Within R 2 0.00011 1.2× 10 − 5 0.00032 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. Controls include rain 90th percentlie indicator and gender. Fixed ef- fects include municipality and year My paper focuses on elementary school shocks. But is there evidence that learning is inhibited later on in the educational journey? I examine this in Table 1.5. Column (1) is the 22 standard elementary school shock, while Columns (2) and (3) has temperature bins dur- ing middle and highschool as the independent variables. Each regression only includes individuals who completed the level of education specified. I find significant effects dur- ing elementary school, but not for middle or highschool. There are a few interpretations behind this. Firstly, it could be that students at a younger age are more susceptible to heat and distracted more easily therein. An elementary school student is less cognizant of the potential lifetime value of learning, and thus more likely to allow heat to reduce their motivation. Individuals who progress to middle or highschool differ from those that do not, in that they had and took the opportunity to do so. This selection is another potential reason for my results. It could be that those that are less perturbed by impediments to their learning complete more education, so it is not the case that heat during later years of education are not detrimental, but those that take more education are better at dealing with heat while in school. A third, final reason for this finding may simply be due to poor estimation of the years of exposure of my empirical strategy. I am using an individual’s year of birth to assign their years of education. It was common for students to retake years 21 , so the further into their education I estimate, the less likely they are at the stage of education I have assigned them. I do note that this does not mean that they are not in education, simply that they are not assigned the correct educational level for their temperature exposure. This however does not change the fact the coefficients are insignificant for later periods of study, even if mislabeled. D Decomposing the Effect by Year of Elementary School The literature points towards two main schooling channels through which heat may affect outcomes: learning and signaling. My paper argues for the former, and Table 1.6 further 21 This form a part of my decision to focus on the latest years of data, where such concerns are less prevalent 23 Table 1.6: Elementary Schooling Days Temperature on Wage by Region, School Year Split Dependent Variable: Wage percent Region Brazil Central North-Northeast South-Southeast Model: (1) (2) (3) (4) Variables 26-28 0.0018 0.0647 -0.0169 -0.0099 (0.0214) (0.0583) (0.0420) (0.0246) 28-30 -0.0263 -0.0477 -0.0286 -0.0142 (0.0253) (0.0479) (0.0376) (0.0316) e1>30 -0.0550 ∗∗ 0.0493 -0.0064 -0.1287 ∗∗∗ (0.0261) (0.0714) (0.0375) (0.0384) e2>30 -0.1159 ∗∗∗ -0.0522 -0.0612 -0.1950 ∗∗∗ (0.0336) (0.0833) (0.0449) (0.0505) e3>30 -0.0619 ∗∗ -0.0995 0.0330 -0.1434 ∗∗∗ (0.0302) (0.0681) (0.0369) (0.0473) e4>30 -0.0773 ∗∗∗ -0.0195 -0.0185 -0.1608 ∗∗∗ (0.0283) (0.0793) (0.0461) (0.0422) e5>30 -0.0067 0.0652 -0.0089 -0.0294 (0.0280) (0.0716) (0.0420) (0.0424) Fit statistics Controls Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Observations 11,018,997 926,901 1,748,088 8,344,008 R 2 0.01961 0.03607 0.02369 0.02085 Within R 2 0.00020 0.00162 0.00100 0.00011 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. Controls include Low temperature bins, rain 90th percentlie indicator, and gender. Fixed effects include municipality and year 24 supports this. In these results I split the schooling effect by grades 1-5 for bins over 30C 22 . As expected, the effects are significant for Brazil as a whole and the South-Southeast, but not the Central or North-Northeast regions. However in the fifth and final year of elemen- tary the effects dissipate. This is where it is determined if students progress to middle school education or retakes the year. Were the mechanism through exam signals, one would expect the opposite set of results - only heat during exam season (a subset of heat in the final fifth year) would have had significant effects. Looking more broadly at the years of elementary we see in column (4) the strongest effect in the second year for the South-Southeast. An extra hot day during the second year of elementary reduces earnings at age 30 by 0.195%. The first year of elementary differs from the others since the focus is on achieving literacy, which may be why the effect is larger in the second year and onwards. However, I note that other than the fifth, each year remains individually significant. There does not appear to be a ”building blocks” story in which the first year is most critical and heat shocks have caused students to fall behind. High temperatures are detrimental throughout most of elementary school. E Further Evidence E.1 Robustness Checks I run numerous robustness checks to validate my main specification. Table A.8 tests for sensitivity to rain and gender controls. The results remain largely significant throughout, although the inclusion of rain controls tends to reduces the magnitude of the coefficients. Table A.9 includes specifications that control for temperatures during middle, highschool and higher education, as well as subsampling to only include individuals who finish ele- mentary school. Once again the effects largely remain significant, especially at the higher end of the temperature range. 22 With my variation split across 5 years instead of one, I do not have the power to further subdivide this bin in the manner I do for the main specification. 25 My main specification relied on 2C temperature bins. I also present 1C and 3C bins. Figure A.4a and Figure A.4b shows the expected effects from Table A.10 and Table A.11. At 1C, the bins are too small, and start to become highly correlated to one another. Each bin now has a smaller number of observations, and is trying to estimate a more precise effect (of a single degree rather than two degrees). On the other hand, while the effects are more obvious with the larger bin sizes, the curvature of the temperature impact is lost. This led me to the natural middle ground of 2C bins. While the temperature variation across individuals born in different years are plausi- bly exogenous with respect to each other, it is still possible that there is some correlation of an individual’s temperature during elementary to other critical periods in their lives or the data in question; namely their birth temperature exposure or exposure when at work when their wage was observed. For example, perhaps individuals in my dataset who had high schooling temperatures also had high birth temperatures, making it difficult to dis- entangle the causal effect. Demeaning by municipality and year average, I calculate the correlation coefficients of average temperature for elementary schooling, birth, and work temperatures across individuals born in different years in Table A.2. The correlations are small: 0.286 for elementary and birth, and 0.043 for elementary and work. This suggests that the temperature shock during elementary schooling is also independent of other tem- perature shocks an individual may have faced. To provide further evidence of the lack of correlation of shocks over time, I split the averaged wage back into its composite parts. For wages at ages 29, 30 and 31, I still find effects, albeit with less precision, as expected, as shown in Table A.12 and Figure A.5. E.2 Non-Linearities The main regression specification assumes linearity of effects within a bin. That is, an extra day in a temperature bin reduces long term earnings by the same value throughout the support. Rather than across bins, in this section I estimate a more flexible within-bin 26 functional form. Table A.4 explores the results. For Brazil as a whole in column (1), an extra hot schooling day reduces earnings by 0.07% at the margin at the average number of hot schooling days (129 days). Columns (2)-(4) give results for the macroregions. The North-Northeast is the hottest region with students experiencing 281 days over 30C during their schooling period. The effect is insignificant here. Meanwhile in the South-Southeast, where such high temperatures are much less prevalent (85 schooling days), I find the marginal effect to be 0.14%. The positive sign on the square coefficient imply the effect of heat tapers the more hot days a student experiences. This contrasts with the main set of results which showed that a relatively hotter day is more damaging than a less hot day, and suggests an element of within-bin adaptation. One interpretation is that the psychological effect of a heat shock is greater than the damage from persistence - an explanation in line with the differing effect across regions. One should note, however, that my specification is unable to determine if different regions are impacted differently due to differing heat exposure, or adaptation. For example, if the cooler south were to experience heat at the same level as the north, would they also find the effects of heat to be less problematic (adaptation leading to rel- atively smaller coefficients, i.e. positive coefficient on square term)? Or would such heat exposure in areas unaccustomed to constant heat be unable to cope (larger coefficients, i.e. negative coefficient on square term). What I can and have discussed is, however is that local to the number of days of heat typically experienced, the marginal effect falls in magnitude with each extra day of heat. E.3 The Effect of Extra Education Returning to the elementary school heat shocks, Table A.5 examines the effect for indi- viduals who have attained higher levels of education thereafter. Column (1) includes the whole sample, inclusive of those who did not finish elementary school. Columns 27 (2)-(4) provide coefficients for individuals who have completed elementary, middle, or highschool at a minimum. The coefficients increase in magnitude as individuals who have not completed elemen- tary school are dropped - if they are not in school, they are not learning and so cannot have their learning inhibited by heat. As the level of total education increases, the coefficients decrease in size, only remaining significant for the 32-34C bin, but still negative in effect. Extra education may allow students who were affected by heat during elementary to catch up to their peers during middle and highschool. However it is also possible selection going on in those that decide to take up further education. I do note that 91% of my sample completed middle school, so there is less selection on that front and it is easier to argue students are catching up. 28 1.5 Heterogeneity in the South A Wages by Gender Previous research ([30]) suggests the effects of heat are greater amongst the male pop- ulation, and my findings support this. Table Table 1.7 examines my evidence, finding significant effects only in the south. An extra day of hot weather between 32-34C reduces female’s earnings by 0.13%, while for males it is almost double in magnitude at 0.22%. For the 30-32C bin, only the coefficient for males is significant. Reasons for the difference may be postulated. If male students are more likely to engage in informal labour outdoors, or spend time outside during recess playing soccer, they will be more exposed to high heat than female students. Returning to class more exhausted than otherwise, the lack of concentration may inhibit their learning. While such differ- entials (especially recess spent outdoors) may have reduced in more recent years, for the time period of schooling involved in my sample (1986-1999), this channel is still likely to be significant. Another strand of the literature explains this difference through the higher metabolic rate of males compared females ([44]), and thus a higher body temperature for the former group. This results in a higher temperature of comfort for women ([45]), and there is evidence that in an experimental setting females handle heat better when it comes to tests ([46]). Indeed, in the workplace it is men who have the advantage as temperature are set for their comfort ([47]) - so a stronger negative effects for males is evidence that the effect I find is not correlated to temperatures experienced in later life during work. There is also potential selection at play. One should recall that my sample consists of individuals with a long term attachment to the formal sector. Perhaps the barrier to entry to to the formal sector is higher for women, and those who make it are individuals most resilient to a lapse of of concentration from high heat. 29 On a final note, the difference by gender rules out certain mechanisms that would be largely identical between the sexes. If the teacher’s discomfort reduced the efficiency by which they taught, one would not find heterogeneity by the gender of the student. Male and female students also tend to share the same school, so differences in school-level char- acteristics are unlikely to be the cause. Table 1.7: Elementary Schooling Days Temperature on Wage by Region and Gender Dependent Variable: Wage percent Region Central North-Northeast South-Southeast Gender F M F M F M Model: (1) (2) (3) (4) (5) (6) Variables 26-28 -0.0222 0.0975 -0.0223 -0.0016 0.0130 -0.0233 (0.0707) (0.0701) (0.0574) (0.0530) (0.0291) (0.0316) 28-30 -0.0968 0.0763 -0.0243 -0.0272 0.0218 -0.0687 ∗ (0.0638) (0.0603) (0.0494) (0.0465) (0.0406) (0.0356) 30-32 -0.0373 0.0495 0.0411 -0.0185 -0.0481 -0.1234 ∗∗∗ (0.0765) (0.0730) (0.0487) (0.0490) (0.0534) (0.0413) 32-34 0.0173 0.1338 -0.0621 -0.0179 -0.1376 ∗∗ -0.2227 ∗∗∗ (0.1067) (0.0946) (0.0543) (0.0508) (0.0564) (0.0592) >34 0.0738 0.2822 ∗∗∗ -0.0200 0.0105 -0.0450 -0.2408 ∗∗ (0.1089) (0.0968) (0.0583) (0.0647) (0.1004) (0.0947) Fit statistics Controls Yes Yes Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Yes Yes Observations 302,776 624,125 561,421 1,186,667 3,204,820 5,139,188 R 2 0.03243 0.04013 0.02768 0.02488 0.02559 0.02198 Within R 2 0.00018 0.00017 0.00017 8.94× 10 − 5 7.53× 10 − 5 9.81× 10 − 5 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct per- centage interpretation. Controls include rain 90th percentlie indicator and gender. Fixed effects include municipality and year 30 B Wages by Education Level I now turn to heterogeneity by the level of education an individual has achieved. Table Table A.6 documents the results. I split by those who have not completed high school, and those who have 23 . The effects are not significantly different except in the south/southeast. It should be noted that the level of education attained is endogenous. That is, this set of results must be interpreted with care: individuals who had a heat shock may be less likely to complete highschool, and it may be this rather than their elementary school shock that results in lower wages. An explanation for the earnings gap may be found in the parents of the children. Wealthier parents are more likely to be able to mitigate the effects of heat, and are also more likely to educate their children to a higher level. Indeed, those who have more years of education rely less on their formative years; more education is perhaps the most straight- forward way to catch up with one’s hot day-less peers. The natural assumption is that poor schooling outcomes is the main channel through which this wage disparity propagates. I note that the level of education achieved is endogenous and so this my regressions are not true, unbiased analysis of heterogeneity. In my supplementary results on schooling outcomes, I return to the idea that there is also a gap by socioeconomic level, which may be correlated with and the interpretation of this set of results. B.1 Robustness Checks I run a few different specifications using data from the south and southeast as robustness checks. Table A.13 estimates specifications without rain controls or gender. I find the re- moval of the rainfall control decreases the magnitude of the coefficients, but coefficients remain significant. Gender meanwhile has no appreciable effect, as can be seen by com- paring columns (1) and (3), or (2) and (4). Table A.14 splits the wages into its yearly 23 My sample excludes individuals who have never attended elementary school, as such individuals by definition were not subject to any heat shock during their ”time in elementary school”. 31 components - Figure A.6 plots these. The effects found are not driven by a particular year of data, and while not as precisely estimated the trend remains as with the average.. 