Bank Distress and Manufacturing: Evidence from the
Great Depression
James Lee and Filippo Mezzanotti∗
November 2015
Abstract
Using newly-digitized, city-industry-year level records from the 1923-1937 US Censuses of Manufactures, we examine the influence of the financial sector on the real
economy. We do so in the context of the Great Depression, a period in which many
banks were suspended and the manufacturing sector, which comprised 30 percent of
the US economy, declined significantly. In our research design, we measure whether industries with high levels of pre-Depression external finance dependence declined in employment, output, and establishment growth by more than industries with low levels of
pre-Depression external finance dependence from 1929-1933, in cities with higher 19291933 bank suspension rates. We control for city-time shocks, industry-time shocks, and
other, non-finance industry characteristics interacted with bank suspension rates. We
find that high external finance dependence industries contracted by 21 to 27 percent
more than low external finance dependence industries following bank suspensions. For
robustness, we instrument for bank suspensions with a measure of trust—religious
fragmentation—and a measure of real estate price growth. Given the pre-Depression
share of high external finance industries and the magnitude of the 1929-1933 bank
suspensions, we estimate that approximately 22 percent of the total decline in manufacturing employment from 1929-1933 was due to the banking crisis. We conclude with
evidence that the negative effects of the banking shock persisted into the late 1930s on
the establishment outcome.
∗
We thank Claudia Goldin, Samuel Hanson, Victoria Ivashina, Josh Lerner, Robert Margo, Kris Mitchener, Diana Moreira, Ramana Nanda, Tom Nicholas, Nathan Nunn, Hugues Pirotte, Daniel Pollmann, David
Scharfstein, Adi Sunderam, and seminar participants at Harvard University and the EABH Workshop for
New Scholars in Financial History for helpful comments. We thank the Lab for Economic and Policy Applications at Harvard University and the Harvard Business School Historical Collections at Baker Library
for financial support. All errors are our own. Cornerstone Research and Harvard University, Department of
Economics and Business School. Email: [email protected], [email protected].
1
Introduction
After nearly a decade of expansion in the 1920s, the US manufacturing sector contracted
by more than 30 percent from 1929 to 1933 (see Figure 1). At the same time, overall US
gross domestic product fell by 28 percent and the national unemployment rate rose from 3
to 25 percent.1 By almost any measure, the real economy suffered.
During those same years, more than 7,600 banks—32 percent of the 1929 total—were
suspended, and total deposits fell by 16 percent (see Figure 2).2 Never before—or since—
had the financial sector seen so many institutions close (see Figure 3). Hence, from 1929-1933,
the financial sector suffered as well.
With our newly digitized data, we are interested in understanding how much of the manufacturing decline was a result of the financial crisis. More broadly, we seek to understand
the influence of the financial sector on the real economy.
In theory, financial distress can affect the real economy through three channels. Fewer
banks reduce the money supply thereby increasing borrowing costs for individuals and nonfinancial firms.3 Fewer or distressed banks also increase the cost of credit intermediation by
reducing the net worth of lenders and borrowers.4 Third, and relatedly, bank distress and
failure destroy valuable firm-bank relationships. This too raises the cost of intermediation.5
Through any of these three channels financial distress can affect the real economy. We test
for such effects using newly digitized data covering the US Great Depression.
We choose the Depression as our laboratory because the period contained such an unprecedented—
and un-repeated—level of financial distress, which also varied across space in ways we can
measure with county level bank data from the FDIC (see Figure 4). Moreover, this spatial
1
Gross domestic product data are from the US Bureau of Economic Analysis. Employment data are from
Margo [1993].
2
Suspended is defined by the FDIC as closed temporarily or permanently by supervisory authorities or
the banks’ board of directors on account of financial difficulties.
3
Friedman and Schwartz [1963].
4
Bernanke [1983].
5
Petersen and Rajan [1994].
1
variation was meaningful because financing remained a fairly local practice due to technological and regulatory constraints such as the McFadden Act, which allowed states to prohibit
banks from operating branches beyond state lines.6
We use the manufacturing sector as our measure of real economic activity because it
comprised 30 percent of the US economy at the time and we can similarly measure outcomes
across space—and industry—using newly-digitized, highly-detailed, city-industry-year level
records from the 1923-1937 Census of Manufactures and industry level data from Moody’s
Manual of Industrials.7
Our research design exploits cross-city variation in the financial distress data and crosscity, cross-industry variation in the manufacturing data to identify econometrically the causal
effect of financial distress on manufacturing activity. In particular, we test whether from 1929
to 1933 manufacturing industries with high levels of pre-Depression external finance dependence (bank loans as a percentage of assets) experienced larger declines in establishment,
employment, and valued added growth than industries with low levels of pre-Depression external finance dependence in cities with larger bank suspension rates (fraction of 1929 banks
suspended between 1930 and 1933). We control for city-time shocks, industry-time shocks,
and non-finance industry characteristics such as capital intensity, skill intensity, contract
intensity, and establishment size. For robustness, we instrument for bank suspensions with
an inverse measure of trust—religious fragmentation—as well as a real estate instrument
similar to prior work.8
We believe this combination of a triple-difference estimator and instrumental variables
robustness checks significantly advances the identification of the effect of bank distress on
the real economy. First, the design allows us to overcome the concern in previous papers
of a geography-level unobservable driving results.9 Indeed, our city-industry-year level data
6
Nestor [1992], Mitchener and Wheelock [2013].
The Moody’s Manual of Industrials data are from Nanda and Nicholas [forthcoming].
8
Our real estate instrument follows the logic of the instrument in Calomiris and Mason [2003] and is
similar in construction to the instrument in Mladjan [2012].
9
See, for example, Calomiris and Mason [2003].
7
2
permit us to control flexibly for not only city-level unobservables, but also city-time and
industry-time level unobservables. This removes many threats to identification. In addition,
compared to other work studying this question using geography-industry-year level data, our
paper is a substantial improvement because we capture more than 70 percent of 1929 US
manufacturing—a far larger proportion than earlier research covered—and a much richer set
of industries, particularly the most technologically advanced and innovative industries found
in cities.10 Finally, the newly digitized, multi-dimensional, comprehensive outcome data
we have on establishments, employment, and value added enable us to develop a broader
understanding of the effects of Great Depression financial distress on the manufacturing
sector. Earlier papers focused on primarily the output outcome.11 Hence, our novel design
and new data move forward substantially the identification of the impact of bank distress
on real economic activity.
Our results using this rigorous design on the newly-digitized data both confirm and
meaningfully extend the conclusion of previous work showing that the financial sector has
a significant effect on the real economy. In confirmation, we find that high external finance
dependence industries contracted by 21 to 27 percent more than low external finance dependence industries as a result of bank distress during the Great Depression. Given the
pre-Depression share of high external finance manufacturing industries and the magnitude
of the 1929-1933 bank suspensions, we estimate that approximately 22 percent of the total
drop in manufacturing employment from 1929-1933 was due to bank distress. This provides
a potential source of the bank distress-induced reduction in income and construction activity Calomiris and Mason [2003] find. It is also consistent with the findings of Mladjan
[2012] and Nanda and Nicholas [forthcoming] who find negative effects of financial distress
on manufacturing output and R&D, respectively.
In extension, we are the first to find a significant relationship between bank distress
10
See, for example, Ziebarth [2013] who studied manufacturing in only a subset of industries all located
within the state of Mississippi, and Mladjan [2012], who also used only a subset of all manufacturing industries, measured at state-level.
11
See, for example, Mladjan [2012].
3
and manufacturing employment. Using a subset of manufacturing industries in Mississippi,
Ziebarth [2013] did not find such a relationship. We believe the difference lies in our sample
covering all manufacturing industries in the 29 largest US cities versus his sample covering
a subset of industries in Mississippi. In addition, exploiting the longer-term nature of our
data, we also are the first to show that the negative effect of bank distress persisted in
reducing the number of manufacturing establishments through 1937. We do not find the
same results for employment or value added. To the extent that establishments represent new
monetary outlays for firm creation, which may be larger than outlays for simply increasing
utilization among existing plant capacity, we believe these results suggest that manufacturing
investment may not have recovered as quickly as manufacturing output following the financial
crisis. This has policy implications, which we discuss. Hence, in addition to confirming the
conclusion of previous work that the financial sector mattered greatly for real economic
activity during the Great Depression, we meaningfully extend that work by showing which
dimensions of manufacturing—establishments, employment, and value added—were most
affected and for how long.
The rest of the paper proceeds as follows. Section 2 reviews relevant literature. Section 3
describes our newly-digitized data. Section 4 explains our empirical strategy. Section 5 discusses results and performs robustness checks. Section 6 addresses the long-term persistence
of the negative effects of financial distress on the manufacturing sector. Section 7 concludes.
2
Related Literature
Our examination of the importance of financial distress on the real economy in the context
of the US Great Depression relates closely to three literatures.
The first uses microeconomic data to understand the effect of financial contraction on
real economic outcomes during the US Great Depression. Calomiris and Mason [2003] began
this literature by aggregating bank-level balance sheet data to the county and state levels
4
to estimate the effect of loan supplies on state and county income and construction activity.
