Banking crises

The events of the Great Recession (2007–9), a financial crisis followed by a deep economic recession and a slow recovery, renewed interest in the study of how financial market distress affects households and firms. The Great Depression (1929–37) serves as the most pertinent historical comparison to the economic challenges faced in the late 2000s. Initiated by a stock market crash in 1929 and exacerbated by subsequent banking failures, the Depression remains a focal point in US economic history and macroeconomic policy. Notably, Friedman and Schwartz (Reference Friedman and Schwartz1963) and Bernanke (Reference Bernanke1983) identify bank panics as pivotal in shaping the economic downturn and determining the Depression’s duration and severity. From a policy standpoint, discerning the causes of financial institution failures and the efficacy of preventive measures is crucial for financial market regulators and central banks.

Two main hypotheses underpin the causes of bank failures during the Depression in the United States. The first hypothesis posits that banks faced insolvency as the assets they held (primarily mortgages, business loans and bonds) diminished in value (Temin Reference Temin1976; White Reference White1984; Calomiris and Mason Reference Calomiris and Mason2003). In essence, banks had made risky investments before 1929: excessive loans to businesses that would later collapse, to stock speculators who suffered significant losses in the 1929 crash, and mortgages during the post-World War I construction boom. Empirical support for this solvency hypothesis manifests in various ways, such as state-level bank failures correlated with economic and loan characteristics, bank-level data on loan quality, and trends in bond yields. Typically, these studies consider the probability or severity of bank suspensions, concluding that underlying economic shocks are potent predictors of bank failure. They suggest that liquidity-assistance policy interventions would likely have had limited impact during the Depression.

The second hypothesis, the illiquidity hypothesis, contends that a surge in withdrawals by panic-stricken depositors crippled the banking sector. The stock market crash sowed doubts about future economic prospects (Romer Reference Romer1990). Additionally, news about the failures of major, interconnected institutions stirred concerns about the banking sector’s stability. Consequently, public trust eroded, leading to a rush on banks, which, unable to liquidate assets swiftly, had no choice but to halt operations (Friedman and Schwartz Reference Friedman and Schwartz1963; Wicker Reference Wicker2000). Under this theory, the crisis might have been mitigated had the Federal Reserve acted as a lender of last resort, offering cash to banks in exchange for illiquid assets at non-fire sale prices. Some critiques center on the Federal Reserve’s perceived inaction during the Depression based on this perspective.

Both hypotheses hold conceptual and empirical merit, contingent on the specific timeframe, geographical context and granularity of the data. No singular cause can encapsulate the magnitude of banking panics during the Great Depression. The most insightful evidence on the solvency and liquidity channels’ relative importance comes from Richardson (Reference Richardson2007). Utilizing quarterly bank-level data from 1929 to 1933, he observes a temporal evolution in bank failures: initially, small rural banks faced increasing failure rates. The collapse of Caldwell and Company and the Bank of the United States in 1930 triggered bank runs. By 1931, following Britain’s exit from the gold standard and declining asset values, the majority of failing banks were indeed insolvent. Nearly three-fourths of these institutions were deemed insolvent, with one-fourth solvent and either resuming business or merging. Among those that suspended operations, half did so due to depositor withdrawals.

What was the scope of Federal Reserve intervention?

The initial response of the Federal Reserve to the Depression has been a subject of criticism, especially its support (or lack thereof) for the money supply and banking institutions. However, pinpointing its direct impact is intricate due to potential responses from households and local bodies. The ideal approach would involve comparing regions with analogous economic trajectories but divergent policy responses during the Depression. Mississippi serves as a pertinent case study in this regard, with its northern and southern counties subjected to distinct policy regimes up until 1931 (Richardson and Troost Reference Richardson and Troost2009; Ziebarth Reference Ziebarth2013; Jalil Reference Jalil2014).

The southern counties of Mississippi fell under the purview of the Atlanta Federal Reserve Bank (6th District). Atlanta leaders adhered to Bagehot’s rule, a doctrine advocating that central bankers should provide credit to illiquid but solvent institutions during financial upheavals, thereby preventing losses from runs on otherwise healthy banks. Gamble (Reference Gamble1989), in his historical overview of the Atlanta Federal Reserve Bank, recounts instances where bank officials physically transported currency into banks to reassure apprehensive depositors of the bank’s solvency. Conversely, in the north, the St. Louis Federal Reserve Bank (8th District) championed the ‘Real Bills’ perspective, asserting that credit supply should diminish during downturns, as reduced economic activity necessitated less credit. This stance persisted until the summer of 1931.Footnote ² Therefore, during the Depression’s initial two years, banks in Mississippi experienced two contrasting policy approaches. What were the implications?

Drawing from bank-level and county-level data, Richardson and Troost (Reference Richardson and Troost2009) unveil compelling findings. Banks in the 6th District counties of Mississippi consistently outperformed their counterparts in the 8th District in terms of survival rates, credit availability and commercial activity, particularly during the panics of 1930-1. Their research suggests that broad application of LLR policies might have mitigated the onset of banking panics. Broadening the analysis to encompass the entire border of the 6th District, Jalil (Reference Jalil2014) investigates whether bank performance in counties within 50 miles of this boundary depended on the specific Federal Reserve policy regime. Utilizing county-level banking data, he discerns that bank suspension rates in 1929 and 1930 were consistently lower inside the 6th District compared to adjacent counties just beyond the border.Footnote ³

Banking panics and local economic outcomes

Bank closures carry direct costs, such as the wealth loss for depositors who only recover a portion of their claims after a bank’s liquidation. However, this is merely one avenue through which bank failures can adversely affect local economies. An alternative channel is the escalation of financial intermediation costs (Bernanke Reference Bernanke1983). In an atmosphere riddled with uncertainty, banks lean towards risk aversion, investing predominantly in secure assets and showing hesitancy in extending credit, even to creditworthy businesses (Cornett et al. Reference Cornett, McNutt, Strahan and Tehranian2011). Consequently, these businesses curtail hiring and production. When these firms seek funds for investment or debt refinancing, the absence of willing lenders leads to a decline in their output.

The experimental scenario in Mississippi, combined with the Atlanta Federal Reserve border, offers a framework to investigate the influence of bank failures on local economic conditions. Ziebarth (Reference Ziebarth2013) assembles plant-level data from the Census of Manufactures during the Depression and employs a difference-in-differences approach to juxtapose plants in northern and southern Mississippi. His findings reveal a 37 percent reduction in physical output in the north with no discernible impact on total workforce numbers. This effect predominantly arises from the intensive margin. When viewed at the county level, he finds a pronounced negative impact on the number of workers. Related work by Hansen and Ziebarth (Reference Hansen and Ziebarth2017) examines how banking distress altered firms’ credit relationships and business dynamics during the Depression. Using detailed firm-level data, they show that bank failures disrupted existing lending relationships and increased business exit among financially dependent firms, highlighting an additional channel through which local banking crises propagated to real economic activity.

