Local banking supply and private firm activity: Evidence from branch closures

Fang, Francis Haoyu; Vlaicu, Razvan Working Paper Local banking supply and private firm activity: Evidence from branch closures IDB Working Paper Series, No. IDB-WP-1605 Provided in Cooperation with: Inter-American Development Bank (IDB), Washington, DC Suggested Citation: Fang, Francis Haoyu; Vlaicu, Razvan (2024) : Local banking supply and private firm activity: Evidence from branch closures, IDB Working Paper Series, No. IDB-WP-1605, InterAmerican Development Bank (IDB), Washington, DC, https://doi.org/10.18235/0013076 This Version is available at: https://hdl.handle.net/10419/302207 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/3.0/igo/ Local Banking Supply and Private Firm Activity: Evidence from Branch Closures Francis Haoyu Fang Razvan Vlaicu WORKING PAPER No IDB-WP-1605 InterA merican Development Bank Department of Research and Chief Economist July 2024 * University of Minnesota ** Inter-American Development Bank Local Banking Supply and Private Firm Activity: Evidence from Branch Closures Francis Haoyu Fang* Razvan Vlaicu** InterA merican Development Bank Department of Research and Chief Economist July 2024 Cataloging-in-Publication data provided by the Inter-American Development Bank Felipe Herrera Library Fang, Francis Haoyu. Local banking supply and private firm activity: evidence from branch closures / Francis Haoyu Fang, Razvan Vlaicu. p. cm. — (IDB Working Paper Series ; 1605) Includes bibliographical references. 1. Branch banks-Brazil. 2. Banks and banking-Brazil. 3. Community banksBrazil. 4. Business enterprises-Brazil. 5. Unemployment-Brazil. I. Vlaicu, Razvan. II. Inter-American Development Bank. Department of Research and Chief Economist. III. Title. IV. Series. IDB-WP-1605 http://www.iadb.org Copyright © 2024 Inter-American Development Bank ("IDB"). This work is subject to a Creative Commons license CC BY 3.0 IGO (https://creativecommons.org/licenses/by/3.0/igo/legalcode). The terms and conditions indicated in the URL link must be met and the respective recognition must be granted to the IDB. Further to section 8 of the above license, any mediation relating to disputes arising under such license shall be conducted in accordance with the WIPO Mediation Rules. Any dispute related to the use of the works of the IDB that cannot be settled amicably shall be submitted to arbitration pursuant to the United Nations Commission on International Trade Law (UNCITRAL) rules. The use of the IDB's name for any purpose other than for attribution, and the use of IDB's logo shall be subject to a separate written license agreement between the IDB and the user and is not authorized as part of this license. Note that the URL link includes terms and conditions that are an integral part of this license. The opinions expressed in this work are those of the authors and do not necessarily reflect the views of the Inter-American Development Bank, its Board of Directors, or the countries they represent. Abstract Private firms establish relationships with banks in local markets to obtain adequate financing for their operations through credit and loans. As major banks reduced their branch networks in recent years, many firms have lost access to their local bank. We evaluate the impact of a large number of branch closures on firm operations, wages and employment, and economic output in Brazil from 2011 to 2021. We adopt a differencein-differences strategy with staggered treatment timing, employing both two-way fixed effects and Callaway-Sant’Anna estimators. Our study finds that bank branch closures result in a reduction in establishments with active operations from 1.2% initially to 8.1% within 4-7 years, a 0.5 decline in weekly hours of formal employment, and a compression in the real wage distribution. Micro firms, trade and service firms, and agricultural firms are found to be the most vulnerable. Our study highlights the importance of physical bank branches that provide financial access and meet the localized financial demand of several types of firms. JEL classifications: G21, R11, J21 Keywords: Bank branch closures, Employment, Firm activity, Economic impact, Financial access, Brazil F.H. Fang: Department of Applied Economics, University of Minnesota, 1994 Buford Ave, Saint Paul, MN 55108. Email: fang0[email protected]. R. Vlaicu: Research Department, Inter-American Development Bank, 1300 New York Ave NW, Washington, DC 20577. Email: [email protected]. We thank Jason Kerwin, Paul Glewwe, Elizabeth Davis, Stephen Polasky, Joe Ritter, Janna Johnson, Andrew Goodman-Bacon, Richard Thakor, Cynthia Kinnan, Priya Mukherjee, Tamara McGavock, Vinicius Vidigl, Jianxuan Lei, Rachel (Chundan) Lan, Sarah Wahby, Giang Thai, Leo Zucker for research assistance, and participants at the PhD Lab, Glewwe Lab and 2023 second year paper seminar in the Department of Applied Economics. 1 Introduction The importance of local bank access for private firms cannot be overstated, as local banks provide essential financial services, including credit, loans, and payments. Local access is particularly vital for small and medium-sized enterprises (SMEs) that may not have the same resources as larger corporations to secure financing through alternative channels and face higher costs to switching lenders (Beck et al., 2008; Nguyen, 2019). Through personal interactions, local banks obtain soft information about market conditions and business owners that allows them to better tailor financial products to the specific needs of local firms (Berger et al., 2008). Moreover, the proximity of banks to firms can lead to improved credit availability, as banks are more likely to lend to businesses with which they have established relationships (Cole and Gunther, 1998). The presence of banks can significantly influence a firm’s labor demand and productivity, thereby affecting the labor market equilibrium and local economic output. Banks’ presence can have a mixed impact on a firm’s financing cost depending on different factors like bank ownership or market power (Ryan et al., 2014; Barth et al., 2013). Furthermore, studies often find credit availability has a positive impact on labor demand and wages by lowering the firm’s financial frictions and eventually facilitating job creation and productivity (Popov and Rocholl, 2018; Fonseca and Van Doornik, 2022). However, the landscape