University of Southern California Dissertations and Theses

Essays in climate change adaptation: role of market power in incentivizing adaptation behavior

(USC Thesis Other)

Essays in climate change adaptation: role of market power in incentivizing adaptation behavior

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ESSAYS IN CLIMATE CHANGE ADAPTATION
by
Rajat Kochhar
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
May 2023
Copyright 2023 Rajat Kochhar
I dedicate this thesis to myself, my family, and the people who have made me
who I am.
ii
Acknowledgements
I am incredibly grateful to my supportive advisors. Matthew Kahn has been a role model
and taught me how to see the bigger picture and dig deeper into rigorous research simul-
taneously. Paulina Oliva has been generous with her time, guided me, and opened the
door of environmental economics for me. Robert Metcalfe’s sharp vision of the direction
of the field has greatly improved my work. John Matsusaka’s broad interest and research
agenda have inspired me to continue research with curiosity.
I would like to thank other faculties at the University of Southern California who have
been always encouraging and helpful, including Guofu Tan, Jonathan Libgober, Pablo
Kurlat, Monica Morlacco, Jeff Weaver, Vittorio Bassi, Rodney Ramcharan, Geert Ridder
and Andrii Parkomenkho. Together you helped me to overcome difficulties and enjoy
research. I would also like to thank Alwyn Young and Swati Dhingra for being wonderful
mentors at LSE.
This dissertation would not have been possible without my fiance, Ruozi Song, and
caring and supportive friends and peers, particularly Clement Boulle, Xiongfei Li, Liying
Yang, Michele Mary Bernadine, and Adam Dessouky. I also thank my brother for always
being there for me during the ups and downs and my parents for their unlimited love and
support.
iii
Contents
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Chapter 1: Does Market Power in India’s Agricultural Markets Hinder Farmer
Climate Change Adaptation? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Background on Agricultural Markets in India . . . . . . . . . . . . . . . . . . 16
1.2.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2.2 Agricultural Produce Market Committee: Regulations . . . . . . . . 17
1.2.3 Monopsony Power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.1 Yields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.2 Weather . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.3.3 Intermediary Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3.4 Quantity Arrivals and Prices . . . . . . . . . . . . . . . . . . . . . . . 28
1.4 Empirical Methods and Results . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.4.1 Effect of Climate Shocks on Yields . . . . . . . . . . . . . . . . . . . . 29
iv
1.4.1.1 Panel Approach . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.4.1.2 Long Differences Approach . . . . . . . . . . . . . . . . . . 33
1.4.1.3 Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.4.2 Effect of Competition on Mitigation of Climate Shocks . . . . . . . . 36
1.4.2.1 Measuring Market Power . . . . . . . . . . . . . . . . . . . . 37
1.4.2.2 Panel Approach . . . . . . . . . . . . . . . . . . . . . . . . . 38
1.4.2.3 Panel Approach: Arrivals Data . . . . . . . . . . . . . . . . 41
1.4.2.4 Hybrid Border Discontinuity Design . . . . . . . . . . . . . 44
1.4.3 Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
1.4.3.1 Analytical Framework . . . . . . . . . . . . . . . . . . . . . . 51
1.4.3.2 Effect of Competition on Prices: Heterogeneous Impact by
Weather . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
1.4.3.3 Changes in Input Use . . . . . . . . . . . . . . . . . . . . . . 57
1.5 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
1.5.1 Basic Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
1.5.2 Competitive Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . 65
1.6 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
1.6.1 Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
1.6.2 Supply . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
1.7 Counterfactual Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
1.8 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Chapter 2: Adapting to Flood Risk: Evidence from a Panel of Global Cities . . . . 86
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.2 Adaptation to Place Based Shocks . . . . . . . . . . . . . . . . . . . . . . . . . 93
2.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.3.1 Cities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
2.3.2 Night Lights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
v
2.3.3 Flood Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
2.3.4 Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
2.3.5 Elevation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
2.3.6 Dams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
2.4 Adaptation by Migrating Away from Flood Prone Cities . . . . . . . . . . . . 102
2.5 The Death Toll from Floods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
2.6 Urban Resilience and Flood Shocks . . . . . . . . . . . . . . . . . . . . . . . . 108
2.6.1 The Effect of Floods on Economic Activity . . . . . . . . . . . . . . . 108
2.6.2 The Effect of Extreme Rain on Economic Activity . . . . . . . . . . . 111
2.6.3 Recovery Dynamics From Floods . . . . . . . . . . . . . . . . . . . . . 113
2.7 Testing Flood Adaptation Hypotheses . . . . . . . . . . . . . . . . . . . . . . 116
2.7.1 Do Richer Cities Suffer Less? . . . . . . . . . . . . . . . . . . . . . . . 116
2.7.2 Does Repeated Experience with Flooding Reduce the Marginal Ef-
fect of the Next Flood? . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
2.7.3 Does Flood Protection Infrastructure Protect Cities? . . . . . . . . . . 118
2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Chapter 3: Do Water Audits Promote Economic Welfare?: Evidence from a Nat-
ural Field Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.2 Background and Experimental Design . . . . . . . . . . . . . . . . . . . . . . 132
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
3.3.1 Likelihood of Engagement . . . . . . . . . . . . . . . . . . . . . . . . . 135
3.3.2 Effect of the Behavioural Interventions on Water Consumption . . . 136
3.3.3 Effect of Diagnostic Completion on Water Consumption . . . . . . . 141
3.3.4 External Validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
3.3.5 Effect of Reminders on Diagnostic Completion . . . . . . . . . . . . . 146
3.4 Welfare Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
vi
3.4.1 Cost Effectiveness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
3.4.2 Benefit-Cost Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
3.4.3 MVPF framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
F Appendix for Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
F.1 Appendix Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
F.2 Appendix Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
F.3 Appendix Derivations . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
F.3.1 Joint Distribution of TFP and Labor Intensity . . . . . . . . 189
F.3.2 Probability of Choosing Market . . . . . . . . . . . . . . . . 192
F.3.3 Profit Function of Farmer . . . . . . . . . . . . . . . . . . . . 194
F.3.4 Land Allocation Problem . . . . . . . . . . . . . . . . . . . . 195
F.3.5 Quantity Supplied to Market . . . . . . . . . . . . . . . . . . 196
F.3.6 Average Conditional Productivity . . . . . . . . . . . . . . . 198
F.3.7 Consumers Utility Maximisation . . . . . . . . . . . . . . . 199
G Appendix for Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
G.1 Appendix Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
G.2 Appendix Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
H Appendix for Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
H.1 Baseline Balance and Additional Results . . . . . . . . . . . . . . . . . 206
H.1.1 Balance Table . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
H.1.2 Heterogeneity Based on Pre-Treatment Water Consumption 208
H.1.3 Characteristics of Households that Complete the Diagnostic 213
H.1.4 Interaction of Households with Reminders . . . . . . . . . . 215
vii
H.2 Welfare Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
H.2.1 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
H.2.2 Cost Effectiveness Calculations . . . . . . . . . . . . . . . . 218
H.3 Calculation of Pre- and Post-Treatment Water Consumption . . . . . 218
H.4 Sample Letters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
List of Tables
1.1 Effect of Temperature on Yields: Panel and Long Difference Estimates . . . . 32
1.2 Competition and Mitigation of Climate Shocks: Panel Approach with Yields 40
1.3 Competition and Mitigation of Climate Shocks: Panel Approach with Arrivals 43
1.4 Competition and Mitigation of Climate Shocks: Border Discontinuity . . . . 49
1.5 Effect of Competition on Prices Post Climate Shocks . . . . . . . . . . . . . . 56
1.6 Heterogeneous Impact of Climate Shocks on Input Usage and Crop Mix . . 61
2.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
2.2 Effect of Extreme Events on Population Growth . . . . . . . . . . . . . . . . . 103
2.3 The Death Toll from Floods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
2.4 Effect of Floods on Economic Activity . . . . . . . . . . . . . . . . . . . . . . 111
2.5 Effect of Extreme Rain on Economic Activity . . . . . . . . . . . . . . . . . . 112
2.6 Recovery Dynamics for Floods and Extreme Rain . . . . . . . . . . . . . . . . 115
2.7 Heterogeneous Effect of Floods based on Wealth and Risk . . . . . . . . . . 117
2.8 Dams and Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.1 ATE Estimates of Letters on Diagnostic Completion . . . . . . . . . . . . . . 137
3.2 ATE Estimates of Letters on Post-Treatment Consumption . . . . . . . . . . . 139
3.3 LATE Estimates of Diagnostic Completion on Post-Treatment Consumption 141
viii
3.4 Reweighted LATE Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
3.5 ATE Estimates of Reminders on Diagnostic Completion . . . . . . . . . . . . 147
3.6 Different Measures of Cost Effectiveness . . . . . . . . . . . . . . . . . . . . . 150
3.7 Comparison of Cost-Effectiveness with Other Studies . . . . . . . . . . . . . 152
3.8 Simple Benefit-Cost Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
3.9 MVPF Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
3.10 Effect of Temperature on Yields (Panel Approach): Robustness Tests . . . . 188
3.11 Effect of Temperature on Yields (Long-Differences): Robustness Tests . . . . 188
3.12 Effect of Out-of-State Competition on Mitigation of Climate Shocks . . . . . 189
3.13 Competition and Mitigation of Climate Shocks: Robustness Tests . . . . . . 190
3.14 Competition and Mitigation of Climate Shocks—Arrivals: Robustness Tests 191
3.15 Effect of Out-of-State Competition on Mitigation—Arrivals . . . . . . . . . . 192
3.16 List of countries and share of cities . . . . . . . . . . . . . . . . . . . . . . . . 202
3.17 Summary Statistics: Dams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
3.19 Effect of Floods based on Productivity . . . . . . . . . . . . . . . . . . . . . . 204
3.18 Effect of Floods on Economic Activity over Time . . . . . . . . . . . . . . . . 205
3.20 Baseline Balance Across Treatment Groups . . . . . . . . . . . . . . . . . . . 207
3.21 Statistics on Diagnostic Completion and Metered Households . . . . . . . . 209
3.22 Heterogeneous Treatment Effects Based on Pre-Treatment Usage . . . . . . . 211
3.23 LATE Estimates of Heterogeneous Treatment Effects . . . . . . . . . . . . . . 212
3.24 Characteristics of Households which Complete the Diagnostic . . . . . . . . 214
3.25 ATE Estimates of Letters on Interaction with Reminders . . . . . . . . . . . . 216
3.26 Parameters and Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
3.27 Cost Effectiveness in Ansink, Ornaghi, and Tonin 2021 . . . . . . . . . . . . . 219
3.28 Cost Effectiveness Calculations for Other Studies . . . . . . . . . . . . . . . . 220
3.29 Format of Consumption Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
ix
List of Figures
1.1 Climatological Changes Over India Between 1960-70 and 2010-20 . . . . . . 8
1.2 APMC Market Yards or Mandis in India . . . . . . . . . . . . . . . . . . . . . 18
1.3 Geographic Distribution of APMC Markets . . . . . . . . . . . . . . . . . . . 27
1.4 Percentage of Short-Run Impacts Offset by Adaptation . . . . . . . . . . . . . 36
1.5 Geographic Distribution of Competition Aggregated to District Level . . . . 38
1.6 Interpreting the Border Discontinuity Design . . . . . . . . . . . . . . . . . . 46
1.7 Geographical Distribution of Markets Selected Using 50 kms Bandwidth . . 46
1.8 Impact of Extreme Heat Offset by Competition: Border Discontinuity . . . . 50
2.1 Night Lights Before and After Floods in Wuhan: 2016 . . . . . . . . . . . . . 97
2.2 Pre-Trend Analysis of Night Lights Intensity for high- & low-income countries109
2.3 Night Lights Before and After Floods in Chennai: 2015-16 . . . . . . . . . . . 114
3.1 Fraction of Cropland in each ECMWF (Weather) Gridcell . . . . . . . . . . . 185
3.2 ICRISAT Districts and APMC States in Sample . . . . . . . . . . . . . . . . . 186
3.3 Coefficient Plot, GDD Distribution, and Extreme Heat Exposure by Season . 187
3.4 Urban Extent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
3.5 Vanilla (Status Quo) Mailer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
3.6 Simplified Mailer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
3.7 Altruism Mailer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
3.8 £10 Incentive Mailer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
3.9 £15 Incentive Mailer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
3.10 Moral Cost Mailer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
3.11 Reminder Email . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
x
Abstract
This dissertation comprises three chapters in environmental economics, in particular on
climate change adaptation and what factors positively influence or impede adaptation be-
havior. The first chapter documents that the pre-existing market power distortion in In-
dia’s intermediary market can hinder farmer climate change adaptation. Using the vari-
ation in the market power of local crop intermediaries due to historical agricultural laws,
I show that (1) farmers selling in the intermediary markets with less market power suf-
fer substantially less from extreme heat; (2) the farmer’s economic loss due to extreme
weather could be mitigated by 13.8 percent if the restrictions on cross-border trading are
removed.
The second chapter uses night lights data as a proxy for economic activity to test several
flood risk adaptation hypotheses. We first document heterogeneity in the economic impact
of floods across 9,500 cities around the globe. We document that floods cause less damage
to richer cities, showing the potential of economic growth in mitigating climate risk. Cities
with more past experience with floods, and cities with protective dams suffer less from
flooding. Finally, and this is by no means a policy recommendation, we also find that
authoritative regimes recover faster after floods. Population growth is lower in cities that
suffer from more floods. Richer cities suffer fewer deaths from flood events.
The third chapter documents how behavioral interventions can incentivize energy con-
servation. The paper is based in the United Kingdom where we attempt to incentivize
households to take up a water audit. We use financial incentives as a behavioral nudge
and compare them to other nudges such as those that evoke altruism and moral suasion.
We find that financial incentives work best, with participants in this group having a higher
xi
rate of take-up of the audit and also substantially reducing water usage. However, we also
find that in spite of these water savings, the financial incentives intervention does not pass
a benefit cost test because the greenhouse gas benefits associated with these reductions are
not large enough to compensate for the loss in revenue of utilities.
xii
Introduction
This dissertation is composed of three essays in environmental economics, particularly on
climate change adaptation and what factors positively influence or impede adaptation be-
havior. It seeks to identify how firms, households and farmers can mitigate or adapt to
environmental and climate challenges, what factors incentivize or distort adaptation be-
havior amongst economic agents, and what the implications of mitigation and adaptation
policies are for economic growth.
In Chapter 1, ”Does Market Power in Agricultural Markets Hinder Farmer Climate
Change Adaptation?” co-authored with Ruozi Song, we investigate the role that govern-
ment policies which distort market competition play in impeding farmers’ climate change
adaptation. We study this question in the context of India, where longer-run adaptation
to climate change has been inadequate — posing a considerable risk to its 250 million agri-
cultural workers. We exploit spatial discontinuities in intermediary market power, created
by state-level laws that restrict farmer- intermediary transactions to the same state, to de-
termine how spatial competition affects farmers’ adaptation. We find that a farmer selling
in the 75th percentile of the competition index compared to one that faces the 25th per-
centile of the competition index achieves a 4.9 percent higher output for each additional
day of extreme heat. This effect is driven by increased input usage by farmers in anticipa-
tion of higher prices after climate shocks, an effect limited only to high competition areas.
We then propose and estimate a quantitative spatial trade model with intermediary mar-
ket power to examine the welfare implications of higher competition for adaptation. Our
structural estimates suggest that the farmer’s economic loss due to extreme weather (i.e.
1
their climate damage function) could be mitigated by 13.8 percent if government regula-
tion distorting market competition is dismantled. These results highlight the importance
of understanding the political economy of reforming these competition-distorting laws to
accelerate climate change adaptation.
Apart from extreme heat, rising greenhouse gas emissions also raise the likelihood
of more extreme precipitation events that can increase the risk of local flooding. Urban
flooding poses danger to people and places. People can adapt to this risk by moving to
safer areas or by investing in private self-protection. Places can offset some of the risk
through urban planning and infrastructure investment. Chapter 2, titled ”Adapting to
Flood Risk: Evidence from a Panel of Global Cities”, which is joint work with Sahil Gandhi,
Matthew E. Kahn, Somik Lall, and Vaidehi Tandel, we test several flood risk adaptation
hypotheses. We construct a global city data set of 9500 cities for the years 2012 to 2018. We
find that population growth is lower in cities that suffer from more floods. Richer cities
suffer fewer deaths from flood events. Over time, the death toll from floods is declining.
Cities protected by dams experience faster population growth. Using lights at night to
measure short run urban economic dynamics, we document that floods cause less damage
to richer cities and cities with protective dams. Cities with more past experience with
floods suffer less from flooding.
Finally, even though adaptation is necessary, we also need to strive towards mitigating
carbon emissions. Energy utilities, specifically water suppliers in water-scarce regions are
showing greater interest in using non-price mechanisms that can help encourage conser-
vation. One such mechanism is a water audit, which involves assessing water use in the
home and providing tailored suggestions regarding how to conserve water. Yet, very lit-
tle is known about the efficacy, efficiency, and cost-effectiveness of water audits. Chapter
3, titled ”Do Water Audits Promote Economic Welfare? Evidence from a Natural Field
Experiment”, co-authored with Jesper Akesson, Robert W. Hahn, and Robert D. Metcalfe,
2
helps fill this research gap by implementing a natural field experiment in the United King-
dom. We implement a natural field experiment by randomly allocating 45,000 water cus-
tomers to a control group or to groups that are provided with different encouragements
to take-up an online water audit. Our analysis yields three main findings. First, provid-
ing participants with financial incentives to participate in the audit significantly increases
audit take-up, with an elasticity of around 0.5. Using the positive encouragements, we
find that the water audit reduces household water consumption by 17 percent for about
two months. Second, notwithstanding these large improvements in water conservation,
incentivizing uptake of the audit does not appear to pass a benefit-cost test. We also im-
plement a marginal value of public funds approach to considering benefit and costs, and
reach a similar conclusion. Third, we find that targeting of high users could double the
effectiveness of the financial incentive interventions.
3
Chapter 1
Does Market Power in India’s Agricultural Markets Hinder
Farmer Climate Change Adaptation?
1
1.1 Introduction
This paper analyzes whether market distortions induced by a country’s institutions in-
hibit its adaptation to climate change. Though the negative impacts of a departure from
perfect competition are well documented Arrow 1962; Ashenfelter, Hosken, and M. Wein-
berg 2014, these detrimental effects risk being exacerbated by the climate crisis. One major
source of these distortions is government regulations which can concentrate market power
in the hands of a few economic agents. In light of this, do institutional policies, that dic-
tate an agents’ market power, impede climate change adaptation? And would eliminating
these distortions by establishing free markets enhance welfare by aiding adaptation? In
this article, we address these questions in the context of competition in India’s agricultural
markets, and the role of market power, in facilitating farmer adaptation to climate change.
Our analysis is motivated by a simple observation: in a world where climate change
will result in crop production losses and fall in agricultural productivity IPCC 2022, farmer
adaptation is crucial, and critically depends on institutional and policy constraints. A
1
Co-authored with Rajat Kochhar
4
country’s agricultural policy can play a dominant role in building resilience and reduc-
ing exposure to the impacts of climate change—with the potential to either advance or
distort (through the imposition of soft limits
2
) adaptation behavior Mees 2017; Valdivieso,
Andersson, and Villena-Rold´ an 2017; Oo, Van Huylenbroeck, and Speelman 2017.
3
For
instance, Annan and Schlenker 2015 show that federal crop insurance policy in the United
States creates a moral hazard problem, disincentivizing adaptation and consequently ex-
acerbating losses. Similarly, agricultural laws in India, which create market power for
agricultural intermediaries Chatterjee 2019, may also disincentivize adaptation. Consider
the case where post climate shock adaptation may be dependent on higher input usage,
which in turn is contingent on higher expected prices. The market power of intermedi-
aries may, however, constrain farmer prices from rising beyond a level that impels farmers
to adapt to climate shocks. Thus, the impact of climate change on agriculture is inherently
dependent on the capacity to effectively adapt. But how and to what extent this capacity is
constrained by government-induced distortions to market competition remains an open
question.
Addressing these questions empirically poses three challenges: first, competition is
not directly observable, making it difficult to credibly measure its intensity OECD 2021;
second, causal identification of competition on adaptation suffers from both potential en-
dogeneity in competition and in adaptation response; and lastly, limited simultaneous
variation in climate shocks and market competition makes it challenging to detect any sig-
nificant causal effects, should they exist. We tackle these challenges by studying the impact
of spatial competition between intermediaries on farmer adaptation in India, focusing on a
law that restricts farmers to selling their produce to intermediaries within their own state.
The Indian context affords us progress on all three challenges.
2
Soft adaptation limit is defined as the existence of adaptive options to avoid intolerable risks, but which
are currently unavailable.
3
See Kuruppu and Willie 2015 and S.-a. Robinson 2018 for a discussion on how the governance archi-
tecture can act as a bottleneck to adaptation in small island developing states.
5
Central to our approach are the state-specific Agriculture Produce and Marketing Com-
mittee (APMC) Acts which regulate the first sale and purchase of agricultural commodi-
ties within each state in India. Two provisions in these laws are noteworthy for our pur-
pose: first, farmers in a state are restricted to sell their produce at government designated
physical markets (known as mandis) within their own state; second, output can only be
sold to government-licensed intermediaries, each of whom requires a market-specific li-
cense to operate in the respective APMC mandi.
4
Importantly, other wholesalers, retail
traders, or food processing companies cannot buy directly from the farmer. Therefore, the
spatial arbitrage constraints imposed on farmers by the law — restricting access to licensed
intermediaries within a state border — reduce the competition faced by intermediaries. In
essence, state-specific institutional setup governing the sale of agricultural output gener-
ates spatially varying monopsony power for licensed intermediaries, a source of variation
which we can exploit to address the empirical challenges.
The context allows us, first, to accurately measure competition at the mandi level, of-
fering spatially granular variation in competition intensity between intermediaries. We
collect novel microdata from India on the geolocations of mandis, and combine it with
the daily quantity arrivals and prices of agricultural produce there-within. Subsequently,
drawing on the standard measure of market access in trade literature Donaldson and
Hornbeck 2016; Allen and Atkin 2016, we measure the competition intensity faced by each
intermediary as the inverse-distance weighted sum of the value of trade at all other mar-
kets near a origin market site, but in the same state.
5
4
Intermediaries, or middlemen, tend to be the principal buyers of farmers’ output in developing coun-
tries Reardon 2015. The license to operate in a mandi is provided to them by the APMC board under whose
jurisdiction the mandi falls. Unlike farmers, there are no sale restrictions on the intermediary, who is free to
transport the purchased produce and sell it to retailers all over the country.
5
Chatterjee 2019 defines spatial competition as the number of markets in the neighborhood of each mar-
ket weighted by the inverse of their distance. Thus, there is no variable controlling for the size of the markets
in his measure. This is similar to the market potential measure in C. D. Harris 1954, who defines it as the sum-
mation of markets accessible to a point divided by their distances from that point. Similarly, Macchiavello
and Morjaria 2021 use the number of proximate competitors as a measure of competition. We employ these
different measures for testing the robustness of our results.
6
Second, the interstate trade restrictions on farmers help us overcome the potential en-
dogeneity in the location of intermediaries. Potential bias in estimating the impact of mar-
ket power on adaptation can arise if, for instance, markets were placed in areas with higher
predisposition for farmer innovation. The APMC Acts establish a discontinuity at the bor-
der in the competition faced by intermediaries. This allows us to employ a hybrid border-
discontinuity design with market pairs, akin to Chatterjee 2019. We form market pairs by
matching mandis that are in close proximity with each other but lie on different sides of a
state border, thereby allowing us to difference out unobserved factors, other than compe-
tition, that affect adaptation.
Third, India, with an estimated 263 million agricultural workers Census of India 2011
spread across 15 agro-climatic regions Ahmad, Habib Kanth, Parvaze, and Sheraz Mahdi
2017—each with substantial spatial heterogeneity in competition intensity—offers signifi-
cant variation to study the effect of market power on farmer adaptation to climate change.
6
Agricultural households, which account for 48 percent of total households in India NABARD
2018, have been incentivized to invest in an adaptation portfolio owing to an unprece-
dented increase over the past several decades in both maximum temperature, and fre-
quency and intensity of extreme heat days (R. Krishnan, Sanjay, Gnanaseelan, Mujumdar,
Kulkarni, and Chakraborty 2020; also see 1.1).
7
Notably, this effect is expected to worsen,
with India projected to have the highest climate-change induced increase in heat expo-
sure and vulnerability to crop production losses relative to other nations IPCC 2022; Jones,
6
India has the highest number of agricultural workers in the world. 263 million people (54.6 percent
of India’s total workforce) are employed in agriculture. The figure of 263 million comprises of 119 million
cultivators/farmers and 144 million agricultural laborers. There is, however, some debate about the total
farmer population in India, with official figures ranging between 100 and 150 million. The main source of
contention is the absence of a standard definition of who constitutes a farmer. See Damodaran 2021 and
Narayanan and Saha 2021 for a detailed discussion.
7
An agricultural household is defined as a household that received some value of produce more than 5000
(equivalent to US$63 using the average USD-INR exchange rate in calendar year 2021) from agricultural
activities (e.g., cultivation of field crops, horticultural crops, fodder crops, plantation, animal husbandry,
vermiculture, sericulture, etc.) and had at least one self-employed member in agriculture, either in principal
status or in subsidiary status during last 365 days.
7
Tebaldi, O’Neill, Oleson, and Gao 2018.
8
This is relevant because it motivates our use of
extreme heat (defined as temperatures ≥ 35
◦
C or 95
◦
F) as a proxy for climate shocks.
(a) Change in Maximum Temperature (b) Change in Extreme Heat Days
Figure 1.1: Climatological Changes Over India Between 1960-70 and 2010-20
Note: The weather data for Panel (a) comes from Terraclimate 2018, which has a monthly temporal resolution and a 4-km (1/24th
degree) spatial resolution. Change in maximum temperature is calculated by taking the average maximum temperature for each grid
point within Indian boundaries for two time periods: 1960-70 and 2010-2020; and then differencing the two. The weather data for
Panel (b) is sourced from India Meteorological Department 2009 which uses 395 weather stations to provide a 1× 1 degree gridded
daily temperature dataset starting from 1951 up until 2020. Extreme heat days were defined for each grid cell as days with maximum
temperature greater than the 95th percentile of the temperature distribution in the respective grid cell between 1950-2020. Change in
number of extreme heat days is calculated by taking the total number of extreme heat days between 1960-70 and comparing the same
to the total number of extreme heat days between 2010-20.
The empirical analysis proceeds in three steps. 1.4.1 motivates our core question —
the role of market power in adaptation — by asking if Indian farmers have adapted in
the long-run. Evidence to the contrary would indicate that constraints imposed by distor-
tionary institutional policies on adaptation may be persistent and binding; 1.4.2 explores
8
India’s average temperature has risen by approximately 0.7
◦
C between 1901–2018. By the end of this
century, the average temperature across India is projected to rise by 4.4
◦
C relative to the 1976–2005 average,
under the RCP8.5 scenario R. Krishnan, Sanjay, Gnanaseelan, Mujumdar, Kulkarni, and Chakraborty 2020.
Furthermore, the frequency of summer heat waves over India is projected to be 3 to 4 times higher by the end
of the 21
st
century under the RCP8.5 scenario, as compared to the 1976–2005 baseline period. The average
duration of heat wave events is also projected to approximately double Rohini, Rajeevan, and Mukhopadhay
2019. Finally, Mishra, Smoliak, Lettenmaier, and Wallace 2012 and A. G. Turner and Annamalai 2012 project
a steady decline in the total precipitation during monsoon months.
8
whether intermediary market power mitigates the deleterious impact of climate shocks;
1.4.3 investigates the mechanisms.
We begin by documenting evidence of limited long-run yield-stabilizing adaptation in
India. Following Burke and Emerick 2016, we measure long-run adaptation as the differ-
ence between panel and long-differences estimates of the effect of extreme heat on yields.
The panel estimates capture short-run within-year adjustments by farmers, while the long-
differences estimates encapsulate long-run transformational adaptations. Their difference,
thus, reflects the share of the short-run impacts that are offset in the longer run. Using fine
geospatial crop yields and weather data from 1968 to 2017, we find that both methods yield
significant but similar estimates — each additional day of extreme heat reduces yields by
1.0 to 2.7 percent — indicating that long-run adaptations have likely offset none of the
short-run impacts of adverse climate. Therefore, the bottlenecks farmers face in adopting
short-run strategies have a direct and cumulative impact on their ability to adapt in the
long-term, making it imperative to recognise and address these constraints.
Our core result is that market power of intermediaries arising out of institutional poli-
cies acts as a major constraint on the farmers’ post-climate shock adaptation efforts. Using
a hybrid border-discontinuity design, we find that a farmer selling in the 75
th
percentile of
competition compared to one that faces the 25
th
percentile of competition achieves a 4.5
to 5.2 percent higher output on average for each additional degree-day of extreme heat.
This result is robust to different distance thresholds, ranging from 25 to 50 kilometers, be-
tween market pairs. We corroborate these findings using (i) a panel approach, and (ii)
a panel approach but changing the spatial unit of analysis from a mandi to a district. As
before, the results unequivocally indicate that monopsony power thwarts adaptation — a
one standard deviation increase in competition helps a farmers alleviate between 15 and
37 percent of the negative impact of extreme heat.
9
Next, in order to investigate the mechanisms underlying the relationship between mar-
ket power and adaptation, we build a simple agricultural household model with incom-
plete input markets Arag´ on, Oteiza, and Rud 2021. This allows us to derive predictions on
how and when farmers would invest in their adaptation portfolio in the event of an exoge-
nous negative shock. Subsequently, we provide evidence consistent with the predictions
highlighted in the model.
The model yields the following prediction — in the event of a negative weather shock,
farmers could increase their input usage if the output prices are expected to rise beyond a
certain threshold. This could happen, for instance, if extreme heat reduces yields and,
hence, aggregate supply. However, the magnitude of increase in prices is likely to be
greater in high competition areas as a large set of intermediaries now compete for lower
output. Lower spatial competition between intermediaries translates into lower farmer
prices Chatterjee 2019. We hypothesize that climate shocks interact with market power to
further exacerbate these pre-existing distortions, thereby incentivizing only the farmers in
high competition areas to adjust their input usage as a response to extreme heat. This, in
turn, helps alleviate the crop production losses associated with heat stress.
We find evidence consistent with the mechanisms highlighted in the model: (i) the
positive effect of higher competition on intermediary prices is compounded after a cli-
mate shock, and (ii) farmers in high competition areas increase their input usage, indi-
cating post climate shock adaptation. Specifically, a one standard deviation increase in
competition causes the pre-existing difference in crop prices to increase by 0.5 to 0.6 per-
centage points, conditional on both areas being exposed to a week of extreme heat. Next,
we use household level survey data to show that this rise in prices incentivizes farmers
to increase their input use within the growing season. Our estimates suggest that a one
standard deviation increase in competition leads to a 1.2 and 1.7 percent increase in land
and labor inputs, respectively, for each additional day of extreme heat. Furthermore, in-
put costs associated with labor, irrigation, fertilizers, and farm equipment also experience
10
a significant increase. Consistent with the adaptation portfolio, we also find evidence
of crop diversification at a macro-scale (i.e., district-level) in high competition areas, in-
dicating crop-mix as a potential avenue for increased resilience. In summary, productive
adjustment, incentivized by increasing prices, attenuates undesirable drops in output, but
is limited to high competition areas.
One way to counter the market power distortions generated by archaic institutional
policies is to remove the inter-state trade restrictions. However, the welfare impacts of
such a policy change cannot be deduced directly from the data, and require a structural
model. Specifically, our empirical strategy is inadequate in encapsulating three pivotal
general equilibrium effects. First, removing trade restrictions will not only affect prices
in mandis near the state borders, but also have a knock-on effect on prices in markets that
are not in close proximity to state borders. Second, change in intermediary prices will
incentivize farmers to re-optimize their choice of crops, intermediate inputs, and mar-
ket for sale. This will alter supply, thereby impacting retail prices, which will eventually
feed back into mandi prices. Finally, and more importantly, climate change will influence
productivity differences across crops and fields, altering comparative advantage between
different regions of India. This evolution of comparative advantage will interact with a
change in market power of intermediaries, shaping the adaptation portfolio of farmers.
A model, therefore, helps us understand how the policy change aids in mitigating the
consequences of climate change.
To estimate adaptation gains from removing interstate trade restriction, we develop
a spatial general equilibrium model of trade in the agricultural markets, drawing on the
work of Costinot, Donaldson, and C. Smith 2016 and Chatterjee 2019. In our framework,
every state consists of a large number of fields with heterogeneous productivity across
multiple crops. Each field is represented by a farmer who makes two decisions: (i) crop
and input choice, and (ii) intermediary market for sale post-harvest. The former decision
11
is influenced by the relative productivity differences across crops and fields, i.e. com-
parative advantage, which determines the pattern of specialization within and between
states. The latter decision relies on the farmers’ transportation costs between competing
markets — each of which is represented by an intermediary — and determines the level of
market power. In order to ensure the model resembles reality, we add three key features.
First, farmers cannot cross state border. To incorporate these trade restrictions, we assume
transportation costs are infinite if the farm and market lie on different sides of the border.
Second, intermediaries are price makers. This is modeled through a Bertrand competition
which ensures intermediaries act strategically when purchasing crops, internalizing their
market power. Third, intermediaries are allowed to sell across state borders. Therefore,
geography and trade restrictions create spatial heterogeneity influencing farmers’ arbi-
trage opportunities, and consequently, creating spatial variation in the monopsony power
that intermediaries can exert.
The competitive equilibrium of our model and any subsequent counterfactual analy-
sis will depend on five key parameters: (i) the elasticity of substitution between different
varieties of the same crop; (ii) the elasticity of substitution between different crops; (iii)
within-field heterogeneity in productivity; (iv) trade costs and; (v) dispersion of idiosyn-
cratic shocks to the trade cost. All these parameters are estimated using a rich micro-level
data set on field level crop productivity, inland trade data on agricultural commodities,
geolocation of markets, as well as prices and quantity arrivals of different crops in each
market. Finally we use the estimated parameters and information on the pattern of com-
parative advantage across fields and crops to simulate our model under the no–climate
change scenario, and compare it to two counterfactual scenarios. In the first, we study the
welfare consequences of a decline in crop and field productivity due to climate change,
but in the presence of trade restrictions on farmers. In the second counterfactual scenario,
we study the welfare consequences of climate change but without trade restrictions. The
12
difference between the two counterfactual helps us ascertain the magnitude of mitigation
that transpires once trade restrictions are removed.
Our model suggests that the welfare impact of climate change is substantially miti-
gated once inter-state trade restrictions are lifted. Specifically, we find that climate change
reduces welfare in India by 2.1 percent of total GDP, assuming no policy change. How-
ever, increase in competition arising out of abolishing trade barriers enables farmers to
receive a higher price, which changes the source and magnitude of adjustment, allowing
a 13.8 percent alleviation in the welfare losses. This illustrates how market distortions cre-
ated by government policies could hinder adaptation, and how removing the same could
expand the adaptation portfolio of farmers, thus helping countries mitigate the negative
consequences of climate change.
Related Literature: This paper contributes to several strands of literature. First, it
contributes to the broader literature on the impact of market concentration on economic
outcomes. A large body of research shows that market power has negative consequences
for consumer surplus Dafny, Duggan, and Ramanarayanan 2012; N. H. Miller and M. C.
Weinberg 2017, economic inequality Comanor and Smiley 1975, employee welfare Prager
and Schmitt 2021, as well as productivity and innovation Aghion, C. Harris, Howitt, and
Vickers 2001; Aghion, Bloom, Blundell, Griffith, and Howitt 2005; Holmes, Levine, and
Schmitz 2012. Interestingly, impeding competition is also linked with anti-democratic out-
comes like concentrated economic and political power, political instability, and corruption
Becker 1958; J. A. Robinson and Acemoglu 2012. We instead focus on the role of market
competition in incentivizing adaptation to climate change. In line with previous studies,
we find that (intermediary) market power has harmful implications, and can put soft lim-
its on adaptation in agriculture.
Second, this work relates to the literature on inefficiencies generated by government
policies and institutional features. An extensive literature has documented the adverse ef-
fects of government regulations: labor regulations hurt output and productivity Holmes
13
1998; Besley and Burgess 2004; licensing regulations which restrict firm entry lead to mar-
ket concentration, decelerate employment growth, and increase corruption Djankov, La
Porta, Lopez-de-Silanes, and Shleifer 2002; Bertrand and Kramarz 2002; product market
regulations (e.g. trade tariffs) adversely impact competition, average firm size and profits
Blanchard and Giavazzi 2003, and; cost-of-service regulations in the utilities sector reduce
efficiency Fabrizio, N. L. Rose, and C. D. Wolfram 2007; Cicala 2022. We complement this
literature by finding evidence that regulations governing the sale and purchase of agri-
cultural products can distort competition and disincentivize adaptation. In this regard,
our study is closest to Annan and Schlenker 2015 who find that a highly subsidized crop
insurance program in the United States, providing coverage to farmers against crop losses,
inhibits adaptation. However, the disincentive to adapt in their setup is a result of moral
hazard, while the disincentive in our setting is driven by government induced distortions
in market power of intermediaries. Thus, our paper documents how government regu-
lations, intended to protect farmers from exploitation by middlemen, have inadvertently
distorted competition and hindered adaptation, thereby exacerbating the dead-weight loss
arising from climate change.
Third, we contribute to the literature on adaptation to climate change, and the mecha-
nisms that underpin it. There is mounting evidence on the deleterious impact of climate
change on several economic indicators like productivity, education, health, etc. Burgess,
Deschenes, Donaldson, and Greenstone 2017; Park, Goodman, Hurwitz, and J. Smith 2020;
E. Somanathan, R. Somanathan, Sudarshan, and Tewari 2021.
9
As a natural progression,
subsequent studies have focused on adaptation efforts, i.e. how these damaging effects
can be mitigated. Researchers have documented the positive role of air conditioners Bar-
reca, Clay, Deschenes, Greenstone, and Shapiro 2016; Zivin and Matthew E Kahn 2016,
9
There are numerous studies on the potential impact of climate change on agriculture in India Guiteras
2009; Mall, R. Singh, Gupta, Srinivasan, and Rathore 2006; Economic Survey of India 2018 and the United
States Mendelsohn, W. D. Nordhaus, and Shaw 1994; Schlenker, Hanemann, and Fisher 2005; Deschˆ enes and
Greenstone 2007; Fisher, Hanemann, M. J. Roberts, and Schlenker 2012; Schlenker and M. J. Roberts 2009. A
review of the impact of global warming on agriculture in developing countries is provided by Mendelsohn
2009.
14
expansion of bank branches Burgess, Deschenes, Donaldson, and Greenstone 2017, and
relocation Deschenes and Moretti 2009 in combating mortality and productivity losses
caused by climate change. Agricultural adaptation has been linked with changing crop-
mix Auffhammer and Carleton 2018; Taraz 2018, using drought-tolerant seeds Boucher,
Carter, Flatnes, Lybbert, Malacarne, Marenya, et al. 2021, labor input adjustments Arag´ on,
Oteiza, and Rud 2021, and migration Feng, Oppenheimer, and Schlenker 2015; Hermans
and McLeman 2021. In a similar vein, this article finds that farmers rely on input adjust-
ments and changing the crop mix to attenuate losses arising from climate shocks. Impor-
tantly, however, this adaptation portfolio is only accessible to farmers in high competition
areas — on account of higher expected prices — and not where government policies have
distorted market power.
Finally, our paper is related to the growing body of literature focusing on trade and
adaptation to climate change Costinot, Donaldson, and C. Smith 2016; Reilly and Hohmann
1993; Randhir and Hertel 2000. While the literature has focused on how international trade
can help alleviate climate change losses, we show that removing domestic trade barriers
would also go a long way in accelerating adaptation efforts. In this regard, we build on the
quantitative spatial general equilibrium model of Costinot, Donaldson, and C. Smith 2016
by moving away from the assumption of perfectly competitive environment, and adding
spatial variation in market power of intermediaries. This allows us to quantify the wel-
fare gains from adaptation to climate change through a reduction in intermediary market
power, an outcome of dismantling domestic trade barriers.
Roadmap: The organization of the paper is as follows. 1.2 provides an overview of the
institutional background of agricultural trade in India, particularly the APMC markets.
In 1.3, we describe our data sources and the construction of variables. 1.4 presents the
empirical strategy and the results from our econometric analysis. A theoretical model of
climate change and trade is laid out in 1.5, its estimation in 1.6, and the counterfactual
analysis in 1.7. Conclusions and areas for future research are discussed in 3.5.
15
1.2 Background on Agricultural Markets in India
To help understand how government regulations concerning agriculture marketing cre-
ated spatial competition distortions amongst intermediaries, i.e. the paper’s specific con-
text, we provide a detailed overview of the origin of these laws. 1.2.1 delves into the his-
tory, 1.2.2 details the provisions, while 1.2.3 provides insight into the unintended conse-
quences of these provisions.
1.2.1 History
The regulation of agricultural marketing in India has its roots in pre-independence poli-
cies introduced during the British Raj. The British government wanted to ensure sustained
supplies of cotton at reasonable prices for textile mills in the United Kingdom. In order
to facilitate this, the first regulated cotton market was set up in Karanja (Maharashtra)
in 1886. Subsequently, the Berar Cotton and Grain Market Act, 1887 was introduced which
empowered the British to establish a trading supervisory committee and, thereafter, des-
ignate any place as a market for sale and purchase of agricultural produce within a district.
In 1928, under the chairmanship of Lord Linlithgow,
10
the British government’s Royal
Commission on Agriculture in India expanded the scope of regulated markets. Simply, the
commission recommended: (i) extending regulation of marketing practices to all crops,
and (ii) establishment of regulated markets. To quote from the report:
“It is only in Berar that the constitution of markets is regulated by special leg-
islation and that the management is in the hands of elected committees. ...
The most hopeful solution of the cultivator’s marketing difficulties seems to lie
in the improvement of communications and the establishment of regulated
markets, and we recommend for the consideration of other provinces, the es-
tablishment of regulated markets on the Berar system. ... The Bombay Act is,
10
Lord Linlithgow was Governor General and Viceroy of India from 1936 to 1943.
16
however, definitely limited to cotton markets and the bulk of the transactions
in Berar markets is also in that crop. We consider that the system can conve-
niently be extended to other crops. ... We consider that the management of
these markets should be vested in a market committee.”
In pursuance of these ideals, the British Government in India circulated a Model Bill in
1931 to regulate trade practices and establish market yards in the countryside. However,
only a few provinces adopted these laws (Central Provinces, Madras, Baroda, Bombay,
Punjab, and Mysore). At its core, however, the establishment of regulated markets under
the British was intended to control the price, quantity, buyer, and type of goods sold, with
the direct aim of ensuring cheap supplies for England.
Post independence, the focus of the government shifted towards incentivizing farmers
to heighten agricultural production. Moreover, the government sought to protect cultiva-
tors from exploitative middlemen who often forced farmers to sell at low prices. In pursuit
of this objective, government regulation was seen as an effective instrument to facilitate
fair and competitive compensation for farmers. Consequently, a large number of states
enacted and enforced the Agriculture Produce Marketing Regulation (APMR) Acts from
the late 1960s to the early 1980s.
11
The provisions within these Acts, and how they create
monopsony power for intermediaries, is explained in the following subsection.
1.2.2 Agricultural Produce Market Committee: Regulations
Agriculture is a state subject under the Indian constitution, i.e. states have the power and
responsibility to legislate on agricultural marketing. In accordance with these legislative
powers, each state has enacted laws under the APMR Act to regulate agricultural trade
within its boundaries. These laws permit state governments to designate certain areas
within the geographical confines of the state as market areas ( mandis). Each market area
is governed by an Agricultural Produce Marketing Committee (APMC) — constituted of
11
All states, except Kerala, Jammu and Kashmir and Manipur, enacted such laws.
17
elected traders, farmers, and government representatives from the area — which is tasked
with framing and enforcing the rules governing agricultural marketing. The committee is
also responsible for setting up market yards where agricultural trade takes place.
These state-specific Acts mandate that the sale or purchase of agricultural commodi-
ties can only be executed in specified market areas, yards, or sub-yards located within the
state (see images in 1.2). In particular, it requires that all food produce should be brought
by the farmers to a market yard in their region and then sold through an auction. Fur-
thermore, intermediaries (buyers) who wish to trade in a certain market area are required
to obtain a license from the market committee. Additionally, the Act also mandates that
sellers and traders pay a market fee on all trade that takes place within the market area.
This institutional setup was designed to ensure that farmers had access to organized mar-
kets operating under the supervision of the government; such oversight was intended to
minimize the risks of exploitation by traders and middlemen. However, the provisions
distorted market competition, which we discuss in detail below.
(a) APMC Market in Bhatinda, Punjab (b) APMC Market in Yavatmal, Maharashtra
Figure 1.2: APMC Market Yards or Mandis in India
Note: Panel (a) and Panel (b) show two designated APMC market yards, also known as mandis, which were established under the
state-specific Agricultural Produce Marketing Committee (APMC) Acts. These yards are the first point of contact between the
farmers and intermediaries. All agricultural produce must be brought to these mandis by farmers in that region, and sales are made
through auction. Intermediaries require a license to operate within a mandi, but are free to transport the purchased produce and sell
it across the country.
18
1.2.3 Monopsony Power
Though noble in their intentions, the APMC laws introduced an unintended consequence:
monopoly power for market committees in their respective area. APMC legislation crim-
inalizes setting up competing markets and buying agricultural produce from outside the
designated market yards. Importantly, as the APMC laws are state-specific, their jurisdic-
tion does not stretch beyond state boundaries. This, coupled with the requirement that
farmers can only sell their produce in the APMC of their region, implies that farmers can-
not cross state boundaries to sell their produce. In essence, jurisdictional boundaries and
strict market regulations distort competition. This negatively impacts farmers’ bargaining
power and, consequently, lowers the probability of receiving fair prices for their produce.
Along with between-market competition, within-market competition is also impacted
by collusion amongst traders. Market committees, responsible for granting licenses, are
usually dominated by the trader lobby. This creates a conflict of interest as existing traders
prevent market entry to preserve their profits. The licensing regime, therefore, artifi-
cially reduces the number of buyers in the market. Furthermore, since wholesalers, retail
traders, and large processors cannot buy directly from the farmers, they rely on licensed
traders to act as intermediaries. This behavior also impacts prices: various studies Abhi-
jit Banerji and Meenakshi 2004; Meenakshi and Banerji 2005 document non-transparent
price discovery processes resulting from trader collusion. This ultimately renders farm-
ers subject to exploitation by intermediaries who act as financiers, information brokers,
and traders. Notably, farmer exploitation creates further opportunities for rent-seeking as
intermediaries can buy low and sell high, capturing the difference as profits.
In summary, while the APMC laws were intended to protect farmer exploitation by reg-
ulating agricultural marketing, many exploitative conditions have gradually resurfaced,
mostly as unintended consequences of these laws. They limit between-market competi-
tion by creating legal barriers to entry, prohibiting farmers to sell outside APMC markets,
and restricting the set of buyers to licensed intermediaries within the state. One way to
19
counter this would be through within-market competition among intermediaries. How-
ever, collusion among traders is rampant, with evidence of price manipulation and re-
stricted buyer entry, effectively creating a monopsony. The net result is an exploitative
system of interlocked transactions that robs farmers of discretion across important selling
decisions.
1.3 Data
Our goal is to study the extent of long-run adaptation, and the role of institution-led dis-
tortions in market competition on adaptation to climate change. To this end, we need four
main types of data, which we draw from varied sources: (i) estimates of yields for our
sample of crops; (ii) weather data to construct estimates of climate shocks; (iii) location
of intermediary markets (mandis) to construct the competition measure, and; (iv) daily
prices and arrivals (quantity of crop brought to a mandi) data for the sample of crops.
Below, we provide detailed information on all the data sources.
1.3.1 Yields
Our agricultural data on yields comes from the International Crop Research Institute for
the Semi-Arid Tropics ICRISAT 2018. In collaboration with the Tata Cornell Institute of
Agriculture and Nutrition (TCI), ICRISAT 2018 provides district level data on area (’000
ha), production (’000 tons), and yields (kg/ha) for 19 major crops in 313 districts of 20
states of India at an annual level from the year 1966 to 2017.
12
Our unit of analysis is,
thus, the crop-district-year. There were 313 districts in 16 states in 1966.
13
Over the next
50 years, four new states were created from the 16 states to make it 20 states, with the
number of districts in the 20 states increasing to 571. The database is, therefore, divided
12
As of 2022, India is divided into 28 states and 8 union territories, with the states being further subdi-
vided into 776 districts. Year refers to the agricultural year, i.e. June 1
st
to May 31
st
Sanghi, K. K. Kumar, and
McKinsey Jr 1998.
13
This excludes northeastern states (except Assam) and Jammu and Kashmir.
20
into 2 datasets: apportioned and unapportioned. Apportioned includes only the 1966 base
districts, with data on districts formed after 1966 given back to their parent district. This
has resulted in a consistent and comparable time series data for all the districts since 1966.
Unapportioned, on the other hand, includes all the districts formed until 2015 in 20 states
of India, but it only spans the years 1990 to 2015. We, thus, use the apportioned dataset for
our analysis given its longer time horizon.
14
We divide the crops based on the growing season, of which there are two main ones
in India: Kharif and Rabi. The Kharif cropping season is from July–October during the
south-west monsoon, with crops harvested from the third week of September to October.
The Rabi cropping season is from October–March during winter, with harvesting in the
spring months between April and May. The ICRISAT 2018 database does not separate
the agricultural data by growing season (except for sorghum or millets). Therefore, we
do not have estimates of what proportion of the crop was grown in each season. This is,
however, necessary to ascertain as otherwise there is a risk of yielding spurious estimates
of the relationship between climate variables and agricultural production. For instance,
modeling yearly yields of a crop as a function of annual number of extreme heat days will
be invalid if the production was, predominantly, limited to one growing season. Favor-
ably for us, the production of most crops is concentrated to one of the two seasons, with
negligible cultivation in the other. Agricultural Statistics at a Glance 2020, released by the
Directorate of Economics and Statistics, Government of India, provides all India estimates
of agricultural production of crops by season, averaged between the years 2014 to 2019.
We use the share of production in each season to classify commodities into either of the
two cropping seasons. For e.g., if a majority of the total production of a crop was con-
densed to the Kharif season, we classify the crop as Kharif. Of course, given India’s varied
topography and climate, there could be variation in the cropping season for the same crop
across different regions. We address this issue in 1.3.4.
14
For a description of the methodology for apportioning newly formed districts to their parent district,
and a list of districts formed after 1966, see Appendices 1 and 2 of ICRISAT 2018.
21
1.3.2 Weather
Our climate data are drawn from European Centre for Medium-Range Weather Forecasts
(ECMWF), an independent intergovernmental organisation and research institute head-
quartered in the United Kingdom. We use the fifth generation of ECMWF atmospheric
reanalyses of the global climate ERA5-Land 2021 dataset that provides gridded temper-
ature (Kelvin) and precipitation (depth in metres) data at a 0.1
◦
× 0.1
◦
(9km) horizontal
resolution.
15
The data is made available at an hourly temporal resolution with coverage
from January 1950 to present.
There was a mismatch in spatial resolution between weather and agricultural data: the
former was available at a very high spatial resolution (9km× 9km grid cells), while the
resolution of the latter was coarser and aggregated to a bigger administrative unit (district
level). This implies that several weather grid cells fell within the boundaries of each dis-
trict. To address this, we take a weighted mean of the temperature (weighted sum in case
of precipitation) across all cells within the district. In order to calculate weights, note first
that districts in India can be fairly large with heterogeneous geographical features, and
contain areas with little to no agricultural activity (e.g. Himalayas in North and East In-
dia, or deserts in Gujarat and Rajasthan). Consequently, weather conditions in such parts
of the district may be irrelevant for agricultural production within that unit. Therefore, we
rely on fine scale land cover data to use as an aggregation weight. Specifically, we use the
Global Food Security-support Analysis Data at 30m resolution GFSAD30 2017 which pro-
vides satellite-derived cropland extent maps in collaboration with National Aeronautics
and Space Administration (NASA) and the United States Geological Survey (USGS) for
South Asia for the year 2015. The database divides land into three categories: water (ocean
15
Temperature of air measured at 2m above the surface of land, sea or in-land waters. Temperature
measured in Kelvin was converted to degrees Celsius (
◦
C) by subtracting 273.15.
22
and water bodies), non-cropland, and cropland. For our purpose, all weather variables
were aggregated based on weights proportional to the cropland extent (see 3.1).
16
Next, we provide details on the construction of the weather variables used in the em-
pirical analysis. Schlenker and M. J. Roberts 2009 have documented strong non-linearities
in the relationship between exposure to weather conditions and agricultural outcomes. To
capture this, we use the concept of Growing Degree Days (GDD), which measures cumula-
tive temperature exposure between two temperature thresholds during a period of time.
The process of creating exposure bins for all district-month-year combinations involved
the following steps. First, we use the hourly cropland-weighted weather data, aggregated
to a district level, to calculate the daily minimum and maximum temperature for each
district in India. Next, we derive how much time is spent at each temperature bin for all
districts. These bins were 1
◦
C wide, ranging from−10
◦
C to 50
◦
C. Finding the number of
hours a district is exposed to each 1
◦
C interval requires intra-daily distribution of temper-
ature, which required making assumptions about the temperature-time path. Specifically,
the distribution of temperatures within each day was approximated using a sinusoidal
curve Ortiz-Bobea 2021, which generates a series of points at 15-minute intervals, between
minimum and maximum temperatures of each day. Following this, we computed the ex-
posure bins (measured in hours) by determining the frequency of these 15-minute interval
points throughout the month.
17
As a final step, we compute growing degree days from
16
Cropland extent was defined as lands cultivated with plants harvested for food, feed, and fiber, includ-
ing both seasonal crops (e.g., wheat, rice, corn, soybeans, cotton) and continuous plantations (e.g., coffee,
tea, rubber, cocoa, oil palms). Cropland fallows are lands uncultivated during a season or a year but are
farmlands and are equipped for cultivation, including plantations (e.g., orchards, vineyards, coffee, tea,
rubber). Further details are available at globalcroplands.org.
17
By construction, summing over all bins across a month for a district equals the number of hours in that
month.
23
these exposure bins by converting the number of hours in each exposure bin to days (di-
vide by 24), and subsequently aggregating them between a low thresholdh and a high
thresholdh using the expression:
GDD
htoh
=
h−1
X
k=h
z
k
(1.1)
wherez
k
is the exposure in days to thek
th
temperature bin. Essentially,GDD
htoh
mea-
sures the amount of time a crop was exposed to temperatures between a given lower and
upper bound.
1.3.3 Intermediary Markets
An empirical analysis of the impact of competition on mitigation of climate shocks requires
information on market power, which is a function of the number, size, and location of
intermediary markets. Our primary measure of competition is defined at a wholesale
market level, and is calculated as an inverse distance weighted sum of total trade across
all neighboring markets in the same state (see 1.4.2 for details). Given the spatial nature
of this statistic, it was important to determine the exact geospatial location of each market.
The steps employed to create this dataset are detailed below.
First, we needed a comprehensive list of all wholesale intermediary markets in the
country. For this purpose, we used the Directory of Wholesale Agricultural Produce Assem-
bling Markets in India published in 2004 by the Directorate of Marketing and Inspection
(DMI), Ministry of Agriculture, Government of India Chimalwar, Tabhane, Verma, H.
Singh, and Bhatia 2004.
18
The directory lists 5,983 markets in the country, and provides
information on the name of the market, name of and distance to the nearest railway sta-
tion, district and state of each market, and the commodities traded therewith. These 5,983
markets form our universe of wholesale intermediaries (mandis) in India.
18
We used the latest version published in 2004. There are also three older directories published in the
years 1963, 1992 and 2000.
24
However, not all of these markets observed active trade and/or reported the daily quan-
tities and prices of commodities arriving in the marketplace. Therefore, as a second step,
we remove from our initial sample the subset of markets for which there did not exist any
price or quantity data since 2001. The assumption here is that data does not exist because
these markets did not see any trade during this time period.
19
For this exercise, we use the
Agmarknet dataset provided by the Ministry of Agriculture and Farmers Welfare in India,
which collates data on daily arrivals and producer prices for all government-regulated
agricultural markets in India since 2001.
20
We match the DMI list of 5,983 markets with
the list of markets in the Agmarknet data, and include a market in our sample if there was
even a single day of trading at the market for any of the 19 major commodities (selected
from ICRISAT 2018) from 2001 onwards. Next, we remove all markets in the state of Bi-
har, which dismantled the APMC markets in 2006, and markets in Kerala, Jammu and
Kashmir, and Manipur, which never enacted the APMC Act. We also remove markets in
the north-eastern states, certain Union Territories, and islands, where agriculture is not
practiced on any substantial scale.
21
This gives us a final sample of 2,938 markets in 20
states.
The third step involved a significant undertaking of finding the exact geolocation of
these 2,938 markets. The problems with using a Google API to identify the coordinates of
a market in India are manifold. First, India’s linguistic diversity means APMC markets are
19
Mandis can be of three types: primary, secondary, and non-regulated Chimalwar, Tabhane, Verma,
H. Singh, and Bhatia 2004. The missing trade data pertains to the latter two. Our analysis is focused on
primary markets, which are large yards where the first trade between farmers and intermediaries takes
place. In essence, these yards are the first point of contact with the farmers. Secondary and non-regulated
markets are smaller with rarely, if any, farmers participating. They are mostly used for further trading of
the agricultural produce purchased by the intermediaries from the primary markets. Given that our chief
focus is on farmers, and we have data on quantities and prices for all primary markets, the missing data for
secondary markets is not a major concern.
20
The Agmarknet data can be accessed at https://agmarknet.gov.in/.
21
This includes Andaman and Nicobar Islands, Arunachal Pradesh, Chandigarh, Dadra and Nagar
Haveli, Daman and Diu, Lakshadweep, Meghalaya, Mizoram, Nagaland, Puducherry, Sikkim, and Tripura.
The states included in the final analysis are shown in 3.2.
25
denoted on Google Maps by local names in different states.
22
This implies that there does
not exist a uniform text string which could be used to search the latitude and longitude
coordinates of the markets. Second, though using the coordinates of the village centroid is
a potential proxy for the geolocation of markets, Indian towns and villages can be expan-
sive, and sometimes have multiple markets in the vicinity. Ignoring these distances and
markets could lead to an erroneous competition measure. Furthermore, various village
names are repeated, sometimes even within the same state, which could lead to inaccu-
racies in the collation of spatial location data. Given these complications, we, therefore,
conducted a search on Google Maps using unique keywords for each market. For each
market in a state, our keywords included the market name, postal address, and district
followed by the commonly used designation for wholesale markets in that state. We re-
placed the designation with different monikers of APMC markets if our search did not
turn up a valid result. In case of uncertainty, we further refined our search by calculating
the distance between the market identified by our search results and the nearest railway
station mentioned in the directory by DMI. We then compared our figure with the distance
to the same railway station given in the directory, and only if the difference was minuscule
(less than 10 percent) was the market selected.
As a final step, we corroborated our findings, wherever possible, with a dataset by the
Pradhan Mantri Gram Sadak Yojana (PMGSY) which provides information on approxi-
mately 770,000 geo-tagged rural facilities, 20 percent of which are agricultural.
23
We did
not use this as our primary source for geolocation of markets because the dataset is only
available for rural India, and does not cover facilities in urban centers. Moreover, in most
22
Examples of the most common names in each state include: Agricultural Market Committee, Agriculture
Market Yard, Rythubazar, or Farmer Grain Market in Andhra Pradesh; Regulated Market Committee or Noti-
fied Mandi in Assam and Orissa; Krishi Upaj Mandi or Galla Mandi in Chhattisgarh, Rajasthan and Madhya
Pradesh; Khetiwadi Utpadan Samiti Market in Gujarat; Anaaj Mandi or Grain Market in Haryana and Punjab;
RMC Yard in Karnataka; Krushi Utpanna Bazar Samiti in Maharashtra; Regulated Market or Weekly Shandi in
Tamilnadu; Galla Mandi Samiti or Naveen Mandi Sthal in Uttar Pradesh; Krishak Bajar, Anaj Hat Tala or Kisan
Mandi in West Bengal.
23
The dataset is provided by the Online Management, Monitoring and Accounting System (OMMAS)
arm of PMGSY and is available at http://omms.nic.in/. The agricultural facilities include cold storages,
collection centres, mandis, warehouse, etc.
26
states, it classifies smaller retail markets as also mandis, making it difficult to differentiate
wholesale markets from retail markets. However, it proved useful in validating — and
confirming in case of uncertainty — our Google Maps search results in rural areas.
Notwithstanding different searches involving various strings and the use of PMGSY
dataset, 13 percent (386 markets) of the markets could not be precisely geocoded. In such
cases, we used the centroid coordinates of the village or town. The geographic distribution
of all wholesale markets in the country is plotted in 1.3.
Figure 1.3: Geographic Distribution of APMC Markets
Note: The map shows the geographic distribution of 2,938 APMC markets by district and state. Each dot represents an APMC market.
Geographic coordinates were found through Google Maps, using data from the Directory of Wholesale Agricultural Produce Assembling
Markets in India published by the Directorate of Marketing and Inspection, Ministry of Agriculture, Government of India Chimalwar,
Tabhane, Verma, H. Singh, and Bhatia 2004, and further corroborated with a dataset on geo-tagged rural facilities provided by the
Pradhan Mantri Gram Sadak Yojana (PMGSY).
27
1.3.4 Quantity Arrivals and Prices
The Ministry of Agriculture and Farmers Welfare aggregates commodity level, daily quan-
tity arrivals and producer prices received by farmers across government-regulated agri-
cultural markets in India. Information is available starting 2001 for 344 agricultural and
livestock commodities from approximately 4,000 markets spread across more than 650
districts of India. Though this data is available on the government’s Agmarknet portal, we
downloaded the same from the portal maintained by the Centre for Economic Data and
Analysis (CEDA) of Ashoka University, as they have collated the data in a format that is
easily downloadable and also corrected for certain inconsistencies.
24
Our sample is comprised of 52 major commodities which mirror the 19 crops in the
ICRISAT 2018 dataset.
25
One potential concern with the latter dataset is that it does not
classify regional crop production based on season. However, growing season for the same
crop may differ across regions.
26
For instance, if we classify an agricultural commodity as
a Kharif crop in a region where it is, in fact, grown in the winter months, the weather
conditions ascribed to the crop yields will be erroneous, leading to spurious results. In
this regard, high frequency arrivals data helps us attribute the right growing season for
crops traded in a market as we can deduce the time of harvest based on its arrival date in
an APMC market. Therefore, to correctly classify the growing season for each crop in each
market, we use the following algorithm: for every crop-market pair, we first aggregate all
arrivals up to a monthly level, and compute the monthly average across all years (2001
onwards). This gives us the average quantity traded in a market for every month across
all years. Second, we use this monthly average to find the proportion of quantity traded
24
The data from CEDA can be accessed at https://agmarknet.ceda.ashoka.edu.in/
25
ICRISAT 2018 tends to aggregate several crops under a single head. For instance, it contains minor
pulses as a crop, but this classification includes numerous pulses for which we have disaggregated data at
the market level in the Agmarknet database. This is the source of discrepancy between the number of crops
in ICRISAT 2018 and Agmarknet (19 versus 52).
26
To give an example, rice growing season in India varies depending upon climatic conditions, soil types,
and water availability. Eastern and southern regions of the country have favorable temperature for rice
cultivation throughout the year, leading to two or three crops of rice every year. Northern and western
regions, on the other hand, grow only one crop of rice from May to November M. P. Singh 2009.
28
in each month. Finally, we determine the growing season based on the month with the
maximum proportion of arrivals. Accordingly, if the peak arrivals was between October
to February, we classify the crop as Kharif, if it was between March to June, we classify it
as Rabi, and Zaid (summer season) otherwise.
The market-wise growing season classification is then used to construct the price and
quantity traded variables at a market-crop-year level. For quantity in each year, we sum
the daily arrivals in a market across the agricultural season, while for prices, we use the
modal price of the crop in the market across the growing season.
1.4 Empirical Methods and Results
The empirical section is divided into three parts. We start with 1.4.1 which motivates our
question on the distortionary impact of market power on adaptation by examining if there
is any evidence of long-run yield-stabilizing adaptation to extreme heat in India. If farmers
were able to neutralize the negative impact of institutional challenges over the long-term,
then studying their distortionary impact in the short-run would just be a cursory exercise.
This is followed by 1.4.2, which estimates the effect of market competition in mitigating the
damaging impact of extreme heat. We approach this question using panel data methods
and then proceed to strengthen our identification strategy through a border regression
discontinuity design. Finally, in 1.4.3, we identify the potential mechanisms driving the
impact of market competition on adaptation.
1.4.1 Effect of Climate Shocks on Yields
This subsection estimates the share of the negative short-run impacts of extreme heat that
are offset in the longer run. We run two separate regressions. First, 1.4.1.1 uses a panel
approach, akin to Deschˆ enes and Greenstone 2007, to estimate the effect of random year-
to-year variation in district weather conditions on agricultural yields for the time period
29
from 1968 to 2017. Second, 1.4.1.2 uses a long-differences approach proposed by Burke
and Emerick 2016 to model long-run district-level changes in yields between two different
points in time as a function of long-run changes in temperature and precipitation. Finally,
in 1.4.1.3, we compare panel and long differences coefficients which offers a test of whether
the shorter run damages of climatic variation on agricultural outcomes are mitigated in
the longer run.
1.4.1.1 Panel Approach
The panel approach uses short-run variation in climate, which is plausibly random, within
a given area to estimate the effect of extreme heat on agricultural productivity. Our econo-
metric model takes the following form:
log(Yields)
cdsy
=α +
6
X
j=1
β
j
GDD
{j}dsy
+ θPrecip
dsy
+ δ (Precip
dsy
)
2
+ π
cd
+
γ
y
+ f
s
(y) + ξ
cdsy
(1.2)
wherelog(Yields)
cdsy
refers to the log of yields (in kg/ha) for cropc in districtd of state
s in agricultural year y (July-June). The key explanatory variable is GDD
{j}dsy
, which
captures the daily distribution of daily temperatures in district d of state s in year y. It
denotes the number of days in district d of state s in agricultural year y on which the
daily mean temperature fell in thej
th
of the six temperature bins (in
◦
C), namely< 15
◦
C,
> 35
◦
C, and four 5
◦
C wide bins in between. Precip
dsy
and (Precip
dsy
)
2
denote the linear
and quadratic polynomial function of total rainfall (in m of water equivalent per day)
for district d in state s and year y. For our main specification, we include crop-district,
π
cd
, and agricultural year, γ
y
, fixed effects, while f
s
(y) refers to state-specific linear and
quadratic time-trend. The fixed effects imply that identification comes only from weather
variation across years within a particular district for each crop after differencing out any
state-specific time trends and macro variations across all states in a year. ξ
cdsy
denotes
30
the error term. Note that the model above is run separately for Kharif and Rabi crops, so
agricultural year refers to the particular cropping season in that agricultural year.
27
We estimate separate coefficients β
j
for each of the temperature bin regressors. Since
the number of days in a particular standardized cropping season always sums to the same
amount, we have to use one bin as a reference category.
28
We use 20−25
◦
C as the reference
category for Kharif crops, and 15− 20
◦
C as the reference category for Rabi crops, with
the coefficients for the reference categories consequently normalized to zero.
29
We use
two-way clustered robust standard errors, with clustering at the crop-state and year level.
Results are presented in columns (1) and (2) of 1.1.
The results from the panel regression indicate that extreme heat has a significant nega-
tive impact on productivity, with each additional degree-day of heat above 35
◦
C reducing
yields by 1 percent and 1.8 percent for Kharif and Rabi crops, respectively. For Rabi crops,
an additional degree-day between 30−35
◦
C is also detrimental, with yields experiencing
a sharp decline by 1 percent in comparison to the yields during the optimal temperature of
15−20
◦
C degrees. The larger impact on Rabi crops is expected, as they are sown in winter
and harvested in early spring and, therefore, will be more sensitive to extreme heat.
Given that panel estimates capture within-year adjustments by farmers Guiteras 2009
— such as modification of inputs or cultivation techniques — the negative results indicate
that short-run adjustment are unable to mitigate the harmful effects of extreme heat.
30
27
Specifically, the weather variables for Kharif crops are defined as the sum of the growing degree days
or precipitation in the months of June, July, August, and September for a particular agricultural year, while
the weather variables for Rabi crops pertain to the months of October, November, December, January, and
February of the agricultural year.
28
For Kharif season, the number is 122, calculated as the total number of days between June-September.
For Rabi season, the number is 151, calculated as the total number of days between October-February.
29
Reference category was selected based on the optimal temperature during the ripening (grain filling)
stage for the key crop in the season — rice for Kharif, and wheat for Rabi. Grain filling is one of the most
sensitive temperature stages for rice and wheat, with a strong bearing on final yields P. Krishnan, Ramakr-
ishnan, K. R. Reddy, and V. Reddy 2011. The mean optimum temperature during this stage is between 21.2
to 24.2 for rice S´ anchez, Rasmussen, and Porter 2014, and 15-20 for wheat Jenner 1991; Wardlaw 1974
30
The ability of panel models to capture long-run climatic adaptation remains a subject of active research.
See McIntosh and Schlenker 2006 and M´ erel and Gammans 2021 for a discussion.
31
Panel Long Differences
Kharif Rabi Kharif Rabi
(1) (2) (3) (4)
Bin<15
dsy
0.005 0.001 0.016 −0.015
(0.005) (0.002) (0.039) (0.014)
Bin 15-20
dsy
0.007 0.002
(0.004) (0.018)
Bin 20-25
dsy
−0.004 −0.012
(0.002) (0.012)
Bin 25-30
dsy
−0.001 −0.003 0.005 −0.016
(0.002) (0.002) (0.005) (0.012)
Bin 30-35
dsy
−0.005
∗
−0.010
∗∗∗
−0.004 −0.029
∗∗∗
(0.002) (0.002) (0.008) (0.011)
Bin>35
dsy
−0.010
∗∗
−0.018
∗∗
−0.023
∗∗∗
−0.027
∗
(0.004) (0.004) (0.008) (0.015)
Fixed Effects
Crop× District ! !
Crop× State ! !
Year ! !
State Time-Trend ! !
Num. obs. 125,279 56,436 2,189 1,082
Adj. R
2
0.743 0.809 0.563 0.580
Table 1.1: Effect of Temperature on Yields: Panel and Long Difference Estimates
Notes: Clustered robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
Columns (1) and (2) provide estimates of the effect of climate shocks on yields using a panel approach, as specified in eq:panel. The
dependent variable is the natural logarithm of yields (in kg/ha) for cropc in districtd of states in agricultural yeary. Columns (3)
and (4) provide estimates of the effect of climate change on agricultural yields from a long-differences approach Burke and Emerick
2016, as specified in 1.3. The dependent variable is the change in logged value of yields for crop c in districtd of states between two
periods, wherein the two periods are 1970 and 2015, with endpoints calculated as five-year average. Data, sourced from ICRISAT 2018,
are for 313 Indian districts of 20 states at an annual level from the year 1966 to 2017. The independent variables,Bin
htoh
, measure
the amount of time, in days, a crop was exposed to temperatures between a given lower and upper bound. The coefficient of interest is
the estimate on Bin>35
dsy
, which represents extreme heat. Columns (1) and (3) provide estimates for Kharif crops (July–October),
while columns (2) and (4) provide estimates for Rabi crops (October–March). Standard errors for panel estimates are clustered at the
crop-state and year level, while standard errors for long-difference estimates are clustered at the district-level.
32
Next, we estimate the effect of climate shocks over the long-run which allows for the pos-
sibility of transformational adaptations, for e.g. crop switching or exit from farming.
1.4.1.2 Long Differences Approach
The Long Differences model uses the approach developed by Burke and Emerick 2016 to
identify the effect of climate change (as opposed to shocks) on agricultural productiv-
ity. Long differencing uses variation in longer run climate change and, therefore, helps
to account for long run adjustments to temperature. Specifically, we construct longer run
yield and weather averages at two different points in time for each location, and calculate
changes in average yields as a function of changes in average temperature and precipita-
tion. The model is as follows:
∆ log

