A. The Banking Sector in Brazil and M&A Activity Involving Large Private Banks

We consider M&As of large private banks that took place in the late 2000s. Prior to these M&As, the banking sector in Brazil was already dominated by a small number of private and government-owned banks (Cortes and Marcondes (Reference Cortes, Marcondes, Amann, Azzoni and Baer2018)). In December 2006, the five largest commercial banks accounted for 57.8% of total assets held by financial institutions, 58.6% of total credit, and 63.8% of total deposits.Footnote ⁵ Government-owned commercial banks were relevant in terms of size. The two largest government-owned commercial banks accounted for 28% of total assets held by financial institutions. The largest banks had an extensive branch network, with most of them having more than 1,000 branches and some of them having more than 4,000 branches.

The banking industry experienced the consolidation of large private banks starting in 2007. Four private commercial banks, which accounted for 30.8% of the total credit granted by financial institutions in 2006, took part in M&A transactions. The share of total credit granted by the target banks alone was 13.5%. The first transaction happened in 2007 and was the result of a consolidation that involved two parent companies headquartered in other countries. The second occurred in 2008 and involved large national banks headquartered in Brazil. The second transaction was a consolidation that had been negotiated for years and was accelerated by the Lehman Brothers collapse in September 2008.Footnote ⁶ These two M&As sparked increases in concentration measures, which remained stable thereafter (Figure A1 in the Supplementary Material).

The great financial crisis primarily defined the timing of these consolidations. In Brazil, the main effects of the crisis manifested themselves after the Lehman Brothers bankruptcy and were short-lived in comparison to other countries (De Mello and Garcia (Reference De Mello and Garcia2012)). GDP only fell for two quarters (the last quarter of 2008 and the first quarter of 2009) and then started to grow again (Figure A2 in the Supplementary Material). The stock of credit to firms was stable for two quarters and then resumed its growth trajectory. Banks did not fail, in part because of measures taken after the Lehman collapse by the Central Bank to alleviate the lack of liquidity (Mesquita and Torós (Reference Mesquita and Torós2010)). Moreover, large banks—the ones we study—were perceived as too-big-to-fail, and, as a result, the liquidity drought sparked by the crisis was less strict for them (Oliveira, Schiozer, and Barros (Reference Oliveira, Schiozer and C. Barros2015)).

Table 1 shows characteristics of the M&A episodes considered in our analyses.Footnote ⁷ We focus on large private banks with extensive branch networks and operations in multiple local markets. These characteristics are essential for addressing our research question, as they enable us to account for variations in local market power and study the internal reallocation of resources. To mitigate concerns about selection into consolidation, the control group consists of large private banks that participated in an M&A in 2015—after the period of our analysis, which ranges from 2004 to 2014. We do not include government-owned banks in our sample because there is ample evidence suggesting their objectives differ significantly from those of private banks in Brazil (Coelho, De Mello, and Rezende (Reference Coelho, De Mello and Rezende2013), Coleman and Feler (Reference Coleman and Feler2015), Cortes, Silva, van Doornik et al. (Reference Cortes, Silva and van Doornik2019), Mariani (Reference Mariani2020), and Garber, Mian, Ponticelli, and Sufi (Reference Garber, Mian, Ponticelli and Sufi2021)). The banks in our sample represent 48% of credit, 44% of assets, 45% of deposits, and 51% of bank branches in the financial system. All banks had a presence in hundreds of municipalities, which is our definition of a local market and follows previous research in our setting (Sanches, Silva Junior, and Srisuma (Reference Sanches, Junior and Srisuma2018), Joaquim et al. (Reference Joaquim, van Doornik and Ornelas2019)). Figure 1 depicts the geographical location of the branches. It shows that the banks in our sample had a significant presence in many municipalities throughout the country.

TABLE 1M&As Involving Large Private Banks, 2007–2015

TABLE 1 Long description

The table is organized into four main columns under the heading Banks in the Sample.

Column 1: Bank 1 – Bank 2 (Treated). M and A announcement: October 2007. Market share assets: 12%. Market share credit: 12%. Market share demand deposits: 10%. Market share savings deposits: 7%. Market share of time deposits: 16%. Market share deposits: 12%. Number of branches: 2,168. Number of municipalities: 747.

