THE LOCAL ECONOMIC IMPACTS OF AMAZON Evidence from Fulfillment Centers and Lessons for Local Economic Development A DISSERTATION SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL OF UNIVERSITY OF MINNESOTA BY Evan R. Cunningham IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF Doctor of Philosophy Jeremy Lise, Advisor July 2025 © 2025 Evan R. Cunningham ALL RIGHTS RESERVED Acknowledgments First and foremost, I would like to thank my wife, Katerina Gribbin, for her unwavering support and friendship during our time in Minnesota. I am also grateful for the support of my parents William and Patricia, my brother Dylan, our dog Maddie, and all my friends and extended family. I would like to thank my advisor, Jeremy Lise, along with Mariacristina De Nardi and Jo Mullins for their excellent feedback and advice. I am grateful to have been part of the Research Department at the Federal Reserve Bank of Minneapolis. I am especially thankful for the support and mentorship of my su- pervisor, Anusha Nath; the positive encouragement of Morris Kleiner; and the economists and pre-docs of the Economic Analysis Group and the Opportunity and Inclusive Growth Institute, who treated me as one of their own. I am thankful for the support of the Minnesota Economics community. I am especially grateful for the class of 2025, participants in the Micro-Macro Workshop, and the many friends in the years above and below. I would also like to thank the organizers and participants of the great conferences I was lucky enough to present at. I am especially grateful for my time with members of the Urban Economics Association and the North American Regional Science Council. A spe- cial thank you is in order for Nikos Terzidis at the University of Groningen, Orsa Kekezi at the Swedish Institute for Social Research at the University of Stockholm, and Charlotta Mellander, Lina Bjerke, and Peter Njekwa Ryberg at the Centre for Entrepreneurship and Spatial Economics at Jonkoping University, for welcoming me to their respective institu- tions. Finally, I am thankful for Mill City Running, especially the Tuesday Night Dinner Crew and the Nation’s Only Race Team Pep Band (probably), for helping me maintain a healthy balance. Last but not least, I would like to thank my running buddy Stella for her emotional support during my final year (and Alessandra Fogli and Fabrizio Perri for i letting us watch her). ii Abstract Does the entry of a large employer to a local labor market increase welfare for residents? What are the implications for local economic development policy? To answer these ques- tions, this dissertation conducts the first, comprehensive analysis of the dramatic expan- sion of Amazon’s fulfillment center (FC) network from 2010 onward. In Chapter 1, I exploit the staggered roll-out of FCs across large U.S. metros in a difference-in-difference framework. I find Amazon’s entry to a new metro increases the total employment rate by 1.0 percentage points and average wages by 0.7 percent. The industrial composition of employment shifts from retail and wholesale trade to warehousing and tradeable ser- vices, primarily driven by younger workers. Employment gains are concentrated among non-college workers. Local rents increase by 1.1 percent, utility costs by 6.0 percent, and home values by 5.6 percent. In Chapter 2, I present a detailed spatial equilibrium model of the U.S. economy. My model is an extension of the classic urban economics models of Rosen (1979) and Roback (1982), combined with elements of modern quantitative spatial equilibrium models, such as Kline and Moretti (2014a) and Redding and Rossi-Hansberg (2017). It incorporates worker heterogeneity in education and home ownership, produc- tivity spillovers across sectors, local government-financed public goods, and unobserved non-wage amenities. The model yields welfare expressions that transparently map the impact of eight distinct channels, all of which have been shown to be empirically impor- tant for the welfare incidence of local labor demand shocks. These channels include (1) direct employment effects, (2) cross-sector spillovers, (3) local cost of living, (4) average home values, (5) local public goods, (6) corporate subsidies, (7) non-wage amenities, and (8) migration. In the third and final chapter, I calibrate my spatial equilibrium model to the U.S. economy before and after Amazon’s expansion. I estimate welfare effects in the aggregate and by education and home ownership status. I find a net increase in welfare: the average worker is willing to pay $329 per year (0.8 percent of income) to live in a large U.S. city after Amazon’s entry. The welfare gains are primarily driven by rising home val- iii ues; the increase in employment, wages, and sectoral shifts are partially offset by rising local costs of living and a declining average value of non-wage amenities in large cities. Corporate subsidies have a negligible impact on welfare as they are a small share of state and local budgets. I conclude with recommendations for state and local leaders when evaluating local economic development policies. iv Contents Acknowledgments i Abstract iii 1 The Local Impact of Amazon: Empirical Estimates 1 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Related Literature and Contributions . . . . . . . . . . . . . . . . . . . . . . . 4 1.3 Amazon’s Distribution Network and Expansion . . . . . . . . . . . . . . . . 9 1.4 Data and Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.4.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.4.2 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 1.5.1 Heterogeneous Effects Across Education-Age Groups . . . . . . . . . 23 1.5.2 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 1.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2 A Spatial Equilibrium Model for Welfare Analysis 28 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 2.3 Definition of Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.4 Welfare Expression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.5 Outline of Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 3 Welfare Effects and Lessons for Local Economic Development Policy 43 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.2 Calibration in the Amazon Context . . . . . . . . . . . . . . . . . . . . . . . . 45 3.3 Welfare Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 v 3.3.1 ”Data-Driven” Approach . . . . . . . . . . . . . . . . . . . . . . . . . 50 3.3.2 Structural Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.4 Policy Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 Bibliography 63 A ACS Rental and Home Value Indices 113 B State and Local Government Finances 114 C Good Jobs First Corporate Subsidy Tracker 116 D Callaway and Sant’Anna (2020) 117 vi List of Tables 1 Summary Statistics, 2009-2019 . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 2 Summary Statistics by Education, 2009-2019 . . . . . . . . . . . . . . . . . . . 73 3 Summary Statistics by Education (ctd.), 2009-2019 . . . . . . . . . . . . . . . 74 4 Balance Tests for Economic Outcomes in Levels by Treatment Group . . . . 75 5 Balance Tests for Demographics in Levels by Treatment Group . . . . . . . . 76 6 Balance Tests for Economic Outcomes in First-Differences by Treatment Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 7 Balance Tests for Demographics in First-Differences by Treatment Group . . 78 8 Effects of Amazon FC Network Expansion, 2009-2019 . . . . . . . . . . . . . 94 9 Effects of Amazon Network Expansion by Education, 2009-2019 . . . . . . . 95 10 Effects of Amazon Network Expansion by Education (ctd.), 2009-2019 . . . . 96 11 Effects of Amazon FC Network Expansion with 2007-09 Recession Con- trols, 2009-2019 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 12 Externally Set Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 13 Federal Income Tax Brackets, 2015 . . . . . . . . . . . . . . . . . . . . . . . . 99 14 Effective Federal Income Tax Rates τijm . . . . . . . . . . . . . . . . . . . . . 99 15 Internally Calibrated Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 100 16 Internally Calibrated Parameters (ctd.) . . . . . . . . . . . . . . . . . . . . . . 101 17 Cross-Sector Spillover Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 101 18 Extreme Value Shocks and Goodness of Fit . . . . . . . . . . . . . . . . . . . 102 19 Data and Parameters for Rest of Country, 2009-2019 . . . . . . . . . . . . . . 103 20 Internally Calibrated Parameters - ”Data-Driven” Approach . . . . . . . . . 104 21 Internally Calibrated Parameters - ”Data-Driven” Approach (ctd.) . . . . . . 105 22 Extreme Value Shocks in ”Data-Driven” Approach . . . . . . . . . . . . . . . 106 23 Welfare Effects of Amazon FC Expansion by Education and Home Ownership107 24 Welfare Effects of Amazon FC Expansion by Education . . . . . . . . . . . . 107 25 Welfare Effects of Amazon FC Expansion by Channel . . . . . . . . . . . . . 108 vii 26 Shocks to Transportation and Warehousing Sector . . . . . . . . . . . . . . . 109 27 Model-Generated Outcome Changes - Structural Approach . . . . . . . . . . 109 28 Welfare Effects of Amazon FC Expansion by Education and Home Owner- ship - Structural Appraoch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 29 Welfare Effects of Amazon FC Expansion by Education - Structural Approach110 30 Extreme Value Shocks in ”Structural” Approach . . . . . . . . . . . . . . . . 111 31 Sample Averages, Corporate Subsidies, 2009-2019 . . . . . . . . . . . . . . . 112 32 Sample Medians, Corporate Subsidies, 2009-2019 . . . . . . . . . . . . . . . . 112 viii List of Figures 1 New FC Opening by Year, 2005-2021 . . . . . . . . . . . . . . . . . . . . . . . 70 2 Warehousing Employment, 2005-2021 . . . . . . . . . . . . . . . . . . . . . . 70 3 New Non-FC Warehouse Openings by Year, 2005-2021 . . . . . . . . . . . . . 71 4 Staggered Roll-out of Amazon’s FC Network, 2010-2020 . . . . . . . . . . . . 71 5 Effect of Amazon Entry on Employment and Industry Shares (Event Study) 79 6 Effect of Amazon Entry on Industry Shares (Event Study)(ctd.) . . . . . . . . 80 7 Effect of Amazon Entry on Unemployment and Population (Event Study) . 81 8 Effect of Amazon Entry on Average Weekly Wages by Industry (Event Study) 82 9 Effect of Amazon Entry of Local Price Indexes (Event Study) . . . . . . . . . 83 10 Effect of Amazon Entry on Rents (Event Study) . . . . . . . . . . . . . . . . . 84 11 Effect of Amazon Entry on State and Local Government Finances . . . . . . 85 12 Effect of Amazon Entry on State/Local Corporate Subsidies (per cap.) (Event Study) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 13 Effect of Amazon Entry on Employment and Industry Shares by Education- Age Group (Event Study) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 14 Effect of Amazon Entry on Industry Shares by Education-Age Group (Event Study)(ctd.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 15 Effect of Amazon Entry on Unemployment and Population by Education- Age Group (Event Study) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 16 Effect of Amazon Entry on Average Weekly Wages by Industry and Education- Age Group (Event Study) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 17 Effect of Amazon Entry on Rents and Home Values by Education-Age Group 91 18 Effect of Amazon Entry on Employment, Selected Industry Shares and Home Values Controlling for Great Recession Severity . . . . . . . . . . . . . . . . . 92 19 Roth and Rambachan (2023) Sensitivity Tests . . . . . . . . . . . . . . . . . . 93 ix Chapter 1 The Local Impact of Amazon: Empirical Estimates 1.1 Introduction In the United States, state and local governments spend $45 billion per year on local job creation incentives. (Bartik 2019) To evaluate if this outlay is a responsible use of taxpayer dollars, the key question is how the entrance of a large employer to new city affects the welfare of residents. Which groups of workers stand to gain (or lose) the most? What mechanisms are driving these welfare impacts? Typically, positive local labor demand shocks lead to increases in aggregate employ- ment and wages via direct effects or positive spillovers across sectors. (Greenstone, Horn- beck, and Moretti 2010) On the other hand, certain new employers, in industries such as retail trade or grocery stores, can have negative impacts on incumbent competitors. (Neu- mark, Zhang, and Ciccarella (2008), Arcidiacono et al. (2020)) These labor market impacts can affect the local cost of living, and home values could rise or fall, depending on the type of employer. (Moretti and Hornbeck 2024) If there is an increase in economic ac- 1 tivity, there may be an influx of tax revenue. These funds can be reinvested back into the community in order to improve the provision of local public goods, such as schools, roads, and parks. (Feler and Senses 2017) On the other hand, local governments must fund the aforementioned potentially costly corporate subsidies. (Slattery and Zidar 2020) The new employer could also affect the non-wage amenity values of living and work- ing in different locations and industry sectors. (Diamond 2016) Finally, the distributional consequences depend on whether these new jobs are filled by previously unemployed locals or new arrivals. In light of these competing channels, the magnitude and incidence of welfare impacts from this new employer are uncertain. To answer these questions, I study one of the largest employers in the United States, and the world: Amazon.com. Amazon.com, Inc. (henceforth ”Amazon”) is the second largest employer in the United States. As of October 2022, the company employed 1.1 million workers domestically (2021 Employer Information Report EEO-1 2022), up from a mere 74,000 in 2009. Nearly 1 in 150 U.S. workers are currently employed by Amazon. This workforce growth has been driven by a dramatic expansion in the company’s online retail warehouse network from 2010 on- ward. In 2010, Amazon operated a mere 17 warehouses, or fulfillment centers (FCs), in a handful of disparate markets. By 2021, the company operated more than 1,200 ware- houses in over 150 metros. (Amazon Global Supply Chain and Fulfillment Center Network 2022) According to Amazon, fulfillment centers are engines of job creation, often hiring thousands of workers. Based on this premise, state and local governments have provided nearly $2.6 billion in subsidies, grants, and tax rebates. (Amazon Tracker: How Much the Public is Subsidizing One of the Largest Retailers 2022) This, combined with concerns over harm to traditional brick and mortar retailers and excessive monopsony power in the la- bor market, has fueled a heated debate over whether Amazon should be welcomed into American communities.1 1See Moss (2022) and O’Neal (2022), for instance. 2 In spite of the substantial public interest, there is a surprising lack of academic research on the impact of these fulfillment centers on local labor markets and economies.2 The fun- damental question at hand is whether workers are better off when Amazon expands in their community. This dissertation strives to provide the first comprehensive answer to this question. In this first chapter, exploiting the variation in timing of Amazon’s en- try across large U.S. metropolitan areas in a staggered adoption difference-in-difference framework (Callaway and Sant’Anna 2021), I estimate the effect of Amazon’s warehouse network on a variety of local economic outcomes, including employment, wages, indus- trial composition, prices, rents, home values, and local government finances. I find Amazon’s entry increases the total metro-level employment rate3 by 1.0 per- centage point. Average weekly wages increase by 0.7 percent, driven by a shift in in- dustrial composition from retail and wholesale trade toward the warehousing and trade- able services (information, financial activities, professional services, etc.) sectors, primar- ily driven by younger workers. Employment gains are concentrated among non-college workers. These positive labor market effects are offset by rising costs of living. Amazon’s entry increases rents by 1.1 percent and the cost of utilities by 6.0 percent. However, Amazon’s entry had a limited effect on aggregate local goods and services prices. Average home values increase by 5.6 percent. State and local government spending on public goods and services increased by 3.1 percent, funded by corresponding increases in local sales tax, property tax, and income tax receipts. The cost of local job creation incentives increased as well, but account for only 1 percent of the state/local budget. In the following sections, I discuss previous research on Amazon along with tradi- tional and online retail sectors, local labor demand shocks, and place-based economic 2The only exceptions I am aware of are Chava et al. (2023) and Pathania and Netessine (2022). I discuss this work and my contribution in more detail below. 3Employment rate refers to payroll employment per prime-working-age adult (16-64 years). 3 policies, especially local economic development. I then describe the expansion of Ama- zon’s warehouse network from 2010 onward, discuss data sources and my empirical strat- egy, and present the results. 1.2 Related Literature and Contributions This dissertation contributes to three strands of literature. First, it extends research ana- lyzing the impact of Amazon and online retail on labor markets and the distribution of economic activity across space. The definitive work on Amazon’s FC network thus far is Houde, Newberry, and Seim (2023). The authors study the impact of nexus tax laws - rules that require online retailers to pay sales taxes to a state if they hold a physical pres- ence - on the extensive and intensive margin of Amazon’s warehouse network. They find replacing nexus tax laws with a nationwide uniform tax would cause Amazon to expand and decentralize their network, but charge higher prices, increasing market share and profits for competitors and overall welfare. I build on this work by demonstrating the impact of Amazon’s FC network on labor markets, housing markets, local public finances (as opposed to state), and non-wage amenities. To my knowledge, there are only two additional papers estimating the local employ- ment and wage effects of Amazon’s FC network. Chava et al. (2023) analyze the effect on the retail sector of the staggered roll-out of a ”major e-commerce firm’s fulfillment centers” from 2010 to 2016. (While the authors do not refer to Amazon by name, the e-commerce firm’s identity is clear from the context.) The authors find an FC opening re- duces retail worker income 2.4 percent, and employment in geographically proximate es- tablishments declines 2.1 percent. While I use a similar empirical approach, my estimates differ in data sources, scope, sample period, level of geographic aggregation, and time horizon. Pathania and Netessine (2022) study the impact of new Amazon warehouses 4 on county-level outcomes. Using Callaway and Sant’Anna (2021) with additional match- ing, classic synthetic controls, and a stacked bias-corrected synthetic control approach, the authors find Amazon entry increased the employment-population ratio between 0.2 and 0.9 percentage points, while median household income increased by 1.8-2.4 percent. Even though Pathania and Netessine (2022) were affiliated with Amazon and had access to additional internal data, allowing them to correct in more detail for site selection, our empirical results are quite similar. Chun et al. (2023) study the expansion of e-commerce within South Korea, and find local retail employment declines. Dolfen et al. (2023) and Li (2024) use structural approaches to assess the gains from e-commerce, finding consumers, especially the wealthy and those in dense cities, benefit from lower prices, additional con- venience, and greater variety. Relihan (2024) finds the expansion of the online grocery market leads consumers to substitute towards offline, local services. Given Amazon’s status as one the nation’s largest retailers and employers, my work connects to that on local impacts of Walmart superstores such as Neumark, Zhang, and Ciccarella (2008). The authors study the effects of Walmart’s expansion from 1977 to 1995, using an instrumental variables strategy informed by their systematic outward expansion from Bentonville, Arkansas. (Holmes 2011) They find Walmart’s entry reduces retail em- ployment by 2.7 percent and total earnings by 1.5 percent after one year. More recently, Arcidiacono et al. (2020) find a new Walmart superstore lowers revenue of proximate gro- cery stores by 16 percent. Wiltshire (2023) confirms the negative employment and earn- ings effects by comparing counties Walmart entered with counties where super centers were blocked by local stakeholders. My contribution to this literature is two-fold. This dissertation is the first to estimate Amazon’s effect on additional local economic variables, including prices, rents, home val- ues, and local government finances, showing impacts beyond the traditional retail labor market. These results, along with my spatial equilibrium model in Chapter 2, allow me to 5 conduct a comprehensive welfare analysis of Amazon’s local impact on resident workers, including labor and housing market channels, public finances, non-wage amenities, and potential migration. Second, this work speaks to the urban economics literature studying local labor de- mand shocks. In this primarily empirical body of work, researchers use a source of plau- sibly exogenous variation in labor demand to study the responsiveness of employment, wages, home values, and other local economic variables along with the associated welfare implications.4 Modern examples of such variation include the 1970’s coal boom (Black, McKinnish, and Sanders 2005), ”million dollar” manufacturing plants (Greenstone, Horn- beck, and Moretti 2010), the fracking revolution (Feyrer, Mansur, and Sacerdote 2017), Bartik shift-share instruments (Moretti (2013), Notowidigdo (2020)), exposure to Chinese import competition (Autor, Dorn, and Hanson (2013), Feler and Senses (2017)) and the relocation of the West German capital city (Becker, Heblich, and Sturm 2021). This paper is most closely related to Qian and Tan (2022) and Moretti and Hornbeck (2024). Qian and Tan (2022) estimate the effects of about 400 high-skill firm entries into local labor markets from 1990 to 2010. They find the incidence of welfare gains is largely determined by home ownership status, with higher rents neutralizing the hire wages among renters. High-skill workers benefited more than low-skill workers, though this reflects in part the high skill firm entry studied. In this work, I study a labor demand shock for non-college and younger workers, and incorporate spillovers across sectors, local public goods, and municipal finances in my welfare analysis. My spatial equilibrium model also includes the entire U.S. economy and some general equilibrium impacts, while Qian and Tan (2022) focus on within-city residential and workplace choices. In spite of these differences, our estimated welfare effects are quite similar in magnitude. 4Earlier work in this stand of research includes Topel (1986), Bartik (1991), Katz and Blanchard (1992), Bound and Holzer (2000), andSaks, Wong, and Hwang (2008). 