Instrumenting for the 1929-1933 drop in loan supplies with pre-crisis assets (bank size) and
the share of assets in real estate (bank exposure to 1920s agricultural land price declines),
the authors found that both income and construction activity fell when local loan supplies
contracted. Similarly, Mladjan [2012] used state level manufacturing and banking data
on a subsample of US industries and found that bank failures led to at least a third of
the Great Depression drop in manufacturing output. Like Calomiris and Mason [2003],
Mladjan [2012] instrumented for bank failures using two pre-Depression, state-level bank
vulnerability measures: the change in farmland value from 1910-1920 and the percentage of
1920 banks that were branch offices of a bank headquartered somewhere else in the state.
More recently, Ziebarth [2013] tested whether bank distress affected manufacturing during
the Great Depression using plant-level manufacturing records from Mississippi and exploiting
the quasi-experimental administrative arrangement of two regional Federal Reserve offices—
St. Louis and Atlanta—supervising different portions of the state.12 He found that bank
failures had a large, negative impact on firm revenue, but no effect on employment. Finally,
Nanda and Nicholas [forthcoming] collected firm-level R&D data to show that bank distress
negatively affected innovation among firms smaller in size and operating in industries with
higher levels of pre-Depression dependence on external finance. Together, these four papers
begin to illustrate the effects of the financial sector on real economic activity during the US
Great Depression. With our newly-digitized, city-industry-year level data covering 70 percent
of the manufacturing sector in 1929, we bring highly-detailed, comprehensive information
and a rigorous identification strategy to this literature.
As a result, we are also able to contribute to a second, related literature which studies the
influence of the local credit supply on real economic activity more generally—outside of just
the Great Depression. Peek and Rosengren [2000] and Ashcraft [2003] initiated this literature
by showing a negative response of local income and construction activity to adverse credit
12
Richardson and Troost [2009] describe how the St. Louis and Atlanta regional Federal Reserve offices
handled the 1929 banking crisis differently.
5
shocks in the 1980s and 1990s. More recently, Gilje [2013] used the shale gas and oil boom to
show that at the county level, sectors more dependent on external finance grow faster than
sectors less dependent on external finance following a sharp increase in county bank deposits.
Most recently, Chodorow-Reich [2014] used the 2008 failure of the Lehman Brothers bank to
measure the impact of a disruption in lending on employment at manufacturing firms. He
found that small- and medium-sized manufacturing firms were more prone than large firms
to cut employment following a shock to their banking partner. We share a triple difference
identification strategy with Gilje [2013] and a manufacturing focus with Chodorow-Reich
[2014]. With both, we confirm and extend into new dimensions and time periods the result
that the local credit supply matters for local economic activity. In doing so, we contribute
meaningfully to the literature examining the influence of the local credit supply on real
economic activity generally.
Finally, because our data contain specific information on the number of establishments
in each manufacturing industry within each city, we also add to a third, burgeoning literature examining the relationship between banking sector dynamics and entrepreneurship.
Evans and Jovanovic [1989] were early contributors to this literature when they showed the
importance of liquidity constraints on business formation. Kerr and Nanda [2009] and Kerr
and Nanda [2011] extended this finding to banking in particular—showing that bank lending
serves as an important funding source for new firms. Our examination of the impact of bank
distress on manufacturing establishments specifically—in both the short and long runs—
complements all three of these papers and thus adds to the literature on the importance of
banking for entrepreneurship.
Hence, our paper relates and contributes importantly to three literatures examining the
influence of financial distress on the real economy.
6
3
Data
To carry out our analysis, we use three data sources.
The first is the US Census of Manufactures. We digitize all city-industry level records from
the bi-annual (except for 1931) 1923-1937 Censuses.13 We limit our sample to the 29 cities
with industry level data in both 1929 and 1933. A list of these cities is in Table 9. Figure
5 shows their spatial distribution across the US. Within these cities, the Census reports
283 unique manufacturing industries.14 The data report for each city-industry-year triplet
the number of establishments, value added, and employment. These are our manufacturing
outcomes.
Onto these manufacturing data we merge bank data from the FDIC. The bank data report
for each year from 1920-1936 the number of banks and number of bank suspensions in every
US county.15 We aggregate these county level data to the city level using the county-to-city
mappings provided in the Censuses of Manufactures.16 We create our primary measure of
bank distress as the fraction of all 1929 banks in the city that were suspended between 1930
and 1933:
P1933
Shockc =
t=1930
SuspendedBankct
Banksc1929
13
(1)
The city-industry records from the 1931 Census of Manufactures cannot be located by the US Census
Bureau, the National Archives, or the Library of Congress.
14
The 283 industries are based on 1929 industry classifications. In order to have a consistent set of
industry classifications across years, we map all industry classifications from years outside of 1929 into 1929
classifications. We do this using the year-to-year industry mappings and national industry employment data
found in each Census. For our regression analysis, we also use the higher level industry groupings provided
in the 1929 Census. There are two such groupings. The first places each industry into one of 16 broad
manufacturing groups such as “Food and Kindred Products.” These groups are pre-cursors to the 2-digit
Standard Industrial Classification (“SIC”) codes, which the Census of Manufactures began using in 1939.
The second places each industry into one of 34 groups. These groups are roughly pre-cursors to the 3-digit
SIC codes. We call these groups the 2-digit and 3-digit SIC groups, respectively, going forward.
15
A bank suspension is defined by the FDIC as the closure of a bank to the public either temporarily or
permanently by supervisory authorities or by the banks’ board of directors on account of financial difficulties
(see Reserve [September 1937]).
16
Each city in the Census of Manufactures is a set of contiguous counties. We simply sum the county level
banking data across each city’s component counties to construct the city’s banking data. Results are robust
to using city definitions from different Census years. We use city definitions from a later year—1977—in
our main analysis because we want to capture bank lending opportunities that may have existed outside the
1929 central city counties. Using only the 1929 central cities counties does not significantly change results.
7
Lastly, to see whether these suspensions affected different industries differently based on
pre-Depression industry financing technology, we merge onto our manufacturing and banking
data an industry-level measure of pre-Depression external finance dependence. The measure
is the fraction of assets that are bank loans. It is constructed by Nanda and Nicholas [forthcoming] who aggregate firm-level data from the 1920s Moody’s Manual of Industrials into
industry-level measures for 16 different industries. We map these 16 industries into our 17
2-digit SIC classifications based on industry name, and then assign each of our 283 industries
an external finance dependence value based on the industry’s 2-digit SIC classification. We
create an indicator variable equal to 1 if an industry’s external finance dependence measure
is above the median measure of 0.06. We use an indicator rather than a continuous measure
to avoid specifying a parametric relationship between external finance dependence and our
manufacturing outcomes (see Figure 12 for the distribution of external finance dependence
across industries in our sample).
With these manufacturing, banking, and external finance dependence data, we carry out
our analysis examining the importance of bank distress on the manufacturing sector during
the US Great Depression.
4
Empirical Strategy
Establishing a causal relationship of bank distress on manufacturing outcomes is empir-
ically challenging. The negative correlation seen in Figure 6 between 1929-1933 city bank
suspension rates and the contemporaneous change in value added could result from either
an omitted variable or reverse causality rather than from bank distress causing the manufacturing decline. Two particularly worrisome omitted variables are city and industry specific
shocks. An example of a city shock is municipality distress. Municipality distress could
lead to both bank suspensions and manufacturing declines if banks hold the municipality’s
bonds and the municipality suspends bond payments while also reducing the quality and
8
quantity of city services local manufacturers use such as roads, water provision, and other
infrastructure.17 An example of an industry shock is a major drop in demand for a particular
industry’s output. If the industry is more prevalent in some cities than in others, such a
drop would lead to a larger output decline in cities with a larger presence of the industry.
If these cities also experienced greater bank distress because of larger exposure to this declining industry, then a spurious relationship could exist between the manufacturing decline
and bank distress.
Perhaps even more worrisome than a spurious relationship is one of reverse causality:
manufacturing decline leading to bank distress. This could happen through an asset or
liability channel. The asset channel would be manufacturers with outstanding loans to
local banks suffering, which then impairs the banks’ balance sheets and in turn leads to
bank distress. On the liability side, if manufacturers suffer and the local economy generally
weakens, demand for deposits may increase as unemployed or otherwise cash-constrained
individuals demand liquidity for consumption. This could drain liquid resources from banks
and thereby increase the cost of funding and likelihood of distress. Hence, through either an
asset or liability channel, declines in local manufacturing could lead to local bank distress—a
reverse causality story opposite of bank distress leading to manufacturing declines.
We address both the omitted variable and reverse causality concerns in our empirical
design below. In doing so, we seek to provide causal evidence of the importance of the credit
supply channel in driving manufacturing outcomes during the Great Depression.
4.1
Fixed Effects Estimator
We begin by constructing an OLS estimator which compares the relative growth from 1929
to 1933 of manufacturing industries with high pre-Depression dependence on external finance
to those with low pre-Depression dependence on external finance, across cities characterized
17
Joffe [2012] documents the commonality of municipality distress during the Great Depression by noting
that more than 4,800 municipalities defaulted on bond payments.
9
by different bank suspension rates. We include city-year and 3-digit SIC industry group-year
fixed effects to control non-parametrically for potential omitted variables at the city-year and
industry group-year levels. This includes any city or industry shocks discussed above, as well
as any of those shocks which also vary over time.