of bank access is changing. Recent World Bank data2highlights a decrease in commercial bank branches per capita from its peak in 2016 and identifies a global trend towards digital banking and away from traditional brick-and-mortar. This shift should have implications for firms that rely heavily on local banking relationships for their financial needs. The decline in physical bank branches can lead to reduced access to financial services for firms located in rural or underserved urban areas, potentially hampering their survival, operations, and growth. This leads us to the core questions of our study: What is the impact of closing bank branches on local firm operations, employment, and economic output? And what kinds of firms are most vulnerable to experiencing such shifts? To study this question, Brazil provides a suitable context. The number of bank branches experienced an increasing trend after the Brazilian government implemented its “Banks for All” program (“Banco para Todos”) in 2004. However, this upward trend was disrupted in 2014, and a downward trend has persisted since then. As of the end of 2022, Brazil had lost more than 5,000 bank branches, which is one-fourth of its peak number. Many municipalities lost all of their bank branches in this process. This creates quasi-experimental variation in 2See: World Bank, Number of Commercial Bank Branches per 100,000 Adults, at https://data. worldbank.org/indicator/FB.CBK.BRCH.P5 1 local bank access that can be used to study the effects of banking supply on firm activity. Figure 1 presents the evolution of the bank branch landscape in Brazil spanning the last two decades from 2002 to 2022. Our empirical analysis relies on a primary sample of private sector firms and employees in Brazil spanning the years from 2011 to 2021. We exploit establishment-level cross-sectional data over time in addition to municipality-level panel data on economic and employment conditions across Brazil. We link bank branch-level data with firm-level labor census data and municipality-level data on population and local economic accounts. Considering the geographic variation in bank branch concentration, as well as the heterogeneity arising from different closure patterns, we chose to concentrate on municipalities with only one branch. This subset constitutes the largest group among banked municipalities and exhibits broad geographic variation in Brazil. To identify the effects of bank branch closures on local formal employment, it is imperative to establish both a treatment group and a suitable comparison group. We adopt a staggered treatment timing design by retaining in the sample municipalities where the local branch closure was not reversed during the study period. We then utilize a difference-in-differences (DiD) framework with variation in treatment timing to identify the effects of bank branch closures on local employment, exploiting the closures of bank branches as a source of plausible exogenous variation. We discuss potential endogeneity challenges in this context, including issues such as reverse causality, omitted variables, anticipation, and spillover effects. To address these issues, several identification assumptions must be made, including the parallel trends assumption and limited anticipation assumption. Our subsequent estimates will be valid provided these assumptions hold, particularly that treatment-related omitted variables do not produce systematic post-treatment effects in one group. To provide an overview of the impact of bank branch closures and construct robust estimates, we provide three estimation approaches in our analysis: the two-way fixed effect (TWFE) model, the Callaway-Sant’Anna (CS) estimator with never-treated observations as a control group, and the CS estimator with not-yet-treated observations as a control group. Considering that issues of treatment effect heterogeneity and staggered timing may affect the reliability of the the two-way fixed effect estimator, we compared several newly developed staggered DiD estimators and selected CS estimators; those estimators are discussed in recent literature and can overcome the identification issues associated with TWFE. Moreover, the CS estimator has other benefits that apply in our context: It can accommodate less stable parallel trends (often a concern in long study periods) and it provides transparent control group options (never-treated versus not-yet-treated.) We present CS estimates using both types of control groups to address potential selection concerns. 2 Figure 1: Bank Branch Trends in Brazil, 2002-2022 Notes: This graph presents the number of bank branches and municipalities with bank branches that appeared in Brazil’s ESTBAN datasets each December from 2002 to 2022. The red line indicates the trend in the number of bank branches, corresponding to the y-axis on the left side. The blue line indicates the number of municipalities with at least one bank branch, corresponding to the y-axis on the right side. Source: ESTBAN dataset, Banco Central do Brasil. We conduct several analyses to investigate the impact of local bank branch closures. First, we conduct an establishment-level analysis to see if bank branch closures impact firm operations, including market entry and exit. We find that roughly 1 percent of establishments become inactive between 0 and 3 years after a bank branch closure, and the share of inactive establishments increases from 1.2 percent to 8.4 percent between 4 and 7 years after the closure. However, we find no effect on the overall or specific type of firm entering or exiting. We then evaluate the impact on local formal employment, using a municipality-level panel constructed from worker census data. Our estimates suggest that bank branch closures could have a larger and quicker impact on wages than on employment. While statistically imprecise, our point estimates show bank branch closures decrease the average wage by 1.5 to 1.7 percent in the first three years. For employment, our estimates suggest bank branch closures do not have a significant short-term effect but potentially a long-term negative effect. Studying wage disparities, we find short-term decreases in the standard deviation of wages, but long-term decreases in the difference between the 90th and 10th percentile wage brackets. This could suggest that branch closures disproportionately affect different wage