Inequality and Environmental Policy: Essays on Climate Adaptation, Health, and Renewable Energy by Emmett Bradley Reynier A dissertation accepted and approved in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Economics Dissertation Committee: Mark Colas, Co-Chair Edward Rubin, Co-Chair Eric Zou, Core Member Nicole Ngo, Institutional Representative University of Oregon Winter 2025 © 2025 Emmett Bradley Reynier This work is openly licensed via CC BY-NC 4.0 2 https://creativecommons.org/licenses/by-nc/4.0/ DISSERTATION ABSTRACT Emmett Bradley Reynier Doctor of Philosophy in Economics Title: Inequality and Environmental Policy: Essays on Climate Adaptation, Health, and Renewable Energy This dissertation covers three topics related to inequality and environmental policy, covering, in particular, the distributional effects of climate change, the health effects of pesticides, and solar panel subsidies. Chapter I provides a summary of the work. Chapter II, co-authored with John Morehouse, examines the distributional effects of climate change. We use a quantitative spatial equilibrium model to simulate the effects of climate change, allowing households to adapt via migration and energy use. We find that low-income and minority households have larger welfare losses from climate change to date and that those disparities in welfare effects will widen under future emissions scenarios. We then simulate the effects of a place-based, means-tested subsidy inspired by an Inflation Reduction Act program. Targeting subsidies to high-climate-exposure cities does a good job of helping the people who are worst off from a climate change perspective. However, it also incentivizes households to move into those high-exposure areas. Chapter III, co-authored with Edward Rubin, quantifies the causal effect of a common herbicide, glyphosate, on perinatal health. To identify this causal effect, we leverage a natural experiment resulting from the release of genetically modified seeds designed to be resistant to glyphosate. We pair temporal variation in glyphosate use induced by the release of genetically modified seeds with spatial 3 variation in the suitability of land for growing crops with genetically modified varieties (corn, soy, and cotton). We find glyphosate causes significant decreases in birthweight and gestation length and increases in the probability of low birthweight. We find considerable heterogeneity in this effect, with the most at-risk births having the largest effects. Chapter IV, co-authored with Mark Colas, determines the optimal spatial distribution of subsidies for residential solar panels. The benefits of solar panel subsidies vary considerably across space due to differences in sunlight, the dirtiness of the electricity grid, and household installation behavior. We pair a model of household solar panel installation behavior with a model of the electricity grid. This model allows us to quantify the optimal schedule of subsidies across states. We find that the current system of subsidies leads to a severe spatial misallocation of solar installations. Additionally, the overall level of subsidies is too high to be justified by the environmental benefits alone. This result suggests that there could be gains from reallocating funds to alternative programs. This dissertation includes both previously published and unpublished and co-authored material. 4 CURRICULUM VITAE NAME OF AUTHOR: Emmett Bradley Reynier GRADUATE AND UNDERGRADUATE SCHOOLS ATTENDED: University of Oregon, Eugene, OR, USA University of Virginia, Charlottesville, VA, USA DEGREES AWARDED: Doctor of Philosophy, Economics, 2025, University of Oregon Master of Science, Economics, 2021, University of Oregon Bachelor of Science, Commerce, 2017, University of Virginia Bachelor of Arts, Economics, 2017, University of Virginia AREAS OF SPECIAL INTEREST: Environmental Economics Public Economics Urban Economics PROFESSIONAL EXPERIENCE: ORISE Research Fellow, US EPA Office of Water, 2023-2025 Senior Business Analyst, Carmax, 2019 Business Analyst, Carmax, 2017-2019 Feeds and Search Engine Marketing Intern, Merkle, 2016 GRANTS, AWARDS AND HONORS: Kleinsorge Summer Research Fellowship, 2022 Dale Underwood Outstanding Graduate Student Scholarship, 2020 Best First Year Econometrics Performance, 2020 First Year Fellowship, 2019-2020 5 PUBLICATIONS: Colas, M. and E. Reynier (Forthcoming). “Optimal Subsidies for Residential Solar.” Journal of Political Economy Microeconomics. Reynier, E. and E. Rubin (2025). “Glyphosate exposure and GM seed rollout unequally reduced perinatal health.” Proceedings of the National Academy of Sciences. 122, 3. Colas, M. and E. Reynier (2023). “Vertical Migration Externalities.” Regional Science and Urban Economics. 101, 103900. 6 ACKNOWLEDGEMENTS I owe many thanks to the co-chairs of my committee Mark Colas and Ed Rubin. Their years of guidance and support have been a major impetus for my growth as an economist and person. I would also like to thank Eric Zou, Nicole Ngo, Andrew Dickinson, Philip Economides, Tamara Ren, Giorgi Nikolaishvili, Keaton Miller, Jon Davis, Woan Foong Wong, Benjamin Hansen, Glen Waddell, Todd Doley, Julie Hewitt, Joel Corona, and Kristen Swedberg for their help in many aspects of my work. The work in this disseration was supported in part by an appointment to the Research Participation Program at the Water Economics Center, US Environmental Protection Agency, administered by the Oak Ridge Institute for Science and Education through an interagency agreement between the US Department of Energy and EPA. Any views expressed are solely those of the authors and do not necessarily represent the opinions of the US EPA. Finally, I’d like to thank my wife, Jocelyn, and our family. They have been incredibly supportive of me as I have crafted this dissertation. 7 To Jocelyn and Rainbow. 8 TABLE OF CONTENTS Chapter Page I. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . 25 II. THE DISTRIBUTIONAL IMPACTS OF CLIMATE CHANGE ACROSS US LOCAL LABOR MARKETS . . . . . . . . . 28 2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.2. Data and Descriptive Evidence . . . . . . . . . . . . . . . . . 35 2.2.1. Data . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.2.2. Descriptive Evidence . . . . . . . . . . . . . . . . . . 38 2.3. Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 2.3.1. Households . . . . . . . . . . . . . . . . . . . . . . . 43 2.3.1.1. The impact of climate change on household decisions . . . . . . . . . . . . . . . 47 2.3.2. Firms . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.3.3. Rents . . . . . . . . . . . . . . . . . . . . . . . . . 51 2.4. Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 52 2.4.1. Energy and Housing Demand . . . . . . . . . . . . . . . 52 2.4.1.1. Elasticities of substitution . . . . . . . . . . . . 52 2.4.1.2. Effect of climate on comfort production . . . . . . 55 2.4.2. Labor Supply . . . . . . . . . . . . . . . . . . . . . . 56 2.5. Parameter Estimates . . . . . . . . . . . . . . . . . . . . . . 58 2.6. The welfare effects of climate change . . . . . . . . . . . . . . . 68 2.6.1. Effect of climate change to-date . . . . . . . . . . . . . . 69 2.6.2. The black-white gap simulated under future emissions scenarios . . . . . . . . . . . . . . . . . . . 74 2.7. IRA inspired place-based, means-tested subsidies . . . . . . . . . 75 9 Chapter Page 2.8. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 85 III. GLYPHOSATE EXPOSURE AND GM SEED ROLLOUT UNEQUALLY REDUCED PERINATAL HEALTH . . . . . . . . . . 87 3.1. Main . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 3.1.1. Background and motivation . . . . . . . . . . . . . . . 89 3.1.1.1. Exposure . . . . . . . . . . . . . . . . . . . 90 3.1.1.2. Health impacts . . . . . . . . . . . . . . . . . 90 3.1.2. Empirical approach . . . . . . . . . . . . . . . . . . . 92 3.1.2.1. Data . . . . . . . . . . . . . . . . . . . . . 93 3.1.2.2. Estimation . . . . . . . . . . . . . . . . . . . 96 3.1.3. Results . . . . . . . . . . . . . . . . . . . . . . . . 99 3.1.3.1. Reduced-form evidence of GM rollout’s health effects . . . . . . . . . . . . . . . . . . 99 3.1.3.2. The effect of glyphosate on perinatal health . . . . 103 3.1.3.3. Heterogeneity in glyphosate’s health impacts . . . . 105 3.2. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 107 3.3. Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 3.3.1. Infant health data . . . . . . . . . . . . . . . . . . . . 116 3.3.2. Pesticide use estimates . . . . . . . . . . . . . . . . . . 117 3.3.3. Attainable yield . . . . . . . . . . . . . . . . . . . . 117 3.3.4. Additional data . . . . . . . . . . . . . . . . . . . . . 118 3.3.5. Empirical strategy . . . . . . . . . . . . . . . . . . . 119 3.3.6. Identification . . . . . . . . . . . . . . . . . . . . . . 120 3.3.6.1. Reduced-form DID . . . . . . . . . . . . . . . 120 3.3.6.2. Two-stage least squares . . . . . . . . . . . . . 122 3.3.7. Birth-outcome index . . . . . . . . . . . . . . . . . . . 124 10 Chapter Page 3.3.8. Predicting birthweight . . . . . . . . . . . . . . . . . . 124 3.3.9. Exposure to glyphosate sprayed upstream through water . . . . . . . . . . . . . . . . . . . . . 127 IV. OPTIMAL SUBSIDIES FOR RESIDENTIAL SOLAR . . . . . . . . 129 4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 129 4.2. Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 4.2.1. Households . . . . . . . . . . . . . . . . . . . . . . . 138 4.2.2. Electricity Production . . . . . . . . . . . . . . . . . . 142 4.2.2.1. Background . . . . . . . . . . . . . . . . . . 142 4.2.2.2. Model: Electricity Production . . . . . . . . . . 143 4.2.2.3. Damages . . . . . . . . . . . . . . . . . . . 146 4.2.3. Government’s Problem and Nationally-Optimal Subsidies . . 146 4.3. Quantitative Model . . . . . . . . . . . . . . . . . . . . . . 150 4.3.1. Household preferences . . . . . . . . . . . . . . . . . . 150 4.3.2. Dispatchable Power Plant Production . . . . . . . . . . . 152 4.3.3. Damages . . . . . . . . . . . . . . . . . . . . . . . . 154 4.4. Data and Estimation . . . . . . . . . . . . . . . . . . . . . 156 4.4.1. Data Sources . . . . . . . . . . . . . . . . . . . . . . 156 4.4.2. Descriptive Patterns . . . . . . . . . . . . . . . . . . . 159 4.4.3. Estimation . . . . . . . . . . . . . . . . . . . . . . . 165 4.5. Estimation Results and Model Fit . . . . . . . . . . . . . . . . 168 4.5.1. Households . . . . . . . . . . . . . . . . . . . . . . . 168 4.5.1.1. Parameter Estimates . . . . . . . . . . . . . . 168 4.5.1.2. Model Fit (Installations) . . . . . . . . . . . . . 169 4.5.1.3. Comparison to Existing Estimates . . . . . . . . 171 4.5.2. Power Plants . . . . . . . . . . . . . . . . . . . . . . 172 11 Chapter Page 4.6. Counterfactuals and Nationally-Optimal Subsidies . . . . . . . . . 176 4.6.1. Welfare-Maximizing Subsidies . . . . . . . . . . . . . . 176 4.6.2. Damage-Minimizing Reforms . . . . . . . . . . . . . . . 180 4.6.3. Tract-level Subsidies . . . . . . . . . . . . . . . . . . . 182 4.6.4. Unconstrained Reforms . . . . . . . . . . . . . . . . . 182 4.6.5. Marginal Subsidy Increases . . . . . . . . . . . . . . . . 185 4.7. Extensions, Robustness and Further Issues . . . . . . . . . . . . 186 4.7.1. Alternative Specifications of Household Utility . . . . . . . 186 4.7.2. Alternative Discounting of Future Subsidy Payments . . . . 186 4.7.3. Line Losses . . . . . . . . . . . . . . . . . . . . . . . 187 4.7.4. Transmission Constraints . . . . . . . . . . . . . . . . 188 4.7.5. Improved Storage of Nondispatchable Technology . . . . . . 189 4.7.6. Cleaner Electricity Production . . . . . . . . . . . . . . 190 4.7.7. Distributional Effects . . . . . . . . . . . . . . . . . . 191 4.7.8. Installation Elasticities with Historical Subsidy Measures . . . 193 4.8. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 194 APPENDICES A. APPENDIX FOR THE DISTRIBUTIONAL IMPACTS OF CLIMATE CHANGE ACROSS US LOCAL LABOR MARKETS . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 A.1. Data Appendix . . . . . . . . . . . . . . . . . . . . . . . . 195 A.1.1. Sample Selection . . . . . . . . . . . . . . . . . . . . 195 A.1.2. City-Level Indices . . . . . . . . . . . . . . . . . . . . 