We compare across pre-Depression high and low external finance dependence industries
for two reasons—one theoretical, one empirical. Theoretically, as Rajan and Zingales [1998]
note, the more developed a local financial sector, the lower the costs of local asymmetric information. This fosters investment and growth disproportionately in industries whose
funding technology relies more on external sources. Hence, when a local financial sector
experiences a negative shock, local industries more dependent on external financing are generally more negatively affected than local industries less dependent on external financing.18
Empirically, comparing across high and low external finance dependence industries within a
city relaxes the identification assumption that manufacturing industries would have evolved
similarly across cities but for the bank shock—an assumption on which previous work relied.19 Instead, our setup requires only that the differential growth rates between high and
low external finance dependent industries would have evolved similarly across cities, but for
the bank shock. This is a much weaker, and therefore more plausible, assumption.
This framework generates the estimating equation:
ln Ycit = βHF inj(i) P ostt Shockc + γHF inj(i) Shockc + αct + αj(i)t + εcit
(2)
where Ycit denotes manufacturing outcome Y = {establishments, value added, employment}
in city c and industry i at time t; HF inj(i) is an indicator equal to 1 if 3-digit SIC industry
group j, which is a function of the observation’s industry i, is a high external finance dependence industry;20 P ostt is in indicator equal to 1 if the year is 1933 (the other year is 1929,
18
Petersen and Rajan [1994] and Ivashina and Kovner [2011] show that when this financial shock occurs
in the banking industry in particular—as opposed to another part of the financial sector—the real economy
suffers as firm-bank specific relationships are disrupted.
19
See, for example, Calomiris and Mason [2003].
20
We follow exactly Nanda and Nicholas [forthcoming] in the definition of this variable.
10
for which P ostt = 0); Shockc is the fraction of 1929 banks in city c suspended between 1930
and 1933, as detailed in equation (1); αct are city-year fixed effects; αj(i)t are industry group
j-year fixed effects; and εcit is the error term. The sample is all city-industry-year triplets
in the 29 cities with industry-level data in both sample years, 1929 and 1933.21 We cluster
standard errors at the city level, the level of the financial shock treatment. Theory predicts
β < 0.
In addition to removing the effects of city-time and industry-time shocks, as mentioned
above, this specification reduces the potential for reverse causality. Specifically, it reduces
the liability channel of reverse causality—that manufacturing declines increase the likelihood
of a bank run—because the only way a deteriorating manufacturing sector can increase the
likelihood of a bank run in equation (2) is by increasing the likelihood of a run differentially across high and low external finance dependence industry deteriorations. This seems
unlikely. Instead, it seems more likely that a deterioration in either group would lead to a
similar likelihood of a bank run. Thus, our specification also significantly reduces the reverse
causality threat coming from the liability channel.
Finally, to ensure that our external finance dependence measure is capturing true external
finance dependence and not some other industry characteristic, we estimate equation (2)
with four separate, non-external finance dependence industry characteristics interacted with
the bank suspension rate and an indicator for the year being 1933. The first is contract
intensity. It is taken from Nunn [2007] and pertains to 4-digit SIC code industries in 1963.
We map this 1963 measure back to our 1929 4-digit SIC code industries using information
on Census-to-Census industry mappings published in the Censuses of Manufactures from
1929 to 1963. The measure is the fraction of an industry’s inputs that are neither sold on
an organized exchange nor listed in a reference price trade journal. The higher the fraction,
the more contract-intensive the industry because input purchases are more likely to occur
in a relationship-specific—and therefore contract-inducing—way. We create an indicator
21
We include all city-industry-year triplets in these cities, even if the city-industry is in the data in only
year 1929 or only year 1933.
11
for whether an industry is above the median contract intensity value across all industries.
We then interact this indicator with the interaction of our bank suspension measure and
an indicator for the year being 1933—just as we did with our external finance dependence
measure.
Our second non-external finance dependence industry measure that we interact with the
bank suspension rate and an indicator for year 1933 is capital intensity. We obtain this
at the 4-digit SIC industry level from the 1919 Census of Manufactures. The 1919 Census
of Manufactures was the last pre-Depression Census of Manufactures to report total capital
stock at the industry level. We divide total capital stock by industry value added to construct
our industry-level capital intensity measure. We map the 1919 industries into 1929 industries
using the industry mapping described above. We then interact the bank suspension rate and
an indicator for year 1933 with an indicator for whether an industry is above the median
capital intensity, as we did with our other two industry measures.
Our third non-external finance dependence industry measure is skill intensity. We construct this directly from the 1929 Census of Manufactures at the 4-digit SIC industry level.
We calculate it as the fraction of an industry’s employees that is not production workers.
Such workers are either “proprietors and firm members” or “salaried officers and employees.”
We consider these workers skilled. Hence, the more of these workers an industry has relative
to its production workers, the more skill-intensive the industry. As with our other industry
measures, we interact an indicator for whether this measure is above the median measure
value with the bank suspension rate and an indicator for year 1933.
Our fourth and final non-external finance dependence industry measure is establishment
size. As with the skill intensity measure, we construct this directly from the 1929 Census of
Manufactures at the 4-digit SIC industry level. We calculate it as the number of industry
employees per establishment. Hence, higher levels mean larger establishments, on average.
We interact an indicator for whether this measure is above the median measure value with the
bank suspension rate and an indicator for year 1933. Combined with the contract intensity,
12
capital intensity, and skill intensity measures, this establishment size measure thus provides
a further test that our external finance dependence measure is indeed capturing industry
dependence on external finance and not some other characteristic.
4.2
Robustness: Instrumental Variables Estimator
Even with these interacted controls, the asset channel of reverse causality may still
threaten identification. Banks could hold disproportionately more loans from firms in high
external finance dependence industries than from firms in low external finance dependence
industries. Hence, when high external finance dependence industries experience a negative
shock, banks may also be more likely to experience a negative shock. The industry groupyear fixed effects in equation (2) control for such shocks. If such shocks, however, are also
city-specific—that is, they vary at the city-industry level, not just the industry level—then β
in equation (2) could be biased. This could happen if two different high external finance dependence industries—say automobiles and textiles—experienced differential demand shocks
and the importance of the two industries varied greatly across two cities—say with Detroit,
MI having a relatively larger share of its manufacturing sector in the automobile industry
than Providence, RI, which had a relatively larger share of its manufacturing sector in the
textiles industry. For then bank distress in the city with the more negatively shocked high external dependence industry could result from the negative industry shock. These are strong
assumptions, though, since they require high degrees of heterogeneity in both high external finance dependence industry mixes across cities and high external finance dependence
industry shocks.
Nevertheless, to address this potential asset channel reverse causality concern, we instrument for city bank suspension rates from 1929-1933 with a measure of city trust: religious
fragmentation. Like Nanda and Nicholas [forthcoming], we construct religious fragmentation
using the 1906 US Census of Religious Bodies. This survey collected the church capacity
of 91 different religious denominations in each county as of 1906. We aggregate the county
13
data to the city level using the same method as before—summing across the counties within
each city. We next construct a Herfindahl Index of religious concentration. Our instrument
measure of religious fragmentation is then one minus the Herfindahl index:
RelF ragc = 1 −
X
h2cr
(3)
r
where hcr is the share of denomination r in city c.
The use of group fragmentation as a measure of trust is common in academic research,
including economics and in relation to financial decision making specifically.22 Alesina and
LaFerrara [2002] show that group identity is one of the four strongest determinants of trust.
In financial decision-making, Guiso et al. [2008] find trust to be an important factor in
individuals’ decisions to invest in the stock market. Similarly, Botazzi et al. [2012] show that
trust plays a role in venture capital investment. Most closely related to our use of religious
fragmentation as a measure of trust, Nanda and Nicholas [forthcoming] also use religious
fragmentation based on the 1906 Census of Religious Bodies as an instrument for local
bank distress. Hence, the use of group fragmentation generally, and religious fragmentation
specifically, as a measure of trust is common in academic literature.
As in Nanda and Nicholas [forthcoming], we expect religious fragmentation to be positively related to the bank suspension rate. This is because higher fragmentation makes bank
runs more likely when uncertainty exists over banks’ illiquid assets.23 We show evidence of
this in Figure 7.24
Our instrument requires that we satisfy the exclusion restriction—namely, that conditional on our city-time and industry group-time fixed effects, the only way in which our
religious fragmentation index affects the relative change in manufacturing across high and
22
Outside of economics, see Tajfel and Turner [1979], which discusses social identity theory.
Because our instrument relies on bank runs causing suspensions, it requires a sufficiently high fraction
of all suspensions to be due to runs rather than insolvency. Only if this is the case will we find a first-stage
relationship between our trust index and bank suspensions.
24
This plot is not the first stage regression used in the analysis. That regression contains fixed effects. We
show this picture for clarity—so we can display each city as a separate data point. The first stage results
that include fixed effects are reported below.
23
14
low external finance dependence industries is through the bank suspension channel. Because
of our fixed effects, religious fragmentation affecting any city-year or industry group-year
level characteristic directly does not threaten the exclusion restriction. For example, our
instrument is still valid if it is correlated with some unobservable, city-level characteristic
such as the quality of institutions, which may influence how the real economy responds to a
negative city shock. We need only that religious fragmentation is not differentially related
to high versus low external finance dependence industries within a city.
