Essays on Banking and Local Credit Markets
by Hoai-Luu Q. Nguyen
S.B. Economics, Minor in Mathematics, Massachusetts Institute of Technology (2007)
Submitted to the Department of Economics in partial fulfillment of the requirements for the degree of MASSACHUSETTS INSTITUTE Doctor of Philosophy OF TECHNOLOLGY at the JUN 09 2015 MASSACHUSETTS INSTITUTE OF TECHNOLOGY LIBRARIES· June 2015 © 2015 Hoai-Luu Q. Nguyen. All rights reserved. The author hereby grants to MIT permission to reproduce and distribute publicly paper and electronic copies of this thesis document in whole or in part. Signature redacted Author ...... y Department of Economics May 15, 2015 Signature redacted Certified by ...... Michael Greenstone Milton Friedman Professor of Economics, University of Chicago Signature redacted Thesis Supervisor Certified by ...... Robert M. Townsend Elizabeth & James Killian Professor of Economics Thesis Supervisor Signature------redacted Accepted by .. /~ Ricardo J. Caballero Ford International Professor of Economics Chairman, Departmental Committee on Graduate Studies
Essays on Banking and Local Credit Markets by Hoai-Luu Q. Nguyen
Submitted to the Department of Economics on May 15, 2015, in partial fulfillment of the requirements for the degree of Doctor of Philosophy
Abstract
This thesis consists of three chapters on banking and local credit markets. The first chapter studies the relationship between bank-specific capital and credit access in a new setting: bank branch closings in markets where the branch network is dense. Existing regulation in the U.S. is targeted toward areas with few branches where closings inhibit physical access to the branch network. I show that, even in crowded markets, closings can have large effects on local credit supply. To generate plausibly exogenous variation in the incidence of closings, I use Census tract level data paired with a novel identification strategy that exploits within-county variation in exposure to post-merger consolidation. This instrument identifies the effect of closings that occur in close proximity to other branches. I find that closings have a prolonged negative impact on credit supply to local small businesses, but only a temporary effect on local mortgage lending. The number of new small business loans is 13% lower for several years, and this decline persists even after the entry of new banks. The decline in lending is highly localized, dissipating 8 miles out, and is concentrated in low-income and high-minority neighborhoods. These results show closings have large effects on local credit supply when lending is information-intensive and lender-specific relationships are difficult to replace. The second chapter (co-authored with Michael Greenstone and Alexandre Mas) estimates the effect of the reduction in credit supply that followed the 2008 financial crisis on the real economy. We predict county lending shocks using variation in pre-crisis bank market shares and estimated bank supply-shifts. Counties with negative predicted shocks experienced declines in small business loan originations, indicating that it is costly for these businesses to find new lenders. Using confidential microdata from the Longitudinal Business Database, we find that the 2007-2009 lending shocks accounted for statistically significant, but economically small, declines in both small firm and overall employment. Predicted lending shocks affected lending but not employment from 1997-2007. The third chapter uses a cash demand framework to model household credit decisions when there are both fixed and marginal costs associated with borrowing. In standard models of credit demand, the price associ ated with a loan is simply the interest rate. In reality, however, loan contracts encompass many dimensions that contribute to the effective price a household pays to borrow. Understanding how these other factors influence households' credit decisions is important for evaluating the impact of policy on household credit demand. I show, using data from Thailand, that the cash demand model matches many observed patterns of household behavior while providing a framework for understanding how tradeoffs between different costs drive borrowing decisions.
Thesis Supervisor: Michael Greenstone Title: Milton Friedman Professor of Economics, University of Chicago
Thesis Supervisor: Robert M. Townsend Title: Elizabeth & James Killian Professor of Economics
3 Contents
Acknowledgements 7
List of Tables 9
List of Figures 11
1 Do Bank Branches Still Matter? The Effect of Closings on Local Economic Outcomes 13
1.1 Introduction . 13
1.2 Data 16
1.3 Identification and Empirical Framework 18
1.3.1 External Validity 21
1.4 Results ...... 21
1.4.1 Exposure to Consolidation and Branch Closings . 21
1.4.2 Closings and Local Credit Supply 22
1.4.2.1 Alternative Explanations 24
1.4.2.2 Varying the Size of the Local Banking Market 25
1.4.3 Heterogeneity Across Borrowers 25
1.4.4 Geographic Spillovers 27
1.4.5 Real Economic Effects 27
1.5 Welfare Implications 29
1.6 Conclusion 31
Figures 33
Tables . 39
4 2 Do Credit Market Shocks Affect the Real Economy? Quasi-Experimental Evidence from the Great Recession and 'Normal' Economic Times 46
2.1 Introduction . 46
2.2 Background . 49
2.3 Data Sources 50
2. 4 Research Design 52
2.4.1 Isolating Supply Shocks in Lending . 52
2.4.2 Summary of the Predicted Supply Shock . 54
2.5 Econometric Models and Results ...... 55
2.5.1 The Relationship Between the Predicted Lending Shock and Actual Loan Originations 55
2.5.2 The Relationship Between the Predicted Lending Shocks and Economic Activity Dur- ing the Great Recession ...... 58
2.5.2.1 Small Standalone Firms . 58
2.5.2.2 Small Establishments in Multi-Unit Firms 58
2.5.2.3 County-Level Economic Outcomes ..... 59
2.5.3 The Role of Small Business Loans in "Normal" Economic Times. 60
2.6 Interpretation 61
2. 7 Conclusion 63
Figures 65
Tables . 68
3 Credit is Cash: A Model of Household Borrowing 77
3.1 Introduction ...... 77
3.2 Household Credit in Thailand 79
3.2.1 Data ...... 79
3.2.2 Loan Products 80
3.2.3 Household Borrowing Patterns 82
3.3 Credit Demand as Cash Demand 83
3.3.1 Model ...... 84
5 3.3.2 Simulations 87
3.4 Discrete Choice 88
3.5 Conclusion 90
Figures 91
Tables . 94
References 99
Appendices for Chapter 1 106
Appendix Figures 112
Appendix Tables 115
Appendices for Chapter 2 121
Appendix Figures 125
Appendix Tables 127
6 Acknowledgements
One of the greatest pleasures of finishing my Ph.D. is having the opportunity to acknowledge and give thanks to all the individuals who helped me along the way. I have been extremely touched and humbled by their support, and I hope I am able to do them justice in these few paragraphs.
I have had many occasions to be thankful that Michael Greenstone was the thesis writers' registration officer
in my third year. I used his initial flicker of interest in my research as fodder for so many meetings that he was eventually forced to ask if he had become my advisor. Michael has been a mentor in the truest sense of the word, and I am especially grateful that he steered me firmly back on track the spring before my job market year. I would be in a very different place were it not for him.
Rob Townsend reached out to me while I was still a prospective student, and has guided me with the same warmth and generosity ever since. His encouragement and enthusiasm gave me the confidence I needed when starting on my own work, and yet he never shied from challenging me to go in directions I would not have otherwise considered. Our conversations have been an invaluable part of my time at MIT.
I am also deeply indebted to the rest of the MIT Economics faculty, whose dedication to supporting their
graduate students is truly unrivaled. I am especially thankful to the development faculty, who continued to welcome me as part of their community even as my research drifted farther and farther afield. I particularly thank Abhijit Banerjee, whose comments during my seminars always baffled me until I'd had sufficient time to think about them and realize that he was right, and Ben Olken, especially for his support during my job market year.
When I returned to MIT for grad school, everyone told me that the biggest benefit would be my peer group. However, it is only now, at the end, that I am able to step back and truly appreciate how lucky I've been to be a part of this community. The students I've met here have been simultaneously the most brilliant
and the most humble individuals I've had the privilege to work with. While I am grateful to many, three deserve special mention. Adrien Auclert and I cemented our friendship over long sessions at the squash
courts, which provided a much-needed break and a chance to vent about the stresses of grad school. He was also unfailingly generous with his time and expertise on the numerous occasions I was instructed to "write a model" and had no idea where to begin. Adam Sacarny and I share the bond unique to New York Fed RAs who are scheduled for their pre-employment drug screening on the same day. He has been extremely tolerant of my continued promise to invite him over for pho... someday. Finally, Manasi Deshpande was the first person I met from our cohort and she has been my strongest support ever since. For me, she is the one person who best epitomizes that combination of brilliance and graciousness, and I cannot imagine what grad school would have been without her.
The MIT Economics department is much more than just its faculty and students, and I am very thankful to the support staff who power it everyday. I am especially grateful to Gary King, our de facto leader; to Kim
7 Scantlebury, for making sure I never took things too seriously; and to Beata Shuster, whose bright spirits made my fifth year so much more enjoyable.
The Capital Markets Function at the New York Fed played a huge role in sparking my interest in economic research. Tobias Adrian was my first mentor in this profession and the single biggest reason I decided to go to grad school. Amidst all the quips, Richard Crump provided many nuggets of valuable life advice. Daniel Green and Ariel Zucker are fellow CM alums who came to MIT, and who I hope will continue the pattern and join me on the West Coast soon.
While my friends in the department carried me through the day-to-day, I was also very fortunate that Alice Fan, my college roommate, moved back to Boston during my second year. Our monthly dinners were welcome reprieves from my grad school bubble, and our anniversary trips are a tradition that I hope will continue for many years to come.
It is wildly insufficient to say that I am grateful to my family for their support while I pursued my Ph.D., because I am grateful to them for so much more than that. To my parents, I can only say that they have always been everything to me. My sister, while being one of the busiest and most productive people I've ever known, has never failed to make time for me. My brother cooked me dinner every night during my first year, and has tolerated having his little sister follow him to every city he's ever lived in. To all of them, I am more grateful than I can say.
Finally, Peter Lee has been my partner and best friend for the last 11 years. More than anyone else, he has been forced to contend with the uglier side of grad school, and yet he has been unwavering throughout. Now that we are finally at the end, I cannot wait for what lies ahead.
8 List of Tables
1.1 M erger Sam ple ...... 3 9
1.2 Merger Summary Statistics ...... 3 9
1.3 Summary Statistics for Exposed and Control Tracts ...... 4 0
1.4 Complier Characteristics ...... 41
1.5 Exposure to Consolidation and Branch Closings ...... 42
1.6 The Effect of Closings on Local Credit Supply ...... 43
1.7 Heterogeneity by Tract-Level Income and Fraction Minority ...... 4 4
1.8 Exposure to Consolidation and ZIP-Level Outcomes ...... 4 5
2.1 Changes in Lending Between 2007-2009 for Selected Large Bank Holding Companies . 68
2.2 County Characteristics ...... 69
2.3 Testing for Spatial Sorting in Bank Lending Shocks ...... 70
2.4 Relationship between Predicted Lending Shock and ln(loan originations) ...... 71
2.5 Effect of Predicted Lending Shock on Employment and Establishment Growth Rates for Small Standalone Firm s ...... 72
2.6 Effect of Predicted Lending Shock on Employment Growth Rates for Small Establishments that are Part of Multi-Unit Firms ...... 73
2.7 Effect of Predicted Lending Shock on County Aggregate Outcomes ...... 74
2.8 Effect of Predicted Lending Shock on Small Business Employment by Year ...... 75
2.9 Two Stage Least Squares Models of the Relationship between Economic Activity and Small Business Loan Originations ...... 76
3.1 Household Summary Statistics ...... 94
9 3.2 Variation in Loan Characteristics ...... 9 4
3.3 Location of Active Lending Institutions ...... 9 5
3.4 Variation in Loans Across Institutions ...... 9 5
3.5 Household Borrowing Patterns ...... 9 6
3.6 Borrowing Patterns Across Sectors ...... 9 6
3.7 Characteristics Associated with Borrowing Ou tcomes ...... 9 7
3.8 Funding Source by Project Type ...... 9 7
3.9 Sources of Emergency Funding ...... 9 7
3.10 Household A's Cost of Borrowing ...... 9 8
3.11 Household B's Cost of Borrowing ...... 9 8
A1.1 Geocoding Summary Statistics ...... 1 15
A1.2 Failing/Crisis Mergers ...... 1 15
A1.3 Buyer and Target Small Business Lending Intensity ...... 116
A1.4 The Effect of Closings on Credit Supply in Target Only Tracts ...... 116
A1.5 Robustness to Varying Size of the Local Banking Market ...... 117
A1.6 Correlation between Tract Demographics and Branch Levels ...... 117
A1.7 Summary Statistics for Exposed and Control ZIP s ...... 118
A1.8 Industry Dependence on External Finance ...... 1 19
A1.9 Summary Statistics for Households Living in Exposed and Control Tracts ...... 120
A2.1 Main Effects of the Predicted Lending Shocks ...... 12 7
A2.2 Relationship between Predicted Lending Shock and ln(loan originations) f or non-CRA Banks 128
A2.3 Effect of Predicted Lending Shock on Employment and Establishment Growth Rates for Small Establishments, NETS Data ...... 128
A2.4 OLS Models of the Relationship Between Economic Activity and Small Business Loan Origi- nations...... 129
10 9W_
List of Figures
1-1 U.S. Bank Branches, 1994-2014 ...... 33
1-2 Defining Exposed and Control Tracts - Wake County, NC ...... 34
1-3 Exposure to Consolidation and the Incidence of Branch Closings ...... 35
1-4 Exposure to Consolidation and Local Branch Levels ...... 35
1-5 Exposure to Consolidation and the Volume of New Lending ...... 36
1-6 The Effect of Subsequent Bank Entry on Local Credit Supply ...... 36
1-7 The Geographic Spillover of Bank Branch Closings ...... 37
1-8 Exposure to Consolidation and ZIP-Level Business Outcomes ...... 37
1-9 The Effect of Closings on Growth in Industries with a High Dependence o n External Finance 38
2-1 Geographic Distribution of the Predicted Lending Shock . 65
2-2 Regression-Adjusted Difference in ln(loan originations) between Top Quartile and Lower Quar- tile C ounties ...... 66
2-3 Effect of the Predicted Lending Shock on Loan Originations, by Year ...... 66
2-4 Effect of the Predicted Lending Shock on Small Business Employment Growth, by Year . . 67
3-1 Lending Sources per Village ...... 91
3-2 Geographic Ubiquity of Different Lenders ...... 9 1
3-3 Percentage of Households Holding At Least One Loan ...... 92
3-4 Sample Profile for surplush ...... 92
3-5 Household A's Solution ...... 93
3-6 Household B's Solution ...... 93
11 A1.1 Branch Closings in Buyer Only and Target Only Tracts versus Unexposed Tracts ...... 112
A1.2 Exposure to Consolidation and the Incidence of Branch Closings, ZIP-Level ...... 112
A1.3 Exposure to Consolidation and the Volume of New Small Business Lending, ZIP-Level ... . 113
A1.4 Salop Circle Model: Effect of a Branch Closing on Lending to Local Borrowers ...... 113
A1.5 Salop Circle Model: Price Effects Following a Branch Closing ...... 114
A1.6 Impact on Household Financial Stability ...... 114
A2.1 CRA-Disclosed Loan Originations to Firms with Less than $1 Million in Annual Revenue .. 125
A2.2 Effect of the Crisis on County-Level Small Business Loan Originations ...... 125
A2.3 ln(loan originations) Relative to 2007 ...... 126
12 Chapter 1
Do Bank Branches Still Matter? The Effect of Closings on Local Economic Outcomes*
1.1 Introduction
Banks are among the most heavily regulated industries in the economy, and a primary policy objective is to ensure local access to the banking system. The FDIC, for example, requires that banks provide 90-day notice ahead of any intention to close a branch; this is intended to facilitate public discussion of "the adverse effect the closing may have on the availability of banking services in the affected area."' This and other policies focus on closings that hinder the physical accessibility of the branch network, but much less emphasis is placed on the disruptive effects of closings that occur in crowded banking markets.
Figure 1-1 shows that after fifteen years of uninterrupted expansion, the U.S. branch network has been shrinking since 2010. This trend is widely expected to continue, and the wave of closings has prompted widespread concern regarding the impact on local communities. In large part, this discussion has emphasized the role of closings in reducing physical access to the banking system: commentators have detailed the
*1 am extremely grateful to Michael Greenstone, Robert Townsend, and Abhijit Banerjee for their advice and guidance throughout this project. I also thank Adrien Auclert, Jean-No8d Barrot, Nittai Bergman, Asaf Bernstein, David Colino, Manasi Deshpande, Ludovica Gazze, Daniel Green, Sally Hudson, Raj Iyer, Dirk Jenter, Scott Nelson, Brendan Price, Natalia Rigol, Matt Rognlie, Adam Sacarny, Antoinette Schoar, Ashish Shenoy, Daan Struyven, Hongkai Zhang, participants in the MIT Development seminar, the MIT Development, Finance, and Labor lunches, and the Federal Reserve Bank of Boston's brown bag series for their comments and suggestions. Special thanks to Equifax and the Boston Fed, especially Jos6 Fillat, Chris Foote, Kaili Mauricio, Ana Patricia Munioz, Scott Schuh, and Oz Shy, for allowing me to use the FRBNY Consumer Credit Panel data, and to Alex Oberg and the staff members at the MIT GIS Lab for their help in working with the GIS data. 'See https://www.fdic.gov/regulations/laws/rules/5000-38 3 0.html. Similarly, under the Community Reinvestment Act, reg- ulators may organize public forums between banks and community groups in low- and middle-income areas when there is concern regarding the effect of a closing on local accessibility to bank services (Barr (2005), Skillern (2002)).
13 emergence of "banking deserts" and chronicled the effect of closings that leave neighborhoods or towns without ready access to another branch.2 Data from the FDIC show that 20% of branch closings since 2010 have been cases in which the closed branch was the only one in its Census tract (the median tract is 2 square miles).3
Mirroring this emphasis, existing regulation vis-A-vis branch closings is geared almost exclusively toward helping communities where closings lead to a substantial decline in the number of local branches. The FDIC's 90-day rule is waived in cases of consolidation where the branches involved are "within the same neighborhood." Yet according to the FDIC data, at least 80% of closings occur in areas where there is no meaningful impact on physical access as measured by the number of remaining branches. Could closings still have an impact on lending in these cases?
This paper evaluates that question by estimating the local economic effects of bank branch closings in areas where the branch network is dense. I estimate the impact of closings on local credit supply as measured by the volume of small business and mortgage lending. The empirical challenge is that banks choose which branches to close, and those decisions are related to local economic conditions that are correlated with credit demand. Branches will close in areas where current or forecasted profitability is expected to be low, and a naive comparison between areas where branches close and areas where they do not would likely overestimate the impact of the closing itself.
As a solution to the endogeneity problem, I use exposure to post-merger consolidation as an instrument for branch closings. Many mergers are followed by a period of retrenchment during which branches are closed in areas where the two previously-separate networks overlap. I therefore define "exposure" to be a binary variable equal to 1 for neighborhoods that had branches from both buyer and target banks prior to the merger, and 0 for neighborhoods that had branches from only one or neither. To use only plausibly exogenous variation, I focus on mergers between large banks (i.e., banks with at least $10 billion in pre- merger assets) and use Census tract level data that allow me to exploit within-county variation in exposure to consolidation. Since the median tract is only 1.5 square miles - compared to 586 square miles for the median county - this level of geographic disaggregation allows me to compare economically similar areas with and without closings and to measure the effects of a closing at a very local level.
Figure 1-2 illustrates this identification strategy for a sample merger and a sample county in the data. The empirical framework compares the pre- and post-merger level of lending in "exposed" tracts relative to a set of control tracts that (i) are located in the same county and (ii) had branches belonging to at least two large banks who did not merge with one another. The spirit of this approach is to compare tracts that, a priori, were equally likely to have been exposed to a large bank merger. This instrument identifies the effect of closings that occur in substantially crowded markets (the average Exposed tract has 6 branches prior to the 2 See, for example, the March 31, 2013, Wall Street Journal article titled "After Years of Growth, Banks are Pruning Their Branches," and the November 13, 2013, story from NPR titled "'Banking Deserts' Spread Across Low-Income Neighborhoods." 3 This figure is obtained by geocoding branch locations and closings as reported in the FDIC Summary of Deposits and the FDIC Report of Changes.
14 merger) and are precisely those excluded from the FDIC's 90-day rule. As such, the results are informative for whether closings have disruptive effects even when the local banking market is very dense.
This paper yields three primary findings. First, closings are associated with a substantial and prolonged decline in credit supply to local small businesses. The number of new small business loans is 13% lower for several years after a closing. Notably, lending remains depressed despite the entry of new banks, which shows the decline is not driven by the competitive effects of the merger. In constrast, there is only a temporary decline in mortgage lending. Second, the decline in lending is concentrated within low-income and high-minority tracts, indicating that closings are most disruptive in disadvantaged neighborhoods. Third, I provide evidence that the impact of a closing is very localized: the magnitude of the effect decreases monotonically as distance from the closed branch increases, and ultimately dissipates 8 miles out.
These results suggest that, even in crowded markets, closings can have large effects on local credit supply when lending is information-intensive and lender-specific relationships are difficult to replace. These dynamics are relatively less important in the mortgage market, where rates of securitization are very high and the process of loan approval has become largely automated. Small business lending, on the other hand, is an information-intensive market. If personnel-specific soft information is destroyed when a branch is closed, borrowers can face a prolonged decline in credit supply until they are able to build a relationship with a new lender.4 Similarly, low-income and minority borrowers may be particularly dependent on soft information and lender-specific relationships.
The welfare implications of this decline hinge on the characteristics of the marginal borrower. If closings restrict credit access for positive NPV borrowers, then the decline in local credit supply is welfare-reducing. If, however, these are negative NPV borrowers, then the decline in lending may actually be welfare-enhancing. The data sources used in this paper do not include borrower and loan characteristics, such as default rates, that can distinguish empirically between these possibilities, and so the welfare implications are ultimately ambiguous. However, I provide a conceptual framework to illustrate that, even if borrowers with positive NPV projects lose access to credit, closings may still be efficient from the viewpoint of total welfare once we account for the banks' forgone costs.
This paper has several important policy implications. First, understanding the heterogeneous effects of branch closings is highly policy-relevant. Existing regulation is heavily focused on lending to low-income and minority borrowers, who generally face high barriers to credit access. I show that closings are more disruptive in disadvantaged neighborhoods even though the number of branches does not vary systematically between lower- and upper-income tracts in my sample. This suggests that the same factors that are believed to restrict credit supply in marginalized neighborhoods (i.e., a greater dependence on soft information and lender-specific relationships) may also increase their vulnerability to adverse shocks. These findings also
4 Drexler and Schoar (2012) provide evidence that soft information is difficult to transfer even between employees in the same institution. They show that shocking the relationship between an individual borrower and her loan manager can disrupt credit access.
15 show that physical access is not the only dimension along which closings can have substantial impacts on local communities, suggesting that the current focus of banking regulation vis-A-vis branch closings may be overly narrow. More broadly, this paper shows that there are segments of the U.S. banking system today where a borrower's access to credit is still defined by her local credit market.
This paper makes several contributions relative to the existing literature. A rich body of work has explored how the infrastructure of the local banking market matters for local outcomes in both developing and developed countries (Jayaratne and Strahan (1996), Black and Strahan (2002), Burgess and Pande (2005), Cetorelli and Strahan (2006), Kerr and Nanda (2009), Gilje (2012), Gilje, Loutskina and Strahan (2013), Townsend and Zhorin (2014)). To the best of my knowledge, this paper is the first to study the local effects of branch closings and, in particular, their effects in already-crowded markets. Papers that have studied the effects of bank consolidation on small business lending have found either negative or neutral effects (Strahan and Weston (1996), Strahan and Weston (1998), Berger et al. (1998), Peek and Rosengren (1998), Sapienza (2002)). These papers are motivated by the concern that small business lending will fall when large banks acquire smaller ones since large banks are less well-suited to relationship-intensive lending (Stein (2002), Berger et al. (2005)). This paper shows the destruction of branch-level soft information is an important factor even in mergers between large banks. Finally, while an existing literature has used state- or county- level data to estimate the effects of negative local credit supply shocks (Peek and Rosengren (2000), Ashcraft (2005), Greenstone, Mas and Nguyen (2015)), the sources of variation used in these papers cannot identify the effects of branch-level shocks. This paper provides a novel identification strategy paired with tract-level data that have not previously been used in this context to show that branch closings have large effects on their local communities.
The paper proceeds as follows. Section 1.2 describes the data. Section 1.3 discusses the identification strategy and empirical framework. Section 1.4 presents the empirical findings, and Section 1.5 interprets them as part of a broader framework that considers the welfare impact of branch closings. Section 1.6 concludes with policy implications.
1.2 Data
The primary unit of observation in this paper is the Census tract. These are defined by the U.S. Census Bureau to be small, relatively permanent statistical subdivisions of a county. Tracts are defined to optimally contain 4,000 inhabitants, and therefore vary in size across urban and rural areas. As discussed in greater detail in Section 1.3, I construct a sample of tracts based on exposure to large bank mergers. The median tract in this sample is 1.5 square miles, while the median county is 586 square miles (these numbers are comparable to those for the U.S. overall). Tract boundaries are slightly revised with each Census, and this paper uses boundaries as of the 2000 Census. For variables reported using 2010 boundaries, the Census
16 provides a set of relationship files that allows researchers to merge geographic entities over time.
To construct the exposure instrument, I use the FDIC Summary of Deposits, which provides an annual enumeration of all branches belonging to FDIC-insured institutions. These data link each branch to its parent bank, and provide a limited amount of branch-level information including deposits, street address, and, since 2008, the branch's latitude and longitude. I use data from 1999-2012, and map branch locations to their Census tract using GIS software. Some observations are dropped because their latitude and longitude data are missing and their recorded street address is either invalid or incomplete. Appendix Table A1.1 provides summary statistics for this geocoding procedure: the percentage of unmapped observations is 7.5% in 1999 and declines to 0.6% in 2012.
As the only bank-level information available in the Summary of Deposits is total assets, I also use balance sheet data on total lending from the FDIC Report of Conditions and Income. Data on merger activity and branch closings are from the FDIC Report of Changes.
To gauge the impact of closings on local lending, I use Community Reinvestment Act (CRA) and Home Mortgage Disclosure Act (HMDA) data published by the Federal Financial Institutions Examination Council (FFIEC). Under the CRA, all banks with assets greater than $1 billion are required to disclose annual tract- level data on the number and dollar volume of loans originated to businesses with gross annual revenues less than or equal to $1 million. While these data only capture small business loans originated by CRA-eligible banks, Greenstone et al. (2015) use call report data to estimate that these account for 86% of all loans under $1 million.
Under similar HMDA reporting criteria, financial institutions are also required to publish data on their local mortgage lending activity. 5 HMDA data are at the loan application level and include not only the Census tract associated with the application, but also its amount, whether it was approved/denied, its type (i.e., home purchase / home equity / refinancing), and applicant characteristics such as income. I drop mortgages subsidized by the Federal Housing Authority, the U.S. Department of Veterans Affairs, or other government programs, which constitute approximately 10% of the full HMDA sample, and aggregate the remaining data to create an annual measure of tract-level mortgage originations. Both tract-level small business loan and mortgage originations are winsorized at the 1% level.
It is important to note that both CRA and HMDA data are based on the location of the borrower, as opposed to the location of the bank. For a given tract, the data measure the total number of loans made to borrowers located in that tract, regardless of the location of the originating branch. This data structure allows me to estimate the impact of a branch closing on total credit supply to borrowers located in the same tract.
Finally, to provide evidence on the real economic effects of branch closings, I use the ZIP Business Patterns
5 According to the 2014 reporting criteria published by the FFIEC, institutions required to disclose under HMDA are banks, credit unions, and savings associations that have at least $43 million in assets, have a branch office in a metropolitan statistical area or metropolitan division, originated at least one home purchase loan or refinancing of a home purchase loan in the preceding calendar year, and are federally insured or regulated.
17 data published by the U.S. Census. These provide annual, ZIP-level measures of total establishments, employment, and annual payrolls. I also discuss results obtained using measures of household credit outcomes (such as bankruptcy and delinquency rates) constructed from the Federal Reserve Bank of New York / Equifax Consumer Credit Panel. Tract-level demographic characteristics such as population and median family income are from the 2000 Census. All other data are for the 1999-2012 period.
1.3 Identification and Empirical Framework
The structural relationship of interest is the effect of a branch closing on local lending:
yit = Ci + Yt + AXit + /cCloseit + fit, (1.1) where yit is total lending to borrowers located in tract i in year t, ai are tract fixed effects, 'yt are year fixed effects, Xit is a vector of tract characteristics, and Closeit is an indicator equal to 1 if a branch closes in tract i in year t. The OLS estimate for i3c is unbiased if Closest is orthogonal to cit: i.e., if the incidence of the closing is unrelated to local factors that would also affect the level of lending. In general, this assumption is unlikely to hold as shocks to credit demand will affect both the level of lending as well as the profitability of local bank branches.