brackets—in the short run, potentially the middle-income bracket, and in the long run, the 3 lower wage bracket. We conduct a heterogeneity analysis to characterize which types of firms are more vulnerable; we find that micro firms, trade and service firms, and agriculture firms are disproportionately impacted by local bank branch closures. Finally, we analyze economic output accounts and find a decreasing trend in service output due to bank branch closures but not in industrial and agricultural output. This paper contributes to several strands of literature. First, it adds to the existing literature on the economic effects of banking supply. The effects of bank deregulation have been extensively studied across various periods in different countries. Among the outcomes considered are income inequality, new incorporation growth, real income, output growth, financial inclusion, and household wealth accumulation in the United States (Beck et al., 2010; Black and Strahan, 2002; Jith Jayaratne et al., 1996; Celerier and Matray, 2019); poverty in India (Burgess and Pande, 2005); income levels in Mexico (Bruhn and Love, 2014); and branch occupation, income and employment in Thailand (Ji et al., 2021). In Brazil, studies discuss shocks that could impact bank branch supply and their effects, including bank mergers and acquisitions affecting financial variables, loans, and employment (Joaquim and van Doornik, 2019); national policies for bank branch expansion to rural areas (Fonseca and Matray, 2022); and bank robberies affecting adoption of digital finance tools (Argentieri Mariani et al., 2023). 3 Our study relates to existing work investigating the effects of credit supply on employment, consumption, and productivity. Stiglitz and Weiss (1981) and Kochar (1997) discuss the equilibrium in the loan market, specifically credit rationing, household demand for credit, and its role in agricultural development. Credit market shocks have been found to negatively impact employment growth in small businesses in the United States (Greenstone et al., 2020) and result in decreased consumption, earnings, and employment in India (Breza and Kinnan, 2021). Field experiments have yielded similar results, with Karlan and Zinman (2010) demonstrating the positive outcomes of expanding credit supply on employment, income, food consumption, and overall well-being. Banerjee and Duflo (2014) utilize a targeted direct lending program to test credit constraints among Indian firms and suggest severe constraints and high marginal rates of return on capital. Finally, the finance literature has demonstrated the significance of bank branches in local credit supply. Gilje et al. (2013) demonstrate how bank branch networks help integrate lending markets in the United States, even in the presence of the securitization market. Hasan et al. (2020) study the role of local bank branches and conclude that they serve as irreplaceable lenders for small and medium enterprises. Khwaja and Mian (2008) show that liquidity shocks to banks result in decreased loan amounts to firms and a reduced probability 3See also Meslier et al. (2022) and Bragoli et al. (2022). 4 estimator (Dube et al., 2023). In our study, since different municipalities have bank branch closures in different years, we adopt the CS estimator as an alternative robust estimator as it can provide flexibility in choosing control groups and robust confidence intervals via bootstrapping, according to Baker et al. (2022) and Roth et al. (2023). The CS estimator calculates the treatment effect by choosing different control groups and assigning them weights different from those of the TWFE model. Their estimated parameter can be expressed as follows: ATTnev(g, t;ω) = E[Yt−Yg−ω−1|Gg= 1] −E[Yt−Yg−ω−1|C= 1] (6) ATTny(g, t;ω) = E[Yt−Yg−ω−1|Gg= 1] −E[Yt−Yg−ω−1|Dt+ω= 1] (7) In equations (7) and (8), AT T nev measures the group-time treatment effect with the limited treatment anticipation assumption imposed. It compares the difference between the outcome at time tand the outcome at time g−ω−1 in the treatment group for group Ggand the same difference in the never-treated group, i.e., those municipalities with a single branch open during the study period. ATTny calculates the group-time average effect using the difference within the same group Ggminus the difference in the not-yet-treated samples, i.e., samples that enter the treatment group after period t+ω. Choosing different control groups can have different benefits in the context of our study: the never-treated group can provide a stable comparison regarding the time effects, while the not-yet-treated group helps mitigate selection concerns by comparing groups that all eventually get treated. Furthermore, the CS estimator aggregates ATT(g, t) to AT T (T) using the following equation: ATT(e) = X g∈G I{g+e≤τ}P{G=g|G+e≤τ}ATT(g, g +e) (8) where estands for the time since the treatment and τstands for the end of the study period, 2021. It multiplies the treated share of the sample by the group average treatment effect to recover AT T (e). The CS estimator requires the following key assumptions for identification: (i) limited treatment anticipation—the treated firms or municipalities should have limited anticipation for ω≥0 terms ahead; and (ii) conditional parallel trends based on the “not-yet-treated” or “never-treated” groups—had they not been treated, the treatment group firms or municipalities would have maintained a parallel trend compared with their control group peers. The empirical challenge in estimating the impact of a local bank branch closure is that the strict exogeneity of the treatment can be questioned in two dimensions: reverse causality and omitted variable bias. Reverse causality occurs when the closure of local bank branches 11 follows worsening economic conditions; omitted variable bias represents the existence of exante unobserved factors that affect the treatment groups. For example, this can happen when banks base their closure decisions on predictions of future branch performance. If banks can observe ex ante local factors and use them to form predictions, while we cannot, these factors become omitted variables, biasing our estimates. We base our identification on the parallel trends (PT) assumption. The first part of PT requires that the treatment and control groups maintain a parallel trend. If banks close branches in response to deteriorating