Column 2: Bank 3 – Bank 4 (Treated). M and A announcement: November 2008. Market share assets: 17%. Market share credit: 18%. Market share demand deposits: 18%. Market share savings deposits: 16%. Market share of time deposits: 11%. Market share deposits: 13%. Number of branches: 3,584. Number of municipalities: 1,128.

Column 3: Bank 5 – Bank 6 (Control). M and A announcement: August 2016. Market share assets: 15%. Market share credit: 17%. Market share demand deposits: 21%. Market share savings deposits: 16%. Market share of time deposits: 18%. Market share deposits: 16%. Number of branches: 3,975. Number of municipalities: 1,527.

Column 4: Other Banks. Market share assets: 56%. Market share credit: 52%. Market share demand deposits: 51%. Market share savings deposits: 61%. Market share of time deposits: 55%. Market share deposits: 59%. Number of branches: 9,322. Number of municipalities: 2,918.

FIGURE 1Local Private Market Share in 2006

Figure 1 shows the geographical presence of the branches that comprise our sample as of December 2006.

FIGURE 1 Long description

A three-panel geographical map of Brazil.

1. Left Panel: Titled Target Banks. Data points are primarily concentrated in the Southeast and South regions, with sparse clusters in the Northeast and along the Amazon River in the Northwest.

2. Middle Panel: Titled Acquirer Banks. This map shows a higher density of data points compared to the first, with a heavy concentration in the Southeast and South, and moderate presence in the Central-West and Northeast regions.

3. Right Panel: Titled Control Banks. This map displays the highest density of data points across all three panels. While the Southeast and South remain the most densely populated, there is significant coverage extending into the Central-West, Northeast, and North regions, including a large dark blue area in the North.

A shared legend at the bottom uses five shades of blue to represent percentage ranges:

- Darkest blue: 75–100%

- Medium-dark blue: 50–75%

- Medium blue: 30–50%

- Light blue: 10–30%

- Palest blue: 0–10%

B. Data

Our analysis draws on microdata from several sources. The first is the RAIS (Relação Anual de Informações Sociais), an annual employer–employee matched data set provided by the Brazilian Ministry of Labor. It contains establishment-level labor data for the universe of firms and workers in the formal sector, covering the full population of bank employees. It includes information on job position characteristics, such as salary and occupation, as well as workers’ characteristics, including age, education, gender, and race. The information on the salary and starting and separation dates allows us to compute the total labor costs of each branch. As both branches and employees have a unique taxpayer identifier, we can follow workers across branches.

Branch- and bank-level financial data come from the Central Bank of Brazil. The branch-level yearly information comes from the ESTBAN (Estatística Bancária Mensal) database. We focus on the period between 2004 and 2014.Footnote ⁸ This data set is derived from a mandatory form that all banks are required to complete, and it includes the branch municipality, the taxpayer identifier, balance sheet information (such as assets, loans, and deposits), and aggregate revenues, costs, and profits. We provide additional details about these data and the variables we construct in Section A3 of the Supplementary Material.

The branch taxpayer identifier is a 14-digit number, out of which the first eight digits are the taxpayer identifier of the bank. This identifier enables us to merge the labor data with branch- and bank-level data. It also allows us to identify the branches that participated in the M&As that occurred in 2007 and 2008 (treated branches) and the ones that participated in the large M&A that occurred in 2015 (control branches). We classify as target branches those that changed their taxpayer number after the M&As.

We obtain detailed branch address information from the Central Bank of Brazil’s Register of Branches, Offices, and Subsidiaries of Consortium Administrators. The data contain the bank branch taxpayer identifier and detailed address information, which is crucial to our article in that it allows us to keep following a branch even when its taxpayer identifier changes because of the M&A. Finally, we utilize bank-level information from IF.data, provided by the Central Bank of Brazil. These data are available on a quarterly frequency and include detailed balance sheet and income statement information. This data set also contains the 8-digit taxpayer identifier of each bank.

C. Descriptive Statistics

In Table 2, we present descriptive statistics of the branches of banks in the treated and control groups before the M&As.Footnote ⁹ Control and treated banks had 9,341 active branches in December 2006, which we define as branches that have at least one employee and a positive stock of credit and deposits. The average stock of lending and deposits was 23 million Brazilian reais (BRL) and 24 million BRL in December 2006, respectively. Branches had on average 16 workers (median 11). The average market share of a branch in the lending market is 10% (median 1.2%) and 12% in the deposit market (median 1.5%). The average profit per employee is 0.11 million BRL.