To generate plausibly exogenous variation in the incidence of branch closings, I use exposure to post-merger consolidation as an instrument for closings. Bank mergers are often followed by a period of retrenchment during which the merged institution closes branches in areas where the two previously-separate networks overlap. This implies that areas with both Buyer and Target bank branches are at greater risk of a post- merger closing. I therefore supplement Equation 1.1 with the following first stage regression:
Closest = hi + Ot + pXit + /eExposeit + wit, (1.2) where Exposeit is an indicator equal to 1 if two banks with branches in tract i undergo a merger in year t.
Mergers themselves are motivated by several considerations, including expansion into new geographic or product markets, the synthesis of complementary business functions, an increase in market power, or cost savings from consolidation. In the context of this identification strategy, this may be problematic if the incidence of the merger is itself driven by factors specific to areas where Buyer and Target branches overlap.
To use only plausibly exogenous exposure to consolidation, I focus on mergers where both Buyer and Target banks held at least $10 billion in pre-merger assets, which roughly corresponds to the top 1% of the size distribution of U.S. banks. For mergers in this category, only 1.4% (3.5%) of Buyer (Target) banks' deposits are located in Exposed tracts prior to the merger. It is unlikely that any factors specific to these areas would be an important determinant in Buyer and Target banks' decision to merge.
18 The full set of criteria for inclusion in my merger sample are those that (i) occurred between 2001-2010, (ii) involved Buyer and Target banks that each held at least $10 billion in pre-merger assets, and (iii) where the merging institutions had overlapping retail branch networks in at least one Census tract. This yields a sample of 20 mergers. To further minimize the possibility that the decision to merge may be related to a decline in economic conditions specific to areas where the banks' branches are located, I also drop mergers that were either classified as failing (i.e., they required financial assistance from the FDIC) or that occurred during the financial crisis.6 The final sample comprises the 13 mergers listed in Table 1.1. The failing / crisis mergers are listed in Appendix Table A1.2.
Table 1.2 reports summary statistics for the Buyer and Target banks in the merger sample. By construction, these are very large institutions (the median Buyer holds $82 billion in assets, while the median Target holds $26 billion) with very extensive branch networks (the median Buyer controls 721 branches and operates in 8 states, while the median Target controls 292 branches and operates in 7 states). For comparison, the median bank in the U.S. holds $100 million in assets and controls only 3 branches.
For each of these 13 mergers, I define Exposed tracts to be those that had branches from both Buyer and Target banks in the year prior to the merger. Control tracts are those that did not have branches from both the Buyer and the Target, but did have branches from at least two large banks that did not merge with one another. The identification strategy is based on within-county comparisons between Exposed and Control tracts.
Figure 1-2 shows how Exposed and Control tracts are defined for a sample merger and a sample county in my data. The top panel shows a map of Wake County, NC, with Census tracts delineated and the geographic distribution of bank branches in the year prior to the 2004 merger between Wachovia and SouthTrust. Red squares are Wachovia branches, green triangles are SouthTrust branches, and any tract containing both is an Exposed tract.7 These branches tend to be clustered around the two urban centers of the county, which suggests that using all other tracts as a Control would amount to a comparison between urban and rural areas. To identify tracts that are more comparable to Exposed tracts, I map the locations of branches belonging to other large banks (i.e., other banks that also held at least $10 billion in assets). As my Control group, I take any tract that did not have both a Wachovia and a SouthTrust branch, but did have branches from at least two large banks who did not merge with one another. This group consists of three different kinds of tracts: Buyer Only tracts who only had a Wachovia branch, but not a SouthTrust; Target Only tracts who only had a SouthTrust branch, but not a Wachovia; and Unexposed tracts who had neither a Wachovia nor a SouthTrust, but did have branches belonging to two other large banks.
The spirit of this approach is to define a set of tracts that, a priori, had similar potential to be exposed to 6 The results are qualitatively similar when these mergers are included but, consistent with these concerns, the outcomes display pre-trends that are absent in the primary sample. 7 The figure shows branches are often located on, or very near, tract boundaries, even though the geocoding procedure maps each branch to a unique tract. This is because boundaries are often determined by major roads where branches are also likely to locate. This introduces some measurement error to the definition of the instrument, but should, if anything, reduce the magnitude of the estimate.
19 a large bank merger. This translates into the set of Exposed and Control tracts shown in the bottom panel of Figure 1-2. I use a difference-in-differences (DD) framework to compare lending in Exposed and Control tracts within the same county, before and after a merger.
Table 1.3 provides summary statistics for the tracts in my sample. The set of 13 mergers shown in Table 1.1 translates into a sample of 394 Exposed tracts and over 3,000 Control tracts. As the identification strategy is based on within-county comparisons, I present summary statistics by estimating regressions of the form:
fic = a + fExposeic + 0 -c+ Cic, (1.3) where fic is a pre-merger characteristic for tract i in county c, and Exposeic is a dummy equal to 1 if tract i is an Exposed tract. Conditional on purging county fixed effects, a is the Control group mean (shown in Column 2 of Table 1.3), and # is the difference in means between Exposed and Control (shown in Column 1).
Table 1.3 shows the average Exposed tract in the sample has roughly 6 branches prior to the merger, which indicates that the instrument identifies the effect of closings that occur in crowded markets. Exposed and Control tracts are similar on most dimensions, but Exposed tracts tend to be more populated, have a higher fraction of college-educated inhabitants, and have a higher number of pre-merger branches. In all specifications, I allow for differential trends based on these pre-merger characteristics.
While Exposed and Control tracts differ on levels, the validity of the DD framework hinges on the assumption of parallel trends. I therefore estimate a year-by-year version of the DD, and present event study plots that allow for visual examination of pre-trends in the data. The primary specification is:
YicMt = ai + rm + (Yt x O-c) + Xiot + 6, (D't x Exposeicm) + 6 jcmt, (1.4)
where yicmt is an outcome for tract i in county c for merger m in year t; a2 are tract fixed effects; rim are merger fixed effects; (7t x o-,) are county-by-year fixed effects; Xi is a vector of pre-merger tract characteristics whose effects are allowed to vary by year; DM' is a dummy equal to 1 if year t is r years after merger m is approved by federal regulators; and Exposeicm is a dummy equal to 1 if tract i is an Exposed tract for merger m. The pre-merger tract characteristics in Xi are population, population density, fraction minority, fraction college-educated, median family income, the number of branches as of the year preceding the merger, and average annual growth in the number of branches for the two years preceding the merger. T ranges from -8 to 10, and standard errors are clustered at the tract level. The coefficient of interest is 6,, which measures the difference, conditional on controls, in outcome y between Exposed and Control tracts T years after the merger.
20 1.3.1 External Validity
The internal validity of the DD framework hinges on the assumption of parallel trends, but assessing external validity is also informative in the context of this identification strategy. While the set of tracts exposed to post-merger consolidation may be exogenously determined, banks still choose which branches to close. This does not invalidate the instrument, which requires that exposure to consolidation is as good as randomly assigned. It does, however, affect the interpretation of the local average treatment effect (LATE) identified by the merger instrument.
In a general framework with heterogeneous treatment effects, the LATE identified by a particular instrument is the effect of treatment on compliers, where compliers are observations whose treatment status is changed by the instrument. In other words, compliers are neither "always-takers" (tracts where a branch would have closed regardless of whether or not there was any merger) nor "never-takers" (tracts where no branch is closed even when a merger occurs). Instead, compliers are tracts where a branch closes if and only if there is a merger. To interpret the LATE identified by the merger instrument, we need to know who the compliers are.
Table 1.4 shows the complier characteristics for my sample. 8 Relative to the median tract in the sample, compliers tend to be less densely populated, have a lower median income, and have a higher number of pre-merger branches, all of which suggests that banks tend to concentrate their closings in areas deemed to be "overbranched." This emphasizes that the merger instrument does not identify the effect of closings that move neighborhoods from 1 to 0 branches. It identifies the effect of taking an already-crowded market and removing one branch from it.
1.4 Results
1.4.1 Exposure to Consolidation and Branch Closings
This section presents evidence for the first stage relationship between exposure to consolidation and the incidence of branch closings. Figure 1-3 provides the template used for the event study results. It plots the 6, estimated from Equation 1.4, where the dependent variable is the number of branch closings in tract i in
year t. The bars show the 95% confidence intervals, and the lines at T = -4 and T = 6 denote the range over
8While it is not possible to identify the compliers in the sample, Angrist and Pischke (2009) describe a procedure for summarizing their characteristics. Briefly, the first step is to calculate the proportion of Always-Takers (irA) and Never-Takers (7rN) in the data. In the context of this paper, the former is calculated by estimating the fraction of Control tracts who experienced a closing after the merger, while the latter is calculated by estimating the fraction of Exposed tracts who did not experience a closing. From these two numbers, one can calculate the proportion of compliers 7rC = 1 - ir^ - rN. With this information, one can back out the average characteristics of compliers by first estimating the average characteristics over the set of Always-Takers and compliers (i.e., Exposed tracts that did experience a closing) and then the average characteristics over Always-Takers only (i.e., Control tracts that had closings).
21 which there is a balanced panel. 6, > 0 indicates a higher incidence of branch closings in Exposed tracts relative to Controls r years after a merger.
Up to several years prior to the merger, Exposed tracts are no more likely than Controls to experience a closing. However, the relative incidence increases in the year the merger is approved, spikes in the year after, and then falls back to zero. Column 1 of Table 1.5 presents the corresponding point estimates, and shows the sum of 60 and 61 is 0.284. As there is generally a maximum of one closing per tract, this can be roughly interpreted as a 28 percentage point increase in the relative probability of a closing in Exposed tracts in the 2 years following the merger.
Note that because the Control group includes Buyer Only and Target Only tracts, the results in Figure 1-3 are not driven by a tendency for merged banks to close branches across the board. Appendix Figure A1.1 confirms this directly by showing the merger has no effect on the incidence of branch closings in Buyer and Target Only tracts relative to Unexposed tracts.9
Figure 1-4 shows the higher incidence of closings in Exposed tracts translates into a decline in the total number of branches, and illustrates the importance of estimating the year-by-year coefficients. There is no evidence of pre-trends, and the plot reveals that the post-merger decline is only temporary. By r = 4, the number of branches in Exposed tracts is again level with Control tracts. The corresponding point estimates are shown in Column 2 of Table 1.5. The dependent variable is the total number of branches, but the results are similar when using the total number of banks.
The results in Figure 1-4 are consistent with Garmaise and Moskowitz (2006), who find the market structure effects of mergers last approximately 3 years before other banks enter. This pattern suggests that while it is in the merged bank's interest to consolidate on its fixed costs by closing an overlapping branch, profits are then high enough to accommodate a new entrant. The fact that we observe subsequent entry in these tracts will play an important role in interpreting the credit supply results presented in Section 1.4.2.
1.4.2 Closings and Local Credit Supply
I now address the question of interest: do closings in dense banking markets have an impact on local credit supply? The dependent variables are drawn from the FFIEC data, and measure the total number of new small business and mortgage loans made to borrowers located in tract i in year t, regardless of the location of the originating branch.
Figure 1-5 shows the reduced form relationship between exposure to consolidation and the volume of new lending. The left panel shows that, coincident with branch closings, there is a decline in new mortgages that lasts approximately 3 years, though the year-by-year coefficients are not significant. The right panel reveals
9 I look at both Buyer Only and Target Only tracts since the data indicate that post-merger closings are split fairly evenly between Buyer and Target bank branches. 60% of post-merger closings involve a Target branch, while 40% involve a Buyer branch.
22 a larger decline in the small business lending market. Relative to Controls, Exposed tracts experience a decline in the number of new small business loans that persists up to 6 years after the closing.
This comparison suggests closings have a more substantial effect in the small business lending market, but the contrast becomes especially striking when we compare the reduced form estimates in both markets with the first stage relationship between exposure to consolidation and the total number of branches. Figure 1-6 plots the reduced form estimates from Figure 1-5 overlaid with the first stage coefficients from Figure 1-4. The left panel shows the decline in mortgage lending is temporary and recovers before the number of branches. The right panel, however, shows closings have a much longer-term impact on credit supply to local small businesses. Small business lending declines when a branch closes, and remains depressed even after the entry of new banks.
To more easily interpret the magnitude of these effects, Table 1.6 provides estimates from less flexible versions of the DD. I estimate:
6 Yicmt = ai + 77m + (-Yt X 0) + Xi3t + POST (POSTmt x Exposeicm) + eicmt, (1.5) where POSTmt is a dummy equal to 1 if year t occurs after merger m is approved by federal regulators and all other variables are as previously defined. 6 POST measures the post-merger mean shift in the level of lending. Given the patterns observed in Figure 1-5, I also allow a post-merger linear trend in event year for the mortgage results by estimating:
3 6 Yicrt = ci+7M+(Yt x 9-c)+Xi/ t+ PosT (POSTt x Exposeicm)-+ 6, (POSTmt x Exposeicm x T)+eicmt, (1.6) where T is the event year.
The reduced form estimates in Column 1 of Table 1.6 show the decline in the number of new loans is mirrored by a decline in the dollar volume of new lending in both markets. While not statistically significant, the point estimates in Panel A indicate that mortgage lending declines temporarily following the closing. Column 3 shows the decline has dissipated by 6 years after the merger. In contrast, Panel B shows closings are associated with a statistically significant 13% annual decline in new small business loans. Over the six years following the closing, this amounts to a total of nearly $2 million in forgone loans. To provide a sense of scale, the average closing involves a branch that controls 16% of tract-level deposits.
The contrast between small business and mortgage lending suggests closings are more disruptive in markets where lending is information-intensive. A large literature in finance has studied the role of soft information and relationships, and Drexler and Schoar (2012) provide evidence that severing the relationship between an individual borrower and her loan manager can lead to disruptions in credit access. In cases of post- merger consolidation, the staff at the closed branch are often let go while the accounts are transferred to the neighboring branch of the merged bank. To the extent this process destroys personnel-specific soft
23 information that is difficult to transfer, borrowers may face a prolonged restriction in credit supply until they are able to establish new relationships.
These dynamics are less important in the mortgage market where rates of securitization are very high and the process of loan approval has become largely automated.10 The fact that lending in this market recovers even before the number of branches suggests the decline is driven by the short-term disruptive effects of the closing. Borrowers may delay their applications until any uncertainty over consolidation is resolved, or there may be administrative delays due to the process of transferring accounts. In contrast, small business lending is typically seen as the prototypical example of an information-intensive market where borrowers are heavily reliant on lender-specific relationships."1 The prolonged decline in small business lending displayed in Figure 1-6 - and, importantly, its persistence despite the entry of new banks - suggests closings disrupt lending relationships in that market that take time to rebuild.
1.4.2.1 Alternative Explanations
The Appendix outlines a model of spatial competition in local banking markets, and shows that lending may decline after a merger or closing if reducing the number of competitors from n to n - 1 places upward pressure on prices. Indeed, Garmaise and Moskowitz (2006) provide empirical evidence that merger-induced increases in local concentation lead to higher prices and less credit. Under this hypothesis, the results in
Figure 1-5 may reflect not that small business lending is more relationship-dependent than mortgage lending, but that it tends to be more locally concentrated and, therefore, more sensitive to changes in local market structure.1 2
While plausible in theory, the patterns in Figure 1-6 provide evidence that the direct effects of a change in tract-level concentration are empirically negligible. Mortgage lending recovers before the number of branches, which shows the initial decline cannot be attributed to the change in local market structure. Similarly, small business lending does not respond to the entry of new banks; the decline in lending persists even after the competitive environment has returned to its previous equilibrium. One reason the competitive effects may be limited in this context is that, as discussed in Section 1.3, this instrument identifies the effect of closings that occur in very crowded markets.
An alternative explanation for the results in Section 1.4.2 is that the decline in small business lending is 0 1 To wit, an October 2014 New York Times article reported that Ben Bernanke had recently been unable to refinance his mortgage because the program used to screen his application detected that he had had a recent change in employment. 11 Petersen and Rajan (1994) and Berger and Udell (1995b) both emphasize the importance of relationship lending for small businesses. Amel and Brevoort (2005) and Brevoort, Holmes and Wolken (2010) show small business lending markets tend to be very local, and Agarwal and Hauswald (2010) argue this is because geographic proximity facilitates the collection of soft information. Greenstone et al. (2015) provide evidence that small businesses who faced restrictions in credit supply during the Great Recession were unable to substitute toward other lenders. 12 1n addition to the price effects, several papers have shown that a change in the competitive environment can have a direct impact on the amount of relationship lending banks choose to engage in. However, the direction of the effect is ambiguous. Petersen & Rajan (1995) argue increased credit market competition will impose constraints on the strength of lending rela- tionships since banks are less able to extract rents from future surplus. Conversely, Boot and Thakor (2000) argue increased competition will lead banks to engage in more relationship lending since this will insulate them from pure price competition.
24 driven by a change in organizational focus induced by the merger. Peek and Rosengren (1998) show that Buyer banks tend to recast Targets in their own image, which leads to post-merger convergence toward the behavior of the Buyer. If Buyers engage in less small business lending than Targets, this may be one reason small business lending declines in Exposed tracts after a merger. A related possibility is that Target banks may engage in more risky lending than Buyers (hence, contributing to their eventual acquisition), which is eliminated after they are acquired.
There are several pieces of evidence that refute this hypothesis. First, Appendix Table A1.3 shows the lending intensity of each Buyer and Target bank in the sample, as measured by the ratio of the dollar volume of small business loans over total assets in the year before the merger. In most cases, Buyer and Target intensities are of similar magnitude. If not, Buyers are often more engaged in small business lending than Targets, which would lead any post-merger convergence to run in the direction opposite to the results.
Moreover, Appendix Table A1.4 shows there is no evidence of a decline in lending in Target Only tracts. Branches in these tracts would be affected by any organizational change resulting from the merger, but are not exposed to the greater risk of post-merger closings.
1.4.2.2 Varying the Size of the Local Banking Market
The standard practice in much of the finance literature is to define local banking markets at the level of the MSA or non-MSA county. Garmaise and Moskowitz (2006) argue that this convention has been driven by data availability, and that evidence suggests local markets are likely to be much smaller. As my identification strategy relies on within-county comparisons, this may be a concern if my results are driven by comparisons between tracts located very far apart. To address this, I re-estimate the reduced form results for small business lending using varying definitions for the size of sub-county local banking markets. For each Exposed tract, I define the market to be all Control tracts located within 10, 15, or 25 miles." Identification is then based on within-market comparisons between Exposed and Control tracts.
Appendix Table A1.5 shows the estimate for the post-merger decline in small business lending is robust to these variations. The estimate obtained when the market is defined using a 15-mile radius (the definition used by Garmaise and Moskowitz (2006)) is -2.414 compared to -2.504 when the market is defined at the county level. Even with a 10-mile radius, the estimate is still -2.051. This suggests the results are not affected in any meaningful way by treating counties as the local market.
1.4.3 Heterogeneity Across Borrowers
Section 1.4.2 provided evidence that the impact of a branch closing varies according to the information- intensity of different loan products. This section addresses whether there are heterogeneous effects across
13 Distances are measured based on tract centroids.
25 different borrowers. This is highly policy-relevant given that U.S. banking regulation is heavily geared toward lending to low-income and minority borrowers. These policies, as evidenced by the Community Reinvestment Act, the Equal Opportunity Credit Act, and the Home Mortgage Disclosure Act, are primarily focused on increasing the level of lending in disadvantaged neighborhoods. However, it is important to know whether the sensitivity of lending to adverse shocks may also be higher in these areas.
I split my sample into terciles based on tract-level median family income, and separately estimate Equations 1.5 and 1.6 for each one. The thresholds are chosen so as to ensure a near equal distribution of observations across each group: Low-Income tracts are those with median income below $40K, Middle-Income are those between $40-58K, and High-Income are those with median income greater than $58K.
Columns (1) through (3) of Table 1.7 present the corresponding IV estimates, and show the post-closing declines in lending are entirely concentrated amongst the lowest-income tracts in the sample. In fact, Panel B shows closings have no statistically significant effect on credit supply to small businesses in Middle- and High-Income tracts, but Low-Income tracts experience a nearly 40% decline in new small business loans. Columns (4) through (6) show the results of splitting the sample according to fraction minority. While the confidence intervals are larger, the point estimates suggest a similar story: the decline in lending is most severe in tracts with the highest fraction of minority households.
What might explain these heterogeneous effects? One possibility is that there may be fewer branches located in low-income and minority tracts. In this case, each closing will represent a more substantial decline in the availability of bank services. However, Appendix Table A1.6 shows the correlation between the number of branches and tract-level median income and fraction minority is extremely low (only 0.0079 and -0.0984, respectively) in this sample. Conditional on having branches from at least two large banks, banking markets in low-income neighborhoods are just as crowded as those of wealthier neighborhoods in this sample.
An alternative explanation is that closings are more disruptive in disadvantaged neighborhoods because these are precisely the borrowers for whom soft information and relationships are most important. Munoz and Butcher (2013) show that credit histories for low-income borrowers tend to be thinner and patchier, meaning there is less hard information available to evaluate a borrower's creditworthiness. Bond and Townsend (1996) provide evidence that borrowers in low-income and minority neighborhoods in Chicago rely more heavily on informal sources of credit, and posit this may be because informal lenders have cheaper access to relevant information about borrowers within the same community. These issues are not particular to the U.S. context and resonate throughout the literature on barriers to credit access in developing countries. 14 In this sense, the same factors that are believe to restrict credit supply in low-income and minority neighborhoods may also increase their vulnerability to adverse shocks.
14Fisman, Paravisini and Vig (2012), for example, use data from India to show that soft information (transferred, in their case, via cultural proximity between borowers and lenders) can be important in ensuring access to credit in settings where problems of asymmetric information would otherwise give rise to substantial credit rationing. Banerjee and Duflo (2010) provide a broader overview of the development literature on this topic.
26 1.4.4 Geographic Spillovers
How local are the effects of a branch closing? The results have shown there is a substantial decline in credit supply to small businesses located in the same tract, but surrounding areas are likely to be affected as well. The median tract in this sample is only 1.5 miles, and survey evidence shows small businesses search up to several miles away for a credit provider (Amel and Brevoort (2005), Brevoort et al. (2010)).
To measure these geographic spillovers, I categorize tracts according to their distance from a branch closing. For each Exposed tract, let Rt denote the set of tracts located between x - 1 and x miles away. R0 contains only the Exposed tract. R1 consists of all tracts whose centroids are located at most 1 mile away from the Exposed tract, but excludes the Exposed tract itself. R 2 consists of all tracts whose centroids are located at most 2 miles away, but excludes all tracts contained in R 1 and R0 . And so on and so forth.
I define Rx for all x E {0, 10}. For each x, I estimate Equation 1.5 where the dependent variable is the number of new small business loans, Rtx is the "exposed" group, and the Control group consists of all tracts
6 located in the same county but at least 10 miles away from the branch closing. POST measures the post- merger decline in lending observed in tracts who did not themselves experience a closing, but who were located x miles away from one.
6 Figure 1-7 plots the POST for each x E {0, 10}, and shows that the effects of a closing are very localized. The impact is most severe in the tract where the branch is located and, strikingly, the magnitude of the effect decreases nearly monotonically as distance from the closed branch increases. Ultimately, the impact on lending dissipates at about 8 miles.
These results are remarkably consistent, both qualitatively and quantitatively, with existing evidence on the local nature of small business lending markets. Amel and Brevoort (2005) and Brevoort et al. (2010) use survey evidence to show the median distance between small firms and their supplier of credit is around 3-5 miles. Figure 1-7 uses actual firm behavior and provides a measure that falls exactly within that range.
1.4.5 Real Economic Effects
Finally, there is a larger question of the extent to which the decline in local lending has real economic effects. Greenstone et al. (2015) show that county-level declines in small business lending during the Great Recession led to lower employment growth amongst small establishments, which suggests there may be similar dynamics in the context of branch closings.
The ideal dataset would provide data on real outcomes at the tract level. This is especially important given the evidence from Section 1.4.4 that the effects of a closing are very localized. Unfortunately, the most finely disaggregated, publicly available data on business outcomes are the ZIP Business Patterns data published by the U.S. Census (the median ZIP in the sample is 9 square miles). I use these data to provide suggestive
27 evidence for the effect of branch closings on the number of establishments and total employment.1 5 Future work will use confidential Census microdata to construct tract-level measures of establishment entry, exit, and employment.' 6
For each merger, I define Exposed ZIPs to be those that contain at least one Exposed tract, while Control ZIPs are those that contain only Control tracts.1 7 In the majority of cases, an Exposed ZIP contains only one Exposed tract. Appendix Table A1.7 provides summary statististics for the Exposed and Control ZIPs. Appendix Figures A1.2 and A1.3 show the tract-level results on branch closings and small business lending hold qualitatively at the ZIP level, albeit more noisily.
Figure 1-8 shows the reduced form relationship between exposure to consolidation and log establishments and log employment at the ZIP level.' 8 There is no notable change immediately following the merger, and the plots reveal substantial pre-trends that were absent in the tract-level results. Panel A of Table 1.8 estimates the less flexible version of the DD, allowing for both a mean shift and trend break in the post- merger period. The results are only marginally significant for log employment, and indicate that annual growth in employment is approximately 4 percent lower in Exposed ZIPs relative to Controls 6 years after the merger. This is a very large effect for a single branch closing and likely reflects other differences between Exposed and Control ZIPs.
While the effects of a single closing may be too diluted to be reflected in ZIP-level aggregates, we can also exploit within-ZIP comparisons across different industries. To the extent there are real economic effects of branch closings, these should be most pronounced in industries that are heavily reliant on bank credit. Rajan and Zingales (1998) provide a method for classifying industries according to their dependence on external finance. I use the classification provided in Gilje (2012) (and reported in Appendix Table A1.8), who shows that industries with a high dependence are more sensitive to local credit supply shocks. I estimate the following triple-difference specification:
yizcmt = az + 7im + (yt x oic) + X,3t + 61Highi + 62Treat, + 63POSTmt (1.7)
+64 (POSTmt x Treat,) + 65 (POSTmt x Highi) + 66 (Highi x Treatz)
+67 (POSTmt x Treatz x Highi) + Eizcmt, where yizcmt is log of the total number of establishments in industry i in ZIP z (this is the only variable broken down by industry in the Census data); Highi is a dummy equal to 1 if industry i is classified as having a high dependence on external finance; Treat, is a dummy equal to 1 if ZIP z is a Treatment ZIP; 5 1 Appendix Section 3.5 discusses the impact of closings on household credit outcomes such as delinquency rates and credit scores. 16 Kerr and Nanda (2009) use these data to examine the impact of branching deregulation in the U.S. on entrepreneurship and incumbent firm displacement. 1 7 This is not an exact match since tract boundaries, which are defined by the U.S. Census, do not correspond to ZIP boundaries, which are defined by the U.S. Postal Servce. In practice, if a tract is located in more than one ZIP code, I assign it to the ZIP in which the majority of its population lives. 18Results on annual payrolls are not shown, but are consistent with those for establishments and employment.
28 POSTmt is equal to 1 if year t occurs after merger m is approved by regulators; and all fixed effects are defined as before. I estimate the year-by-year version of this triple difference. If industries with a greater dependence on external finance are more severely affected by branch closings, we would expect 67 < 0 in the post-merger period.
Figure 1-9 shows that, while there are still pre-trends, there is a decline in the relative growth of industries with a high dependence on external finance that coincides with the incidence of branch closings. Panel B of Table 1.8 shows this translates into 3 percent lower annual growth in the number of establishments in high dependent industries relative to low dependent industries 6 years after the closing.
1.5 Welfare Implications
Section 1.4 shows that closings have large, negative effects on local credit supply in markets where lending is information-intensive. The first-order issue for determining the impact on welfare, however, is: who is the marginal borrower? If closings sever relationships that facilitate credit access for positive NPV borrowers, then the decline in lending is welfare-reducing. In general, this need not be the case. Hertzberg, Liberti and Paravisini (2010) show that when loan managers are responsible not only for maintaining a relationship with their borrowers, but also for monitoring their repayment prospects, they may suppress negative signals about the firm's ability to repay since it will reflect negatively on their own reputation. If managers siphon funds to borrowers with negative NPV projects, the observed decline in lending may be welfare-enhancing. The last several years have also revealed ample evidence of lax lending standards and their role in fueling the credit boom that preceded the 2008 financial crisis. The evidence from Section 1.3.1 that merger-induced closings are concentrated in "overbranched" areas may suggest that the marginal borrower affected by these closings is especially likely to have benefited from overlending. The data sources used in this paper do not include borrower and loan characteristics, such as default rates, that can distinguish empirically between these possibilities, and so the welfare implications are ultimately ambiguous.
Even if borrowers with positive NPV projects lose access to credit, however, closings may still be efficient from the viewpoint of total welfare once we account for the banks' forgone costs. To illustrate this, consider the following simple model. There is a banking market where banks must pay a cost e in order to enter. 19 Upon entry, banks pay a per-period fixed cost of operation F for each branch, which covers the cost of renting a storefront and hiring staff and does not vary with the number of customers served. For the moment, I assume that each bank operates only one branch. Banks are Nash price setters and engage in only one activity, which is lending. There is zero marginal cost to lending, but banks must charge an interest rate that is high enough to cover their fixed costs. Consumers borrow from banks to invest in projects with a
return w. 19 As an example, one could consider the Salop circle model discussed in the Appendix. However, as the spatial component of the Salop circle is not central in this context, I opt for a more general setting.