6 Moretti and Hornbeck (2024) study the incidence of local manufacturing TFP shocks using confidential micro data from the U.S. Census of Manufacturing. Constructing four different shift-share instruments, the authors find over a 20-year period, a one-percent increase in city level manufacturing TFPR leads to a 1.5 percent increase in otal earnings, a 4.0 percent increase in employment (driven in part by in-migration), a 1.5 percent increase in rents, and a 2.3 % increase in home values. A worker’s position in the housing market is critical, with owners benefiting more than renters. They also propose a novel method of estimating spillover effects across metro areas. Moretti and Hornbeck (2024) approximate welfare using average worker ”purchasing power”, reflecting changes in total earnings, rents, and home values. In my data-driven welfare calculations, I extend this approach incorporating local public goods, municipal finances, and unobserved amenities. My contribution to this literature is the study of a novel, widespread, and policy- relevant ”Amazon” shock. By studying the impact on a wider variety of local economic outcomes, my analysis allows me to quantify the individual contribution to welfare of eight different channels: (1) direct employment effects, (2) cross-sector spillovers, (3) local cost of living, (4) average home values, (5) local public goods, (6) corporate subsidies, (7) non-wage amenities, and (8) migration. While the welfare implications of these channels has been studied before, to my knowledge no paper has analyzed all of them in one, comprehensive analysis. Third, I contribute to the place-based policies literature. I build heavily on Busso, Gre- gory, and Kline (2013), who study the impact of the federal urban Empowerment Zone (EZ) program. They find EZ neighborhoods, relative to rejected and future applicants, saw increases in employment and wages, with no concurrent cost of living or population rise. I follow their approach of using a spatial equilibrium model to evaluate the welfare implications of empirical estimates; my ”data-driven” welfare calculations are informed by this work. Busso, Gregory, and Kline (2013) also find the EZ program created limited 7 distortions in the economy, a result consistent with my findings that local job creation incentives have limited welfare costs. Kline and Moretti (2014a) similarly use empirical results and a structural spatial equi- librium approach to conduct a long-term cost-benefit analysis of the Tennessee Valley Authority (TVA), finding positive and persistent spillovers to manufacturing productiv- ity from agricultural subsidies and infrastructure improvements. Rather than using their model merely as a guide to calculate welfare impacts directly from observed data, as in Busso, Gregory, and Kline (2013), Kline and Moretti (2014a) derive welfare implications from additional structure, and can replicate their reduced-form estimates of the TVA with a shock to labor productivity via public infrastructure investments. In my structural ap- proach, I replicate the qualitative impact of Amazon’s expansion via a shock to produc- tivity and the relative efficiency of non-college and younger workers in the transportation and warehousing sector. My work is also related to the sub-genre of place-based policy research studying local economic development incentives. Measurement of job creation incentives has improved in recent years. (Bartik (2018), Bartik (2019)) Slattery and Zidar (2020) evaluate the ex- tent of state and local corporate subsidies in the U.S, finding mega-deals tend to go to manufacturing and high skilled technology and service firms. These firms typically ac- cept offers from richer locations; poorer locations need to provide larger incentives. The authors find evidence of direct employment effects but limited spillovers to the broader economy. There is also empirical work evaluating specific job creation subsidy programs a la Hyman et al. (2023). I contribute to this strand of research by evaluating the welfare impact of the geographic expansion of a corporation - Amazon - that is arguably the mod- ern poster child of state/local business incentives. I find given the relative magnitude of aggregate welfare gains and the fiscal cost of these subsidies, such deals may benefit local communities. 8 1.3 Amazon’s Distribution Network and Expansion The data I use to study Amazon’s fulfillment center (FC) network is from MWPVL Inter- national, Inc., a supply chain logistics consulting company. In their online report, ”Ama- zon Global Supply Chain and Fulfillment Center Network,” MWPVL provides informa- tion on every Amazon FC or warehouse in the United States. This includes the address, opening date, size (in square footage), and type of warehouse. The data covers all fulfill- ment centers, sortation centers, delivery stations, and airport and import/export hubs.5 To provide background, Amazon.com, Inc. is a global technology and e-commerce company that began operations in 1994 as an online book retailer. While the company has expanded into numerous markets including cloud computing, media and streaming, artificial intelligence, and groceries (through their acquisition of Whole Foods), they are perhaps most well known as a general purpose online retailer. Anyone can shop on Ama- zon.com, but Amazon Prime members, for a monthly subscription fee, receive free one- or two-day shipping. Their business model is to allow customers to buy a wide range of goods via the Internet, and deliver them quickly and cheaply. To this end, Amazon’s distribution network, including FCs and other warehouses, is essential. In the 1990’s and 2000’s, Amazon only operated a handful of smaller warehouses. After 2010, the company began slowly expanding their network. Previously, Amazon primarily outsourced delivery of products sold on their website. Independent shippers, such as Fed-Ex or UPS handled packages via their own distribution networks. How- ever, Amazon began building out its own network in the early 2010’s for three main rea- sons. (Houde, Newberry, and Seim 2023) By bringing distribution in-house, the company would reduce reliance on other firms while realizing economies of scale. Additionally, a 5While information is also available on Amazon Fresh warehouses along with Whole Foods distribution centers, we exclude these from our analysis due to their smaller footprint and the fact that their business operations were originally independent. 9 large, geographically dispersed network was integral for 2-day and 24-hour shipping. Within their network, Amazon operates three main types of distribution centers: ful- fillment centers (FCs), sortation centers, and delivery stations. Fulfillment centers - the flagship warehouses of the system - are ”upstream”. When a customer places an order on Amazon.com, it is routed to an FC, where a worker will locate and package the requested item. Packages are then shipped to a sortation center - a ”middle mile” warehouse in the system - where they will be sorted by workers based on zip code. From there, the pack- ages are often taken to local delivery stations - the ”last mile” - where they are picked up by van drivers and taken to the customer’s address. Finally, airport and import/export centers serve as delivery points for products being shipped within the United States or from abroad, respectively, and are typically located near major airports or shipping hubs. (Amazon Global Supply Chain and Fulfillment Center Network (2022), Amazon Tracker: How Much the Public is Subsidizing One of the Largest Retailers (2022)) Figure 1 shows that prior to 2010, new fulfillment centers were opening infrequently. After 2010, however, the pace of expansion slowly increased to about 8 to 10 FCs per year. From 2016 onward, the expansion was rapid, with 25 new FCs beginning operation each year. This pace continued in 2020 and 2021, with nearly two hundred new FCs opening across the two years. Openings of other types of Amazon warehouses increased steadily beginning in 2013, with a similar spike in 2020.6 (See Figure 3.) According to Amazon, FCs hire a substantial number of workers. While MWPVL employment data is sporadic, and therefore not used in my analysis, they report many FCs hire around 2,000 new workers. Amazon representatives, speaking on a virtual fulfillment center tour, reported their fulfillment centers, in Massachusetts and Maryland, employed 4,000 workers year round, with up to 1,000 seasonal jobs. (Amazon Virtual Fulfillment Center 6This concurrent mass openings of fulfillment centers as well as complementary warehouses informs my decision to analyze the impact of Amazon’s entire warehouse network within a metro, rather than one FC. (See Section 5.) 10 Tour 2022) Figure 2 shows that this expansion in Amazon’s FC network corresponds to a similarly dramatic increase in employment in the warehousing sector. From 2010 to 2021, the warehousing sector nearly tripled in size, going from 638,000 to 1.7 million workers. Note this is in line with Amazon’s self-reported increase in employment. (2021 Employer Information Report EEO-1 2022) Figure 4 illustrates the staggered roll-out of Amazon’s FC network throughout the United States from 2010 to 2021. Each mark represents a metropolitan statistical area, or metro, while the color and shape correspond to the year Amazon opened their first local FC or warehouse.7 There are two key takeaways from this series of maps. First, there is substantial variation in the timing of when Amazon opened their first FC across different metros. Second, the expansion of Amazon’s distribution was experienced in communities all across the U.S. These effects were not confined to a particular region or the largest cities. Moreover, even within geographic regions, there is substantial variation in the timing of Amazon entry. This setting provides an exceptional natural experiment to identify the causal effects of Amazon’s FC network on local labor markets. 1.4 Data and Empirical Strategy 1.4.1 Data Next, I describe the rest of my data sources. All data has been collected for the years 2005 to 2021. As mentioned above, data on the location, opening date, size, and type of Amazon FCs and warehouses is from MWPVL International, a supply chain logistics con- sulting firm. Data on payroll employment and average weekly wages by NAICS sector for metropolitan areas is from the Quarterly Census of Employment and Wages (QCEW), 7Technically, the color refers to one year before the official opening of the first FC. This date roughly corresponds to when plans for the warehouse would have been publicly announced, and ensures there are no anticipation effects in my empirical strategy. 11 a program of the U.S. Bureau of Labor Statistics (BLS). Employment and wage data for three different education-age groups – young workers (age 16-24), non-college workers (age 25 and over), and college workers (age 25 and over) – is accessed from the Quar- terly Workforce Indicators (QWI) program run by the U.S. Census Bureau’s Center for Economic Studies. Data on local goods/services prices, rents, and utility costs is from the Regional Price Parities (RPP) program of the Bureau of Economic Analysis (BEA). Other demographic and economic information, including population, home owner- ship rates, and unemployment rates by education-age group for metropolitan areas, are calculated from the U.S. Census Bureau’s Population Estimates, and public use micro data files of the American Community Survey (ACS), accessed via the IPUMS program at the Minnesota Population Center. (Ruggles et al. 2022) ACS data is also used in the construc- tion of rental price indices and average home value indices. Details on the construction of these indices can be found in Appendix A. I supplement my ACS home value indices using the metro-level Zillow Home Value Index (ZHVI) database. These data represent the typical value of a home between the 35th and 65th percentile within a metro, adjusted for size.8 Data on state and local government finances is from the Individual Unit (micro data) Files of the Annual Survey of State and Local Government Finances, a program of the U.S. Census Bureau. I construct a novel, panel dataset of state and local direct public ex- penditures, tax receipts (sales, property and state income), and federal intergovernmental revenue. Using the micro data, I aggregate local government activity9 to the metro level. I allocate state direct revenues and expenditures to counties within their border propor- tionally by the size of intergovernmental transfers, and aggregate to the metro level as well. Finally, I collect data on the value of state and local corporate subsidies from the 8Information on Zillow housing data can be found at https://www.zillow.com/research/data/. 9I include expenditures and revenues from all local governments, including counties, cities, townships, independent school districts, and special districts. 12 https://www.zillow.com/research/data/ Good Jobs First Corporate Subsidy Tracker, from which I calculate that annual value of incentives provided to local employers.10 1.4.2 Empirical Strategy For my empirical strategy, I use a staggered adoption difference-in-differences approach, as implemented in Callaway and Sant’Anna (2021). My goal is to estimate the effect of a binary treatment Dmt on local economic outcome Ymt in a panel of metropolitan areas m and years t. I define Dmt as an indicator for whether Amazon has ”entered” a market, by which I mean is operating at least one FC or warehouse. My effects reflect the cumulative impact of all Amazon FCs and warehouses operating in a metro, rather than just one establishment as in Chava et al. (2023), Neumark, Zhang, and Ciccarella (2008), and Wiltshire (2023). To avoid anticipation effects, I consider a metro treated one year prior to first FC or warehouse opening. My sample is a balanced panel of all metro areas where Amazon did not operate a warehouse prior to 2010, but entered between 2010 and 2021. (N = 138) My period of analysis, however, is from 2005 to 2019. (T = 11) This choice allows me to include metros Amazon entered in 2020 and 2021 in the control group for all treatment groups. I also avoid potential complications for the COVID-19 pandemic, which began affecting U.S. labor markets in early 2020, while capturing as much of Amazon’s recent expansion as possible. Using the Callaway and Sant’Anna (2021) estimation procedure, I can calculate two types of average treatment effects on treated units (ATTs). The first is an overall ATT, virtually equivalent to an estimate one would get from a standard two-by-two difference- in-difference approach. 10Details on the construction of these local finance datasets are in Appendices B and C, respectively. 13 Ymt = θt + αm + βAMZmt + ϵmt (1.1) AMZmt is an indicator equal to one if an Amazon FC or warehouse is operating within a metro. I also estimate ATTs by length of exposure to the entrance of Amazon’s FC network. These estimates are comparable to those from traditional two-way fixed effects event study frameworks, but notably avoid known shortcomings including the ”negative weight problem” as described in Goodman-Bacon (2021) and Sun and Abraham (2021). Ymt = θt + αm + ∑ j∈{−k,...,0,...,n} βjDmt−j + ϵmt (1.2) Dmt−j is an indicator variable equal to one if Amazon entry occurred j years ago. For identification, I am exploiting variation in the timing of Amazon’s entrance to lo- cal labor markets, rather than comparing metros Amazon did and did not enter. Amazon eventually enters all metros in my sample. Thus, the key identifying assumption is par- allel trends between treated and not-yet treated metros in the absence of Amazon’s entry. While this assumption is not directly testable, I will assess its veracity in a number of ways. First, examining pre-Amazon trends in my event-study style plots provides suggestive evidence parallel trends is likely to hold after treatment. Second, following Sant’Anna and Marcus (2021), the parallel trends assumption implies a series of moment restrictions, which I can test directly in the data. These moment restrictions amount to a requirement that changes in economic outcomes must be equal between each treatment group g and all other units treated after g. In other words, treatment and control groups must be balanced in first-differences. (See Section 6.) Third, I assess the robustness of my results to post-treatment violations of parallel trends as in Roth and Rambachan (2023). (See ”Robustness”.) 14 Note that in my baseline ATT estimates, I do not include control variables. In Callaway and Sant’Anna (2021), time-varying controls can only be included if they are independent of treatment. I instead view nearly all local economic variables as potentially affected by Amazon’s entrance. Additionally, I show treated and not-yet-treated metros are balanced with respect to the over-the-year changes in metro-level demographics including the age, race, ethnicity, foreign-born status, and family structure. (See Section 6.) However, I do control for the severity of local labor market responses to the 2007-09 recession for robust- ness. Finally, standard errors are clustered at the metro level, and are calculated using a bootstrap procedure. For more detail on the Callaway and Sant’Anna (2021) staggered adoption DID approach, see Appendix D. 1.5 Results In this section, I present my estimates of the effect of an expansion of Amazon’s FC net- work on employment, wages, industry shares, prices, rents, home values, and local gov- ernment finances at the metro level. I then examine estimated effects for workers in three difference education-age groups: young workers (age 16 to 24), non-college workers (age 25 and over), and college workers (age 25 and over). Table 1 contains summary statistics of my sample of metropolitan areas, grouped by not-yet-treated and treated observations.11 My sample consists of most medium and large metros throughout the United States. The average employment rate, which I calculate as payroll employment per adult age 16 to 64, is 64.3 percent before Amazon’s expansion. Non-tradeable services, including diverse sectors such as education, health care, public administration, and accommodation/food services, accounts for about half of employ- 11While my period of analysis is from 2005 to 2019, when presenting summary statistics, I only use data from 2009 to 2019, to coincide with the timing of Amazon’s expansion. Data from 2005-2008 are used in my analysis, but primarily inform my evaluation of pre-trends. 15 ment. Tradeable goods (manufacturing, mining, agriculture), tradeable services (informa- tion, financial activities, professional and business services), and retail trade account for 10 percent each, while wholesale trade, transportation, and warehousing have relatively small footprints. In terms of wages, tradeable services, wholesale trade, and tradeable goods are the highest paying sectors on average, while retail trade is the lowest. Averages for the regional price parities are close to one hundred in the pre-treatment periods, which represents the average price level for the nation as a whole. The average home value in my sample is roughly $179,000 prior to Amazon’s expansion. It is also worth noting that my custom rent price and home values indices constructed from the ACS, while differing in levels or units, show a similar trend between the pre- and post-Amazon observations. Indeed, my empirical results will be similar using both rent and home value data sources. Finally, within the average metro area, and state and local governments spend roughly $8,331 annually per adult on local public goods such as schools, parks, and roads, among other services. These numbers reflect all spending by local governments within a metro, as well as direct state expenditure. Federal intergovernmental transfers, sales taxes, and property taxes are the primary means by which state and local governments fund them- selves, with state income taxes playing a smaller role.12 In an average metro, state and local governments combined spend roughly $79 per adult per year on corporate subsi- dies. While in an average metro, this would amount to upwards of $60 million, it is a small share of total state and local government spending.13 Tables 2 and 3 presents summary statistics for metros pre- and post-Amazon by educa- tion group. Data is presented for the three education-age groups, reflecting the published 12Revenues from these four sources cover roughly 80 percent of state and local expenditures. The re- mainder is covered by charges for services, municipal debt issuance, and other miscellaneous and interest income. 13My calculation of the value of state and local corporate subsidies per capita in an average metro is in line with those of Slattery and Zidar (2020), who find in 2014, U.S. states spent between $5 and $216 per capita. 16 data of the QWI. College workers have at least a four-year bachelor’s degree. In my sample, non-college workers make up the majority of the adult population. The average employment rate for college workers (77 percent) is higher than the rate for non-college workers (69.1 percent) and youth workers (43.4 percent). Of course, many potential young workers may be enrolled in school full-time, so their lower employment rates are unsurprising. For all groups, most workers are employed in the broad tradeable services sector. However, college graduates find work in the tradeble services sector more often, while non-college workers are more often employed in the transportation/warehousing and tradeable goods sectors. Finally, over 80 percent of young workers are in either the retail or the non-tradeable services sector, which includes restaurants. Unsurprisingly, av- erage weekly wages increase with education. Young workers typically earn much lower wages on average, reflecting both a lower hourly rate, and an increasing propensity to work part-time. To assess the likelihood parallel trends holds between treated and not-yet-treated units, I conduct balance tests for four treatment groups: metros where Amazon entered from 2010 to 2013, 2014 to 2016, 2017 to 2019, and 2020. For each group, I compare pre- Amazon outcomes with all remaining not-yet-treated metros. Tables 4 and 5 show there is significant selection on levels. Amazon is more likely to enter larger metros with higher wages, prices, and home values earlier in the time period studied. Metros in earlier treat- ment groups are also home to more college graduates and are more racially and ethnically diverse. These findings are consistent with public statements from Amazon executives describing their rational for choosing where and when to expand. Akash Chauhan, Ama- zon’s VP of North American Operations, said, ”There are several factors we consider when deciding where to place a new fulfillment center. Most importantly, we look to see where we can improve Prime benefits with faster shipping speeds for customers and where there is a dedicated workforce.” (Amazon Fulfillment Center to Open in Fresno, Cali- 17 fornia 2017) Amazon Prime customers are disproportionately located in larger, wealthier metros. (Good Jobs First 2021) Crucially, this selection on levels does not necessarily threaten my identification strat- egy. In a staggered adoption DID framework, it is balance in first-differences that is es- sential.14 Table 6 presents balance tests in first-differences for each treatment group. For most economic variables, there is no significant difference in changes across treatment groups and their respective control groups prior to Amazon’s entry. There is some lim- ited evidence that Amazon may have chosen to enter metros where employment and home values were declining more first. I hypothesize these metros were hit hardest by the 2007-09 Recession. To ensure this possible selection is not driving my results, I control for the severity of local labor market responses to the 2007-09 Recession. (See ”Robust- ness”.) Finally, Table 7 shows treated and not-yet treated metros exhibit balance in the first-differences of demographics, suggesting time-varying controls are not necessary. In the rest of this section, I will present my empirical results on the effect of Amazon’s FC network on the local economic variables described above. Event-study style estimates of treatment effects by time since Amazon’s entrance are shown graphically, while overall average treatment effect estimates will be discussed in the text, and presented for refer- ence in Table 8. Figure 5 presents event study style plots for total payroll employment (per adult age 16 to 64) and shares of employment for selected industries. An expansion of Amazon’s FC network has a positive, statistically significant effect on the overall em- ployment rate that is increasing over time. On average, the employment rate in a metro with Amazon is 1.0 percentage-points higher than a market where Amazon has yet to expand, from a base value of 64.3 percent. Figure 5b shows that the share of employment in the warehousing sector expands by 40 percent following Amazon’s entry, from 0.5 per- 14These are the moment restrictions implied by the parallel trends assumption discussed by Sant’Anna and Marcus (2021). 