The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Working Paper Series
Network Contagion and Interbank
Amplification during the Great
Depression
WP 16-03
Kris James Mitchener
Santa Clara University
CAGE
CEPR
NBER
Gary Richardson
Federal Reserve Bank of Richmond
University of California at Irvine
NBER
This paper can be downloaded without charge from:
http://www.richmondfed.org/publications/
Network Contagion and Interbank Amplification
during the Great Depression
Gary Richardson
Federal Reserve Bank of
Richmond, University of
California at Irvine, and
NBER *
Kris James Mitchener
Santa Clara University
CAGE, CEPR
and NBER
March 2016
Working Paper No. 16-03
Abstract
Interbank networks amplified the contraction in lending during the Great Depression.
Banking panics induced banks in the hinterland to withdraw interbank deposits from Federal
Reserve member banks located in reserve and central reserve cities. These correspondent
banks responded by curtailing lending to businesses. Between the peak in the summer of
1929 and the banking holiday in the winter of 1933, interbank amplification reduced
aggregate lending in the U.S. economy by an estimated 15 percent.
Keywords: Bank networks, Great Depression, banking panics, contagion, interbank market
JEL Codes: E44, G01, G21, L14, N22
*
Mitchener: Department of Economics, Leavey School of Business, Santa Clara University;
kmitchener@scu.edu. Richardson: Department of Economics, UC Irvine and Federal Reserve Bank of
Richmond; garyr@uci.edu. We thank four anonymous referees, the editor, and seminar and conference
participants at Banca d’Italia, Stanford University, the University of Colorado – Boulder, Southern Denmark
University, Lund University, the London School of Economics, the Board of Governors of the Federal Reserve,
CEPR-Bank of Italy, the Cleveland Federal Reserve Bank, the Atlanta Federal Reserve Bank, Dartmouth, the
San Francisco Federal Reserve Bank, Solvay Brussels School of Economics and Management, and the NBER
summer institute for helpful comments and suggestions. Joseph Johnson, Joseph Henry, and Arjola Cuko
provided invaluable research assistance. The views in this paper are the authors’ and not necessarily those of the
Federal Reserve Bank of Richmond or the Federal Reserve System.
1.
Introduction
Understanding how financial networks propagate shocks, increase systemic risk, and
potentially magnify economic downturns are questions of long-standing interest to
policymakers and scholars. Economic theory suggests many channels through which
networks may transmit shocks (e.g., Allen and Gale, 2000; Allen, Babus, and Carletti, 2010;
Feixas, Parigi, and Rochet, 2000; Lagunoff and Schreft, 2001; Dasgupta, 2004; Cabellero and
Simesek, 2013). Empirical research has provided some evidence of contagious failures
flowing through interbank markets, particularly for the recent financial crisis in the United
States and Europe (Cohen-Cole, Patacchini, and Zenou, 2011; Puhr, Seliger, and Sigmund,
2012; Fricke and Lux, 2012). 1 Yet, empirical analysis of how interbank connections
amplified downturns is a surprisingly lightly studied aspect of past crises, even for wellresearched episodes such as the Great Depression.
In this paper, we document how the interbank network transmitted deposit
withdrawals through the nation’s banking system and amplified the contraction in real
economic activity during the Great Depression. Interbank balances exposed correspondent
banks to shocks afflicting banks in the hinterland. Interbank deposits were a liquid source of
funds that could be employed to meet sudden demands by depositors to convert claims to
cash, and the removal of these deposits from correspondent banks peaked during periods that
contemporary commentators described as, and that our analysis confirms were, banking
panics. During these panics, withdrawals of interbank deposits forced correspondent banks to
reduce lending to businesses. We estimate that these interbank outflows led to a substantial
decline in aggregate lending, equal to approximately 15 percent of the total decline in
commercial bank lending from the peak in 1929 to the trough in 1933.
1
See Allen and Babus (2009) for a survey of the theoretical and empirical literature on bank networks.
1
Our paper contributes to the growing literature on financial networks and the real
economy, illuminating both a mechanism for transmission (interbank deposits) as well as a
source of amplification (balance-sheet effects). It also introduces an additional channel
through which banking distress deepened the Great Depression and complements existing
research on how bank distress during the Great Depression influenced the real economy.
Friedman and Schwartz (1963) focused on how distress reduced money multipliers and
monetary aggregates, triggering deflations, altering real interest rates, distorting consumers’
and firms’ choices, and fueling debt deflation (Fisher 1933). The key mechanism transmitting
bank shocks to the aggregate economy was changes in the aggregate price level. In
examining what they called “a contagion of fear,” which altered the behavior of bank
depositors and managers, they excluded interbank deposits from their analysis since at the
aggregate level these net to zero (i.e., one bank’s asset was another bank’s liability). We
show that interbank networks transfered shocks from one banks’ balance sheet to their
interbank partners, often in cities far away. These transmission chains could change the
quantity and composition of credit, even if they did not change the money supply, the price
level, or other aggregate variables at the heart of Friedman and Schwartz’s analysis.
Bernanke (1983) focused on the liquidation of banks themselves, which raised the
costs of acquiring credit and reduced the supply of loans from banks. The key mechanism
transmitting this shock to the aggregate economy was the destruction of relationships and
information that facilitated financial intermediation when banks failed. We find that the
destruction of banks not only reduced credit available to their own customers, but also to
customers of their correspondents, who had to alter the composition of their balance sheets to
accommodate increased and increasingly volatile interbank outflows. We also find that bank
failures were not necessary for banking panics and central-bank policies to have a substantial
impact on aggregate lending. Runs on banks (i.e., periods when individuals and banks
2
demanded prompt repayment of deposits) were sufficient to induce a cascade of interbank
withdrawals that reduced aggregate lending. The Fed’s tepid response to these interbank
withdrawals during panic periods forced banks in reserve and central-reserve cities –
particularly correspondent banks in major money centers – to reduce illiquid investments,
such as loans to businesses, and hold more liquid assets, such as cash, reserves at the Fed, and
government bonds. Our result extends Calomiris and Mason (2003b), who demonstrated that
bank distress reduced loan supply and economic activity in the location in which a bank
operated. Our research shows that these types of shocks had an additional effect, which was
transmitted through the interbank network, from the hinterland where bank distress originated
to the reservoirs of reserves in money centers such as New York and Chicago.
The next section of the paper describes the pyramid-like structure of interbank
deposits that developed in the nineteenth century, how the founding of the Fed altered the
holdings of these deposits, and the potential consequences of this structure during periods of
severe distress, such as banking panics. Section 3 defines banking panics and identifies the
panics that occurred during the early 1930s. In Section 4, event studies using high-frequency
data show that banking panics triggered flows of interbank deposits. Panel estimates on callreport data identify a causal relationship between panics (as measured by the number, cause,
and clustering of bank suspensions) and interbank outflows. Section 5 estimates how Fed
member banks in financial centers altered the asset side of their balance sheets in response to
interbank deposit flows. The penultimate section calculates the total decline in credit due to
interbank amplification during the Great Depression. We conclude by discussing the
implications of our findings for Federal Reserve policymaking in the 1930s and the effects of
financial networks in banking crises in general.
3
2. The Interbank Deposit Network
The structure of the interbank network that existed on the eve of the Depression
developed during the nineteenth century. Interior banks sought correspondent linkages with
banks in larger cities. Regulations, particularly the national banking acts of the 1860s,
cemented this pyramid structure, requiring country banks to meet legal reserve requirements
by keeping a portion of their reserves as cash in their vaults and the remainder (originally up
to three-fifths) in correspondent banks in reserve or central reserve cities. State laws
reinforced the pyramid structure by requiring state-chartered banks to also split their reserves
between vault cash and interbank balances. This reserve pyramid proved ineffective during
large financial crises of the nineteenth century, when reserves became difficult, and at times
impossible, to access. When faced with widespread demands for cash and credit, reserve city
banks hoarded funds for their own defense and left country banks to fend for themselves. As
a result, banking panics periodically shut down the payments system (Kemmerer, 1918;
Sprague, 1910).
The authors of the Federal Reserve Act aimed to eliminate these crises by creating an
elastic currency and a lender of last resort. They also sought to change the structure of
reserves, which they saw as crucial for reducing the incidence of panics. If all commercial
banks joined the Federal Reserve System, required reserves would be consolidated in one of
the Fed’s 12 regional reserve banks instead of being scattered among hundreds of commercial
banks in scores of reserve cities. Reserves at the Fed could be shifted towards areas afflicted
by local shocks or released en masse by relaxing reserve requirements to alleviate local
shocks (Kemmerer 1918). A consequence of this consolidation would be a reduction in
correspondent balances invested in the stock and call-loan markets, which were seen at the
4
time as contributing to the banking panics of the nineteenth century (James 1978, Myers
1931, Sprague 1910). 2
The Federal Reserve Act (as amended in 1917) required national and state-member
banks to meet their reserve requirements entirely by holding deposits at their regional Federal
Reserve Bank. By 1929, however, only 10% of state-chartered commercial banks had joined
the Federal Reserve System. Since roughly 90% declined to join, much of the reserve
pyramid remained in place. Non-member banks continued to meet their state-mandated
reserve requirements by holding interbank deposits at banks in reserve and central reserve
cities; they also held their excess reserves in these reserve-city banks, which they used to
manage liquidity and offer a broader array of services to their clientele. Federal Reserve
member banks also deposited excess reserves at correspondent banks in reserve and central
reserve cities, since commercial correspondent accounts paid a higher interest rate (typically
2%) than Federal Reserve Banks (typically 0%). This was particularly true of reserve-city
banks, which deposited their excess reserves in money-center banks in New York and
Chicago. All major money-center banks belonged to the Federal Reserve System, as did
almost all banks in the United States that conducted a correspondent banking business. 3 In
June 1929, member banks held 93% of all interbank deposits in the United States. Nonmember banks held only 7%.
2
Withdrawals from non-central reserve city banks happened regularly, and if they were of sufficient magnitude,
they could put pressure on call loan rates to rise and stock prices to fall, triggering panic selling of assets and
inducing a financial panic that could reach well beyond New York City. The standard story for explaining why
country banks and reserve city banks withdrew their interbank deposits in this era was due to the seasonal
demand for money arising from planting and harvest cycles (Calomiris and Gorton, 1991). Indeed, all of the
major panics of the pre-Federal Reserve era were marked by withdrawals of funds from country and reserve-city
banks located in New York City (Bordo and Wheelock, 2011).
3
In New York, for example, the 15 banks with largest number of correspondents belonged to the Federal
Reserve. This group included nine national banks (Chase, Chatham-Phenix, Chemical, City, Commerce, First,
Hanover, Park, and Seaboard), and six state member bank and trust companies (Bankers, Chemical, Guaranty,
Irvine, Manhattan, and New York). In Chicago, the banks doing a substantial correspondent banking business
either belonged to the Federal Reserve, were owned by a national bank that belonged to the Federal Reserve
(e.g., all of the stock of First Union Trust and Savings was owned by First National Bank of Chicago), or were
combined with a national bank in a holding company or similar corporate structure (e.g., Continental Illinois
was a holding company that controlled the Continental National Bank and the Illinois Merchant and Trust
Company).
5
Interbank deposits comprised a substantial share of the deposits of Federal Reserve
member banks. On the eve of the Great Depression, in June 1929, the 8,707 member banks
held $35.9 billion in total deposits, $18.7 billion in demand deposits, and $3.7 billion in
interbank deposits. 4 The share of interbank deposits was highest in reserve and central
reserve cities, where member banks held $1 of interbank deposits for every $4 of demand
deposits, and interbank deposits exceeded 60% of aggregate reserves. The June 1929 call
report of the Federal Reserve shows that excess reserves held at the Fed, which could be used
without triggering regulatory intervention, were low and interbank balances exceeded them
by a substantial multiple. The implication is that money-center banks could satisfy
unexpected declines in interbank balances only by liquidating investments or borrowing from
the Fed.
Thus, although the Federal Reserve Act altered where national and state-member
banks held almost all of their required reserves, an inverted pyramid of interbank balances
remained in place on the eve of the Great Depression. The top layer of the pyramid consisted
of member banks in central reserve cities. These banks held 46% of aggregate interbank
balances, with 37.1% in New York and 8.8% in Chicago. These money-center banks held
deposits from non-member banks as well as from member banks in reserve cities and country
locales. The middle layer of the pyramid consisted of banks in the 59 reserve cities located
throughout the United States. These banks held 37.6% of all interbank deposits, with each
reserve city on average holding 0.6% of aggregate interbank balances. The base of the
reserve pyramid consisted of country banks located in cities and towns throughout the United
States. Most of these banks were state chartered, did not belong to the Federal Reserve
System, and could not borrow directly from the Federal Reserve’s discount window.
4
For comparison, in June 1929, the 15,797 non-member banks held $12.9 billion in time and demand deposits
and $291 million in interbank deposits.
6
Polk’s Bankers Encyclopedia (1929), a compendium of all banks operating in the
United States, documents the correspondent linkages between the foundation and pinnacle of
the pyramid. Almost all country banks possessed correspondents within their own Federal
Reserve district. The majority of country banks also possessed correspondents in either New
York or Chicago, but few country banks possessed correspondents in reserve cities in other
districts, and few reserve-city banks served as correspondents for more than a handful of
banks from other districts. Almost every reserve-city bank deposited funds in a correspondent
in New York or Chicago, and many worked with correspondents in both cities. 5
These network linkages formed a channel through which country banks influenced the
behavior of Fed members in financial centers. During banking panics, country non-member
banks withdrew reserves from their correspondents, since these interbank balances were
some of their most liquid assets. Country Fed member banks responded in similar manner
since accessing excess reserves deposited at their correspondents was cheaper than going to
the discount window or violating their legal reserve requirement and paying the penalties for
doing so. 6 Interbank withdrawals from banks in reserve cities, in turn, induced further
responses within the system that worked their way up to the Fed member banks at the top of
the pyramid in Chicago and New York. Our analysis focuses on how this chain reaction
induced reserve and central reserve-city banks to alter their balance sheets, reducing the
quantity of credit available to commerce and industry.
