Fundamentals, Panics, and Bank Distress During the Depression

Home , Panic of 1930

Fundamentals, Panics, and Bank Distress Duringthe Depression

By CHARLESW. CALOMIRISAND JOSEPHR. MASON*

We assemble bank-level and other data for Fed member banks to model determinants of bankfailure. Fundamentals explain bankfailure risk well. The first two Friedman-Schwartzcrises are not associated with positive unexplained residual failure risk, or increased importanceof bank illiquidityfor forecasting failure. The third Friedman-Schwartzcrisis is more ambiguous, but increased residualfailure risk is small in the aggregate. Thefinal crisis (early 1933) saw a large unexplained increase in bankfailure risk. Local contagion and illiquiditymay have played a role in pre-1933 bank failures, even though those effects were not large in their aggregate impact. (JEL N22, G21, N12, E32, E5)

The central unresolved question about the or sudden crises of systemic illiquidity that may causes of bank distress duringthe Depression is have forced viable banks to fail. The causes of the extent to which the waves of bank failures bank distress are particularlyrelevant from the and deposit contraction (which together define perspective of modem macroeconomictheories bank distress) reflected "fundamental"deterio- of the relationship between bank distress and ration in bank health, or alternatively,"panics" economic fluctuations, and public policy debates about the appropriateresponses of central banks to financialcrises. To the extent that bank distress was not due to fundamentalbank weak- * Calomiris: Graduate School of Business, 601 Uris Hall, ColumbiaUniversity, 3022 Broadway,New York, NY ness, policy actions to protect threatenedbanks 10027, National Bureau of Economic Research, and Amer- via Fed or governmentloans or other assistance ican Enterprise Institute (e-mail: [email protected]); might have preventedfailures and deposit con- Mason: LeBow College of Business, Drexel University, traction. If the collapse of the 3141 Chestnut PA and banking system Street, Philadelphia, 19104, Wharton was driven events within the Financial Institutions Center (e-mail: [email protected]). by banking system We gratefully acknowledge supportfrom the National Sci- (rather than shocks to banks from the "real" ence Foundation,the University of Illinois, and the Federal economy), that would also have importantim- Reserve Bank of St. Louis. In particular,we would like to plications for macroeconomictheory-namely, thank William of the St. Dewald, formerly Louis Fed, and the implication that the financial sector itself the late William Bryan of the University of Illinois, for facilitating the process of transcribingour data from the can be an importantsource of shocks, not just a original microfilm.Jo Ann Landen, formerly of the Federal victim or a propagatorof shocks (see Douglas Reserve Board library, was instrumental in locating the W. Diamond and Phillip H. Dybvig, 1983; microfilm of call reports from which bank balance sheets Franklin Allen and Douglas Gale, 2000; Dia- and income statementswere assembled. We are also grate- ful to Jim Coen and his staff at the Columbia Business mond and RaghuramRajan, 2002). School library for their help in assembling additionaldata. The list of fundamentalshocks that may have Cary Fitzmaurice, Aaron Gersonde, Inessa Love, Jennifer weakened banks is a long and varied one. It Mack, BarberaShinabarger, George Williams, and Melissa includes declines in the value of bank loan Williams provided excellent research assistance. We thank default risk Mike Bordo, Barry Eichengreen, Glenn Hubbard, Peter portfolios producedby rising in the Temin, Elmus Wicker, two anonymous referees, and par- wake of regional, sectoral, or national macro- ticipants in seminars at the NBER Development of the economic shocks to bank borrowers,as well as American Economy Summer Institute meetings, Indiana monetary-policy-induceddeclines in the prices University, Northwestern University, Tulane University, of the bonds held by banks. There is no doubt and the Universityof Pennsylvaniafor helpful comments on an earlier draft. This paper is a revised version of "Causes that adverse fundamental shocks relevant to of U.S. Bank Distress During the Depression" (Calomiris bank solvency were contributorsto bank dis- and Mason, 2000). tress; the controversy is over the size of these 1615 1616 THE AMERICANECONOMIC REVIEW DECEMBER2003 fundamentalshocks-that is, whetherbanks ex- But there are reasons to question Friedman periencing distress were truly insolvent or sim- and Schwartz's view of the exogenous origins ply illiquid. of the banking crises of the Depression. As Milton Friedman and Anna J. Schwartz Calomiris and Gary Gorton (1991) show, pre- (1963) are the most prominentadvocates of the Depression panics were moments of temporary view that many bank failures resulted from un- confusion about which (of a very small number warranted"panic" and thatfailing banks were in of banks) were insolvent. In contrast, as Peter large measure illiquid rather than insolvent. Temin (1976) and many others have noted, the Friedmanand Schwartzattach great importance bank failures during the Depression marked a to the banking crisis of late 1930, which they continuation of the severe banking sector dis- attributeto a "contagion of fear" that resulted tress that had gripped agricultural regions from the failure of a large New York bank, the throughout the 1920's. Of the nearly 15,000 Bank of United States, which they regard as bank disappearances that occurred between itself a victim of panic. 1920 and 1933, roughly half predate 1930. And They also identify two otherbanking crises in massive numbersof bank failures occurreddur- 1931-from March to August 1931, and from ing the Depression era outside the crisis win- Britain's departurefrom the gold standard(Sep- dows identified by Friedman and Schwartz tember 21, 1931) through the end of the year. (notably, in 1932). Wicker (1996, p. 1) esti- The fourthand final bankingcrisis they identify mates that "[b]etween 1930 and 1932 of the occurredat the end of 1932 and the beginning of more than 5,000 banks that closed only 38 per- 1933, culminatingin the nationwidesuspension of cent suspended during the first three banking banks in March.The 1933 crisis and suspension crisis episodes."2 Recent studies of the condi- was the beginningof the end of the Depression, tion of the Bank of United States indicate that it but the 1930 and 1931 crises (becausethey did not too was insolvent, not just illiquid, in December result in suspension) were, in Friedman and 1930 (Joseph Lucia, 1985; Friedman and Schwartz'sjudgment, important sources of shock Schwartz, 1986; Anthony P. O'Brien, 1992; to the real economy that turned a recession in 1929 into the GreatDepression of 1929-1933. Friedmanand Schwartz's (1963) summaryof of the banking crisis of 1933, Barrie A. Wigmore (1987) the aggregatetrends for the macroeconomyand sees the risk of abandoningthe gold standardas an impor- the banking sector focuses on the extreme se- tant exogenous motivator of depositor flight from solvent of the crises (the incidence of banks.Wigmore emphasizesexternal currency drain and the verity banking not and the de- expectation of the departurefrom the gold standard, bank suspension) accompanying concerns over domestic bank solvency, as the precipitating clines in deposits and the money multiplier. event that led to the March 6 declarationof a nationalbank They argue that Federal Reserve errorsof com- holiday. Elmus Wicker (1996) accepts the importanceof the mission (decisions to tighten) and omission external drain in early 1933, but argues that Wigmore un- derestimates the of the crisis that to address the problem of banking importance regional (failures grippedmidwestern banks (beginningwith Michiganbanks) "panic" and bank illiquidity) were central in early 1933. causes of the economic collapse of the Depres- 2 Furthermore,banking distress in the 1930's did not sion. Our interest is in the second aspect-the provokecollective action by banks (clearinghouseactions to of whetherthe were share risks or suspend convertibility),as had been the case question banking collapses Schwartz that "... the solvent but il- in the pre-Federa. Friedmanand argue unwarrantedpanics that forced existence of the Reserve System prevented concerted re- liquid banksto fail. The Friedmanand Schwartz striction... by reducing the concern of stronger banks, argumentis based upon the suddennessof bank- which had in the past typically taken the lead in such a distress the that they identify, concertedmove ... and indirectly,by supportingthe general ing during panics the and the absence of in relevant macro- assumptionthat such a move was made unnecessaryby collapses establishmentof the System" (1963, p. 311). Another pos- economic time series prior to those banking sibility is that collective action was not warranted (i.e., crises (see Charts 27-30 in Friedman and solvent banks were not threatenedby the failures of insol- Schwartz, 1963, p. 309).1 vent banks). Collective action remained feasible, as illus- tratedby the behavior of Chicago banks in June 1932, but Friedmanand Schwartzsee these as exceptions. See F. Cyril James (1938) and Calomiris and Mason (1997) for details 1 Exaggeratedfears of bankinsolvency were not the only on the Chicago panic and the role of collective action in potential contributorsto runs on solvent banks. In the case resolving it. VOL. 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESS DURING THE DEPRESSION 1617