Yields

cds
=α +
6
X
j=1
β
j

∆ GDD
{j}ds

+ θ

∆ Precip
ds

+ δ

∆ Precip
ds

2
+
π
cs
+ ξ
cds
(1.3)
where ∆ log

Yields

cds
is the change in logged value of yields for cropc in districtd of state
s between two periods. In our main specification, the two periods are 1970 and 2015, with
endpoints calculated as five-year averages for each variable to smooth out any idiosyn-
cratic noise. ∆ GDD
{j}ds
is the average difference in the number of growing degree days
in thej
th
temperature bin in districtd between the two periods, while ∆ Precip
ds
refers to
the change in average rainfall between the two periods in a given district. We also include
crop-state fixed effects, π
cs
, to account for any crop- and state-specific trends. The identi-
fying variation, therefore, comes from temperature changes within different districts in a
state after differencing out crop-specific trends. The key coefficient of interest is β
6
, which
measures how yields are affected by exposure to extreme heat, i.e. > 35
◦
C. Like before,
the analysis is run separately for Kharif and Rabi crops, and the coefficients for the refer-
ence categories are normalized to zero. Error terms are assumed to be correlated within
33
districts, and consequently, the standard errors are clustered at the district-level. Results
are presented in columns (3) and (4) of 1.1.
The long differences estimates are higher in magnitude than estimates from the panel
approach, and suggest that one unit increase in exposure to heat above 35
◦
C results in a
significant 2.3 and 2.7 percent decline in yields for Kharif and Rabi crops, respectively. As
before, exposure to degree days between 30−35
◦
C are also damaging for crops in the Rabi
season, with yields dropping by 2.9 percent relative to one additional day in the reference
bin of 15− 20
◦
C. Note that the magnitude of these effects is net of any transformational
adaptations made by farmers over the 45-year estimation period, for e.g. crop switching
or exit from farming.
1.4.1.3 Adaptation
We can compare panel and long differences coefficients, in the style of Burke and Emerick
2016, to estimate adaptation to extreme heat in the long-run. The logic is as follows: panel
models identify the short-run responses to weather, while long differences models iden-
tify the impact of long-run changes in climate, embodying any adaptation that farmers
have undertaken over the estimation period. Comparing the two estimates can, therefore,
allow us to test whether the shorter run detrimental effects of extreme heat on agricul-
tural outcomes are in fact mitigated over a longer horizon. We quantify the magnitude of
adaptation as 1−β
LD
j
/β
FE
j
, withj = 6, i.e. > 35
◦
C temperature bin, and it gives us the
percentage of the negative short-run impact of extreme heat on yields that is offset in the
long-run. β
LD
j
here refers to the estimate ofβ
{j=6}ds
in the long differences model in 1.3,
andβ
FE
j
refers to the estimate ofβ
{j=6}dsy
in the panel model in 1.2. A positive estimate
would signify adaptation, with farmers demonstrating better adjustability to rising tem-
peratures over the long-run, compared to shorter run heat shocks. Contrarily, a null or
34
negative result provides evidence of a failure to alleviate short-term yield losses from ex-
posure to extreme heat through adaptation in the long-run; worse still, this could indicate
mitigation measures available in the short-run prove untenable in the long term.
Given thatβ
FE
j
andβ
LD
j
are estimated using separate regressions, we need to quantify
the uncertainty in the adaptation estimate. We bootstrap our data 5,000 times, sampling
districts with replacement to preserve the within-cluster features of the error Cameron,
Gelbach, and D. L. Miller 2008. Therefore, if thed
th
cluster (district) is selected, then all
data (dependent and regressor variables) in that cluster appear in the resample.
31
This
procedure is run separately for Kharif and Rabi crops for two time periods: 1970-2015 and
1990-2015. We then use the distribution of bootstrapped adaptation estimates to test, for
each season and time period of interest, the null hypothesis of ”no adaptation” to extreme
heat—i.e., that 1−β
LD
j
/β
FE
j
= 0. Results are presented in 1.4.
Results suggest that long-run adaptation to extreme heat has been absent, and in fact,
the deleterious impact of weather shocks over the long-run is higher relative to the short-
run when adaptation avenues could be more limited. Median estimates (dark black lines)
from the bootstrap distribution are negative for all the cases. Long-run point estimates are
higher than short-run estimates by 48 to 132 percent for Kharif crops, and by 45 to 85 per-
cent higher for Rabi crops. However, even though the estimates are negative, the 2-sided
confidence intervals (red dashed lines) for all cases span zero. Thus, longer run adap-
tations appear to have mitigated none of the large negative short-run impacts of extreme
heat on productivity. More likely, short-term adaptation measures mitigate a portion of
the damaging effects, but the same measures prove to be unsustainable over the long-run.
In summary, our results on adaptation in the long-run indicate that the bottlenecks
farmers face in adopting short-run strategies have a direct and cumulative impact on their
ability to adapt in the long-term. It should be noted that various studies document that
31
That is, we take a draw of districts with replacement, estimate both long differences and panel model for
those districts, compute the extreme heat coefficients for the two models, calculate the adaptation measure,
and repeat 5,000 times for a given time period.
35
Figure 1.4: Percentage of Short-Run Impacts Offset by Adaptation
Note: 1.4 shows the percentage of the short-run impacts of extreme heat on agricultural productivity for Kharif and Rabi cropping
seasons that are mitigated in the longer run. Each box plot corresponds to a particular season and time period as labeled on the left,
and represents 5,000 bootstrap estimates of 1−β
LD
j
/β
FE
j
for that time period. The dark line in each distribution is the median, the
blue dot the mean, the grey box the interquartile range, and the whiskers represent the fifth to ninety-fifth percentile. The red dashed
lines in each box plot represents the 2-sided confidence intervals for the test that 1−β
LD
j
/β
FE
j
= 0.
Indian farmers correctly perceive climatic changes, which makes the lack of adaptation
we find in the long-run puzzling P. Datta, Behera, et al. 2022. A clearer understanding of
the effect of distortionary policies on year-to-year adaptation is, therefore, crucial to assess
their persistent impact and also shed light on heterogeneity in adaptation across regions
with high and low competition.
1.4.2 Effect of Competition on Mitigation of Climate Shocks
We start by constructing a local spatial competition measure in 1.4.2.1. This measure is
then used in 1.4.2.2 to estimate, using a panel regression, the effect of competition on adap-
tation. In our analysis thus far, adaptation has been measured by the magnitude of the
fall in district-wise yields mitigated by competition. We modify this definition in 1.4.2.3,
where we now use spatially disaggregated market arrivals data to measure adaptation.
36
Finally, in 1.4.2.4, to address endogeneity concerns, we implement a border discontinuity
design with market pairs to causally identify these effects.
1.4.2.1 Measuring Market Power
We construct a measure of local competition at the market level, Comp
1m
, by taking a
weighted sum of the total value of trade at all other markets near the origin market site,
provided they are all in the same state. The weights are the inverse of distances of the
neighboring markets (n) to the origin market (m), while the total value of trade (Y
n
)
refers to the sum of the value of agricultural produce traded in the neighboring marketn
between the years 2000 to 2021. For any marketm,
Comp
1m
=
X
n∈M\{m}

1
distance
mn

Y
n
× 1{state of m = state of n} (1.4)
whereM is the set of all markets in India. Competition in any marketm is driven by three
factors: the number of neighboring markets (n), a farmer’s ease of access to alternative
markets, which we incorporate throughdistance
mn
, and the size of the alternative mar-
kets, which we proxy using the value of trade over the last two decades (Y
n
). Therefore,
theComp
1m
measure will assign a greater weight to a proximal market. Furthermore, a
neighboring market with large trade volumes will lead to a higher competition measure,
as opposed to a market with limited trade.
We also create an analogous local competition measureComp
2m
, similar to Chatterjee
2019, by taking an inverse distance weighted sum of other markets near a particular market
site but in the same state. The only difference between the two measures is that we do
not include the value of trade (Y
n
) in the latter, and competition is only defined by the
proximity of markets. Finally, since our unit of analysis is crop-district-year, we aggregate
competition to a district level by averaging the competition measure for all markets in a
districtd of states. This also gives us an opportunity to define a third, more crude measure
37
of competition,Comp
3ds
, which equals the density of markets, i.e. the number of markets
per square km in a districtd of states. The geographic distribution of competition using
theComp
1m
measure, aggregated to a district level, is illustrated in 1.5.
Figure 1.5: Geographic Distribution of Competition Aggregated to District Level
Note: The map shows the geographic distribution of competition at a district level. Competition is measured for each of the 2,938
APMC markets (represented by black dots) as the weighted sum of the total value of trade at all other markets near the origin market
site, provided they are all in the same state (see 1.4). The weights are the inverse of distances of the neighboring markets (n) to the
origin market (m), while the total value of trade refers to the sum of the value of agricultural produce traded in the neighboring
marketn between the years 2000 to 2021. Since our unit of analysis is crop-district-year, we aggregate competition to a district level by
averaging the competition measure for all markets in a districtd of states.
1.4.2.2 Panel Approach
We run a panel model to estimate how market competition, measured at a district level,
mitigates the adverse effects of extreme heat on crop yields. Our main specification takes
the following form:
log(Yields)
cdsy
=α +
6
X
j=1
η
j
GDD
{j}dsy
+
6
X
j=1
Ω j
(GDD
{j}dsy
×Comp
ds
) +
ϕPrecip
ds
+δ (Precip
ds
)
2
+λ
cy
+λ
dct
+λ
sy
+ξ
cdsy
(1.5)
38
where Comp
ds
is the aggregate measure of competition at the district level, and equals
either the mean value of the market level competition measure, Comp
im
∀i∈{1,2}, for
all marketsm in districtd, or the number of markets per square km in districtd of states
(Comp
3ds
). Since our baseline competition measure (Comp
1m
) is calculated as an inverse
distance weighted sum of total value of trade across all years in neighboring markets, and
we do not have data on the date of construction of markets,Comp
ds
is time invariant across
the length of our sample.
To control for confounds, we include multiple fixed effects, including the following in
our most rigorous specification: a crop-year fixed effect, λ
cy
, that controls for changes in
national or world prices of the commodity; a district-crop-decade fixed effect, λ
dct
, that
controls for slow-moving changes in crop-specific costs, in the area allocated to the crop,
in preferences, or in technologies; and a state-year fixed effect, λ
sy
, that controls for state-
specific cost or demand shocks common to all crops. Certain specifications also include
f
s
(y), which is a state-specific linear and quadratic time-trend. Note that the inclusion
of any form of district fixed effects implies that the level effect of time-invariant district
specific competition ( Comp
ds
) is swept out and cannot, therefore, be estimated. Finally,
we compute robust standard errors clustered at the state-year and crop level to account for
cropping decisions and other shocks which are likely to be spatially and serially correlated.
Results are presented in 1.2.
Our results suggest that there is significant mitigation of the effect of extreme heat ow-
ing to increased competition. Depending on the specification, each additional degree-day
of heat above 35
◦
C reduces yields by 1.4 to 3.2 percent. Importantly, in areas with lower
intermediary market power, this effect is attenuated, with the coefficient on the interac-
tion term between extreme heat and competition significantly positive and ranging from
0.001 to 0.003. To interpret the scale of this number, we can compute the impact of a one
standard deviation increase in competition on the effect of heat shocks on yields. Divid-
ing this byη
j
gives us the percentage of impact mitigated. We find that a one standard
39
Dependent Variable: log(Yields)
cdsy
(1) (2) (3) (4) (5) (6)
Bin 30-35
dsy
−0.004
∗
−0.004
∗
−0.011
∗∗∗
−0.002 −0.013
∗∗∗
−0.010
∗∗∗
(0.002) (0.002) (0.003) (0.002) (0.003) (0.002)
Bin>35
dsy
−0.026
∗∗∗
−0.025
∗∗∗
−0.014
∗∗
−0.033
∗∗∗
−0.018
∗∗∗
−0.015
∗∗∗
(0.005) (0.006) (0.006) (0.007) (0.006) (0.004)
Bin<15
dsy
× Comp
ds
−0.002 −0.002 −0.000 −0.000 0.000 0.001
(0.001) (0.002) (0.001) (0.002) (0.001) (0.001)
Bin 15-20
dsy
× Comp
ds
0.003 0.003 0.002 0.002 0.000 −0.000
(0.002) (0.002) (0.001) (0.002) (0.001) (0.001)
Bin 25-30
dsy
× Comp
ds
−0.000 0.000 0.002 −0.000 0.001 0.001
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Bin 30-35
dsy
× Comp
ds
−0.001 −0.001 0.001 −0.001 0.001 0.001
(0.001) (0.002) (0.001) (0.002) (0.001) (0.001)
Bin>35
dsy
× Comp
ds
0.004
∗∗∗
0.004
∗∗∗
0.003
∗∗
0.005
∗∗
0.003
∗∗
0.002
∗
(0.001) (0.001) (0.002) (0.002) (0.001) (0.001)
Fixed Effects
Crop !
District ! !
Year ! !
Crop× District !
Crop× Year ! ! ! !
District× Year !
State× Year !
District× Crop× Decade ! !
Effect Mitigated (in %) 18.5 20.9 31.7 18.8 20.5 15.1
Num. obs. 59,593 59,593 59,593 59,593 59,593 59,593
Adj. R
2
0.624 0.615 0.805 0.635 0.829 0.844
Table 1.2: Competition and Mitigation of Climate Shocks: Panel Approach with Yields
Notes: Clustered robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
Columns (1) to (6) provide estimates of how market competition, measured at a district level, mitigates the adverse effects of extreme
heat on crop yields (eq:panelcompadaptation).Thedependentvariable,log(Yields)
cdsy
, refers to the log of yields (in kg/ha) for crop
c in districtd of states in agricultural yeary (July-June). Data, sourced from ICRISAT 2018, are for 313 Indian districts of 20 states at
an annual level from the year 1966 to 2017.Comp
ds
is the aggregate measure of competition at the district level. For this purpose, we
first calculate competition for each of the 2,938 APMC markets as the weighted sum of the total value of trade at all other markets near
the origin market site, provided they are all in the same state. The weights are the inverse of distances of the neighboring markets (n)
to the origin market (m), while the total value of trade refers to the sum of the value of agricultural produce traded in the neighboring
marketn between the years 2000 to 2021. Second, we aggregate competition to a district level by averaging the competition measure
for all markets in a districtd of states. The independent variables related to temperature,Bin
htoh
, measure the amount of time, in
days, a crop was exposed to temperatures between a given lower and upper bound. The coefficient of interest is the estimate on the
interaction term betweenBin> 35
dsy
(extreme heat) andComp
ds
. It can be interpreted as the supplementary impact of an additional
degree day of extreme heat for a given level of competition. The antepenultimate row, titled Effect Mitigated (in %), provides estimates
of the impact of extreme heat mitigated by a one standard deviation increase in district competition. Coefficients related to the effect
of temperatures less than 30
◦
C on yields have been omitted for brevity. Standard errors are clustered at the state-year and crop level.
40
deviation increase in market competition can help farmers mitigate the impact of extreme
heat by 13.2 percent in our most rigorous specification in column (6) with crop-year, state-
year and district-crop-decade fixed effects. The effect is substantially larger in column (3),
where we control for crop-district and year fixed effects and add in state time trends, with
one standard deviation increase in competition leading to an attenuation of 31.3 percent.
The rest of the specifications give us a number between these two extreme values.
1.4.2.3 Panel Approach: Arrivals Data
To this point in our paper, we have used the attenuation in district-wise crop-specific yields
as a measure of adaptation. However, potential concerns could arise regarding mismea-
surement of the district level competition variable as the same was constructed by av-
eraging the market competition across all mandis in the district. Particularly, if farmers
regularly cross district borders within the state boundaries to sell their produce, then the
average competition across district mandis may not be a true indicator of the monopsony
power faced by farmers. Therefore, we address this by using microdata on the daily quan-
tity arrivals of each crop at a market. Arrivals reflect the daily quantity traded of a crop in
a particular mandi, and the sum across the growing season acts as a proxy for the total
production of the crop during the agricultural year.
Our econometric specification closely follows 1.5, except that our unit of analysis is
now market-crop-year, and we replace yields at the district level with quantity arrivals at
each market. Specifically,
log(Arrivals)
cmdsy
=α +
6
X
j=1
η
j
GDD
{j}dsy
+
6
X
j=1
Ω j
(GDD
{j}dsy
×Comp
mds
) +
ϕPrecip
ds
+δ (Precip
ds
)
2
+λ
m
+λ
cy
+λ
dt
+λ
sy
+ξ
cmdsy
(1.6)
wherelog(Arrivals)
cmdsy
refers to the natural logarithm of the quantity of cropc arriving
in marketm situated in districtd of states in agricultural yeary. Comp
mds
is the market
41
level measure of competition, calculated as either the weighted sum of the total value
of trade at all other markets in the same state near the origin market site (Comp
1m
), or
the inverse distance weighted sum of other markets in the same state near a particular
market site (Comp
2m
). We include crop-year fixed effects ( λ
cy
) to account for changes in
national or world prices of commodities, and district-decade fixed effects ( λ
dt
) to factor
out slow moving district-specific technological changes. We also control for state-specific
cost or demand shocks common to all crops by including state-year fixed effects ( λ
sy
), and
individual market time-invariant idiosyncrasies by adding individual market fixed effects
(λ
m
). The inclusion of market fixed effects implies that the level effect of time-invariant
market specific competition ( Comp
mds
) is swept out and cannot, therefore, be estimated.
η
6
can now be interpreted as the effect of an additional degree-day of extreme heat in
the district on quantity arrivals, while the coefficient of interest, Ω 6
, indicates the magni-
tude of impact mitigated by competition. We cluster our standard errors two-way, both at
the state-decade level and the crop level. Results are presented in 1.3.
The results mirror the estimates from the previous subsection — in fact, the mitiga-
tion effects are larger. Depending on the specification, each additional degree-day of heat
above 35
◦
C reduces quantity arrivals by 2.3 to 3.0 percent. However, as before, this effect
is significantly allayed in high-competition areas. A one standard deviation increase in
market competition can help farmers mitigate the impact of extreme heat by 36.2 percent
in our most rigorous specification in column (4) with market, crop-year, state-year, and
district-decade fixed effects. In the remaining columns, the effect sizes range from 29.6
to 36.9 percent. This suggests that our results using district-level yields as a measure of
adaptation were biased downwards, and using arrivals data as a proxy helps correct this
bias.
42
Dependent Variable: log(Arrivals)
cmdsy
(1) (2) (3) (4)
Bin 30-35
dsy
0.001 0.001 0.002 0.001
(0.006) (0.006) (0.006) (0.006)
Bin>35
dsy
−0.023
∗
−0.023
∗
−0.030
∗
−0.023
∗
(0.013) (0.013) (0.016) (0.014)
Bin<15
dsy
× Comp
mds
−0.000 −0.000 −0.001 −0.000
(0.002) (0.002) (0.002) (0.002)
Bin 15-20
dsy
× Comp
mds
0.002 0.002 0.002 0.002
(0.001) (0.001) (0.002) (0.001)
Bin 25-30
dsy
× Comp
mds
0.000 −0.000 0.000 −0.000
(0.001) (0.001) (0.001) (0.001)
Bin 30-35
dsy
× Comp
mds
−0.000 −0.000 −0.001 0.000
(0.001) (0.001) (0.002) (0.001)
Bin>35
dsy
× Comp
mds
0.002
∗∗∗
0.002
∗∗∗
0.003
∗∗
0.002
∗∗∗
(0.001) (0.001) (0.001) (0.001)
Fixed Effects
Market ! !
Crop× Year ! ! ! !
District× Decade !
Market× Decade !
Market× Year !
State× Year ! ! ! !
Effect Mitigated (in %) 36.9 35.4 29.6 36.2
Num. obs. 148,814 148,814 148,814 148,814
Adj. R
2
0.433 0.450 0.450 0.437
Table 1.3: Competition and Mitigation of Climate Shocks: Panel Approach with Arrivals
Notes: Clustered robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
Columns (1) to (4) provide estimates of how market competition mitigates the adverse effects of extreme heat on quantity arrivals
at each market (eq:arrivalspanelcompadaptation).Thedependentvariable,log(Arrivals)
cmdsy
, refers to the natural logarithm of the
quantity of cropc arriving in marketm situated in districtd of states in agricultural yeary. Data, sourced from Centre for Economic
Data and Analysis (CEDA) of Ashoka University, comprises of quantity arrivals of 52 major commodities in 2,938 APMC markets from
2001 to 2021. Comp
mds
is the measure of competition at the market level, and equals the weighted sum of the total value of trade at
all other markets near the origin market site, provided they are all in the same state. The weights are the inverse of distances of the
neighboring markets (n) to the origin market (m), while the total value of trade refers to the sum of the value of agricultural produce
traded in the neighboring marketn between the years 2000 to 2021. The independent variables related to temperature, Bin
htoh
,
measure the amount of time, in days, a crop was exposed to temperatures between a given lower and upper bound. The coefficient of
interest is the estimate on the interaction term betweenBin> 35
dsy
(extreme heat) andComp
mds
. It can be interpreted as the effect
of an additional degree-day of extreme heat in the district on quantity arrivals. The antepenultimate row, titled Effect Mitigated (in %),
provides estimates of the impact of extreme heat mitigated by a one standard deviation increase in market competition. Coefficients
related to the effect of temperatures less than 30
◦
C on quantity arrivals have been omitted for brevity. Standard errors are clustered at
the state-decade and crop level.
43
1.4.2.4 Hybrid Border Discontinuity Design
Although our results are consistent across different empirical specifications, one can still
be concerned about other forms of unobserved heterogeneity. For example, if a large num-
ber of markets were set up in regions where farmers had a higher potential for innovation,
then the coefficient on the interaction between competition and weather could just be cap-
turing the effect of farmer ingenuity. To overcome this issue, we implement a hybrid bor-
der discontinuity design with market pairs. We match all markets which are less thanx
kilometers apart (bandwidth) but lie on different sides of a state boundary. We try dif-
ferent values of the bandwidth ranging from 25 kms to 50 kms, and all multiples of five
therein. For each bandwidth, we obtain a sample of market pairs, with markets belonging
to a pair lying in close proximity spatially but divided by a state border. The empirical
strategy involves regressing — for each market pair — the difference in arrivals on: (i) the
difference in competition; (ii) the average weather conditions across the two markets, and
(iii) the interaction between the two. We call it hybrid because even though there is a dis-
continuity in competition at the border, the regression involves weather variables which
are continuous.
The basic rationale behind employing the border discontinuity design is that other de-
terminants of arrivals like demand, weather, productivity via soil quality, farmer ingenu-
ity, and transportation costs will vary continuously across a state boundary. This should,
therefore, help to assuage concerns about unobserved heterogeneity. One could be con-
cerned that geographical conditions change discontinuously at the border. However, post
independence in 1947, Indian states were redrawn along linguistic principles, rather than
44
administrative, economic, or geographic factors Chari 2016; Samaddar 2020.
32
Neverthe-
less, an important determinant of farmer adaptation which could change discontinuously
at the state border is each state’s policy on weather shocks. To address this confound-
ing effect, we add market pair-year fixed effects. Thus, the only remaining discontinuity
across state borders which could potentially aid in attenuating the impact of extreme heat
is local competition, as farmers are not allowed to sell their output across state borders.
33
In essence, the advantage of this design is that we can difference out unobserved fac-
tors other than competition that affect adaptation by choosing market pairs in close geo-
graphical proximity to each other. To better illustrate the design, 1.6 presents a graphical
representation of our hybrid border discontinuity design. Along the same state border,
we have two market pairs (Pair 1 and Pair 2), with markets in each pair within 25 kms of
each other but lying on different sides of the border. For Pair 1, there is no difference in
competition, while for Pair 2, market C has a higher competition than market D. Now, in
the event of extreme heat (bad weather), the difference in arrivals should not change for
Pair 1, as both markets have the same competition and will be equally affected. However,
for Pair 2, the difference in arrivals should increase because farmers in Market C have been
able to attenuate some of the impact through higher competition. The spatial feature of the
design is illustrated in 1.7, which presents the geographical distribution of all 652 markets
selected using the bandwidth of 50 kms.
34
Note that only markets less than 50 kms apart
and situated in different states will be considered as a pair. This results in 1,210 market
pairs for the said bandwidth.
32
The Government of India appointed the States Reorganisation Commission in December 1953 which
advocated the following: To consider linguistic homogeneity as an important factor but not to consider it as an
exclusive and binding principle Parameswaran and Chattopadhyay 2014. In August 1956, the Indian Parliament
enacted the States Reorganisation Act, which remains India’s largest collective administrative reorganisation.
While due consideration was given to administrative and economic factors, it recognized for the most part
the linguistic principle and redrew state boundaries on that basis A. Kumar 2019.
33
As Chatterjee 2019 mentions, Indian languages change gradually over distance. Therefore, farmers and
intermediaries in close geographical proximity but settled on different sides of a state border should be able
to communicate with each other.
34
Our preferred bandwidth is 25 kms. We illustrate the geographical distribution of markets selected
with the 50 kms bandwidth as that leads to more market pairs, offering a vivid visualization.
45
Market A
(low comp)
Market B
(low comp)
Market C
(high comp)
Market D
(low comp)
Pair 1 Pair 2
State Border
Distance≤ 25km
∆ Comp = 0 ∆ Comp> 0
Good Weather : ∆ Arrivals =x
Good Weather : ∆ Arrivals =y
Bad Weather : ∆ Arrivals =x
Bad Weather : ∆ Arrivals>y
Figure 1.6: Interpreting the Border Discontinuity Design
Note: 1.6 presents a graphical representation of the hybrid border regression discontinuity design in 1.7; see text for details.
Figure 1.7: Geographical Distribution of Markets Selected Using 50 kms Bandwidth
Note: The map shows the geographic distribution of markets used in the hybrid border discontinuity design with a bandwidth of 50
kms. The dots represent the sample of market pairs which lie in close spatial proximity but are divided by a state border. There are
652 markets for the distance threshold of 50 kms. The empirical strategy involves regressing — for each market pair — the difference
in arrivals on: (i) the difference in competition; (ii) the average weather conditions across the two markets, and (iii) the interaction
between the two. Competition refers to the weighted sum of the total value of trade at all other markets near the origin market site,
provided they are all in the same state (see Equation (4)). The weights are the inverse of distances of the neighboring markets (n) to
the origin market (m), while the total value of trade refers to the sum of the value of agricultural produce traded in the neighboring
market n between the years 2000 to 2021.
46
The hybrid border discontinuity model linking difference in arrivals to the interaction
between differences in competition and weather variables is as follows:
∆ log(Arrivals)
cmm
′
y
=α +
6
X
j=1
η
j
GDD
{j}cmm
′
y
+
6
X
j=1
Ω j
(GDD
{j}cmm
′
y
× ∆ Comp
mm
′) +
ϕPrecip
mm
′
y
+δ (Precip)
2
mm
′
y
+λ
cy
+λ
mm
′
y
+λ
bct
+ξ
cmm
′
y
(1.7)
where ∆ log(Arrivals)
cmm
′
y
is the difference in the natural log of arrivals of crop c in agri-
cultural yeary between marketsm andm
′
which lie on different sides of the state boundary
b. ∆ Comp
mm
′ is the time-invariant difference in competition measure Comp
1m
between
marketsm andm
′
. GDD
{j}cmm
′
y
denotes the number of days in the cropping season, for
cropc in yeary, on which the daily mean temperature fell in thej
th
of the six temperature
bins (in
◦
C) at the district border betweenm andm
′
. To arrive at this variable, we average
the GDD’s in the districts containing the markets which, by design, lie on either sides of
the state border. The precipitation variables, too, are constructed in a similar manner.
We control for confounding factors by adding three fixed effects in our most stringent
specification: a crop-year fixed effect ( λ
cy
) that controls for changes in global or national
prices of the crop c; a market pair-year fixed effect ( λ
mm
′
y
) that controls for differences
in market specific infrastructure, policies and cost or demand shocks that are common to
all crops; and a state border-crop-decade fixed effect ( λ
bct
) that accounts for differences in
slow moving changes in crop-specific costs, in the area allocated to the crop, in preferences,
or in technology. As an aside, we can include market pair-year fixed effects as there are
multiple crops within that dimension. Importantly, these different crops within the same
market pair-year level are not subjected to the same weather. For example, kharif and rabi
arrivals in a market will be exposed to different weather conditions in the same agricultural
year.
35
Also note that the data for each market pair only includes crops which had the same
cropping season across both markets. Therefore, if rice in market A is a Kharif crop and
35
The identification of cropping season for each market-crop is made possible using the time series on
arrivals data, and is explained in detail in 1.3.4.
47
rice in Market B is a Rabi crop, we dropped rice as a commodity for market pair A-B. Hence,
the identifying variation comes from the differing weather conditions that different crops
within a market pair and year were exposed to, after differencing out any crop specific
fixed effect. Finally, the inclusion of any form of market pair fixed effects implies that the
level effect of time-invariant difference in market competition ( ∆ Comp
mm
′) is swept out
and cannot, therefore, be estimated.
The interpretation of the coefficients changes slightly as compared to previous speci-
fications. Though previously η
j
measured the effect of spending an additional day in the
j
th
temperature bin on arrivals, it now measures the effect on the difference in arrivals.
If our discontinuity in competition assumption is correct, then the only thing impacting
the difference in arrivals during extreme heat should be competition, which is captured
by the coefficient Ω j
. In other words, η
6
should not be significantly different from 0. If
that is not the case, it would indicate the presence of extraneous factors affecting arrivals
during extreme heat which, if correlated with competition, could bias our results. Thus,
the coefficient on η
6
acts as a placebo check. However, this implies that we cannot calcu-
late the percentage of impact mitigated by competition, as we do not obtain an estimate of
the marginal effect of climate shocks on arrivals. We cluster our standard errors two-way,
both at the border-year level and crop level. Results are presented in 1.4 for our preferred
bandwidth of 25 kms.
Our results with the border discontinuity design are in harmony with our previous
specifications and deliver the same message: competition helps in fostering adaptation
to climate shocks. All estimates of the interaction term between extreme heat and differ-
ence in competition are positive and significant, irrespective of the fixed effects used. A
farmer selling in the 75
th
percentile of competition compared to one that faces the 25
th
percentile of competition achieves a 4.5 to 5.2 percent higher yield on average for an ad-
ditional degree-day of extreme heat. Reassuringly, the estimates on the temperature bins
themselves are insignificant, as was expected if the regression was correctly specified. To
48
Dependent Variable: ∆ log(Arrivals)
c{mm
′
}by
Markets≤ 25km Apart
(1) (2) (3) (4)
Bin 30-35
cmm
′
y
−0.029 −0.016 0.016 0.011
(0.048) (0.016) (0.030) (0.031)
Bin>35
cmm
′
y
0.029 0.037 0.025 0.036
(0.034) (0.026) (0.031) (0.027)
Bin<15
cmm
′
y
× ∆ Comp
mm
′ 0.005 0.012 0.012 0.014
∗
(0.008) (0.008) (0.008) (0.008)
Bin 15-20
cmm
′
y
× ∆ Comp
mm
′ 0.012 0.010 0.010 0.007
(0.014) (0.009) (0.012) (0.012)
Bin 25-30
cmm
′
y
× ∆ Comp
mm
′ 0.011 0.012 0.013 0.012
(0.008) (0.008) (0.009) (0.009)
Bin 30-35
cmm
′
y
× ∆ Comp
mm
′ 0.003 0.007 0.006 0.006
(0.006) (0.006) (0.006) (0.007)
Bin>35
cmm
′
y
× ∆ Comp
mm
′ 0.015
∗∗
0.018
∗∗
0.018
∗∗
0.017
∗∗
(0.007) (0.006) (0.008) (0.008)
Fixed Effects
Market-Pair× Year ! ! ! !
Border× Crop ! !
Crop× Year ! !
Border× Crop× Decade !
75th — 25th Percentile (in %) 4.5 5.2 5.2 5.1
Num. obs. 2,899 2,899 2,899 2,899
Adj. R
2
0.454 0.545 0.536 0.534
Table 1.4: Competition and Mitigation of Climate Shocks: Border Discontinuity
Notes: Clustered robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
Columns (1) to (4) provide estimates, using a 25 kms hybrid border discontinuity approach, of how
market competition mitigates the adverse effects of extreme heat on quantity arrivals at each market
(eq:arrivals
b
order
d
isccompadaptation).Thedependentvariable,∆ log(Arrivals)
c{mm
′
}by
, refers to the difference in the natu-
ral log of arrivals of crop c in agricultural year y between markets m and m’ which lie on different sides of the state boundary b.
Data, sourced from Centre for Economic Data and Analysis (CEDA) of Ashoka University, comprises of quantity arrivals of 52 major
commodities in 2,938 APMC markets from 2001 to 2021. ∆ Compmm’ is the time-invariant difference in competition measure between
marketsm andm’. GDD
{j}cmm
′
y
denotes the number of days in the cropping season, for cropc in yeary, on which the daily mean
temperature fell in thej
th
of the six temperature bins (in
◦
C) at the district border betweenm andm
′
. The coefficient of interest is
the estimate on the interaction term betweenBin>35
cmm
′
y
(extreme heat) and ∆ Comp
mm
′. The antepenultimate row, titled 75th
— 25th Percentile (in %), provides estimates of the higher yield experienced by a farmer selling in the 75
th
percentile of competition
compared to one that faces the 25
th
percentile of competition for an additional degree-day of extreme heat. Coefficients related to the
effect of temperatures less than 30
◦
C on yields have been omitted for brevity. Standard errors are clustered two-way at the border-year
and crop level.
49
test the robustness of the adaptation results, we also present the coefficient on Ω 6
from
regressions using different bandwidths in 1.8. Like before, the effect sizes are positive
and significant, although smaller in magnitude. One additional degree day in the highest
temperature bin leads to a difference in the range of 1.9 to 3.6 percent in yields between
farmers in the 75
th
percentile of competition relative to farmers in the 25
th
percentile. Note
that the confidence intervals for the 25 kms bandwidth are larger, which is expected given
the low number of market pairs due to the shorter distance.
Figure 1.8: Impact of Extreme Heat Offset by Competition: Border Discontinuity
Note: 1.8 provides estimates, using a hybrid border-discontinuity design with different distance bandwidths, of how market
competition mitigates the adverse effects of extreme heat on quantity arrivals at each market. The distance thresholds used for each
estimate are labeled on the X-axis. The point estimates (red dots) on the Y-axis correspond to the coefficient of interest — the
estimate on the interaction term betweenBin>35
cmm
′
y
(extreme heat) and ∆ Comp
mm
′ — in 1.7. ∆ Comp
mm
′ is the
time-invariant difference in competition measure between markets m andm’.GDD
{j}cmm
′
y
denotes the number of days in the
cropping season, for cropc in yeary, on which the daily mean temperature fell in thej
th
of the six temperature bins (in
◦
C) at the
district border betweenm andm
′
. To interpret the point estimates, we calculate the difference in yields experienced by a farmer
selling in the 75
th
percentile of competition compared to one that faces the 25
th
percentile of competition, for each additional
degree-day of extreme heat. This is indicated by the labels next to the red dots. The whiskers represent the 90
th
percentile confidence
intervals, with standard errors clustered two-way at the border-year and crop level.
50
1.4.3 Mechanisms
Given the result that market competition leads to higher adaptation, we now turn our
attention towards ascertaining the mechanisms behind our findings. 1.4.3.1 presents an
analytical framework that uses a simple agricultural model to derive predictions on input
usage post a climate shock. We then test empirically whether these predictions hold in the
data, the results of which are presented in 1.4.3.2 and 1.4.3.3.
1.4.3.1 Analytical Framework
In this subsection, we present a simple agricultural household model to examine how
subsistence farmers would adjust their input decisions in the event of an exogenous heat
shock. Closely following the work of J. E. Taylor and Adelman 2003 and Arag´ on, Oteiza,
and Rud 2021, we present a framework where production and consumption decisions are
linked. This transpires because the farmer is both a producer, choosing the allocation of
inputs to crop-production, and a consumer, choosing the allocation of income to consump-
tion.
We start with an agricultural production function with two inputs, land (T) and labor
(L). The household has an endowment of landT
e
, which can be used for production or
non-productive activities like leisure.
36
Household utilityU(c,t) is a function of consump-
tion of a market good (c) and land used for leisure (t). Households are price takers and
obtain income by renting their land, and selling their produce in the market at pricep. The
production function is defined by F(A,L,T ), whereA is farmer’s total factor productivity.
Specifically, we use A to capture the productivity shock due to exposure to extreme heat.
Consistent with our results on the relationship between crop yields and temperature, we
36
We follow Arag´ on, Oteiza, and Rud 2021 in this regard. The inclusion of land directly in the utility
function is a modeling device to create a positive shadow price (i.e., an opportunity cost of using land) and
should not be taken literally. Since land cannot be sold or rented out, without this device, the model would
predict that farmers will always use all available land. This prediction is inconsistent with the empirical
observation that as a proportion of cultivable area, 13.4 percent of the land was left fallow in 2010–2011, an
increase from 10.6 per cent in 1970–1971 Ranganathan and Pandey 2018.
51
assume that extreme heat has a detrimental effect on productivity. Each growing season,
the household maximizes utility by choosing simultaneously the amount of land allocated
to productive and nonproductive uses, and the labor to be employed. Finally, we assume
that both the utility and the production functions are increasing and strictly concave.
Under the extreme assumption that all input markets exist and are well functioning,
the household’s production and consumption decisions can be decoupled Benjamin 1992.
This separation result is driven by the possibility of trade in complete markets. In this
scenario, the farmer’s optimal input usage can simply be inferred by solving the profit
maximization problem, max
{T,L}
π =pF (A,T,L)−rK−wL, where r and w refer to input
prices. Under such a setting, a negative productivity shock, such as extreme heat, would
always reduce input usage.
37
The prediction above changes in the case of incomplete markets, which is a more re-
alistic setting in the context of India Rosenzweig and Wolpin 1993. To illustrate this, we
consider a mixed market scenario. Specifically, we assume that there is no input market
for land, but there is a well-functioning input market for labor. In this simplified setting,
the farmer’s problem becomes:
max
{T,L}
U(c,t)
subject to
c≤pF (A,T,L)−wL
T +t≤T
e
(1.8)
37
Assuming a Cobb-Douglas technologyf =AT
α
L
β
, the optimalT equals

pA

α
r

1−β

β
w

β

1/γ
, while
the optimalL equals

pA

α
r

α

β
w

1−α

1/γ
, whereγ = (1−α−β). Differentiating these two terms with
respect toA, we get
dT
dA
=

pA
α+β

α
r

1−β

β
w

β

1/γ
/γ, and
dL
dA
=

pA
α+β

α
r

α

β
w

1−α

1/γ
/γ. Both of
these are positive as long asα +β< 1.
52
The two first order condition are U
t
=pU
c
F
t
andpF
L
=w. Taking total derivatives of
the first order conditions with respect to A, followed by some algebra, we obtain:
dT
dA
=
(F
LA
F
TL
U
c
/F
LL
) +F
A
U
tc
−pF
A
F
T
U
cc
−F
TA
U
c
pF
2
t
U
cc
− 2F
T
U
ct
+F
TT
U
c
+ (U
tt
/p)− (F
2
TL
U
c
/F
LL
)
(1.9)
Assuming a Cobb-Douglas technologyF (A,T,L) =AT
α
L
β
, we can show that the nec-
essary and sufficient condition for dT/dA< 0, i.e. land usage increases with a negative
productivity shock, is:
p>
1
αF