In theory, the Federal Reserve System possessed tools that could be used to provide
liquidity, even to non-member institutions, including rediscounting and open market
5
This information comes from a stratified random sample of commercial banks drawn from Polk’s Bankers’
Encyclopedia for 1929. We randomly selected two pages per state from the encyclopedia. Those pages
contained data on 1,393 banks, or about one of 20 banks operating at that time. In our sample, 58.3% of banks
had correspondents in New York, 24.8% had correspondents in Chicago, 84.3% had correspondents in a reserve
city in their own district, 15.8% had a correspondent in a reserve city in another districts.
6
To access the reserves they were required to hold at one of the Federal Reserve System banks, member banks
had to pay a fee, potentially be subject to an examination, and could not pay dividends as long as they were
below the reserve requirement. Going to the discount window meant paying a penalty rate, potentially not being
allowed to borrow, or posting double collateral.
7
operations, but leaders of the System disagreed on the extent to which the Fed could and
should aid non-member banks (Richardson and Troost, 2009). These debates partly reflected
concerns about the System’s obligations to financial institutions that did not contribute to its
upkeep or submit to its regulations as well as concerns about whether credit policies should
be pro-cyclical as dictated by the doctrine of real bills. The leaders also disputed whether
collateral originated by or passed through non-member banks was eligible at the Federal
Reserve Banks’ discount windows or for purchase on the open market. State non-member
banks could borrow from their correspondents. These loans were secured by collateral, which
typically took the form of short-term commercial loans originated by the borrowing bank.
The correspondent could, in turn, use this collateral to borrow from the Federal Reserve. The
legality of this practice was questionable. Some Federal Reserve Banks – particularly in
Atlanta and New York – did on some occasions accept collateral originated by non-member
institutions. Other Federal Reserve Banks and the Federal Reserve Board, however, refused
to accept such collateral and pressured Atlanta and New York to cease doing so and to end
efforts to pass liquidity through member banks to the banks at the base of the reserve
pyramid.
3. Banking Panics and Abrupt Changes in Depositor Behavior
Because we seek to determine how interbank deposit withdrawals reduced lending,
we begin by discussing the reasons for these withdrawals and their link to banking distress.
During the Great Depression, there were periods when depositors abruptly changed behavior
and rapidly withdrew deposits, perhaps in response to growing concerns over the safety of the
financial system, increased uncertainty about the solvency of banks or government policy, or
due to the “contagion of fear” that plays such a prominent role in the account by Friedman
and Schwartz (Calomiris and Gorton, 1991; Calomiris and Mason, 1997; Calomiris and
8
Mason, 2003b; Friedman and Schwartz, 1963; Postel-Vinay, forthcoming). We therefore
estimate interbank deposit withdrawals that occurred during banking panics versus those in
other periods – when households and firms may have gradually drawn down their bank
deposits to finance consumption and meet payrolls in response to the declining state of the
economy. Estimating panic-period withdrawals helps to pin down the direction of causality in
our argument. During panics, individuals’ and firms’ willingness to deposit funds in financial
institutions shifted suddenly and sizably. Country banks accommodated these withdrawals in
turn by withdrawing interbank balances. We can clearly measure the impact of changes in the
supply of interbank balances on the quantity and composition of lending by correspondent
banks. At other times, when supply and demand for deposits and loans moved gradually and
in response to similar stimuli, endogeneity makes it difficult to infer the relationships that
interest us from patterns in the data.
Economists have clearly defined banking panics in theory and practice (Calomiris and
Gorton, 1991; Diamond and Dybvig, 1984; Freixas and Rochet, 1997; Gorton, 2012; Jorda et.
al., 2013; Leavan and Valencia, 2013; Reinhart and Rogoff, 2009; Rochet, 2008). During
panics, depositors (including consumers, firms, and financial institutions) abruptly change
their behavior en masse, quickly seeking to convert debt claims to cash. Depositors, in other
words, simultaneously run many banks. Banks respond by taking action to satisfy depositors’
demands or mitigate the consequences of their failure to do so. Banks raise cash to repay
depositors by calling loans, selling assets, drawing down interbank balances, or borrowing
funds, either from financial institutions that remain on solid footing or from a lender of last
resort. During the Great Depression, banks which failed to obtain sufficient liquidity resorted
to enforcing the 30-day clause on savings deposits and suspending conversion of demand
deposits.
9
Panics expand when depositors at additional institutions begin drawing down
deposits. Banks that can neither repay depositors nor defray their demands must suspend
operations. Those that do not re-open usually have their affairs resolved by regulatory
authorities or in bankruptcy court. Sub-optimal outcomes can occur even for institutions that
would have been solvent in calmer states of the world, but whose lack of liquidity forced
them into liquidation.
Economists’ definition of banking panics illuminates symptoms that can often be
observed in historical records: a rush by depositors to convert claims to cash; efforts by banks
to satisfy depositors’ demands or mitigate the costs of their failure to do so; the consequences
of failing to repay depositors, including bank suspensions and liquidations; the suspension of
normal activities by healthy banks, which could have operated without interruption, in
absence of panic; and the clustering of all of these events in both time and space. Scholars
seeking to identify historical panics look for evidence of these symptoms using two methods.
One method, often referred to as the narrative approach, draws inferences from aggregate
time series and qualitative accounts. Aggregate data – such as time series of the number of
bank suspensions, deposits trapped in suspended banks, changes in currency held by the
public, or total deposit outflows from operating banks – reveal periods when the national
financial system experienced substantial, sudden distress. Narrative sources – such as articles
in newspapers and periodicals, memoirs of bankers, businessmen, and policymakers, and
reports of regulatory agencies – provide confirmation as to whether informed observers
believed that the bank distress observed in aggregate data was caused by sudden changes in
depositors’ behavior.
Scholars who have employed this narrative method have identified a consistent series
of national and regional events in the United States during the Great Depression (Friedman
and Schwartz, 1963; Jalil, 2015; Richardson, 2007; Wicker, 1996). The panics begin with the
10
failure of Caldwell in November 1930, which triggered runs in Tennessee and surrounding
states. A series of regional and national panics followed. These occurred in June 1931 (a
regional panic emanating from Chicago), July 1931 (triggered by the banking crisis in
Germany and troubles in Europe), September 1931 (a national panic induced by Britain’s
departure from gold), June 1932 (a second Chicago panic), and February 1933 (the
nationwide panic preceding the bank holiday). 7 Wicker (1996) also discusses additional,
smaller events, which he describes as local or mini-panics, such as events in Toledo in
August 1931 and Philadelphia and Pittsburg in September 1931.
The second method uses micro data from examiners’ reports of bank suspensions to
identify patterns consistent with the symptoms of banking panics. This approach has been
successfully employed (and cross-checked with the narrative method) to identify periods of
bank distress during the 1930s (Richardson, 2008), to analyze banking panics on the eve of
the Great Depression (Carlson, Mitchener, and Richardson, 2010), and to identify “local”
panics during the 1920s (Davison and Ramirez, 2014).
We employ both methods to identify panics during the Great Depression. The first
method – employing aggregate data on suspensions and narrative accounts – clearly identifies
events that effected large numbers of banks in multiple, and in some cases, most Federal
Reserve districts. These national and super-regional panics can be easily observed in data
from weekly reporting banks, a national aggregate which comes from a sample drawn from
large (and reportedly representative) institutions in about 100 sizeable cities (including all
reserve and central reserve cities). The second method – which uses micro data and cluster
analysis – shows that the national and super-regional events consisted of scores of runs
clustered in time and space. Importantly, this method also identifies local panics that were
7
For more information on the Chicago banking Panics, see Calomiris and Mason (1997) and Postel-Vinay
(forthcoming).
11
confined to smaller geographical areas such as individual Federal Reserve district, counties,
or cities, and that are difficult to detect in aggregate data.
For both methods, we use data from Federal Reserve Board’s Division of Bank
Operations ST 6386 data, described in detail in Richardson (2007, 2008) and in Appendix 1.
This database indicates the date and location of all bank suspensions, liquidations, and
mergers under duress. A suspension is defined as a bank closing its doors to depositors for at
least one business day, whether temporarily or permanently. The database also indicates the
reasons why banks suspended operations, based on examiners’ conclusions at the time of
suspension; whether banks reopened; and who decided to suspend operations, typically either
regulatory authorities or a bank’s own board of directors. For certain calculations, we merge
this data with data on all banks in operation in the United States on July 1, 1929, as reported
in Rand McNally Bankers Directory, with all observations given the latitude and longitude of
the center of the town in which the bank operated.
Using these data, Figure 1 plots aggregate U.S. bank suspensions on a weekly basis,
from July 1929 through February 1933. From July 1929 through October 1930, an average of
15 banks suspended each week, a figure slightly higher than during the 1920s (Davison and
Ramirez, 2014). A Bai-Perron test identifies a structural break in the time-series of weekly
suspensions in the middle of November 1930, just after the collapse of Caldwell and
Company. 8 Weekly suspensions increased dramatically at that point in time. The weekly
average rose from 15.1 to 39.1. The standard deviation increased from 6.6 to 29.4. This
period of heightened banking distress lasted from November 1930 through the Banking
Holiday of 1933.
8
The break in November 1930 does not appear sensitive to the choice of the time interval or the minimum
number of weeks for defining break length. Similar results arise when aggregating suspensions by day, week,
month, quarter, or call-report interval, or when seeking to identify a single break or unknown number of breaks.
Similar results are also found when examining subsets of banks, such as state and national, or member and nonmember.
12
In this period, we search for periods when a panic’s symptoms are clearly evident. We
recognize panics as occurring in spans of consecutive weeks that cross four quantitative
thresholds. First, suspensions in each week are above average for the period following the
collapse of Caldwell and Company, when the Bai-Perron test indicated the shift to a regime
of heighted bank suspensions. Second, for at least one week in the interval, weekly
suspensions must be more than one standard deviation above weekly suspensions in the highsuspension regime and at least five standard deviations above weekly suspensions for the
low-suspension regime. In most cases, this five-standard-deviation spike occurred during the
initial weeks; in several cases, peak weekly suspensions exceeded average pre-Caldwell
weekly suspensions by more than 10 standard deviations. 9 Third, for the entire interval, the
proportion of banks closed by a decision of their board of directors, rather than by regulators,
must exceed the proportion during the pre-Caldwell period. Fourth, for the entire interval, the
proportion of suspensions that examiners attributed to runs, rather than other causes, must
exceed the proportion during the pre-Caldwell period.
Using this data-driven definition, we identify seven panics. Table 1 lists these events
and the tests of proportions described by criteria three and four above. Our list includes all of
the regional and national panics typically identified by economic historians using the
narrative method and aggregate data. Our method also identifies an additional interval that
has received little attention in the literature. This unnamed event began in December 1931
and continued through the first week of February 1932.
Our second method uses geocoded data from examiners’ reports of bank suspensions
to identify geographic clusters of bank suspensions. Our approach extends the method of
9
Intervals with peaks above five but below 10 standard deviations above the pre-Caldwell mean typically
involved: (1) larger banks, especially in the two events centered in Chicago; (2) a new crisis beginning before its
predecessor concluded, such as in the surge in suspensions that coincided with the financial crisis in Germany;
or (3) the widespread suspension of payments imposed by state and/or county governments, such as the crisis in
the winter of 1933, which are not recorded in our data set.
13
Davison and Ramirez (2014), who identified local banking panics during the 1920s. The
Davison-Ramirez procedure involves identifying clusters of bank suspensions by calculating
the number of banks that suspend within a limited distance (measured in miles) and a rolling
window (measured in days since the last suspension of a bank within the potential cluster).
Their algorithm also defines a minimum cluster size. Their analysis focuses on clusters with a
minimum of four banks suspending within 10 miles of each other with no more than 30 days
between failures. In the 1920s, they identify 35 “4-10-30” clusters, or roughly 3 per year. We
apply their method to the contraction from 1929 through 1933 and find hundreds of clusters,
with some spanning multiple call periods and multiple Federal Reserve districts. Many of
these clusters occurred during the regional and national panics listed in Table 1, but scores of
clusters occurred outside of these periods.
To determine whether clustering exceeded that which would normally be expected if
bank suspensions were randomly distributed across all operating banks, we turn to the joincount statistic. The join count indicates the number of pairs of banks that suspended
operations within a set distance (measured in miles) and a fixed time interval (such as a
week). 10 Appendix 2 presents formulas for calculating the join-count, the count expected
under the null hypothesis that bank suspensions were randomly distributed across operating
banks, and the variance of the null. Figure 2 graphs the join count for banks suspending
within 70 miles of each other during each week from July 1929 until the Banking Holiday in
March 1933. The figure also plots the threshold at which we can reject the null hypothesis at
the 5% significance level. In the majority of weeks during the Depression, the null hypothesis
10
Our intervals begin each Sunday and end the following Saturday. Similar results arise for weeks spanning
Thursday to Wednesday and Wednesday to Tuesday, which are consistent with the tabulation of weekly
reporting data (usually based on Wednesday evening figures). Because a significant increase in withdrawals at
nearby institutions may not occur instantaneously and nearby institutions may be able to satisfy depositor
withdrawals for some time before suspending, we calculated joins over a variety of time horizons (including 7,
10, and 30 days; monthly, and quarterly) and distances (including 10, 30, 70, and 100 miles).