Paul B. Trescott, 1992; Wicker, 1996). So there few locales, althoughhe regardsthe nationwide is some prima facie evidence that the banking increase in the tendency to convert deposits into distress of the Depression era was more than a cash as evidence of a possible nationwidebank- problem of panic-inspireddepositor flight. ing crisis in September and October 1931. But how can one attributebank failures dur- Wicker agrees with Friedmanand Schwartzthat ing the Depression to fundamentalswhen Fried- the final banking crisis (of 1933), which re- man and Schwartz's evidence indicates no prior sulted in universal suspension of bank opera- changes in macroeconomic fundamentals?One tions, was nationwide in scope. The banking possibility is that Friedmanand Schwartzomit- crisis that culminated in the bank holidays of ted importantaggregate measures of the state of February-March1933 resulted in the suspen- the economy relevant for bank solvency. For sion of at least some bank operations (bank example, measures of commercial distress and "holidays")for nearly all banks in the country constructionactivity may be useful indicatorsof by March 6. fundamentalshocks. From the regionally disaggregated perspec- A second possibility is that aggregation of tive of Wicker's findings, the inability to ex- fundamentals masks important sectoral, local, plain the timing of bank failuresusing aggregate and regional shocks that buffeted banks with time-series data (which underlaythe Friedman- particularcredit or market risks. The most im- Schwartz view that banking failures were an portant challenge to Friedman and Schwartz's unwarrantedand autonomous source of shock) aggregate view of bank distress during the De- would not be surprising even if bank failures pression has come from the work of Wicker were entirely due to fundamental insolvency. (1980, 1996). Using a narrativeapproach simi- Failuresof bankswere local phenomenain 1930 lar to that of Friedmanand Schwartz, but rely- and 1931, and so may have had little to do with ing on data disaggregated to the level of the national shocks to income, the price level, in- Federal Reserve districts and on local newspa- terest rates, and asset prices. per accounts of banking distress, Wicker argues The unique industrial organization of the that it is incorrectto identify the banking crisis American banking industryis of central impor- of 1930 and the first banking crisis of 1931 as tance to the Wicker view of the process of bank national panics comparableto those of the pre- failure during the Depression. Banks in the Fed era. According to Wicker,the properway to United States (unlike banks in other countries) understandthe process of banking failure dur- did not operate throughout the country. They ing the Depression is to disaggregate, both by were smaller, regionally isolated institutions. region and by bank, because heterogeneitywas In the United States, therefore, large region- very importantin determiningthe incidence of specific shocks might producea sudden wave of bank failures. bank failures in specific regions even though no Once one disaggregates, Wicker argues, it evidence of a shock was visible in aggregate becomes apparentthat at least the first two of macroeconomictime series (see the cross-country the three banking crises of 1930-1931 identi- evidence in Ben S. Bemanke and HaroldJames, fied by Friedman and Schwartz were largely 1991, and Richard S. Grossman, 1994). regional affairs. Wicker (1980, 1996) argues Microeconomic studies of banking distress that the failures of November 1930 reflected have provided some useful evidence on the re- regional shocks and the specific risk exposures actions of individual banks to economic dis- of a small subset of banks, linked to Nashville- tress, which bears on these macroeconomic based Caldwell & Co., the largest investment debates. Eugene N. White (1984) showed that bank in the South at the time of its failure. the failures of banks in 1930 are best explained Temin (1989, p. 50) reaches a similar conclu- as a continuationof the agriculturaldistress of sion. He arguesthat the "panic"of 1930 was not the 1920's, and were traceable to fundamental really a panic, and that the failure of Caldwell & disturbancesin agriculturalmarkets. Calomiris Co. and the Bank of United States reflected and Mason (1997) studied the Chicago banking fundamentalweakness in those institutions. panic of June 1932 (a locally isolated phenom- Wicker's analysis of the third banking crisis enon). They found that the panic resulted only (beginning September 1931) also shows that in a temporaryunwarranted contraction of de- bank suspensions were concentratedin a very posits; local fundamentalsdetermined both the 1618 THE AMERICANECONOMIC REVIEW DECEMBER2003 long-run contraction of bank deposits and of individualbank failures to the characteristics which Chicago banks failed before and during of individual banks, and to the changing local, the panic. Calomiris and Berry Wilson (1998) regional, and nationaleconomic environmentin studied the behavior of New York City banks which they operated.A detailed, disaggregated during the interwar period, and in particular, model of the fundamentaldeterminants of bank analyzed the contractionof their lending during failure makes possible the evaluation of the the 1930's. They found that banking distress relative importanceof contagion for generating was an informedmarket response to observable banking distress. weaknesses in particularbanks, traceableto ex To summarizeour objectives, we seek (1) to ante bank characteristics. gauge the extent to which the attributesof spe- Taken together,these studies suggest that lo- cific banks, in concert with the fundamental cal fundamentalsplayed a large role in gener- local or national shocks that buffeted those ating banking distress during the Depression. banks, can explain the timing and incidence of From the standpoint of the larger macroeco- bank failures, (2) to evaluate the importanceof nomic questions that underlie much of the in- panic or contagion-nationally or locally-as a terest in the origins of banking distress during cause of bank failureduring the Depression, and the Depression, however, existing microecono- (3) to identify the extent to which particular metric contributions suffer from three weak- bankingcrises were nationalor regional events. nesses. First, they rely upon limited samples. Our investigation of the causes of banking Analysis of banks in particularlocations, or at distress relies upon the fact that the U.S. bank- particulartimes, may paint a misleading picture ing system was geographically fragmented. In of the causes of bankingdistress for the country most states, banks were not free to operate as a whole duringthe Depression. Second, some branches (the so-called "unit"banking restric- of the previous microeconomic studies have tion). Even in states that permitted branching used sources that contain a limited set of bank within the state, branching was often limited, characteristics,and which exclude characteris- and in all cases, branching was not allowed tics that are likely to be importantin modeling outside the state.3 Geographic fragmentation bank distress (as indicated by the results of of banking permits one to identify location- Calomiris and Mason, 1997, which show the specific and bank-specific determinantsof fail- advantage of including a relatively rich set of ure for a large sample of banks, and to characteristics). investigatewhether the failures of banks located Third, none of the microeconometricstudies nearby affected the probabilityof a bank's fail- has tried to measure the relative importanceof ure (a local contagion effect). fundamentals and "contagion" for explaining The chief limitation of our data set is that it bank failures at the regional or national level. only covers Fed member banks (nationalbanks This is an important omission. The fact that plus state-charteredbanks that belonged to the regional shocks were important(as argued by Federal Reserve System). Most bank failures Wicker and others) does not in itself disprove during the Depression were nonmemberbanks, the Friedman-Schwartz view that runs on so there is some question as to whether our banks resulted in large part from panic. Indeed, results offer an adequateportrayal of the expe- Wicker-who disputes the existence of nation- rience of all banks. We discuss this issue in wide panics in 1930 and early 1931-argues more detail in Section I below. that local and regional panics contributed to The remainder of this paper is organized bank failures over and above fundamentalre- as follows. Section I briefly describes the data gional shocks. set and defines and explains the limits of our This paper assembles a rich disaggregated investigation-that is, why we confine our data set capable of linking fundamentalsources attentionto certain measures of economic per- of bank weakness-individual Fed member formance, and to Fed memberbanks' behavior. bank's portfolio and liability structureand con- Section II contains our analysis of the causes of dition, and local, regional, and national economic shocks-to the process of bank failure. 3 We construct a survival duration model of See Calomiris(2000) for a review of the history of unit banks that relates informationabout the timing banking restrictionsand their costs. VOL. 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESSDURING THE DEPRESSION 1619 bank failure using data on individual banks. tor that shows sudden change similar to that of Specifically, in Section II we construct a sur- bank failures is the liabilities of failed busi- vival durationmodel for banks and consider the nesses, and it does not show increases prior to significance of bank characteristics,shocks to the firstthree panic episodes identifiedby Fried- the economic environment, and various mea- man and Schwartz,although it often does move sures of "contagion"or "panic"for reducingthe in parallel to bank failure risk. The evidence probability of bank survival. Section III sum- presentedin Tables 2-3 and Figures 1-3 shows marizes our results and concludes. that our sample of Fed member banks provides pictures of the timing of total bank failures, the I. Data relationship between aggregate bank failures and macroeconomic aggregates, and the re- The sources and definitions of the data used gional and temporal distributionof bank fail- in our empirical work are discussed in detail in ures that are similar to those in Friedman and the Data Appendix. Our data set combines data Schwartz (1963) and Wicker (1996). Visual in- on individualbank characteristicsfor Fed mem- spection of aggregate variables indicates that ber banks observed in December 1929 and they are not very helpful in predictingthe Fried- December 1931 with county-, state-, and man and Schwartz crisis windows, and the national-level data at monthly, quarterly, and cross-sectional variation emphasized by Wick- annual frequencies. These data permit us to er's discussion of suspensions at the Fed Dis- measure bank distress by date of failure at var- trict level is quite visible in the patternof bank ious levels of disaggregation, and to capture a failures at the state level. These tables and fig- variety of influences on bank distress. Ta- ures provide prima facie evidence for the desir- ble 1 summarizesthe measures of bank charac- ability of disaggregating the analysis of bank teristics we constructed and the measures we failure and examining connectionsbetween fun- employ to capture variation in the local, re- damental determinantsof bank weakness and gional, and national economic environment. the probabilityof bank failures. Tables 2 and 3 provide information about Despite the fact that the nationaland regional variation over time and across regions in the aggregatetime series of suspension rates for all incidence of bank failure, which we define as banks coincides with the national and regional bank closure and liquidation.Tables 2a and 2b average survival hazardsfor our sample of Fed reportsemiannual numbers and deposits of Fed member banks, the absence of nonmember member banks that failed, by region. Tables 3a banks from our sample is an importantlimita- and 3b express these regional-level measuresof tion of our analysis of bank failure, which may bank failure as fractions of total Fed member matter for more disaggregated results. As of banks, or total Fed member bank deposits, in June 30, 1929, nonmember banks comprised each region at the end of 1929. The data re- 15,797 of the 24,504 banks in existence (of portedin Tables 2 and 3 have not been collected which 7,530 were national banks and 1,177 or reported in previous studies (more detailed were state-chartered member banks). Non- data are described in Calomiris and Mason, member banks were smaller on average, ac- 2000). These data clearly show the remark- counting for 27 percent of total bank deposits. able heterogeneity in regional experiences of Failure rates were higher for nonmember bank distress and deposit growth during the banks. Nonmember banks fell as a proportion Depression. of total banks from 63 percent of the number Figures 1-3 report various macroeconomic of banks in June 1929 to 57 percent by June time series alongside our measure of Fed mem- 1933. In Calomiris and Mason (2000), we ber banks' conditional failure hazard. These found that indicators of the condition of Fed data provide a similarpicture of aggregatebank member banks within the county were useful distress over time to the evidence on bank sus- indicators of annual suspension rates or de- pension rates in Friedman and Schwartz, and posit growth rates at the county level for all confirmFriedman and Schwartz's view that ag- banks. Despite that evidence for the represen- gregate macroeconomic indicators provide a tativeness of Fed member banks, it is possible poor explanation for the timing of waves of that nonmember banks had different sensi- bank failures. The only macroeconomicindica- tivities to panic events, so our conclusions 1620 THE AMERICANECONOMIC REVIEW DECEMBER2003