local economic conditions, the outcomes for municipalities with closed branches would likely display a downward pre-closure trend, diverging from the control group. For the post-treatment period, PT assumes no omitted variables change the current trend except for the treatment. PT cannot be strictly tested, as we cannot observe an untreated counterpart for each treated firm and municipality. However, we do provide pre-trend analysis to argue PT does hold and note that our estimates will remain unbiased only if PT is satisfied. We also consider issues of anticipation and spillovers in our study and are able to conclude the concerns are minimal. Anticipation refers to the possibility that local businesses may anticipate the closure of local bank branches and adjust their employment patterns in advance. Local bank branch closures are due to bank strategic decisions, and it is reasonable to assume most firms will not have access to inside information about an upcoming branch closure in a specific location. So, we assume no anticipation and set ω= 0. Spillover effects occur when firms that lost access to the local bank branch seek credit in a neighboring municipality, which can potentially affect that municipality’s firms’ performance. Figure 1 shows that treated and comparison groups are geographically separated; it is thus likely that spillovers will be relevant in this context. However, nationwide bank branch closures can still have an impact on both groups, leading to potential underestimates of the impact. 4 Empirical Results We first report results regarding firm operations, using both establishment-level and municipalitylevel data. Then we look at employment variables. Finally, we conduct a heterogeneity analysis by firm type and economic sector. 4.1 Firm Operations To evaluate the impact of losing access to bank branches for firm operations and number of firms, we employ the labor census data that covers the universe of formal firms. The 12 publicly available version of the RAIS dataset contains an “active establishment” variable— an indicator for firms/establishments conducting any economic activity during the year—and the number of employees. 4We adopt the “active establishment” as our primary outcome measure for firm operations. Additionally, since the RAIS dataset itself comes from the annual labor census in Brazil and requires all registered firms to report, we can track the numbers of firms registered in the dataset or proportions of certain types of firms to measure whether firms are entering or leaving the market. Henceforth, we report either the overall effect or the event study effects for the variables of interest, but we sometimes choose to focus on one kind of effect for different variables under discussion. In the main text we display the CS estimates using the never-treated group for event studies and report the event study plots with all three estimators in Appendix B. The estimates are generally very similar, suggesting selection issues are minor, and the PT assumption holds well across different control groups. Figure 3 shows the event study results, namely dynamic regression coefficients for firm operations via the binary variable “active establishment.” The revealed pattern suggests that bank branch closures make roughly 1 percent of establishments become inactive in the following 3 years, with the effect increasing from 1.1 to 1.4 percent between the first and second years after a closure. Our estimates suggest that bank closures could have a severe impact in the long run: The share of inactive establishments increases from 1.2 percent to 8.1 percent between the 4th and 7th years after a closure. Roth (2024) indicates the default way to calculate the pre-treatment effect with the CS estimator is to use short gaps and display violations of PT as kinks in pre-treatment estimates. We follow this approach and calculate the pre-treatment effect using long gaps instead, thus making the pre-treatment estimates comparable with the dynamic TWFE model and visual checks for pre-treatment effects possible. We find no evidence of an existing pre-trend. While some firms may respond to shrinking credit supply by scaling back operations, others may close down. At the same time, lack of local bank access can create a barrier to entry for new firms. In the absence of data on firm exit and entry, we report estimates using the number of firms within the municipality by aggregating establishment-level data to the municipality level. Furthermore, we calculate the proportions of SMEs, agricultural firms, and non-zero-employee firms. We report an overall ATT effect in Table 1. In general, in these data we do not find evidence that a bank branch closure leads to a change in the total number of local firms or 4Over 60 percent of establishments report zero active employees at the end of the year in our subsample. For this reason, firm-level employee numbers are not a reliable measure of firm activities. 13 Figure 3: Effect of Exposure to Bank Branch Closure on Firm Operations at the Local Level Note: This figure shows estimates from equation (7) regarding the binary “active establishment” variable, using CS estimators with never-treated municipalities as the control group. The bar shows the 95% confidence interval. Years since bank branch closure equals 0 in the year the municipality loses its sole bank branch. We report pre-treatment trends starting from period -3. The panel is balanced between periods -3 and 0; other periods are estimated using incomplete samples. The standard errors are computed by wild bootstrap. We use the “long2” option under “csdid” in Stata according to Roth’s (2014) suggestion, making the pre-treatment estimates comparable with the dynamic TWFE model and visual checks for pre-treatment effects possible. We use establishment-level data to conduct this analysis. Source: RAIS, authors’ calculations. a compositional change in the types of firms. In column (1), the static TWFE estimates indicate a slightly positive or small effect on the number of firms and shares of firms. However, by investigating with the more robust CS estimators and analyzing dynamic patterns (not reported), we observe a lack of consistent evidence of significant impact. 