14
cannot be rejected. This result indicates that in the majority of weeks, the distribution of
suspensions among operating banks across the landscape approaches that expected from
random chance. In many weeks, however, the null can be rejected at the 5% level. In these
weeks, bank suspensions were geographically correlated. Periods of geographic clustering
coincide with all of the intervals that our data-driven definition identified as banking panics,
which the figure indicates with gray horizontal bars. Each of these intervals began with a
spike in the join count. For each initial spike (and for most other weeks in each of these
intervals), the null hypothesis can be rejected at the 1% level. The large national and regional
panics that we identify, in other words, began with and consisted of bank suspensions that
were clustered in both time and space. Smaller statistically-significant upticks in the join
count reveal local panics, often confined within a single Federal Reserve district. An example
is the first spike apparent in Figure 2, in July 1929, which represents Florida’s fruit-fly panic
analyzed in Carlson, Mitchener, and Richardson (2010).
The time series and spatial data presented in this section establish that during the
Great Depression, commercial banks experienced periods of national, regional, and local
distress that were consistent with economists’ definitions of banking panics. During these
panics, bank suspensions differed in many ways from suspensions at other times. During
panics, suspensions were more concentrated in time and space. Suspended banks were more
likely to be run by their depositors. Suspensions were more frequently initiated by a bank’s
board of directors. In many of these events, the fraction of banks suspending temporarily
increased. Temporary suspensions are clear indicators of illiquidity, since these institutions
must have been solvent; they eventually repaid depositors, and in most cases, continued to
pay dividends to stockholders. From related studies on the Great Depression, we know that in
at least some of these incidents, suspended banks that were liquidated eventually had higher
15
recoveries on assets (Richardson and Troost, 2009). All of this evidence is consistent with the
observation that at some points during the Depression, depositors ran banks.
4. Interbank Deposits During Banking Panics
This section assesses whether bank runs triggered flows of interbank deposits. We
begin by examining weekly aggregate data and using an event-study approach to determine
how interbank balances behaved during the large regional and national panics identified in
the previous section. Next, we examine a panel of quarterly call-report data and estimate the
decline in interbank balances associated with suspensions of banks, during both large and
local panics as well as both inside and outside panic periods.
A. Event Studies on Regional and National Panics
The previous section identified seven banking panics. For each, we define an event
window based upon the criteria for determining a panic described in the previous section.
From the date the panic began, t=0, we trace changes in deposits for eight weeks, until
interbank balances began to rebound or until the beginning of the next event. Our data-driven
definition, based upon spikes in the spatial and temporal clustering of bank suspensions
relative to trend, makes it unlikely that the movements of interbank deposits during these
periods, which included large outflows and followed by smaller return flows, arose from
longer-run trends in the data, such as gradual declines in business activity.
To track how deposits behaved during these events, we examine data on the assets
and liabilities of reporting member banks. The data report end-of-Wednesday balances each
week for two groups: reporting member banks in New York City, and reporting other banks
in 100 other cities, which includes Chicago, all reserve cities, and about 40 other cities, as
described in our data appendix. Figure 3 plots changes in demand deposits for reporting cities
16
outside of New York for the six major banking panics that occurred from 1930 to 1932.
During each panic, demand deposits dropped substantially. Figure 4 plots changes in
interbank deposits, which also dropped substantially during all of the panics. The regional
crises following the collapse of Caldwell in 1930 and panics in Chicago in 1930 and 1931
witnessed smaller, shorter declines, as did the panic in December 1931, which followed the
Fed’s announcement that it would further raise discount rates. The banking panics following
the financial crisis in Germany and Britain’s departure from gold witnessed longer, larger
declines in interbank deposits. The earliest inflection point in interbank balances occurred in
the third week, with two more rebounding in the sixth week, and the two in the seventh week.
Figure 5 shows banks in reporting cities reacted to outflows of interbank and demand deposit
by reducing their deposits in domestic banks, primarily interbank deposits in central reserve
cities. The magnitude of the reductions indicates that, during these panics, banks in reporting
cities acquired from one-quarter to one-half of the deposits they paid out from their centralreserve-city correspondent accounts. For example, during the first four weeks of the postCaldwell crisis, demand deposits fell by $159 million and interbank deposits fell by $176
million. Balances with domestic banks fell by $161 million or about one-half the total deposit
outflow. 11
The severity of the panic that occurred in the winter of 1933 is illustrated in Figures 6
through 9. Figure 6 shows the accelerating drain of interbank, demand, and time deposits
from reporting banks outside of New York City. Demand deposits declined by $1,078
million, time deposits by $973 million, and interbank deposits by $887 million. Figure 7
reports how banks outside of New York gathered cash to satisfy depositors’ demands. The
11
Note that similar patterns appear in HP detrended data and when we limit the sample to events that occurred
at least six months after the preceding panic. In the latter sample, we can determine that deposit outflows began
four to eight weeks prior to the waves of bank suspensions that contemporary commentators described as fullblown banking panics. At that point (our t=0), interbank outflows accelerated rapidly. Federal Reserve officials
at the time referred to the pre-panic withdrawals conducted largely by check and wire transfer as invisible runs
(Richardson, 2008).
17
initial reaction was the withdrawal of bankers’ balances from New York City. These
withdrawals began in February, with $261 million withdrawn by February 15th. Withdrawals
accelerated as the panic progressed, with $1,197 million withdrawn by March 8. In late
February and early March, banks began calling loans and selling bonds in large volumes.
$432 million out of the $531 million total decline in bonds came in the last week before the
holiday. For loans to firms and individuals (not for the purpose of purchasing securities),
$610 million of the total decline of $709 million occurred in the last week. It should be noted
that after banks drew down their balances in New York, further response was muted by the
state-banking holidays, which began in the second half of February and spread rapidly during
the first week of March.
The crisis of 1933 affected banks in New York City to a greater extent than earlier
events. Figure 8 plots weekly changes in deposits in New York City. From January 18 to
March 8, demand deposits declined by $1,364 million, time deposits by $165 million, and
interbank deposits by $919 million. Figure 9 shows how New York banks reacted to the drain
from the interior as well as large foreign withdrawals. New York’s money-center banks drew
down their loans, called loans to firms and individuals, and borrowed from the Fed. Reserves
fell first; between January 18 and February 15, they declined by $319 million. They fell by
another $68 million by the end of the first week in March, and after the Fed suspended
reserve requirements, fell another $90 million before rebounding. Lending contracted later.
From February 15 to March 1, loans to firms and individuals (not for the purpose of
purchasing securities) fell by $419 million. Borrowing from the Fed began in the last week of
February when banks borrowed $183 million. In the first week of March, banks borrowed
another $449 million. By that point, banks in New York City had discounted essentially all of
their eligible paper, withdrawn almost all of their excess reserves, and were clamoring for
additional liquidity support. The New York Federal Reserve Bank, however, could not
18
provide that liquidity support as its stocks of free gold had fallen too low for it to distribute
additional cash and credit to member banks. This constraint forced the New York Fed to seek
a banking holiday in its state and for the nation as a whole.
B. Panel Estimates using Call Report Data
A new, panel database (described in Appendix 1) constructed from commercial bank
call reports enables us to use an additional method for generating causal estimates of the
effects of suspensions on interbank deposit flows during the Great Depression. The data
come from call reports, which are roughly quarterly in frequency. No call took place in the
first quarter of 1933, when the national banking holiday was declared, so these estimates do
not include the effects of the national panic in the winter of 1933. The panel contains
information aggregated by Federal Reserve district and for the three tiers of the banking
system: country banks, reserve-city banks, and central-reserve-city banks. The panel’s
structure enables us to compare interbank deposit flows associated with banks that suspended
operations due to bank runs or during panics with those of a control group – banks whose
suspensions occurred outside of panics or banks that did not experience runs – from June
1929 through December 1932.
Our estimation strategy incorporates the structure of the reserve pyramid described in
Section 2 via a series of assumptions. First, flows of interbank deposits to and from reserve
cities within a district reflect distress among country banks in that district and not in other
districts. Second, flows of interbank deposits to and from New York City reflect distress
among all banks in the United States. Third, flows of interbank deposits to and from Chicago
reflect distress among all banks in the Seventh, Ninth, and Tenth Federal Reserve Districts,
which is the region from which Chicago derived the bulk of its interbank deposits.
Using the 15 calls spanning our sample period, we estimate the following equation:
19
(1)
𝐷𝐷𝑖𝑖𝑖𝑖 = 𝛼𝛼𝑖𝑖 + 𝛾𝛾𝑡𝑡 + ∑𝑘𝑘 𝛽𝛽𝑘𝑘 𝑆𝑆𝑘𝑘𝑘𝑘𝑘𝑘 + ∑𝑘𝑘 𝛽𝛽𝑘𝑘𝑘𝑘 (𝑆𝑆𝑘𝑘𝑘𝑘𝑘𝑘 ∗ 𝑅𝑅) + 𝜑𝜑𝑋𝑋𝑖𝑖𝑖𝑖 + 𝜀𝜀𝑖𝑖𝑖𝑖 .
Dit indicates the change in interbank deposits in district or central reserve city i from call t-1
to call t. The subscript i = 1, 2,…, 14 corresponds to the 12 Federal Reserve districts plus the
central reserve cities of New York (i=13) and Chicago (i=14).
Skit measures bank suspensions in district i from call t-1 to call t. 12 The coefficient of
interest, β, indicates the average flow of interbank deposits in response to the suspension of a
bank in that district or in the region served by a central reserve city. To capture the
identification strategies for panics discussed in section 3, in many specifications, we classify
suspensions by type k, a dichotomous variable that splits the sample into suspensions during
panics and all other suspensions. In these specifications there are thus two β coefficients –
one for suspensions during panics and one for all other suspensions.
As also shown in equation (1), Skit can be interacted with R, a variable indicating
reserve cities, which we include in some of our specifications. In these specifications, β,
indicates the flow of interbank deposits from a central reserve city associated with the
suspension of a bank, and β + βR, indicates the flow from the reserve cities associated with
the suspension of a bank.
Equation (1) includes additional controls that are meant to capture other time and
district influences on bank suspensions, including changes in economic conditions, Fed
policies, and fundamentals, which previous scholarship has identified as potentially important
(Calomiris and Mason, 2003b, Bernanke and James, 1991, Grossman, 1994). Xit includes the
Federal Reserve’s consumption index, derived from department store sales in each Federal
Reserve district and measured by the change in the index between the month of call t-1 and t
12
For New York, S is the sum of bank suspensions across the 12 reserve districts. For Chicago, S is the sum of
bank suspensions in reserve districts 7, 9, and 10.
20
and the change in the discount rate in effect i each Federal Reserve district from date t-1 to t.
For the central reserve city of New York, consumption is measured by the national index. For
Chicago, consumption is measured by the average of the indices for the 7th, 9th, and 10th
Districts. The discount rates for New York and Chicago are their respective district rates.
Some specifications of Equation (1) also include intercept terms for each Federal Reserve
district and each central reserve city, αi, as well as time fixed effects, 𝛾𝛾𝑡𝑡 , or time trends,
captured by a fourth order polynomial, ∑𝑝𝑝 𝑡𝑡 𝑝𝑝 , where p= 1, …, 4.
The first column of Table 2 provides a baseline OLS regression that includes only
three explanatory variables: the number of bank suspensions, a constant, and an indicator for
reserve cities. The estimated coefficients on our key variable of interest indicate that when a
suspension forced a bank to shut its doors: (1) interbank deposits fell by $230,563 in reservecity banks located in the suspended bank’s district; (2) declined by the same amount in banks
in the central reserve city of New York; and (3) fell by the same amount in the central reserve
city of Chicago, if the bank belonged to the 7th, 9th, or 10th Federal Reserve Districts.
Columns 2-6 display “within,” fixed-effects models using a variety of definitions of
suspensions and approaches for controlling for unobserved influences. All specifications
include time-varying, district-level control variables. The specifications control for timevarying factors common across the districts in one of two ways: a polynomial time trend
(column 2) or time fixed effects (columns 3-6). The inclusion of these variables ensures that
our findings are not driven by changes in Federal Reserve policies, changes in economic
conditions in each district, and factors changing over time. Standard errors are Huber-White
corrected and clustered at the level of Federal Reserve districts and central reserve cities.
Column 2 reports estimates for regressions that include an interaction between all
bank suspensions and reserve cities This specification indicates that when a suspension
forced a bank to shut its doors, interbank deposits in New York and Chicago fell by
21
approximately $294,683. Column 3 replaces the polynomial time trend with time-fixed
effects. This reduces the influence of observations in periods when bank suspensions and
interbank flows increased simultaneously in all districts, such as the fourth quarter of 1931
(after Britain abandoned gold). This specification indicates that when a suspension forced a
bank to shut its doors, interbank deposits in New York and Chicago fell by $287,594.
To pin that causality runs from banking panics to reductions in interbank deposits,
using the definitions of panics discussed in Section 3, columns 4-6 separate suspensions into
two classes: banks that suspended operations during panics and those that suspended
operations during other periods,. We then compare the estimates for suspensions of these two
types. Column 4 defines the panic group as banks in 10-day, 30-mile joins. Column 5 defines
the panic group as banks using the Davison and Ramirez (2014) clustering algorithm. The
non-panic group in these two specifications consists of suspensions not clustered in time and
space. Column 6 defines the panic group as banks that suspended payments temporarily,
resumed payments, and in most cases survived the Depression plus banks whose suspensions
examiners attributed to runs. Suspensions in non-panic periods could have also triggered
reductions in interbank balances because one of the first actions taken by the court-appointed
receiver would be to withdraw the failed institutions’ correspondent balances; these would
typically be placed in a preferred custodial account under the receiver’s control prior to
disbursing funds to creditors. However, our regression estimates reveal that, on average,
these suspensions were not consistently correlated with substantial flows of interbank
balances. In most specifications, the null-hypothesis that the coefficient on non-panic
suspension equals zero cannot be rejected.