TABLE1-VARIABLE DEFINITIONS

BANK CHARACTERISTICS,Measured Biannually (December 1929, December 1931) Basic bank characteristics: LTotAss = log (Total Assets) STBANK = State-CharteredIndicator (equal to 1 for State-CharteredBank) LNBRANCH = log [ max (numberof branches,0.0010) ] MKTPWR = Total Deposits / Deposits of All Banks in the County Bank Asset Composition NonCash_TotAss = "Non-Cash"Assets / Total Assets "Non-Cash"Assets = Total Assets - (U.S. Govt. Securities + Reserves + Cash Due from Banks + Outside Checks and Other Cash Items) Loans_OtherNonCash= Loans and Discounts / (Noncash Assets - Loans and Discounts) LIQLOANS = Loans Eligible for Rediscount / Loans and Discounts DFB_CashAss = Cash due from Banks / (U.S. Govt. Securities + Reserves + Cash Due from Banks + Outside Checks and Other Cash Items) Asset quality measures: Losses_Exp = Losses on Assets and Trading/ Total Expenses (Including Losses) REO_NonCashAss= Real Estate Owned / Noncash Assets (BONDYLD)X(SEC) = (Change in U.S. Govt. Bond Yield)X(Bonds and Other Securities) Change in U.S. Govt. Bond Yield = (This Month's Bond Yield - Bond Yield of Same Month in Previous Year) Liability mix and cost: TD = Total Deposits = Due to Banks + Demand Deposits + Time Deposits + U.S. GovernmentDeposits + Bills Payable and Rediscounts NW_TA = (Capital + Surplus + Undivided Profits + Contingency Reserve) / TA (DD + DTB)_TD = Demand Deposits + Due to Banks / TD DTB_TD = Due to Banks / TD BPR_TD = Bills Payable and Rediscounts / TD PrivBPR_BPR = Private Bills Payable and Rediscounts / BPR INTCOST = Interestand Discount Expenses on TD / TD COUNTY CHARACTERISTICS,Measured in 1930, Unless Otherwise Noted PCT_CROPINC30= Crop Value /(Crop Value + ManufacturingValue Added) PCT_ACRES_PAST30= Acreage in Pasture/ Total Acreage in Farms VALGR_INC_CROP30= Value of Cereals, Oats, Grains, Seeds / Total Crop Value UNEMP30 = (Persons Out of Work + Persons Laid Off) / Number of Gainful Workers SMLFM30 = Farms of Less Than 100 Acres / Total Number of Farms (DAGLBE) X (PCT_CROPINC30)= (PCT_CROPINC30)x (Growth in Value of Farm Land, Buildings, and Equipmentfrom 1920 to 1930) PCT_STBANK (annual data) = Number of State-CharteredBanks, Including NonmemberBanks/Total Number of Banks STATE ECONOMICENVIRONMENT STBUILDPERM(monthly) = Value of Buildings with New Permits in Cities within the State / State Income in 1929 STBUSFAIL (quarterly)= Value of Liabilities of Failed Businesses / State Income in 1929 NATIONAL ECONOMICENVIRONMENT NATDAGP (monthly) = Log Difference, AgriculturalPrice Index, Seasonally Adjusted NATDBUSFAIL (monthly) = Log Difference (CurrentLog Value Less Log Value for Same Month in Previous Year), Value of Liabilities of Failed Businesses DISTRESS INDICATORVARIABLES FSPANIC-30 = 1 for November and December 1930 and January1931, and 0 Otherwise FSPANIC-31a = 1 for May-June 1931, and 0 Otherwise FSPANIC-31b = 1 for September-November 1931, and 0 Otherwise DUM_JAN-33 = 1 for January1933, and 0 Otherwise DUM_FEB-33 = 1 for February1933, and 0 Otherwise DUM_MAR-33 = 1 for March 1933, and 0 Otherwise WICKER-30 = 1 for November 1930-January 1931 for Banks in Tennessee, Kentucky, Arkansas,North Carolina, and Mississippi, and 0 Otherwise WICKER-31a = 1 for April-July 1931 for Banks in Illinois and Ohio, and 0 Otherwise WICKER-31b = 1 for September-October1931 for Banks in West Virginia, Ohio, Missouri, Illinois, and Pennsylvania,and 0 Otherwise Chicago-6-32 = 1 for Banks in Chicago for June 1932, and 0 Otherwise NEARFAILS = Log (Deposits in Other Banks That Failed in that Month in the Same State)

Sources: See Data Appendix. VOL.93 NO. 5 CALOMIRISAND MASON:BANK DISTRESS DURING THE DEPRESSION 1621

TABLE2a-NUMBER OF FAILED SAMPLE BANKS, SEMIANNUALLY, 1930-1933

Region 1930-1 1930-2 1931-1 1931-2 1932-1 1932-2 1933-All Total Central 15 29 45 75 59 25 130 378 Mid-Atlantic 7 14 28 51 33 16 58 207 Mountain 8 7 5 16 9 9 18 72 New England 1 1 4 9 6 0 22 43 Northwestern 25 30 42 68 37 43 114 359 Pacific 3 3 10 10 23 14 21 84 South Atlantic 21 16 21 36 20 9 26 149 South Central 28 35 49 42 35 32 50 271 Total 108 135 204 307 222 148 439 1,563

Notes: Tables are constructedfrom 1,716 failed banks in the authors'total data set of 8,470 FederalReserve MemberBanks. Deposits are those recordedfrom the last available call reportprior to failure. Regions are defined as follows: New England contains ME, NH, VT, MA, RI, and CT. Mid-Atlanticcontains NY, NJ, and PA. South Atlantic contains MD, DE, DC, VA, WV, NC, SC, GA, and FL. South Centralcontains KY, TN, AL, MS, AR, OK, LA, and TX. Centralcontains OH, IL, IN, MI, and WI. Northwesterncontains MN, IA, MO, ND, SD, NE, and KS. Mountaincontains MT, ID, WY, CO, NM, AZ, UT, and NV. Pacific contains WA, OR, and CA. State-level quarterlyversions of these tables are available in Calomiris and Mason, 2000.