1.6 Heat and Schooling Outcomes A The Region Divide, Revisited Table Table A.15 presents estimation results by region. For the most part, it appears as though this is mostly noise, with potentially some effect in the South and Southeast. While not as convincing as the result on wages, the region where I find an effect is the same. Is there evidence of detrimental outcomes earlier than for earnings at age 30 that could support my results thus far? This section documents the evidence using 5’th grade SAEB test scores. There were no significant effects for Portuguese test scores, consistent with findings from other researchers ([30]). I focus on the results in Mathematics. B Test Scores by Socioeconomic Level I run regressions on subsamples split by the socioeconomic level of the school across all regions. Table Table A.16 presents my results. I find significant effects only in the south, mirroring the results from the main specification on wages, as in Figure A.7. This is a distinct dataset to RAIS, so the results were not a given. Indeed this would suggest that my main results on wages are not driven by an artefact of the RAIS dataset (the datasets cover different periods), however it is still possible that unobservables therein are correlated. Wages were most impacted for the least educated in the south, and similarly here in this specification I find educational achievement impacts the largest for the lowest socioe- conomic classes in the same areas; an extra day of high heat reduces standardized test scores by up to 0.0031 standard deviations at over 34C. The coefficients do not have as much of a curvature towards more extreme values, however I note other papers on test scores that follow my empirical strategy are similar in this regard [24] 32 While not conclusive, the results are suggestive. There is a well know correlation be- tween lower educational outcomes and poorer students. As such my results support the hypothesis that it is a wealth gap, proxied by educational level, rather than the endoge- nous educational level that is driving the gap in wages seen. Furthermore, the educational SAEB dataset is from students in a much more recent time frame - and so inclusive of adap- tation over the time difference. This suggests Brazil continues suffer from the detrimental impacts of heat on education (and thus long term human capital accumulation and earn- ings) to this day. 1.7 Conclusion I present the first set of evidence, to the best of my knowledge, of the long term effects of heat exposure during formative years of school. I look one step further than previous re- search, corroborating the findings that learning affects the accumulation of human capital; individuals suffer in the job market through lower realised wages. I find effects for elementary school, but do not discount the possibility that middle and highschool are also important, though my paper does not find evidence for this. I postulate that early schooling is when students are most susceptible to heat, whether due to a lack of concentration, selection out at higher levels of education, or otherwise. The effect remains consistently significant for all but the final year of elementary education. The long term nature of my outcome variable weakens the opportunity for me to invoke causality amongst all possible routes. However through results on summer and week- end temperatures, as well as the inclusion of rainfall and municipality fixed effects I can present results inconsistent with certain other channels such as health, economic condi- tion, and any time-invariant unobservables at the municipality level. My findings suggest that there is a vast range of heterogeneity in the effects. It appears that males and areas unaccustomed to hot weather are more susceptible to heat. The gender disparity is well 33 known in the literature, and regional heterogeneity has been postulated and identified before. Further, while endogenous due to elementary outcomes potentially affecting high school completion rates, I also document how the least educated seem to suffer greater wage re- ductions. Using a separate dataset I find a similar trend on educational outcomes for the poorest schools. While not conclusive, the typical overlap between these two sets of groups suggests it is the poor who suffer the effects of heat the most. There are simple adaptation strategies such as investment in air conditioning or extra schooling days that policy makers may take to help mitigate these very real effects. But care must be taken - there is no need to continuously cool classrooms if it is only extreme fluctuations that are detrimental. If it is the case students start to get accustomed to hotter temperatures (which my results cannot reject due to my null findings in the northen and central regions of Brazil), then air conditioning may not be necessary. From the perspec- tive of gender disparities they may be reversed rather than reduced, if air conditioning is set in the classroom lower and optimally for males as tend to be the case in the workplace. The historical temperatures experienced are still manageable, but if recent trends of record breaking heat continue, there is a real risk of death and physical impairment, even amongst the healthy section of the population. No doubt in such cases if weeks of 40C rather than 30C begin to arise the impacts will be greater still. Furthermore, with the breakdown of thermoregulation, it may no longer be feasible for the north and central regions of Brazil to acclimatize to to heat, implying large and economically significant effects throughout the whole of Brazil. I hope that future researchers can shed more light on the impact of more extreme heat shocks on long term outcomes, with particular focus on exploring the interactions between the formal and informal sectors. 34 Chapter 2 Last Hired, First Fired: Droughts and Labor Market Inequality in Brazil - with Antonio Bento and Edson Severnini 2.1 Introduction and Motivation Brazil is a country that has experienced multiple droughts over the past few decades, with two highly publicized droughts occurring in 2001 and 2012. While the impact of these droughts on Brazil’s agricultural sector has been well-studied, their effects on the labor market remain less explored. A separate set of literature has found that minorities are the first to be fired and last to be hired. This proposal aims to examine how droughts in Brazil affect labor market inequality, dovetailing with the environmental justice literature, and searches for evidence that environmental concerns may exacerbate underlying inequalities across race and gender. Droughts do not hit all sectors equally and some such a farming or construction may be harder hit, while others sectors are indirectly affected. The mechanisms of firing and rehiring is another potential objective of this research. What are the stated reasons for fir- ing workers, and are they able to find jobs quicker by being rehired into other less affected sectors? Separations can also be avoided by taking lower wages - do we find this substi- tution going on? We also intend to study the heterogeneity in effects across regions; for 35 the tourism sector is largely situated across the north-east of Brazil which may result it in inequalities exacerbated disproportionately by extremely dry seasons. 2.2 Research Design A Data Two major datasets are combined in this paper to provide us with a measure of jobs status, earnings, and droughts. As with my Job Market Paper, earnings are from the RAIS dataset, which covers the entirety of the formal sector in Brazil since 1985. Race is only measured after 2002, so we take our sample as the subset from 2003-2018. We use monthly level wages, taking the highest earning for a worker per month if they have multiple jobs. We do not include those on weekly or other non-monthly payment routines. Using hiring and firing dates, we are able to determine the month individuals are in work. If an individual is outside the workforce for a year they do not appear. Thus we assign them as unemployed for any period they are not in the dataset. Race is measured inconsistently, so we use the race most commonly reported across years. There is a non-trivial number of increases, and more questionably decreases in the education level reported by workers. We once again assume the most common reported level over the period of study is correct and assign this to each worker. The RAIS dataset also provides other pertinent and useful information, such as the municipality of each job and the reason for firings. We also calculate workers’ months unemployed, and months-to-employment for each month. Working with the large size of the RAIS dataset becomes tricky, so preliminary analysis is based on a random 1% sample of workers that are present in the latest data (2013). From this, the sample is back-filled by selecting workers from previous years who were also in the 1% sample in 2013. Brazil’s municipalities have shifted over the years, so we use 36 Minimum Comparable Areas (AMCs) as our unit of geographic observation 1 . Workers are assigned AMCs based on their municipality codes. We use two measures of droughts. The first and primary measure is the SPEI index which measures moisture deficit in a given location relative to its 100 year average and is based on local precipitation and temperature data. It compares the amount of precipita- tion in a given area with its evapotranspiration needs, which are a function of temperature. We measure the index across different time horizons: one month, 6 months, and the past 12 months. According to the classifications of McKee et al. (1993) and Paulo et al. (2012) for the SPEI drought classes, we have 5 classes, namely: 1. Non-Drought (in this class the value of SPEI greater than -0.5), 2. Mild (the value of SPEI is between -0.5 and -1), 3. Moderate (SPEI is between -1 and -1.5), 4. Severe (SPEI is between -1.5 and -2), 5. Extreme (Less than -2). The second measure of droughts is rainfall from the ERA-5 Land dataset. This is a re- analysis dataset that integrates weather stations and satellite data into a climate model to project climate variables at an hourly level, on a 9km grid (0.1 global grid) from 1950 to the near-present. We aggregate to the monthly AMC level. For each AMC, we calculate standard deviations and create dummies for rain that is one, two, or three standard devi- ations below the AMC average 2 . This is also repeated each month for measures created on rainfall in the past 6 months. 1 There are approximately 5600 municipalities, which are mapped to around 4000 AMC, which are con- sistent from 1991 onwards. There was a large change in 2010 2 Average rainfall per month over the period 1950-2021 37 B Empirical Strategy We use a fixed-effects approach, taking advantage of the natural variation in droughts across months and AMCs. The baseline regression is as follows: y i,m =βDrought a,m− 1 + f(γ s,a,r,m,y )+ε i,m Wherey is the individuali’s employment status in monthm, number of months unem- ployed to date, or months until employed again from the current month. We also exam- ine monthly wages as an outcome. Drought a,m− 1 is an indicator for droughts in previous month(s), and γ is a flexible set of fixed effects over sectors, AMC, region 3 , month, and year respectively. We also consider monthly wages as an outcome variable. Our primary measure of droughts uses the SPEI index multiplied by (-1) in the previous month, pre- vious 6 months and previous year. This multiplication helps ease the reading of results and interpretability; a positiveβ coefficient implies poorer labor market outcomes due to excess dryness. Standard errors are clustered at the AMC level, inline with the level of variation in treatment. Employment status captures the ”firing” element of workers in response to droughts, while months unemployed aims to answer the ”first fired” question 4 . Months-to-employment is focused on the ”last rehired” aspect. There are priors that some sectors are more vulnerable to droughts than others. We examine the effect by sector in the following interaction framework: y i,m = ∑ s (β s Drought a,m− 1 )+λ s + f(γ a,r,m,y )+ε i,m Where we now have individual coefficients for each sector β s , and explicitly include differing intercepts for sectors, λ s . We also join individual sectors into a ”high impact” 3 The five regions of Brazil are: North, North-East, Central, South, and South-East. 4 Months unemployed will be 0 if workers are still in work after a drought period. 38 sector, which is expected to be construction and farming, amongst others, for the final triple interaction regression examining the environmental justice question of disparate outcomes by race and gender: y i,m = ∑ s,c (β s,c Drought a,m− 1 ∗ genderXrace c )+λ s,c + f(γ a,r,m,y )+ε i,m Our indicator variable genderXrace spans over characteristics: c∈{femaleXblackBrown, femaleXwhite, maleXblackBrown, maleXwhite} and we allow interceptsλ s,c to vary accordingly. The crux of this proposal is answered by this final regression. We look for evidence that women and minorities are the first fired and last to be rehired when a drought occurs, and expect this effect to be strongest in industries most impacted by droughts. 39 2.3 Preliminary Results We run initial regressions and present our results for our second equation. The measure of droughts is the negative of SPEI 12. This the dryness of the rolling previous 12 month period. The outcome variables are indicators for unemployment status, and count vari- ables for the number of months unemployed to the current month, and the months to the next employment. Two fixed effect specifications are examined, with the results largely remaining significant in the more stringent scenario. Using the months unemployed (column 4), we split the sectors into ”affected” and ”unaffected” by whether the coefficients are significant. Our final set of results splits the results further into ”high” effect and ”low” effect sectors, using a cutoff of a coefficient of 0.25 to distinguish the two. The sectors most vulnerable to droughts are: Farming, Con- struction, and Financial-Health-Insurance. The sectors less vulnerable but still affected are: TransformInd, BusinessRepairMotor, Communication, OtherServices. For further coverage, we also include sectors in ”low” that have large, significant coefficients in the unemployment regression: ExtractInd and ElectricitGas. The results in Table 2.3 suggest droughts do have an effect on unemployment and this varies greatly by sector. For a highly vulnerable sector, a 1 standard deviation increase in dryness over the past 12 months increase the probability of being unemployed by 2.3 percentage point. On average, individuals who experience such drought conditions have been unemployed for 0.48 months longer than they would have been otherwise, and they take 0.37 months longer to find their next job. Our results suggest droughts have an economically significant effect on the labor mar- ket. The next step we intend to accomplish is a triple interaction that studies the differences by race and gender. We expect that the previous effects are strongest amongst women and minorities, and that droughts may further widen gaps across different demographics. 40 Table 2.1: Droughts by Individual Sectors Dependent Variables: unemp months unemployed months to employment Model: (1) (2) (3) (4) (5) (6) Variables spei12 n× sector = FarmForestFishing 0.0062 ∗ 0.0074 ∗ 0.3480 ∗∗ 0.4044 ∗∗∗ -0.1738 -0.1583 (0.0037) (0.0038) (0.1394) (0.1401) (0.1249) (0.1281) spei12 n× sector = ExtractInd 0.0160 ∗∗ 0.0169 ∗∗∗ 0.1564 0.2018 0.4935 ∗∗ 0.4957 ∗∗ (0.0063) (0.0063) (0.1661) (0.1687) (0.2184) (0.2086) spei12 n× sector = TransformInd 0.0040 0.0048 ∗∗ 0.0665 0.1177 ∗∗∗ -0.0020 -0.0061 (0.0025) (0.0023) (0.0515) (0.0441) (0.0322) (0.0464) spei12 n× sector = ElectricityGas 0.0131 ∗∗∗ 0.0141 ∗∗∗ 0.1779 0.2378 0.5236 ∗∗∗ 0.5288 ∗∗∗ (0.0036) (0.0038) (0.1383) (0.1551) (0.1327) (0.1264) spei12 n× sector = WaterWaste 0.0045 0.0055 0.0823 0.1344 0.3508 ∗∗∗ 0.3492 ∗∗∗ (0.0038) (0.0042) (0.1422) (0.1606) (0.0985) (0.0897) spei12 n× sector = Construction 0.0124 ∗∗∗ 0.0133 ∗∗∗ 0.2831 ∗∗∗ 0.3330 ∗∗∗ -0.0137 -0.0130 (0.0020) (0.0020) (0.0524) (0.0464) (0.1227) (0.1301) spei12 n× sector = BusinessRepairMotor -0.0005 0.0003 0.0270 0.0762 ∗∗ 0.0013 -0.0071 (0.0017) (0.0014) (0.0372) (0.0311) (0.0363) (0.0477) spei12 n× sector = TransportStorage 0.0044 0.0052 0.1143 0.1629 0.2152 ∗∗∗ 0.2088 ∗∗∗ (0.0030) (0.0034) (0.0930) (0.1094) (0.0717) (0.0690) spei12 n× sector = FoodAccomodation 0.0013 0.0020 -0.0134 0.0348 -0.0506 -0.0616 (0.0017) (0.0022) (0.0530) (0.0707) (0.0531) (0.0640) spei12 n× sector = Communication 0.0060 ∗∗∗ 0.0067 ∗∗∗ 0.1404 ∗∗∗ 0.1895 ∗∗∗ 0.2731 ∗∗∗ 0.2596 ∗∗∗ (0.0014) (0.0017) (0.0330) (0.0463) (0.0348) (0.0399) spei12 n× sector = FinancialInsuranceHealth 0.0099 ∗∗∗ 0.0106 ∗∗∗ 0.2506 ∗∗∗ 0.2977 ∗∗∗ 0.5967 ∗∗∗ 0.5817 ∗∗∗ (0.0019) (0.0023) (0.0432) (0.0606) (0.0674) (0.0645) spei12 n× sector = RealEstate 0.0020 0.0027 0.0109 0.0566 0.0801 0.0644 (0.0027) (0.0026) (0.0703) (0.0720) (0.0737) (0.0787) spei12 n× sector = ScienceTech 0.0044 ∗∗∗ 0.0050 ∗∗∗ 0.0536 0.0988 0.1777 ∗∗∗ 0.1632 ∗∗∗ (0.0013) (0.0017) (0.0520) (0.0698) (0.0489) (0.0560) spei12 n× sector = Administrative -0.0072 ∗ -0.0066 -0.0864 -0.0404 -0.0965 -0.1086 (0.0040) (0.0045) (0.1266) (0.1420) (0.0906) (0.0943) spei12 n× sector = PublicAdmin -0.0008 0.0002 0.0980 0.1543 -0.0098 -0.0030 (0.0029) (0.0031) (0.1047) (0.1128) (0.0659) (0.0755) spei12 n× sector = Education 0.0016 0.0023 0.0051 0.0525 0.0441 0.0322 (0.0015) (0.0020) (0.0408) (0.0477) (0.0550) (0.0644) spei12 n× sector = HealthSocial 0.0081 ∗∗ 0.0088 ∗∗ 0.1256 0.1732 0.3967 ∗∗∗ 0.3833 ∗∗∗ (0.0035) (0.0039) (0.1129) (0.1310) (0.1247) (0.1228) spei12 n× sector = Arts -0.0028 -0.0021 -0.1081 -0.0575 -0.2504 ∗∗ -0.2641 ∗∗ (0.0032) (0.0030) (0.1047) (0.0992) (0.1170) (0.1283) spei12 n× sector = OtherService 0.0048 ∗∗ 0.0056 ∗∗ 0.1086 ∗∗ 0.1570 ∗∗ 0.1473 ∗∗∗ 0.1384 ∗∗ (0.0023) (0.0027) (0.0547) (0.0687) (0.0427) (0.0554) spei12 n× sector = DomesticService -0.0365 -0.0357 -0.7556 -0.7066 -3.105 ∗∗ -3.100 ∗∗ (0.0332) (0.0327) (1.155) (1.141) (1.295) (1.300) spei12 n× sector = InternationalBodies -0.0096 -0.0088 0.3353 0.3793 -0.1247 -0.1214 (0.0123) (0.0127) (0.3478) (0.3628) (0.6540) (0.6590) Fixed-effects sector Yes Yes Yes Yes Yes Yes amc-month Yes Yes Yes Yes Yes Yes month-year Yes Yes Yes month-year-region Yes Yes Yes Fit statistics Observations 61,182,576 61,182,576 61,182,576 61,182,576 61,182,576 61,182,576 R 2 0.04498 0.04509 0.02889 0.02903 0.04206 0.04218 Within R 2 0.00018 0.00018 7.78× 10 − 5 8.48× 10 − 5 0.00023 0.00022 Clustered (amc) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 41 Table 2.2: Droughts by Sector Dependent Variables: unemp months unemployed months to employment Model: (1) (2) (3) (4) (5) (6) Variables spei12 n× sector treat any = 0 -0.0022 -0.0013 -0.0109 0.0396 0.0284 0.0181 (0.0031) (0.0036) (0.0861) (0.1032) (0.0713) (0.0728) spei12 n× sector treat any = 1 0.0054 ∗∗∗ 0.0063 ∗∗∗ 0.1201 ∗∗∗ 0.1727 ∗∗∗ 0.0927 ∗∗∗ 0.0859 ∗∗ (0.0017) (0.0015) (0.0305) (0.0257) (0.0284) (0.0412) Fixed-effects sector treat any Yes Yes Yes Yes Yes Yes amc-month Yes Yes Yes Yes Yes Yes month-year Yes Yes Yes month-year-region Yes Yes Yes Fit statistics Observations 61,182,576 61,182,576 61,182,576 61,182,576 61,182,576 61,182,576 R 2 0.02876 0.02889 0.02493 0.02508 0.03771 0.03783 Within R 2 8.67× 10 − 5 8.88× 10 − 5 3.63× 10 − 5 4.34× 10 − 5 1.6× 10 − 5 1.12× 10 − 5 Clustered (amc) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Table 2.3: Droughts by Sector - High/Low Dependent Variables: unemp months unemployed months to employment Model: (1) (2) (3) (4) (5) (6) Variables spei12 n× sector treat HL = C -0.0025 -0.0017 -0.0169 0.0321 0.0228 0.0110 (0.0030) (0.0034) (0.0851) (0.1022) (0.0699) (0.0708) spei12 n× sector treat HL = H 0.0220 ∗∗∗ 0.0229 ∗∗∗ 0.4319 ∗∗∗ 0.4827 ∗∗∗ 0.3798 ∗∗∗ 0.3747 ∗∗∗ (0.0021) (0.0021) (0.0331) (0.0333) (0.0344) (0.0417) spei12 n× sector treat HL = L 0.0011 0.0020 0.0415 0.0926 ∗∗∗ 0.0204 0.0114 (0.0016) (0.0014) (0.0314) (0.0279) (0.0292) (0.0407) Fixed-effects sector treat HL Yes Yes Yes Yes Yes Yes amc-month Yes Yes Yes Yes Yes Yes month-year Yes Yes Yes month-year-region Yes Yes Yes Fit statistics Observations 61,182,576 61,182,576 61,182,576 61,182,576 61,182,576 61,182,576 R 2 0.02913 0.02926 0.02513 0.02528 0.03788 0.03801 Within R 2 0.00030 0.00031 0.00013 0.00013 9.24× 10 − 5 8.9× 10 − 5 Clustered (amc) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 42 2.4 Conclusion In this proposed paper, we ask and attempt to answer the question of whether labor mar- ket disparities seen during economic downturns are also evidenced from the result of environment-related declines in activity. Using a fixed-effects approach with variation in droughts both temporally and spatially we intend to estimate the impact on the hiring and firing of workers by race and gender. Afterwards, we hope to delve into the mechanisms, such as the shift to different sectors, and if sector plurality negatively impacts the ability to adapt therein. Our past experience using RAIS and intimate knowledge of the dataset makes this a feasible endeavor. With adequate time we plan to present further results and continue to examine the economic channels therein. 