4.2 Employment Through firms, bank branch closures could have an impact on local employment. In this section, we evaluate the impact of bank branch closures on formal sector employment. In conducting our analysis, we find that some municipalities have very few paid formal workers registered. To avoid boundary effects, we reframe the employment analysis around those municipalities where paid private sector employees totaled more than one percent of the municipal population during our study period. It appears that municipalities with low 14 shares of paid workers are randomly distributed and not affected by the treatment. This exclusion rules out similar numbers of observations across groups: 43 (6.3%) of control group municipalities and 36 (8.3%) of treatment group municipalities. Table 1: Estimates of the Impact of Bank Branch Closure on Firms N (Firms) P(SME) P(Ag Firms) P(Nonzero Firms) (1) (2) (3) (4) Panel A. TWFE δtwfe 5.057* -0.000 0.002 0.000 (2.390) (0.000) (0.002) (0.003) Panel B. CS-Never δcsnev 2.411 -0.000 0.001 0.001 (7.927) (0.000) (0.011) (0.007) Panel C. CS-Notyet δcsny 1.883 0.000 0.001 0.001 (7.971) (0.000) (0.011) (0.008) Baseline Mean 281.285 0.996 0.176 0.357 Observations 12,199 12,199 12,199 12,199 Note: This table shows overall ATT estimates for the treatment effects on numbers and shares of firms from equations (6) and (9). The standard errors are reported in parentheses. The standard errors for static TWFE estimates are clustered at both the municipality and year levels. The standard errors for the CS estimates are computed by wild bootstrap. The significance level in this table is represented by: *** p<0.01, ** p<0.05, * p<0.10. The baseline mean represents the mean of the outcome variables for all observations in the sample during 2011-2013. We use a municipality-level panel constructed from establishment-level data to conduct this analysis. Source: Authors’ calculations Two key dimensions for understanding the equilibrium effects of bank branch closures on employment are quantity and price. We calculated the natural log of numbers of employed workers and the natural log of worker real wages. We find that employment and wages display significant variation—they react to a broad range factors regarding workers and their working environment—so adjusting for covariates can improve estimation precision. We thus include workers’ demographic profiles within municipalities as covariates, along with job and establishment information. We include the following variables: female share of workers, average education level (11 levels, adapting the scale from the RAIS dataset), average weekly hired hours, average establishment size in groups, time in employment, and whites as a share of workers. Figure 4 Panel A shows the results from this analysis. The point estimates are relatively robust to the inclusion of covariates, which narrow the confidence intervals (see Figure 12 for comparison). However, the confidence intervals do not reach statistical significance at the 95% level. The impact of bank branch closure on employment can be read as either a null or 15 Panel A. Local Level of Employment and Real Wage (a) (b) Panel B. Wage Disparities (c) (d) Figure 4: Effect of Exposure to Bank Branch Closure on Local Levels of Employment and Wages Note: This figure shows estimates from equation (7) regarding the Log(Number of Employed People), Log(Real Wages), SD (Real Wages), and P90-10 Real Wages. The bar shows the 95% confidence interval. Years since bank branch closure equals 0 in the year the municipality loses its sole bank branch. We report pre-treatment trends starting from period -3. The panel is balanced between periods -3 and 0; other periods are estimated using incomplete samples. The standard errors are computed by wild bootstrap. We use the “long2” option under “csdid” in Stata according to Roth’s (2014) suggestion, making the pre-treatment estimates comparable with dynamic TWFE model and visual checks for pre-treatment effects possible. We use municipality-level data to conduct this analysis. We include the following covariates to narrow down the confidence intervals: female share of workers, average education level, average weekly hired hours, average establishment size in groups, time in employment, and whites as a share of workers. Real wages are calculated by adjusting nominal wages by Brazil’s annual CPI. The unit for real wages is 2011 reais. Source: RAIS, authors’ calculations. 16 negative effect. For the number of employed workers, the CS estimator with never-treated as controls generally shows a potential long-term decreasing trend despite its point estimates being statistically insignificant. For the average wage, it shows a long-term decrease of 1.5 to 1.7 percent and an inconclusive short-term effect. In this analysis, both CS estimators show a similar trend, while the TWFE estimator has a slightly different pattern. The TWFE suggests a null effect on the number of employed workers and a small positive effect on wages. We do not find evidence of violation of PT. To better understand the impact of bank branch closures on local employment dynamics, we delve into wage disparities. We employ two distinct measures to assess the effects on wage disparities: the standard deviation of real wages, to gauge overall wage variability among workers, and the difference between the 90th and 10th percentiles of real wages, to examine disparities between highand low-paid jobs. We present our estimates in Figure 4 Panel B. For the standard deviation of real wages, we note a short-term decrease of 142 reais one year after a closure, persisting for two years before gradually diminishing over approximately five years. In contrast, the gap between the 90th and 10th percentile wages initially shows no significant short-term effect. This hints at a potential long-term widening, beginning three years post-closure, with an increase of 28 reais, and reaching a difference of 231 reais six years after a closure. These observations suggest that in the short term, bank branch closures do not significantly affect the wage gap between highand low-paid jobs, but they do lead to a notable reduction in overall wage disparities among workers, potentially through the middle-income wage bracket. In the long term, however, the loss of a bank branch appears to contribute to increasing wage disparities between highand low-paid positions, even as overall wage disparities among workers become less pronounced. Finally, we are interested in determining whether bank branch