In Columns 4-6, the coefficients indicate that when banks suspended operations
during panics – as defined by joins, clusters, or temporary suspensions plus runs – interbank
deposits in New York and Chicago fell by $500,000 to $2,000,000. Further, interbank
22
deposits in the reserve cities located in the distressed bank’s district fell by roughly $120,000
to $180,000. The largest estimates arise from definitions of panics that focus attention on
periods when runs were widespread, even among banks that remained in operation. Column 6
reports the regression with the most appealing statistical properties in terms of credible
identification. Panics are defined by examiners’ reports of runs and by the fact that many
banks that suspended operations during these events remained solvent and resumed
operations. Causality clearly runs from sudden changes in depositors’ behavior to outflows of
interbank deposits. Other factors that might influence this relationship, both observed and
unobserved, are controlled for with time and district fixed effects as well as data on the
discount rate, consumption, and the number of banks suspending operations due to factors
other than sudden changes in the public’s demand for cash relative to deposits (all of which
are time-varying, district-level control variables).
The coefficients in Table 2 represent withdrawals of interbank balances by banks that
suspended operations as well as withdrawals of interbank balances by banks that did not
suspend but, during the panic, withdrew substantial quantities of interbank balances, either
because they needed to pay depositors who withdrew funds or they wished to hold more cash
in their vaults in case of future withdrawals. They also represent correspondent banks in
reserve cities passing through the effect of interbank withdrawals as they withdrew their own
excess reserves deposited in money-center banks in New York and Chicago. Interbank
withdrawals for each group can be calculated with the following formulas.
(2) 𝐼𝐼𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 = 𝛽𝛽1 𝑆𝑆1 + 𝛽𝛽1 𝑆𝑆1,𝐶𝐶ℎ𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 + (𝛽𝛽1 + 𝛽𝛽1𝑅𝑅 )𝑆𝑆1
(3) 𝐼𝐼𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃ℎ𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟ℎ = (𝛽𝛽1 + 𝛽𝛽1𝑅𝑅 )𝑆𝑆1 ∗ 𝛿𝛿5,2
(4) 𝐼𝐼𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 = 𝐷𝐷1,𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 ∗ 𝑓𝑓𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 + 𝐷𝐷1,𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 ∗ 𝑓𝑓𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚
Where 𝐼𝐼𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 equals total interbank withdrawals estimated by equation (1) for a particular
specification in Table 2. 𝛽𝛽1 is the coefficient on all suspensions (if all suspensions is the only
23
type of suspension in the regression) or panic suspensions (if the specification subdivided
suspensions into type 1 during panics and type 2 not during panics) in that specification for
central reserve cities. (𝛽𝛽1 + 𝛽𝛽1𝑅𝑅 ) is the impact in reserve cities. 𝑆𝑆1 is the total number of that
type of suspensions in the United States. 𝑆𝑆1,𝐶𝐶ℎ𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖 is the total number of suspensions of that
type in the 7th, 9th, and 10th Federal Reserve Districts. 𝐼𝐼𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃ℎ𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟ℎ is the amount of
interbank deposits withdrawn from reserve cities that banks in those cities in turn withdraw
from central reserve cities. 𝛿𝛿5,2, the coefficient indicating the share of interbank deposits
passed through, which the next section estimates.
𝐼𝐼𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 equals withdrawals of interbank balances by banks of type 1 that
suspended operations. This amount cannot be directly observed. We estimate it with a series
of assumptions. We do not know the balance sheets of banks that suspend operations at the
beginning of the event, such as a panic, which precipitated their closure. For the majority of
banks, however, we know their deposits on the date of suspension, which is often noted on
their St. 6386 form. When this information is missing, we can determine their deposits at the
June call prior to their closure, which is listed either on their St. 6386 or in Rand McNally
Bankers Directory. 13 𝐷𝐷1 is the sum of deposits measured near the date of suspension for all
type 1 banks, member and non-member. 𝑓𝑓𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 is the ratio of interbank balances to
deposits on aggregate for all state commercial banks. 𝑓𝑓𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 is the ratio of interbank
balances to deposits on aggregate for all state commercial banks, excluding reserves with the
Federal Reserve. 14
13
A series of robustness checks (available from the authors upon request) involves rerunning all of the
regressions using deposits in suspended banks as the key independent variable. Those regressions yield signs
and significance levels of coefficients almost identical to the regressions that we report using the number of
bank suspensions and estimates of the interbank amplifier that differ by only a few percent. We report the latter
because of the potential error in variables problem that this paragraph reports for data on deposits at the date of
suspension.
14
These values were measured by the OCC each year from the balance sheets of commercial and reported in
their annual report.
24
These formulas allow us to calculate three important ratios. One is the fraction of
interbank outflows caused by banks in reserve cities passing withdrawals of interbank
balances through to central reserve cities, 𝐼𝐼𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝ℎ𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟ℎ ⁄ 𝐼𝐼𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 . In our preferred specification
from Table 2, that fraction is 9.9%. Two is the fraction of total interbank outflows due to
withdrawals of banks that suspended operations, 𝐼𝐼𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 ⁄ 𝐼𝐼𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 . In our preferred
specification from Table 2, that fraction is 8.6%. Three is the fraction of interbank
withdrawals during panics by banks that survived the event 1 − 𝐼𝐼𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝ℎ𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟ℎ ⁄ 𝐼𝐼𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 −
𝐼𝐼𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 ⁄ 𝐼𝐼𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 . In our preferred specification from Table 2, that fraction is 80.7%. Note
that we list the results of these calculations for all banks in Table 6, which summarizes the
overall impact of the estimates presented in Tables 2 through 4.
To address uncertainty about which specifications in Table 2 best fit the data, we
conduct two Bayesian-averaging exercises. Appendix 3 describes the details of our
calculations. Table 3 presents the results. The left-hand columns report the results of
Bayesian model averaging over all specifications reported in Table 2 and over all possible
permutations of the control variables and fixed effects for those specifications. The results
strongly favor the inclusion of district fixed effects, time fixed effects, and time varying
controls. The results strongly favor the specification that divides suspensions into a panic
group composed of suspensions that were temporary or attributed to runs and a non-panic
group of all other suspensions. The posterior inclusion probability for that model exceeds
93%. The inclusion probability for the reserve city interaction receives moderate support: an
inclusion probability of about 7%. The inclusion probabilities for all other specifications of
suspensions are below 3%, and in most cases, less than 1%. 15 The coefficients from this
Bayesian average exercise resemble those from our classical regression. Each bank
15
The inclusion probability, coefficients, and other statistics for the variables and specifications that we do not
report can be found in an online appendix, along with the data and code used to conduct this exercise.
25
suspension is associated with an outflow of about $472,000 from each central reserve city
and about $442,000 from the relevant reserve cities. The Bayesian estimates provide a more
precise estimate of the flows associated with suspensions outside of panics. As we saw in the
aggregate data presented earlier, interbank balances tended to rebound when panics ended.
The rebound is represented by the positive sign of the coefficient on non-panic suspensions in
our Bayesian estimates.
The right-hand columns in Table 3 report the results of a Bayesian-averaging exercise
over a broader set of models, including all specifications that we examined when analyzing
the data, all specifications suggested by seminar audiences and reviewers, and all
specifications that appeared in papers written by other authors using similar methodologies to
quantify banking panics. The exercise strongly favors including district fixed effects, time
fixed effects, and time varying controls, particularly the policy rate. The model with the
highest inclusion probability, 99.9%, defines suspension during panics as suspensions that
examiners attributed to runs and non-panic suspensions as all other suspensions. The
inclusion probability for the reserve-city interaction is 43%. Both Bayesian models indicate
that interbank deposits flowed back into reserve and central reserve cities during non-panic
periods. The coefficients from this Bayesian exercise falls in the middle of those estimated in
the comparable columns (4 through 6) in table 2. During panics, each bank suspension was
associated with an outflow of about $1,049,000 from central reserve cities and an outflow of
about $617,000 from reserve cities.
5. Asset Rebalancing and Interbank Amplification
We now examine how reserve and central-reserve-city banks, the two upper layers of
the interbank network, altered their portfolios in response to deposit flows, particularly once
the period of severe banking distress began. We analyze the same data panel described in the
26
preceding section. Our econometric model differentiates deposit inflows from outflows
because banks may not have responded symmetrically to reductions and increases in deposits.
We divide banks’ assets into seven comprehensive and mutually exclusive categories:
(1) loans to the private sector
(2) government bonds
(3) corporate bonds
(4) cash and reserves held at the Federal Reserve
(5) interbank assets (a bank’s deposits in other commercial banks)
(6) fixed assets such as the value of the bank building plus furniture and fixtures
(7) all other assets.
For each asset category, we regress the change in that asset type on inflows and outflows of
interbank and public (demand plus time) deposits. Since the error terms for the seven
equations corresponding to the seven asset categories may be correlated, we simultaneously
estimate the seven equations using Zellner’s (1962) method of seemingly unrelated
regressions (SUR), which improves the efficiency of the estimates. We can summarize the set
of regressions as:
(5) 𝑦𝑦𝐴𝐴 = 𝑋𝑋𝐴𝐴 𝛿𝛿𝐴𝐴 + 𝜀𝜀𝐴𝐴 ,
where X is a matrix. y, δ, and ε are vectors, respectively indicating the dependent variables,
coefficients, and error terms. The subscript, A = {1, …, 7}, indexes these vectors, where the
numbers indicate the category of assets from the list above. Each individual asset-category
regression takes the following form:
(6) 𝑦𝑦𝑖𝑖𝑖𝑖 =∝ + ∑𝑧𝑧 𝛿𝛿𝑧𝑧 𝑋𝑋𝑖𝑖𝑖𝑖𝑖𝑖 + ∑𝑧𝑧 𝛿𝛿𝑧𝑧 𝑋𝑋𝑖𝑖𝑖𝑖𝑖𝑖 (𝑋𝑋𝑖𝑖𝑖𝑖𝑖𝑖 ∗ 𝑅𝑅) + 𝜀𝜀𝑖𝑖𝑖𝑖
where yit indicates the dependent variable (i.e. changes in dollar values of one of the asset
classes listed above), measured from t-1 to t. The letter z indicates either inflows or outflows.
Xitz indicates deposit flows of type z in district I from call t-1 to t. R is an indicator variable
for central reserve cities. We interact the indicators for central reserve city and panic period
so that we can obtain separate marginal effects for the two upper layers of the pyramid during
the period of banking distress that began with the collapse of Caldwell and company.
27
The results reported in Table 4 show, on average, how the number of dollars invested
in a type of asset changed when a dollar of interbank deposits flowed in (or out) of banks
over our sample period, June 1929 to December 1932. To simplify the presentation, the table
reports linear combinations of the underlying estimated coefficients and standard errors and
hypothesis tests for those linear combinations. The table does not report asset categories (6)
and (7), for which the coefficients were uniformly insignificant statistically and
economically. Each row indicates the average response to a $1 flow of interbank deposits
(either in or out) of the portfolio of a type of banks for a particular period. For example, the
table’s third row indicates how central reserve city banks responded to inflows of interbank
deposits. On average, when $1 flowed in, loans increased by $0.56, government bonds
increased by $0.48, and corporate bonds increased by $0.33, while reserves and interbank
deposits fell by $0.37 and $0.14 respectively. These patterns probably reflect money-center
banks’ efforts to put to productive use cash and reserves that they had accumulated during
periods of distress.
The second half of Table 4 focuses on outflows of interbank deposits. (Positive
coefficients in this part of the table indicate that outflows were associated with reductions in
assets, since in the underlying data, outflows are negative numbers). The estimates indicate
that reserve-city banks acted as conduits, passing country-bank withdrawals up the pyramid
to central-reserve-city banks. When country banks withdrew a dollar of interbank deposits
from reserve-city banks, reserve-city banks in turn reduced deposits in central-reserve-city
banks by $0.59. Central-reserve-city banks responded to interbank outflows by reducing
holdings of loans and bonds and by increasing reserves held in their vaults or at the Fed. The
size of the coefficients indicates that the buck stopped in central reserve cities. When a dollar
of interbank deposits flowed out, money-center banks reduced lending by $0.54, holdings of
28
corporate bonds by $0.67, and holdings of government bonds by $0.34. They also
accumulated an additional $0.47 in reserves.
By comparing coefficients on outflows and inflows, we can determine how banks
changed the allocation of funds after a cycle in which interbank balances flowed out during a
panic and returned when calm was restored. In reserve cities, the net effect of a $1 outflow
followed by a $1 inflow was a $0.32 reduction in loans, a $0.03 decline in interbank balances
held in central reserve cities, and a $0.55 increase in holdings of government bonds. In
central reserve cities, the increasing volatility of interbank balances principally altered the
composition of the bond portfolio. A $1 outflow followed by a $1 inflow resulted in a net
increase in holdings of government bonds by $0.14, an increase in reserves of $0.10, and a
reduction in holdings of corporate bonds by $0.34.
Overall, these results indicate that, as banking distress intensified and country banks
faced runs, they pulled deposits out of reserve-city banks. These banks in the middle layer of
the pyramid responded by drawing down their deposits at banks in central reserve cities. The
buck stopped in New York and Chicago, where money-center banks accommodated
interbank outflows by reducing lending and stockpiling cash.
6. Aggregate Interbank Amplifier
We can now quantify the aggregate reduction in lending associated with interbank
deposit outflows during the Great Depression. We call this reduction interbank amplification.