Recent work also provides complementary evidence on the microeconomic transmission of banking distress to real activity in earlier periods. Using newly assembled national data from the Progressive Era (1900–29), del Angel et al. (Reference del Angel, Richardson and Gou2024) show that bank failures causally increased bankruptcies among firms that depended on commercial banks for short-term credit, while firms that were largely bank-independent (primarily large manufacturers with access to retained earnings and capital markets) were largely insulated. They demonstrate that credit supply disruptions operated through clearly identifiable bank-firm linkages, rather than through broad aggregate demand channels. Their findings underscore that bank failures mattered precisely because of firms’ reliance on relationship banking for working capital, reinforcing the importance of financial intermediation in shaping real outcomes even before the Great Depression.

Banking and finance

To assess local credit conditions during the Depression, I draw on a 1934 US Commerce Department survey of over 6,000 small and medium-sized manufacturers, which found widespread credit use and credit difficulties: over 80 percent reported that commercial banks provided the main source of external funding, and 45 percent of borrowing firms reported trouble accessing loans (US Census Bureau 1935). I construct an industry-level measure of credit constraint (the proportion of all firms that reported credit difficulty) from this survey and aggregate it to the county level using output-weighted averages.Footnote ⁴ While informative, the survey reflects conditions several years into the Depression and may understate constraints in industries where many firms had already failed. To address this limitation, I complement it with the measure of industry-level external financial dependence from Rajan and Zingales (Reference Rajan and Zingales1998), which captures structural financing needs computed using accounting data from publicly traded firms in the 1980s. I further incorporate county-level data on banking disruptions using bank suspension data from the Federal Deposit Insurance Corporation and newly digitized balance sheet data from the Office of the Comptroller of the Currency (Federal Deposit Insurance Corporation 1992; United States Office of the Comptroller of the Currency 1920–32).

Manufacturing and industry composition

To construct measures of local industrial activity and composition, I digitize county-level output, employment and establishment count data from the Census of Manufactures for 1927–35, including a 1927 special tabulation of plant counts by industry and county published in the Market Data Handbook (US Census Bureau 1937; Stewart Reference Stewart1929). I transform the establishment counts by industry to the 15-sector classification used in the 1935 Credit Survey and convert these counts into economic shares using state-industry averages of output, employment and wages, thereby accounting for variation in establishment size across sectors. Due to census confidentiality thresholds, coverage is incomplete in low-manufacturing counties; I address this by restricting to a balanced panel across census years. I construct outcomes including log output, log employment and log wages. Baseline county characteristics are well balanced across the border, though manufacturers just outside the boundary were marginally larger on average.

Linked decennial censuses, 1930–40

Data regarding manufacturing workers comes from the 100 percent count US census records. This article focuses on working-age adults (ages 18–55) in Mississippi, Tennessee and Louisiana – the three states bisected by the Atlanta Federal Reserve border. I used the crosswalks provided by the Census Tree Project (Price et al. Reference Price, Buckles, Haws and Wilbert2023) and IPUMS publicly available census data (Ruggles et al. Reference Ruggles, Nelson and Sobek2024) to link records over time.Footnote ⁵ In total, the 1930–40 sample includes over 1.88 million individuals, with 65 percent living in counties within the Atlanta district. This linked sample covers roughly 51 percent of the manufacturing workforce in 1930 in those states but is not perfectly representative. To address this issue, I use inverse probability weighting in my empirical analysis, creating weights after predicting the characteristics associated with a successful link. I construct four primary outcome variables: a binary variable taking the value of 1 if the individual reported working in the manufacturing sector in 1940, the logarithm of the change in the individual’s occupational income score, and two binary variables indicating inter-county and inter-state migration.

Other data

Using the Geographic Information Systems (GIS) software, I identify counties within a 50-mile radius of all Federal Reserve districts. These bordering counties were further segmented based on their proximity to specific Federal Reserve districts (e.g. Atlanta – St. Louis segment) as shown in Figure 1. My approach closely mirrors the method employed by Jalil (Reference Jalil2014). For robustness checks, I transcribed the 1927 consumer markets map from the US Department of Commerce Market Data Handbook. This map categorizes counties into 632 distinct consumer markets, which were crafted from a consumer-centric viewpoint based on an initiative by the International Magazine Company to optimize sales efficiency. The selection of trading centers was based on parameters like population, geography, economic sources, transportation and trade channels. Lastly, I use New Deal county-level Reconstruction Finance Corporation (RFC) and Agricultural Adjustment Administration (AAA) expenditure data collected by Fishback et al. (Reference Fishback, Kantor and Wallis2003) in various robustness exercises.

Figure 1.

Counties around the Atlanta Federal Reserve district border

Figure 1 Long description

The map shows the southeastern United States with state outlines. North is at the top. Counties within 50 miles of the Atlanta (6th) district border are shaded by border segment and by side of the border: Atlanta–St. Louis segment. This is the longest segment. It runs west across Tennessee, then south through Mississippi, then west across southern Mississippi. Atlanta-side counties lie to the east and south of the line. St. Louis-side counties lie to the north and west.Atlanta–Cleveland segment. This is a short segment in eastern Tennessee and southeastern Kentucky.Atlanta–Richmond segment. This segment runs southeast from the Tennessee–North Carolina area along the Georgia–South Carolina border to the Atlantic coast.Atlanta–Dallas segment. This segment is in Louisiana. Atlanta-side counties lie to the east and Dallas-side counties to the west. Uncoloured areas lie more than 50 miles from the district border.

Note: This map shows the border regions of the Atlanta Federal Reserve district. The four border regions are: Atlanta – St. Louis; Atlanta – Dallas; Atlanta – Cleveland; and Atlanta – Richmond. The 50-mile buffer was generated using Geographic Information System (GIS) software.

Summary statistics

Table 1 presents the summary statistics. Overall, county-level banking variables are available for 365 border counties across nine years and for 364 counties for a single year (1934, due to missing bank data for one county). The roughly 1,000 observations for manufacturing span between 190 and 210 border counties reporting manufacturing activities biennially from 1927 to 1935.

Table 1.