29 Let xi (pi, r) denote the total loan demand for bank i when i charges an interest rate pi and all other banks charge a rate r. An equilibrium consists of an interest rate r and a number of banks n such that (i) each bank i earns maximum profit: pi = arg max [qxi (q,r) - F] q and (ii) there is no entry: 7r (n) 0 >7r (n +1), where 7r (n) are per-bank profits, net of the cost of entry, when n banks are in the market. As all banks are identical, consider the symmetric equilibrium (r*, n*) where each bank has equal market share, L*. Conditional on this equilibrium loan portfolio, let p denote the minimum interest rate a bank would have to charge in order to cover the fixed cost F. Since there are barriers to entry, banks can earn positive profits in equilibrium and the interest rate r* will lie between banks' minimum interest rate and borrowers' maximum willingness to pay: i.e., p < r* < w.
To consider the welfare implications of consolidation, suppose there is an unanticipated merger where Bank A acquires Bank B, who both operate a branch in this market. Post-merger, A has the option of closing B's branch and absorbing its loan portfolio to consolidate on the fixed cost F. However, some percent p of B's loan portfolio is lost during consolidation, and those borrowers who are dropped are shut out of the credit market (for example, due to the destruction of personnel-specific soft information). To simplify notation, let A (r) = pL*r denote the revenue the bank earns from lending to this p percent of borrowers at an interest rate r.
Ceteris paribus, A will close B if the revenue lost from dropping these loans is less than the savings accrued from consolidation on the fixed cost: i.e., if A (r*) < F. The borrowers who would be dropped, however, are willing to pay up to w to prevent the branch from closing. This means that as long as A (W) > F, the bank can renegotiate a higher interest rate i < w where (i) those consumers who would lose access as a result of consolidation are willing to borrow at that higher rate and (ii) A (i) > F, so A prefers to keep B open. 20 This implies there is no room for policies geared at preventing branch consolidation. Banks' profit-maximizing behavior will dictate that a branch closes only when the cost savings from doing so exceed consumers' maximum willingness to pay to keep the branch open: i.e., when A (w) < F.
This argument relies, however, on the implicit assumption that banks can bargain with borrowers over the full social surplus from their lending. To the extent this is violated, consolidation may occur even when that is the socially inefficient outcome. As a simple example, suppose there are positive spillovers to bank lending. These may take the form of agglomeration economies where the profits of non bank-dependent firms
2 0 The model implicitly assumes loan contracts are characterized by one-sided commitment. Banks cannot break their ex ante contract with customers, except by closing a branch. However, customers can ask to renegotiate the contract if they are want to pay a higher rate. In this case, only those borrowers whose credit access is threatened by the closure are willing to bargain. All others have nothing to gain from doing so since, post-consolidation, they can still borrow at r*.
30 are positively correlated with those of bank-dependent firms.2 1 In this scenario, the social value of capital, wSP, will exceed the private value, w, and there may be cases where A (w) < F < A (wSP) and the merged bank consolidates its branches even though this results in a loss of social surplus.
This discussion has focused on welfare in the market where the branch closing occurs. At a more aggregate level, banks may reallocate the resources from a closed branch across their remaining network, and so the within-market decline in lending may be perfectly offset by an increase in lending elsewhere. From the viewpoint of social efficiency, however, what matters is the net change in welfare and not the net change in lending. In that vein, it is important to emphasize that Section 1.4.3 shows the decline in lending is concentrated in low-income areas where the marginal utility of consumption is high. Unless banks reallo- cate their resources to equally poor neighborhoods, and conditional on the earlier discussion regarding the characteristics of the marginal borrower, this would imply a net reduction in welfare.
1.6 Conclusion
This paper uses a novel identification strategy paired with Census tract level data to estimate the local economic effects of bank branch closings. I show that, even in crowded banking markets, closings have large effects on local credit supply when lending is information-intensive and lender-specific relationships are difficult to replace. The effects are concentrated in low-income and minority neighborhoods, which are areas that have historically faced high barriers to credit access and are highly relevant in the context of U.S. banking regulation. I also characterize the geographic spillovers of branch closings and show their effects are very localized.
There are two important policy implications. First, existing regulation in the U.S. is heavily focused on increasing the availability of banking services in low-income and minority neighborhoods, which tend to be less heavily-branched than wealthier areas. However, I show that closings are more disruptive in these disadvantaged areas even though the number of branches does not vary systematically between lower- and upper-income tracts in my sample. This suggests that the same factors that are believed to restrict credit supply to marginalized borrowers may also make it harder for them to adjust to credit market disruptions. This implies that financial shocks, even those that affect only the largest financial institutions, may ultimately have disproportionate effects on already-disadvantaged groups.
Second, these findings also suggest the current approach to regulating branch closings and evaluating the impact of bank mergers may be overly narrow. The focus on the availability of other branches fails to recognize that if closings destroy lender-specific information, borrowers will be unable to obtain credit at equal terms even in dense banking markets. More broadly, these conclusions show that in the U.S. banking
2 1 Pashigian and Gould (1998), Gould, Pashigian and Prendergast (2005), and Benmelech, Bergman, Milanez and Mukharlyamov (2014) provide evidence of local agglomeration economies.
31 system today, there are some markets and some segments of the population for whom local credit markets still play an important role in determining local credit access.
32 Figures
Figure 1-1: U.S. Bank Branches, 1994-2014
Number of Branches (1,000)
Cf) 0)
000~ 1995 2000 2005 2010 2015 Year Source: FDIC
Figure displays the total number of bank branches reported in the FDIC Summary of Deposits from 1994-2014. These are annual data that enumerate all branches belonging to FDIC-insured institutions.
33 Figure 1-2: Defining Exposed and Control Tracts - Wake County, NC
em
Ar 7
79
4
Legend * Wachovia (Buyer) * SouthTrust (Target) 0 Other Large Banks
0 5 10 20 Miles
I
Legend Exposed Control
0I I 5I I 10I I I I 20i Miles
Source: FDIC. Figure uses the example of Wake County, NC, to show how Exposed and Control tracts are defined. The top panel shows the Census tract boundaries in Wake County along with the geographic distribution of bank branches in the year prior to the 2004 merger between Wachovia and SouthTrust. Red squares are Wachovia (Buyer) branches, green triangles are SouthTrust (Target) branches, and blue circles are branches belonging to other large banks (i.e., other banks with at least $10 billion in assets). Tracts with both a Wachovia and a SouthTrust branch are Exposed tracts. Tracts that did not have both a Wachovia and a SouthTrust branch, but did have branches belonging to at least two large banks are the Control group. This corresponds to the set of Exposed and Control tracts shown in the bottom panel.
34 Figure 1-3: Exposure to Consolidation and the Incidence of Branch Closings
Number of branch closings
N
AIf
0
-10 -5 0 5 10 Years since merger Source: FDIC, author's own calculations. Plot shows estimated deltatau coefficients.
Figure shows the first stage relationship between exposure to consolidation and the incidence of branch closings. The figure plots the 6, estimated from the event study specification, along with the 95% confidence intervals. The dependent variable is the number of branch closings in tract i in year t. r = 0 is the year the merger was approved by federal regulators, and and all coefficients are normalized relative to r = -1. The vertical lines at r = -4 and r = 6 denote the range over which the panel is balanced. Robust standard errors are clustered at the tract level.
Figure 1-4: Exposure to Consolidation and Local Branch Levels
Total branches
n 4~- ii~4
U I -10 -5 0 5 10 Years since merger Source: FDIC, author's own calculations. Plot shows estimated deltatau coefficients.
Figure plots the 6, estimated from the event study specification, along with the 95% confidence intervals. The dependent variable is the total number of branches in tract i in year t. -r = 0 is the year the merger was approved by federal regulators, and all coefficients are normalized relative to r = -1. The vertical lines at r = -4 and r = 6 denote the range over which the panel is balanced. Robust standard errors are clustered at the tract level.
35 Figure 1-5: Exposure to Consolidation and the Volume of New Lending
New mortgages New small business loans CD-
4 C. to -
(0.
-10 -5 0 5 10 -10 -5 0 5 Years since merger Years since merger Source: FFIEC, author's own calculations. Plot shows estmated delta-tau coefficients. Source: FF1EC, author's own calculations. Plot shows estimated delta tau coefficients.
Figure displays reduced form estimates of the relationship between exposure to consolidation and lending to borrowers located in that tract. The figure plots the 6, estimated from the event study specification, along with the 95% confidence intervals. The dependent variables are, respectively, the number of new mortgages and new small business loans made to borrowers located in tract i in year t. -r = 0 is the year the merger was approved by federal regulators, and all coefficients are normalized relative to -r = -1. The vertical lines at r = -4 and r = 6 denote the range over which the panel is AA balanced. Robust standard errors are clustered at the tract level.
Figure 1-6: The Effect of Subsequent Bank Entry on Local Credit Supply
New mortgages New small business loans _cq
.50CD to
cj- 0 C~ A It AI1* . -0 A A 0 :2
CO 0
-10 -5 0 5 10 -10 -5 0 10 Years since merger Years since merger SLos anches Loans A Branches
Source: FDIC, FFIEC, author's own calculations. Figure plots the 6, estimated from the event study specification. For red triangles, the dependent variable is the total number of branches in tract i in year t. For blue circles, the dependent variables are, respectively, the number of new mortgages (left panel) and new small business loans (right panel) made to borrowers located in tract i in year t. r = 0 is the year the merger was approved by federal regulators, and all coefficients are normalized relative to -r = -1. The vertical lines at r = -4 and 7 = 6 denote the range over which the panel is balanced. Robust standard errors are clustered at the tract level.
36 MM
Figure 1-7: The Geographic Spillover of Bank Branch Closings
New small business loans LO
0- ijiji I)
0D
0 2 4 6 8 10 Distance from Exposed Tract Source: FFIEC, author's own calculations. Plot shows estimated deltaPOST coefficients.
Figure displays reduced form estimates of the post-merger decline in new small business loans in tracts sorted according to their distance from an Exposed tract. The Control group is tracts located at least 10 miles away from an Exposed tract. Section 1.4.4 provides more details. Estimates are from the version of the difference-in-differences that allows for a single post-merger mean shift in the level of lending. The bars show the 95% confidence intervals. Robust standard errors are clustered at the tract level.
Figure 1-8: Exposure to Consolidation and ZIP-Level Business Outcomes
Log establishments Log employment LO
CM q_
C) cli LO
-10 -5 0 5 10 -10 -5 0 5 10 Years since merger Years since merger Source: U.S. Census, author's own calculations. Plot shows estimated delta tau coefficients. Source: U.S. Census, author's own calculations. Plot shows estimated della_tau coefficients.
Figure shows the reduced form relationship between exposure to consolidation and ZIP-level log establishments and log employment. The figures plot the 6, estimated from the event study specification, along with the 95% confidence intervals. r = 0 is the year the merger was approved by federal regulators, and all coefficients are normalized relative to -r = -1. The vertical lines at r = -4 and -r = 6 denote the range over which the panel is balanced. Robust standard errors are clustered at the ZIP level.
37 Figure 1-9: The Effect of Closings on Growth in Industries with a High Dependence on External Finance
Log establishments
cmJ {.+V~ -. ~ (0
-10 -5 0 10 Years since merger Source: U.S. Census, author's own calculations. Plot shows estimated deltatau coefficients.
Figure shows the results of estimating the triple difference specification described in Section 1.4.5. The figure plots the coefficients on the triple interactions between indicators for the post-merger period, being located in an Exposed ZIP, and belonging to one of the industries classified as having a high dependence on external finance in Appendix Table A1.8. The dependent variable is log establishments in industry i in ZIP z in year t. r = 0 is the year the merger was approved by federal regulators, and all coefficients are normalized relative to r = -1. The vertical lines at r = -4 and r = 6 denote the range over which the panel is balanced. Robust standard errors are clustered at the ZIP level.
38 Tables
Table 1.1: Merger Sample
Buyer Target Year Approved
Manufacturer and Traders Trust Company Allfirst Bank 2003 Bank of America Fleet National Bank 2004 National City Bank The Provident Bank 2004 Regions Bank Union Planters Bank 2004 JPMorgan Chase Bank Bank One 2004 North Fork Bank Greenpoint Bank 2004 SunTrust Bank National Bank of Commerce 2004 Wachovia Bank SouthTrust Bank 2004 Sovereign Bank Independence Community Bank 2006 Regions Bank AmSouth Bank 2006 Bank of America United States Trust Company 2007 The Huntington National Bank Sky Bank 2007 Bank of America LaSalle Bank 2007
Source: FDIC. Table shows the 13 mergers included in the primary merger sample and the year they were approved by federal regulators. The criteria for inclusion in this sample are all mergers that (i) occurred between 2001-2010, (ii) involved Buyer and Target banks with at least $10 billion each in pre-merger assets, and (iii) where both banks had overlapping retail branch networks in at least one Census tract. Of the remaining 20 mergers, I drop those that were either classified as failing (i.e., they required assistance from the FDIC) or occurred during the 2008 financial crisis. These excluded mergers are listed in Appendix Table A1.2.
Table 1.2: Merger Summary Statistics
Median Min Max
Panel A: Buyer
Total assets (billion $) 82 26 1,250 No. of branches 721 259 5,781 States of operation 8 1 31 Counties of operation 183 18 694
Panel B: Target
Total assets (billion $) 26 10 246 No. of branches 292 29 1,563 States of operation 7 1 13 Counties of operation 54 7 204
Source: FDIC. Table displavs summary statistics for the 13 Buyer and 13 Target bank s in the merger sample. All variables are as of the year in which the intention to merge was announced.
39 Table 1.3: Summary Statistics for Exposed and Control Tracts
(1) (2) Variable Exposed Minus Control Control Mean
Population 307.5* 5,408 (180.8)
Population Density -2.931 5,826 (316.5)
Fraction Minority -0.005 0.238 (0.012)
Fraction College-Educated 0.0242** 0.333 (0.010)
Percent MSA Median Income 3.712 118.3 (2.667)
Median Income (000s) -0.135 51.3 (1.135)
Fraction Mortgage 0.005 0.715 (0.008)
Pre-Merger Branches 2.069*** 3.845 (0.216)
Pre-Merger Branch Growth -0.007 0.058 (0.009)
Joint F-test 17.53 p-value 0.00
Number Exposed 394 Number Control 3,129
Source: FDIC, U.S. Census, author's own calculations. Table provides summary statistics for Exposed and Control tracts. Values are obtained from a regression of each tract-level characteristic on an indicator for being an Exposed tract and county fixed effects. Population density is per square mile. Percent MSA median income is the ratio of tract median income to MSA median income. Demographic variables are as of the 2000 Census; "pre-merger" variables are as of the year preceding each merger. Pre-merger branch growth is the average annual growth in the number of branches for the two years preceding the merger. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p 40 Table 1.4: Complier Characteristics (1) (2) Variable Compliers Ratio: Compliers to Sample Population 0.58 1.16 Population Density 0.18 0.36 Fraction Minority 0.58 1.16 Fraction College-Educated 0.48 0.96 Percent MSA median income 0.44 0.88 Median Income (000s) 0.29 0.58 Fraction Mortgage 0.45 0.90 Pre-Merger Branches 0.89 1.78 Pre-Merger Branch Growth 0.39 1.10 Source: FDIC, FFIEC, U.S. Census, author's own calculations. Table uses the methodology outlined in Angrist & Pischke (2009) to show how Complier tracts compare to the median tract in the sample. For more details, see Footnote 8 in Chapter 1. Column 1 shows the fraction of Compliers who lie above the median tract in the sample. For example, the first row shows 58% of Compliers are more populated than the median tract in the sample. Column 2 calculates the ratio of Compliers to Sample by dividing each entry in the second column by 0.50, since 50% of tracts in the sample will, by definition, lie above the median tract. 41 Table 1.5: Exposure to Consolidation and Branch Closings (1) (2) Number of Closings Total Branches -0.018 0.031 (0.018) (0.056) 60 0.060** -0.028 (0.025) (0.067) 61 0.224*** -0.318*** (0.034) (0.086) 62 0.021 -0.267** (0.028) (0.111) 63 0.041 -0.293*** (0.031) (0.100) 6>3 -0.016 -0.003 (0.013) (0.135) 2Y Cum. Effect 0.284*** (0.042) Control Mean 0.129 4.048 Obs. 49,630 49,630 Source: FDIC, author's own calculations. Table shows estimates of the first stage relationship between exposure to consolidation and the incidence of branch closings. All coefficients are normalized relative to r < -3, where r = 0 is the year in which the merger was approved by federal regulators. The 2Y Cumulative Effect is the sum of 6o and 5J. The control mean is calculated for T - 1. Robust standard errors are clustered at the tract level and are in parentheses. *** p<0.01, ** p<0.05, * p 42 Table 1.6: The Effect of Closings on Local Credit Supply (1) (2) (3) (4) Coefficient Reduced Form IV 6Y Cum. Effect Control Mean Panel A: Mortgages 6 # Loans POST -7.234 -21.49 12.71 157.26 (4.807) (13.80) (14.86) 1.743* 5.701* (1.049) (3.401) Obs. 47,931 47,931 6 $ Volume (000s) POST -1,308 -3,823 1,614 30,690 (968.5) (2,758) (2,815) 6, 288.4 906.1 (205.8) (652.1) Obs. 47,975 47,975 Panel B: Small Business Loans 6 # Loans POST -2.504*** -9.373*** -56.24*** 67.87 (0.903) (3.291) (19.75) Obs. 46,631 46,631 $ Volume (000s) cPOST -83.91* -318.6* -1,912* 2,505 (49.61) (179.8) (1,078) Obs. 46,601 46,601 Source: FFIEC, author's own calculations. Table presents difference-in-differences estimates where the dependent variable is either the number or dollar volume of new loans in tract i in year t. Based on the patterns observed in Figure 1-6, I allow for a post-merger mean shift in the level of lending in both mortgage and small business lending markets, and include a linear trend in event year for the former. Column 1 provides the reduced form estimates, Column 2 the IV estimates, Column 3 the cumulative effect over the 6 years following the merger (6PosT + 6 x 6, for mortgages; 6 x 6POST for small business lending), Column 4 is the per-year control group mean averaged over the post-merger period. Robust standard errors are clustered at the tract level and are in parentheses. *** p<0.01, ** p<0.05, * p 43 Table 1.7: Heterogeneity by Tract-Level Income and Fraction Minority (1) (2) (3) (4) (5) (6) Coefficient Low Middle High Low Middle High Median Income Fraction Minority Panel A: Mortgages 6POST -55.42** -39.74* 1.665 -13.14 -9.531 -78.67** (24.21) (20.43) (32.81) (15.17) (25.77) (33.07) 11.23** 5.977 0.145 5.060 -4.034 14.66* (5.276) (4.495) (8.816) (3.813) (4.997) (8.446) Control Mean 86.28 154.74 227.09 183.72 180.58 108.18 Obs. 16,383 16,165 15,383 16,179 15,836 15,916 Panel B: Small Business Loans 6POST -19.53** -3.264 -3.773 -4.397 -8.790 -11.84 (9.478) (4.012) (7.270) (4.416) (6.203) (8.156) Control Mean 53.09 66.10 84.60 68.16 77.33 58.44 Obs. 15,956 15,749 14,926 15,826 15,340 15,465 Source: FFIEC, U.S. Census, author's own calculations. Table presents IV estimates of the effect of closings on local credit supply in tracts split according to median family income and fraction minority households. The dependent variable in Panel A is the number of new mortgages in tract i in year t; in Panel B, it is the number of new small business loans. Estimates are based on the same specifications used for Table 1.6. Control Mean is the per-year control group mean averaged over the post-merger period. Tercile thresholds are chosen to ensure a nearly equal distribution of observations across each group. Low-Income tracts have median family income in 2000 less than $40K, Middle- have $40-$58K, and High- have greater than $58K. Low-Minority tracts have less than 0.09, Middle- have 0.09-0.24, and High- have greater than 0.24. Robust standard errors are clustered at the tract level and are in parentheses. *** p<0.01, ** p<0.05, * p<0.l 44 Table 1.8: Exposure to Consolidation and ZIP-Level Outcomes (1) (2) Coefficient Log Establishments Log Employment Panel A: Difference-in-Differences 6POST 0.0019 0.0052 (0.0040) (0.0084) 6r -0.0027 -0.0073* (0.0022) (0.0040) 6Y Cum. Effect -0.0142 -0.0386* (0.0108) (0.0203) Obs. 25,295 25,295 Panel B: Triple Difference: Post x Exposed x High 6 POST -0.0038 (0.0093) 67- -0.0050* (0.0027) 6Y Cum. Effect -0.0336** (0.0160) Obs. 51,072 Source: U.S. Census, author's own calculations. Table presents reduced form estimates of the effect of exposure to consolidation on ZIP-level outcomes. Panel A shows the estimates from the diff-in-diff framework, while Panel B shows the result of the triple difference described in Section 1.4.5. The dependent variables are log establishments and log employment in ZIP z in year t. The total number of establishments is the only variable available at the industry-level and, therefore, the only outcome used for the triple difference. Robust standard errors are clustered at the tract level and are in parentheses. *** p<0.01, ** p<0.05, * p<0.1 45 Chapter 2 Do Credit Market Shocks Affect the Real Economy? Quasi-Experimental Evidence from the Great Recession and 'Normal' Economic Times* 2.1 Introduction The 2008 financial crisis caused an extraordinarily sharp decline in employment and, despite extensive fiscal and monetary policy interventions, the subsequent recovery was slow compared to the recovery from typical recessions. The range of explanations for this deep decline and slow pace of recovery include reduced aggregate demand (Mian and Sufi (2014)), uncertainty (Baker, Bloom and Davis (2013); Bloom et al. (2012)), and structural factors (Charles, Hurst and Notowidigdo (2012)). The list of explanations certainly also includes the tightening of bank lending standards and, at a high level, this theory is supported by the disproportionate employment losses incurred by small firms that are more reliant on bank lending than other firms (Charnes and Krueger (2011); CBO (2012); Fort et al. (2013)). Based on this observation, some policymakers (e.g., Bernanke (2010); Krueger (2010)) have suggested that fractured credit markets played a major role in overall employment declines. Indeed, restoring access to credit was a key feature of the policy *This chapter is co-authored with Michael Greenstone and Alexandre Mas. We are grateful to Laurien Gilbert, Felipe Goncalves, Ernest Liu, and Steven Mello for excellent research assistance. We thank Daron Acemoglu, Pat Kline, Lawrence Summers, Ivan Werning, and seminar participants at the NBER Summer Institute, Columbia, Brookings, Boston Federal Reserve, and Bank of Mexico for helpful comments. We also thank Abigail Cooke, Javier Miranda, and Lars Vilhuber for enabling our use of the Census LBD microdata. The first version of this paper was released in November 2012. 46 response following the financial crisis. 1 The academic literature (e.g., Brunner and Meltzer (1963); Bernanke (1983)) has emphasized that banks can play a central role in the functioning of the economy because small and medium-sized businesses do not have ready substitutes for bank credit. For this reason, banks have been labeled "special". And, it is thought that their health can be an important determinant of macroeconomic fluctuations (Bernanke and Gertler (1995); Peek and Rosengren (2000); Ashcraft (2005)). This paper gauges the degree to which the economic consequences of shocks to small business credit con- tributed to the employment losses of the 2007-2009 recession, in addition to providing new evidence on the credit channel's empirical importance. We also examine whether the relationship between local lend- ing supply shocks and economic activity differs between the recent crisis and the less turbulent 1997-2007 period.2 Our identification strategy leverages substantial heterogeneity in the extent to which different national banks cut their small business lending during the financial crisis. It then isolates the portion of these cuts that can be attributed to supply factors. Specifically, we predict the change in county-level small business lending over the 2007-2009 period using interactions of banks' pre-crisis county market shares and their national change in lending. Between 2007 and 2009, for example, Citigroup reduced small business lending by 84%, while U.S. Bancorp's small business lending declined by just 3%. In our procedure, we purge each bank's national change in lending of its exposure to local markets in order to isolate supply, rather than demand, shocks to lending. There is sufficient variation in banks' market shares across counties in the same state that our results are based on within-state comparisons. The essence of our approach is to ask whether, within a given state, counties with more Citigroup branches than U.S. Bancorp branches before the crisis experienced sharper declines in their economies over the 2007-2010 period. There are three primary findings. First, counties with banks that cut lending had declines in small business loan originations over the 2008-2009 period. For example, a one standard deviation reduction in predicted lending in 2009 is associated with a 17% reduction in total county-level small business loan originations from 2009 through 2010. In the short term, at least, it appears that the costs to switching lenders are meaningful for small firms. Second, this same predicted negative shock in lending depresses 2009-2010 employment growth rates for small standalone firms (single-unit establishments with fewer than 20 employees) by 0.4 percentage points. Third, while there is a significant relationship between predicted lending shocks and bank 'Speaking in July 2010 at the Federal Reserve Meeting Series, "Addressing the Financing Needs of Small Businesses," Chairman Ben Bernanke stated that "making credit accessible to sound small businesses is crucial to our economic recovery and so should be front and center among our current policy challenges," and that "the formation and growth of small businesses depends critically on access to credit, unfortunately, those businesses report that credit conditions remain very difficult." 2 This paper is an updated version of Greenstone and Mas (2012). The substantive difference between versions is that we primarily use restricted-use LBD data, whereas Greenstone and Mas (2012) primarily relied on NETS data, which is compiled by Walls and Associates using Dun and Bradstreet's proprietary Market Identifier files. Smaller changes include incorporating a symmetric growth measure, weighting the sample by each county's employment count in 2006 instead of the number of establishments in 2006, and extending the sample back to 1997. 47 loans to small businesses during the 1997-2007 period, these predicted shocks are not associated with changes in economic activity. This finding suggests that, at least with this identification strategy, the credit channel is not empirically important in normal times. To gauge the degree to which the reduction in credit supply contributed to the decline in employment over the 2008-2010 period, we conduct a simple bounding exercise by assuming that the entire 22% and 33% reductions in small business lending in 2008 and 2009, respectively, were due to banks' credit supply decisions. 3 Under this polar assumption, the point estimates imply that these shocks led to a maximum of a 0.3 percent decline in employment in establishments with fewer than 20 employees nationally; this is roughly 3 percent of the overall decline in small business employment over this period. Taking the upper end of the 95 percent confidence interval of the estimates, and assuming that all credit reductions were supply-driven, gives a maximum contribution of 12% of the overall change. As this is the combination of two extreme assumptions, the true aggregate effect is very likely to be considerably smaller. We conclude that the credit channel contributed to the significant changes in the real economy during the Great Recession, but its effects were modest. Of course, we cannot reject that the employment effects would have been larger in the absence of the extraordinary interventions undertaken by the Federal Reserve and the U.S. Government to aid banks. Our paper is most closely related to Chodorow-Reich (2014). Chodorow-Reich examines disruptions in the syndicated loan market following the collapse of Lehman Brothers in 2008, and finds that borrowers of weaker banks faced restrictions in credit supply, which translated into greater cuts in employment at these firms. He estimates the withdrawal of credit accounts for between one-third and one-half of the employment decline at small and medium firms in his sample over this period. This paper adds to the Chodorow-Reich paper and much of the rest of the literature in several ways. First, our study is nationally representative, which allows us to consider the aggregate implications of our estimates without external validity concerns. Bentolila et al. (2015) use national credit registry data to estimate the impact of credit supply shocks on employment in Spain, but papers studying the U.S. have focused on narrower samples. The Chodorow-Reich paper, for example, studies a particular market of (generally) larger firms, and other work has been focused on particular regions and episodes (e.g., Peek and Rosengren (2000); Ashcraft (2005)).4 Second, while other papers have also focused on small firms, which are more likely to be affected by bank supply decisions (e.g., Duygan-Bump et al. (2014)), our paper measures the impacts on overall county-level employment so our estimates incorporate establishment entry, exit, and expansion/shrinkage, as well as any multiplier-style effects or indirect effects via competitor responses. Third, we utilize a research design that allows us to control for confounding demand factors that may have affected employment growth. Fourth, our paper also contributes to the literature on the causes of the Great Recession and the subsequent slow recovery. 3 These figures refer to small business loan originations from the Community Reinvestment Act disclosure data. 4 The Chodorow-Reich sample is comprised of 2040 firms that have an average employment size in 2008 of 2985 employees (median 620). 48 The remainder of the paper is organized as follows. Section 2.2 provides some brief background on the financial crisis and the role of small businesses in the U.S. economy. Section 2.3 describes the data sources. Section 2.4 explains the research design and how it is implemented. Section 2.5 outlines the econometric models and presents the results. Section 2.6 interprets the findings, and Section 2.7 concludes. 