TABLE 2Branch-Level Descriptive Statistics

TABLE 2 Long description

The table is divided into two sections: Panel A (Unrestricted Sample, N = 9,341) and Panel B (Excluding the 50 Largest Branches, N = 9,291).

Columns are numbered 1 through 7 and represent:

1. N (sample size)

2. Mean

3. Std. Dev. (Standard Deviation)

4. p1 (1st percentile)

5. p50 (median)

6. p99 (99th percentile)

7. p50/Mean ratio.

Key data from Panel A:

- Assets: Mean 285, Median 31, p99 598.

- Loans: Mean 23, Median 5, p99 176.

- Ratio of loans to assets: Mean 0.23, Median 0.18.

- Number of workers: Mean 16, Median 11, p99 82.

- Profit per employee: Mean 0.11, Median 0.11, p1 minus 0.07.

- Loans local market share: Mean 10, Median 1, p99 100.

Key data from Panel B (Excluding 50 largest branches):

- Assets: Mean 55, Median 31, p99 464.

- Loans: Mean 12, Median 5, p99 129.

- Ratio of loans to assets: Mean 0.23, Median 0.18 (identical to Panel A).

- Number of workers: Mean 15, Median 11, p99 77.

- Profit per employee: Mean 0.13, Median 0.11, p1 minus 0.06.

- Loans local market share: Mean 10, Median 1, p99 100.

Several variables exhibit substantial right skewness.Footnote ¹⁰ In particular, the mean-to-median ratio is 9.1 for assets, 4.4 for the stock of credit, and 2.1 for the stock of deposits. To provide a clearer picture of the average branch, we also report statistics excluding the 50 largest branches (0.5% of the sample) in Panel B of Table 2.Footnote ¹¹ In this restricted sample, the average stock of lending falls to 12 million BRL, the average profit per employee is 0.13 million BRL, the average stock of credit per employee is 0.74 million BRL, and the average stock of deposits per employee is 1.33 million BRL. The average ratio of loans to assets is 0.23.Footnote ¹²

As we use municipality-by-time fixed effects in our regressions, municipalities that do not contain both treatment and control branches end up being dropped. When we restrict the sample to branches located in municipalities with both control and treatment branches, the characteristics of the branches do not change significantly (Table A1 in the Supplementary Material). Branches are slightly larger in terms of lending, deposits, and the number of employees, and have smaller market shares.

In Table 3, we report summary statistics by branch status—acquirer, target, and control. Although equality of means is rejected for the vast majority of comparisons, the economic magnitudes of many differences remain relatively modest. For example, after excluding the 50 largest branches, acquirer and target branches each hold an average of 17 million BRL in deposits, compared with 23 million BRL in control branches. In this subsample, control branches exhibit lower profits per employee: profit per employee averages 0.17 million BRL in acquirer branches, 0.14 million BRL in target branches, and 0.08 million BRL in control branches. Control branches, however, have larger deposits per employee—1.5 million BRL per worker, compared with 1.05 million BRL in target branches and 1.31 million BRL in acquirer branches. Lending is also larger in target branches relative to control branches.Footnote ¹³

TABLE 3Acquirer, Target, and Control Branches

TABLE 3 Long description

The table is divided into two sections: Panel A (Unrestricted Sample) and Panel B (Excluding the 50 Largest Branches). Both panels use the same columns: Mean for Acquirer (A), Target (T), and Control (C) branches, followed by P-Values for comparisons A minus C, T minus C, and A minus T.

Panel A key data points:

* Assets: Mean values are 296 (A), 469 (T), and 181 (C).

* Loans: Mean values are 21 (A), 47 (T), and 13 (C).

* Number of workers: Mean values are 13.0 (A), 18.7 (T), and 16.4 (C).

* Ability of all workers (A K M): Mean values are minus 0.28 (A), minus 0.29 (T), and minus 0.20 (C).

* Number of observations: 3,417 (A), 2,002 (T), and 3,922 (C).

Panel B key data points:

* Assets: Mean values are 48 (A), 62 (T), and 57 (C).

* Loans: Mean values are 11 (A), 22 (T), and 7 (C).

* Profit per employee: Mean values are 0.17 (A), 0.14 (T), and 0.08 (C).