Summary statistics

Table 1 Long description

Two panels report summary statistics using columns for count, mean, sd, p5, p25, p50, p75, and p95. Panel A covers county level banking, manufacturing, and related measures, with most banking rows having count 3,650. Active banks have mean 3.51, sd 2.57, p5 1.00, p25 2.00, p50 3.00, p75 5.00, and p95 9.00. Suspended banks are mostly zero at the median, with means of 0.18 for all, 0.03 for national, and 0.14 for state; p95 is 1.00 for all and state and 0.00 for national. Active deposits in millions have mean 2.60 and p50 0.91, while suspended deposits in millions have p50 0.00 across all, national, and state; p95 is 0.36 for all, 0.00 for national, and 0.22 for state. Capitalization ratio has count 1,840 with mean 0.15 and p50 0.13; loan growth for 1924 to 1929 has count 168 with mean 0.11 and p50 0.10. Manufacturing measures include output in millions with count 1,013 and p50 1.33, wages in millions with count 1,013 and p50 0.28, establishments with count 1,023 and p50 16.00, and wage earners in thousands with count 1,021 and p50 0.50; borrower difficulty and RZ difficulty each have count 356 with p50 values 0.48 and 0.21. Panel B summarizes linked census variables with observations 1,883,289, including indicator rows with p50 values of 0.00 for manufacturing in 1930 and 1940, 1.00 for Atlanta, native, male, and white, and 0.00 for urban, moved counties 1930 to 1940, and moved states 1930 to 1940; years of schooling has count 1,856,465 with mean 7.59 and p50 8.00, and age in 1930 has mean 32.86 and p50 32.00.

Notes: Panel A presents the summary statistics for all counties within 50 miles of the Atlanta Federal Reserve district border (see Figure 1). Banks (count) and Deposits (millions of nominal dollars) are reported annually between 1926 and 1936. Capitalization ratio is defined as the total surplus and profits divided by assets for nationally chartered banks in 1928. Loan growth is change between the log of all loans between 1924 and 1929 for nationally chartered banks. Loans, surplus and total assets come from the Office of the Controller of Currency. Output (millions of nominal dollars), Wages (millions of nominal dollars), Establishments (count) and Workers (thousands) come from the Census of Manufacturing, reported biennially between 1927 and 1937. Borrower-Difficulty is an industry-weighted share of small and medium manufacturers facing credit difficulties. RZ-Difficulty is an industry-weighted external capital dependency measure from Rajan and Zingales (Reference Rajan and Zingales1998). Panel B presents the summary statistics of linked census data in Mississippi, Tennessee and Louisiana between 1930 and 1940. See text for details.

Across counties, there is a pronounced negative correlation between manufacturing activity and bank failures. Figure 2 illustrates a binned scatter plot of county-level changes in the logarithm of nominal per-capita manufacturing output between 1933 and 1929 (y-axis) against the cumulative number of bank suspensions as a proportion of all 1927 banks (x-axis) in all counties with at least one suspended bank and manufacturing activity in both 1933 and 1929 (totaling 1,623 counties). On average, a 10 percentage point increase in bank failure is associated with 2 log point decrease in manufacturing output across counties.

Figure 2.

Manufacturing decline during the Great Depression and banking failures

Note: This figure plots a binscatter plot of county-level change in the log of manufacturing output in nominal dollars between 1933 and 1929 (y-axis) and the cumulative number of bank suspensions (1929–33) as a proportion of all banks in 1927 (x-axis). Only the counties with at least one suspended bank during 1929–33 and manufacturing activity in both 1933 and 1929 are included (1,623 counties). Manufacturing data comes from Census of Manufactures and the banking data comes from the FDIC. Standard error in parenthesis.

Empirical design

The primary outcome of interest is the rate of bank failure. At the county level, I gauge this using two metrics: suspensions and the proportion of deposits retained by active banks at year-end, with the latter assessed against pre-Depression (1927) data on banks and total deposits. Both of these metrics are key, as banks might recommence operations after a brief suspension. However, the suspensions dataset does not differentiate between permanent liquidations and temporary suspensions. Hence, the end-of-year deposit values, reflecting more enduring banking sector shocks, act as a supplementary measure. Additionally, I create a binary variable, which is set to 1 if any bank within the county was suspended within the year and 0 otherwise, as a mechanism to mitigate the influence of outliers using the continuous variables.

I introduce various control variables in my analysis. I compute two measures to account for unobserved time-varying confounders due to fundamental banking differences between counties. First, I define the pre-Depression (1927) ‘capitalization ratio’ as the total surplus and capital divided by total assets. Higher capitalization ratios reflect lower leverage of the banking sector and a higher probability of withstanding depositor withdrawals. Second, I compute the logarithm of loan growth between 1924 and 1929. Higher loan growth could potentially correlate with decreased loan quality and a higher default rate in the 1930s. Finally, to control for non-financing industry-level time-varying confounders, I use the 1927 manufacturing revenue by industry shares in order to identify the dominant industry in each county. I interact the dominant industry dummy variable with time dummies to account for national industry trends.

I do not find strong underlying ex ante differences between counties across the border. Table 3 shows that counties on the border of the Atlanta district were similar.Footnote ⁶ Notably, the counties did not differ in their suspension rates as of 1927, had the same (estimated) proportion of manufacturing firms facing financial constraints, and their industrial structure was similar. There are some differences, but they are small. For example, although fewer banks were in the average county inside Atlanta, the total amount of deposits in 1928 was the same.

Table 3.

Covariate balance in 1927

Table 3 Long description

County averages are reported for counties within 50 miles of the Atlanta Federal Reserve district border, with columns for, All, In, Out, and, Difference. For, Deposits active minus all, the averages are 3546.18 in and 2876.15 out, with a difference of minus 670.02. For, Log pop, the averages are 9.78 in and 9.94 out, with a difference of 0.17 marked with one asterisk. For, Log Output, Log Wages, Log Wage Earners, and, Log number of Establishments, the differences are minus 0.12, minus 0.17, minus 0.30, and minus 0.17, respectively. For, Bank Suspension Rate all, Borrower Difficulty, and, RZ Difficulty, the in and out averages are 0.01 and 0.01, 0.48 and 0.48, and 0.21 and 0.20, with differences of 0.00, 0.00, and minus 0.00. Observations are 365 for All, 170 for In, 195 for Out, and 365 for Difference. Industry shares are also listed with the same columns and include, Paper and allied, Food and kindred products, Textiles, Rubber, Petroleum and coal, Forest products, Printing and publishing, Leather, Stone clay glass, Nonferrous metals, Transportation equipment, and, Miscellaneous. In the industry section, the difference is marked with two asterisks for, Forest products, at minus 0.014, and with one asterisk for, Printing and publishing, at 0.012; other listed industry differences range from minus 0.013 for, Textiles, to 0.023 for, Food and kindred products. Industry observations are 356 for All, 164 for In, 192 for Out, and 356 for Difference.

Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01

Notes: This table reports variable averages among counties within 50 miles of the Atlanta Federal Reserve district border. Column ‘All’ reports the averages for all counties along the border and columns ‘In’ and ‘Out’ report them only for those in the district and for those outside the district, respectively. Column ‘Difference’ computes the difference and reports the T-test on the equality of means. The variables ‘Banks (active - all)’, ‘Deposits (active - all)’ and ‘Bank Suspension Rate (All)’ reported here are as of 1927 and come from the FDIC. ‘Borrower-Difficulty’ is an industry-weighted measure of credit access difficulty that uses the 1927 manufacturing industry by county establishment data and the 1935 credit survey of manufacturing industries. ‘RZ-Difficulty’ is an industry-weighted measure of external finance dependence that uses the 1927 manufacturing industry by county establishment data and industry measures derived in Rajan and Zingales (Reference Rajan and Zingales1998). Chemicals industry – zero establishments – is omitted.

I compare county-level outcome variables before and after the onset of the Great Depression across the Atlanta boundary using the following specification:

(2)\begin{equation}S_{jk}=\alpha_j+\beta_{s(j)k}+\sum_{i=1926}^{1933}\left(Atl_jT_i\right)\cdot\gamma_i+X_{jk}+\epsilon_{jk}\end{equation}

where ${T_i}$ is a year dummy taking the value of 1 if $i = k$ and 0 otherwise and $At{l_j}$ takes the value of 1 if the county belongs to the Atlanta district and 0 otherwise. Like in the previous section, I use county ( ${\alpha _j}$) and border region by year fixed effects to account for all unobserved but static county variables and region-level trends in bank failure rates. The control variables in ${X_{jk}}$ include, in various specifications, proxies for baseline banking and manufacturing interacted with year dummies as described above. The coefficients of interest are ${\gamma _i}$, which capture the time-varying difference in outcome $S$ in counties inside Atlanta compared to average outcomes within border regions. The omitted interaction year is 1927, and I cluster the standard errors at the county level.

Baseline results

Table 4 presents the results. Panel A gives the estimates of the coefficients of interest when the outcome variables are suspension rates (proportion of suspended to non-suspended banks) and panel B presents them for active rates (proportion of active to 1927 active banks and deposits).

Table 4.

Bank suspension and active rates around the Atlanta border

Table 4 Long description

Two panels report coefficients for in ATL interaction with each year, for suspension outcomes and active rate outcomes, split into all banks, national, and state columns with standard errors in parentheses. Panel A lists Bank Suspension Rate columns 1 to 3 and I bank suspended columns 4 to 6 for years 1926 and 1928 to 1933. In 1926, coefficients are near zero, including Bank Suspension Rate all 0.007 and I bank suspended all 0.001. In 1928, Bank Suspension Rate is 0.019 for all, minus 0.032 for national, and 0.025 for state, while I bank suspended is 0.071 for all, minus 0.030 for national, and 0.065 for state. In 1929, coefficients are negative across all six suspension columns, including Bank Suspension Rate minus 0.068 for all, minus 0.064 for national, and minus 0.070 for state, and I bank suspended minus 0.143 for all, minus 0.077 for national, and minus 0.139 for state. From 1930 to 1933 in Panel A, Bank Suspension Rate remains negative in 1930 and near zero thereafter, while I bank suspended is near zero in 1930 and slightly negative or positive through 1933 depending on charter type. Panel B lists Bank Active Rate columns 1 to 3 and Deposit Active Rate columns 4 to 6 for the same years. In 1930, active rate coefficients are positive across all six columns, including Bank Active Rate 0.101 for all, 0.142 for national, 0.102 for state, and Deposit Active Rate 0.073 for all, 0.096 for national, 0.084 for state; other years in Panel B are closer to zero, with 1933 showing Deposit Active Rate 0.048 for all and 0.082 for state. Model rows show R squared values, N counts, and checkmarks for Border region by Year fixed effects and County fixed effects in every column.

Notes: This table reports the estimated coefficients of the in-ATL × year fixed effects in Equation (2). Controls include boundary region (e.g. Atlanta – St. Louis border) by year fixed effects and the omitted baseline interaction is 1927. Outcome variables are the county × year level and indicated by the column header. ‘Bank suspension rate’ is defined as the number of banks suspended divided by end of year total number of banks in operation. ‘I(Bank Suspended)’ is a binary variable taking the value of 1 if at least one bank suspended operations during the year. ‘Bank Active Rate’ is the number of banks in operation at the end of the year divided by number of banks in operation in the same county in 1927. ‘Deposit Active Rate’ is defined analogously. ‘National’ and ‘State’ refer to nationally chartered vs. state-chartered banks. The sample period is 1926–33. The standard errors are clustered at the county level.

In both panels, the estimates show that county bank failure occurred at similar rates around the Atlanta border before the Depression, relative to 1927 levels and controlling for local shocks inside a border region. The results in column 1 indicate that, relative to 1927 levels, the share of suspended banks was 6.0 and 5.2 percentage points lower in Atlanta counties in 1929 and 1930 (s.e. = 0.025 and 0.023), respectively, which is consistent with the results in Jalil (Reference Jalil2014). At the mean number of banks, this translates to 0.24 and 0.2 fewer suspended banks in each year, or approximately 0.45 more banks remaining on average in a county after 1930 in the Atlanta district. After 1931, the coefficients are not significantly different than zero. These years are also when more banks closed due to solvency issues and when there was a convergence of policy between Atlanta and its neighboring districts. Columns 2 and 3 show that the effect on the suspension rate is similar for both state and national banks. Columns 4–6 show that a county in Atlanta was 14 percent less likely to experience any bank failure in 1929, but the effect does not extend to 1930.

In panel B, I show that the results are consistent when using the number of active banks and total deposits as outcomes. By the end of 1930, counties within the Atlanta district retained 10.1 percentage points more active banks and held 7.3 percentage points more deposits relative to 1927 levels compared to neighboring counties (s.e. = 0.029 and 0.03), conditional on county fixed effects and region-specific trends. These results are qualitatively similar to those using suspensions as the outcome variable in panel A, and more pronounced for nationally chartered banks, where the effect is present even at the end of 1931.

Robustness results

Despite generally balanced counties on either side of the border and the plausibly exogenous historical placement of Federal Reserve boundaries, concerns about interpreting the results in Table 4 as the causal effects of the Atlanta Federal Reserve policies remain. In this section, I discuss the results with additional controls and different samples, and argue that the stability of the baseline estimates provides strong evidence in favor of a causal interpretation.Footnote ⁷

The first concern is that the small pre-Depression differences in county characteristics across the Atlanta border as shown in Table 3 could contribute to the divergence in bank failure rates during 1929 and 1930. For example, if smaller farms tend to default at higher rates on mortgages, the presence of relatively smaller farms just outside Atlanta may explain larger bank failure rates in 1929 and 1930. I control for these pre-Depression differences by interacting the non-balanced covariates with time dummies, and I do not find any change in the baseline estimates (column 2 of Online Appendix Tables A1 and A2).