Table 1 shows the first-stage results. As in Figure 7, column 1 indicates a positive and
statistically significant relationship between religious fragmentation and bank suspension
rates: a 10 percentage point increases in religious fragmentation is associated with a 5
percentage point higher bank suspension rate.
The remaining columns of Table 1 show suggestive evidence of the exclusion restriction.
While this evidence is neither necessary nor sufficient in proving an inherently untestable
exclusion restriction, it does show that cities with greater religious fragmentation are not
systematically different than cities with less religious fragmentation on dimensions other
than bank suspensions. In particular, religious fragmentation does not appear to be strongly
related to pre-Depression banking sector size, the manufacturing growth rate, or the share
of manufacturing in high external finance dependence industries.
5
5.1
Results
Fixed Effects Estimator
Table 2 shows our primary results, which estimate equation (2) for our three manufacturing outcomes: establishments, employment, and value added. All columns contain industry group-time fixed effects. Even-numbered columns contain city fixed effects and oddnumbered columns contain city-time fixed effects. We normalize the bank suspension rate
to have mean zero and unit standard deviation so that the estimated β coefficients can be
15
interpreted as the difference in the conditional (on fixed effects) growth rate between high
and low external finance dependent industries following a one standard deviation increase in
the bank suspension rate.
Consistent with the credit supply channel theory of bank shocks having a relatively larger
impact on high versus low external finance dependence industries, the estimated β coefficients are negative, significant at the 5 and 1 percent levels, and economically meaningful.
Industries with greater external finance dependence contracted by 11 log points more than
industries with less external finance dependence following a one-standard deviation in the
bank suspension rate on the establishment measure. The decline in the employment measure
was 14 log points, and for value added, 15 log points. At the median city bank suspension
rate of 30 percent, this translates into a 21 percentage point larger decline in the number
of establishments, a 25 percentage point larger decline in employment, and a 27 percentage
point larger decline in value added for high external finance dependence industries than for
low external finance dependence industries. Figure 8 shows these magnitudes graphically.
Figure 9 shows the regression coefficients graphically. Thus, manufacturing industries relatively more dependent on external finance declined markedly more than industries relatively
less dependent on external finance following bank shocks during the US Great Depression.
Moreover, the relative declines were driven by dependence on external finance, not other
industry characteristics. Tables 3-5 add to equation (2) interactions of different, non-external
finance dependence industry characteristics. The coefficient on the external finance dependence interaction remains large, statistically significant, and stable across all specifications
and outcomes. As a result, we conclude that bank distress greatly exacerbated the 1929-1933
manufacturing contraction.
Tables 3-5 also show that industry capital intensity played a role in the manufacturing
decline. Industries with above-median capital intensity (dollars of capital stock per dollar
value added) contracted relatively more than industries with below-median capital intensity
in cities where with larger bank suspension rates. This is consistent with the credit supply
16
model, which predicts that all else equal—including external finance dependence—more capital intensive industries which require greater levels of capital investment will decline by more
than less capital intensive industries when the banking sector experiences distress. In the
face of the declining demand of the Great Depression, this was especially true, as Tables 3-5
indicate. Hence, industry capital intensity also played a role in leading to the manufacturing
decline of the early 1930s in the US.
To estimate the aggregate effect of external finance dependence, we combine our regression coefficients in Table 2 with the pre-Depression share of high external finance industries
and the overall magnitude of the 1929-1933 bank suspensions. We conclude that approximately 22 percent of the overall manufacturing decline during the US Great Depression can
be attributed to bank distress.25
5.2
Robustness: Instrumental Variables Estimators
5.2.1
Religious Fragmentation Instrument
To address the asset channel reverse causality concern, we proceed in Table 6 to estimate
equation (2) instrumenting for bank suspensions with our religious fragmentation measure
as a robustness check. All coefficients remain negative and significant.
They do, however, increase in absolute value. We see two possible explanations for this.
First, if our measure of bank shocks is noisy, classical measurement error will attenuate the
OLS estimates. This seems plausible because our suspension rate measure collapses together
suspensions of varying types and time lengths. In this case, our IV estimates will be larger
because they reduce the measurement error attenuation in our OLS estimates.
Second, our IV approach estimates a local average treatment effect (“LATE”, in the
25
In the Appendix we detail our construction of this value, which we see as a lower bound.
17
language of Angrist and Imbens [1994]). If the treatment effect (bank suspension rate) is
constant across all cities, then our instrument (religious fragmentation) identifies the average
treatment effect across cities. If, instead, the treatment effect is heterogeneous across cities,
then our instrument identifies a LATE. A LATE is a weighted-average of individual city
effects, where the weights depend positively on the sensitivity of the instrumented variable
to the instrument. In our case, this means that the LATE assigns greater weight to cities in
which religious fragmentation makes bank suspensions more likely. Such a weighting would
make the estimated IV effect larger than the estimated OLS effect if the cities in which
religious fragmentation is associated with more bank suspensions were also cities in which
non-financial firms too were weaker pre-Depression. This seems plausible if cities with weak
bank balance sheets were also cities with weak non-bank balance sheets. In this case, the
onset of a negative panic in the banking sector would lead to both greater bank suspensions
and non-financial firm decline. This will increase the size of the estimated effect of bank
suspensions on non-financial firm decline when that effect weights these weak cities more
heavily. Hence, the LATE estimate will be larger than the OLS estimate.26
Either LATE or measurement error could explain why our IV estimates are larger than
our OLS estimates. Regardless, our IV estimates confirm the sign and economic and statistical significance of our OLS results. As such, they provide further evidence using these
newly digitized data of the importance of bank distress in causing the large decline in US
manufacturing during the Great Depression.
5.2.2
Land Value Instrument
Previous examinations of the impact of bank distress on the real economy during the US
Great Depression have used real estate related instruments for bank distress. Calomiris and
26
In labor economics, the concept of LATE has been used to reconcile larger estimates of the return to
education in IV settings relative to OLS settings when intuition predicts the opposite. See, for example,
Card [2001].
18
Mason [2003] used the share of bank assets in real estate in 1929. The authors asserted that
a larger holding of real estate in 1929 was a good predictor of 1929-1933 bank distress for
two reasons. First, real estate is a relatively illiquid asset compared to financial securities,
so the more real estate a bank held, the less liquid it was, and therefore the more likely to
suffer difficulties during a run or negative shock. Second, real estate values—particularly
those of agricultural real estate, which comprised a large share of bank real estate assets in
the first quarter of the twentieth century—declined markedly during the 1920s as demand
for agricultural goods from World War I subsided. Consequently, banks with large real
estate portfolios—especially in agriculture—suffered large losses. As a result, these banks
had weaker balance sheets going into the Great Depression than banks with relatively less
real estate holdings. This made the more real estate heavy banks more likely to experience
distress in a negative shock. Hence, because of the illiquidity of real estate assets and the
decline in the value of such assets during the 1920s, Calomiris and Mason [2003] argue that
the 1929 share of bank assets in real estate is a good predictor of 1929-1933 bank distress.
Using the same, underlying 1920s decline in agricultural real estate prices, Mladjan [2012]
constructed a similar, state-level instrument for Great Depression bank distress: the 19101920 growth in agricultural land values. His idea was that states with the largest increases in
agricultural land prices from 1910-1920 had the largest decreases in agricultural land prices
from 1920-1929. This made the banks in those states the most vulnerable to suspension at
the onset of the 1929 nationwide panic. Thus, he instrumented for state bank distress with
state 1910-1920 growth in agricultural land values.
We construct this same instrument at the city level for our sample cities. Because banks
within our cities may have had exposure to agricultural land in counties near but not technically within the city boundaries, we include in our calculations all counties within specified
distances (20, 25, and 30 miles) of our city centroids.27 We then calculate for each city,
a weighted average of the 1910-1920 agricultural land value growth rates for the counties
27
See Figure 10 for a geographical depiction of the counties included in our instrument.
19
within the specified distance, where the weights are the number of county acres in farmland
in 1910:
P
g p lp
p∈d lp
Growthc = Pp∈d
(4)
where gp is the growth rate in agricultural land values from 1910-1920 for county p; lp
the number of acres of agricultural land in county p as of 1910; and d the distance—20, 25,
or 30 miles—from the city centroid.
As in Mladjan [2012], the exclusion restriction here requires that the 1910-1920 growth
rate in nearby agricultural land values affects the differential performance between high and
low external finance dependence manufacturing industries within a city—conditional on the
city-time and industry-time fixed effects—through only increasing the vulnerability of the
city’s banks to suspensions from 1930-1933. Similar to the religious fragmentation index
robustness instrument, this land value growth robustness instrument can be correlated with
city-time and industry-time characteristics and still produce unbiased estimates. It need
only be uncorrelated with city-industry characteristics.
Unlike the religious fragmentation instrument, however, this land value growth instrument operates through a balance sheet channel rather than a bank run channel. That is,
while greater religious fragmentation leads to more bank suspensions because a lack of community trust makes bank runs more likely, large growth in agricultural land values from
1910-1920 leads to more bank suspensions because bank balance sheet are weaker at the
outset of the 1929 financial crisis. Hence, this instrument operates through the insolvency
rather than the illiquidity channel. Nevertheless, its affect on bank suspensions is similarly
to increase them.
Figure 11 shows a graphical representation of the first stage. As expected, cities with
larger increases in agricultural land value from 1910 to 1920 experienced more bank suspensions from 1930-1933. Table 7 reports the second stage results using all counties within 25
20
miles of the city centroid.28 Across all specifications, we confirm our previous results.