18 cent to 0.7 percent. On the other hand, Figure 5c and 5d show retail and wholesale trade employment shares each declined by 0.1 percentage points, from bases of 12.1 and 3.5 percent, respectively. This translates to a 0.9 percent decline in the retail share, and a 3.1 percent decline in wholesale. Importantly, this does not mean jobs in retail trade were de- stroyed, but merely that employment growth did not keep pace with other sectors. This is consistent with Amazon’s expansion applying downward pressure on retail employment as Neumark, Zhang, and Ciccarella (2008) finds for Walmart and in Chava et al. (2023). In all four panels, the effect of Amazon’s expansion appears to grow over time. This is because when Amazon entered a metro, the company did not typically open only one FC - they opened a number of warehouses over a multi-year period, until there was sufficient capacity to serve the market. The overall average treatment effects may appear smaller in magnitude, but this reflects the larger number of observations in my sample for shorter exposure lengths. Taking into account the stark increase in the size of confidence inter- vals as length of exposure increases, I view the overall ATTs as conservative estimates of Amazon’s effect. Figures 6a-6d show Amazon’s expansion leads to a 0.2 percentage point (1.7 per- cent) increase in the tradeable services employment share, but has no significant effect on the transportation, tradeable goods, and non-tradeable services sectors.15 As men- tioned above, tradeable services includes the information, financial services, and pro- fessional and business services sectors. While some jobs in the information sector may reflect concurrent expansion in Amazon’s network of data centers for their cloud com- puting business, my results imply there are spillovers to other sectors from the expansion of Amazon’s FC network.16 Taken in its totality, these results show the expansion of Ama- 15Given that I find limited aggregate employment and wage effects for the transportation sector, I will refer to the combined NAICS transportation and warehousing sector for the remaineder of the paper. 16Anecdotal evidence is consistent with this story, as a Bessmmer, Alabama local development official noted in the Economist that an Amazon FC can serve as a ”seal of approval” encouraging other business to follow suit and expand in the market. 19 zon’s FC network has led to an increase in the aggregate employment rate, and a shift in the composition of employment from lower-paying to higher-paying sectors on average. A key question is whether the positive employment effects are driven by previously unemployed workers finding jobs or a population inflow in response to Amazon’s expan- sion. I find the former to be most important. (See Figure 7.) Amazon’s entry causes the local unemployment rate in a metro to decline 0.3 percentage-point, suggesting improved labor market conditions are bringing some workers off the sidelines. On the other hand, there is limited evidence Amazon’s expansion leads to in-migration, especially from rural areas or smaller metros. The event study plot may display a modestly upward trend, but it is not significant at any conventional level. Additionally, there are not differential pre- trends in population between treated and not-yet treated metros. This further suggests any systematic selection where Amazon selects faster-growing metros to enter first is of limited concern with regard to identification. Figure 8 displays event-study style plots for the effect of Amazon’s expansion on aver- age weekly wages overall and for selected industries. Figure 8a shows an Amazon entry exerts upward pressure on overall average weekly wages in a market. Wages increase by 0.7 percent, though this estimate falls just below the 95 percent level of confidence. The results for sector-specific wages suggest a substantial share of this increase is driven by the changes in industrial composition discussed above. Figure 8b implies Amazon’s expansion has caused a 4.9 percent decline in average wages for warehousing workers. This is consistent with media reports that have argued Amazon fulfillment centers have lowered wages for warehouse workers (See “What Amazon Does to Wages” (2018) and “What Happens When Amazon Comes to Town” (2022).), but neither these analyses nor my own (so far) account for potential composition effects. Much of the expansion in ware- house employment has been among entry-level positions, which receive lower pay. Thus, my results may not reflect wage effects conditional on tenure. It may also reflect the rel- 20 ative youth (as later results will show) and inexperience of Amazon’s workforce, or the substitution of cash wages with fringe benefits. Figures 8c and 8d show that Amazon has a limited effect on average weekly wages in the retail and wholesale trade sectors. Wage effects in sectors not displayed are also small and statistically insignificant. Figure 9 shows treatment effects for various measures of local prices. Figure 9a shows the effect of an Amazon FC network expansion on an aggregate regional price index for metro areas. In general, the treatment effects are not statistically different from zero. Ag- gregate price levels may be trending up after more years of exposure to Amazon, but these estimates are highly imprecise. Amazon’s expansion appears to have little discernible ef- fect on the local of price of goods. While buying off Amazon.com may be more convenient to local customers closer to an FC, these results suggest it is not lowering costs meaning- fully for consumers. This is in line with the consumer-focused research on e-commerce and online retail (Dolfen et al. (2023), Relihan (2024)) that argues consumers benefit from increased variety and time savings. Figure 9c shows the effect on services prices is gener- ally insignificant, with some highly speculative evidence of an upward trend the longer Amazon has operated in a market. This may be driven by the concurrent rise in local wages. These results highlight the importance of analyzing longer term trends, as changes in local economic conditions induced by these shocks may take time to develop or grow and compound over time. Figure 9d shows Amazon’s expansion into a market leads to a significant increase in the price index of utilities; the overall average treatment effect is 5.1 percentage points. This likely reflects increased demand for electricity, gas, and water from Amazon itself as well as households and other businesses. Figure 10 displays results for outcomes related to the housing market - rents and home values - from multiple data sources. Figures 10a and 10b show the effect of Amazon on rent prices using Regional Price Parities data from BEA, and a custom rental price 21 index I constructed using ACS micro data.17 I find local rent prices increase by 1.1 to 1.2 percentage points following Amazon’s entry. This result, combined with treatment effects for utility prices, imply Amazon’s FC network exerts upward pressure on cost of living when it expands into a new market. Figures 10c and 10d display results for home values. The bottom left panel shows results for the Zillow Typical Home Value Index (ZHVI), while the right panel uses a custom home value index constructed from ACS micro data. I find the expansion of Amazon FC network leads to significant home value appreciation. While the overall average treatment effect is higher using Zillow data (5.6 percent) than for my custom index (3.6 percent), the qualitative patterns are quite similar. Differences likely result from Zillow using higher quality administrative data on home transactions than are available in ACS micro data. These results foreshadow that the incidence of Amazon’s expansion may depend crucially on home ownership. Finally, Figure 11 summarises the effect of Amazon’s expansion on state and local government finances. The top left panel shows that Amazon’s entry has a significant positive effect on state and local government spending and revenues; expenditures per person were 3.1 percent higher on average, while revenues18 per person increased by 1.5 percent. The value of corporate subsidies promised by state and local government also increased by 0.8 percent. (See Figure 12.) Figures 11b-11c show Amazon’s effects on the different revenues sources: property, sales, and personal/corporate income taxes. An expansion of Amazon’s FC network also leads to significant increases in all revenue streams. On average, sales tax revenues (+3.3 percent) increased the most, followed by income taxes (+2.5 percent) and property taxes (+1.9 percent). Federal intergovernmental 17The BEA draws heavily on ACS micro data when constructing their price indices. However, there are differences between the indices. For instance, BEA imputes housing costs for homeowners, where as I only use renters. 18Here, state/local government revenues includes only taxes and federal intergovernmental transfers. Current charges for services, receipts from municipal debt issuance, and other miscellaneous income is excluded in order to align my measures of state/local finances with those in my spatial equilibrium model. 22 revenue also increased by 3.0 percent. These results show that to the extent governments are spending funds on local public goods that are valued by residents, Amazon’s effect on local budgets must be incorporated to accurately assess the welfare consequences of their entry into a market. Moreover, differential changes in tax revenues matter for how costs are allocated to different groups of workers. 1.5.1 Heterogeneous Effects Across Education-Age Groups Next, I present estimates of the effect of an expansion of Amazon’s FC network on em- ployment, wages, rents, and home values for workers in three different education-age groups: young workers (age 16 to 24), non-college workers (age 25 and over), and college workers (age 25 and over). I find substantial heterogeneity across these education-age groups. (See Tables 9 and 10.) For instance, Figure 13a shows the increase in the over- all employment rate is concentrated among non-college workers, and to a lesser extent, young workers. The non-college employment rate in markets after Amazon entered was 2.1 percentage points (3.0 percent) higher, while the rate for young workers increased by 0.6 percentage points (1.4 percent). This finding is to be expected, as anecdotally, these are the workers Amazon tends to hire at their new warehouses. However, most new jobs for these groups are not within the transportation and warehousing sector, but distributed throughout the entire labor market. The employment rate among college workers was essentially unchanged by Amazon’s expansion. Figures 13b-13c show that the change in the industrial composition of employment varied by education and age. For instance, the share of employment in the transportation and warehousing sector increased for all groups, but the change was most pronounced for young workers (+0.5 percentage points; 24 percent). Non-college (+0.2 percentage points; 6 percent) and college workers (+0.2 percentage points; 9 percent) saw more modest in- creases. One can see a similar story in the retail trade sector. The share of young workers 23 employed in retail declined 2.8 percent, compared with declines of 1.3 and 1.9 percent for non-college and college workers, respectively. Wholesale trade, on the other hand, saw consistent declines in the employment share across groups. Among the remaining sectors, there were not drastic differences in the effect of Amazon’s expansion across education-age groups, with the possible exception of the non-tradeable services sector, where non-college workers saw a decline in employment share. (See Figure 14.) These results are not surprising; younger workers have traditionally been considered more mo- bile since they have not had long careers to build up sector specific human capital yet. Figure 15 shows among all education-age groups, Amazon’s entry led to a decline in unemployment. Among non-college and college workers, unemployment fell by a statis- tically significant 0.3 to 0.4 percentage point; among young workers, the point estimate was smaller (0.2 percentage point) and less precise. Panel b of Figure 15 provides more detail on potential migration effects, showing college workers are the only group to see a positive point estimate, though it is highly imprecise and not statistically significant. The population of young and non-college workers was essentially unchanged after treat- ment. While not conclusive, these results are consistent with prior research (e.g. Jia et al. (2023)) showing higher mobility rates for more highly educated workers, and implies labor market improvements were felt by incumbent young and non-college workers. Figure 16 shows the the effect of Amazon’s FC network expansion on average weekly wages overall and for selected sectors by education-age group. The top left panel shows all groups saw fairly similar wage increases in percentage terms. Young workers (1.3 per- cent) and college workers (1.4 percent) had slightly higher wage growth after Amazon’s entry than non-college workers (0.8 percent). In the transportation and warehousing sec- tor, however, it was young workers who saw significantly higher weekly wages (+2.9 per- cent), whereas wages changed little for non-college and college workers. The same was true in the tradeable goods and services sectors, where young workers had 1.7 percent 24 and 2.7 percent higher wages, respectively, while college and non-college workers saw little change. These results suggest that aggregate wage gains from Amazon’s expansion were primarily captured by younger workers. Finally, Figure 17 shows how the effects of Amazon on the housing market varied by education and age. The left-hand panel shows that young renters saw a larger increase in average prices (2.6 percentage points) than college or non-college renters. On the other hand, Figure 13b shows that home values increased for all education-age groups. Of course, college-workers were more likely to own their homes during the 2010-2019 period than non-college and especially younger workers. Thus, Amazon’s effect on the housing market probably benefited older and more highly educated workers. Thus, there were meaningful differences across education groups, suggesting welfare impacts may vary as well. 1.5.2 Robustness In this section, I provide robustness checks to support my conclusions. In staggered adop- tion difference-in-difference settings, a key concern is that parallel trends are violated, bi- asing results. While the necessary assumption - outcomes for treated and not-yet-treated units would have evolved in tandem if not for Amazon’s entry - is impossible to test di- rectly, there are multiple methods to infer the likelihood trends are indeed parallel. I can also evaluate how sensitive my key results are to potential violations. First, when calculating event-study style average treatment effects, it is natural to ex- amine pre-trends. In my event study plots, I find very limited evidence of violations of parallel pre-trends.19 That I present results for a wide range of local economic outcomes 19Crucially, in my event study plots, I use a universal base period of t = −1, or ”long-differences” when estimating pre-treatment coefficients. Roth (2024) has shown that when using the Callaway and Sant’Anna (2021) R package did, the default is a variable base period, which does not allow results to be interpreted using the same visual heuristics as standard event studies. In this case, plots can be interpreted in the usual way. 25 further strengthens my case. Thus, not only does parallel pre-trends hold for each local economic variable of interest, it holds for a battery of likely confounders. While useful, evaluating pre-trends is a placebo-type test. Sant’Anna and Marcus (2021) note that the parallel trends assumption implies a series of moment restrictions that can be directly tested. These moment restrictions are equivalent to a requirement that each treatment group and their respective controls be balanced in first-differences. Earlier in this section, I showed balance in first-differences holds for nearly all local eco- nomic variables and treatment groups. However, there was some evidence that, early in the 2010’s, Amazon was choosing metros with larger declines in employment and home values as a result of the 2007-09 recession. If Amazon first entered metros that were the hardest hit by the recession, positive employment effects could reflect a national labor market rebound, not Amazon’s impact. To ensure this is not driving my results, following Callaway and Sant’Anna (2021), I can re-run my regressions controlling for the local severity of the 2007-09 Recession. I calculate severity using the percent change in the metro-level unemployment rate from 2007 to 2009. Table 11 presents the overall treatment effect estimates including this con- trol. For most economic variables, the controls do not change the results. For some key results, including the overall employment rates, utility prices, and average home values, the magnitudes decline slightly, but remain positive and statistically significant. (See Fig- ure 18.) Next, I assess how sensitive my main results are to potential violations of parallel trends following the methods in Roth and Rambachan (2023). Specifically, I re-evaluate the significance of key results assuming the true post-treatment violation of parallel trends is no more than a constant M larger than the maximum violation in the pre-treatment pe- riod. This sensitivity check answers the question of how large the actual parallel trends violation would need to be, relative to the pre-treatment period to threaten my results. 26 Figure 19 conducts this analysis for the overall employment rate and average home val- ues. The left panel shows the post-treatment violation of parallel trends needs to be 1.2 times the largest pre-treatment violation, for my estimated employment effect to not be significant at the 95 percent confidence level. For home values, the threshold is 0.7. Note this is a conservative way of assessing robustness, because the largest pre-treatment de- viation is often many years in the past, and the exercise assumes this deviation remains constant through the entire post-treatment period. Were I to limit my focus to a smaller number of pre-treatment years, my results would appear more robust. 1.6 Conclusion In this chapter, I documented the unprecedented expansion of Amazon’s U.S. FC network from 2010 onward, highlighting how the company transformed the warehousing sector. This Amazon ”shock” had a significant impact on local communities. Exploiting the stag- gered roll-out of the network, I estimated Amazon’s entry led to a significant increase in economic activity, in terms of a higher local employment rate, higher average wages, and an increase in state and local government public good spending. The employment gains were largest among non-college workers, but younger workers drove the shift in indus- trial composition, from traditional retail and wholesale to warehousing and tradeable services sectors. Crucially, my analysis is the first to comprehensively demonstrate that the local economic impact of e-commerce extends far beyond the traditional retail sector to the rest of the labor market, the housing market, and public finances. 27 Chapter 2 A Spatial Equilibrium Model for Welfare Analysis 2.1 Introduction In Chapter 1, I showed the expansion of Amazon’s FC network had a significant, causal effect on a variety of local economic outcomes. In order to assess the welfare implications of these results, a unifying framework or economic model is required. To this end, I will present a spatial equilibrium model of the U.S. economy in the spirit of Rosen (1979) and Roback (1982). My model will also draw heavily from the literature on modern quantitative spatial equilibrium models, such as Kline and Moretti (2014a) and Redding and Rossi-Hansberg (2017). The primary unique ingredients are worker heterogeneity in education and home ownership, productivity spillovers across sectors, local government-financed public goods, and unobserved non-wage amenities. The criti- cal value of this model is its ability to evaluate the channels driving welfare impacts at a higher level of detail then previous work. Most previous research on local labor demand shocks is exclusively empirical, but even the minority of papers that conduct a welfare 28 analysis only focus on one or two key mechanisms.1 This model, on the other hand, al- lows for a transparent mapping of the welfare impacts of eight different channels, all of which the empirical literature have shown to be important. These channels are (1) direct employment effects, (2) cross-sector spillovers, (3) local cost of living, (4) average home values, (5) local public goods, (6) corporate subsidies, (7) non-wage amenities, and (8) migration. Finally, I briefly outline a general calibration strategy. 2.2 Model In Chapter 1, I showed the expansion of Amazon’s warehouse network had a significant impact on various local economic outcomes, stretching from the labor market to the hous- ing market and state and local finances. Next, I will present a spatial equilibrium model in the spirit of Rosen (1979) and Roback (1982) to evaluate the welfare implications of these effects. My model draws from the literature on modern quantitative spatial equilibrium models, such as Kline and Moretti (2014b) and Redding and Rossi-Hansberg (2017). My model is designed to represent the U.S. economy. There are two locations: a repre- sentative large city, which stands in for all large U.S. metros (i.e. metros Amazon entered), and the rest of the country, which stands in for all remaining medium and small metros as well as rural areas. There is a continuum of workers of unit mass one. Workers are ex-ante heterogeneous in education (young, non-college, college), and ex-post in home ownership (owner, renter). They choose where to live and in which sector to work. The utility of a worker with education i living in location m and working in sector j takes the following Cobb-Douglas form:2 uijm = cβ i ijmh (1−βi) ijm gλi ϵijm (2.1) 1See Notowidigdo (2020) or Greenstone, Hornbeck, and Moretti (2010), for instance. 