The second concern is that omitted underlying differences in bank conditions, such as bank leverage or historical loan growth before the Depression, could be causing the differences in bank failures. Using the county-level OCC data on nationally chartered banks, I define the pre-Depression (1927) ‘capitalization ratio’ as the total surplus and capital divided by total assets. Higher capitalization ratios reflect lower leverage of the banking sector and a higher probability of withstanding depositor withdrawals. I also compute logarithmic loan growth between 1924 and 1929. Higher loan growth could potentially correlate with decreased loan quality and a higher default rate in the 1930s. I interact both county-level measures with time dummies and include them in the specification and, again, I do not find any meaningful differences with the baseline (column 3 of Online Appendix Tables A1 and A2).

The third concern is that firms in counties just outside the border are in more pro-cyclical industries than are those just inside the Atlanta region and thus caused more strain on the banking sector in 1929 and 1930. To control for industry-level time-varying confounders, I use the 1927 manufacturing revenue by industry shares in order to identify the dominant industry in each county. I interact the dominant industry dummy variable with time dummies to capture dynamic effects and include them in the specification. The results remain significant (column 4 of Online Appendix Tables A1 and A2).

Lastly, I address concerns of sample selection and standard error clustering. Because the original study of the Atlanta Fed’s LLR behavior, Richardson and Troost (Reference Richardson and Troost2009), focuses solely on Mississippi – a design that controls, by construction, for unobservable differences across states such as the regulatory environment – I first limit the analysis to the counties in Mississippi. I find larger, but noisier, point estimates for both bank suspensions and deposits, consistent with Richardson and Troost (Reference Richardson and Troost2009) and Ziebarth (Reference Ziebarth2013) (column 5 of Online Appendix Tables A1 and A2). Next, to account for unobserved differences in state-level policies in the full border sample, I limit the analysis to counties that belong to consumer market areas bisected by the Federal Reserve border, and find little change in the estimates (column 6). I then show that potential spillovers across the region are not driving the results by estimating the specification with only counties within 25 miles of the border and with only counties in the 25–100-mile range (columns 7 and 8 of Online Appendix Tables A1 and A2). Furthermore, the results are robust to using spatial standard errors via the methods of Colella et al. (Reference Colella, Lalive, Sakalli and Thoenig2019) (column 9) and to sequentially excluding one border segment at a time (columns 10–13).

Placebo tests and other policy regimes

To further assess the validity of the identification strategy and rule out spurious spatial correlations, I now turn to a series of placebo tests that randomize treatment assignment and compare effects across other Federal Reserve district borders. I conduct a randomization test in which I resample placebo borders within each border region of all counties in the sample using tools developed in Heß (Reference Heß2017). Specifically, at each permutation, the share of counties inside the Atlanta region in each border segment remains fixed, but each individual county is randomly chosen to be either ‘in Atlanta’ or ‘out of Atlanta’. I do this 1,000 times and estimate Equation (2) using the main outcome variable, bank suspension rates, and collect the estimated coefficients in $\gamma $ for 1929 and 1930. I plot the two distributions in Figure 4. The vertical lines indicate the baseline effects estimated using the true Atlanta Federal Reserve borders. As is clear from the figure, the true estimates lie in the tail of the distribution (98th percentile) of the placebo estimates, and the distribution of placebo estimates for both years is centered around zero. These results provide supporting evidence that the estimated baseline impacts of the differing policies of the Atlanta Fed on banking failures is unlikely to have occurred by chance.

Figure 4.

Distribution of estimated coefficients of 1,000 placebo border permutations

Notes: This figure plots the distributions of placebo effects computed using a randomization test as follows: using the tools developed in Heß (2017), I resample each county’s Atlanta status (‘in’ vs. ‘out’) within each border region, keeping the share of ‘in’ and ‘out’ counties constant. I conduct this 1,000 times, re-estimate Equation (2), and store the 1929 and 1930 interaction terms of interest. The outcome variable is the annual suspension rate of all banks within a county. Control variables in Xjk include only border region by year fixed effects. The vertical line shows the point estimates using actual borders.

Finally, instead of permuting counties into placebo borders, I extend the analysis to actual border counties in other Federal Reserve regions. If the differences in LLR are driving these outcomes, as the robustness exercises suggest, then it must also be true that the absence of these differences should result in little or no change in bank failures. In districts that did not differ in their policies from their neighbors, what is the prevalence of significant differences in bank failures in their border counties? Using the border regions of the Kansas City, Dallas, Cleveland, Richmond and St. Louis Federal Reserve Bank regions, I re-estimate the main specification and plot the results in Figure 5. The outcome variable is the total bank suspension rate. The variable $Atl_j$ in Equation (2) is now defined as follows: it takes the value of 1 if the county is located inside the Federal Reserve district specified and 0 if it is located in the border region of that same district but outside the border. For example, the Cleveland regression uses all the Cleveland border regions, and all counties within the Cleveland region around the border are assigned $At{l_k} = 1$ and those just outside Cleveland are assigned $At{l_k} = 0$. In the same figure, I plot the Atlanta and St. Louis estimates in red and purple, respectively, for reference. While some coefficients are statistically large – namely, Dallas and Cleveland counties in 1932 and Richmond counties that appeared worse off throughout – the baseline results for the Atlanta district are still relatively larger in 1929 and 1930 than in the other regions. Taken together with the permutation results above and the host of robustness exercises in the preceding section, it is likely that banking conditions in Atlanta were more favorable because of LLR policies.

Figure 5.

Estimates across Federal Reserve boundaries

Figure 5 Long description

Axes and markers. The vertical axis lists the years 1928 to 1933 from top to bottom. The horizontal axis shows the estimated difference in bank suspension rates between counties inside and outside each district’s border, from −0.2 to 0.25. A dashed vertical line marks zero. Each estimate is a dot with a horizontal 90 percent confidence interval.Districts. Six districts are shown: Kansas City, Dallas, Cleveland, Richmond, St. Louis and Atlanta. Atlanta and St. Louis are joined by lines across years.Atlanta. Atlanta is about +0.02 in 1928. It falls to about −0.07 in 1929 and −0.05 in 1930, then returns to near zero from 1931 onward.St. Louis. St. Louis moves the opposite way: about +0.07 in 1929 and +0.03 in 1930, then near zero.Richmond. Richmond is positive in most years, at about +0.05 in 1929, +0.07 in 1930 and +0.09 in 1932.Dallas and Cleveland. Dallas is about −0.13 in 1932, with a wide interval. Cleveland is about −0.05 in 1932.Kansas City. Kansas City is about +0.05 in 1928 and −0.04 in 1929.1933. All districts are close to zero.