2.2 Background The mechanisms behind the unraveling of the financial system in 2008 are complex, and have been analyzed in depth by Brunnermeier (2009) and Shleifer & Vishny (2011) among others. Most relevant to our context is that the liquidity crisis translated into less available credit across the economy, and the decline in commercial bank lending was especially severe for small business loans. According to data from the Federal Reserve Survey of Senior Loan Officers, the net percentage of loan officers reporting tightening standards for medium and large firms was 64 percent in the first quarter of 2009 as compared to zero percent in the first quarter of 2007. Data from banks reporting under the Community Reinvestment Act show loan originations to small businesses fell by 52% between 2007 and 2010.5 A similar pattern is seen in the survey of members of the National Federation of Independent Business: loan availability began to decline in the beginning of 2007, did not reach its nadir until 2009, and has been on a slow recovery since then (Dunkelberg and Wade (2012)). A number of papers explore the underlying mechanisms for this decline in lending and conclude that it was, in large part, "supply-driven." Ivashina and Sharfstein (2010), for example, document that new loans to large borrowers fell by 79% between the second quarter of 2007 and the fourth quarter of 2008. They argue that an important mechanism behind this decline was banks' reduced access to short-term debt following the failure of Lehman, coupled with a drawdown of credit lines by their borrowers.6 This severe contraction in small business lending may have led to significant real economic effects given both the importance of small businesses in the U.S. economy (in 2007, firms with less than 100 employees represented approximately 36% of employment and 20% of net job creation in the U.S. 78 ) and their depen- dence on local bank credit. Using data from the Survey of Small Business Finances through 2003, Brevoort, Holmes and Wolken (2010) estimate the median distance between firms and their suppliers of credit to be 3 5 Appendix Figure A2.1 plots the log of constant dollar loan originations to small businesses-defined here as businesses with gross revenues of less than $1 million-from banks reporting under the Community Reinvestment Act. It is apparent that the 2008 financial crisis led to an enormous decline in originations to small businesses. Appendix Figure A2.2 shows a kernel density plot of the change in log (nominal) small business loan originations between 2007 and 2009 across counties, weighted by the number of establishments in each county in 2006. It reveals substantial geographic dispersion in the decline in small business lending, though the pervasive nature of the recession is also evident in the fact that almost all establishments were in counties that experienced some decline. 6 See also Huang and Stephens (2014), Berrospide and Edge (2010), Goetz and Gozzi (2010), and Almeida et al. (2012). 7 Calculated using Census Business Dynamic Statistics. 8 We note, however, that there is considerable debate regarding the importance of small firms for net job creation. For example, using different datasets on firm employment dynamics Neumark, Wall and Zhang (2009) and Haltiwanger, Jarmin and Miranda (2013) both find evidence supporting an inverse relationship between net growth rates and firm size, but the latter study notes that it is really new businesses, rather than small businesses, that disproportionately contribute to net job creation. 49 miles, and find only 14.5 percent of small firms borrowed from an institution that was more than 30 miles from their headquarters. 9 A number of empirical studies have investigated the benefits of long-term lending relationships as a way to overcome information asymmetries in the lending market (e.g., Cole (1998), Berger and Udell (1995a), Hoshi et al. (1990), Petersen and Rajan (1994)). Berger et al. (2002) argue that firms that borrow from large banks tend to be more credit rationed, suggesting that firms that are cut off from credit from larger banks (as in our study) may not be able to obtain credit elsewhere. Nguyen (2014) provides direct evidence of this by showing that branch closings that follow mergers between large banks lead to prolonged declines in local small business lending, indicating that borrowers who lose access to credit have difficulty obtaining credit from other (bank) lenders. In the macroeconomics literature, credit market frictions have been suggested as a channel for the transmission of monetary policy, specifically through the effect of interest rates on the external finance premium, which arises through imperfections in credit markets (Bernanke and Gertler (1995)). 2.3 Data Sources Our analysis is conducted with what we believe to be the most comprehensive data ever assembled to investigate the role of bank lending in the real economy. The predicted lending shock is constructed using Community Reinvestment Act (CRA) disclosure data from the Federal Financial Institutions Examination Council (FFIEC). The CRA requires banks above a certain asset threshold to report small business lending each year and by Census tract. The asset threshold was $1.033 billion in 2007 and is adjusted with CPI.1 0 We estimate that, in 2007, CRA eligible banks accounted for approximately 86% of all loans under $1 million.12 The FFIEC provides data by bank, county, and year. Two definitions of small business lending are available: the total dollar amount of small business loan originations, defined as loans under $1 million (~30% of total originations in 2007), and the dollar amount of small business loan originations to businesses with $1 million 9 Amel and Brevoort (2005) use survey data from the National Federation of Independent Business Research Foundation to show the median distance over which small firms search for credit is only 4.3 miles. Using data from a single large commercial bank, Agarwal and Hauswald (2010) find a similar median distance between the lending branch and the firm (2.6 miles), and argue this is because geographic proximity facilitates the acquisition of "soft" information. Using data from the Community Reinvestment Act, Laderman (2008) finds that only about 10 percent of small business lending is from banks with no branch in the local market. loBefore 2005, the asset threshold was $250 million. 1'We use FDIC Call Report data from 2007 to compute the fraction of all loan balances held by banks below the asset threshold. This is an inexact estimate since loan balances in the FDIC Call Reports are a stock measure, while CRA originations are a flow. 2 1 FDIC Call Reports are not designed to study regional lending because the balance sheet data are only available nationally at the bank-level. For small community banks, however, it may not be a bad approximation to assign the location of the bank' headquarters as the market in which the bank lends (something that would clearly not work for, say, Bank of America). As discussed in Section 2.5, we use these data to better understand the implications of excluding the smaller banks that do not meet CRA reporting thresholds from our analysis. We find no evidence that small banks change lending balances in response to lending shocks of larger banks, and we conclude that our analysis is not greatly affected by the exclusion of these smaller banks. 50 or less in annual gross revenue (~13% of total originations in 2007). As our focus is on small firms, we use the second measure throughout the paper. These data are available from 1997 through 2010. To calculate changes in a bank's lending over time without including changes due to acquisitions, we employ the standard correction (e.g., Bernanke and Lown (1991)), which is to identify acquisitions over every pair of years and treat the acquired and acquiring bank as a single entity over that span. Following this procedure, we roll banks up to the holding company level. 13 This leaves us with 654 bank holding companies that are in the data for at least one year over the 1997-2010 period. While these are a relatively small fraction of all banks, they are the largest banks nationally and thus account for a large share of all lending. To study establishment-level dynamics, we use confidential microdata from the near universe of establish- ments in the U.S. Census Longitudinal Business Database (LBD). A key advantage of using these microdata is that we can compute growth rates over a given period based on establishments' sizes at the beginning of that period.1 4 Specifically, for a given size category i (e.g., establishments with fewer than 20 employees), we define employment growth between t - 1 and t in a given county as: Employment growth rateit [jobs created by new establishmentsit - jobs lost f rom closing (2.1) establishmentsit + employment in continuing establishmentsit - employment in continuingestablishmentsi,t- 1] [0.5 * employmenti,ti + 0.5 * employmentit]. Subscript i denotes establishments that are in size category i at the end of period t - 1 so that we are only measuring the change in employment for establishments that were in the relevant size class in the base period, as well as new establishments. Note that the above growth measure is symmetric, ranging between -2 and 2, as we use the average of t - 1 and t employment in the denominator, and is a second-order approximation to ln differences. Similarly, we compute the establishment growth rate as: Establishmentgrowth rateit [new establishmentsit- closing establishmentsit|| (2.2) [0.5 * establishmentsi.t_1 + 0.5 * establishmentsit]. We use the LBD microdata to compute these measures. In addition, we also use a special extract of the NETS database, which is compiled by Walls and Associates using Dun and Bradstreet's Market Identifier files.1 5 13 We use the FDIC institution directory to identify acquisitions and the FDIC Call reports to link banks to their holding companies. 1 4 An example may help to clarify this approach to calculating the growth rate. Consider calculating the growth rates of establishments with 20 or fewer employees, and with 21 to 150 employees. Suppose that an establishment had 100 employees in 2007, shrank to 10 employees in 2008, and then increased to 15 employees in 2009. This establishment would contribute to the 2007-8 growth rate for the 21 to 150 employee category and to the 2008-9 growth rate for the 20 or fewer category. 15 See Walls (2007) for an in-depth description of these data. 51 ,Pgrqrrr$ m %,wift-mw- -Mmm_ From these microdata, we construct employment and establishment growth rates for all small standalone firms (single-unit establishments with fewer than 20 employees) in each county and year, as well as for establishments that are part of multi-state firms (defined as operating in at least three states). Estimates using the NETS database are primarily used to assess robustness. County-level outcomes are constructed from the County Business Patterns (CBP) and the Quarterly Census of Employment and Wages (QCEW). The CBP are derived from the Census Business Registrar, while the QCEW are derived by the Bureau of Labor Statistics from state unemployment insurance records. Since the CBP and QCEW have very limited information on firm size, we use them exclusively for county-level analyses. 16 Finally, our main estimating equations also include county-level controls derived from Census data, the QCEW, and county debt-to-income ratios from the Federal Reserve Bank of New York.17 2.4 Research Design 2.4.1 Isolating Supply Shocks in Lending Our research design is based on the observation that some banks cut small business lending more than others following the crisis, and that bank market shares vary substantially across local areas. Table 2.1 shows the percent change in the nominal dollar amount of small business lending between 2007 and 2009 according to FFIEC CRA disclosures. While small business lending declined by 48% nationally over this period, the table reveals considerable differences across individual banks. Our identification strategy exploits heterogeneity in counties' exposure to these banks (as measured by their pre-shock market shares) under the testable assumption that firms can only incompletely substitute for a reduction in the supply of credit from their bank. Accordingly, a supply shock to a subset of banks in a given region will affect aggregate lending in that area. We test this assumption empirically, but the numerous papers cited above provide evidence of such frictions. To implement our strategy, we develop a modified version of the shift-share approach that has been used to identify local labor demand shocks, as in Bartik (1991).18 A standard application of the approach to our setting would involve constructing an instrument for bank lending in county i and year t as the sum across all banks in i of the interaction between bank market shares in t - 1 and changes in those banks' national lending between t - 1 and t. The identifying assumption in this standard approach is that shocks to banks' national lending reflect supply restrictions, rather than demand conditions, in the areas where the banks 6 1 The County Business Patterns has information on the number of establishments by size-category of firm, but this breakdown is inadequate for our purposes. The QCEW does not break down the data by firm size. 17County-level debt-to-income ratios are posted on Amir Sufi's website at http://faculty.chicagobooth.edu/amir.sufi/data.html. 18 See also Blanchard and Katz (1992), Card (2001), Autor and Duggan (2003), and Notowidigdo (2013) for other applications of this approach. 52 operate. This assumption is unlikely to be valid in this setting because firms in depressed areas will both reduce employment and demand less credit. The problem is that loan volume is an equilibrium outcome of supply and demand factors, and the standard approach is susceptible to confounding supply shocks with demand ones. We use a modified approach to address this identification problem. The presence of branches of multiple bank holding companies in each county provides an opportunity to purge the common county, or demand, effects from banks' national changes in lending. Specifically, we estimate the following equation that attempts to divide the contribution of demand and supply shocks to changes in bank lending: A ln (Qig) = di + si + e%. (2.3) Khwaja and Mian (2008) use a similar methodology to purge firm-specific credit demand shocks for matched bank-firm lending data from Pakistan. 19 The outcome variable in Equation 2.3 is the log change in small business lending by bank j in county i between two years. We weight the sample by each bank's base period lending in county i so that an observation's influence is proportional to its lending in that year. The bank fixed effects, sj, are re-centered so that their (bank asset size weighted) mean is zero. The county fixed effects, di, capture the variation in banks' change in lending that is due to the condition of the local economy, which we interpret as measuring the effects of local demand for credit. The parameters of interest are those associated with the vector of bank fixed effects, sj. They are estimates of banks' supply response that are purged of their differential exposure to regional variation in demand for small business loans. We estimate the sj for every pair of consecutive years beginning in 1997. Finally, we use these estimated bank-specific supply shocks (the sj's) to construct a county-level measure of the predicted lending supply shock. For each county and year, we take the weighted average of the estimated bank fixed effects from Equation 2.3, weighting by the bank's base period market share in that county20 : Ai = mijs. (2.4) sj is the estimated bank fixed effect from Equation 2.3, and mi is bank j's market share in county i (as measured by CRA small business loans) in the year prior to the estimated shock. We standardize the county- level predicted shock, pi, using its mean and standard deviation, and weighting by county-level lending in the base year. As with the sj, we compute the predicted lending shock across every pair of consecutive years beginning in 1997. The advantage of this approach is that the identifying assumption now only requires that banks with above- 19 See also Amiti and Weinstein (2013), who apply the Khwaja-Mian methodology to Japanese data to show that supply-side on firm investment. financial20 shocks can have large impacts To calculate year t's value of the instrument, the estimated bank fixed effects are obtained from fitting Equation 2.3 for changes in bank lending between t and t - 1. The relevant bank market share is taken from year t - 1. 53 or below-average supply-shifters are not systematically sorted into counties with worse-than-average shocks to outcomes. This assumption could be invalid if, for example, managerial skill in choosing branch locations is correlated with skill in choosing investments for the bank's portfolios. 21 In the following section, we provide empirical evidence that suggests banks are not systematically sorting in this fashion. Nevertheless, to confront the possibility of a violation of this assumption, we also exploit establishments' differential dependence on bank credit. As a specification check, we separately examine (non-franchise) establishments that are part of larger multi-unit and multi-state firms. These firms plausibly have greater access to alternative sources of financing than the local banks in a given county. This group serves as a useful check for whether our specification is adequately controlling for confounding factors that affect all establishments located in the same area. There may, of course, be indirect effects of local lending shocks on establishments who are not reliant on local banks. For example, the shock may enable them to take market share from establishments who are affected, or there may be a multiplier from the shock that negatively affects all firms in the area. Such indirect effects would complicate the interpretation of these intra-county comparisons. 2.4.2 Summary of the Predicted Supply Shock Figure 2-1 is a map of the United States with counties' shading reflecting the quartile of their predicted lending shock. Here, we use a version of the predicted lending shock that is computed for the entire 2007-2009 period. The regional correlation is evident, though not always in obvious ways. Florida and Massachusetts, for example, appear to have experienced worse-than-average shocks, while Georgia and Tennessee fared better. The figure also shows substantial within=state variation in thc value of the predicted shock, thereby allowing for within-state comparisons. Table 2.2 reports summary statistics of county characteristics based on whether the county is above- or below- median in terms its predicted lending shock. Columns (1) and (2) are the raw means with no adjustments, and column (4) is the within-state difference after purging out state fixed effects. Columns (3) and (5) report the p-values from tests on equality of means. Columns (1) and (2) show that counties with worse predicted lending shocks (i.e., below-median) have different characteristics compared to counties with better predicted shocks. This is not surprising given the spatial patterns seen in Figure 2-1. These columns compare areas that are far-removed from one another, and, since different regions of the country were likely differentially affected by the economic downturn for a variety of reasons (such as exposure to the construction or manufacturing sectors) that are not directly related to the supply of bank credit, a richer model that controls for regional effects is desirable. Column (4) shows that, when looking within states, counties with above- and below-median predicted shocks look much more 2 1 We provide a more precise description of the underlying assumptions in the Appendix. 54 similar. The only characteristics for which there remains a significant difference are population, population density, and the debt-to-income ratio. Consequently, our main analysis will emphasize specifications that include state fixed effects and that control for county population and debt-to-income ratio, as well as a larger set of county characteristics interacted with year dummies. We also investigate whether banks with larger cuts in lending systematically sort into particular types of counties. The presence of such sorting might indicate that banks are able to observe something about a county's future prospects that is both unrelated to the predicted supply shocks and unobservable to the econometrician. Evidence of such sorting would undermine the validity of our research design. Table 2.3 assesses the degree to which unhealthy banks non-randomly sorted into certain counties. In Column (1) we regress the fixed effect of the bank with the largest market share in a county against the fixed effect of the bank with the second largest market share in the same county. We do not find a significant correlation between the two. In Column (2) we take a more systematic approach by regressing bank j's fixed effect against the average fixed effect of other banks in markets where j operates, weighted by y's lending in each county. This specification also shows no significant relationship between the lending change of a bank and the lending changes of other banks in the same market. These spatial patterns are consistent with the presence of unhealthy banks in particular counties being due to "the luck of the draw", rather than to a systematic sorting of banks into certain counties as a function of their lending policy over the 2007-2009 period. 2.5 Econometric Models and Results 2.5.1 The Relationship Between the Predicted Lending Shock and Actual Loan Originations This section provides evidence that our predicted lending shock measure is predictive of realized county-level loan originations. We begin with a graphical analysis where we divide counties into a top quartile, middle 50%, and bottom quartile according to the value of their 2007-2009 predicted lending shock. The bottom quartile consists of those counties who experienced the largest negative supply shock. We then estimate the following model: In (lit) = 6 st + /tXit + Tt,<25Pi,<25 + Tt,25-75Pi,25-75 + fit (2.5) where lit denotes small business loan originations in county i and year t, Pi,<25 is an indicator for whether the county is below the 25th percentile according to the value of its 2007-2009 predicted lending shock, and Pi,25-75 is an indicator for whether the county lies in the middle 50%. The effects of these shocks are all allowed to vary by year, including pre-shock years, in order to investigate trends. The model includes a full set of state-by-year fixed effects, 6St, and 2006 county characteristics whose effects are allowed to vary 55 by year. 22 The state-by-year fixed effects mean that comparisons between the groups of counties are made within-state for each year. Finally, we weight the sample by each county's 2006 employment count.23 It is worth noting that the use of loan originations as the dependent variable is similar in spirit to using changes in total loans outstanding as an outcome variable. In this respect, since the outcome is a flow variable, the subsequent models are interpretable as first-differences. Thus, these are relatively rich models where the covariates predict changes, rather than levels, of the outcome variable. The coefficients of interest are the Tt,k, which capture the annual within-state difference in loan originations between the counties with top quartile values of the predicted lending shock and counties in the bottom and middle two quartiles, respectively. In Figure 2-2, the line with triangle data points plots the coefficients associated with the bottom quartile and year interactions (i.e., Tt,<25), while the line with square data points plots the coefficients from the middle quartiles and year interactions (i.e., Tt,25-75). The figure confirms there is a strong first-stage relationship even after these regression adjustments. 24 Al- though there are differences in the level of loan originations between the three groups, the regression adjust- ment removes most of the difference in pre-existing trends, especially during the 2000-2007 period. In the subsequent analysis, we primarily rely on the continuous version of the predicted lending shock. In these models, we focus on the 2008 and 2009 shocks separately, and estimate versions of the following model: In (lit) = St + OtXit + N8Pi,2008 + Y9Pi,2009 - 08,8 (V2008 X Pi,2008) (2.6) +68,9 (v 2 0 0 9 x Pi,2008) + 08,10 (V 2 0 10 X Pi,2008) +09,9 (V2009 X Pi,2009) + 09,10 (V2010 X Pi,2009) - Eit, where pi,, is the predicted lending shock in county i in year r, and vt are year dummies. The predicted shock main effects control for differences in county-level annual loan originations as a function of the 2008 and 2009 predicted supply shocks. 25 For ease of interpretation, the pi,,'s are standardized to have a mean of zero and standard deviation of one. We report standard errors clustered at the county level to account for serial correlation. 26 The parameters of interest are the O's. They are the coefficients on the interactions of the 2008 predicted lending shock with year indicators for 2008, 2009, and 2010, and on the interactions of the 2009 predicted lending shock with year indicators for 2009 and 2010. The 0's measure the impact of the lending shocks on 2 2 The controls are log per capita income, construction share, manufacturing share, log population, log population density, and debt-to-income ratio. 2 3 Unless otherwise specified, all models are weighted by the county's 2006 employment count. 2 4 Appendix Figure A2.3 shows the corresponding unadjusted estimates. It plots the Ot,k from estimating In (lit) = Pi,<25 + Pi,>75+ t,<25Pi,<25+ t,25-75Pi,25-75+0t,>75Pi,>75+Eit. These represent the annual means of small business loan originations for each group, relative to the 2007 value, which is constrained to be equal across 2 5 all groups. We have also estimated models that include county fixed effects, which is another way to control for differences in annual loan originations across counties. This alternative approach produced almost identical results. Due to the strong similarity of the results, we emphasize the more parsimonious specification going forward. 2 6 We have also experimented with clustering by state, but this did not have a notable effect on the standard errors. 56 loan originations in the year of the shock and all subsequent years, relative to the rate of loan originations in the years before the shock and in other counties. Thus, this is a difference-in-differences style estimator. We emphasize linear combinations of the estimated coefficients that are both easier to interpret and useful for summarizing the magnitudes. Specifically, we report the cumulative effect of the 2008 shock over the 2008-2010 period, and (the larger) 2009 shock over the 2009-2010 period. We define #8 and #9 to be the cumulative effect of a county having a one standard deviation increase in its 2008 and 2009 predicted lending shocks: 8 -- 08,8 - 08,9 + 08,10 #9 -- 09,9 + 09,10. For example, #9 is the cumulative effect over the 2009-2010 period of a county that is +1 standard deviation in the 2009 distribution of predicted lending shocks on log loan originations. The results in Table 2.4 confirm a robust and statistically significant relationship between the predicted lending shock and loan originations.2 7 Column (1) presents estimates from the specification that controls for state-by-year fixed effects. Column (2) adds the interaction of 2006 values of county covariates and year dummies. Column (3) adds to those covariates the interaction of year and the county's debt-to-income ratio in 2006. We add debt-to-income in a separate specification since it is not available for all counties. The point estimates in Column (2) imply that a county with a one standard deviation decline in predicted lending in 2008 experiences a large and persistent decline in loan originations of approximately 7.4% in 2008, 7.7% in 2009, and 8.9% in 2010.28 The estimate for #9 suggests that a county with a one standard deviation decline in the predicted lending shock in 2009 is predicted to have a 17% reduction in loan originations over 2009-2010 relative to pre-crisis levels, as compared to the mean county. Overall, these estimates provide evidence that there are important frictions in the small business lending market. When firms lose access to credit from their bank, it appears that there are meaningful costs that prevent them from immediately switching to other banks, thus leading to a decline in aggregate lending in that area.29 30 27 Appendix Table A2.1 presents the corresponding 'y estimates. 28 More precisely, these are log points. 29 1n Appendix Table A2.2, we use FDIC Call Report data to test whether non-CRA banks are offsetting the effects of lower lending from the larger banks. For small banks, defined as those that are not subject to CRA disclosure, we assign banks to the county where they are headquartered. We then estimate whether loan balances of small banks are affected by the predicted lending supply shocks of larger banks in that county. We find virtually no evidence that small banks change lending balances in response to the lending shocks of larger banks. Thus, the omission of small banks from the analysis is not likely to be a major problem for our analysis. 30 They may still borrow from non-bank sources, which is one reason the effect of a bank lending shock may not have real effects. We discuss this in Section 2.5.3. 57 2.5.2 The Relationship Between the Predicted Lending Shocks and Economic Activity During the Great Recession Having established a strong relationship between predicted and actual loan originations, we turn to examining the effects of these predicted shocks on measures of economic activity using the same specifications. Before describing these results, we note that the dependent variables are all measured in growth rates or log differences. Thus, the controls in the statistical models can be interpreted as controls for growth rates. This is not a change in focus from the previous subsection since loan originations are an approximation to the preferred, but unobserved, outcome of changes in the outstanding value of loans to small businesses. 2.5.2.1 Small Standalone Firms Table 2.5 provides estimates for the effects of the supply shocks on the growth rates of small standalone firms, which are defined to be single-unit establishments with fewer than 20 employees. We estimate Equation 2.6 where the dependent variable is either the employment or establishment growth rate for small standalones. The estimates for #8 and #9 (recall these represent the total effect of the predicted lending shocks occurring in 2008 and 2009, respectively) show the 2008 effect (#8) is close to zero and statistically insignificant, while a one standard deviation in the 2009 shock corresponds to a 0.43 percent cumulative reduction in the employment growth rate over 2009-2010 without county covariates (Column 1), and a 0.64 percent reduction with covariates (Column 2). Adding the debt-to-income control in Column (3) does not change the estimated coefficient relative to the estimate in Column (2), suggesting that the predicted lending shock is not picking up the effects of deleveraging as emphasized in Mian and Sufi (2014). As a basis for comparison, this set of firms experienced a 10% decline in employment between the end of 2007 and the end of 2010.31 In Columns (4)-(6) we report the results from estimating the same models, but using the small standalone growth rate as the dependent variable, following the equation (2) definition of the growth rate and versions of the equation (6) specification. These estimates range from marginally significant to insignificant across the specifications, though the standard errors are quite large. In the NETS data reported in Appendix Table A2.3 we find smaller but more precise (and statistically significant) effects on business births and deaths. Therefore, while there is some suggestive evidence that credit affects the formation and destruction of small businesses, the imprecise estimate in the LBD dataset prevents us from making a firm conclusion. 2.5.2.2 Small Establishments in Multi-Unit Firms As a specification check, Table 2.6 examines a set of establishments that should not be as sensitive to local lending shocks: namely, small (non-franchise) establishments that are part of larger multi-unit firms. These 3 1 This figure is based on data from the Census Business Dynamics Statistics. 58 establishments are less likely to be affected by the lending conditions in a particular county since multi-unit firms tend to have broader geographic coverage. We find, across all specifications, that the estimated effect of lending shocks for employment growth rates are insignificant. Moreover, the estimates on the interactions of the predicted lending shocks and year dummies are jointly insignificant. A caveat is that we were not able to verify in the LBD data whether the multi-unit firm was geographically diversified or concentrated in a single county. In the latter case we might still pick up some effect of the credit shock. As an additional check, we therefore present estimates from the NETS database where we limit the sample to small establishments of multi-unit firms that operate in at least three states. This sample of firms should have very limited, if any, exposure to changes in the supply of lending from banks in a particular county. These estimates range from slightly negative and insignificant to negative and borderline significant. We conclude from this analysis that the credit shock variables are not picking up differential business cycle effects across regions. 2.5.2.3 County-Level Economic Outcomes Table 2.7 explores the relationship between the predicted small business lending shock and county-level employment and establishment growth. These estimates provide an opportunity to gauge the full county- level effect of credit supply shocks beyond the category of small firms, including any general equilibrium effects. We use employment growth from the CBP and the QCEW. These datasets are designed to measure the same thing, but they are not perfectly correlated, and the literature does not offer guidance on which is better suited for analyses of county employment. To reduce measurement error in these analyses, we therefore use the average of the growth rates from these datasets for each county and year. In the specification with state fixed effects and baseline controls (Column (2)), the cumulative effect of one standard deviation in the 2009 shock is a statistically significant 0.34 percentage points. 