2For ease of notation, the individual index has been suppressed. 29 Workers derive utility from the consumption of goods and services cijm, housing hijm, an amenity value from local government public goods spending gm, and an idiosyncratic preference shock ϵijm. Workers choose cijm and hijm to maximize (3) subject to the fol- lowing budget constraint: (1 + τSm)pmcijm + (1 + τPm)rmhijm = w̃ijm (2.2) The right-hand side of the budget constraint reflects after-tax disposable income less utility costs plus transfers: w̃ijm = (1 − τFEDijm − τIm)wijm − pU mutilm + TF j (2.3) Wages, denoted wijm, are constant for each education type within each sector, and subject to federal and state/local income tax.3 All workers must purchase a fixed quan- tity of utilities (electricity, gas, water, etc.) at price pU m. This price is set at the marginal cost of supply, which is a convex function of residential and business demand (represented by total employment): pU m = f(Nm + ∑ i,j̸=0,N nijm). Goods/services consumption prices and housing rents are given by pm and rm, respectively. These prices are distorted by local sales (τSm) and property (τPm) taxes. As a result of Cobb-Douglas preferences, work- ers optimally devote a constant share βi(1 − βi) of disposable income to consumption (housing). The preference shock is of the Type 1 extreme value form; the scale parameter is identical across sectors, but the location parameter is allowed to vary, capturing loca- tion and sector-specific amenities independent of wages, prices, and local government spending. Next, I describe the worker’s location and sector choice decision. Workers know their education type, as well as realizations of the preference shock for each option. While 3In practice, almost all non-federal income tax is collected at the state level. In this model, however, there is no distinction between state and local. 30 workers do not know their individual home ownership status, they know the relevant ownership rate for each location and the profits they would receive. For each option, workers determine their optimal spending mix. I impose workers allocate spending as if they were renters, to ensure the chosen bundle is affordable in any state of the world.4 Workers further assume if they become a home owner, profits they receive are used to finance additional tax-free consumption.5 Then, workers calculate their expected utility of each option. In each location, options include different employment sectors, or a not in the labor force option. If a worker chooses employment, their job search fails at a known exogenous rate θim, and they become unemployed. Unemployed workers receive UI benefits from the federal government, which I assume replaces 50 percent of average wage income for their education level. Thus, if workers solve the following problem: max j,m  (1 − θi1)E[Ui11] + θi1E[Ui01], ..., (1 − θi1)E[UiJ1] + θi1E[Ui01],E[UiN1], (1 − θi2)E[Ui12] + θi2E[Ui02], ..., (1 − θi2)E[UiJ2] + θi2E[Ui02],E[UiN2]  (2.4) Ui0m represents maximized utility when unemployed, while UiNm represents max- imized utility when not in the labor force. In the equation above, m = 1 refers to the representative large city, while m = 2 refers to the rest of the country. As mentioned above, the expectations are over the home ownership shocks. In each sector j and location m, there is a representative firm with a constant returns to scale production function. Taking prices Pj, productivity Ajm, the relative efficiency of young (εYjm) and non-college (εNC jm ) workers, the local subsidy rate τSUB jm , and federal UI payroll tax τUI as given, the firm chooses the quantity of labor of each type to maximize profits. 4This assumption also simplifies the clearing of the housing market. 5The assumption that additional consumption financed by housing profits is tax-free simplifies the balancing of government budgets. 31 max n1jm,n2jm,n3jm PjAjm(εYjmn1jm+ εNC jm n2jm+n3jm)α j −(1−τSUB jm +τUI) ∑ i=1,2,3 wijmnijm (2.5) Note 1, 2, and 3 are used to denote young, non-college, and college workers respectively. Effective units of young, college, and non-college labor are perfect substitutes. I assume goods or services produced by firms in each sector and location are sold on a perfectly integrated global market. In this sense, my model is partial equilibrium.6 Finally, I assume productivity and relative efficiency shocks to any sector k spillover to other sectors l in the same location with elasticity σA kl and σεi kl. In each location, there is a representative landlord firm that supplies rental housing in order to maximize profits: max hs m ΠL m = ∫hs m 0 (rm−(1+τPLm )zmxκm)dx+∆νm = rmhs m− (1 + τPLm )zm(hs m)κ+1 κ+ 1 +∆νm (2.6) Housing supply is denoted by hs m. The inverse of the local housing supply elasticity is κm, and zm is a cost shifter. The local property tax rate faced by landlords is τPLm . My model allows for the burden of local property tax to fall on either households, landlords, or a combination of the two. ∆νm represents the change in local home values, where E[∆νm] = 0. The average home value within a location is the present value of receiving maximized landlord profits in perpetuity with discount factor δ. νm = ΠL m 1 − δ = 1 1 − δ ( 1 (1 + τPLm )zm ) 1 κm κm 1 + κm r 1+κm κm m (2.7) By taking the derivative of νm with respect to rm, I can formalize the relationship 6The decision to not force goods and services markets to clear was primarily based on a lack of data on goods produced and consumed by different sectors within individual metros. 32 between changes in rents and home values. ∂νm ∂rm = 1 1 − δ ( 1 (1 + τPLm )zm ) 1 κm r 1 κm m (2.8) I assume representative landlord firms are risk neutral, and do not observe ∆νm until after making their decisions. This, combined with E[∆νm] = 0, means potential home value fluctuations do not affect housing supply decisions. In the model, home ownership is represented by workers holding shares of the land- lord firm in their location. Landlord firms, taking rent prices, the housing supply elastic- ity, the cost shifter, and local property tax rates as given, choose the amount of housing to supply to maximize profits. After making this choice, ∆νm is realized. The landlord immediately realizes profits from rental income and home value changes, and distributes these funds to their shareholders (i.e. home owners). As discussed above, households use these funds to finance additional, tax-free consumption. Finally, there are two different levels of government in my model. A federal gov- ernment collects income and UI payroll taxes from workers and firms in both locations. This income is used to fund the UI system, lump-sum transfers, and grants to state/local governments. In order to capture elements of the U.S. welfare system, I assume federal transfers are only given to the non-employed (TF j = 0 if j ̸= 0,N). The federal government budget balance condition is the following: ∑ i,m ni0mwi0m+ ∑ ijm nijmTF j + ∑ m IGRm = ∑ i,j,m τFEDijm nijmwijm+ ∑ i,m,j̸=0,N τUInijmwijm (2.9) Note IGRm stands for federal grants to the combined state/local government in location m. I assume federal tax rates, the UI replacement rate, and the value of lump sum trans- fers to the non-employed are fixed. Since this is a static model, the federal government 33 budget must balance, so I assume all remaining funds are disbursed to state and local governments.7 Within each location, there is a representative, combined state/local government. This entity levies property and sales tax on workers and landlords within its jurisdiction. These self-generated funds, along with grants from the federal government, are used to fund local public goods and to subsidize local production firms. The local government budget balance condition for each location m is the following: gm + ∑ jm τSUB jm ( ∑ i nijmwijm) = ∑ i,j pmτSmnijmcijm + ∑ i,j rmτPmnijmhijC+ τPLm zm(hs m)κm + ∑ i,j τImnijmwijm + IGRm (2.10) From the state/local government’s perspective, local tax rates are given, as is federal grant revenue. I assume local subsidy rates for each sector are exogenous and fixed. The government chooses gm such that their balanced budget condition and fixed obligations to production firms are met. 2.3 Definition of Equilibrium Next, I formally define an equilibrium in this model. An equilibrium allocation consists of worker location-sector choices, labor demand, goods/services consumption, housing demand, housing supply, federal grants, and local government spending on public goods. {{πijm} , {nijm} , {cijm} , {hijm} , {hs m} , {IGRm} , {gm}} 7Alternatively, the federal government chooses IGRm such that their budget balances and UI and trans- fer obligations are met. 34 Prices include wages, rents, goods/services prices, the price of final goods/services received by production firms, and utility prices. { {wijm} , {rm} , {pm} , {Pj} , { pUTIL m }} Finally, (exogenous) government policy includes federal income and UI payroll tax rates, UI replacement rates, transfer amounts to the non-employed, state/local sales, property, and income taxes, and local subsidy rates. {{ τFEDijm } , τUI, {wi0m} , Tijm, τSm, τPm, τIm, { τSUB jm }} Taking prices and government policy as given, households choose locations and sec- tors {πijm}, consumption {cijm}, and housing {hijm} to maximize utility, production firms choose young, non-college, and college labor {nijm} to maximize profits, and landlords choose housing supply {hs m} to maximize profits. The federal government chooses {IGRm}, and the state/local governments choose {gm} such that their respective budgets balance. Goods/services for consumption and production are assumed to be freely traded on the international market. Finally, labor markets clear for each education type, sector, and location, while housing markets clear for each location. (1 − θim)πijm = nijm ∀i, j,m (2.11) hs m = ∑ i,j nijmhijm ∀m (2.12) In my specific context, I model the expansion of Amazon’s FC network as a shock to productivity and relative efficiency in the transportation and warehousing sector. The impact of this shock will spillover across sectors and affect the housing market and pub- lic finances. I also assume Amazon can have a direct effect on non-wage amenities, as 35 reflected by expected values of the idiosyncratic preference shocks for each location and sector. It is important to note, however, that nearly all aspects of this model are generalizable, and can be readily applied to other contexts. For instance, the number and nature of sec- tors and education levels, can be adjusted to fit the data and or particular labor demand shock at hand. The model can also be used to evaluate changes in federal or state transfer policy, or unemployment insurance. 2.4 Welfare Expression Using my model, I derive expressions for the indirect utility of renters and home owners with education type i working in sector j and living in location m. URENT ijm = ( βiw̃ijm (1 + τsm)pm )β i ( (1 − βi)w̃ijm (1 + τPm)rm )1−βi gλi mE[ϵijm|Uijm > Uikl∀k, l] (2.13) UOWN ijm = ( βiw̃ijm (1 + τsm)pm + γmΠm pm )β i ( (1 − βi)w̃ijm (1 + τPm)rm )1−βi gλi mE[ϵijm|Uijm > Uikl∀k, l] (2.14) Note the only difference between these two expressions is γmΠm pm , reflecting profits from the representative landlord firm. The home ownership rate in location m is γm. Aggregate ex ante welfare for workers in location m is the following: Wm = 1 Nm ∑ ij nijm(γimUOWN ijm + (1 − γim)URENT ijm ) (2.15) Equations (2.13)-(2.15) show worker welfare is determined by real after-tax income, 36 housing market returns, the value of local public goods, non-wage amenities, and the ex- tent of mobility. These expressions also reflect how each of the eight channels discussed earlier impact worker welfare. Direct employment effects and cross-sector spillovers change {nijm} and {w̃ijm}. An increase in the cost of living and rising home values af- fect rm, pU m, and Πm. An increase in local government tax revenues clearly affects {gm}, while an increase in { τSUB jm } would affect both {gm} and {nijm}. A change in non-wage amenities would affect {ϵijm}, while migration would affect {Nm}. My spatial equilibrium model maps the local economic outcomes affected by Amazon’s entry to their welfare im- plications in a clear and transparent way. 2.5 Outline of Calibration To conclude, I provide an overview of how this model could be calibrated to match the U.S. economy. In Chapter 3, I will discuss specific details for the context of Amazon. This model can be calibrated to match the U.S. economy in three main steps. First, we fix a number of parameters external to the model using additional data sources or prior research. Second, we can use observed data on local economic outcomes and the equi- librium conditions of the model to create a system of equations, which can be solved for virtually all remaining unknowns in the model, save for those governing the sector- location specific idiosyncratic preference shocks. Third, we can use a simulated method of moments procedure to set the location parameters of the preference shocks so the model generated choice probabilities8 match empirically observed choice probabilities (i.e. em- ployment rates, industry shares, and population) in the data. The natural candidate parameters to be set outside of the model are those governing the consumer’s utility function, the representative firm production functions in each sec- 8Choices over locations and sectors, to be exact. 37 tor, federal personal and corporate tax rates, as well as the elasticity of housing supply in each sector. In the second step, we set values of the following parameters such that given our observed data, the system of equations below holds:9 {{ τSUB jm ,Ajm, εYjm, εNC jm } j=1,...,J,m=1,...,M , { TF j } j=0,...,J , { τSm, τPm, τLPm , τIm } m=1,...,M , {zm}m=1,..,M } (2.16) For my current example formulation, my observed data would need to obtain the following local economic outcomes: { {nijm,wijm}i=1,2,3,j=1,...,6,m=1,...,M , { pm, rm,pU m,νm,gm, sSm, sPm, sPLm , sIm } m=1,...,M , {SUBjm}j=1,...,J,m=1,2 } (2.17) Thus, the model would be calibrated to match data on wages by education and sector, consumption prices, rents, home values, local public goods spending, and the share of revenues collected from local sales, property, and income taxes.10 The parameters in 2.16 would be chosen to solve the following system of equations: τSUB jm = SUBjm∑ i wijmnijm Nm j = 1, .., J;m = 1, ...,M (2.18) Ajm = ( (1 − τSUB jm + τUI)w3jm Pjαj ) 1 αj−1 1∑ i wijmnijm j = 1, ..J;m = 1, ...,M (2.19) εYjm = w1jm w3jm εNC jm = w2jm w3jm j = 1, ..., J;m = 1, ...,M (2.20) 9Note in this discussion, I have reverted to a general case of M locations and J sectors. I have kept worker types I = 3 for ease of display. 10An alternative approach would be to fix local tax rates in Step 1, which would remove the need for local government spending data. 38 cijm = βiw̃ijm (1 + τSm)pm hijm = (1 − βi)w̃ijm (1 + τPm)rm i = 1, 2, 3; j = 1, ..., J;m = 1, ...,M (2.21) rm = (1 + τLPm )zm(hS m)κm m = 1, ...,M (2.22) ∑ ij pmτSmnijmcijm = sSmGm m = 1, ...,M (2.23) ∑ ij pmτPmnijmhijm = sPmGm m = 1, ...,M (2.24) τPLm zm(hS m)κm = sPLm Gm m = 1, ...,M (2.25) ∑ ij pmτImnijmwijm = sImGm m = 1, ...,M (2.26) The final equations in this system are the federal and local government budget bal- ances, as well as the labor and housing market clearing conditions (2.9)-(2.12). To provide intuition, (2.18) ensures the subsidy rate times total payroll in sector j equals the observed subsidy provided. Equations (2.19) and (2.20) are first order conditions from the produc- tion firms’ maximization problem, (2.21) comes from the worker utility maximization, and (2.22) comes from landlord profit maximization. Equations (2.23-2.26) ensure that the shares of revenue local governments collect from each different source - sales taxes, property taxes on households and landlords, and income taxes - match the observed data. With steps one and two complete, I can calculate the ex-ante expected utility of each location-sector option; expectation are over the idiosyncratic preference shock and home ownership status. From equations (2.13)-(2.15), and the independence of the preference shocks, I can write ex-ante expected utility as: E[Uijm] = Ũijm︷ ︸︸ ︷ (γimUOWN ijm + (1 − γim)URENT ijm )E[ϵijm] (2.27) The probability a given worker of type i chooses sector j and location m is the likeli- 39 hood that option has the highest expected return. πijm = Pr(E[Uijm] ⩾ E[Uikl] ∀k, l) (2.28) Applying a log transformation and plugging in for E[Uijm], I can rewrite the choice probability as: πijm = Pr(ϵ̇ikl − ϵ̇ijm ⩽ log Ũijm − log Ũikl ∀k, l) (2.29) where ϵ̇ijm is a log-transformed preference shock. Because these shocks are Gumbel distributed with scale parameter one, the CDF of ϵ̇ijm is: Fijm(x) = e−e −(x−µijm) (2.30) Since { Ũijm } are known for all i, j, and m, as are the observed choice probabilities {πijm}, I use a simulated method of moments procedure to estimate the location param- eters of each preference shock for each location-sector option and each education group. I choose {µijm} to minimize the distance between model generated choice probabilities {π̂ijm} and observed choice probabilities. Formally, for each education group i, I solve: µi = argmin(π̂i(µi) − πi) ′(π̂i(µi) − πi) (2.31) With estimates of {µijm} in hand, I calculate E[ϵ̇ijm|Uijm > Uikl∀k, l] for each location- sector option via a Monte Carlo simulation. Thus, I use the structure of my spatial equi- librium model to identify the average value of non-wage amenities in a location. The intuition is expected utility E[Uijm] is comprised of a known component, Ũijm, and an unknown (thus far) component E[ϵijm]. If Ũijm > Ũikl, yet πijm = πikl, then E[ϵijm] < E[ϵikl] implying µijm < µikl. 40 With all three steps in the calibration complete, we can calculate welfare overall and for each worker group using (2.15). Welfare can be converted to a dollar amount using the following formula: Wijm = ( βix (1 + τSm)pm )β i ( (1 − βi)x (1 + τPm)rm )1−βi gλi m (2.32) This is a ”wage-equivalent” measure which represents the after-income-tax disposable income required to provide a worker with a given level of welfare. Using this approach to calculate the welfare impact of a local shock or policy change is straightforward if one has quasi-experimental evidence at their disposal. We can apply the estimated treatment effect, when statistically significant, to the observed data, and recalibrate the model in this post-treatment scenario.11 Once the model is calibrated in the pre- and post-treatment scenarios, we calculate the change in wage-equivalent welfare. This can be interpreted as a willingness-to-pay of consumers to exist in the post-treatment scenario relative to the baseline. Finally, this approach can be used to isolate the relative contributions of different channels to welfare changes. For instance, if I wish to only measure the welfare impact of increasing home values, I create a new set of observed data, applying the estimated effect of the shock on home values, but omitting impacts of other variables. WTP = xPOST − xPRE (2.33) 11For instance, suppose the employment rate is 64.0 percent, and my empirical estimates show a labor demand shock has increased 0.5 percent, then in the post-treatment ”observed” data set, I would replace 64.0 percent with 64.5 percent. 41 2.6 Conclusion In this chapter, I have presented a detailed spatial equilibrium model that can be used to evaluate welfare impacts when quasi-experimental evidence of the impacts of local economic or policy shocks is available. Crucially, if causal effects are estimated for a variety of outcomes, my approach can quantify the individual contributions of a number of different channels in a transparent manner. There are disadvantages, however. By separately calibrating a model to pre- and post- treatment scenarios, this approach lacks the structure that is common to most macroeco- nomic models. I argue this is a selling feature, and a complementary approach. Using this model, with the exception of parameters set externally, welfare impacts are driven by observed data and empirical estimates, rather than relying heavily on model structure. In fact, I view this approach as having slightly more structure than the model-motivated cost benefit analysis in Busso, Gregory, and Kline (2013). It can provide a transparent framework for calculating welfare effects from local economic shocks, without the need for fully calibrating a structural spatial macroeconomic model. 42 Chapter 3 Welfare Effects and Lessons for Local Economic Development Policy 3.1 Introduction In this chapter, I combine my empirical results on the impact of the expansion of Ama- zon’s FC network from Chapter 1 and my spatial equilibrium model from Chapter 2 to calculate the welfare impact of this shock on residents of large U.S. cities overall, and by education and home ownership status. First, I discuss the details of calibrating my spatial equilibrium model to the U.S. econ- omy and the context of Amazon. Then, I present my estimates of the welfare effects. I find that Amazon’s expansion in large U.S. cities during the 2010’s increased welfare on net. The average worker is willing to pay $329 per year to live in a large U.S. city after Ama- zon’s entry, relative to a counterfactual U.S. economy where Amazon did not expand. This increase was primarily driven by rising home values, implying the lion’s share of the benefits accrued to home owners. Among renters, welfare declined slightly. The im- provement in local labor market conditions – the increase in employment, wages, and 43 shift in industrial composition – accounted for only 10 percent of the welfare gains. These positive impacts were partially offset by rising local costs of living and a decline in the average value of non-wage amenities experienced by residents of large U.S. cities. Finally, corporate subsidies, by virtue of representing a small share of the state/local budget, had a negligible welfare impact. Finally, I discuss the policy implications of my results. I argue from a purely economic growth standpoint, Amazon’s entry to local economies across the U.S. was a success. However, a justified critique of Amazon is that its expansion has applied upward pres- sure to inequality. Home owners, who tend to be more highly educated and better paid already, captured all of the net benefits. I argue if state and local leaders wish to ensure the positive impacts of local labor demand shocks are more evenly distributed, they should enact policies that improve the accessibility and affordability of housing. Finally, I con- clude skepticism of large corporate subsidy packages may be overblown, given that in the case of Amazon, additional tax revenue far exceeded any costs. However, whether those taxpayer dollars could have been spent more productively depends on one’s opinion of Amazon’s outside option. I leave policymakers with three principles to keep in mind when evaluating potential local economic development projects. One, on economic grounds, leaders should prior- itize incentivizing the largest firms to move to their region, as they are the most likely candidates to generate widespread increases in economic activity. Second, a proper eval- uation of a new employer should extend beyond the new jobs they directly create. Third, since benefits of a large, new employer are spread throughout an entire metro area, local leaders should collaborate with their neighbors as much as possible, in order to increase scale and avoid costly inter-regional bidding wars. 44 3.2 Calibration in the Amazon Context First, I discuss the details of my calibration for the U.S. economy and the Amazon con- text. Recall from Chapter 2 there are two locations in my model: a representative large city standing in for all medium and large metros, and the rest of the country. In the rep- resentative large city, I assume workers can choose one of six sectors: transportation and warehousing, retail trade, wholesale trade, tradable goods (mining, manufacturing, etc.), tradeable services (information, professional services, finance, etc.), and non-tradeable services (restaurants, education, health care). In the rest of the country, I simplify the cal- ibration by assuming there is one representative production sector.1 I assume the burden of property tax is split evenly between households and landlords. Finally, I assume the logarithm of the preference shocks follow a Gumbel distribution, with a scale parameter equal to one. I calibrate my spatial equilibrium model to match the U.S. economy prior to Amazon’s entry in the three main steps in Chapter 2. First, I fix a number of parameters externally based on additional data sources or prior research. Second, I use observed data on pre- Amazon local economic outcomes and the equilibrium conditions of my model to create a system of equations, which I solve for virtually all remaining unknowns in the model. Third, I use a simulated method of moments (SMM) procedure to set the location param- eters of the preference shocks so the model generated choice probabilities match those in the data. Table 12 presents values for parameters set externally using additional data sources or prior work. In the household utility functions, I set the βi’s, the Cobb-Douglas preference weights for consumption, equal to the share of total household income not spent on rent or housing costs in the American Community Survey (ACS). I arbitrarily set the λi’s, the 1As a reminder, in each location, workers can choose to either work for wages or be not in the labor force, where the only income is federal transfers. 