T
U
tc
U
cc
−
U
c
U
cc

(1.10)
Intuitively, the inequality suggests that, in the presence of incomplete markets, farm-
ers could increase their input usage post a negative weather shock if the output price is
expected to rise. The increase in output prices could occur because of two reasons: first, a
negative effect on aggregate supply coupled with inelastic demand could increase prices;
and second, local competition in the markets could interact with a fall in supply to drive
the prices even higher. The former effect will be common to all areas, but the latter would
be restricted to high competition areas.
Another alternative, but not mutually exclusive, mechanism that could cause this phe-
nomenon is high risk aversion amongst farmers. This can be seen in 1.10 where an in-
crease in the coefficient of absolute risk aversion, −U
cc
/U
c
, increases the probability of
satisfying the inequality. In this context, high risk aversion would imply that farmers are
more likely to use supplementary inputs to attenuate the fall in agricultural output and
minimize the drop in consumption. This response is analogous to coping mechanisms
to smooth consumption, such as selling disposable assets. The key distinction is that it
involves adjustments in productive decisions.
53
The model predicts that an increase in land usage post a negative productivity shock
also increases the likelihood of an increase in the use of labor inputs. To see this, note that
the necessary and sufficient condition for labor inputs to increase post a weather shock is:
dT
dA
<−
T
αA
(1.11)
Therefore, if the increase in land usage following a negative productivity shock is large
enough, labor inputs on the farm will also rise.
With this framework in mind, our empirical analysis focuses on examining the effect
of competition on prices, and how the same varies across different weather conditions.
We subsequently test whether input usage increases in areas which experience price rise
during heat stress, as predicted by the model.
1.4.3.2 Effect of Competition on Prices: Heterogeneous Impact by Weather
This subsection aims to causally identify the effect of competition on prices during in-
clement weather. Besides its intrinsic interest, the heterogeneous effect of weather on the
correlation between competition and prices could also help inform the mechanism behind
the adaptive behavior documented in previous sections. The econometric specification
takes the following form:
log(Prices)
cmdsy
=α +
6
X
j=1
η
j
GDD
{j}dsy
+
6
X
j=1
Ω j
(GDD
{j}dsy
×Comp
mds
) +
ϕPrecip
dsy
+δ (Precip
dsy
)
2
+λ
m
+λ
cy
+λ
dt
+λ
sy
+ξ
cmdsy
(1.12)
wherelog(Prices)
cmdsy
refers to the natural logarithm of the price of cropc in marketm
situated in district d of state s in agricultural year y. This is the price during the main
agricultural season pertaining to the crop-market pair, and is calculated as the mean of
the daily modal price. The regressors and fixed effects have the exact same definition as
54
in 1.6. η
6
can now be interpreted as the effect of an additional degree-day of extreme heat
in the district on prices in markets with low competition, while the coefficient of inter-
est, Ω 6
, indicates the supplementary impact of high competition on prices during heat
stress. As before, the inclusion of market fixed effects ( λ
m
) implies that the level effect of
time-invariant market specific competition ( Comp
mds
) is swept out and cannot, therefore,
be estimated. However, Chatterjee 2019 shows, in a similar setting, that increasing spa-
tial competition by one standard deviation causes prices received by farmers to increase
between 2.7 and 6.4 percent. Though not shown here, we also calculate the difference
in prices between high and low competition areas and our results are similar, with a one
standard deviation increase in competition leading to a 5.8 to 6.6 percent increase in prices.
Thus, there is a positive effect of competition on farmer prices on average, holding con-
stant the weather. Our aim is to establish how this relationship changes during inclement
weather. Results on heterogeneous impact by weather are presented in 1.5.
Our results indicate that the positive effect of local competition on prices is exacerbated
during periods of extreme heat. The estimate of Ω 6
is positive and significant: markets
with higher level of competition experience a larger gain in price with each additional
degree day of extreme heat. If the effect was being driven solely by a fall in supply, we
would expect prices in low competition markets during heat stress to also increase relative
to prices in the same markets during good weather. Nevertheless, as indicated by the
coefficient on Bin> 35
dsy
, the effect sizes are positive but not significant, irrespective of
the specification. Notably, the effect sizes are large, but the lack of significance most likely
stems from low power caused by very few markets with competition tending to zero.
To interpret the coefficients, we calculate the difference in prices between high and
low competition areas when exposed to the mean number of extreme heat days during
the growing season (7.3 days). Note that this is in addition to the positive difference in
prices that exists during good weather. We find that a one standard deviation increase in
competition causes the difference in prices to increase by 0.53 to 0.57 percentage points,
55
Dependent Variable: log(Price)
cmdsy
(1) (2) (3) (4)
Bin 30-35
dsy
0.297 0.248 0.295 0.251
(0.827) (0.828) (0.923) (0.835)
Bin>35
dsy
1.180 1.323 1.563 1.297
(0.758) (0.796) (1.071) (0.774)
Bin<15
dsy
× Comp
mds
−0.131 −0.134 −0.150 −0.124
(0.205) (0.209) (0.214) (0.206)
Bin 15-20
dsy
× Comp
mds
0.188 0.186 0.176 0.190
(0.117) (0.114) (0.122) (0.118)
Bin 25-30
dsy
× Comp
mds
0.124 0.125 0.120 0.128
(0.079) (0.079) (0.085) (0.082)
Bin 30-35
dsy
× Comp
mds
0.026 0.022 −0.009 0.020
(0.120) (0.121) (0.125) (0.121)
Bin>35
dsy
× Comp
mds
0.227
∗∗
0.220
∗∗
0.230
∗
0.236
∗∗
(0.111) (0.106) (0.115) (0.107)
Fixed Effects
Market ! ! !
Crop× Year ! ! ! !
District× Decade !
District× Year !
Market× Decade !
State× Year ! ! !
Increase in Prices (in pp) 0.555 0.538 0.563 0.577
Num. obs. 147,005 147,005 147,005 147,005
Adj. R
2
0.877 0.879 0.879 0.878
Table 1.5: Effect of Competition on Prices Post Climate Shocks
Notes: Clustered robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
Columns (1) to (4) provide causal estimates of the effect of competition on prices following a period of extreme heat
(eq:pricespanelcompweather).Thedependentvariable,log(Price)
cmdsy
, refers to the natural logarithm of the price of cropc in mar-
ketm situated in districtd of states in agricultural yeary. Data, sourced from Centre for Economic Data and Analysis (CEDA) of
Ashoka University, comprises of prices of 52 major commodities in 2,938 APMC markets from 2001 to 2021.Comp
mds
is the measure
of competition at the market level, and equals the weighted sum of the total value of trade at all other markets near the origin market
site, provided they are all in the same state. The weights are the inverse of distances of the neighboring markets (n) to the origin
market (m), while the total value of trade refers to the sum of the value of agricultural produce traded in the neighboring marketn
between the years 2000 to 2021. The independent variables related to temperature,Bin
htoh
, measure the amount of time, in days, a
crop was exposed to temperatures between a given lower and upper bound. The coefficient of interest is the estimate on the interaction
term betweenBin> 35
dsy
(extreme heat) andComp
mds
. It can be interpreted as the supplementary impact of high competition on
prices during heat stress. The antepenultimate row, titled Increase in Prices (in pp), provides estimates of the effect of a one standard
deviation increase in competition on the difference in prices, given that both areas were exposed to a week of extreme heat. Coefficients
related to the effect of temperatures less than 30
◦
C on prices have been omitted for brevity. All coefficients have been multiplied by
1000 for illustrative purposes. Standard errors are clustered at the state-decade and crop level.
56
given that both areas were exposed to a week of extreme heat. Therefore, monopsony
power tends to aggravate the already existing price distortions. A simple back of the en-
velope calculation suggests that this translates to an extra yearly income in the range of 172
($3) and 31,642 ($608), depending on the crop being grown by the average farmer.
38
This
is equivalent to an increase of 0.4-69 percent in yearly net receipts from crop production
for an average agricultural household in India.
We have shown that during extreme heat, the prices received by farmers in high com-
petition areas increase, while prices in low competition areas do not. The analytical frame-
work presented in 1.4.3.1 predicts that this increase in prices should lead to higher input
usage in high competition areas, which in turn could help alleviate the crop production
losses associated with heat stress. The next section tests this hypothesis.
1.4.3.3 Changes in Input Use
We examine changes in input use as a potential margin of adjustment to high temperatures,
and how this differs between low and high competition areas. We combine household
survey with spatial weather and competition data to construct a comprehensive dataset
containing agricultural, socioeconomic, competition, and weather variables. The unit of
observation is the household-year. The household data is a repeated cross section from
the India Human Development Survey (IHDS), a nationally representative multi-topic
survey conducted in 2005 and 2011-12 Desai, Vanneman, and National Council of Ap-
plied Economic Research 2005, 2012. Our primary focus is on the income, social capital
and agricultural part of the survey, which asks questions on input usage and expendi-
ture in the last one year. Using the date of interview, we can construct household specific
weather variables, i.e. the number of growing degree days in each temperature bin and
precipitation over the last 12 months is specific to each household.
39
38
The to $ conversion was based on the historical average USD-INR exchange rate of 52.004 from 1
st
January, 2000 to 31
st
December, 2020, as published by Investing.com.
39
Households in the same district and interviewed in the same month-year will have identical values for
the weather variables.
57
The generic estimating equation is as follows:
Y
hdsy
=α +
6
X
j=1
η
j
GDD
{j}hdsy
+
6
X
j=1
Ω j
(GDD
{j}hdsy
×Comp
ds
) +
ϕPrecip
dsy
+δ (Precip
dsy
)
2
+ψZ
hdsy
+λ
d
+λ
sy
+ξ
hdsy
(1.13)
whereY
hdsy
refers to either input usage or input costs of householdh situated in district
d of state s in agricultural year y, GDD
{j}hdsy
refers to the number of growing degree
days in thej
th
temperature bin which the household was exposed to over the course of
the 12 months prior to the interview,Precip
dsy
is the analogous rainfall counterpart, and
Comp
ds
denotes the mean value of the market level competition measure across all mandis
in the district. Z
hdsy
is a vector of household characteristics, and includes religion, caste,
main income source, total land endowment, and permanent fallow land of the household,
in addition to the occupation and education of the household head. Finally, we control
for district and state-year fixed effects to account for, first, district specific determinants of
TFP as well as other drivers of input, and second, changes in agricultural prices at the state
level. Standard errors are clustered at the state-year level to allow for spatial dependence.
If the model prediction is accurate, then we expect the interaction term between high tem-
perature and competition (Ω 6
) to be positive and significant, indicating increased input
usage in high competition areas during heat stress. Results are presented in 1.6.
As predicted by the model, input usage and expenditure increases in high competition
areas during periods of high temperature. Columns (1) and (2) focus on changes in land
and labor inputs. For each additional degree day of extreme heat, a one standard deviation
increase in competition increases the land cultivated and labor employed by 1.2 and 1.7
percent, respectively. This estimate already controls for household endowments and per-
manent fallow land, and thus is not simply picking up changes in the size composition of
farmers. Columns (3) to (6) relate to input costs, specifically expenditure on labor, irriga-
tion, equipment, and fertilizers over the past 12 months. For each of these categories, the
58
effect sizes on Ω 6
are positive and significant, indicating farmers in high competition areas
expend more when faced with inclement weather. A one standard deviation increase in
competition would cause a farmer, experiencing an additional day in the extreme temper-
ature bin, to increase their labor expenditure by 122 ($2.3), irrigation expenditure by 31
($0.6), expenditure on farm equipment by 98 ($1.9), and expenditure on fertilizers by 157
($3.0).
In addition to adjustments in input usage and costs, we also find evidence of crop
diversification at a macro-scale ( i.e., district-level) in high competition areas, indicating
crop-mix as a potential avenue for increased resilience. To measure crop diversity, we
follow Auffhammer and Carleton 2018 and construct an indicator of concentration, the
Herfindahl–Hirschman Index (HHI), based on area planted to different crops in a given
year and district. The HHI for districtd and yeary is defined as follows:
HHI
dy
=
C
X
c=1
s
2
dcy
(1.14)
wheres
dcy
=a
dcy
/
C
P
c=1
a
dcy
is the share of total planted area in districtd dedicated to crop
c in yeary. C is the total number of crops, which in our data set comprises the 19 major
and minor crops available in ICRISAT 2018. The regression specification in 1.13 changes
slightly, with the unit of observation now districtd and agricultural yeary. Furthermore,
the regressandY
dsy
now denotes the crop-mix, while we dispense with the household con-
trol variables. Results presented in column (7) imply that each additional day of extreme
heat reduces the HHI significantly in areas with higher competition, indicating higher
crop diversity. The point estimates suggest that for each additional day of extreme heat, a
one standard deviation increase in competition leads to a 0.13 percent fall in HHI. Coupled
with the evidence that farmers adjust their land during the growing season, we interpret
these findings as suggestive evidence that the additional land is planted with distinct crops
in order to diversify the weather risk.
59
Our main results suggest that farmers adjust input use within the growing season as
a mechanism to cope with the negative effects of extreme temperatures, but only in high
competition areas. Farmers in these areas adjust their use of land, both in terms of area
planted and crop composition, as a response to extreme heat. Furthermore, they increase
labor usage, reflected both in the number of workers hired and total wages paid. Addi-
tionally, the expenditure on irrigation, equipment hired to work on the farm, and fertilizer
and manure also rises. These margins of adjustment attenuate undesirable drops in out-
put and consumption caused by heat. Importantly, the mechanism for these productive
adjustments are prices, which rise in high competition areas during heat stress, further in-
flating the pre-existing monopsony distortions. In this sense, our findings are consistent
with models of subsistence farmers in a context of incomplete markets J. E. Taylor and
Adelman 2003; De Janvry, Fafchamps, and Sadoulet 1991, which predict a rise in input
usage if prices increase following a negative productivity shock.
1.5 Theory
1.5.1 Basic Environment
Our setup closely follows the environment assumed by ?. We consider a national econ-
omy comprising multiple states, indexed byi∈I≡{1,...,I}. Within each state there are
two factors of production, labor and land, which can be used to produce multiple crops,
indexed byk∈K≡{1, ..., K}, and an outside good. The outside good can be thought
of as a composite of manufactured goods and services. Labor is homogeneous, perfectly
mobile within a state, and immobile across states. The termN
i
denotes the total endow-
ment of labor, and w
it
denotes the wage in state i at time t. Land comes in the form of
heterogeneous fields, indexed by f∈F
i
≡{1, ..., F
i
}, each comprising a continuum of
heterogeneous parcels, indexed byω∈ [0,1]. We lets
f
i
denote the area in hectares of field
f in statei.
60
Inputs Input Costs () Crop Mix
log(Land)
hdsy
log(Labor)
hdsy
Labor
hdsy
Irrigation
hdsy
Equipment
hdsy
Fertilizers
hdsy
HHI
dsy
(1) (2) (3) (4) (5) (6) (7)
Bin 30-35
dsy
−0.007 −0.026 −186.367 −10.872 67.362 −110.610
∗
0.511
∗
(0.012) (0.021) (134.622) (19.305) (124.858) (63.290) (0.294)
Bin>35
dsy
−0.020 −0.014 −137.790 −16.740 −60.710 −176.421
∗
0.480
(0.012) (0.015) (101.053) (16.826) (89.583) (92.065) (0.481)
Bin<15
dsy
× Comp
ds
0.012 0.021 100.256
∗∗
9.924 −46.148 −76.222 0.000
(0.009) (0.013) (44.907) (29.171) (89.536) (127.411) (0.000)
Bin 15-20
dsy
× Comp
ds
0.014
∗∗
0.018 −16.132 21.854 8.843 46.186 −0.197
(0.005) (0.012) (48.905) (26.599) (90.759) (105.920) (0.155)
Bin 25-30
dsy
× Comp
ds
0.005 0.026
∗∗
109.807
∗∗
44.160 56.101 −13.769 −0.194
(0.006) (0.011) (55.445) (29.828) (67.600) (100.552) (0.133)
Bin 30-35
dsy
× Comp
ds
0.008 0.025
∗∗
125.444
∗
23.881
∗
−28.514 −98.107 −0.039
(0.006) (0.011) (66.089) (13.473) (79.673) (56.889) (0.087)
Bin>35
dsy
× Comp
ds
0.009
∗∗
0.014
∗
93.363
∗
24.120
∗
73.561
∗
122.220
∗∗
−0.267
∗
(0.004) (0.008) (51.790) (13.887) (41.445) (46.165) (0.142)
Fixed Effects
District ! ! ! ! ! ! !
State× Year ! ! ! ! ! ! !
Num. obs. 25,592 20,517 24,652 27,654 28,256 21,179 4,624
Adj. R
2
0.580 0.352 0.243 0.187 0.187 0.380 0.944
Table 1.6: Heterogeneous Impact of Climate Shocks on Input Usage and Crop Mix
Notes: Clustered robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All columns represent estimates from different versions of the estimating 1.13, which examines changes in input use or crop-mix as a
potential margin of adjustment to high temperatures, and how this differs between low and high competition areas. The dependent
variable in Columns (1) and (2) represents land and labor inputs used by householdh situated in districtd of states in agricultural year
y, respectively. The dependent variable in columns (3) to (6) relates to input costs, specifically expenditure (in ) on labor, irrigation,
equipment, and fertilizers over the past 12 months. The dependent variable in column (7) represents an indicator of crop concentration,
the Herfindahl–Hirschman Index (HHI), based on area planted to different crops in a given year and district (1.14). The independent
variable related to competition intensity,Comp
ds
, is the measure of competition at the district level. For this purpose, we first calculate
competition for each of the 2,938 APMC markets as the weighted sum of the total value of trade at all other markets near the origin
market site, provided they are all in the same state. The weights are the inverse of distances of the neighboring markets (n) to the origin
market (m), while the total value of trade refers to the sum of the value of agricultural produce traded in the neighboring marketn
between the years 2000 to 2021. Second, we aggregate competition to a district level by averaging the competition measure for all
markets in a districtd of states. The independent variables related to temperature,Bin
htoh
, measure the amount of time, in days, a
crop was exposed to temperatures between a given lower and upper bound. The coefficient of interest is the estimate on the interaction
term betweenBin> 35
dsy
(extreme heat) andComp
ds
. It can be interpreted as the supplementary impact of high competition on
household input usage (for columns (1)-(6)), or crop mix (for column (7)), during heat stress. The household data is a repeated
cross section from the India Human Development Survey (IHDS), a nationally representative multi-topic survey conducted in 2005
and 2011-12 Desai, Vanneman, and National Council of Applied Economic Research 2005, 2012. Data for crop mix is sourced from
ICRISAT 2018, which has data on area under cultivation for 19 crops in 313 Indian districts of 20 states at an annual level from the year
1966 to 2017. Coefficients related to the effect of temperatures less than 30
◦
C on the dependent variable have been omitted for brevity.
Standard errors are clustered at the state-year level to allow for spatial dependence.
61
Preferences—Each statei at timet has a representative agent who derives utility from
consuming the outside good,C
0
it
, and a composite of all crops,C
it
:
U
it
=C
0
it
+β
i
ln(C
it
) (1.15)
The quasi-linear form of the utility function in 3.1 implies that there are no income effects.
Moreover, the total demand for crops depends only on a state-specific and time-invariant
demand shifter,β
i
≥ 0. Assuming that the crops in our analysis account for a small fraction
of consumers’ expenditure across states, the absence of income effects acts as a minor
limitation of our study.
Aggregate crop consumption at timet,C
it
, depends on the consumption of each crop,
C
k
it
, which itself depends on the consumption of varieties from different origins, C
k
jit
:
C
it
=


X
k∈K
(β
k
i
)
1/φ
(C
k
it
)
(φ−1)/φ


φ/(φ−1)
(1.16)
C
k
it
=


X
j∈I
(β
k
ji
)
1/σ
(C
k
jit
)
(σ−1)/σ


σ/(σ−1)
(1.17)
where φ> 0 denotes the elasticity of substitution between different crops (e.g., rice vs
wheat), and σ > 0 denotes the elasticity of substitution between different varieties of a
given crop (e.g., West Bengal vs Punjab rice). Finally,β
k
i
≥ 0 denotes crop and state spe-
cific demand shocks, whereas β
k
ji
≥ 0 denotes crop and origin-destination specific demand
shocks. The functional form implies that all states export each crop that they produce to
all other states (as long asβ
k
ji
> 0).
Technology—The outside good is produced under constant returns to scale using labor
only. The termA
0
it
> 0 denotes labor productivity in statei’s outside sector at timet. In the
agriculture sector, we assume that labor and parcels of land are perfect complements in
the production of each crop. CombiningL
fk
it
(ω) hectares of parcelω withN
fk
it
(ω) workers
enables a representative farmer to produce:
62
Q
fk
it
(ω) =A
fk
it
(ω) min{L
fk
it
(ω),N
fk
it
(ω)/ν
f
i
(ω)} (1.18)
whereA
fk
it
(ω)≥ 0 denotes the total factor productivity (TFP) of parcelω in field f if allo-
cated to cropk in statei at timet, andν
f
i
(ω)> 0 measures the time-invariant labor intensity
of the production process. Inspired by ? gravity model of trade, we assume that TFP and
labor intensity are independently drawn for each (i,f,ω,t) from a Fr´ echet distribution:
Pr{A
f1
it
(ω)≤a
1
, ..., A
fK
it
(ω)≤a
K
,ν
f
i
(ω)≤ν}
= exp



−γ


X
k∈K
(a
k
/A
fk
it
)
−θ
+ (ν/ν
i
)
−θ





(1.19)
where the constantγ is set such thatA
fk
it
=E[A
fk
it
(ω)] andν
i
=E[ν
fk
i
(ω)].
40
The termA
fk
it
≥
0 captures the average productivity of field f for growing cropk in statei at timet and is,
thus, shared by all plotsω∈f. A highA
fk
it
implies that on average all plots in farmf have
high productivity for growing cropk. In other words, it measures the comparative and
absolute advantage of a field in producing particular crops. The parameter θ> 1 measures
the extent of technological heterogeneity within each field. A higher value of θ will imply
higher specialization across different farms. Since we do not have access to disaggregated
data on labor intensity, we require average labor intensityν
i
> 0 to be identical across crops,
fields, and time. However, agriculture is allowed to be more labor intensive in some states
than in others.
Market Choice—This part of the model takes inspiration from the market setup of ?.
Upon harvest, farmers optimally choose the market where they want to sell, indexed by
40
Formally, we setγ≡ Γ

θ− 1
θ
!
−θ
, where Γ(·) denotes the gamma function; i.e.,
Γ(t) =
R
+∞
0
u
t−1
exp(−u)du for anyt> 0.
63
m∈M≡{1,...,M}. We assume that farmers are subject to iceberg trade costs, such that
the quantity of cropk actually reaching any marketm from farmf at timet is:
Q
fk
mit
(ω) =
Q
fk
it
(ω)
τ
f
mt
(1.20)
Trade costs between farmf and marketm at timet are constant for all parcelsω∈f,
and are defined as:
τ
f
mt
= (1 +ζd
f
m
)·ξ
f
mt
(1.21)
whered
f
m
is the geodesic distance between farmf and marketm, andζ is a scale param-
eter. The shock term,ξ
f
mt
, represents origin farm-market specific costs like broken roads,
availability of a truck, or a strike among intermediaries, which are not observable to the
econometrician but are known to the farmers. We follow ?, and assume for tractability
thatξ
f
mt
is drawn from a Weibull distribution such that:
Pr
h
ξ
f
mt
≤ξ
i
= 1− exp

−Υξ
λ

(1.22)
λ> 0 is the shape parameter and can be interpreted as an inverse measure of the dispersion
of shocks. Υ> 0 is the scale parameter and controls the efficiency of transporting goods
to a market. The distribution of shocks is i.i.d. across crops and over time, and shocks are
independent across markets. To incorporate trade restrictions,τ
f
mt
is set to∞ if farmf and
marketm lie in different states.
Intermediary—Each marketm can be thought of as an intermediary, a chain linking the
farmer to the consumer. Though each market can have multiple intermediaries, only a few
are active and cartelization among intermediaries is common. Incumbent intermediaries
also prevent new entrants Chand 2012. This fact makes our simplifying assumption that
each market is served by a single intermediary not too unrealistic.
64
An intermediarym in statei can purchase multiple cropsk∈K at timet, and sells the
same to retailers/consumers at priceP
rk
it
. Unlike the farmer, the intermediary is allowed
to cross state borders. However, interstate trade in crops may be subject to iceberg trade
costs. In order to sell one unit of a cropk in statej, intermediaries from statei must ship
Ψ
k
ij
units. Non-arbitrage therefore requires the price of a cropk produced in statei and
sold in statej to be equal to
P
rk
ijt
= Ψ
k
ij
P
rk
it
(1.23)
whereP
rk
it
denotes the local price of the domestic variety of cropk in statei.
1.5.2 Competitive Equilibrium
In a competitive equilibrium, all consumers maximize their utility, all farmers and inter-
mediaries maximize their profits, and all markets clear.
Farmer Profit Maximisation —In the outside sector, profit maximization requires that
w
it
=A
0
it
whenever the outside good is produced. Throughout this model, we assume that
labor endowments,N
it
, are large enough for the outside good to be produced in all states.
Thus, we can useA
0
it
in place of the wagew
it
and treat it as an exogenous parameter.
In the agricultural sector, profit maximization requires that the farmer first choose a
cropk, and subsequent to harvest, choose a marketm to sell. The price that farmers get in
marketm for cropk at timet is denoted byP
k
mit
. We can use backward induction to solve
for the farmer’s choice. Let
Ω fk
mit
≡ Pr{P
k
mit
Q
fk
mit
(ω) = max{P
k
1it
Q
fk
1it
(ω),...,P
k
Mit
Q
fk
Mit
(ω)}} (1.24)
65
denote the probability that a farmer, tilling parcel ω of a field f located in state i and
growing cropk at timet, chooses marketm. Given distributional assumptions:
41
Ω fk
mit
=




P
k
mit
1 +ζd
f
m




λ
P
m
′
∈M




P
k
m
′
it
1 +ζd
f
m
′




λ
(1.25)
This expression has an intuitive explanation. The probability of choosing a marketm
for cropk depends on how large the distance adjusted price of the crop inm is relative
to the distance adjusted price index of the crop. A higher price of crop k in market m
increases the probability of farmers selling their output inm, whereas an increase in the
price in other marketsm
′
relative tom reduces this probability. Similarly, if the distance to
m is large, that will depress the probability of choosingm. Currently, the farmers in state
i only take into account the prices in markets situated ini. An opening of trade borders
would lead the farmer to also factor in the prices in all other statesj∈I.
Conditional on choosing marketm, the farmer decides the cropk to grow at timet. Let
π
fk
it
(ω) denote the profits from parcel ω∈f in statei when farmer decides to growk at
timet. It can be expressed as:
42
π
fk
it
(ω) =A
fk
it
(ω)L
fk
it
(ω)P
fk
it
−w
it
N
fk
it
(ω) (1.26)
41
See F.3.2 for derivation.
42
See F.3.3 for derivation.
66
where
P
fk
it
=
X
m
′
∈M
Ω fk
m
′
it
P
k
m
′
it
=
P
m
′
∈M
(P
k
m
′
it
)
λ+1

1 +ζd
f
m
′

λ
P
m
′
∈M




P
k
m
′
it
1 +ζd
f
m
′




λ
(1.27)
denotes a probability weighted price of cropk for farmerf at timet, aggregated across all
markets. Profit maximisation requires that all parcels of land are (i) allocated to the crop
that maximizes the value of their marginal product if such value is greater than the wage
bill associated with operating that parcel, or (ii) left unused if the maximum value of their
marginal product is less than the wage bill. Given the production function in 3.6, the land
allocation can be solved as a simple discrete choice problem.
43
Let
∆ fk
it
≡ Pr{A
fk
it
(ω)P
fk
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
f1
it
, ...,A
fK
it
(ω)P
fK
it
}} (1.28)
denote the probability that a parcelω of a field f located in statei is allocated to cropk
at timet. Since there is a continuum of parcels within each field, ∆ fk
it
also corresponds to
the share of parcels allocated to that crop.
Given our distributional assumptions, standard algebra implies:
44
∆ fk
it
=
(A
fk
it
P
fk
it
)
θ
(α
it
)
θ
+
P
k
′
(A
fk
′
it
P
fk
′
it
)
θ
(1.29)
where α
it
≡A
0
it
ν
i
parameterizes cross-state differences in labor costs, because of differ-
ences in either wages or labor intensity. The higher α
it
is, the more costly it is to hire
workers to produce crops, and the smaller the share of a field f allocated to any given
43
See F.3.4 for derivation.
44
We use the property that given n draws{z
1
,...,z
n
}, where z
i
is distributed Fr´ echet with F
i
(z) =
exp{−(T
i
z
−θ
)}, the probability thatz
i
= max{z
1
,...,z
n
} is ∆ i
=T
i
/
n
P
j=1
T
j
M. Turner 2019.
67
cropk. Likewise, the higher the average value of the marginal product of land,A
fk
it
P
fk
it
,
the higher the share of field f allocated to cropk. In our model, the extent of technological
heterogeneity,θ, determines the elasticity of the relative supply of land to various crops.
Whenθ is higher, parcels are more homogeneous within a field, which makes the supply
of land more sensitive to changes in prices,P
fk
it
, or productivity,A
fk
it
.
LetQ
k
mit
=
P
f∈F
i
Ω fk
mit
R
1
0
Q
fk
it
(ω)dω denote the total output of cropk supplied to market
m in statei at timet. Intuitively, it is the expected output of cropk across all parcels of
land inf, weighted by the probability of choosing marketm, and this expression is then
summed across all the fields f in statei. Using the production function in 3.6 and the law
of iterated expectations, we must have:
45
Q
k
mit
=
X
f∈F
i
s
f
i
∆ fk
it
Ω fk
mit
E[A
fk
it
(ω)|A
fk
it
(ω)P
fk
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
f1
it
, ...,A
fK
it
(ω)P
fK
it
}] (1.30)
Given our distributional assumptions, one can also check that:
46
E[A
fk
it
(ω)|A
fk
it
(ω)P
fk
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
f1
it
, ...,A
fK
it
(ω)P
fK
it
}]
=A
fk
it
× (∆
fk
it
)
−1/θ
(1.31)
Note that because of the endogenous selection of fields into crops, the average productivity
conditional on a crop being produced is strictly greater than the unconditional average,
i.e. A
fk
it
× (∆
fk
it
)
−1/θ
>A
fk
it
45
See F.3.5 for derivation.
46
See F.3.6 for derivation.
68
Combining the above two equations, we obtain the following expression for the supply
of cropk in marketm in statei at timet:
Q
k
mit
=
X
f∈F
i
s
f
i
A
fk
it
Ω fk
mit
(∆
fk
it
)
(θ−1)/θ
=
X
f∈F
i
s
f
i
A
fk
it


















P
k
mit
1 +ζd
f
m




λ
P
m
′
∈M




P
k
m
′
it
1 +ζd
f
m
′




λ


















(A
fk
it
P
fk
it
)
θ
(α
it
)
θ
+
P
k
′
(A
fk
′
it
P
fk
′
it
)
θ




(θ−1)/θ
(1.32)
The quantity supplied of cropk at marketm is, thus, a function of the average TFP of
the crop, the price of the crop in other markets within the state, and also the productivity
and price of crops other thank.
Intermediary Price Setting—Each intermediarym can purchase multiple cropsk
′
∈K,
offering price P
k
′
mit
. They purchaseQ
k
′
mit
units of cropk
′
from the farmer, and sell the same
to retailers/consumers in different parts of the country at a price P
rk
′
ijt
. We assume that
there is no restriction on where the intermediary can sell, but transportation costs are
incurred only if the produce is sold outside the state.
Every intermediary exerts market power over farmers, which we model as Bertrand
competition for crops.
47
When deciding what price to offer for a crop, intermediaries form
expectations about how farmers respond. In other words, they internalize the upward
sloping crop supply curve in 1.32: each additional unit they purchase increases the price
of every other unit. ? has a similar setting but models exporters instead of local interme-
diaries. Additionally, he assumes that each exporter only buys a single crop, whereas in
our case, we assume that an intermediary can purchase all crops supplied in the market.
47
Market power can also be modeled as Cournot competition, but 1.32 does not lend itself to a closed
form inverse supply curve.
69
An intermediarym maximises the following profit function
max
n
P
k
′
mit
∀k
′
o
X
k
′
∈K

P
rk
′
it
−P
k
′
mit

Q
k
′
mit
(1.33)
subject to the supply curve in 1.32, whereP
rk
′
it
represents the retail price of commodityk
′
in statei at timet. The first order condition for price P
k
mit
can be expressed as:

P
rk
mit
−P
k
mit





X
f∈F
i
Q
fk
mit




λ




1− Ω fk
mit
P
k
mit




+ (θ− 1)Ω
fk
mit




λ + 1
P
fk
′
it
−
λ
P
k
mit












| {z }
tialQ
k
mit
/tialP
k
mit
−
(θ− 1)
X
k
′
∈K
X
f∈F
i

P
rk
′
mit
−P
k
′
mit





Q
fk
′
mit
Ω fk
mit
∆ fk
it




λ + 1
P
fk
′
it
−
λ
P
k
mit








| {z }
−tialQ
k
′
mit
/tialP
k
mit
=Q
k
mit
(1.34)
? states that the extent of oligopsony power of an intermediarym over an inputk can
be parametrized through an inverse input supply elasticityη
k
mit
, defined as:
η
k
mit
≡
tialP
k
mit
tialQ
k
mit
×
Q
k
mit
P
k
mit
If an intermediary has oligopsony power over input k, the input priceP
k
mit
increases if
more inputs are purchased. This, thus, has the interpretation of an input price ‘markdown
ratio’. Also, we can define inverse cross input supply elasticity as:
η
kk
′
mit
≡
tialP
k
mit
tialQ
k
′
mit
×
Q
k
′
mit
P
k
mit
which reflects how the price of commodity k in marketm changes if there is a change in
the supply of commodityk
′
to the said market.
70
Additionally, we define markup µ as the ratio of retail prices over marginal costs:
µ k
mit
≡
P
rk
mit
P
k
mit
Using these three definitions, 1.34 can be rewritten as:
P
k
mit
=




X
k
′
̸=k




µ k
′
mit
− 1
η
kk
′
mit




P
k
′
mit
Q
k
′
mit








η
k
mit
1 +η
k
mit
−µ k
mit




1
Q
k
mit
(1.35)
Thus, the price for cropk paid by an intermediarym is a function of the markdown
for not onlyk, but also the markdown fork caused by quantity supplied of other crops. It
also depends on the markup the intermediary may expect to receive in the retail market.
Finally, the intermediaries sell the produce to the consumer/retailer, with the retail
price of cropk in statei,P
rk
it
, set such that all the intermediaries selling in statei (including
from statej̸=i) sell at the same price, i.e.P
rk
jit
=P
rk
it
∀j∈I. Utility Maximisation—Given
equations (3.1), (3.2), (3.5) and (1.23), utility maximization by the representative agent
in each state requires that:
48
C
k
jit
=β
i
β
k
i
(
ˆ
P
rk
it
)
1−φ
P
l∈K
β
l
i
(
ˆ
P
rl
it
)
1−φ
β
k
ji
(Ψ
k
ji
P
rk
jt
)
−σ
P
n∈I
β
k
ni
(Ψ
k
ni
P
rk
nt
)
1−σ
∀ i,j∈I,k∈K (1.36)
where
ˆ
P
rk
it
≡


X
n∈I
β
k
ni

Ψ
k
ni
P
rk
nt

1−σ


1/1−σ
(1.37)
denotes the CES retail price index associated with cropk in statei at timet.
Market Clearing—Define Q
k
it
as the total output of cropk produced in statei at time
t. Since farmers are only allowed to sell their produce in statei,Q
k
it
=
P
m∈M
Q
k
mit
. Trade in
48
See F.3.7 for derivation.
71
crops is subject to iceberg trade costs, which implies market clearing for all varieties of all
crops requires
Q
k
it
=
X
j∈I
Ψ
k
ij
C
k
ijt
∀{i,j}∈I andk∈K (1.38)
Parcels of land may remain idle if the value of their marginal product is below the labor
cost required to produce on these parcels. Thus, by construction, land demand is weakly
less than land supply at all locations. Finally, under the assumption that the outside good
is produced in all states, the amount of labor demanded by the outside sector adjusts to
guarantee labor market clearing at the wage equal toA
0
it
.
DEFINITION 1. Given parameters β
i
,β
k
i
,β
k
ji
(demand shifters), φ,σ (elasticities of
substitution), λ
t
,ζ,Ψ
k
ij
(trade cost for farmers and intermediaries), θ (technological het-
erogeneity), andµ (intermediary markup), a competitive equilibrium consists of, for each
statei∈I≡{1, ..., I} and each time periodt:
1. inputs for crops{L
fk
it
(ω),N
fk
it
(ω)}
k∈K,f∈F
i
, and outside good{N
0
it
},
2. output of crops{Q
fk
it
(ω)}
k∈K,f∈F
i
, and outside good{Q
0
it
},
3. optimal market choice at each farm{m(f)},
4. domestic trade flows {X
k
ijt
}
k∈K,j∈I
, which is the total value of exports of cropk∈K
from statei to statej, expressed in ,
5. consumer prices{P
rk
it
}
k∈K
, intermediary prices,{P
k
mit
}
k∈K,m∈M
, and outside good
price{P
0
it
},
6. final crop consumption {C
k
it
}
k∈K
and outside good consumption{C
0
it
},
such that:
1. farmers maximise their profits by choosing the optimal crop (1.29) and market (1.25);
2. intermediaries maximise their profits according to 1.34
72
3. consumers maximise their utility to solve 1.36
4. market for all crops clears, which requires:
X
f∈F
i
Q
fk
it
=
X
j∈I
Ψ
k
ij
C
k
ijt
∀i∈I andk∈K : (1.39)
In the remainder of this paper we will use the model outlined in this section to study
the consequences of climate change. We will compute competitive equilibria for states
with contemporary agricultural productivities and trade restrictions, compute competi-
tive equilibria for counterfactual economies with post–climate change productivities and
open trade borders, and then compare welfare levels across equilibria. However, we first
need to estimate the unknown structural parameters of our model, and we describe below
the methodology and data used.
1.6 Estimation
To simulate the model described in 1.5 and run counterfactual, we require estimates of
demand- and supply-side parameters. 1.6.1 details the estimation methodology for de-
mand side parameters, while 1.6.2 focuses on the supply side parameters.
1.6.1 Demand
We follow Costinot, Donaldson, and C. Smith 2016 closely for our demand side estimation.
Similar to their methodology, it involves three steps, each pertaining to a different level of
the nested demand system. The first step uses data on bilateral shipment flows (total
quantity and not total value) of crops between states (N
k
ijt
), retail prices (P
rk
it
), and crop
yields at the district level (A
k
dit
) to estimate the elasticity of substitution between different
state varieties of a given crop,σ. In addition, it allows us to estimate a composite of the
lower-level demand shifters (β
k
ijt
) and trade costs for intermediaries (ψ
k
ijt
). Second, we
73
use the estimates from the previous step to construct crop-specific retail price indices,
ˆ
P
rk
it
. This, combined with data on crop quantity, N
k
jt
=
P
i∈I
N
k
ijt
, allows us to estimate φ
— the elasticity of substitution between different crops — and mid-level demand shifters,
β
k
jt
. Finally, we construct data on total crop expenditures, X
jt
=
P
X
k
jt
, to estimate the
upper-level demand shifters,β
jt
.
Step 1—Define the value of exports of crop k at time t from state i to state j, X
k
ijt
=
P
rk
ijt
C
k
ijt
. Using the non-arbitrage condition in 1.23, we can rewrite the value of exports as:
X
k
ijt
=

Ψ
k
ijt
P
rk
it

C
k
ijt
=β
jt
β
k
jt

ˆ
P
rk
jt

1−φ
P
l∈K
β
l
jt

ˆ
P
rl
jt

1−φ
β
k
ijt

Ψ
k
ijt
P
rk
jt

1−σ
P
n∈I
β
k
njt

Ψ
k
njt
P
rk
nt

1−σ
∀{i,j}∈I andk∈K (1.40)
When estimating the lower level of our demand system, we consider the cases of zero and
nonzero inter-state trade flows separately. If X
k
ijt
= 0, we simply setβ
k
ijt

Ψ
k
ijt

1−σ
= 0. If
X
k
ijt
> 0, we take logs and rearrange equation 1.40 as:
ln

X
k
ijt
/X
k
jt

=M
k
jt
+ (1−σ)ln

P
rk
it

+ε
k
ijt
(1.41)
where the first term on the right-hand side,
M
k
jt
≡−ln



X
n∈I;X
k
njt
>0
β
k
njt

P
rk
nt
Ψ
k
njt

1−σ



can be treated as an importer-crop-year fixed effect while the final term ε
k
ijt
≡ ln

β
k
ijt

Ψ
k
ijt

1−σ

reflects idiosyncratic year-specific demand shocks across varieties of different crops as well
as trade costs. Without loss of generality, we normalize these shocks such that
X
i∈I;X
k
ijt
>0
ε
k
ijt
= 0 (1.42)
74
Equilibrium retail prices of crop (P
rk
it
) depend on demand shocks, ε
k
ijt
. To address this
endogeneity in 1.41, we need exogenous supply shocks that are correlated withP
rk
it
but
uncorrelated withε
k
ijt
. We construct the following instrument based on the ICRISAT 2018
data,
Z
rk
it
= ln


1
D
i
X
d∈D
i
A
k
dit


(1.43)
which corresponds to the log of the arithmetic average of cropk’s yields across all districts
in statei at timet. Our exclusion restriction is thatE
h
Z
rk
it
ε
k
ijt
i
= 0.
Note that our data contains information on quantity of crops traded between states,
and not their value. However, our use of the Eaton and Kortum 2002 model allows us to
overcome this missing data problem. This is because their model predicts that the fraction
of quantity imported byj originating fromi,
N
k
ijt
P
i∈I
N
k
ijt
, should equal the fraction ofj’s value
of imports fromi,
X
k
ijt
P
i∈I
X
k
ijt
, in expectation. This implies that we can replace the left-hand
side of 1.41 with ln

N
k
ijt
,
P
i∈I
N
k
ijt
!
to estimate the model.
The results from the instrumental variable regression are reported in xxx. Our estimate
of the elasticity of substitution between different varieties of the same crop, σ, is 25.66, with
a standard error of 14.6 when clustered at the crop-importer and crop-exporter levels. Fur-
thermore, our instrument has a strong first stage (F-stat of 34.28) and has the expected
negative sign with a coefficient of -0.052 — implying a one percent increase in yields leads
to a 5.2 percent fall in retail prices. Though our elasticity estimate is higher than Costinot,
Donaldson, and C. Smith 2016, this is expected given we are looking at substitution be-
tween different varieties of a crop but produced in the same country. Therefore, the quality
differentiation across varieties will be lower, making it easier to substitute between them.
Having estimatedσ, we subsequently solve forβ
k
ijt

Ψ
k
ijt

1−σ
as residuals. Specifically,
we find β
k
ijt

Ψ
k
ijt

1−σ
for alli,j∈I andk∈K for whichX
k
ijt
> 0 so that equations (1.41)
and (1.42) simultaneously hold for all crops, states and years. This estimation procedure
does not allow us to identify separately lower-level demand shifters,β
k
ijt
, from trade costs,
75
Ψ
k
ijt
. However, the composite shock,β
k
ijt

Ψ
k
ijt

1−σ
, is sufficient to construct equilibria in
1.7.
Step 2—The second step of our demand estimation is similar to the first one: the retail
price index,
ˆ
P
rk
jt
, plays the role of the individual crop price,P
rk
it
, whereas crop expenditure,
X
k
jt
, plays the role of bilateral trade flows, X
k
ijt
. Note that unlikeP
rk
it
, we do not observe
ˆ
P
rk
jt
in the data and construct it as
ˆ
P
rk
jt
=


X
i∈I
β
k
ijt

Ψ
k
ijt
P
rk
it

1−σ


1/1−σ
using data on crop prices,P
rk
it
, as well as our estimates ofσ andβ
k
ijt

Ψ
k
ijt

1−σ
from Step
1.
For all crops and states with positive quantity traded in yeart,X
k
jt
> 0, we can again
use 1.40 and take logs to get
ln

X
k
jt
/X
jt

=M
jt
+ (1−φ)ln

ˆ
P
rk
jt

+ε
k
jt
(1.44)
where the first term on the right-hand side,
M
jt
≡−ln



X
l∈K;X
l
jt
>0
β
l
jt

ˆ
P
rl
jt

1−φ



can now be treated as an importer-time fixed effect, and the final term, ε
k
jt
≡ ln

β
k
jt

, re-
flects idiosyncratic year-specific demand shocks across crops. Without loss of generality,
we again normalize these shocks such that
X
k∈K;X
k
jt
>0
ε
k
jt
= 0 (1.45)
76
There still exists endogeneity issues between demand shocks (ε
k
jt
) and prices (
ˆ
P
rk
jt
) at
this higher level of aggregation, which could potentially bias our estimates ofφ. To address
this, we now instrument
ˆ
P
rk
jt
withZ
rk
jt
. The exclusion restriction now equalsE
h
Z
rk
jt
ε
k
jt
i
= 0.
As before, since we do not have data on either the value of specific crops imported from all
exporters or the total value of crops imported, we replaceln

X
k
jt
/X
jt

withln

N
k
jt
/N
jt

.
Results, reported in xxx, indicate that the IV estimate for the elasticity of substitution
between crops,φ, equals 9.39 with standard errors of 2.3 when clustered at the importer
level. Also, the first stage estimate equals -0.055, which can be interpreted as a one percent
increase in yields leading to a 5.5 percent fall in prices. As in Step 1, once the elasticity of
substitution,φ, is known, we can solve forβ
k
jt
for allj∈I andk∈K such thatX
k
jt
> 0, as
residuals using equations (1.44) and (1.45).
Step 3—The final step of our procedure estimates the upper-level demand shifters, β
jt
.
The assumption of log preferences at the upper level implies thatβ
jt
’s can be read directly
from data on total expenditure across crops. Specifically, using 1.40, we can show that
β
jt
=X
jt
for allj∈I at timet.
Since we only have data on the quantity of crops imported, and not on the value of
imports, we need a proxy for the price of imports to constructX
jt
. To this end, we assume
that the value of exports of cropk fromi toj, X
k
ijt
, equals the average price ofk across
all the markets m within state i, multiplied by the quantity exported from i to j, N
k
ijt
.
Summing this value across alli∈I andk∈K for statej provides us withX
jt
for yeart.
1.6.2 Supply
There are four supply side parameters we need to estimate: the inverse measure of the
dispersion of shocks (λ), the scale parameter for the trade costs (ζ), the extent of tech-
nological heterogeneity (θ), and the state-specific labor cost shifters ( α
i
). We proceed in
two steps. First, we use data on crop prices in different markets ( P
k
mit
), distance between
farms and markets (d
f
m
), and crop quantity produced in each state (Q
k
it
) to estimateλ and
77
ζ using a generalized method of moments (GMM) estimation procedure. Then, we use
the previous estimates along with data on farm productivity (A
fk
i
) in a nonlinear least
squares (NLS) framework to estimateθ andα
i
.
Step 1—We know that Ω fk
mit
from 1.25 represents the probability that a farmerf located
in statei and growing cropk at timet, chooses marketm. This probability can be used to
calculate the share of cropk produced in statei that reaches a marketm at timet. Denoting
the same byS
k
mit
, we can calculate it as the share of cropk that each farmer takes to market
m at timet (Q
fk
it
), summed across all farmers, and divided by the total quantity of cropk
produced in statei at timet (Q
k
it
). The expression takes the following form:
S
k
mit
=
P
f∈F
i
Ω fk
mit
Q
fk
it
Q
k
it
=
X
f∈F
i


















P
k
mit
1 +ζd
f
m




λ
Q
fk
it
P
m
′
∈M




P
k
m
′
it
1 +ζd
f
m
′




λ














,
Q
k
it
(1.46)
Now, we use 1.46 to carry out a GMM procedure to estimateλ andζ. In particular, we
choose Θ ={λ,ζ}, with true parameter value Θ
0
, to minimize the distance between mo-
ments of the data and their estimated counterparts. Letg(Y
m
,θ) be a continuous and con-
tinuously differentiable function of θ andY
m
, where the latter is a market-specific vector
of parameters like prices and distance to farms. Then the population moment conditions
are such that:
E[g(Y
m
,Θ
0
)] =E






S
k
mit
−
P
f∈F
i
Ω fk
mit
Q
fk
it
Q
k
it






= 0
78
The corresponding sample moments are given by:
g
m
(Θ) =
1
M
X
m∈M
g(Y
m
,Θ) = 0
Our GMM estimator can, therefore, be written as:
ˆ
Θ = arg min
(Θ)


1
M
X
m∈M
g(Y
m
,Θ)


T
ˆ
W


1
M
X
m∈M
g(Y
m
,Θ)


(1.47)
where
ˆ
W is the optimal weighting matrix. We use numerical methods to find the required
gradients. Note that we do not have data on total crop produced per farm in any year,
so we proxy that using ICRISAT 2018 data on district level yearly output of crops (Q
k
dit
).
Specifically, we assume that all the farms f that fell within the district had the same output,
which we can calculate as follows:
Q
fk
it
=
Q
k
dit
F
i
Finally,Q
k
it
is the sum of output of cropk across all districtsd in statei at timet.
Our estimates ofλ andζ equal 1.84 and 0.07, respectively. Sinceλ is inversely related
to the dispersion of transportation cost shocks, a low value ofλ implies that farmers face
substantial heterogeneity in trade costs across different states. To interpret the scale pa-
rameterζ, we use H.1 to calculate the elasticity of trade costs with respect to distance:
tialτ
tiald
×
d
τ
=
ζd
1 +ζd
.
The equation above implies that the change in trade costs with respect to change in distance
is not uniform, and depends on the original distance being traversed. For instance, if the
distance between the farm and the market is 10 kms, a 10 percent increase in distance — or
1 km — leads to a 4.1 percent increase in trade costs for the farmer. On the other hand, if the
distance was 100 kms, increasing the same by 10 percent will increase the transportation
costs by approximately 8.7 percent.
79
Step 2—The remaining supply-side parameters that need to be estimated are the extent
of technological heterogeneity,θ, as well as the state-specific labor cost shifters in different
years,α
it
. We do not need to estimate the productivity of fields for different crops, A
fk
i
—
the main variable that changes in our model under a climate change scenario — as it is di-
rectly observable in the GAEZ data. However, GAEZ data is not available at a yearly level;
rather there is only one observation per field for the time period 1980-2010. Therefore,
we only use data pertaining to the year 2010 for estimating the supply parameters in this
step.
49
Naturally, the labor cost shifters too will only pertain to the year 2010. Henceforth,
we remove the time subscript (t) in this subsection.
Using 1.32, we can denote the predicted supply of cropk in statei as a function of the
unknown parametersθ andα
i
:
Q
k
i
(θ,α
i
) =
X
m∈M
i
X
f∈F
i
s
f
i
A
fk
i


















P
k
mi
1 +ζd
f
m




λ
P
m
′
∈M




P
k
m
′
i
1 +ζd
f
m
′




λ



















(A
fk
i
P
fk
i
)
θ
(α
i
)
θ
+
P
k
′
∈K
(A
fk
′
i
P
fk
′
i
)
θ





(θ−1)/θ
(1.48)
Next, letL
i
(θ,α
i
) denote the predicted land allocated to all crops in statei as a function
ofθ andα
i
. This is calculated as the share of field f which is allocated to cropk (∆ fk
i
from
1.28), multiplied by the field size, s
f
i
, and sum across all crops and fields. Specifically,
L
i
(θ,α
i
)≡
X
k
X
f
s
f
i