43 Chapter 3 Religious Figures and their Impact on Non-Religious Decisions, for the Religious and ”A”-religious: An Experiment on Environmental Donations 3.1 Introduction The ability of religious figures to influence the devout on matters of religion is expected. On perceived non-religious matters, probable. For the rest, previous literature suggests causal effects to be weak or non-existent. One would be led to believe that the impact of religious figures on individuals is non-malevolent, conditional on a benign message; the religious pay attention and others simply ignore. This paper studies the heterogeneous impact of religious figures by religiosity in an ex- perimental setting. Using a real measure of preferences I study the question with regards to environmental concerns. Participants were automatically enrolled in a$100 prize draw for survey completion, and asked to donate to a well known non-religious environmen- tal charity. Environmental decisions are by nature of higher social benefit than monetary private. Free rider problems are abound; it is perfectly rational even if one cares for the environment to continue driving a highly polluting vehicle in a city choking on smog. A single car has a negligible impact on pollution levels in the city. Thus non-monetary incentives such as social pressure play an important role. 44 I use an online platform to recruit participants who are paid a fixed fee for their time as well as automatic entry into a prize draw. On such a platform there will of course be individuals strictly trying to maximise their earnings. Using the duration of the surveys I attempt to identify such individuals. This adds a further layer to the locality of my average treatment effect. It does not seem prudent to include individuals who either didn’t read the question or have a script answering for them in the sample. As expected, given such individuals answering randomly were themselves randomly assigned between treatment groups, removing them increase the magnitude of the treatment effect. Such decisions to work with a subsample was pre-specified in my preanalysis plan and the paper includes robustness checks pertaining to this. There were three treatment arms and one control. All arms included quotes from a religious figure, juxtaposed by either Pope Francis, Pope Benedict, Ayatollah Khamenei, or a ”world leader”. The control was not simply the quote without attribution. This was to avoid the concern that I would be unable to conclude if any difference was due to the religiosity that a religious leader brings to the statement, or the fact that the statement simply had text surrounding it. The treatment arms aim to differentiate between effects drawn from specific quotes, and specific leaders that may have other political considera- tions at play. The heterogeneity analysis is split into three groups which behaved similarly: ”Very Religious”, ”Somewhat Religious or Non-Religious”, and ”Atheist or Agnostic”, as self-identified. The Oxford English dictionary defines areligious as ”Not influenced by, concerned about, or practising religion; having no religious beliefs.” While not entirely accurate, for the sake of brevity I will be using ”A”-religious to refer to atheists and agnos- tics, and ”Not Religious” for the somewhat religious and non-religious individuals. One point to note: non-religious individuals behaved more like somewhat than the areligious, even though from the standpoint of actions the aforementioned group should be more similar to them. This seems to imply that it is the belief rather than actions undertaken due to the belief that is key in determining the reaction to the treatment. 45 For both popes I find large, significant, and positive effects for the very religious, and a small, positive but insignificant effect for the ayatollah. There are no significant effects for the not religious. As for the areligious, there is significant backlash, similar in magnitude between popes and the ayatollah. An important caveat to keep in mind for this paper; given religiosity is not randomly assigned, such heterogeneous treatment effects are not casually identified. However suggestive evidence that these results do not hold as well for heterogeneity across religion or political leaning do tend to favour this explanation. Taken together this suggests that moral imperatives from religious figures, even benign ones, should be treated with caution. While they may affect the religious as intended, poor targeting or indiscriminate messaging may result in backlash from the areligious and an overall negative effect for an otherwise well meaning statement. Indeed, as an example, religion is often used as a means to elicit donations for the less well off, whether by the simple collection of donations by a religious charity such as The Salvation Army or more explicit links. In the UK Muslim Aid has in the past taken out public adverts at bus stations during Ramadan, a time when Muslims increase their charitable giving. While this likely increases charitable contributions by Muslims, if it dissuades those it was not aimed at, and one as a planner incorporates inequality in their social welfare function, this may lead to a situation where overall contributions to the less well off declines. A similar situations may arise in awareness and charitable drives regarding the environment, where market failures are already prevalent. The paper proceeds as follows. Section 3.1 continues on with a literature review. Sec- tion 3.2 discusses the experiential design and Section 3.3 the main results. Section 3.4 expounds upon the results through robustness checks and Section 3.5 concludes. 46 A Study Motivation & Contributions Bursztyn et al (UChicago) ([48]) ran a field experiment in which text messages were sent to those who owe debt to an Islamic bank in Indonesia. They found they are able to in- crease debt repayment through moral appeals. To very briefly summarise their findings: moral appeals in the form of a text message including a religious quote reduces delin- quency, while reminders, religious priming, message novelty and customer view on the bank’s determination to collect debt are ruled out as potential channels. The main signif- icant effect does not differ when the source of the quote is excluded - moral appeal rather than the religious nature of the message seems to be the key here. While there was a stronger effect for the very religious, the authors argue that the main effect is driven by moral appeals. I differ in this regard by testing effects by the source of the appeal. i.e. Does it matter who says the statement? I also build upon the literature in my choice of locality. My sample is from a more secu- lar US population that also did not self-select into the equivalent of an Islamic bank. That is, I am able to access a wide spectrum of religiosity which may be why my results differed, in particular regarding the backlash effect from the areligious. Furthermore the source of moral appeal is from a different religion for most participants in my case - Catholics and Muslim Shia leaders to a predominantly non-Catholic/Shia audience instead of an Islamic appeal to Muslims. I am able to see if religious figures can influence beyond their own fol- lowers. My research relates to a few areas of the literature, with non-monetary incentives being the most relevant ([49], [50], [51]). The outcome in my study is donations to an environ- mental charity. Appealing to morality is instrumental in this sector due to the nature of negative externalities. Behavioural and environmental economics are other related sets of literature, Hunt Allcott is one individual who has has written extensively on. Allcott ([52]) demonstrates that social norms (comparing consumption to that of neighbours) can reduce energy consumption. Private costs typically outweigh the tangible private benefit 47 when it comes to care for the environment so it is typical to appeal to other dimensions of a person’s utility function. I also add to the smaller set of literature on religion and economic behaviour ([53], [54], [55]). In particular I study the effect on an outcome that is not religious in nature. In this framework I search for heterogeneous effects of moral appeals across a broader swathe of the population than one religion and its adherents. Field work typically encom- passes a single religion and also makes it difficult to disentangle moral appeals from social concerns, especially since many acts are communal. Shifting to an online survey which is almost completely anonymous lets me tackle both of these issues together. 3.2 Experimental Design The study design is a simple randomised controlled trial. The randomisation is in the information presented to individuals during an online survey prior to answering the ques- tion related to the outcome variables of interest. Individuals were paid to complete ”tasks”. They received payment for completion while also being entered into a prize draw with the option to select how much of their potential winnings to automatically donate to an envi- ronmental charity. This allowed me to elicit a real measure of preferences in addition to pure opinion and hypothetical answers. The latter form of questions were simple ratings on a scale of 1-10 of agreement. I do not have the sample size to measure any differences if there were any, and my primary reason for including them was to try and stop partic- ipants second-guessing the researcher’s intention while not jarring the individual with a brief demographic survey to only result in a single question related to the research. The prize draw was for$100. Too small a figure and individuals may not consider the question seriously. With the presence of very large lottery winnings in real life, too high a figure may have caused participants to draw parallels and assume a very low win rate - again perhaps not giving the question too much thought. My selection of a ”reasonable” 48 amount is based on this. For reference, individuals were paid $0.45-$0.50 for successful survey completion. A Sample Population and Randomisation It is important to note that Amazon Mturk users are not representative of citizens of a country as a whole. They tend to be middle aged or otherwise low paid workers or stu- dents, using the platform to supplement their income. External validity will rest on the assumption that any treatment effect is orthogonal to unobserved and observed traits that differ between Amazon Mturk users that select into the survey and the superset of this group (i.e. the rest of the population that we are seeking external validity for). The question on the donation of potential winnings was as follows: You will be entered into a prize draw for $100. You may donate some of the winnings to Friends of the Earth U.S., a charity focused on protecting the en- vironment. This will be automatic if you win. How much would you like to specify for donation? Individuals were randomised by computer, stratifying by religiosity for additional power. Randomisation was on an individual level so identification is simple and robust standard errors sufficient. In total I had 1582 responses from across the US. I also tried the same ex- periment in Portugal and Italy (areas with a high catholic population) but I only received 1 and 29 responses respectively. These were removed from the sample. B Treatment Arms Note that Francis, Ayatollah and Control all use the same quote by Pope Francis, while Benedict uses a quote by Pope Benedict. As such, differences between the first three cannot be attributed to a difference in quotes used, only the influential figure involved, while the fourth arm listed gives some measure of confidence that any effect was not particular to the exact quote chosen. The statements presented to each treatment arm is as follows. 49 B.1 Control (Generic Leader) ”Climate change is a global problem with grave implications: environmental, social, economic, political and for the distribution of goods. It represents one of the principal challenges facing humanity in our day.” Many world leaders have noted such concerns and advocated for pro-environmental policies, including on the monitoring of the environment and climate change. They have highlighted the imperative for humans to act in a way that protects the Earth’s resources. B.2 Pope Francis Pope Francis (the current pope), in his second encyclical ’on care for our com- mon home’, describes the dangers facing the environment. To quote: ”Climate change is a global problem with grave implications: environmental, social, economic, political and for the distribution of goods. It represents one of the principal challenges facing humanity in our day.” He further discusses the imperative for humans to act in a way that protects God’s creation. B.3 Shia Arm ”Climate change is a global problem with grave implications: environmental, social, economic, political and for the distribution of goods. It represents one of the principal challenges facing humanity in our day.” Iran’s supreme leader has noted such concerns and advocated for pro-environmental policies, including on the monitoring of the environment and climate change. He highlights the imperative for humans to act in a way that protects God’s creation. 50 B.4 Pope Benedict Pope Benedict (the previous pope), in his letter to the Ecumenical Patriarch of Constantinople, describes the dangers facing the environment. To quote: “Preservation of the environment, promotion of sustainable development and particular attention to climate change are matters of grave concern for the entire human family.” He further discusses the imperative for humans to act in a way that protects God’s creation. C Balance and Stratification Table Table 3.1 tests for join orthogonality by covaraite across all four groups. The ran- domisation appears successful by the lack of joint significance. There is a significant dif- ference over Islam (Sunni/Other), however with only 8 observations in the category and the number of tests involved I do not consider this an issue in my experiment. Duration in absolute terms does vary greatly across arms which can be attributed to a few extreme observations. 51 Table 3.1: Balance Table on Covariates (1) (2) (3) (4) F-test Control Benedict Francis Khamenei for joint Variable Mean/SE Mean/SE Mean/SE Mean/SE orthogonality Duration (secs) 581.291 (234.278) 208.641 (30.478) 226.596 (38.301) 430.856 (182.828) 0.277 Age 41.003 (0.683) 41.466 (0.646) 42.371 (0.624) 41.717 (0.639) 0.508 Male 0.529 (0.025) 0.527 (0.025) 0.508 (0.025) 0.548 (0.025) 0.729 Agnostic 0.177 (0.019) 0.175 (0.019) 0.177 (0.019) 0.177 (0.019) 1.000 Atheist 0.149 (0.018) 0.154 (0.018) 0.154 (0.018) 0.157 (0.018) 0.994 Non-Religious 0.159 (0.018) 0.167 (0.019) 0.167 (0.019) 0.167 (0.019) 0.990 Somewhat Religious 0.352 (0.024) 0.347 (0.024) 0.338 (0.024) 0.338 (0.024) 0.972 Very Religious 0.162 (0.019) 0.157 (0.018) 0.164 (0.019) 0.162 (0.019) 0.994 Buddhism 0.015 (0.006) 0.020 (0.007) 0.020 (0.007) 0.020 (0.007) 0.927 Christianity (Catholic) 0.210 (0.021) 0.208 (0.020) 0.232 (0.021) 0.207 (0.020) 0.803 Christianity (Orthodox) 0.066 (0.012) 0.051 (0.011) 0.056 (0.012) 0.058 (0.012) 0.836 Christianity (Other) 0.256 (0.022) 0.268 (0.022) 0.242 (0.022) 0.253 (0.022) 0.871 Hinduism 0.013 (0.006) 0.013 (0.006) 0.008 (0.004) 0.003 (0.003) 0.187 Islam (Shia) 0.008 (0.004) 0.003 (0.003) 0.005 (0.004) 0.000 (0.000) 0.111 Islam (Sunni/Other) 0.000 (0.000) 0.005 (0.004) 0.010 (0.005) 0.005 (0.004) 0.046** Judaism 0.020 (0.007) 0.043 (0.010) 0.018 (0.007) 0.028 (0.008) 0.183 No Religion 0.413 (0.025) 0.390 (0.025) 0.409 (0.025) 0.427 (0.025) 0.768 N 395 395 396 396 Notes: F tests display their p values The value displayed for t-tests are the differences in the means across the groups. Standard errors are robust. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. 