45 preference parameters for local public goods, at one-half of the respective βi’s. The intu- ition is that public goods (e.g. a tennis court) provides positive utility for local residents, but have less value that a comparable private good, as sharing is required. Since this logic bounds λi between zero and βi, λi = βi 2 seems a reasonable approximation. Utility spending utilm is set to match the average monthly cost of utilities reported in the ACS: roughly $180 in large U.S. cities. Population shares and home ownership rates for each education type in large cities and in the rest of the country are also calculated from the ACS.2 In terms of production, I calibrate the labor intensity αj to match the aggregate labor share nationwide in sector j. Using the BEA’s National Income and Product Accounts (NIPA), I set αj equation to the average ratio of total worker compensation in sector j to gross national income generated by sector j:3 αj = TOTALCOMPj GNIj Turning to the housing market, I set κm equal to match the inverse of housing supply elasticities from Saiz (2010). For the representative large city, I use the median of the 150 largest cities. For the rest of the country, I use the median of all other cities for which data is available. The effective federal tax rates { τFEDijm } are set to match 2015 tax brackets from the U.S. Internal Revenue Service. The federal UI payroll tax rate, τUI, is set to 5.4% as calculated from the Department of Labor Employee State UI Contributions data from 2010-2019. In the second step, I set values of the following parameters such that given my ob- 2When calculating population shares living in large cities, I use the averages across untreated metro observations in my sample. For the rest of the country, I use the average from 2010-2019. I find Amazon has no significant effect on local home ownership rates by education, so I use averages for 2010-2019 and hold them constant throughout. 3For most sectors, I find this ratio is quite stable during the 2010 to 2019 period, even in the transporta- tion and warehousing sector. 46 served data on the U.S. economy prior to Amazon’s expansion, the equilibrium condi- tions of my model hold. {{ τSUB jm ,Ajm, εYjm, εNC jm } j=1,...,6,m=1,2 , { TF j } j=0,N , { τSm, τPm, τLPm , τIm } m=1,2 , {zm}m=1,2 } (3.1) My observed data contains the following local economic outcomes: { {nijm,wijm}i=1,2,3,j=1,...,6,m=1,2 , { pm, rm,pU m,νm,gm, sSm, sPm, sPLm , sIm } m=1,2 , {SUBjm}j=1,...,6,m=1,2 } (3.2) Thus, my model is calibrated to match data on U.S. wages by education and sector, goods/services prices, rents, home values, local public goods spending, the share of rev- enue collected from local sales, property, and income taxes, and the value of corporate subsidies to each sector. For the representative large city, I use the averages on untreated observations from my sample. For the rest of the country, I use averages from 2010- 2019. Tables 15 and 16 contain the parameter values that solve this system of equations.4 Since the number of equations is equal to the number of unknowns, the solution perfectly matches the observed data. I calibrate the elasticities of cross-sector productivity and relative efficiency spillovers using the internally calibrated pre-Amazon production parameters (See Table 17.) and my estimated effects of Amazon’s entry on employment and industrial composition. First, I construct { nAMZ ijm ,wAMZ ijm , τSUB jm } by adding my statistically significant estimates from Sec- tion 6 to the pre-treatment baseline.5 (See Table 19.) I then solve for post-Amazon pro- 4Note for all sectors j, I have assumed Pj = 1. 5Note that no observed data for the rest of the country will change. 47 duction parameters { AAMZ ijm , (εYjm)AMZ, (εNC jm )AMZ } using equations (2.19) and (2.20) for all j and m. The elasticities of cross-sector productivity and relative efficiency spillovers can be calculated as: σA 1j = %∆Aj1 %∆A11 σεY 1j = %∆εYj1 %∆εY11 σεNC 1j = %∆εNC j1 %∆εNC 11 j = 2, ..., 5 (3.3) %∆x = xAMZ − x x Note here j = 1 refers to the transportation and warehousing sector. I assume that there are no spillovers in the opposite direction or between other sectors.6 σA jk = σεY jk = σεNC jk = 0 if j ̸= 1 Table 18 presents the results of my SMM approach. I present log-expected utility and the estimated location parameters, along with the model-generated and observed choice probabilities. The location parameters are defined only in relative terms, so I normalize their values such that the average value of non-wage amenities for each education type is equal to zero. Note for all three education types, I match the observed choice probabilities quite well. 3.3 Welfare Effects Having fully calibrated my spatial equilibrium model to match the U.S. economy, I can use it to evaluate the welfare implications of the expansion of Amazon FC network. I estimate welfare impacts using two distinct yet complementary approaches, as described in Chapter 2. 6I also do not need to calculate the elasticity cross-sector spillovers in the rest of the country, as I have assumed there is only one sector. 48 The first, which I term a ”data-driven” approach, is an extension of Busso, Gregory, and Kline (2013).7 I calculate a new post-Amazon set of observed data by applying sta- tistically significant empirical results to the pre-treatment baselines. (See Table 8.) This new economy reflects pre-treatment U.S. economic conditions plus the causal impact of entry in large cities. Holding fixed only parameters set externally (Step 1 above), I re- calibrate the model to match this new economic scenario. I calculate worker welfare in each scenario, and estimate changes. A pro of my ”data-driven’ approach is by construc- tion, I match the pre- and post-Amazon observed data on all local economic outcomes in my data set perfectly. This comes at the cost of structure; a potentially unrealistic large number of parameters are allowed to change in response to Amazon’s entry. A second advantage is that I can quantify Amazon’s effect on non-wage amenities in different loca- tions and sectors, for which I need to use all data at my disposal. This motivates my second approach, which I term the ”structural” approach. This version of my analysis follows Kline and Moretti (2014a) more closely. I assume the ex- pansion of Amazon’s FC network only affects the overall productivity and the relative efficiency of non-college and young workers in the transportation and warehousing sec- tor. I set the value of this shock to replicate the estimated employment and wage gains in the transportation and warehousing sector. Then, I allow the structure of the model to replicate the impacts on the rest of the labor market, the housing market, and public finances. While this approach does not exploit all available data, and the magnitudes of estimated welfare impacts vary slightly, I am able to replicate most qualitative impacts of Amazon’s entry from my empirical results.8 However, a key weakness of my structural 7Busso, Gregory, and Kline (2013) present a spatial equilibrium model to illustrate the mechanisms by which the U.S. federal urban Empowerment Zone (EZ) program could impact the welfare of zome resi- dents and workers. However, instead of fully calibrating their model, the authors use the model only for guidance in conducting a cost-benefit analysis and calculating the total dead-weight loss. They calculate the aggregate benefits and costs of the EZ program directly from their empirical estimates and observed data. 8To replicate the effects on non-wage amenities from my ”data-driven” approach, I must additionally 49 approach is I am unable to replicate the large increases in home values and rents. To accurately capture these changes, I would likely need to add more detail to the housing market. While this would be a fruitful avenue for future study, it highlights the strength of my ”data-driven” approach. I can estimate credible welfare effects without having to fully capture every relevant aspect of a local economy in a structural setting. In both my ”data-driven” and ”structural” approaches, I calculate welfare using equa- tions (2.32) and (2.33) above, and convert each into a wage-equivalent measure. I express changes as a willingness to pay, and as a share of total post-federal tax income. Finally, I divide my WTP estimates by 5 years in order to convert to an annual frequency. 3.3.1 ”Data-Driven” Approach In this sub-section, I present results from my ”data-driven” approach to calculating wel- fare impacts. Table 8 highlight observed dataset for the pre- and post-Amazon scenar- ios. Only the effects of Amazon’s entry that are statistically significant at the 90 percent confidence level are included. Any economic outcome with a statistically insignificant response to Amazon or a precise zero is held constant. The observed dataset for the rest of the country is also held constant. (See Table 19.) Tables 20 and 21 compare the internally estimated parameter values for the two sce- narios. Other than those set externally (See Step 1 in Calibration.), all parameters can take on different values to perfectly match the pre- and post-Amazon data. This includes local government tax rates, { τSm, τPm, τIm } that would normally be held constant. In this context, these tax rates increase ever so slightly in order to fund additionally local govern- ment public goods spending observed in the data. This illustrates the trade off between these two approaches: perfectly matching observed data versus imposing structure. Ta- ble 22 shows the estimated average non-wage amenity values across the two scenarios. In assume Amazon’s entry has a direct effect on the location parameters of the Gumbel preference shocks. 50 most employment sectors, amenity values shift only slightly. Among the NILF, however, average amenity values decline substantially for all education groups. The NILF prefer- ence shock is most representative of location-specific factors, suggesting Amazon’s entry lowers the amenity value of living in large cities for workers of all groups. Tables 23 and 24 present the welfare effects of Amazon’s entry for workers in large U.S. cities from my ”data-driven” approach. Effects are expressed as an annual willingness- to-pay (WTP) to live in a large U.S. city after Amazon’s FC network expanded, relative to a large city in a counterfactual U.S. economy where Amazon had not expanded. These WTPs are also expressed as a share of average annual wages. I find the aggregate welfare effects of Amazon’s expansion for resident workers in large cities are positive. The average worker would be willing to pay $329 per year, or 0.8 percent of annual income. As a group, non-college workers see the largest proportional welfare gain (1.1%) relative to young (+0.6%) and college (+0.5%) workers. Table Y shows the lion’s share of the welfare gains accrue to home owners. Proportionally, non-college home owners benefit the most, followed by college and younger home owners. On the other hand, welfare declines slightly for renters of all education levels. This gap in welfare impacts between home owners and renters echoes the findings of Qian and Tan (2022) and Moretti and Hornbeck (2024). Table 25 displays the individual contributions of each of the eight channels referenced earlier in the paper: direct employment effects, cross-sector spillovers, cost of living, home values, local public goods, corporate subsidies, non-wage amenities, and migra- tion. For these calculations, I conduct the same procedure as for total welfare gains. When constructing the post-Amazon data set, however, I only include the statistically signifi- cant effects relevant to that channel, holding other local economic outcomes constant. For instance, when estimating the contribution of the direct employment effects, I only add significant effects of Amazon entry related to the transportation and warehousing sector. 51 My results provide novel insights about the mechanisms driving welfare impacts of local economic shocks. First, I find direct employment effects have a minimal welfare impact on their own. Employment and wage gains in the transportation and warehousing sector following Amazon’s expansion are worth only $6 per year to the average worker, or 0.01 percent of income. This reflects, in part, the small scope of the warehousing sector. The contribution to welfare of employment and wage increases in other sectors, resulting from spillovers, is larger by a factor of four. These spillover effects are worth $24, or 0.06 percent of income. However, the two labor market channels combined only account for roughly 10 percent of the aggregate welfare gain. Turning to the housing market channels, I find increases in the local cost of living, including rent and utility prices, cost resident workers $39 annually, or 0.09 percent of income. Younger workers bear this cost the most, as they tend to spend a larger share of income on rent and utilities. On the other hand, home values are the primary driver of welfare gains. The rise in home values is worth $419 per year (1.02 percent of income) to the average worker, surpassing the aggregate welfare impact. Among non-college work- ers, this home value appreciation is worth $571, or 1.46 percent of income, reflecting a higher home ownership rate relative to younger workers, and lower wages relative to college graduates. Next, I examine the role of public finances. I find the increase in state/local govern- ment spending on public goods has a positive welfare impact of $11 per worker per year, or 0.03 percent of income. This holds even though I account for a rise in tax rates required to fund this increase. The rise in spending on corporate subsidies has a negligible welfare impact; this outlay costs the average worker $1 per year. Finally, Amazon’s FC network expansion leads to a decline in the average value of non-wage amenities among workers in large U.S. cities. Location-sector amenity values reflect all costs and benefits of an option not reflected in wages, employment rates, cost 52 of living, and local government spending. This decline costs the average worker $108 per year, or -0.26 percent of income. Thus, if the average value of amenities did not change, the welfare gain from Amazon’s entry would have been roughly 33 percent higher. The largest decline in amenity value was for the not in the labor force (NILF) choice. In ad- dition to these sector-specific amenity declines, the average experienced amenity value for a location can decline if residents move from high amenity to low amenity sectors. Since NILF provides the highest amenity value,9, an increase an employment or search from those previously NILF would mechanically lower the average experienced amenity value. Proportionally, this channel impacted non-college workers the most, reflecting their relatively higher concentration in retail trade and NILF, both options where amenity values declined. Migration, because of the lack of significant effects from my empirical section, has no welfare impact. There are two additional takeaways from this analysis. First, in isolation, the number of jobs created by a new employer contains limited information. Especially in smaller industries, the magnitude of cross-sector spillovers has much larger welfare implications. Second, in the case of an employer moving to a new city, the primary drivers of welfare gains are housing market channels. The resulting increase in home values accounts for more than 100 percent of the overall welfare gain. This suggests large employers, when describing their potential benefits to prospective cities and residents, should emphasize how they will affect home values, in addition to labor market benefits.10 This finding is also consistent with a standard spatial equilibrium result that if a population is suffi- ciently immobile, productivity shocks will be fully capitalized into home values. (Albouy 2016) Second, these results show incorporating non-wage amenities into the analysis of a 9Amenity values are assigned based on the sectoral choice they make. If a resident fails to find a job and becomes unemployed, they receive the amenity value of the sector they searched in. 10There is ample evidence home owners participate more in local politics. (Hall and Yoder 2022) 53 local economic shock is essential. Depending on the positive or negative amenity impact of an employer, the magnitude of welfare effects could shift substantially. While many lo- cal amenities, such as government spending on public goods, can be measured, a spatial equilibrium framework such as mine can capture the impact on all unobserved ameni- ties. Finally, I find the welfare costs of corporate subsidies and other local job creation incentives are minimal. Even though values cited in the media may seem large, they are often quite small relative to the aggregate state/local budgets. If the new employer leads to any increase in economic activity, it will lead to influx of tax revenue; the subsidies pay for themselves. This is not to say money spent to incentivize large corporations could not have been put to better use, but concerns about these outlays may be overblown. 3.3.2 Structural Approach In my second, complementary method of evaluating the welfare impacts of Amazon’s FC network expansion, I take a more structural approach, as in Kline and Moretti (2014a). I assume that Amazon’s entry only directly affects productivity (A11) and the relative efficiency of youth (εY11) and non-college (εNC 11 ) workers in the transportation and ware- housing sector in the representative large city. I set the values of these shocks such that given post-Amazon wages, the employment gains in transportation and warehousing match my empirical estimates. (See Table 26.) As a result of my model assumptions (See (30).), these shocks will spillover to all remaining sectors, and replicate perfectly the ag- gregate employment and industrial composition effects I estimated above. Then, utility prices will adjust to reflect increasing demand, and rent prices will adjust to clear the housing market. Home values will increase following (10). The federal and state/local governments will then adjust their spending in order to ensure all tax revenues are used and their budgets balance. I can then use my SMM procedure to estimate the location 54 parameters of my preference shocks, and calculate changes in welfare.11 Tables 28 and 29 present the results of the welfare calculations from my structural approach. I am able to qualitatively replicate the key results from my ”data-driven” ap- proach. Aggregate welfare increases, and the lion’s share of the gains accrue to home owners. Renters merely break even. However, the estimated magnitudes of my welfare effects are smaller. Using the structural approach, the average worker is only willing to pay $72 (0.2 percent of income) per year to live in a large u.S. city after Amazon’s entry, compared to $329 in my data driven method. To help understand why my structural approach results in smaller welfare gains, Table 27 present the changes in economic outcomes generated by my model. I can generate the direction of impact seen in the data. I match the state/local government finance impacts quite well. Considering no underlying government policy parameters change, this im- plies that the expansion in state/local government revenues and expenditures is driven by the increase in economic activity. However, my structural approach substantially underestimates the magnitudes of housing market impacts. In my model, rents and home values only change when they adjust to clear the housing market after an increase in demand. These results suggest that Amazon’s entry is affecting the housing market merely stimulating additional demand. Note in this analysis, due to the static nature of my model, I do not allow expectations of future rent increases to affect housing supply or home values. Were I to incorporate expectations of future rent increases, I could generate a larger increase in home values relative to rents. My estimates of Amazon’s impact on non-wage amenities are virtually identical across my data-driven and structural approaches. The underlying intuition, at play in both 11To proceed, I set the housing market discount factor δ = 0.9 arbitrarily. However, because this parame- ter is held constant in both the pre- and post-Amazon equilibrium, it does not impact the elasticity of home values to rents. 55 methods, is that if large cities see an increase in economic activity, we should expect to see in-migration, all else equal. Since I do not observe an increase in migration in the data, this must imply a change in the amenity value of living in each location. 