33 It is not surprising that this coefficient is smaller than the small establishment sample as we expect that a credit shock will have a larger impact on smaller firms. Note, also, that the interactions between the predicted lending shocks and year dummies are only marginally significant in the joint test (p-value = 0.18). As in the case for the small establishment sample, the 2008 shock has an employment effect that is close to zero. We do not find a significant impact of the lending shocks on total county establishments from the 2008 shock, but there is a significant effect from the 2009 shock. The magnitude is smaller than the one estimated using small establishments in Table 2.5, but it is much more precisely estimated. 3 2 0ne reason why we see a negative response is that chain establishments might be indirectly affected by the local lending shock. For example, these firms' products may be substitutes for those sold by the small standalones. In this case, the decline in credit for the standalones may allow the chain establishments to expand their operations if a standalone's inability to replace old equipment or expand its operations results in more business for establishments that are part of multi-state firms. 33 1t is not possible to obtain reliable estimates of population changes over these years, so it is unclear whether the shocks affected outmigration or employment to population ratios. 59 2.5.3 The Role of Small Business Loans in "Normal" Economic Times Up to now, we have considered the effects of the credit shocks that occurred over the 2007-2009 period. However, the methodology we use to construct the predicted lending shock can also be used to assess how shocks affected the real economy during less volatile times. To this end, we extend the analysis to include shocks dating back to 2000 and employ a model that incorporates all shocks simultaneously. We estimate a model that constrains the effect of the predicted lending shock to be the same for all years, but allows for a shift in 2008 and 2009. For loan originations, the estimating equation is: 0 In (lit) = OlPit + 2Pit-1 + 03 (v 2008 x Pit) + 04 (v 200 9 x Pit-1) (2.7) +05 (v 200 9 X Pit) + 06 (v 201o x Pit-,) - fXit + Ai + 6,t + Eit, where vt is a dummy for year t, and the lending shocks for county i in year t, pit, are calculated as in Equation 2.4. This specification assumes that a shock has an effect over two periods, in t and t + 1. In addition to reporting the estimated 0 parameters, we also report the total effect of the 2008 shock (01 + 02 + 03 + 04), the total effect of the 2009 shock (01+ 02 + 05 06), and the excess effect of the 2008 and 2009 shocks, which are (03 + 04) and (05 + 06), respectively. We estimate the model separately for small establishment and total county employment growth rates. Table 2.8 presents these estimates. Column (1) shows there is a strong relationship between predicted small business lending and actual small business lending in all years of the sample. This means we can use the predicted lending shocks to test the relationship between lending and employment in non-crisis years. The interaction terms show that there is a much larger and more precise effect of predicted lending on actual lending over the 2008-2009 period. We believe this asymmetry is an inherent feature of the shift-share approach, as it is easier to predict where lending will decline than where it will grow. Column (2) presents estimates from Equation 2.7 with the LBD small standalone employment growth out- come. The shock terms for the pre-2008 period are small and insignificant. The estimated impact of the 2009 shock, however, is similar in magnitude to the estimate in Table 2.5: one standard deviation in the predicted lending shock is associated with a 0.4 percent reduction in the small establishment growth rate over the 2009-2010 period. The difference in effects between 2009 and the earlier years is significant, as shown by the coefficient on the "Excess effect of the 2009 shock." As in Table 2.5, there is no relationship between the 2008 predicted lending shocks and small standalone employment outcome. Column (3) shows the predicted lending shocks have no significant effect on county-level outcomes in non- crisis years. The effects of the 2008 and 2009 shocks are close to zero and insignificant. We also present visual evidence of the year-by-year differences in the effect of predicted lending shocks on 60 outcomes. To do this, we estimate the following model: In (lit) = WitPt + W2tPt-l + W3tPt+1 + 6st + fXit + Vt + Ai + Eit. (2.8) This model includes the interactions of the year t shock, the lagged t - 1 shock, and the lead t + 1 shock with calendar year dummies. Therefore, it allows the effect of a shock to persist over two periods and to differ by calendar year. We include the lead term as a specification check since a shock in year t + 1 should not affect lending in year t. Figure 2-3 plots the effect of the shock originating in each year. For each year t, we plot the sum of wit and W2t+1, which is the effect of a one standard deviation lending shock that occurred in year t on lending in years t and t + 1. The dotted lines show the 95 percent confidence interval. As seen in Table 2.8, the relationship between predicted lending and actual lending is highly significant in all years, but displays a counter-cyclical pattern with a point estimate that is almost 5 times larger in 2009 than in 2004." As previously discussed, we believe this pattern derives mechanically from the construction of the predicted shock. Figure 2-4 is analogous to Figure 2-3 except the outcome variable is now small standalone employment growth. The figure visually confirms that the 2009 shock has a larger effect on small business employment relative to any other year. We conclude that while the effect of small business lending shocks on employment is relatively small during the 2008 financial crisis, it is larger than in earlier years. The lack of a relationship during the pre-crisis period may be because economic activity is responsive to changes in total credit availability, not just credit from banks. Alternative sources of credit for small businesses, such as home equity loans, were also restricted during the 2007-2009 period, which could be why the effect during the 2008 financial crisis is larger. It is also worth noting that credit was largely booming during the pre-crisis period. As such, our results may reflect that the real effects of negative credit supply shocks are asymmetric relative to those of positive credit supply shocks. 2.6 Interpretation Applying estimates of the cross-sectional effects of these supply shocks to time-series variation in aggregate small business lending is not straightforward. One must determine how much of the change in small business lending nationally was due to supply shifts rather than to demand shifts, and incorporating the general equilibrium effects of these shocks is non-trivial. Instead, we conduct the following simple bounding exercise. We obtain an upper bound estimate of the 34 The estimated coefficient (standard error) for the average of the lead terms of the predicted credit supply shock, wit, is -0.0010 (0.0007), indicating that future credit supply shocks do not significantly affect loan originations. The lack of significance of the lead terms supports the validity of our specification. 61 aggregate effects by assuming that the entire reduction in small business lending between 2007 and 2009 was driven by the credit supply decisions of banks. Clearly this will overestimate the effect of reduced credit supply since some of the observed reduction in lending was due to lower demand for credit as a result of the recession and a more elevated risk of business default. However, we still believe this to be a useful exercise for the purpose of assessing the magnitudes of our estimates. The first step is to note that CRA-disclosed small business lending declined by 22% in 2008 and 33% in 2009. Assuming that these represent supply shifts, we can apply these shifts to our estimates to assess the magnitude of the aggregate impact of the 2007-2009 lending shocks on small business employment growth and county-level economic activity. The second step is to estimate a two stage least squares (2SLS) model where employment is the dependent variable, and the regressors of interest are contemporaneous and lagged log loan originations. The instru- ments for these regressors are the interactions of the 2008 lending shock with 2008, 2009 and 2010 dummies, and the interaction of the 2009 shock with 2009 and 2010 dummy. The model also includes all main effects, state-by-year fixed effects, and the standard set of county-level control variables interacted with year dum- mies. Thus, the first-stages are versions of Equation 2.6 where the dependent variables are contemporaneous and lagged log loan originations. These models are estimated on data from 1997 through 2010.31 The 2SLS estimates are reported in Table 2.9. All entries can be interpreted as elasticities since the outcomes are expressed as growth rates or natural log differences and the endogenous variables are the natural log of the loan origination rate. Column (1) reports the estimates for employment growth of small establishments in the LBD. As this is a bounding exercise, we are less concerned with statistical significance than with magnitudes and confidence bands. Nevertheless, we find that employment growth has a marginally significant positive relationship with contemporaneous shocks and an insignificant relationship with lagged shocks. The elasticity of employment growth with respect to current loan originations is 0.009 (se = 0.008), while the elasticity with respect to lagged originations is -0.003 (se = 0.007). The first stage Angrist-Pischke F-statistics are well above conventional thresholds. The 2SLS estimates imply that the national changes in small business lending (22% in 2008 and 33% in 2009) resulted in 0.3 percentage points (= 0.22 x 0.009 + 0.33 x 0.009 + 0.22 x -0.003+0.33 x -0.003) lower small business employment at the end of 2010 due to the reduction in small business lending. Thus, this upper bound estimate accounts for 3 percent of the 10 percent decline in small business employment between the end of 2007 and 2010 for firms with fewer than 20 employees. The same exercise using the values at the upper end of the 95% confidence interval for the time t and t-1 lending effects results in a reduction of small business employment of 1.2 percentage points, or an upper-bound of 12 percent of the overall change over the period.3 6 However, this estimate is the combination of two extreme assumptions and the true contribution 3 5 Appendix Table A2.4 reports the corresponding OLS models. 3 6 Calculating the 95% confidence interval associated with the 0.005 point estimate requires calculating the standard error of the linear combination of coefficients on In (lit) and in (lit-). This, in turn, requires knowing the covariance between these 62 MIWWMUWWikiMWIWJRNEPillilipiWPluiltill'Milsileilliiniill"I'''"l'illii 'i'''"''''11'1""1'11?''ill''lille is very likely much lower. Column (2) reports on aggregate county employment. A similar calculation on the effect of total employment shows a modest cumulative upper bound effect of 0.3 percentage points from 2008-2009. Thus, the point estimates suggest that the reduction in small business lending accounted for a decline in total employment of up to 5 percent of the 6 percent decline in total employment between the end of 2007 and the end of 2010. Using the values at the upper end of the 95% confidence interval implies the decline in small business lending contributed to no more than 15% of the overall reduction in total employment. Here, too, it is apparent that the decline in small business loans was not a primary contributor to the employment decline during the Great Recession. 2.7 Conclusion This paper has used a new identification strategy and what we believe to be the most comprehensive data set ever assembled to investigate the role of bank lending on the real economy. We develop a new measure of local credit supply shocks for small businesses that is based on (i) the market shares of the banks that served a county before the crisis and (ii) the national change in each bank's lending that is attributable to supply factors (e.g., due to differences in the crisis' effect on their balance sheets). The analysis finds that the 2008 and 2009 measures of local credit supply shocks are associated with sharp declines in total county-level small business loan originations. We find that a one standard deviation reduction in the 2009 measures of local credit supply shocks is associated with a 17% reduction in total county-level small business loan originations from the end of 2008 through the end of 2010, indicating that, at least in the near term, it is costly for small businesses to switch lenders. The paper then assesses the effect of these credit supply shocks on the real economy. A one standard deviation reduction in the 2009 predicted lending shock depresses 2010 levels of small establishment employment growth by 0.4 percentage points. Upper bound estimates suggest that the 2007-2010 decline in small business lending accounted for up to 3 percent of the 10 percent decline in small firm employment over the same period, and up to 5 percent of the 6 percent decline in total employment. Finally, we note that while there is a significant relationship between predicted lending shocks and bank loans to small businesses during the 1997-2007 period, these predicted shocks are not associated with changes in economic activity. This finding suggests that, at least when using the variation from this paper's identification strategy, the credit channel is not empirically important in normal times. These results are also informative, although unlikely to be dispositive, about a series of policy issues. The banking industry is heavily regulated and, as the extraordinary response to the recent financial crisis demon- strates, governments are willing to extend significant aid to banks in moments of financial stress. In the regression coefficients. As the LBD data are restricted access, we are unable to calculate this directly, and therefore use the covariance calculated from the CBP/QCEW estimates. 63 United States, these policies included capital injections through the Toxic Assets Relief Program, nearly costless loans from the Federal Reserve to banks, and stress tests. It is possible, and indeed perhaps likely, that the credit shocks and resulting impacts on the broader economy would have been more severe in their absence. We close with two final observations. First, if the limited impact of credit shocks on the real economy during the "normal" period of 1997-2007 reflects that borrowers are usually able to adjust to a restriction in bank credit without impacting the real economy, it may be appropriate to take these findings as evidence that proposals for new programs that aim to increase banks' credit supply are unnecessary if the goal is to affect county-level employment. Second, as this paper does not present evidence on the moral hazard consequences of government interventions in the banking market, it does not provide a complete picture of their welfare impacts. 64 Figures Figure 2-1: Geographic Distribution of the Predicted Lending Shock Adjusted Predicted Credit Shock ftupma a = Ma-%4 0.39 to14 0 0.14to 0.3 0 -018 t 0.1 -5.47n to -0 -8 No datan.. Figure 2-2: Regression-Adjusted Difference in ln(loan originations) between Top Quartile and Lower Quartile Counties In 0- 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 0 -0.1 -0.2 06 -0.3 0. 0 -0.4 -0.5 -0.6 C 0 -0.7 0 -0.8 " ottom quartile counties C 'U 0 -0.9 -M iddle-50 -1 Notes: Figure is based on estimation of Equation 2.5. Quartiles are based on the value of the 2007-2009 predicted lending shock. See text for further details. Figure 2-3: Effect of the Predicted Lending Shock on Loan Originations, by Year 0.3 0.25 0.2 0.15 e00 -I 0.1 0.05 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 Notes: Figure is based on estimation of Equation 2.8 where the dependent variable is log small business loan originations. Y-axis shows effect of a one standard deviation change in predicted log lending on log loan originations in t and t+1. See text for further details. 66 Figure 2-4: Effect of the Predicted Lending Shock on Small Business Employment Growth, by Year 0.006 0.004 0.002 0- 2000 2001 3 2004 2005 6 2007 008 / 9 -0.002 -0.004 \0I'~ / -0.006 \ -0.008 'I -0.01 Notes: Figure is based on estimation of Equation 2.8 where the dependent variable is the employment growth rate for small standalone firms, defined as single unit establishments with fewer than 20 employees. Y-axis shows effect of a one standard deviation change in predicted log lending on small business employment growth in t and t+1. See text for further details. 67 Tables Table 2.1: Changes in Lending Between 2007-2009 for Selected Large Bank Holding Companies (1) (2) (3) Percent change in small Percentile Percentile net of county business lending fixed effects Bank of New York Mellon -89.9 2 3 JP Morgan Chase -88.5 2 2 Citigroup -83.6 4 6 Bank of America -77.2 6 9 Wachovia -57.0 18 22 Capital One Financial -79.5 5 5 Suntrust Banks -41.8 34 36 Regions Financial -37.7 38 34 Wells Fargo -33.1 44 59 HSBC -31.9 45 71 BB&T -19.5 60 58 PNC Financial -33.2 43 44 U.S. Bancorp -3.3 76 78 Median across all CRA reporting banks -32 All banks combined -48 Note: Column (1) is the percent change in lending to firms with less than $1m in gross revenue between 2007-2009 as reported in CRA disclosures published by the FFIEC. Column (2) is the percentile of the change in CRA lending across all holding companies that meet the criteria for CRA disclosure (a lower percentile is worse). Column (3) is the percentile in the change in CRA lending after partialing out county fixed effects. 68 Table 2.2: County Characteristics (1) (2) (3) (4) (5) Above median - Below Above median in Below median in median in predicted p-value on predicted lending predicted lending p-value on lending shock (within- within-state shock shock difference state) difference Employment growth 2002- 0.042 0.042 0.970 0.005 0.351 2006 (n=3117) [0.114] [0.132] (0.005) Wage growth 2002-2006 0.142 0.154 0.000 0.000 0.965 (n=3117) [0.0781 [0.087] (0.003) Home price appreciation 0.327 0.449 0.000 -0.014 0.321 2002-2006 (n=571) [0.020] [0.203] (0.014) % change total bank lending 0.022 0.108 0.001 -0.039 0.168 2002-2006 (n=3138) [0.6091 [0.772] (0.028) Log median per capita 10.585 10.580 0.624 0.012 0.172 income 2006 (n=3140) [0.220] 10.270] (0.009) Poverty rate 2006 (n=3141) 15.363 15.507 0.555 -0.064 0.767 [5.950] [6.728] (0.217) Construction share 2006 0.066 0.064 0.484 0.002 0.335 (n=3115) [0.043] [0.0511 (0.002) Manufacturing share 2006 0.180 0.149 0.000 0.006 0.208 (n=3073) [0.129] [0.136] (0.005) ln(density in 2006) (n=3113) -10.860 -11.190 0.000 0.309 0.000 [1.567] [1.929] (0.055) ln(population in 2006) 10.347 10.098 0.000 0.340 0.000 (n=3114) [1.303] [1.670] (0.052) Debt-to-income ratio 2006 1.540 1.651 0.000 -0.090 0.000 (n=2219) [0.529] [0,689] (0.025) Notes: Standard deviations in brackets. Employment growth, wage growth, construction share, and manufacturing share are from the QCEW. Change in lending is from the FFIEC. Per capita income, poverty rates, population and density are from the Census. Home values are from Zillow. County debt-to-income ratios are from the Federal Reserve Bank of New York. Column 4 is obtained from a regression of the county characteristic on an indicator for above median with state fixed effects. 69 Table 2.3: Testing for Spatial Sorting in Bank Lending Shocks (1) (2) County-level data Bank-level data Dependent variable: Fixed effect of the bank with the largest marketshare in the Dependent variable: Bank county fixed effect Constant 0.493 0.618 (0.044) (0.031) Fixed-effect of the bank with the second -0.043 largest marketshare in the county (0.069) Average competitor bank fixed-effect in 0.078 counties where the bank operates (0.111) Observations 2352 654 R-squared 0.0025 0.001 Notes: Robust standard errors in parentheses. Model 1 is an OLS regession of the fixed effect of the bank with the largest market share in the county on the fixed effect of the bank with the second highest market share, weighted by the number of establishments in the county in 2006. The bank fixed effects are estimated from a regression of the log change in small business lending by county and bank between 2007 and 2009 on county and bank holding company fixed effects weighted by 2007 lending. Model 2 is a regression of each bank holding company's fixed effect on the average bank fixed effect in counties where the bank operates. To compute the average bank fixed effect, we calculate the dollar weighted average bank fixed effect in every county excluding bank i, and then aggregate these averages to the bank holding company level weighting by the share of bank i's lending in the county. 70 Table 2.4: Relationship between Predicted Lending Shock and ln(loan originations) (1) (2) (3) 2009 shock * 2010 0.0744 0.0812 0.0822 (0.0089) (0.0090) (0.0096) 2009 shock * 2009 0.0809 0.0879 0.0882 (0.0087) (0.0088) (0.0091) 2008 shock * 2010 0.0549 0.0893 0.0937 (0.0108) (0.0105) (0.0117) 2008 shock * 2009 0.0496 0.0766 0.0815 (0.0106) (0.0111) (0.0123) 2008 shock * 2008 0.0478 0.0735 0.0847 (0.0092) (0.0089) (0.0093) Cumulative effect of 2008 shock 0.1523 0.2394 0.2599 (0.0258) (0.0249) (0.0273) Cumulative effect of 2009 shock 0.1553 0.1691 0.1704 (0.0164) (0.0166) (0.0175) F-test of joint significance of shock interactions (p-value) 0.000 0.000 0.000 Observations 43358 42224 30884 State-by-year fixed effects X X X Baseline controls X X Debt-to-income ratio X Notes: Entries are based on estimation of Equation 2.6 where the dependent variable is log loan originations. Standard errors clustered on county in parentheses. An observation is a county-by-year cell. Shocks refer to predicted loan originations as specified in Equation 2.4. Baseline controls are 2006 log density, log population, construction share, manufacturing share, and log per capita income. All controls are interacted with year dummies. All main effects are included. See the text for further details. 71 Table 2.5: Effect of Predicted Lending Shock on Employment and Establishment Growth Rates for Small Standalone Firms Employment growth rate Establishment growth rate (1) (2) (3) (4) (5) (6) 2009 shock * 2010 0.0015 0.0018 0.0019 0.0081 0.0073 0.0073 (0.0009) (0.0009) (0.0010) (0.0046) (0.0043) (0.0046) 2009 shock * 2009 0.0011 0.0022 0.0023 0.0073 0.0063 0.0058 (0.0010) (0.0010) (0.0012) (0.0053) (0.0050) (0.0055) 2008 shock * 2010 -0.0003 -0.0019 -0.0018 0.0073 -0.0033 -0.0055 (0.0011) (0.0010) (0.0010) (0.0051) (0.0046) (0.0051) 2008 shock * 2009 -0.0011 0.0002 -0.0004 0.0079 -0.0061 -0.0101 (0.0014) (0.0012) (0.0013) (0.0064) (0.0060) (0.0067) 2008 shock * 2008 -0.0024 -0.0002 -0.0006 -0.0066 -0.0105 -0.0137 (0.0009) (0.0009) (0.0009) (0.0200) (0.0202) (0.0224) Cumulative effect of 2008 Shock -0.0038 -0.0019 -0.0029 0.0086 -0.0199 -0.0292 (0.0028) (0.0024) (0.0025) (0.0265) (0.0260) (0.0288) Cumulative effect of 2009 Shock 0.0026 0.0040 0.0042 0.0154 0.0136 0.0131 (0.0014) (0.0015) (0.0018) (0.0096) (0.0088) (0.0098) F-test of joint significance of shock interactions (p-value) 0.03 0.03 0.10 0.10 0.55 0.47 Observations 43540 42420 30842 43540 42420 30842 State-by-year fixed effects X X X X X X Baseline controls X X X X Debt-to-income ratio X X Note: Entries are based on estimation of Equation 2.6. The dependent variable in Columns (1)-(3) is the employment growth rate for small standalone firms calculated according to Equation 2.1. The dependent variable in Columns (4)-(6) is the establishment growth rate for small standalone firms calculated according to Equation 2.2. Small standalone firms are defined to be single-unit establishments with fewer than 20 employees. Standard errors clustered on county in parentheses. An observation is a county-by-year cell. Shocks refer to predicted loan driginations as specified in Equation 2.4. Baseline controls are 2006 log density, log population, construction share, manufacturing share, and log per capita income. All controls are interacted with year dummies. All main effects are included. See the text for further details. Table 2.6: Effect of Predicted Lending Shock on Employment Growth Rates for Small Establishments that are Part of Multi-Unit Firms LBD: Establishments that are Part of NETS: Establishments that are Part of Multi-Unit Firms Multi-State Firms (1) (2) (3) (4) (5) (6) 2009 shock * 2010 0.0023 0.0015 0.0018 (0.0017) (0.0014) (0.0016) 2009 shock * 2009 0.0011 0.0012 0.0012 0.0038 -0.0034 -0.0028 (0.0011) (0.0011) (0.0012) (0.0050) (0.0035) (0.0037) 2008 shock * 2010 0.0017 -0.0029 -0.0034 (0.0016) (0.0015) (0.0016) 2008 shock * 2009 0.0020 0.0000 -0.0003 0.0151 -0.0005 -0.0015 (0.0014) (0.0013) (0.0013) (0.0039) (0.0033) (0.0036) 2008 shock * 2008 -0.0003 -0.0010 -0.0009 -0.0041 -0.0100 -0.0100 (0.0015) (0.0014) (0.0015) (0.0060) (0.0057) (0.0060) Cumulative effect of 2008 shock 0.0034 -0.0039 -0.0047 0.0110 -0.0105 -0.0115 (0.0034) (0.0030) (0.0031) (0.0070) (0.0063) (0.0067) Cumulative effect of 2009 shock 0.0035 0.0028 0.0030 0.0038 -0.0034 -0.0028 (0.0023) (0.0021) (0.0024) (0.0050) (0.0035) (0.0037) F-test of joint significance of shock interactions (p-value) 0.18 0.26 0.29 0.00 0.25 0.29 Observations 43503 42406 30842 40184 39142 28678 State-by-year fixed effects X X X X X X Baseline controls X X X X Debt-to-income ratio X X Notes: Entries are based on estimation of Equation 2.6 where the dependent variable is the employment growth rate for small establishments that are part of multi-unit firms. Small establishments are defined to be those with less than 20 employees. Columns (1)-(3) use the LBD data, which extends through 2010. Columns (4)-(6) use the NETS data, which extends only through 2009. Standard errors clustered on county in parentheses. An observation is a county-by-year cell. Shocks refer to predicted loan originations as specified in Equation 2.4. Baseline controls are 2006 log density, log population, construction share, manufacturing share, and log per capita income. All controls are interacted with year dummies. All main effects are included. See the text for further details. Table 2.7: Effect of Predicted Lending Shock on County Aggregate Outcomes Employment growth Establishment growth (1) (2) (3) (4) (5) (6) 2009 shock * 2010 0.0006 0.0008 0.0007 0.0006 0.0012 0.0013 (0.0009) (0.0009) (0.0010) (0.0005) (0.0004) (0.0005) 2009 shock * 2009 0.0027 0.0026 0.0027 0.0007 0.0015 0.0016 (0.0011) (0.0011) (0.0012) (0.0005) (0.0005) (0.0006) 2008 shock * 2010 0.0017 0.0002 0.0004 -0.0012 -0.0013 -0.0015 (0.0011) (0.0010) (0.0011) (0.0007) (0.0005) (0.0006) 2008 shock * 2009 -0.0012 -0.0003 -0.0010 -0.0002 -0.0003 -0.0007 (0.0014) (0.0012) (0.0013) (0.0008) (0.0007) (0.0007) 2008 shock * 2008 -0.0002 0.0012 0.0007 -0.0002 0.0006 0.0004 (0.0010) (0.0009) (0.0010) (0.0005) (0.0004) (0.0005) Cumulative effect of 2008 shock 0.0004 0.0012 0.0001 -0.0016 -0.0011 -0.0018 (0.0028) (0.0024) (0.0026) (0.0018) (0.0014) (0.0015) Cumulative effect of 2009 shock 0.0033 0.0034 0.0034 0.0013 0.0027 0.0028 (0.0017) (0.0017) (0.0019) (0.0010) (0.0009) (0.0010) F-test of joint significance of shock interactions (p-value) 0.017 0.176 0.313 0.111 0.001 0.010 Observations 42947 41973 30830 42947 41973 30830 State-by-year fixed effects X X X X X X Baseline controls X X X X Debt-to-income ratio X X Notes: Entries are based on estimation of Equation 2.6 where the dependent variables are, respectively, county-level employment and establishment growth. We use the average of the growth rates from the CBP and QCEW. Standard errors clustered on county in parentheses. An observation is a county-by-year cell. Shocks refer to predicted loan originations as specified in Equation 2.4. Baseline controls are 2006 log density, log population, construction share, manufacturing share, and log per capita income. All controls are interacted with year dummies. All main effects are included. See the text for further details. Table 2.8: Effect of Predicted Lending Shock on Small Business Employment by Year Small All Private Log Standalones Employment Originations (LBD) (CBP/QCEW) (1) (2) (3) Shock (t) 0.0492 -0.0006 0.0005 (0.0030) (0.0005) (0.0004) Shock (t-1) 0.0337 0.0002 0.0007 (0.0031) (0.0003) (0.0003) Shock (t) * 2008 0.0285 0.0005 -0.0010 (0.0083) (0.0011) (0.0010) Shock (t-1) * 2009 0.0410 0.0001 -0.0008 (0.0109) (0.0013) (0.0008) Shock (t) * 2009 0.0465 0.0028 0.0009 (0.0088) (0.0011) (0.0010) Shock (t-1) * 2010 0.0860 0.0010 -0.0008 (0.0091) (0.0009) (0.0012) Total effect of 2008 shock 0.1524 0.0002 -0.0006 (0.0144) (0.0018) (0.0014) Total effect of 2009 shock 0.2154 0.0035 0.0013 (0.0152) (0.0015) (0.0018) Excess effect of the 2008 shock 0.0695 0.0006 -0.0018 (0.0158) (0.0021) (0.0014) Excess effect of the 2009 shock 0.1326 0.0039 0.0001 (0.0160) (0.0016) (0.0018) F-test for joint significance of interactions (p-value) 0.00 0.09 0.24 Observations 30160 30300 29945 Notes: Entries are based on estimation of Eqution 2.7. Standard errors clustered on county in parentheses. An observation is a county-by-year cell. Shocks refer to predicted lending shocks as calculated in Equation 2.4. The total effect of the 2008 shock is: shock(t) + shock(t-1) + shock(t)*2008 + shock(t-1)*2009. The total effect of the 2009 shock is: shock(t) + shock(t-1) + shock(t)*2009 + shock(t-1)*2010. The excess effect of the 2008 shock is: shock(t)*2008 + shock(t-1)*2009. The excess effect of the 2009 shock is: shock(t)*2009 + shock(t-1)*2010. All models include baseline controls (2006 log density, log population, construction share, manufacturing share, and log per capita income) interacted with year dummies. All main effects are included. See text for further details. 75 Table 2.9: Two Stage Least Squares Models of the Relationship between Economic Activity and Small Business Loan Originations LBD CBP/QCEW (1) (2) ln(loan originations) (t) 0.0089 0.0203 (0.0078) (0.0080) ln(loan originations) (t-1) -0.0032 -0.0143 (0.0073) (0.0077) Point estimate: upper bound impact of 2008-2009 credit supply reduction on 2008-2010 employment growth -0.003 -0.003 Upper 95supply reduction on 2008-2010 employment growth -0.012 -0.009 Angrist Pischke First-Stage F-stat (t) 52.23 51.88 Angrist Pischke First-Stage F-stat (t-1) 77.00 77.40 Observations 39359 39001 Notes: Entries show two stage least squares estimates of the relationship between small business lending and employment. The dependent variable in Column (1) is small business employment growth. The dependent variable in Column (2) is county-level employment growth. All models include state-by-year fixed effects along with baseline controls (2006 log density, log population, construction share, manufacturing share, and log per capita income) interacted with year dummies. All main effects are included. The upper bound impact of the 2008-2009 credit supply reduction on 2008-2010 employment growth is obtained by assuming the entire decline in small business lending observed over this period was supply-driven. See text for further details. 