* Ability loan officers (A K M): Mean values are minus 0.05 (A), 0.05 (T), and 0.07 (C).

* Number of observations: 3,401 (A), 1,988 (T), and 3,902 (C).

Other metrics included in both panels are Deposits, local market shares for loans and deposits, loans and deposits per employee, and the share of loan officers. P-values of 0.00 are frequent for market share and worker count comparisons across both panels.

The fact that characteristics at the bank branch levels are not strikingly different is reassuring. However, to further guarantee that differences in observables do not threaten the plausibility of our identification hypothesis (discussed in Section II.D), that treatment and control branches would follow similar trajectories had the M&As not occurred, we include controls for baseline characteristics interacted with time dummies in our empirical analysis.

D. Empirical Strategy

Our empirical strategy exploits branch-level information and the timing of the M&A events to establish a link between the restructuring that follows a consolidation and branch outcomes. We also exploit variation in branches’ ownership before the consolidation to investigate heterogeneous effects on target and acquirer branches. As our treatment timing is staggered, with events occurring in 2007 and 2008, we implement a stacked regression estimation (Baker, Larcker, and Wang (Reference Baker, Larcker and Wang2022)). We construct two event-specific data sets (one for each M&A) and stack them (Figure A3 in the Supplementary Material). We refer to each event-specific data set as a cohort. For each cohort, the control group consists only of private bank branches that did not participate in a bank M&A during the time window of our estimation, but that participated in these events in the future. We define the control variables for each cohort by interacting them with cohort-specific indicators. We use an estimation window of 10 years: 3 years before and 6 years after the event. We restrict the main analyses to branches that remain open during the window and for which we can measure outcome variables. Using the same stacked approach, we also check the parallel trends assumption and the dynamic treatment effects after the M&A events.

We estimate the following model:

(1) $$ {\displaystyle \begin{array}{l}{y}_{i,g,t}={\delta}_1{Post}_{i,t}\times {Target}_i+{\delta}_2{Post}_{i,t}\times {Acquirer}_i+{\alpha}_{i,g}+{\alpha}_{g,m,t}\\ {}\gamma {MP}_i\times {Post}_{i,t}\times {Treated}_i+{\beta}_{g,t}{X}_i++{\epsilon}_{i,g,t}\end{array}} $$

in which $ i $ represents the bank branch, $ g $ the cohort, $ t $ calendar year, and $ m $ the municipality. The dummy variable $ {Target}_i $ ( $ {Acquirer}_i $ ) takes the value 1 when branch $ i $ belongs to a bank that was the target (acquirer) party in one of the M&A transactions of our sample. The dummy variable $ {Treated}_i $ takes the value 1 when branch $ i $ is either an acquirer or a target branch. The coefficient $ {\delta}_1 $ represents the M&A effect on the target branches, and the coefficient $ {\delta}_2 $ represents the M&A effect on the acquirer branches. We control for any fixed unobserved branch characteristics by using branch fixed effects $ {\alpha}_{i,g} $ (interacted with cohort dummies). To account for time-varying changes at the local level that are common to all branches, such as demand shocks, we include municipality-by-time-by-cohort fixed effects $ {\alpha}_{g,m,t} $ .Footnote ¹⁴ We also control for other possible confounding effects that might be correlated with post-consolidation performance by using branch pre-M&A characteristics (measured in 2006) ( $ {X}_i $ ) interacted with cohort and time dummies.Footnote ¹⁵

We account for the heterogeneous market power gains resulting from the consolidation, which are measured by the sum of the market shares in total lending of the target and acquirer branches in a given municipality prior to the M&A ( $ {MP}_i $ ).Footnote ¹⁶ Controlling for these gains is essential in our setting since it enables us to isolate the effects of resource restructuring from direct impacts led by changes in local market power. Since both the treated and control banks in our sample are large and diversified retail banks, it is reasonable to assume they operate in the same market. If banks were small and specialized in serving particular types of clients or services (e.g., Blickle et al. (Reference Blickle, Parlatore and Saunders2023), Paravisini et al. (Reference Paravisini, Rappoport and Schnabl2023)), this assumption would no longer hold, leading to mismeasurement in our metrics of market power gains. M&As involving large banks offer the added advantage of creating significant variation in market power gains across municipalities (Figure A4 in the Supplementary Material). Without such variation, controlling for market power gains would be infeasible. Finally, as banks were already large and diversified pre-consolidation, market power gains that are common across all markets (such as those resulting from a large network of branches or concerns about banks being perceived as “too big to fail”) play a smaller role in this context.