195 A.1.2.1. Energy Use . . . . . . . . . . . . . . . . . . 195 A.1.2.2. Wages . . . . . . . . . . . . . . . . . . . . . 197 12 Chapter Page A.1.2.3. Housing Quantity and Rent . . . . . . . . . . . 198 A.2. Additional Descriptive Results . . . . . . . . . . . . . . . . . 199 A.2.1. Energy demand and climate . . . . . . . . . . . . . . . 201 A.3. Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 A.3.1. Derivation of Comfort and Energy Prices . . . . . . . . . . 207 A.3.2. Marginal effect of climate on comfort demand . . . . . . . 210 A.3.3. Firm FOC derivation . . . . . . . . . . . . . . . . . . 211 A.3.4. Labor Demand Parameters . . . . . . . . . . . . . . . . 211 A.3.5. Housing Supply Calibration . . . . . . . . . . . . . . . 212 A.3.6. Equilibrium Definition . . . . . . . . . . . . . . . . . . 213 A.3.7. Equilibrium Outline . . . . . . . . . . . . . . . . . . . 215 A.4. Estimation Appendix . . . . . . . . . . . . . . . . . . . . . 216 A.4.1. Alternative specifications for effect of climate on comfort . . . 216 A.4.2. Energy and comfort prices . . . . . . . . . . . . . . . . 218 A.4.3. Energy Efficiency: Evidence . . . . . . . . . . . . . . . 222 A.4.4. Moving Cost Parameters for All Years . . . . . . . . . . . 227 A.4.5. Climate amenities . . . . . . . . . . . . . . . . . . . . 227 A.4.6. Model fit . . . . . . . . . . . . . . . . . . . . . . . . 227 A.5. Additional Results . . . . . . . . . . . . . . . . . . . . . . . 229 A.5.1. IRA Subsidies . . . . . . . . . . . . . . . . . . . . . 233 B. APPENDIX FOR GLYPHOSATE EXPOSURE AND GM SEED ROLLOUT UNEQUALLY REDUCED PERINATAL HEALTH . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240 B.1. Background . . . . . . . . . . . . . . . . . . . . . . . . . . 240 B.2. Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 B.3. OLS results . . . . . . . . . . . . . . . . . . . . . . . . . . 245 13 Chapter Page B.4. Comparing rescaled reduced-form DID results with 2SLS results . . . 245 B.5. Shift-share specification . . . . . . . . . . . . . . . . . . . . 245 B.6. Robustness of first-stage and reduced form results . . . . . . . . . 246 B.7. Robustness of 2SLS results for other outcomes . . . . . . . . . . 246 B.8. Prediction performance . . . . . . . . . . . . . . . . . . . . 247 B.9. 2SLS results and the shape of the glyphosate damage function . . . 248 B.10.Demographic trends . . . . . . . . . . . . . . . . . . . . . . 250 B.11.Other forms of heterogeneity . . . . . . . . . . . . . . . . . . 250 B.12.Effect of GM on acreage and yield . . . . . . . . . . . . . . . . 251 B.13.Other socioeconomic outcomes . . . . . . . . . . . . . . . . . 253 B.14.Effects of upstream glyphosate in water . . . . . . . . . . . . . 254 B.14.1. Predicting glyphosate in water with machine learning . . . . 254 B.14.2.Results: Effect from upstream glyphosate in water . . . . . 258 C. APPENDIX FOR OPTIMAL SUBSIDIES FOR RESIDENTIAL SOLAR . . . . . . . . . . . . . . . . . . . . . . 306 C.1. Data Appendix . . . . . . . . . . . . . . . . . . . . . . . . 306 C.1.1. Deep Solar . . . . . . . . . . . . . . . . . . . . . . . 306 C.1.2. Google Project Sunroof . . . . . . . . . . . . . . . . . 306 C.1.3. Tracking the Sun . . . . . . . . . . . . . . . . . . . . 307 C.1.4. System Advisor Model . . . . . . . . . . . . . . . . . . 308 C.1.5. State Electricity Prices . . . . . . . . . . . . . . . . . 309 C.1.6. Subsidies . . . . . . . . . . . . . . . . . . . . . . . . 309 C.1.7. Power Plants . . . . . . . . . . . . . . . . . . . . . . 311 C.2. Theory and Quantitative Appendix: For Online Publication . . . . 316 C.2.1. States Without Net Metering . . . . . . . . . . . . . . . 316 14 Chapter Page C.2.2. Maximum Likelihood Estimation of Power Plant Policy Functions . . . . . . . . . . . . . . . . . . . . 317 C.2.3. Details: Cost-Neutral Reforms . . . . . . . . . . . . . . 318 C.2.4. Numerical Algorithm for Calculating Optimal Subsidies . . . 320 C.2.5. Details: Damage-Minimizing Subsidies . . . . . . . . . . . 320 C.3. Results Appendix: For Online Publication . . . . . . . . . . . . 322 C.3.1. Installation Prices . . . . . . . . . . . . . . . . . . . . 322 C.3.2. Relationship Between Installations and Monetary Incentives . 322 C.3.3. Installation Size Regressions . . . . . . . . . . . . . . . 324 C.3.4. Border Discontinuities in Household Characteristics . . . . . 326 C.3.5. Border Regressions with Controls for Non- Subsidy Incentives . . . . . . . . . . . . . . . . . . . 326 C.3.6. Power Plant Model . . . . . . . . . . . . . . . . . . . 327 C.3.7. Type of Subsidy in Nationally-Optimal System . . . . . . . 327 C.3.8. Average Panel Size Across Counterfactuals . . . . . . . . . 333 C.3.9. State-Level Results . . . . . . . . . . . . . . . . . . . 333 C.3.10.Unconstrained Nationally-Optimal Subsidies with λ = 1 . . . . . . . . . . . . . . . . . . . . . . . . . 333 C.4. Extensions and Robustness Appendix: For Online Publication . . . 337 C.4.1. Alternative Specifications of Household Utility . . . . . . . 337 C.4.2. Alternative Discounting of Future Subsidy Payments . . . . 337 C.4.3. Line Losses . . . . . . . . . . . . . . . . . . . . . . . 343 C.4.4. Transmission Constraints . . . . . . . . . . . . . . . . 344 C.4.5. Improved Storage of Nondispatchable Electricity . . . . . . 347 C.4.6. Cleaner Electricity Production . . . . . . . . . . . . . . 348 REFERENCES CITED . . . . . . . . . . . . . . . . . . . . . . . . 352 15 LIST OF FIGURES Figure Page 1. Change in the number of hot and cold days between 1990 and 2019. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2. Relationship between income and climate change. . . . . . . . . . . 40 3. Relationship between share non-white and climate. . . . . . . . . . . 41 4. Estimation results for the effect of climate on comfort production . . . . 61 5. Moving cost estimates for Black and White households . . . . . . . . 62 6. Simulations use actual climate data across all years. . . . . . . . . . 68 7. General equilibrium changes in welfare, energy prices, and population across states . . . . . . . . . . . . . . . . . . . . . . 71 8. Compensating Variation by income decile and race in 2019 . . . . . . 72 9. Time series of the Black-white welfare gap under different emissions scenarios. . . . . . . . . . . . . . . . . . . . . . . . . 75 10. CV from subsidies relative to CV from climate change to date for different spatial distributions of IRA subsidies. . . . . . . . . 81 11. Change in climate with subsidies relative to without subsidies for households in the lowest income decile. . . . . . . . . . 84 12. GM-crop suitability predicts glyphosate increases after GM-seed introduction. . . . . . . . . . . . . . . . . . . . . . . 95 13. GM-seed introduction increased glyphosate intensity in GM- crop suitable areas; birthweight reductions were also higher in GM-suitable counties and match GM-seed timing. . . . . . . . . . 100 14. Glyphosate’s direct and policy effects reduced birth outcomes for the average rural birth. . . . . . . . . . . . . . . . . 103 15. Birthweight losses due to glyphosate and GM are largest for births with the lowest expected birthweights. . . . . . . . . . . . . . 106 16 Figure Page 16. Female infants, children of Black and non-White parents, and children of unmarried parents have lower predicted birthweights. . . 110 17. Expected subsidies and monetary benefit for a 15-panel system in each state. . . . . . . . . . . . . . . . . . . . . . . . 160 18. Installed solar systems per 1000 individuals. . . . . . . . . . . . . . 161 19. Border Discontinuities in Monetary Benefits and Installations. . . . . . 162 20. Household solar installation model fit. . . . . . . . . . . . . . . . . 170 21. Model fit at the interconnection level. . . . . . . . . . . . . . . . . 173 22. Model fit at the interconnection level by hour and season. . . . . . . . 174 23. Fuel mix of production by interconnection. . . . . . . . . . . . . . 175 24. Estimated marginal damage of electricity production by region. . . . . 177 25. State-level nationally-optimal subsidies and misallocation for welfare-maximizing reforms. . . . . . . . . . . . . . . . . . . 180 26. State-level nationally-optimal subsidies and misallocation for damage-minimizing reforms. . . . . . . . . . . . . . . . . . . 181 27. Damages offset per additional dollar of government funds . . . . . . . 185 A.1. Mean fire and flood risk scores from the First Street Foundation data by census tract. . . . . . . . . . . . . . . . . . . 200 A.2. Change in the no rain days and annual precipitation. . . . . . . . . . 200 A.3. Relationship between share non-white and climate. . . . . . . . . . . 201 A.4. Relationship between share non-white and climate. . . . . . . . . . . 202 A.5. Black-white energy expenditure gap by income decile. . . . . . . . . . 203 A.6. Heterogeneity in the effect of degree days on energy expenditures by income decile. . . . . . . . . . . . . . . . . . . . 205 A.7. Heterogeneity in the effect of degree days on energy expenditures by income decile and race. . . . . . . . . . . . . . . . 206 A.8. Estimation results for the effect of climate on comfort production . . . . 217 17 Figure Page A.9. The effect of an additional day at temperature on comfort price by race. . . . . . . . . . . . . . . . . . . . . . . . . . . 218 A.10.Rent and Energy Prices versus Comfort Price by demographic group . . . . . . . . . . . . . . . . . . . . . . . . 219 A.11.Energy-efficiency by education group across years and CBSAs . . . . . 223 A.12.Energy-efficiency by education-group across years and CBSAs . . . . . 224 A.13.Energy-efficiency by demographic-group across years and CBSAs . . . . 225 A.14.Estimate of MWTP as a percent of income for an additional day at the given temperature . . . . . . . . . . . . . . . . . . . 228 A.15.Model fit when using observed climate to simulate the model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229 A.16.Change in prices and demand across states. . . . . . . . . . . . . . 230 A.17.Compensating Variation by race in 2019 if the climate was the same as in 1990. . . . . . . . . . . . . . . . . . . . . . . . 231 A.18.College and non-college wages comparing present day climate to that of 1990. . . . . . . . . . . . . . . . . . . . . . . 233 A.19.Share of the population living in a disadvantaged community. . . . . . 234 A.20.Distribution of baseline CV by income group. . . . . . . . . . . . . 234 A.21.CV for 2019 climate with and without subsidies relative to the 1990 climate without subsidies. . . . . . . . . . . . . . . . . . 235 A.22.CV for 2019 climate with and without subsidies relative to the 1990 climate without subsidies for different magnitudes, G. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236 A.23.Percent change in population under different subsidy program magnitudes and distributions. . . . . . . . . . . . . . . . 237 A.24.Percent change in rents under different subsidy program magnitudes and distributions. . . . . . . . . . . . . . . . . . . . 238 A.25.Percent change in wages under different subsidy program magnitudes and distributions. . . . . . . . . . . . . . . . . . . . 239 B.1. GM crop suitability and increases in glyphosate for rural counties. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262 18 Figure Page B.2. GM seeds were rapidly adopted after their 1996 release. . . . . . . . 263 B.3. Interpolation of fertilizer data. . . . . . . . . . . . . . . . . . . . 264 B.4. Counties with high suitability for GM crops increased glyphosate intensity and reduced non-glyphosate pesticides with the introduction of glyphosate-resistant seeds. . . . . . . . . . 265 B.5. Perinatal health declined in GM-crop suitable counties after the introduction of glyphosate-resistant seeds . . . . . . . . . . . . 266 B.6. The estimated effect of glyphosate on birthweight is robust to alternative specifications. . . . . . . . . . . . . . . . . . . . . 269 B.7. Glyphosate effects for infants born to non-white mothers are larger for birthweight and for the probabilities of preterm birth, LBW, and VLBW. . . . . . . . . . . . . . . . . . . . . . 270 B.8. Predicted birthweights closely match actual birthweights across the predicted birthweight distribution. . . . . . . . . . . . . 