3.4 Policy Implications My empirical results and welfare calculations concerning the local economic impact of Amazon’s FC network provide crucial lessons for policy makers. From a purely eco- nomic growth standpoint, Amazon’s entry to new local markets throughout the U.S. was a success. FCs and other warehouses hired a significant number of local workers, many of whom were young or lacking a college degree. However, direct employees of Amazon only account for a small share of the total jobs created. Spillover effects from this economic activity created jobs in other sectors including construction, administrative services, accommodation and food services, and health care. Notably, even though job growth in retail sector slowed, causing its share to decline, the level of employment held steady. Fewer jobs may have been created, but pre-existing positions were not eliminated at scale. This widespread improvement in labor market conditions led to higher average wages for all groups of workers. Higher levels of disposable income meant higher levels of state and local tax revenue, that could be reinvested back into the community. However, one justified critique of Amazon’s expansion is that it has increased inequal- ity. The increase in economic activity was coupled with a rise in the local cost of living and a significant appreciation in home values. This combination of local impacts meant the lion’s share of the benefits of Amazon’s expansion were captured by home owners, a group which includes college-educated and highly paid workers more often. Renters, while not worse off, merely broke even in real terms. State and local leaders may wonder if there are policies which can ensure the bene- 56 fits from large, local labor demand shocks are more widely spread. My analysis points to the housing market as a crucial policy lever. When home values rise in response to increasing local economic activity, the culprit is almost always a lack of sufficient new construction to meet demand. Thus, state and local policy makers should enact policies that increase accessibility and affordability of housing for residents. If housing supply keeps pace with demand, rent and home price increases will be more subdued. Thus, lower-income houses will have more leftover to spend on local goods and services, and a greater likelihood of transitioning to home ownership in the future. If increasing housing supply meets determined resistance from NIMBY (not-in-my-backyard) groups who wish to preserve single-family neighborhoods, focusing on higher density projects near city or regional centers can be useful. Local policymakers can also increase their local housing supply elasticity by paring back burdensome or unnecessary environmental regulations or encouraging standardized approvals. An advantage of these approaches is they do not require large public outlays. Rent subsidies, on the other hand, are a relatively inefficient way of reducing cost of living burdens. To a first order, they provide relief to lower-income families, the subsidies distort the housing market. Landlords will likely raise rents in response, so every $1 spent on rental subsidies will yield significantly less in benefits to families. In practice, renters are frequently subsidized by property tax refunds or credits on their annual return.12 While less distorting, these measures are still costly, and once-a-year refunds after the fact may not meaningfully impact spending behavior. This analysis also provides clear guidance on when local job creation incentives or corporate subsidy packages can deliver sizable positive returns for local taxpayers. Much of the public discourse surrounding Amazon has portrayed subsidies provided to them as wasteful handouts. As with any negotiation, incomplete information plays a role. If 12See https://www.revenue.state.mn.us/renters-credit, for instance. 57 https://www.revenue.state.mn.us/renters-credit a city knew with one hundred percent certainty Amazon would open an FC even in the absence of support, that money can be better spent elsewhere. However, if the subsidy package would be the difference between Amazon staying or going, my results suggest that investment is, on average, worth it. The subsidy would pay for itself in terms of increased tax revenue from the resulting economic activity in short order. It is not optimal, however, for cities within the same local labor market to bid against each other for an Amazon warehouse or other large employer. The city or town where Amazon locates will suffer a winner’s curse. Though they will bear all of the cost of any subsidy or tax credit, they will not receive all of the additional tax revenue generated, as the increase in economic activity spreads across the entire metro. It is possible leaders of neighboring individual cities complete financially for large companies for mainly polit- ical reasons. A new warehouse is physical evidence that a leader ”created” jobs. Ama- zon has often exploited these tendencies, pressuring local officials to sign non-disclosure agreements (NDAs) concerning ongoing negotiations, with the goal of extracting as many concessions as possible. This makes a strong case for the centralization of local economic development policy at the metro level. In hindsight, many corporate subsidy packages provided to Amazon were probably unnecessary. Taking into account Amazon’s stated goal of providing one or two day de- livery across the country, the online retailer would need warehouses in nearly every city of a decent size. In fact, the company has opened at least one warehouse in nearly every medium and large metro. Thus, subsidy packages were almost certainly not essential for an FC’s viability, especially given Amazon’s strong cash flow from other business lines, including their lucrative cloud computing service (AWS). Having discussed the case of Amazon in depth, what principles should local leaders use when evaluating potential local economic development projects? First, I found the vast majority of job gains as a result of Amazon’s entry were not in warehouses; they 58 were spread throughout nearly all sectors of the local economy. Thus, leaders should pri- oritize the largest firms, whose arrival could potentially lead to a widespread increase in economic activity. Programs targeted at small businesses, while desirable for the amenity value they provide, are not as efficient if the goal is purely to drive economic growth. Second, evaluate the potential economic impacts beyond just new jobs. This analysis has shown much of the net economic benefit of a new employer is reflected in higher home values. A positive local shock can lead to improvements in local public goods - e.g. schools, parks, and roads - or other amenities. The benefits of improved labor market conditions could be cancelled out by a rising cost of living, especially if there is not sufficient new residential construction. It is essential to keep in mind all aspects of the local economy that could be affected. Third, local leaders should collaborate with their neighbors and wider metro area as much as possible. Proactive institutions that coordinate local economic development ac- tivities can increase scale and expand the scope of potential suitor companies, and ensure that municipalities are not bidding against each other. If local economic development policy was overseen at the metro level throughout the U.S., Amazon would have likely received much less in financial support. 3.5 Conclusion In this chapter, and dissertation writ large, I have shown the dramatic expansion of Ama- zon’s FC network across large U.S. cities during the 2010’s increased welfare for resident workers. Amazon’s entry leads to a significant increase in economic activity, in terms of a higher employment rate and higher wages, driven by a shift in employment from retail and wholesale trade to transportation and warehousing and tradeable services. These im- pacts were experienced most strongly by non-college and younger workers. Cross-sector 59 agglomeration spillovers, rather than direct employment effects in warehousing, were re- sponsible for most labor market improvements. However, these labor market channels combined only accounted for a small fraction of the total welfare impact. Amazon’s entry also drove up the local cost of living, including rents and utility costs, and led to a substantial home value appreciation: roughly one percent annually. These were the primary factors behind the aggregate welfare gains, and the gap in benefits between home owners and renters. State and local government spending on local public goods increased, along with spending on job creation incentives. However, the size of these incentives relative to total state/local spending, and an influx of tax revenue meant incumbent workers did not face a higher tax burden. Finally, Amazon’s expansion led to a decline in the value of non-wage amenities in large cities, lowering the total welfare gain by one-third. Critics of Amazon often highlight the steep cost of state and local subsidies for Ama- zon’s FC network around the country. My calculations, however, suggest these concerns may be overblown. Further, my work finds no evidence Amazon’s arrival is depressing aggregate wages for workers at the local labor market level. The ”monopsony power in the labor market” critique that is often leveled at large employers like Walmart may not apply to the same extent. The biggest shortcoming of Amazon’s expansion may be that it is regressive. Since home owners capture the lion’s share of the benefits, and college-educated, well-paid workers own homes at a higher rate, Amazon may be exacerbating inequality, to a de- gree. This may provide a political economy explanation for why governments subsidize Amazon to such an extent: home owners tend to be more active politically, especially at the local level. Policies that increase housing affordability, especially new residential de- velopment, or provide support to first-time home buyers may hold the key to ensuring the benefits of new, large employers are widespread. 60 However, it is important to acknowledge the dimensions in which this analysis falls short. I mostly abstract from negative environmental consequences of Amazon’s expan- sion.13 While the company has recently taken steps to reduce its carbon footprint (us- ing electric vans, for instance), there are reports of significant negative externalities from these new hubs of commercial operations, especially at the neighborhood level. (Waddell et al. 2021) New warehouses may lead to increased air pollution from freight traffic, more dangerous streets for other drivers, pedestrians, and cyclists, and near-constant noise. All of these factors may contribute to worse health outcomes, including a higher prevalence of asthma in children and impaired lung function in adults. (Traffic-Related Air Pollution: A Critical Review of the Literature on Emissions, Exposure, and Health Effects 2010) Where in each market Amazon locates warehouses is likely crucial for the direction and incidence of welfare effects. New FCs are more likely to be located in predominantly Black or Hispanic neighborhoods, suggesting Amazon may exacerbate racial inequities as well. Evaluating the welfare impacts of Amazon at the neighborhood level, and measur- ing the extent of these potential negative externalities is an important avenue for future study. Though Qian and Tan (2022) find qualitatively similar results from high-skilled firm entries, it is also important to note my conclusions may not necessarily generalize to large corporate headquarters. My results, however, do hint at a motivation for residents of Long Island City, New York scuppering plans for Amazon’s HQ2. If the neighborhood was primarily renters, they might have correctly feared being priced out due to a ris- ing cost of living and gentrification. A detailed accounting of the costs and benefits of Amazon’s HQ2 in Alexandria, Virginia would likely further our understanding of this backlash. Finally, in recent years, Amazon and other technology companies have been dramat- 13The decline in non-wage amenities in large cities may partly reflect these negative impacts. 61 ically expanding their network of data centers, especially following the generative AI boom. While data centers anecdotally have a smaller employment footprint, and are more geographically concentrated, they may have an even larger impact on the utilities sector and regional power grids, with profound environmental and cost of living impli- cations. Applying the tools in this paper to evaluate the local impact of data centers is another promising direction for future work. 62 Bibliography Albouy, David. 2016. “What are Cities Worth? Land Rents, Local Productivity, and the Total Value of Amenities.” The Review of Economics and Statistics 98 (3): 477–487. Amazon Fulfillment Center to Open in Fresno, California. 2017. Technical report. Amazon.com, Inc. Amazon Global Supply Chain and Fulfillment Center Network. 2022. Technical report. MW- PVL International. https://www.mwpvl.com/html/amazon com.html. Amazon Tracker: How Much the Public is Subsidizing One of the Largest Retailers. 2022. Tech- nical report. Good Jobs First, August. https://goodjobsfirst.org/amazon-tracker/ ?activity=Distribution%20center. Amazon Virtual Fulfillment Center Tour. 2022. Video, July. https://amazontours.com/. 2021 Employer Information Report EEO-1. 2022. Technical report. Amazon.com, Inc. https: //assets.aboutamazon.com/ff/dc/30bf8e3d41c7b250651f337a29c7/2021-amazon- consolidated-eeo-1-report-2p.pdf. Arcidiacono, Peter, Paul B. Ellickson, Carl F. Mela, and Jon D. Singleton. 2020. “The Com- petitive Effects of Entry: Evidence from Supercenter Expansion.” American Economic Journal: Applied Economics 12 (3): 175–206. 63 https://www.mwpvl.com/html/amazon_com.html https://goodjobsfirst.org/amazon-tracker/?activity=Distribution%20center https://goodjobsfirst.org/amazon-tracker/?activity=Distribution%20center https://amazontours.com/ https://assets.aboutamazon.com/ff/dc/30bf8e3d41c7b250651f337a29c7/2021-amazon-consolidated-eeo-1-report-2p.pdf https://assets.aboutamazon.com/ff/dc/30bf8e3d41c7b250651f337a29c7/2021-amazon-consolidated-eeo-1-report-2p.pdf https://assets.aboutamazon.com/ff/dc/30bf8e3d41c7b250651f337a29c7/2021-amazon-consolidated-eeo-1-report-2p.pdf Autor, David, David Dorn, and Gordon Hanson. 2013. “The China Syndrome: Local La- bor Market Effects of Import Competition in the United States.” American Economic Review 103 (6): 2121–68. Bartik, Timothy J. 1991. Who Benefits from State and Local Economic Development Policies. Technical report. W.E. Upjohn Institute for Employment Research. . 2018. “Improving Economic Development Incentives.” Upjohn Institute Policy and Research Briefs. . 2019. Should Place-Based Policies Be Used to Help Distressed Communities? Technical report. Upjohn Institute for Employment Research. Becker, Sascha O, Stephan Heblich, and Daniel M Sturm. 2021. “The Impact of Public Employment: Evidence from Bonn.” Journal of Urban Economics 122:103291. Black, Dan, Terra McKinnish, and Seth Sanders. 2005. “The Economic Impact of the Coal Boom and Bust.” The Economic Journal 115 (503): 449–476. Bound, John, and Harry J. Holzer. 2000. “Demand Shifts, Population Adjustments, and Labor Market Outcomes during the 1980s.” Journal of Labor Economics 18 (1): 20–54. Busso, Matias, Jesse Gregory, and Patrick Kline. 2013. “Assessing the Incidence and Ef- ficiency of a Prominent Place Based Policy.” American Economic Review 103, no. 2 (April): 897–947. Callaway, Brantly, and Pedro H.C. Sant’Anna. 2021. “Difference-in-Differences with Mul- tiple Time Periods.” Journal of Econometrics 225 (2): 200–230. Chava, Sudheer, Alexander Oettl, Manpreet Singh, and Linghang Zeng. 2023. “Creative Destruction? Impact of E-Commerce on the Retail Sector.” Management Science. 64 Chun, Hyunbae, Hailey Hayeon Joo, Jisoo Kang, and Yoonsoo Lee. 2023. “E-Commerce and Local Labor Markets: Is the ’Retail Apocalypse’ Near?” Journal of Urban Economics 137. Diamond, Rebecca. 2016. “The Determinants and Welfare Implications of US Workers’ Diverging Location Choices by Skill: 1980-2000.” American Economic Review 106 (3): 479–524. Dolfen, Paul, Liran Einav, Peter J. Klenow, Jonathan D. Levin, Larry Levin, and Wayne Best. 2023. “Assessing the Gains from E-Commerce.” American Economic Journal: Macroe- conomics 15 (1): 342–370. Feler, Leo, and Mine Z. Senses. 2017. “Trade Shocks and the Provision of Local Public Goods.” American Economic Journal: Economic Policy 9 (4): 101–143. Feyrer, James, Erin T Mansur, and Bruce Sacerdote. 2017. “Geographic Dispersion of Eco- nomic Shocks: Evidence from the Fracking Revolution.” American Economic Review 107 (4): 1313–34. Good Jobs First. 2021. “Mapping Amazon 2.0,” https://storymaps.arcgis.com/stories/ 144d21045a794cf8b7834b0c49fdd0c0. Goodman-Bacon, Andrew. 2021. “Difference-in-Differences with Variation in Treatment Timing.” Journal of Econometrics 225 (2): 254–277. Greenstone, Michael, Richard Hornbeck, and Enrico Moretti. 2010. “Identifying Agglom- eration Spillovers: Evidence from Winners and Losers of Large Plant Openings.” Jour- nal of Political Economy 118 (3). Hall, Andrew B., and Jesse Yoder. 2022. “Does Homeownership Influence Political Behav- ior? Evidence from Administrative Data.” Journal of Politics 84 (1). 65 https://storymaps.arcgis.com/stories/144d21045a794cf8b7834b0c49fdd0c0 https://storymaps.arcgis.com/stories/144d21045a794cf8b7834b0c49fdd0c0 Holmes, Thomas J. 2011. “The Diffusion of Walmart and Economics of Density.” Econo- metrica 79 (1): 253–302. Houde, Jean Fracois, Peter Newberry, and Katya Seim. 2023. “Nexus Tax Laws and Eco- nomics of Density in E-Commerce: A Study of Amazon s Fulfillment Center Network.” Econometrica. Hyman, Benjamin, Matthew Freedman, Shantanu Khanna, and David Neumark. 2023. “Firm Responses to Hiring and Investment Subsidies: Regression Discontinuity Evi- dence from the California Competes Tax Credit.” NBER Working Paper, no. 30664. Jia, Ning, Raven Molloy, Christopher Smith, and Abigail Wozniak. 2023. “The Economics of Internal Migration: Advances and Policy Questions.” Journal of Economic Literature 61 (1): 144–180. Katz, Lawrence F., and Olivier Blanchard. 1992. “Regional Evolutions.” Brookings Papers on Economic Activity. Kline, Patrick, and Enrico Moretti. 2014a. “Local Economic Development, Agglomeration Economies, and the Big Push: 100 Years of Evidence from the Tennessee Valley Au- thority.” Quarterly Journal of Economics 129 (1): 275–331. . 2014b. “People, Places, and Public Policy: Some Simple Welfare Economics of Local Economic Development Programs.” Annual Review of Economics 6:629–662. Li, Elmer. 2024. “E-Commerce and Regional Inequality: A Trade Framework and Evi- dence from Amazon s Expansion.” https://elmerli.github.io/ecommerce.pdf. Moretti, Enrico. 2013. “Real Wage Inequality.” American Economic Journal: Applied Eco- nomics 5 (1): 65–103. 66 https://elmerli.github.io/ecommerce.pdf Moretti, Enrico, and Richard Hornbeck. 2024. “Estimating Who Benefits from Productiv- ity Growth: Local and Distant Effects of City Productivity Growth on Wages, Rents, and Inequality.” The Review of Economics and Statistics 106 (3): 587–607. Moss, Trefor. 2022. “How a Small Town Learned to Stop Worrying and Love Amazon.” Wall Street Journal (January). https://www.wsj.com/articles/how-a-small- town- learned-to-stop-worrying-and-love-amazon-11642761003. Neumark, David, Junfu Zhang, and Stephen Ciccarella. 2008. “The Effects of Wal-Mart on Local Labor Markets.” Journal of Urban Economics 63 (2): 405–430. Notowidigdo, Matthew J. 2020. “The Incidence of Local Labor Demand Shocks.” Journal of Labor Economics 38 (3): 687–725. O’Neal, Lydia. 2022. “Americans Are Pushing Back on the Warehouse Construction Boom.” Wall Street Journal (April). https://www.wsj.com/articles/americans-are-pushing- back - on - the - warehouse - construction - boom - 11649422800 ? mod = Searchresults pos4&page=4. Pathania, Vikram, and Serguei Netessine. 2022. “The Impact of Amazon Facilities on Local Economies.” Qian, Franklin, and Rose Tan. 2022. “The Effects of High-Skilled Firm Entry on Incumbent Residents.” Stanford Institute for Economic Policy Research (SIEPR) Working Paper, nos. 21-039. Redding, Stephen J., and Esteban Rossi-Hansberg. 2017. “Quantitative Spatial Economics.” Annual Review of Economics 9 (1): 21–58. Relihan, Lindsay. 2024. “Clicks and Bricks: The Complementarity of Online Retail and Urban Services.” 67 https://www.wsj.com/articles/how-a-small-town-learned-to-stop-worrying-and-love-amazon-11642761003 https://www.wsj.com/articles/how-a-small-town-learned-to-stop-worrying-and-love-amazon-11642761003 https://www.wsj.com/articles/americans-are-pushing-back-on-the-warehouse-construction-boom-11649422800?mod=Searchresults_pos4&page=4 https://www.wsj.com/articles/americans-are-pushing-back-on-the-warehouse-construction-boom-11649422800?mod=Searchresults_pos4&page=4 https://www.wsj.com/articles/americans-are-pushing-back-on-the-warehouse-construction-boom-11649422800?mod=Searchresults_pos4&page=4 Roback, Jennifer. 1982. “Wages, Rents, and the Quality of Life.” Journal of Political Economy 90, no. 6 (December): 1257–1278. Rosen, Sherwin. 1979. “Current Issues in Urban Economics.” Chap. Wage-Based Indexes of Urban Quality of Life, edited by Peter Miezkowski and Mahlon R. Straszheim, 74– 104. Johns Hopkins University. Roth, Jonathan. 2022. “Pre-Test with Caution: Event-Study Estimates After Testing for Parallel Trends.” American Economic Review: Insights 4 (3): 305–322. . 2024. “Interpreting Event Studies from Recent Difference-in-Difference Methods.” Roth, Jonathan, and Ashesh Rambachan. 2023. “A More Credible Approach to Parallel Trends.” Review of Economic Studies 90 (5): 2555–2591. Ruggles, Steven, Sarah Flood, Ronald Goeken, Megan Schouweiler, and Matthew Sobek. 2022. IPUMS USA: Version 12.0 [dataset]. https://usa.ipums.org/usa/cite.shtml. Saiz, Albert. 2010. “The Geographic Determinants of Housing Supply.” Quarterly Journal of Economics 125 (3): 1253–1596. Saks, Raven E, Grace Wong, and Min Hwang. 2008. “Reassessing the Role of National and Local Shocks in Metropolitan Area Housing Markets.” Brookings-Wharton Papers on Urban Affairs, 95–126. Sant’Anna, Pedro H.C., Dalia Ghanem, and Kaspar Wuthrich. 2023. “Selection and Paral- lel Trends.” Sant’Anna, Pedro H.C., and Michelle M. Marcus. 2021. “The Role of Parallel Trends in Event Study Settings: An Application to Environmental Economics.” Journal of the Association of Environmental and Resource Economics 8 (2): 235–275. 68 https://usa.ipums.org/usa/cite.shtml Slattery, Cailin, and Owen Zidar. 2020. “Evaluating State and Local Business Incentives.” Journal of Economic Perspectives 34 (2): 90–118. Sun, Liyang, and Sarah Abraham. 2021. “Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects.” Journal of Econometrics. “What Amazon Does to Wages.” 2018. The Economist (January). https://www.economist. com/united-states/2018/01/20/what-amazon-does-to-wages. “What Happens When Amazon Comes to Town.” 2022. The Economist (March). https : //www.economist.com/united-states/what-happens-when-amazon-comes- to- town/21808308. Topel, Robert H. 1986. “Local Labor Markets.” Journal of Political Economy 94 (3, Part 2): S111–S143. Traffic-Related Air Pollution: A Critical Review of the Literature on Emissions, Exposure, and Health Effects. 2010. Technical report 17. Health Effects Institute. Waddell, Kaveh, Maanvi Singh, Brian Vines, Aaron Brezel, and Andy Bergmann. 2021. “When Amazon Expands, These Communities Pay the Price.” Consumer Reports (De- cember). https : / / www . consumerreports . org / corporate - accountability / when - amazon-expands-these-communities-pay-the-price-a2554249208/. Wiltshire, Justin C. 2023. “Walmart Supercenters and Monopsony Power: How a Large Low-Wage Employer Impacts Local Labor Markets.” 