Our “within market” approach is convenient because the large M&As we study coincide with the onset of the 2007–2008 financial crisis. As a result, in the absence of the M&A events, different markets could have had distinct performances. For instance, firms in markets with more branches per population could be more leveraged or exposed to disruptions in international markets through trade relationships. Our municipality-by-time-by-cohort fixed effects approach deals with those possibilities.

The main concern with our empirical strategy is that banks participating in consolidations are not randomly selected, which is reflected in some differences in baseline characteristics, as shown in Table 3. Therefore, unobserved factors can cloud the magnitude of our coefficients. For instance, relative to control branches, target and acquirer branches may have experienced changes in productivity even in the absence of the M&A, a possibility that a parallel trends test would not detect. This concern is common to all M&A studies, and no perfect solution has been proposed in the literature. We seek to minimize this issue in several ways. First, our control branches belong to banks that participated in a large M&A after our sample period. Second, we add baseline branch characteristics interacted with time dummies to alleviate concerns that heterogeneity across bank branches affects our results. Third, we implement tests proposed by Rambachan and Roth (Reference Rambachan and Roth2023) to assess the robustness of our results to violations of the parallel trends assumption. Finally, certain features of the M&As we study further mitigate these concerns. One M&A was the result of an operation that involved companies with headquarters in foreign countries. The other had been negotiated for years and was accelerated by the collapse of Lehman Brothers. As a result, its timing was relatively exogenous.

A second confounder could arise from a different reaction of treated and control banks during the financial crisis. However, as explained in Section II.A, the crisis in Brazil was short-lived, mild, and did not spark bank failures. Moreover, our sample comprises large banks only, and these banks experienced an inflow of funding because they were perceived as too-big-to-fail (Oliveira et al. (Reference Oliveira, Schiozer and C. Barros2015)).

Finally, because M&As often lead to branch closures (e.g., Nguyen (Reference Nguyen2019)), our results could be affected by survivorship bias if the branches that remain open are positively selected based on future outcomes. A related concern is that the assets, credit, and deposits of closed branches may be transferred to surviving branches, mechanically inflating these variables in ways unrelated to productivity improvements. We believe these concerns are limited for two reasons. First, branch closures are relatively modest in our setting. Second, our results are robust even when we restrict the sample to markets without branch closures. Lastly, we apply tests following Lee (Reference Lee2009) to account for potential selection bias, and the results further confirm the robustness of our findings.

A. Dynamic Effects and Resource Reallocation

In this section, we report estimates from an augmented version of equation (1) that includes leads and lags of the treatment indicator to examine the dynamic effects of M&As on our main variables.Footnote ²¹ The goals of this exercise are twofold. First, by showing that treated and control branches followed similar trajectories pre-M&A, we alleviate concerns of selection into consolidation. For example, we show that, in comparison to control branches, acquirer branches are not becoming more productive in the years that precede the transaction. Second, even though our estimations take into account local market power gains prompted by the M&A episodes, the results could be driven by increases in market power at a more aggregate level, such as the size of the ATM and branch network, brand reputation, and safety. As these market power gains can be more timely exercised than restructuring gains, a gradual improvement in productivity variables is a further indication that restructuring gains are playing an important part.

In Figure 2, we analyze the branches’ labor dynamics around the consolidation announcement. We observe that employment decreases slowly over the years in target branches, driven by the increased number of internal transfers of employees, which starts right after the announcement. At the same time, we see an increase in the number of employees in acquirer branches, which is consistent with the transfers from the target branches to acquirer branches. Moreover, we provide evidence that the loan officer’s ability increases in acquirer branches, and the dynamics of this change align with the dynamic effects of internal transfers. Additionally, our dynamic results show that in the long run, the loan officer’s ability in target branches remains unchanged.

FIGURE 2Employment Reallocation

Figure 2 reports the dynamic effects of M&As on labor reallocation, estimating an augmented version of equation (1) that includes leads and lags of the treatment indicator. Reported 99% confidence intervals are based on standard errors clustered at the branch-cohort and bank-time-cohort levels. Dependent variables are inverse hyperbolic sine transformations of the original variables. All regressions include branch-by-cohort fixed effects, municipality-by-year-by-cohort fixed effects, baseline branch characteristics interacted with time-by-cohort fixed effects, and market power controls interacted with the Post M&A indicator. The estimation sample includes branches that remained open throughout the estimation window (3 years before and 6 years after the events).