76 Chapter 3 Credit is Cash: A Model of Household Borrowing* 3.1 Introduction In standard models of credit demand, the price associated with a loan is simply the interest rate. In reality, however, loan contracts encompass many dimensions that contribute to the effective price a household pays to borrow. These include the burden of providing collateral, the time required to travel to the bank and the administrative costs incurred in applying for the loan. Understanding how these other factors influence households' credit decisions is important for evaluating the impact of policy on household credit demand. Consider, for example, the focus in many developing countries on helping households transition from informal to formal sources of credit. While formal lenders, such as commercial banks, often charge low interest rates, there are other factors that make them relatively inaccessible: bank branches are often more geographically dispersed than informal institutions, and banks typically have more restrictive lending policies such as strict collateral requirements. In this case, determining how sensitive households are to factors other than the interest rate is extremely relevant for policies geared at increasing formal access. In this paper, I use a cash demand framework to model borrowing decisions when loans are associated with two kinds of costs: fixed and marginal costs. Fixed costs are incurred only at the time of taking out the loan and do not scale with the amount borrowed - e.g, the cost of traveling to the bank. Marginal costs, on the other hand, are incurred for the duration of the loan and scale in proportion with the amount borrowed; these will capture the burden of complying with loan policies such as the interest rate or collateral requirement. *1 am extremely grateful to Rob Townsend for his feedback and guidance on this project. Special thanks as well to Abhijit Banerjee, Michael Greenstone, Adrien Auclert, Manasi Deshpande and Yuhei Miyauchi as well as to seminar participants at MIT for their insights and comments, and to Jiaqiang Chen for help with the Thai data. I also gratefully acknowledge support from the Legatum Center at MIT. 77 Households in the model will face multiple lenders offering different tradeoffs between these two costs. Given a need for financing, they will decide which institution(s) to borrow from by choosing the lowest-cost loan portfolio. The fact that households can borrow from multiple lenders is an important feature of the model. It is a well-documented fact that formal and informal credit sources coexist in many countries and that households often borrow simultaneously from both. This is demonstrated by, among others, Jain & Mansuri (2003) in Bangladesh and by Collins et al. (2009) in Bangladesh, India and South Africa. Correspondingly, several models that incorporate formal and informal credit have been proposed. Bell et al. (1997) and Kochar (1997) both focus on rationing in the formal credit market, while Gine (2011) uses a contract theory model to explain the coexistence of formal and informal credit in Thailand. A limitation of this existing literature is that much of the focus has been on the broad distinction between formal and informal lenders, even though institutions within each category might differ quite markedly from each other.1 The model proposed in this paper tractably incorporates multiple lenders and allows us to distinguish between specific institutions. To provide an empirical framework for the model, I focus on the example of Thailand. The Thai credit market is typical of those in many countries in that the set of lending institutions spans the spectrum from formal to quasiformal to informal. Loan contracts tend to differ at the level of the institution, and so the question of how households trade off between different costs is closely related to the question of how they choose which institution(s) to borrow from. Using household survey data collected from urban communities across Thailand, I characterize the choice sets faced by households in this environment, describe the kinds of borrowing behavior they give rise to, and subsequently show that simulations from the model match the observed patterns. These results show that incorporating the distinction between fixed and marginal costs can yield models that match empirical patterns of household behavior while providing a framework for understanding how tradeoffs between different costs drive borrowing decisions. Ultimately, however, the distinction between fixed and marginal costs is still too coarse to be useful for policy evaluation. Are households more sensitive to changes in collateral requirements or to interest rates? Would opening new bank branches be an effective way to increase credit access, or should policymakers focus on easing the lending restrictions at existing banks? These questions require a model that distinguishes between individual components of fixed and marginal costs. I therefore conclude the paper with a discussion of how the framework presented here could be used as the foundation for a discrete choice model of credit demand. This model would decompose the demand for loans into the demand for particular loan attributes (e.g., collateral requirement, interest rate and distance from the bank) and therefore be ideally suited to inform policy. 1Gin6 (2011), for example, includes both private commercial banks as well as government-subsidized microfinance organiza- tions in the formal category. 78 This paper relates to the development literature that has focused on credit demand in markets where formal and informal lenders coexist. In contrast to the existing literature, this model tractably incorporates an environment with multiple lenders rather than two broad sectors. Given the focus on Thailand, this paper also supplements a growing literature that has examined the financial system in that country. Of particular note is Townsend & Zhorin (2014), who model the supply side of the Thai credit market; they endogenize the location and contract choice of banks and model banks' location decisions. This paper can be viewed as a complementary exercise where I instead take the supply side as fixed and model households' credit demand. As I use a modification of cash demand to model credit demand, this paper also relates to Alvarez et al. (2013) who examine cash management among Thai households. Finally, this paper builds loosely on the classic model of cash demand proposed in Baumol (1952). To the best of my knowledge, however, this is the first paper to use this framework in the context of credit demand rather than traditional cash management. The paper proceeds as follows. In Section 3.2, I characterize the Thai credit market, show that loans differ along many dimensions in addition to the interest rate, and describe the corresponding patterns of household borrowing behavior. Building on these empirical observations, in Section 3.3, I present a model where loans are associated with fixed and marginal costs, and show that simulations from the model qualitatively match many of the patterns observed in the data. Section 3.4 discusses the benefits and challenges of moving from the cash demand framework to a discrete choice model of credit demand. Section 3.5 concludes. 3.2 Household Credit in Thailand In this section, I use survey data to show that loans in the Thai credit market differ along many dimensions in addition to the interest rate, and that observed patterns of borrowing suggest these factors play an important role in driving household credit demand. While I focus on the particular case of Thailand, the empirical facts observed here are similar to those in other countries and markets where households face numerous sources of credit. 3.2.1 Data I use panel data from the 2005-2011 waves of the Urban sample of the Townsend Thai Data. These are annual survey data collected from 96 urban communities distributed across 6 provinces in Thailand. The household-level dataset is a stratified, clustered, random sample of 15 households in each community, yielding an annual sample size of roughly 1,440. The initial motivation for the surveys was to understand how households in Thailand deal with risk through either informal or formal channels. Consequently, the surveys span a broad range of modules with particular 79 focus on tracking the financial lives of these households. The data include information on education, assets and investment, income and expenditures, saving, consumption, household composition and, most impor- tantly for this paper, a complete enumeration of the household's loan portfolio in every year, including both formal and informal sources of credit. For each loan in a household's portfolio, we know the source of the loan as well as other characteristics including term, amount, interest rate, and collateral requirement. Table 3.1 provides summary statistics for the pooled 2005-2011 sample. 50% of households report owning a business and, reflecting the urban setting, only 9% are involved in agriculture. Average annual income is 465,000 Thai baht (roughly 16,000 USD) and nearly 75% of households are landowners. 64% of the pooled sample is active in the credit market and, conditional on borrowing, only a small percentage of households borrow only from informal lenders. In addition to the household survey, a key informant/headman survey is administered annually to track community-level changes, such as changes in technology or population flows. Of greatest relevance for this paper, the 2005 baseline survey asks key informants to enumerate the financial institutions currently active in the community. Unfortunately, this question was not included in subsequent resurveys, but I will use the 2005 responses to gain some insight into the relative ubiquity of different lenders. 3.2.2 Loan Products I first characterize the loan products offered in this market, and then describe the corresponding patterns of household borrowing in Section 3.2.3. A limitation of the data is that we only observe loans that were actually extended to households; loan contracts that were available to households but not taken up are not included. This limits the extent to which I can enumerate the menu of contracts faced by a given household, but, asi will discuss below, institutional knowledge about lending policies can alleviate some of these concerns. Tables 3.2 and 3.3 show that, in any given year, loans differ along several dimensions that render them more or less costly to households. In addition to the interest rate, there is variation in the collateral requirement and the consequence of default associated with different loans (Table 3.2), as well as the distance households must travel to access loans from different institutions (Table 3.3). This variation in loan characteristics is largely driven by differences in lending policies across institutions. As in many countries, the credit market in Thailand consists of multiple institutions spanning from formal to quasiformal to informal. Formal lenders include organizations such as private commercial banks and gov- ernment banks that are heavily institutionalized and centralized, but who operate local branches throughout the country. In Thailand, the main government banks are the Bank for Agriculture and Agricultural Co- operatives (BAAC) and the Government Savings Bank (GSB). Quasiformal lenders, such as cooperatives and microfinance groups, also operate according to institutionalized standards, but have greater scope for local autonomy than formal lenders. The primary quasiformal lenders in Thailand are the village funds, 80 government-subsidized microfinance banks that exist in every village and whose policies are (largely) deter- mined by local committees.' Finally, the informal sector includes moneylenders as well as loans between family and friends or business acquaintances. Table 3.4 shows how loan characteristics vary across institutions for a representative year from the sample. At the sector level, formal institutons tend to offer lower interest rates but impose stricter collateral requirements relative to informal lenders. Formal bank branches are also geographically less accessible, while quasiformal institutions, namely the village funds, are located in every village. In fact, Table 3.4 shows that village fund loans are relatively inexpensive along several dimensions: low interest rates, low collateral requirements, and conveniently located for most households. It is not surpising, then, that village funds are the single largest source of household credit in Thailand with a market share of 61%. Table 3.4 is indicative of variation across the whole sample, but we can also show that most households actually have acccess to, and therefore the discretion to choose between, multiple lenders. Households' choice sets are not directly observed in the data (since we only see the loans they took out), but the 2005 key informant survey solicits a list all institutions who are active lenders in each community. Figure 3-1 shows that most communities have at least four active lenders, in addtion to informal sources such as moneylenders and friends/family. This suggests that, given a need for external finance, most households have scope for choosing between multiple lenders and, therefore, multiple loan contracts. Moreover, Figure 3-2 shows the most ubiquitous institutions in the sample are the formal lenders (BAAC, commercial banks, GSB) and village funds. This is important because these institutions offer standardized loan products: formal institutions' policies are standardized at the national level, while village fund policies are standardized at the village level. This makes it more plausible that a household's choice set, or some large part of it, can be inferred based on the characteristics of loans taken out by other households in the sample. For the remainder of the paper, I treat the choice between different loan contracts as equivalent to the choice between lending institutions. This is, of course, a simplification since a single institution may offer different contracts; however, it seems reasonable given that most variation occurs across, rather than within, institutions. In the same vein, I treat institutional type as the relevant distinction between lenders: if a household borrows from two commercial banks with different names, I treat this as borrowing from a single lender/institution. 2 The village funds were established as part of the Million Baht Fund program in 2001. This was a large-scale credit intervention that granted 1 million baht to every village in Thailand to found a local microfinance bank. These funds are subject to operating guidelines imposed by the central government (e.g., the maximum loan size is 20,000 baht and funds must charge a positive interest rate), but, within these broad directives, local policies are entirely determined by village committees. For more on the structure of these funds and their impact on the Thai economy, see Kaboski & Townsend (2011, 2012). 81 3.2.3 Household Borrowing Patterns Having characterized the supply side of the market, I now describe households' borrowing behavior. Recall from Table 3.1 that 64% of households in the pooled sample are active in the credit market in the sense of holding outstanding loans from either formal or informal lenders. Figure 3-3 shows this aggregate number hides a declining trend in borrowing activity over the survey period. Other results, not shown here, show this decline has occurred across all types of credit (formal, quasiformal, and informal), rather than being driven by reduced borrowing in any single sector. Though the following tables will abstract from this and describe borrowing behavior conditional on being active in the credit market, we should keep the overall trend of declining participation in mind. Table 3.5 shows that, over the course of a year, households tend to hold multiple loans and borrow from multiple institutions: conditional on borrowing, nearly 70% of households borrow more than one loan over the course of a year and nearly 35% borrow from more than one institution. Moveover, as shown in Table 3.6, while the majority of households borrow entirely within one sector of the credit market, roughly 25% borrow from different sectors. For example, 14% of households borrow from both a formal and quasiformal institution in the same year, while 0.7% borrow from both formal and informal institutions. Given these patterns, two questions arise: (i) what household characteristics are correlated with borrowing from formal, quasiformal, or informal institutions, and (ii) why might households borrow multiple loans from multiple lenders in the same year? Table 3.7 looks at the correlation between household characteristics and having any formal / quasiformal informal loans or no loans at all. We see that business owners are more likely to be active in the credit market, and that wealthier households tcnd to borrow from formal institutions, iess wealthy ones borrow from quasiformal institutions, and those without land borrow in the informal sector. Households with kin nearby are more likely to borrow from the formal and quasiformal sectors. Non-active households tend to be non-business owners who have high levels of education and high income. These results are consistent with the loan contracts offered in different sectors. Formal institutions have low interest rates, but are more inaccessible in terms of imposing higher collateral requirements and being located farther away. These barriers may be easier for households with higher income to surmount. Likewise, households with kin nearby may have an easier time finding the resources (or guarantors) needed to satisfy the collateral requirements imposed by quasiformal and formal lenders. The composition of non-active households suggests they do not borrow simply because they have no need for it. Of course, these cross-sectional patterns are only part of the picture since the same household often borrows simultaneously in multiple sectors. An explanation for this behavior that has been explored in other papers is credit rationing: certain lenders may impose restrictions on the amount that can be borrowed, leading households to borrow up to the limit of their most preferred lender and then supplementing from additional 82 sources. This surely plays a role in Thailand since, in some cases, we have direct knowledge of such limits: by federal regulation, for example, the maximum amount for a village fund loan is 20,000 baht. For other institutions, these limits likely exist even though they are not explicitly stated. An alternative (and not mutually exclusive) explanation for borrowing from multiple lenders is that house- holds borrow for many reasons (e.g., to buy a house or to pay emergency medical bills) and the structure of different contracts may render them more or less suitable for financing different projects. Therefore, we might observe a correlation between the identity of the lender and the reason for borrowing. For each loan in their portfolio, households in the Thai data are asked to provide their "Reason for bor- rowing" that particular loan. Responses are chosen from a list of prepared options including livestock, business investment, build/buy house, for ceremony, and many others. I sort each of these responses into the Investment or Consumption category based on my judgment of the size of the project and its time horizon. For example, "build/buy house" is categorized as Investment, while "for ceremony" is categorized as Consumption. This is admittedly rough, and made even more so by the fact that households can provide multiple reasons for borrowing a single loan. Nonetheless, the evidence is suggestive. Table 3.8 shows that Investment projects (i.e., large projects with a long time horizon) tend to be funded from formal sources, while Consumption projects (i.e., small projects with a short time horizon) are largely funded by quasiformal loans. Informal lending, on the other hand, does not play a large role in funding either Investment or Consumption projects. Instead, it appears that households primarily rely on informal lenders for emergency situations. Table 3.8 shows households' responses to the following question: "Suppose you encounter an unavoidable emergency and you needed 20,000 baht (roughly 700 USD) right away. How would you get 20,000 baht? If that source wasn't available, what would you do? And if that source wasn't available, what would you do?" Most households resort to savings (this is the first response for 50% of households) but, when that is unavailable, the majority rely on borrowing from family/friends or a moneylender. By contrast, the percentage who cite borrowing from either a bank or quasiformal lender in any of their three responses is very low. These data suggest that both credit rationing and contract/project matching are factors affecting household borrowing decisions in Thailand. The model presented in Section 3.3 will accommodate both as mechanisms that lead households to borrow from multiple lenders. 3.3 Credit Demand as Cash Demand Having used the data to describe the relevant environment, I now present a model of household credit demand that incorporates the empirical observations. Using the fact that credit is ultimately just cash, I modify a standard framework used in the literature on cash demand to model household borrowing decisions. 83 In the classic Baumol (1952) model of the transactions demand for cash, households face a stream of perfectly foreseen transactions for which they need cash on hand. However there is a cost i to holding cash (which weighs against holding too much) and a fixed cost b for withdrawals (which weighs against holding too little). Facing this tradeoff, households choose the frequency and size of withdrawals that minimizes total costs subject to meeting their transaction needs. I augment this basic framework by introducing multiple lenders who offer diferent combinations of i and b. The focus is on modeling households' decisions between different sources of credit or, analagously, between different sources of cash. Households perfectly foresee their future funding needs inclusive of all non-credit cash flows including income, consumption, and investment. Given these needs, the only decision households make is who to borrow from, and they simply choose the lowest-cost combination of loans that satisfies their budget constraint. I describe the model in detail in Section 3.3.1, then show in Section 3.3.2 that simulations match many of the patterns observed in the Thai data. 3.3.1 Model There are T periods and H households. Each household h is characterized by an exogenously determined, fully known vector of cash surpluses in every period: surplush = [surplush, surplush,..., surplus . surplust can be negative or positive, and is the cash position of household h in period t after netting out all non-credit inflows (e.g., wage income, business revenue, payouts from government insurance programs) and outflows (e.g., business expenses, consumption expenditures, agricultural investments). In periods where surplust > 0, the household's inflows (weakly) exceed its outflows and no further action is needed. When surplust < 0, however, cash outflows exceed inflows and the household will have to borrow externally or spend down savings to meet the difference. The set of lending institutions is denoted J. Each institution j E J offers a single loan contract, and the cost to household h of borrowing from institution j has two components: 1. a marginal cost, P, incurred every period in which the loan is outstanding, and 2. a fixed cost, F, incurred only in the period in which the loan is withdrawn. In addition, institutions can impose a limit L4 on the principal amount outstanding in all periods. There are several ways to think about the mapping between the contract offered by institution j and the costs P4 and F. On the one hand, institutions may directly offer triplets of P4, F, and L to household 84 h. Alternatively, and more realistically, each institution offers a contract that consists of many dimensions including, among other factors, the distance the household will have to travel to get to the lender, the administrative costs of applying for a loan, the collateral requirement imposed, the interest rate offered, and the maximum amount that can be borrowed. Collectively, these impose some aggregate cost on household h, and Ph and F4 capture the marginal and fixed components of this cost in a reduced form way. J 3 Let loanin denote the vector of net loan flows from institution j to household h in every period: loaninh = [loanin,1,loanin,. .., loanin . loanint > 0 indicates that household h is a net borrower in period t; loanin4 <0 indicates household h is a net repayer. For reasons that will be clear later on, I define an indicator variable for whether household h is a net borrower in period t: I loaii27' > 0 0 otherwise The outstanding balance owed to institution j in period t is balance4,t. I assume that balance),o is exogenously given for all j E J and may be greater than zero. In all periods t > 0, t balanceht = balance,+ Z loanin . Marginal costs are paid on the amount outstanding as of the end of the previous period, and fixed costs are incurred only when taking out a new loan. In period t, the total amount owed by household h to institution j is therefore Fx4.J + P balancegt 1 . Savings is treated analogously to external sources of cash and is denoted j = S. The previous notation translates exactly with a few adjustments. loahin, represents the net flow from savings in period t: household h is a net dis-saver when loaninh, > 0, and is a net saver when loanih<0. balance), is the amount of savings that remains at the end of period t and, as above, I assume that balarce%, is exogenously given (i.e., households enter with exogenously given wealth). As loan flows in subsequent periods now diminish rather than augment the outstanding balance, in all periods t > 0 we have t balances, = balances,0 - loaninirt. i= 1 ph denotes household h's returns to saving, which may be positive or negative. The model s is agnostic about the particular form that savings takes, and it is allowed to vary across different households. Accordingly, for any given household, Ph may be negative (the household earns interest on its savings) or positive (savings is 85 costly). In any period t, the cost associated with saving is therefore given by the same expression as before: Fhxh + ph balanCehg1 We can now formulate the household's problem. Let costh be the total cost associated with borrowing and saving for household h in period t: costt = F xjt + P balancegJ . jE. Each household chooses the pattern of borrowing and saving that minimizes total cost and satisfies all funding needs: T min 3 costt { }oanJ{ T} t1 subject to cost = surplush + loaninhVt balance > 0, Vj, t balanceh < L, Vj, t A few comments before showing simulations from the model: first, note the constraints only require that outstanding balances in every period be non-negative and respect the borrowing limits for each institution. In particular, there is no constraint that they must be run down to zero by period T. One way to think of this is that we only observe a snapshot of T periods for a longer-lived household. Households may enter and leave with outstanding balances but, in the T periods for which we observe them, their borrowing and saving decisions must be chosen to minimize total cost. Second, this framework easily accommodates the possibility that households may decide to borrow from multiple institutions, even within the same period. As long as households satisfy the budget constraint in every period (and respect the constraints on outstanding balances), there are no restrictions on the combination of loans that can be used to meet funding needs at minimal cost. Finally, there are two reasons we might observe that borrowing decisions differ across households in this model. The first is that surplush and balance.J0 vary across households: clearly, households who have different funding needs and different resources available will make different borrowing decisions. Of greater interest is the variation that arises from heterogeneity in the costs associated with any given lender: P and F are specific to a given household h and may vary in the population. Parellel to the previous discussion, one way to think about this is that institution j offers different triplets of P4, F4, and L4 to different households. Moneylenders, for example, almost surely tailor their loan terms to the circumstances of each particular borrower. As discussed in Section 3.2, however, we know that many quasiformal and formal 86 institutions in Thailand offer standardized loan contracts to all customers. In this case, heterogeneity in P4 and F4 is driven by the fact that the burden of complying with institution j's lending policies differs across households. As a simple example, suppose the only cost associated with borrowing is the collateral requirement, and that institution j requires all borrowers to provide physical collateral. Household A owns a plot of land as well as a car, while Household B owns neither, and so we would expect P4 < PP since it is easier for Household A to comply. As I show in the following section, this means that even if A and B face identifical funding needs, their borrowing patterns may differ due to heterogeneity in the costs they associate with different lenders. 3.3.2 Simulations To show this model captures many of the patterns observed in the Thai data, I solve for the optimal loan portfolio in an illustrative setting. Let T = 8 and suppose there are four institutions, Savings (S), Bank (BA), Village Fund (VF) and Moneylender (ML), all of whom inherit the characteristics of their empirical counterparts. The Bank charges a relatively low interest rate, but is located far away and requires collateral. The Village Fund, in comparison, offers a very attractive source of funding as it charges a lower interest rate, is located nearby and has less restrictive collateral policies. Village Fund loans are, however, constrained by a maximum loan amount. Finally, the Moneylender charges a high interest rate, but is very accessible and has no collateral requirements. Consider Household A, whose cash profile is shown in Figure 3-4. A has deficits in periods 1, 4, and 7 that will need to be funded either externally or out of savings. For simplicity, I assume balanceio = 0, VJ E {S, BA, VF, ML}. A's cost of borrowing from each institution is given in Table 3.10. To focus on the choice between external lenders, I assume the returns to saving are zero. Reflecting that the Moneylender charges high interest rates but is very accessible, A has a high marginal cost of borrowing from the Moneylender but zero fixed cost. In contrast, the Bank charges lower interest rates (and so PA is only 0.08), but imposes a high fixed cost. Finally, while borrowing from the Village Fund is relatively inexpensive for Household A in terms of both fixed and marginal costs, there is a loan cap of L AF = 50. We can already see that the size of the funding need will play a crucial role in determining who A borrows from in a given period. When the amount required is small enough, it is cheaper to pay the high marginal cost to the Moneylender than to pay the fixed costs charged by the Bank and the Village Fund. Conditional on the loan amount being above this threshold, however, A will always borrow as much as possible from the Village Fund first and then supplement as necessary from either the Bank or the Moneylender. This interaction between funding need and the household's choice of lender illustrates how incorporating fixed and marginal costs allows the model to capture some of the contract/project matching dynamic observed in Tables 3.8 and 3.9. Those results suggested that one reason we observe simultaneous borrowing from 87 multiple lenders is that households borrow for many reasons and the structure of different loan contracts renders them more or less suitable for different purposes. Analogously, the tradeoff between fixed and marginal costs associated with different lenders in the model renders them more or less attractive as a source of funding depending on the amount needed. To the extent that households face funding needs of different magnitude in different periods, this leads the same household to borrow from different lenders over time. Figure 3-5 illustrates this more concretely by showing Household A's optimal path for borrowing and saving. At t = 1, A borrows up to the limit allowed from the Village Fund and then supplements with a small loan from the Moneylender. At t = 4, the funding need is still small enough that borrowing from the Moneylender is the cheapest option. At t = 7, however, A needs a large enough loan that the cost-minimizing choice is to pay the fixed cost at the Bank in order to access the low marginal rate. Notice that the model also captures credit rationing as an explanation for borrowing from multiple lenders. At t = 1, the cheapest source of credit is the Village Fund, but A is constrained by the borrowing limit and so has to borrow from the Moneylender as well. While variation in loan size drives the result that households borrow from different lenders in different periods, credit limits lead to borrowing from multiple lenders within the same period. This simple example reflects many patterns observed in the Thai data: in the optimal portfolio, Household A borrows multiple loans from multiple lenders, and this behavior is driven by variation in funding needs over time as well as by credit rationing. To show how the model also rationalizes cross-sectional variation in borrowing patterns, consider Household B whose cash profile is identical to A's, but whose cost of borrowing is given in Table 3.11. The only difference is that B associates a higher fixed cost with borrowing from the Bank: FA = 12 > FA = 6. Recalling the discussion in Section 3.3.1, a possible interpretation is that A owns a car while B does not, and so the burden of traveling to the Bank is more costly for B. Figure 3-6 shows this single alteration gives rise to a different solution for Household B. Increasing the fixed cost of going to the Bank drives Bank borrowing to zero; B still borrows the maximum amount allowed from the Village Fund, but subsequently borrows only from the Moneylender. Viewed in the cross-section (and given the interpretation for why FBA > FBA), the comparison of A and B would show that wealth (which is likely correlated with owning a car) is positively correlated with Bank borrowing. These examples, while ad hoc, illustrate the model's ability to rationalize the empirical patterns observed in the Thai data while also providing a framework for understanding how tradeoffs between different costs affect household credit demand. 