272 B.9. First-stage event study coefficients are similar across predicted BW quintiles, reduced form shows larger effects in lower quintiles. . . . . . . . . . . . . . . . . . . . . . . . . . 274 B.10.Heterogeneity in Glyphosate Effect is consistent across various predicted birthweight bin sizes, greater disparities among birthweight outcomes. . . . . . . . . . . . . . . . . . . . 275 B.11.Limited evidence of heterogeneous marginal effects by sex within predicted BW quintile. . . . . . . . . . . . . . . . . . . . 276 B.12.Robustness of birthweight effect to alternative economic controls. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279 B.13.Robustness of birthweight effect to alternative farm controls. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280 B.14.Robustness of birthweight effect to alternative instruments. . . . . . . 281 B.15.Robustness of birthweight effect to alternative fixed effects. . . . . . . 282 B.16.Robustness of birthweight effect to alternative fixed effects—within state. . . . . . . . . . . . . . . . . . . . . . . . 283 B.17.Heterogeneity in Birthweight Effect by Geographic Subsets. . . . . . . 284 19 Figure Page B.18.Birthweight event studies by rural and non-rural counties. . . . . . . 285 B.19.The estimated effect of glyphosate on gestation is robust to alternative specifications. . . . . . . . . . . . . . . . . . . . . . 286 B.20.The estimated effect of glyphosate on the health index is robust to alternative specifications. . . . . . . . . . . . . . . . . 287 B.21.Robustness to spatial subsets, all outcomes. . . . . . . . . . . . . . 288 B.22.Demographic event studies. . . . . . . . . . . . . . . . . . . . . 289 B.23.Reduced form heterogeneity by mother’s race. . . . . . . . . . . . . 290 B.24.Reduced form heterogeneity by mother’s education. . . . . . . . . . 291 B.25.Heterogeneity in effect by different month of birth, all outcomes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 292 B.26.Effect of local Max GM attainable yield on crop acreage. . . . . . . . 293 B.27.Effect of local GM max attainable yield on crop yield. . . . . . . . . 294 B.28.Coefficients from an event study regression of various socioeconomic variables on GM suitability for rural counties. . . . . . 295 B.29.Predicted glyphosate disaggregated into watersheds in 2004. . . . . . 296 B.30.Spatial variation in water ML predictors. . . . . . . . . . . . . . . 297 B.31.Capturing upstream and downstream watersheds. . . . . . . . . . . 298 B.32.Cross Validation Results. . . . . . . . . . . . . . . . . . . . . . 299 B.33.Out-of-sample prediction performance for LASSO and Random Forest models. . . . . . . . . . . . . . . . . . . . . . 300 B.34.Density of out-of-sample predictions relative to the actual values. . . . . 300 B.35.Aggregating watersheds to counties. . . . . . . . . . . . . . . . . 301 B.36.Predicted county-level AMPA in July of 2004. . . . . . . . . . . . . 302 B.37.Effect of upstream glyphosate by distance bin. . . . . . . . . . . . . 304 B.38.Effect of upstream glyphosate by distance bin by high and low soil erodibility and precipitation. . . . . . . . . . . . . . . . . 305 C.1. Daily solar radiation (kWh/m2/day) by census tract from Deepsolar. . . 306 20 Figure Page C.2. Hourly electricity generation for a standard solar panel for six example counties. . . . . . . . . . . . . . . . . . . . . . . . 309 C.3. State electricity prices ($/kWh) . . . . . . . . . . . . . . . . . . 310 C.4. Expected subsidy for a 15-panel system by subsidy type in 2014 dollars. . . . . . . . . . . . . . . . . . . . . . . . . . . . 310 C.5. Map of region service areas . . . . . . . . . . . . . . . . . . . . 314 C.6. Estimation results for solar system price regression . . . . . . . . . . 324 C.7. Border Discontinuities in Household Characteristics. . . . . . . . . . 326 C.8. Model fit at the region level in 2020. . . . . . . . . . . . . . . . . 330 C.9. Model fit at the region level by hour and season in 2020. . . . . . . . 331 C.10.Fuel mix of production by region. . . . . . . . . . . . . . . . . . . 332 21 LIST OF TABLES Table Page 1. Elasticity of substitution estimates. . . . . . . . . . . . . . . . . . 59 2. Estimation results for non-temperature climate amenities. . . . . . . . 64 3. Estimated climate amenity rankings by city for 2019. . . . . . . . . . 65 4. Change in climate amenities from 1990 to 2019. . . . . . . . . . . . . 66 5. Decomposition of welfare effects. . . . . . . . . . . . . . . . . . . 74 6. Regression of log installations per capita on the net present value of subsidies for a 15-panel installation within 10 miles of state borders. . . . . . . . . . . . . . . . . . . . . . . . . . 165 7. Parameter estimates for household utility function. . . . . . . . . . . 168 8. Welfare maximizing results . . . . . . . . . . . . . . . . . . . . . 178 9. Unconstrained Nationally-Optimal Subsidies. . . . . . . . . . . . . . 183 10. Regression of log installations per capita on “historically- adjusted” measures of subsidy generosity. . . . . . . . . . . . . . . 192 A.1. Effect of CDD and HDD on Energy Expenditures. . . . . . . . . . . 204 A.2. Effect of CDD and HDD on Energy Demand. . . . . . . . . . . . . . 208 A.3. Cities ranked by the percent difference between their average comfort expenditure and average housing expenditure in 2019. . . . . . 220 A.4. Cities ranked by the percent difference between their average comfort price and average housing price in 2019. . . . . . . . . . . . 221 A.5. Cities ranked by their average aggregate energy demand E and comfort demand C in 2019. . . . . . . . . . . . . . . . . . . . 222 A.6. Household moving costs estimated from Equation (2.31) for 1990. . . . . 225 A.7. Household moving costs estimated from Equation (2.31) for 2000. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 22 Table Page A.8. Household moving costs estimated from Equation (2.31) for 2010. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 A.9. Household moving costs estimated from Equation (2.31) for 2019. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 A.10.Decomposition of welfare effects. Values are CV as a percent of income. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 A.11.Changes in population using 1990 climate as a counterfactual in 2019 . . . . . . . . . . . . . . . . . . . . . . . 232 B.1. Summary statistics for high- and low-yield rural counties and urban counties. . . . . . . . . . . . . . . . . . . . . . . . . . . 260 B.2. Correlation between GAEZ suitability measures and pre- period acerage for GM crops. . . . . . . . . . . . . . . . . . . . . 261 B.3. 2SLS estimates of the policy and direct GLY effects on perinatal health. . . . . . . . . . . . . . . . . . . . . . . . . . 267 B.4. Comparing difference-in-differences and 2SLS results . . . . . . . . . 268 B.5. Effect of glyphosate on birthweight, gestation, and health index: Robustness to age and race share controls . . . . . . . . . . . 271 B.6. ML prediction performance . . . . . . . . . . . . . . . . . . . . 273 B.7. OLS estimates of the policy and direct GLY effects on perinatal health. . . . . . . . . . . . . . . . . . . . . . . . . . 277 B.8. Effect of GLY on perinatal health estimated with 2SLS shift-share instrument. . . . . . . . . . . . . . . . . . . . . . . 278 B.9. Effect of predicted GLY or AMPA in water on birthweight. . . . . . . 303 C.1. Summary statistics on power plants by region . . . . . . . . . . . . 314 C.2. Summary statistics on power plants by fuel category . . . . . . . . . . 315 C.3. Solar system installation prices . . . . . . . . . . . . . . . . . . . 323 C.4. Regression of log installations on the net present value of total monetary benefits associated with solar panel installations. . . . . 325 C.5. Regression of average panel size on the total monetary benefits associated with solar panel installations. . . . . . . . . . . . 325 23 Table Page C.6. Regression of log installations per capita on the net present value of subsidies . . . . . . . . . . . . . . . . . . . . . . . . . 328 C.7. Regression of log installations per capita on the net present value of subsidies . . . . . . . . . . . . . . . . . . . . . . . . . 329 C.8. Percent of total subsidy value from each type of subsidy. . . . . . . . 329 C.9. Average number of solar panels in a solar installation by Census region. . . . . . . . . . . . . . . . . . . . . . . . . . . 333 C.10.Cost neutral state level subsidy results. . . . . . . . . . . . . . . . 334 C.11.Damage miminizing state level subsidy results. . . . . . . . . . . . . 335 C.12.Unconstrained state level subsidy results. . . . . . . . . . . . . . . 338 C.13.Counterfactual results under alternative model specifications. . . . . . . 339 C.14.Nationally-optimal subsidies with alternative household discounting. . . . 340 C.15.Percent of total subsidy value from each type of subsidy . . . . . . . . 341 C.16.Damages offset per additional dollar of government funds . . . . . . . 342 C.17.Nationally-optimal subsidies when accounting for line losses. . . . . . . 345 C.18.Nationally-optimal subsidies when accounting for transmission constraints. . . . . . . . . . . . . . . . . . . . . . . 346 C.19.Nationally-optimal cost-neutral and unconstrained subsidies with improved storage technology. . . . . . . . . . . . . . . . . . . 349 C.20.Nationally-optimal cost-neutral and unconstrained subsidies under alternative assumptions about central generation energy production. . . . . . . . . . . . . . . . . . . . . . . . . 351 24 CHAPTER I INTRODUCTION My dissertation provides insights into environmental policy-relevant questions. It contains previously published and unpublished co-authored material. Chapter II is joint with John Morehouse, Chapter III with Edward Rubin is published in the Proceedings of the National Academy of Sciences, and Chapter IV with Mark Colas is forthcoming in The Journal of Political Economy Microeconomics. Chapter II quantifies the distributional effects of climate change. The spatial distribution of the population is such that low-income and minority households face greater exposure to the changing climate. Further, there is evidence that the same disadvantaged households who face greater exposure to climate change are also less able to adapt. We construct a quantitative spatial equilibrium model where households choose where to live and, conditional on that choice, how much electricity, natural gas, and housing to consume. Thus, households are able to adapt via changes in their energy and housing consumption or by migrating. The climate affects the amenity value of each location and household energy demand. We find that welfare losses for the lowest income decile are twice as large as those of the highest income decile. Black households have 20% larger welfare losses relative to White households, and these occur throughout the income distribution. These gaps will continue to grow with further emissions and more extreme climate change. We perform a model-based decomposition and find the spatial distribution of the population and differential ability to migrate drive these gaps. Finally, we evaluate a policies that could be used to address these gaps, adding a place-based, means-tested subsidy to the model. Giving places with high climate exposure 25 the largest subsidies does a good job of targeting the households with the largest welfare losses from climate change but also incentivizes households previously living in other parts of the country to move into these high climate exposure areas. Chapter III investigates the perinatal health effects of glyphosate. The widespread adoption of genetically modified crops engineered for herbicide tolerance led to a dramatic rise in glyphosate use. Despite glyphosate use increasing by over 750% since the early 90s, we know relatively little about the human health effects of glyphosate. This chapter estimates the causal effect of glyphosate on several infant health outcomes: birthweight, gestation length, and indicators of low and very low birthweight. To identify the causal effect, we leverage temporal variation driven by the release of genetically modified seeds in 1996, paired with spatial variation in suitability for crops with genetically modified varieties. We find that glyphosate reduces average birthweight and gestation length and increases the probability of low and very low birthweight. We find considerable heterogeneity in these effects along the expected birthweight dimension. The most vulnerable births, with the lowest expected birthweight, have effects that are an order of magnitude larger than the least vulnerable births. Chapter IV studies the optimal design of spatially differentiated subsidies for residential solar panels. We build a structural model of solar panel demand and electricity production across the US and estimate the model by combining 1) remotely sensed data on residential solar panels, 2) power-plant-level data on hourly production and emissions, and 3) a state-of-the-art air pollution model. The current subsidies lead to severe spatial misallocation. National funding for subsidies under the current system exceeds the unconstrained optimum by over 26 70%. Our results suggest that there could be large welfare gains to redistributing funds towards other programs. 