52 The sample was stratified by religiosity - each new individual was randomly placed in the treatment arm amongst those with the lowest number of observations to maintain balance. Table Table 3.2 shows the samples split by religiosity and treatment. With an almost equal number of each religiosity per arm stratification appears successful. The difference sometimes being greater than one is due to multiple batches of the survey being run. A preview of the results can be found in Table 3.3. Donations decrease from the control for the areligious, do not change much for the not religious, and increase for the very religious, except in the case of Khamenei. Section 3.3 incorporates this into a regression format with standard errors and controls to help derive a firmer conclusion. Table 3.2: Number of Participants by Strata and Religiosity Religiosity Treatment arm Agnostic Atheist Non-Religious Somewhat Religious Very Religious Total Benedict 69 61 66 137 62 395 Control 70 59 63 139 64 395 Francis 70 61 66 134 65 396 Khamenei 70 62 66 134 64 396 Total 279 243 261 544 255 1582 Table 3.3: Average donations conditional on winning the$100 prize draw Religiosity Treatment arm Agnostic Atheist Non-Religious Somewhat Religious Very Religious Total Benedict 18.12 15.49 14.58 22.51 29.61 20.45 Control 25.83 19.51 17.87 24.23 23.17 22.62 Francis 15.31 16.28 18.77 25.62 28.65 21.71 Khamenei 14.77 15.05 17.89 26.62 21.41 20.42 Total 18.51 16.55 17.27 24.73 25.69 21.30 D Regression Specifications Randomisation was on an individual level thus the coefficients on the treatment arms are causal. The regression is as follows, 53 Y i =α+ ∑ t β t T t,i +γ ′ X i +ε i (3.1) whereY i is the dependent variable - a measure of donation commitment, T t,i is an in- dicator for individual i being in treatment group t, and X i is a set of controls included in some specifications. The excluded category captured by α is the control group for all regressions. Randomisation was stratified by religiosity. Individual was randomly assigned to an arm so there is no clustering required - all standard errors presented are robust. Given the number of treatment arms and religiosity strata (even after combining the five into three), a single regression with all the permutations is unwieldy. Thus I run separate regression; for the very religious, the not religious, and the areligious. As prespecified I also reduce the sample to those that completed the survey in a rea- sonable amount of time. My preferred specification is 60-240 seconds. The survey was expected to take 2 minutes and participants were told as such. With a median completion time of 114 seconds this seems to have been a good approximation. Section 3.4 delves into robustness checks for a range of durations to show that the results are not dependent on this particular interval. Individuals were randomly assigned to treatment arms. A naive prediction with only this information would be that truncating the sample only increases standard errors. How- ever, individuals can be split into two mutually exclusive and exhaustive groups: those that read before answering, and those that did not. As they are paid to complete surveys there is an incentive to rush through and answer randomly so one may move onto the next paid task. This was one reason for keeping the survey very short. Those that did not read the information provided should behave the same irrespective of treatment arm, bringing the coefficients closer. Once these individuals are removed as best as possible through the survey duration proxy the magnitude of difference from the control does indeed increases. 54 I excluded individuals who took less than 60 seconds to finish a two minute survey. Scripts answering randomly is a possibility as well as manual skipping without reading. As for those that took longer than 240 seconds the reasoning is more platform-specific. Taking a long time to finish the survey is potentially due to them engaged in other activ- ities, such as collecting a bunch of surveys to complete, rather than simple reading com- prehension. These individuals are likely to be the kind who are focused purely on speed of completion and again are likely to skip a paragraph if not questioned directly about it. As mentioned, excluding these individuals results in a greater dispersion of treatment co- efficients, supporting my belief that I have excluded those answering randomly who were causing the difference to be underestimated. What this all means is my preferred estimates are Local Average Treatment Effects. They are the effects for survey participants on a particular platform, who care enough to read all the information provided in surveys, proxied by their survey completion time. Nonetheless they present interesting findings which one may wish to test in the future over a more natural setting. I also present regressions using an indicator for above/below mean donations, and graphs of different cutoffs around it. My main results remain significant for the most part, however the loss of power does push some out of the much coveted 5% significance level. 3.3 Main Results A Full Sample Tables I begin the analysis by providing regression estimates of the full sample without hetero- geneous effects. Table 3.4 has the results. Column (1) shows the simplest regression of treatment dummies on donations committed from winnings. All arms give insignificant results. Since the assignment of treatment status was randomised within strata the coeffi- cients are causal independent of controls. As such, the remaining Columns (2)-(4) add in 55 control variables for theoretically smaller standard errors and an increase in power. Col- umn (2) adds in Religiosity which changes the estimates very little - unsurprising given this was what treatment was stratified on. Column (3) supplements with age, sex and religion dummies. Column (4) finally adds politics and state dummies. While non linear in the decrease (especially Column (3)-(4), likely due to the high number of additional state dummies), the general trend is indeed to lower standard er- rors. The magnitude in change, however, is not nearly sufficient to detect any effects that may be present. Ending the analysis here would have resulted in a null results: that in the population studied, there is no effect of religious figures on preferences for the environ- ment as proxied through potential donations. B Heterogeneous Treatment Effects by Religiosity Table Table 3.5 presents my main set of results. Regression were run on subsamples split by self-identified religiosity. Columns (1)-(3) are for the full sample while Columns (4)- (6) takes a the subset that completed the 2 minute survey in one to four minutes, which is my preferred specification. All regressions included the full set of controls. Starting with Column (1), donations by the very religious increased by $13.28 when attribution was given to Pope Benedict over a generic world leader. Pope Francis saw a $8.46 increase in donations but this was not statistically significant. Khamenei’s point estimate was much smaller and also suggests no effect. As briefly touched on this ATE is made up of two ATEs; of those that read the question, and those that did not. ATE f =ATE r α r +ATE s (1− α r ) (3.2) Where ATE f is the average treatment effect of the full sample, ATE r the average for those that read the quote. ATE s for the skippers, andα r the proportion that read the quote. ATE s is zero - individual were randomly assigned and if they didn’t read the statement then they have all ”seen” the same survey. Thus the equation simplifies to 56 Table 3.4: Treatment Effect: Donation of Potential $100 Lottery (1) (2) (3) (4) donate donate donate donate Pope Benedict -2.175 -2.088 -1.740 -1.763 (1.788) (1.779) (1.747) (1.751) Pope Francis -0.908 -0.822 -0.977 -0.647 (1.877) (1.868) (1.817) (1.798) Ayatollah -2.206 -2.096 -1.568 -2.149 (1.852) (1.845) (1.835) (1.847) Control (Base) 22.62 ∗∗∗ 22.55 ∗∗∗ 17.71 ∗∗∗ 17.42 ∗∗∗ (1.299) (1.303) (2.543) (2.560) Age No No Yes Yes Sex No No Yes Yes Religion No No Yes Yes Religiosity No Yes Yes Yes Politics No No No Yes State No No No Yes N 1582 1582 1582 1581 R 2 0.001 0.023 0.062 0.107 Standard errors in parentheses Note: Standard errors are robust. Note that the treatment randomisation is stratified by religiosity. Column 1 is a simple OLS regression of treatment indicators on amount selected for donation. Column 2 adds in religiosity, while Column 3 adds in the full set of controls. ∗ p<0.05, ∗∗ p<0.01, ∗∗∗ p<0.001 57 ATE f =ATE r α r (3.3) Note that I have \ ATE f but only a proxy of α r from reducing the sample by duration. Regardless, 0 < α r < 1 so point the estimates in the reduced equations should naturally increase. The fact that it does seems to suggest that I have indeed removed a sizeable number of ”skippers” from my sample and that [ ATE s <ATE r . To reiterate, this is simply suggestive - I cannot say E( [ ATE s ) = 0, which would imply that my estimate of α r was correct. Further, it is very possible that α r varies by religiosity as well, though I would be reluctant to say such a relationship was causal and not a function of unobserved and excluded corvariates. With these caveats in mind, Column (4) provides the estimate for the individuals who likely read the question as implied from the duration of their survey response. Both Pope Benedict and Pope Francis have large and significant effects of $17.18 and$15.73 above the control mean. The Ayatollah has a larger effect at $3.66 but still remains insignificant. This I interpret as the very religious increasing their preference for environmental welfare when targeted by an individual they generally feel amiable towards (US population towards popes). Such effects are effectively cancelled out if they dislike the religious figure, but there is no backlash effect where donations are reduced. Columns (2) and (5) look at the not religious: the self-identified ”Somewhat Reli- gious” or ”Non-Religious”. No treatment arms produce significant results, in the full or reduced sample. The fact that Non-Religious are more similar in treatment effect to the Somewhat Religious than Atheist or Agnostics seems to suggest it is the beliefs held rather than actions acted upon said belief that determines the effect of this experiment. The areligious, the ”Atheists” and ”Agnostics” are perhaps the most interesting group to look at. For all religious figures they reduce donations by between$7.23-$9.39 in the full sample and$9.33-$10.41 in the reduced sample. This seems to suggest religiosity should be seen on a scale similar to ”left vs right”, rather than ”0 to 1”, colloquially speaking. That 58 is, religion, or in this case the sayings of religious figures are not something that simply results in no effect on the wrong target, but backlash to a position this individual would otherwise hold themselves. One final note of interest - it was the areligious rather than the very religious for whom there was a backlash. This is in spite of the religious typically siding with the Republicans over the Democrats and a higher general feeling of animosity one may expect to Iran from such supporters. This further supports my hypothesis that it is the religiosity of the figure that is of primary importance and can trump politics to some degree. 59 Table 3.5: Heterogeneous Treatment Effects: Donation of Potential $100 Lottery Full Sample Realisitc Durations (1) (2) (3) (4) (5) (6) Very Religious Some/Non Religious Atheist/Agnostic Very Religious Some/Non Religious Atheist/Agnostic Pope Benedict 13.28 ∗ -1.898 -7.277 ∗ 17.18 ∗ -1.846 -10.41 ∗∗ (6.548) (2.460) (3.031) (7.458) (2.677) (3.567) Pope Francis 8.464 1.831 -7.139 ∗ 15.73 ∗ 0.626 -9.325 ∗∗ (5.991) (2.526) (3.079) (6.942) (2.779) (3.406) Ayatollah 0.711 1.332 -9.387 ∗∗ 3.663 -0.298 -9.719 ∗∗ (5.959) (2.639) (3.100) (6.869) (2.781) (3.646) Control (Base) 8.542 18.28 ∗∗∗ 18.26 ∗∗∗ 5.936 14.26 ∗∗∗ 18.94 ∗∗∗ (8.120) (3.485) (4.563) (9.626) (4.017) (5.432) Controls Yes Yes Yes Yes Yes Yes N 249 799 513 191 640 385 R 2 0.261 0.130 0.159 0.311 0.157 0.175 Standard errors in parentheses Note: Standard errors are robust. Note that the treatment randomisation is stratified by religiosity. Cols (1) - (3) pertain to the full sample while (4)-(6) truncate to 60-240 second survey durations. Controls include age, sex, religion, politics and state ∗ p<0.05, ∗∗ p<0.01, ∗∗∗ p<0.001 60 C Alternative Heterogeneity Regressions Is the heterogeneity due to differences in political backgrounds? Table 3.7 provides es- timates of regression subsampled by politics. ”Strongly Republican” and ”Republican Leaning” individuals were classified as Republican, similarly for Democrats, and those that selected ”Independent/Other” in the final category. There are only significant effect for Democrats in two of the three treatment arms. The magnitude of the significant coeffi- cients are generally smaller than when split by religiosity. This I would argue supports my primary specification to split heterogeneity by religiosity. That political leanings should be a weaker matter that determines how one reacts to religious leaders compared to reli- giosity is also a valid point to make. A second, more nuanced question is as follows. Are the results being driven by reli- gion rather than religiosity? That is, is the effect coming from Catholics highly motivated environmentally due to attribution to the pope who is from their own religion? While there were nine religions (no religion inclusive) selected, only No Religion, Christianty (Catholic), and Christianity (Other) had a large enough number of observations to study. Christianity (Other) refers to non-Catholics and non-Orthodox Christians. For reference, 91 individuals identifed as Orthodox. Table 3.7 provides estimates by religion. Subsampling Catholics, there is no effect in any treatment as shown in Columns (1) and (3), not even with the popes. The same holds true for Christian (Other) in Columns (2) and (4). All the coefficients on the treatment are in addition small and with high standard errors relative to the magnitude. This is all suggestive evidence that is is indeed religiosity and not adherence to the religion of the religious figure that is driving the results. Religious figures, this study suggests, may have a reach beyond their own adherents. Finally, those who selected ”No Religion” had some significant negative effects. This is expected as agnostics and atheists are the natural respondents here. The coefficients were 61 less significant and smaller compared to the areligious subsample in Table 3.5. It seems likely that at least some who identified as non-religious for religiosity selected no religion for religion, which would explain these sets of results. 62 Table 3.6: Heterogeneous Treatment Effects: Donation of Potential $100 Lottery by Politics Full Sample Realisitc Durations (1) (2) (3) (4) (5) (6) Republican Democrat Independent Republican Democrat Independent Pope Benedict 4.073 -6.124 ∗ -5.545 5.463 -8.342 ∗∗ -5.205 (3.592) (2.458) (4.653) (3.777) (2.863) (5.238) Pope Francis 3.266 -4.043 -4.253 3.805 -5.250 -5.368 (3.419) (2.577) (4.303) (3.505) (2.943) (4.818) Ayatollah 3.970 -6.601 ∗ -2.676 2.330 -6.478 ∗ -5.013 (3.552) (2.647) (4.408) (3.634) (3.014) (5.150) Control (Base) 17.74 ∗∗∗ 20.86 ∗∗∗ 15.39 ∗ 14.83 ∗ 20.50 ∗∗∗ 18.04 ∗∗ (5.347) (3.537) (6.087) (6.010) (4.152) (6.750) Controls Yes Yes Yes Yes Yes Yes N 460 825 271 397 640 213 R 2 0.198 0.140 0.209 0.215 0.160 0.228 Standard errors in parentheses Note: Standard errors are robust. Cols (1) - (3) pertain to the full sample while (4)-(6) truncate to 60-240 second survey durations. Controls include age, sex, religion, religiosity and state ∗ p<0.05, ∗∗ p<0.01, ∗∗∗ p<0.001 63 Table 3.7: Heterogeneous Treatment Effects: Donation of Potential $100 Lottery by Religion Full Sample Realisitc Durations (1) (2) (3) (4) (5) (6) Catholic Christian No Religion Catholic Christian No Religion Pope Benedict -2.385 0.658 -6.502 ∗ -2.431 3.387 -9.417 ∗∗ (4.254) (3.783) (2.693) (5.279) (3.903) (3.031) Pope Francis 2.538 -2.844 -4.672 1.022 0.751 -7.390 ∗ (4.663) (3.642) (2.787) (5.536) (3.818) (3.011) Ayatollah -2.645 0.245 -4.847 -7.660 3.537 -5.589 (4.881) (3.427) (2.846) (5.502) (3.486) (3.148) Control (Base) 23.41 ∗∗∗ 13.65 ∗ 15.48 ∗∗∗ 21.87 ∗∗ 6.956 14.67 ∗∗ (5.374) (5.567) (4.058) (7.175) (5.479) (4.530) Controls Yes Yes Yes Yes Yes Yes N 334 396 641 241 326 496 R 2 0.215 0.169 0.095 0.270 0.170 0.130 Standard errors in parentheses Note: Standard errors are robust. Cols (1) - (3) pertain to the full sample while (4)-(6) truncate to 60-240 second survey durations. Controls include age, sex, politics, religiosity and state ∗ p<0.05, ∗∗ p<0.01, ∗∗∗ p<0.001 64 3.4 Robustness Checks A Indicator for Donations Above the Mean Table 3.8 provides estimates of heterogeneous effects where the dependent variable is an indicator for being above the mean of 21. The results remain significant in this linear probability model - the very religious are between 23.5-24.6 p.p. more likely to donate above the mean when presented with attribution to a pope. The areligious are between 17.3-22.5 p.p. less likely to donate. There is no significant effect for the not religious. Table 3.8: Heterogeneous Treatment Effects: Donations Above the Mean of 21 (1) (2) (3) Very Religious Some/Non Religious Atheist/Agnostic Pope Benedict 0.235 ∗ 0.00253 -0.173 ∗ (2.16) (0.05) (-2.57) Pope Francis 0.246 ∗ 0.0215 -0.175 ∗∗ (2.25) (0.39) (-2.66) Ayatollah 0.0676 -0.0527 -0.225 ∗∗∗ (0.59) (-0.98) (-3.36) Control (Base) -0.00310 0.205 ∗∗ 0.291 ∗∗ (-0.02) (2.73) (3.09) Controls Yes Yes Yes N 191 640 385 R 2 0.340 0.140 0.223 t statistics in parentheses ∗ p<0.05, ∗∗ p<0.01, ∗∗∗ p<0.001 B Survey Durations Appendix B.B.2 graphs the full set of regression treatment coefficients and their 95% con- fidence intervals by treatment arm and religiosity for different survey durations. Figure 3.1 provides a sample of such graphs. Graphs are either fixed with a lower bound of 60 65 seconds and mapped across an upper bound between 180-300 seconds, or fixed at 240 sec- onds for the upper bound and mapped across 40-80 on the lower bound. My preferred specification is 60-240 seconds - both reasonable given a 2 minute survey and keeping the bulk of the sample. The results by and large hold for different interval sizes. −10 0 10 20 30 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Benedict − Very Relgious −20 −15 −10 −5 0 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Benedict − Atheist/Agnostic 0 20 40 60 80 100 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Benedict − Very Relgious −30 −20 −10 0 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Benedict − Atheist/Agnostic Figure 3.1: Coefficients and 95% confidence intervals as the upper and lower bounds vary for Benedict treatment arm, very and areligious. There are two competing effects as the boundaries change. Decreasing the size of the window for durations included in regressions helps to remove the individuals who likely did not answer in good faith (i.e. at the very least read the information) and so the magni- tude of the effect increases. On the other hand, smaller sample sizes may increase standard errors. This is especially prominent for the very religious - the sample size was smallest here, and as the window shrinks estimates begin to become imprecise. Together with the 66 loss of degrees of freedom from including all controls in the specification and the estimates begin to become imprecise if taken too far. 3.5 Conclusion In this paper I report the results of an online survey to determine the impact of religious figures on a non-religious topic, namely the environment. I measure the effect by compar- ing the amount individuals select to donate from their potential prize draw winnings. I find that there is no overall effect in my sample, but once the prespecified dimension of hetrogenity (religiosity) is considered the underlying effects becomes apparant. The very religious increase donations in two of the three treatment groups - only the Ayatollah has no significant effect. The somewhat and non-religious are not impacted, while atheists and agnostics significantly reduce their donations. Heterogeneity across other covariates are less convincing, suggesting that it is indeed the religiosity of the individual that is driv- ing my results. The results are not driven by catholics responding to the popes, or shias responding to the ayatollah, and the backlash also goes across (non)-religious lines, sug- gestive that religious figures may have an influence, negative and positive, beyond their own congregations. An important caveat to note, however. Religiosity is not randomly assigned, only the treatment is - I cannot say the effect of religiosity on reaction to religious figure is causal. Simply the weaker statement that the effect within each religiosity strata is causal. That is, if an atheist were wake up one day and feel very religious or vice versa, my experiment cannot provide an unbiased estimation as the identification is not there. What I do say is, given an individual is an atheist or very religious, they react in different ways to the treatment. 