69 https://www.economist.com/united-states/2018/01/20/what-amazon-does-to-wages https://www.economist.com/united-states/2018/01/20/what-amazon-does-to-wages https://www.economist.com/united-states/what-happens-when-amazon-comes-to-town/21808308 https://www.economist.com/united-states/what-happens-when-amazon-comes-to-town/21808308 https://www.economist.com/united-states/what-happens-when-amazon-comes-to-town/21808308 https://www.consumerreports.org/corporate-accountability/when-amazon-expands-these-communities-pay-the-price-a2554249208/ https://www.consumerreports.org/corporate-accountability/when-amazon-expands-these-communities-pay-the-price-a2554249208/ Figure 1: New FC Opening by Year, 2005-2021 0 20 40 60 2005 2007 2009 2011 2013 2015 2017 2019 2021 # of F C s Figure 2: Warehousing Employment, 2005-2021 750 1000 1250 1500 1750 2005 2007 2009 2011 2013 2015 2017 2019 2021 (T ho us an ds ) 70 Figure 3: New Non-FC Warehouse Openings by Year, 2005-2021 Figure 4: Staggered Roll-out of Amazon’s FC Network, 2010-2020 2010-2013 2010-2013 2014-2016 2010-2013 2014-2016 2017-2019 2010-2013 2014-2016 2017-2019 2020 71 Table 1: Summary Statistics, 2009-2019 Variable Pre-Amazon Post-Amazon Population, Age 16-64 751,306 1,923,256 Employment Rate (%) 64.3 68.9 Tran. and Ware. Share (%) 3.3 4.3 Retail 12.1 11.1 Wholesale 3.5 3.7 Tradeable Goods 10.9 9.1 Tradeable Services 11.6 14.3 Non-tradeable Services 50.3 50.8 Average Weekly Wages ($) 869 1,064 Tran. and Ware. 900 1,005 Retail 526 610 Wholesale 1,157 1,399 Tradeable Goods 1,133 1,359 Tradeable Services 1,341 1,762 Nontradeable Services 747 866 Unemployment Rate (%) 6.1 4.4 RPP - Goods (100 = U.S. Average) 99 99 RPP - Services 100 100 RPP - Utilities 98 102 RPP - Housing 97 111 Rent Price Index (ACS) 122 135 Typical Home Values (Zillow) ($) 178,778 257,168 Home Value Index (ACS) 120 140 State/Local Govt. Spending (per cap.) ($) 8,331 9,153 State/Local Govt. Revenues 5,916 7,298 Sales Tax 2,132 2,577 Property Tax 1,131 1,675 Income Tax 178 423 Federal I.G. Revenue 2,481 2,636 Corporate Subsidies 79 66 72 Table 2: Summary Statistics by Education, 2009-2019 Variable Pre-Amazon Post-Amazon Population - Youth (Age 16-24) 144,940 347,358 Population - Non-College (Age 25+) 438,998 1,067,792 Population - College (Age 25+) 175,419 516,010 Employment Rate - Youth 43.4 48.4 Employment Rate - Non-College 69.1 76.7 Employment Rate - College 77 73.4 Unemployment Rate - Youth 9.7 7.3 Unemployment Rate - Non-College 5.8 4.1 Unemployment Rate - College 3.6 2.7 Youth Tran. and Ware 2.1 3.5 Retail 22 20.9 Wholesale 2.2 2.2 Tradeable Goods 6.2 4.9 Tradeable Services 7.4 8.8 Non-Tradeable Services 60.2 59.7 Non-College Tran. and Ware 3.9 5.2 Retail 11.7 10.8 Wholesale 4.4 4.7 Tradeable Goods 13 10.5 Tradeable Services 10.5 13.1 Non-Tradeable Services 56.5 55.8 College Tran. and Ware 2.2 3.1 Retail 6.9 6.9 Wholesale 4.3 4.6 Tradeable Goods 9.4 8 Tradeable Services 18.9 23.2 Non-Tradeable Services 58.3 54.2 73 Table 3: Summary Statistics by Education (ctd.), 2009-2019 Variable Pre-Amazon Post-Amazon Average Weekly Wages - Youth 280 331 Tran. and Ware 354 401 Retail 227 257 Wholesale 441 522 Tradeable Goods 479 558 Tradeable Services 404 529 Non-Tradeable Services 257 294 Average Weekly Wages - Non-College 728 868 Tran. and Ware 761 873 Retail 552 628 Wholesale 986 1,162 Tradeable Goods 940 1,109 Tradeable Services 978 1,251 Non-Tradeable Services 642 743 Average Weekly Wages - College 1,269 1,532 Tran. and Ware 1,060 1,245 Retail 804 903 Wholesale 1,738 2,045 Tradeable Goods 1,583 1,900 Tradeable Services 1,729 2,205 Non-Tradeable Services 1,070 1,191 Rent Price Index (ACS) - Youth 118.4 132.3 Rent Price Index (ACS) - Non-College 121.5 131.9 Rent Price Index (ACS) - College 121.8 138.9 Home Price Index (ACS) - Youth 116.6 131.9 Home Price Index (ACS) - Non-College 115.0 129.8 Home Price Index (ACS) - College 120.6 145.5 74 Ta bl e 4: Ba la nc e Te st s fo r Ec on om ic O ut co m es in Le ve ls by Tr ea tm en tG ro up 20 10 -2 01 3 20 14 -2 01 6 20 17 -2 01 9 V ar ia bl e Tr ea t. C on tr ol P V al ue Tr ea t C on tr ol P V al ue Tr ea t C on tr ol P V al ue Em pl oy m en tR at e (% ) 0. 66 1 0. 66 0 0. 92 8 0. 66 1 0. 62 1 0. 00 0 0. 65 5 0. 63 1 0. 05 1 W ho le sa le 0. 04 2 0. 03 6 0. 00 1 0. 03 7 0. 03 4 0. 01 1 0. 03 4 0. 03 3 0. 53 2 R et ai l 0. 11 3 0. 12 1 0. 00 0 0. 11 5 0. 12 4 0. 00 0 0. 12 2 0. 12 7 0. 04 0 Tr an .A nd W ar e. Sh ar e (% ) 0. 04 1 0. 03 2 0. 00 0 0. 03 4 0. 03 1 0. 01 9 0. 03 3 0. 02 8 0. 02 6 Tr ad ea bl e G oo ds 0. 10 2 0. 11 4 0. 08 7 0. 08 4 0. 11 9 0. 00 0 0. 11 6 0. 11 8 0. 80 7 Tr ad ea bl e Se rv ic es 0. 14 8 0. 12 0 0. 00 0 0. 14 5 0. 10 6 0. 00 0 0. 11 4 0. 08 9 0. 00 0 N on -t ra de ab le Se rv ic es 0. 47 7 0. 49 6 0. 00 9 0. 52 8 0. 49 4 0. 00 0 0. 50 6 0. 49 8 0. 33 9 A ve ra ge W ee kl y W ag es ($ ) 6. 79 6. 66 0. 00 0 6. 82 6. 67 0. 00 0 6. 78 6. 71 0. 00 0 R PP 10 2. 1 98 .3 0. 00 0 10 0. 5 97 .5 0. 00 0 98 .0 96 .6 0. 04 4 R PP -G oo ds 10 1. 1 98 .9 0. 00 9 99 .5 98 .1 0. 00 0 98 .0 98 .0 0. 94 5 R PP -S er vi ce s 10 0. 6 99 .7 0. 03 1 10 0. 5 99 .9 0. 01 0 99 .6 10 0. 6 0. 00 1 R PP -H ou si ng 11 5. 2 96 .8 0. 00 0 10 6. 0 92 .2 0. 00 0 97 .1 85 .6 0. 00 1 R PP -U ti lit ie s 97 .9 98 .9 0. 55 3 97 .7 99 .5 0. 10 0 97 .2 10 2. 0 0. 00 0 A ve ra ge H om e V al ue s (Z ill ow ,$ ) 22 0, 73 0 18 2, 73 3 0. 01 7 17 5, 70 0 15 5, 46 4 0. 00 4 19 4, 51 0 16 3, 34 3 0. 01 6 N ot e: Pr e- tr ea tm en tb al an ce te st s sh ow n fo r th re e di ff er en tt re at m en tg ro up s: 20 10 -2 01 3, 20 14 -2 01 6, an d 20 17 -2 01 9. Ba la nc e te st s re fle ct av er ag es in 20 08 -2 00 9, 20 10 -2 01 3 an d 20 14 -2 01 6 re sp ec ti ve ly . 75 Ta bl e 5: Ba la nc e Te st s fo r D em og ra ph ic s in Le ve ls by Tr ea tm en tG ro up 20 10 -2 01 3 20 14 -2 01 6 20 17 -2 01 9 V ar ia bl e Tr ea t. C on tr ol P V al ue Tr ea t C on tr ol P V al ue Tr ea t C on tr ol P V al ue Po pu la ti on ,A ge 16 -6 4 2, 42 1, 08 5 58 4, 80 0 0. 00 0 1, 17 0, 14 6 36 6, 55 5 0. 00 0 44 1, 12 2 29 4, 00 4 0. 00 0 M al e 0. 49 9 0. 49 8 0. 49 6 0. 49 5 0. 49 5 0. 63 4 0. 49 6 0. 49 6 0. 99 3 A ge 16 -1 9 0. 08 4 0. 09 1 0. 00 0 0. 08 3 0. 08 9 0. 00 0 0. 08 7 0. 08 6 0. 54 0 A ge 20 -2 4 0. 10 0 0. 11 4 0. 00 0 0. 10 8 0. 12 0 0. 00 0 0. 12 1 0. 11 8 0. 42 5 A ge 25 -3 4 0. 21 0 0. 20 4 0. 02 0 0. 20 2 0. 20 3 0. 73 7 0. 20 7 0. 21 2 0. 09 5 A ge 35 -4 4 0. 21 9 0. 20 4 0. 00 0 0. 19 8 0. 19 3 0. 00 1 0. 19 0 0. 19 0 0. 88 1 A ge 45 -5 4 0. 22 1 0. 21 8 0. 14 4 0. 21 9 0. 21 0 0. 00 0 0. 20 0 0. 19 9 0. 63 7 A ge 55 -6 4 0. 16 5 0. 17 0 0. 08 6 0. 18 9 0. 18 5 0. 09 9 0. 19 4 0. 19 5 0. 97 7 N on -H is pa ni c W hi te 0. 59 8 0. 68 6 0. 00 0 0. 68 4 0. 65 2 0. 03 3 0. 66 8 0. 59 2 0. 00 2 N on -H is pa ni c Bl ac k 0. 13 4 0. 12 0 0. 26 9 0. 12 0 0. 12 6 0. 58 5 0. 12 5 0. 13 2 0. 69 3 N on -H is pa ni c A si an 0. 06 3 0. 03 7 0. 00 6 0. 05 0 0. 03 6 0. 00 2 0. 04 7 0. 02 9 0. 00 9 H is pa ni c 0. 18 6 0. 13 6 0. 01 9 0. 11 9 0. 16 2 0. 00 1 0. 13 1 0. 22 2 0. 00 1 O th er 0. 01 9 0. 02 2 0. 09 1 0. 02 6 0. 02 4 0. 20 7 0. 02 8 0. 02 5 0. 30 4 Fo re ig n Bo rn 0. 18 7 0. 11 6 0. 00 0 0. 13 4 0. 11 8 0. 04 1 0. 11 8 0. 12 1 0. 78 1 M ar ri ed 0. 47 5 0. 47 6 0. 79 5 0. 45 0 0. 45 9 0. 00 8 0. 44 5 0. 44 4 0. 92 5 Le ss th an H ig h Sc ho ol 0. 17 5 0. 16 1 0. 02 8 0. 13 8 0. 16 2 0. 00 0 0. 14 4 0. 16 9 0. 00 0 H ig h Sc ho ol ,N o C ol le ge 0. 25 7 0. 26 4 0. 29 2 0. 25 4 0. 26 2 0. 04 5 0. 25 5 0. 27 0 0. 01 3 So m e C ol le ge or A ss oc ia te ’s D eg re e 0. 30 5 0. 33 1 0. 00 0 0. 32 6 0. 33 8 0. 00 3 0. 33 2 0. 34 0 0. 11 3 Ba ch el or ’s D eg re e 0. 26 4 0. 24 5 0. 04 2 0. 28 3 0. 23 7 0. 00 0 0. 26 9 0. 22 1 0. 00 0 En ro lle d in Sc ho ol 0. 15 7 0. 17 5 0. 00 0 0. 17 2 0. 18 0 0. 01 1 0. 17 8 0. 16 7 0. 05 3 N ot e: Pr e- tr ea tm en tb al an ce te st s sh ow n fo r th re e di ff er en tt re at m en tg ro up s: 20 10 -2 01 3, 20 14 -2 01 6, an d 20 17 -2 01 9. Ba la nc e te st s re fle ct av er ag es in 20 08 -2 00 9, 20 10 -2 01 3, an d 20 14 -2 01 6 re sp ec ti ve ly . 76 Ta bl e 6: Ba la nc e Te st s fo r Ec on om ic O ut co m es in Fi rs t- D iff er en ce s by Tr ea tm en tG ro up 20 10 -2 01 3 20 14 -2 01 6 20 17 -2 01 9 V ar ia bl e Tr ea t. C on tr ol P V al ue Tr ea t C on tr ol P V al ue Tr ea t C on tr ol P V al ue Em pl oy m en tR at e (p .p C hg .) -0 .0 42 -0 .0 37 0. 06 5 0. 00 2 -0 .0 01 0. 16 2 0. 00 9 0. 00 7 0. 07 5 Tr an .A nd W ar e. Sh ar e (p .p .C hg .) -0 .0 01 -0 .0 01 0. 80 5 0. 00 0 0. 00 0 0. 49 2 0. 00 0 0. 00 0 0. 95 5 R et ai l -0 .0 01 0. 00 0 0. 76 6 0. 00 0 0. 00 0 0. 78 4 0. 00 0 0. 00 0 0. 14 2 W ho le sa le -0 .0 01 -0 .0 01 0. 14 8 0. 00 0 0. 00 0 0. 09 1 0. 00 0 0. 00 0 0. 81 0 Tr ad ea bl e G oo ds -0 .0 05 -0 .0 08 0. 00 1 0. 00 0 0. 00 0 0. 49 5 -0 .0 01 -0 .0 02 0. 26 3 Tr ad ea bl e Se rv ic es 0. 00 0 0. 00 1 0. 93 2 0. 00 0 -0 .0 01 0. 05 3 -0 .0 01 -0 .0 01 0. 59 7 N on -t ra de ab le Se rv ic es 0. 00 4 0. 00 7 0. 00 3 0. 00 1 0. 00 2 0. 35 7 0. 00 3 0. 00 3 0. 90 3 A ve ra ge W ee kl y W ag es (P ct .C hg .) 0. 00 2 0. 00 7 0. 13 7 0. 02 0 0. 01 8 0. 07 6 0. 02 3 0. 02 0 0. 24 8 R PP -0 .1 25 0. 08 2 0. 59 3 -0 .1 24 0. 04 0 0. 24 1 -0 .0 09 0. 05 6 0. 66 2 R PP -G oo ds -0 .6 28 -0 .0 11 0. 38 9 -0 .1 04 -0 .0 19 0. 68 3 -0 .1 71 -0 .0 64 0. 48 5 R PP -S er vi ce s 0. 10 5 0. 10 5 0. 99 9 -0 .0 92 0. 13 4 0. 17 2 0. 10 5 0. 36 0 0. 15 3 R PP -H ou si ng -0 .4 60 0. 08 9 0. 36 6 -0 .2 28 -0 .0 85 0. 69 7 -0 .1 82 -0 .3 96 0. 73 9 R PP -U ti lit ie s 0. 90 7 -0 .1 67 0. 17 7 -0 .1 60 0. 09 1 0. 37 0 -0 .0 15 -0 .1 48 0. 70 7 A ve ra ge H om e V al ue s (Z ill ow ,P ct .C hg .) -0 .1 51 -0 .0 96 0. 04 7 -0 .0 16 -0 .0 09 0. 33 1 0. 05 6 0. 04 9 0. 09 3 N ot e: Pr e- tr ea tm en tb al an ce te st s sh ow n fo r th re e di ff er en tt re at m en tg ro up s: 20 10 -2 01 3, 20 14 -2 01 6, an d 20 17 -2 01 9. Ba la nc e te st s re fle ct av er ag es in 20 08 -2 00 9, 20 10 -2 01 3, an d 20 14 -2 01 6 re sp ec ti ve ly . 77 Ta bl e 7: Ba la nc e Te st s fo r D em og ra ph ic s in Fi rs t- D iff er en ce s by Tr ea tm en tG ro up 20 10 -2 01 3 20 14 -2 01 6 20 17 -2 01 9 V ar ia bl e Tr ea t. C on tr ol P V al ue Tr ea t C on tr ol P V al ue Tr ea t C on tr ol P V al ue Po pu la ti on ,A ge 16 -6 4 0. 01 2 0. 00 9 0. 21 8 0. 00 4 0. 01 0 0. 09 4 0. 00 6 0. 00 4 0. 07 6 M al e 0. 00 0 0. 00 0 0. 48 4 -0 .0 01 -0 .0 01 0. 84 6 0. 00 0 0. 00 0 0. 50 6 A ge 16 -1 9 -0 .0 02 -0 .0 01 0. 24 3 -0 .0 01 -0 .0 01 0. 32 5 0. 00 1 0. 00 0 0. 42 8 A ge 20 -2 4 0. 00 0 0. 00 4 0. 01 6 0. 00 2 0. 00 0 0. 05 6 -0 .0 02 -0 .0 01 0. 35 8 A ge 25 -3 4 0. 00 9 -0 .0 02 0. 00 4 0. 00 0 0. 00 1 0. 21 0 0. 00 1 0. 00 2 0. 61 7 A ge 35 -4 4 -0 .0 06 -0 .0 07 0. 50 4 -0 .0 03 -0 .0 02 0. 03 1 0. 00 0 0. 00 0 0. 89 8 A ge 45 -5 4 -0 .0 01 0. 00 0 0. 36 3 -0 .0 03 -0 .0 03 0. 82 6 -0 .0 02 -0 .0 03 0. 57 7 A ge 55 -6 4 0. 00 0 0. 00 5 0. 00 5 0. 00 5 0. 00 5 0. 60 4 0. 00 3 0. 00 2 0. 97 3 N on -H is pa ni c W hi te -0 .0 06 -0 .0 04 0. 03 0 -0 .0 08 -0 .0 07 0. 58 1 -0 .0 05 -0 .0 06 0. 27 3 N on -H is pa ni c Bl ac k 0. 00 1 0. 00 0 0. 27 6 0. 00 1 0. 00 1 0. 53 2 0. 00 1 0. 00 1 0. 92 0 N on -H is pa ni c A si an 0. 00 1 0. 00 0 0. 10 2 0. 00 2 0. 00 1 0. 05 8 0. 00 1 0. 00 1 0. 68 7 H is pa ni c 0. 00 4 0. 00 2 0. 08 6 0. 00 4 0. 00 4 0. 29 8 0. 00 3 0. 00 3 0. 11 0 O th er 0. 00 0 0. 00 1 0. 10 4 0. 00 1 0. 00 1 0. 27 2 0. 00 1 0. 00 1 0. 86 8 Fo re ig n Bo rn 0. 00 1 0. 00 1 0. 94 9 0. 00 2 0. 00 2 0. 96 8 0. 00 1 0. 00 1 0. 97 3 M ar ri ed -0 .0 07 -0 .0 06 0. 82 1 -0 .0 06 -0 .0 06 0. 87 2 -0 .0 04 -0 .0 03 0. 82 0 Le ss th an H ig h Sc ho ol -0 .0 03 -0 .0 03 0. 84 5 -0 .0 03 -0 .0 02 0. 42 2 -0 .0 01 -0 .0 02 0. 72 6 H ig h Sc ho ol ,N o C ol le ge 0. 00 0 0. 00 0 0. 90 9 -0 .0 01 -0 .0 01 0. 68 2 -0 .0 02 0. 00 0 0. 59 1 So m e C ol le ge or A ss oc ia te ’s D eg re e -0 .0 01 0. 00 3 0. 06 9 0. 00 0 0. 00 1 0. 69 8 -0 .0 03 -0 .0 01 0. 46 2 Ba ch el or ’s D eg re e 0. 00 3 0. 00 0 0. 10 1 0. 00 3 0. 00 2 0. 41 8 0. 00 6 0. 00 3 0. 33 4 En ro lle d in Sc ho ol 0. 00 2 0. 00 4 0. 28 6 0. 00 0 -0 .0 01 0. 48 1 -0 .0 02 0. 00 0 0. 15 2 N ot e: Pr e- tr ea tm en tb al an ce te st s sh ow n fo r th re e di ff er en tt re at m en tg ro up s: 20 10 -2 01 3, 20 14 -2 01 6, an d 20 17 -2 01 9. Ba la nc e te st s re fle ct av er ag es in 20 08 -2 00 9, 20 10 -2 01 3, an d 20 14 -2 01 6 re sp ec ti ve ly . 78 Figure 5: Effect of Amazon Entry on Employment and Industry Shares (Event Study) (a) Total Employment 0.0 2.5 5.0 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Warehousing (Pct. of Emp.) 0.0 0.5 1.0 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (c) Retail Trade (Pct. of Emp.) −0.8 −0.6 −0.4 −0.2 0.0 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (d) Wholesale Trade (Pct. of Emp.) −0.4 −0.2 0.0 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 79 Figure 6: Effect of Amazon Entry on Industry Shares (Event Study)(ctd.) (a) Transportation (Pct. of Emp.) −0.4 −0.2 0.0 0.2 0.4 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Tradeable Goods (Pct. of Emp.) −1.0 −0.5 0.0 0.5 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (c) Tradeable Services (Pct. of Emp.) −0.5 0.0 0.5 1.0 1.5 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (d) Non-Tradeable Services (Pct. of Emp.) −1.0 −0.5 0.0 0.5 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 80 Figure 7: Effect of Amazon Entry on Unemployment and Population (Event Study) (a) Unemployment Rate −2 −1 0 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Population, Age 16 to 64 −2 0 2 4 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . 81 Figure 8: Effect of Amazon Entry on Average Weekly Wages by Industry (Event Study) (a) Average Weekly Wages 0 2 4 6 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . (b) Warehousing (Wages.) −20 −10 0 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . (c) Retail Trade (Wages.) −5 0 5 10 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . (d) Wholesale Trade (Wages) −2 0 2 4 6 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . 82 Figure 9: Effect of Amazon Entry of Local Price Indexes (Event Study) (a) Regional Price Parities (RPP) −1 0 1 2 3 4 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Goods −3 −2 −1 0 1 2 3 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (c) Services −1 0 1 2 3 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (d) Utilities 0 10 20 30 40 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 83 Figure 10: Effect of Amazon Entry on Rents (Event Study) (a) Rent Prices (RPP) 0 5 10 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Rent Prices (ACS) 0 5 10 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (c) Home Values (Zillow) 0 10 20 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . (d) Home Values (ACS) 0 10 20 30 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 84 Figure 11: Effect of Amazon Entry on State and Local Government Finances (a) Total Expenditures −5 0 5 10 15 20 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . (b) Property Taxes 0 5 10 15 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . (c) Sales Tax −5 0 5 10 15 20 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . (d) Income Tax 0 5 10 15 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . 85 Figure 12: Effect of Amazon Entry on State/Local Corporate Subsidies (per cap.) (Event Study) −1 0 1 2 3 4 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P ct . C hg . 86 Figure 13: Effect of Amazon Entry on Employment and Industry Shares by Education- Age Group (Event Study) (a) Total Employment Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) 0 5 10 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Tran. and Warehousing (Emp.) Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) 0 1 2 3 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (c) Retail Trade (Emp.) Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −3 −2 −1 0 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (d) Wholesale Trade (Emp.) Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −0.75 −0.50 −0.25 0.00 0.25 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 87 Figure 14: Effect of Amazon Entry on Industry Shares by Education-Age Group (Event Study)(ctd.) (a) Tradeable Goods (Emp.) Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −1 0 1 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Tradeable Services (Emp.) Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −1 0 1 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (c) Non-Tradeable Services (Emp.) Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −1 0 1 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 88 Figure 15: Effect of Amazon Entry on Unemployment and Population by Education-Age Group (Event Study) (a) Unemployment Rate Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −2 −1 0 1 2 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Population, Age 16 to 64 Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −5 0 5 10 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 89 Figure 16: Effect of Amazon Entry on Average Weekly Wages by Industry and Education- Age Group (Event Study) (a) Average Weekly Wages Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) 0 4 8 12 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Tran. and Warehousing (Wages) Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −10 0 10 20 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (c) Tradeable Goods (Wages Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −5 0 5 10 15 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (d) Tradeable Services (Wages Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) 0 5 10 15 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 90 Figure 17: Effect of Amazon Entry on Rents and Home Values by Education-Age Group (a) Rent Prices (ACS) Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −5 0 5 10 15 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Home Values (ACS) Youth (Age 16−24) Non−College (Age 25+) College (Age 25+) −20 0 20 40 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 91 Figure 18: Effect of Amazon Entry on Employment, Selected Industry Shares and Home Values Controlling for Great Recession Severity (a) Total Employment −1 0 1 2 3 4 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (b) Tran. and Ware. (Pct. of Emp.) 0.0 0.5 1.0 1.5 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (c) Retail (Pct. of Emp.) −0.75 −0.50 −0.25 0.00 0.25 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . (d) Average Home Values (Zillow) 0 10 20 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 P .P . C hg . 92 Figure 19: Roth and Rambachan (2023) Sensitivity Tests (a) Employment Rate (%) 0 1 2 0.0 0.5 1.0 1.5 2.0 M P ct . C hg . (b) Home Values (Zillow) −10 0 10 20 0.0 0.5 1.0 1.5 2.0 M P ct . C hg . 93 Table 8: Effects of Amazon FC Network Expansion, 2009-2019 Variable Pre-Amazon Overall ATT SE Population, Age 16-64 751,306 0.67% 0.52 Employment Rate (%) 64.3 1.02 pp. 0.36 Tran. and Ware. Share (%) 3.3 0.25 0.07 Retail 12.1 -0.12 0.05 Wholesale 3.5 -0.11 0.03 Tradeable Goods 10.9 -0.12 0.11 Tradeable Services 11.6 0.19 0.10 Non-tradeable Services 50.3 -0.07 0.12 Average Weekly Wages ($) 869 0.65 % 0.38 Tran. and Ware. 900 -0.39 0.61 Retail 526 0.40 0.88 Wholesale 1,157 0.78 0.65 Tradeable Goods 1,133 0.30 0.71 Tradeable Services 1,341 0.69 0.55 Nontradeable Services 747 0.12 0.37 Unemployment Rate (%) 6.1 -0.28 pp. 0.12 RPP - Goods (100 = U.S. Average) 99 -0.32 0.38 RPP - Services 100 0.06 0.24 RPP - Utilities 98 5.85 2.29 RPP - Housing 97 1.09 0.86 Rent Price Index (ACS) 122 1.30 0.81 Typical Home Values (Zillow) ($) 178,778 5.57 % 1.71 Home Value Index (ACS) 120 3.55 1.94 State/Local Govt. Spending (per cap.) ($) 8,331 2.96 % 1.58 State/Local Govt. Revenue 2,481 1.49 1.26 Sales Tax 2,132 3.18 1.74 Property Tax 1,131 1.97 1.40 Income Tax 178 2.39 1.27 Federal I.G. Revenue 2,481 3.01 2.23 Corporate Subsidies 79 0.75 0.25 94 Table 9: Effects of Amazon Network Expansion by Education, 2009-2019 Variable Pre-Amazon Overall ATT SE Population - Youth (Age 16-24) 144,940 0.56% 0.71 Population - Non-College (Age 25+) 438,998 -0.29 0.58 Population - College (Age 25+) 175,419 0.93 0.88 Employment Rate - Youth 43.4 0.62 0.37 Employment Rate - Non-College 69.1 2.05 0.55 Employment Rate - College 77.0 -0.09 0.65 Unemployment Rate - Youth 9.7 -0.16 0.19 Unemployment Rate - Non-College 5.8 -0.34 0.14 Unemployment Rate - College 3.6 -0.38 0.15 Youth Tran. and Ware 2.1 0.49 0.12 Retail 22 -0.62 0.15 Wholesale 2.2 -0.09 0.03 Tradeable Goods 6.2 0.15 0.16 Tradeable Services 7.4 0.10 0.09 Non-Tradeable Services 60.2 -.04 0.18 Non-College Tran. and Ware 3.9 0.24 0.07 Retail 11.7 -0.15 0.05 Wholesale 4.4 -0.14 0.03 Tradeable Goods 13 0.05 0.11 Tradeable Services 10.5 0.31 0.11 Non-Tradeable Services 56.5 -0.32 0.13 College Tran. and Ware 2.2 0.21 0.05 Retail 6.9 -0.13 0.05 Wholesale 4.3 -0.15 0.04 Tradeable Goods 9.4 -0.09 0.11 Tradeable Services 18.9 0.12 0.15 Non-Tradeable Services 58.3 0.00 0.14 95 Table 10: Effects of Amazon Network Expansion by Education (ctd.), 2009-2019 Variable Pre-Amazon Overall ATT SE Average Weekly Wages - Youth 280 1.25% 0.74 Tran. and Ware 354 2.89 1.29 Retail 227 0.28 0.55 Wholesale 441 0.31 0.68 Tradeable Goods 479 1.65 1.05 Tradeable Services 404 2.74 1.04 Non-Tradeable Services 257 -0.08 0.68 Average Weekly Wages - Non-College 728 0.76 0.46 Tran. and Ware 761 -0.14 0.54 Retail 552 0.57 0.54 Wholesale 986 0.30 0.46 Tradeable Goods 940 -0.26 0.65 Tradeable Services 978 0.69 0.67 Non-Tradeable Services 642 0.22 0.41 Average Weekly Wages - College 1,269 1.36 0.55 Tran. and Ware 1,060 1.24 1.01 Retail 804 0.57 0.54 Wholesale 1,738 1.66 0.82 Tradeable Goods 1,583 0.52 1.17 Tradeable Services 1,729 0.81 0.72 Non-Tradeable Services 1,070 0.37 0.43 Rent Price Index (ACS) - Youth 118.4 2.71 1.21 Rent Price Index (ACS) - Non-College 121.5 0.50 0.63 Rent Price Index (ACS) - College 121.8 1.20 1.01 Home Price Index (ACS) - Youth 116.6 1.79 4.51 Home Price Index (ACS) - Non-College 115.0 2.66 1.48 Home Price Index (ACS) - College 120.6 3.49 2.20 96 Table 11: Effects of Amazon FC Network Expansion with 2007-09 Recession Controls, 2009-2019 Variable Pre-Amazon Overall ATT SE Population, Age 16-64 751,306 0.35% 0.49 Employment Rate (%) 64.3 0.66 pp. 0.32 Tran. and Ware. Share (%) 3.3 0.26 0.07 Retail 12.1 -0.11 0.05 Wholesale 3.5 -0.10 0.03 Tradeable Goods 10.9 -0.14 0.10 Tradeable Services 11.6 0.18 0.11 Non-tradeable Services 50.3 -0.1 0.11 Average Weekly Wages ($) 869 0.44 % 0.38 Tran. and Ware. 900 -0.64 0.55 Retail 526 0.16 0.88 Wholesale 1,157 0.79 0.83 Tradeable Goods 1,133 -0.16 0.73 Tradeable Services 1,341 0.61 0.56 Nontradeable Services 747 -0.03 0.35 Unemployment Rate (%) 6.1 -0.22 pp. 0.13 RPP - Goods (100 = U.S. Average) 99 -0.23 0.38 RPP - Services 100 0.06 0.26 RPP - Utilities 98 5.25 2.33 RPP - Housing 97 0.58 0.97 Rent Price Index (ACS) 122 0.87 0.84 Typical Home Values (Zillow) ($) 178,778 4.32 % 1.56 Home Value Index (ACS) 120 3.08 1.92 State/Local Govt. Spending (per cap.) ($) 8,331 3.16 % 1.70 Sales Tax 2,132 2.77 1.82 Property Tax 1,131 1.77 1.41 Income Tax 178 2.48 1.23 Federal I.G. Revenue 2,481 4.01 2.13 Corporate Subsidies 79 0.66 0.23 97 Table 12: Externally Set Parameters Desc. Source Value βY C.D. Preferences (Cons.) - Youth ACS 0.58 βNC C.D. Preferences (Cons.) - Non-College ACS 0.63 βC C.D. Preferences (Cons.) - College ACS 0.71 λY C.D. Preferences (Public Goods) - Youth βY 2 0.29 λNC C.D. Preferences (Public Goods) - Non-College βNC 2 0.31 λC C.D. Preferences (Public Goods) - College βC 2 0.35 α1 C.D. Production (Labor) - Trans. & Ware. NIPA 0.66 α2 C.D. Production (Labor) - Retail NIPA 0.55 α3 C.D. Production (Labor) - Wholesale NIPA 0.56 α4 C.D. Production (Labor) - T. Goods NIPA 0.59 α5 C.D. Production (Labor) - T. Serv. NIPA 0.50 α6 C.D. Production (Labor) - NT. Serv. NIPA 0.80 αR C.D. Production (Labor) - R.O.C. NIPA 0.63 κC Housing Supply Elas. (Inv.) - City Saiz (2010) 0.63 κR Housing Supply Elas. (Inv.) - R.O.C. Saiz (2010) 0.20 γYC Home Ownership - Youth, City ACS 50.7% γNCC Home Ownership - Non-College, City ACS 58.5 % γCC Home Ownership - College, City ACS 72.4 % γYR Home Ownership - Youth, R.O.C. ACS 49 % γNCR Home Ownership - Non-College, R.O.C. ACS 66.7 % γCR Home Ownership - College, R.O.C. ACS 79.2 % utilm Utility Costs (monthly) ACS $180 τFED Income Tax (Fed.) IRS (2015) Tables 13 & 14 τUI UI Payroll Tax (Fed.) DOL 0.054 98 Table 13: Federal Income Tax Brackets, 2015 Lower Bound Upper Bound Rate $0 $9,225 10% 9,226 37,450 15 37,451 90,750 25 90,751 189,300 28 Source: Internal Revenue Service (IRS) Table 14: Effective Federal Income Tax Rates τijm Location-Sector Youth Non-College College Tran. and Ware. 0.126 0.148 0.177 Retail 0.112 0.135 0.153 Wholesale 0.131 0.171 0.206 Tradeable Goods 0.132 0.167 0.201 Tradeable Services 0.129 0.170 0.206 Non-Tradeable Services 0.117 0.137 0.177 Unemployment 0.100 0.127 0.137 NILF 0.000 0.000 0.000 Rest of Country - Employment 0.120 0.137 0.175 Rest of Country - Unemployment 0.100 0.124 0.134 Rest of Country -NILF 0.000 0.000 0.000 Note: Effective tax rates calculated using 2015 Federal income tax brackets. 