Notes: This figure plots the estimated coefficients γi from Equation (2) using the Atlanta, St. Louis, Richmond, Cleveland, Dallas and Kansas City Federal Reserve border regions in separate regressions. The outcome variable is the annual suspension rate of all banks within a county. Control variables in Xjk include only border region by year fixed effects. Standard errors are clustered at the county level. 90 percent confidence intervals shown.

County manufacturing

Having established that credit constraints were strongly associated with manufacturing decline and that LLR policies improved banking stability, I now examine whether these financial interventions translated into better outcomes for manufacturing establishments. Using county-level data from the Census of Manufactures, I analyze economic activity along the Atlanta Federal Reserve border between 1929 and 1935. Because local banks were a critical source of working capital for manufacturers, particularly for small and medium-sized firms, we would expect to observe relatively better outcomes within the Atlanta district, where banks remained operational.

The empirical strategy follows the framework similar to Ziebarth (Reference Ziebarth2013), with county fixed effects and a comparison between 1929 and 1931. I extend the geographic scope beyond Mississippi and incorporate the entire border region. Specifically, I estimate the following specification:

(3)\begin{equation}y_{it}=\alpha_i+\alpha_1\,Atl_i\times{1931}_t+X_{it}+\epsilon_{it}\end{equation}

where $i$ denotes county, $t$ denotes the year, and $At{l_i}$ is an indicator variable that takes the value of 1 if the county is located within the Atlanta Federal Reserve district. The control variables in ${X_{it}}$ include state by year fixed effects in the main specification and, additionally, industry share by year fixed effects and 1927–9 trends by year in robustness exercises.

Table 5 reports the main results using just Mississippi counties (panel A) and the entire region (panel B).Footnote ⁸ Relative to counties just outside the Atlanta district, manufacturing output in counties within the Atlanta district was 21 to 26 log points higher in 1931 compared to 1929 (s.e. = 0.14). Given that the average US county experienced a 50 log point decline in manufacturing output between 1929 and 1931, this differential implies that LLR exposure offset roughly half of the national contraction. These effects are even more pronounced along other margins of firm activity: the number of wage earners was higher by 25–39 log points (s.e. = 0.13), and the number of establishments, in Mississippi, by 57 log points (s.e. = 0.21). These patterns suggest that LLR policies played a crucial role in preserving the extensive margin of production – keeping plants open and workers employed.Footnote ⁹

Table 5.

Manufacturing results at the county level

Table 5 Long description

Regression coefficients are reported for three outcomes, log output, log workers, and log establishments, each estimated over 1929 to 31 and over 1929 to 35, shown separately for Panel A Mississippi and Panel B all Atlanta Fed states. Panel A lists Year 1931 interaction with In ATL 1 as 0.26 with one star for log output in both samples with standard error 0.14, 0.39 with three stars for log workers in both samples with standard error 0.14, and 0.57 with three stars for log establishments in both samples with standard error 0.17. In Panel A for the 1929 to 35 columns, Year 1933 interaction with In ATL 1 is 0.16 for log output with standard error 0.17, 0.25 for log workers with standard error 0.17, and 0.64 with three stars for log establishments with standard error 0.17; Year 1935 interaction with In ATL 1 is 0.01 for log output with standard error 0.19, 0.17 for log workers with standard error 0.19, and 0.48 with three stars for log establishments with standard error 0.18. Panel A reports R sq values of 0.71 and 0.59 for log output, 0.61 and 0.33 for log workers, and 0.65 and 0.56 for log establishments, with N of 94 for 1929 to 31 and about 187 to 188 for 1929 to 35, and check marks for Year FE and County FE in all columns. Panel B lists Year 1931 interaction with In ATL 1 as 0.21 with three stars for log output in both samples with standard error 0.06, 0.25 with three stars for log workers in both samples with standard error 0.08, and 0.15 for log establishments in both samples with standard errors 0.09 and 0.10. In Panel B for the 1929 to 35 columns, Year 1933 interaction with In ATL 1 is 0.18 with two stars for log output with standard error 0.07, 0.23 with three stars for log workers with standard error 0.07, and 0.14 for log establishments with standard error 0.11; Year 1935 interaction with In ATL 1 is 0.08 for log output with standard error 0.07, 0.18 with two stars for log workers with standard error 0.07, and 0.12 for log establishments with standard error 0.09. Panel B reports R sq values of 0.66 and 0.01 for log output, 0.43 and 0.01 for log workers, and 0.46 and 0.00 for log establishments, with N of 717 for 1929 to 31 and about 1,423 to 1,435 for 1929 to 35, and check marks for State x Year FE and County FE in all columns.

Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01

Notes: This table reports the results of the regression in Equation (3). The balanced sample includes US counties with at least \$5,000 nominal dollars of manufacturing output in 1927–35, biennially, within Mississippi (panel A) and the entire Atlanta Federal Reserve border region (panel B). Standard errors are clustered at the county level.

When extending the panel through 1935, the magnitude of these differences declines over time, particularly in Mississippi, as policy differences across Federal Reserve districts diminished and the national economy began a slow recovery. Nevertheless, the coefficients remain positive and statistically significant through 1933, highlighting the early and localized nature of Atlanta’s intervention. Panel B of Table 5 expands the analysis to the full Atlanta border region and confirms these findings: the estimated impacts on output and employment remain stable, while the establishment effect, although still present, is somewhat smaller than in Mississippi alone.

These estimates are robust to concerns about differential industry composition, pre-existing growth trajectories, as well as potential interactions with New Deal programs instituted after 1932. I conduct various robustness checks and include year-by-county industry share fixed effects, control for manufacturing trends between 1927 and 1929, use 1927 as the base year, and exclude counties within 25 miles of the border to account for spillovers. I further directly control for New Deal spending programs, specifically from the AAA and RFC. Results are qualitatively unchanged under these alternative specifications (see Online Appendix Table A4).

Taken together, these results show that LLR policy exposure was associated with higher manufacturing output, employment and firm counts during the early Depression period. This stabilization of local industrial activity suggests that emergency liquidity support helped preserve labor demand in affected counties. In the next section, I examine whether these effects were reflected not only in persistently higher retention and entry, but also in changes to the composition of the manufacturing workforce.

Labor response

The preceding sections establish that the Atlanta Federal Reserve’s lender of last resort (LLR) policies supported the manufacturing firms at the onset of the Great Depression. However, these findings reflect outcomes over a relatively short window (1929–33) and do not speak directly to the persistent labor market consequences of LLR intervention. To assess whether early financial stabilization translated into durable gains for local workers, I turn to individual-level panel data from the full-count 1930 and 1940 US censuses.