Hence, when using a real estate related instrument for 1930-1933 bank distress, we confirm
our OLS and religious fragmentation IV robustness findings, and the findings of previous
researchers who have used real estate instruments: bank distress led to a significant decline
in real economic activity during the US Great Depression.
6
Persistence
Whether that decline was short-lived or persisted is a question we are uniquely able to answer
with our newly-digitized, city-industry-year level manufacturing data, which extends into the
late 1930s. Moreover, with information on three different measures of manufacturing—the
number of establishments, employment, and value added—we can further test whether the
persistence was concentrated along certain dimensions of the sector. This is something
previous researchers have not been able to do.
Learning the answers to these questions is important for understanding the total costs of
a financial crisis and designing government policy. Failing to account for persistence in the
negative effects could lead to an underestimation of financial crisis costs. This, in turn, could
lead to ill-informed policy decisions—particularly in the immediate post-crisis period. For
example, if the negative effects on the real economy of a credit supply shock dissipate quickly,
then policymakers might be wise to intervene less in the economy once the credit market has
returned to pre-crisis health. In the context of the Great Depression, with bank suspension
rates in our sample cities returning to 1920s levels by 1935 (see Figure 4), to the extent
that the rates reflect a healthier credit market, a finding of non-persistence could suggest a
perhaps less-aggressive role for government policy in the late 1930s. Either way, investigating
the persistence of negative credit market effects on the real economy will enhance lead the
28
Results are similar when including counties within 20- and 30-mile distances from the city centroids. See
Table 7.
21
understanding of the costs of financial crises as well as any government policy that could
seek to address those costs.
We test for persistence by running the same OLS specification in equation (2), except
instead of interacting our high external finance and bank shock variables with an indicator
for 1933, we use 1937 data and interact it with an indicator for 1937:
ln Ycit = β1937 HF inj(i) 1 [t = 1937] Shockc + γHF inj(i) Shockc + αj(i)t + αct + εcit
(5)
Thus, we are now comparing the differential growth rates between high and low external
finance dependence industries in cities with large and small bank suspensions from 1929 to
1937 instead of from 1929 to 1933. If β1937 < 0, we would see evidence of persistence.
Before estimating β1937 , however, we make two comments about identification. First, to
identify β1937 requires an additional assumption beyond the assumptions required to identify
β in equation (2)—namely, that any city-level shocks from 1933 to 1937 not caused by the
original 1929-1933 bank shocks are uncorrelated with the 1929-1933 bank shock. A post1933 shock correlated with but not caused by the 1929-1933 bank shock would cause β1937 to
include the effects from the post-1933 shock in addition to the 1929-1933 bank shock effects.
This would bias β1937 . A post-1933 shock correlated with the earlier bank shock but also
caused by the bank shock, on the other hand, would not pose a problem. For this post-1933
shock would be attributable to the bank shock. In that case, β1937 would remain unbiased.
Secondly, even if β1937 remains unbiased, it could be attenuated by post-1933, targeted
government intervention. If politicians devoted recovery resources to cities with high levels
of bank distress and within those cities to manufacturing industries that declined the most
from 1930-1933, then the recovery resources might have mitigated persistent effects of local
bank distress. If, on the other hand, politicians distributed resources orthogonally to 19301933 bank distress, then recovery resources would simply add noise to our model, which, all
22
else equal, would just increase our standard errors. This too will decrease the likelihood that
we find persistent effects.
With these two identification points in mind, we estimate equation (5) in Table 8. We do
not find strong evidence of persistence on the value added and employment outcomes. On
the establishment outcome, however, we do see the negative effects of the 1930-1933 banking
shock present in 1937—particularly when controlling for other industry characteristics. The
magnitude of the effect is similar to the effect in 1933, which suggests little recovery in the
number of manufacturing establishments between 1933 and 1937. We confirm our results using our religious fragmentation robustness instrument—also shown in Table 8.29 Combined,
the OLS and IV robustness results using 1937 outcomes suggest that the negative effects
of the 1930-1933 local credit supply shocks on the manufacturing sector persisted on the
establishment dimension.
That the effects persisted the most along the establishment dimension is unsurprising if
establishments are a proxy for significant investment in fixed costs.30 Existing plants can
increase employment and output relatively cheaply compared a new plant beginning production. This is what our persistence results suggest. Moreover, the non-recovery among
establishments we estimate in our persistence regressions alongside the recovery of employment and value added is reflected in the national data. As Figure 1 shows, establishments
had not achieved their pre-Depression levels by 1939, whereas employment and value added
had. Hence, aggregate data patterns are consistent with our sample micro data estimates:
the creation of new manufacturing plants simply did not recover at the same rates that employment and output recovered. We believe this is likely due to the higher fixed investment
cost required to open a new establishment relative to scaling up existing plant capacity.
These persistent, negative effects of financial distress on establishments are particularly
29
We obtain a more precise estimate on the employment outcome using the IV approach, but the magnitudes of the coefficient and the precision of the estimate are both less than the corresponding figures on the
establishment outcome.
30
We would expect a similar result if we used capital investment as an outcome. Unfortunately, that
outcome is not available during our sample period. Hence, establishments is as close as we can come to some
measure of fixed cost investment.
23
important for policy for two reasons. First, to the extent that establishments do indeed
proxy for investment, their slower recovery suggests that policymakers may want to design
policies to encourage investment in the aftermath of a financial crisis even if other recovery
policy measures are easing. Second, if as Chinitz [1961] suggests, the number of firms in an
area is correlated with innovation, competition, and industrial diversity, policymakers may
also want to assist directly small and new establishments to foster these local outcomes as
well. Either way, our result that the negative effects of financial distress are particularly
persistent in reducing the number of new firms has policy implications.
7
Conclusion
Exploiting the comprehensive outcomes of newly-digitized, city-industry-year level data covering US manufacturing during the country’s largest financial crisis, the Great Depression,
we show that distress in the financial sector has a significant, negative, and—along certain
dimensions—persistent effect on the real economy. We estimate that the unprecedented
wave of bank suspensions from 1929-1933 led to approximately one fifth of the contemporaneous decline in manufacturing employment, establishments, and value added. Such a
finding provides a channel through which the Depression-era, bank distressed-induced income and construction declines shown by Calomiris and Mason [2003] could have operated.
It also is the first to document a negative effect of financial distress on manufacturing employment.31 Moreover, our findings that the negative effects of financial distress lingered into
the late 1930s for establishments, which in the Census of Manufactures data did not achieve
pre-Depression levels until 1947, are among the first micro-evidence on the persistence of
31
Ziebarth [2013] did not find a negative effect of financial distress on manufacturing employment in his
Mississippi subsample.
24
financial shocks. Thus, with our new data, which allow for a uniquely rigorous identification strategy, we have shown that US manufacturing, which comprised nearly a third of the
country’s economic activity in 1929, suffered greatly as a result of financial sector distress,
and that the effects persisted along the establishment channel.
We encourage future researchers to investigate in greater detail the nature of the recovery in the financial and non-financial sectors following the Great Depression—and financial
crises, generally—using similar microeconometric techniques and new data. Since Calomiris
and Mason [2003], scholars have increasingly exploited area, industry, or firm level data to
document the causes of financial crises and answer macroeconomic questions more generally.
A similar approach on recoveries would be equally informative, particularly for policymakers
seeking ex-ante and ex-post policies to mitigate economic damage caused by financial crises.
Our analysis into the persistence of the negative effects of Depression-era bank suspensions
is a start, but as more micro-level data from the Great Depression era and the 2008-2009
financial crisis become available, scholars can augment our understanding of how economies
rebound from credit supply shocks.
25
References
Alberto Alesina and Elinana LaFerrara. Who trusts others? Journal of Public Economics,
85(2):207–234, 2002.
Joshua Angrist and Guido Imbens. Identification and estimation of local average treatment
effects. Econometrica, 62(2):467–475, 1994.
Adam Ashcraft. Are banks really special? new evidence from the fdic-induced failure of
healthy banks. Federal Reserve Bank of New York Staff Reports, 176:1–45, 2003.
Benjamin Bernanke. Nonmonetary effects of the financial crisis in the propagation of the
great depression. American Economic Review, 73(3):257–276, 1983.
Laura Botazzi, Marco Da Rin, and Thomas Hellmann. The importance of trust for investment: Evidence from venture capital. NBER Working Paper No. 16923, 2012.
Charles Calomiris and Joseph Mason. Consequences of us bank distress during the depression. American Economic Review, 93(3):937–947, 2003.
David Card. Estimating the return to schooling: Progress on some persistent econometric
problems. Econometrica, 69(5):1127–1160, 2001.
Benjamin Chinitz. Contrasts in agglomeration: New york and pittsburgh. American Economic Review: Papers and Proceedings, 51(2):279–289, 1961.
Gabriel Chodorow-Reich. The employment effects of credit market disruptions: Firm-level
evidence from the 2008-2009 financial crisis. Quarterly Journal of Economics, 129(1):1–59,
2014.
David Evans and Boyan Jovanovic. An estimated model of entrepreneurial choice under
liquidity constraints. Journal of Political Economy, 97(4):808–827, 1989.
Milton Friedman and Anna Schwartz. A Monetary History of the United States. Princeton
University Press, Princeton, NJ, 1963.