TABLE2b-DEPOSITS OF FAILED SAMPLE BANKS, SEMIANNUALLY, 1930-1933 ($ THOUSANDS)

Region 1930-1 1930-2 1931-1 1931-2 1932-1 1932-2 1933-All Total Central 30,587 66,660 92,809 304,891 118,229 216,993 898,204 1,728,374 Mid-Atlantic 57,515 34,606 101,747 146,340 310,550 16,631 135,573 802,962 Mountain 5,965 1,908 1,839 6,532 2,754 12,719 16,475 48,193 New England 913 1,686 9,542 49,280 137,536 0 56,228 255,184 Northwestern 29,825 21,765 24,041 36,104 33,411 32,573 116,033 293,752 Pacific 1,018 5,521 5,067 8,318 29,195 13,686 70,224 133,029 South Atlantic 42,184 54,850 16,535 38,854 29,555 5,814 127,409 315,200 South Central 26,140 77,399 60,867 39,933 37,865 61,537 28,311 332,051 Total 194,146 264,395 312,447 630,252 699,095 359,953 1,448,457 3,908,746

Note: See notes for Table 2a. below about Fed member banks may not hold One of the advantages of the survival hazard for nonmember banks. model is its flexibility in using data observed at differentlevels of aggregationand differentfre- II. ModelingBank Failure:Fundamentals and quencies. County-level variables (which are Contagion only observed once during the sample period) exert a constant effect on the hazardrate, bank- Our bank failure data, which track the spe- specific variables (observed biannually at call cific dates of each Fed member bank failure, report dates) affect the hazard rate for two allow us to model each bank's daily failure years, and state- and national-level monthly or hazard as a function of various fundamentals, quarterly series affect the hazard rate on a including bank-specific variables observed at monthly or quarterlybasis. earlier call report dates, county characteristics, Our model of the determinants of failure and state- and national-level time series ob- starts with many of the same bank-level deter- served at relatively high frequency.All survival minants that were found to be useful in Calo- durationmodels we report are estimated using miris and Mason (1997) to explain bank failures the log-logistic distribution. Detailed descrip- during the Chicago panic of June 1932. Our tions of the survival durationmethodology can model of bank failures throughoutthe country be found in Nicholas M. Kiefer (1988), Tony over several years differs, however, from that Lancaster(1990), and Guido W. Imbens (1994). earlier paper (which focused on failures 1622 THE AMERICANECONOMIC REVIEW DECEMBER2003

TABLE3a-NUMBER OF FAILED SAMPLE BANKS, SEMIANNUALLY, 1930-1933, AS PERCENTOF TOTAL NUMBER OF SAMPLE BANKS, 1929

Region 1930-1 1930-2 1931-1 1931-2 1932-1 1932-2 1933-All Total Central 0.27 0.53 0.81 1.36 1.07 0.45 2.35 6.85 Mid-Atlantic 0.23 0.46 0.92 1.67 1.08 0.52 1.90 6.77 Mountain 1.01 0.88 0.63 2.01 1.13 1.13 2.26 9.05 New England 0.15 0.15 0.59 1.32 0.88 0.00 3.22 6.30 Northwestern 0.40 0.47 0.66 1.07 0.58 0.68 1.80 5.67 Pacific 0.31 0.31 1.02 1.02 2.34 1.43 2.14 8.55 South Atlantic 0.89 0.68 0.89 1.53 0.85 0.38 1.10 6.32 South Central 0.64 0.80 1.12 0.96 0.80 0.73 1.15 6.22 Total 0.45 0.56 0.84 1.27 0.92 0.61 1.82 6.47

TABLE3b-DEPOSITS OF FAILEDSAMPLE BANKS, SEMIANNUALLY,1930-1933, AS PERCENTOF TOTALDEPOSITS IN SAMPLEBANKS, 1929

Region 1930-1 1930-2 1931-1 1931-2 1932-1 1932-2 1933-All Total Central 0.30 0.64 0.90 2.94 1.14 2.09 8.67 16.68 Mid-Atlantic 0.28 0.17 0.49 0.71 1.50 0.08 0.65 3.87 Mountain 0.72 0.23 0.22 0.79 0.33 1.54 2.00 5.84 New England 0.03 0.05 0.30 1.52 4.25 0.00 1.74 7.89 Northwestern 0.69 0.51 0.56 0.84 0.78 0.76 2.70 6.83 Pacific 0.02 0.12 0.11 0.19 0.66 0.31 1.58 3.00 South Atlantic 1.49 1.94 0.58 1.37 1.05 0.21 4.51 11.15 South Central 0.69 2.04 1.61 1.05 1.00 1.62 0.75 8.76 Total 0.38 0.52 0.61 1.24 1.37 0.71 2.84 7.67

Note: See notes for Table 3a. occurringin one city during a brief time inter- only one lagged value of each, based on some val); in the empirical analysis here we include experimentationto find the lag length with the county-level, state-level, and national-level greatest explanatory power. Below we report variablesin additionto bank-specificcharacter- results using lags of five months for state-level istics. Our sample period for dating bank fail- building permits, national-level agricultural ures is from January 1930 to December 1933, prices, and national-level liabilities of failed and our fundamentals(on which the predictions businesses, and three quarters for state-level of survival or failure are based) are observed liabilities of failed businesses. We also experi- from January1930 throughMarch 1933. mented with using moving averages of these Our explanatory variables are expressed as variables. The results described below for the ratios (ratherthan log ratios) to avoid omitting influence of other variables are robust to varia- from the sample observations with a value of tion in the specific lag structuresof the high- zero. In resultsnot reportedhere, we definedour frequency variables. The definitions of the regressorsas log ratios, and this transformation variables used in the regressions are given in did not affect our results much, but did reduce Table 1 and summary statistics for these vari- our sample size. For the high-frequency state- ables are provided in Table 4. level and national-level variables we included Table 5 (a and b) presents survival duration VOL. 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESSDURING THE DEPRESSION 121623

o= o 0 0 I - V4 V- C14 r4 (1 cn e en e cn en ~~~~~~~en

.-inLog NationalLiabilities of FailedBusinesses (LH Scale) -n*-*ConditionalProbability of Failure (RH Scale)

FIGURE 1. LIABILITIES OF FAILED BUSINESS AND FAILURE PROBABILITY, JANUARY 1930-DECEMBER 1933

0.018 4.8001 0.014 0.012 0.010

0.006 3.3 ~~~~~~~~~~~~~~~~~~~~0.004 0.002

2.8 w 9w a 0.000 o0 0 0 ~~-"V4 - -4 N (C CC4 e41 rfen n

-~~ b Q~~~~ 0 0. -1 C0

-*Log AgriculturalPrice Index (LH Scale) 4Conditional Probabilityof Failure (RH Scale)

FIGURE 2. AGRICULTURAL PRICE INDEX AND FAILURE PROBABILITY, JANUARY 1930-DECEMBER 1933 results for the period January 1930 through exercise judgment about the "6correct"timing of March 1933. Includingbank failures in 1933 in the failure of banks. Bank holidays were de- our study posed a problem that requiredus to dlaredat the state and nationallevels in February 1624 THE AMERICANECONOMIC REVIEW DECEMBER2003

-112 _0.018 0.016

-1 13 0.014 0.012 0.010 -114 0.008 0.006 0.004

-1 0.002 16 0.000

-*-Log BuildingPermits Divided by State Income (LH Scale) ml-ConditionalProbability of Failure (RH Scale)