3.4 Discrete Choice Notwithstanding the results shown in Section 3.3, the cash demand model has limited value as a tool to inform policy. The distinction between fixed and marginal costs is very coarse, and the policy proscriptions 88 that would arise from the model are correspondingly too broad to be useful. In this final section, I therefore discuss the possibility of using a discrete choice model. In many ways, discrete choice is a natural fit for this context. In the classic random coefficient discrete choice model, households have preferences over product attributes, and these preferences are allowed to vary cross-sectionally based on household characteristics. Products, in turn, are modeled as bundles of attributes and households choose the product which yields the highest utility. The policy relevance of these models is immediately clear: by decomposing the demand for any given product into the demand for individual attributes, we not only isolate the dimensions along which demand is most sensitive (and along which policy changes would have the greatest effect), but can also evaluate the introduction of entirely new products. As a result, discrete choice has been used to model demand in many product markets ranging from cars (Berry, Levinsohn & Pakes (1995, 2004)) to cereal (Nevo (2000)). The much wider range of potential policy simulations is the primary benefit of moving from cash demand to discrete choice. Loans can be viewed as bundles of attributes (e.g., interest rate, collateral requirement, distance) over which households have preferences that vary according to characteristics such as wealth or occupation. We can think of these preferences as being derived from the cash demand model described in Section 3.3, which describes the underlying relationship between loan attributes and household behavior. Estimating the preference parameters then opens the door to evaluating policies as varied as mandating the geographic dispersion of commercial banks, granting land rights (which enable more households to meet collateral requirements) or imposing higher interest rates on microfinance groups. While there are clearly benefits, applying discrete choice to credit demand is not entirely straightforward. A central feature of the cash demand model, reflecting observations from the Thai data, is that households can borrow multiple loans from multiple lenders. In most discrete choice models, however, households make a single choice from the menu of products available. Some papers have developed models to accommodate multiple discrete choices - e.g., Hendel (1999) and Koulayev et al. (2012). To ensure tractability, however, these models assume that, while a given household makes multiple choices, each one can be treated, to some extent, as an independent discrete choice. This is clearly inappropriate for credit where, as formalized in the cash demand model, loan decisions in one period have direct implications for funding needs, and therefore loan decisions, in future periods. Applying discrete choice in this setting therefore requires a dynamic model where households make multiple discrete choices, and choices in different periods are linked and interdependent. Developing such a framework will be the focus of future work. 89 3.5 Conclusion Credit market policies form the heart of development strategy in many countries, and there are several channels through which governments attempt to improve financial access either directly or indirectly: these include mandating the expansion of commercial banks into rural areas, subsidizing formal credit, or strength- ening property rights. These polices are aimed at reducing the myriad costs associated with borrowing and, to evaluate their potential impact, we need a framework that examines how costs other than the interest rate affect household credit demand. In this paper, I propose a cash demand framework for evaluating how the tradeoff between fixed and marginal costs determine households' borrowing decisions when faced with a market of multiple lenders. The structure of the model is heavily informed by the Thai credit market, and I show, in turn, that the model is able to rationalize many of the observed patterns of household behavior. Nonetheless, the broad distinction between fixed and marginal costs is very coarse. A promising avenue for future work is to investigate whether the framework for household behavior proposed here could be extended to a discrete choice model of credit demand. 90 "I Figures Figure 3-1: Lending Sources per Village U- 0~ 0- 2 3 4 5 8 7 Note: Data are from the 2005 key informant survey; sample size is 96 communities. Each bar shows the number of communities with x active lenders. Includes agricultural cooperatives, BAAC, commercial banks, financial companies, GSB, Production Credit Groups, and village funds. Excludes informal sources such as moneylenders and friends/family. Figure 3-2: Geographic Ubiquity of Different Lenders 0 0- 0* S U- 0 0 .- Note: Data are from the 2005 key informant survey; sample size is 96 communities. Each bar shows the number of communities who report that institution x is an active lender. AC=Agricultural Cooperative; BAAC=Bank for Agricul- ture and Agricultural Cooperatives; CB=Commercial Bank; FC=Financial Company; GSB=Government Savings Bank; PCG=Production Credit Group; VF=Village Fund 91 Figure 3-3: Percentage of Households Holding At Least One Loan 2004 2006 2008 2010 2012 Note: Data are from the 2005-2011 household surveys. Figure 3-4: Sample Profile for surplush Surplus 500 250 -250- - 500I - I I LI 1 2 3 4 5 6 7 8 92 - Figure 3-5: Household A's Solution S Balance BA Balance 300 300 200 200 100 100 U 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 VF Balance ML Balance 3 'A. 300 200 200 100 100 " A 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 Note: The figures show balanceA obtained from solving for A's optimal loan portfolio. Figure 3-6: Household B's Solution S Balance BA Balance 300 300 200 200 100 100 n 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 VF Balance ML Balance 300' 300 200- 200 100- 100 U 2 1 2 3 4 5 6 7 8 1 2 345 67 8 ' Note: The figures show balanceB obtained from solving for B's optimal loan portfolio. 93 . Tables Table 3.1: Household Summary Statistics Mean SD N Panel A: Demographic Variables Age of household head 52.2 14.3 10,336 Years of education 6.5 4.5 10,288 % with male household head 59.6 - 10,336 No. of household members 4.2 1.9 10,060 Panel B: Occupation % who own a business 50.6 - 10,336 % involved in agriculture 8.8 - 10,336 Panel C: Income/Wealth Income (1,000 baht) 465.3 964.3 10,057 Wealth (1,000 baht) 1,042.4 2,048.8 10,336 % who own land 73.1 - 10,336 Value of landholdings 669.7 1,780.0 10,336 Panel D: Credit % with outstanding loans 63.7 - 10,336 % who borrow only in informal market 3.3 - 10,336 Pooled data from 2005-2011 household surveys. Wealth includes the value of all household, business, and agricultural assets in addition to the value of landholdings and current savings. Informal lenders include moneylenders, business partners, and friends/family. Table 3.2: Variation in Loan Characteristics Min Max Mean SD N AnnualI interest rate 0 240 12.2 29.7 1,397 Amount (1,000 baht) 1 2,000 53.0 147.1 1,532 Term (months) 1 360 20.3 34.2 1,481 Collateral requirement 0 6 2.1 1.1 1,517 Consequences of default 1 3 2.0 0.1 1,488 Data are from the 2011 household surveys. Collateral requirements are measured on an index where higher values correspond to stricter requirements: 0=no collateral, 1-2=guarantors, 3-7=physical collateral. Consequence of default is measured on an index where higher values correspond to harsher consequences: 1=no consequence, 2=not able to borrow again from this lender, 3=not able to borrow again from this lender and others. 94 Table 3.3: Location of Active Lending Institutions Location Percent In village 12.7 Not in village, in tambon 68.3 Not in tambon, in amphoe 17.4 Not in amphoe, in changwat 1.4 In another changwat 0.3 These data are from a question in the 2005 baseline survey that asked the key informant in each community to provide the location of all active lending institutions. This question was not included in subsequent resurveys. Institutions include Agricultural Cooperatives, BAAC, Commerical Banks, Finance Companies, Government Savings Bank and Production Credit Groups, but excludes village funds and moneylenders, which are located in every village. Administrative units in Thailand increase in size from village to tambon to amphoe to changwat (province). Table 3.4: Variation in Loans Across Institutions Sector Institution Market Share Amount Term Interest Collateral Location (%) (1,000 baht) (Months) (APR) (Index) F BAAC 10.9 83 40 8.0 2.7 Tam Commercial Banks 3.0 600 147 8.8 3.7 Tam GSB 4.2 304 88 8.1 2.8 Tam Q Village Fund 61.0 20 12 7.2 2.0 Vill Ag. Coop. 4.2 81 24 9.3 2.6 Tam PCG 3.7 21 13 11.9 2.3 Vill Fin. Companies 4.9 440 57 6.9 1.2 Tam I Moneylenders 3.7 42 8 142.7 0.7 - Relatives 4.4 74 26 11.1 0.6 - Data are from the 2011 household survey, except location, which is from the 2005 headman survey. The table display mean attributes with the exception of location, which is the modal response. Section is either formal, quasiformal, or informal. Market Share is based on the number of loans. Collateral requirements are measured on an index where higher values correspond to stricter requirements: 0=no collateral, 1-2=guarantors, 3-7=physical collateral. For location, Tam=Not in village, in tambon; Vill=In village. The headman survey does not ask for the locations of moneylenders and friends/family, though it seems reasonable to expect that these are usually located nearby. 95 Table 3.5: Household Borrowing Patterns Number % of Households Panel A: Loans Held Annually 1 31.6 2 42.1 3 13.0 4 8.5 5 2.2 6+ 2.6 Panel B: Lenders Used Annually 1 66.9 2 26.9 3 5.4 4+ 0.8 Based on snapshot of all loans held annually. Pooled data from 2005-2011 household surveys. Panel A shows the number of loans held annually by individual households, while Panel B shows the number of lenders that a given household borrows from over the course of a year. Lenders are differentiated according to institutional type: if a household borrows from two different commercial banks, this is counted as a single lender. Table 3.6: Borrowing Patterns Across Sectors Formal Quasiformal Informal Formal 7.9 - - Quasiformal 14.0 60.3 - Informal 0.7 8.7 6.1 This table shows the percentage of all borrowing households who borrow from both the row and column sectors. 2./o uorrow froim all three sectors. Pooled data from 2005-2011 household surveys. 96 Table 3.7: Characteristics Associated with Borrowing Outcomes (1) (2) (3) (4) VARIABLES Any Formal Any Quasiformal Any Informal No Credit Income 0.0357*** -0.127*** 0.0165 0.0644*** (0.0120) (0.0166) (0.0108) (0.0160) Business owner dummy 0.0603*** 0.0501*** 0.0132** -0.0730*** (0.00750) (0.0103) (0.00670) (0.00997) Agriculture dummy 0.101*** 0.0207 -0.0409*** -0.0280 (0.0132) (0.0182) (0.0118) (0.0175) Landowner dummy 0.109*** 0.0139 -0.0466*** -0.0117 (0.00850) (0.0117) (0.00760) (0.0113) Education -0.00110 -0.00842*** -0.00390*** 0.00865*** (0.000847) (0.00117) (0.000757) (0.00113) Presence of kin 0.0199** 0.0271** 0.00645 -0.0202 (0.00953) (0.0131) (0.00853) (0.0127) Constant 0.0181 0.575*** 0.166*** 0.354*** (0.0125) (0.0172) (0.0112) (0.0166) Observations 10,010 10,010 10,010 10,010 R-squared 0.039 0.019 0.009 0.018 Robust standard errors in parentheses. Pooled 2005-2011 data. *** p<0.01, ** p<0.05, * p<0.1 Table 3.8: Funding Source by Project Type Project Type % Funded by Formal % Funded by Quasiformal % Funded by Informal Investment 59.0 31.7 9.3 Consumption 12.4 74.3 13.3 Pooled sample of 2005-2011 data. See text for details on how projects are categorized into Investment and Con- sumption projects. Table 3.9: Sources of Emergency Funding Source Response 1 Response 2 Response 3 Borrow from bank 1.52 1.35 2.84 Borrow from family/friends 30.22 63.54 57.66 Borrow from moneylender 8.30 15.23 19.13 Borrow from quasiformal lender 0.48 0.84 1.03 Sell assets 9.47 18.67 19.13 Use savings 50.01 0.36 0.22 Pooled 2005-2011 data. Figures are percent of surveyed households who gave these responses to the following question: "Suppose you encounter an unavoidable emergency and you needed 20,000 baht (roughly 700 USD) right away. How would you get 20,000 baht? If that source wasn't availble, what would you do? And if that source wasn't available, what would you do?" 97 Table 3.10: Household A's Cost of Borrowing S BA VF ML P4 0 0.08 0.06 0.9 F. 0 6 2 0 J - - LA:1 50 Table 3.11: Household B's Cost of Borrowing S BA VF ML P4 0 0.08 0.06 0.9 F.A 0 12 2 0 Le 50 98 References Agarwal, Sumit, and Robert Hauswald. 2010. "Distance and Private Information in Lending." Review of Financial Studies, 23 (7): 2757-2788. Almeida, Heitor; Campello, Murillo; Laranjeira, Bruno, and Scott Weisbenner. 2012. "Corporate Debt Maturity and the Real Effects of the 2007 Credit Crisis." Critical Finance Review, 1: 3-58. Alvarez, Fernando, Anan Pawasutipaisit, and Robert M. Townsend. 2013. "Cash Management in Village Thailand." Working Paper. Amel, Dean F. and Kenneth P. Brevoort. 2005. "The Perceived Size of Small Business Banking Markets." Journal of Competition Law and Economics, 1(4): 771-784. Amiti, Mary and David E. Weinstein. 2013. "How Much Do Bank Shocks Affect Investment? Evidence from Matched Bank-Firm Loan Data." NBER Working Paper 18890. Angrist, Joshua D. and Jorn-Steffen Pischke. 2009. Mostly Harmless Econometrics, Princeton, NJ: Princeton University Press. Ashcraft, Adam B. 2005. "Are Banks Really Special? New Evidence from the FDIC-Induced Failure of Healthy Banks." The American Economic Review, 95(5): 1712-1730. Autor, David, and Mark Duggan. 2006. "The Rise in Disability Rolls and the Decline in Unemploy- ment." The Quarterly Journal of Economics 118(1): 157-205. Baker, Scott; Bloom, Nicholas, and Steven Davis. 2013. "Measuring Economic Policy Uncertainty." Unpublished. Banerjee, Abhijit V. and Esther Duflo. 2010. "Giving Credit Where It Is Due." Journal of Economic Perspectives, 24(3): 61-80 Bartik, Timothy J. 1991. "Who Benefits From State and Local Economic Development Policies." W.E. Upjohn Institute for Employment Research: Kalamazoo, Mich. Barr, Michael S. 2005. "Credit Where it Counts: The Community Reinvestment Act and its Critics." New York University Law Review, 80(2): 513-652. Barros, Pedro Pita. 1999. "Multimarket Competition in Banking, with an Example from the Portuguese Market." InternationalJournal of Industrial Organization, 17(3): 335-352. Baumol, William J. 1952. "The Transactions Demand for Cash: An Inventory Theoretic Approach." The Quarterly Journal of Economics, 66(4): 252-255. Bell, Clive, T. N. Srinivasan, and Christopher Udry. 1997. "Rationing, Spillover, and Interlinking in Credit Markets: The Case of Rural Punjab." Oxford Economic Papers, 49(4): 557-585. 99 Benmelech, Efraim, Nittai Bergman, Anna Milanez, and Vladimir Mukharlyamov. 2014. "The Agglomeration of Bankruptcy." Unpublished. Bentolila, Samuel; Jansen, Marcel; Jimenez, Gabriel, and Sonia Ruano. 2015. "When Credit Dries Up: Job Losses in the Great Recession." Unpublished. Berger, Allen and Gregory Udell. 1995a. "Universal Banking and the Future of Small Business Lending." Center for Financial Institutions Working Papers 95-17, Wharton School Center for Financial Institutions, University of Pennsylvania. Berger, Allen and Gregeory Udell. 1995b. "Relationship Lending and Lines of Credit in Small Firm Finance." The Journal of Business, 68(3): 351-381. Berger, Allen, Anthony Saunders, Joseph M. Scalise, and Gregory Udell. 1998. "The Effects of Bank Mergers and Acquisitions on Small Business Lending." Journal of Financial Economics, 50(2): 187-229. Berger, Allen; Miller, Nathan; Petersen, Mitchell; Rajan, Raghuran, and Jeremy Stein. 2005. "Does Function Follow Organizational Form? Evidence From the Lending Practices of Large and Small Banks." Journal of Financial Economics, 76(2): 237-269. Bernanke, Ben S., and Cara S. Lown. 1991. "The Credit Crunch." Brookings Papers on Economic Activity 22(2): 205-248. Bernanke, Ben. 1983. "Nonmonetary Effects of the Financial Crisis in the Propagation of the Great Depression." The American Economic Review, 73(3): 257-276. Bernanke, Ben and Mark Gertler. 1995. "Inside the Black Box: The Credit Channel of Monetary Policy Transmission." Journal of Economic Perspectives, 9(4): 27-48. Bernanke, Ben S. 2010. "Addressing the Financial Needs of Small Businesses." Federal Reserve Meeting Series, July 12. Available at http://www.federalreserve.gov/newsevents/speech/bernanke20lOO7l2a.htm. Berrospide, Jose M., and Rochelle M. Edge. 2010. "The Effects of Bank Capital on Lending: What Do We Know, and What Does it Mean?" Board of Governors of the Federal Reserve System Finance and Economics Discussion Series Working Paper No. 2010-44. Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. "Automobile Prices in Market Equilibrium." Econometrica, 63(4): 841-890. Berry, Steven, James Levinsohn, and Ariel Pakes. 2004. "Differentiated Products Demand Systems from a Combination of Micro and Macro Data: The New Car Market." Journal of PoliticalEconomy, 112(1): 68-105. Black, Sandra E. and Philip E. Strahan. 2002. "Entrepreneurship and Bank Credit Availability." The Journal of Finance, 57(6): 2807-2833. 100 Blanchard, Olivier Jean. and Larry Katz. 1992. "Regional Evolutions." Brookings Papers on Economic Activity 1: 1-75. Bloom, Nicholas, Max Floetotto, Nir Jaimovich, Itay Saporta-Eksten, and Stephen J. Terry. 2012. "Really Uncertain Business Cycles." NBER Working Paper 18245. Bond, Philip and Robert M. Townsend. 1996. "Formal and Informal Financing in a Chicago Ethnic Neighborhood." Federal Reserve Bank of Chicago Economic Perspectives, 20(4): 3-27. Boot, Arnoud W. A. and Anjan V. Thakor. 2000. "Can Relationship Banking Survive Competition?" The Journal of Finance, 55(2): 679-713. Brevoort, Kenneth P., John A. Holmes, and John D. Wolken. 2010. "Distance Still Matters: The Information Revolution in Small Business Lending and the Persistent Role of Location, 1993-2003." Board of Governors of the Federal Reserve System Finance and Economics Discussion Series Working Paper No. 2010-08. Brunner, Karl, and Allan H. Meltzer. 1963. "The Place of Financial Intermediaries in the Transmission of Monetary Policy." The American Economic Review, 1963: 372-382. Brunnermeier, Markus K. 2009. "Deciphering the Liquidity and Credit Crunch 2007-2008." Journal of Economic Perspectives, 23(1): 77-100. Burgess, Robin and Rohini Pande. 2005. "Do Rural Banks Matter? Evidence from the Indian Social Banking Experiment." The American Economic Review, 95(3): 780-795. Card, David. 2001. "Immigrant Inflows, Native Outflows and the Local Labor Market Impacts of Higher Immigration." Journal of Labor Economics, 19(1): 22-64. Cetorelli, Pierre-Andre, David Perez-Castrillo, and Thierry Verdier. 1995. "Spatial Competition in the Banking System: Localization, Cross Subsidies, and the Regulation of Deposit Rates." European Economic Review, 39(5): 889-918. Charles, Kerwin; Hurst, Erik; and Matthew Notowidigdo. 2012. "Masking and Aggregate Employ- ment Changes: Housing Booms and Manufacturing Declines During the 2000s." Unpublished. Charnes, Sarah, and Alan Krueger. 2011. "JOLTS as a Timely Source of Data by Establishment Size." Monthly Labor Review, 134(5):16-24. Chodorow-Reich, Gabriel. 2014. "The Employment Effects of Credit Market Disruptions: Firm-level Evidence from the 2008-9 Financial Crisis." The Quarterly Journal of Economics, 129(1): 1-59. Cole, Rebel A. 1998. "The Importance of Relationships to the Availability of Credit." Journal of Banking and Finance, 22(6-8): 959-997. Collins, Daryl, Jonathan Morduch, Stuart Rutherford, and Orlanda Ruthven. 2009. "Portfolios of the Poor: How the World's Poor Live on $2 a Day." Princeton University Press, Princeton, New Jersey. 101 Congressional Budget Office. 2012. "Small Firms, Employment, and Federal Policy." Available at < http://www.cbo.gov/publication/43029>. Drexler, Alejandro and Antoinette Schoar. 2012. "Do Relationships Matter? Evidence from Loan Officer Turnover." Unpublished. Dunkelberg, William C., and Holly Wade. 2012. "NFIB Small Business Economic Trends: January 2012." NFIB publication available at http://www.nfib.com/Portals/0/PDF/sbet/sbet201201.pdf. Duygan-Bump, Burcu, Alexey Levkov, and Judit Montoriol-Garriga. 2015. "Financing Con- straints and the Great Recession." Journal of Monetary Economics. Fisman, Raymond, Daniel Paravisini, and Vikrant Vig. 2012. "Cultural Proximity and Loan Outcomes." Fort, Teresa C., John Haltiwanger, Ron S. Jarmin and Javier Miranda. 2013. "How Firms Respond to Business Cycles: The Role of Firm Age and Firm Size." IMF Economic Review, Palgrave Macmillan, vol. 61(3): 520-559. Garmaise, Mark J. and Tobias J. Moskowitz. 2006. "Bank Mergers and Crime: The Real and Social Effects of Credit Market Competition." The Journal of Finance, 61(2): 495-538. Gilje, Erik. 2012. "Does Local Access to Finance Matter?" Unpublished. Gilje, Erik, Elena Loutskina, and Philip E. Strahan. 2013. "Exporting Liquidity: Branch Banking and Financial Integration." Unpublished. Gine, Xavier. 2011. "Access to Capital in Rural Thailand: An Estimated Model of Formal vs. Informal Credit." Journal of Development Economics, 96(1): 16-29. Goetz, Martin R., and Juan C. Gozzi. 2010. "Liquidity Shocks, Local Banks, and Economic Ac- tivity: Evidence from the 2007-2009 Crisis." Unpublished manuscript available at http://www.martin- goetz.com/GG _Liquidity_ 11 _12.pdf. Gould, Eric D., B. Peter Pashigian, and Candice J. Prendergast. 2005. "Contracts, Externalities, and Incentives in Shopping Malls." The Review of Economics and Statistics, 87(3): 411-422. Greenstone, Michael, Alexandre Mas, and Hoai-Luu Q. Nguyen. 2015. "Do Credit Market Shocks Affect the Real Economy? Quasi-Experimental Evidence from the Great Recession and 'Normal' Economic Times." Unpublished. Haltiwanger, John C., Ron S. Jarmin, and Javier Miranda. 2013. "Who Creates Jobs? Small vs. Large vs. Young." The Review of Economics and Statistics, 95(2): 347-361. Hendel, Igal. 1999. "Estimating Multiple-Discrete Choice Models: An Application to Computerization Returns." The Review of Economic Studies, 66(2): 423-446. 102 Hertzberg, Andrew, Jose Maria Liberti, and Daniel Paravisini. 2010. "Information and Incentives Inside the Firm: Evidence from Loan Officer Rotation." The Journal of Finance, 65(3): 795-828. Hoshi, Takeo; Anil Kashyap, and David Scharfstein. 1990. "The Role of Banks in Reducing the Costs of Financial Cistress in Japan." Journal of FinancialEconomics, 27(1): 67-88. Huang, Haifang, and Eric Stephens. 2014. "From Housing Bust to Credit Crunch: Evidence from Small Business Loans." Canadian Journal of Economics, forthcoming. Ivashina, Victoria, and Sharfstein, David. 2010. "Bank Lending During the Financial Crisis of 2008." Journal of FinancialEconomics, 97(3): 319-338. Jain, Sanjay and Ghazala Mansuri. 2003. "A Little at a Time: The Use of Regularly Scheduled Repayments in Microfinance Programs." Journal of Development Economics, 72(1): 253-279. Jayaratne, Jith and Philip E. Strahan. 1996. "The Finance-Growth Nexus: Evidence from Bank Branch Deregulation." The Quarterly Journal of Economics, 111(3): 639-670. Kaboski, Joseph P. and Robert M. Townsend. 2011. "A Structural Evaluation of a Large-Scale Quasi-Experimental Microfinance Initiative." Econometrica, 79(5): 1357-1406. Kaboski, Joseph P. and Robert M. Townsend. 2012. "The Impact of Credit on Village Economics." American Economic Journal: Applied Economics, 4(2): 98-133. Kerr, William R. and Ramana Nanda. 2009. "Democratizing Entry: Banking Deregulation, Financing Constraints, and Entrepreneurship." Journal of Financial Economics, 94(1): 124-149. Khwaja, Asim Ijaz and Atif Mian. 2008. "Tracing the Impact of Bank Liquidity Shocks: Evidence from an Emerging Market." The American Economic Review, 98(4): 1413-1442. Kochar, Anjini. 1997. "An Empirical Investigation of Rationing Constraints in Rural Credit Markets in India." Journal of Development Economics, 53(2): 339-371. Koulayev, Sergei, Marc Rysman, Scott Schuh, and Joanna Stavins. 2013. "Explaining Adoption and Use of Payment Instruments by U.S. Consumers." Working Paper. Krueger, Alan. 2010. Speech to Charlotte Chamber of Commerce's Summit on Access to Capital for small Businesses and Entrepreneurs. June 29, 2010. Available at Laderman, Elizabeth S. 2008. "The Quantity and Character of Out-of-Market Small Business Lending." Federal Reserve Bank of San Francisco Economic Review : 31-39. Lee, Donghoon and Wilbert van der Klaauw. 2010. "An Introduction to the FRBNY Consumer Credit Panel." 103 Mian, Atif and Amir Sufi. 2014. "What Explains the 2007-2009 Drop in Employment?" Econometrica, 82(6): 2197-2223. Munoz, Ana Patricia and Kristin F. Butcher. 2013. "Using Credit Reporting Agency Data to Assess the Link between the Community Reinvestment Act and Consumer Credit Outcomes." Neumark, David, Brandon Wall, and Junfu Zhang. 2011. "Do Small Businesses Create More Jobs? New Evidence for the United States from the National Establishment Time Series." The Review of Economics and Statistics 93(1): 16-29. Nevo, Aviv. 2000. "Mergers with Differentiated Products: The Case of the Ready-to-Eat Cereal Industry." The RAND Journal of Economics, 31(3): 395-421. Nguyen, Hoai-Luu Q. 2014. "Do Bank Branches Still Matter? The Effect of Closings on Local Economic Outcomes." Unpublished. Notowidigdo, Matthew. 2013. "The Incidence of Local Labor Demand Shocks." Unpublished. Park, Kwangwoo and George Pennacchi. 2009. "Harming Depositers and Helping Borrowers: The Disparate Impact of Bank Consolidation." Review of Financial Studies, 22(1): 1-40. Pashigian, B. Peter and Eric D. Gould. 1998. "Internalizing Externalities: The Pricing of Space in Shopping Malls." Journal of Law and Economics, 41(1): 115-142. Peek, Joseph and Eric Rosengren. 1998. "Bank Consolidation and Small Business Lending: It's Not Just Bank Size that Matters." Journal of Banking & Finance, 22(6-8): 799-819. Peek, Joseph and Eric Rosengren. 2000. "Collateral Damage: Effects of the Japanese Bank Crisis on Real Activity in the United States." The American Economic Review 90(1): 30-45. Petersen, Mitchell and Raghuram Rajan. 1994. "The Benefits of Lending Relationships: Evidence from Small Business Data." The Journal of Finance 49(1): 3-37. Petersen, Mitchell and Raghuram Rajan. 1995. "The Effect of Credit Market Competition on Lending Relationships." The Quarterly Journal of Economics, 110(2): 407-443. Rajan, Raghuram and Luigi Zingales. 1998. "Financial Dependence and Growth." The American Economic Review, 88(3): 559-586. Salop, Steven C. 1979. "Monopolistic Competition with Outside Goods." The Bell Journal of Economics, 10(1): 141-156. Sapienza, Paula. 2002. "The Effects of Banking Mergers on Loan Contracts." The Journal of Finance, 57(1): 329-367. Shleifer, Andrei and Robert Vishny. 2011. "Fire Sales in Finance and Macroeconomics." Journal of Economic Perspectives 25(1): 29-48. 104 Skillern, Peter. 2002. "When Your Bank Leaves Town: How Communities Can Fight Back." Shelterforce, December. Stein, Jeremy C. 2002. "Information Production and Capital Allocation: Decentralized versus Hierarchical Firms." The Journal of Finance, 57(5): 1891-1921. Strahan, Philip E. and James P. Weston. 1996. "Small Business Lending and Bank Consolidation: Is There Cause for Concern?" Current Issues in Economics and Finance, 2(3). Strahan, Philip E. and James P. Weston. 1998. "Small Business Lending and the Changing Structure of the Banking Industry." Journal of Banking & Finance, 22(6-8), 821-845. Townsend, Robert M. and Victor V. Zhorin. 2014. "Spatial Competition among Financial Service Providers and Optimal Contract Design." Working Paper. Walls, Donald W. 2007. "National Establishment Time-Series Database: Data Overview." Presented at the 2007 Kauffman Symposium on Entrepreneurship and Innovation Data. Available at http://papers.ssrn.com/sol3/papers.cfm' 105 Appendices for Chapter 1 106 Salop Circle Model Following the framework of Salop (1979), let a local banking market be represented as a circle of length 1, where borrowers are uniformly distributed along the circle and have linear transportation costs. There are n banks, spaced evenly along the circle as shown in the left panel of Appendix Figure A1.4. Papers that have used Salop circles to model banking markets typically interpret them as representing geographic space (see, for example, Chiappori, Perez-Castrillo and Verdier (1995) and Park and Pennacchi (2009)). Following the example of Barros (1999), however, I interpret positions along the circle as locations in characteristics space. This reflects the fact that distance, while important, is not the only dimension over which consumers have preferences when choosing a credit provider: factors such as services offered and customer satisfaction will also matter. This implies that while consumers have a central tendency to borrow from a bank that is located nearby (since distance is one of the characteristics that determine borrowing preferences), there is no reason to expect they will always borrow from the bank that is geographically closest. This is a more accurate description of actual borrowing patterns than a model where only physical distance matters. As the empirical analysis is based on the location of borrowers and lenders in geographic, not characteristics, space, it is worth discussing how one maps into the other. In the data, I observe locations at the Census tract level, and compare the evolution of lending in tracts where branches close with that of similar, and often neighboring, tracts located in the same county. Let v denote the percentage of a tract's potential borrowers who borrow from a given bank A. If the distribution of borrower preferences is identical across tracts, the framework implies v will be highest in the tract where A is located, since distance is one of the factors that borrowers care about. But we would also expect that v is (i) greater than 0 in tracts neighboring A, since some borrowers are willing to travel if A is closely aligned with their other preferences and (ii) decreasing monotonically in distance from A, since preference for A decreases as distance from it increases. At the end of this section, I discuss what implications this will have for the spillover effects if A closes. First, to illustrate the more general impact of a closing in this market, suppose that lending is relationship- dependent and consumers must invest in building relationship capital with a specific lender before they are able to borrow. This process is costless, but time-consuming. It takes k > 0 periods to build a relationship, and this is the sense in which relationships are "sticky" and difficult to replace. Relationship capital is lender-specific, nontransferable, and permanent (i.e., having established a relationship with a given lender, a firm never has to repeat the process of building a relationship with that lender even if they temporarily stop borrowing). Conditional on having a relationship with a bank located x away along the circle, firms borrow if the net surplus from doing so is positive: i.e., if F (x,p) = s - p - tx > 0, 107 where s is the surplus from borrowing, p is the price of the loan, and t is the "transportation" cost, which can be interpreted as the cost of borrowing from a bank located farther from one's ideal preference. Banks are Nash price setters. They pay a fixed cost f to enter the market, and have marginal cost of lending equal to c. Restricting attention to configurations where banks are spaced evenly around the circle, as in the left panel of Appendix Figure A1.4, a consumer located x E (0, 1/n) from bank i is indifferent between borrowing from i or its neighbor if F (x, p) = F (1/n - X, p) - i.e., if p - pi + 1/n 2t In other words, x is the location of the consumer who is indifferent between borrowing from i or its neighbor. Given the symmetric configuration, bank i's demand is equal to 2x. An equilibrium consists of a price p and number of banks n such that (i) each bank i earns maximum profit: pi = arg max (q - c) - f- and (ii) there is no entry: ir (n) > 0 > 7r (n +), where 7r (n) are per-bank profits, net of the fixed cost of entry, when n banks are in the market. The market begins in an equilibrium (p*, n*) where all borrowers have stable, established relationships and borrow from their closest bank. The market is fully covered, banks are spaced i/n* apart, and the marginal consumer is located 1/2n* from the closest bank. Suppose a bank is closed at time t = 0, and the locations of remaining banks are fixed. Ignoring, for a moment, the possible effect on prices, the right panel of Figure A1.4 illustrates the impact on lending. The dashed segment shows consumers who previously borrowed from the closed bank: those on the thick, blue segment switch to borrowing from one of the neighboring banks, while those on the thin, red segment exit the market as they are too far to earn positive surplus. Lending declines by the length of the dashed segment: those on the thin, red segment are inactive, and those on the thick, blue segment appear inactive while in the process of building a new relationship. Note that this model provides positive rather than normative predictions, and that the scenario depicted in Appendix Figure A1.4 represents just one possible outcome of the branch closing. In particular, it requires the assumption that F (1, p*) > 0 and so the market is initially covered, but F (I, p*) < 0 and so some borrowers exit the market after their branch closes. Given that positions along the circle represent locations in characteristics space, how does the pattern depicted in Appendix Figure A1.4 translate to the tract-level identification strategy? The dashed segment of the circle represents consumers who previously borrowed, and had a relationship with, the closed bank. 108 As previously discussed, if distance is one of the factors that determine firms' borrowing preferences, the fraction of borrowers with a relationship should be higher in tracts located closer to the bank branch. Geographic proximity therefore serves as an imperfect proxy for having a relationship, and the decline in lending represented by the dashed segment in Appendix Figure A1.4 should be most concentrated in the tract where the branch is located. This will be reflected in a lower level of aggregate lending when we compare Exposed tracts with Control tracts in the same county. Up to this point, I have ignored how prices in the market may be affected by the closing, but remaining banks may respond to the change in market structure. As I show below, the pattern of competition suggests the effect on prices will vary according to remaining banks' locations (in characteristics space) relative to the closed branch. To illustrate this, consider one of the bank branches directly neighboring the closed one - for clarity, call this bank A. After the closing, A's market share expands to M = n + , where x is defined by: F (x, p*) 0 s - P* - ff 0 - s - p t If A increases its price to p* + c, it earns c more on each of its inframarginal consumers, but loses those at the margin who drop out due to the higher price. The loss in market share corresponding to an c increase in the price is i. To see this: at p*, the marginal consumer is located s-P away from A. At p* + E, the marginal consumer is located S-P*' away. Increasing the price by e therefore decreases A's market share yt This means the range of E for which the benefit of increasing the price outweighs the cost is given by: e(M - > (P* -C) 6 < Mt - (p* -c). That is, it is optimal for A to increase its price if Mt > (p * - c). I derive the parameter restrictions under which this condition is satisfied by substituting in the equilibrium expressions n* = /f and p* = c + V (these are derived in Salop (1979)). t is then given by: S s - p t t 109 A's market share after the closing is: 1 M = +.