67 My findings are in the context of environmental concerns as related to my other work, but this was a case of picking a non-religious topic that feelings for could easily be mea- sured in the context. I could have equivalently detailed treatment arms with quotes related to healthcare and general fitness and measured the effect on monthly gym membership by offering discounts and seeing how many accept. It is likely these results may generalize to different contexts which remains an avenue for future research. With experiments there is always a question of generalizability. This was a short in- tervention, the treatment only being 15-30 seconds of text, and preferences elicited almost straight after. On the one hand, given the magnitude of the effect one may expect it to be short lived. On the other, repeated messaging may work to reinforce the effect. Indeed, this is commonly used across human society, from politics to religion to the monotonous repetition in studying for exams. My experiment suggests that in certain cases, even of non-religious activities, religious leaders can have a positive, but perhaps more surpris- ingly, negative effect, even with a benign statement. 68 References 1. Christidis, N., McCarthy, M. & Stott, P. A. The increasing likelihood of temperatures above 30 to 40 °C in the United Kingdom. en. Nature Communications 11. Number: 1 Publisher: Nature Publishing Group, 3093. doi:10.1038/s41467-020-16834-0 (2020). 2. Bathiany, S., Dakos, V., Scheffer, M. & Lenton, T. M. Climate models predict increasing temperature variability in poor countries. Science Advances 4. Publisher: American Association for the Advancement of Science, eaar5809. doi:10.1126/sciadv.aar580 9 (2018). 3. Hansen, J., Sato, M., Ruedy, R., Lo, K., Lea, D. W. & Medina-Elizade, M. Global tem- perature change. Proceedings of the National Academy of Sciences 103. Publisher: Pro- ceedings of the National Academy of Sciences, 14288–14293. doi:10.1073/pnas.060 6291103 (2006). 4. Ebi, K. L., Vanos, J., Baldwin, J. W., Bell, J. E., Hondula, D. M., Errett, N. A., et al. Extreme Weather and Climate Change: Population Health and Health System Impli- cations. Annual Review of Public Health 42. eprint: https://doi.org/10.1146/annurev- publhealth-012420-105026, 293–315. doi:10.1146/annurev-publhealth-012420-105 026 (2021). 5. Carleton, T., Jina, A., Delgado, M., Greenstone, M., Houser, T., Hsiang, S., et al. Valuing the Global Mortality Consequences of Climate Change Accounting for Adaptation Costs and Benefits*. The Quarterly Journal of Economics 137, 2037–2105. doi:10.1093/q je/qjac020 (2022). 6. Garc´ ıa-Le´ on, D., Casanueva, A., Standardi, G., Burgstall, A., Flouris, A. D. & Nybo, L. Current and projected regional economic impacts of heatwaves in Europe. en. Nature Communications 12. Number: 1 Publisher: Nature Publishing Group, 5807. doi:10.10 38/s41467-021-26050-z (2021). 7. Karhula, A., Lehti, H. & Erola, J. Intergenerational scars? The long-term effects of parental unemployment during a depression. en. Finnish Journal of Social Research 10, 87–99. doi:10.51815/fjsr.110768 (2017). 8. Corak, M. Death and Divorce: The Long-Term Consequences of Parental Loss on Adolescents. en. Journal of Labor Economics. Publisher: The University of Chicago Press. doi:10.1086/322078 (2001). 9. Gandhi, S., Kahn, M. E., Kochhar, R., Lall, S. & Tandel, V. Adapting to Flood Risk: Evi- dence from a Panel of Global Cities Working Paper. 2022. doi:10.3386/w30137. 69 10. Wen, J. & Burke, M. Lower test scores from wildfire smoke exposure. en. Nature Sus- tainability 5. Number: 11 Publisher: Nature Publishing Group, 947–955. doi:10.1038 /s41893-022-00956-y (2022). 11. Park, R. J., Goodman, J., Hurwitz, M. & Smith, J. Heat and Learning. en. American Economic Journal: Economic Policy 12, 306–339. doi:10.1257/pol.20180612 (2020). 12. Keivabu, R. C. Too hot to study? Gender and SES differences in the effect of temperature on school performance en-us. 2020. doi:10.31235/osf.io/whtf5. 13. Lavy, V., Ebenstein, A. & Roth, S. The Impact of Short Term Exposure to Ambient Air Pollution on Cognitive Performance and Human Capital Formation Working Paper. 2014. doi:10.3386/w20648. 14. Pham, L. & Roach, T. Particulate pollution and learning. en. Economics of Education Review 92, 102344. doi:10.1016/j.econedurev.2022.102344 (2023). 15. Graff Zivin, J., Hsiang, S. M. & Neidell, M. Temperature and Human Capital in the Short and Long Run. Journal of the Association of Environmental and Resource Economists 5. Publisher: The University of Chicago Press, 77–105. doi:10.1086/694177 (2018). 16. Park, R. J., Behrer, A. P. & Goodman, J. Learning is inhibited by heat exposure, both internationally and within the United States. en. Nature Human Behaviour 5. Number: 1 Publisher: Nature Publishing Group, 19–27. doi:10.1038/s41562-020-00959-9 (2021). 17. Li, X. & Patel, P. C. Weather and high-stakes exam performance: Evidence from student- level administrative data in Brazil. en. Economics Letters 199, 109698. doi:10.1016/j .econlet.2020.109698 (2021). 18. Masuda, Y. J., Garg, T., Anggraeni, I., Wolff, N. H., Ebi, K., Game, E. T., et al. Heat ex- posure from tropical deforestation decreases cognitive performance of rural workers: an experimental study. en. Environmental Research Letters 15. Publisher: IOP Publish- ing, 124015. doi:10.1088/1748-9326/abb96c (2020). 19. Maula, H., Hongisto, V., ¨ Ostman, L., Haapakangas, A., Koskela, H. & Hy¨ on¨ a, J. The effect of slightly warm temperature on work performance and comfort in open-plan offices - a laboratory study. eng. Indoor Air 26, 286–297. doi:10.1111/ina.12209 (2016). 20. Mackworth, N. H. Effects of heat on wireless telegraphy operators hearing and record- ing Morse messages. eng. British Journal of Industrial Medicine 3, 143–158 (1946). 21. Bleakley, H. Malaria Eradication in the Americas: A Retrospective Analysis of Child- hood Exposure. en. American Economic Journal: Applied Economics 2, 1–45. doi:10.125 7/app.2.2.1 (2010). 22. Shah, M. & Steinberg, B. M. Drought of Opportunities: Contemporaneous and Long- Term Impacts of Rainfall Shocks on Human Capital. Journal of Political Economy 125. Publisher: The University of Chicago Press, 527–561. doi:10.1086/690828 (2017). 23. Alberto, I. C., Jiao, Y. & Zhang, X. Too hot or too cold to study? The effect of temper- ature on student time allocation. en. Economics of Education Review 84, 102152. doi:10 .1016/j.econedurev.2021.102152 (2021). 70 24. Garg, T., Jagnani, M. & Taraz, V. Temperature and Human Capital in India. Journal of the Association of Environmental and Resource Economists 7. Publisher: The University of Chicago Press, 1113–1150. doi:10.1086/710066 (2020). 25. Melo, A. P. & Suzuki, M. Temperature, effort, and achievement: Evidence from a large-scale standardized exam in Brazil. en, 48. 26. Park, R. J. Hot Temperature and High-Stakes Performance. en. Journal of Human Re- sources 57. Publisher: University of Wisconsin Press, 400–434. doi:10.3368/jhr.57.2 .0618-9535R3 (2022). 27. Fishman, R., Carrillo, P. & Russ, J. Long-term impacts of exposure to high tempera- tures on human capital and economic productivity. en. Journal of Environmental Eco- nomics and Management 93, 221–238. doi:10.1016/j.jeem.2018.10.001 (2019). 28. Isen, A., Rossin-Slater, M. & Walker, R. Relationship between season of birth, tem- perature exposure, and later life wellbeing. en. Proceedings of the National Academy of Sciences 114. Publisher: National Academy of Sciences Section: Social Sciences, 13447– 13452. doi:10.1073/pnas.1702436114 (2017). 29. Cil, G. & Kim, J. Extreme temperatures during pregnancy and adverse birth out- comes: Evidence from 2009 to 2018 U.S. national birth data. en. Health Economics 31. eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/hec.4559, 1993–2024. doi:1 0.1002/hec.4559 (2022). 30. Conte Keivabu, R. Too hot to study? Gender and SES differences in the effect of tempera- ture on school performance en. Tech. rep. whtf5. Publication Title: SocArXiv (Center for Open Science, 2020). 31. Hsu, A., Sheriff, G., Chakraborty, T. & Manya, D. Disproportionate exposure to urban heat island intensity across major US cities. en. Nature Communications 12. Number: 1 Publisher: Nature Publishing Group, 2721. doi:10.1038/s41467-021-22799-5 (2021). 32. Chyn, E. & Katz, L. F. Neighborhoods Matter: Assessing the Evidence for Place Ef- fects. en. Journal of Economic Perspectives 35, 197–222. doi:10.1257/jep.35.4.197 (2021). 33. Heilmann, K., Kahn, M. E. & Tang, C. K. The urban crime and heat gradient in high and low poverty areas. en. Journal of Public Economics 197, 104408. doi:10.1016/j.jp ubeco.2021.104408 (2021). 34. Albert, C., Bustos, P. & Ponticelli, J. The Effects of Climate Change on Labor and Capital Reallocation Working Paper 28995. Series: Working Paper Series (National Bureau of Economic Research, 2021). doi:10.3386/w28995. 35. Arteaga, C. The effect of human capital on earnings: Evidence from a reform at Colom- bia’s top university. en. Journal of Public Economics 157, 212–225. doi:10.1016/j.jpub eco.2017.10.007 (2018). 36. Clark, D. & Martorell, P. The Signaling Value of a High School Diploma. Journal of Political Economy 122. Publisher: The University of Chicago Press, 282–318. doi:10.1 086/675238 (2014). 71 37. Mu˜ noz-Sabater, J., Dutra, E., Agust´ ı-Panareda, A., Albergel, C., Arduini, G., Balsamo, G., et al. ERA5-Land: a state-of-the-art global reanalysis dataset for land applications. English. Earth System Science Data 13. Publisher: Copernicus GmbH, 4349–4383. doi:1 0.5194/essd-13-4349-2021 (2021). 38. Oliveira, R. & Silveira Neto, R. d. M. Brazil’s south–north labour income gap re- examined: evidence across purchasing power adjusted wage distributions. Regional Studies 0. Publisher: Routledge eprint: https://doi.org/10.1080/00343404.2021.1962005, 1–21. doi:10.1080/00343404.2021.1962005 (2021). 39. Hovdahl, I. Deadly Variation: The Effect of Temperature Variability on Mortality eng. Work- ing paper. Accepted: 2020-02-26T15:03:57Z ISSN: 1892-2198 Publication Title: 58 (BI Norwegian Business School, 2020). 40. Meghir, C., Narita, R. & Robin, J.-M. Wages and Informality in Developing Countries. en. American Economic Review 105, 1509–1546. doi:10.1257/aer.20121110 (2015). 41. Engbom, N., Gonzaga, G., Moser, C. & Olivieri, R. Earnings Inequality and Dynamics in the Presence of Informality: The Case of Brazil en. SSRN Scholarly Paper ID 3778408 (Social Science Research Network, Rochester, NY, 2022). doi:10.2139/ssrn.3778408. 42. Kang, T. H. & Felix, N. F. A. The evolution of schooling in Brazil, 1940-2010: insights from a new dataset. en, 18. 43. History — Inep. SAEB, Aneb, Prova Brasil 44. Karjalainen, S. Thermal comfort and gender: a literature review. en. Indoor Air 22. eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1600-0668.2011.00747.x, 96– 109. doi:10.1111/j.1600-0668.2011.00747.x (2012). 45. Schweiker, M., Huebner, G. M., Kingma, B. R. M., Kramer, R. & Pallubinsky, H. Drivers of diversity in human thermal perception – A review for holistic comfort models. Tem- perature 5. Publisher: Taylor & Francis eprint: https://doi.org/10.1080/23328940.2018.1534490, 308–342. doi:10.1080/23328940.2018.1534490 (2018). 46. Chang, T. Y. & Kajackaite, A. Battle for the thermostat: Gender and the effect of tem- perature on cognitive performance. en. PLOS ONE 14. Publisher: Public Library of Science, e0216362. doi:10.1371/journal.pone.0216362 (2019). 47. Kingma, B. & van Marken Lichtenbelt, W. Energy consumption in buildings and fe- male thermal demand. en. Nature Climate Change 5. Number: 12 Publisher: Nature Publishing Group, 1054–1056. doi:10.1038/nclimate2741 (2015). 48. Bursztyn, L., Fiorin, S., Gottlieb, D. & Kanz, M. Moral Incentives in Credit Card Debt Repayment: Evidence from a Field Experiment. Journal of Political Economy 127, 1641– 1683. doi:10.1086/701605 (2019). 49. Frey, B. S. & Oberholzer-Gee, F. The Cost of Price Incentives: An Empirical Analysis of Motivation Crowding- Out. The American Economic Review 87, 746–755 (1997). 50. B´ enabou, R. & Tirole, J. Intrinsic and Extrinsic Motivation. The Review of Economic Studies 70, 489–520. doi:10.1111/1467-937X.00253 (2003). 51. B´ enabou, R. & Tirole, J. Incentives and Prosocial Behavior. American Economic Review 96, 1652–1678. doi:10.1257/aer.96.5.1652 (2006). 72 52. Allcott, H. Social norms and energy conservation. Journal of Public Economics 95. Spe- cial Issue: The Role of Firms in Tax Systems, 1082–1095. doi:https://doi.org/10.1 016/j.jpubeco.2011.03.003 (2011). 53. Iannaccone, L. R. Introduction to the Economics of Religion. Journal of Economic Lit- erature 36, 1465–1495 (1998). 54. Clingingsmith, D., Khwaja, A. I. & Kremer, M. Estimating the Impact of The Hajj: Re- ligion and Tolerance in Islam’s Global Gathering*. The Quarterly Journal of Economics 124, 1133–1170. doi:10.1162/qjec.2009.124.3.1133 (2009). 55. Benjamin, D., Choi, J. & Fisher, G. Religious Identity and Economic Behavior. The Review of Economics and Statistics 98, 617–637 (2016). 73 Appendices A Heat and Long Term Earnings: Evidence from Schooling in Brazil A.1 Figures and Summary Statistics Figure A.1: Regions of Brazil 74 Table A.1: Means by Region Region Male Asinh Wage 30 Highschool Middle Elementary High Rain 28-30C 30-32C 32-34C 34C+ 1 C 0.67 7.54 0.72 0.89 0.97 0.08 249.04 152.13 60.02 24.33 2 N-NE 0.68 7.50 0.75 0.89 0.96 0.04 315.11 179.68 66.87 34.97 3 S SE 0.62 7.61 0.76 0.92 0.98 0.09 126.45 60.59 19.56 5.52 Table A.2: Correlation Across Birth, Elementary, and Work Year Temperatures Birth temp Schooling temp Work temp Birth temp 1.00000 0.28625 0.04348 Schooling temp 0.28625 1.00000 0.10899 Work temp 0.04348 0.10899 1.00000 75 Figure A.2: Population Density, Brazil 76 A.2 Regressions Table A.3: School Days Temperature on Wage by School Exposure Period Dependent Variable: Wage Percent from Exposure During: Exposure Period Elementary Middle Highschool Higher Education Model: (1) (2) (3) (4) Variables 26-28 0.0121 0.0228 -0.0171 0.0384 (0.0212) (0.0251) (0.0291) (0.0314) 28-30 -0.0373 0.0295 0.0001 0.0063 (0.0239) (0.0210) (0.0292) (0.0289) 30-32 -0.0564 ∗∗ 0.0566 ∗∗ 0.0439 ∗ -0.0249 (0.0236) (0.0227) (0.0254) (0.0273) 32-34 -0.0938 ∗∗∗ 0.0261 -0.0577 0.0842 ∗∗ (0.0253) (0.0309) (0.0481) (0.0420) >34 -0.0769 ∗∗ 0.0624 ∗∗ 0.0954 ∗∗∗ 0.0459 (0.0335) (0.0282) (0.0327) (0.0315) Fit statistics Controls Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Observations 11,019,010 11,019,010 11,019,010 11,019,010 R 2 0.01960 0.01958 0.01958 0.01958 Within R 2 0.00018 0.00017 0.00017 0.00017 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. Controls include rain 90th percentlie indicator and gender. Fixed effects include municipality and year 77 Table A.4: Elementary Schooling Days Temperature on Wage by Region - Non-Linear Dependent Variable: Wage percent Region Brazil Central North-Northeast South-Southeast Model: (1) (2) (3) (4) Variables >30 -0.1134 ∗∗∗ -0.0697 ∗ -0.0292 -0.1813 ∗∗∗ (0.0199) (0.0384) (0.0294) (0.0336) >30 sq 0.0001 ∗∗∗ 0.0001 ∗∗ 4.6× 10 − 5 0.0002 ∗∗∗ (2.85× 10 − 5 ) (6.56× 10 − 5 ) (2.95× 10 − 5 ) (8.26× 10 − 5 ) Fit statistics Controls Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Average n Hot Days 129.43 236.48 281.53 85.671 Total Effect at Mean n Hot Days -0.07606 -0.00578 -0.00334 -0.14283 R 2 0.01960 0.03596 0.02362 0.02083 Observations 11,018,997 926,901 1,748,088 8,344,008 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is below 30C. Coefficients multiplied by 100 to allow for direct percentage interpretation. Controls include rain 90th percentlie indicator and gender. Fixed effects include municipality and year 78 Table A.5: Elementary Schooling Days Temperature on Wage by School Completion Dependent Variable: Wage percent Highest Level of Education Any Elementary Middle Highschool Model: (1) (2) (3) (4) Variables 26-28 0.0034 -0.0021 0.0017 0.0074 (0.0213) (0.0211) (0.0206) (0.0210) 28-30 -0.0353 -0.0401 ∗ -0.0366 -0.0286 (0.0240) (0.0240) (0.0248) (0.0266) 30-32 -0.0504 ∗∗ -0.0538 ∗∗ -0.0372 -0.0258 (0.0240) (0.0239) (0.0244) (0.0255) 32-34 -0.0857 ∗∗∗ -0.0935 ∗∗∗ -0.0852 ∗∗∗ -0.0830 ∗∗∗ (0.0255) (0.0253) (0.0252) (0.0259) >34 -0.0556 -0.0597 ∗ -0.0423 -0.0428 (0.0350) (0.0347) (0.0347) (0.0351) Fit statistics Controls Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Observations 11,018,997 10,799,531 10,041,846 8,334,822 R 2 0.01960 0.01777 0.01434 0.01314 Within R 2 0.00019 0.00012 3.39× 10 − 5 0.00034 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. Controls include rain 90th percentlie indicator and gender. Fixed effects include municipality and year 79 Table A.6: Elementary Schooling