27 CHAPTER II THE DISTRIBUTIONAL IMPACTS OF CLIMATE CHANGE ACROSS US LOCAL LABOR MARKETS This chapter is co-authored with John Morehouse. We started working on this together while we were both graduate students at the University of Oregon. We have jointly worked on idea formulation, data collection, writing code, analyzing results, and writing the paper. 2.1 Introduction Climate change has led to adverse economic outcomes across the globe as households face rising temperatures, sea levels, and natural disaster risks (U. S. Global Change Research Program, 2023). However, expsore to climate change differs considerably across space, both globally and within the United States (Cruz & Rossi-Hansberg, 2023; Hsiang et al., 2017; Wing et al., 2022). For example, climate change has given cold areas like New England milder winters and hot regions like the Southeast more extreme summers. These changes have significant distributional implications, as hotter locations within the US are generally poorer and less white than the cold locations. Assessing the welfare effects of climate change requires accounting not only for exposure but also for households’ capacity to adapt. Adaptation may exacerbate disparities in climate-related damages across demographic groups if economically advantaged households have access to more effective mitigation strategies. Meanwhile disadvantaged households with limited resources to adapt may experience larger impacts even when facing similar levels of exposure (e.g., (Heilmann, Kahn, & Tang, 2021)). These differences in exposure and adaptive capacity may then deepen the environmental and social inequalities documented in 28 the environmental justice literature (Cain, Hernandez-Cortes, Timmins, & Weber, 2023). In this paper, we use a quantitative spatial equilibrium model of US local labor markets in the spirit of Rosen (1979) and Roback (1982) to quantify heterogeneity in the effects of climate change, allowing households to adapt through migration and home-energy use. In the model, heterogeneous households make static choices over (1) where they live and (2) how much energy and housing they consume. Potential wages earned, housing and energy costs, and amenities available in each city affect household’s location choices. Wages and rents are determined in equilibrium. Firms with varying productivity levels across cities demand labor, using college- and non-college-educated workers as imperfect substitutes. Each city has a housing supply curve, with rents responding endogenously to household location choices. Households adapt to climate change in two ways in the model—through home energy demand and migration. Air conditioning and residential energy use are critical forms of adaptation to climate change (Barreca, Clay, Deschenes, Greenstone, & Shapiro, 2016). We model demand for “comfort,” which households produce using electricity, natural gas, and housing. This notion of comfort production captures the process by which people combine a house’s physical characteristics and the use of appliances such as air conditioners or heaters powered by electricity or natural gas to create a comfortable indoor space (Quigley & Rubinfeld, 1989). The local climate is an additional input to the comfort production function, generating differences in the marginal productivity of electricity, natural gas, and housing. For example, if outdoor temperatures increase, households must use their air conditioners more intensely to maintain the same 29 indoor temperature, thus increasing the cost of maintaining that temperature. Recent energy justice work suggests that the energy costs associated with climate change will be higher for minority and low-income households (e.g., Lyubich (2020); Reames (2016)). We capture these observed differences in energy demand, allowing households to endogenously adjust their housing, electricity, and natural gas demand in response to changes in outdoor climate. The climate also affects the shared amenities of each location. These amenities capture how households value the climate when assessing the desirability of a location—for example, households may prefer milder summers. We characterize the amenity value of each city’s climate using heating and cooling degree days, precipitation patterns, and natural disaster risk. Spatial heterogeneity in exposure to climate change will alter some households’ relative rankings of locations—allowing them to mitigate the impact of climate change by moving to a new location (Cruz & Rossi-Hansberg, 2023). However, there is considerable heterogeneity in household mobility, with the literature consistently finding that lower-income households are less mobile than higher-income households (e.g., Bound and Holzer (2000); Colas and Hutchinson (2021); Depro, Timmins, and O’Neil (2015); Piyapromdee (2021)). Thus, low-income households may have a more difficult time migrating to mitigate the welfare effects of climate change. We quantify the model with publicly available Census data from Ruggles, Flood, Goeken, Schouweiler, and Sobek (2022), historical climate data from PRISM (PRISM Climate Group, 2022), and natural disaster risk data from First Street Foundation (First Street Foundation, 2022a, 2022b). We specify the household “comfort” production function and estimate its parameters using derived demand functions, which are affected by prices of each good and the climate. The estimated 30 comfort production function allows us to predict energy and housing demand in counterfactual climates. We then estimate the model’s key household location choice parameters using the two-step estimator proposed in Berry, Levinsohn, and Pakes (2004), including the amenity value of climate. The first step uses household-level location choices to estimate moving costs and the component of city-level utility common among households of the same demographic group, which we call the “mean utilities.” The second step decomposes the mean utilities into contributions from wages, rents and energy bills, and climate amenities. We use our estimated model to simulate the welfare effects from climate change across households in the US, both that have already occurred and that we predict to occur over the next century using state-of-the-art downscaled future climate models shared by the Climate Impact Lab (Gergel et al., 2023). We find significant distributional consequences of climate change both presently and in the future. Black and low-income households are particularly vulnerable, experiencing negative impacts not only due to their initial exposure to climate change but also due to disparities in their adaptive capacity. Black households have suffered 20% greater climate damages compared to white households from climate change to date, and we project this disparity to widen under future emissions scenarios. In the most extreme future emissions projections, we estimate that the Black- white damage gap will increase by more than eightfold by the end of the century. Additionally, the welfare losses from climate change accrue disproportionately to lower-income households. The lowest income decile has welfare losses twice as large as the highest income decile. Notably, the Black-white gap remains throughout the income distribution. 31 We then decompose the factors contributing to the unequal impacts. First, we simulate a counterfactual where we fix household locations, wages, and rents. We call this the “Mechanical” effect since it does not allow households to reoptimize across space, thus capturing differences in exposure and comfort costs. However, we can eliminate differences in exposure by comparing welfare within cities, therefore isolating differences in comfort costs. Next, we simulate counterfactuals where households can sort across locations, but we fix wages and rents at their baseline levels. We call this the “Sorting” effect as it reflects how differences in mobility impact inequality relative to the Mechanical effect. We can again condition on location in order to isolate differences in mobility from differences in exposure. Finally, we allow complete flexibility—households can sort across locations, and all prices are determined in equilibrium—to demonstrate the effect of endogenous changes in rents and wages on the welfare effects of climate change. Our model decomposition reveals that the ex-ante distribution of the population is the most significant contributor to these gaps, followed by differences in mobility and, to less of an extent, differences in energy efficiency. When we fix locations, we find that Black households are worse off than white households by 0.8% of income. However, this gap disappears after conditioning on baseline location, suggesting that differences in comfort costs are not contributing to the average gap. When we allow households to sort across locations, the average gap remains at the same level as when fixing locations, while the gap conditional on baseline location increases to 0.5% of income. This result suggests that Black households are less able to mitigate welfare losses through migration relative to similarly exposed white households. 32 We then use the model to evaluate the effects of a means-tested, place- based policy inspired by the US Environmental Protection Agency’s Community Change Grant program. The Inflation Reduction Act (IRA) allocated $3 billion to the Community Change Grant program, which funds projects in disadvantaged communities with the stated goal of improving climate resiliency (US EPA, 2023). Our results suggest that while the distribution of funds under the current program helps disadvantaged households generally, it does not target households with the largest climate damages. We then test alternative spatial distributions of the subsidies. Subsidizing places with high climate exposure benefits the households with the greatest losses from climate change. However, these subsidies attract marginal households from elsewhere in the country to high-climate-damage cities, thus increasing total exposure to climate change, increasing rents, and decreasing wages in equilibrium. This spatial reallocation results in some of the intended benefits of the program dissipating as deadweight loss (Kline & Moretti, 2014). Alternatively, we test distributing the subsidies to “climate havens,” places where climate amenities have improved. While this distribution of subsidies reduces exposure to climate change, most of the funds go to households that are already relatively well off from a climate perspective. Policymakers must weigh the benefits of spatially targeted, place-based subsidies with their associated effect on the distribution of households across space, thus increasing climate exposure and resulting in deadweight loss. Our paper follows a rich literature using spatial equilibrium models to estimate the economic impact of the climate or other environmental goods (e.g., Albouy, Graf, Kellogg, and Wolff (2016); Bayer, Keohane, and Timmins (2009); Hamilton and Phaneuf (2015); Wrenn (2023)). In particular, Albouy et al. 33 (2016) estimate the amenity value of days across the temperature distribution in the US, a