FIGURE 3. BUILDING PERMITS AND FAILURE PROBABILITY, JANUARY 1930-DECEMBER 1933

and March 1933, which entailed the partialsus- In our survival analysis reported below, we pension of bank operationsfor periods of time. truncateboth the determinantsof failure and the Many banks failed duringand immediatelyafter observed failure dates in March 1933. We iden- the bank holidays. Some banks that did not tify only those banks that officially failed in reopen in March 1933 after suspension re- March as March failures. We also tried several mained in a state of regulatorylimbo for several alternativeapproaches to dealing with the prob- months. Many of these banks failed in late 1933 lem of bank holidays. One alternativeapproach after the regulators and the ReconstructionFi- would be to assume that the banks that officially nance Corporation (RFC) decided not to ap- failed between April and December 1933 had prove them for membership in the Federal actually failed in March 1933. The results for Deposit Insurance Corporation(FDIC), which that approachare also similar to those reported began operationin January1934. The decision below, except that (by construction)there is a to permit banks to reopen sometimes followed large and significant residual for the month of approval of assistance from the RFC, and March 1933. We chose to reportthe firstversion Mason (2001a) finds empirical evidence that over this alternativeapproach because we think preferredstock assistance from the RFC (which that despite its limitations, the first approach began in 1933) did help banks to avoid failure. distinguishesto some extent between banks that Thus the meaning and the timing of bank failed in March and those that failed later in failures become less clear after February1933. 1933, which were arguably stronger. Another In particular,some banks that officially failed alternativeapproach is to truncateall observa- after March 1933 could be deemed reasonably tions of regressorsand failures in January1933. to have failed in March, and some banks that The coefficients derived for the determinantsof did not fail officially could be deemed to have failure using that approachare very similar to failed in March but been rescued by the RFC's those we report below. The problem with that new preferredstock program.We experimented third approachis that it does not permit us to with many alternativeways of dealing with the examine whetherthere are unexplainedresidual problem of the bank holidays. failures during the alleged panic of early 1933 VOL. 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESSDURING THE DEPRESSION 1625

TABLE4-SUMMARY STATISTICS

Standard Variable N Mean deviation Minimum Maximum Survival Model (Full Sample) Dependent variable Log(DAYS UNTIL FAILURE) 269,683 5.913 1.320 0.000 7.078 MONTHLY BANK FAILURE RATE 269,683 0.005 0.068 0.000 1.000 Bank data, December 31, 1929 LTotAss 7,553 13.974 1.265 10.960 21.312 STBANK 7,553 0.112 0.316 0.000 1.000 LNBRANCH 7,553 -8.858 1.861 -9.210 4.934 MKTPWR 7,553 0.993 0.067 0.038 1.000 NonCash_TotAss 7,553 0.766 0.107 0.064 0.965 Loans_OtherNonCash 7,553 0.744 0.186 0.030 0.997 LIQLOANS 7,553 0.284 0.216 0.000 0.999 Losses_Exp 7,553 0.165 0.145 0.000 0.911 REO_NonCash 7,553 0.013 0.025 0.000 0.340 (BONDYLD) X (SEC) 7,553 -0.007 0.004 -0.023 0.000 (DD + DTB)_TD 7,553 0.520 0.229 0.000 1.000 DTB_TD 7,553 0.033 0.061 0.000 0.748 DFB_CashAss 7,553 0.281 0.172 0.000 1.000 BPR_TD 7,553 0.028 0.049 0.000 0.504 PrivBPR_BPR 7,553 0.052 0.165 0.000 0.993 NW_TA 7,553 0.149 0.061 0.031 0.601 INTCOST 7,553 0.011 0.010 0.000 0.598 Bank data, December 31, 1931 LTotAss 6,857 13.887 1.325 10.752 21.197 STBANK 6,857 0.126 0.332 0.000 1.000 LNBRANCH 6,857 -8.856 1.872 -9.210 4.934 MKTPWR 6,857 0.990 0.082 0.026 1.000 NonCash_TotAss 6;857 0.760 0.109 0.130 0.978 Loans_OtherNonCash 6,857 0.701 0.192 0.015 0.997 LIQLOANS 6,857 0.253 0.198 0.000 0.999 Losses_Exp 6,857 0.298 0.203 0.000 0.926 REO_NonCash 6,857 0.014 0.024 0.000 0.385 (BONDYLD) X (SEC) 6,857 0.080 0.043 0.001 0.258 (DD + DTB)_TD 6,857 0.467 0.233 0.000 1.000 DTB_TD 6,857 0.028 0.055 0.000 0.683 DFB_CashAss 6,857 0.244 0.171 0.000 1.000 BPR_TD 6,857 0.048 0.072 0.000 0.588 PrivBPR_BPR 6,857 0.069 0.180 0.000 1.000 NW_TA 6,857 0.166 0.069 0.010 0.635 INTCOST 6,857 0.012 0.019 0.000 0.995 Distress variables FSPANIC-30 269,683 0.081 0.272 0.000 1.000 FSPANIC-31a 269,683 0.052 0.222 0.000 1.000 FSPANIC-31b 269,683 0.075 0.263 0.000 1.000 DUM_JAN-33 269,683 0.023 0.151 0.000 1.000 DUM_FEB-33 269,683 0.023 0.151 0.000 1.000 DUM_MAR-33 269,683 0.023 0.150 0.000 1.000 (FSPANIC-30) X (DFB_CashAss) 269,683 0.023 0.091 0.000 1.000 (FSPANIC-31a) X (DFB_CashAss) 269,683 0.015 0.074 0.000 1.000 (FSPANIC-31b) X (DFB_CashAss) 269,683 0.021 0.088 0.000 1.000 WICKER-30 269,683 0.003 0.054 0.000 1.000 WICKER-31a 269,683 0.008 0.090 0.000 1.000 WICKER-31b 269,683 0.006 0.079 0.000 1.000 (WICKER-30) X (DFB_CashAss) 269,683 0.001 0.021 0.000 1.000 (WICKER-31a) X (DFB_CashAss) 269,683 0.002 0.025 0.000 1.000 (WICKER-31b) X (DFB_CashAss) 269,683 0.001 0.021 0.000 1.000 Chicago-6-32 269,683 0.000 0.015 0.000 1.000 NEARFAILS 269,683 0.972 15.235 -16.118 20.026 1626 THEAMERICAN ECONOMIC REVIEW DECEMBER2003

TABLE4-Continued

Standard Variable N Mean deviation Minimum Maximum Survival Model (215 City Sample) Log(DAYS UNTIL FAILURE) 53,032 5.922 1.317 0.000 7.078 MONTHLY BANK FAILURE RATE 53,032 0.004 0.065 0.000 1.000 Bank data, December 31, 1929 LTotAss 1,470 15.057 1.506 11.645 20.862 STBANK 1,470 0.196 0.397 0.000 1.000 LNBRANCH 1,470 -7.970 3.342 -9.210 4.934 MKTPWR 1,470 0.974 0.124 0.044 1.000 NonCash_TotAss 1,470 0.786 0.097 0.245 0.965 Loans_OtherNonCash 1,470 0.727 0.175 0.030 0.996 LIQLOANS 1,470 0.189 0.172 0.000 0.980 Losses_Exp 1,470 0.143 0.123 0.000 0.799 REO_NonCash 1,470 0.007 0.014 0.000 0.121 (BONDYLD) x (SEC) 1,470 -0.007 0.004 -0.022 0.000 (DD + DTB)_TD 1,470 0.511 0.215 0.000 1.000 DTB_TD 1,470 0.058 0.082 0.000 0.559 DFB_CashAss 1,470 0.256 0.162 0.000 1.000 BPR_TD 1,470 0.029 0.045 0.000 0.294 PrivBPR_BPR 1,470 0.058 0.170 0.000 0.993 NW_TA 1,470 0.148 0.065 0.045 0.601 INTCOST 1,470 0.012 0.006 0.000 0.151 Bank data, December 31, 1931 LTotAss 1,383 15.004 1.570 11.462 20.720 STBANK 1,383 0.205 0.404 0.000 1.000 LNBRANCH 1,383 -7.976 3.352 -9.210 4.934 MKTPWR 1,383 0.961 0.159 0.037 1.000 NonCash_TotAss 1,383 0.764 0.110 0.237 0.962 Loans_OtherNonCash 1,383 0.678 0.177 0.015 0.993 LIQLOANS 1,383 0.161 0.143 0.000 0.998 Losses_Exp 1,383 0.306 0.200 0.000 0.926 REO_NonCash 1,383 0.010 0.016 0.000 0.144 (BONDYLD) x (SEC) 1,383 0.080 0.042 0.001 0.239 (DD + DTB)_TD 1,383 0.454 0.223 0.000 1.000 DTB_TD 1,383 0.053 0.077 0.000 0.589 DFB_CashAss 1,383 0.219 0.162 0.000 0.877 BPR_TD 1,383 0.044 0.063 0.000 0.376 PrivBPR BPR 1,383 0.088 0.207 0.000 0.956 NW TA 1,383 0.159 0.070 0.010 0.635 INTCOST 1,383 0.013 0.035 0.000 0.995 Distress variables FSPANIC-30 53,032 0.079 0.271 0.000 1.000 FSPANIC-31a 53,032 0.051 0.221 0.000 1.000 FSPANIC-31b 53,032 0.074 0.262 0.000 1.000 DUM_JAN-33 53,032 0.024 0.152 0.000 1.000 DUM_FEB-33 53,032 0.024 0.152 0.000 1.000 DUM_MAR-33 53,032 0.023 0.151 0.000 1.000 (FSPANIC-30) X (DFB_CashAss) 53,032 0.020 0.083 0.000 1.000 (FSPANIC-31a) x (DFB_CashAss) 53,032 0.013 0.067 0.000 1.000 (FSPANIC-31b) x (DFB_CashAss) 53,032 0.019 0.080 0.000 1.000 WICKER-30 53,032 0.001 0.036 0.000 1.000 WICKER-31a 53,032 0.008 0.088 0.000 1.000 WICKER-31b 53,032 0.007 0.081 0.000 1.000 (WICKER-30) x (DFB.CashAss) 53,032 0.000 0.013 0.000 0.649 (WICKER-31a) x (DFB_CashAss) 53,032 0.002 0.027 0.000 1.000 (WICKER-31b) x (DFB_CashAss) 53,032 0.001 0.020 0.000 1.000 Chicago-6-32 53,032 0.001 0.033 0.000 1.000 NEARFAILS 53,032 0.729 15.422 -16.118 20.026 VOL.93 NO. 5 CALOMIRISAND MASON:BANK DISTRESS DURING THE DEPRESSION 1627