- 2n* 1 s - c - if 2 t/f t Which means it is optimal for A to increase its price if: Mt > p*-c t +s-c- tf > c+ t-c 2t/ 2 -U2 tf > c-s, which can be rearranged as: 3It t In other words, it is optimal for A to increase.its price if firms' surplus from borrowing is high enough relative to banks' marginal costs. The same analysis applies to the bank neighboring the closed branch on the other side. As these banks increase their prices, this provides their upstream neighbors with flexibility in increasing their prices (since some of their customers now face a more expensive outside option), which in turn allows their neighbors to increase prices, and so on and so forth. Ultimately, this can lead to a cascade of price increases across the market, with the magnitude of the increase being larger for banks located closer to the closed branch. This can lead to the scenario depicted in Appendix Figure A1.5, where the orange dotted segments show that consumers who are not directly exposed to the closing may nonetheless drop out in response to the post-closing higher prices. It is not entirely straightforward to map the pattern displayed in Appendix Figure A1.5 to the tract-level identification strategy, but to the extent that distance in geographic space is roughly correlated with distance in characteristics space, we would expect that tracts very far from branch closings are less affected by changes in local concentration than tracts located nearby. This suggests that lending may be lower in Exposed tracts relative to Controls after a closing because these are the areas where prices increase more. While this is a concern in theory, I show in Figure 1-6 that lending does not respond to the entry of new banks in Exposed tracts. This suggests that, in this context, these concentration effects are empirically negligible. 110 Household Credit Outcomes In this section, I use data from the Federal Reserve Bank of New York/Equifax Consumer Credit Panel (CCP) to gauge the impact of branch closings on household credit outcomes such as bankruptcy and delinquency rates. These are quarterly, individual-level panel data provided by Equifax, and constitute a 5% random sample of all individuals with a credit history, along with all members of their household. 3 For each individual, the data include the information contained in their history including loan account data, public record and collection agency data, and a limited amount of individual background data. I use only fourth-quarter data to generate an annual dataset, and aggregate variables to the household level. Households in the CCP data are linked to their Census tract of residence, so these results rely on the tract- level identification strategy described in Section 1.3. Approximately 20% of households in the sample move at least once over the sample period so, for each merger, Exposed and Control tracts are associated with households living there in the year before the intention to merge was announced. The results are robust to including/excluding those who subsequently move. I estimate a household-level version of Equation 1.4, where the right hand side includes a vector of household-level pre-merger characteristics, including whether the household has a mortgage, any delinquent accounts (those at least 30 days past due), any bankruptcy or foreclosure on file, and a proxy for holding a small business credit card. The full CCP sample consists of over 13 million households. Once restricted to those living in Exposed and Control tracts, the sample shrinks to 233,701 households. Summary statistics are shown in Appendix Table A1.9. Households living in Exposed and Control tracts are similar along most dimensions, though those living in Exposed tracts are slightly more likely to hold a small business credit card. These cards are not directly identified in the data, but, as they are characterized by high limits, I proxy for them by identifying households where the average credit limit over all open credit cards is at least $20,000. Appendix Figure A1.6 shows closings have no impact on the financial stability of surrounding households, as measured by delinquency rates, collection rates, credit scores, and bankruptcy rates. The results are similar when the sample is restricted to low-income tracts, and when the Control group consists of tracts located at least 5 miles away from an Exposed tract. There are several possible explanations for this. The first is that Section 1.4.2 shows closings are associated with a prolonged decline in credit supply to local small businesses, and not to local consumers. Data from the 2010 Survey of Consumer Finances show that only 4% of households took out a loan to finance a business they owned in that year. Second, given the CCP consists of only a 5% random sample, the number of observations in any tract may not be large enough to pick up the direct effects of a decline in small business lending. Third, to the extent there are indirect effects stemming from depressed local economic activity, this may be better reflected in the financial stability of households who work, rather than live, in these areas. 3 For a detailed explanation of the randomization procedure, see Lee and van der Klaauw (2010). 111 ...... Appendix Figures Figure A1.1: Branch Closings in Buyer Only and Target Only Tracts versus Unexposed Tracts Number of branch closings Ct, - N4 - 0 +- -10 -5 0 5 10 Years since merger Source: FDIC, author's own calculations. Plot shows estimated deltatau coefficients. Figure plots the 6, estimated from the event study specification, along with the 95% confidence intervals. The treated group is tracts that only had branches from either the Buyer or the Target bank (but not both) prior to the merger, and the control group is unexposed tracts. r = 0 is the year the merger was approved by federal regulators, and all coefficients are normalized relative to -r = -1. The vertical lines at -r = -4 and -r = 6 denote the range over which the panel is balanced. Robust standard errors are clustered at the tract level. Figure A1.2: Exposure to Consolidation and the Incidence of Branch Closings, ZIP-Level Number of branch closings C') - N1 - 0~l4-- - N- 10 -5 0 10 Years since merger Source: FDIC, author's own calculations. Plot shows estimated deltatau coefficients. Figure shows the first stage relationship between exposure to consolidation and the incidence of branch closings at the ZIP level. The figure plots the 5, estimated from the event study specification, along with the 95% confidence intervals. The dependent variable is the number of branch closings in ZIP z in year t. r = 0 is the year the merger was approved by federal regulators, and all coefficients are normalized relative to r = -1. The vertical lines at - = -4 and r = 6 denote the range over which the panel is balanced. Robust standard errors are clustered at the ZIP level. 112 E ! .-- --- __ -- , - __ ...... =214E Figure A1.3: Exposure to Consolidation and the Volume of New Small Business Lending, ZIP-Level New small business loans C - C - 0 N~j- J4i CD, -10 -5 6 5 10 Years since merger Source: FFIEC, author's own calculations. Plot shows estimated deltatau coefficients. Figure shows the reduced form relationship between exposure to consolidation and small business lending at the ZIP level. The figure plots the 6, estimated from the event study specification, along with the 95% confidence intervals. The dependent variable is the number of new small business loans made to borrowers located in ZIP z in year t. r = 0 is the year the merger was approved by federal regulators, and all coefficients are normalized relative to r = -1. The vertical lines at r = -4 and r = 6 denote the range over which the panel is balanced. Robust standard errors are clustered at the ZIP level. Figure A1.4: Salop Circle Model: Effect of a Branch Closing on Lending to Local Borrowers Initial Equilibrium After Closing Figure illustrates the effect of a branch closing on lending to local borrowers following the Salop circle framework described in the Appendix. The left panel shows the initial equilibrium with banks spaced evenly along the circle. The right panel shows the effect of closing a branch. The dashed segment shows consumers who previously borrowed from the closed bank: those located on the thick, blue segment switch to borrowing from one of the neighboring banks, while those on the thin, red segment exit the market as they are too far away to earn positive surplus. Lending declines by the length of the dashed segment immediately after the closing, since those on red exit and those on blue appear inactive while in the process of building a new relationship. 113 Figure A1.5: Salop Circle Model: Price Effects Following a Branch Closing Figure displays the potential market-wide price effects of a branch closing using the Salop circle framework presented in the Appendix. As described in greater detail in that section, a single closing can result in a cascade of increasing prices across the market, with the largest increases occurring for banks located closest to the location of the closing. Consumers on the thick, blue, dashed segment are those stranded by the closing who switch to borrowing from one of the neighboring banks; those on the thin, red, dashed are stranded by the closing but are located too far away from neighboring banks to continue borrowing; those on the dotted orange segments are not directly exposed to the closing, but drop out in response to higher prices. Figure A1.6: Impact on Household Financial Stability Delinquent accounts Third party collections RD *0 -10~~ -5 0T 1 I 10 -5 10 10 -5 0 5 1,0 Event Time Event Time Credit score Bankruptcy CU V - *0 I~ ~~ - -10 -5 0 5 10 -10 -5 0 5 10 Event Time Event Time Sources: FRBNY Consumer Credit Panel / Equifax, author's own calculations Figure shows the reduced form relationship between living in a tract that is exposed to post-merger consolidation and various measures of household financial stability. The figure plots the 6, estimated from the event study specifications, along with the 95% confidence intervals. The dependent variables are indicators for having at least one delinquent account (i.e., an account at least 30 days past due), having any third party collections on file, and having either a bankruptcy or foreclosure on file. r = 0 is the year the merger was approved by federal regulators, and all coefficients are normalized relative to -r = -1. Robust standard errors are clustered at the tract level. 114 Appendix Tables Table A1.1: Geocoding Summary Statistics Year Total Branches Mapped Unmapped % Unmapped 1999 84,312 77,971 6,341 7.5 2000 85,492 79,713 5,779 6.8 2001 86,069 80,919 5,150 6.0 2002 86,578 82,001 4,577 5.3 2003 87,790 85,297 2,493 2.8 2004 89,784 87,598 2,186 2.4 2005 92,042 90,083 1,959 2.1 2006 94,752 93,016 1,736 1.8 2007 97,274 95,847 1,427 1.5 2008 99,163 98,211 952 1.0 2009 99,550 98,856 694 0.7 2010 98,520 97,812 708 0.6 2011 98,204 97,657 547 0.6 2012 97,337 96,774 563 0.6 Source: FDIC, author's own calculations. Table shows summary statistics for the geocoding procedure used to map branch locations from the FDIC Summary of Deposits to their Census tract. Branch locations can be geocoded either by plotting their latitude and longitude, or by matching their street address to those stored in a GIS repository. I rely on the former whenever possible as it is the most reliable, but latitude and longitude data are only available beginning in 2008 and can only be matched to a limited number of observations prior to that. As a result, in every year there are observations that cannot be mapped because they have no lat/long data and their street address was either incomplete or invalid and could not be matched to an address in the GIS repository. Table A1.2: Failing/Crisis Mergers Buyer Target Year Approved FDIC Assistance TD BankNorth Commerce Bank 2008 JPMorgan Chase Bank Washington Mutual Bank 2008 X Wells Fargo Bank Wachovia Bank 2008 U.S. Bank Downey Savings and Loan 2008 X PNC Bank National City Bank 2008 Branch Banking and Trust Company Colonial Bank 2009 X East West Bank United Commercial Bank 2009 X Source: FDIC. Table shows the seven mergers excluded from the primary sample because they were either classified as failing (i.e., they required financial assistance from the FDIC) or occurred during the 2008 financial crisis. 115 Table A1.3: Buyer and Target Small Business Lending Intensity Buyer Target Buyer Target Year Approved Intensity Intensity Manufacturer & Traders Allfirst 2003 6.5 5.5 Bank of America Fleet 2004 1.5 4.3 National City Provident 2004 5.0 0.8 Regions Union Planters 2004 37.6 14.2 JPMorgan Chase Bank One 2004 1.6 1.8 North Fork Greenpoint 2004 12.9 0 SunTrust National Bank of Commerce 2004 11.0 10.2 Wachovia SouthTrust 2004 4.9 13.7 Sovereign Independence Community 2006 5.2 2.0 Regions AmSouth 2006 22.2 30.0 Bank of America United States Trust 2007 1.5 0.3 Huntington National Sky 2007 9.5 11.8 Bank of America LaSalle 2007 1.2 0.8 Source: FDIC, FFIEC, author's own calculations. Table shows the intensity of small business lending for each Buyer and Target bank in the merger sample. Intensity is defined as the ratio of the dollar volume of small business loan originations (as reported through CRA disclosures) over total assets in the year prior to the merger. Table A1.4: The Effect of Closings on Credit Supply in Target Only Tracts (1) (2) Coefficient Exposed Tracts Target Only Tracts # Loans 6POST -2.504*** -1.032 (0.903) (0.727) Obs. 46,631 35,517 $ Volume (thousands) 8 POST -83.91* -37.87 (49.61) (42.23) Obs. 46,601 35,801 Source: FFIEC, author's own calculations. Table compares reduced form estimates of the post-merger decline in new small business loans in Exposed tracts (Column 1) versus tracts that only had branches from the Target bank prior to the merger (Column 2). Robust standard errors are clustered at the tract level and are in parentheses. *** p<0.01, ** p<0.05, * p<0.1 116 Table A1.5: Robustness to Varying Size of the Local Banking Market (1) (2) (3) (4) Variable County 25-Mile 15-Mile 10-Mile 6 POST -2.504*** -2.474*** -2.414*** -2.051** (0.903) (0.895) (0.908) (0.916) Control Mean 90.70 89.5 89.3 89.7 Obs. 46,631 95,177 77,643 57,436 Source: FDIC, FFIEC, author's own calculations. Table shows reduced form estimates of the post-merger mean shift in the level of new small business loans for different definitions of the local banking market. Column 1 is the baseline estimate based on within-county comparisons between Exposed and Control tracts. Column 2 defines the local market for each Exposed tract to be all Control tracts located within a 25-mile radius. Columns 3 and 4 use analogous definitions for markets of 15- and 10-mile radii. The number of observations is higher in Columns 2 through 4 since the same Control tract may be defined as part of the local market for several Exposed tracts. Robust standard errors are clustered at the tract level and are in parentheses. *** p<0.01, ** p<0.05, * p Table A1.6: Correlation between Tract Demographics and Branch Levels Variable Branches Median Income Fraction Minority Branches 1.00 Median Income 0.0079 1.00 Fraction Minority -0.0984 -0.4369 1.00 Source: FDIC, U.S. Census, author's own calculations. Table presents the correlation matrix between tract-level median income and fraction minority (as of the 2000 Census) and the number of pre-merger branches. 117 Table A1.7: Summary Statistics for Exposed and Control ZIPs Variable Exposed Minus Control Control Mean Population 1,622 28,538 (1,111) Population Density 210.6 4,477 (235.5) Fraction Minority -0.037*** 0.280 (0.014) Fraction College-Educated 0.051*** 0.300 (0.011) Median Income (000s) 2.928** 49.06 (1.154) Fraction Mortgage 0.018** 0.721 (0.009) Pre-Merger Branches 2.721*** 9.040 (0.459) Pre-Merger Branch Growth -0.001 0.057 (0.015) Joint F-test 7.10 p-value 0.00 Number Exposed 353 Number Control 1,588 Source: FDIC, U.S. Census, author's own calculations. Table provides summary statistics for Exposed and Control ZIPs. The former are ZIPs that contain at least one Exposed tract; the latter are those that contain only Control tracts. Table values are obtained from a regression of each ZIP-level characteristic on an indicator for being an Exposed ZIP and county fixed effects. Population density is per square mile. Demographic characteristics are as of the 2000 Census; "pre-merger" variables are as of the year preceding each merger. Pre-merger branch growth is the average annual growth in the number of branches for the two years preceding the merger. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p 118 Table A1.8: Industry Dependence on External Finance External Dependence External Dependence Two Digit NAICS Two Digit NAICS Name Measure Flag 62 Health Care and Social Assistance -0.87 0 42 Wholesale Trade -0.79 0 11 Agriculture, Forestry, Fishing, and Hunting -0.43 0 61 Educational Services -0.43 0 81 Other Services (except Public Administration) -0.29 0 44 Retail Trade -0.20 0 22 Utilities -0.14 0 56 Administrative and Support and Waste Management and Remediation Services -0.01 1 48 Transportation and Warehousing -0.01 1 31 Manufacturing 0.16 1 72 Accommodation and Food Services 0.18 1 71 Arts, Entertainment, and Recreation 0.45 1 54 Professional, Scientific, and Technical Services 0.81 1 51 Information 0.95 1 Source: Gilje (2012). Table shows the industry-level measure of dependence on external finance computed by Gilje (2012) using Compustat data for the period 1999- 2008 and the methodology outlined by Rajan and Zingales (1998). Industry groups are based on the two digit North American Industry Classification System. The external dependence measure is the industry median dependence on external finance, and industries with above median dependence on external finance have an External Dependence Flag equal to 1. !!!IIIEHhEhI ______Table A1.9: Summary Statistics for Households Living in Exposed and Control Tracts Variable Exposed Minus Control Control Mean Has a mortgage 0.00486 0.508 (0.00317) Credit score 0.990 683.3 (0.617) Any delinquent accounts 0.00287 0.273 (0.003) Any third-party collections -0.00108 0.220 (0.003) Bankruptcy on file -0.0027 0.127 (0.002) Foreclosure on file 0.00135 0.0242 (0.0010) Small business credit card 0.00475*** 0.0300 (0.0011) Joint F-test 3.85 p-value 0.0003 Number Exposed 32,285 Number Control 199,416 Source: FRBNY Consumer Credit Panel / Equifax, author's own calculations. Table provides summary statistics for households living in Exposed and Control tracts prior to each merger. Table values are obtained from a regression of household-level characteristics on an indicator for living in an Exposed tract and county fixed effects. Credit score is the Equifax risk score, which is correlated with the FICO score and has the same range. Delinquent accounts are those that are at least 30 days past due. I proxy for having a small business credit card by using a dummy equal to one if the average limit over all open credit cards is at least $20,000. The median number of households in each tract is 245, and the median length of time a household appears in the data is 13 years. Robust standard errors are in parentheses. *** p<0.01, ** p<0.05, * p 120 Appendices for Chapter 2 121 Assumptions Underlying Modified Instrument Our estimating equation of interest is of the form: A ln (Y) = yXj + 3In (Qj) + Ej, (1) where i indicates a county and Y is a county-level outcome variable. The outcome is a function of covariates, X2 , and loan originations in county i, ln (Qj). A standard formulation of the Bartik instrument is: Zi (msij x (A In (Qj) - A In (Q))) , (2) where county i's value of the instrument is the product of the beginning of period market shares of banks in its county and the difference between those banks' national change in lending and the economy-wide change in lending. The exclusion restriction for this instrumental variables identification strategy requires that COV (Zi, Ej) = 0. In other words, a bank's national change in lending, relative to the national average, does not covary with factors that determine the county-level outcomes. To understand the implications of this assumption, it is helpful to decompose Q into the contribution of demand and supply shifters. Consider a simple model of credit demand and supply where credit supply of a bank to a county is perfectly elastic and there is county heterogeneity in the elasticity of credit demand. The change in the log "price" of credit for bank j in county i can be written as: A In (pij) = where c. is a bank/county specific supply shock. Demand for credit is: A In (Qij) = Ej - f3A ln (pij) = + i where /3# > 0 is the county-specific demand elasticity for credit. Note that we assume there are county- specific demand shocks but not bank-specific demand shocks. The implications of bank-specific shocks will be discussed later. Given this model, then the Bartik instrument can be written as: zi = ms 3 msij x (( i +#cE() A ln (Q)) T ads reSinstons X + s d n ms ofche - sA tn (Q) . The validity of the exclusion restriction can now be considered in terms of the separate components of Zi. 122 The first term is the average exposure of banks in a county to demand shocks, in places where the banks operate. The second term is the average response to supply shocks (mediated through the demand elasticity) in places where the bank operates. It is necessary to assume that both of these terms are uncorrelated with ci. With respect to the former, this condition implies that markets with worse than average labor market shocks do not have disproportionate representation of banks with greater than average exposure to negative demand-side shocks nationally. However, it is immediately evident that Zi is a function of di so local demand shocks enter directly and are a possible threat to validity. For example, banks that operate in markets that were especially hard- hit by the recession (e.g., Florida, Nevada, and Arizona) may make fewer loans than banks with branches in other areas simply because of the larger negative demand shocks in these areas. The traditional approach to addressing this problem is to exclude the own state or county when constructing the instrument; but, this is unlikely to be useful in this setting, because there are regional economies of scale in banking and there is spatial correlation in demand shocks. We instead employ a shift-share approach that is purged of local demand shocks for lending as a partial solution to this problem. Specifically, we estimate an equation that decomposes the contribution of the change in equilibrium credit to county and bank components: A In (Qij) = di + s3 + eij, where the outcome variable is the log change in small business lending by bank j in county i. The vector di is a full set of county fixed effects and the parameters of interest are those associated with the vector of bank fixed effects, sj. They are estimates of changes in bank credit purged of banks' differential geographic exposure to local lending shocks. We complement the standard shift-share (Bartik) approach by replacing the change in aggregate bank lending, A ln (Qj), in the construction of Zi in Equation 2 with these estimated bank-specific supply shocks. To see what we are identifying in estimating this equation with respect to the simple supply and demand model specified above, note that: s3. = (A In (Qij) - A In (Qu)) CS 1 j=1 The estimated county effect identifies the average supply response for bank j relative to the average supply response in counties where j operates. Note that the demand effect has been differenced out. The resulting 123 instrument is: j=1 i= j=1 The identifying assumption is now weaker than when using the unadjusted approach, and requires only that counties exposed to banks with above or below average supply shocks relative to county averages (weighted by demand elasticities) not systematically have above or below average shocks to outcomes. This assumption would be violated if banks with negative supply shocks, or banks with more nationwide exposure to markets with elastic credit supply, were more likely to be located in areas that were hard-hit by the recession. We made the assumption that county demand shocks have no bank specific component. Suppose that credit demand is instead: Aln (Qij) = c + E - i Aln (pij) D D q = C. + E-'+'-c meaning that there is a bank-specific demand shock that cannot be separately identified from the supply shock. This case entails an added identifying assumption that banks that have a bigger demand shock are not more likely to be located in hard-hit areas. An example of a violation of this assumption is if banks specialize in lending to certain industries, and if these industries then decline relative to others. This would represent a national demand shock to the bank that might be correlated to county outcomes. As such, while the modified-Bartik instrument safeguards against some confounders, it does not safeguard against all possible ones. We therefore employ several specification checks to ensure that our estimates are being driven by supply rather than demand shocks. First, we estimate all models on a sample of small establishments that are part of larger multi-unit firms. Since these establishments should be unaffected by local credit shocks, estimates from this model will pick up any relationship between the instrument and local economic conditions. We find no significant relationship. Second, we show that the estimates are unaffected by inclusion of a rich set of county characteristics which are known to be predictors of the severity of the economic downturn in a local area. Third, we show that the correlation between the county fixed effect and the market-share-weighted bank fixed effects in a county is close to zero. This suggests that banks with negative shocks are not systematically sorted into areas with negative shocks, which is consistent with the assumption that the cumulative supply response of banks in a county is uncorrelated to local economic shocks. Fourth, we show that the estimated supply responses of individual banks in a county are largely uncorrected with one another. This suggests that there is not some unobserved factor which might be correlated with local economic shocks, and that attracted banks who eventually cut supply. 124 Appendix Figures Figure A2.1: CRA-Disclosed Loan Originations to Firms with Less than $1 Million in Annual Revenue 160,000 14 O 140,000 , 120,000 100,000 0 80,000 i C 60,000 0 *A S 40,000 o 20,000 C 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Source: Authors' calculation from FFIEC Community Reinvestment Act disclosure data. Figure A2.2: Effect of the Crisis on County-Level Small Business Loan Originations Kemel Density Estimate of the Log Change in Small Business Loan Originations, 2008-2009 LO -2 -1 0 1 2 In(originations in 2009) - In(originations In 2008) kernel = epanhchnikov, bandwidth = 0.0560 125 Figure A2.3: ln(loan originations) Relative to 2007 0.2 0- 1998 1999 2000 2 2002 2003 2004 6 2 008 2009 2010 -0.2 -0.4 -0.6 -0.8 00ottom quartile counties Middle-50 0Top Quartile counties -1.2 Notes: Figure shows small business loan originations in counties divided according to the value of their 2007-2009 predicted lending shock. See the text for further details. 126 Appendix Tables Table A2.1: Main Effects of the Predicted Lending Shocks (1) (2) (3) Panel A: ln(loan orginations) 2008 shock -0.3033 0.2754 0.2724 (0.0753) (0.0180) (0.0199) 2009 shock -0.1044 0.0128 0.0066 (0.0612) (0.0171) (0.0185) Panel B: Small standalone firms (LBD) Employment growth 2008 shock -0.0038 -0.0010 -0.0007 (0.0012) (0.0010) (0.0010) 2009 shock -0.0011 -0.0010 -0.0009 (0.0008) (0.0007) (0.0008) Establishment growth 2008 shock -0.0094 0.0027 0.0049 (0.0052) (0.0047) (0.0052) 2009 shock -0.0082 -0.0069 -0.0068 (0.0045) (0.0042) (0.0045) Panel C: County-level aggregates Employment growth 2008 shock -0.0004 0.0005 0.0005 (0.0006) (0.0005) (0.0006) 2009 shock -0.0002 -0.0007 -0.0007 (0.0005) (0.0005) (0.0005) Establishment growth 2008 shock -0.0010 0.0001 0.0001 (0.0005) (0.0005) (0.0005) 2009 shock -0.0006 -0.0007 -0.0007 (0.0005) (0.0004) (0.0004) State-by-year fixed effects X X X Baseline controls X X Debt-to-income ratio X Notes: Entries are based on estimation of Equation 2.6. The dependent variables are, respectively, log small business loan originations, small standalone firm employment and establishment growth rates, and county-level aggregate employment and establishment growth rates. Standard errors clustered on county in parentheses. An observation is a county-by-year cell. Shocks refer to predicted loan originations as specified in Equation 2.4. Baseline controls are 2006 log density, log population, construction share, manufacturing share, and log per capita income. All controls are interacted with year dummies. All main effects are included. See the text for further details. 127 Table A2.2: Relationship between Predicted Lending Shock and ln(loan originations) for non-CRA Banks ln(loan originations) 2009 shock * 2009 0.026 (0.026) 2008 shock * 2009 -0.024 (0.030) 2008 shock * 2008 -0.055 (0.031) Observations 29284 Notes: This table tests whether areas with larger credit shocks experienced increased lending from banks not covered by the CRA. The unit of analysis is commercial banks that are below the CRA asset threshold. The dependent variable is small loan balances from FDIC Call Reports. Standard errors clustered on county in parentheses. See text for further details. Table A2.3: Effect of Predicted Lending Shock on Employment and Establishment Growth Rates for Small Establishments, NETS Data Employment growth rate Establishment growth rate (1) (2) (3) (4) 2009 shock * 2009 0.003 0.002 -0.005 0.003 (0.001) (0.001) (0.004) (0.001) 2008 shock * 2009 0.009 0.005 0.021 0.005 (0.002) (0.001) (0.004) (0.002) 2008 shock * 2008 -0.001 0.002 -0.005 0.002 (0.001) (0.001) (0.002) (0.001) Cumulative effect of 2008 Shock 0.008 0.007 0.016 0.007 (0.002) (0.002) (0.004) (0.002) Cumulative effect of 2009 Shock 0.003 0.002 -0.005 0.003 (0.001) (0.001) (0.004) (0.001) F-test of joint significance of shock interactions (p-value) 0.00 0.00 0.00 0.00 Observations 40287 28678 40287 28678 State-by-year fixed effects X X X X Baseline controls X X Debt-to-income ratio X X Note: Entries are based on estimation of Equation 2.6 where the dependent variable is, respectively, the employment or establishment growth rate for small establishments. Small establishments are defined to be those with less than 20 employees. These estimates use the NETS data, which only extends through 2009. Standard errors clustered on county in parentheses. An observation is a county-by-year cell. Shocks refer to predicted loan originations as specified in Equation 2.4. Baseline controls are 2006 log density, log population, construction share, manufacturing share, and log per capita income. All controls are interacted with year dummies. All main effects are included. See the text for further details. 128 Table A2.4: OLS Models of the Relationship Between Economic Activity and Small Business Loan Origina- tions LBD CBP/QCEW (1) (2) ln(loan originations) (t) 0.0007 0.0010 (0.0010) (0.0009) ln(loan originations) (t-1) -0.0026 -0.0028 (0.0011) (0.0008) Observations 39359 39001 Source: Notes: Entries show OLS estimates of the relationship between small business lending and employment. The dependent variable in Column (1) is small business employment growth. The dependent variable in Column (2) is county-level employment growth. All models include state-by-year fixed effects along with baseline controls (2006 log density, log population, construction share, manufacturing share, and log per capita income) interacted with year dummies. All main effects are included. See text for further details. 129