methodology rely on in our own estimation approach. Wrenn (2023) estimates household marginal willingness to pay for natural disasters using a Rosen- Roback model but does not run counterfactuals or explore environmental justice considerations. Another set of papers considers migration an adaptation mechanism to climate change in global spatial equilibrium models (e.g., Cruz and Rossi- Hansberg (2023); Desmet et al. (2021); Desmet and Rossi-Hansberg (2015)), but is more focused on quantifying differences in the welfare impacts across space rather than by demographic group. Rudik, Lyn, Tan, and Ortiz-Bobea (2021) estimate a dynamic spatial equilibrium model with daily temperature affecting amenities and firm productivity. Their estimation strategy relies on migration flows between states, limiting their ability to quantify differential effects between households. However, they incorporate additional mechanisms through which the climate affects the economy—namely it’s effect on firm productivity. Thus allowing for additional adaptation through trade and sectoral switching. We complement their work by formally analyzing the distributional effects of climate change, modeling heterogeneity by race, education, and income. This paper is also related to work about the effects of climate on energy use (e.g., Auffhammer (2022); Davis and Gertler (2015); Doremus, Jacqz, and Johnston (2022); Rode et al. (2021)). This work inspires our strategy for estimating the impact of climate on energy demand. Auffhammer and Mansur (2014) emphasize a gap in our understanding of climate change’s long-run extensive margin effects on energy demand. Our estimation strategy is able to capture these extensive margin effects, as it reflects changes in energy use due to cliamte over a nearly 30 year period. Notably, we also address two concerns with the existing intensive margin 34 literature, (1) we explicitly deal with sorting across locations and (2) we do not assume a constant interior temperature to measure welfare effects. We also contribute to the environmental justice literature, recently reviewed in Cain et al. (2023). Several papers in this literature study the interaction between household migration decisions and environmental justice (e.g., Bayer et al. (2009); Depro et al. (2015); Hausman and Stolper (2021)). These analyses focus on how sorting contributes to differences in measured exposure to environmental pollutants, as described in Banzhaf, Ma, and Timmins (2019a). We extend this sorting mechanism to climate change. Methodologically, our model builds on the static quantitative spatial equilibrium literature with heterogeneous workers (e.g., Colas and Morehouse (2022); Diamond (2016); Piyapromdee (2021)). These papers analyze other settings with spatial consequences, allowing rents and wages respond endogenously in general equilibrium to household location choices. Using a similar framework, Morehouse (2022) demonstrates that a fear of distributional consequences is first- order concern for enacting serious climate-change regulation (e.g., a carbon tax). Here, we show that there are also significant distributional effects from a lack of climate policy. 2.2 Data and Descriptive Evidence We begin by describing the main data sources used throughout the analysis. We break these into three categories: individual household data, climate data, and energy use data. We then show descriptive evidence of the spatial heterogeneity in climate change to-date and how the changes correlate with demographics. 2.2.1 Data. 35 Household Data. We use repeated cross-sections from the 1990 and 2000 censuses and the 2010 and 2019 5-year aggregated ACS surveys (Ruggles et al., 2022). These give us household demographic information, city, income, employment status, housing, electricity, and natural gas expenditures. We follow standard sample selection and data-cleaning techniques described in Appendix A.1. From these data, we estimate city-education-level wages, city-level rents, and city- demographic group-level energy expenditure for each year. As these indices are standard in the literature, we leave the details in Appendix A.1.2. Climate Data We collect historical weather data from the PRISM climate group (PRISM Climate Group, 2022). Specifically, we use the 4-kilometer grid of daily average temperature and precipitation to create a panel of daily temperature and precipitation for each of the CBSAs in our model.1 We aggregate the gridded data to the CBSA level using the weighted average of grid cells within each CBSA, where we weight by the fraction of the CBSA covered by the grid cell and the 1990 population in the grid cell. Population rasters come from SEDAC (CIESIN, 2017). We then construct various annual summary measures of climate in each city. Following the recent literature, we count the number of days each year in a set of discrete temperature bins (Auffhammer, 2022). We also calculate the percent of days with no rain as those with less than one mm of precipitation. Since, in some cases, we are interested in the effects of changes in climate, not short-run weather shocks, we take the 5-year moving averages of each of our annual weather variables. 1Since the PRISM data only cover the continental US, we collect daily weather station observations from NOAA for stations in Honolulu (Menne et al., 2012), then take the population- weighted average of those observations. 36 We pair this factual climate data with future climate simulations from the Climate Impact Lab’s global downscaled projections for climate impacts research, CIP-GDPCIR (Gergel et al., 2023). These data result from downscaling and debiasing2 25 climate models for four emissions scenarios, each with daily precipitation and maximum and minimum surface temperature to 2100. We aggregate these to CBSA-level values by taking a population-weighted average as we do with the PRISM data. We then take the average of the 25 climate models to form an ensemble prediction of daily average temperature and total precipitation. Finally, we use data on natural disaster risks—specifically, First Street Foundation’s fire and flood scores aggregated to the zip-code level (First Street Foundation, 2022a, 2022b). They use state-of-the-art fire and flood models to calculate a risk score between 1 and 10 associated with the respective disasters for every property in the US. We calculate various aggregations of the scores for each CBSA—for example, the median score or the percent of properties in specific score ranges. Appendix A.2 has maps of the average risk scores by census tract. Unfortunately, we only have access to a single set of scores—we use these scores in estimation but cannot vary them in our counterfactual simulations. Energy Data. We supplement data on household energy expenditure from the census with state-level annual electricity prices from the EIA. We use these prices to back out energy usage from total expenditures. To address the concern that some households may not report energy expenditures separately from rent3, we follow Glaeser and Kahn (2010) in using the Residential Energy Consumption 2The original models produce output on a 1-degree grid, and the CIP-GDPCIR data downscale this to a 0.25-degree grid. 3If their utilities are included as part of rent, for example. 37 (a) Change in the number of hot days per year (b) Change in the number cold days per year Figure 1. Change in the number of hot and cold days between 1990 and 2019. Hot days are those with an average temperature above the 90th percentile of the temperature distribution, or about 80F. Cold days are below the 10th percentile, or about 32F. Difference taken between 5 year moving averages and are censored at the 5th and 95th percentiles of grid cells. Survey (RECS) to correct this issue. The details of this estimation are in Appendix A.1.2.1. We first estimate city-demographic group-level energy use for single-family homeowners from the Census and ACS data, as those households are most likely to report their expenditures accurately. We then use the RECS to estimate the relationship between energy use for single-family homeowners and multi-family homeowners, single-family renters, and multi-family renters. 2.2.2 Descriptive Evidence. Geography of climate change Climate change is often talked about as increases in average global temperatures, but this masks significant spatial heterogeneity in the degree and impact of change. Figure 1 shows the change in the number of hot and cold days, above the 90th and below the 10th percentiles of the 1990 temperature distribution, respectively, between 1990 and 2019. The hotter portions of the US have seen significant increases in hot days, while the cold places 38 in the US have seen significant decreases in cold days. Appendix A.2 shows that the eastern US has generally gotten wetter—higher annual precipitation and fewer days with no rain. Heterogeneity in the impact of climate change is important because it means climate change alters the relative attractiveness of cities, not just the absolute level across all cities. Some cities, such as those in the Northeast, may have become more attractive to households as their winters are now milder than previously. Other cities, such as Oklahoma City or Dallas, might have become less attractive now that their already-hot summers have become hotter. These distortions in climate will lead to changes in the share of the population that chooses to live in each city in the long run, with households taking climate into account when considering the utility they would get from living in a particular city. Demographics of climate change We are interested in determining whether climate change has differentially affected demographic groups. Here we look at how variation in demographic makeup between cities in 1990 compare to changes in the climate between 1990 and 2019. By looking at demographics in 1990, we avoid endogenous city-choice sorting that has resulted from changes in the climate. Figure 2 shows a strong positive association between the share of a city’s population in the lowest income quintile in 1990 and how cold and hot days have changed since 1990. Cities with larger shares of low-income households have seen the largest increases in hot days and little change in the number of cold days. Meanwhile, cities with a smaller share of low-income households have seen bigger decreases in cold days and a smaller increase in hot days. Both of these suggest that climate change is regressive—cities with more poor people have had changes in their climate that require spending more on heating and cooling to maintain 39 Figure 2. Relationship between income and climate change. Each point is a city, where the size of the point represents the city’s 1990 population. Low income defined as households in the bottom quintile of the national income distribution in 1990. Regression lines are weighted by population. the same indoor temperature relative to cities with fewer poor people. A similar relationship holds for the share of non-white households in cities, as seen in Figure 3. Appendix A.2 also demonstrates that a similar relationship holds for total precipitation. We can also summarize the change in climate experienced by the average white and non-white household. A common measure of cooling demand is cooling degree days (CDD), which is calculated based on the difference between the daily average temperature and 65◦F, summed over all days above 65◦F. The average non-white household experienced 38 (5.1%) more CDDs than the average white household in 1990. Between 1990 and 2019, CDDs increased by 110 (14.7%) for white households and 124 (15.9%) for non-white households. Thus, the average 40 Figure 3. Relationship between share non-white and climate. Each point is a city, where the size of the point represents the city’s 1990 population. Low income defined as households in the bottom quintile of the national income distribution in 1990. Regression lines are weighted by population. non-white household experienced an additional 14 (13.5%) CDD increase relative to white households, while already having a higher baseline CDD level. These results suggest that disadvantaged demographic groups have faced disproportionate exposure to the adverse effects of climate change. Lower-income and less white cities have more hot days, while higher-income and more white cities have fewer cold days. However, the welfare effects of climate change may be more or less equitable once we account for adaptation. While the ability to adapt is greater among more economically advantaged households, households that are more exposed have a greater need for adaptation. Energy use and demographics There is robust evidence from the energy justice literature that Black households spend more on energy than observably similar white households (Lyubich, 2020; Reames, 2016). This gap suggests 41 that there may also be differences across demographic groups in the response of household energy demand to climate. We explore this in depth in Appendix A.2.1. We first replicate the results in Lyubich (2020), finding a conditional energy expenditure gap between Black and white households throughout the income distribution. We then find that energy expenditures for Black households are more responsive to additional cold weather and less responsive to additional hot weather than white households. 