closures impact firms’ employment patterns in other ways. For instance, we examine whether firms maintain the same number of employees but reduce their working hours, or if part-time positions are being used as substitutes for full-time jobs. Our analysis utilizes weekly contracted hours and the proportion of part-time work as metrics. The results are reported in Figure 5. We observe a decrease in contracted hours 5 to 6 years following a branch closure. However, upon examining the event-study plot for part-time work metrics, we do not find a clear impact due to bank branch closures. Finally, we study whether local bank branch closures have an impact on economic output. If bank branch closures lead to reduced firm operations and employment levels, it is possible that impacted firms would produce less output. Here we rely on municipality-level GDP data and value-added accounts in services (excluding public administration, education, etc.), 17 (a) (b) Figure 5: Effect of Exposure to Bank Branch Closure on Other Local Employment Patterns Note: This figure shows estimates from equation (7) regarding weekly hired hours and a binary variable of part-time work. The bar shows the 95% confidence interval. Years since bank branch closure equals 0 in the year the municipality loses its sole bank branch. We report pre-treatment trends starting from period -3. The panel is balanced between periods -3 and 0; other periods are estimated using incomplete samples. The standard errors are computed by wild bootstrap. We use the “long2” option under “csdid” in Stata according to Roth’s (2014) suggestion, making the pre-treatment estimates comparable with the dynamic TWFE model and visual checks for pre-treatment effects possible. We use municipality-level data to conduct this analysis. We include the following covariates to narrow down the confidence intervals: female share of workers, average education level, average weekly hired hours, average establishment size in groups, time in employment, and whites as a share of workers. Source: RAIS, authors’ calculations. industry, and agriculture to conduct our analysis. We adjust prices to 2011 reais and present the effect pattern in Figure 10. In services, we find a short-term decreasing trend in the first 3 years after closure, although the estimates are imprecise. We estimate that bank branch closures decrease valued added in services by 3,366 thousand reais by the 3rd year. In industry and agriculture, we do not find significant short-term patterns but a potential long-term decreasing trend. 4.3 Heterogeneity Analysis In this section, we evaluate how the impact of bank branch closures differs by firm size and economic sector. Knowing the average effect of closing bank branches on firm operations, it will be valuable to learn what kind of firms are more vulnerable to these effects. We consider two dimensions: firm size and main economic sector. There are different ways to categorize firms by number of employees in Brazil (Carneiro et al., 2020). IBGE classifies the size of firms in certain sectors using the number of employees. We use IBGE’s employee count-based firm size categorization for firms/establishments in manufacturing and services (see Table 4). In the 18 absence of a settled classification for the remaining sectors, such as agriculture, we adopt Cravo et al. (2018)’s classification scheme. 5 We classify firms into four main sectors: industry and construction, referred to as manufacturing by Carneiro et al. (2020); trade and services, referred to as services; agriculture; and all other sectors. Table 2: Estimates of the Effect of Bank Branch Closure on Firm Operations, by Firm Size Outcome Variable Active Establishments Micro Firms Small Firms Medium Firms Large Firms (1) (2) (3) (4) Panel A. TWFE δtwfe -0.009 -0.000 -0.001 0.002 (0.005) (0.000) (0.002) (0.002) Panel B. CS-Never δcsnev -0.014*** -0.000 -0.000 0.000 (0.002) (0.000) (0.000) (0.000) Panel C. CS-Notyet δcsny -0.012*** -0.000 -0.000 -0.000 (0.002) (0.000) (0.000) (0.000) Baseline Mean 0.785 0.996 0.997 0.996 Observations 3,158,958 208,168 26,725 5,189 Note: This table presents the impact of a bank branch closure on firm operations by firm size, using equations (5), (7) and (8). The standard errors are reported in parentheses. The standard errors for the static TWFE estimates are clustered at both the municipality and year level. The standard errors for the CS estimates are computed by wild bootstrap. The significance level in this table is represented by: *** p<0.01, ** p<0.05, * p<0.10. The baseline mean represents the mean of the outcome variables for all observations in the sample during 2011-2013. We use establishment-level data to conduct this analysis. Observations denote the number of firms per size category. Source: Authors’ calculations We first study the effect of a bank branch closure on firm operations across different firm sizes. Table 2 presents the results. We find that micro firms bear almost the full impact on operations from local bank branch closures, while the effects are roughly zero for firms of all other sizes. This could be due to their high level of active operation status: small, medium, and large firms are all more than 99.6% active in the baseline years (between 2011 5An alternative approach for firms in agriculture and all other sectors is to apply IBGE’s categorization for manufacturing firms. The results are qualitatively similar. 19 to 2013). The results confirm our previous hypothesis that larger firms have greater financial access and are less likely to be disrupted by a local bank branch closure. Micro firms play a dominant role (92.9%) in our sample and are sensitive to local branch closures, which could imply that micro firms’ banking demands are more localized. We include an event-study plot of this effect in Appendix B (see Figure 14). The effect is similar to the impact pattern for our entire sample (see Figure 11). Table 3: Estimates of the Effect of Bank Branch Closure on Firm Operations, by Sector Outcome Variable Active Establishments Trade & Service Industry & Construction Agriculture All Other (1) (2) (3) (4) Panel A. TWFE δtwfe -0.008 -0.003 -0.013 0.017 (0.005) (0.006) (0.008) (0.052) Panel B. CS-Never δcsnev -0.014*** 0.002 -0.018*** 0.000 (0.002) (0.005) (0.003) (0.053) Panel C. CS-Notyet δcsny -0.013*** 0.003 -0.016*** 0.000 (0.002) (0.005) (0.003) (0.054) Baseline