To calculate the gross amplifier, we multiply the number of bank suspensions by the
estimated deposit outflows associated with suspensions during periods of panic and then
multiply that product by the estimated decline in lending associated with interbank outflows.
In other words, the gross interbank amplifier, A, is:
(7)
A = s1 B1 δ1.
29
Here, s1 is a 1 by 12 vector, indicating suspensions in Federal Reserve districts 1 through 12.
B1 is a 12 by 14 matrix whose elements indicate how suspensions during panics in district i
influenced interbank outflows in district j. It is estimated in Tables 2 and 3. Using the
notation from Equation (1), the elements of the matrix are: 𝐵𝐵𝑖𝑖𝑖𝑖 = 𝛽𝛽1 + 𝛽𝛽1𝑅𝑅 if 𝑖𝑖 = 𝑗𝑗 and 𝑗𝑗 ≤
12. 𝐵𝐵𝑖𝑖𝑖𝑖 = 0 if 𝑖𝑖 ≠ 𝑗𝑗 and 𝑗𝑗 ≤ 12. 𝐵𝐵𝑖𝑖,13 = 𝛽𝛽1.
And,
𝐵𝐵𝑖𝑖,14 = 𝛽𝛽1 if 𝑖𝑖 = 7, 9, 10 and 𝛽𝛽𝑖𝑖,14 =
0 otherwise. δ1 is a 14 by 1 vector that indicates how interbank outflows influenced lending
to businesses and purchases of corporate bonds in the 12 Federal Reserve districts and two
central reserve cities, which we estimated in Table 4.
For all estimates of interbank amplification, we report bootstrapped standard errors,
following the methodology of Cameron, Gelback and Miller (2008). For our models
estimated using classical regression methods (columns 1 through 6), we resample with
replacement by district 10,000 times, and for each sample containing at least one Fed district
and one central reserve city (both are necessary to estimate the aggregate amplifier), we
estimate Tables 2 and 4, calculate the gross and net amplifier, and then calculate the
distribution of those estimates.
The first part of Table 5 first row shows gross amplification for each model estimated
in Tables 2 and 3 for July 1929 through December 1932. Our estimate of the interbank
amplifier in that period ranges from $2.3 billion to $4.9 billion. For example, in our preferred
specification shown in column 6, the estimate is $2.5 billion. In the specification selected by
Bayesian model averaging over all potential models (column 8), the estimate is $4.9 billion.
We also calculate the net interbank amplifier, which accounts for both the reduction
in lending due to interbank outflows during panics and the rebound in lending when
interbank funds returned after each panic subsided. The net amplifier, Anet, is:
(8)
Anet = s1 B1 δ1 + s2 B2 δ2
30
Here, s2 is a 1 by 12 vector, indicating suspensions in Federal Reserve districts 1 through 12.
B2 is a 12 by 14 matrix, whose elements indicate how suspensions during panics in district i
influenced interbank outflows in district j, which we estimate in Tables 2 and 3. Using the
notation from Equation (1), the elements of the matrix are: 𝐵𝐵𝑖𝑖𝑖𝑖 = 𝛽𝛽2 + 𝛽𝛽2𝑅𝑅 if 𝑖𝑖 = 𝑗𝑗 and 𝑗𝑗 ≤
12. 𝐵𝐵𝑖𝑖𝑖𝑖 = 0 if 𝑖𝑖 ≠ 𝑗𝑗 and 𝑗𝑗 ≤ 12. 𝐵𝐵𝑖𝑖,13 = 𝛽𝛽2. 𝐵𝐵𝑖𝑖,14 = 𝛽𝛽2 if 𝑖𝑖 = 7, 9, 10 and 𝛽𝛽𝑖𝑖,14 =
0 otherwise. δ2 is a 14 by 1 vector that indicates how interbank inflows influenced lending to
businesses and purchases of corporate bonds in the 12 Federal Reserve districts and two
central reserve cities, which we estimated in the top portion of Table 4. Estimates of the net
amplifier range from $1.3 in our preferred specification to $3.3 billion in the specification
based upon Davison and Ramirez (2014).
Table 5 shows that our estimates of interbank amplification amounted to a substantial
fraction of the decline in all commercial bank lending. From June 1929 through December
1932, the total fall in commercial bank lending amounted to $14.055 billion. Of this, $3.335
billion consisted of loans and investments trapped in failed commercial banks, and $10.4
billion was the decline of credit by banks still in operation. Given these figures, the net
interbank amplifier amounted to 13.2% of the total decline in commercial bank credit
outstanding and 55.7% of the loans and investment trapped in failed commercial banks, in
our preferred classical estimate, and about the same percentage in the Bayesian specification
with the highest posterior inclusion probability. 16
The second part of Table 5 expands our estimate to include the impact of the banking
panic of winter 1933. As noted earlier, the effects of the final panic during the contraction
cannot be calculated using call report data, since the comptroller of the currency and state
authorities did not call for reports of condition between December 1932 and June 1933. So,
16
Data on total loans and investments trapped in suspended banks are from the Federal Reserve’ compilation of
examiners’ reports for suspended banks as described in Richardson (2008). Data on total credit extended by
commercial banks from Federal Reserve Board of Governors (1943).
31
for this period, we compute the interbank amplifier from weekly reporting data. We assume
that, during this period, the share of the decline in an asset category due to interbank
withdrawals was equal to the share of total withdrawals made by interbank depositors. We
scale up the amount withdrawn to account for the fact weekly reporting banks represented
only a fraction of all banks in New York and around the nation. 17 Based upon these
assumptions, we calculate that during the panic in January, February, and March of 1933,
interbank withdrawals resulted in a $268 million reduction in lending by banks in New York
City and a $646 million reduction in lending by banks outside of New York City.
Incorporating this information yields our estimate of gross and net interbank amplification
from the peak of the business cycle in the summer of 1929 through the trough in the winter of
1933. In our preferred specification shown in column 6, the net amplifier is $2.773 billion.
The standard error is $1,277, indicating that we can reject the null hypothesis that our
estimate equals zero at the 1% level.
To restate our estimate in terms of the total decline in commercial bank credit from
peak to trough, we need to make another calculation, which is complicated because the
precise figure for the total credit decline during the banking panic of the winter of 1933 is
unknown. Little data exists since, as noted above, most states and the federal government
suspended the collection of call reports during the crisis. We therefore estimate the
contraction of credit by assuming that loans and investments for non-member and countrymember banks declined at the same rate as reporting member banks for which data exists. 18
17
The Federal Reserve does not indicate the identities of the reporting banks or the exact proportion of all assets
represented by these banks. For the nation as a whole, the Federal Reserve reports that they attempted to keep
that proportion around 70 percent. When we compare the weekly reporting data for New York City for the last
week of 1932 to the call report data from the same week (though from a different day of the week), we find that
the interbank balances for the weekly reporting banks represented 82 percent of all interbank balances in the
city. The fraction of other components of the balance sheet represented by weekly reporting banks ranges from
74 percent (net demand deposits) to 96 percent (reserves with the Fed). Since the fraction of interbank balances
falls near the median of this range, we use this figure as our scaling factor.
18
Note that this may be an overestimate of their rate of loss for several reasons. First, reserve city banks
typically lost funds at higher rates than country banks during panics. Second, during the winter of 1933, more
32
That rate was 13.2%. At the December 1932 call, member banks in reserve and central
reserve cities possessed $17.862 billion in loans and investments. Country member banks
held $9.607 billion in loans and investments and non-member banks held $7.924 billion in
loans and investments. Using these figures, we estimate that, during the 1933 panic, total
loans and investments declined by $4.603 billion ($2.365 billion among member banks
reserve and central reserves cities and $2.238 billion in country member and all-nonmember
banks). The total decline in credit from commercial banks during the contraction from
summer 1929 to winter 1933 therefore equaled $18.658 billion. Net interbank amplification,
therefore, accounted for 12.6% to 29.9% of the total decline in credit during the contraction
from summer 1929 through winter 1933. The decline was 14.9% in our preferred estimate
and 12.6% in the Bayesian estimate with the highest inclusion probability over all
specifications. 19
Table 6 estimates the sources of withdrawals during panic periods, using the methods
outlined in equations 2 through 4 and coefficients reported in tables 2 through 4. The
majority of interbank withdrawals came from banks that remained open during panics and
survived the contraction. Estimates of the survivors’ share range from 66.2% to 94.7%, with
80.7% in our preferred classical specification and 70.6% in the Bayesian estimate with the
highest inclusion probability over all models. The share due to banks that suspended
operations was smaller, ranging from 3.0% to 14.6%, with 8.6% in our preferred
specification. Estimates of interbank withdrawals passed through reserve to central reserve
cities range up to 23.6%, with 10.7% in our preferred classical estimate and 16.8% in the in
than two dozen states declared banking holidays in the latter half of February and first week of March, and
moratoria effecting country banks occurred in most cases before moratoria effecting reserve and central reserve
cities. This may be an underestimate, however, because country bank’s post-Banking-Holiday reopening rate
was substantially lower than that of large banks operating as correspondents in reserve and central-reserve
cities; early banking holidays may mask, therefore, the date and impact of their demise.
19
Data limitations prevent us from stating this as a fraction of failed banks from peak to trough of the
contraction, because we lack information on the contraction of credit due to banks forced to suspend operations
by local and state moratoria in February and March of 1933 and which then subsequently ceased operations.
33
the Bayesian estimate with the highest inclusion probability over all models. These estimates
indicate that the reserve pyramid operated, in important ways, as intended. It enabled
respondent banks to access interbank funds during periods of distress, and when regional
demands for funds exceeded regional capacities, it enabled regional reserve cities to drawn
on central reserve cities for additional funds. This result explains a key finding in Calomiris
and Mason (2003b), which showed that member banks’ interbank balances decreased their
risk of failure to the same extent as holding cash. We show that when the reserve pyramid
operated effectively prior to the banking holiday, interbank bank balances were liquid.
Surviving banks used them as a source of funds whenever they needed to satisfy depositors’
demands. Their effectiveness in reducing respondents’ risk of failure, however, depended
upon correspondent banks’ ability to convert assets – such as loans and bonds – into cash.
Our result also reconciles Calomiris and Wilson (2003a), who examined failure risk of Fed
member banks during the early 1930s; Calomiris and Wilson (2004), who showed that banks
in New York City reduced lending in response to increases in solvency risk; and Van Horn
and Richardson (2011), who showed that New York’s money-center banks maintained
lending as usual in response to shocks to likely repayment of international loans in central
Europe, particularly Germany. We show that money-center banks experiencing liquidity
drains (whether or not related to the quality of their own loan portfolios) remained in
operation by reducing lending and selling corporate bonds. Country member banks,
according to Calomiris and Mason (2003a), could not do this. Differential liquidity of loans
made by country and reserve-city banks, therefore, explains the different trajectories of bank
distress and credit contraction across the reserve pyramid.
7. Conclusion
Theory suggests that network linkages can transmit shocks and exacerbate financial
crises. We demonstrate that network linkages played an important role in amplifying the
34
contraction of lending during the Great Depression. 20 In the 1930s, commercial banks
responding to sudden, large depositor withdrawals – those typically associated with bank runs
and panics – and rapidly removed interbank balances from correspondent banks in reserve
and central-reserve cities in order to meet depositors’ demands for cash. Reserve-city banks,
almost all of which were Fed members, responded by re-balancing their portfolios, passing
the buck up the reserve pyramid, and reducing lending to businesses. The aggregate
contraction in bank credit from this “interbank amplifier” caused overall lending during the
Great Depression to decline by approximately 15 percent up to the Banking Holiday of 1933.
Ironically, the Federal Reserve System had been created with the purpose of
preventing crises, such as those that had regularly plagued the banking system in the 19th
century, but the Fed’s failure to convince roughly half of all commercial banks to join the
system allowed a pyramided-structure of reserves to persist into the third decade of the 20th
century and created a channel through which the interbank deposit could influence real
economic activity. In theory, pyramided reserves could be deployed to help troubled banks,
but during the banking panics of the 1930s, just as in the panics of the late nineteenth century,
the total size of these withdrawals overwhelmed correspondent banks, leaving those banks
with the choice of either contracting on the asset side of their balance sheets or borrowing
from the Fed. With the Fed unable or unwilling to provide sufficiently liquidity to support
distressed correspondent banks, they were forced to react to interbank outflows by reducing
lending, thus amplifying the decline in investment spending. Although the mechanism is new,
20
By contrast, during the recent financial crisis, the interbank market evaporated in August 2007, with many
banks unable to borrow at a range of short-term maturities; however, central banks responded to the collapse in
interbank lending by injecting massive amounts of liquidity (Allen and Babus, 2009). As most countries
experiencing troubles had floating exchange rates, governments were not constrained in ways that may have
prevented them from injecting liquidity into the banking system in the 1930s (Eichengreen, 1992); moreover,
central banks more actively coordinated their response to the crisis through agreements such as swap facilities
for foreign exchange (Moessner and Allen, 2011). This event was preceded by lax credit conditions that fueled
the rise in house prices in several countries and the subsequent collapse in mortgage-backed securities prices in
the U.S. that particularly affected commercial and investment banks.
35
our results corroborate other studies on the Depression, which emphasize how banking
distress reduced loan supply (Bernanke, 1983; Calomiris and Mason, 2003b).
What might have alleviated this problem? One solution would have been for the
Federal Reserve to extend sufficient liquidity to the entire financial system. The Fed could
have done this by lending funds to banks in reserve centers. In turn, those banks could have
loaned funds to their interbank clients. To do this, banks in reserve centers would have had to
accept as collateral loans originated by non-member banks. Banks in reserve centers would,
in turn, need to use those assets as collateral at the Federal Reserve’s discount window.