2.3 Model In this section, we develop a Rosen (1979) and Roback (1982) style spatial equilibrium model where households make a static, discrete choice of where to live. Conditional on that location, households pick consumption of electricity, natural gas, housing, and the numeraire. Cities vary in their wages, rents, and amenities. The climate impacts household choices through two mechanisms: the amenities available in each city and the marginal benefits of electricity, gas, and housing consumption. Amenities in the model reflect non-market characteristics of cities that make them more or less desirable, including the climate. We use a rich characterization of the climate as amenities to capture a broad range of values that households have for the climate, for example, outdoor recreation or the risk of physical damage from natural disasters. Additionally, households have an imperfect ability to transform the outdoor environment into a comfortable indoor environment with housing and heating or cooling appliances. In the model, households use housing, electricity, and natural gas as inputs to produce indoor “comfort”.4 The climate affects the marginal benefits of inputs to the comfort production function. For example, if 4We use a similar notion of comfort to that found in Quigley and Rubinfeld (1989). 42 there are fewer cold days in a year, households require less natural gas to heat their homes, which we capture as a decrease in the marginal benefit of natural gas use. We allow for rich heterogeneity in the parameters governing household mobility. We specify an individual-level moving cost function to capture the psychic costs of relocating.5 Additionally, households have a location-specific idiosyncratic preference shock that allows us to relax the spatial indifference condition of traditional Rosen-Roback models.6 The variance of this preference shock governs the elasticities of household location choices. A higher variance means that households have stronger individual preferences for locations relative to the shared components of utility. Firms in each city produce the tradable numeraire with imperfectly substitutable skilled and unskilled labor. Each city also has an upward-sloping housing supply curve, the slope of which reflects differences in difficulty in developing new housing. These firms and housing supply curves complete the labor and housing markets that allow wages and rents to respond endogenously to household location choices.7 2.3.1 Households. Let j ∈ J index cities and d ∈ D index demographic groups. We omit t subscripts for exposition, but re-introduce them when describing the estimation. Household i living in city j receives utility from a composite numeraire good, X, “comfort in dwelling” C, and location-specific amenities A. We specify this utility in traditional Cobb-Douglas form, 5Bayer et al. (2009) emphasize the importance of including these moving costs. 6The spatial indifference condition requires households to receive the same utility from all cities in equilibrium. Busso, Gregory, and Kline (2013) and Kline and Moretti (2014) demonstrate the importance of relaxing such a constraint. 7We define an equilibrium for our model in Appendix A.3.6. 43 Uij = X1−αC dCαC d dj exp(Aij). Households produce comfort using electricity, gas, and housing, which we specify as a nested CES production function. This production function captures the utility services provided by electricity, gas, and housing. For example, electricity does not give people utility directly, but an air conditioner can use electricity to provide cool indoor temperatures in a bedroom. Specifically, let comfort C be: C(H, E|d, j) = (Hρc + Eρc)1/ρc . (2.1) where H is housing and E is energy. The measure of housing, H, represents both the quantity and quality of housing—i.e. both an additional room in a house and a more recently built house could be interpreted as higher values of H. σC = 1 1−ρC is the elasticity of substitution between housing and energy. This aggregation reflects the imperfect complementarity or substitutability between energy and housing. For example, larger houses may require more energy to heat or cool, or a newer home may be better insulated, requiring less energy to heat or cool, all else equal. E(·) is the household’s energy production function, which we parameterize as E(E,G|d, j) = ( θEdjE ρE dj + θGdjG ρE dj )1/ρE , (2.2) where σE = 1 1−ρE is the elasticity of substitution between electricity, E, and natural gas, G. Climate impacts the production of comfort through θmdj , where θmdj = f(Zj;κ m d ) for m ∈ {E,G}, where Zj is a vector of local climate variables and κm d is a parameter vector governing the potentially non-linear shape of f(·). Since we do not normalize these factor intensities to sum to one, we can think of them as 44 affecting the productivity of electricity and gas in both relative and absolute terms, as well as the importance of energy relative to housing. We provide more intuiton on the effect of climate on housing and energy demand in Section 2.3.1.1. Finally, we specify amenities as Aij = αZ d ·Zj + ξdj + g(j, bi) + σdεij, where αZ d is a vector of parameters on a vector of climate variables Zj, ξdj is a shared, unobservable component of amenities, g(j, bi) are moving costs as a function of the household’s birth state, bi and location j, and εij is an idiosyncratic preference draw with variance σd. A higher variance in the idiosyncratic utility parameter reduces household mobility, as larger changes in the other components of utility would be required to overcome the idiosyncratic term. We parameterize moving costs, g(·), as: g(j, bi) = γstd I ( j ∈ bsti ) + γdistd ϕ ( j, bsti ) + γdist2d ϕ2 ( j, bsti ) , (2.3) where I (j ∈ bsti ) is an indicator for j being in household i’s birth state, ϕ(j, bsti ) is the Euclidean distance between location j and the agent’s birth state bsti , and ϕ2(j, bsti ) is the squared Euclidean distance between j and bsti . Thus, households get a utility premium, γstd , for living in their birth state, and pay an increasing utility cost the further they move from their birth state, captured by the quadratic in distance. Households are subject to the following budget constraint, conditional on choosing location j, 45 Idj +Υd = X + PE j E + PG j G+ PH j H, where Idj is earned income for demographic group d in city j, Υd is unearned income for demographic group d, PE j and PG j are prices of electricity and gas, and PH j is the price of housing. We normalize the price of the composite to one. We can find the household’s utility maximizing choices of electricity, gas, housing, and the numeraire by first solving the nested cost minimizations associated with comfort production, the details of which are in Appendix A.3.1. Households choose the cost-minimizing combination of electricity and gas to produce a given amount of energy, which we can then use to solve for a unit cost function for energy. We call this unit cost function the ”price of energy”, P E dj = ( θEdj σEPE j 1−σE + θGdj σEPG j 1−σE ) 1 1−σE . (2.4) Given this price of energy, households choose a cost-minimizing combination of housing and energy to produce a given amount of comfort. Similarly, we can use the conditional demand functions for housing and energy to solve for the unit cost function of comfort, which we’ll call the ”price of comfort,” P C dj = ( PH dj 1−σc + P E dj 1−σc ) 1 1−σc . (2.5) The price of comfort is a useful and rarely quantified object that reflects not just the total costs of housing and energy, but also captures how households ability to substitute between housing, electricity, and gas when they face changes in relative prices or productivities. 46 We can now transform the problem into a simpler form where, conditional on their location choice, households choose comfort and composite good consumption subject to a simplified budget constraint, max X,C X1−αC dCαC d exp(Aij) s.t. Idj +Υd = X + P C djCdj. This gives us the familiar demand functions for consumption and comfort, X∗ dj = (1 − αc dj)(Idj + Υd) and C∗ dj = αc dj(Idj + Υd)/P C dj. These demand functions yield the following conditional indirect utility function after taking logs, vij = log (Idj +Υd)− αC dj log(P C dj) +αZ d ·Zj + ξdj + g(j, bi) + σdεij. (2.6) Assuming εd is distributed type 1 extreme-value, we can write the probability household i chooses location j as: Pij = exp (ṽij)∑ j exp(ṽij) , (2.7) where ṽij = vijσ −1 d − εij is the non-idiosyncratic portion of utility. 2.3.1.1 The impact of climate change on household decisions. In order to clarify the mechanisms through which climate affects household choices in the model, we provide some partial equilibrium derivations here. Consider the effect of climate variable zlj ∈ Zj on household conditional indirect utility, holding income, rent, electricity, and gas prices fixed: ∂vij ∂zlj = αl d − αC d P C dj × ∂P C dj ∂zlj . (2.8) There are two channels through which a change in climate affects utility, the amenity value of a location and the price of producing comfort. The change 47 in amenity value of city j is reflected by αl d ∈ αZ d . If the change in climate is not desirable, such that αl d is negative (i.e., temperatures get hotter in the summer), then the utility households get from living in city j decreases. The second term αC d PC dj × ∂PC dj ∂zlj reflects the effect of the change in the cost of producing comfort inside one’s home. Suppose, that the climate variable is one that affects cooling demand, where ∂θE/∂zl = κE is the marginal effect on the electricity share parameter and ∂θG/∂zl = 0. Taking the derivative of the price of comfort from equation (2.5), we have ∂P C dj ∂zlj = −κE ρE PE j Edj Cdj . (2.9) The sign of this derivative hinges on the elasticity of substitution between electricity and gas, σE = 1 1−ρE . If electricity and gas are gross substitutes, such that ρE > 0 and σE > 1, then ∂P C dj/∂z l j < 0. If electricity and gas are gross complements, such that ρE < 0 and σE < 1, then ∂P C dj/∂z l j > 0. In the gross complements case, increases in the intensity parameters θE or θG caused by changes in the climate lead to a decrease in produced energy for given levels of electricity and gas, thus decreasing the productivity of the household’s comfort function, and increasing the price of comfort. Plugging equation (2.9) into (2.10) and simplifying, we get ∂vij ∂zlj = αl d + κE ρE PE j Edj Idj +Υd . (2.10) 48 We generally expect both of the terms in Equation (2.10) to have the same sign, since a change in climate that makes outdoor amenities less desirable typically also makes it more expensive to produce comfort inside.8 Importantly note that ∂vij/∂z l j is decreasing in income Idj and increasing in electricity use Edj. Thus, all else equal, households with lower incomes, or higher electricity usage conditional on income will be more affected by changes in the climate. Appendix A.3.2 has derivations for the marginal effect of climate on electricity, gas, and housing demand. Now, we consider the effect of the climate variable on a household’s city choice probability Pij. Taking the derivative of equation (2.7), again holding income, rent, electricity, and gas prices constant, we have, dPij dzlj = 1 σd ∂vij ∂zlj Pij(1− Pij). (2.11) This derivative makes it clear that larger values for variance of the idiosyncratic utility shock, σd decrease mobility. Higher values of σd mean that the unobserved, idiosyncratic portion of utility an individual has a higher relative weight as compared to wages, rents, and shared amenities. If an individual has a particularly strong affinity for a city, for example due to their social connections or cultural ties, this will make them less sensitive to changes in other aspects of the city, such as wages, rent, or the climate. Note that in general equilibrium, a change in the climate in one city can affect indirect utility in all other cities through endogenous changes in wages and rents. Colas and Saulnier (2023) provide a general derivation 8The sign of the second term is determined by the sign of κE and ρE , the rest of the variables in the second term will always be positive. Using our previous example, κE > 0 captures hotter weather requiring more electricity maintain comfort. Thus, it would be opposite sign as αl d. If electricity and gas are gross complements, then ρE < 0. 