Eric Gilje. Does local access to finance matter? evidence from us oil and natural gas shale
booms. Working Paper, 2013.
Luigi Guiso, Paola Sapienza, and Luigi Zingales. Who trusts the stock market. Journal of
Finance, 63(6):2557–2600, 2008.
Victoria Ivashina and Anna Kovner. The private equity advantage: Leveraged buyout firms
and relationship banking. Review of Financial Studies, 24(7):2462–2498, 2011.
Marc Joffe. Drivers of municipal bond defaults during the great depression. Working Paper,
2012.
William Kerr and Ramana Nanda. Democratizing entry: Banking deregulations, financing
constraints, and entrepreneurship. Journal of Financial Economics, 94(1):124–149, 2009.
26
William Kerr and Ramana Nanda. Handbook of Research on Innovation and Entrepreneurship, chapter Financing Constraints and Entrepreneurship, pages 88–103. Edward Elgar
Publishing, 2011.
Robert Margo. Employment and unemployment in the 1930s. Journal of Economic Perspectives, 7(2):41–59, 1993.
Kris Mitchener and David Wheelock. Does the structure of banking markets affect economic
growth? evidence from us state banking markets. Explorations in Economic History, 50
(2):161–178, 2013.
Mrdjan Mladjan. Accelerating into the abyss: Financial dependence and the great depression.
Working Paper, 2012.
Ramana Nanda and Tom Nicholas. Did bank distress stifle innovation during the great
depression. Journal of Financial Economics, forthcoming.
JoAnn Senzler Nestor. Interstate branch banking reform: Preserving the policies underlying
the mcfadden act. Boston University Law Review, 72:607, 1992.
Nathan Nunn. Relationship-specificity, incomplete contracts, and the pattern of trade. Quarterly Journal of Economics, 122(2):569–600, 2007.
Joe Peek and Eric Rosengren. Collateral damage: Effects of the japanese bank crisis on real
activity in the united states. American Economic Review, 90(1):30–45, 2000.
Mitchell Petersen and Raghuram Rajan. The benefits of lending relationships: Evidence
from small business data. Journal of Finance, 49(1):3–37, 1994.
Raghuram Rajan and Luigi Zingales. Financial dependence and growth. American Economic
Review, 88(3):559–586, 1998.
U.S. Federal Reserve. Bank suspensions, 1921-1936. Federal Reserve Bulletin, September
1937.
Gary Richardson and William Troost. Monetary intervention mitigated banking panics
during the great depression: Quasi-experimental evidence from a federal reserve district
border, 1929-1933. Journal of Political Economy, 117(6):1031–1073, 2009.
Henri Tajfel and John Turner. The Social Psychology of Intergroup Relations, chapter Evidence on the Nature and Sources of agglomeration economies, pages 33–47. North Holland,
1979.
Nicolas Ziebarth. Identifying the effects of bank failures from a natural experiment in mississippi during the great depression. American Economic Journal: Macroeconomics, 5(1):
81–101, 2013.
27
Appendix: Aggregate Effects of Financial Distress
In the paper, we claim that the credit supply shock caused approximately 22 percent of the
drop in manufacturing employment during the Great Depression. We arrive at this estimate
using the following procedure.
First, we define the unconditional percentage point decline in employment as αH for
industries with high external finance dependence and αL for industries with low external
finance dependence. We say unconditional because αj , with j = {H, L}, is the drop in
employment coming from all sources—the credit supply channel and others. Second, we
define the share of employment in high external finance dependence industries as θ. In our
sample, θ is approximately 40 percent. Hence, the total unconditional drop in employment
is θαH + (1 − θ)αL . Using the values of αH and αL that we see in our data, roughly 60% and
40%, respectively, we obtain the total unconditional drop in employment 40 percent. Third,
we define as β j with j = {H, L} the drop in employment coming from only the credit supply
channel. This is the decline we estimate in equation (2). Thus, the the drop in employment
due to the credit shock is θβ H + (1 − θ)β L . We divide this by the total unconditional drop
to obtain the fraction of the total unconditional drop coming from the credit shock:
σ=
θβ H + (1 − θ)β L
θαH + (1 − θ)αL
We assume β L = 0. This means that low external finance industries do not experience
any drop from the credit supply shock. We view this as a conservative assumption. It is
supported by estimating equation (2) on the sub-sample of low external finance dependent
firms only, which yields a negative coefficient on the interaction of bank suspension rate and
an indicator for 1933. The alternative assumption would be β L > 0. This could be plausible if
a general equilibrium effect existed of declines in high external finance dependence industries
leading to expansions of low external finance dependence industries within a city, even after
controlling for industry group demand. We are not aware of any evidence of this—including
28
in our data. Even if that were the case however—that β L > 0—the magnitude of the effect
would have to be large enough to overpower the negative effect of β H < 0, which we have
precisely estimated. We think such a large effect is unlikely. Hence, we proceed conservatively
under the assumption that β L = 0—even though the evidence we have suggests β L < 0.
Using the βˆH = 0.22 we obtain at the median shock level, this yields σ =
(0.40)·(−0.22)
(−0.40)
θβˆH +(1−θ)β L
θαH +(1−θ)αL
=
= 0.22.32 Hence, the fraction of the total drop in manufacturing employment
during the Great Depression that can be explained by financial distress is approximately 22
percent.
We understand the limitations of this exercise, but believe it is a useful starting point
for quantifying aggregate effects. In particular, because our data do not capture the roughly
30 percent of US manufacturing activity taking place outside of the 29 cities in our sample,
this aggregate calculation is somewhat of a partial equilibrium exercise. If plants in high
relative to low external finance dependence industries differentially moved out of our sample
cities and into smaller cities or more rural areas, then our calculation could overstate the
aggregate manufacturing decline. With manufacturing activity increasingly concentrating
in metropolitan areas over the twentieth century and the threat to our calculation requiring plant migration to areas outside of our sample to be differential by external finance
dependence, we think this is not a significant concern.
32
The median and mean shocks are very similar, so results are not sensitive to using the median rather
than the mean.
29
Figure 1: US Manufacturing Establishments, Employment, and Value Added, 1923-1939
Notes: Each line reflects the natural log of the indicated outcome, less the 1929 value. All data are from the
US Census of Manufactures, 1923-1929.
Figure 2: US Manufacturing Establishments, Employment, and Value Added, and US Bank
Deposits, 1923-1939
Notes: Each line reflects the natural log of the indicated outcome, less the log of the 1929 value. All
manufacturing data are from the US Census of Manufactures, 1923-1929. Deposit data are from the FDIC.
30
Figure 3: US Bank Suspensions, 1880-2010
Notes: A bank suspension is defined by the FDIC as the temporary or permanent closure of a bank by supervisory authorities or the banks’ board of directors on account of financial difficulties. Data are constructed
combining two sources. For the older part of the series (pre-1970), the data come from the “Historical
Statistics of the United States: Colonial Times to 1970”, Table bf01. For the more recent part of the series,
the data come from the FDIC website—in particular the “Failures and Assistance Transactions” page.
Figure 4: US Bank Suspensions Rates, by City, 1921-1936
Notes: Each dot represents the fraction of the indicated city’s bank’s that were suspended in the two years
leading up to the shown data point. All data are from the FDIC. Cities are defined as contiguous counties
based on definitions from the US Census Bureau.
31
Figure 5: Manufacturing Value Added in 29 Sample Cities, 1929
Panel A: All 29 Sample Cities
Panel B: Northeast Zoom
Notes: Manufacturing data and city boundaries are from the US Census of Manufactures. The light gray
lines are county boundaries within cities.
32
Figure 6: Change in Manufacturing Value Added from 1929-1933, by 1930-1933 Bank Suspension Rates
Notes: Bank data are from the FDIC. Manufacturing data are from the US Census of Manufactures.
Figure 7: First-Stage: Religious Fragmentation, by Fraction of 1929 Banks Suspended by
1933
Notes: Religious fragmentation is defined by equation (3) in the text. Religious data are from the US Census
of Religious Bodies, 1906. Bank data are from the FDIC.
33
Figure 8: Average Difference in 1929-1933 Manufacturing Growth at the Median Bank Suspension Rate, by External Finance Dependence
Notes: Each red (blue) bar represents the average change in the log outcome value among high (low)
external finance dependence industries at the median bank suspension rate (29.9 percent), controlling for
city-year and industry-year fixed effects as in equation (2). Manufacturing data are from the US Census of
Manufactures. External finance dependence data are from Moody’s Manual of Industrials and Nanda and
Nicholas [forthcoming]. Bank data are from the FDIC.
34
Figure 9: Difference in 1929-1933 Manufacturing Growth, by External Finance Dependence
Panel A: Establishments
Panel B: Employment
Panel C: Value Added
Notes: Each red (blue) dot represents the average outcome value among data points in the given quintile of
the share of banks suspended among high (low) external finance dependence industries, after residualizing
for city and year. The red (blue) line represents the linear fit to all red (blue) data points. Manufacturing
data are from the US Census of Manufactures. External finance dependence data are from Moody’s Manual
of Industrials and Nanda and Nicholas [forthcoming]. Bank data are from the FDIC.