Mean 0.768 0.790 0.914 0.858 Observations 2,315,624 415,022 665,012 3,757 Note: This table presents the impact of a bank branch closure on firm operations by firm sector, using equations (5), (7) and (8). The sectors are: trade and services, industry and construction, agriculture, and all other firms. The standard errors are reported in parentheses. The standard errors for the static TWFE estimates are clustered at both the municipality and year level. The standard errors for the CS estimates are computed by wild bootstrap. The significance level in this table is represented by: *** p<0.01, ** p<0.05, * p<0.10. The baseline mean represents the mean of the outcome variables for all observations in the sample during 2011-2013. We use establishment-level data to conduct this analysis. Observations denote the number of firms per sector. Source: Authors’ calculations We now turn to the impact of bank branch closures on firm operations by sectors. The results are presented in Table 3. While the TWFE estimates suggest null effects across different sectors, the CS estimates imply that bank branch closures depress the operations of firms in trade and services as well as in agriculture. Overall effects on agriculture firms are somewhat more severe than on trade and service firms despite agriculture firms’ higher rate of active establishment status in the baseline year. We present event-study plots for the trade and service sector as well as the agriculture sector in Figure 6. 20 Table 5: Summary Statistics for Sample Variables Across Groups Never-Treated Group Treatment Group Difference Mean N Mean N (1) (2) (3) (4) (5) Establishment-Level Data Active Establishment 0.939 2,337,785 0.926 1,138,674 0.013 *** (1.146) (1.115) (0.001) Establishment Size (Groups) 1.585 2,337,785 1.573 1,138,674 0.012 *** (0.965) (0.969) (0.001) Municipality-Level Data Number of Establishments 307.770 7,436 233.222 4,763 74.548 *** (176.866) (128.248) (2.963) P(SME) 0.995 7,436 0.996 4,763 0.000 ** (0.007) (0.008) (0.000) P(Ag Firm) 0.175 7,436 0.196 4,763 -0.021 *** (0.163) (0.198) (0.003) Log(Real Wages) 6.812 6,963 6.791 4,367 0.021 *** (0.172) (0.170) (0.003) SD (Real Wages) 845.087 6,963 795.49 4,367 49.597 *** (434.860) (410.458) (8.216) Real Wages P90 1591.892 6,963 1514.931 4,367 76.961 *** (496.044) (525.198) (9.796) Real Wages P10 627.451 6,963 628.28 4,367 -0.83 (69.203) (71.615) (1.354) P(Female) 0.309 6,963 0.302 4,367 0.007 * (0.096) (0.101) (0.002) Time Employed (Months) 35.384 6,963 34.891 4,367 0.493 * (11.027) (10.803) (0.211) Weekly Hired Hours 42.913 6,963 42.974 4,367 -0.061 *** (0.957) (0.899) (0.018) Age 33.789 6,963 33.868 4,367 -0.080 * (1.828) (2.011) (0.037) Average EstSize 3.672 6,963 3.593 4,367 0.080 *** (0.910) (0.939) (0.018) P(Part-Time Work) 0.002 6,963 0.002 4,367 0.000 (0.008) (0.009) 0.000 Education Level 5.922 6,963 5.87 4,367 0.052 *** (0.601) (0.664) (0.012) Number of Employed People 1017.736 6,963 736.796 4,367 280.94 *** (886.796) (669.004) (15.632) Race: White 0.522 6,963 0.533 4,367 -0.011 * (0.267) (0.264) (0.005) Population 10237.16 6,963 8624.363 4,367 1612.797 *** (6884.674) (6277.038) 0.000 Valued Added: Ag 30494.07 7,436 21364.62 4,763 9129.45 *** (49670.360) (27651.030) (787.917) Value Added: Ind 21514.61 7,436 13793.65 4,763 7720.96 *** (59695.790) (35852.060) (959.715) Value Added: Services 35166.9 7,436 25079.69 4,763 10087.21 *** (35294.550) (29811.660) (617.299) 27 [Continue for table notes] Note: This table presents summary statistics for all the variables in our sample. Firm size is categorized into 9 groups in terms of number of employees. Firm sector categories are industry and construction, trade and services, and all other sectors. Real wages are deflated using the annual consumer price index; their units are 2011 Brazilian reais. Education level consists of 11 groups according to categories in the RAIS dataset. The partial work status is only available from 2017. Value added is deflated using the annual consumer price index; its units are thousands of 2011 Brazilian reais. The significance level in this table is represented by: ∗p < 0.1∗ ∗p < 0.5,∗∗∗p < 0.01. Source: Authors’ calculations Table 6: Population and GDP Comparisons Across Groups Year Never-Treated Treatment Other Mun (1) (2) (3) Population 2011-2013 10688.15 9037.55 41421.61 (159.66) (177.69) (2002.94) 2014-2017 11075.77 9295.07 43427.09 (144.78) (159.00) (1805.63) 2018-2021 11138.72 9268.61 44695.30 (149.08) (160.02) (1854.50) Real GDP 2011-2013 117566.50 85692.04 976888.10 (2300.79) (1808.36) (79505.82) 2014-2017 123471.40 91513.17 965289.10 (2159.20) (1650.16) (66289.58) 2018-2021 135178.90 96514.38 972079.40 (2518.69) (1889.33) (61537.83) RGDP per capita 2011-2013 13.99 13.12 14.92 (0.29) (0.35) (0.16) 2014-2017 14.10 13.31 14.82 (0.25) (0.30) (0.12) 2018-2021 15.20 14.02 15.95 (0.27) (0.33) (0.14) Note: This table presents the average municipal population, real GDP, and real GDP per capita of each sample and time period. The simple average is presented and the standard deviations within groups are provided in parentheses. The unit of population is number of individuals, and the unit for real GDP and real GDP per capita are thousands of reais. Real GDP is deflated using Brazil’s GDP deflator from FRED and IMF; the base year is 2011. Municipal populations are IBGE estimates, not counts. RGDP per capita is calculated by the authors using the estimated population. Source: IBGE, authors’ calculations 28 Table 7: Involved Banks: Treatment and Never-Treated Group BANK Treatment Never-Treated Total Public-Owned BCO BANESTES S.A. 0 18 18 X BCO BRADESCO S.A. 9 157 166 BCO DO BRASIL S.A. 332 271 603 X BCO DO EST. DE SE S.A. 1 12 13 X BCO DO EST. DO PA S.A. 0 1 1 X BCO DO ESTADO DO RS S.A. 0 57 57 X BCO DO NORDESTE DO BRASIL S.A. 0 1 1 X BCO SANTANDER (BRASIL) S.A. 29 51 80 CAIXA ECONOMICA FEDERAL 0 11 11 X ITA`o UNIBANCO S.A. 62 97 159 Total 433 676 1,109 703 Note: This table presents the banks and the number of their branches in the treatment and control groups. The right column indicates whether a bank is publicly owned (i.e., owned by the state or federal government). We include the “Total” row and column to indicate the number of bank branches in each group and each bank. Source: Authors’ calculations Table 8: Patterns of Bank Branch Closures by Year Year Freq. Percent Cum. 2014 29 2.61 2.61 2015 21 1.89 4.5 2016 27 2.43 