Why didn’t the Fed do this? At the beginning of the Depression, the leaders of the
Federal Reserve disagreed about the efficacy and legality of such action. During the stock
market crash in the fall of 1929 and the initial banking panics in the fall of 1930, the Federal
Reserve Banks of New York and Atlanta aggressively extended credit to member banks,
which in turn extended loans to their financial counterparties. The Fed Banks did this by
accepting as collateral eligible paper held but not originated by their member banks. Other
Federal Reserve banks and some members of the Federal Reserve Board criticized these
actions, which they deemed to be unwise and potentially illegal. The debate continued until
legislation in 1932 broadened the Fed’s lending authority, clearly authorizing New York and
Atlanta’s earlier actions (Chandler, 1971). Then, however, the Fed lacked the will and may
have lacked the resources to alleviate the situation.
Another potential solution would have been to compel all commercial banks to join
the Federal Reserve System and require all commercial banks to hold their reserves at a
Federal Reserve Bank. Due to powerful political lobbies representing state and local bankers,
however, Congress was unwilling to contemplate legislation that would have effected such
changes. Had they done so, the pyramid structure of required reserves would have ceased to
exist, and the interbank amplifier, as defined here, would have been dramatically diminished.
36
That said, given the inaction of some Federal Reserve Banks during the 1930s, had such
changes taken place, they may have magnified banking distress as more banks would have
depended on obtaining funds through Federal Reserve Banks that adhered to the real bills
doctrine. As we show, the costs of the pyramid in terms of a contraction in lending were
substantial, but banks still met some of their short-term needs through this structure during
the turbulent periods of banking distress.
The notion that linkages of Federal Reserve member banks to financial institutions
outside the Federal Reserve System were an important source of contagion in the 1930s
echoes recent commentary on the role of shadow banks in the financial crisis of 2007-08.21
The parallels include the susceptibility of shadow banks and non-member banks to runs, the
lack of access of non-member banks and shadow banks to the discount window, the flight to
liquidity and quality during periods of acute uncertainty and financial distress (Moessner and
Allen, 2012), and intercontinental contagion (the last of which remains a clear avenue for
future research). 22 That said, the timing of the two financial crises in relation to the business
cycle, the precise mechanisms for propagating network distress, and the network structures
appear somewhat different.
By the mid-2000s, shadow banks had accumulated assets that were at least as large as
those in the commercial banking system. Some scholars have argued that the initial distress in
financial markets originated from transactions in which shadow banks were intimately
involved, such as in the asset-backed commercial paper and repurchase agreement (“repo”)
markets (Gorton, 2010 and Gorton and Metrick, 2012). The run on the shadow banking
system began in 2007, prior to the economic downturn in early 2008, and distress was spread
21
Shadow banks of the early twenty-first century carry out credit and maturity-transformation activities that are
similar to those of commercial banks, but they are not subject to the regulation and supervision traditional banks
face.
22
In the earlier crisis, many scholars have pointed to the collapse of Creditanstalt in May 1931 as the pivotal
event that spread contagion across Europe (Eichengreen, 1992; Schubert, 1991; Accominotti, 2012)
37
from these markets to other parts of the financial system partly through shadow-bank
linkages with traditionally regulated banking and insurance firms. 23 By contrast, in the
United States. Great Depression, output fell prior to severe financial institution distress.24
Industrial production in the United States peaked in the summer of 1929, with the first
banking panic coming more than a year later, in October 1930.
Further, the precise way in which network linkages led to disintermediation differs in
the two crises. In the 1930s, the decline in interbank balances was sufficient to reduce
aggregate lending; the effect was sizeable, even when compared to the loss of lending due to
the failure of commercial banks. On the other hand, in the recent crisis, Acharya (2013)
argues that the destruction of shadow banks has been the source of overall disintermediation,
suggesting that commercial banks have been unable to fill the void in credit resulting from
their disappearance.
Given the widespread use of deposit insurance today, the behavior of commercial
bank deposits also differed during the two crises: they fell during the Great Depression
whereas they rose during the recent financial crisis (Moessner and Allen, 2011). It is possible
that the presence of deposit insurance during the 1930s could have mitigated runs on banks
and thus reduced the large strain on the reserve pyramid. The reduction in interbank deposit
outflows, in turn, might have then alleviated the need for the Fed to act as a lender of last
resort.
23
Adrian and Ashcroft (2012, p.3) state that these network linkages include “back up lines of credit, implicit
guarantees to special purpose vehicles and asset management subsidiaries, the outright ownership of securitized
assets on bank balance sheets, and the provision of credit puts by insurance companies.”
24
The precise timing of banking distress and economic distress, however, varied across countries (Grossman,
1994, Grossman and Meissner, 2010).
38
Appendix 1: Data
Our principal source for constructing information on the reserve pyramid is Banking
and Monetary Statistics of the United States, 1914 to 1941 (Federal Reserve Board of
Governors, 1943). It presents information from the call reports of Federal Reserve member
banks aggregated by Federal Reserve district, including counts of banks in each district as
well as detailed summaries of assets (15 categories) and liabilities (17 categories) for member
banks located in reserve cities and for banks located outside of reserve cities (called country
banks). Book values of loans and investments are reported, making the interpretation of
regression coefficients straightforward. It also contains detailed classifications of the loans,
investments, and deposits of banks from 1928 through 1941. For the Second and Seventh
Federal Reserve Districts, we calculate the balance sheets of banks in the central reserve
cities of New York and Chicago by subtracting reserve and country banks from all banks.
For our analysis, we then aggregate bank asset information into three categories: (1)
lending to businesses (the sum of loans, acceptances, and corporate bonds); (2) lending to the
U.S. government (the sum of government securities of varying maturity); and (3) reserves
(the sum of cash in the vault and deposits at Federal Reserve banks). We calculate reserves in
this manner to conform to the approach used by Friedman and Schwartz, who excluded from
their calculations balances at domestic banks (which counted as part of a bank’s legally
required reserves if deposited in a bank in a reserve or central reserve city) and balances at
foreign banks. We also excluded cash items in the process of collection from banks’ reserves,
because the slow pace of intercity check clearing left these items simultaneously on the
balance sheets of multiple banks, leading to a double counting of reserves presumed reserves.
During periods of distress, banks found items in the process of collection generally illiquid
and uncollectible (see Richardson, 2007, for details).
39
Data are for each call date. The nature of the calls raises statistical issues. Many
modern time series tests assume observations arise from stable data generating processes with
consistent spacing, which is not characteristic of these data. The spacing of the calls was long
and variable. Calls occurred in 1929 on March 27, June 29, October 4, and December 31; in
1930, on March 27, June 30, September 24, and December 31; in 1931, on March 25, June
30, September 29, and December 31; in 1932, on June 30, September 30, and December 31;
and in 1933, not until June 30. Because the comptroller did not call for reports of condition
during the Banking Holiday in March 1933, call report data reveal little about the collapse of
the commercial banking system in the winter of 1933. Overall, calls occurred on average
every 96 days, but the standard deviation of that average, 35, was high. Restricting the
analysis to regularly spaced calls, December and June, eliminates more than half the
observations from the data set, leaving six during the Great Contraction – far too few to
employ statistical tests based upon asymptotic arguments. Moreover, the December and June
calls almost always occurred on the last day of the month. Banks’ balance sheets on these
dates may have differed systematically from balance sheets on other dates, when the calls
were intentionally unpredictable. We overcome these complications with the methods typical
in the literature. We present results assuming the calls were equally spaced, but we check this
assumption, by re-estimating our results with daily rates of change between calls rather than
change between calls, and recovering similar results.
Bank data covering the events in the first quarter of 1933 and periods between call
reports can be obtained using weekly data on a subset of reporting member banks. The
information appears in Banking and Monetary Statistics of the United States, 1914 to 1941
(Federal Reserve Board of Governors, 1943), Tables 49 and 50. Table 49 presents
information for weekly reporting banks in New York City. Table 50 presents information for
weekly reporting banks in 100 other cities, including Chicago, all other reserve cities, and
40
approximately 40 other municipalities. The advantages of these data are that they are reported
consistently over time and at a higher frequency. The disadvantages are that the tables do not
present complete balance sheets – only figures for selected assets and liabilities. Moreover,
the data come from a non-random set of banks. The chosen banks were meant to create a
representative sample, but the sample changes in size over time, as banks merge, fail, or, for
some other reason, depart the data set. Changes in the sample occasionally represent changes
in the sample composition or sampling methodology. It is impossible to know the size of this
problem because the Federal Reserve does not indicate the identities of the reporting banks or
the exact proportion of all assets represented by these banks. For the nation as a whole, the
Federal Reserve reports that they attempted to keep that proportion around 70%. When we
compare the weekly reporting data for New York City for the last week of 1932 to the call
report data from the same week (although on a different day), we find that the interbank
balances for the weekly reporting banks represented 82 percent of all interbank balances in
the city. The fraction of other components of the balance sheet represented by weekly
reporting banks ranges from 74 percent (net demand deposits) to 96 percent (reserves with
the Fed).
The most accurate source for information about bank failures during this period is the
micro-level data from the Board of Governors’ bank suspension study. These are described in
Richardson (2007, 2008). The Board’s Division of Bank Operations completed a form ST
6386b for each bank that suspended operations. From these forms, we extract an array of
information: a bank’s location, whether it was a Fed member or non-member, whether it
possessed a state or national charter, the date of its suspension, the date of its reopening (if
any), the deposits it possessed on the date of suspension, whether it was suspended by a
decision of its board of directors or under the authority of a state or national bank examiner,
and whether its suspension was triggered by a run. The latter piece of information was
41
elicited by asking opinions of examiners and other authorities and according to the
assessment procedures used by the division of bank operations. Documents describing these
procedures indicate that while depositors lining up outside a bank pleading to withdraw funds
was one symptom of a run, the determination of whether a run occurred should be based upon
the volume of withdrawals and their impact on the bank as well as evidence of significant
withdrawals via check or wire transfer, which usually occurred before ordinary individuals
panicked over the safety of their funds. Researchers at the Federal Reserve called these
events “invisible runs.” For 1929-32, we tabulate the ST 6386 micro data by call date and
Federal Reserve district, creating an accurate analog for our panel of bank balance sheets by
call date.
Appendix 2: Join-Count Statistic
We use the Join-count statistic as a measure of spatial autocorrelation to assess
whether bank failures are geographically clustered together, and to examine how this
clustering varies over time. For the sample of all banks in the United States, we calculate a
separate join-count statistic for every week in the sample between July 1st, 1929 and
February 27th, 1932 based on whether the bank failed or remained open in that week. We
represent the spatial relationship between banks using a spatial weights matrix Wij, where the
ijth element of the matrix contains one if bank i is within an L mile radius of bank j and zero
otherwise. We examined join counts for L equals 10, 20, 30, 70, and 100 miles and find that
70 miles best fits the data. We represent the status of each bank as being either failed or open
with vector y, where an element of the vector equals one if a bank has failed in that week and
zero if a bank remains open in that week. We use Wij and y to count the number of joins as the
number of cases where two banks that are both within L miles of one another both fail in the
same week. The number of observed joins, FF, equals
42
1
� � 𝑤𝑤𝑖𝑖𝑖𝑖 𝑦𝑦𝑖𝑖 𝑦𝑦𝑗𝑗
2
𝐹𝐹𝐹𝐹 =
𝑖𝑖
𝑗𝑗
We can test for whether bank failures are clustered near one another by comparing the
observed number of joins with the expected number of joins under the null hypothesis that
bank failures are randomly dispersed. Define NF as the total number of failed banks in that
week, and N as the total number of banks. The expected number of joins under the null
hypothesis is:
𝔼𝔼[𝐹𝐹𝐹𝐹] =
1
𝑁𝑁𝐹𝐹
�� � 𝑤𝑤𝑖𝑖𝑖𝑖 � � �
2
𝑁𝑁
𝑖𝑖
The variance of the expected join-counts is:
2
𝜎𝜎𝐹𝐹𝐹𝐹
=
where
𝑗𝑗
1 2𝑠𝑠2 𝑁𝑁𝐹𝐹 (𝑁𝑁𝐹𝐹 − 1) (𝑠𝑠3 − 𝑠𝑠1 )𝑁𝑁𝐹𝐹 (𝑁𝑁𝐹𝐹 − 1)(𝑁𝑁𝐹𝐹 − 2)
�
+
4
𝑁𝑁(𝑁𝑁 − 1)
𝑁𝑁(𝑁𝑁 − 1)(𝑁𝑁 − 2)
2
4(𝑠𝑠1 + 𝑠𝑠2 − 𝑠𝑠3 )𝑁𝑁𝐹𝐹 (𝑁𝑁𝐹𝐹 − 1)(𝑁𝑁𝐹𝐹 − 2)(𝑁𝑁𝐹𝐹 − 3)
+
� − 𝔼𝔼[𝐹𝐹𝐹𝐹]2
𝑁𝑁(𝑁𝑁 − 1)(𝑁𝑁 − 2)(𝑁𝑁 − 3)
𝑠𝑠1 = � � 𝑤𝑤𝑖𝑖𝑖𝑖
𝑖𝑖
𝑙𝑙
𝑠𝑠2 = � ��𝑤𝑤𝑖𝑖𝑖𝑖 + 𝑤𝑤𝑗𝑗𝑗𝑗 �
𝑖𝑖
𝑙𝑙
2
𝑠𝑠3 = � �� 𝑤𝑤𝑖𝑖𝑖𝑖 + � 𝑤𝑤𝑗𝑗𝑗𝑗 �
𝑖𝑖
𝑗𝑗
2
𝑗𝑗
This can be used to calculate the test statistic, 𝑍𝑍𝐹𝐹𝐹𝐹 , which will be asymptotically normally
distributed and can be used to conduct a one-sided test for geographical clustering of bank
failures.