Days Temperature on Wage by Region and High School Completion Dependent Variable: Wage percent Region Central North-Northeast South-Southeast High School Completion noHS HS noHS HS noHS HS Model: (1) (2) (3) (4) (5) (6) Variables 26-28 0.0702 0.0433 0.0222 0.0279 -0.0468 -0.0016 (0.1096) (0.0625) (0.0699) (0.0401) (0.0395) (0.0246) 28-30 0.0600 0.0178 0.0134 0.0142 -0.0681 ∗ -0.0278 (0.1012) (0.0526) (0.0605) (0.0363) (0.0355) (0.0330) 30-32 0.1413 -0.0243 -0.0269 0.0581 -0.1143 ∗∗ -0.0619 (0.1135) (0.0645) (0.0646) (0.0365) (0.0457) (0.0392) 32-34 0.0156 0.0991 -0.0138 0.0081 -0.2080 ∗∗∗ -0.1952 ∗∗∗ (0.1478) (0.0828) (0.0693) (0.0396) (0.0683) (0.0483) >34 0.2209 0.1743 ∗∗ -0.0511 0.0322 -0.3628 ∗∗∗ -0.1025 (0.1538) (0.0836) (0.0935) (0.0455) (0.0846) (0.0857) Fit statistics Controls Yes Yes Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Yes Yes Observations 260,698 666,203 433,633 1,314,455 1,989,844 6,354,164 R 2 0.04427 0.02797 0.03843 0.01575 0.02571 0.01407 Within R 2 0.00018 0.00026 0.00028 4.28× 10 − 5 0.00059 0.00063 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percent- age interpretation. Controls include rain 90th percentlie indicator and gender. Fixed effects include municipality and year 80 Table A.7: High School Completion by Region Dependent Variable: High School Completion pct All C N NE S SE Model: (1) (2) (3) (4) Variables 26-28 0.0006 -0.0045 -0.0338 ∗∗ 0.0031 (0.0041) (0.0117) (0.0153) (0.0045) 28-30 -0.0045 -0.0121 -0.0280 ∗∗ 0.0003 (0.0041) (0.0104) (0.0117) (0.0048) 30-32 -0.0049 -0.0050 -0.0273 ∗∗ -0.0101 ∗ (0.0042) (0.0120) (0.0114) (0.0056) 32-34 0.0125 ∗∗ 0.0193 -0.0190 0.0132 ∗ (0.0049) (0.0148) (0.0116) (0.0072) >34 0.0394 ∗∗∗ 0.0546 ∗∗∗ -0.0030 0.0486 ∗∗∗ (0.0064) (0.0159) (0.0130) (0.0112) Fit statistics Controls Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Observations 11,018,997 926,901 1,748,088 8,344,008 R 2 0.11009 0.13433 0.15585 0.10086 Within R 2 0.03683 0.04594 0.04733 0.03334 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. Controls include rain 90th percentlie indicator and gender. Fixed effects include munic- ipality and year 81 Figure A.3: High School Completion by Region 82 A.3 Regressions - Main Specification Robustness 83 Table A.8: Elementary Schooling Days Temperature on Wage - Robustness to Controls, Full Sample Dependent Variable: Wage percent Model: (1) (2) (3) (4) Variables 26-28 0.0120 0.0033 0.0121 0.0034 (0.0212) (0.0213) (0.0212) (0.0213) 28-30 -0.0375 -0.0355 -0.0373 -0.0353 (0.0239) (0.0241) (0.0239) (0.0240) 30-32 -0.0564 ∗∗ -0.0504 ∗∗ -0.0564 ∗∗ -0.0504 ∗∗ (0.0236) (0.0240) (0.0236) (0.0240) 32-34 -0.0937 ∗∗∗ -0.0856 ∗∗∗ -0.0938 ∗∗∗ -0.0857 ∗∗∗ (0.0253) (0.0255) (0.0253) (0.0255) >34 -0.0769 ∗∗ -0.0556 -0.0769 ∗∗ -0.0556 (0.0335) (0.0351) (0.0335) (0.0350) Fixed-effects municipality Yes Yes Yes Yes year Yes Yes Yes Yes Controls Low Temp Bins Yes Yes Yes Yes Rain 90th Yes Yes Gender Yes Yes Fit statistics Observations 11,019,010 11,018,997 11,019,010 11,018,997 R 2 0.01945 0.01946 0.01960 0.01960 Within R 2 3.34× 10 − 5 4.22× 10 − 5 0.00018 0.00019 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. 84 Table A.9: Elementary Schooling Days Temperature on Wage with Other School Temperature Bin Controls Dependent Variable: Wage Percent Sample Full Elementary Completed Only Model: (1) (2) (3) (4) (5) (6) (7) (8) Variables 26-28 0.0121 0.0214 -0.0083 0.0019 0.0064 0.0125 -0.0106 0.0001 (0.0212) (0.0226) (0.0260) (0.0264) (0.0210) (0.0223) (0.0257) (0.0261) 28-30 -0.0373 -0.0310 -0.0411 ∗ -0.0374 -0.0422 ∗ -0.0367 ∗ -0.0419 ∗ -0.0383 ∗ (0.0239) (0.0223) (0.0241) (0.0228) (0.0239) (0.0223) (0.0241) (0.0227) 30-32 -0.0564 ∗∗ -0.0465 ∗∗ -0.0424 ∗ -0.0325 -0.0599 ∗∗ -0.0514 ∗∗ -0.0420 ∗ -0.0328 (0.0236) (0.0223) (0.0228) (0.0224) (0.0235) (0.0222) (0.0226) (0.0222) 32-34 -0.0938 ∗∗∗ -0.1008 ∗∗∗ -0.1065 ∗∗∗ -0.1037 ∗∗∗ -0.1016 ∗∗∗ -0.1076 ∗∗∗ -0.1083 ∗∗∗ -0.1029 ∗∗∗ (0.0253) (0.0261) (0.0276) (0.0274) (0.0251) (0.0259) (0.0274) (0.0271) >34 -0.0769 ∗∗ -0.0829 ∗∗ -0.0920 ∗∗∗ -0.0763 ∗∗ -0.0808 ∗∗ -0.0893 ∗∗∗ -0.0942 ∗∗∗ -0.0774 ∗∗ (0.0335) (0.0330) (0.0325) (0.0318) (0.0332) (0.0329) (0.0324) (0.0318) Fixed-effects municipality Yes Yes Yes Yes Yes Yes Yes Yes year Yes Yes Yes Yes Yes Yes Yes Yes Controls Low Temp Bins Yes Yes Yes Yes Yes Yes Yes Yes Gender Yes Yes Yes Yes Yes Yes Yes Yes Middle School Temp Bins Yes Yes Yes Yes Yes Yes Highschool Temp Bins Yes Yes Yes Yes Higher Edu Temp Bins Yes Yes Fit statistics Observations 11,019,010 11,019,010 11,019,010 11,019,010 10,799,544 10,799,544 10,799,544 10,799,544 R 2 0.01960 0.01961 0.01964 0.01965 0.01776 0.01777 0.01780 0.01781 Within R 2 0.00018 0.00020 0.00023 0.00024 0.00011 0.00013 0.00015 0.00016 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. 85 Table A.10: Elementary Schooling Days Temperature on Wage by Region - 3C Bins Dependent Variable: Wage percent Region Brazil Central North-Northeast South-Southeast Model: (1) (2) (3) (4) Variables 26-29 -0.0161 0.0425 -0.0168 -0.0225 (0.0213) (0.0492) (0.0391) (0.0236) 29-31 -0.0499 ∗∗ 0.0090 -0.0160 -0.0624 ∗∗ (0.0240) (0.0573) (0.0373) (0.0303) 32-35 -0.0788 ∗∗∗ 0.1332 ∗ -0.0253 -0.1876 ∗∗∗ (0.0258) (0.0766) (0.0416) (0.0386) >35C -0.0660 0.2425 ∗∗ -0.0131 -0.1847 ∗ (0.0402) (0.0998) (0.0570) (0.0957) Fit statistics Controls Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Observations 11,018,997 926,901 1,748,088 8,344,008 R 2 0.01960 0.03609 0.02367 0.02084 Within R 2 0.00019 0.00164 0.00098 9.17× 10 − 5 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. Controls include rain 90th percentlie indicator and gender. Fixed effects include municipality and year 86 Table A.11: Elementary Schooling Days Temperature on Wage by Region - 1C Bins Dependent Variable: Wage percent Region Brazil Central North-Northeast South-Southeast Model: (1) (2) (3) (4) Variables 26 -0.0414 0.0403 -0.0347 -0.0552 (0.0358) (0.0698) (0.0639) (0.0414) 27 0.0520 ∗ 0.0983 -0.0148 0.0505 (0.0286) (0.0719) (0.0416) (0.0319) 28 -0.0664 ∗ -0.0048 -0.0142 -0.0884 ∗∗ (0.0350) (0.0576) (0.0502) (0.0420) 29 -0.0184 0.0872 -0.0542 0.0147 (0.0267) (0.0680) (0.0381) (0.0378) 30 -0.0407 0.0742 -0.0048 -0.0875 (0.0317) (0.0696) (0.0400) (0.0600) 31 -0.0680 ∗ -0.0124 -0.0238 -0.0961 (0.0410) (0.0763) (0.0530) (0.0658) 32 -0.1050 ∗∗∗ 0.0564 -0.0023 -0.2459 ∗∗∗ (0.0405) (0.0975) (0.0471) (0.0742) 33 -0.0403 0.2114 ∗ -0.0835 -0.0688 (0.0446) (0.1109) (0.0546) (0.0840) >34 -0.0553 ∗ 0.2249 ∗∗∗ -0.0069 -0.1493 ∗∗ (0.0323) (0.0837) (0.0544) (0.0647) Fixed-effects municipality Yes Yes Yes Yes year Yes Yes Yes Yes Controls Low Temp Bins Yes Yes Yes Yes Rain 90th Yes Yes Yes Yes Gender Yes Yes Yes Yes Fit statistics Observations 11,018,997 926,901 1,748,088 8,344,008 R 2 0.01961 0.03614 0.02369 0.02086 Within R 2 0.00020 0.00169 0.00101 0.00011 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage interpretation. 87 Table A.12: Elementary Schooling Days Temperature on Wage - Robustness to Age Dependent Variables: Wage age 29 percent Wage age 30 percent Wage age 31 percent Wage percent Model: (1) (2) (3) (4) Variables 26-28 -0.0051 0.0193 0.0339 0.0034 (0.0315) (0.0331) (0.0394) (0.0213) 28-30 -0.0916 ∗∗∗ -0.0354 0.0157 -0.0353 (0.0331) (0.0347) (0.0435) (0.0240) 30-32 -0.0697 ∗∗ -0.0678 ∗∗ -0.0287 -0.0504 ∗∗ (0.0340) (0.0334) (0.0442) (0.0240) 32-34 -0.1060 ∗∗∗ -0.1191 ∗∗∗ -0.0385 -0.0857 ∗∗∗ (0.0378) (0.0394) (0.0461) (0.0255) >34 -0.0778 ∗ -0.0441 -0.0507 -0.0556 (0.0456) (0.0510) (0.0759) (0.0350) Fit statistics Controls Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Observations 9,212,800 11,018,997 8,869,169 11,018,997 R 2 0.01465 0.00917 0.01199 0.01960 Within R 2 0.00135 0.00058 0.00067 0.00019 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage inter- pretation. Controls include rain 90th percentlie indicator and gender. Fixed effects include municipality and year 88 (a) Main Specification, 1C Bins (b) Main Specification - 3C Bins Notes: Figure A.4: Elementary Svhooling Days Temperature on Wage by Region - 1C and 3C Bins 89 Figure A.5: Brazil Earnings Split by Age 90 A.4 Regressions - South and Southeast Robustness Figure A.6: South-Southeast by Age 91 Table A.13: Elementary Schooling Days Temperature on Wage - Robustness to Controls in the South Dependent Variable: Wage percent Model: (1) (2) (3) (4) Variables 26-28 -0.0041 -0.0089 -0.0040 -0.0089 (0.0245) (0.0246) (0.0245) (0.0245) 28-30 -0.0372 -0.0347 -0.0372 -0.0346 (0.0289) (0.0291) (0.0289) (0.0291) 30-32 -0.0984 ∗∗∗ -0.0918 ∗∗∗ -0.0982 ∗∗∗ -0.0917 ∗∗∗ (0.0346) (0.0351) (0.0345) (0.0351) 32-34 -0.1947 ∗∗∗ -0.1863 ∗∗∗ -0.1949 ∗∗∗ -0.1866 ∗∗∗ (0.0454) (0.0454) (0.0454) (0.0454) >34 -0.1783 ∗∗ -0.1593 ∗∗ -0.1783 ∗∗ -0.1594 ∗∗ (0.0719) (0.0743) (0.0718) (0.0743) Fixed-effects municipality Yes Yes Yes Yes year Yes Yes Yes Yes Controls Low Temp Bins Yes Yes Yes Yes Rain 90th Yes Yes Gender Yes Yes Fit statistics Observations 8,344,021 8,344,008 8,344,021 8,344,008 R 2 0.02082 0.02083 0.02084 0.02085 Within R 2 7.69× 10 − 5 8.08× 10 − 5 9.53× 10 − 5 9.92× 10 − 5 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to al- low for direct percentage interpretation. 92 Table A.14: Elementary Schooling Days Temperature on Wage - Robustness to Age in the South Dependent Variables: Wage age 29 percent Wage age 30 percent Wage age 31 percent Wage percent Model: (1) (2) (3) (4) Variables 26-28 -0.0073 0.0097 0.0447 -0.0089 (0.0373) (0.0374) (0.0460) (0.0245) 28-30 -0.1174 ∗∗∗ -0.0324 0.0317 -0.0346 (0.0418) (0.0420) (0.0537) (0.0291) 30-32 -0.1094 ∗∗ -0.0839 ∗ -0.0275 -0.0917 ∗∗∗ (0.0492) (0.0447) (0.0631) (0.0351) 32-34 -0.1805 ∗∗∗ -0.2972 ∗∗∗ -0.1483 ∗ -0.1866 ∗∗∗ (0.0649) (0.0657) (0.0847) (0.0454) >34 -0.1051 -0.1823 -0.1947 -0.1594 ∗∗ (0.0916) (0.1123) (0.1783) (0.0743) Fit statistics Controls Yes Yes Yes Yes Fixed Effects Yes Yes Yes Yes Observations 7,008,878 8,344,008 6,755,784 8,344,008 R 2 0.01603 0.00984 0.01254 0.02085 Within R 2 0.00089 0.00032 0.00036 9.92× 10 − 5 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Reference temperature bin is 24-26C. Coefficients multiplied by 100 to allow for direct percentage inter- pretation. Controls include rain 90th percentlie indicator and gender. Fixed effects include municipality and year 93 A.5 Regressions - SAEB Scores Table A.15: Elementary Schooling Days Temperature on Mathematics Score by Region Dependent Variable: Standardized Score (Mathematics) C N NE S SE Model: (1) (2) (3) Variables 24-26 0.0011 -0.0005 -0.0003 (0.0013) (0.0007) (0.0003) 26-28 0.0019 ∗ 0.0003 -0.0008 ∗∗∗ (0.0011) (0.0006) (0.0003) 28-30 0.0018 ∗ -0.0002 -0.0006 ∗∗ (0.0010) (0.0006) (0.0003) 30-32 0.0017 − 8.59× 10 − 5 0.0004 (0.0011) (0.0006) (0.0004) 32-34 0.0011 -0.0008 0.0001 (0.0011) (0.0006) (0.0004) >34 0.0018 0.0001 -0.0002 (0.0012) (0.0006) (0.0005) Fixed-effects School ID Yes Yes Yes year Yes Yes Yes Controls Low Temp Bins Yes Yes Yes Fit statistics Observations 371,644 1,748,846 2,343,499 R 2 0.15227 0.24273 0.14954 Within R 2 0.00034 0.00020 0.00049 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 94 Table A.16: Elementary Schooling Days Temperature on Mathematics Score by Region and SEC Dependent Variable: Standardized Score (Mathematics) region C N NE S SE SEC high low high low high low Model: (1) (2) (3) (4) (5) (6) Variables 24-26 0.0012 -0.0054 -0.0005 -0.0008 -0.0002 -0.0019 ∗∗ (0.0013) (0.0049) (0.0013) (0.0008) (0.0003) (0.0008) 26-28 0.0020 ∗ -0.0011 0.0005 0.0005 -0.0008 ∗∗∗ -0.0025 ∗∗∗ (0.0011) (0.0046) (0.0010) (0.0007) (0.0003) (0.0007) 28-30 0.0019 ∗ -0.0010 -0.0004 − 8.15× 10 − 5 -0.0007 ∗∗ -0.0027 ∗∗∗ (0.0010) (0.0043) (0.0011) (0.0007) (0.0003) (0.0006) 30-32 0.0019 -0.0005 -0.0007 0.0002 0.0005 -0.0027 ∗∗∗ (0.0011) (0.0044) (0.0011) (0.0007) (0.0004) (0.0008) 32-34 0.0009 -0.0032 -0.0007 -0.0008 0.0003 -0.0016 ∗∗ (0.0012) (0.0045) (0.0011) (0.0007) (0.0004) (0.0008) >34 0.0017 0.0001 -0.0006 0.0006 0.0001 -0.0031 ∗∗∗ (0.0012) (0.0043) (0.0011) (0.0007) (0.0005) (0.0007) Fixed-effects School ID Yes Yes Yes Yes Yes Yes year Yes Yes Yes Yes Yes Yes Controls Low Temp Bins Yes Yes Yes Yes Yes Yes Fit statistics Observations 324,832 46,812 547,162 1,201,684 2,155,042 188,457 R 2 0.14688 0.20185 0.20152 0.25029 0.14745 0.17105 Within R 2 0.00046 0.00078 0.00026 0.00029 0.00054 0.00045 Clustered (municipality-year) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 95 Figure A.7: SAEB Mathematics Test Scores by SEC 96 B Religious Figures and their Impact on Non-Religious Decisions, for the Religious and ”A”-religious: An Experiment on Environmental Donations B.1 Robustness to Survey Durations Appendix Appendix B.B.1 graphs the output of the hetereogeneous regressions for differ- ent values of cutoff durations. 97 −10 0 10 20 30 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Benedict − Very Relgious −10 0 10 20 30 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Francis − Very Relgious −20 −10 0 10 20 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Khamenei − Very Relgious (a) Coefficients and 95% confidence intervals for the very religious as the lower bound varies. 0 20 40 60 80 100 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Benedict − Very Relgious −40 −20 0 20 40 60 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Francis − Very Relgious −20 0 20 40 60 80 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Khamenei − Very Relgious (b) Coefficients and 95% confidence intervals for the very religious as the upper bound varies. Notes: Figure B.1: Coefficients and 95% confidence intervals for the very religious: Upper and Lower Bound 98 −10 −5 0 5 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Benedict − Somewhat/Non Religious −5 0 5 10 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Francis − Somewhat/Non Religious −10 −5 0 5 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Khamenei − Somewhat/Non Religious (a) Coefficients and 95% confidence intervals for the somewhat and non-religious as the lower bound varies. −10 −5 0 5 10 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Benedict − Somewhat/Non Religious −10 −5 0 5 10 15 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Francis − Somewhat/Non Religious −10 −5 0 5 10 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Khamenei − Somewhat/Non Religious (b) Coefficients and 95% confidence intervals for the somewhat and non-religious as the upper bound varies. Notes: Figure B.2: Coefficients and 95% confidence intervals for the somewhat and non-religious: Upper and Lower Bound 99 −20 −15 −10 −5 0 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Benedict − Atheist/Agnostic −20 −15 −10 −5 0 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Francis − Atheist/Agnostic −20 −15 −10 −5 0 Donation difference from control 40 50 60 70 80 Lower bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The upper bound is fixed at 240. Khamenei − Atheist/Agnostic (a) Coefficients and 95% confidence intervals for agnostic and atheists as the lower bound varies. −30 −20 −10 0 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Benedict − Atheist/Agnostic −25 −20 −15 −10 −5 0 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Francis − Atheist/Agnostic −40 −30 −20 −10 0 Donation difference from control 100 150 200 250 300 Upper bound on survey duration (secs) Regressions include the full set of controls and robust standard errors. The lower bound is fixed at 60. Khamenei − Atheist/Agnostic (b) Coefficients and 95% confidence intervals for the agnostic and atheists as the upper bound varies. Notes: Figure B.3: Coefficients and 95% confidence intervals for agnostic and atheists: Upper and Lower Bound 100 B.2 Robustness to Donation Indicator Appendix Appendix B.B.2 documents different cutoffs around the mean. The results that are significant continue to be so within the ranges specified. 101 0 .2 .4 .6 Difference from Control 17 18 19 20 21 22 23 Cutoff Regressions include the full set of controls and robust standard errors. Using the subsample with survey duration between 60 and 240. Benedict − Very Relgious 0 .1 .2 .3 .4 .5 Difference from Control 17 18 19 20 21 22 23 Cutoff Regressions include the full set of controls and robust standard errors. Using the subsample with survey duration between 60 and 240. Francis − Very Relgious −.2 −.1 0 .1 .2 .3 Difference from Control 17 18 19 20 21 22 23 Cutoff Regressions include the full set of controls and robust standard errors. Using the subsample with survey duration between 60 and 240. Khamenei − Very Relgious (a) Coefficients and 95% confidence intervals for dependent indicator variable cutoff for the very religious. −.15 −.1 −.05 0 .05 .1 Difference from Control 17 18 19 20 21 22 23 Cutoff Regressions include the full set of controls and robust standard errors. Using the subsample with survey duration between 60 and 240. Benedict − Somewhat/Non Religious −.1 −.05 0 .05 .1 .15 Difference from Control 17 18 19 20 21 22 23 Cutoff Regressions include the full set of controls and robust standard errors. Using the subsample with survey duration between 60 and 240. Francis − Somewhat/Non Religious −.2 −.1 0 .1 .2 Difference from Control 17 18 19 20 21 22 23 Cutoff Regressions include the full set of controls and robust standard errors. Using the subsample with survey duration between 60 and 240. Khamenei − Somewhat/Non Religious (b) Coefficients and 95% confidence intervals for dependent indicator variable cutoff for the somewhat and non-religious. Notes: Figure B.4: Coefficients and 95% confidence intervals for dependent indicator variable cutoff: for the very religious, somewhat, and non-religious. 102 −.3 −.2 −.1 0 Difference from Control 17 18 19 20 21 22 23 Cutoff Regressions include the full set of controls and robust standard errors. Using the subsample with survey duration between 60 and 240. Benedict − Atheist/Agnostic −.3 −.2 −.1 0 Difference from Control 17 18 19 20 21 22 23 Cutoff Regressions include the full set of controls and robust standard errors. Using the subsample with survey duration between 60 and 240. Francis − Atheist/Agnostic −.4 −.3 −.2 −.1 Difference from Control 17 18 19 20 21 22 23 Cutoff Regressions include the full set of controls and robust standard errors. Using the subsample with survey duration between 60 and 240. Khamenei − Atheist/Agnostic Figure B.5: Coefficients and 95% confidence intervals for dependent indicator variables cutoff: for Agnostics and Atheists 103