TABLE4-Continued

Standard Variable N Mean deviation Minimum Maximum County Data PCT_CROPINC30 2,187 0.991 0.059 0.000 1.000 PCT_ACRES_PAST30 2,249 0.386 0.205 0.000 1.000 VALGR_INC_CROP30 2,259 0.416 0.281 0.000 0.982 UNEMP30 2,252 0.044 0.031 0.000 0.271 SMLFM30 2,254 0.534 0.292 0.000 1.000 (DAGLBE) X (PCT_CROPINC30) 2,187 -0.223 1.009 -1.534 25.805 PCT_STBANK 2,259 0.583 0.251 0.000 1.000 Quarterly State Data STBUSFAIL 565 -7.006 2.551 -24.488 -3.640 Monthly State Data STBUILDPERM5 1,693 -14.423 1.188 -19.290 -11.716 STBUILDPERM3 1,693 -13.781 1.083 -15.056 26.185 Monthly National Data NATDAGP 39 0.003 0.036 -0.070 0.078 NATDBUSFAIL 39 -0.003 0.172 -0.349 0.432

(that is, failures significantly greater than pre- ables for specific panic windows identified by dicted by a stable model of failure determinants Friedmanand Schwartz(1963). Second, we add for the whole period). regional panic indicatorvariables to capturethe Our model of bank survival posits that the regional panics identified by Wicker (1996), durationof survival (measuredin days) depends and the Chicago 1932 panic. Calomirisand Ma- on fundamentals,which are measured at up to son (1997) show that Chicago did indeed suffer monthly frequency.The survival statusof banks a panic in June 1932, but that runs on banks after March 1933 is treated as unknown. For during the panic did not result in the failures of each month from January 1930 until March solvent banks. We include the Chicago panic 1933 the future survival paths of banks are variable not to test for contagion-inducedfail- regressed on fundamentalsto compute the pre- ures there (since our tests are less informative dicted survival hazardfunction (i.e., the coeffi- for answering that question than our earlier pa- cients for the model). per) but rather to gauge the extent to which Table 5a reportsresults for what we term the indicatorvariables may exaggeratethe extent to "basic model," which includes fundamentals which panics induced bank failures because of and a time trend. The eight columns in Ta- missing location-specific fundamental indica- ble 5b report coefficient values for eight addi- tors, as we discuss further below. Third, we tional specifications that include variables include a measure of local contagion (NEAR- intended to capture the possible presence of FAILS) to capture the effect of the failure of panic, contagion, or illiquidity crises. For the nearbybanks (otherbanks within the same state most part, the coefficients on fundamentalsin that failed in that same month) for predicting a Table 5b do not change importantlywhen the bank's probabilityof failure. various panic variables are added to the basic Fourth, we consider "interactioneffects" re- specification, and to conserve space we do not lated to panics. Specifically, we investigate reportthose coefficients. The exceptions are the whether measures of bank liquidity or linkages coefficients on (BONDYLD) X (SEC) and among banks through interbank deposits had NATDAGP_Lag5M, which do vary across special effects on bank failure hazard during specifications. episodes identified as panics by prior authors. We consider four types of variables to cap- For example, the ratio of interbank deposits ture illiquidity crises, contagion, or panics. owed to total bank deposits (DTB_TD) may First, we include national-level indicator vari- capturea bank's susceptibility to liquidity risk. 1628 THEAMERICAN ECONOMIC REVIEW DECEMBER2003

TABLE5a-SURVIVAL REGRESSIONS FOR INDIVIDUAL FED MEMBER BANKS, DEPENDENT VARIABLE: LOGDAYS UNTIL FAILURE AFTER DECEMBER 31, 1929 FULLSAMPLE OF FED MEMBER BANKS (StandardErrors in Parentheses)

(1) (1) (Continued) Constant 6.044 BPR_TD -1.490 (0.283) (0.146) LTotAss 0.105 PrivBPR_BPR -0.126 (0.011) (0.050) STBANK 0.136 INTCOST -0.671 (0.031) (0.428) LNBRANCH -0.012 PCT_INC_CROP30 0.317 (0.006) (0.093) MKTPWR 0.259 PCT_ACRES_PAST30 0.063 (0.099) (0.063) NonCash_TotAss -0.845 VALGR_INC_CROP30 -0.016 (0.124) (0.058) Loans_OtherNonCash -0.229 UNEMP30 -1.204 (0.058) (0.315) LIQLOANS 0.115 SMFARM30 -0.075 (0.054) (0.052) Losses_Exp 0.027 (DAGLBE) X (PCT_CROPINC30) 0.139 (0.049) (0.036) REO_NonCash -3.415 PCT_STBANK -0.288 (0.331) (0.047) (BONDYLD) X (SEC) -0.247 STBUILDPERM_Lag5M 0.054 (0.239) (0.010) NW_TA 1.700 STBUSFAIL_Lag3Q -0.005 (0.184) (0.004) (DD + DTB)_TD -0.164 NATDAGP_Lag5M -0.086 (0.059) (0.264) DTB_TD -0.478 NATDBUSFAIL_Lag5M -0.057 (0.203) (0.054) DFB_CashAss 0.059 TIME 0.044 (0.060) (0.001) Number of observations(bank-months) 269,683 Log-likelihood -11,704

Sourcesand definitions:Definitions of variablesare providedin Table 1 and sourcesare describedin the Data Appendix.Indicator variablesfor individualmonths appear as DUM, followed by the monthand year of the indicatorvariable. Lags are indicatedby appending-Lag,followed by an indicationof the lag length(3M = threemonths, 3Q = threequarters). Time is a monthlytime trend.

Evidence of a significantnegative coefficient on of two possibilities: (1) an increasedprobability this variablemay suggest that liquidity risk was of failure that is unrelated to long-run funda- a significantcontributor to failure risk through- mentals (i.e., an unwarrantedfailure related to out our period. Our test of interaction effects temporary illiquidity or contagion), or (2) an examines whether alleged panic episodes were incomplete model of fundamentals,where the times of unusual sensitivity to liquidity risk. elements missing in the model mattermore for The use of panic indicatorvariables, interac- the failures of banks in some times and places tion effects, or nearby failures to test for conta- than for others. For example, finding a negative gion in producing unwarrantedbank failures is residual in our survival model for a particular a "one-sided"test, by which we mean that it is month may mean that a panic in that month capable of rejecting, but not proving, the pres- caused failures, or it may mean that our model ence of a contagion effect. A statistically sig- lacks a fundamentalthat was importantduring nificant negative coefficient for any of the four that month. Finding no significant negative re- types of panic/contagionindicators implies one sidual or special liquidity interaction effects VOL. 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESS DURING THE DEPRESSION 1629

TABLE5b-MODIFIED SURVIVAL REGRESSIONS FOR INDIVIDUAL FED MEMBER BANKS, PANIC VARIABLE RESULTS (StandardErrors in Parentheses)