(A
fk
i
P
fk
i
)
θ
(α
i
)
θ
+
P
k
′
∈K
(A
fk
′
i
P
fk
′
i
)
θ





(1.49)
To estimate θ and α
i
, we follow the same procedure as Costinot, Donaldson, and C.
Smith 2016, i.e. choose a value of θ, and conditional on it, find the vector of labor cost
shifters,α
i
, such that the total amount of land allocated to crops as predicted by the model,
49
We could have used yields data from ICRISAT 2018 but GAEZ is more accurate and spatially disaggre-
gated (district versus farm).
80
L
i
(θ,α
i
), exactly matches the total amount of land allocated to crops in the data, L
i
, for
all states. Next, given the vector of labor cost shifters (α
i
) for all states, we search forθ
such that the difference between the output predicted by the model, Q
k
i
(θ,α
i
), and output
observed in the data,Q
k
i
, is minimised. This algorithm can be formally expressed as the
following non-linear least squares problem:
min
θ,α
i
X
i∈I
X
k∈K
h
lnQ
k
i
(θ,α
i
)− lnQ
k
i
i
2
subject to
L
i
(θ,α
i
) =L
i
for all i∈I.
Our estimate ofθ equals 1.82, which suggests that within-field, within-crop productiv-
ity dispersion in Indian agriculture is large. This is reassuringly close to the estimate of
Sotelo 2020, who finds a value of 1.658 for θ using data from Peru.
1.7 Counterfactual Analysis
We now use the estimated parameters to simulate our model and run policy counterfactu-
als that involve climate change, with and without inter-state trade barriers. Specifically, we
can use the model to run the following two counterfactuals: (i) welfare impact of climate
change under a policy with inter-state trade restrictions, and; (ii) welfare impact of climate
change under a policy without inter-state trade restrictions. Under both counterfactuals,
we allow for production and trade patterns to fully adjust.
To run either of these counterfactuals, we first need to solve for the competitive equi-
librium before climate change. This equilibrium is characterized by the market supply
curve in 1.32, and the following three conditions: (i) intermediary profit maximization in
1.34; (ii) consumer utility maximisation in 1.36, and; (iii) market clearing in 1.38. Crop
productivity for different farms, A
fk
i
, is the structural parameter which will change un-
der the climate change scenario. Therefore, for the first equilibrium, we use pre-climate
81
change GAEZ data. Subsequently, for the counterfactual equilibrium with climate change,
we use the exact same equations and structural parameters except for crop yields, which
is replaced with post-climate change productivity, (A
fk
i
)
′
, as measured by GAEZ.
The key mechanism driving any differences between the equilibrium with and without
climate change is the change in productivity across different farms and crops caused by
global warming. Climate change will affect comparative advantage in crop yields across
different regions of the country. This, in turn, alters supply as farmers change the use of
intermediate inputs and substitute between crops, which then impacts mandi and retail
prices — an effect that feeds back into the prices farmers can get.
It is worth noting that the welfare consequences of changes in comparative advantage
will also depend crucially on the spatial competition faced by farmers, as it directly in-
centivizes farmer adaptation. To see this, consider the following example: a farmer near
a state border grows rice, but climate change shifts their comparative advantage towards
wheat. The price offered to wheat farmers in the nearby mandi, however, is not competitive
due to intermediary market power. Thus, despite rice yields falling, the farmer does not
substitute. Lifting border restrictions would necessarily improve the welfare outcome by
increasing the farmer’s choice set, both in terms of accessible markets and crop choices.
The channel for this welfare improvement is intuitive. First, opening state borders di-
rectly impacts the probability of the farmer choosing a market, as seen in 1.25. This occurs
because reducing distance — and, thus, transportation costs — between farms and po-
tential markets increases farmers’ arbitrage opportunities. Second, the change in market
choice probability subsequently affects farmers’ probability of allocating land to a crop
through changes in the average value of the marginal product of the field (1.29). These
changes in farmer input decisions, in turn, change quantities supplied to each market
(1.32). Now, the prices received by farmers will be affected through two sources: the
change in quantity, and the change in bargaining power, as the intermediary faces in-
creased competition now from across the border. This increase in competition affects each
82
intermediaries share in the market, which will affect the markdowns. Importantly, the
changes in intermediary market power near the borders has ripple effects across inte-
rior markets through this change in quantity 1.34. Finally, changes in production, market
choice of farmers and intermediary market power could also adjust retail prices which in
turn will feed back into the prices farmers’ receive. This change would eventually incen-
tivize the farmer at the border to substitute from rice to wheat, as predicted by comparative
advantage. Therefore, change in production and incomes brought about by opening trade
borders could aid in mitigating the climate change impact.
Our assumption of quasi-linear preferences allows us to compute welfare changes as
changes in social surplus, expressed as a fraction of GDP in the initial equilibrium:
∆ W
i
=
(Y
i
)
′
−Y
i
+ (β lnC
i
−P
i
C
i
)
′
− (β lnC
i
−P
i
C
i
) + (π)
′
−π
Y
i
(1.50)
whereY
i
andπ
i
are the GDP and intermediary profits in the initial equilibrium, respec-
tively, while primes denote the analogous variable in the counterfactual equilibrium.
We find that climate change reduces welfare in India by 2.1 percent of total GDP, as-
suming border restrictions for farmers remain in place. Note that up until now, we have
set the distance between farms in state i and markets in state j, for i̸=j, as∞. In the
subsequent counterfactual where we remove the trade barriers, the distance is set to the
actual geodesic distance, similar to if farms and markets were in the same state. Under
this counterfactual, where farmers can access markets across state borders, the country
still experiences a 1.81 percent fall in GDP. However, this is 13.8 percent lower, implying a
mitigation of the negative impacts. This illustrates how market distortions created by gov-
ernment policies could hinder adaptation, and how removing the same could expand the
adaptation portfolio of farmers, thus helping countries mitigate the negative consequences
of climate change.
83
1.8 Discussion and Conclusion
Extreme and frequent heat events, induced by climate change, are predicted to acceler-
ate crop failures, leading to increased food prices and greater food insecurity IPCC 2022.
Given this portentous scenario, a higher magnitude and rate of farmer adaptation are cru-
cial to flatten the slope of the climate damage function. However, to what extent does the
effectiveness of adaptation responses depend on a country’s institutional framework? We
offer an insight into this question by studying the impact of institution-led distortions in
market competition on farmer climate-change adaptation in India. Using spatial variation
in intermediary market power — an unintended consequence of regulations governing
agricultural marketing — we show that higher competition among buyers of agricultural
produce helps farmers alleviate the detrimental impact of extreme heat. This effect is pri-
marily driven by an increase in input usage in more competitive areas, a response to higher
expected prices post climate shocks. Subsequently, we structurally estimate a spatial trade
equilibrium model to test the implications of eliminating these market distortions under a
climate change scenario. Our results indicate that there is potential for substantial welfare
gains if government policies distorting market competition are removed, highlighting the
positive role of free markets in facilitating adaptation.
Though our setting — distortions in intermediary competition emanating from Indian
agricultural laws — is a specific one, we believe that many of its characteristics, and the
lessons derived from it, apply more broadly. First, we show that well-intended govern-
ment policies can distort adaptation behavior. A similar result is outlined by Annan and
Schlenker 2015, who show that federal crop insurance can lead to moral hazard, and thus,
discourage private adaptation efforts. Similarly, Matthew E. Kahn and Lall 2021 hypothe-
size that government investment in resilience infrastructure can encourage migration into
risky areas, increasing the population’s overall risk exposure. Second, intermediary power
in agricultural value chains is ubiquitous in developing economies, for e.g. Ecuador Zavala
2020, Kenya Dhingra and Tenreyro 2020, and Rwanda Macchiavello and Morjaria 2021,
84
among others. Our results indicate that adaptation to climate change in such countries
will therefore, be rendered more challenging as farmers would also need to overcome the
distortions to adaptation incentives caused by market power.
While State intervention in case of market failures is valuable, there needs to be recog-
nition that government and private individuals respond to each other Kousky, Luttmer,
and Zeckhauser 2006. Therefore, these strategic interactions need to be internalized when
designing policies, else the resulting distortions arising from unintended consequences
could have negative implications in a world afflicted by climate change.
85
Chapter 2
Adapting to Flood Risk: Evidence from a Panel of Global
Cities
1
2.1 Introduction
Rising greenhouse gas emissions raise the likelihood of more extreme precipitation events
that increase the risk of local flooding (Trenberth 2005; AghaKouchak, Chiang, Huning,
Love, Mallakpour, Mazdiyasni, et al. 2020). Such floods often kill, displace people, and
destroy capital. Areas that experience such disasters suffer from a disruption in economic
activity (Hsiang and A. S. Jina 2014; Elliott, Strobl, and Sun 2015; Kocornik-Mina, McDer-
mott, Michaels, and Rauch 2020). Roughly 2,600 floods took place each year from 2012 to
2018, leading to an estimated 31,000 total deaths and US$240 billion in damages (Guha-
Sapir, Below, and Hoyois 2016).
Households and firms are not passive victims in the face of such natural disasters. Peo-
ple can adapt through three different strategies. First, they can invest in self-protection by
avoiding living in increasingly risky areas and by investing in strategies to reduce their
place based risk exposure (see Collado and Wang 2020; Douglas, K. Alam, Maghenda,
Mcdonnell, McLean, and Campbell 2008). Second, government can invest in local public
goods to offset risk (Kousky, Luttmer, and Zeckhauser 2006). Third, they can purchase
1
Co-authored with Sahil Gandhi, Matthew E. Kahn, Somik Lall, and Vaidehi Tandel
86
insurance or demand public insurance such that they receive financial transfers if a disas-
ter does take place (I. Ehrlich and Becker 1972). Economic theory emphasizes that these
strategies can sometimes be regarded as complements, and at other times as substitutes.
Publicly provided insurance for flood damage or fire damage can crowd out purchase
of private insurance. Similarly, public investment in resilience infrastructure can crowd
out private self protection strategies (Kousky, Luttmer, and Zeckhauser 2006). On the
other hand, a complementary strategy would involve the government adopting a match-
ing grant formula such that for every private dollar invested in resilience, the government
spends a dollar on local public goods. At a given point in time, an urban population’s risk
exposure and the place’s risk exposure depends on the complex interplay of these various
factors.
We use both time series and cross-sectional approaches to test several natural disas-
ter adaptation hypotheses. We use a global city data set of 9,468 cities, complemented by
data on night lights and disasters from 2012 to 2018, to shed light on the net effects for
cities of private self protection efforts and public investment in resilience. We first anal-
yse the population growth rate in flood prone cities to test if people are migrating away
from vulnerable areas. Second, we study the death toll from flood events across different
cities, and ask if deaths per disaster have been declining over time. Third, we examine the
impact of floods on economic activity, as proxied by intensity of night lights
2
, and how
that differs across developed and developing nations. Additionally, we test whether ge-
ographical and topographical characteristics of cities matter in determining the extent of
impact. Fourth, we analyse the pattern of recovery in the aftermath of a flood to ascertain
how long it takes for cities to come back to their pre-disaster level of night lights. Finally,
we test three different flood adaptation hypotheses — whether higher income or produc-
tivity helps a city to mitigate the effect of floods; whether repeated exposure to flooding
2
A growing number of studies use night lights intensity as a proxy for economic activity (V. Henderson,
Storeygard, and D. N. Weil 2011; Donaldson and Storeygard 2016).
87
reduces the negative impact of subsequent events (the novelty factor as defined by Guit-
eras, A. Jina, and A Mushfiq Mobarak 2015); and if flood protection infrastructure does
in fact attenuate the effect of floods. Places with a higher likelihood of flooding can invest
in costly infrastructure, such as dams, to offset risk. Such public investments may attract
more people to move to the city because the area is now perceived to be safer. They may
also reduce the marginal effect on economic activity of a flood in such a city. We quantify
these effects using data from four major nations – China, India, Mexico, and the United
States.
We report three main findings. First, we document that urban population growth is
significantly lower in cities that experienced more severe floods in the recent past (cate-
gorised asRisky cities). The results hold when we separately analyse high income and
low income countries, but are only significant for the former group. However, the effects
are small, with population growth in risky cities lower by 0.4-0.5 percentage points. This
provides suggestive evidence that cities with recurrent floods do lose some of their de-
sirability for potential and current residents. Furthermore, deaths from urban floods are
lower (1.1 percent) in risky cities, but only for high income countries. For vulnerable cities
in low income countries, the death rate is actually 2.7 percent higher.
Second, using monthly night lights data, we show that floods have a significant neg-
ative impact on economic activity. The effect is unsurprisingly higher in cities in low in-
come countries, with night lights falling by 8.3 percent, as opposed to 1.4 percent in high
income countries. Importantly, we find that high altitude cities suffer more, but only in
low income countries. Our findings are robust to using extreme precipitation events as a
measure of floods. In addition, the recovery dynamic results indicate that economic activ-
ity is restored to pre-disaster levels within one month in cities in high income countries.
However, it takes two months for night lights to recover in low income countries, with the
effect size still a significantly negative 4.9 percent after the first month.
88
Third, we find some evidence of adaptation and resilience to climate shocks. Richer
cities, as measured by city level GDP per capita, experience a lower fall in night lights
during a flood event, controlling for country specific income. Specifically, the effect of
floods on low income cities is 9.3 percent, but the same is attenuated by 75-86 percent in the
case of medium and high income cities, respectively. Furthermore, high risk cities or cities
that experienced recurrent severe floods in the past suffer less from flooding by almost half
than cities that don’t face recurrent floods. Lastly, cities protected by dams suffer more
floods, but the effect of each flood is mitigated by a substantial 40 percent. Thus, flood
protection infrastructure does aid in reducing the negative impact on economic activity.
In terms of population growth, we find that high risk cities with dams experience a fall
in population growth, but low risk cities with dams experience a 9.5 percentage point
increase in growth over cities with no dams. Together, our various pieces of empirical
work support the adaptation progress.
Our work is most closely related to Kocornik-Mina, McDermott, Michaels, and Rauch
2020, who study the short-run effects of 53 large flood events around the world. They use
flood maps from 2003 to 2008 for 1,868 cities located primarily in developing countries.
They find that urban economic activity tends to concentrate in low-lying areas and is vul-
nerable to flood risks. Using annual lights at night data, they document that large floods
lead to a decline in the intensity of night lights by 2 to 8 percent in the year of the flood.
However, economic activity recovers to pre-flood levels in the year immediately following
the flood event. Finally, they document that economic activity does not relocate to safer
areas in the aftermath of floods, with the exception of newly populated parts of cities.
We build on their work by focusing on adaptation, and expanding the scope and extent
of the analysis in five different ways. First, our study is based on a large, globally repre-
sentative sample of 9,468 cities from 175 countries. Second, we focus on a more recent time
period from 2012 to 2018, and include the universe of floods (18,420) during this time pe-
riod in our study. Third, we proxy economic activity using monthly night lights data from
89
Visible and Infrared Imaging Suite (VIIRS) instruments, which have a higher spatial and
radiometric resolution than its predecessors.
3
Importantly, the monthly frequency of the
night lights data allows us to study short run effects and recovery dynamics, which is not
possible with data at an annual frequency. Fourth, we look at the heterogeneous effects of
floods based on income classification of countries, which gives us an insight into the dif-
ferential impact and recovery from disasters based on economic development. Finally, our
major focus is adaptation, and we study the role of income, infrastructure and familiarity
in mitigating the impact of disasters.
Our paper contributes to three strands of literature. First, we contribute to the litera-
ture on adaptation to climate shocks. Desmet and Rossi-Hansberg 2015 have shown using
a dynamic spatial model that the consequences of global warming can be mitigated by the
ability of agents and goods to move across space. Therefore, migration is an important
and effective adaptation method to limit the negative economic impact of climate change,
and many studies have extensively documented the same. Boustan, Matthew E Kahn, and
Rhode 2012 use US migration data from 1920’s and 1930’s to document evidence of pri-
vate self protection by men who move away from tornado hit areas. Using grid-level data
on temperatures from 1970-2000 for a panel of countries, Peri and Sasahara 2019 find that
rising temperatures induce rural-urban migration in middle income countries. Hornbeck
2012 analyzes the aftermath of American Dust Bowl in 1930’s, and finds that the main
margin of economic adjustment was out-migration from affected areas. Strobl 2011 anal-
yses the economic impact of hurricanes between 1970-2005 in the US, and estimates that a
quarter of the economic effect of an average hurricane is due to richer people moving out
as a consequence of the hurricane. In contrast, Deryugina 2011 looks at the effects of hur-
ricanes in the 1980’s and 1990’s in the US, and finds no impact on population. Similarly,
Kocornik-Mina, McDermott, Michaels, and Rauch 2020 find little evidence of adaptation,
3
See Section 1.3 for a detailed discussion on night lights.
90
at least in the sense of a relocation of economic activity away from the most vulnerable
locations, except for newly populated parts of cities.
Migration, however, is just one instrument in the adaptation toolbox.
4
Barreca, Clay,
Deschenes, Greenstone, and Shapiro 2016 study the effect of high temperatures on mor-
tality in the US, and find that the diffusion of residential air conditioning has helped fa-
cilitate a decline in hot day–related fatalities by 75 percent since 1960’s. Likewise, Park,
Goodman, Hurwitz, and J. Smith 2020 show that heat inhibits learning and that school air
conditioning helps mitigate this effect. Bunten and Matthew E Kahn 2017 argue in favor
of building less durable structures as an adaptation technique to preserve an option value
to walk away from areas facing a higher climate risk. Arag´ on, Oteiza, and Rud 2021 ex-
amine the adaptation response of Peruvian farmers to extreme heat. They document that
the farmers adapt by increasing the area planted, using more domestic labor on the farm,
and changing the crop mix to attenuate the effect of extreme heat on output.
Hsiang and Narita 2012 estimate the extent of adaptation to tropical cyclones using the
global cross-section of countries. They find evidence that countries with more intense
tropical cyclone climates suffer lower marginal losses from an actual event, indicating
adaptation to climatological risk. Burke and Emerick 2016 study adaptation in the con-
text of agriculture in the US, and find that longer run adaptation only partially mitigate
the adverse impacts of heat on agricultural productivity. Our paper extends the scale of
the analysis to the entire globe, and indicates that migration is an adaptation tool, but
it is more pertinent in high income countries. Additionally, evidence of richer countries
adapting can also be seen in the form of lower deaths per flood. Moreover, productive
cities are better adept at mitigating the impact of floods. Familiarity with floods also helps
as a tool to reduce the impact, and dams are an effective infrastructure to reduce the effect
of floods.
4
See Klein, Midgley, Preston, M. Alam, Berkhout, Dow, et al. 2015 for an extensive survey of the key
adaptation opportunities available in response to climate change.
91
Second, we contribute to the nascent literature on Climate Justice which predicts that
disruptions from extreme weather will disproportionately affect the developing world,
particularly the poor and most vulnerable (Mendelsohn, Morrison, Schlesinger, and An-
dronova 2000; World Bank, ADB, AfDB, BMZ, DFID, DGIS, et al. 2003; Mendelsohn, Dinar,
and Williams 2006; Stern 2006; Tol 2009). Various reports by the Intergovernmental Panel
on Climate Change (IPCC) (Portner, D. C. Roberts, Adams, Adler, Aldunce, Ali, et al.
2022; Houghton, Ding, Griggs, Noguer, Linden, Dai, et al. 2001) estimate that poor coun-
tries will suffer the bulk of the damages from climate change, with economic damages per
capita from climate change for developing countries higher as a fraction of income. This
is mainly due to the economic importance of climate-sensitive sectors for these countries.
Moreover, the limited capacity to anticipate and respond to climate change can also impact
adaptation in poor nations. Our results tend to support this discouraging hypothesis. We
don’t find any significant movement of people away from vulnerable cities in low income
countries, even though the death rate is higher. Also, the economic impact as measured
by fall in night lights is 6 times higher (8.3 percent as opposed to 1.4 percent). Further-
more, economic activity in low income nations take longer to return to pre-flood levels
post a disaster (two months versus one). Low income nations also tend to be have a large
proportion of low productivity cities, and we find that these cities are worse affected dur-
ing floods. Thus, we do find strong evidence of disproportionate effects of floods on low
income countries, and slower adaptation progress.
A key equation in global Integrated Assessment Models is the “climate damage func-
tion” (Barrage 2019; W. Nordhaus 2019). This function relates changing climate condi-
tions, often proxied for by using the world’s average temperature, to economic outcomes.
Most of these parametric models do not incorporate adaptation progress over time as they
assume a stationary climate damage function. Our paper’s third contribution, by docu-
menting that the flood damage function flattens over time, is to this emerging empirical
literature.
92
Section 2.2 introduces rising place based natural disaster risk in a spatial equilibrium
where individuals and locations can invest in adaptation strategies. Section 2.3 discusses
the data used in this paper. Section 2.4 reports our population growth regressions and
Section 2.5 looks at the death toll from flood regressions. Section 2.6 studies night light
dynamics. Section ?? reports additional adaptation hypothesis tests based on the night
light dynamics. Section 2.8 concludes.
2.2 Adaptation to Place Based Shocks
Cities differ with respect to their local amenities and physical features. In the hedonic
spatial equilibrium, more people will live in a city and its real estate rent will be higher if
the area is more productive and features better amenities (J. V. Henderson 1974; Roback
1982). If an area’s quality of life declines because of extreme weather, the afflicted area will
suffer population loss and home prices will decline (E. L. Glaeser and Gyourko 2005).
Revealed preference logic teaches us that if people choose to locate in ”harm’s way”,
there must be offsetting factors that attracted them to the location. Such individuals can
protect themselves either by avoiding risky areas or by making private and local public in-
vestments to offset the risks. Whether private self protection and defensive public goods
investments are complements or substitutes plays a key role in determining how natural
risks actually impact people and the local economy (Kousky, Luttmer, and Zeckhauser
2006). Some cross-country studies find that richer countries are at some advantage in
terms of coping with natural disasters (Matthew E Kahn 2005; Kellenberg and Ahmed
Mushfiq Mobarak 2008). Kocornik-Mina, McDermott, Michaels, and Rauch 2020 doc-
ument surprisingly little adaptation (defined as movement within cities from riskier to
safer areas) in the aftermath of floods. M ˚ ard, Di Baldassarre, and Mazzoleni 2018 find
that high protection levels (in the form of dams and levees) in flood-prone urban areas
in rich countries are effective in mitigating the damage in the aftermath of floods. On the
93
other hand, a study by Ferdous, Di Baldassarre, Brandimarte, and Wesselink 2020 finds
that flood fatalities are higher in areas in the floodplains where flood protection measures
like building levees were undertaken.
We do not observe the individual level, city level, and national expenditures on disaster
risk offsetting. Such cost data is both difficult to collect and will depend on many location
specific and micro-economic factors. Given this data challenge, we proceed with a reduced
form approach that captures the net effects of a variety of choices that people, firms and
governments have made that together determines an urban population’s disaster exposure
risk and ex-post damage realizations.
We test several place based adaptation hypotheses as we study the circumstances such
that flooding and extreme rainfall causes less economic damage. Our unit of analysis
is either the city/year or the city/year/month. The benefit of our aggregate approach
is that we can track how different places cope with shocks. Some places will be better
able to cope because they are richer. Other places may be better able to cope due to their
geography and the investments in the place’s infrastructure. Other places may have central
governments who can act faster both ex-ante and ex-post after a disaster hits to invest in
resilience strategies. We assume that the quality of government is directly related to the
nation’s degree of economic development.
2.3 Data
In this section, we first explain what is our unit of analysis and how we measure economic
activity. Next, we discuss how we measure flood events and the data sources used to
measure public investment in flood protection infrastructure
94
2.3.1 Cities
To conduct our longitudinal analysis, we must define the boundary of each city around
the world. The Global Human Settlement Urban Center Database (GHS-UCDB), created
in 2015, applies one such definition to identify urban centers and their boundaries. This
definition uses “ contiguous grid cells with a density of at least 1,500 inhabitants per km2 of
permanent land or with a built-up surface share on permanent land greater than 0.5, and has at least
50,000 inhabitants in the cluster with smoothed boundaries.” (Florczyk, Melchiorri, Corbane,
Schiavina, Maffenini, Pesaresi, et al. 2019, p.3). The urban extent of cities so identified
include the city centres and suburbs. Figure 3.4 shows the urban extents using this method
for the cities of Mumbai and Los Angeles. The GHS-UCDB database covers 13,135 cities
having 50,000 people or more in 2015 and provides data on population for the years 1975,
1990, 2000, and 2015. Further, we only include cities that had at least 10,000 people in 1990
and 2000. This gives us a sample of 9,468 cities from 175 countries belonging to high-,
middle-, and low-income groups.
5
The GHS-UCDB database uses underlying data on built-up areas and population from
the Global Human Settlement Layer (GHSL) database. The GHSL primarily uses satellite
remote sensing to identify built-up area grids and thus delineate the physical extent of
settlements (Florczyk, Melchiorri, Corbane, Schiavina, Maffenini, Pesaresi, et al. 2019).
The GHSL combines the built-up area of settlements with countries’ official census data
to produce absolute population at a grid of 1 km resolution. This is done for four points
in time: 1975, 1990, 2000, and 2015. Summing up the population over all grids within the
urban extent boundaries gives the total population of cities for each of four periods. The
city GDP in the GHS-UCDB database is calculated by summing up the total GDP value
(in PPP values expressed in US dollars in 2007) for each grid cell provided in Kummu,
Taka, and Guillaume 2018 over all the grids within the urban extent.
5
Throughout this study, we use the country income groups and regional classifications as defined by
the World Bank.
95
2.3.2 Night Lights
Our key outcome variable for measuring economic dynamics is night lights (NTL). Night
light data are collected by satellites at a uniform and disaggregated spatial scale for the
whole world. This allows for a comparison of economic activity across time and place at
a finer spatial scale and circumvents the issue of poorly measured or missing estimates of
GDP at a local level. For these reasons, a number of studies use night lights as a proxy for
economic activity (V. Henderson, Storeygard, and D. N. Weil 2011; Donaldson and Storey-
gard 2016; J. V. Henderson, Squires, Storeygard, and D. Weil 2018). Research studies make
use of two major sources of night lights data: Defence Meteorological Satellite Program
(DMSP) and Visible Infrared Imaging Radiometer Suite (VIIRS) Day-Night Band on the
Suomi satellite. The DMSP night light data has been shown to have some flaws. The
values are top coded, leading to saturation in core cities (F.-C. Hsu, K. E. Baugh, Ghosh,
Zhizhin, and Elvidge 2015), and the data do not correlate well with output in less dense
areas (Chen and W. D. Nordhaus 2011). While some of the issues were rectified in an
updated Radiance Calibrated Nighttime Light data set, the data are only available annu-
ally and only for seven years up until 2010. We use VIIRS night lights data
6
, which have
been calibrated and do not feature top coding (Elvidge, K. Baugh, Zhizhin, F. C. Hsu, and
Ghosh 2017). The data have been shown to be accurate and reliable (Gibson, Olivia, Boe-
Gibson, and Li 2021). It has a spatial resolution of 465m X 465m (grids) and provides
monthly frequency since April 2012. This measure is a proxy for urban economic activity,
population/ density, and built-up area.
7
6
The data are provided by the Earth Observation Group, Payne Institute for Public Policy, Colorado
School of Mines
7
There are two caveats to note. First, low lit areas could have negative pixel values if the areas are darker
than the background light that is subtracted from them (Beyer, Hu, and Yao 2022). Second, monthly data
are missing for high-latitude countries during summer months because the data are contaminated by solar
illumination. Data could also be affected due to heavy cloud coverage (Beyer, Hu, and Yao 2022).
96
(a) April 2016 (b) May 2016
Figure 2.1: Night Lights Before and After Floods in Wuhan: 2016
Notes: Average monthly night light intensity in Wuhan of Hubei province, China. Wuhan suffered from
major floods between May-July 2016, causing economic losses estimated to be greater than $350m. Figure
2.1a and 2.1b show the light intensity in April and May of 2016, respectively.
Figure 2.1 shows VIIRS night light for Wuhan before and during the months the city
faced floods in 2016. The light is dimmer for the month of May when it was hit with floods
(see Figure 2.1b).
In the aftermath of a flood, lights at night can dim for several reasons that include
temporary power failures, disruption of essential services, damage to property, temporary
closure of offices and factories. If a specific geographic neighborhood is evacuated in the
aftermath of a flood, this displacement effect may increase economic activity in another
part of the city where the people move to. Our city level aggregate measures will capture
this if the people remain within the geographic boundary as we have defined them.
8
8
Economic activity can be displaced by a shock to areas outside the city’s boundaries. In this case, people
adapt to the shock but the place’s lights at night metric shrinks and will not recover if this economic activity
is permanently displaced.
97
2.3.3 Flood Events
Our flood data are from The Geocoded Disasters (GDIS) dataset, which is an open-source
database that provides GIS locations and polygons for disaster-affected areas in the Emer-
gency Events Database (EM DAT). It includes the dominant geophysical, meteorologi-
cal, hydrological, and climatological disaster types: floods, storms, earthquakes, volcanic
activity, extreme temperatures, landslides, droughts, and (dry) mass movements. This
dataset includes all disasters between 1960 and 2018. Our main focus is on storms and
floods, which are the majority of total disasters during the time period of the study for
our sample cities. EM DAT classifies storms as meteorological disasters “ caused by short-
lived, micro- to meso-scale extreme weather and atmospheric conditions that last from minutes to
days” and floods as hydrological disasters “ caused by the occurrence, movement, and distribu-
tion of surface and subsurface freshwater and saltwater”.
9
Previous studies have used EM-DAT
data at the nation level to explore various issues related to the impact of natural disasters
(see for example Matthew E Kahn 2005; Cavallo, Galiani, Noy, and Pantano 2013).
The GDIS provides geolocation information at various levels of administrative divi-
sions – city level, province or state level, and country level. However, as our analysis is
at the city level, using administrative divisions that are larger than the city poses a risk of
misclassifying a flood as having taken place in a city where it did not take place. To address
this issue, we conducted an extensive Google search for newspaper articles, government
reports, maps by aid agencies, etc. to find the cities affected by the disaster amongst all
urban areas in the larger administrative division. However, when no clear information
was available from the search, we classified all cities in our database that fell within the
affected administrative division as being hit by the disaster.
9
https://www.emdat.be/classificationHydrologicals
98
2.3.4 Precipitation
Precipitation intensity data allow us to create a metric of the severity of a natural disaster
based on a consistent criteria. To measure rainfall intensity, we use data from TerraClimate,
which provides monthly climate and climatic water balance for global terrestrial surfaces
from 1958-2019. All data have a monthly temporal resolution and a c.4-km (1/24th degree)
spatial resolution.
We use the TerraClimate data to create a distribution of precipitation for each city. We
classify months that witnessed a precipitation greater that 95th or 90th percentile of the
city specific distribution as an extreme precipitation event. In the results we report below,
we document the positive correlation between flood events and extreme rainfall and we
report results where our measure of a disaster is an extreme rainfall event. This shock’s
intensity is city specific as we measure an outlier event based on the city’s past empirical
distribution of rainfall.
10
.
2.3.5 Elevation
We calculate the mean elevation for each city by averaging the elevation (in meters) for
all 30 arc second grids (which is approximately one kilometre) in a city. The data are
available in the GTOPO30 dataset.
11
2.3.6 Dams
We identify cities that were protected by dams in four major countries; China, India, Mex-
ico and the United States. Due to the time costs of accurately identifying cities protected by
dams and the requirement of having comprehensive country maps of rivers, we focus on
10
Due to the topography and hydrology of cities and due to their investments in dam flood protection,
there may not be a one to one mapping of extreme local rainfall events with local flooding. The water may
accumulate in nearby geographic areas (see Guiteras, A. Jina, and A Mushfiq Mobarak 2015)
11
https://developers.google.com/earth-engine/datasets/catalog/USGSGTOPO30. This data has been
compiled with the help of a number of organizations. The team was led by U.S. Geological Survey’s Center
for Earth Resources Observation and Science (EROS).
99
these four major nations. For the location of dams, we rely on a comprehensive geocoded
global database of 7,320 large dams.
12
Identifying cities in our sample that are downstream from a dam involved first iden-
tifying the rivers on which the dams were built and then identifying whether a city was
near a river. We used four different sources for geocoded maps of rivers for each of the
four countries.
13
We assume that the dam was built before 2013.
We define a dam as protecting a city if the following two conditions are satisfied; first,
the dam is located upstream of ”the nearby rivers” of this city. If the distance from any
point of a river to the geometric center of a city is greater than 15km, we consider this river
is not near the city; otherwise, this river is near the city. Second, the distance between the
upstream dam and the geometric center of the city is less than or equal to 100km. If a city
has one or more dam that has protection power, it is a city ”protected by dams”; otherwise,
this city is ”not protected by dams.”
12
The dams database is compiled by the Global Water System Project as part of the Global Reservoir and
Dam Database (GRanD) which is available at: https://hub.arcgis.com/datasets/panda::global-dams-and-
reservoirs/about?layer=2.
13
Mexico from ArcGIS hub(https://hub.arcgis.com/), the US from Esri ArcGIS online
(https://www.esri.com/en-us/arcgis/products/arcgis-online/overview), India from Stanford Lib Earth-
works and China from 1998 China River Location Map(downloaded in a Chinese website)
100
Table 2.1: Summary Statistics
Variable All Cities High Income Cities Low Income Cities
Coastal Inland
Low
Elevation
High
Elevation
All High
Income Cities
Coastal Inland
Low
Elevation
High
Elevation
All Low
Income Cities
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Panel A
Number of Cities
(% of Total)
9,468 777
(8.21)
4,325
(45.68)
2,859
(30.19)
2,243
(23.69)
5,102
(53.89)
437
(4.62)
3,929
(41.49)
1,947
(20.56)
2,419
(25.55)
4,366
(46.11)
Mean Elevation (m) 370.53
[514.41]
38.59
[45.00]
426.96
[564.29]
55.15
[44.30]
766.35
[610.78]
385.54
[557.99]
35.28
[47.25]
388.84
[470.89]
54.39
[46.68]
594.17
[499.20]
354.49
[462.75]
Night Lights (nW/cm2/sr) 14.23
[16.12]
27.43
[22.30]
20.21
[16.83]
20.69
[18.24]
22.08
[17.56]
21.31
[17.96]
8.84
[12.71]
5.63
[6.89]
6.49
[8.24]
5.52
[7.27]
5.95
[7.73]
GDP per capita (US$) 8,237.39
[9,156.55]
17,966.07
[14,380.46]
11,470.26
[9,249.45]
13,791.01
[11,200.41]
10,761.95
[9,163.29]
12,459.52
[10,460.82]
4,305.93
[4,311.26]
3,192.04
[2,750.52]
3,919.16
[3,326.65]
2,808.02
[2,527.03]
3,303.53
[2,962.48]
Population Growth (%) 20.51
[38.01]
19.71
[34.50]
12.86
[27.52]
13.15
[27.38]
14.86
[30.49]
13.90
[28.793]
35.26
[78.53]
27.46
[39.89]
23.76
[49.63]
31.84
[41.18]
28.24
[45.32]
Built-up Area (%) 34.68
[18.15]
45.99
[14.04]
44.41
[16.02]
46.52
[15.16]
42.28
[16.15]
44.65
[15.75]
25.20
[13.61]
22.78
[13.06]
22.52
[12.91]
23.43
[13.30]
23.03
[13.13]
Panel B
Number of Floods
(% of Total)
18,420 958
(5.20)
10,726
(58.23)
6,183
(33.57)
5,501
(29.86)
11,684
(63.43)
585
(3.18)
6,151
(33.39)
3,033
(16.46)
3,703
(20.10)
6,736
(36.57)
Avg. Floods 1.95
[3.08]
1.23
[1.85]
2.48
[3.85]
2.16
[3.24]
2.45
[4.09]
2.29
[3.64]
1.34
[1.72]
1.57
[2.24]
1.56
[2.01]
1.53
[2.33]
1.54
[2.19]
Extreme Precip. Events
(% of Total)
32,946 3,177
(9.64)
13,450
(40.82)
9,617
(29.19)
7,007
(21.27)
16,627
(50.47)
1,874
(5.69)
14,445
(43.84)
7,050
(21.40)
9,269
(28.13)
16,319
(49.53)
Avg. Extreme Precip. Events 3.48
[2.31]
4.09
[2.08]
3.11
[2.19]
3.36
[2.21]
3.12
[2.21]
3.26
[2.21]
4.29
[3.01]
3.68
[2.31]
3.62
[2.62]
3.83
[2.20]
3.74
[2.40]
Notes: Column 1 provides the summary statistics for all 9,468 cities, while columns (2)-(6) and (7)-(11) provide summary statistics for High Income and Low Income cities, respectively. High Income and Low Income cities have been further classified into coastal or inland cities
(columns (2)-(3) and (7)-(8)), and low elevation and high elevation cities (columns (4)-(5) and (9)-(10)). Columns (6) and (11) provide summary statistics for the universe of High Income and Low Income cities, respectively. The sum of coastal and inland cities equals the
total number of cities in the respective income group, as do the sum of low and high elevation cities. Panel A provides information on geographical and economic characteristics, while Panel B focuses on floods and extreme precipitation events. GDP per capita is measured
in PPP 2015 US$. Population Growth is computed as the change in the population of a city between years 2015 and 2000. Built-up Area is defined as the percentage of the total area of the city (km
2
) that contains built-up structures. Polity is a continuous variable computed by
subtracting the Autocracy score from the Democracy score. The resulting unified polity scale ranges from +10 (strongly democratic) to -10 (strongly autocratic).Both the Autocracy and the Democracy scores are an additive eleven-point scale (0-10). The operational indicators of
Autocracy and Democracy are derived from codings of the competitiveness of political participation, the openness and competitiveness of executive recruitment, and constraints on the chief executive. Since Polity scores are only available at a country level, we cannot calculate
them separately based on the geography of the city. Extreme Precipitation Events are a a dummy indicating whether the precipitation in the monthm and yeary in cityc in countryj (between the years 2012-2018) was greater than the 95
th
percentile of the city-specific
distribution of precipitation, which was created using data from 1958-2018. The Average Floods and Average Extreme Precipitation Events refer to the average number of such events during the years 2012-2018. Standard deviation for all variables are reported in brackets.
101
Table 2.1 reports the summary statistics. The upper panel reports the count of cities in
our data set overall and divided into high income and low income nations. We also report
the count of cities in different geographic categories including coastal cities and inland
cities and cities at high elevation. The bottom panel of Table 2.1 reports the flood disaster
conditional means. In our sample of cities, over 18,000 flood events took place and roughly
63 percent took place in richer nations. The average city experienced two floods during
our sample period. Below we will discuss how this fact affects our econometric estimation
strategy.
2.4 Adaptation by Migrating Away from Flood Prone Cities
We test whether cities that have experienced more floods in recent years experience lower
population growth.
We report population growth regressions using data from 2000-2015. Our regression
specification takes the following form:
∆ Population
cj
=α +β
1
Risky
cj
+β
2
ln(GDP/capita)
cj
+β
3
ln(Builtup Area)
cj
+
β
4
ln(Population)
cj
+δ
c
+ξ
cj
(2.1)
where, the dependent variable ∆ Population
cj
is the population growth (in percent) be-
tween years 2000 and 2015 in cityc in countryj. The variable of interest is Risky
cj
, which
is a proxy for the vulnerability of the city. It is a continuous variable that measures the
number of extreme precipitation events between 2000-15, where we define an extreme
event based on two different cutoffs. For each city, we construct a distribution of monthly
rainfall between 1958-2015, and then sum the total number of 90
th
(or 95
th
) percentile
events that struck the city between 2000-15. Thus, a city that was frequently hit with ex-
treme events, relative to its own distribution, over the course of the first 15 years of the
new millennium will have a higher value for the Risky
cj
variable. We control for baseline
102
economic opportunity measured by the the natural log of the year 2000 values of GDP
per capita, built-up area (per square km.), and population of the city, and also include
country fixed effects ( δ
c
) to account for time-invariant country characteristics. Table 2.2
presents the results, first at the aggregate level, and then divided by income groups.
Table 2.2: Effect of Extreme Events on Population Growth
Dependent Variable: ∆ Population
cj
All High Income Low Income
(1) (2) (3) (4) (5) (6)
Risky (90
th
Perc)
cj
−0.004
∗
−0.003 −0.005
(0.002) (0.002) (0.004)
Risky (95
th
Perc)
cj
−0.005
∗
−0.004
∗
−0.006
(0.003) (0.002) (0.005)
GDP/capita (2000)
cj
0.033
∗∗
0.033
∗∗
0.061
∗∗∗
0.061
∗∗∗
0.015 0.016
(0.014) (0.014) (0.013) (0.013) (0.016) (0.016)
Builtup Area (2000)
cj
−0.038
∗
−0.039
∗∗
−0.018 −0.020 −0.050
∗∗
−0.049
∗∗
(0.020) (0.019) (0.023) (0.022) (0.024) (0.023)
Population (2000)
cj
−0.020
∗∗
−0.020
∗∗
−0.024
∗∗∗
−0.024
∗∗∗
−0.018 −0.017
(0.009) (0.009) (0.009) (0.009) (0.017) (0.017)
Fixed Effects
Country ! ! ! ! ! !
Num. obs. 9,468 9,468 5,102 5,102 4,366 4,366
Adj. R
2
0.261 0.261 0.328 0.329 0.193 0.193
Notes: clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
The dependent variable in all regressions, ∆ Population
cj
, refers to the population growth between years 2000 and 2015 in cityc in countryj. Risky
cj
is a
continuous variable that measures the number of extreme precipitation events in cityc in countryj between years 2000 and 2015, where extreme precipitation
is a dummy indicating whether the precipitation in the monthm and yeary in cityc in countryj fell in the 90
th
(or 95
th
) percentile of the distribution of
rainfall in the said city. The distribution was created using precipitation data from 1958-2015. Log values of GDP per capita, Builtup Area (per sq. km), and
Population pertain to the year 2000. Models (3) and (4) only include observations from High Income and Upper Middle Income countries, whereas models (5)
and (6) only include observations from Low Income and Lower Middle Income countries. All standard errors are clustered by country.
We find that cities that experienced higher frequency of extreme events between 2000-
15 saw lesser growth in population in the same period (columns 1-2 in Table 2.2). This
is only significant for cities in high-income countries (column 4) and not for cities in
low-income countries (columns 5-6). The effect is quantitatively small as the percentage
change in population in high flood risk cities is .5 percent lower than cities facing less flood
risk.
Recent empirical research studying localized shocks such as bombings generally does
not find strong evidence of population decline in shocked areas. Studies set in Japan
103
and Vietnam have documented that cities bombed during war time have proved to be
resilient in the face of extreme shocks in the long run (Davis and Weinstein 2002; Miguel
and Roland 2011).
Unlike bombed areas, natural disasters prone areas are likely to be repeatedly shocked.
Such a time series persistence should act as a tax on investing in capital in the affected area
because investors expect that the area will be struck again. Given this logic it is surprising
that recent empirical work on disasters has found that incumbents tend to remain in the
shocked area unless their housing is destroyed (Kocornik-Mina, McDermott, Michaels,
and Rauch 2020; Boustan, Matthew E Kahn, and Rhode 2012). This could be explained by
positive migration costs and built up local social capital. Existing residents, especially the
poor, may find it difficult to finance the migration costs to move to a safer city. Cattaneo
and Peri 2016 and Peri and Sasahara 2019 find that in low income countries, such costs
inhibit migration out of rural areas when increasing temperatures cause decline in agri-
cultural productivity. Shocks can trigger huge federal transfer payments in richer nations.
The expectation of such ex-post relief can create a moral hazard effect that acts to anchor
people to risky places.
Recent empirical studies (Deryugina, Kawano, and Levitt 2018; Nakamura, Sigurds-
son, and Steinsson 2016) have documented a ”silver lining” such that people displaced
from their origin due to a natural disaster actually enjoy an improvement in their material
standard of living in subsequent years. Access to family and social capital may anchor
people to risky places (E. L. Glaeser, Laibson, and Sacerdote 2002). If the poor are less
likely to ”vote with their feet” to move to higher ground, then environmental justice is-
sues are exacerbated as climate change will cause the poor to be exposed to greater risks
but their rents for living in such places will be lower.
104
2.5 The Death Toll from Floods
One important natural disaster adaptation metric is the death count. During our sample
period, this is a highly skewed variable with many zero counts and some truly deadly
disasters.
To test whether floods cause less deaths in richer cities, or in cities with more past
experience with flood events, we run the following regression:
ln(Deaths per disaster
cjr
) =α +β
1
Risky
cjr
+β
2
ln(GDP/capita)
cjr
+β
3
ln(Population)
cjr
+
X
cjr
+δ
r
+ξ
cjr
(2.2)
Here, the dependent variable is the natural log of the ratio of total deaths caused by floods
and the total number of floods between the years 2010 and 2018, for each city c in country
j and regionr. Cities that suffered no disasters between 2010-18 were dropped from the
analysis
14
. The independent variable Risky
cjr
reflects vulnerability, but we slightly change
its definition as compared to Equation 2.1. Firstly, the distribution of monthly rainfall is
constructed using data between 1958-2018, and secondly, we sum the number of extreme
precipitation events between 1970-2010. Thus, cities that have been most affected by nat-
ural disasters over the course of the 40 years between 1970-2010 will have a higher value
for Risky
cjr
. We control for city characteristics as measured by the natural log of the year
2015 values of GDP per capita and population.
We add in a vector of city-specific time-invariant dummies ( X
cjr
) for geography and
topography. The first, High Elevation
cjr
, indicates whether cityc in countryj has an ele-
vation
15
that falls in the top 50
th
percentile of the distribution of elevations across all the
9,468 cities. The second, Coastal
cj
, signifies whether the city has a coast line. We also con-
trol for Capital
cj
, which is a dummy that indicates whether cityc is the capital of country
14
For cities with disasters but no deaths, the dependent variable islog

1+deaths
total disasters
!
.
15
Calculated as median elevation across all 30 arc second grids in a city.
105
j. Finally, we add in World Bank region
16
fixed effects ( δ
r
) to control for any time invari-
ant heterogeneity between cities in different regions. Standard errors are clustered at the
World Bank region level. Results are presented in Table 2.3.
17
Table 2.3: The Death Toll from Floods
Dependent Variable: ln(Deaths per disaster
cjr
)
All High Income Low Income
(1) (2) (3) (4) (5) (6)
Risky (90th Perc)
cjr
0.008 −0.008
∗∗
0.017
∗∗
(0.006) (0.003) (0.005)
Risky (95th Perc)
cjr
0.011 −0.011
∗∗
0.027
∗∗
(0.009) (0.003) (0.008)
GDP/capita (2015)
cjr
−0.097
∗∗∗
−0.095
∗∗∗
−0.042
∗
−0.043
∗
−0.036
∗∗
−0.035
∗
(0.014) (0.015) (0.018) (0.019) (0.014) (0.014)
Population (2015)
cjr
0.074
∗∗∗
0.074
∗∗∗
0.055
∗∗∗
0.055
∗∗∗
0.094
∗∗∗
0.093
∗∗∗
(0.014) (0.013) (0.015) (0.015) (0.006) (0.007)
High Elev
cjr
−0.075 −0.073 0.050 0.052 −0.215
∗
−0.212
(0.102) (0.102) (0.038) (0.036) (0.101) (0.105)
Capital
cjr
0.424 0.415 0.171 0.188 0.703
∗
0.719
∗
(0.279) (0.281) (0.166) (0.158) (0.303) (0.288)
Coastal
cjr
0.209
∗∗
0.210
∗∗
0.024 0.026 0.418
∗
0.425
∗
(0.061) (0.063) (0.045) (0.044) (0.171) (0.175)
Fixed Effects
WB Region ! ! ! ! ! !
Num. obs. 7,032 7,032 3,874 3,874 3,158 3,158
Adj. R
2
0.131 0.131 0.137 0.137 0.113 0.114
Notes: clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
The dependent variable in all regressions, ln(deaths per disaster)
cjr
, refers to the natural log of deaths per disaster for each cityc in countryj and regionr between
the years 2010 and 2018. Risky
cj
is a continuous variable that measures the number of extreme precipitation events in cityc in countryj between years 1970
and 2015, where extreme precipitation is a dummy indicating whether the precipitation in the monthm and yeary in cityc in countryj fell in the 90
th
(or 95
th
)
percentile of the distribution of rainfall in the said city. The distribution was created using precipitation data from 1958-2018. Log values of GDP/capita and
Population pertain to the year 2015. High Elev
cj
is a dummy indicating whether cityc in countryj has a median elevation that ranks in the top 50
th
percentile
of the distribution of the median elevation across all cities. Coastal
cj
and Capital
cj
are dummies that indicate whether cityc in countryj is a coastal or capital
city, respectively. Models (3) and (4) only include observations from High Income and Upper Middle Income countries, whereas models (5) and (6) only include
observations from Low Income and Lower Middle Income countries. All standard errors are clustered by World Bank Region.
16
There are 7 World Bank regions, namely East Asia & Pacific, Europe & Central Asia, Latin America &
Caribbean, Middle East & North Africa, North America, South Asia, and Sub-Saharan Africa.
17
One caveat related to the death toll results is the measurement error in the dependent variable. EM
DAT provides deaths and injuries for each disaster, and not the affected cities therein. This is not a concern
if the administrative unit affected by the disaster comprises a single city. However, when the death toll is
provided for an administrative unit larger than the city, we have made the assumption that total deaths were
divided equally between all the cities in our sample that are located within the disaster zone. This implies a
certain degree of measurement error in the regressand, as there is bound to be variation in deaths between
cities in the same region. However, it is important to note that this only makes the estimates less precise, but
the coefficients would still remain unbiased.
106
We find that cities in high-income countries having higher frequency of extreme events
between 1970-2010 saw fewer deaths per disaster between 2010-2018 (columns 3-4 in Ta-
ble 2.3). Conversely, cities in low-income countries that had a higher number of extreme
events in the past saw higher deaths per disaster (columns 5-6). Cities in high-income
countries have been able adapt to recurring shocks of extreme flooding in the past. This
could be due to people moving to safer areas in high-income countries as seen in Table
2.2. India, which is classified as a lower middle-income country, has far fewer large cities
than what the Zipf’s law would suggest (Chauvin, E. Glaeser, Ma, and Tobio 2017).
We find that GDP/capita
cjr
has a negative relationship to deaths per disaster in both
high- and low-income countries. Comparing the coefficient values in columns (3) and
(6), we find that the negative impact of GDP/capita
cjr
on deaths per disaster is marginally
greater in high-income countries than in low-income countries. This result supports the
hypothesis that the relationship between income and damage caused by flooding is non-
linear and depends on the stage of development (see Kellenberg and Ahmed Mushfiq
Mobarak 2008).
We include a time trend to capture overall trends in adaptation (see Table 3.18). Many
Integrated Assessment Models (IAM) assume that the climate damage function is station-
ary over time (Pindyck 2013). We reject this pessimistic hypothesis as we find that richer
cities suffer less death and the death gradient with respect to flood events flattens over
time.
18
18
We also test if deaths from disasters are declining over time. We run a regression of the log of deaths per
disaster on a linear and quadratic function of time, with city and country-month fixed effects. The sample
only includes city-month combinations in which there was a flood. Clustering the standard errors at the
city level, we find a negative and significant coefficient on the time trend. Running the same regression
separately for cities in high income and low income countries yields a negative and significant coefficient
for the former, but a positive and insignificant coefficient for the latter.
107
2.6 Urban Resilience and Flood Shocks
We now turn to presenting our findings on how flood shocks affect economic activity as
based on lights at night dynamics. Floods could lead to temporary power failures, disrup-
tion of essential services, damage to property, temporary closure of offices and factories,
and in all cases will affect the normal functioning of economic life for sometime. The
flood’s impact on economic activity depends on the resilience of the city’s core infrastruc-
ture. Richer cities should have more resilient infrastructure. We now test whether cities
in high income countries suffer less from floods than cities in low income countries and
we test for whether cities in low income countries take longer to bounce back after flood
events.
2.6.1 The Effect of Floods on Economic Activity
We examine the impact of floods on economic outcomes by running the following regres-
sion:
ln(Night Lights
cjmy
) =α +
2
X
i=−2