6.93 2017 155 13.98 20.91 2018 34 3.07 23.98 2019 104 9.38 33.36 2020 2 0.18 33.54 2021 61 5.5 39.04 Not Lost 676 60.96 100 Total 1109 Note: This table shows the pattern of bank branch closures in our sample. Our sample consists of municipalities with one bank branch from 2011 to 2013. We exclude municipalities whose branch count doesn’t fall into the absorbing stage from 2014. A branch is considered lost when the only branch closes and doesn’t reopen till the end of the study period. ”Not Lost” municipalities are those that retain one bank branch from 2011 to 2021. The frequency in the timing of branch loss is shown for our sample, including their shares and cumulative percentage. Source: Authors’ calculations 29 Figure 8: Distribution of Municipalities by Number of Bank Branches, 2011-2021 Note: This figure shows the distribution of municipalities by number of bank branches located in each municipality in each year. The colors for each category are marked in the right legend. This figure only presents the distributions for banked municipalities. During the study period, Brazil had 5,570 municipalities in total. Source: Authors’ calculations, ESTBAN dataset. 30 Figure 9: Effect of Exposure to Bank Branch Closure on Firm Operations: Micro Firms Note: This figure shows estimates from equation (7) regarding the “active establishment” variable among micro firms. The bar shows the 95% confidence interval. Years since bank branch closure equals 0 in the year the municipality loses its sole bank branch. We report pre-treatment trends starting from period -3. The panel is balanced between periods -3 and 0; other periods are estimated using incomplete samples. The standard errors are computed by wild bootstrap. We use the “long2” option under “csdid” in Stata according to Roth’s (2014) suggestion, making the pre-treatment estimates comparable with the dynamic TWFE model and visual checks for pre-treatment effects possible. We use municipality-level data to conduct this analysis. Source: RAIS, authors’ calculations. 31 (a) (b) (c) Figure 10: Effect of Exposure to Bank Branch Closure on Economic Output Note: This figure shows estimates from equation (7) regarding value added in services (excluding administration, education, public health, and social security), industry, and agriculture. The bar shows the 95% confidence interval. Years since bank branch closure equals 0 in the year the municipality loses its sole bank branch. We report pre-treatment trends starting from period -3. The panel is balanced between periods -3 and 0; other periods are estimated using incomplete samples. The standard errors for the dynamic TWFE estimates are clustered at both the municipality and year level. The standard errors for the CS estimates are computed by wild bootstrap. Prices are deflated using the reference year’s average consumer price index. The unit of measurement is thousands of 2011 Brazilian reais. Source: Brazil’s GDP by Municipality, authors’ calculations. 32 Appendix B. Event-Study Plots with Three Estimators In this section, we provide event-study plots with results from all three estimators for equations (5), (7) and (8). The estimators are dynamic TWFE, CS-never-treated, CS-not-yet-treated, respectively. Figure 11: Effect of Exposure to Bank Branch Closure on Firm Operations at the Local Level, Three Estimators Note: This figure shows estimates from equations (5), (7) and (8) regarding the “active establishment” variable. The bar shows the 95% confidence interval. Years since bank branch closure equals 0 in the year the municipality loses its sole bank branch. We report pre-treatment trends starting from period -3. The panel is balanced between periods -3 and 0; other periods are estimated using incomplete samples. The standard errors for the TWFE estimates are robust standard errors clustered by both year and municipality. The standard errors for the CS estimates are computed by wild bootstrap. Source: RAIS, authors’ calculations. 33 Panel A. Local Level of Employment and Real Wage (a) (b) Panel B. Wage Disparities (c) (d) Figure 12: Effect of Exposure to Bank Branch Closure on Local Levels of Employment and Wages, Three Estimators Note: This figure shows estimates from equations (5), (7) and (8) regarding the variables Log(Number of Employed People), Log(Real Wages), SD (Real Wages), and P90-10 Real Wages. The bar shows the 95% confidence interval. Years since bank branch closure equals 0 in the year the municipality loses its sole bank branch. We report pre-treatment trends starting from period -3. The panel is balanced between periods -3 and 0; other periods are estimated using incomplete samples. The standard errors for the TWFE estimates are robust standard errors clustered by both year and municipality. The standard errors for the CS estimates are computed by wild bootstrap. Source: RAIS, authors’ calculations. 34 (a) (b) Figure 13: Effect of Exposure to Bank Branch Closure on Other Local Employment Patterns, Three Estimators Note: This figure shows estimates from equations (5), (7) and (8) regarding weekly hired hours and a binary variable of part-time work. The bar shows the 95% confidence interval. Years since bank branch closure equals 0 in the year the municipality loses its sole bank branch. We report pre-treatment trends starting from period -3. The panel is balanced between periods -3 and 0; other periods are estimated using incomplete samples. The standard errors for the TWFE estimates are robust standard errors clustered by both year and municipality. The standard errors for the CS estimates are computed by wild bootstrap. Source: RAIS, authors’ calculations. 35 Figure 14: Effect of Exposure to Bank Branch Closure on Firm Operations: Micro Firms, Three Estimators Note: This figure shows estimates from equations (5), (7) and (8) regarding the “active establishment” variable among micro firms. The bar shows the 95% confidence interval. Years since bank branch closure equals 0 in the year the municipality loses its sole bank branch. We report pre-treatment trends starting from period -3. The panel is balanced between periods -3 and 0; other periods are estimated using incomplete samples. The standard errors for the TWFE estimates are robust standard errors clustered by both year and municipality. The standard errors for the CS estimates are computed by wild bootstrap. Source: RAIS, authors’ calculations. 36