(2) (3) (4) (5) (6) (7) (8) (9) (BONDYLD) X (SEC) -1.334 -0.168 -1.072 -0.720 -0.115 -0.338 -1.220 -2.139 (0.280) (0.244) (0.265) (0.202) (0.234) (0.218) (0.290) (0.979) NATDAGP_Lag5M 0.930 -0.181 0.806 0.794 -0.058 0.924 0.696 1.005 (0.295) (0.270) (0.282) (0.216) (0.259) (0.234) (0.304) (0.999) FSPANIC-30 0.073 0.122 0.140 0.101 (0.035) (0.035) (0.047) (0.051) FSPANIC-31a 0.046 0.050 0.106 0.135 (0.037) (0.036) (0.048) (0.052) FSPANIC-31b -0.086 -0.043 0.066 0.053 (0.029) (0.028) (0.040) (0.043) DUM_JAN33 -0.619 -0.570 -0.510 -0.478 -0.568 -0.369 (0.063) (0.060) (0.045) (0.049) (0.067) (0.268) DUM_FEB33 -0.452 -0.412 -0.415 -0.401 -0.411 -0.588 (0.070) (0.066) (0.051) (0.055) (0.074) (0.226) DUM_MAR33 -0.060 -0.042 -0.173 -0.135 0.112 0.511 (0.088) (0.084) (0.064) (0.070) (0.093) (0.497) (FSPANIC-30) X (DFB_CashAss) -0.028 0.140 (0.151) (0.163) (FSPANIC-31a) X (DFB_CashAss) -0.049 -0.087 (0.140) (0.153) (FSPANIC-31b) X (DFB_CashAss) -0.277 -0.245 (0.113) (0.123) WICKER-30 -0.464 -0.439 -0.419 -0.150 -0.327 -0.625 (0.085) (0.078) (0.117) (0.121) (0.082) (0.326) WICKER-31a 0.055 0.047 0.215 0.193 (0.084) (0.074) (0.123) (0.133) WICKER-31b -0.307 -0.190 -0.136 -0.093 -0.230 -0.034 (0.073) (0.065) (0.121) (0.132) (0.070) (0.255) (WICKER-30) X (DFB_CashAss) 0.298 -0.429 (0.301) (0.326) (WICKER-31a) X (DFB_CashAss) -0.677 -0.514 (0.451) (0.487) (WICKER-31b) X (DFB_CashAss) -0.126 -0.236 (0.433) (0.474) Chicago-6-32 -1.378 -1.078 -1.259 -0.430 (0.727) (0.504) (0.601) (0.353) NEARFAILS -0.004 -0.006 -0.009 (0.001) (0.001) (0.002) Number of observations (bank-months) 269,683 269,683 269,683 269,683 269,683 269,683 269,683 53,032 Log-likelihood -11,644 -11,681 -11,628 -11,643 -11,679 -11,569 -11,568 -2,076