β
i
Flood
cj{m+i}y

+γ
1
Flood
cjmy
× High Elev
cj
+
γ
2
Flood
cjmy
× Coastal
cj
+γ
3
Flood
cjmy
× Capital
cj
+X
cjmy
+δ
c
+ Γ
jmy
+ϵ
cjmy
(2.3)
where ln(Night Lights)
cjmy
refers to the natural log of the average value of night lights in
cityc in countryj in monthm and yeary. This is calculated as the mean of the radiance
values of all grids within the city in a given month and year. Flood
cjmy
is a dummy vari-
able that equals 1 if cityc in countryj was hit by a flood in month m of yeary. We include
two month lags and leads of the flood dummy to account for (i) early flood warning or
heavy rainfall prior to floods affecting economic outcomes, and (ii) recovery dynamics
post floods. Similar to the death toll regressions, we add in a dummy for capital cities,
and use High Elev
cj
and Coastal
cj
as controls for city-specific time-invariant dummies for
108
(a) High Income Cities: Floods (b) High Income Cities: Extreme Precip.
(c) Low Income Cities: Floods (d) Low Income Cities: Extreme Precip.
Figure 2.2: Pre-Trend Analysis of Night Lights Intensity for high- & low-income countries
Notes: Figure 3.3a and 3.3e show the 6 month lag and 2 month lead around a flood event for cities in high- and low-income countries.
The coefficients in the two plots are estimated by running a regression of ln(Night Lights)
cjmy
, which is the natural log of mean light
intensity in cityc in countryj in monthm of yeary, on the contemporaneous and 8 month leads and lags of the flood dummy, where
Flood
cjmy
is a dummy indicating whether cityc in countryj was hit by a flood in month m of yeary. All three models include
the controls Storm
cjmy
and Landslide
cjmy
, dummies indicating whether city c in country j was hit by a storm or landslide,
respectively, in monthm of yeary. 8 month leads and lags for these two disaster types have also been included as controls. Figure
3.3b and 3.3f show the 6 month lag and 2 month lead around an extreme precipitation event for cities in high- and low-income
countries, which is a dummy indicating whether the precipitation in the monthm and yeary in cityc in countryj was greater
than the 95
th
percentile of the city-specific distribution of precipitation, which was created using data from 1958-2018. The shaded
ribbons in each plot represent the 95
th
confidence interval band. In all regressions, observations include city-country-month-year
observations which had a non-zero value of nightlights. Each observation was weighted by the mean of the cloud free coverage for
the city-country-month-year observation. Standard errors are clustered at the city and month-year level.
109
geography and topography. X
cjmy
represents a vector of city-specific controls, specifi-
cally whether the city was affected by storms or landslides in month m of yeary, and the
corresponding two month lags and leads. δ
c
and Γ
jmy
represent city, and country-month-
year fixed effects, respectively. Finally, to account for possible noise in the measurement
of night lights intensity in the city due to cloud cover, we weight each observation by the
proportion of cloud free cover images used to create a monthly composite for the city. To
account for spatial and temporal correlation, we cluster the standard errors at the city and
month-year level. The results are presented in columns (1) and (2) of Table 2.4.
We find that on average cities that suffer from floods see a decline in mean night lights
by around 3%.
19
This decline is associated with disruptions to economic activity caused
due to floods. Counter-intuitively, the effect is attenuated for cities located on coasts com-
pared to cities located away from the coast. Similarly, cities in low elevation areas on av-
erage see a smaller decline in mean night lights.
We posit that richer cities suffer less from floods than poorer cities. To test this, we
ran Equation 2.3 separately for high income and low income countries. The results are
presented in columns (3)-(6). We find cities in both high and low income countries see,
on average, a decline in night lights after being hit by a flood. Further, the coefficients for
cities in high income countries are much lower compared to the coefficients for cities in
low income countries, which are around 0.05-0.08 (columns 5-6 in Table 2.4).
19
Our empirical methodology involves using two-way fixed effects (TWFE) regression specification. If
the treatment effects are heterogeneous across time or units, the coefficients from a standard TWFE model
may not be robust due to a negative weighting problem (see Roth, Sant’Anna, Bilinski, and Poe 2022, for a
review of this literature). This is a valid concern for our setting as we have 9,468 groups (each group is a city)
and 81 periods (each month, from April 2012 to December 2018, is a period). As long as the effect of a flood
varies across cities and/or changes over time, the standard common trends assumption may be violated,
which makes it plausible that
ˆ
β
fe
may not be robust to heterogeneous effects. To address this, we estimate
the number of negative weights attached to the two-way fixed effects regressions for each specification using
the TwoWayFEWeights package in R. Second, we estimate the degree of heterogeneity in treatment effects
that would be necessary for the estimated treatment effect to have the wrong sign. Specifically, as shown
by De Chaisemartin and d’Haultfoeuille 2020, the ratio of the absolute value of the expectation of
ˆ
β
fe
and
the standard deviation of the weights corresponds to the minimal value of the standard deviation of the
treatment effect across the treated groups and time periods under which beta and the average treatment
effect on the treated (ATT) could be of opposite signs. We find that the degree of heterogeneity in treatment
effects is not substantially large enough for the beta and the ATT to have opposite signs.
110
Table 2.4: Effect of Floods on Economic Activity
Dependent Variable: ln(Night Lights
cjmy
)
All High Income Low Income
(1) (2) (3) (4) (5) (6)
Flood
cjmy
−0.034
∗∗∗
−0.070
∗∗∗
−0.014
∗∗
−0.020 −0.083
∗∗∗
−0.162
∗∗∗
(0.009) (0.017) (0.007) (0.027) (0.021) (0.041)
Flood
cjmy
× High Elev
cj
−0.019
∗∗∗
−0.002 −0.049
∗∗∗
(0.007) (0.011) (0.014)
Flood
cjmy
× Coastal
cj
0.029
∗∗∗
0.027
∗∗
0.032
(0.009) (0.011) (0.023)
Flood
cjmy
× Capital
cj
0.005 −0.006 0.049
(0.028) (0.017) (0.076)
Flood
cjmy
× Time Trend
my
0.002
∗
0.000 0.003
∗
(0.001) (0.001) (0.002)
Fixed Effects
City ! ! ! ! ! !
Country× Month× Year ! ! ! ! ! !
Num. obs. 663,161 663,084 341,899 341,822 321,262 321,262
Adj. R
2
0.952 0.952 0.935 0.935 0.930 0.930
Notes: two-way clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
The dependent variable in all regressions,ln(NightLights)cjmy , is the natural log of mean light intensity in cityc in countryj in monthm of yeary. Floodcjmy is a dummy indicating
whether cityc in countryj was hit by a flood in month m of yeary. High Elev
cj
is a dummy indicating whether cityc in countryj has a median elevation that ranks in the top 50
th
percentile of the distribution of the median elevation across all cities.Coastalcj andCapitalcj are dummies that indicate whether cityc in countryj is a coastal or capital city, respectively.
Time Trendmy is a continuous variable that takes the value from 1 to 81, with 1 representing the first month of the sample (April 2012), and 81 representing the last month (September
2018. Models (3) and (4) only include observations from High Income and Upper Middle Income countries, whereas models (5) and (6) only include observations from Low Income and Lower
Middle Income countries. All regressions include the controlsStormcjmy andLandslidecjmy , dummies indicating whether cityc in countryj was hit by a storm or landslide, respectively,
in monthm of yeary. One month lead and two month lags for all three disaster types have also been included in controls. Observations include all city-country-month-year observations
which had a non-zero value of nightlights. Each observation was weighted by the mean of the cloud free coverage for the city-country-month-year observation. Standard errors are clustered
at the city and month-year level.
We plot the estimated coefficients of the effect of floods on night lights for all cities six
months prior and two months after a flood event in Figure ??. We observe no pre-existing
trend prior to the flood event and a negative effect in the month of the flood followed by
an upward trend in the months after the flood event. Figures for high income and low
income countries are available in appendix Figures 3.3a and 3.3e respectively.
2.6.2 The Effect of Extreme Rain on Economic Activity
In this section, we change the specification and now examine the impact of extreme rainfall
events on lights at night dynamics. To construct the variable, we first calculate the city
specific distribution of rainfall for each of the 9,468 cities in our sample, using monthly
precipitation data from TerraClimate for the time period 1958-2018. Next, for each city, we
classify all months with precipitation greater than the 95
th
percentile of the city specific
distribution as having experienced an extreme event. This gives us a list of months for
111
each city when it experienced precipitation that was extreme relative to its recent 60 year
history.
The regression specification is the same as in Equation 2.3, except that the regressor
is Extreme Rain
cjmy
, and we don’t include the vector of controls, X
cjmy
. Results from this
specification are presented in Table 2.5, with the estimates broken down by high income
countries (columns 3 and 4) and low income countries (columns 5 and 6).
Table 2.5: Effect of Extreme Rain on Economic Activity
Dependent Variable: ln(Night Lights
cjmy
)
All High Income Low Income
(1) (2) (3) (4) (5) (6)
Extreme Rain
cjmy
−0.051
∗∗∗
−0.081
∗∗∗
−0.046
∗∗∗
−0.037
∗∗∗
−0.059
∗∗∗
−0.144
∗∗∗
(0.010) (0.008) (0.009) (0.009) (0.018) (0.016)
Extreme Rain
cjmy
× High Elev
cj
−0.022
∗∗∗
−0.012
∗∗
−0.033
∗∗∗
(0.005) (0.005) (0.009)
Extreme Rain
cjmy
× Coastal
cj
0.013
∗∗
0.022
∗∗∗
−0.016
(0.006) (0.006) (0.014)
Extreme Rain
cjmy
× Capital
cj
−0.039 −0.019 −0.096
(0.026) (0.020) (0.075)
Extreme Rain
cjmy
× Time Trend
my
0.003
∗∗∗
0.001 0.005
∗∗∗
(0.000) (0.000) (0.001)
Fixed Effects
City ! ! ! ! ! !
Country× Month× Year ! ! ! ! ! !
Num. obs. 663,161 663,084 341,899 341,822 321,262 321,262
Adj. R
2
0.952 0.952 0.935 0.935 0.930 0.930
Notes: two-way clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
The dependent variable in all regressions,ln(Night Lights)cjmy , is the natural log of mean light intensity in cityc in countryj in monthm of yeary.ExtremeRaincjmy is a dummy indicating whether
the precipitation in the monthm and yeary in cityc in countryj was greater than the 95th percentile of the city-specific distribution of precipitation, which was created using data from 1958-2018.
High Elev
cj
is a dummy indicating whether cityc in countryj has a median elevation that ranks in the top 50th percentile of the distribution of the median elevation across all cities. Coastalcj and
Capitalcj are dummies that indicate whether cityc in countryj is a coastal or capital city, respectively. Time Trendmy is a continuous variable that takes the value from 1 to 81, with 1 representing
the first month of the sample (April 2012), and 81 representing the last month (September 2018) Models (3) and (4) only include observations from High Income and Upper Middle Income countries,
whereas models (5) and (6) only include observations from Low Income and Lower Middle Income countries. One month lead and two month lags for the Extreme Rain dummy have been included as
controls. Observations include all city-country-month-year observations which had a non-zero value of nightlights. Each observation was weighted by the mean of the cloud free coverage for the
city-country-month-year observation. Standard errors are clustered at the city and month-year level.
The coefficient for all cities are negative and significant and, at 4%. Extreme rain affect
cities in both rich and poor nations. For cities in high income countries, extreme rain is
associated with a 5% decline on mean night lights — higher than the average impact of
flood events. Given that our variable for precipitation measures extreme events, the results
show that high income countries are also vulnerable to such disasters.
We plot the coefficient estimates of the effect of extreme rain on night lights for all
cities six months prior and two months after a flood event in Figure ??. We observe no pre-
existing trend prior to the flood event and a negative coefficient in the month of the extreme
112
precipitation event followed by an upward trend in the months after the flood event. Pre-
trends figures for high income and low income countries are available in appendix Figure
3.3b and 3.3f, respectively.
2.6.3 Recovery Dynamics From Floods
In the aftermath of a flood, how many months does it take for the city to recover? In
Figure 2.3a, we report one case study. Chennai, a major city in India, suffered from major
floods between November-December 2015. Figure 2.3a shows the light intensity in October
before the flood events. This intensity reduced during the the months of flooding (Figure
2.3b & Figure 2.3c). By January 2016, the intensity of lights in Chennai started recovering
to pre-flood levels (see Figure 2.3d).
In this section, we analyse the pattern of recovery following a flood event, i.e. the
length of time it takes cities to recover from floods on average. To test for this, we estimate
the following regression:
ln(Night Lights
cjmy
) =α +
3
X
i=−6