Abstract (if available)

Abstract This dissertation consists of three chapters, with different approaches and insights into topics related to Environmental Economics.

The first chapter provides causal evidence of a negative impact of heat exposure during school age on long term earnings in Brazil. Using natural variation in temperature exposure over the first five years of elementary education, I find that a one standard deviation increase in the number of elementary schooling days between 30-32C reduces long term earnings by 0.94% in the South and Southeast. Although ex-ante these impacts could be operating through a number of mechanisms, the heterogeneity in the effects of shocks across school and non-school days supports a learning channel: there are no long-term earnings effects of heat exposure during summer and weekends. No effects are found for middle or high-school exposure, but heat shocks remain significant in all but the final year of elementary education. Consistent with some degree of adaptation, effects are concentrated in climates unaccustomed to continuous heat, and are most prominent amongst the least educated. A secondary set of data and results examining schooling outcomes further complements the evidence that the effect occurs through disrupted education and falls disproportionately on the most disadvantaged socioeconomic class. These results suggest the learning inhibited by heat is sizeable enough to be captured in later life earnings, and that public policy could potentially mitigate these impacts by supporting human capital accumulation during childhood.

The second chapter, joint work with Antonio Bento and Edson Severnini, proposes a paper in which we ask and attempt to answer the question of whether labor market disparities seen during economic downturns are also evidenced from the result of environment-related declines in activity. Using a fixed-effects approach with variation in droughts both temporally and spatially we intend to estimate the impact on the hiring and firing of workers by race and gender. This chapter documents promising preliminary results.