𝑍𝑍𝐹𝐹𝐹𝐹 =
𝐹𝐹𝐹𝐹 − 𝔼𝔼[𝐹𝐹𝐹𝐹]
2
�𝜎𝜎𝐹𝐹𝐹𝐹
43
Appendix 3: Bayesian Model Averaging
We address uncertainty over the specification of equation (1) with a Bayesian Model
Averaging (BMA) approach, which enables us to determine the specification that best fits the
data. Our BMA analysis considers all models that potentially illuminate the causal link
between interbank-deposit flows and banks suspensions. This class of models includes, on
the right-hand side, a series of variables that divide all bank suspensions into those that
occurred during panics, as defined by the methods in Section 3, and those that occurred
outside of panics. We consider all time varying controls, and allow our averaging procedure
to choose between including them in levels, changes, or seasonally adjusted. We consider
specifications including either none or a full set of unit fixed effects and specifications
including time trends, a full set of time fixed effects, both, or neither. With this set of models,
we estimate the posterior probability for every model under consideration.
We consider Q models across the model space 𝑀𝑀 = [𝑀𝑀1 , 𝑀𝑀2 , … , 𝑀𝑀𝑄𝑄 ]. Every model
has a linear form:
(A.1) 𝐷𝐷 = 𝜃𝜃𝑞𝑞 𝑍𝑍𝑞𝑞 + 𝜀𝜀𝑞𝑞 with 𝜀𝜀𝑞𝑞 ~𝑁𝑁(0, 𝜎𝜎𝑞𝑞2 𝐼𝐼) for q = 1, 2,…,Q
where 𝐷𝐷 is interbank deposit flow. For model q, 𝜃𝜃𝑞𝑞 is vector of coefficients for all variables
included in the model. 𝑍𝑍𝑞𝑞 is the data matrix. 𝜀𝜀𝑞𝑞 is an error term, and 𝜎𝜎𝑞𝑞2 is the error variance.
For each model, both 𝜃𝜃𝑞𝑞 and 𝑍𝑍𝑞𝑞 are a subset of the full set of possible regressors. We use a
Normal-Inverse-Gamma prior distribution for our models and parameters, which enables us
to estimate posterior distributions. We place an equal prior probability on every model, so
𝜋𝜋�𝑀𝑀𝑞𝑞 � = 1/𝑄𝑄. For each 𝜃𝜃𝑞𝑞 , we assign a prior probability 𝜋𝜋�𝜃𝜃𝑞𝑞 �𝜎𝜎𝑞𝑞2 � = 𝑁𝑁(0, 𝜎𝜎𝑞𝑞2 𝐼𝐼). Thus, the
prior for 𝜃𝜃 is centered around zero. We select 𝜎𝜎𝑞𝑞2 depending on our anticipated scale of the
parameters. We select 𝜎𝜎𝑞𝑞2 = 100,000. For 𝜎𝜎𝑞𝑞2 , we assign a prior probability 𝜋𝜋(𝜎𝜎 2 |𝛾𝛾) =
44
𝐼𝐼𝐼𝐼(𝛼𝛼, 𝛽𝛽), where IG denotes the Inverse Gamma distribution. For all models we set 𝛼𝛼 = 𝛽𝛽 =
0.5. By combining this prior distribution with our set of data, D, we can obtain an estimate of
the posterior probability of each model as:
(A.2) 𝑝𝑝�𝑀𝑀𝑞𝑞 �𝐷𝐷� =
𝑝𝑝�𝐷𝐷�𝑀𝑀𝑞𝑞 �𝜋𝜋(𝑀𝑀𝑞𝑞 )
∑𝑄𝑄
𝑞𝑞=1 𝑝𝑝�𝐷𝐷�𝑀𝑀𝑞𝑞 �𝜋𝜋(𝑀𝑀𝑞𝑞 )
, 𝑞𝑞 = 1,2, … , 𝑄𝑄.
A higher posterior model probability indicates a model that better fits the data. We calculate a
particular coefficient’s probability of inclusion as:
(A.3) 𝑝𝑝(𝛽𝛽𝑗𝑗 ≠ 0|𝐷𝐷) = ∑𝛽𝛽𝑗𝑗𝜖𝜖𝜖𝜖 𝑝𝑝�𝑀𝑀𝑞𝑞 �𝐷𝐷�.
The posterior expected value and variance of each regression coefficient is an average of 𝛽𝛽𝑗𝑗
weighted by the model probabilities:
𝑄𝑄
(A.4) 𝐸𝐸(𝛽𝛽𝑗𝑗 |𝐷𝐷) = ∑𝑞𝑞=1 𝑝𝑝�𝑀𝑀𝑞𝑞 �𝐷𝐷�𝐸𝐸( 𝛽𝛽𝑞𝑞 |𝐷𝐷, 𝑀𝑀𝑞𝑞 )
and
(A.5) 𝑉𝑉𝑉𝑉𝑉𝑉�𝛽𝛽𝑗𝑗 |𝐷𝐷) = ∑𝑄𝑄𝑞𝑞=1 𝑝𝑝�𝑀𝑀𝑞𝑞 �𝐷𝐷�[𝑉𝑉𝑉𝑉𝑉𝑉( 𝛽𝛽𝑗𝑗 |𝐷𝐷, 𝑀𝑀𝑞𝑞 � + 𝐸𝐸(𝛽𝛽𝑗𝑗 |𝐷𝐷, 𝑀𝑀𝑞𝑞 )2 ] + [𝐸𝐸(𝛽𝛽𝑗𝑗 |𝐷𝐷)]2.
45
References
Acharya, Viral. (2013). “Understanding Financial crises: Theory and evidence from the crisis
of 2007-2008” NBER Reporter (April).
Accominotti, Olivier. (2012). “London Merchant Banks, the Central European Panic and the
Sterling Crisis of 1931.” Journal of Economic History vol.72(1): 1-43.
Adrian, Tobias and Adam B. Ashcraft. (2012). “Shadow Banking Regulation.” Federal
Reserve Bank of New York Staff Report No. 559 (April).
Allen, Franklin and Douglas Gale. (2000). “Financial Contagion,” Journal of Political
Economy 108:1-33.
Allen, Franklin and Ana Babus. (2009). “Networks in Finance.” In The Network Challenge:
Strategy, Profit, and Risk in an Interlinked World, Paul R. Kleindorfer and Robert E.
Gunther (eds.), 367-82.
Allen, William A. and Richhild Moessner. (2012). “The international propagation of the
financial crisis of 2008 and a comparison with 1931.” Financial History Review
19(2): 123-147.
Alston, Lee, Wayne Grove, and David Wheelock. (1994). “Why do banks fail? Evidence
from the 1920s.” Explorations in Economic History 31: 409–431.
Beckhart, Benjamin. (1922). “Outline of Banking History from the First Bank of the United
States Through the Panic of 1907.” Annals of the American Academy of Political and
Social Science 99 (January): 1-16.
Beckhart, Benjamin (1932) and James Smith. The New York Money Market, Volume II,
Sources and Movement of Funds. New York: Columbia University Press.
Beckhart, Benjamin Haggott (1932). The New York Money Market, Volume III, Uses of
Funds. New York: Columbia University Press.
Bernanke, Ben S. (1983). “Nonmonetary Effects of the Financial Crisis in the Propagation of
the Great Depression.” American Economic Review 73(3): 257-276.
Bernanke, Ben S. and Harold James. (1991). “The Gold Standard, Deflation, and Financial
Crisis in the Great Depression: An International Comparison.” In R. Glenn Hubbard,
ed., Financial Markets and Financial Crises. University of Chicago Press, 33-68.
Board of Governors of the Federal Reserve System. (1943). Banking and Monetary Statistics
1914-1941. Federal Reserve Board: Washington, D.C.
Board of Governors of the Federal Reserve System. (1959). All Bank Statistics 1896 – 1955.
Federal Reserve Board: Washington, D.C.
46
Board of Governors of the Federal Reserve System. (1932). Committee on Branch, Group,
and Chain Banking, 1932, Dual Banking System in the United States Federal Reserve
Board: Washington, D.C.
Bordo, Michael D. and David C. Wheelock. (2011). “The Promise and Performance of the
Federal Reserve as Lender of Last Resort 1914-1933.” NBER Working Papers 16763.
Caballero, R.J. and A. Simsek. (2013). "Fire Sales in a Model of Complexity." The Journal of
Finance, 68(6): 2549-2587
Calomiris, Charles and Gary Gorton. (1991). “The Origins of Banking Panics: Models, Facts,
and Bank Regulation.” In Financial Markets and Financial Crises, edited by R. Glen
Hubbard. Chicago: University of Chicago Press, pp. 109-174.
Calomiris, Charles W., and Joseph R. Mason (1997). “Contagion and Bank Failures During
the Great Depression: The June 1932 Chicago Banking Panic.” American Economic
Review 87: 863-83.
Calomiris, Charles W., and Joseph R. Mason (2003a). “Fundamentals, Panics and Bank
Distress During the Depression.” (with Joseph Mason) American Economic Review
93: 1615-47.
Calomiris, Charles W., and Joseph R. Mason (2003b). “Consequences of Bank Distress
During the Great Depression.” American Economic Review 93: 937-47.
Calomiris, Charles W., and Berry Wilson (2004). “Bank Capital and Portfolio Management:
The 1930s ‘Capital Crunch’ and Scramble to Shed Risk.” Journal of Business 77:
421-55.
Cameron, A. Colin, Jonah B. Gelback, and Douglas L. Miller (2008). “Bootstrap-Based
Improvements for Interenence with Clustered Standard Errors,” The Review of
Economics and Statistics, 90(3): 414–427
Carlson, Mark, Kris James Mitchener, and Gary Richardson. (2011). “Arresting Banking
Panics: Federal Reserve Liquidity Provision and the Forgotten Panic of 1929.”
Journal of Political Economy 119(5): 889-924.
Chandler, Lester V. (1971). American Monetary Policy, 1928 -1941. New York: Harper and
Row.
Chipman, H.A., E.I. George, and R.E. McCulloch. (2001). “The practical implementation of
Bayesian model selection (with discussion)” In Model Selection (P. Lahiri, ed.) 65–
134. IMS, Beachwood, OH
Cohen-Cole, Ethan, Eleonora Patacchini, and Y ves Z enou . (2010). ”Systemic Risk and
Network Formation in the Interbank Market.”
47
Dasgupta, Amil. “Financial Contagion through Capital Connections: A Model of the Origin
and Spread of Bank Panics.” (2004). Journal of the European Economic Association
2(6): 1049-1084.
Davison, Lee K. and Ramirez, Carlos D. (2014). “Local Banking Panics of the 1920s:
Identification and Determinants.” Journal of Monetary Economics 66 (September):
164–177.
Eichengreen, Barry. (1992). “The Origins and Nature of the Great Slump Revisited.”
Economic History Review 45(2): 213–39.
Freixas, Xavier, Bruno Parigi, and Jean-CharlesRochet. (2000). “Systemic Risk, Interbank
Relations, and Liquidity Provision by the Central Bank.” Journal of Money, Credit
and Banking 32: 611-638.
Fisher, Irving. (1933). “The Debt-Deflation Theory of Great Depressions,” Econometrica
1(4): 337-56.
Fricke, D. and T. Lux (2012). “Core-periphery structure in the overnight money market:
Evidence from the e-MID trading platform.” Kiel Institute for the World Economy
Working Paper No. 1759.
Friedman, Milton, and Anna J. Schwartz. (1963). A Monetary History of the United States,
1867–1960. Princeton, NJ: Princeton University Press.
Gorton, Gary. (1988). “Banking Panics and Business Cycles." Oxford Economic Papers 40:
751- 81.
Gorton, Gary B. (2010). Slapped by the Invisible Hand: The Panic of 2007. New York:
Oxford University Press.
Gorton, Gary B., and Andrew Metrick. (2012) “Securitized Banking and the Run on Repo,”
Journal of Financial Economics 104: 425-451.
Grossman, Richard. (2004). “The Shoe That Didn’t Drop: Explaining Banking Stability
During the Great Depression.” Journal of Economic History 54(3): 654–82.
Grossman, Richard and Christopher M. Meissner. (2010). “International aspects of the Great
Depression and the crisis of 2007: similarities, differences, and lessons.” Oxford
Review of Economic Policy 26(3): 318–338
Hamilton, David (1985). “The Causes of the Banking Panic of 1930: Another View.” Journal
of Southern History 51(4): 581–608.
Hansen, Bruce. (2001). “The New Econometrics of Structural Change: Dating Breaks in U.S.
Labor Productivity.” Journal of Economic Perspectives 15(4): 117-128.
Iyer, R. and Peydró, J.L. (2011). “Interbank Contagion at Work: Evidence from a Natural
Experiment.” Review of Financial Studies 24(4): 1337–1377.
48
Jalil, Andrew. (2015). “A New History of Banking Panics in the United States, 1825-1929:
Construction and Implications.” American Economic Journal: Macroeconomics.
James, John. (1978). Money and Capital Markets in Postbellum America. Princeton:
Princeton University Press.
Kemmerer, Edwin Walter. (1918). The A B C of the Federal Reserve System. Princeton:
Princeton University Press.
Lagunoff, Roger and Stacey Schreft. (2001). “A Model of Financial Fragility.” Journal of
Economic Theory 99: 220-264.
Montgomery, J.M. and Nyhan, B. (2010). Bayesian model averaging: theoretical
developments and practical applications. Political Analysis. 18 (2): 245-270.