β
i
Flood
cj{m+i}y

+X
cjmy
+δ
c
+ Γ
jmy
+ϵ
cjmy
(2.4)
where Flood
cj{m+i}y
is a dummy that equals 1 if there was a flood in month {m+i}. The co-
efficient on this variable can, thus, be interpreted as the effect of a flood in month {m+i} on
night lights in monthm. We include lags for six months, allowing us to trace the economic
recovery immediately after the flood event. X
cjmy
represents a vector of city-specific con-
trols, specifically whether the city was affected by storms or landslides in month m of year
y, and the corresponding lags and leads.
The results are presented in Panel A of Table 2.6. As a robustness check, we also test
for recovery dynamics after extreme rain events, with results presented in Panel B. In the
interest of brevity, we have only presented the estimates on lags of floods and extreme
113
(a) October 2015 (b) November 2015
(c) December 2015 (d) January 2016
Figure 2.3: Night Lights Before and After Floods in Chennai: 2015-16
Notes: Average monthly night light intensity in Chennai, capital of Tamil Nadu state, India. Chennai
suffered from major floods between November-December 2015, with economic losses estimated to be
US$1bn. Figure 2.3a and Figure 2.3b show the light intensity in October and November of 2015, respec-
tively, whereas Figure 2.3c and Figure 2.3d show the light intensity in December 2015 and January 2016,
respectively.
rain for three months. However, the coefficients on leads and the remaining lags for all
the specifications are insignificant.
We see that negative impact of floods persist for at least one month after the disaster
and is more severe for low-income countries relative to high-income countries.
114
Table 2.6: Recovery Dynamics for Floods and Extreme Rain
Dependent Variable: ln(Night Lights
cjmy
)
All High Income Low Income
Panel A
Flood
cj{m}y
−0.033
∗∗∗
−0.016
∗∗
−0.079
∗∗∗
(0.009) (0.007) (0.021)
Flood
cj{m−1}y
−0.017 −0.002 −0.049
∗
(0.011) (0.009) (0.025)
Flood
cj{m−2}y
0.010 0.007 0.016
(0.011) (0.009) (0.028)
Flood
cj{m−3}y
0.020 0.013
∗
0.029
(0.013) (0.008) (0.030)
Panel B
Extreme Rain
cjmy
−0.049
∗∗∗
−0.047
∗∗∗
−0.052
∗∗∗
(0.011) (0.009) (0.018)
Extreme Rain
cj{m−1}y
−0.022
∗∗∗
−0.014
∗∗
−0.034
∗∗
(0.007) (0.006) (0.014)
Extreme Rain
cj{m−2}y
0.003 0.001 0.005
(0.008) (0.007) (0.015)
Extreme Rain
cj{m−3}y
0.011 0.015
∗∗
0.007
(0.009) (0.006) (0.017)
Fixed Effects
City ! ! !
Country× Month× Year ! ! !
Num. obs. 606,353 311,287 295,066
Adj. R
2
0.953 0.938 0.932
Notes: two-way clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
The dependent variable in all regressions,ln(Night Lights)
cjmy
, is the natural log of mean light in-
tensity in cityc in countryj in monthm of yeary. Flood
cj{m−t}y
(Extreme Rain
cj{m−t}y
) indi-
cates whether cityc in countryj was hit by a flood (extreme precipitation) event t months prior.
ExtremeRain
cjmy
is a dummy indicating whether the precipitation in the monthm and yeary in
cityc in countryj was greater than the 95
th
percentile of the city-specific distribution of precipitation,
which was created using data from 1958-2018. Observations include all city-country-month-year ob-
servations which had a non-zero value of nightlights. Panel A presents the recovery dynamics for
floods, and Panel B focuses on extreme precipitation. All regressions in Panel A include the controls
Storm
cjmy
andLandslide
cjmy
, dummies indicating whether cityc in countryj was hit by a storm
or landslide, respectively, in monthm of yeary. The controls in Panel B include 7 lags and 3 leads
for each of the three disaster types. Controls in Panel B include 7 lags and 3 leads for extreme pre-
cipitation events. Each observation was weighted by the mean of the cloud free coverage for the
city-country-month-year observation. Standard errors are clustered at the city and month-year level.
115
2.7 Testing Flood Adaptation Hypotheses
We have found that economic activity in cities declines in the immediate aftermath of
floods and extreme rain and that poorer nations suffer more. In this section, we test sev-
eral resilience hypotheses. We ask: Are richer cities with productive capital resilient to
shocks? Does past experience with extreme events result in better preparedness for fu-
ture disasters? Do investments in protective infrastructure such as flood control dams
reduce the effect of floods?
2.7.1 Do Richer Cities Suffer Less?
Cities that are economically productive have better infrastructure and resources to cope
with extreme events. We hypothesize that richer cities, as measured by their per capita
GDP, will be less affected by disasters. We test this using the following regression equation:
ln(Night Lights
cjmy
) =α+
2
X
i=−2
β
i
Flood
cj{m+i}y
+
X
k∈{mid,high}
γ
k
Flood
cjmy
×GDP/capita
kcj
+X
cjmy
+δ
c
+ Γ
jmy
+ϵ
cjmy
(2.5)
We divide the log of per capita GDP into three quantiles, namely Low (omitted category),
Medium and High. Thus, GDP/capita
kcj
is a factor variable that represents whether the log
of per capita GDP of cityc in countryj in 2015 fell into thek
th
quantile. The coefficient
of interest in Equation 2.5 isγ
k
, which represents the mitigating effect of being in the k
th
income group on the impact of floods. The result is presented in column (1) of Table 2.7.
Cities in middle income and high income categories see a much lower effect of floods on
night light intensity compared to cities that have low incomes.
116
Table 2.7: Heterogeneous Effect of Floods based on Wealth and Risk
Dependent Variable: ln(Night Lights
cjmy
)
All
(1) (2)
Flood
cjmy
−0.093
∗∗∗
−0.042
∗∗∗
(0.017) (0.010)
Flood
cjmy
× Medium GDP/capita
cj
0.070
∗∗∗
(0.015)
Flood
cjmy
× High GDP/capita
cj
0.080
∗∗∗
(0.020)
Flood
cjmy
× High Risk
cj
0.020
∗
(0.010)
Fixed Effects
City ! !
Country× Month× Year ! !
Num. obs. 663,161 663,161
Adj. R
2
0.952 0.952
Notes: two-way clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
The dependent variable in all regressions,ln(Night Lights)
cjmy
, is the natural log of mean light intensity
in cityc in countryj in monthm of yeary.Flood
cjmy
is a dummy indicating whether cityc in country
j was hit by a flood in month m of yeary. ln(GDP/capita)
cj
, measured in PPP US$ (2007), pertains
to the year 2015. To create factor variables,ln(GDP/capita)
cj
was divided into three quantiles of equal
size, with the omitted category in regressions being quantile 1, i.e. Low GDP/capita
cj
. High Risk
cj
is
a dummy variable that equals 1 if the number of extreme precipitation events in city c in country j,
between years 1970 and 2011, were greater than the median number of extreme events across all 9,468
cities. Extreme precipitation is a dummy indicating whether the precipitation in the monthm and yeary
in cityc in countryj fell in the 90
th
percentile of the distribution of rainfall in the said city. The distribution
was created using precipitation data from 1958-2018. All regressions include the controlsStorm
cjmy
andLandslide
cjmy
, dummies indicating whether cityc in countryj was hit by a storm or landslide,
respectively, in monthm of yeary. Two month leads and lags for all three disaster types have also been
included as controls. Observations include all city-country-month-year observations which had a non-
zero value of nightlights. Each observation was weighted by the mean of the cloud free coverage for the
city-country-month-year observation. Standard errors are clustered at city and month-year level.
2.7.2 Does Repeated Experience with Flooding Reduce the Marginal Ef-
fect of the Next Flood?
In the previous sections, we focus on the occurrence of a flood event without considering
whether the city had previous experience of such flood events. The impact of a disaster
could also depend on whether it was an unexpected event or something that people were
used to dealing with in the past (Guiteras, A. Jina, and A Mushfiq Mobarak 2015).
Cities that are faced with recurring weather shocks may be better able to cope with
future shocks. Public authorities and citizens with past experience with extreme weather
117
events know what to expect and can plan better for future events. An alternative hypoth-
esis is that places that face repeat flooding experience disinvestment as the repeat events
act as a tax on capital investment.
To test whether cities facing recurring extreme disasters in the past have become re-
silient, we estimate the following regression equation:
ln(Night Lights
cjmy
) =α +
2
X
i=−2
β
i
Flood
cj{m+i}y
+β
2
Flood
cjmy
× High Risk
cj
+X
cjmy
+δ
c
+ Γ
jmy
+ϵ
cjmy
(2.6)
where High Risk
cj
is a dummy variable that equals 1 if the number of extreme precipita-
tion events in city c in country j between 1970 to 2012 were above the median number
of extreme precipitation events across all cities in the sample. To determine the number
of extreme precipitation events that struck a city, we construct a distribution of monthly
rainfall between 1958-2015 for each city, and then sum the total number of 90
th
percentile
events between 1970-2012. The assumption here is that the extreme precipitation events
led to flooding, and therefore, a higher number of events between 1970-2012 would indi-
cate recurrent shocks which we hypothesize should enable a city to adapt.
The result is presented in column (2) in Table 2.7. Cities that had flood events in the
past see a lower impact of floods relative to cities that had no such prior experience.
2.7.3 Does Flood Protection Infrastructure Protect Cities?
Using data on dams for four countries – China, India, Mexico, and the United States –
we examine whether cities protected by dams are more likely to experience flood events,
higher population growth, and importantly, if dams help to mitigate the impact of a flood
shock. The four countries include 3,820 cities out of our total sample of 9,468 cities.
First, we analyse the heterogeneity in the number of dams between different countries.
China and US, which fall in the high income category, have 71 percent and 67 percent of their
118
river cities protected by a dam.
20
Surprisingly, India, which is classified as a low income
country, has the exact same percentage of cities with dams as the US. Only for Mexico
does this number drop slightly to 58 percent. In terms of new dams that were built during
the time period of our analysis, i.e. post 2012, the numbers are negligible with 10 new
dams in China, 3 in India, 2 in Mexico, and none in the US. Thus, most of the cities in
these countries already had an established flood protection infrastructure before the time
period of our analysis.
There are some interesting differences within countries between cities with and with-
out a protective dam. We provide summary statistics in Table 3.17 on country specific
geographical, economic, and disaster related characteristics, classified based on the pres-
ence or absence of a protective dam. Cities with protective dams in high income countries
have experienced a lower population growth as compared to cities without dams in these
countries. However, for India and Mexico, the difference in population growth in cities
with and without dams is minimal. Indian cities without dams are considerably richer
than cities with dams, with GDP per capita 25 percent higher in the former. Another
sharp distinction between high and low income nations exists in terms of number of flood
events. While cities protected by dams in high income countries faced twice as many floods
as cities without dams, in low income countries, the ratio was close to 1.
To analyse these differences more formally, we run the following regression:
Flood
cjmy
=α +β
1
Dams
cj
+β
2
ln(GDP/capita)
cj
+β
3
ln(Builtup Area)
cj
+
β
4
ln(Population)
cj
+δ
j
+ Γ
my
+ϵ
cjmy
(2.7)
whereDams
cj
is a dummy that equals 1 if the river flowing through city c in countryj
has a dam upstream within a 100km radius. We control for GDP per capita, built-up area
and population, and also include country and month-year fixed effects. The coefficient
20
We define a city as protected by a dam if the river flowing though the city has a dam upstream of the
city and the geodesic distance between the city, and the dam is less than or equal to 100 kilometers
119
Table 2.8: Dams and Adaptation
Flood
cjmy
∆ Population
cj
ln(Night Lights
cjmy
)
(1) (2) (3) (4) (5)
Dams
cj
0.003
∗∗
−0.015
∗∗
0.095
∗∗
(0.001) (0.007) (0.037)
Risky
cj
−0.002
(0.002)
Risky
cj
× Dams
cj
−0.006
∗∗∗
(0.002)
Flood
cjmy
−0.042
∗∗∗
(0.006)
Extreme Rain
cjmy
−0.071
∗∗∗
(0.005)
Flood
cjmy
× Dams
cj
0.017
∗∗
(0.008)
Extreme Rain
cjmy
× Dams
cj
−0.008
(0.006)
Fixed Effects
City ! !
Country ! ! !
Month× Year !
Country× Month× Year ! !
Num. obs. 292,459 3,820 3,820 277,179 277,179
Adj. R
2
0.107 0.123 0.132 0.919 0.919
Notes: clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
The dependent variable in column (1) isFlood
cjmy
, a dummy indicating whether cityc in countryj was hit by a flood in month m of yeary. The regression
includes log values of GDP per capita, Builtup Area (per sq. km), and Population pertaining to the year 2015 as controls. The dependent variable in columns (2)
and (3), ∆ Population
cj
, refers to the population growth between years 2000 and 2015 in cityc in countryj. Risky
cj
is a continuous variable that measures
the number of extreme precipitation events in cityc in countryj between years 2000 and 2015, where extreme precipitation is a dummy indicating whether
the precipitation in the monthm and yeary in cityc in countryj fell in the 90
th
percentile of the distribution of rainfall in the said city. The distribution was
created using precipitation data from 1958-2015. Both regressions include log values of GDP per capita, Builtup Area (per sq. km), and Population pertain to the
year 2015 as controls. The dependent variable in columns (5) and (6),ln(NightLights)
cjmy
, is the natural log of mean light intensity in cityc in countryj
in monthm of yeary. Both regressions include the controlsStorm
cjmy
andLandslide
cjmy
, dummies indicating whether cityc in countryj was hit by a
storm or landslide, respectively, in monthm of yeary. Two month lead and lags for all three disaster types have also been included as controls. Observations
include all city-country-month-year observations which had a non-zero value of nightlights. Each observation in the two regressions was weighted by the
mean of the cloud free coverage for the city-country-month-year observation. Standard errors are clustered at the city level.
of interest is β
1
which signifies the likelihood of floods in cities with dams. Results are
presented in Table 2.8.
Column (1) presents the correlation between cities with dams and flood events. As
expected based on the summary statistics (see Table 3.17), cities with dams face a higher
number of floods, and this result is primarily driven by the high income countries. This
is not surprising given that dams are more likely to be placed in areas at greater risk.
Columns (2) and (3) report population growth regressions and use a similar cross-section
120
specification to the one in Equation 2.1 except that we add the covariate Dams
cj
, and in-
teract it with the proxy for vulnerability of the city (Risky
cj
). The coefficient on Dams
cj
in
column (2) informs us about the population growth in cities with dams, while the coeffi-
cient on the interaction term (Risky
cj
×Dams
cj
) in column (3) indicates whether vulner-
able cities with dams have a differential population growth rate as compared to low risk
cities with dams. On average, cities protected by dams have a significant 1.5 percentage
point lower population growth rate (column 2). The summary statistics indicate that this
result too is being driven mostly by cities in high income countries. However, as column
(3) shows, this masks considerable heterogeneity between high risk and low risk cities.
Low risk cities with dams actually experience a higher population growth rate, and it is
the risky cities with dams which experience a lower population growth.
The final two columns focus on adaptation, as we explore the extent to which dams
mitigate the impact of floods and extreme rain. To estimate this, in column (4) we run a
regression similar to Equation 2.3 without the geography and capital dummies, and add
Dams
cj
and its interaction with the contemporaneous flood dummy as regressors of in-
terest. The coefficient on the interaction term signifies the magnitude by which the effect
of the floods on economy activity is mitigated by the presence of a dam within a 100km
radius. Finally, we replace Flood
cjmy
with Extreme Rain
cjmy
in column (5). Our results
suggest that cities protected by dams experience a lesser decline in night lights during
floods. As column 4 shows, the effect of floods is negative and significant, with night
lights falling by a 4.2 percent in the case of a flood. However, the effect is attenuated by
1.7 percentage points in cities with dams. Surprisingly, we do not observe any mitigating
effect of dams during extreme precipitation events. We hypothesize that dams are benefi-
cial in mitigating the effects of riverine flooding which may be caused by factors upstream.
This is not the case with extreme precipitation, which falls directly on the city, and may
cause waterlogging in low-lying areas. This last result, therefore, represents a puzzle that
merits future research.
121
2.8 Conclusion
Climate change raises the risk of extreme weather events such as floods. The damage
caused by these events can be partially offset through adaptation investments at the indi-
vidual and government level. We have studied the correlates of flood adaptation progress
by estimating how thousands of flood events that have occurred around the world from
2012 to 2018 have affected urban population growth rates, the death count from such
shocks, and lights at night dynamics. We find that floods cause a decline in short term
economic activity, with a larger negative effect in poorer nations. Though these floods
also lead to a higher death rate in vulnerable cities in low income countries, we do not
find evidence of significant out-migration from such cities. Cities in poorer nations take a
longer time to recover from these disasters, relative to disasters in high income countries.
However, the damage caused by flooding in poor nations is declining over time.
Using a global panel of cities, this paper provides evidence of flood adaptation tak-
ing place. Cities are becoming more resilient with time. Repeated experience with floods
does tend to reduce the marginal economic impact of subsequent floods, while flood pro-
tection infrastructure, specifically dams, also helps mitigate a significant portion of the
negative impact. The reasons for the differential impact of disasters across cities in low in-
come and high income countries merits further research. Shi, Chu, and Debats 2015 point
out that strong political leadership, high municipal expenditures, and awareness about
climate change are associated with adaptation planning among environmentally progres-
sive cities.
Poorer nations also suffer from poor urban planning and lack of investment in infras-
tructure. A large proportion of the urban population in poorer nations lives in slum set-
tlements. These countries are also seeing more growth in settlements in flood-risk ar-
eas (Rentschler, Avner, Marconcini, Su, Strano, and Hallegatte 2022). Much of the urban
population growth in developing countries over the last two decades has been in slums
(Marx, Stoker, and Suri 2013). Slums feature more low quality buildings and are located
122
in areas most vulnerable to climate risk and hence suffer more due to flooding. All these
factors could potentially contribute to the large income based heterogeneity of the impact
of floods that we document. Detailed flood maps coupled with knowledge of high and
low income areas within each city could help to shed light on the mechanisms behind the
differences.
A fruitful area of further research involves analysing the interplay between how private
and public decisions jointly determines disaster resilience. In this regard, the I. Ehrlich and
Becker 1972 framework offers a model for improving our understanding of producing re-
silience. In cases where governments anticipate that resilience infrastructure investments
could actually encourage greater risk taking by the public, the crowding out effect induced
by ”climate proofing” an area, the government must consider introducing complementary
policies to limit this substitution effect.
123
Chapter 3
Do Water Audits Promote Economic Welfare?: Evidence
from a Natural Field Experiment
1
3.1 Introduction
Water scarcity has become an issue across the world, and there are increasing concerns that
climate change could exacerbate this situation (IPCC 2022; Vorosmarty, Green, Salisbury,
and Lammers 2000). In response to concerns about scarcity, water suppliers and regu-
lators are showing greater interest in identifying non-price mechanisms that could help
encourage conservation. One increasingly common mechanism is a water audit, which
helps identify behavioral and technological inefficiencies in the home, and provides tai-
lored recommendations for promoting conservation. The U.S. Environmental Protection
Agency considers water audits to be the critical first step in identifying and quantifying
water uses and losses (Environmental Protection Agency, USA 2013).
2
However, very lit-
tle is known about the efficacy, efficiency, and cost-effectiveness of audits.
This paper helps fill this gap by implementing a natural field experiment that encour-
ages residential customers to take a home water audit. We partnered with Northumbrian
1
Co-authored with Jesper Akesson, Robert Hahn, and Robert D. Metcalfe
2
Based on data from the Water Research Foundation, a total of 4,575 audits were conducted by water
utilities in five US states between 2011-2014 (Sturm, Gasner, and Andrews 2015).
124
Water Limited (NWL), a water utility in the United Kingdom, to examine the effective-
ness of these audits. We randomly encourage some customers to take up the audit, and
we randomize the type of encouragement that customers receive — from various financial
incentives to environmental appeals and moral suasion. Our experimental design allows
us to study how different encouragements affect take up, and how water audits affect con-
sumption.
Our paper contributes to recent research on water and energy conservation in three
ways. First, to the best of our knowledge, this is the first paper to use a natural field exper-
iment to estimate the causal impact of water audits on consumption. Second, most pre-
vious research on water conservation studies the effect of non-pecuniary interventions,
such as moral suasion and social comparisons (see Nauges and Whittington 2019 for a
review of studies). We introduce several treatments that include both financial incentives
and non-financial incentives, which allows us to compare the effectiveness of different
kinds of treatments. Third, while researchers have noted the need for rigorous benefit-
cost analysis of water policies based on causal estimates, we develop and operationalize
two frameworks for implementing such analysis: one is a standard benefit-cost framework;
and a second is a less traditional approach based on the marginal value of public funds
(MVPF) (Hendren and Sprung-Keyser 2020).
The experimental design involves sending letters to residential customers that encour-
age them to take a water audit. This audit consists of logging into the company’s online
water audit, answering questions on water use habits and home features, and receiving
recommendations for reducing consumption. The online tool provided information on
free water-saving devices offered by the utility, and helped customers book an in-home
audit if appropriate. We measure water consumption after the interventions and compare
it with the water consumption of a control group.
125
We randomly allocate 45,000 customers to a control group and one of 6 treatment
groups. The control group received no communication. Treatment group 1 (Vanilla) re-
ceived an encouragement letter that was in use by NWL prior to the trial, while the re-
maining five groups received newly designed letters, each catering to a different motiva-
tion for water conservation. Treatment group 2 (Simplified ) received a simplified version
of the letter sent to treatment group 1, which made the call to action more salient. The
third treatment group (Altruism) received letters reminding them of the scarcity of wa-
ter, while treatment group 4 (Moral Cost) was sent a letter comparing the household’s
consumption to that of their neighbors (i.e., moral suasion). Treatment groups 5 and 6
(Incentives) received letters that provided different levels of monetary incentives ( £10 and
£15) to encourage completion of the home water audit.
We have three main findings. First, the take-up of the audit is affected by the inter-
ventions. Relative to the Vanilla letter, all letters led to a significant increase in the take-up
of the diagnostic, with the Incentives treatment having the maximum impact. Specifically,
households exposed to the Incentives 10 treatment had a 4.5 percent higher rate of take-
up relative to the Vanilla group, while the increase was 5.7 percent for households in the
Incentives 15 group. Because the impact of the two Incentive treatments are statistically
different from each other, we calculate a price elasticity of audit demand to be 0.53. Thus,
increasing the amount of the financial incentive could be a fruitful strategy to increase
participation.
Next, we estimate the causal impact of audits on water consumption using an encour-
agement design with two-stage least squares. For the first stage, we estimate the impact
of the randomized encouragement on the take-up of the audit. Using the results from the
first-stage, we then examine the impact of the audit on consumption, which yields a local
average treatment effect (LATE). Our analysis suggests that there is a meaningful effect,
ranging from 17 to 23 percent of pre-treatment consumption. Limiting our focus to the
financial incentives intervention, we estimate the £15 treatment reduces consumption for
126
metered households by 44 liters per day, while the £10 treatment reduces consumption
for the same subgroup by 43 liters per day. This suggests that the size of the subsidy for
completing the audit may not be that important for water conservation, unlike take-up.
We also consider external validity by examining how our results would generalize to
customers who currently do not have metered water consumption. Weighting each house-
hold by the inverse probability of being metered revises our estimates on water conserva-
tion downwards to 34 liters per day, though the results still remain significant with our
preferred specification. These effects persist for at least two months post-treatment.
3
Second, notwithstanding the substantial improvements in water conservation, no in-
terventions appear to pass a benefit-cost test. Because our analysis does not quantify other
potentially important benefits, such as ecosystem improvements and the value of availabil-
ity, we define a lower bound on other benefits needed to pass a benefit-cost test. We find
that a cubic meter of water conservation would need to yield other benefits or reduced in-
vestment costs of about$8, or ten times the marginal price of water, for this intervention to
be worthwhile.
4
For example, we find that the social cost of carbon in the base case would
need to be 22 times higher for benefits to just equal costs. Using our base case estimate of
$51 per ton (Interagency Working Group, USG 2021), this means the social cost of carbon
would need to be about$1200 per ton for benefits to just equal costs if there were no other
benefits.
5
Our benefit-cost framework builds on using the marginal value of public funds ap-
proach (Hendren and Sprung-Keyser 2020; Finkelstein and Hendren 2020). A key advan-
tage of this approach is that it separates the problem of estimating the welfare impact of
the subsidy from the problem of estimating the welfare impact of the intervention that
3
A limitation of our study is that we have household data on water consumption for only 65 days post-
treatment. See discussion in sec:cost
e
ffectiveness.
4
We asked NWL to provide a willingness to pay for a cubic meter increase in water conservation, but
they were unable to furnish a value.
5
In what follows, we will round all estimates in the text to two significant digits, while all estimates in
regression tables will be rounded to three decimal places. All estimates are either in 2020 pounds or dollars.
Estimates of dollars from earlier studies have also been converted from original year dollars to 2020 dollars
using the Consumer Price Index (CPI) for Urban Consumers US Bureau of Labor Statistics 2021.
127
could pay for the subsidy, such as a tax. While it is convenient to assume a lump sum
tax will be used for analytical simplicity, it is not necessary. Recent work on audits in the
energy area, discussed below, uses the assumption of lump sum transfers.
We calculate the MVPF using both short-run and long-run marginal costs for the utility.
Our results suggest that the MVPF in the base case is -0.074 using the short-run marginal
cost, and increases to 0.0048 using the long-run marginal cost. The value under the short-
run cost assumption implies that the government would be spending$1 to generate neg-
ative benefits. Using the long-run case is not much better in that $1 of net costs to the
government generates less than one cent in welfare benefits.
Third, we consider the issue of targeting interventions and explore whether such tar-
geting allows the benefits of an intervention to exceed the cost (Allcott 2011; Ayres, Rase-
man, and Shih 2013; Ferraro and Miranda 2013; Brent, J. H. Cook, and Olsen 2015; Wich-
man, L. O. Taylor, and Von Haefen 2016; Knittel and Stolper 2019; Brent, Lott, M. Taylor,
J. Cook, Rollins, and Stoddard 2020; Gerarden and Yang 2021). We examine the targeting
of high users, who are defined as users with pre-treatment consumption higher than the
median consumption. We find that targeting of high users that receive financial incentives
roughly doubles the reduction in consumption (83 liters per day versus 43 liters per day).
This suggests that audits can be targeted to improve their efficiency.
Taking this analysis a step further, we ask whether targeting could pass a benefit-cost
test. We find that targeting is not sufficient for benefits to exceed costs in our base case.
Though cost-effectiveness improves by 38 percent, we estimate that a cubic meter of water
conservation would need to yield other benefits or reduced investment costs of at least $4
for this intervention to pass a benefit-cost test.
6
The basic intuition behind our results can be explained simply. The short-term reduc-
tions in greenhouse gas emissions from water conservation are comparatively small, on
the order of 1.6 tons for 65 days for our base case. And while the experimental cost per
6
The social cost of carbon in the base case would need to be 11 times higher (as compared to 22 times
before) for benefits to just equal costs.
128
person is also relatively small, on the order of$1.2 per consumer (not including the pro-
ducer surplus loss), this leads to a cost effectiveness (CE) of $1000 per ton, which is much
higher than most estimates for the SCC (Interagency Working Group, USG 2021). If we
assume our results persist over a long period of time, the cost-effectiveness calculus looks
more attractive because the benefits from conservation increase. See sec:welfare for a more
extended discussion of welfare.
Our results build on a growing literature that examines the impacts of different kinds of
interventions on resource use, such as water and energy. There are, however, relatively few
rigorous estimates of the impacts of audits on water or energy use. A useful summary of
studies analyzing water audits is provided by Ansink, Ornaghi, and Tonin 2021. However,
they do not identify any natural field experiments that address online audits.
Apart from audits, several other interventions related to rates, rebates, and enforce-
ment regulations have been used to induce behavioral change to promote water conserva-
tion, with effect sizes comparable to ours. O. R. Browne, Gazze, and Greenstone 2019 dis-
entangle the effect of different residential water conservation policies adopted by a utility
during the 2011-2017 California drought. They find large effect of rate changes (elasticity
between .22 and .41)
7
and outdoor water schedule regulations (water use decreased by
21-24 percent). These findings are similar in magnitude to our result that participation in
audits leads to a 17 percent decline in consumption relative to pre-treatment consumption.
West, Fairlie, Pratt, and L. Rose 2021 examine the effects of automating the enforcement of
water conservation regulations, and find similar large effects, with treated households cur-
tailing their water consumption by 31 percent. Therefore, price change and enforcement
policies also lead to effect sizes in the same range as our results. The one exception is O.
Browne, Gazze, Greenstone, and Rostapshova 2019, who implement a field experiment in
California, randomly assigning visual or automated enforcement methods to detect water-
use violations.
7
The elasticity refers to the absolute value here.
129
Our study also relates to the literature on behavioral nudges. There have been sev-
eral experiments and quasi-experiments examining the implementation of social norm
messaging (e.g., Ferraro and M. K. Price 2013, Brent, J. H. Cook, and Olsen 2015, Jaime
Torres and Carlsson 2016, S. Datta, S. Datta, Josı, Zoratto, Calvo-Gonzı, Darling, et al.
2015). Nauges and Whittington 2019 provide a review of the literature on the impact of
information treatment on water and energy use.
8
Most studies, whether in the energy
or the water sector, find that social norm information treatments reduce consumption by
about 2 to 5 percent for a period of time, with greater reductions typically observed when
the intervention includes social norm comparisons as opposed to interventions providing
technical advice or raising awareness. Our paper integrates social norms messaging with
online audits (Moral Cost letter), allowing us to study the effect of the former on diagnos-
tic completion and their combined effect on water consumption. We find that though the
Moral Cost letter has a significant effect on take-up of the audit, their combined effect on
consumption is relatively small — a 1.1 percent decline in consumption compared with the
Vanilla treatment group. This is somewhat lower than the impact of several interventions
that use only social norms in related contexts.
A natural question that arises with interventions aimed at incentivizing natural re-
source conservation is whether they can be scaled up — a subject that requires investigat-
ing whether the benefits exceed the costs. Several authors have used their estimates of the
expected change in water or energy use to estimate the cost effectiveness and economic
welfare implications of water conservation strategies. These include studies on the im-
pacts of metering, social norm messaging, and the nature of the regulatory intervention.
Ferraro and M. K. Price 2013 find that social norm messaging augmented by technical ad-
vice reduces consumption by 4.8 per cent, which implies a cost of $0.17 per cubic meter
8
Nauges and Whittington 2019 argue using illustrative calculations that that social norm messaging in-
struments may not pass a benefit-cost test, especially in low- and middle-income countries. Our results
indicate that the same could hold true for high-income countries for certain kinds of behavioral interven-
tions, such as audits. In contrast, Mansur and Olmstead 2012 suggest there could be potential welfare gains
of switching from non-market to market-based regulation of water supply during periods of drought.
130
reduced for the utility. Bernedo, Ferraro, and M. Price 2014 demonstrate that persistent
long-term impacts of the policy studied by Ferraro and M. K. Price 2013 imply that the cost
per gallon saved is 60 percent lower ($0.07 per cubic meter) than the figure derived using
only contemporaneous treatment effects. This is substantially lower than the estimates for
our experiment, which range from$1.30 to$3.70 per cubic meter reduced.
9
The evidence on the cost effectiveness of water audits is limited. To the best of our
knowledge, Ansink, Ornaghi, and Tonin 2021 provide the only cost effectiveness assess-
ment of water audits, and find that technology is more cost-effective than information
provision by a factor of two for a water audit program in England. However, the selection
into their audit program was not random, and the focus was exclusively on households
with above-average water use.
In contrast to water, the experimental literature on the cost-effectiveness of conserva-
tion measures in the energy sector is well developed. Insightful examples include Allcott
and Greenstone 2017 and Fowlie, Greenstone, and C. Wolfram 2018, who study the wel-
fare impact of audits. Their results are similar to ours. The former study models home
energy efficiency investment decisions to evaluate two large residential energy efficiency
programs in Wisconsin. These programs involved a home energy audit followed by de-
cisions on which recommended investments to undertake. They implement a large field
experiment in Wisconsin, and find that the programs reduced economic welfare. A com-
parison of the observed investment costs with the present discounted value of energy sav-
ings indicates the programs has an internal rate of return of -4.1 percent, while a revealed
preference model finds that the programs reduce welfare by $0.18 per dollar of subsidy.
Our finding of a negative MVPF of -0.074 has a similar implication. The costs to the gov-
ernment of the intervention are higher than the social benefits.
9
One area that we do not address is spillovers that may occur due to water conservation, for example,
in terms of energy use. This could increase the attractiveness of the interventions we study. Jessoe, Lade,
Loge, and Spang 2021 experimentally test the effect of social norms messaging about residential water use on
electricity consumption. Taking into account the electricity conservation spillover increases the net benefits
of their intervention from$2.9 per household to$4.0, an increase of 39 percent. If we assume that benefits
increase by this amount in our application, the interventions would still not pass a benefit-cost test.
131
In Fowlie, Greenstone, and C. Wolfram 2018, the authors measure the welfare gains
from the Weatherization Assistance Program, a residential energy efficiency program in
Michigan. The program involves conducting an energy audit of the home before imple-
menting a weatherization retrofit, with the purpose of recommending specific efficiency
improvements. The paper uses experimental and quasi-experimental variation in partic-
ipation to identify the returns to investments. Their results suggest that the upfront in-
vestment costs are about twice the actual energy savings, and the projected savings are
more than three times the actual savings. This again implies that the costs outweigh the
benefits, which supports the results from our welfare analysis.
The paper proceeds as follows. Section 3.2 provides the details on the audit program
and randomized trial. In Section 3.3, we describe our empirical strategy and present the
results from the experiment. Section 3.4 presents a welfare analysis, including information
on cost effectiveness. Conclusions and areas for future research are discussed in Section
3.5.
3.2 Background and Experimental Design
In 2018, Northumbrian Water commissioned Save Water Save Money (SWSM) to provide
its online water audit tool for Northumbrian’s customers.
10
The tool, hosted on the com-
pany website, asked customers questions about their water use habits and homes. The
main purpose of the tool was to help customers understand their water consumption, and
identify ways in which they can save water and money. The tool also informed customers
about free water-saving devices that NWL offers, and helped them book an in-home water
audit if appropriate. The questionnaire on the platform took approximately ten minutes
to complete.
10
The water audit can be accessed at this url: https://www.getwaterfit.co.uk/questions/ (last accessed:
March 09, 2023)
132
NWL was interested in getting customers to take their online water audit, and under-
standing the impact of the audits on consumption. We were interested in helping NWL
with these objectives, and, in addition, understanding the impact of different behavioral
interventions on economic welfare. In order to encourage the use of the SWSM platform,
we designed a set of customer communications using theories from economics and behav-
ioral science. We used one of NWL’s existing direct mailers as a template, and designed
5 new direct mailers (see the discussion in Section H.4). The only difference between the
five communications was the application of different behavioral science ideas.
We implemented a natural field experiment (Harrison and List 2004) to test the effec-
tiveness of the redesigned letters, and to understand how the SWSM platform influences
water consumption. This field experiment included 44,757 NWL customers, spread across
three post code areas. The customers that participated in the trial were randomly allo-
cated to one of six treatment groups that received letters or a control group that received
no letter. Subsequently, customers for whom NWL had email contact details were also
randomly allocated to groups that either received or did not receive an email reminder
about the online audit tool. The reminder emails followed the same theme as the initial
letters that customers received. This design allows us to estimate the effects of particular
letters and reminders on take-up of the audit.
11
There were six letter treatments. Treatment 1 (Vanilla) was NWL’s initial letter, and
informed customers that they can save water and money by using the free online platform.
It also noted that many other customers had saved money with the platform, and told
them how to access it. Treatment 2 (Simplified ) was similar to the Vanilla communication
but it simplified the content, making the main message and the call to action more salient.
Treatment 3 (Altruism) added to the message of the Simplified mailer by reminding the
consumers that water is a scarce resource, and asked them to help conserve it in their local
area. Treatment group 4 (Moral Cost) received a letter that complemented the Simplified
11
We are not aware of other studies of water audits that have estimated the effect of email reminders.
133
Mailer by telling customers that people in their region were making a change in an effort
to save water, and invited them to join their neighbors. Furthermore, for consumers with
relatively high water consumption, it informed them that they were in the top 50 percent of
consumption, whereas for the bottom 50 percent, it congratulated them on being efficient.
The final two treatment groups, Treatment 5 and Treatment 6, were offered pecuniary
incentives (£10 Incentive and £15 Incentive) for completing the water audits. The former
supplemented the Simplified mailer by emphasizing monetary savings, and offered a £10
incentive for using the platform, while the latter communication changed the incentive
from£10 to£15.
The data used to randomize the trial participants and to measure outcomes came from
three anonymized sources: NWL’s administrative data on meter readings; the SWSM plat-
form, which was used to code responses to the diagnostic questionnaire; and Customer Re-
lationship Management (CRM) data identifying whether reminder emails were opened.
12
The experiment took place over four months between December 2018 and March 2019.
We collected baseline data for purposes of randomization and analysis of pre-treatment
consumption from January 2017. All direct mailers were posted on 8
th
December 2018,
and email reminders were sent on 6
th
February 2019.
Table 3.20 presents summary statistics on the observable characteristics of the house-
holds across different treatment groups. We have data on whether the household was in a
rural or urban area, whether they had a water meter, whether they provided NWL with an
email, and their consumption before the experiment. Using an F-test of joint significance,
we find that the differences across different treatment groups are not statistically signifi-
cant at conventional levels. This suggests that that the various treatments are balanced on
pre-treatment observable variables.
12
CRM is a tool to help manage and analyze customer interactions and data on websites.
134
3.3 Results
We begin by reporting the effect of the letters on the take-up of the audit program, and
then analyze the impact of the interventions on water consumption. We then use a LATE
analysis to measure the effect of completing the diagnostic on water conservation. Our
analysis of consumption is limited to households with meters. To measure the likely im-
pact of the interventions if scaled up to include non-metered households, we reweight
our estimates to reflect the broader population of consumers. We also use the data on
reminders to study the effect of reminders on completing the audit. However, because
the reminder emails were sent near the end of the study period, we cannot analyze their
impact on consumption.
3.3.1 Likelihood of Engagement
To examine the effects of the behavioral interventions on the share of households that
complete the diagnostic, we run the following regression:
13
y
i
=α +
X
j
β
j
T
ij
+γX
i
+ϵ
i
(3.1)
where,y
i
is a dummy variable that equals 1 if householdi completed the water audit, and 0
otherwise.α represents the average take-up of the audit for the excluded treatment group.
T
ij
is a dummy that equals 1 if householdi received treatmentj, and 0 otherwise, where
j refers to the different treatment groups. The coefficient of interest, β
j
, is the average
treatment effect (ATE) of the different letters on the likelihood of completing the audit.
X
i
represents a vector of dummy controls,γ is a vector of estimates of their impact, andϵ
i
is an error term.
13
The raw data from the field experiment on the number of households that completed the diagnostic,
and how that differs across metered and unmetered households, is presented in Table 3.21. We do not have
data on the water-saving devices ordered by different households, and if they booked an in-home audit.
135
Table 3.1 presents the estimates. The excluded group in models (1) and (2) is the
Vanilla letter, and the excluded group in models (3) and (4) is the Simplified letter. The
control vector here includesRural
i
, which is a dummy that equals 1 if householdi lived
in a rural area, andMeter
i
, which is also a binary variable that equals 1 if householdi has
a water meter.
14
We present the results both with and without controls included in the
regressions. Our results indicate that relative to the Vanilla treatment arm, all the letters
increased the take-up of the diagnostic significantly, with the Incentive treatment arm per-
forming the best. Within the Incentives treatment arm, the higher financial incentive of £15
had a marginally greater impact (5.7 per cent versus 4.5 per cent).
15
The effect of the Al-
truism letter becomes insignificant when the reference treatment arm is the Simplified letter
(column 3), but the impact of the Incentives and Moral Cost letters continues to be positive
and significant. The results do not differ when we control for whether a household is situ-
ated in a rural area and has a water meter. We conclude that behavioral interventions can
help to promote the use of audit tools, with financial incentives being more effective than
others.
3.3.2 Effect of the Behavioural Interventions on Water Consumption
Although the letters were successful in promoting the take-up of the water audit tool,
the main objective was to encourage water conservation. In this section, we estimate the
effects of the different communications on household water consumption. The letters can
work in one of two ways: first, by directly encouraging an individual to conserve water
after being influenced by the content of the letter; and second, through take-up of the
audit. Unfortunately, we cannot estimate the direct effect of the letters because the time
period between receiving the letter and completing the audit is too small. This section,
thus, focuses on the overall impact on consumption of the direct encouragement and the
take-up of the audit.
14
We do not have data on household covariates such as income and family size
15
These impacts are statistically different from each other at a 1 per cent level of significance
136
Table 3.1: ATE Estimates of Letters on Diagnostic Completion
Completed Diagnostic
Vanilla Simplified
(1) (2) (3) (4)
Simplified 0.007
∗∗∗
0.007
∗∗∗
(0.002) (0.002)
Altruism 0.005
∗∗
0.005
∗∗
−0.002 −0.002
(0.002) (0.002) (0.003) (0.003)
Incentives£10 0.045
∗∗∗
0.045
∗∗∗
0.039
∗∗∗
0.038
∗∗∗
(0.004) (0.004) (0.004) (0.004)
Incentives£15 0.057
∗∗∗
0.057
∗∗∗
0.050
∗∗∗
0.051
∗∗∗
(0.005) (0.005) (0.005) (0.005)
Moral Cost 0.016
∗∗∗
0.016
∗∗∗
0.009
∗∗∗
0.009
∗∗∗
(0.003) (0.003) (0.003) (0.003)
Intercept 0.019
∗∗∗
0.008
∗∗∗
0.025
∗∗∗
0.013
∗∗∗
(0.002) (0.002) (0.002) (0.003)
Controls No Yes No Yes
Observations 37,298 37,298 29,838 29,838
Notes: Robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All regressions report the average treatment effect estimates of different behavioral interventions on diagnostic com-
pletion (eq:1). The dependent variable for all models is Completed Diagnostic, a dummy variable that equals 1 if the
household completed the water diagnostic, and 0 otherwise. Models (1) and (2) exclude the observations in the
Control group, with the Vanilla letter comprising the reference treatment arm. Models (3) and (4) exclude the ob-
servations in the Control and Vanilla groups, with the Simplified letter serving as the reference group. Models (2)
and (4) include the dummy variables Meter and Rural as controls. The former equals 1 if the household has a water
meter attached to it, and the latter equals 1 if the household is located in a rural area.
To estimate the effect of the treatment on consumption, we run the following regression
of post-treatment water consumption (y
i
) on an indicator for whether the household was
treated (T
i
), and a vector of controls:
y
i
=α +βT
i
+γX
i
+ϵ
i
(3.2)
where T
i
equals 1 if the household received any treatment letter, and 0 if it was in the
control group. With respect to water consumption, we have two data points for each
household: one pre-treatment and one post-treatment. Consumption information was not
available at a uniform frequency for all participants, which makes calculation of monthly
137
consumption data difficult. Though pre-consumption data are available since 2017, post-
consumption data were available only up until February 2019 for a majority of the sample.
Therefore, we are only able to estimate the short-term impact of the audit. Furthermore,
the data on water consumption is available for the 30 percent of households who had a
meter installed.
16
However, this is not a concern for our econometric identification strat-
egy because, as shown in Table 3.20, all treatment groups were balanced on the proportion
of households with a water meter, pre-diagnostic consumption, and the number of con-
sumers in each treatment group in the top 50
th
percentile of consumption (high-use house-
holds).
17
Furthermore, we run balance tests on observable characteristics just for metered
households in Table 3.20 (columns (6) and (7)) and find no significant differences be-
tween different treatment groups. Additionally, to address concerns about possible bias
in sample selection due to the focus on metered households, we reweight our LATE esti-
mates in Section 3.3.4 so that the metered sample matches the demographic composition
of the general population of NWL customers.
To analyze the heterogeneity among different treatments, we ran a regression similar
to Equation 3.1, withy
i
now denoting post-treatment water consumption for household
i. The vector of covariates, X
i
, now consists of Rural
i
and an additional covariate, Pre-
Treatment Water Consumption
i
, which measures the daily water usage of a household be-
fore the letters were sent. For all regressions, variables related to water consumption are
measured in liters per day. The effect of receiving a letter on consumption (Equation 3.2)
is presented in column (1) of Table 3.2, while the heterogeneity results are reported in
columns (2)-(4).
We find evidence that all behavioral interventions, except Vanilla, reduced water con-
sumption, though results are significant only for the Incentives group. Column (1) pro-
vides the average treatment effect of receiving any letter on post-treatment consumption.
16
Details on the computation of pre- and post-treatment water consumption are provided in Section H.3.
17
Calculations on pre-diagnostic consumption and the number of high-use customers are only feasible for
households that have meters.
138
Table 3.2: ATE Estimates of Letters on Post-Treatment Consumption
Post-Treatment Water Consumption
Control Control Vanilla Simplified
(1) (2) (3) (4)
Treated −1.392
(1.264)
Vanilla 1.627
(1.600)
Simplified −1.595 −3.217
∗∗
(1.635) (1.592)
Altruism −1.593 −3.216
∗∗
−0.031
(1.625) (1.581) (1.613)
Incentives£10 −3.543
∗
−5.177
∗∗∗
−2.077
(1.923) (1.886) (1.916)
Incentives£15 −4.687
∗∗
−6.328
∗∗∗
−3.259
∗
(1.919) (1.882) (1.910)
Moral Cost −1.302 −2.925
∗
0.191
(1.592) (1.546) (1.581)
Intercept 117.635
∗∗∗
116.782
∗∗∗
119.045
∗∗∗
44.042
∗∗∗
(43.032) (42.439) (42.634) (6.524)
Controls Yes Yes Yes Yes
Observations 11,700 11,700 9,770 7,795
Notes: Robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All regressions report the average treatment effect estimates of different behavioral interventions on post-treatment
water consumption (eq:2), measured in liters per day. The dependent variable for all models is Post-Treatment Water
Consumption, a continuous variable that measures the water consumption of a household after the treatment date
December 8, 2018. Pre-treatment consumption and post-treatment consumption were available for only a subset
(30 percent) of the households. Households with unreasonably large differences between pre- and post-treatment
consumption (absolute value greater than 50 percent) were dropped from the sample. The data were trimmed at 1
and 99 percentile of pre-treatment consumption. The model names reflect the reference group for each regression.
The regressor of interest in Model (1), Treated, is a dummy variable that equals 1 for all households that received
any letter. Models (1) and (2) include all observations, with the control treatment arm constituting the reference
group. Model (3) excludes the observations in the control group, with the Vanilla letter comprising the reference
group. Model (4) excludes the observations in the control and Vanilla group, with the Simplified letter acting as the
reference group. All models include Rural and Pre-Treatment Consumption as controls. Rural is a dummy variable
that equals 1 if the household is located in a rural area. Pre-Treatment Consumption is a continuous variable that
measures the water consumption of a household before the treatment date of December 8, 2018.
139
Though the estimate is negative, it is insignificant. Columns (2) through (4) estimate the
effect for each behavioral intervention, with the reference group as the control, Vanilla,
and Simplified letter, respectively. With reference to the control group, all treatment arms
except Vanilla experienced a fall in consumption after letters were sent out; however, only
the monetary incentives led to a statistically significant decrease. Though point estimates
suggest that Incentives 15 had a larger impact than Incentives 10 (3.5 versus 4.7 liters per
day), the two are not significantly different from each other. When we exclude the control
group, and the Vanilla letter becomes the omitted category (column (3)), the drop in con-
sumption is significant across all remaining categories, with the decrease in consumption
ranging from 2.9 liters per day under Moral Cost to 6.3 liters per day under Incentives 15. In
percentage terms, this decrease amounts to between 1.1 percent and 2.5 percent of the pre-
treatment water consumption, a small but economically meaningful impact.
18
The effect
of the Incentives 15 treatment is more than twice the effect of the Moral Cost one, and the
effect sizes are statistically different from each other. In general, pecuniary incentives lead
to a significantly larger decrease in consumption when compared with other behavioral
interventions.
Finally, dropping both the control group and the Vanilla group with the Simplified let-
ter as the reference category (column 4) leads to only the£15 financial incentive remain-
ing significant. In summary, the Incentives group significantly reduced their consumption
regardless of the reference group, while the other treatments had a significant negative
impact only when we compare them to the Vanilla arm. Again, it is important to note that
these estimates represent the overall impact of the treatment letters, and so the numbers
are bound to be small as we average across all households, many of which never completed
the audit. We attempt to disentangle the effect of completing the audit in the following
section.
18
The small decrease in consumption due to the Moral Cost letter (which also combined a social com-
parison message) stands in contrast to the literature (Ferraro and M. K. Price 2013), which finds that social
comparison messages have a greater impact on water conservation than prosocial messages or technical
information alone.
140
3.3.3 Effect of Diagnostic Completion on Water Consumption
We now turn our attention to estimating the impact of completing the audit on water
consumption. To do this, we employ an Instrumental Variable (IV) strategy using two
stage least squares (2SLS).
Table 3.3: LATE Estimates of Diagnostic Completion on Post-Treatment Consumption
Post-Treatment Water Consumption
(1) (2) (3) (4)
Complete Diagnostic −45.446
∗∗∗
−43.442
∗∗∗
−58.922
∗
−49.147
(15.335) (16.691) (34.400) (93.367)
Intercept 10.402
∗∗∗
10.276
∗∗∗
13.769
∗∗∗
15.040
(1.525) (2.122) (4.041) (12.469)
Instruments All Treatment Incentives+Simplified Incentives Incentives 15
F-stat in First Stage 39 66 16 3
Controls Yes Yes Yes Yes
Observations 11,700 5,830 3,900 1,974
Notes: Robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All regressions report the local average treatment effect estimates of diagnostic completion on post-treatment water consumption. The depen-
dent variable for all models is Post-Treatment Water Consumption, a continuous variable that measures the water consumption of a household,
in liters per day, post the treatment date of 08-Dec-2018. Pre-treatment consumption and post-treatment consumption were available for only
a subset (30 per cent) of the households. Households with unreasonably large differences between pre and post-treatment consumption (ab-
solute value greater than 50 per cent) were dropped from the sample. The data was then trimmed at 1 and 99 percentile of pre-treatment
consumption. The regressor of interest is Complete Diagnostic, which is a dummy variable that equals 1 for all households who completed the
water diagnostic. The IV in Model (1) is a vector of dummies for all the different treatment arms. The IV in Model (2) is a vector that includes
dummies for Incentives 10, Incentives 15, and Simplified treatment arms. The IV in Model (3) is a vector of dummies for Incentives 10 and
Incentives 15 groups, while the IV in Model (4) is only the Incentives 15 group. The sample in model (2) consists of the Incentives, Simplified
and control group, while the sample in model (3) includes only Incentives and the Simplified group. Model (4) only includes the Incentives
group. All models include Rural and Pre-Treatment Consumption as controls. Rural is a dummy variable that equals 1 if the household is located
in a rural area. Pre-Treatment Consumption is a continuous variable that measures the water consumption of a household, in liters per day, before
the treatment date of 08-Dec-2018.
The first stage involves running the following regression:
Diagnostic Completion
i
=α +
X
j
β
j
Z
ij
+γX
i
+ϵ
i
(3.3)
where Diagnostic Completion
i
is a dummy that equals 1 if householdi completed the online
diagnostic, andZ
ij
is the instrument used. The number of instruments vary depending on
the specification, and the subscript j refers to the different instruments. β
j
is the estimate
of the j
th
instrument. X
i
is a vector of household covariates as before, and consists of
Rural
i
and Pre-Treatment Water Consumption
i
. γ
i
is a vector of estimates of the impact of
141
the household covariates, and ϵ
i
is the error term. The second stage involves using the
residuals from Equation 3.3,
ˆ
Z
i
, to run the following regression:
y
i
=α +β
ˆ
Z
i
+γX
i
+ϵ
i
(3.4)
wherey
i
represents post-treatment water consumption, and X
i
is the same vector of house-
hold covariates used in the first stage.
We use different combinations of instruments for our LATE estimates, all of which give
similar results. The results are presented in Table 3.3. The model in column (1) uses all
the letters as instruments. Therefore,Z
i
is a vector of lengthj = 6, with each element of
the vector a dummy variable for the different letters. This model satisfies the relevance
condition as letters do tend to increase adoption of the water audit tool (see Table 3.1).
The estimates in column (1) suggest that completing the diagnostic led to a significant
fall in consumption of 45 liters per day (17 percent). However, a potential problem with
the instrument in column (1) is that the exclusion restriction may not strictly be satisfied, as
certain letters could directly impact water consumption through their message of altruism
or moral suasion (the direct impact). Therefore, in column (2), we restrict the sample to
the following four groups: Incentive £10, Incentive £15, Simplified and the control group.
Z
i
now represents a vector of 3 instruments, namely Incentives £10, Incentives £15, and
Simplified groups.
19
We are reasonably confident of satisfying the exclusion restriction here
because there were few differences between the Incentives and the Simplified letter, with
the exception that the former used a monetary incentive. These letters simply asked the
customers to download the water audit application, without any inducement to an envi-
ronmental or altruistic cause, and therefore our assumption is that they should not affect
water consumption directly. Our results under this specification indicate that completing
the diagnostic reduces water consumption by 43 liters per day (17 percent).
19
This formulation appears to satisfy the relevance condition i.e., the correlation between the endogenous
regressor and the IV is significantly different from 0. As column (2) in Table 3.2 shows, the letters do tend
to increase the likelihood of completing the audit.
142
Next, in column (3) we run a regression that is even more likely to be consistent with
the exclusion restriction by restricting our sample to Incentives £10, Incentives £15, and Sim-
plified groups. Our set of instruments now include the Incentives £10 and Incentives £15
groups. This specification is more robust because, if the very act of receiving a letter influ-
enced water consumption, then the presence of the control group in the regression would
violate the exclusion restriction. Model (3) avoids this problem by only including customers
who received the Incentives or Simplified letters. The effect size increases to 59 liters per day
(23 percent), but it is significant only at the 10 percent level. This is most likely due to is-
sues of statistical power as our sample size has decreased considerably. Finally, there is a
possibility that the Simplified letter affected the customers attitude towards water conser-
vation directly, and therefore, including it would be a violation of the exclusion restriction,
making our LATE estimates biased. In column (4), we address this issue by restricting
our sample to the Incentives £10 and Incentives £15 groups. Z
i
is now a single instrument
represented by a dummy for the Incentives£15 group. Though we get a negative coefficient
of 49 liters per day, which is similar to our earlier specifications, it is insignificant because
of low statistical power. Given the small sample size in the last two models, we use model
(2) as our preferred specification.
Our results suggest that there is a meaningful effect of completing the water audit
tool on water consumption, ranging from 17 to 23 percent of pre-treatment consumption.
However, the duration of this effect is not known; nor do we know whether the audit may
have stimulated the adoption of new technologies over time.
In Section H.1.2, we examine whether there were any heterogeneous treatment effects
of the behavioral interventions. Specifically, we test whether households with consump-
tion greater than the median (high-use households) conserved more after receiving the let-
ter. This is relevant because if audits differ in terms of their impact across groups, it may be
more effective to target a behavioral intervention based on a customer’s attributes. We find
that the effect of the Incentive £10 and Incentive £15 letters on consumption is 7.5 and 8.4
143
liters per day, respectively, for the high users, and both effects are statistically significant.
On the other hand, the effect is indistinguishable from 0 for the low users, irrespective of
the intervention. A related analysis for the LATE, presented in Section H.1.2 (Table 3.23),
reveals a similar pattern. Completion of the audit had a large impact on the high users
(between 78 liters to 89 liters per day), but it was not statistically different from 0 for the
low users.
20
Thus, average treatment effects mask crucial heterogeneity.
21
3.3.4 External Validity
The main results of our experiment are for metered houses in NWL’s service region. In
this section we explore how the results could generalize to other NWL consumers who
do not currently have meters. Our reweighting exercise reduces the estimates of water
savings from 43 liters per day (17 percent) to 34 liters per day (14 percent), but they still
remain statistically significant.
The reweighting exercise is important because the sample used for estimating the effect
sizes of the interventions consists solely of metered households. This may lead to concerns
about the extent to which these findings generalize to other populations. We cannot say
whether our numerical estimates generalize to populations outside the region that NWL
serves; however, within our sample we can explore the extent to which the sample might
be affected by including all customers as opposed to just those customers that have me-
ters. Though we show that all observable covariates for the metered households are bal-
anced across all treatment groups (Section H.1.1), we can test the sensitivity of the results
by reweighting the study sample to match the demographic composition of the general
population of NWL customers. We reweight the metered sample so that it looks like the
general population that was sampled, and that yields a reweighted LATE. One important
20
The difference between the ATE and LATE estimate (7.5-8.4 liters per day versus 78-89 liters per day) is
large because the former measures the effect of the letter on water consumption for all treated households,
while the latter looks at the impact on households who completed the audit.
21
We also checked whether different interventions encouraged different categories of households to take-
up the water-audit tool, but did not find any significant differences in take-up between high and low users.
144
caveat is that the reweighted LATE is conditional on unmetered households getting a me-
ter. If this is not the case, the impact of an intervention on a metered household is likely to
be very different from the same intervention for an unmetered one because information
on water use via meters could significantly alter water consumption.
To implement the reweighting, we conduct the following four steps (similar to Stuart,
Cole, Bradshaw, and Leaf 2011). First, we determine the household demographics (X
i
)
we use to reweight. We choose all of the observable variables that were provided to us by
NWL: rural-urban classification, availability of email address, and residential post-code.
22
Second, we use a logistic regression to model the probability (ˆ p
i
) of being metered with
the covariates as predictors. ˆ p
i
thus denotes the estimated probability of sample selection
for householdi. Third, we follow inverse probability of treatment weighting (IPTW) to
weight each household.
23
IPTW gives each household their own weight, which is calcu-
lated as the inverse propensity scores, i.e., in our setting, the inverse probability of being
metered: w
i
(X
i
) = 1/ˆ p
i
(X
i
). Lastly, we estimate the LATE using the weightsw
i
we gener-
ated as a population weight.
To estimate the re-weighted LATE, we run the same regression as in Section 3.3.3, but
include the weights in the estimation. The reweighted LATE estimates are presented in
Table 3.4. For our preferred specification in column (2) ( Incentives£10, Incentives£15 and
Simplified group as IV), re-weighting slightly reduces the point estimates from 43 liters per
day to 34 liters per day (13 per cent), and the coefficients still remain statistically signif-
icant. The reweighted LATE in column (1), where we use all the letters as instruments,
is also very similar to the unweighted LATE in Table 3.3, with a significant effect size of
42 liters per day. For models (3) and (4), where the instruments are both Incentives £10
and Incentives £15, and only Incentives 15, respectively, the effects sizes are insignificant.
We do not read too much into these coefficients because of low statistical power, but they
22
We do not have data on household income or the number of household members, so we use the data
on post codes as a proxy.
23
See Hahn and Metcalfe 2021 for weighting using sub-classification , which is a coarser method than IPTW.
145
Table 3.4: Reweighted LATE Estimates
Post-Treatment Water Consumption
(1) (2) (3) (4)
Complete Diagnostic −41.634
∗∗
−33.854
∗∗
6.403 −220.795
(20.607) (16.184) (40.670) (358.316)
Intercept 7.819
∗∗∗
7.125
∗∗∗
4.991 33.334
(1.790) (1.845) (4.356) (43.110)
Instruments All Treatment Incentives+Simplified Incentives Incentives 15
F-stat in First Stage 39 66 21 1
Controls Yes Yes Yes Yes
Observations 11,700 5,830 3,900 1,974
Notes: Robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All regressions report the reweighted local average treatment effect estimates of diagnostic completion on post-treatment water consumption.
The dependent variable for all models is Post-Treatment Water Consumption, a continuous variable that measures the water consumption of
a household, in liters per day, post the treatment date of 08-Dec-2018. Pre-treatment consumption and post-treatment consumption were
available for only a subset (30 per cent) of the households. Households with unreasonably large differences between pre and post-treatment
consumption (absolute value greater than 50 per cent) were dropped from the sample. The data was then trimmed at 1 and 99 percentile
of pre-treatment consumption. The regressor of interest is Complete Diagnostic, which is a dummy variable that equals 1 for all households
who completed the water diagnostic. The IV in Model (1) is a vector of dummies for all the different treatment arms. The IV in Model (2) is
a vector that includes dummies for Incentives 10, Incentives 15, and Simplified treatment arms. The IV in Model (3) is a vector of dummies
for Incentives 10 and Incentives 15 groups, while the IV in Model (4) is only the Incentives 15 group. The sample in model (2) consists of
the Incentives, Simplified and control group, while the sample in model (3) includes only Incentives and the Simplified group. Model (4) only
includes the Incentives group. All models include Rural and Pre-Treatment Consumption as controls. Rural is a dummy variable that equals
1 if the household is located in a rural area. Pre-Treatment Consumption is a continuous variable that measures the water consumption of a
household, in liters per day, before the treatment date of 08-Dec-2018.
have been presented for comparison. In conclusion, we are reasonably confident that our
results are not driven by sample selection, and can be scaled up with similar effects to
unmetered households, provided they are metered before the intervention.
24
3.3.5 Effect of Reminders on Diagnostic Completion
Customers that provided NWL with email contact details (13,989 households) were also
randomly allocated to groups that either received or did not receive an email reminder.
The randomization was limited to households who had not completed the water audit by
February 2019. The reminder emails followed the same themes as the initial direct mailers
that the customers received. This allows us to estimate the impact of receiving a reminder
email. We run the following regression to estimate the effect of reminders:
y
i
=α +ϕR
i
+
X
j
β
j
T
ij
+
X
j
π
j
R
i
×T
ij
+γX
i
+ϵ
i
(3.5)
24
Metered and unmetered households may differ on unobserved variables, in which case the results may
not generalize.
146
Table 3.5: ATE Estimates of Reminders on Diagnostic Completion
Vanilla Simplified
(1) (2) (3)
Reminder 0.026
∗∗∗
0.015
∗∗∗
0.026
∗∗∗
(0.002) (0.004) (0.005)
Reminder× Simplified 0.011
∗
(0.006)
Reminder× Altruism 0.003 −0.007
(0.006) (0.006)
Reminder× Incentives£10 0.015
∗
0.004
(0.008) (0.009)
Reminder× Incentives£15 0.011 0.001
(0.008) (0.009)
Reminder× Moral Cost 0.030
∗∗∗
0.019
∗∗
(0.007) (0.008)
Intercept −0.007
∗∗∗
−0.008
∗∗∗
−0.009
∗∗∗
(0.002) (0.002) (0.002)
Controls Yes Yes Yes
Observations 11,031 11,031 8,752
Note: Robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All regressions report the average treatment effect estimates of reminders on diagnostic completion (eq:3). The
dependent variable for all models is Completed Diagnostic, a dummy variable that equals 1 if the household com-
pleted the water diagnostic, and 0 otherwise. Models (1) and (2) exclude the observations in the control group,
with the Vanilla letter comprising the reference treatment arm in model (2). Model (3) excludes the observations
in both the control and Vanilla group, with the Simplified letter constituting the reference group. The estimates
on the various treatment arms (Simplified, Altruism, Incentives 10, Incentives 15 , and Moral Cost) are omitted from
the table in the interest of space, but are all statistically insignificant. Observations only include households for
whom NWL had email contact details, provided they did not complete the diagnostic before the reminder emails
were sent. Therefore, 631 households for whom NWL had email details, but who had completed the water diag-
nostic before the reminders were sent, were excluded from the analysis. All regressions include Meter and Rural
as controls. The former equals 1 if the household has a water meter attached to it, and the latter equals 1 if the
household is located in a rural area.
147
where,y
i
is a dummy for diagnostic completion,R
i
is a dummy that equals 1 if the house-
hold i received a reminder email, and T
ij
is a dummy that that equals 1 if household i
initially received treatmentj. X
i
is a vector of household covariates, specifically dummies
for whether the household was located in a rural area, and whether it had a water me-
ter attached to it. The constant α represents the average diagnostic completion rate for
households that were not sent a reminder and that belonged to the excluded group in the
regression analysis,ϕ is the estimate for the average effect of any reminder on diagnostic
completion, andβ
j
represents the effect of the initial treatment allocation on diagnostic
completion. Our main coefficient of interest is π
j
which is the estimate on the interaction
term. It represents the effect of reminders belonging to the j
th
treatment group on diag-
nostic completion. ϵ
i
signifies the error term. The sample only includes households who
had not completed the audit on the date the reminder emails were sent out. Table 3.5
presents the results.
In column (1) in Table 3.5, we estimate the direct impact of any reminder on the likeli-
hood of completing the diagnostic. To do so, we modify Equation 3.5 and run the model
without the effect of initial treatment groups ( T
ij
), and the interaction terms between the
treatment groups and reminders (R
i
×T
ij
). Our findings suggest that reminders increased
the likelihood of completing the diagnostic by 2.6 percent as compared to the group that
did not receive the reminders. Next, in columns (2) and (3), we estimate the impact of
each specific reminder. The omitted category in column (2) is the Vanilla letter, with con-
trol group excluded from the sample.
25
With reference to the Vanilla group, reminders
to the Moral Cost group have the highest additional impact of 3 percentage points, while
the magnitude of impact for Incentives £10 and Simplified groups is also significant. No-
tably, the impact of the Moral Cost reminder is significantly different from the impact of
these other two treatments. In the final specification in column (3), the omitted group is
Simplified , with both control and Vanilla groups excluded from the sample.
25
We omit estimates of effect of the initial treatment assignment ( β
j
) from Table 3.5 for brevity. However,
all of the estimates ofβ
j
are insignificant for all specifications.
148
3.4 Welfare Analysis
In this section, we examine whether promoting online water audits improves economic
welfare. We consider the impacts of different interventions from our experiment on var-
ious measures of economic welfare. We have three main findings. First, there is not suf-
ficient information on water conservation benefits to say whether specific interventions
would pass a benefit-cost test. Second, the cost effectiveness of these interventions does
not appear to be attractive relative to interventions studied by other researchers for pro-
moting water conservation. Third, a benefit-cost analysis based on greenhouse gas bene-
fits alone, and a comparable MVPF analysis, do not appear to suggest that the investment
would increase net economic benefits. However, without information on the full range of
quantitatively significant benefits from water conservation, and the impact of our inter-
ventions over time, it is difficult to make informed statements about whether audits are
likely to pass different benefit-cost tests. Nonetheless, we do a bounding exercise that al-
lows us to estimate what other benefits would need to be for some of our interventions
to just pass a benefit-cost test. Details of our welfare analysis and additional sensitivity
analyses are presented in Section H.2.
3.4.1 Cost Effectiveness
We consider the cost effectiveness of this intervention and then compare it with other in-
terventions in the literature. Cost effectiveness is defined as cost divided by effectiveness.
We measure three categories of costs: the cost of sending letters, the lost producer surplus
associated with the decline in production, and the value of time in filling out the survey.
Effectiveness is measured by the per capita reduction in water consumption relative to the
case of no letter. Our base case is the Incentive £10 treatment. Dividing total cost by effec-
tiveness yields a cost effectiveness of £2.9 per cubic meter for the base case. Results are
presented in Table 3.6.
149
Table 3.6: Different Measures of Cost Effectiveness
Case Base Case
No Producer
Surplus Loss
Vanilla
Letter
Targeting
High Users
Two Year
Duration
Parameter (1) (2) (3) (4) (5)
Cost of Mailing 420 420 - 200 420
Producer Surplus Loss 200 - 290 200 2200
Time Cost 68 68 68 30 68
[A]: Total Cost (in£) 690 490 360 430 2,700
[B]: Effectiveness (in m
3
) 240 240 340 240 2,600
Cost Effectiveness ( £/m
3
) 2.9 2.1 1.0 1.8 1.0
Notes: This table shows how the cost effectiveness changes using different assumptions. Cost effectiveness is measured in terms of pounds
per cubic meter of water conserved in 2020 £. It is computed as the total cost divided by the effectiveness (A/B). See text for details on
the various cases.
The other four cases are variations on the base case. They lead to cost effectiveness
numbers that range between £1.0 and £2.9 per cubic meter. The first variation labeled
No Producer Surplus Loss sets producer surplus losses to zero. We present the case of No
Producer Surplus because many studies do not consider changes in producer surplus in
computing cost effectiveness. The second variation changes the benchmark for compari-
son from the control group to the Vanilla letter. We do this analysis because NWL planned
to send out this letter to their customers without our intervention. The third variation tar-
gets only high users, who are defined as users above the median consumption threshold.
This leads to an increase in the average reduction in consumption from 3.5 to 7.5 liters
per household per day. This suggests that targeting could be an important strategy for
improving cost effectiveness and increasing net benefits, which is consistent with other
studies (e.g., see Ferraro and M. K. Price 2013, Ferraro and Miranda 2013, and Brent, Lott,
M. Taylor, J. Cook, Rollins, and Stoddard 2020). The fourth variation considers the impact
of a change in duration of the persistence of the effects due to the intervention. This leads
to an improvement in cost effectiveness relative to the base case because the reduction in
water use increased by more than the producer surplus losses.
150
We compared our results on cost effectiveness with other studies.
26
We found that
cost effectiveness estimates vary over a large range, from $0.06 per cubic meter in Ferraro
and Miranda 2013 to$8.1 per cubic meter in Ansink, Ornaghi, and Tonin 2021. Our study
appears to fall in the mid-range of existing estimates. We found that the persistence of
the treatment effect is important for cost effectiveness, as can be seen from the difference
between the cost effectiveness numbers in Ferraro and M. K. Price 2013 and Bernedo, Fer-
raro, and M. Price 2014 ($0.17 versus $0.07 per cubic meter). Both studies analyze the
same field experiment, but the former assumes the effect lasts for four months, while the
latter estimates that effects are statistically detectable six years later.
3.4.2 Benefit-Cost Analysis
The previous section considered the cost effectiveness of our intervention. In principle,
one could do a full-blown benefit-cost analysis (BCA). We start with a simplified BCA,
and then consider a Marginal Value of Public Funds (MVPF) approach in the next section
that is more detailed. Our purpose in this section is to present a framework for a BCA
that allows us to ask one simple question: how large do other benefits ( i.e., those not
quantified in our analysis) need to be to just offset costs that we estimate? Other benefits
could include ecosystem benefits as well as reductions in investment costs (see discussion
below).
The benefits in our analysis result from greenhouse gas emission reductions associated
with a reduction in water consumption. Non-carbon greenhouse gas emissions have been
converted to CO
2
-equivalents for use in our analysis.
27
The carbon footprint numbers
for the water supply, use, and disposal system have been sourced from the Environment
Agency, a leading public body for the environment in England and Wales (Emma, Feiferi,
Fida, and Paul 2008).
26
See discussion in the appendix.
27
The contribution of different greenhouse gases to total water industry emissions are: carbon dioxide
(74 per cent), nitrous oxide (14 per cent), and methane (12 per cent) (Emma, Feiferi, Fida, and Paul 2008).
151
Table 3.7: Comparison of Cost-Effectiveness with Other Studies
Variables
Studies
CE Estimate
($/m
3
)
Intervention
CE Provided in
Initial Study
Quantity Estimated
(’000 m
3
)
Direct Costs
Estimated
Other Costs
Estimated
Ferraro and M. K. Price 2013 0.12 - 0.17 Social norm letter Yes 700 $1.2 per hh
Forgone Revenues:
$1.5mn to$1.6mn
Ferraro and Miranda 2013 0.06 - 0.11 Social norm letter Yes No $1.2 per hh No
Bennear, Lee, and L. O. Taylor 2013 2.2 - 7.6
Rebate on high efficiency
toilets
Yes 3.0 $170 per hh
Forgone Revenues:
$9,800
Bernedo, Ferraro, and M. Price 2014 0.07 Social norm letter Yes 1,700 $1.2 per hh No
Brent, J. H. Cook, and Olsen 2015 0.50 - 0.75
Social comparison letter
and home water report
Yes No
$11 per hh
per year
No
S. Datta, S. Datta, Josı, Zoratto, Calvo-Gonzı, Darling, et al. 2015 - Social comparison letter No 6.7 $440 -
Brent, Lott, M. Taylor, J. Cook, Rollins, and Stoddard 2020 - Social comparison messages No 27 - 37 No -
Ansink, Ornaghi, and Tonin 2021 3.8 - 8.1
Water audit with information
and technological component
No No
Information arm:
$38 per hh;
Technology arm:
$29 per hh
No
NWL Study 1.3 - 3.7 Online audit with£10 incentive Yes 0.24 $0.5 per hh $0.3 per hh
Notes: Notes: hh refers to household. Details of the calculation for this table are provided in tab:effectiveness ansinkandtab :cecalcotherstudies.Costeffectivenessismeasuredintermsofdollarspercubicmeterofwaterconservedin2020dollars.Quantityismeasuredintermsofthousandsofcubicmeters.FerraroAndM.K.
Price2013,FerraroAndMiranda2013,andBernedo,Ferraro,AndM.Price2014analyzethesamefieldexperimentandtherefore,thedirectcostestimatesarethesame.
152
To define benefits formally, we introduce some notation. Let ∆ g be the total change
in water consumption due to the intervention over the time period of our analysis. LetV
be the incremental greenhouse gas benefits that result from one cubic meter reduction in
water consumption
28
. The benefits from the intervention, B, are then−V ∆ g. The costs,
C, are given by the direct incremental costs (cost of mailing the letters and financial in-
centives provided to the households who complete the audit) of the experiment,E, and
any other losses in producer surplus that may result.
29
These losses can be represented
by the difference in the price of water, p, and the marginal cost of water,c (presumed to
be constant for simplicity), multiplied by the change in water consumption, ∆ g. That is,
the producer surplus losses are (p−c)∆ g. We can now estimate net benefits as follows:
Net benefits =B−C
=−V ∆ g + (p−c)∆ g−E (3.6)
Note that we have not included any measure of consumer surplus. This is because we
invoke the envelope theorem, and assume that consumers who switch are just as well off as
they were before.
30
If they are better off, then the measure of B−C is an underestimate of
net benefits.
31
We also do not include other benefits from water conservation, which may
be substantial, but for which we do not have an estimate. These include possible savings
from reduced capital and operating costs associated with expanded supply (Maddaus
2011). In addition, ecosystem services, such as habitat, biodiversity, fishing, recreation,
erosion protection, aesthetic value, and non-use values that can result from conservation
28
See Table 3.26 for a full breakdown of the greenhouse gas benefits based on different stages of water
supply and use.
29
These costs may be better approximated by the prices charged by a private sector firm for doing these
tasks. We consider a sensitivity on costs below to address this issue.
30
For a model that motivates our welfare equation based on nudge theory, see Allcott and Kessler 2019.
These authors assume a lump sum tax finances the nudge and quasi-linear utility for consumers. For an
application that includes externalities, see Akesson, Hahn and Metcalfe (draft).
31
We explore this issue in a sensitivity analysis below.
153
are not included (See Bishop and Weber 1996 for a more extended discussion on the impact
of demand reduction on water utilities and the environment).
Estimates for the various parameters in eq:4 are shown in Table 3.8 along with the de-
tailed results on net benefits. We perform the analysis using two different assumptions
about cost: a short-run marginal cost (SRMC) of £0.44 per cubic meter, and a long-run
marginal cost (LRMC) of£0.98 per cubic meter. The cost numbers were estimated based
on sources from NWL (NWL Financial Statements 2009, 2021). The SRMC, in our case,
is equivalent to the base operating expenditure per cubic meter of water, or the marginal
operating cost. It takes capacity as given, and includes costs associated with electricity for
water transport, storage and treatment, and abstraction charges by environmental agen-
cies.
32
LRMC, on the other hand, is the sum of marginal operating and marginal capacity
costs.
33
The LRMC was calculated based on the annualized cost of the last major water
resource investment undertaken by NWL – expanding Abberton reservoir in 2009 – and
equals£0.54 per cubic meter.
34
Before explaining the results, it is useful to highlight one key point. The price for water
in this application appears to exceed the estimated marginal social cost (MSC) based on
quantified benefits. The current price is £1.3 and the estimated MSC is the sum of marginal
private costs (either£0.44 if we use the SRMC or£0.98 if we use the LRMC) and marginal
external costs (£0.27). This gives an estimated MSC of either £0.71 if we use the SRMC
(£0.44 +£0.27) or£1.3 if we use the LRMC (£0.98 +£0.26). This observation implies that
any conservation measure, even if it had no costs attached, would not pass a narrowly
prescribed benefit-cost test because price already exceeds the estimated marginal social
32
Marsden Jacob 2004 state that for all practical purposes in the water industry, estimating SRMC by
reference to operating costs is reasonable. Moreover, conversations with NWL representatives suggested
that setting SRMC equal to the short-run average costs was a reasonable assumption.
33
Costs associated with investments as a result of an incremental increase in demand.
34
This is similar to the concept of long-run incremental cost in Mann, Saunders, and Warford 1980, and
includes both the capital costs associated with a change in capacity and volume sensitive costs. However,
in this case, it may be an underestimate because it does not appear to include investments in raw water
and wastewater treatment facilities, and water and sewer networks. Such costs could increase the LRMC
substantially, but NWL did not have an estimate.
154
cost. Stated another way, because price is greater than the estimated marginal social cost,
consumers may be consuming too little water relative to what might be viewed as eco-
nomically efficient. This, of course, assumes that the estimated marginal social cost is a
valid measure. Below, we argue that is unlikely to be the case in many instances because
other benefits associated with water conservation have not been quantified. This is the rea-
son we perform the bounding exercise contained in Table 3 to estimate what those other
benefits would need to be to just offset costs.
We consider three different cases for estimating net benefits associated with the SRMC
and the LRMC. The first uses the base case with the £10 Incentive, and it is compared to
the case of no letter. The second uses£15 Incentive intervention with the same comparison
group. The third uses the Vanilla letter as the benchmark with the £10 Incentive. We use
the two incentive treatments because those interventions are the ones that resulted in an
economically significant reduction in water consumption. The rationale for considering
a different benchmark is that NWL was going to send out the Vanilla letter anyway. V is
computed based on the Social Cost of Carbon (SCC), which is the monetary value of the
net harm to society associated with adding a small amount of carbon to the atmosphere
in a given year. We use an estimate of the SCC of $51 per metric ton of CO
2
(in 2020
dollars), which assumes a discount rate of 3 percent (Interagency Working Group, USG
2021). Below, we also consider how using different values for the SCC would affect the
benefit-cost analysis.
The table shows that the measured benefits fall short of the measured costs in all three
scenarios under both the cost structures. This may not be particularly surprising in light
of the fact that we are not quantifying many benefits. The measured benefits are slightly
higher, albeit still negative with the LRMC because the producer surplus loss due to re-
duced consumption is lower as a result of assuming the higher cost. The second to last
row of each panel in the table shows that other benefits would need to be between £2.5 to
£7.2 per cubic meter reduced for the total benefits to just offset the total costs. Though this
155
Table 3.8: Simple Benefit-Cost Analysis
Case Units
Base Case
(£10 Incentive)
£15 Incentive
Vanilla Letter
as Benchmark
Parameter (1) (2) (3)
V £/m
3
0.27 0.27 0.27
p £/m
3
1.3 1.3 1.3
∆ g m
3
-240 -290 -340
−V ∆ g £ 64 80 94
E £ 1,300 1,900 850
Panel A (SRMC)
c £/m
3
0.44 0.44 0.44
(p−c)∆ g £ -200 -250 -290
B−C (eq:4 above) £ -1,400 -2,100 -1,000
Breakeven Other Benefits =−(B−C) £ 1,400 2,100 1,000
Breakeven Other Benefits / ∆ g £/m
3
6.0 7.2 3.0
Breakeven Other Benefits / GHG benefits multiple 22 26 11
Panel B (LRMC)
c £/m
3
0.98 0.98 0.98
(p−c)∆ g £ -64 -80 -94
B−C (eq:4 above) £ -1,300 -1,900 -860
Breakeven Other Benefits =−(B−C) £ 1,300 1,900 860
Breakeven Other Benefits / ∆ g £/m
3
5.4 6.7 2.5
Breakeven Other Benefits / GHG Benefits multiple 20 24 9
Notes: We implement the equation for net benefits, eq:4. Panel A shows the results for short-run marginal costs (c=£0.44 per cubic meter). Panel B shows
the results for long-run marginal cost (c=£0.98 per cubic meter). See tab:welfareparametersfordetailsonparametersusedforwelfarecalculations.
gives us the required monetary value of other benefits needed to break even, we can also
calculate how much the other benefits have to be in relation to the greenhouse gas bene-
fits. The final row of both the panels show that other benefits would need to be 9 times (in
the case of LRMC; 11 times in the case of SRMC) as great as carbon emission benefits for
benefits to justify costs.
One could also ask how much the SCC would have to increase for benefits to just equal
costs when other benefits are excluded (or assumed to be zero). The answer is that in the
base case with LRMC, the SCC would need to increase by about 2,000 percent to $1,100
per ton, and to$1,200 per ton using the SRMC. These numbers are much higher than most
estimates for the SCC.
35
35
In the appendix, we explore a number of other sensitivities, including varying persistence, the relation-
ship between price and costs, and varying the LRMC.
156
3.4.3 MVPF framework
In this section we apply an MVPF approach to assessing benefits and costs. The core of the
MVPF approach is to consider the after-tax benefits to all groups in society from a small
change in expenditure on a particular intervention and compare that with the net cost to
the government (Hendren and Sprung-Keyser 2020, Finkelstein and Hendren 2020). In
general, the higher the benefits and the lower the net cost to the government, the more
attractive the intervention is, other things equal.
Define after-tax benefits as WTP or willingness to pay, and define G as the net cost to
the government. The measure of MVPF isWTP/G.
36
Table 3.9: MVPF Calculations
Case Base Case £15 Incentive Vanilla Letter
(£10 Incentive) as Benchmark
Parameter (1) (2) (3)
Panel A (SRMC)
Cost 0.44 0.44 0.44
WTP −0.076 −0.062 −0.17
G 1.0 1.0 1.1
MVPF =
WTP
G
−0.074 −0.060 −0.16
Panel B (LRMC)
c 0.98 0.98 0.98
WTP 0.0049 0.0040 0.011
G 1.01 1.01 1.02
MVPF =
WTP
G
0.0048 0.0039 0.010
Notes: This table computes the MVPF for the three scenarios described in Table 3.8 using Equation ??.
Panel A shows the results for the short-run marginal cost. Panel B shows the results for long-run marginal
cost. The values forV andp are the same as those in Table 3.8. See appendix for details.
Table 3.9 summarizes three MVPF calculations. It mirrors the net benefit calculations.
For the short-run marginal cost scenario, MVPF ranges from -0.16 to -0.062. The negative
sign here arises becauseWTP is negative and the net cost to the government is positive.
This analysis is similar to our benefit-cost analysis in that it suggests the investment is not
worth making unless other benefits not included here are significant. Using LRMC instead
36
See appendix for the implementation of the MVPF equation.
157
of SRMC increases the after-tax benefits due to a fall in producer surplus loss. The MVPF
is positive in this case, but still remains small in absolute terms.
3.5 Conclusion
Water suppliers and regulators are showing greater interest in assessing non-price mech-
anisms to encourage conservation as scarcity becomes more of an issue. One approach
that is being used is water audits, which offer customers recommendations on how they
could reduce their water consumption.
This paper explores how online water audits affect cost effectiveness and economic ef-
ficiency using a natural field experiment. We have three main findings. First, encouraging
subjects to participate in an online water audit with financial incentives reduces household
consumption by 43 liters per day, or about 17 percent. However, we also find that the size
of the financial incentive used to encourage conservation only matters marginally, with
£10 and£15 incentives having roughly the same effect. This finding suggests that it may
be worth testing lower levels of financial incentives in future experiments to see how such
incentives affect both water consumption and the willingness to take the audit. Second,
notwithstanding these improvements in water conservation, the intervention does not ap-
pear to pass a benefit-cost test that only includes the benefits of reducing greenhouse gas
emissions. Because our analysis does not quantify other potentially important benefits and
cost savings of conservation, such as ecosystem benefits and reductions in infrastructure
costs, we define a lower bound on other benefits needed for benefits to just offset costs. We
find that other benefits would need to be about 22 times as high as greenhouse gas benefits
for benefits to just offset costs.
37
Using a marginal value of public funds approach for mea-
suring benefits and costs yields similar conclusions. Third, we find that targeting of high
users could roughly double the effectiveness of interventions with financial incentives.
37
If the social cost of carbon increases over time, as much research suggests, such interventions could
become more attractive.
158
There are several areas for future research that we think could be fruitful. First, we
think it would be useful to develop better measures of the cost effectiveness and net ben-
efits associated with different kinds of interventions aimed at promoting water conser-
vation. Table 3.7, which reviews behavioral economics research in this area, reveals how
little we know about the cost effectiveness of different interventions. It would be useful
for decision makers in charge of water conservation to know something about the likely
costs and effectiveness of the group of interventions they are considering. The same is true
of net benefits. Very few studies using causal methods for estimating water conservation
have tried to address the net benefit question. We think using both a standard net benefit
framework as well as the MVPF framework could provide useful inputs to decision mak-
ing. Just as Hendren and Sprung-Keyser 2020 developed and compared several estimates
of MVPFs in the education and health areas, it could be useful to undertake a similar exer-
cise for water and energy interventions. Second, it would be very useful to try to quantify
some of the other benefits associated with water conservation in monetary terms, such as
the willingness to pay for greater reliability of supply. Related to that, it would be useful
to get better measures of the full marginal external cost of water consumption, and how
this varies over time and space (Hanemann et al. 2006, Garrick, Hall, Dobson, Damania,
Grafton, Hope, et al. 2017). Third, better information is needed on private costs, in partic-
ular the short-run and long-run marginal costs associated with water supply in different
regions, as these will also be critical in assessing the net benefits of conservation. Armed
with more accurate information on the marginal social cost and its relationship to price,
policy makers will be in a better position to design more equitable and efficient policies
that promote conservation when it is needed.
159
References
Arrow, Kenneth (1962). “Economic Welfare and the Allocation of Resources for Inven-
tion”. In: The rate and direction of inventive activity: Economic and social factors. Princeton
University Press, pp. 609–626.
Ashenfelter, Orley, Daniel Hosken, and Matthew Weinberg (2014). “Did Robert Bork Un-
derstate the Competitive Impact of Mergers? Evidence from Consummated Mergers”.
In: The Journal of Law and Economics 57.S3, S67–S100.
IPCC (2022). “IPCC, 2022: Summary for Policymakers”. In: Climate Change. Ed. by HO
P¨ ortner, DC Roberts, ES Poloczanska, K Mintenbeck, M Tignor, A Alegrıa, et al.
Mees, Heleen (2017). “Local Governments in the Driving Seat? A Comparative Analysis of
Public and Private Responsibilities for Adaptation to Climate Change in European and
North-American Cities”. In: Journal of Environmental Policy & Planning 19.4, pp. 374–390.
Valdivieso, Patricio, Krister P Andersson, and Benjamin Villena-Rold´ an (2017). “Institu-
tional Drivers of Adaptation in Local Government Decision-Making: Evidence from
Chile”. In: Climatic Change 143.1, pp. 157–171.
Oo, Aung Tun, Guido Van Huylenbroeck, and Stijn Speelman (2017). “Determining Fac-
tors for the Application of Climate Change Adaptation Strategies Among Farmers in
Magwe District, Dry Zone Region of Myanmar”. In: International Journal of Climate
Change Strategies and Management.
160
Kuruppu, Natasha and Reenate Willie (2015). “Barriers to Reducing Climate Enhanced
Disaster Risks in Least Developed Country-Small Islands Through Anticipatory Adap-
tation”. In: Weather and Climate Extremes 7, pp. 72–83.
Robinson, Stacy-ann (2018). “Climate Change Adaptation Limits in Small Island Devel-
oping States”. In: Limits to climate change adaptation. Springer, pp. 263–281.
Annan, Francis and Wolfram Schlenker (2015). “Federal Crop Insurance and the Disin-
centive to Adapt to Extreme Heat”. In: American Economic Review 105.5, pp. 262–66.
Chatterjee, Shoumitro (2019). “Market Power and Spatial Competition in Rural India”. In.
OECD (2021). “Methodologies to Measure Market Competition https://oe.cd/mmmc”.
In: OECD Competition Committee Issues Paper.
Reardon, Thomas (2015). “The Hidden Middle: The Quiet Revolution in the Midstream of
Agrifood Value Chains in Developing Countries”. In: Oxford Review of Economic Policy
31.1, pp. 45–63.
Donaldson, Dave and Richard Hornbeck (2016). “Railroads and American Economic Growth:
A “Market Access” Approach”. In: The Quarterly Journal of Economics 131.2, pp. 799–858.
Allen, Treb and David Atkin (2016). Volatility and the Gains from Trade. Tech. rep. National
Bureau of Economic Research.
Harris, Chauncy D (1954). “he Market as a Factor in the Localization of Industry in the
United States”. In: Annals of the association of American geographers 44.4, pp. 315–348.
Macchiavello, Rocco and Ameet Morjaria (2021). “Competition and relational contracts in
the Rwanda coffee chain”. In: The Quarterly Journal of Economics 136.2, pp. 1089–1143.
Census of India (2011). 2011 Census of India. Data Retrieved from Census Tables.
161
Ahmad, Latief, Raihana Habib Kanth, Sabah Parvaze, and Syed Sheraz Mahdi (2017).
“Agro-Climatic and Agro-Ecological Zones of India”. In: Experimental agrometeorology:
A practical manual. Springer, pp. 99–118.
Damodaran, Harish (2021). “Agriculture Policy Should Target India’s Actual Farming Pop-
ulation”. In: The Indian Express.
Narayanan, Sudha and Shree Saha (2021). “How Many Farmers are Really There in India?
A Critical Assessment of Nationally Representative Data”. In.
NABARD (2018). NABARD All India Rural Financial Inclusion Survey, 2016–17.
Krishnan, R, J Sanjay, Chellappan Gnanaseelan, Milind Mujumdar, Ashwini Kulkarni, and
Supriyo Chakraborty (2020). Assessment of Climate Change Over the Indian Region: A
Report of the Ministry of Earth Sciences (MOES), Government of India. Springer Nature.
Jones, Bryan, Claudia Tebaldi, Brian C O’Neill, Keith Oleson, and Jing Gao (2018). “Avoid-
ing Population Exposure to Heat-Related Extremes: Demographic Change vs Climate
Change”. In: Climatic change 146.3, pp. 423–437.
Rohini, P, M Rajeevan, and P Mukhopadhay (2019). “Future Projections of Heat Waves
over India from CMIP5 Models”. In: Climate dynamics 53.1, pp. 975–988.
Mishra, Vimal, Brian V Smoliak, Dennis P Lettenmaier, and John M Wallace (2012). “A
prominent Pattern of Year-to-Year Variability in Indian Summer Monsoon Rainfall”. In:
Proceedings of the National Academy of Sciences 109.19, pp. 7213–7217.
Turner, Andrew G and Hariharasubramanian Annamalai (2012). “Climate Change and
the South Asian Summer Monsoon”. In: Nature Climate Change 2.8, pp. 587–595.
Terraclimate (2018). “TerraClimate, A High-Resolution Global Dataset of Monthly Cli-
mate and Climatic Water Balance from 1958–2015”. In: Scientific data 5.1. Ed. by John T
162
Abatzoglou, Solomon Z Dobrowski, Sean A Parks, and Katherine C Hegewisch, pp. 1–
12.
India Meteorological Department (2009). “Development of a High Resolution Daily Grid-
ded Temperature Data Set (1969–2005) for the Indian Region”. In: Atmospheric Science
Letters 10.4. Ed. by AK Srivastava, M Rajeevan, and SR Kshirsagar, pp. 249–254.
Burke, Marshall and Kyle Emerick (2016). “Adaptation to Climate Change: Evidence from
US Agriculture”. In: American Economic Journal: Economic Policy 8.3, pp. 106–40.
Arag´ on, Fernando M, Francisco Oteiza, and Juan Pablo Rud (2021). “Climate Change and
Agriculture: Subsistence Farmers’ Response to Extreme Heat”. In: American Economic
Journal: Economic Policy 13.1, pp. 1–35.
Costinot, Arnaud, Dave Donaldson, and Cory Smith (2016). “Evolving Comparative Ad-
vantage and the Impact of Climate Change in Agricultural Markets: Evidence from 1.7
Million Fields Around the World”. In: Journal of Political Economy 124.1, pp. 205–248.
Dafny, Leemore, Mark Duggan, and Subramaniam Ramanarayanan (2012). “Paying a Pre-
mium on Your Premium? Consolidation in the US Health Insurance Industry”. In:
American Economic Review 102.2, pp. 1161–85.
Miller, Nathan H and Matthew C Weinberg (2017). “Understanding the Price Effects of
the MillerCoors Joint Venture”. In: Econometrica 85.6, pp. 1763–1791.
Comanor, William S and Robert H Smiley (1975). “Monopoly and the Distribution of
Wealth”. In: The Quarterly Journal of Economics, pp. 177–194.
Prager, Elena and Matt Schmitt (2021). “Employer Consolidation and Wages: Evidence
from Hospitals”. In: American Economic Review 111.2, pp. 397–427.
163
Aghion, Philippe, Christopher Harris, Peter Howitt, and John Vickers (2001). “Competi-
tion, Imitation and Growth with Step-by-Step Innovation”. In: The Review of Economic
Studies 68.3, pp. 467–492.
Aghion, Philippe, Nick Bloom, Richard Blundell, Rachel Griffith, and Peter Howitt (2005).
“Competition and Innovation: An Inverted-U Relationship”. In: The Quarterly Journal
of Economics 120.2, pp. 701–728.
Holmes, Thomas J, David K Levine, and James A Schmitz (2012). “Monopoly and the
Incentive to Innovate when Adoption Involves Switchover Disruptions”. In: American
Economic Journal: Microeconomics 4.3, pp. 1–33.
Becker, Gary S (1958). “Competition and Democracy”. In: The Journal of Law and Economics
1, pp. 105–109.
Robinson, James A and Daron Acemoglu (2012). Why Nations Fail: The Origins of Power,
Prosperity and Poverty. Profile London.
Holmes, Thomas J (1998). “The Effect of State Policies on the Location of Manufacturing:
Evidence from State Borders”. In: Journal of political Economy 106.4, pp. 667–705.
Besley, Timothy and Robin Burgess (2004). “Can Labor Regulation Hinder Economic Per-
formance? Evidence from India”. In: The Quarterly Journal of Economics 119.1, pp. 91–
134.
Djankov, Simeon, Rafael La Porta, Florencio Lopez-de-Silanes, and Andrei Shleifer (2002).
“The Regulation of Entry”. In: The quarterly Journal of economics 117.1, pp. 1–37.
Bertrand, Marianne and Francis Kramarz (2002). “Does Entry Regulation Hinder Job Cre-
ation? Evidence from the French Retail Industry”. In: The Quarterly Journal of Economics
117.4, pp. 1369–1413.
164
Blanchard, Olivier and Francesco Giavazzi (2003). “Macroeconomic Effects of Regulation
and Deregulation in Goods and Labor Markets”. In: The Quarterly journal of economics
118.3, pp. 879–907.
Fabrizio, Kira R, Nancy L Rose, and Catherine D Wolfram (2007). “Do markets Reduce
Costs? Assessing the Impact of Regulatory Restructuring on US Electric Generation
Efficiency”. In: American Economic Review 97.4, pp. 1250–1277.
Cicala, Steve (2022). “Imperfect Markets Versus Imperfect Regulation in US Electricity
Generation”. In: American Economic Review 112.2, pp. 409–41.
Burgess, Robin, Olivier Deschenes, Dave Donaldson, and Michael Greenstone (2017). “Weather,
Climate Change and Death in India”. In: University of Chicago.
Park, R Jisung, Joshua Goodman, Michael Hurwitz, and Jonathan Smith (2020). “Heat and
Learning”. In: American Economic Journal: Economic Policy 12.2, pp. 306–39.
Somanathan, Eswaran, Rohini Somanathan, Anant Sudarshan, and Meenu Tewari (2021).
“The Impact of Temperature on Productivity and Labor Supply: Evidence from Indian
Manufacturing”. In: Journal of Political Economy 129.6, pp. 1797–1827.
Guiteras, Raymond (2009). “The Impact of Climate Change on Indian Agriculture”. In:
Manuscript, Department of Economics, University of Maryland, College Park, Maryland.
Mall, Rajesh Kumar, Ranjeet Singh, Akhilesh Gupta, G Srinivasan, and LS Rathore (2006).
“Impact of Climate Change on Indian Agriculture: A Review”. In: Climatic change 78.2,
pp. 445–478.
Economic Survey of India (2018). Climate, Climate Change, and Agriculture.
165
Mendelsohn, Robert, William D Nordhaus, and Daigee Shaw (1994). “The Impact of Global
Warming on Agriculture: a Ricardian Analysis”. In: The American economic review, pp. 753–
771.
Schlenker, Wolfram, W Michael Hanemann, and Anthony C Fisher (2005). “Will US Agri-
culture Really Benefit from Global Warming? Accounting for Irrigation in the Hedonic
Approach”. In: American Economic Review 95.1, pp. 395–406.
Deschˆ enes, Olivier and Michael Greenstone (2007). “The Economic Impacts of Climate
Change: Evidence from Agricultural Output and Random Fluctuations in Weather”.
In: American economic review 97.1, pp. 354–385.
Fisher, Anthony C, W Michael Hanemann, Michael J Roberts, and Wolfram Schlenker
(2012). “The Economic Impacts of Climate Change: Evidence from Agricultural Out-
put and Random Fluctuations in Weather: Comment”. In: American Economic Review
102.7, pp. 3749–60.
Schlenker, Wolfram and Michael J Roberts (2009). “Nonlinear Temperature Effects Indi-
cate Severe Damages to US Crop Yields Under Climate Change”. In: Proceedings of the
National Academy of Sciences 106.37, pp. 15594–15598.
Mendelsohn, Robert (2009). “The Impact of Climate Change on Agriculture in Developing
Countries”. In: Journal of Natural Resources Policy Research 1.1, pp. 5–19.
Barreca, Alan, Karen Clay, Olivier Deschenes, Michael Greenstone, and Joseph S Shapiro
(2016). “Adapting to Climate Change: The Remarkable Decline in the US Temperature-
Mortality Relationship Over the Twentieth Century”. In: Journal of Political Economy
124.1, pp. 105–159.
166
Zivin, Joshua Graff and Matthew E Kahn (2016). Industrial Productivity in a Hotter World:
The Aggregate Implications of Heterogeneous Firm Investment in Air Conditioning. Tech. rep.
National Bureau of Economic Research.
Deschenes, Olivier and Enrico Moretti (2009). “Extreme Weather Events, Mortality, and
Migration”. In: The Review of Economics and Statistics 91.4, pp. 659–681.
Auffhammer, Maximilian and Tamma A Carleton (2018). “Regional Crop Diversity and
Weather Shocks in India”. In: Asian Development Review 35.2, pp. 113–130.
Taraz, Vis (2018). “Can Farmers Adapt to Higher Temperatures? Evidence from India”.
In: World Development 112, pp. 205–219.
Boucher, Stephen R, Michael R Carter, Jon Einar Flatnes, Travis J Lybbert, Jonathan G
Malacarne, Paswel Marenya, et al. (2021). Bundling Stress Tolerant Seeds and Insurance
for More Resilient and Productive Small-Scale Agriculture. Tech. rep. National Bureau of
Economic Research.
Feng, Shuaizhang, Michael Oppenheimer, and Wolfram Schlenker (2015). “Weather Anoma-
lies, Crop Yields, and Migration in the US Corn Belt”. In: NBER Working Paper w17734.
Hermans, Kathleen and Robert McLeman (2021). “Climate Change, Drought, Land Degra-
dation and Migration: Exploring the Linkages”. In: Current opinion in environmental sus-
tainability 50, pp. 236–244.
Reilly, John and Neil Hohmann (1993). “Climate Change and Agriculture: The Role of
International Trade”. In: The American Economic Review 83.2, pp. 306–312.
Randhir, Timothy O and Thomas W Hertel (2000). “Trade Liberalization as a Vehicle
for Adapting to Global Warming”. In: Agricultural and Resource Economics Review 29.2,
pp. 159–172.
167
Banerji, Abhijit and JV Meenakshi (2004). “Buyer Collusion and Efficiency of Government
Intervention in Wheat Markets in Northern India: An Asymmetric Structural Auctions
Analysis”. In: American journal of agricultural economics, pp. 236–253.
Meenakshi, JV and A Banerji (2005). “The Unsupportable Support Price: An Analysis of
Collusion and Government Intervention in Paddy Auction Markets in North India”.
In: Journal of Development Economics 76.2, pp. 377–403.
ICRISAT (2018). “Updated District Level Database Documentation of Files: 1966-67 to
2015-16 (Apportioned)”. In.
Sanghi, Apurva, KS Kavi Kumar, and James W McKinsey Jr (1998). “India Agriculture
and Climate Data Set”. In: World Bank.
Agricultural Statistics at a Glance (2020). “Ministry of Agriculture & Farmers Welfare”.
In: Department of Agriculture, Cooperation & Farmers Welfare, Directorate of Economics &
Statistics, Govt. of India.
ERA5-Land (2021). “ERA5-Land Hourly Data from 1950 to 1980”. In: Copernicus Climate
Change Service (C3S) Climate Data Store (CDS). Ed. by J Mu˜ noz-Sabater.
GFSAD30 (2017). “GFSAD NASA Making Earth System Data Records for Use in Research
Environments (MEASURES) Global Food Security-Support Analysis Data (GFSAD)
Cropland Extent 2015 South Asia, Afghanistan, Iran 30 m v001”. In: NASA EOSDIS
Land Processes DAAC. Ed. by MK Gumma, PS Thenkabail, P Teluguntla, A Oliphant,
J Xiong, RG Congalton, et al.
Ortiz-Bobea, Ariel (2021). “Climate, Agriculture and Food”. In: arXiv preprint arXiv:2105.12044.
168
Chimalwar, PJ, CM Tabhane, VK Verma, HP Singh, and GR Bhatia (2004). “Directory of
Wholesale Agricultural Produce Assembling Markets in India”. In: Directorate of Mar-
keting & Inspection, Govt. of India.
Singh, Madan Pal (2009). “Rice Productivity in India under Variable Climates”. In: Country
Report Presented at MARCO Symposium, pp. 5–7.
Krishnan, P, Balasubramanian Ramakrishnan, K Raja Reddy, and VR Reddy (2011). “High-
Temperature Effects on Rice Growth, Yield, and Grain Quality”. In: Advances in Agron-
omy 111, pp. 87–206.
S´ anchez, Berta, Anton Rasmussen, and John R Porter (2014). “Temperatures and the Growth
and Development of Maize and Rice: A Review”. In: Global Change Biology 20.2, pp. 408–
417.
Jenner, CF (1991). “Effects of Exposure of Wheat Ears to High Temperature on Dry Matter
Accumulation and Carbohydrate Metabolism in the Grain of Two Cultivars. I. Imme-
diate Responses”. In: Functional Plant Biology 18.2, pp. 165–177.
Wardlaw, IF (1974). “Temperature Control of Translocation”. In: Mechanisms of Regulation
Growth. Bull. 12, pp. 533–538.
McIntosh, Craig T and Wolfram Schlenker (2006). “Identifying Non-Linearities in Fixed
Effects Models”. In: UC-San Diego Working Paper.
M´ erel, Pierre and Matthew Gammans (2021). “Climate Econometrics: Can the Panel Ap-
proach Account for Long-Run Adaptation?” In: American Journal of Agricultural Eco-
nomics 103.4, pp. 1207–1238.
169
Cameron, A Colin, Jonah B Gelbach, and Douglas L Miller (2008). “Bootstrap-Based Im-
provements for Inference with Clustered Errors”. In: The Review of Economics and Statis-
tics 90.3, pp. 414–427.
Datta, Pritha, Bhagirath Behera, et al. (2022). “Climate Change and Indian Agriculture: A
Systematic Review of Farmers’ Perception, Adaptation, and Transformation”. In: Envi-
ronmental Challenges, p. 100543.
Chari, Mridula (2016). “How the Map of India was Redrawn on Lines of Language”. In:
Scroll.in.
Samaddar, Ranabir (2020). “Rule, Governmental Rationality and Reorganisation of States”.
In: Interrogating Reorganisation of States. Routledge India, pp. 48–65.
Parameswaran, MP and Srikumar Chattopadhyay (2014). “Reorganisation of States: A Dif-
ferent Approach”. In: Economic and Political Weekly, pp. 81–86.
Kumar, Akshat (2019). “A Review of the Issue of the Reorganization of the States in India”.
In: Think India Journal 22.1, pp. 280–286.
Taylor, J Edward and Irma Adelman (2003). “Agricultural Household Models: Genesis,
Evolution, and Extensions”. In: Review of Economics of the Household 1.1, pp. 33–58.
Ranganathan, Thiagu and Ghanshyam Pandey (2018). “Who Leaves Farmland Fallow and
Why? An Empirical Investigation using Nationally Representative Survey Data from
India”. In: The European Journal of Development Research 30.5, pp. 914–933.
Benjamin, Dwayne (1992). “Household Composition, Labor Markets, and Labor Demand:
Testing for Separation in Agricultural Household Models”. In: Econometrica: Journal of
the Econometric Society, pp. 287–322.
170
Rosenzweig, Mark R and Kenneth I Wolpin (1993). “Credit Market Constraints, Con-
sumption Smoothing, and the Accumulation of Durable Production Assets in Low-
Income Countries: Investments in Bullocks in India”. In: Journal of Political Economy
101.2, pp. 223–244.
Desai, Sonalde, Reeve Vanneman, and National Council of Applied Economic Research
(2005). India Human Development Survey (IHDS), 2005, ICPSR22626-v8. Ann Arbor, MI:
Inter-University Consortium for Political and Social Research [distributor], 2010-06-29.http:
//doi.org/10.3886/ICPSR22626.v8.
— (2012). India Human Development Survey-II (IHDS-II), 2011-12, Inter-University Consor-
tium for Political and Social Research [distributor], 2018-08-08. https://doi.org/10.
3886/ICPSR36151.v6.
De Janvry, Alain, Marcel Fafchamps, and Elisabeth Sadoulet (1991). “Peasant Household
Behaviour with Missing Markets: Some Paradoxes Explained”. In: The Economic Journal
101.409, pp. 1400–1417.
Chand, Ramesh (2012). “Development Policies and Agricultural Markets”. In: Economic
and Political Weekly, pp. 53–63.
Turner, Matt (2019). “The Monocentric Model in Discrete Space”. In.
Eaton, Jonathan and Samuel Kortum (2002). “Technology, Geography, and Trade”. In:
Econometrica 70.5, pp. 1741–1779.
Sotelo, Sebastian (2020). “Domestic trade frictions and agriculture”. In: Journal of Political
Economy 128.7, pp. 2690–2738.
Kahn, Matthew E. and Somik V. Lall (2021). “Enhancing Climate Change Resilience Through
Self-Protection, Public Investment, and Market Insurance”. In: The Brookings Institution.
171
Zavala, Lucas (2020). “Unfair Trade? Market Power in Agricultural Value Chains”. In: Yale
Uni.
Dhingra, Swati and Silvana Tenreyro (2020). “The Rise of Agribusiness and the Distribu-
tional Consequences of Policies on Intermediated Trade”. In.
Kousky, Carolyn, Erzo FP Luttmer, and Richard J Zeckhauser (2006). “Private Investment
and Government Protection”. In: Journal of Risk and Uncertainty 33.1, pp. 73–100.
Trenberth, Kevin E (2005). “The impact of climate change and variability on heavy pre-
cipitation, floods, and droughts”. In: Encyclopedia of hydrological sciences 17.
AghaKouchak, Amir, Felicia Chiang, Laurie S Huning, Charlotte A Love, Iman Mallakpour,
Omid Mazdiyasni, et al. (2020). “Climate extremes and compound hazards in a warm-
ing world”. In: Annual Review of Earth and Planetary Sciences 48, pp. 519–548.
Hsiang, Solomon M and Amir S Jina (2014). The causal effect of environmental catastrophe
on long-run economic growth: Evidence from 6,700 cyclones. Tech. rep. National Bureau of
Economic Research.
Elliott, Robert JR, Eric Strobl, and Puyang Sun (2015). “The local impact of typhoons on
economic activity in China: A view from outer space”. In: Journal of Urban Economics
88, pp. 50–66.
Kocornik-Mina, Adriana, Thomas KJ McDermott, Guy Michaels, and Ferdinand Rauch
(2020). “Flooded cities”. In: American Economic Journal: Applied Economics 12.2, pp. 35–
66.
Guha-Sapir, Debarati, Regina Below, and Philippe Hoyois (2016). “EM-DAT: the CRED/OFDA
international disaster database”. In.
172
Collado, Jos´ e Rafael N´ u˜ nez and Han-Hsiang Wang (2020). “Slum upgrading and cli-
mate change adaptation and mitigation: Lessons from Latin America”. In: Cities 104,
p. 102791.
Douglas, Ian, Kurshid Alam, Maryanne Maghenda, Yasmin Mcdonnell, Louise McLean,
and Jack Campbell (2008). “Unjust waters: climate change, flooding and the urban
poor in Africa”. In: Environment and urbanization 20.1, pp. 187–205.
Ehrlich, Isaac and Gary S Becker (1972). “Market insurance, self-insurance, and self-protection”.
In: Journal of Political Economy 80.4, pp. 623–648.
Henderson, Vernon, Adam Storeygard, and David N Weil (2011). “A bright idea for mea-
suring economic growth”. In: American Economic Review 101.3, pp. 194–99.
Donaldson, Dave and Adam Storeygard (2016). “The view from above: Applications of
satellite data in economics”. In: Journal of Economic Perspectives 30.4, pp. 171–98.
Guiteras, Raymond, Amir Jina, and A Mushfiq Mobarak (2015). “Satellites, self-reports,
and submersion: exposure to floods in Bangladesh”. In: American Economic Review 105.5,
pp. 232–36.
Desmet, Klaus and Esteban Rossi-Hansberg (2015). “On the spatial economic impact of
global warming”. In: Journal of Urban Economics 88, pp. 16–37.
Boustan, Leah Platt, Matthew E Kahn, and Paul W Rhode (2012). “Moving to higher
ground: Migration response to natural disasters in the early twentieth century”. In:
American Economic Review 102.3, pp. 238–44.
Peri, Giovanni and Akira Sasahara (2019). The impact of global warming on rural-Urban mi-
grations: Evidence from global big data. Tech. rep. National Bureau of Economic Research.
173
Hornbeck, Richard (2012). “The enduring impact of the American Dust Bowl: Short-and
long-run adjustments to environmental catastrophe”. In: American Economic Review
102.4, pp. 1477–1507.
Strobl, Eric (2011). “The economic growth impact of hurricanes: Evidence from US coastal
counties”. In: Review of Economics and Statistics 93.2, pp. 575–589.
Deryugina, Tatyana (2011). “The dynamic effects of hurricanes in the US: The role of non-
disaster transfer payments”. In.
Klein, Richard JT, Guy Midgley, Benjamin L Preston, Mozaharul Alam, Frans Berkhout,
Kirstin Dow, et al. (2015). “Adaptation opportunities, constraints, and limits”. In: Cli-
mate change 2014: Impacts, adaptation, and vulnerability. Part A: Global and sectoral aspects.
Contribution of working group II to the fifth assessment report of the intergovernmental panel
on climate change. Cambridge: Cambridge University Press., p. 899.
Bunten, Devin Michelle and Matthew E Kahn (2017). “Optimal real estate capital dura-
bility and localized climate change disaster risk”. In: Journal of Housing Economics 36,
pp. 1–7.
Hsiang, Solomon M and Daiju Narita (2012). “Adaptation to cyclone risk: Evidence from
the global cross-section”. In: Climate Change Economics 3.02, p. 1250011.
Mendelsohn, Robert, Wendy Morrison, Michael E Schlesinger, and Natalia G Andronova
(2000). “Country-specific market impacts of climate change”. In: Climatic change 45.3,
pp. 553–569.
World Bank, ADB, AfDB, BMZ, DFID, DGIS, et al. (2003). Poverty and Climate Change: Re-
ducing the Vulnerability of the Poor through Adaptation. Tech. rep. OECD.
174
Mendelsohn, Robert, Ariel Dinar, and Larry Williams (2006). “The distributional impact of
climate change on rich and poor countries”. In: Environment and development economics
11.2, pp. 159–178.
Stern, Nicholas (2006). “Stern Review: The economics of climate change”. In.
Tol, Richard SJ (2009). “The economic effects of climate change”. In: Journal of economic
perspectives 23.2, pp. 29–51.
Portner, Hans O, Debra C Roberts, Helen Adams, Carolina Adler, Paulina Aldunce, Elham
Ali, et al. (2022). “Climate change 2022: impacts, adaptation and vulnerability”. In.
Houghton, John Theodore, YDJG Ding, David J Griggs, Maria Noguer, Paul J van der Lin-
den, Xiaosu Dai, et al. (2001). Climate change 2001: the scientific basis: contribution of Work-
ing Group I to the third assessment report of the Intergovernmental Panel on Climate Change.
Cambridge university press.
Barrage, Lint (2019). “The Nobel Memorial Prize for William D. Nordhaus”. In: The Scan-
dinavian Journal of Economics 121.3, pp. 884–924.
Nordhaus, William (2019). “Climate change: The ultimate challenge for economics”. In:
American Economic Review 109.6, pp. 1991–2014.
Henderson, J Vernon (1974). “The sizes and types of cities”. In: American Economic Review,
pp. 640–656.
Roback, Jennifer (1982). “Wages, rents, and the quality of life”. In: Journal of Political Econ-
omy 90.6, pp. 1257–1278.
Glaeser, Edward L and Joseph Gyourko (2005). “Urban decline and durable housing”. In:
Journal of Political Economy 113.2, pp. 345–375.
175
Kahn, Matthew E (2005). “The death toll from natural disasters: the role of income, geog-
raphy, and institutions”. In: Review of Economics and Statistics 87.2, pp. 271–284.
Kellenberg, Derek K and Ahmed Mushfiq Mobarak (2008). “Does rising income increase
or decrease damage risk from natural disasters?” In: Journal of Urban Economics 63.3,
pp. 788–802.
M˚ ard, Johanna, Giuliano Di Baldassarre, and Maurizio Mazzoleni (2018). “Nighttime light
data reveal how flood protection shapes human proximity to rivers”. In: Science Ad-
vances 4.8, eaar5779.
Ferdous, Md Ruknul, Giuliano Di Baldassarre, Luigia Brandimarte, and Anna Wesselink
(2020). “The interplay between structural flood protection, population density, and
flood mortality along the Jamuna River, Bangladesh”. In: Regional Environmental Change
20.1, pp. 1–9.
Florczyk, AJ, Michele Melchiorri, Christina Corbane, Marcello Schiavina, M Maffenini, M
Pesaresi, et al. (2019). “Description of the GHS urban centre database 2015”. In: Public
Release.
Kummu, Matti, Maija Taka, and Joseph HA Guillaume (2018). “Gridded global datasets
for gross domestic product and Human Development Index over 1990–2015”. In: Sci-
entific data 5.1, pp. 1–15.
Henderson, J Vernon, Tim Squires, Adam Storeygard, and David Weil (2018). “The global
distribution of economic activity: nature, history, and the role of trade”. In: The Quar-
terly Journal of Economics 133.1, pp. 357–406.
Hsu, Feng-Chi, Kimberly E Baugh, Tilottama Ghosh, Mikhail Zhizhin, and Christopher
D Elvidge (2015). “DMSP-OLS radiance calibrated nighttime lights time series with
intercalibration”. In: Remote Sensing 7.2, pp. 1855–1876.
176
Chen, Xi and William D Nordhaus (2011). “Using luminosity data as a proxy for economic
statistics”. In: Proceedings of the National Academy of Sciences 108.21, pp. 8589–8594.
Elvidge, Christopher D, Kimberly Baugh, Mikhail Zhizhin, Feng Chi Hsu, and Tilottama
Ghosh (2017). “VIIRS night-time lights”. In: International Journal of Remote Sensing 38.21,
pp. 5860–5879.
Gibson, John, Susan Olivia, Geua Boe-Gibson, and Chao Li (2021). “Which night lights
data should we use in economics, and where?” In: Journal of Development Economics
149, p. 102602.
Beyer, Robert, Yingyao Hu, and Jiaxiong Yao (2022). “Measuring Quarterly Economic
Growth from Outer Space”. In.
Cavallo, Eduardo, Sebastian Galiani, Ilan Noy, and Juan Pantano (2013). “Catastrophic
natural disasters and economic growth”. In: Review of Economics and Statistics 95.5,
pp. 1549–1561.
Davis, Donald R and David E Weinstein (2002). “Bones, bombs, and break points: the
geography of economic activity”. In: American Economic Review 92.5, pp. 1269–1289.
Miguel, Edward and Gerard Roland (2011). “The long-run impact of bombing Vietnam”.
In: Journal of Development Economics 96.1, pp. 1–15.
Cattaneo, Cristina and Giovanni Peri (2016). “The migration response to increasing tem-
peratures”. In: Journal of development economics 122, pp. 127–146.
Deryugina, Tatyana, Laura Kawano, and Steven Levitt (2018). “The economic impact of
Hurricane Katrina on its victims: Evidence from individual tax returns”. In: American
Economic Journal: Applied Economics 10.2, pp. 202–33.
177
Nakamura, Emi, J´ osef Sigurdsson, and J´ on Steinsson (2016). The gift of moving: Intergenera-
tional consequences of a mobility shock. Tech. rep. National Bureau of Economic Research.
Glaeser, Edward L, David Laibson, and Bruce Sacerdote (2002). “An economic approach
to social capital”. In: The Economic Journal 112.483, F437–F458.
Chauvin, Juan Pablo, Edward Glaeser, Yueran Ma, and Kristina Tobio (2017). “What is
different about urbanization in rich and poor countries? Cities in Brazil, China, India
and the United States”. In: Journal of Urban Economics 98, pp. 17–49.
Pindyck, Robert S (2013). “Climate change policy: what do the models tell us?” In: Journal
of Economic Literature 51.3, pp. 860–72.
Roth, Jonathan, Pedro HC Sant’Anna, Alyssa Bilinski, and John Poe (2022). “What’s Trend-
ing in Difference-in-Differences? A Synthesis of the Recent Econometrics Literature”.
In: arXiv preprint arXiv:2201.01194.
De Chaisemartin, Cl´ ement and Xavier d’Haultfoeuille (2020). “Two-way fixed effects es-
timators with heterogeneous treatment effects”. In: American Economic Review 110.9,
pp. 2964–96.
Shi, Linda, Eric Chu, and Jessica Debats (2015). “Explaining progress in climate adapta-
tion planning across 156 US municipalities”. In: Journal of the American Planning Asso-
ciation 81.3, pp. 191–202.
Rentschler, Jun, Paolo Avner, Mattia Marconcini, Rui Su, Emanuele Strano, and Stephane
Hallegatte (2022). “Global evidence of rapid urban growth in flood zones since 1985”.
In.
Marx, Benjamin, Thomas Stoker, and Tavneet Suri (2013). “The economics of slums in the
developing world”. In: Journal of Economic Perspectives 27.4, pp. 187–210.
178
Vorosmarty, Charles J, Pamela Green, Joseph Salisbury, and Richard B Lammers (2000).
“Global water resources: vulnerability from climate change and population growth”.
In: science 289.5477, pp. 284–288.
Environmental Protection Agency, USA (2013). “Water audits and water loss control for
public water systems”. In: Environmental Protection Agency, Washington DC, USA.
Sturm, Reinhard, Katherine Gasner, and Lucy Andrews (2015). Water audits in the United
States: A review of water losses and data validity. Water Research Foundation.
Nauges, C´ eline and Dale Whittington (2019). “Social norms information treatments in
the municipal water supply sector: some new insights on benefits and costs”. In: Water
Economics and Policy 5.03, p. 1850026.
Hendren, Nathaniel and Ben Sprung-Keyser (2020). “A unified welfare analysis of gov-
ernment policies”. In: The Quarterly Journal of Economics 135.3, pp. 1209–1318.
Interagency Working Group, USG (2021). Technical support document: social cost of carbon,
methane, and nitrous oxide interim estimates under executive order 13990. Tech. rep. Tech.
rep., White House. URL https://www. whitehouse. gov/wp-content/uploads ...
US Bureau of Labor Statistics (2021). “All Urban Consumers (CPI-U)”. In: Washington,
DC: US Department of Labor, Bureau of Labor Statistics.
Finkelstein, Amy and Nathaniel Hendren (2020). “Welfare analysis meets causal infer-
ence”. In: Journal of Economic Perspectives 34.4, pp. 146–67.
Allcott, Hunt (2011). “Social norms and energy conservation”. In: Journal of Public Eco-
nomics 95.9-10, pp. 1082–1095.
179
Ayres, Ian, Sophie Raseman, and Alice Shih (2013). “Evidence from two large field exper-
iments that peer comparison feedback can reduce residential energy usage”. In: The
Journal of Law, Economics, and Organization 29.5, pp. 992–1022.
Ferraro, Paul J and Juan Jos´ e Miranda (2013). “Heterogeneous treatment effects and mech-
anisms in information-based environmental policies: Evidence from a large-scale field
experiment”. In: Resource and Energy Economics 35.3, pp. 356–379.
Brent, Daniel A, Joseph H Cook, and Skylar Olsen (2015). “Social comparisons, household
water use, and participation in utility conservation programs: Evidence from three ran-
domized trials”. In: Journal of the Association of Environmental and Resource Economists 2.4,
pp. 597–627.
Wichman, Casey J, Laura O Taylor, and Roger H Von Haefen (2016). “Conservation poli-
cies: who responds to price and who responds to prescription?” In: Journal of Environ-
mental Economics and Management 79, pp. 114–134.
Knittel, Christopher R and Samuel Stolper (2019). Using machine learning to target treatment:
The case of household energy use. Tech. rep. National Bureau of Economic Research.
Brent, Daniel A, Corey Lott, Michael Taylor, Joseph Cook, Kimberly Rollins, and Shawn
Stoddard (2020). “What Causes Heterogeneous Responses to Social Comparison Mes-
sages for Water Conservation?” In: Environmental and Resource Economics 77.3, pp. 503–
537.
Gerarden, Todd D and Muxi Yang (2021). “Using targeting to optimize program design:
evidence from an energy conservation experiment”. In: Energy, COVID, and Climate
Change, Online-June, pp. 7–9.
Ansink, Erik, Carmine Ornaghi, and Mirco Tonin (2021). “Technology vs information to
promote conservation: Evidence from water audits”. In.
180
Browne, Oliver R, Ludovica Gazze, and Michael Greenstone (2019). Determinants of water
conservation: Evidence from a California drought. Tech. rep. Working Paper.
West, Jeremy D, Robert W Fairlie, Bryan E Pratt, and Liam Rose (2021). Automated enforce-
ment of irrigation regulations and social pressure for water conservation. Tech. rep. National
Bureau of Economic Research.
Browne, Oliver, Ludovica Gazze, Michael Greenstone, and Olga Rostapshova (2019). En-
forcement and deterrence with certain detection: An experiment in water conservation policy.
Tech. rep. Working Paper.
Ferraro, Paul J and Michael K Price (2013). “Using nonpecuniary strategies to influence
behavior: evidence from a large-scale field experiment”. In: Review of Economics and
Statistics 95.1, pp. 64–73.
Jaime Torres, M´ onica Marcela and Fredrik Carlsson (2016). “Social Norms and Informa-
tion Diffusion in Water-saving Programs: Evidence from a Randomized Field Experi-
ment in Colombia”. In.
Datta, Saugato, Saugato Datta, Juan Josı, Laura Zoratto, Oscar Calvo-Gonzı, Matthew Dar-
ling, et al. (2015). A behavioral approach to water conservation: evidence from Costa Rica.
World Bank Washington, DC, USA.
Mansur, Erin T and Sheila M Olmstead (2012). “The value of scarce water: Measuring the
inefficiency of municipal regulations”. In: Journal of Urban Economics 71.3, pp. 332–346.
Bernedo, Maria, Paul J Ferraro, and Michael Price (2014). “The persistent impacts of norm-
based messaging and their implications for water conservation”. In: Journal of Consumer
Policy 37.3, pp. 437–452.
181
Jessoe, Katrina, Gabriel E Lade, Frank Loge, and Edward Spang (2021). “Spillovers from
behavioral interventions: Experimental evidence from water and energy use”. In: Jour-
nal of the Association of Environmental and Resource Economists 8.2, pp. 315–346.
Allcott, Hunt and Michael Greenstone (2017). Measuring the welfare effects of residential en-
ergy efficiency programs . Tech. rep. National Bureau of Economic Research.
Fowlie, Meredith, Michael Greenstone, and Catherine Wolfram (2018). “Do energy effi-
ciency investments deliver? Evidence from the weatherization assistance program”.
In: The Quarterly Journal of Economics 133.3, pp. 1597–1644.
Harrison, Glenn W and John A List (2004). “Field experiments”. In: Journal of Economic
literature 42.4, pp. 1009–1055.
Stuart, Elizabeth A, Stephen R Cole, Catherine P Bradshaw, and Philip J Leaf (2011). “The
use of propensity scores to assess the generalizability of results from randomized tri-
als”. In: Journal of the Royal Statistical Society: Series A (Statistics in Society) 174.2, pp. 369–
386.
Hahn, Robert W and Robert D Metcalfe (2021). “Efficiency and equity impacts of energy
subsidies”. In: American Economic Review 111.5, pp. 1658–88.
Bennear, Lori S, Jonathan M Lee, and Laura O Taylor (2013). “Municipal rebate programs
for environmental retrofits: An evaluation of additionality and cost-effectiveness”. In:
Journal of Policy Analysis and Management 32.2, pp. 350–372.
Emma, R, L Feiferi, C Fida, and R Paul (2008). “Greenhouse gas emissions of water sup-
ply and demand management options”. In: UK Environmental Agency Science Report,
SC070010.
182
Allcott, Hunt and Judd B Kessler (2019). “The welfare effects of nudges: A case study of
energy use social comparisons”. In: American Economic Journal: Applied Economics 11.1,
pp. 236–276.
Maddaus, William O (2011). “Realizing the benefits from water conservation”. In: Journal
of contemporary water research and education 114.1, p. 2.
Bishop, Daniel B and Jack A Weber (1996). Impacts of demand reduction on water utilities.
American Water Works Association.
NWL Financial Statements (2009). “Northumbrian Water Limited Annual Report and Fi-
nancial Statements”. In.
— (2021). “Northumbrian Water Limited Annual Report and Financial Statements”. In.
Marsden Jacob, Associates (2004). “Estimation of Long Run Marginal Cost (LRMC)”. In:
Queensland Competition Authority.
Mann, Patrick C, Robert J Saunders, and Jeremy J Warford (1980). “A note on capital indi-
visibility and the definition of marginal cost”. In: Water Resources Research 16.3, pp. 602–
604.
Hanemann, W Michael et al. (2006). “The economic conception of water”. In: Water Crisis:
myth or reality 61, pp. 74–76.
Garrick, Dustin E, Jim W Hall, Andrew Dobson, Richard Damania, R Quentin Grafton,
Robert Hope, et al. (2017). “Valuing water for sustainable development”. In: Science
358.6366, pp. 1003–1005.
NWL Charges Scheme (2020). “Northumbrian Water Limited Household Charges Scheme”.
In.
183
HM Government (2019). Environmental reporting guidelines: including streamlined energy and
carbon reporting guidance.
Royal Mail (2021). Price Guide: Advertising Mail (including Sustainable), Business Mail, Pub-
lishing Mail, Subscription Mail, Door to Door, Response Services and Inbound Mail Services.
data retrieved from Royal Mail,https://www.royalmail.com/sites/royalmail.com/
files/2020-12/Marketing_services_and_general_correspondence_price_guide_
1_January_2021_v2.pdf.
ASHE, UK (2021). Annual Survey of Hours and Earnings. data retrieved from Office of Na-
tional Statistics,https://www.ons.gov.uk/employmentandlabourmarket/peopleinwork/
earningsandworkinghours/datasets/ashe1997to2015selectedestimates.
White, Vinn (2016). “Revised departmental guidance on valuation of travel time in eco-
nomic analysis”. In: Office of the Secretary of Transportation, US Department of Transporta-
tion, Available at: transportation. gov/sites/dot. gov/files/docs/2016% 20Revised% 20Value%
20of% 20Travel% 20Time% 20Guidance. pdf.
Exchange Rates UK (2020).https://www.exchangerates.org.uk/. Accessed: 2022-04-30.
184
Appendices
F Appendix for Chapter 1
F.1 Appendix Figures
Figure 3.1: Fraction of Cropland in each ECMWF (Weather) Gridcell
Note: The cropland shares are computed based on 30m land cover data from the Global Food Security-support Analysis Data 2015
GFSAD30 2017.
185
(a) Districts in Sample for Analysing Effect of
Extreme Heat
(b) States in Sample for Analysing Effect of
Competition
Figure 3.2: ICRISAT Districts and APMC States in Sample
Note: Panel (a) shows the 313 districts (filled in goldenrod color) covered in ICRISAT 2018’s Apportioned database, which provides
yields for 25 major crops district-wise from 1966-2017. The district boundaries pertain to the year 1966. There were a total of 349
districts in 1966. Therefore, the 36 districts not included in the ICRISAT 2018 database are filled in grey. Panel (b) shows the 19 states
(filled in goldenrod color) which constitute the sample used to analyse the effect of competition on various economic outcomes. These
states include 2,938 wholesale intermediary Mandis geolocated within their boundaries, which forms our final sample of markets.
The state boundaries pertain to the year 2020. 9 States and 8 Union Territories not included in the sample are filled in grey.
186
(a) Effect of Extreme Heat on Yields: Kharif (b) Effect of Extreme Heat on Yields: Rabi
(c) Growing Season Distribution: Kharif (d) Growing Season Distribution: Rabi
(e) Map of Extreme Heat Exposure: Kharif (f) Map of Extreme Heat Exposure: Rabi
Figure 3.3: Coefficient Plot, GDD Distribution, and Extreme Heat Exposure by Season
187
Dependent Variable: log(Yields)
cdsy
Kharif Rabi
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Bin<15
dsy
0.010
∗∗∗
0.002 0.011 0.004 0.004 0.002 0.004
∗
0.002 0.004
∗∗∗
0.003
∗
(0.003) (0.006) (0.007) (0.003) (0.004) (0.001) (0.002) (0.003) (0.001) (0.001)
Bin 15-20
dsy
0.007
∗∗
0.000 0.008 0.004 0.003
(0.003) (0.004) (0.005) (0.003) (0.003)
Bin 20-25
dsy
−0.002 −0.001 −0.004 0.000 0.001
(0.002) (0.001) (0.003) (0.001) (0.002)
Bin 25-30
dsy
−0.001 −0.003
∗∗
−0.002 −0.004
∗∗∗
−0.004
∗∗
−0.001 −0.002 −0.002 −0.002
∗
−0.001
(0.002) (0.001) (0.003) (0.001) (0.001) (0.001) (0.001) (0.002) (0.001) (0.001)
Bin 30-35
dsy
−0.005
∗∗
−0.005
∗∗∗
−0.005 −0.006
∗∗∗
−0.006
∗∗∗
−0.009
∗∗∗
−0.005
∗∗
−0.009
∗∗∗
−0.006
∗∗∗
−0.006
∗∗∗
(0.003) (0.002) (0.003) (0.001) (0.002) (0.002) (0.002) (0.002) (0.001) (0.001)
Bin>35
dsy
−0.011
∗∗∗
−0.011
∗∗∗
−0.011
∗∗
−0.011
∗∗∗
−0.011
∗∗∗
−0.017
∗∗∗
−0.014
∗∗
−0.018
∗∗∗
−0.014
∗∗∗
−0.014
∗∗∗
(0.004) (0.002) (0.004) (0.002) (0.002) (0.004) (0.004) (0.004) (0.004) (0.004)
Fixed Effects
District ! ! ! ! ! !
Crop× Year ! ! ! ! ! ! ! ! ! !
Crop× State ! !
State× Year ! ! ! ! ! !
District× Decade ! !
District× Crop× Decade ! !
State Time-Trend ! !
Num. obs. 125,279 125,279 125,279 125,279 125,279 60,429 60,429 60,429 60,429 60,429
Adj. R
2
0.599 0.622 0.701 0.628 0.836 0.724 0.742 0.754 0.754 0.871
Table 3.10: Effect of Temperature on Yields (Panel Approach): Robustness Tests
Notes: two-way clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
Dependent Variable: log(Yields)
cds
Cross Section 2-Period Panel 3-Period Panel
Kharif Rabi Kharif Rabi Kharif Rabi
Bin<15
dst
−0.038
∗
−0.006 −0.007 −0.019
∗
0.036 0.006
(0.021) (0.018) (0.018) (0.010) (0.027) (0.015)
Bin 15-20
dst
−0.003 0.016 0.027
(0.019) (0.012) (0.022)
Bin 20-25
dst
−0.005 −0.039
∗∗∗
−0.015
(0.012) (0.009) (0.009)
Bin 25-30
dst
−0.001 −0.008 −0.006 −0.035
∗∗∗
−0.018
∗∗∗
0.001
(0.008) (0.016) (0.005) (0.010) (0.005) (0.012)
Bin 30-35
dst
−0.016
∗∗
−0.011 −0.017
∗∗∗
−0.031
∗∗∗
−0.021
∗∗∗
0.000
(0.008) (0.014) (0.006) (0.007) (0.005) (0.008)
Bin>35
dst
−0.017
∗
−0.034
∗
−0.019
∗∗∗
−0.037
∗∗∗
−0.044
∗∗∗
−0.050
∗∗∗
(0.009) (0.019) (0.006) (0.013) (0.006) (0.017)
Time Period
Period 1 1990-2015 1990-2015 1970-1990 1970-1990 1970-1980 1970-1980
Period 2 1995-2015 1995-2015 1985-1995 1985-1995
Period 3 2000-2015 2000-2015
Fixed Effects
Crop× State ! ! ! ! ! !
Time Period ! ! ! !
Num. obs. 2,636 1,267 4,877 2,397 7,283 3,547
Adj. R
2
0.510 0.382 0.219 0.227 0.185 0.121
Table 3.11: Effect of Temperature on Yields (Long-Differences): Robustness Tests
Notes: clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
188
Dependent Variable: log(Yields)
cdsy
(1) (2) (3) (4) (5) (6)
Bin 30-35
dsy
−0.004
∗∗
−0.004 −0.011
∗∗∗
−0.002 −0.014
∗∗∗
−0.010
∗∗∗
(0.001) (0.003) (0.003) (0.002) (0.003) (0.003)
Bin>35
dsy
−0.022
∗∗∗
−0.020
∗∗∗
−0.011
∗
−0.029
∗∗∗
−0.015
∗∗
−0.014
∗∗∗
(0.004) (0.005) (0.006) (0.006) (0.006) (0.004)
Bin<15
dsy
× Comp’
ds
−0.000 −0.000 −0.000 −0.000 −0.000 −0.000
(0.000) (0.001) (0.000) (0.001) (0.000) (0.000)
Bin 15-20
dsy
× Comp’
ds
0.000 0.001 0.000 0.000 0.000 −0.000
(0.000) (0.001) (0.000) (0.000) (0.000) (0.000)
Bin 25-30
dsy
× Comp’
ds
−0.000 −0.000 0.001 −0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Bin 30-35
dsy
× Comp’
ds
−0.000 −0.000 0.000 −0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)
Bin>35
dsy
× Comp’
ds
0.000 0.000 0.001 0.001 0.000 0.000
(0.001) (0.000) (0.000) (0.001) (0.000) (0.000)
Fixed Effects
Crop !
District ! !
Year ! !
Crop× District !
Crop× Year ! ! ! !
District× Year !
State× Year !
District× Crop× Decade ! !
Num. obs. 59,593 59,593 59,593 59,593 59,593 59,593
Adj. R
2
0.623 0.614 0.805 0.635 0.829 0.844
Table 3.12: Effect of Out-of-State Competition on Mitigation of Climate Shocks
Notes: clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
F.2 Appendix Tables
F.3 Appendix Derivations
F.3.1 Joint Distribution of TFP and Labor Intensity
Assume that the total factor productivity (TFP) of parcelω in field f if allocated to cropk
in statei at timet,A
fk
it
(ω)≥ 0, is Fr´ echet distributed with
Pr[A
fk
it
(ω)≤a
k
] = exp