FIGURE 2 Long description

The figure consists of six panels, each with an x-axis representing Time after M & A in Years ranging from negative 3 to 6, and a vertical dashed line at year 0 marking the event.

* Graph A, Target Branch Employment: The y-axis ranges from negative 0.6 to 0.2. Employment is stable near zero before the event, then shows a steady linear decrease post-event, reaching approximately negative 0.3 by year 6.

* Graph B, Acquirer Branch Employment: The y-axis ranges from negative 0.2 to 0.2. Employment shows a slight upward trend from negative 0.02 at year negative 3 to nearly 0.1 by year 6.

* Graph C, Target Branch Net Labor Transfers: The y-axis ranges from negative 1.5 to 0.5. There is a sharp drop to negative 0.6 at year 0 and a further trough at year 2, followed by a recovery toward zero by year 4.

* Graph D, Acquirer Branch Net Labor Transfers: The y-axis ranges from negative 0.2 to 0.8. Transfers spike after year 0, peaking at 0.4 in year 3 before declining and stabilizing around 0.2.

* Graph E, Target Branch Average Loan Officer Ability: The y-axis ranges from negative 0.3 to 0.1. Ability drops sharply to negative 0.12 in year 1 post-M & A, followed by a gradual recovery toward negative 0.05.

* Graph F, Acquirer Branch Average Loan Officer Ability: The y-axis ranges from negative 0.2 to 0.2. Ability starts at negative 0.1 in year negative 3 and shows a consistent asymptotic increase, crossing zero at the event and reaching 0.1 by year 6.

All graphs include vertical error bars representing 99% confidence intervals for each data point.

In Figure 3, we show that lending grows gradually after the event in acquirer branches, peaking at around 4 years after the M&A and remaining at this level thereafter. Lending declines gradually at target branches. These dynamic effects on lending supply are consistent with the changes in the labor force. In Figure 4, we show that increases in productivity take some years to materialize, supporting our hypothesis that M&A-driven resource reallocation is a driver of the observed productivity gains. Figure 5 shows the dynamics of the effects on branch closures, which are in line with previous results presented in Section III.C.

FIGURE 3M&A Effects on Branch Output

Figure 3 reports the dynamic effects of M&As on branch output, estimating an augmented version of equation (1) that includes leads and lags of the treatment indicator. Reported 99% confidence intervals are based on standard errors clustered at the branch-cohort and bank-time-cohort levels. Dependent variables are inverse hyperbolic sine transformations of the original variables. All regressions include branch-by-cohort fixed effects, municipality-by-year-by-cohort fixed effects, baseline branch characteristics interacted with time-by-cohort fixed effects, and market power controls interacted with the Post M&A indicator. The estimation sample includes branches that remained open throughout the estimation window (3 years before and 6 years after the events).

FIGURE 3 Long description

Four line graphs arranged in a two-by-two grid. All graphs share a common x-axis labeled Time after M & A in Years, ranging from negative 3 to 6, with a vertical dashed line at year 0 representing the event. The y-axes represent inverse hyperbolic sine transformations of output variables. Each data point includes vertical error bars representing 99% confidence intervals.

* Graph A. Target Branch Lending (Top-Left): The trend is relatively flat before the event, peaking at year negative 1. After year 0, there is a steady decline, reaching a trough around year 4 before a slight recovery.

* Graph B. Acquirer Branch Lending (Top-Right): Shows a consistent upward trend starting from year negative 3. The growth accelerates after year 0, transitioning from negative values to a plateau of approximately 0.4 by year 4 through 6.

* Graph C. Target Branch Deposits (Bottom-Left): Deposits remain near zero before the event. Post-event, there is a sharp decline reaching its lowest point at year 3, followed by a recovery toward zero by year 6.

* Graph D. Acquirer Branch Deposits (Bottom-Right): Similar to Graph B, there is a steady increase from year negative 3. After the event at year 0, the trend continues upward, stabilizing between 0.2 and 0.3 by year 6.