35
Figure 10: Manufacturing Value Added in Counties within 25 miles of City Centers, 1929
Notes: Darker shaded counties reflect higher levels of manufacturing value added in 1929. Un-shaded
counties reflect counties within 25 miles of a city centroid. Counties within 25 miles of more than one
city are assigned to the city to which they are closest. Manufacturing data are from the US Census of
Manufactures. Agricultural land price data are from the US Census of Agriculture.
Figure 11: Real Estate First-Stage: Change in Nearby Agricultural Land Values from 19101920, by Fraction of 1929 Banks Suspended by 1933
Notes: Each dot reflects the change in agricultural land values from 1910-1920 among counties within 25
miles of the indicated city center. Each county is weighted by its 1910 acres of farmland. Agricultural land
price data are from the US Census of Agriculture. Bank data are from the FDIC.
36
Figure 12: Sample Distribution of External Finance Dependence, by 2-Digit SIC Industry
Notes: Each bar represents the fraction of the sample in the labeled 2-digit SIC industry, which corresponds to
a given level of external finance dependence. In the regressions, high external finance dependence industries
are those with above the median (0.06) fraction of total assets in bank loans as of the mid-1920s, according
to Moody’s Manuals of Industrials.
37
38
Shockc
0.513***
(0.129)
-0.052
(0.093)
0.264
29
log(Banks29 )
-0.468
(1.204)
4.869***
(0.892)
0.008
29
M an. Growth27−29
-0.396
(0.608)
0.981**
(0.373)
0.015
29
M an. Growth23−29
-0.566
(0.594)
1.144***
(0.369)
0.031
29
%High F in29
-0.069
(0.096)
0.705***
(0.067)
0.023
29
Notes: Each column represents a different city-industry-year level regression of the indicated manufacturing outcome on the indicated covariates. The
sample is city level. Manufacturing outcomes are constructed from the US Censuses of Manufactures. High external finance dependence industries
are those with above the median (0.06) fraction of total assets in bank loans as of the mid-1920s, according to Moody’s Manuals of Industrials. Bank
shock is defined at the city level as the fraction of the city’s 1929 banks that were suspended by 1933, according to FDIC data. Robust errors are
applied. *** denotes significance at the 1% level, ** at the 5%, and * at the 10%.
R2
N
Constant
Rel.F rag.
Table 1: City Characteristics 1929 and Religious Fragmentation
39
0.327
4,733
29
0.319
4,733
29
0.320
4,733
29
0.341
4,733
29
0.342
4,733
29
ln (Emp)
ln (V A)
(3)
(4)
(5)
(6)
-0.142** -0.141** -0.151*** -0.153***
(0.063)
(0.063)
(0.054)
(0.054)
-0.024
-0.025
-0.028
-0.028
(0.064)
(0.063)
(0.066)
(0.065)
-0.024
-0.066**
(0.023)
(0.027)
Y
Y
Y
Y
Y
Y
Y
Y
Suspension Rate and Industry External Finance Dependence,
Notes: Each column represents a different city-industry-year level regression of the indicated manufacturing outcome on the indicated covariates. The
sample is all city-industry-year triplets in the 29 cities with industry level data reported in both the 1929 and 1933 US Censuses of Manufactures. High
external finance dependence industries are those with above the median (0.06) fraction of total assets in bank loans as of the mid-1920s, according
to Moody’s Manuals of Industrials. Bank shock is defined at the city level as the fraction of the city’s 1929 banks that were suspended by 1933,
according to FDIC data. Standard errors are clustered at the city level. *** denotes significance at the 1% level, ** at the 5%, and * at the 10%.
R2
0.326
N
4,733
Cities (SE cluster level) 29
Table 2: Difference in Manufacturing Outcomes, by City Bank
1929-1933
ln (Estab)
(1)
(2)
HiF inj P ostt Shockc
-0.112** -0.114**
(0.047)
(0.048)
HiF inj Shockc
-0.027
-0.028
(0.027)
(0.027)
P ostc Shockc
-0.005
(0.015)
City FE
Y
CityXTime FE
Y
IndGrpXTime FE
Y
Y
Table 3: Difference in Establishment Growth, Controlling for Non-Finance Industry Characteristics, 1929-1933
ln (Estab)
(1)
(2)
(3)
(4)
(5)
(6)
HiF inj P ostt Shockc
-0.114** -0.107** -0.120** -0.115** -0.123** -0.125**
(0.048)
(0.047)
(0.045)
(0.050)
(0.047)
(0.054)
Contrj P ostt Shockc
-0.009
-0.017
(0.038)
(0.039)
Skillj P ostt Shockc
-0.046
-0.048
(0.045)
(0.051)
Capitalj P ostt Shockc
-0.067
-0.098**
(0.043)
(0.044)
Sizej P ostt Shockc
0.051
0.055
(0.073)
(0.085)
CityXTime FE
Y
Y
Y
Y
Y
Y
IndGrpXTime FE
Y
Y
Y
Y
Y
Y
R2
0.327
N
4,733
Cities (SE cluster level) 29
0.327
4,733
29
0.333
4,733
29
0.328
4,733
29
0.349
4,733
29
0.337
4,733
29
Notes: Each column represents a different city-industry-year level regression of the indicated manufacturing
outcome on the indicated covariates. The sample is all city-industry-year triplets in the 29 cities with
industry level data reported in both the 1929 and 1933 US Censuses of Manufactures. High external finance
dependence industries are those with above the median (0.06) fraction of total assets in bank loans as of
the mid-1920s, according to Moody’s Manuals of Industrials. Bank shock is defined at the city level as the
fraction of the city’s 1929 banks that were suspended by 1933, according to FDIC data. Contract intensity
(Contrj ) is taken from Nunn [2007] and pertains to 4-digit SIC code industries in 1963. It is then recoded
as a dummy, which takes the value of one for all industries whose contract intensity is above the median
level. Capital Intensity(Capitalj ) is obtained from the 1919 Census of Manufactures. The 1919 Census of
Manufactures is the last pre-Depression Census of Manufactures to report total capital stock at the industry
level. We divide total capital stock by industry value added to construct our industry-level capital intensity.
We then recode the variable as a dummy , which, as with the other industry-level variables,takes the value
of one if the industry level is above the median level. Skill Intensity (Skillj ) is constructed directly from the
1929 Census of Manufactures at the 4-digit SIC industry level. It is the fraction of an industry’s employees
that is not production workers. It is coded as a dummy, which also takes the value of one if the industry
level is above the median level. Average establishment size (Sizej ) is also constructed directly from the 1929
Census of Manufactures at the 4-digit SIC industry level. It is the total number of employees in the industry
divided by the total number of establishments in the industry. We code it as a dummy, which also takes the
value of one if the industry level is above the median industry level. We report only the triple interactions,
but all regressions include all lower-level interactions as well. Standard errors are clustered at the city level.
*** denotes significance at the 1% level, ** at the 5%, and * at the 10%.
40
Table 4: Difference in Employment Growth, Controlling for Non-Finance Industry Characteristics, 1929-1933
ln (Emp)
HiF inj P ostt Shockc
(1)
-0.141**
(0.063)
Contrj P ostt Shockc
Skillj P ostt Shockc
Capitalj P ostt Shockc
Sizej P ostt Shockc
CityXTime FE
IndGrpXTime FE
Y
Y
R2
0.320
N
4,733
Cities (SE cluster level) 29
(2)
(3)
(4)
(5)
(6)
-0.128** -0.138** -0.150** -0.132** -0.130**
(0.059)
(0.060)
(0.067)
(0.058)
(0.063)
-0.065
-0.079*
(0.042)
(0.039)
-0.014
-0.056
(0.056)
(0.069)
-0.097**
-0.115**
(0.046)
(0.055)
0.022
0.030
(0.091)
(0.112)
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
0.321
4,733
29
0.349
4,733
29
0.321
4,733
29
0.407
4,733
29
0.412
4,733
29
Notes: Each column represents a different city-industry-year level regression of the indicated manufacturing
outcome on the indicated covariates. The sample is all city-industry-year triplets in the 29 cities with
industry level data reported in both the 1929 and 1933 US Censuses of Manufactures. High external finance
dependence industries are those with above the median (0.06) fraction of total assets in bank loans as of
the mid-1920s, according to Moody’s Manuals of Industrials. Bank shock is defined at the city level as the
fraction of the city’s 1929 banks that were suspended by 1933, according to FDIC data. Contract intensity
(Contrj ) is taken from Nunn [2007] and pertains to 4-digit SIC code industries in 1963. It is then recoded
as a dummy, which takes the value of one for all industries whose contract intensity is above the median
level. Capital Intensity(Capitalj ) is obtained from the 1919 Census of Manufactures. The 1919 Census of
Manufactures is the last pre-Depression Census of Manufactures to report total capital stock at the industry
level. We divide total capital stock by industry value added to construct our industry-level capital intensity.
We then recode the variable as a dummy , which, as with the other industry-level variables,takes the value
of one if the industry level is above the median level. Skill Intensity (Skillj ) is constructed directly from the
1929 Census of Manufactures at the 4-digit SIC industry level. It is the fraction of an industry’s employees
that is not production workers. It is coded as a dummy, which also takes the value of one if the industry
level is above the median level. Average establishment size (Sizej ) is also constructed directly from the 1929
Census of Manufactures at the 4-digit SIC industry level. It is the total number of employees in the industry
divided by the total number of establishments in the industry. We code it as a dummy, which also takes the
value of one if the industry level is above the median industry level. We report only the triple interactions,
but all regressions include all lower-level interactions as well. Standard errors are clustered at the city level.