−

a
k
/s
k

−θ

∀k∈K (F.1)
whereθ> 0 is a shape parameter, ands> 0 is the scale parameter. DenoteE
h
A
fk
it
(ω)
i
=
A
fk
it
, which is given by
189
Dependent Variable: log(Yields)
cdsy
Comp
2m
Comp
3ds
(1) (2) (3) (4) (5) (6) (7) (8)
Bin 30-35
dsy
−0.002 −0.002 −0.001 −0.006
∗
−0.002 −0.003 −0.001 −0.011
∗∗∗
(0.003) (0.004) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003)
Bin>35
dsy
−0.038
∗∗∗
−0.037
∗∗∗
−0.043
∗∗∗
−0.023
∗∗
−0.034
∗∗∗
−0.032
∗∗∗
−0.040
∗∗∗
−0.021
∗∗
(0.010) (0.006) (0.009) (0.009) (0.007) (0.007) (0.006) (0.007)
Bin<15
dsy
× Comp
ds
−0.004 −0.003 −0.001 −0.005 −2.752 −2.399 −0.874 −0.723
(0.003) (0.003) (0.003) (0.004) (2.416) (2.394) (2.847) (2.055)
Bin 15-20
dsy
× Comp
ds
0.005 0.005 0.004 0.000 3.383 2.851 3.571 1.456
(0.004) (0.003) (0.004) (0.002) (2.918) (2.954) (3.692) (2.545)
Bin 25-30
dsy
× Comp
ds
−0.002 −0.001 −0.001 −0.003 −1.209 −1.327 0.683 −3.176
(0.003) (0.003) (0.003) (0.003) (2.781) (2.817) (3.638) (2.189)
Bin 30-35
dsy
× Comp
ds
−0.003 −0.004 −0.003 −0.004 −2.321 −2.544 −2.490 −0.957
(0.004) (0.004) (0.003) (0.003) (1.957) (1.971) (2.401) (1.829)
Bin>35
dsy
× Comp
ds
0.014
∗∗
0.015
∗∗∗
0.015
∗∗∗
0.007
∗
9.973
∗∗∗
9.926
∗∗∗
11.719
∗∗∗
5.048
∗
(0.005) (0.004) (0.005) (0.004) (2.733) (2.752) (3.185) (2.843)
Fixed Effects
Crop ! !
District ! ! ! !
Year ! !
Crop× Year ! ! ! ! ! !
District× Year ! !
District× Crop× Decade ! !
State Time-Trend ! ! ! ! ! !
Effect Mitigated (in %) 24.5 26.2 23.2 19.6 22.3 23.6 21.9 18.2
Num. obs. 59,783 59,783 59,783 59,783 59,783 59,783 59,783 59,783
Adj. R
2
0.627 0.618 0.637 0.831 0.626 0.617 0.636 0.831
Table 3.13: Competition and Mitigation of Climate Shocks: Robustness Tests
Notes: clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
A
fk
it
=s
k
Γ((θ− 1)/θ) forθ> 1,∀k∈K
where Γ(·) denotes the Gamma function, i.e. Γ(t) =
+∞
R
0
u
t−1
exp(−u)du for anyt> 0.
Using the above definition, and setting γ≡ Γ((θ− 1)/θ)
−θ
, F.1 becomes
Pr[A
fk
it
(ω)≤a
k
] = exp

−γ

a
k
/A
fk
it

−θ

∀k∈K (F.2)
Also, assume labor intensity, ν
f
i
(ω), which is constant across crops and time, is dis-
tributed Fr´ echet such that
Pr[ν
f
i
(ω)≤ν] = exp
n
−γ (ν/ν
i
)
−θ
o
(F.3)
190
Dependent Variable: log(Arrivals)
cmdsy
(1) (2) (3) (4)
Bin 30-35
dsy
−0.009 −0.008 −0.013 −0.007
(0.008) (0.010) (0.012) (0.010)
Bin>35
dsy
−0.035
∗
−0.036
∗∗
−0.056
∗∗
−0.030
∗
(0.019) (0.016) (0.023) (0.015)
Bin<15
dsy
× Comp
mds
0.003 0.003 0.008 0.002
(0.004) (0.005) (0.006) (0.004)
Bin 15-20
dsy
× Comp
mds
0.012
∗
0.012 0.019
∗∗
0.009
(0.006) (0.007) (0.008) (0.007)
Bin 25-30
dsy
× Comp
mds
0.008
∗∗
0.008
∗∗
0.013
∗∗
0.006
(0.004) (0.004) (0.006) (0.004)
Bin 30-35
dsy
× Comp
mds
0.006
∗
0.006 0.010 0.004
(0.004) (0.004) (0.007) (0.004)
Bin>35
dsy
× Comp
mds
0.013
∗
0.013
∗∗
0.023
∗∗
0.010
∗
(0.007) (0.006) (0.010) (0.006)
Fixed Effects
Market ! !
Crop× Year ! ! ! !
District× Decade !
Market× Decade !
Market× Year !
State× Year ! ! !
Effect Mitigated (in %) 65.2 65.4 74.8 60.2
Num. obs. 156,724 156,724 156,724 156,724
Adj. R
2
0.434 0.451 0.449 0.429
Table 3.14: Competition and Mitigation of Climate Shocks—Arrivals:
Robustness Tests
Notes: clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
191
Dependent Variable: log(Arrivals)
cmdsy
(1) (2) (3) (4)
Bin 30-35
dsy
0.001 0.001 0.001 0.001
(0.006) (0.006) (0.006) (0.006)
Bin>35
dsy
−0.019 −0.019 −0.025
∗
−0.019
(0.012) (0.012) (0.014) (0.012)
Bin<15
dsy
× Comp’
mds
0.000 0.000 −0.000 0.000
(0.001) (0.001) (0.001) (0.001)
Bin 15-20
dsy
× Comp’
mds
0.000 0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000)
Bin 25-30
dsy
× Comp’
mds
−0.000 −0.000 −0.000 −0.000
(0.000) (0.000) (0.000) (0.000)
Bin 30-35
dsy
× Comp’
mds
−0.000 −0.000 −0.000 0.000
(0.000) (0.000) (0.001) (0.000)
Bin>35
dsy
× Comp’
mds
0.001 0.000 0.000 0.000
(0.001) (0.000) (0.001) (0.001)
Fixed Effects
Market ! !
Crop× Year ! ! ! !
District× Decade !
Market× Decade !
Market× Year !
State× Year ! ! !
Num. obs. 148,814 148,814 148,814 148,814
Adj. R
2
0.433 0.450 0.449 0.437
Table 3.15: Effect of Out-of-State Competition on Mitigation—Arrivals
Notes: clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
whereν
i
denotesE
h
ν
f
i
(ω)
i
. Given that TFP and labor intensity are independently drawn
for each (i,f,ω,t), and using F.2 and (F.3), the joint CDF can, therefore, be written as
Pr{A
f1
it
(ω)≤a
1
, ..., A
fK
it
(ω)≤a
K
,ν
f
i
(ω)≤ν}
=
Y
k∈K
exp

−γ

a
k
/A
fk
it

−θ

· exp
n
−γ (ν/ν
i
)
−θ
o
= exp



−γ


X
k∈K
(a
k
/A
fk
it
)
−θ
+ (ν/ν
i
)
−θ





F.3.2 Probability of Choosing Market
Derivation of probability of choosing marketm for cropk, 1.24 in text.
192
A farmer in statei chooses marketm∈M at timet if:
P
k
mit
Q
fk
mit
(ω)≥P
k
m
′
it
Q
fk
m
′
it
(ω)∀ m
′
∈M\{m} (F.4)
Our assumption of iceberg trade costs for farmers (1.20 in text) implies that Q
fk
mit
(ω) =
Q
fk
it
(ω)/τ
f
mt
. Using this, we can rewrite the condition above. Therefore, a farmer chooses
marketm∈M at timet if:
τ
f
mt
P
k
mit
= min







τ
f
1t
P
k
1it
, ..,
τ
f
mt
P
k
mit
, ...,
τ
f
Mt
P
k
Mit







(F.5)
Our assumption of Weibull distributed trade cost shocks (1.22), and the distribution’s
property of being closed under scale transformations implies:
τ
f
mt
P
k
mit
∼ Weibull




λ,
Υ
−1/λ
(1 +ζd
f
m
)
P
k
mit




LetG
f
mt
denote the c.d.f. of
τ
f
mt
P
k
mit
. Then:
G
f
mt
(ϵ) = Pr




τ
f
mt
P
k
mit
≤ϵ




= 1− exp


−Υϵ
λ

P
k
mit
1 +ζd
f
m
!
λ


193
The probability of choosing marketm for cropk can now be written as:
Ω fmk
it
= Pr




τ
f
mt
P
k
mit
≤ min
m
′







τ
f
m
′
t
P
k
m
′
it











= Pr




τ
f
mt
P
k
mit
≤ min
m
′
̸=m







τ
f
m
′
t
P
k
m
′
it











=
Z
∞
0
Y
m
′
̸=m

1−G
f
m
′
t
(ϵ)

dG
f
mt
(ϵ)
We can use the c.d.f.G
f
m
′
t
(ϵ) = 1−exp


−Υϵ
λ

P
k
m
′
it
1+ζd
f
m
′
!
λ


, and the corresponding p.d.f.
dG
f
mt
(ϵ) =λϵ
λ−1
Υ

P
k
mit
1+ζd
f
m

λ
exp

−Υϵ
λ

P
k
mit
1+ζd
f
m

λ
!
dϵ to get,
Ω fk
mit
=λΥ

P
k
mit
1 +ζd
f
m
!
λ
Z
∞
0
Y
m
′
exp


−Υϵ
λ


P
k
m
′
it
1 +ζd
f
m
′


λ


ϵ
λ−1
dϵ
=λΥ

P
k
mit
1 +ζd
f
m
!
λ
Z
∞
0
exp


−



X
m
′


P
k
m
′
it
1 +ζd
f
m
′


λ


Υϵ
λ


ϵ
λ−1
dϵ
=λΥ

P
k
mit
1 +ζd
f
m
!
λ








−exp


−


P
m
′

P
k
m
′
it
1+ζd
f
m
′
!
λ


Υϵ
λ


λΥ
P
m
′

P
k
m
′
it
1+ζd
f
m
′
!
λ








∞
0
=




P
k
mit
1 +ζ·d
f
m




λ
P
m
′
∈M




P
k
m
′
it
1 +ζ·d
f
m
′




λ
F.3.3 Profit Function of Farmer
Derivation of profits for a farmer growing crop k in farmf at timet, 1.26 in text.
194
Given the production function in 3.6, the profit for a farmer from parcel ω∈f in state
i who grows cropk at timet is given by:
π
fk
it
(ω) = (P
k
1it
A
fk
it
(ω)L
fk
it
(ω)−w
it
N
fk
it
(ω))· Ω fk
1it
+ ...
+ (P
k
Mit
A
fk
it
(ω)L
fk
it
(ω)−w
it
N
fk
it
(ω))· Ω fk
Mit
=


X
m
′
∈M
Ω fk
m
′
it
P
k
m
′
it


(A
fk
it
(ω)L
fk
it
(ω))−


X
m
′
∈M
Ω fk
m
′
it


| {z }
=1
(w
it
N
fk
it
(ω))
Using the expression for the probability of choosing a market (Ω fk
mit
) in 1.25, we can write
the above as:
π
fk
it
(ω) =A
fk
it
(ω)L
fk
it
(ω)














P
m
′
∈M
(P
k
m
′
it
)
λ+1

1 +ζd
f
m
′

λ
P
m
′
∈M




P
k
m
′
it
1 +ζd
f
m
′




λ














| {z }
=P
k
it
−w
it
N
fk
it
(ω) (F.6)
F.3.4 Land Allocation Problem
Derivation of probability that a parcelω of a field f located in statei is allocated to cropk at time
t, 1.28 in text.
Conditional on choosing to grow a crop, farmer in farmf and statei chooses cropk at
timet if:
π
fk
it
(ω)>π
fk
′
it
(ω) ∀ (k
′
̸=k)∈K
We can use the profit function in F.6 to write the above condition as:
A
fk
it
(ω)L
fk
it
(ω)P
k
it
− w
it
N
fk
it
(ω)>A
fk
′
it
(ω)L
fk
′
it
(ω)P
k
′
it
−w
it
N
fk
′
it
(ω) (F.7)
∀ (k
′
̸=k)∈K
195
The Leontief production function in 3.6 implies L
fk
it
(ω) =
N
fk
it
(ω)
ν
f
i
(ω)
∀k∈K. Also, once a
farmer decides to grow a crop, they will use the entire land area available since profits are
an increasing function of production inputs. Thus,L
fk
it
(ω) =L
fk
′
it
(ω)∀k
′
∈K. F.7 can then
be written as:
A
fk
it
(ω)P
k
it
>A
fk
′
it
(ω)P
k
′
it
∀ (k
′
̸=k)∈K (F.8)
The farmer in statei also has an outside option which entails working in statei’s outside
sector and producing the outside good. With labor productivity denoted byA
0
it
, and pro-
duction under constant returns to scale using only labor, the profit maximisation problem
of statei’s outside sector can be written as:
max
{N
0
it
}
π
0
it
=A
0
it
N
0
it
−w
it
N
0
it
Differentiating the above w.r.t. {N
0
it
}, we find that profit maximisation in state i’s outside
sector requiresw
it
=A
0
it
. Therefore, a farmer chooses to grow cropk over working in state
i’s outside sector if:
A
fk
it
(ω)P
k
it
>A
0
it
ν
f
i
(ω) (F.9)
Combining F.8 and F.9, we can deduce that a farmer in statei will grow cropk in parcel
ω∈f at timet if:
A
fk
it
(ω)P
k
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
1
it
, ...,A
fK
it
(ω)P
K
it
} (F.10)
F.3.5 Quantity Supplied to Market
Derivation of quantity of cropk supplied to marketm in statei at timet, 1.30 in text. LetQ
k
mit
denote the quantity of cropk supplied to marketm in statei at timet. Then
Q
k
mit
=
X
f∈F
i
Ω fk
mit
Z
1
0
Q
fk
it
(ω)dω (F.11)
196
Assume thatω∼U
[0,1]
. Thus, the probability density function ofω is:
f(ω) =









1 for 0≤ω≤ 1
0 for ω< 0 or ω> 1
Also, by law of iterated expectations,
E[Q
fk
it
(ω)] =E
k
[E[Q
fk
it
(ω)|A
fk
it
(ω)P
k
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
1
it
, ...,A
fK
it
(ω)P
K
it
}]]
F.11 can, therefore, be written as
Q
k
mit
=
X
f∈F
i
Ω fk
mit
∆ fk
it
E[Q
fk
it
(ω)|A
fk
it
(ω)P
k
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
1
it
, ...,A
fK
it
(ω)P
K
it
}] (F.12)
Furthermore, note that
E[L
fk
it
(ω)|A
fk
it
(ω)P
k
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
1
it
, ...,A
fK
it
(ω)P
K
it
}]
=E[L
fk
it
(ω)]
=
Z
1
0
L
fk
it
(ω)f(ω)dω
=s
f
i
Using (i) the production function in 3.6; (ii) the fact that conditional on choosing cropk,
A
fk
it
(ω)⊥L
fk
it
(ω), and; (iii) the previous expression, F.12 can be expressed as:
Q
k
mit
=
X
f∈F
i
s
f
i
Ω fk
mit
∆ fk
it
E[A
fk
it
(ω)|A
fk
it
(ω)P
k
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
1
it
, ...,A
fK
it
(ω)P
K
it
}] (F.13)
197
F.3.6 Average Conditional Productivity
Derivation of average productivity conditional on a crop being produced, 1.31 in text.
Using the definition of a c.d.f. and formula for conditional probability, we can write:
Pr{A
fk
it
(ω)≤a|A
fk
it
(ω)P
k
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
1
it
, ...,A
fK
it
(ω)P
K
it
}}
=
1
∆ fk
it
Pr{A
fk
it
(ω)≤a,A
0
it
ν
f
i
(ω)≤P
k
it
A
fk
it
(ω),P
l
it
A
fl
it
(ω)≤P
k
it
A
fk
it
(ω)∀l̸=k}
=
1
∆ fk
it
Pr







A
0
it
ν
f
i
(ω)
P
k
it
≤A
fk
it
(ω)≤a,
P
l
it
P
k
it
A
fl
it
(ω)≤A
fk
it
(ω)≤a∀l̸=k







Define A
fk
it
(ω) =χ as a Fr´ echet distributed random variable. Then,
Pr{A
fk
it
(ω)≤a|A
fk
it
(ω)P
k
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
1
it
, ...,A
fK
it
(ω)P
K
it
}}
=
1
∆ fk
it
Z
a
0
Pr







A
0
it
ν
f
i
(ω)
P
k
it
≤χ,
P
l
it
P
k
it
A
fl
it
(ω)≤χ∀l̸=k







f(χ)dχ
=
1
∆ fk
it
Z
a
0
Y
l̸=k
Pr







P
l
it
P
k
it
A
fl
it
(ω)≤χ







Pr







A
0
it
ν
f
i
(ω)
P
k
it
≤χ







f(χ)dχ (F.14)
Given the p.d.f. for Fr´ echet distributed TFP and labor intensity in 1.19, we can derive the
following c.d.f’s:
Pr