FIGURE 4Productivity

Figure 4 reports the dynamic effects of M&As on productivity, estimating an augmented version of equation (1) that includes leads and lags of the treatment indicator. Reported 99% confidence intervals are based on standard errors clustered at the branch-cohort and bank-time-cohort levels. Dependent variables are inverse hyperbolic sine transformations of the original variables. All regressions include branch-by-cohort fixed effects, municipality-by-year-by-cohort fixed effects, baseline branch characteristics interacted with time-by-cohort fixed effects, and market power controls interacted with the Post M&A indicator. The estimation sample includes branches that remained open throughout the estimation window (3 years before and 6 years after the events).

FIGURE 4 Long description

A multi-panel figure with six line graphs arranged in two columns and three rows. Each graph shares a common x-axis titled Time after M & A in Years, ranging from negative 3 to 6, with a vertical dashed line at year 0 representing the event. The y-axes represent inverse hyperbolic sine transformations of productivity variables.

* Graph A, Top-Left: Target Branch Deposits per Employee. The trend is flat near zero before the event, dips slightly in years 1 through 3, then increases to a higher plateau around 0.3 in years 4 through 6.

* Graph B, Top-Right: Acquirer Branch Deposits per Employee. Shows a steady upward trend starting from negative 0.1 at year negative 3, crossing zero at year negative 1, and reaching nearly 0.2 by year 6.

* Graph C, Middle-Left: Target Branch Lending per Employee. Data points fluctuate between negative 0.1 and 0 before the event and remain relatively stable near zero post-event, with a slight uptick at year 6.

* Graph D, Middle-Right: Acquirer Branch Lending per Employee. Displays a clear linear increase from negative 0.1 at year negative 3 to approximately 0.4 at year 4, then leveling off.

* Graph E, Bottom-Left: Target Branch Profits per Employee. Values are near zero pre-event, spike to 2.0 at year 2, then fluctuate between 1.0 and 1.5 through year 6.

* Graph F, Bottom-Right: Acquirer Branch Profits per Employee. Shows a gradual increase from zero at the event time to nearly 2.0 by year 6.

All data points include vertical error bars representing 99% confidence intervals. The horizontal zero line serves as a baseline for all panels.

FIGURE 5Branch Closures

Figure 5 reports the dynamic effects of M&As on branch closures, estimating an augmented version of equation (1) that includes leads and lags of the treatment indicator. Reported 99% confidence intervals are based on standard errors clustered at the branch-cohort and bank-time-cohort levels. The dependent variable is a dummy equal to 1 if a branch has closed. All regressions include branch-by-cohort fixed effects, municipality-by-year-by-cohort fixed effects, baseline branch characteristics interacted with time-by-cohort fixed effects, and market power controls interacted with the Post M&A indicator.

FIGURE 5 Long description

The image contains two panels, Graph A and Graph B. Both share a common x-axis labeled Time after M & A (in Years) ranging from minus 3 to 6, with a vertical dashed line at year 0 representing the treatment event.

Graph A, titled Target Closure, has a y-axis ranging from minus 0.05 to 0.15. The data points for years minus 3 to 0 are near zero. At year 1, there is a sharp peak to approximately 0.08 with a large vertical error bar extending from 0 to 0.15. From years 2 to 6, the values stabilize between 0.02 and 0.05 with smaller error bars.

Graph B, titled Acquirer Closure, has a y-axis ranging from minus 0.1 to 0.05. The data points for years minus 3 to minus 1 are near zero. Starting at year 0, the line shows a downward trend, dropping to approximately minus 0.04 at year 1 and remaining relatively flat between minus 0.04 and minus 0.06 through year 6. Vertical error bars are present for all points, indicating 99 percent confidence intervals.

B. Robustness Checks

We conduct robustness checks to assess the validity of the assumptions underlying our empirical strategy and the sensitivity of our results to alternative specifications.

First, in Table A2 in the Supplementary Material, we re-estimate our baseline specification excluding very large branches, which proxy for regional headquarters.Footnote ²² These establishments perform functions distinct from standard retail branches and may adjust differently following M&As. Moreover, as shown in Table 2, including these observations yields highly skewed distributions and large differences in magnitudes for several key variables, raising concerns that our baseline results may be partly driven by outliers. When we exclude these very large branches, the results remain essentially unchanged, confirming that our findings are not driven by regional headquarters or other exceptionally large establishments.Footnote ²³

Second, to address concerns about differential growth paths between treated and control banks, we replicate our main results using alternative control groups and including additional controls. In Table A3 in the Supplementary Material, we include only branches of control banks that acted as acquirers in M&A deals after the period of our analysis, reducing concerns that control banks faced different trends. The estimates remain quantitatively similar to our baseline results.Footnote ²⁴ Additionally, in Tables A4 and A5 in the Supplementary Material, we include baseline productivity measures interacted with cohort and time dummies as controls, with results again consistent with our main findings.