The third chapter studies the heterogeneous effect of religious figures on preferences for the environment proxied by donations from a $100 prize draw. Quotes from religious leaders are provided and juxtaposed either with a religious leader or a generic "world leader". In the preferred specification's LATE, I find a backlash effect among those that identify as Atheists or Agnostics who reduce their donations by $8.53-$9.62 across treatment arms compared to the control group. There is no statistically significant effect for the Non-Religious and Somewhat Religious. The Very Religious select to increase donations by $9.83-$11.28 when less controversial religious figures are involved. Robustness checks to specification and subsampling are discussed, as is suggestive evidence that it is indeed the heterogeneity in religiosity rather than a correlated variable driving the results. I finish with a discussion on why I have found such an effect that does not appear very prominently in related literature.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays in environmental economics

PDF

Essays in the economics of education and conflict

PDF

Essays on development economics and adolescent behavior

PDF

Essays on development and health economics: social media and education policy

PDF

Essays on urban and real estate economics

PDF

Essays on the dual urban-rural system and economic development in China

PDF

Three essays on strategic commuters with late arrival penalties and toll lanes

PDF

Essays on narrative economics, Climate macrofinance and migration

PDF

Essays on the economics of climate change adaptation in developing countries

PDF

Essays on political economy and corruption

PDF

Essays on applied microeconomics

PDF

Essays in labor economics: demographic determinants of labor supply

PDF

Essays on economics of education

PDF

Essays in international economics

PDF

Two essays in economic policy: The influence of perceived comparative need on financial subsidy requests; &, Unintended consequences of the FOSTA-SESTA legislation

PDF

Essays on innovation, human capital, and COVID-19 related policies

PDF

Essays on development economics

PDF

Essays on development and health economics

PDF

Essays on education and institutions in developing countries

PDF

Essays in development economics

Asset Metadata

Creator Khan, Taraq (author)

Core Title Essays on environmental economics: education, employment, and experiments

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2023-05

Publication Date 03/23/2023

Defense Date 03/21/2023

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

Tag Charity,economics,Education,employment,Environment,environmental economics,environmental justice,Labor,OAI-PMH Harvest,Religion,wages

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Bento, Antonio (committee chair), Oliva, Paulina (committee chair), Weaver, Jeff (committee member)

Creator Email taraq.khan@hotmail.co.uk,taraqkha@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC112847706

Unique identifier UC112847706

Identifier etd-KhanTaraq-11514.pdf (filename)

Legacy Identifier etd-KhanTaraq-11514

Document Type Dissertation

Format theses (aat)

Rights Khan, Taraq

Internet Media Type application/pdf

Type texts

Source 20230324-usctheses-batch-1011 (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. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

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