49 for this case—the intuition remains the same, higher values of the variance of idiosyncratic portion of utility decrease responsiveness to the shared components of utility. 2.3.2 Firms. Firms competing in perfectly competitive markets combine college-educated labor and non-college-educated labor to produce the composite numeraire good. We parameterize the firms’ production function as: Yj = BjK α j L1−α j , (2.12) where Bj is city-specific total factor productivity, Kj is capital use, and Lj is a CES aggregator between the two labor types. More specifically, Lj = ( λjS ρl j + (1− λj)L ρl j ) 1 ρl , (2.13) where λj is the college labor input use intensity in city j, Sj is efficiency units of college labor, Lj is the efficiency units of non-college labor, and ρl = σl−1 σl is the elasticity of substitution between college and non-college labor. Assuming capital supply is perfectly elastic and supplied on an international market (at rate r̄), we can write the firm’s optimal capital demand as K⋆ j = BjαYj r̄ . Plugging this into the firm’s production function and then solving for the first order-conditions yields the firm’s labor demand curves for college- and non-college educated labor: 50 W S j =B̃jL1−ρl j λjS ρl−1 j WL j =B̃jL1−ρl j (1− λj)L ρl−1 j (2.14) where B̃j = (1 − α)B 1 1−α j ( α r̄ ) α 1−α . Appendix A.3.3 has additional details on the firm’s profit maximization problem. 2.3.3 Rents. Each city is characterized by a long-run, upward sloping rental supply curve, which we parameterize as log(P h j ) = aj + ζj log(Hj), (2.15) where P h j is the price of housing in city j, aj is a city-specific intercept capturing construction costs, ζj is the inverse elasticity of housing supply, and Hj is the total housing demand in the city. The intercept, aj, varies across cities due to differences in baseline construction costs, such as the cost of building materials or labor. The elasticity, ζj, captures how responsive housing prices are to changes in the quantity of housing demanded. The availability of land for development and local regulatory constraints, such as zoning laws and land-use restrictions, affect the inverse elasticity—more restrictive housing development conditions lead to steeper increases in housing prices. We assume that rental profit, Π, is distributed back to landowners lump- sum as unearned income, Υd = sdΠ, where sd is demographic group d’s share of investment income. Rental profits are Π = ∑ j ( PH j Hj − exp aj ζj + 1 H ζj+1 j ) (2.16) 51 2.4 Estimation In this section, we detail our estimation procedure. We focus our exposition on the estimation of the household demand and labor supply parameters. We use city-level energy, wage, and rent indices in estimating various parts of the model. Details for how we construct these indices can be found in Appendix A.1.2. 2.4.1 Energy and Housing Demand. We estimate the parameters of the household’s nested CES comfort function in two steps. First, we use the relative demand functions to estimate the elasticity of substitution between electricity and gas, σE , and then the elasticity of substitution between energy and housing, σC in first differences. Then, we use the implied values for the θ’s to estimate the effect that climate has on comfort production. 2.4.1.1 Elasticities of substitution. We first derive the first- order conditions from households choosing the cost minimizing combination of electricity and gas to produce a certain quantity of energy. The details of this cost- minimization problem are in Appendix A.4.1, which yields the following relative demand function for electricity and gas, log ( Edjt Gdjt ) = σE log ( PG jt PE jt ) + σE log ( θEdjt θGdjt ) . (2.17) Since we observe quantities demanded and prices for both electricity and gas, we can estimate σE using linear regression. We take first differences of Equation (2.17) to arrive at our estimating equation, ∆ log ( Edjt Gdjt ) = σE∆ log ( PG jt PE jt ) +∆udjt, (2.18) 52 As we are concerned with endogeneity when regressing quantities of electricity and gas demanded on the prices of electricity of gas due to the simultaneity of supply and demand, we follow Burke and Abayasekara (2018) and instrument the log ratio of retail electricity and natural gas prices with the log ratio of lagged commercial electricity and natural gas prices, where Pm,Comm jt for m ∈ {E,G} is the price of electricity or natural gas for commercial customers. Formally, the exclusion restriction is E [ ∆ log ( PG,Comm j,t−1 PE,Comm j,t−1 ) ∆udjt ] = 0, (2.19) where udjt is any other factor affecting the change in the ratio of electricity and gas demand. By instrumenting with lagged commercial prices, we use variation in in energy prices that is driven by supply side factors and not correlated with other factors affecting residential energy demand, including the climate. We can then back out θEdjt/θ G djt using equation (2.17), the estimate of σE , and our data on the prices and quantities of both electricity and gas: log ( θEdjt/θ G djt ) = 1 σE log (Edjt/Gdjt) − log ( PG jt /P E jt ) . This allows us to calculate normalized versions of energy, Ẽdjt, and the price of energy, P̃ E djt, where we have normalized energy by θGdjt 1/ρE .9 Next, we estimate σC, the elasticity of substitution between housing and energy from the comfort production function in a similar manner. Agents produce comfort in their dwelling using housing and energy as inputs.10 Once again, we use 9These are Ẽdjt = (( θEdjt/θ G djt ) EρE djt +GρE djt ) 1 ρE and P̃ E dj = (( θEdj/θ G dj )σE PE j 1−σE + PG j 1−σE ) 1 1−σE . 10We estimate the quantity of housing, H, by regressing gross rent reported in the data on a set of observable characteristics about the dwelling and a city fixed effect in logs. We interpret the city fixed effect as the cost per unit of housing, and then back out the quantity of housing by dividing gross rent by this per unit cost of housing. Appendix A.1.2.3 explains this process in detail. 53 the relative conditional demand functions, but now for energy and housing, that result from households choosing the cost-minimizing combination of housing and energy to produce a given amount of comfort, log ( Ẽdjt Hdjt ) = σc log ( PH jt P̃ E jt ) + σc ρc ρE log θGdjt. (2.20) Taking first differences yields our estimating equation, ∆ log ( Ẽdjt Hdjt ) = σC∆ log ( PH jt P̃ E jt ) +∆ϵdjt. (2.21) Note that we do not observe amenities, and many amenities are correlated with both housing demand and rent. For example, consider a city with higher quality restaurants. These restaurants will cause increases in housing prices and housing demand, confounding our parameter estimates. Therefore, we use instrumental variables to isolate exogenous variation in the ratio of rent to energy price. Our instrument interacts labor demand shocks with land availability, where the labor demand shocks are from Katz and Murphy (1992), and the interaction with land availability follows Diamond (2016). We construct the labor demand shocks by multiplying the historical share of workers with the change in the number of workers in the rest of the country. Formally, the labor demand shock is ∆Zdjt = ∑ ι∈n ω1990 djι × (∆Nd,−j,ι,t) , (2.22) where ω1990 djιt is the share of workers in demographic group d from city j that worked in industry ι in 1990 as a fraction of the total number of workers across all industries in demographic group d in city j. ∆Nd,−j,ι,t is the change in the number of workers in industry ι between years t and t − 1 outside of city j. We 54 interact this with the measure of housing supply elasticity from Saiz (2010), which reflects geological and bureaucratic constraints on space to build homes in a city. The exclusion restriction is E [∆Zdjt∆ϵdjt] = 0 (2.23) E [ Elasticityj∆Zdjt∆ϵdjt ] = 0. (2.24) Any factors that effect relative energy and housing demand besides the relative prices are included in ϵdjt, this includes climate and other amenities as described above. The instruments—labor demand shocks interacted with a static measure of land availability in the city—are unlikely to be correlated with changes in these other factors. We can then use our estimate of σC, data on quantities and prices of housing and (normalized) energy to back out θGdjt using equation (2.20). Since we already have estimates of log ( θEdjt/θ G djt ) from the first step, we are able to calculate both θGdjt and θEdjt. 2.4.1.2 Effect of climate on comfort production. The next step of our estimation is to measure the effect of climate on the θmdjt’s, capturing how climate effects the productivity of electricity, gas, and housing. Note that while we normalize both parameters in the outer nest of the comfort production CES to be equal to one, we place no restrictions on θGdjt and θ E djt in the inner nest. Thus, θGdjt and θEdjt tell us not only the relative importance of electricity and gas, but also the importance of energy relative to housing. We parameterize the effect of climate on these intensity parameters as: 55 log θEdjt = ∑ τ κE(τ)Dτjt + ψE j + ψE d + νEdjt (2.25) log θGdjt = ∑ τ κG(τ)Dτjt + ψG j + ψG d + νGdjt, (2.26) where Dτjt is the number of days at temperature τ in city j and year t, ψj are city fixed effects, ψd are demographic group fixed effects, and the νdjt’s are random disturbances. κm(τ) for m ∈ {E,G} is the marginal effect of an additional day at temperature τ on electricity or gas. As we have limited data, we choose to allow for flexible non-linearity with relatively few parameters by estimating the κ(τ)’s with cubic splines. Specifically, for m ∈ {E,G}, let κm(τ) = ∑ s κ m s Ss(τ), where Ss(τ) for s ∈ {1, . . . , S} are the standard basis functions of a cubic B-spline of degree S. Substituting this into the above parameterization of the θ’s gives us our two estimating equations, log θEdjt = ∑ s κEs (∑ τ Ss(τ)Dτjt ) + ψE j + ψE d + νEdjt (2.27) log θGdjt = ∑ s κGs (∑ τ Ss(τ)Dτjt ) + ψG j + ψG d + νGdjt, (2.28) Thus, we identify the marginal impacts of climate on comfort demand using random deviations of weather outcomes from the long-run average climate within cities. For robustness, we use various degrees for the spline, as well as alternative parameterizations of climate. 