Third, we examine whether our controls for market power adequately account for competitive dynamics in local banking markets. In Table A6 in the Supplementary Material, we include changes in market power in deposit and labor markets, as well as in agricultural loans.Footnote ²⁵ In Table A7 in the Supplementary Material, we control for changes in the Herfindahl–Hirschman Index induced by the M&A events. Across all specifications, our baseline results remain robust.

Fourth, given that the main analyses rely on branches that remain open throughout the estimation window and the abnormal pattern of branch closures that followed the M&As, we test if our results are mechanically driven by the transfer of lending and deposits from closed branches to surviving branches. In Table A8 in the Supplementary Material, we augment our baseline specification with controls for the pre-M&A share of loans and deposits associated with branches that subsequently closed, interacted with the post-M&A indicator. The results remain robust: acquirer branches expand lending, deposits, and profitability, while target branches contract. In Tables A9 and A10 in the Supplementary Material, we further restrict the sample to municipalities with no branch closures. Despite the smaller sample size, the results remain very similar to our baseline. Furthermore, we examine whether workers from closed branches contribute to the labor reallocation documented in Section III.A. Table A11 in the Supplementary Material shows that closed branches account for a nontrivial share of worker inflows at surviving target and acquirer branches. These results indicate that closures might also play a role in the restructuring process by generating fixed-cost savings through the elimination of redundant branches and by contributing to the reallocation of workers to surviving branches.

Fifth, we further probe the role of selective branch exit for our results by implementing tests similar to Lee (Reference Lee2009). As M&As trigger branch closures, one may worry that our estimates based on surviving branches overstate post-M&A improvements if poorly performing units are disproportionately closed. The Lee-bounds approach provides best- and worst-case estimates for treatment effects under monotonic selection by trimming the outcome distribution of surviving control branches so that treated and control groups are comparable. The resulting bounds, shown in Table A12 in the Supplementary Material, remain consistent with the sign and magnitude of our baseline estimates, suggesting that strong selection on branch survival would be required to overturn our main conclusions.

Sixth, we test the impact of alternative transformations of the outcome variables. While our baseline uses the inverse hyperbolic sine transformation to accommodate zero and negative values (Bellemare and Wichman (Reference Bellemare and Wichman2020)), Table A13 in the Supplementary Material replicates the main results using the log of the variable plus one. As this specification excludes negative-valued outcomes, we suppress profit regressions. The estimates remain quantitatively similar.

Seventh, to address potential spillovers in municipalities affected by multiple consolidations, we replicate our analysis excluding all municipalities treated in both the 2007 and 2008 M&A events. Table A14 in the Supplementary Material shows that, although the sample size is reduced by over 90% and confidence intervals widen, point estimates are consistent with the baseline. Notably, we observe insignificant changes in employment at acquirer branches and improvements in target productivity metrics, supporting the interpretation that reallocation of human capital is an important driver of productivity gains.

Eighth, the variable branch assets include certain categories that do not correspond to actual assets and therefore inflate the measure (see Section A3 of the Supplementary Material for details). In Table A15 in the Supplementary Material, we show that our results are robust to controlling for an adjusted measure of pre-consolidation assets. In Table A16 in the Supplementary Material, we further demonstrate that the results remain robust when this control is excluded altogether.

Finally, we evaluate the sensitivity of our results to violations of the parallel trends assumption using the methodology of Rambachan and Roth (Reference Rambachan and Roth2023). Specifically, we allow for differences in linear trends between treated and not-yet-treated units and assess how large such violations would have to be to overturn our conclusions. Figures A6–A8 in the Supplementary Material show that our significant results are robust to both linear and substantial non-linear deviations from parallel trends.Footnote ²⁶

Overall, these robustness checks provide strong evidence that our main conclusions are not sensitive to branch closures, alternative control groups, violations of the parallel trends assumption, outcome transformations, potential spillovers, or market power measures.