A
fk
it
(ω)P
k
it
≤x

= exp
(
−γ

x/A
fk
it
P
k
it

−θ
)
Pr

A
0
it
(ω)ν
f
i
(ω)≤ν

= exp

−γ
h
ν/A
0
it
ν
i
i
−θ

198
Using the above, F.14 can be written as
Pr{A
fk
it
(ω)≤a|A
fk
it
(ω)P
k
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
1
it
, ...,A
fK
it
(ω)P
K
it
}}
=
Z
a
0
exp







−γχ
−θ
(A
fk
it
)
θ
∆ fk
it







θγ
∆ fk
it
(χ)
−1−θ
(A
fk
it
)
θ
dχ
= exp









−




a
A
fk
it
(∆
fk
it
)
−1/θ
γ
1/θ




−θ









Thus, the c.d.f. is Fr´ echet distributed with shape parameterθ and scale parameter equiv-
alent toA
fk
it
(∆
fk
it
)
−1/θ
γ
1/θ
. Then,
E[A
fk
it
(ω)|A
fk
it
(ω)P
k
it
= max{A
0
it
ν
f
i
(ω),A
f1
it
(ω)P
1
it
, ...,A
fK
it
(ω)P
K
it
}
=A
fk
it
(∆
fk
it
)
−1/θ
γ
1/θ
Γ




1−
1
θ




=A
fk
it
× (∆
fk
it
)
−1/θ
where Γ(·) denotes the gamma function
F.3.7 Consumers Utility Maximisation
Derivation of representative consumers’ consumption of cropk, imported fromj toi, at timet; 1.36
in text.
199
Consumer solves the following maximisation problem:
max
n
C
it
,C
k
it
,C
k
jit
o
U
it
=C
0
it
+β
i
lnC
it
subject to
C
it
=


X
k∈K
(β
k
i
)
1/φ
(C
k
it
)
(φ−1)/φ


φ/(φ−1)
C
k
it
=


X
j∈I
(β
k
ji
)
1/σ
(C
k
jit
)
(σ−1)/σ


σ/(σ−1)
E
it
≥
X
k∈K
X
j∈I
h
P
rk
jit
C
k
jit
i
+C
0
it
P
rk
jit
= Ψ
k
ji
P
rk
jt
whereE
it
is household income for the representative consumer in statei at timet.
Setting up the Lagrangian and solving, we get
C
k
jit
= (β
i
)
σ
(C
it
)
(1−φ)σ/φ

C
k
it

(σ−φ)/φ

β
k
i

σ/φ
β
k
j
i

Ψ
k
ji
P
rk
jt

σ
(F.15)
Defining the CES price index associated with crop k in staten at timet as:
ˆ
P
rk
it
≡


X
n∈I
β
k
ni

Ψ
k
ni
P
rk
nt

1−σ


1/1−σ
(F.16)
Using F.15 and (F.16) in 3.5 gives us:
C
k
it
= (β
i
)
φ
β
k
i
(C
it
)
1−φ

ˆ
P
rk
it

φ
(F.17)
200
Substituting F.17 in 3.2 implies:
C
it
=β
i


X
k∈K
β
k
i

ˆ
P
rk
it

1−φ


1/(φ−1)
(F.18)
Finally, use F.16, (F.17) and (F.18) in F.15 to get:
C
k
jit
=β
i
β
k
i
(
ˆ
P
rk
it
)
1−φ
P
l∈K
β
l
i
(
ˆ
P
rl
it
)
1−φ
β
k
ji
(Ψ
k
ji
P
rk
jt
)
−σ
P
n∈I
β
k
ni
(Ψ
k
ni
P
rk
nt
)
1−σ
∀ i,j∈I,k∈K
201
Table 3.16: List of countries and share of cities
Country name # of
cities
% Country name # of
cities
% Country name # of
cities
%
Afghanistan 25 0.26 Gambia 4 0.04 North Macedonia 7 0.07
Albania 6 0.06 Georgia 5 0.05 Norway 4 0.04
Algeria 92 0.97 Germany 87 0.92 Oman 11 0.12
Angola 42 0.44 Ghana 48 0.51 Pakistan 168 1.77
Argentina 70 0.74 Greece 10 0.11 Palestine, State of 7 0.07
Armenia 3 0.03 Guatemala 39 0.41 Panama 6 0.06
Australia 27 0.29 Guinea 17 0.18 Papua New Guinea 8 0.08
Austria 6 0.06 Guinea-Bissau 3 0.03 Paraguay 8 0.08
Azerbaijan 15 0.16 Guyana 2 0.02 Peru 41 0.43
Bahamas 1 0.01 Haiti 21 0.22 Philippines 90 0.95
Bahrain 1 0.01 Honduras 13 0.14 Poland 46 0.49
Bangladesh 80 0.84 Hungary 11 0.12 Portugal 9 0.10
Barbados 1 0.01 Iceland 1 0.01 Puerto Rico 3 0.03
Belarus 14 0.15 India 1563 16.51 Qatar 3 0.03
Belgium 12 0.13 Indonesia 311 3.28 Romania 27 0.29
Belize 1 0.01 Iran 172 1.82 Russian Federation 204 2.15
Benin 20 0.21 Iraq 69 0.73 Rwanda 7 0.07
Bolivia 12 0.13 Ireland 5 0.05 Saudi Arabia 43 0.45
Bosnia & Herzegovina 5 0.05 Israel 9 0.1 Senegal 29 0.31
Botswana 7 0.07 Italy 91 0.96 Serbia 13 0.14
Brazil 347 3.66 Jamaica 4 0.04 Sierra Leone 9 0.10
Brunei Darussalam 1 0.01 Japan 109 1.15 Singapore 1 0.01
Bulgaria 7 0.07 Jordan 9 0.10 Slovakia 6 0.06
Burkina Faso 27 0.29 Kazakhstan 27 0.29 Slovenia 2 0.02
Burundi 10 0.11 Kenya 38 0.40 Solomon Islands 1 0.01
Cabo Verde 1 0.01 Korea DPR 76 0.80 Somalia 18 0.19
Cambodia 8 0.08 Korea Republic of 39 0.41 South Africa 77 0.81
Cameroon 45 0.48 Kosovo 7 0.07 South Sudan 13 0.14
Canada 48 0.51 Kuwait 4 0.04 Spain 72 0.76
Central African Republic 6 0.06 Kyrgyzstan 9 0.10 Sri Lanka 20 0.21
Chad 23 0.24 Lao 4 0.04 Sudan 52 0.55
Chile 33 0.35 Latvia 3 0.03 Suriname 1 0.01
China 1776 18.76 Lebanon 7 0.07 Sweden 12 0.13
Colombia 87 0.92 Lesotho 1 0.01 Switzerland 16 0.17
Comoros 2 0.02 Liberia 5 0.05 Syrian Arab Republic 24 0.25
Congo 114 1.20 Libya 15 0.16 Taiwan 21 0.22
Congo 4 0.04 Lithuania 6 0.06 Tajikistan 14 0.15
Costa Rica 3 0.03 Luxembourg 1 0.01 Tanzania 38 0.40
Cˆ ote d’Ivoire 35 0.37 Madagascar 6 0.06 Thailand 41 0.43
Croatia 6 0.06 Malawi 8 0.08 Timor-Leste 1 0.01
Cuba 19 0.20 Malaysia 36 0.38 Togo 13 0.14
Curac ¸ao 1 0.01 Mali 16 0.17 Trinidad and Tobago 4 0.04
Cyprus 3 0.03 Malta 1 0.01 Tunisia 26 0.27
Czechia 12 0.13 Mauritania 4 0.04 Turkey 129 1.36
Denmark 4 0.04 Mauritius 1 0.01 Turkmenistan 10 0.11
Djibouti 1 0.01 Mexico 157 1.66 Uganda 23 0.24
Dominican Republic 16 0.17 Moldova 5 0.05 Ukraine 78 0.82
Ecuador 30 0.32 Mongolia 1 0.01 United Arab Emirates 5 0.05
Egypt 182 1.92 Montenegro 1 0.01 United Kingdom 138 1.46
El Salvador 9 0.10 Morocco 59 0.62 Uruguay 6 0.06
Equatorial Guinea 2 0.02 Mozambique 38 0.40 USA 324 3.42
Eritrea 2 0.02 Myanmar 86 0.91 Uzbekistan 56 0.59
Estonia 2 0.02 Namibia 2 0.02 Venezuela 73 0.77
Eswatini 2 0.02 Nepal 9 0.10 Viet Nam 128 1.35
Ethiopia 86 0.91 Netherlands 37 0.39 Yemen 15 0.16
Fiji 1 0.01 New Zealand 8 0.08 Zambia 35 0.37
Finland 6 0.06 Nicaragua 14 0.15 Zimbabwe 19 0.20
France 76 0.80 Niger 23 0.24 Total 9468 100
Gabon 3 0.03 Nigeria 376 3.97
202
(a) Los Angeles map (b) Los Angeles GHS-UCDB extent
(c) Mumbai map (d) Mumbai GHS-UCDB extent
Notes: Figure 3.4a and 3.4c show the extent of Los Angeles region and Mumbai Metropolitan region. The
shaded region in Figure 3.4b and Figure 3.4d shows the polygons used to define the urban boundaries of
these cities in our analysis. Hence, our geographic coverage includes the core city and also covers a large
part of the suburban areas.
Figure 3.4: Urban Extent
203
Table 3.17: Summary Statistics: Dams
Variable Countries
China India Mexico United States
Dams
(1)
No Dams
(2)
Dams
(3)
No Dams
(4)
Dams
(5)
No Dams
(6)
Dams
(7)
No Dams
(8)
Panel A
Number of Cities
(% of Total)
1,043
(58.72)
733
(41.28)
967
(61.87)
596
(38.13)
61
(38.85)
96
(61.15)
216
(66.67)
108
(33.33)
Mean Elevation (m)
279.36
[421.26]
390.49
[569.15]
312.77
[268.43]
196.26
[227.60]
1187.82
[940.51]
1022.15
[863.07]
287.05
[395.34]
208.35
[294.17]
Night Lights (nW/cm2/sr)
8.97
[5.67]
10.05
[7.17]
6.37
[4.21]
6.03
[4.80]
26.56
[12.03]
23.72
[15.98]
29.14
[8.89]
29.74
[12.21]
GDP per capita (US$)
8,275.40
[3,393.87]
8,755.00
[5,333.17]
3,521.91
[2,042.96]
4,399.99
[3,583.61]
10,075.30
[4,307.90]
10,298.21
[11,502.46]
32,977.64
[4,610.90]
31,979.21
[4,839.27]
Population Growth (%)
3.27
[17.09]
5.70
[19.37]
14.50
[17.02]
14.43
[25.63]
31.44
[25.64]
33.76
[31.35]
18.21
[22.02]
23.67
[29.12]
Built-up Area (%)
38.60
[13.37]
39.99
[14.27]
19.13
[8.98]
19.73
[10.67]
52.57
[14.03]
49.29
[11.66]
66.24
[6.82]
62.99
[8.97]
Panel B
Number of Floods
(% of Total)
5,978
(66.70)
2,985
(33.30)
1,228
(50.74)
1,192
(49.26)
20
(52.63)
18
(47.37)
152
(66.38)
77
(33.62)
Average Floods
5.73
[4.92]
4.07
[4.27]
1.27
[1.41]
2.00
[2.13]
0.33
[0.79]
0.19
[0.47]
0.70
[1.09]
0.71
[1.15]
Extreme Precip. Events
(% of Total)
4,147
(63.15)
2,420
(36.85)
3,346
(61.93)
2,057
(38.07)
302
(39.58)
461
(60.42)
626
(61.61)
390
(38.39)
Avg. Extreme Precip. Events
3.98
[2.14]
3.30
[2.27]
3.46
[1.63]
3.45
[1.66]
4.95
[1.63]
4.80
[1.62]
2.90
[1.86]
3.61
[2.31]
Notes: This table provides summary statistics for cities with and without dams for four countries - China, India, Mexico, and the US. We classify the cities in each country based on whether they are protected
by a dam. A city is defined as protected by a dam if the river flowing though the city has a dam upstream of the city and the geodesic distance between the city, and the dam is less than or equal to 100
kilometers. Panel A provides information on geographical and economic characteristics, while Panel B focuses on floods and extreme precipitation events. GDP per capita is measured in PPP 2015 US$.
Population Growth is computed as the change in the population of a city between years 2015 and 2000. Built-up Area is defined as the percentage of the total area of the city (km
2
) that contains built-up
structures. Extreme Precipitation Events is a a dummy indicating whether the precipitation in the monthm and yeary in cityc in countryj (between the years 2012-2018) was greater than the 95
th
percentile
of the city-specific distribution of precipitation, which was created using data from 1958-2018. The Average Floods and Average Extreme Precipitation Events refer to the average number of such events during the
years 2012-2018. Standard deviation for all variables are reported in brackets.
G Appendix for Chapter 2
G.1 Appendix Figures
G.2 Appendix Tables
Table 3.19: Effect of Floods based on Productivity
Dependent Variable: ln(Night Lights
cjmy
)
All High Income Low Income
(1) (2) (3) (4) (5) (6)
Flood
cjmy
−0.089
∗∗∗
−0.078
∗∗
−0.043 −0.018 −0.161
∗∗∗
−0.118
∗∗∗
(0.031) (0.036) (0.033) (0.056) (0.057) (0.043)
Flood
cjmy
× Productivity
cj
0.012
∗∗
0.004 0.029
∗
(0.006) (0.006) (0.017)
Flood
cjmy
× Medium Productivity
cj
0.029 −0.016 0.036
(0.037) (0.056) (0.052)
Flood
cjmy
× High Productivity
cj
0.055 −0.005 0.112
∗
(0.037) (0.058) (0.060)
Fixed Effects
City ! ! ! ! ! !
Country× Month× Year ! ! ! ! ! !
Num. obs. 61,636 61,636 32,967 32,967 28,669 28,669
Adj. R
2
0.953 0.953 0.926 0.926 0.936 0.936
Notes: two-way clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
The dependent variable in all regressions,ln(Night Lights)cjmy , is the natural log of mean light intensity in cityc in countryj in monthm of yeary.Floodcjmy is a dummy indicating
whether cityc in countryj was hit by a flood in month m of yeary. Productivitycj refers to the average height of the city’s buildings within 2 kms of the highest point in the said
city. To create factor variables,Productivitycj was divided into three quartiles of equal size, with the omitted category in regressions being quartile 1, i.e. LowProductivity
cj
. All
regressions include the controlsStormcjmy andLandslidecjmy , dummies indicating whether cityc in countryj was hit by a storm or landslide, respectively, in monthm of yeary.
Two month leads and lags for all three disaster types have also been included as controls. Models (3) and (4) only include observations from High Income and Upper Middle Income
countries, whereas models (5) and (6) only include observations from Low Income and Lower Middle Income countries. Observations include all city-country-month-year observations
which had a non-zero value of nightlights. Each observation was weighted by the mean of the cloud free coverage for the city-country-month-year observation. Standard errors are
clustered at the city and month-year level.
204
Table 3.18: Effect of Floods on Economic Activity over Time
Dependent Variable: ln(Night Lights
cjmy
)
All High Income Low Income
(1) (2) (3) (4) (5) (6)
Flood
cjmy
−0.079
∗∗∗
−0.018 −0.200
∗∗∗
(0.017) (0.026) (0.040)
Flood
cjmy
× Time Trend
my
0.002
∗∗
0.000 0.004
∗∗
(0.001) (0.001) (0.002)
Extreme Rain
cjmy
−0.062
∗∗∗
−0.021
∗∗∗
−0.136
∗∗∗
(0.006) (0.007) (0.013)
Extreme Rain
cjmy
× Time Trend
my
0.001
∗∗∗
−0.001 0.004
∗∗∗
(0.000) (0.000) (0.001)
Fixed Effects
City ! ! ! ! ! !
Country× Month× Year ! ! ! ! ! !
Num. obs. 663,161 663,161 341,899 341,899 321,262 321,262
Adj. R
2
0.952 0.952 0.935 0.935 0.930 0.930
Notes: clustered robust standard errors in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
The dependent variable in all regressions,ln(Night Lights)cjmy , is the natural log of mean light intensity in cityc in countryj in monthm of yeary. Extreme Raincjmy is a dummy indicating
whether the precipitation in the monthm and yeary in cityc in countryj was greater than the 90 extsuperscriptth percentile of the city-specific distribution of precipitation, which was created
using data from 1958-2018. Time Trendmy is a continuous variable that takes the value from 1 to 81, with 1 representing the first month of the sample (April 2012), and 81 representing the last month
(September 2018). Models (1), (3) and (5) include the controlsStormcjmy andLandslidecjmy , dummies indicating whether cityc in countryj was hit by a storm or landslide, respectively, in
monthm of yeary. Two month leads and lags for all three disaster types have also been included as controls in the three models. Models (2), (4) and (6) include the two month lags and leads for
Extreme Rain. The contemporaneous disaster dummy has been interacted with the linear and quadratic time trend in all models. Models (3) and (4) only include observations from High Income
and Upper Middle Income countries, whereas models (5) and (6) only include observations from Low Income and Lower Middle Income countries. Observations include all city-country-month-year
observations which had a non-zero value of nightlights. Each observation was weighted by the mean of the cloud free coverage for the city-country-month-year observation. Standard errors are
clustered at the city level.
H Appendix for Chapter 3
The appendix for Chapter 3 is divided into four sections. Section H.1 provides balance ta-
bles, statistics on diagnostic completion, and additional results on heterogeneity of treat-
ment effects based on pre-treatment water consumption. We also compare the character-
istics of households who complete the diagnostic versus those who did not. The section
concludes by analyzing the interaction of households with the reminder emails and how
it differed across treatment arms. Section H.2 provides details on the welfare section. We
first report all the parameters and their sources, and subsequently present our calcula-
tions of the cost effectiveness of other studies in the literature. Section H.3 sheds light on
the measurement of pre- and post-treatment water consumption data using an illustrative
example. Finally, in Section H.4, we provide samples of the different letters and reminders
that were sent to households.
205
H.1 Baseline Balance and Additional Results
H.1.1 Balance Table
Our various treatments are balanced on pre-treatment covariates. Table 3.20 provides a
measure of the balance on observed covariates across different treatment groups. Column
(1) reports the number of people in each treatment group. Columns (2) to (4) provide the
percentage of population with a water meter, living in a rural area, and for whom NWL
had an email address, respectively. Column (5) reports balance on the number of con-
sumers for whom we had water data available. We also check for balance within the sub
sample of customers with meters as our LATE estimates only use metered households. In
this regard, columns (6) and (7) report the number of metered households living in rural
areas, and who provided NWL with an email id. The only significant differences (at 10
percent) are: a) Vanilla households have a lower probability (67 percent versus 70 percent
in the control group) of living in rural areas, and b) fewer customers (41 percent versus 44
percent in the control group) in the Incentives 10 group had registered their email ids with
NWL. We, therefore, control for these covariates in our regressions. Finally, columns (8)
and (9) provide balance on pre-treatment water consumption, and how many consumers
within each treatment group fell in the top 50
th
percentile of water consumption for the
entire sample.
We calculated the p-value on t-test of equality of means with control group, and the
same is reported in brackets. We find that the covariates for the treatment arms are not
significantly different from the covariates in the control group.
38
Column (8) reports the
p-values from F-tests of joint significance of all the regressors from an OLS regression
where the dependent variable is a dummy variable taking a value of 0 if the customer
is assigned to the Control group, and it takes a value of 1 for customers assigned to the
38
It is important to note that water consumption data was only available for metered customers and,
therefore, columns (6) to (9) pertain to the sub sample with meters.
206
Table 3.20: Baseline Balance Across Treatment Groups
Number of
Customers
Has a
Water Meter
Lives in
Rural Area
Provided
an Email
Water
Consumption
Data Available
Lives in
Rural Area
(Metered h/h)
Provided
an Email
(Metered h/h)
Pre-Diagnostic
Consumption
(m
3
/day)
Top 50%
Consumers
F-test of Joint
Significance
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
All Customers 44,757
.429
(.002)
.701
(.002)
.313
(.002)
.324
(.002)
0.685
(.003)
0.434
(0.004)
.258
(.003)
.499
(.004)
Control 7,459
.427
(.006)
.706
(.005)
.312
(.005)
.320
(.005)
0.695
(0.008)
0.441
(0.010)
.257
(.007)
.499
(.010)
Vanilla 7,460
.428
(.006)
[.887]
.696
(.005)
[.206]
.316
(.005)
[.638]
.327
(.005)
[.339]
0.674
(0.008)
[0.066]
0.433
(0.010)
[.581]
.266
(.007)
[.360]
.498
(.010)
[.921]
{.296}
Simplified 7,460
.428
(.006)
[.848]
.703
(.005)
[.697]
.313
(.005)
[.920]
.321
(.005)
[.852]
0.692
(0.008)
[.779]
0.435
(0.010)
[0.709]
.256
(.007)
[.855]
.511
(.010)
[.427]
{.501}
Altruism 7,460
.429
(.006)
[.797]
.699
(.005)
[.355]
.314
(.005)
[.809]
.324
(.005)
[.591]
0.685
(0.008)
[.386]
0.442
(0.010)
0.900
.253
(.007)
[.661]
.500
(.010)
[.966]
{.486}
Incentives 10 3,789
.436
(.008)
[.356]
.699
(.007)
[.488]
.320
(.008)
[.372]
.332
(.008)
[.189]
0.685
(0.011)
[.445]
0.409
(0.014)
[0.062]
.248
(.010)
[.451]
.482
(.014)
[.300]
{.252}
Incentives 15 3,670
.423
(.008)
[.670]
.698
(.008)
[.409]
.303
(.008)
[.313]
.324
(.008)
[.633]
0.678
(0.012)
[.219]
0.440
(0.014)
[0.985]
.254
(.010)
[.752]
.472
(.015)
[.121]
{.499}
Moral Cost 7,459
.430
(.006)
[.741]
.701
(.005)
[.519]
.309
(.005)
[.684]
.323
(.005)
[.700]
0.683
(0.008)
[.270]
0.432
(0.010)
[.529]
.263
(.007)
[.557]
.513
(.010)
[.373]
{.656}
F-test of Joint
Significance
{.958} {.916} {.733} {.883} {0.595} {0.581} {.698} {.229}
Notes: Robust standard errors from OLS regressions are in parenthesis.
p-value on t-test of equality of means with control group is in brackets; P-value on F-tests is in braces.
All data was provided by Northumbrian Water Limited. Column (1) reports the number of customers assigned to each treatment. Columns (2) to (5) report the mean value of each customer characteristic, derived from an OLS regression of the characteristic
of interest on a series of dummy variables for each treatment group. The excluded (comparison) group in these regressions is the control group. Robust standard errors are reported in parenthesis throughout. Columns (6) and (7) report balance on customer
characteristics for a specific sub sample: metered households. Columns (8) and (9) provide the balance on pre-treatment water consumption and number of high users, defined as households with pre-treatment water consumption greater than the median
of the metered sample. Column (8) reports the p-values from F-Tests of joint significance of all the regressors from an OLS regression where the dependent variable is a dummy variable taking value 0 if the customer is assigned to the control group, and
it takes value 1 for customers assigned to treatment groupj, and the independent variables are the variables in columns (2) to (4) and (8) to (9). The sample only includes observations for which we have consumption data available, and therefore, Water
Consumption Data Available (column (5)) is excluded as the same will always be 1 for every observation. The p-values reported in the last row are from the F-test of joint significance of the treatment dummies in each column regression where the sample
includes all customers, except for columns (6) to (9), in which case the sample only includes observations for which we have consumption data available.
207
treatment group in each respective row. A significant F-test would represent that covari-
ates can predict participation in a particular group, but all of them are insignificant. Fi-
nally, the p-values reported in the last row are from the F-test of joint significance of the
treatment dummies from an OLS regression where the dependent variable is the observ-
able covariate and the independent variables are dummies for different treatment groups.
A significant F-test would indicate that in at least one treatment group the mean of the
covariate is different than the others. Again, we fail to reject the null hypothesis that all
coefficients are 0.
Table 3.21 provides the raw data from the RCT on the number of households that com-
pleted the diagnostic. These figures are further broken down based on the number of
metered and unmetered households. As reported in column (4), the majority of house-
holds that completed the audit were metered, and this is consistent across all treatment
groups.
H.1.2 Heterogeneity Based on Pre-Treatment Water Consumption
Targeting households based on some pre-treatment covariates may be a more cost-effective
intervention for utilities if there is heterogeneity in response. We find that our letters have a
greater impact on high water users, and the result holds with the LATE estimate of impact
of the online audit.
To show that high-use households are more likely to be influenced by these interven-
tions, we run the following econometric model:
y
i
=α +
X
j
β
j
T
ij
+ϕ High-Use
i
+
X
j
η
j
High-Use
i
×T
ij
+γX
i
+ϵ
i
(H.1)
where, y
i
denotes post-treatment water consumption for household i, T
ij
is a dummy
that equals 1 if household i received treatment j, where j refers to the different treat-
ment groups. High-Use
i
is also a dummy and equals 1 if householdi had a pre-treatment
208
Table 3.21: Statistics on Diagnostic Completion and Metered Households
Number of Customers Completed Audit Metered Customers
Metered Customers who
Completed Audit
(% of Customers) (% of Customers) (% of Completed Audit)
(1) (2) (3) (4)
All Customers 44,757
1,287
(2.9)
19,180
(42.9)
860
(66.8)
Control 7,459
3
(0.0)
3,184
(42.7)
3
(100.0)
Vanilla 7,460
140
(1.9)
3,193
(42.8)
102
(72.9)
Simplified 7,460
189
(2.5)
3,196
(42.8)
133
(75.6)
Altruism 7,460
176
(2.4)
3,200
(42.9)
119
(67.6)
Incentives 10 3,789
242
(6.4)
1,652
(43.6)
136
(56.2)
Incentives 15 3,670
278
(7.6)
1,551
(42.3)
161
(57.9)
Moral Cost 7,459
259
(3.5)
3,204
(43.0)
206
(79.5)
Notes: All data was provided by Northumbrian Water Limited. Column (1) reports the number of customers assigned to each treatment group. Column (2) reports the
number of customers who completed the online diagnostic. Percentage of households which completed the audit relative to total number of households in the treatment
group are reported in parenthesis. Column (3) reports the number of customers who had a water meter installed in their homes. Percentage of metered households
relative to total number of households in the treatment group are reported in parenthesis. Column (4) reports the number of metered households who completed
the audit. Percentage of metered households which completed the audit relative to total number of households which completed the audit in the treatment group are
reported in parenthesis.
209
water consumption greater than the median of the sample. γ is a vector of estimates for
the different dummy controls, represented by X
i
, for householdi. These controls include
Rural
i
and Pre-Treatment Water Consumption
i
in liters per day. Finally,ϵ
i
is the error term.
If households with higher pre-treatment usage were incentivized more to conserve water,
we would expectη
j
to be negative and significant.
Table 3.22 presents the results. With reference to the control group, the interventions
had an additional significant negative impact of 4.9 liters per day on treated consumers
in the high usage category (column (1)). When we include individual dummies for dif-
ferent behavioral communications (column (2)), our findings suggest that the Simplified
and Incentive letters have a significantly higher impact on high-use households. This het-
erogeneity, however, does not persist when our reference group changes to Vanilla or Sim-
plified in columns (3)-(4), but the effect sizes are still negative and large. Thus, we do
find evidence of heterogeneous treatment effects based on water usage prior to treatment,
especially when we compare the interventions to the control group.
We also test for heterogeneity in our LATE estimates by running the following regres-
sion:
y
i
=α +βT
i
+ϕ High-Use
i
+η High-Use
i
×T
i
+γX
i
+ϵ
i
(H.2)
whereT
i
represents Completed Diagnostic, which is an indicator for whether the household
completed the audit or not. The coefficient of interest is η, which represents the additional
impact of completing the diagnostic on high users compared with low users.
As discussed in Section 3.3.3, we need to use IV’s for Completed Diagnostic, with the IV
for the interaction term, High-Use
i
×T
i
, just the IV for Completed Diagnostic
i
interacted with
High-Use
i
. The results are presented in Table 3.23.
For all specifications, the coefficient on Completed Diagnostic is negative but insignifi-
cant. Notably, the coefficient on the interaction term is negative and significant for our
first two specifications, and also much higher than the coefficients in Table 3.3 where we
do not distinguish between high- and low-use households. This implies that audits had
210
Table 3.22: Heterogeneous Treatment Effects Based on Pre-Treatment Usage
Post-Treatment Water Consumption
Control Control Vanilla Simplified
(1) (2) (3) (4)
High-Use 9.716
∗∗∗
9.729
∗∗∗
7.328
∗∗∗
5.361
∗
(2.702) (2.701) (2.822) (2.918)
High-Use× Treated −4.891
∗
(2.503)
High-Use× Vanilla −2.199
(3.196)
High-Use× Simplified −5.821
∗
−3.614
(3.220) (3.171)
High-Use× Altruism −5.223 −3.025 0.679
(3.223) (3.170) (3.194)
High-Use× Incentives 10 −7.863
∗∗
−5.655 −2.053
(3.920) (3.882) (3.899)
High-Use× Incentives 15 −7.398
∗
−5.195 −1.554
(3.977) (3.934) (3.958)
High-Use× Moral Cost −3.960 −1.760 1.913
(3.148) (3.092) (3.123)
Intercept 7.771
∗∗∗
7.770
∗∗∗
10.546
∗∗∗
10.879
∗∗∗
(1.664) (1.663) (1.739) (1.995)
Controls Yes Yes Yes Yes
Observations 11,700 11,700 9,770 7,795
Notes: Robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All regressions report the average treatment effect estimates of different behavioral interventions on post-treatment water con-
sumption (eq:8). The dependent variable for all models is Post-Treatment Water Consumption, a continuous variable that measures
the water consumption of a household, in liters per day, post the treatment date of 08-Dec-2018. Pre-treatment consumption and
post-treatment consumption were available for only a subset (30 per cent) of the households. Households with unreasonably large
differences between pre- and post-treatment consumption (absolute value greater than 50 per cent) were dropped from the sample.
The data was trimmed at 1 and 99 percentile of pre-treatment consumption. The model names reflect the reference group for each
regression. The regressor of interest, High-Use, is a dummy that equals 1 if the household had a pre-treatment water consumption
greater than the median of the sample. Treated is a dummy variable that equals 1 for all households who received any letter. The esti-
mates on Treated and the various treatment arms (Vanilla, Simplified, Altruism, Incentives 10, Incentives 15 , and Moral Cost) are omitted
from the table in the interest of space. Models (1) and (2) include all observations, with the control treatment arm constituting the
reference group. Model (3) excludes the observations in the control group, with the Vanilla letter comprising the reference group.
Model (4) excludes the observations in the control and Vanilla group, with the Simplified letter acting as the reference group. All
models include Rural and Pre-Treatment Consumption as controls. Rural is a dummy that equals 1 if the household is located in a
rural area. Pre-Treatment Consumption is a continuous variable that measures the water consumption of a household, in liters per
day, before the treatment date of 08-Dec-2018.
211
Table 3.23: LATE Estimates of Heterogeneous Treatment Effects
Post-Treatment Water Consumption
(1) (2) (3) (4)
Complete Diagnostic −12.637 −5.097 −11.215 −47.769
(10.745) (11.459) (39.373) (63.047)
High-Use 9.851
∗∗∗
13.475
∗∗∗
13.818
∗∗
8.364
(2.651) (3.420) (6.973) (28.261)
High-Use× −65.567
∗∗
−82.943
∗∗
−129.464 4.134
Complete Diagnostic (32.642) (36.576) (111.668) (241.173)
Intercept 10.292
∗∗∗
10.460
∗∗∗
10.944
∗∗∗
16.654
∗
(1.603) (2.311) (2.077) (9.449)
Instruments All Treatment Incentives+Simplified Incentives Incentives 15
F-stat in First Stage 20, 17 34, 30 40, 16 2, 1
Controls Yes Yes Yes Yes
Observations 11,700 5,830 11,700 1,974
Notes: Robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All regressions report the local average treatment effect estimates of diagnostic completion on post-treatment water consumption. The dependent
variable for all models is Post-Treatment Water Consumption, a continuous variable that measures the water consumption of a household, in liters per
day, post the treatment date of December 08, 2018. Pre-treatment consumption and post-treatment consumption were available for only a subset
(30 per cent) of the households. Households with unreasonably large differences between pre and post-treatment consumption (absolute value
greater than 50 per cent) were dropped from the sample. The data was then trimmed at 1 and 99 percentile of pre-treatment consumption. The
regressor of interest is High-Use× Complete Diagnostic. Complete Diagnostic is a dummy variable that equals 1 for all households who completed
the water diagnostic. High-Use is a dummy that equals 1 if the household had a pre-treatment water consumption greater than the median of
the sample. For all the models, the instrument for the interaction term is the IV for the endogenous variable, Complete Diagnostic, interacted with
High-Use. The IV in Model (1) is a vector of dummies for all the different treatment arms. The IV in Model (2) is a vector that includes dummies for
Incentives 10, Incentives 15, and Simplified treatment arms. The IV in Model (3) is a vector of dummies for Incentives 10 and Incentives 15 groups,
while the IV in Model (4) is only the Incentives 15 group. The sample in model (2) consists of the Incentives, Simplified and control group, while
the sample in model (3) includes only Incentives and the Simplified group. Model (4) only includes the Incentives group. All models include
Rural and Pre-Treatment Consumption as controls. Rural is a dummy variable that equals 1 if the household is located in a rural area. Pre-Treatment
Consumption is a continuous variable that measures the water consumption of a household, in liters per day, before the treatment date of December
08, 2018.
a far greater impact on high-use households than low-use households. The estimate in
column (3), when our sample includes Simplified and Incentives group, is negative but in-
significant. However, as discussed earlier, this is most likely due to low statistical power
because we lose a major portion of our sample. For our preferred specification in column
(2), the fall in consumption is 83 liters per day for the high-use consumer. This represents
a percentage reduction of 24 percent relative to pre-treatment consumption for high-use
households. Thus, the online audit incentivized high-use households to conserve more
water.
212
Both the results above lend credence to the theory that behavioral interventions can
have heterogeneous impacts on consumers depending on their pre-treatment usage. There-
fore, utilities can target the households with high consumption as they seem more likely
to be incentivized.
H.1.3 Characteristics of Households that Complete the Diagnostic
Continuing with our theme of targeting, we find that different behavioral interventions
influenced different set of households to take up the water audit tool. This is relevant
because if the different letters differ in terms of which households they influence, it may
be easier to target the right behavioral intervention based on the customer attributes.
Table 3.24 provides the average value of the household characteristics across different
treatment arms for the subset of households who completed the diagnostic. The columns
represent different interventions and each row represents a household characteristic, rang-
ing from the type of residence and the number of different water-consumption devices
installed, to its water and energy usage. The last column reports the p-value from a joint
F-test of whether the household characteristic varies across the different groups. The re-
sults indicate that the financial incentives treatment influenced a relatively larger number
of unmetered households to commit to the audit. Therefore, households who were un-
able to monitor their daily consumption were more likely to complete the diagnostic if
offered monetary rewards. Furthermore, the average number of basins and toilets were
lower in households that completed the diagnostic owing to the Incentives treatment, sug-
gesting that financial inducement is a strong motivator for smaller or poorer households.
In other words, financial incentives also influenced households that would not reasonably
be expected to use the water audit tool.
213
Table 3.24: Characteristics of Households which Complete the Diagnostic
Vanilla Altruism Simplified Incentives 10 Incentives 15 Moral Cost F-Test
(1) (2) (3) (4) (5) (6) (7)
Rural 0.61 0.66 0.65 0.70 0.68 0.65 0.71
Metered 0.73 0.68 0.70 0.56 0.58 0.80 9.18
∗∗∗
Number of:
Showers 1.29 1.31 1.31 1.19 1.21 1.27 1.76
Toilets 1.99 1.97 1.93 1.73 1.73 1.97 4.29
∗∗∗
Basins 1.95 1.81 1.83 1.64 1.63 1.81 4.05
∗∗∗
Bathtubs 0.90 0.92 0.93 0.88 0.89 0.90 0.37
Kitchen Utility Taps 1.33 1.25 1.44 1.26 1.37 1.28 3.03
∗∗∗
People at Home 2.25 2.11 2.10 2.22 2.23 2.17 0.69
Cost of Water (£/year) 386.72 365.96 402.65 387.85 353.27 383.91 0.70
Frequency (per week):
Showers 10.36 10.22 9.71 9.95 10.83 10.34 0.81
Baths 2.85 2.97 2.89 3.19 2.86 2.81 0.29
Boiling Water 27.39 24.51 25.16 23.79 24.10 26.12 1.78
Wash Up by Hand 12.96 13.26 15.41 14.16 12.87 13.13 1.73
Dishwasher 2.22 2.28 2.15 1.62 1.73 2.17 2.08
∗
Washing Machine 5.22 4.85 4.51 4.98 4.51 4.36 1.16
Watering Garden 2.11 2.26 1.89 1.97 1.80 1.92 0.77
Shower Duration (mins) 6.49 6.83 7.05 7.68 7.04 6.73 2.10
∗
Water Use (’000 litres/yr):
Bathroom 85.11 81.47 83.52 90.22 90.76 86.68 0.87
Kitchen 32.56 30.15 30.72 31.66 30.53 30.46 0.49
Outdoor 1.08 1.49 1.40 1.03 1.03 1.09 1.53
Household 118.74 113.11 115.65 122.92 122.32 118.22 0.63
Per Person 54.12 55.52 55.58 56.22 56.19 55.02 0.27
Energy Use (’000 kWh/yr):
Bathroom 1.22 1.21 1.23 1.34 1.34 1.27 0.65
Kitchen 0.67 0.63 0.64 0.65 0.63 0.64 0.35
Household 1.90 1.85 1.87 1.99 1.97 1.91 0.39
Type of Residence:
Cottage/Bungalow 0.09 0.06 0.10 0.07 0.08 0.12 1.25
Detached 0.31 0.35 0.28 0.21 0.23 0.34 3.83
∗∗∗
Flat 0.04 0.03 0.07 0.02 0.04 0.06 1.33
Semi-Detached 0.41 0.38 0.41 0.48 0.44 0.33 2.65
∗∗
Terrace 0.15 0.18 0.15 0.21 0.21 0.15 1.44
Observations 140 176 189 242 278 259
Notes:
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All data are from the water diagnostic survey, and the number of observations, therefore, include only the homes which completed the diagnostic. Columns 1 to 6 report the mean value of each household
characteristic for the respective treatment groups. Rural, Metered, and all five variables related to Type of Residence are binary. Cost of Water (£/year) is self-reported and only includes homes which pay
for their own water. The variables related to Water Use (litres/year) and Energy Use (kWh/year) are calculated by NWL based on the answers provided by the households in the diagnostic. Energy Use
(kWh/year) is the total amount of energy used by a household in a year to heat water. The final column, F-Test, reports thep-value from a joint orthogonality test of equality of means between the six
treatment groups.
214
H.1.4 Interaction of Households with Reminders
Sending reminders to consumers may be an important method to reinforce the impact
of behavioral interventions. Therefore, it is important to know how customers interact
with reminders and their impact on take-up of the audit. We find that customers inter-
action with the reminder email depends on the content of the reminder, with Moral Cost
reminder doing well in terms of positive engagement.
Email reminders were randomly sent to the subset of customers that had not completed
the diagnostic by February 2019. Using CRM data, we can count the number of people
who opened the reminder emails, or opened the reminder email and clicked on the link to
the audit tool, or simply unsubscribed. Results from this analysis are presented in Table
3.25.
As compared to the Vanilla reminder, all reminders, except Altruism, had a positive
and significant effect on the probability of opening the reminder. The email appealing
to an altruistic motive, however, was opened considerably fewer times. Moreover, fewer
households clicked on the diagnostic link after opening the email if it belonged to the
said category. Surprisingly, the Moral Cost reminder resonated positively, with greater
participation in the audit as compared to households who received the Vanilla reminder.
Finally, receiving a reminder with a monetary incentive was the only intervention which
reduced the probability of unsubscribing from future emails.
H.2 Welfare Calculations
H.2.1 Parameters
The different parameters used in the welfare calculations are specified in Table 3.26, along
with their units and sources.
215
Table 3.25: ATE Estimates of Letters on Interaction with Reminders
Opened Reminder Clicked Reminder Email Unsubscribed
(1) (2) (3)
Simplified 0.060
∗∗∗
0.009 −0.001
(0.021) (0.009) (0.004)
Altruism −0.059
∗∗∗
−0.039
∗∗∗
−0.002
(0.021) (0.006) (0.004)
Incentives 10 0.056
∗∗
0.017 −0.008
∗∗
(0.026) (0.011) (0.003)
Incentives 15 0.072
∗∗∗
0.011 −0.002
(0.027) (0.011) (0.005)
Moral Cost 0.068
∗∗∗
0.028
∗∗∗
−0.000
(0.021) (0.010) (0.004)
Intercept 0.432
∗∗∗
0.028
∗∗∗
0.017
∗∗∗
(0.019) (0.008) (0.004)
Controls Yes Yes Yes
Observations 5,563 5,563 5,563
Notes: Robust standard errors are in parenthesis.
∗∗∗
p< 0.01;
∗∗
p< 0.05;
∗
p< 0.1.
All regressions report the average treatment effect estimates of different behavioral interventions on how customers interacted
with the reminders. Dependent variables, all dummy variables, are presented as column names. Opened Reminder refers to
if the household clicked the email and were shown its content. Clicked Reminder means that the household clicked the link to
the audit tool within the reminder. Email Unsubscribed refers to a situation where the household unsubscribed from receiving
any further reminder emails from NWL. The reference group in each model is the Vanilla group. The data for each regression
includes only households who had not completed the diagnostic by 06-Feb-2019, and had received an email reminder. All
models include the dummy variables Meter and Rural as controls. The former equals 1 if the household has a water meter
attached to it, and the latter equals 1 if the household is located in a rural area.
216
Table 3.26: Parameters and Sources
Variable Unit Value Source Notes
(1) (2) (3) (4)
Consumer Price of Water £/m
3
1.3 NWL Charges Scheme 2020
Short-Run Marginal Cost £/m
3
0.44 NWL Financial Statements 2021
Long-Run Marginal Cost £/m
3
0.98 NWL Financial Statements 2009, 2021
Marginal operating cost of£0.44/m
3
(SRMC) and marginal capacity
cost of£0.54/m
3
from Abberton
Reservoir Water Resource Scheme
Emissions/Ml of Water Supply kgCO2e/MI 140 NWL Financial Statements 2021
Emissions/Ml of Sewage Treated kgCO2e/MI 520 NWL Financial Statements 2021
Includes Scope 3 (other indirect)
GHG emissions. For details, see
HM Government 2019
Emissions/MI from
Household Water Use
kgCO2e/MI 6,200 Emma, Feiferi, Fida, and Paul 2008
Social Cost of CO2 $/tCO2e 51 Interagency Working Group, USG 2021 Assumes a 3 percent discount rate
Time to Complete Diagnostic Minutes 7 Field Experiment Mean of all households
Cost of Posting a Letter £/letter 0.41 Royal Mail 2021
Standard tariff in 2020-21 for orders
containing less than 2,500 items
Corporate Tax Rate percentage 0.19 NWL Financial Statements 2021
UK Median Wage £/hour 14 ASHE 2021
Weight on Leisure Time percentage 0.5 White 2016
$ to£ conversion 0.78 Exchange Rates UK 2020 Average 2020 rate
Notes: 1MI equals 10
6
liters. For emissions from wastewater treatment, NWL provides two measures of GHG intensity ratios: flow to full treatment, and water distribution input. Based on the Environmental
Reporting Guidelines published by HM Government 2019, the latter intensity ratio takes into account Scope 3 emissions, which are defined as emissions that are a consequence of the utility’s actions, which
occur at sources which the utility does not own or control and which are not classed as Scope 2 emissions.
217
H.2.2 Cost Effectiveness Calculations
Table 3.7 in the main text provides a comparison of the cost effectiveness with other studies
in the literature. Calculations related to the comparison with Ansink, Ornaghi, and Tonin
2021 are presented below in Table 3.27. The cost effectiveness calculations in their paper
do not lend themselves easily to comparison with our numbers, and therefore, we provide
a summary of our calculations below. Panel A shows the total water savings from the
information and technology arm for all the months in the one year following the treatment.
In other words, it provides a measure of the effectiveness against which costs need to be
compared. Panel B shows the calculations related to total costs. Subsequently, we divide
the costs in Panel B by the effectiveness in Panel A to arrive at the CE.
For studies other than Ansink, Ornaghi, and Tonin 2021, there was a cost effectiveness
number specified, but in dollars ( $) per gallon. The same has been converted to dollars ($)
per cubic meter and in 2020 dollars (as opposed to dollars in the year of publishing) for
comparison. The inflation adjustment used price data from US Bureau of Labor Statistics
2021
39
. The details of the calculations are presented in Table 3.28.
H.3 Calculation of Pre- and Post-Treatment Water Consumption
We now provide a detailed description of the computation of consumption data for dif-
ferent households. To help illustrate the format of the data shared by NWL, and our data
cleaning process, we use some randomly generated data in Table 3.29
The consumption data from NWL consisted of a series of four meter readings for each
household. Each meter reading includes the date of the reading, and its corresponding
value. Thus, Readdate 1 represents the date of the earliest reading for the household in
our data set, while Readdate 4 represents the date of the latest reading. All households
for which we did not have at least one reading before and after the treatment date (i.e.
08-Dec-2018) were dropped from the sample. Readings for different households were
39
Consumer Price Index for All Urban Consumers (US city average series for all items)
218
Table 3.27: Cost Effectiveness in Ansink, Ornaghi, and Tonin 2021
Panel A
Month
Reduction due
to Information
(liters/day/hh)
Reduction due
to Technology
(liters/day/hh)
Total Reduction
due to Information
(cubic meters)
Total Reduction
due to 1 Device
(cubic meters)
(1) (2) (3) (4)
Month 1 -46 -6 -13,011 -1,403
Month 2 -42 -6 -11,977 -1,407
Month 3 -39 -6 -11,171 -1,341
Month 4 -38 -4 -10,813 -1,053
Month 5 -31 -6 -8,702 -1,345
Month 6 -28 -5 -7,958 -1,240
Month 7 -26 -5 -7,286 -1,276
Month 8 -22 -6 -6,232 -1,387
Month 9 -19 -7 -5,523 -1,570
Month 10 -17 -7 -4,765 -1,564
Month 11 -15 -7 -4,205 -1,713
Month 12 -14 -7 -3,928 -1,717
A: Total Water Conserved in 1 Year (m
3
) -95,570 -17,016
Panel B
Variable Unit
Information
Component
Technology
Component
(1) (2) (3)
Cost
£/hh (column 2)
£/device (column 3)
30 13.5
B: Total Cost £ 284,880 107,685
Cost Effectiveness £/m
3
3.0 6.3
Cost Effectiveness $/m
3
3.8 8.1
Notes: Total number of households in the study were 9,496. For calculating the reduction due to 1 device, the percentage of h/h’s with no water saving
devices (16 percent) were removed from the sample. Reductions due to information and technology component are sourced from Appendix Table A of
Ansink, Ornaghi, and Tonin 2021. Total water conserved in 1 year is the sum of water reductions across all the 12 months. Cost of information component
calculated as the product of time taken per audit (1.5 hours) and average hourly labor cost of£20/hour (as assumed by the authors). Cost of technology
component includes cost of one device (£9 per device) plus delivery costs per household (£4.5). Total Cost calculated as per household cost multiplied by
total number of households. Total number of households in the case of technology component adjusted for percentage of households with no water saving
devices. For conversion rate from£ to$, see tab:welfareparametersforparameters
219
Table 3.28: Cost Effectiveness Calculations for Other Studies
Paper Population / Bound
$ per 1000
gallons reduced
(Year of Paper)
$ per 1000
gallons reduced
(2020)
Cost effectiveness
($/m3)
(1) (2) (3) (4)
Bennear, Lee, and L. O. Taylor 2013
Lower Bound 7.33 8.3 2.2
Upper Bound 26 29 7.6
Ferraro and Miranda 2013
All Households 0.37 0.41 0.11
High-Use Households 0.20 0.22 0.06
Ferraro and M. K. Price 2013
All Households 0.58 0.65 0.17
High-Use Households 0.42 0.47 0.12
Bernedo, Ferraro, and M. Price 2014 All Households 0.24 0.26 0.07
Brent, J. H. Cook, and Olsen 2015
Lower Bound 1.7 1.9 0.50
Upper Bound 2.6 2.9 0.75
Notes: 1000 gallons equals 4.5 cubic meters. For all studies, the cost effectiveness was converted to 2020 values based on the cumulative inflation rate between the year the study
was published and 2020. The inflation adjustment used price data from US Bureau of Labor Statistics 2021. High-use households in Ferraro and Miranda 2013 refer to households
who both have above median consumption and own their homes. High-use households in Ferraro and M. K. Price 2013 refer to households who have above median consumption.
Table 3.29: Format of Consumption Data
Unique ID Readdate 1 Read 1 Readdate 2 Read 2 Readdate 3 Read 3 Readdate 4 Read 4
1 2017-02-21 7438 2018-02-23 7585 2018-12-24 7864 2019-04-20 7986
2 2016-11-03 1184 2017-07-27 1379 2018-07-19 1674 2019-01-14 1803
taken at different times, and therefore, Readdate 1 for Unique ID 1 could be very different
from Readdate 1 for Unique ID 2. Pre-treatment water consumption was calculated by
differencing the two readings immediately prior to the treatment date. In the example,
pre-treatment consumption for Unique ID 1 is the difference between Read 2 and Read 1,
whereas the pre-treatment consumption for Unique ID 2 is the difference between Read 3
and Read 2. If either of the two readings immediately prior to treatment were taken before
01-Jan-2010, the household was dropped as the date is too far back in time to accurately
measure consumption in the present period.
Post treatment water consumption was the difference between the two most recent
readings. Most of the households only had a single reading post treatment, and therefore,
post consumption in that case would be the difference between the reading post treatment
and the reading immediately prior to the treatment. For example, post consumption for
both Unique ID 1 and Unique ID 2 would be the difference between Read 3 and Read 4,
but Readdate 3 in case of Unique ID 2 was prior to the treatment date.
220
The difference between any two readings gives us the water consumption in cubic
meters during the time interval obtained by differencing the two corresponding reading
dates. To standardize this measure across all households, the difference between any two
readings was divided by the number of days between the respective readings to obtain
average water consumption in cubic meters per day. Finally, this measure was multiplied
by a 1000 to obtain water consumption in liters per day.
H.4 Sample Letters
The letters and the reminder emails sent to the different treatment groups by NWL to their
customers are presented below.
221
Figure 3.5: Vanilla (Status Quo) Mailer
222
Figure 3.6: Simplified Mailer
223
Figure 3.7: Altruism Mailer
224
Figure 3.8: £10 Incentive Mailer
225
Figure 3.9: £15 Incentive Mailer
226
Figure 3.10: Moral Cost Mailer
227
Figure 3.11: Reminder Email
228

Abstract (if available)

Abstract This dissertation comprises three chapters in environmental economics, in particular on climate change adaptation and what factors positively influence or impede adaptation behavior. The first chapter documents that the pre-existing market power distortion in India’s intermediary market can hinder farmer climate change adaptation. Using the variation in the market power of local crop intermediaries due to historical agricultural laws, I show that (1) farmers selling in the intermediary markets with less market power suffer substantially less from extreme heat; (2) the farmer’s economic loss due to extreme weather could be mitigated by 13.8 percent if the restrictions on cross-border trading are removed.

The second chapter uses night lights data as a proxy for economic activity to test several flood risk adaptation hypotheses. We first document heterogeneity in the economic impact of floods across 9,500 cities around the globe. We document that floods cause less damage to richer cities, showing the potential of economic growth in mitigating climate risk. Cities with more past experience with floods, and cities with protective dams suffer less from flooding. Finally, and this is by no means a policy recommendation, we also find that authoritative regimes recover faster after floods. Population growth is lower in cities that suffer from more floods. Richer cities suffer fewer deaths from flood events.

The third chapter documents how behavioral interventions can incentivize energy conservation. The paper is based in the United Kingdom where we attempt to incentivize households to take up a water audit. We use financial incentives as a behavioral nudge and compare them to other nudges such as those that evoke altruism and moral suasion. We find that financial incentives work best, with participants in this group having a higher rate of take-up of the audit and also substantially reducing water usage. However, we also find that in spite of these water savings, the financial incentives intervention does not pass a benefit cost test because the greenhouse gas benefits associated with these reductions are not large enough to compensate for the loss in revenue of utilities.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays on the economics of climate change adaptation in developing countries

PDF

Essays in environmental economics

PDF

Essays on narrative economics, Climate macrofinance and migration

PDF

Location choice and the costs of climate change

PDF

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

PDF

Essays on urban and real estate economics

PDF

Sustaining open source software production: an empirical analysis through the lens of microeconomics

PDF

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

PDF

The impact of minimum wage on labor market dynamics in Germany

PDF

Essays in the economics of education and conflict

PDF

Migration, location, and economic opportunity

PDF

Essays on applied microeconomics

PDF

Essays on environmental economics: education, employment, and experiments

PDF

Essays on development economics and adolescent behavior

PDF

Essays on business cycle volatility and global trade

PDF

Essays in behavioral and entrepreneurial finance

PDF

Essays in health economics and provider behavior

PDF

The impact of economic shocks on firm behavior: insights from three studies

PDF

The impact of campus climate on community college student motivation

PDF

Genetic and environmental effects on symbiotic interactions across thermal gradients

Asset Metadata

Creator Kochhar, Rajat (author)

Core Title Essays in climate change adaptation: role of market power in incentivizing adaptation behavior

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Degree Conferral Date 2023-05

Publication Date 05/16/2023

Defense Date 03/28/2023

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

Tag adaptation,climate change,Markets,mitigation,nudges,OAI-PMH Harvest

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Kahn, Matthew E. (committee chair), Chaney, Thomas (committee member), Matsusaka, John G. (committee member), Metcalfe, Robert D. (committee member), Oliva, Paulina (committee member)

Creator Email rajat.kochhar26@gmail.com,rajatkoc@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113127156

Unique identifier UC113127156

Identifier etd-KochharRaj-11857.pdf (filename)

Legacy Identifier etd-KochharRaj-11857

Document Type Thesis

Format theses (aat)

Rights Kochhar, Rajat

Internet Media Type application/pdf

Type texts

Source 20230516-usctheses-batch-1045 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

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

Repository Name University of Southern California Digital Library

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

Repository Email cisadmin@lib.usc.edu