This section uses census records in the three states bisected by the Atlanta federal reserve border – Mississippi, Tennessee, and Louisiana – to examine how 10-year employment trajectories varied depending on whether individuals resided in counties exposed to Atlanta’s LLR policies at the start of the Depression.Footnote ¹⁰ Because the 1940 outcomes are observed well after the banking crisis subsided and after Atlanta’s emergency liquidity provision had ceased, this exercise tests for persistent, long-run effects of LLR on local labor markets and mobility.

Specifically, I estimate the following regression:

(4)\begin{equation}Y_{i(j)}=\beta Atl_j+X_i+\Omega_j+\epsilon_i\end{equation}

where outcome $Y$ is observed in 1940 for individual $i$ who resides in county $j$ as of 1930. The variable $At{l_j}$ takes the value of 1 if county $j$ is within the Atlanta Federal Reserve district. At the individual level, the controls in ${X_i}$ include age, sex, the logarithm of occupational income score in 1930, and state of residence in 1930 and 1940 fixed effects. To account for persistent county-level manufacturing trends driven by unobservable $local$ characteristics, ${{{\Omega }}_j}$ includes county-level retention, entry and migration rates between 1910 and 1920, which I compute using the same linking and sample selection methods as for the main sample. In essence, this specification compares 10-year outcomes of two individuals who were employed in the same state in 1930, in counties with similar historical trends of manufacturing retention and exit, where one lived in a county that received LLR support during the banking crises and the other did not. I cluster the standard errors at the county level.

Table 6 presents the regression results of $\beta $ for three outcome variables: (A) an indicator for employment in manufacturing in 1940, (B) the log change in occupational income score between 1930 and 1940, and (C) indicators for inter-county and inter-state migration. In Panels A and B, columns 1 and 3 restrict the sample to individuals employed in manufacturing in 1930, while columns 2 and 4 focus on those not in manufacturing in 1930. Panel C uses only 1930 manufacturing workers to assess migration outcomes. For each panel, I report results separately for Mississippi (columns 1–2) and for the pooled sample across all three states (columns 3–4).

Table 6.

Labor response, 1930–40

Table 6 Long description

Three panels report regression results for outcomes labeled, manufacturing in 1940, log occscore change, and migration, each with columns split by MS versus MS plus TN plus LA and by Manu 1930 versus Non manu 1930 where shown. Panel A lists In atl one coefficients of 0.029 with standard error 0.017 for MS Manu 1930, 0.007 with 0.003 for MS Non manu 1930, 0.018 with 0.020 for MS plus TN plus LA Manu 1930, and 0.004 with 0.002 for MS plus TN plus LA Non manu 1930; E of Y is 0.30, 0.06, 0.37, and 0.07, with R sq 0.02, 0.04, 0.03, and 0.04, and N 29,458, 273,534, 162,446, and 919,399. Panel B lists In atl one coefficients of minus 0.023 with 0.014 for MS Manu 1930, 0.022 with 0.009 for MS Non manu 1930, minus 0.009 with 0.009 for MS plus TN plus LA Manu 1930, and 0.016 with 0.011 for MS plus TN plus LA Non manu 1930; E of Y is minus 0.07, 0.17, minus 0.04, and 0.14, with R sq 0.14, 0.44, 0.15, and 0.40, and N 27,335, 234,149, 146,770, and 801,170. Panel C reports Manu 1930 only, with outcomes P move county and P move state across four columns; In atl one is minus 0.013 with 0.020 and minus 0.004 with 0.019 for the first two columns, and minus 0.035 with 0.011 and minus 0.029 with 0.011 for the third and fourth columns; E of Y is 0.38, 0.27, 0.17, and 0.13, with R sq 0.35, 0.35, 0.01, and 0.03, and N 29,458, 162,446, 29,460, and 162,447. Rows for Individual Controls, County Controls, State, FE, and State FE show check marks in all columns where listed.

Standard errors in parentheses. *p < 0.10, ^** p < 0.05, ^*** p < 0.01

Notes: This table reports the results of the regression in Equation (4). The sample includes all 1930–40 linked working-age individuals (men and women ages 18–55 in 1930) residing in the three states bisected by the Atlanta Federal Reserve district: Mississippi, Tennessee and Louisiana. Individual-level controls include age, sex, state of residence in 1930 and 1940, the logarithm of occupational income score in 1930. County-level controls include historical manufacturing retention (columns 1/3 in panels A and B), entry (columns 2/4 in panels A and B) and migration rates (panel C) computed using linked census data between 1910 and 1920. Columns ‘Manu 1930’ use only manufacturing workers in 1930 in the regression, while ‘Non-manu 1930’ use those not in manufacturing. Standard errors are clustered at the county level.

Panel A shows that among 1930 manufacturing workers in Mississippi, those residing in Atlanta district counties were 2.9 percentage points more likely to be employed in manufacturing by 1940 (s.e. = 0.017), a 9.7 percent increase relative to the group mean of 30 percent. For non-manufacturing workers in 1930, LLR exposure was associated with a 0.7 percentage point increase in entry into manufacturing by 1940 in Mississippi (column 2), and a 0.4 percentage point increase across the full sample (column 4). Scaled relative to the baseline manufacturing share, these estimates imply increases of 9.4 percent in Mississippi and 4.1 percent in the pooled sample.Footnote ¹¹ Approximately 30 percent of the manufacturing gain in Mississippi was driven by retention, and 70 percent by new entry. Given that the policy treatment occurred nearly a decade prior, these results suggest that LLR support had long-lasting effects on local labor structure.

Panel B reports results for the log change in occupational income score between 1930 and 1940. The occscore is a commonly used proxy for labor earnings based on occupation-specific median income as of 1950, retroactively applied to occupations in previous censuses. For retained manufacturing workers, I do not find statistically significant evidence that LLR exposure led to greater occupational mobility. However, for those in Mississippi, I find a 0.022 log point increase (s.e. = 0.009) for new entrants into manufacturing, or about 13 percent of the average increase in the sample between 1930 and 1940.

Panel C turns to migration outcomes for workers employed in manufacturing in 1930. LLR exposure did not significantly affect inter-county migration (columns 1 and 2) but did significantly reduce interstate migration (columns 3 and 4), suggesting that LLR policies anchored workers locally. These results suggest that LLR intervention during the banking crisis had persistent impacts on individual employment outcomes. Residents of LLR counties were more likely to remain in or transition into manufacturing jobs a decade after the policy was implemented, and less likely to migrate out of state.