2.4.2 Labor Supply. We quantify the rest of the parameters in the household utility function using maximum likelihood estimation and then linear regression. First, we define the mean utility associated with each choice as: 56 δdjt = βI e log(Idjt +Υdt) + βc e log(P c djt) + βZ e · Zjt + ξdjt, (2.29) where βI e = 1 σe , βc e = −αc e σe and βZ e = αZ e σe . We then write the household’s choice probability as: Pijt(δdjt;γdt) = exp ( δdjt + γ̃stdtI (j ∈ bsti ) + γ̃distdt ϕ (j, b st i ) + γ̃dist2dt ϕ2 (j, bsti ) )∑ j′ exp ( δdj′t + γ̃stdtI (j′ ∈ bsti ) + γ̃distdt ϕ (j ′, bsti ) + γ̃dist2dt ϕ2 (j′, bsti ) ) , (2.30) where the tildes over the parameters indicate they are normalized by the variance of the idiosyncratic preference shock, σe. Given these choice probabilities, the likelihood function is: LL(δdjt;γdt) = ∑ i∈d ∑ j Iidjt log(Pijt(δdjt;γdt)), (2.31) where i ∈ D is the set of all households in demographic group d and Iidjt is an indicator equal to one if the household chose city j in the data. We recover the δdjt’s conditional on γdt using the contraction mapping proposed in Berry et al. (2004). We then decompose δdjt from step one to estimate climate amenities. First, we use estimates from Diamond (2016) to calibrate βw e to 5.22 for college and 4.15 for non-college. Additionally, we use the observed comfort expenditure shares for each city, demographic group, and year to calibrate βc djt. With these parameters, we calculate the residual deltas as δ̂djt = δdjt − ( βI e log(Idjt) + βc djt log(P c djt) ) . We then estimate the effects of climate on the mean-utilities using a rich specification of the climate that includes heating and cooling degree days, total precipitation, 57 the number of days without precipitation, and measures of fire and flood risk. Our estimating equation is δ̂djt = βZ ·Zjt + βx ·Xjt + ωd + ωt + ξdjt, (2.32) where Zjt = {Firej,Floodj,CDDjt,HDDjt,Precipjt,NoRainjt} is a vector of climate variables. Firej and Floodj are median fire and flood risk scores in city j that are static over time. CDDjt and HDDjt are heating and cooling degree days. Precipjt is total precipitation and NoRainjt is the proportion of days with no precipitation. Xjt is a vector of geographic and demographic controls including average elevation, average slope, percent of population married, average age, and percent of population Black in each city and year. ωd and ωt are demographic group and year fixed effects. This is a parisimonuous, but rich characterization of the climate that allows us to capture the nuanced heterogeneity in the effects of climate change. 2.5 Parameter Estimates Comfort Demand. Here, we report our parameter estimation for comfort demand. First, Table 1 shows the results from estimating Equations (2.18) and (2.21) with instrumental variables, where the columns of the table have results for alternative fixed effects included in the regressions. While the results are presented together for concision, they were run as two separate regressions. Our preferred specification is in column 1, with demographic group and city fixed effects. We estimate that electricity and gas are net complements, as we estimate the elasticity of substitution, σE , to be less than one. Intuitively, a house that requires more electricity for cooling also requires more gas to heat. The instrument for the log ratio of electricity and gas prices, which is the log ratio of lagged 58 Model: (1) (2) (3) (4) (5) Parameters σE 0.501∗∗∗ 0.525∗∗∗ 0.439∗∗∗ 0.432∗∗∗ 0.432∗∗∗ (0.063) (0.066) (0.070) (0.086) (0.087) σC 0.594∗∗∗ 0.463∗∗∗ 0.316∗∗∗ 0.293 0.570 (0.079) (0.091) (0.092) (0.348) (0.436) Fixed-effects Dem Group Yes Yes Yes City Yes Yes Year Yes Yes Fit statistics Observations 1,260 1,260 1,260 1,260 1,260 F-test (1st stage), log (Pg/Pe) 378.82 363.51 392.65 391.88 391.25 F-test (1st stage), log (Ph/PE) 147.47 65.63 143.79 1.78 0.90 Clustered (City) standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 The elasticity of substitution between electricity and gas is σE and the elasticity of substitution between energy and housing is σC. Coefficients are shown in the same column, but come from separate regressions. Regressions are weighted by 1990 population. We exclude the division aggregation CBSA’s in estimation. Table 1. Elasticity of substitution estimates. 59 commercial electricity and gas prices, is strong, with an F-stat of well over 300 in our preferred specification. Housing and energy are net complements as well. Suppose the price of housing goes down relative to energy prices. In that case, the housing quantity demanded will increase, but so will the energy demanded since more energy is required to keep that increased amount of housing comfortable. Here, we instrument the log ratio of rent to energy price with the same instrument from energy estimation, and the Katz-Murphy index interacted with housing supply elasticities from Saiz (2010). Again, our first stage F-stat is very strong. With the estimated elasticities of substitution in hand, we can calculate values for log θEdjt and log θGdjt, which we then regress on climate variables to determine the impact of the climate on the comfort production function. We use cubic B-splines with 7 degrees of freedom to flexibly estimate the marginal effect of an additional day at a given temperature. For electricity, an additional day at either a hotter or colder temperature than 65 degrees leads to additional energy demand, reflecting that households use electricity for both heating and cooling. Additionally, the demand response is stronger further into temperature extremes. Meanwhile, natural gas has coefficients significantly different from zero only when the temperature is colder than 65 degrees. We investigate heterogeneity in the climate’s effect on comfort production in appendix A.4.1, finding differences that are small in magnitude. We then use the above estimates to calculate the marginal effect of climate on comfort prices, finding that the effect is larger for Black households than white—though not outside of bootstrapped confidence intervals. Appendix A.4.1 provides other figures describing patterns in comfort production across cities. 60 (a) Estimate of κE(τ), the marginal effect of temperature on log θE . (b) Estimate of κG(τ), the marginal effect of temperature on log θG. Figure 4. Estimation results for the effect of climate on comfort production Estimationg parameters κE(τ) and κG(τ) using cubic splines with 7 degrees of freedom. The effect is normalized to a day with an average temperature of 65 degrees F. Standard errors are calculated with clustered bootstrapping using 10,000 draws, clustered by city-year. We have limited the range of temperature to between the 1st and 99th percentile of average daily temperature, weighted by 1990 population. Labor Supply. Figure 5 plots the estimated moving cost function (2.3) using parameters for estimated by equation (2.31). We plot the moving cost estimates over the range of possible distances from a household’s birthstate.11 We find that Black households are generally less mobile than other households. The birth-state premium parameter is positive, indicating a utility premium, and larger for Black households. Furthermore, the distance parameter is negative, indicating a utility penalty for moving further away from one’s birth state, and more negative for Black households. We also find that non-college-educated workers are generally less mobile than college-educated workers. Our estimates are quantitatively very close to those in Colas and Morehouse (2022) and Diamond (2016). We include moving cost estimates for all years in Appendix A.4.4. 11In both Figure 5 and in our simulations, we set the moving cost function equal to its minimum value for distances beyond its estimated minimum. 61 Figure 5. Moving cost estimates for Black and White households We use estimates from Table A.9, but hold moving costs constant at the minimum for distances further than the estimated mininimum. The intercept reflects the parameter estimates of γ̃st. The histogram shows the distribution of potential moves in 2019, excluding the census divisions—so each household in the model shows up 70 times, once for each of the cities we include in the model. We set moving costs equal to the minimum value of the moving costs function at distances further beyond its estimated minimum. 62 Table 2 shows the parameter estimates for climate amenities. We show the robustness of our controls to the inclusion of different geographic and demographic controls. Households dislike both heating and cooling degree days and like days without rain. These estimates are larger in magnitude than Albouy et al. (2016), but smaller than Rudik et al. (2021). Households dislike both fire and flood risk— they are willing to pay 24% of income for a one standard deviation decrease in fire risk and 6% of income for a one standard deviation decrease in flood risk. We emphasize that we do not presently have natural disaster risk indices that vary over time or climate scenarios. Appendix A.4.5 contains estimates for alternative specifications for temperature portion of climate amenities, showing qualitatively similar results to our degree day specification. Table 3 shows the top and bottom ten cities according to their climate amenities in 2019 (β̂Z · Zj,2019), with category-specific rankings also broken out. These rankings demonstrate our model’s ability to richly characterize a city’s climate. For example, the top ten has cities from California with mild summers, winters, and risks, as well as San Antonio—which ranks poorly due to its hot summer weather but makes up for it by having very few cold days. Most striking is the inclusion of Baton Rouge in the top ten and nearby New Orleans in the bottom ten. This is due to the difference in flood risk between the two cities—nearly every property in New Orleans has at least moderate flood risk, whereas less than half do in Baton Rouge. Table 4 shows how climate amenities have changed between 1990 and 2019, using 5-year moving averages. Increased hot temperatures have decreased amenities in most cities, while fewer cold days have improved amenities. In already- hot places, the shape of the temerature-amenity curve means that the decrease in 63 Table 2. Estimation results for non-temperature climate amenities. Dep. Var. Mean Utility Residual Model: (1) (2) (3) (4) Variables HDD (1000’s) -1.73∗∗∗ -1.62∗∗∗ -1.60∗∗∗ -1.99∗∗∗ (0.14) (0.15) (0.15) (0.23) CDD (1000’s) -1.54∗∗∗ -1.44∗∗∗ -1.03∗∗∗ -1.63∗∗∗ (0.28) (0.30) (0.28) (0.43) Median Fire Risk -0.75∗∗∗ -0.60∗∗∗ -0.55∗∗∗ -0.62∗∗∗ (0.13) (0.14) (0.10) (0.11) Median Flood Risk -0.83∗∗∗ -0.94∗∗∗ -0.68∗∗∗ -0.75∗∗∗ (0.15) (0.14) (0.11) (0.11) Annual Precip (m) -2.81∗ -2.30 -2.03 -1.17 (1.69) (1.61) (1.59) (1.75) Annual Precip Squared (sq m) 1.55∗ 1.42∗ 0.75 0.53 (0.81) (0.79) (0.73) (0.77) Pr(No Rain) 7.00∗∗ 7.06∗∗ 7.10∗∗ 7.18∗∗ (2.99) (3.03) (2.86) (2.86) Geographic Controls Yes Yes Demographic Controls Yes Yes Fixed-effects Dem Group Yes Yes Yes Yes Year Yes Yes Yes Yes Fit statistics Observations 1,680 1,680 1,680 1,680 R2 0.94 0.94 0.95 0.95 Within R2 0.27 0.29 0.33 0.34 Heteroskedasticity-robust standard-errors in parentheses Signif. Codes: ***: 0.01, **: 0.05, *: 0.1 Regresions are weighted by population. We exclude the division aggregation CBSA’s in estimation. Geographic controls include city’s average slope and elevation. Demographic controls include percent married, average age, and percent black. 64 amenities from more hot days outweighs the improvement in amenities from fewer cold days. The opposite generally holds for places that are relatively cold. Table 3. Estimated climate amenity rankings by city for 2019. The Hot category is from temperatures over 65F, Cold from temperatures below 65F, Rain from percent no rain days and annual precipitation, and Risks from median fire and flood risk. Full sample degree day specification with geographic and demographic controls. Category Rankings Rank City Hot Cold Rain Risks Best Cities 1 San Jose-Sunnyvale-Santa Clara, CA 5 18 7 32 2 Los Angeles-Long Beach-Santa Ana, CA 40 7 2 54 3 San Diego-Carlsbad-San Marcos, CA 30 10 5 65 4 San Francisco-Oakland-Fremont, CA 2 21 10 59 5 Oxnard-Thousand Oaks-Ventura, CA 18 15 1 67 6 Fresno, CA 53 17 6 53 7 Sacramento–Arden-Arcade–Roseville, CA 38 22 9 57 8 Baton Rouge, LA 59 13 61 4 9 San Antonio, TX 61 12 13 58 10 Phoenix-Mesa-Scottsdale, AZ 69 6 3 64 Worst Cities 61 Grand Rapids-Wyoming, MI 7 67 42 14 62 Buffalo-Cheektowaga-Tonawanda, NY 10 64 58 7 63 Scranton–Wilkes-Barre, PA 12 59 51 43 64 Springfield, MA 16 62 25 44 65 Worcester, MA 8 63 32 48 66 Albany-Schenectady-Troy, NY 14 66 41 37 67 Rochester, NY 9 65 56 42 68 Minneapolis-St. Paul-Bloomington, MN-WI 11 70 20 22 69 New Orleans-Metairie-Kenner, LA 65 8 60 70 70 Syracuse, NY 4 68 67 45 65 Table 4. Change in climate amenities from 1990 to 2