Moessner, Richhild and William A. Allen, William. (2011). “Banking crises and the
international monetary system in the Great Depression and now.” Financial History
Review 18(1): 1-20.
Myers, Margaret. (1931). The New York Money Market: Origins and Development. New
York: Columbia University Press.
Postel-Vinay, Natacha. (forthcoming). “What Caused Chicago Bank Failures in the Great
Depression? A Look at the 1920s.” Journal of Economic History.
Puhr, C., Seliger, R. and M. Sigmund. (2012). “Contagiousness and vulnerability in the
Austrian interbank market. OeNB Financial Stability Report, No 24, Oesterreichische
Nationalbank.
Quandt, Richard. (1960). “Rests of the Hypothesis that a Linear Regression Obeys Two
Separate Regimes.” Journal of the American Statistical Association 55: 324-30.
Richardson, Gary. (2008). “Quarterly Data on the Categories and Causes of Bank Distress
during the Great Depression,” Research in Economic History 25 (January): 37-115.
Richardson, Gary. (2007). “Categories and Causes of Bank Distress during the Great
Depression, 1929-1933: The Illiquidity-Insolvency Debate Revisited,” Explorations
in Economic History 44(4): 586-607.
Richardson, Gary. (2007). “The Check is in the Mail: Correspondent Clearing and the
Banking Panics of the Great Depression,” Journal of Economic History 67(3): 643671.
Richardson, Gary and William Troost. (2009). "Monetary Intervention Mitigated Banking
Panics During the Great Depression: Quasi-Experimental Evidence from the Federal
Reserve District Border in Mississippi, 1929 to 1933." Journal of Political Economy
117(6): 1031-1073.
49
Richardson, Gary and Patrick Van Horn (2011). “When the Music Stopped: Transatlantic
Contagion During the Financial Crisis of 1931.” National Bureau of Economics
Working Paper w17437.
Romer, Christina D. and David H. Romer. (2004). “A New Measure of Monetary Shocks:
Derivation and Implications.” American Economic Review 94 (September): 1055-84.
Romer, Christina D. and David H. Romer. (2010). “The Macroeconomic Effects of Tax
Changes: Estimates Based on a New Measure of Fiscal Shocks.” 100(June): 763-801.
Schubert, Aurel. (1991). The Credit-Anstalt Crisis of 1931. Cambridge: Cambridge
University Press.
Sprague, O.M.W. (1910). History of Crises Under the National Banking System. A Report of
the National Monetary Commission. U.S. Senate, 61st Congress, 3rd Session,
Document 538. Washington: Government Printing Office.
Wicker, Elmus. (1996). The Banking Panics of the Great Depression. Cambridge: Cambridge
University Press.
50
Table 1: National and Regional Banking Panics in the United States during the Great Depression
Name of Period
Pre-panic baseline
Caldwell crisis
First Chicago panic
Crisis after German panic
After England departs gold
Winter ’32 crisis
Second Chicago panic
Winter ‘33 crisis
Weeks
7/1/1929 - 11/1/1930
11/16/1930 - 1/31/1931
6/7/1931 - 6/27/1931
7/26/1931- 9/12/1931
9/13/1931 - 11/7/1931
12/13/1931 - 2/6/1932
6/19/1932 - 8/20/1932
12/18/1932 - 3/4/1933
Number of
Suspensions
NonAll
member
1037
756
133
270
771
662
271
491
904
645
104
206
597
512
203
396
Closed by Bank
Directors
Difference
#
(Pct Pt) Z
770
698
115
229
662
557
252
439
0.18
0.12
0.10
0.11
0.10
0.19
0.15
9.5
3.1
3.5
5.9
4.8
6.6
6.6
Banks Experiencing
Runs
Difference
#
(Pct Pt)
Z
422
402
99
125
511
425
143
265
0.12
0.34
0.06
0.26
0.24
0.12
0.13
5.2
7.4
1.7
10.8
9.4
3.6
4.9
Sources and notes: See the text and Appendix 1 for definitions and data sources. Closed by directors and banks
experiencing runs are tabulations based on examiners’ reports of all suspended banks. Difference (Pct Pt) is the
percentage of banks closed by directors or banks experiencing runs in that period minus the percentage in the pre-panic
baseline value shown in the tables’ initial row. The Z-statistic for the two sample test for equality of proportion of
means indicates that in each case, we can reject the null hypothesis that the percentage in that period equaled the
percentage in the pre-panic period.
Table 2: Bank Suspensions and Interbank Deposit Flows, June 1929- December 1932
Definition of Panic-Period Suspensions and Estimation Method
All:
OLS
All:
Panel FE
All:
Panel FE
10-30 Joins: 10-30 Clusters: Temp + Runs:
Panel FE
Panel FE
Panel FE
Panic Suspensions
-230,563*** -294,683*** -287,594*** -816,996**
(53,759)
(25,003)
(30,750)
(283,229)
Panic Suspensions * Reserve City
318,213*** 316,286***
632,626
(82,851)
(84,894)
(480,682)
Other Suspensions
Other Suspensions * Reserve City
Observations
R-squared
District TVC
Time FE
Time Trend
SE Robust & Clustered
# Fed Districts
196
0.114
No
No
No
No
14
196
0.250
Yes
No
Yes
Yes
14
196
0.307
Yes
Yes
No
Yes
14
-1,996,247**
(747,782)
1,875,112*
(970,537)
-501,302**
(166,766)
338,567
(252,280)
40,955
(186,777)
168,340
(430,397)
80,477
(123,000)
5,513
(215,389)
143,697
(359,695)
242,766
(587,973)
196
0.307
Yes
Yes
No
Yes
14
196
0.310
Yes
Yes
No
Yes
14
196
0.320
Yes
Yes
No
Yes
14
Notes: In Columns 1 to 3, all suspensions are treated as panic suspensions. In columns 4 through 6, 10-30 joins indicates banks suspending within
10 miles of another suspending bank in 30-day windows. 10-30 clusters indicates banks suspending in Davison-Ramirez clusters with parameters
10 days, 30 miles, and 4 banks, as described in text. Temp + Runs is the sum of temporary suspensions plus permanent suspensions that examiners
attributed to runs. Other suspensions are all suspensions minus panic suspensions Standard errors in parentheses calculated using Huber-White
method and are clustered on Federal Reserve districts and central reserve cities. *** indicates p<0.01, ** p<0.05, * p<0.1.
Table 3: Bayesian Model Averaging
Over Specifications shown in Table 2
Suspension Specification
Panic
Temporary + Runs
(Temporary + Runs) * Reserve City
Non-Panic
Other Suspensions
Other * Reserve City
Over All Regression Specifications
Inclusion
Probability Coefficient
0.935
0.067
0.935
0.067
Inclusion
Suspension Specification Probability Coefficient
-472,036
29,579
Panic
Runs
Runs * Reserve City
0.999
0.430
-1,049,586
432,637
206,607
-14,338
Non-Panic
Other Suspensions
Other * Reserve. City
0.999
0.430
935,146
-462,209
Note: Suspension specification reports pairs of panic and non-panic suspensions considered across model specifications.
Calculations of inclusion probabilities and coefficients are described in detail in Appendix 3.
53
Table 4: Estimated Interbank Deposit Flows and Asset Allocations, June 1929- December 1932
Loans
Government
Bonds
Asset Category
Corporate
Bonds
Reserves
Interbank
Deposits
Interbank Deposit Inflows
Reserve City, 1929-1933
Central Reserve City, 1929-1933
-0.137
(0.387)
0.557 ***
(0.180)
0.612 *
(0.319)
0.479 ***
(0.148)
0.143 *
(0.076)
0.325 ***
(0.035)
-0.081
(0.203)
-0.370 ***
(0.093)
0.557 ***
(0.096)
-0.138 ***
(0.045)
0.178
(0.429)
0.542 ****
(0.198)
0.067
(0.354)
0.341 **
(0.16)
0.142 *
(0.084)
0.665 ***
(0.039)
0.038
(0.224)
-0.468 ***
(0.104)
0.590 ***
(0.106)
0.091 *
(0.049)
Interbank Deposit Outflows
Reserve City, 1929-1933
Central Reserve City, 1929-1933
Notes: Regressions are based on equation (6) as described in the text. P-values indicated with asterisks: * p < 0.10; ** p < 0.05; *** p < 0.01.
54
Table 5: Estimates of the Interbank Amplifier during the Great Depression
All:
OLS
Definition of Panic-Period Suspensions and Estimation Method
All:
10-30
10-30
Temp +
Temp +
Panel
All:
Joins:
Clusters:
Runs:
Runs:
FE
Panel FE Panel FE
Panel FE
Panel FE
BMA
Runs:
BMA
July 1929 to December 1932
Gross Amplifier ($millions) 2,287 2,365
Standard Error (Bootstrapped) 1,149
943
% Decline Bank Lending
16.3 16.8
% Lending Failed Banks
68.6 70.9
2,298
1,061
16.4
68.9
Net Amplifier ($millions)
Standard Error (Bootstrapped)
% Decline Bank Lending
% Lending Failed Banks
2,676
913
19.0
80.2
4,842
2,964
34.5
145.2
2,453
992
17.5
73.6
2,603
4,906
18.5
78.1
34.9
147.1
2,560
1,449
18.2
76.8
4,670
3,257
33.2
140.0
1,859
1,277
13.2
55.7
1,755
1,440
12.5
52.6
10.2
43.2
July 1929 to Banking Holiday
Gross Amplifier ($millions) 3,201 3,279
Standard Error (Bootstrapped) 1,149
943
% Decline Bank Lending
17.2 17.6
Net Amplifier ($millions)
Standard Error (Bootstrapped)
% Decline Bank Lending
3,212
1,061
17.2
3,590
913
19.2
5,756
2,964
30.9
3,367
992
18.0
3,517
5,820
18.9
31.2
3,474
1,449
18.6
5,584
3,257
29.9
2,773
1,277
14.9
2,669
2,354
14.3
12.6
Notes: Columns 1-6 are based using estimates from Table 2. Columns 7 and 8 are based on models shown in Table 3. The method for
bootstrapping standard errors is described in the text. Due to missing data, % Lending Failed Banks cannot be calculated for January, February, and
March of 1933.
Table 6: Decomposing Withdrawals During Panics, June 1929 - December 1932
Sample Period
and Estimated Effect
Survivors
Suspensions
Pass-through
All:
OLS
Estimation Method and Definition of Panic-Period Suspensions Employed
10-30
10-30
Temp +
Temp +
All:
All:
Joins:
Clusters:
Runs:
Runs:
FE
Panel FE Panel FE
Panel FE
Panel FE
BMA
66.2
10.2
23.6
89.3
14.0
-3.3
89.6
14.6
-4.2
86.0
6.2
7.8
94.7
3.0
2.3
80.7
8.6
10.7
70.2
6.8
23.0
Runs:
BMA
79.6
3.6
16.8
Notes: Columns 1-6 are based using estimates from Table 2. Columns 7 and 8 are based on models shown in Table 3.
56
0
Suspensions
50
100
150
Figure 1
Suspensions Per Week, 1929 to 1933
Jul 1929
Jul 1930
Jul 1931
Jul 1932
Date
Vertical lines indicate Caldwell crisis, first Chicago panic, German crisis, Britain's departure
from gold, second Chicago panic, winter '32 crisis, and winter '33 crisis.
Join Count
Figure 2: Spatial Autocorrelation of Bank Suspension, Weekly July 1929 through February 1933
Panics identified in Table 1
Note: The solid horizontal bars at level 200 indicate the intervals identified as regional or national banking panics in Table 1.
58
Figure 4: Interbank Deposits, Change $ Million
Reserve Cities Outside New York
Reserve Cities Outside New York
-600
-600
-400
-400
-200
-200
0
0
Figure 3: Demand Deposits, Change $ Million
0
2
Nov 30
Sept 31
4
Weeks After Onset
6
June 31
Dec 31
8
0
Nov 30
Sept 31
July 31
June 32
Figure 5: Balances with Domestic Banks, Change $ Million
-500
-300
-100
0
100
Reserve Cities Outside New York
0
2
Nov 30
Sept 31
4
Weeks After Onset
June 31
Dec 31
8
6
2
July 31
June 32
59
4
Weeks After Onset
June 31
Dec 31
6
8
July 31
June 32
Figure 8: New York City, Change in Deposits $ Million
Weeks Before and After Onset of Panic
Weeks Before and After Onset of Panic
-1500
-1500
-1000
-1000
-500
-500
0
0
Figure 6: Outside New York, Change in Deposits $ Million
-8
-6
-4
-2
0
2
Weeks After Onset
Demand
4
Time
6
-8
8
-6
-4
-2
0
2
Weeks After Onset
Demand
Interbank
4
6
8
Interbank
Time
Vertical lines: 18 January 1933 and 8 March 1933.
Vertical lines: 18 January 1933 and 8 March 1933.
Figure 9: New York City, Change in Assets $ Million
Weeks Before and After Onset of Panic
Weeks Before and After Onset of Panic
-1500
-400 -200
-1000
0
-500
200
0
400
600
500
Figure 7: Outside New York, Change in Assets $ Million
-8
-6
-4
Borrow at Fed
Bonds
-2
0
2
Weeks After Onset
Reserves
Loans
4
6
-8
8
-6
-4
Borrow at Fed
Loans
Loans on Securities
Balances with Banks
-2
0
2
Weeks After Onset
4
6
Loans on Securities
Reserves
Vertical lines: 18 January 1933 and 8 March 1933. Reserves = vault cash + Fed deposits
Vertical lines: 18 January 1933 and 8 March 1933. Reserves = vault cash + Fed deposits
60
8
Bonds
@FedFRASER