Why did this relatively brief episode in 1930–1 generate persistent differences in labor market outcomes observed in 1940? The patterns in Table 6 suggest several complementary mechanisms. Panel A shows that long-run manufacturing gains arise from both retention of incumbent workers and new entry, consistent with the idea that manufacturing production networks are costly to reestablish: early lender-of-last-resort intervention preserved local employment relationships, capital and organizational capacity, while firm exit was permanent and restarting activity substantially more costly. Panel B indicates that new entrants into manufacturing experienced some occupational upgrading, suggesting that surviving firms continued to generate relatively stable, higher-quality employment opportunities over time. Panel C further shows that LLR exposure reduced interstate migration among 1930 manufacturing workers, implying that early stabilization anchored workers locally and limited spatial reallocation. Together, these results are consistent with slow-moving labor market adjustment frictions and local general equilibrium effects, whereby short-run differences in financial conditions and firm survival translate into persistent divergence in individual employment trajectories even as aggregate manufacturing outcomes partially converge.

Compositional effects of the labor response

To understand whether LLR exposure changed the composition of the manufacturing workforce, I test for heterogeneous effects across observable characteristics. I extend Equation (4) by interacting the LLR treatment indicator $At{l_j}$ with individual-level characteristics (e.g. age, education, gender, income score, urban status). I then compute predicted probabilities of remaining in or entering manufacturing for different subgroups. Figures 6 and 7 plot these probabilities.Footnote ¹²

Figure 6.

Predicted probabilities of manufacturing retention by subgroups

Figure 6 Long description

Each panel compares counties inside the Atlanta district (ATL = 1) with counties outside it (ATL = 0), with 95 percent confidence intervals. The vertical axis is the predicted probability, from 0.20 to 0.46. (a) By 1930 occupational score tercile. Both lines form a V shape. Retention is about 0.37–0.38 in the first tercile, 0.26–0.27 in the second and 0.38–0.40 in the third. ATL = 1 is slightly higher throughout.(b) By age. Retention declines with age, from about 0.40 (ATL = 1) and 0.37 (ATL = 0) at ages 18–25 to about 0.32 and 0.31 at ages 41–55.(c) By nativity. Both lines are nearly flat at about 0.36. ATL = 1 is slightly higher for native-born workers.(d) By race. The two groups are equal at about 0.36 for non-white workers. For white workers, ATL = 1 is about 0.38 and ATL = 0 about 0.35.(e) By urban status. The lines cross. Among rural workers, ATL = 0 is higher (about 0.38 against 0.36). Among urban workers, ATL = 1 is higher (about 0.38 against 0.34).(f) By education in 1940. Retention falls from about 0.38 with fewer than 8 years or exactly 8 years of schooling to about 0.30 with more than 12 years. ATL = 1 is slightly higher in each category.(g) By sex. For women, ATL = 1 is about 0.32 and ATL = 0 about 0.26. For men, both are about 0.38–0.39.(h) By location in 1940. The lines cross. Among workers not living in the Atlanta district in 1940, ATL = 0 is about 0.36 and ATL = 1 about 0.26. Among those living in the district in 1940, ATL = 1 is about 0.39 and ATL = 0 about 0.30.

Notes: This figure plots the predicted probabilities of retention for various subgroups of manufacturing workers inside (ATL = 1) and outside (ATL = 0) the Atlanta district border based on the regression in Equation (4). The sample includes all working-age individuals (men and women ages 18–55 in 1930) residing in the three states bisected by the Atlanta Federal Reserve district: Mississippi, Tennessee and Louisiana. Individual level controls include age, sex, state of residence in 1930 and 1940, the logarithm of occupational income score in 1930. County-level controls include historical manufacturing retention computed using linked census data between 1910 and 1920. Standard errors are clustered at the county level.

Figure 7.

Predicted probabilities of manufacturing entry by subgroups

Figure 7 Long description

Each panel compares counties inside the Atlanta district (ATL = 1) with counties outside it (ATL = 0), with 95 percent confidence intervals. The vertical axis is the predicted probability, from 0 to 0.12. (a) By 1930 occupational score tercile. Entry is lowest in the middle tercile. It is about 0.06 in the first tercile, 0.033–0.038 in the second and 0.067–0.075 in the third. ATL = 1 is slightly higher in the second and third terciles.(b) By race. Entry is about 0.07 for both groups. ATL = 1 is slightly higher for white workers.(c) By urban status. Entry rises from about 0.07 for rural workers to about 0.09 for urban workers. ATL = 1 is slightly higher in both.(d) By nativity. Entry is about 0.065–0.076. ATL = 1 is higher, most clearly for non-native workers.(e) By education in 1940. Entry is about 0.06 up to 12 years of schooling. Beyond 12 years it drops to about 0.04 for ATL = 1 and 0.05 for ATL = 0.(f) By age. Entry falls with age. At ages 18–25 it is about 0.117 (ATL = 1) against 0.102 (ATL = 0). At ages 26–40 both are about 0.07. At ages 41–55 it is about 0.04 for ATL = 1 and 0.046 for ATL = 0.(g) By location in 1940. Among workers living in the Atlanta district in 1940, ATL = 0 is about 0.115 and ATL = 1 about 0.076. Among those living elsewhere, both are about 0.07–0.08.(h) By sex. Entry rises from about 0.03 for women to 0.08–0.09 for men. ATL = 1 is slightly lower for women and slightly higher for men.

Notes: This figure plots the predicted probabilities of entry by 1940 into manufacturing for various subgroups inside (ATL = 1) and outside (ATL = 0) the Atlanta district border based on the regression in Equation (4). The sample includes all working-age individuals (men and women ages 18–55 in 1930) residing in the three states bisected by the Atlanta Federal Reserve district: Mississippi, Tennessee and Louisiana. Individual level controls include age, sex, state of residence in 1930 and 1940, the logarithm of occupational income score in 1930. County-level controls include historical manufacturing retention computed using linked census data between 1910 and 1920. Standard errors are clustered at the county level.

Figure 6 presents predicted probabilities of staying in manufacturing between 1930 and 1940 by subgroup. I find limited compositional differences across the Atlanta border among stayers. In both types of counties, workers in the 1st and 3rd terciles of the occupational income distribution had higher retention than those in the middle tercile. Retention was also higher among younger, male and less-educated workers. Notably, urban workers in Atlanta counties saw greater retention, likely due to selective out-migration elsewhere.

In contrast, Figure 7 shows clear compositional effects among new entrants into manufacturing. Entry was significantly more likely for individuals who were younger, male and had lower educational attainment in Atlanta counties. This pattern suggests that LLR support enabled broader access to manufacturing jobs throughout the Depression, especially for low-skilled workers.