*** denotes significance at the 1% level, ** at the 5%, and * at the 10%.
41
Table 5: Difference in Value Added Growth, Controlling for Non-Finance Industry Characteristics, 1929-1933
ln (V A)
HiF inj P ostt Shockc
Contrj P ostt Shockc
Skillj P ostt Shockc
Capitalj P ostt Shockc
Sizej P ostt Shockc
CityXTime FE
IndGrpXTime FE
(1)
(2)
(3)
(4)
(5)
(6)
-0.153*** -0.144*** -0.149*** -0.163*** -0.128** -0.129**
(0.054)
(0.049)
(0.053)
(0.058)
(0.055)
(0.060)
-0.047
-0.065
(0.046)
(0.041)
0.006
-0.069
(0.063)
(0.062)
-0.129***
-0.132**
(0.044)
(0.049)
-0.044
-0.041
(0.110)
(0.129)
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
R2
0.342
N
4,733
Cities (SE cluster level) 29
0.343
4,733
29
0.344
4,733
29
0.359
4,733
29
0.417
4,733
29
0.419
4,733
29
Notes: Each column represents a different city-industry-year level regression of the indicated manufacturing
outcome on the indicated covariates. The sample is all city-industry-year triplets in the 29 cities with
industry level data reported in both the 1929 and 1933 US Censuses of Manufactures. High external finance
dependence industries are those with above the median (0.06) fraction of total assets in bank loans as of
the mid-1920s, according to Moody’s Manuals of Industrials. Bank shock is defined at the city level as the
fraction of the city’s 1929 banks that were suspended by 1933, according to FDIC data. Contract intensity
(Contrj ) is taken from Nunn [2007] and pertains to 4-digit SIC code industries in 1963. It is then recoded
as a dummy, which takes the value of one for all industries whose contract intensity is above the median
level. Capital Intensity(Capitalj ) is obtained from the 1919 Census of Manufactures. The 1919 Census of
Manufactures is the last pre-Depression Census of Manufactures to report total capital stock at the industry
level. We divide total capital stock by industry value added to construct our industry-level capital intensity.
We then recode the variable as a dummy , which, as with the other industry-level variables,takes the value
of one if the industry level is above the median level. Skill Intensity (Skillj ) is constructed directly from the
1929 Census of Manufactures at the 4-digit SIC industry level. It is the fraction of an industry’s employees
that is not production workers. It is coded as a dummy, which also takes the value of one if the industry
level is above the median level. Average establishment size (Sizej ) is also constructed directly from the 1929
Census of Manufactures at the 4-digit SIC industry level. It is the total number of employees in the industry
divided by the total number of establishments in the industry. We code it as a dummy, which also takes the
value of one if the industry level is above the median industry level. We report only the triple interactions,
but all regressions include all lower-level interactions as well. Standard errors are clustered at the city level.
*** denotes significance at the 1% level, ** at the 5%, and * at the 10%.
42
43
0.327
4,733
29
29
Y
Y
-0.205**
(0.099)
0.100
(0.066)
0.324
4,733
Y
Y
-0.209**
(0.099)
0.106
(0.070)
-0.044
(0.033)
29
0.317
4,733
Y
Y
-0.224**
(0.098)
0.217
(0.145)
-0.066*
(0.038)
29
0.318
4,733
Y
Y
-0.223**
(0.097)
0.215
(0.143)
ln (Emp)
(3)
(4)
29
0.339
4,733
Y
Y
4,733
29
0.341
Y
Y
-0.167* -0.160*
(0.092) (0.092)
0.160
0.154
(0.141) (0.137)
-0.129**
(0.050)
ln (V A)
(5)
(6)
Notes: Each column represents a different city-industry-year level regression of the indicated manufacturing outcome on the indicated covariates. The
sample is all city-industry-year triplets in the 29 cities with industry level data reported in both the 1929 and 1933 US Censuses of Manufactures. High
external finance dependence industries are those with above the median (0.06) fraction of total assets in bank loans as of the mid-1920s, according
to Moody’s Manuals of Industrials. Bank shock is defined at the city level as the fraction of the city’s 1929 banks that were suspended by 1933,
according to FDIC data. This is instrumented using the Trust index, constructed from the Census of Religion Bodies (1906), as described in the
paper. The Standard errors are clustered at the city level. *** denotes significance at the 1% level, ** at the 5%, and * at the 10%.
R2
N
Cities (SE cluster level)
City FE
CityXTime FE
IndGrpXTime FE
P ostt Shockc
HighF inj Shockc
HighF inj P ostt Shockc
ln (Estab)
(1)
(2)
Table 6: Difference in Manufacturing Outcomes, by City Bank Suspension Rate and Industry External Finance Dependence,
1929-1933, Instrumenting for City Bank Suspension Rate with Religious Fragmentation
44
0.323
4,733
29
29
Y
Y
0.320
4,733
Y
Y
-0.381***
(0.142)
-0.056
(0.060)
-0.372**
(0.159)
-0.077
(0.062)
29
0.322
4,733
Y
Y
-0.418**
(0.175)
-0.057
(0.067)
30m.
(3)
ln (Emp)
25m.
(5)
29
0.316
4,733
Y
Y
29
0.317
4,733
Y
Y
-0.301* -0.289**
(0.158) (0.139)
-0.211
-0.197
(0.141) (0.146)
20m.
(4)
20m.
(7)
ln (V A)
25m.
(8)
30m.
(9)
29
0.316
4,733
Y
Y
29
0.337
4,733
Y
Y
29
0.338
4,733
Y
Y
29
0.336
4,733
Y
Y
-0.311* -0.241* -0.233* -0.236*
(0.164) (0.138) (0.123) (0.134)
-0.215 -0.306* -0.290* -0.328*
(0.172) (0.164) (0.160) (0.198)
30m.
(6)
Notes: Each column represents a different city-industry-year level regression of the indicated manufacturing outcome on the indicated covariates. The
sample is all city-industry-year triplets in the 29 cities with industry level data reported in both the 1929 and 1933 US Censuses of Manufactures. High
external finance dependence industries are those with above the median (0.06) fraction of total assets in bank loans as of the mid-1920s, according
to Moody’s Manuals of Industrials. Bank shock is defined at the city level as the fraction of the city’s 1929 banks that were suspended by 1933,
according to FDIC data. This is instrumented using the Land Growth between 1910 and 1920, constructed from the Census data. Land Growth is
computed, as described in the paper, over three different distances from the city: 20 miles, 25 miles and 30 miles. Every outcome is presented with
each instrument.Standard errors are clustered at the city level. *** denotes significance at the 1% level, ** at the 5%, and * at the 10%.
R2
N
Cities (SE cluster level)
CityXTime FE
IndGrpXTime FE
HighF inj Shockc
HighF inj P ostt Shockc
ln (Estab)
25m.
(2)
20m.
(1)
Table 7: Difference in Manufacturing Outcomes, by City Bank Suspension Rate and Industry External Finance Dependence,
1929-1933, Instrumenting for City Bank Suspension Rate with Regional Land Value Growth from 1910-1920
45
OLS
-0.114** -0.205**
(0.048)
(0.099)
-0.118*
-0.300**
(0.059)
(0.123)
-0.213**
(0.099)
-0.304**
(0.121)
OLS
-0.141**
(0.063)
-0.088
(0.068)
-0.223**
(0.097)
-0.256*
(0.143)
-0.236**
(0.092)
-0.231*
(0.136)
ln (Emp)
IV
IV &Contr
OLS
-0.153***
(0.054)
-0.065
(0.063)
-0.160*
(0.092)
-0.158
(0.129)
-0.170*
(0.090)
-0.141
(0.126)
ln (V A)
IV
IV &Contr
Notes: Each column represents a different city-industry-year level regression of the indicated manufacturing outcome on the indicated covariates. The
sample is all city-industry-year triplets in the 29 cities with industry level data reported in both the 1929 and 1933 US Censuses of Manufactures.
βt represents the coefficient on the triple interaction between the dummy for high dependence on external finance, the City Bank Suspension Rate
and a dummy that takes value one for the period t. All variables definition is the same as in previous tables. The OLS columns have the standard
least-squares regression with city-time and industry group-time fixed effect. The IV columns are estimated using the Religious Fragmentation index
as instrumental variable. The IV &Contr columns are estimated with Religious Fragmentation index as instrumental variable and controlling for the
other relevant factors in the analysis, as in the previous Table. The Standard errors are clustered at the city level. *** denotes significance at the 1%
level, ** at the 5%, and * at the 10%.
β1937
β1933
ln (Estab)
IV
IV &Contr
Table 8: Persistence Regressions 1929-1933 and 1929-1937, (OLS and IV with Religious Fragmentation)
Table 9: City Bank Suspension Rate and City Manufacturing Outcomes, 1929-1933
The sample is all city in the 29 cities with industry level data reported in both the 1929 and 1933 US
Censuses of Manufactures. Bank shock is defined at the city level as the fraction of the city’s 1929 banks
that were suspended by 1933, according to FDIC. Deposit shock is defined at the city level as the fraction of
the city’s 1929 deposit that were held by a suspended banks by 1933, according to FDIC. Correlations are
simple raw correlations.
46