99 Table 15: Internally Calibrated Parameters Description Identification Value τSUB 1 Local Subsidy Rate - Tran./Ware. (18) 0.001 τSUB 2 Local Subsidy Rate - Retail (18) 0.000 τSUB 3 Local Subsidy Rate - Wholesale (18) 0.000 τSUB 4 Local Subsidy Rate - T. Goods (18) 0.003 τSUB 5 Local Subsidy Rate - T. Services (18) 0.000 τSUB 6 Local Subsidy Rate - N.T. Services (18) 0.000 A1C Productivity - Tran./Ware. Firm F.O.C. 1,470 εY1 Rel. Eff. - Youth - Tran./Ware. Firm F.O.C. 0.334 εNC 1 Rel. Eff. - Non-College - Tran./Ware. Firm F.O.C. 0.718 A2C Productivity - Retail Firm F.O.C. 1,361 εY2 Rel. Eff. - Youth - Retail Firm F.O.C. 0.282 εNC 2 Rel. Eff. - Non-College - Retail Firm F.O.C. 0.687 A3C Productivity - Wholesale Firm F.O.C. 880 εY3 Rel. Eff. - Youth - Wholesale Firm F.O.C. 0.254 εNC 3 Rel. Eff. - Non-College - Wholesale Firm F.O.C. 0.567 A4C Productivity - T. Goods Firm F.O.C. 2,847 εY4 Rel. Eff. - Youth - T. Goods Firm F.O.C. 0.303 εNC 4 Rel. Eff. - Non-College - T. Goods Firm F.O.C. 0.594 A5C Productivity - T. Services Firm F.O.C. 1,378 εY5 Rel. Eff. - Youth - T. Services Firm F.O.C. 0.234 εNC 5 Rel. Eff. - Non-College - T. Services Firm F.O.C. 0.566 A6C Productivity - N.T. Services Firm F.O.C. 4,004 εY6 Rel. Eff. - Youth - N.T. Services Firm F.O.C. 0.241 εNC 6 Rel. Eff. - Non-College - N.T. Services Firm F.O.C. 0.600 A6C Productivity - Rest of Country Firm F.O.C. 3,173 εY6 Rel. Eff. - Youth - Rest of Country Firm F.O.C. 0.273 εNC 6 Rel. Eff. - Non-College - Rest of Country Firm F.O.C. 0.643 100 Table 16: Internally Calibrated Parameters (ctd.) Description Identification Value TF Federal Transfers Fed. GBB $325 τSm Local Sales Tax Rate Local GBB 0.181 τPm Local Property Tax Rate (Households) Local GBB 0.101 τPLm Local Property Tax Rate (Landlords) Local GBB 0.094 τIm Local Income Tax Local GBB 0.009 zY1 Housing Cost Shifter - City - Youth Landlord F.O.C. 0.113 zNC 1 Housing Cost Shifter - City - Non-College Landlord F.O.C. 0.036 zC1 Housing Cost Shifter - City - College Landlord F.O.C. 0.053 z0 Housing Cost Shifter - Rest of Country Landlord F.O.C. 0.404 Table 17: Cross-Sector Spillover Parameters Description Value σA 12 Prod. Spillover Elasticity - Retail 0.287 σA 13 Prod. Spillover Elasticity - Wholesale 0.079 σA 14 Prod. Spillover Elasticity - T. Goods 0.000 σA 15 Prod. Spillover Elasticity - T. Services 0.694 σA 16 Prod. Spillover Elasticity - N.T. Services 0.000 σεY 12 Eff. Spillover Elasticity - Retail -1.014 σεNC 12 Eff. Spillover Elasticity - Retail 1.142 σεY 13 Eff. Spillover Elasticity - Wholesale -1.077 σεNC 13 Eff. Spillover Elasticity - Wholesale 1.212 σεY 14 Eff. Spillover Elasticity - T. Goods 1.077 σεNC 14 Eff. Spillover Elasticity - T. Goods 0.000 σεY 15 Eff. Spillover Elasticity - T. Services 1.702 σεNC 15 Eff. Spillover Elasticity - T. Services 0.000 σεY 16 Eff. Spillover Elasticity - N.T. Services 0.000 σεNC 16 Eff. Spillover Elasticity - N.T. Services 0.000 101 Table 18: Extreme Value Shocks and Goodness of Fit Ũijm µijm E[ϵijm|Uijm > Uikl] π̂ijm πijm Youth Tran. and Ware. 8.15 -2.14 -0.92 0.7% 0.8% Retail 7.74 0.65 -0.51 7.9 8.0 Wholesale 8.35 -1.87 -1.11 1.2 0.8 Tradeable Goods 8.42 -1.30 -1.19 2.3 2.3 Tradeable Services 8.27 -0.90 -1.04 2.9 2.7 Non-tradeable Services 7.86 1.54 -0.63 21.7 21.9 NILF 7.62 2.16 -0.39 31.9 32.1 Rest of Country - Employment 6.22 2.82 1.02 15.0 15.3 Rest of Country - NILF 5.90 3.21 1.34 16.5 16.3 Non-College Tran. and Ware. 9.15 -1.60 -1.01 2.1 2.0 Retail 8.86 -0.26 -0.72 6.2 6.0 Wholesale 9.36 -1.79 -1.22 2.3 2.3 Tradeable Goods 9.32 -0.61 -1.18 6.9 6.7 Tradeable Services 9.36 -0.86 -1.22 5.6 5.4 Non-tradeable Services 9.00 1.15 -0.86 28.8 29.0 NILF 7.92 1.70 0.22 16.9 17.2 Rest of Country - Employment 7.09 2.76 1.05 21.6 21.5 Rest of Country - NILF 5.92 3.12 2.21 9.7 10.0 College Tran. and Ware. 9.76 -1.08 -0.87 1.8 1.3 Retail 9.52 0.00 -0.62 4.0 4.1 Wholesale 10.21 -1.16 -1.31 2.5 2.6 Tradeable Goods 10.13 -0.29 -1.23 5.7 5.6 Tradeable Services 10.21 0.32 -1.31 11.2 11.2 Non-tradeable Services 9.77 1.89 -0.88 34.8 34.7 NILF 8.18 2.57 0.71 14.0 14.3 Rest of Country - Employment 7.58 3.38 1.32 17.2 17.3 Rest of Country - NILF 5.99 4.29 2.91 8.7 8.9 Note: Expected values are normalized so average for each education type is zero. 102 Table 19: Data and Parameters for Rest of Country, 2009-2019 Variable Value Data Population (Pct. of U.S.) 30.4% Total 100.0 Youth 20.2 Non-College 61.0 College 18.8 Employment Rate - Youth 39.4 Employment Rate - Non-College 63.1 Employment Rate - College 63.7 Unemployment Rate - Youth 8.9 Unemployment Rate - Non-College 5.2 Unemployment Rate - College 2.4 Average Weekly Wages - Youth $282 Average Weekly Wages - Non-College 666 Average Weekly Wages - College 1,036 Goods/Services Price Index (U.S. = 100) 98.6 Utility Prices Index 97.6 Utility Spending (per month) $180 Typical Home Values (Zillow) 162,755 Rental Price Index (ACS) 113 Parameters C.D. Production (Labor) (α) 0.63 C.D. Preferences (Cons.) - Youth (βY) 0.58 C.D. Preferences (Cons.) - Non-College (βNC) 0.63 C.D. Preferences (Cons.) - College (βC) 0.71 103 Table 20: Internally Calibrated Parameters - ”Data-Driven” Approach Description Pre-Amazon Post-Amazon τSUB 1 Local Subsidy Rate - Tran./Ware. 0.001 0.001 τSUB 2 Local Subsidy Rate - Retail 0.000 0.000 τSUB 3 Local Subsidy Rate - Wholesale 0.000 0.001 τSUB 4 Local Subsidy Rate - T. Goods 0.003 0.004 τSUB 5 Local Subsidy Rate - T. Services 0.000 0.001 τSUB 6 Local Subsidy Rate - N.T. Services 0.000 0.000 A1C Productivity - Tran./Ware. 1,470 1,530 εY1 Rel. Eff. - Youth - Tran./Ware. 0.334 0.339 εNC 1 Rel. Eff. - Non-College - Tran./Ware. 0.718 0.708 A2C Productivity - Retail 1,361. 1,376 εY2 Rel. Eff. - Youth - Retail 0.282 0.278 εNC 2 Rel. Eff. - Non-College - Retail 0.687 0.676 A3C Productivity - Wholesale 880 883 εY3 Rel. Eff. - Youth - Wholesale 0.254 0.250 εNC 3 Rel. Eff. - Non-College - Wholesale 0.567 0.558 A4C Productivity - T. Goods 2,847 2,868 εY4 Rel. Eff. - Youth - T. Goods 0.303 0.308 εNC 4 Rel. Eff. - Non-College - T. Goods 0.594 0.594 A5C Productivity - T. Services 1,378 1,417 εY5 Rel. Eff. - Youth - T. Services 0.234 0.239 εNC 5 Rel. Eff. - Non-College - T. Services 0.566 0.566 A6C Productivity - N.T. Services 4,004 4,014 εY6 Rel. Eff. - Youth - N.T. Services 0.241 0.241 εNC 6 Rel. Eff. - Non-College - N.T. Services 0.600 0.600 A6C Productivity - Rest of Country 3,173 3,173 εY6 Rel. Eff. - Youth - Rest of Country 0.273 0.273 εNC 6 Rel. Eff. - Non-College - Rest of Country 0.643 0.643 104 Table 21: Internally Calibrated Parameters - ”Data-Driven” Approach (ctd.) Description Pre-Amazon Post-Amazon TF Federal Transfers $325 $328 τSm Local Sales Tax Rate 0.181 0.184 τPm Local Property Tax Rate (Households) 0.101 0.103 τPLm Local Property Tax Rate (Landlords) 0.094 0.098 τIm Local Income Tax 0.009 0.009 zY1 Housing Cost Shifter - City - Youth 0.113 0.116 zNC 1 Housing Cost Shifter - City - Non-College 0.036 0.036 zC1 Housing Cost Shifter - City - College 0.053 0.055 z0 Housing Cost Shifter - Rest of Country 0.404 0.403 105 Table 22: Extreme Value Shocks in ”Data-Driven” Approach Pre-Amazon Post-Amazon µijm E[ϵijm] µijm E[ϵijm] Youth Tran. and Ware. -2.14 -0.92 -1.96 -0.93 Retail 0.65 -0.51 0.62 -0.51 Wholesale -1.87 -1.11 -1.89 -1.09 Tradeable Goods -1.30 -1.19 -1.26 -1.17 Tradeable Services -0.90 -1.04 -0.91 -1.04 Non-tradeable Services 1.54 -0.63 1.57 -0.62 NILF 2.16 -0.39 2.11 -0.44 Rest of Country - Employment 2.82 1.02 2.90 1.09 Rest of Country - NILF 3.21 1.34 3.24 1.36 Non-College Tran. and Ware. -1.60 -1.01 -1.53 -1.01 Retail -0.26 -0.72 -0.32 -0.74 Wholesale -1.79 -1.22 -1.83 -1.19 Tradeable Goods -0.61 -1.18 -0.56 -1.17 Tradeable Services -0.86 -1.22 -0.86 -1.19 Non-tradeable Services 1.15 -0.86 1.20 -0.87 NILF 1.70 0.22 1.45 0.02 Rest of Country - Employment 2.76 1.05 2.93 1.21 Rest of Country - NILF 3.12 2.21 3.30 2.33 College Tran. and Ware. -1.08 -0.87 -1.11 -0.86 Retail 0.00 -0.62 0.01 -0.64 Wholesale -1.16 -1.31 -1.12 -1.30 Tradeable Goods -0.29 -1.23 -0.25 -1.20 Tradeable Services 0.32 -1.31 0.40 -1.28 Non-tradeable Services 1.89 -0.88 1.94 -0.87 NILF 2.57 0.71 2.44 0.49 Rest of Country - Employment 3.38 1.32 3.52 1.40 Rest of Country - NILF 4.29 2.91 4.37 2.95 Note: Expected values are normalized so average for each education type is zero. 106 Table 23: Welfare Effects of Amazon FC Expansion by Education and Home Ownership WTP ($) Pct. of Income Renters Youth -$13 -0.1 % Non-College -93 -0.2 College -92 -0.1 Home Owners Youth 210 1.4 Non-College 933 2.4 College 520 0.8 Total 329 0.8 Note: Expressed at annual frequency. Table 24: Welfare Effects of Amazon FC Expansion by Education WTP ($) Pct. of Income Youth $93 0.6% Non-College 433 1.1 College 337 0.5 Total 329 0.8 Note: Expressed at annual frequency. 107 Ta bl e 25 :W el fa re Ef fe ct s of A m az on FC Ex pa ns io n by C ha nn el W T P ($ ) Pc t. of In co m e To ta l Yo ut h N on -C ol le ge C ol le ge To ta l Yo ut h N on -C ol le ge C ol le ge La bo r M ar ke t D ir ec tE m p. Ef fe ct s $6 $4 $5 $8 0. 01 % 0. 02 % 0. 01 % 0. 01 % C ro ss -S ec to r Sp ill ov er s 24 18 25 23 0. 06 0. 12 0. 06 0. 03 % H ou si ng M ar ke t C os to fL iv in g (R en ts ,U ti lit ie s) -3 9 -2 6 -4 0 -4 5 -0 .0 9 -0 .1 7 -0 .1 0 -0 .0 7 H om e V al ue s 41 9 10 8 57 1 45 1 1. 02 0. 71 1. 46 0. 67 Pu bl ic Fi na nc es Lo ca lP ub lic G oo ds 11 5 11 22 0. 03 0. 03 0. 03 0. 03 C or po ra te Su bs id ie s -1 0 -1 -2 0. 00 0. 00 0. 00 0. 00 O th er N on -W ag e A m en it ie s -1 08 -2 6 -1 50 -1 03 -0 .2 6 -0 .1 7 -0 .3 8 -0 .1 5 M ig ra ti on 0 0 0 0 0. 00 0. 00 0. 00 0. 00 To ta l 32 9 93 43 3 33 7 0. 80 0. 62 1. 11 0. 50 108 Table 26: Shocks to Transportation and Warehousing Sector Parameter Pre-Amazon Shock A11 Productivity 1,470 +4.0% εY11 Rel. Efficiency - Youth 0.334 +1.6% εNC 11 Rel. Efficiency - Non-College 0.718 -1.4% Table 27: Model-Generated Outcome Changes - Structural Approach Outcome Model Data rY1 Rent - Youth +0.3% +2.7% rNC1 Rent - Non-College +0.8% +0.5% rC1 Rent - College -0.4% +1.2% νm Home Values +0.7% +5.6% gm Local Public Goods +2.2% +2.9% Property Tax Revenue (H.H) +0.6% +1.4% Sales Tax Revenue +0.8% +1.7% Local Income Tax Revenue +1.5% +1.3% IGRm Federal Grants +4.8% +2.2% Property Tax Revenue (Landlord) +0.9% +1.4% 109 Table 28: Welfare Effects of Amazon FC Expansion by Education and Home Ownership - Structural Appraoch WTP ($) Pct. of Income Renters Youth $1 0.0% Non-College 15 0.0 College -16 0.0 Home Owners Youth 30 0.2 Non-College 163 0.4 College 65 0.1 Total 72 0.2 Note: Expressed at annual frequency. Table 29: Welfare Effects of Amazon FC Expansion by Education - Structural Approach WTP ($) Pct. of Income Youth $15 0.1% Non-College 97 0.3 College 42 0.1 Total 72 0.2 Note: Expressed at annual frequency. 110 Table 30: Extreme Value Shocks in ”Structural” Approach Pre-Amazon Post-Amazon µijm E[ϵijm] µijm E[ϵijm] Youth Tran. and Ware. -2.14 -0.92 -1.84 -0.94 Retail 0.65 -0.51 0.63 -0.50 Wholesale -1.87 -1.11 -1.85 -1.11 Tradeable Goods -1.30 -1.19 -1.28 -1.20 Tradeable Services -0.90 -1.04 -0.93 -1.05 Non-tradeable Services 1.54 -0.63 1.56 -0.62 NILF 2.16 -0.39 2.14 -0.41 Rest of Country - Employment 2.82 1.02 2.87 1.06 Rest of Country - NILF 3.21 1.34 3.24 1.35 Non-College Tran. and Ware. -1.60 -1.01 -1.39 -1.00 Retail -0.26 -0.72 -0.28 -0.71 Wholesale -1.79 -1.22 -1.73 -1.21 Tradeable Goods -0.61 -1.18 -0.60 -1.17 Tradeable Services -0.86 -1.22 -0.81 -1.21 Non-tradeable Services 1.15 -0.86 1.19 -0.85 NILF 1.70 0.22 1.60 0.17 Rest of Country - Employment 2.76 1.05 2.84 1.11 Rest of Country - NILF 3.12 2.21 3.20 2.25 College Tran. and Ware. -1.08 -0.87 -1.08 -0.88 Retail 0.00 -0.62 0.04 -0.64 Wholesale -1.16 -1.31 -1.10 -1.33 Tradeable Goods -0.29 -1.23 -0.28 -1.23 Tradeable Services 0.32 -1.31 0.37 -1.31 Non-tradeable Services 1.89 -0.88 1.94 -0.88 NILF 2.57 0.71 2.63 0.66 Rest of Country - Employment 3.38 1.32 3.49 1.35 Rest of Country - NILF 4.29 2.91 4.35 2.91 Note: Expected values are normalized so average for each education type is zero. 111 Table 31: Sample Averages, Corporate Subsidies, 2009-2019 Variable Pre-Amazon Post-Amazon Corporate Subsidies, Total $2,352,655 $6,202,088 Tran. and Ware. 434,639 1,418,162 Retail 316,951 1,177,096 Wholesale 312,944 961,045 Tradeable Goods 1,764,471 4,610,178 Tradeable Services 1,032,297 3,387,521 Non-Tradeable Services 913,329 2,599,606 Table 32: Sample Medians, Corporate Subsidies, 2009-2019 Variable Pre-Amazon Post-Amazon Corporate Subsidies, Total $1,110,126 $3,880,595 Tran. and Ware. 0 0 Retail 0 0 Wholesale 0 0 Tradeable Goods 819,717 2,771,476 Tradeable Services 0 1,346,388 Non-Tradeable Services 0 864,091 112 Appendix A ACS Rental and Home Value Indices In order to supplement the Regional Price Parity index and Zillow Home Value data, and to examine heterogeneity by worker type, I create custom rent price and home value indices using micro data from the American Community Survey (ACS), accessed via IPUMS. (Ruggles et al. 2022) To do this, I create a sample of all heads of household between the ages of 16 and 64, who live in dwellings that contain all basic necessities, including electricity, heating, a kitchen, etc. I estimate a hedonic regression, that allows me to adjust rents and home values for observable characteristics such as size, number of rooms, etc. Yimt = αmt + θt + βXmt + ϵmt (A.1) I control for nationwide trends in rent and home values with θt, and Xit contains a number of observable housing characteristics including square footage, number of rooms, number of bedrooms, and decade built, etc. Yimt refers to either the monthly contract rent, or reported home value. I then collect all the αmt’s, which reflect metro-year average rents or home values, adjusted for observable characteristics. I then adjust each value so the average value for the entire U.S. is 100 in each year. 113 Appendix B State and Local Government Finances Here I briefly describe the construction of my dataset on state and local government fi- nances at the metro level. The Annual Survey of State and Local Government Finances is an annual survey conducted by the U.S. Census Bureau. They collect information on revenues, expenditures, debt and assets for all 70,000 state and local governments in the U.S. in years that end in 2 and 7, and for a stratified representative sample of 11,000 gov- ernments in the intervening years. I use the Individual Unit file, which reports responses at the government level, to aggregate local spending up to the metro level. In ”census” years, it is straightforward, as data is collected from all governments. In the intervening years however, I need to use an adjustment factor for metro level estimates. I calculate this factor by calculating the ratio of spending by sample units to total spending within a metro in Census years. Since the ratio is different in each Census year, I use a linear interpolation to calculate adjustment factors for non-census years. When I calculate expenditures, I subtract revenue from current government charges. This is so my measure of state and local government spending is most consistent with my model; in my model, government services funded by user fees as a opposed to dis- torting taxes would fall under standard firm production, as public administration is a 114 part of non-tradeable services. Due to the variety of centralization across U.S. states, I allocate state government spending to metro areas within their state based on shares of intergovernmental remittances. 115 Appendix C Good Jobs First Corporate Subsidy Tracker Data on the value of state and local corporate subsidies comes from the Good Jobs First Corporate Subsidy Tracker. In my Analysis I drop all federal subsidies. In converting the data to monetary observations, I use the total subsidy value, which I consider to be a con- servatively large measure of the fiscal burden local governments face as a result of their subsidy programs. When observations did not include a specific geographic area within a state, they were allocated to metros within a state by the share of total government expenditure. 116 Appendix D Callaway and Sant’Anna (2020) In the Callaway and Sant’Anna (2021) staggered roll-out DID framework, the building blocks are referred to as ”group-time average treatment effects”. ATT(g, t) = E[Ymt(g) − Ymt(0)|Gmg = 1] (D.1) These effects can then be aggregated up to produce summary measures of the effect of Amazon’s expansion. In my context, ATT(g, t) is the average effect in year t for metros where Amazon entered (opened their first FC or warehouse) in year g. I temporarily assume Y is the local employment rate, t is 2014, and g is 2012. In this case, I would want to estimate the difference between the observed employment rate in 2014 for metros receiving their first FC in 2012 and the (unobserved) employment rate that would have prevailed in 2014 in the absence of Amazon. In practice, I estimate this counterfactual using metros where Amazon first enters from 2015-2021. In order to identify these treatment effects, a few assumptions are required. First, note that after opening a new FC, Amazon does not exit from any market in the sam- ple during my period of analysis. Thus, (1) treatment is irreversible. Motivated by my definition of the treatment date above, I assume (2) no anticipation effects. I assume that 117 {Ym2009, Ym2010, ...,Ym2019,Dm2009,Dm2010, ...,Dm2019} - the full panel for each market - is (3) independently and identically distributed. Last and most importantly is a (4) parallel trends assumption based on not-yet-treated groups. Formally, for all g and each (s, t) ∈ {2, .., T }× {2, ..., T } where t > g and t ⩽ s < gmax: E[Ymt(0) − Ymt−1(0)|Gmg = 1] = E[Ymt(0) − Ymt−1(0)|Dms = 0,Gmg = 0] (D.2) Using my example from above, this assumption requires that E[∆Ym2012],E[∆Ym2013], and E[∆Ym2014] be equivalent for metros where Amazon enters in 2012, and from 2015- 2021. It is worth noting that I make this assumption unconditional on covariates. This framework accommodates a conditional version of this assumption with time-varying controls independent of treatment, but my empirical results show that parallel trends likely hold unconditionally for all outcome variables in this study. Drawing on discus- sion in Sant’Anna and Marcus (2021) and Sant’Anna, Ghanem, and Wuthrich (2023), I note that my strategy is robust to selection on time-fixed market characteristics. Amazon choosing to enter larger, wealthier markets first is not a threat, so long as these differences do not induce diverging trends in outcomes. Selection on time-varying observable char- acteristics would require those characteristics to be independent of outcomes for identi- fication, but my findings show no evidence of such differences for a wide range of local economic outcomes. Following Sant’Anna and Marcus (2021), for each treatment group I compare pre-treatment changes in outcomes to those of the relevant control group, and find all moment restrictions implied by my parallel trends assumption are satisfied. This is a direct test, rather than a placebo or falsification test, so the lack of violations is strong supporting evidence my identification strategy is sound. These tests also confirm that for each individual treatment and control group, there is no evidence of selection bias on 118 time-varying observable market characteristics. Under these assumptions, Callaway and Sant’Anna (2021) show the group-time av- erage treatment effects are point-identified, and can be calculated from the following ex- pression: ATTNY unc(g, t) = E[Ymt − Ymg−1|Gmg = 1] − E[Ymt − Ymg−1|Dmt = 0] (D.3) The last step is to aggregate these group-time average treatment effects. I focus on aggregation methods that calculate an overall average treatment effect of an expansion of Amazon’s FC network, and the average treatment effect by length of exposure - a mea- sure akin to those used in traditional two-way fixed effects event study approaches. The group-time ATTs can be aggregated into an overall treatment effect estimate using the following: θ = ∑ g 1 T − g+ 1 T∑ t=g ATT(g, t)P(G = g|G ⩽ T) (D.4) This equation calculates the average effect for each treatment group across all years, and then averages them across all groups. The interpretation is similar to an ATT in a standard two-period, two-group DID framework. To calculate event-study style ATTs for different lengths of exposure, I use: θES(e) = ∑ g ⊮ {g+ e ⩽ T }P(G = g|G+ e ⩽ T)ATT(g,g+ e) (D.5) where e = t − g is ”event-time”. While these estimates may lend themselves to an iden- tical interpretation as the conventional two-way fixed effects event study estimates, this method avoids the shortcoming of the latter including the ”negative weight problem” from Goodman-Bacon (2021) and Sun and Abraham (2021). Further, this framework does 119 not impose any restrictions on the heterogeneity of treatment effects. During estimation, I use the ”doubly robust” version of Callaway and Sant’Anna (2021) and the standard bootstrapped standard errors, clustered at the market level. Though recent research has shown conventional ”pre-tests” of parallel trends in event study frameworks suffer from low power (Roth 2022), graphically confirming pre-treatment coefficients do not differ from zero can provide an additional sanity check. 120 Acknowledgments Abstract The Local Impact of Amazon: Empirical Estimates Introduction Related Literature and Contributions Amazon's Distribution Network and Expansion Data and Empirical Strategy Data Empirical Strategy Results Heterogeneous Effects Across Education-Age Groups Robustness Conclusion A Spatial Equilibrium Model for Welfare Analysis Introduction Model Definition of Equilibrium Welfare Expression Outline of Calibration Conclusion Welfare Effects and Lessons for Local Economic Development Policy Introduction Calibration in the Amazon Context Welfare Effects "Data-Driven" Approach Structural Approach Policy Implications Conclusion Bibliography ACS Rental and Home Value Indices State and Local Government Finances Good Jobs First Corporate Subsidy Tracker Callaway and Sant'Anna (2020)