Sources and definitions:All models utilize the control variable specificationin Table Sa. The results for the control variables do not qualitativelydiffer from those reportedin Table 5a in the presence of the panic variables.Definitions of variables are provided in Table 1 and sources are described in the Data Appendix. Indicatorvariables for individual months appear as DUM, followed by the month and year of the indicator variable. Lags are indicated by appending-Lag, followed by an indication of the lag length (3M = three months, 3Q = three quarters). during a Friedman-Schwartzpanic window, coefficients indicates no residual failures asso- however, provides evidence against the view ciated with particularregions, or occurring in that contagion or illiquidity producedbank fail- the neighborhoodof other failed banks, but the ures in that month that cannot be explained by significance of these effects may simply indi- fundamentals. cate the absence of regressors that capture im- Similarly, regional indicatorsand interaction portant local or regional fundamentals. The effects, and the NEARFAILS variable, provide potentialfor making false inferences from these one-sided tests of local or regional contagion; indicatorswarrants emphasis, especially in light the absence of statistically significant negative of the fact that all of these indicators were 1630 THEAMERICAN ECONOMIC REVIEW DECEMBER2003 constructed based on ex post observations of the stabilizing effects of branching in U.S. bank failures. If our fundamentalmodel is in- banking history (Calomiris, 2000). Neverthe- complete (as it surely is), then indicator vari- less, as contemporariesduring the Depression ables and interactioneffects for specific dates and Calomiris and David C. Wheelock (1995) constructed from ex post observations of fail- note, some of the largest branchingnetworks in ures could prove significanteven in the absence the United States collapsed during the 1930's, of true contagion or illiquidity crises. indicatingthat the 1930's may have been some- It is also important to note that indicator thing of an exception from the standpoint of variablesare uninformativeabout the particular the stability of branching banks. Many large mechanism through which illiquidity or conta- branching banks were active acquirers during gion produces bank failure. Significant unex- the 1920's, taking advantage of windows of plained residualsfor particulartimes and places opportunity provided by the distress of unit may indicate failures caused by an external banks. Many of those acquirers,therefore, were drain (as in a flight from the dollar) that pro- in a vulnerable position (i.e., they had just ac- duces exogenous withdrawalpressure on banks. quired a relatively weak portfolio of assets) at Some historianshave argued that such a mech- the beginning of the 1930's. Furthermore,Mark anism may have been importantin the fall of Carlson (2001) argues that branching made 1931 and in early 1933. Alternatively, unex- banks more vulnerableto the aggregate shocks plained residualeffects may indicate "panic"in of the Great Depression. Branching provides reaction to a "contagion of fear" about bank diversification of sectoral risks, and thus per- solvency (that is, a massive loss of confidence mits branchingbanks to take on more loan risk. in the domestic banking system). While we will But branchingdoes not offer as much protection sometimes refer to the indicator variables as duringan economywide Depression (like that of "panic" or "contagion"indicators, for conve- the 1930's) that affects all sectors. Because nience, it is importantto bear in mind that- branchingbanks believed that they were more particularly in the case of the nationwide protected against loan loss than other banks, indicatorvariables for the fall of 1931 and early they took on more loan risk and were subject to 1933-our measures of possible panic/conta- a greatershock when Depression-eraloan losses gion/illiquidity do not distinguish possible ef- occurred. fects of a loss of confidence in domestic banks State-charteredbanks operate under different from a crisis producedby a run on the currency. regulations, and in general were given greater latitudein lending. Thus, it may be that national A. Indicators of Bank Failure Risk banks were constrainedto have lower asset risk than state banks. Before reviewing the results in Table 5, we Measures of the proportionsof different cat- first explain the logic underlyingthe fundamen- egories of assets (NonCash_TotAss, Loans_ tal predictorsof survival(see also Calomirisand OtherNonCash,LIQLOANS, and DFB_CashAss) Mason, 1997). According to basic finance the- capturethe degree of ex ante asset risk, and the ory, the probabilityof insolvency should be an liquidity of assets. Loan losses (Losses_Exp) increasing function of two basic bank charac- and real estate owned (REO_NonCashAss)are teristics: asset risk and leverage. Liquidity of ex post measures of asset quality. assets relative to liabilities may be an additional Bank net worth relative to assets (NW_TA) factor influencing the risk of failure. measuresthe extent of leverage using book val- Our measuresof fundamentalbank attributes ues. Book values are imperfect measures of net capture variation in bank asset risk, leverage, worth, but market values are not available for and liquidity. Banks that are larger (higher most of the banks in our sample. The structure LTotAss) are better able to diversify their loan of bank liabilities (captured here by various portfolios, reducing their asset risk. Thus, ce- ratios of componentsof deposits relative to total teris paribus, large banks should have lower deposits) also provides informationabout bank failure risk (higher survival hazard).Banks that failure. Calomiris and Mason (1997), among achieve their size througha branchingnetwork others, have found that weak banks were forced (LNBRANCH) should also be more diversified, to expand their reliance on high-cost categories ceteris paribus.There is substantialevidence for of debt (thatis, debt held by relatively informed VOL 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESS DURING THE DEPRESSION 1631 parties), and that the ratio of bills payable to B. Regression Resultsfor the Bank Survival total deposits (BPR_TD) is a useful indicatorof Model fundamental weakness. It may also be that a reliance on demandabledebt (DD + DTB) in- The results for the basic model in Table creased bank liquidity risk, and therebycontrib- 5a show that many fundamentalshave explan- uted to failure. The average interestrate paid on atory power for bank survival (failure). Gener- deposits (INTCOST)is a directmeasure of bank ally, coefficients are of the predicted sign and default risk, but a lagging measure (dependent highly significant. Bank size (LTotAss) is pos- on the frequency of deposit rollover). itively associated with survival. Higher net The bank marketpower variable is included worth is also associated with longer survival. A to capturethe potentialrole of "rents"related to reliance on demandable debt rather than time a bank's marketpower for boosting the market deposits, where the demandabledebt ratio is the value of bank net worth, and therefore,reducing sum of demanddeposits held by the public and the effective leverage ratio of the bank. Carlos interbank deposits relative to total deposits D. Ramirez (2000) found this variablewas use- [(DD + DTB)_TD], lowers survival probabil- ful in predicting failures of banks in Virginia ity. But interbankdeposits have a much larger and West Virginia in the late 1920's. effect than demand deposits of the public. The We also include a measureof the exposure of interbankdeposits effect is given by the sum of the bank's securities portfolio to changes in the coefficients on (DD + DTB)_TD and on bond yields (BONDYLD X SEC), to capture DTB_TD (that is, the sum of -0.164 and what we call the "Temin effect." Temin (1976, -0.478). The effect of interbankdeposits may p. 84) writes that: "The principalreason usually reflect either liquidity risk or the fact that weak given for [post-1930] bank failures is the de- banks were forced to rely more on interbank cline in the capital value of bank portfolios credit, and our results are not able to distinguish coming from the decline in the marketvalue of between these two interpretations.Consistent securities."Wicker (1996, p. 100) disputes that with the latter interpretation,nondemandable view, and argues instead that bank loan quality debt from informed creditors (bills payable or was the dominantsource of fundamentalshock rediscounts),however, has the largest effect on that led to bank failures. Our model includes survival probabilityof any debt category. Bills measures of loan quantity and quality, but we payable or rediscountsfrom official sources en- also include BONDYLD X SEC to capture ters with a coefficient of -1.490, while such bank vulnerabilityto changes in bond yields. debt from privatesources has a somewhatlarger Some county-level characteristics take ac- effect (the sum of the two coefficients, -1.490 count of the shares of various elements of the and -0.126). agriculturalsector in the county economy, and State-charteredbanks (STBANK) were less the extent to which agricultural investment likely to fail, ceteris paribus, than national grew during the 1920's. That emphasis reflects banks. This is a somewhat surprisingresult for the view of White (1984), Wicker (1996), and which we lack a clear interpretation.Neverthe- others that much of the distress suffered by less, we are able to say that constraintson the banks during the 1930's was a continuationof lending of national banks likely were not very the distress suffered in agriculturalareas during important for limiting their relative risk. Our the 1920's. Other county-level, state-level, and interpretationof the state-charteredindicator national-level variables (including unemployment, building permits, business failures, and agricultural prices) capture general economic bank distress, or because of fundamentallinks from eco- conditions in the county, state, and country.4 nomic activity to banking distress). That problem is mitigated, but not eliminated, by our use of lagged values of high-frequencyfundamentals. Calomiris and Mason (2003a, address 4 2003b) the question of the dynamic relationship One potential concern is reverse causation-that is, the among bank failures,business failures, and buildingpermits possibility that business failures or building permits are at the state level. We find little effect of autonomousshocks endogenous to shocks originatingin the banking sector. For to bank failures on other variables, and little serial correla- example, it is possible thatpanics producedeclines in build- tion in the bank failure process. Thus, those results support ing and increases in business failures, which in turn predict the assumed exogeneity of fundamental determinants of future bank distress (either because of serial correlationin bank failure. 1632 THEAMERICAN ECONOMIC REVIEW DECEMBER2003 variable is not complicated by possible selec- Note, however, thatthe size of the coefficient on tivity bias relatedto a bank's choice of location (BONDYLD) X (SEC) is larger and often sig- (i.e., that state banks were more present in cer- nificant in other regressions in Table 5, specif- tain counties) because we include a separate ically in regressions that include indicator variable (PCT_STBANK) to capture the pro- variablesfor the first three months of 1933 [that pensity of banks in a given county to be state- is, regressionsother than (1), (3), and (6)]. This chartered, and therefore, we control for result has an intuitive interpretation;when one location-specific selectivity bias. That control controls for the most importantepisode of na- variable has a negative sign, indicating that tionwide panic or illiquidity crisis (during counties with a greater proportion of state- which a flight to quality would have raised the chartered banks suffered higher bank failure price of government securities, but not other rates, ceteris paribus. securities held by banks), the Temin effect on Branching (LNBRANCH) is negatively re- the average securities portfolio should be lated to survival duration,after controlling for stronger. other effects (including size). This result may Thus our results on the effects of bank port- reflect the unusual vulnerability of branching folio compositionon failurerisk indicatethat, in banks in the early 1930's. In future work, we a sense, both Temin and Wicker were correct: plan to investigate the extent to which prior banks with more lending, and riskier lending, acquisitions of distressed banks by branching were more vulnerablethan other banks, ceteris banks may explain this result. paribus, but to the extent banks had securities Consistent with Ramirez's (2000) findings portfolios, rising bond yields increased their for Virginia and West Virginia in the late failure risk. 1920's, greatermarket power (MKTPWR)low- Some county characteristicsare highly sig- ers failure risk. nificant. Higher unemployment (UNEMP30) Consistentwith Wicker's (1996) emphasis on lowered bank survival rates. More agriculture, loan quality as a source of fundamentalprob- per se, does not appearto have been a problem. lems, more lending and lower bank asset quality In fact, a reliance on agricultureas a source of (measuredeither ex ante by NonCash_TotAss, income was associated with increased survival Loans_OtherNonCash,and LIQLOANS or ex rates. But the relative health of the agricultural post by REO_NonCashAss)is associated with sector made a difference for bank survival. In lower survival. We found no differences in fail- counties where agriculture was an important ure risk associated with the composition of cash and healthy sector, as indicated by the inter- assets (which we define as the sum of cash, action of the percent of income earned from reserves at the Fed, government securities, and crops and the investment in agriculturalcapital deposits due from banks). We reportthe results during the 1920's [(DAGLBE) X (PCT_ for the ratio of due from banks relative to total CROPINC30)],bank survivalrates were higher. cash assets (DFB_CashAss), where the coeffi- The extent that a county's agriculturalincome cient measures the effect of increasing the rel- was based in grains (VALGR_INC_CROP30), ative share of due from banks in total cash as opposed to pasture(PCT_ACRES_PAST30), assets. It has an insignificantpositive effect on did not enter significantly.A greaterpresence of survival duration. small farms in a county (SMFARM30) had a Higher debt interest cost is associated with negative, but not a highly significant or robust, lower survival rates, but this is not a significant effect on bank survival. or robust result. The insignificance of higher At the state level, the effect of lagged debt interest cost reflects the correlation be- monthly building permits (STBUILDPERM_ tween interest cost and other regressors that lag5) on bank survival proves positive and capture asset risk, leverage, and debt composi- highly significant,while lagged quarterlyliabil- tion. In the absence of those other variables,it is ities of business failures does not prove signif- a significant predictorof failure risk. icant. At the national level, monthly liabilities Banks with relatively high securities portfo- of business failures has a negative sign but is lios suffered greater risk of failure when bond not highly significant. Monthly agricultural yields rose, as predicted by Temin (1976), but price change is insignificantin the basic model, the effect is not significant in the basic model. but becomes significant when panic indicator VOL 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESS DURING THE DEPRESSION 1633 variables are added [in regressions (2), (4), (5), failed after March in the definition of March (7), and (8)]. failures. Regression (2) in Table 5b shows the result of The results in regressions (3) and (4) support including indicator variables for Friedman- (but do not prove) Wicker's view that sudden Schwartz national banking crises alongside our waves of bank failure unrelated to observable other variables. Owing to the complexity of the fundamentals (prior to 1933) were largely re- suspension and failure process in early 1933, gional affairs. Two of Wicker's regional indi- the 1933 crisis indicatorsare divided into three cators prove negative and significant (for late separate monthly indicator variables for Janu- 1930 and for September-October 1931). An ary, February, and March. Regression (3) of indicator for the third regional panic identified Table 5b includes indicator variables for the by Wicker (that is, mid-1931) enters with the three regional panics identified by Wicker. Re- wrong sign and is not significant. In regression gression (4) of Table 5b includes both the (4), in the presence of the Wicker regional in- Friedman-Schwartzand the Wicker crisis indi- dicator for the fall of 1931, the Friedman and cator variables. Regression (5) adds interactive Schwartznational indicatorfor that episode de- effects related to due from banks to the speci- clines in magnitude and becomes statistically fication of regression (4). Regression (6) adds insignificant. the June 1932 Chicago panic indicatoralone to Conclusions based on the magnitudeand sig- our basic model. Regression (7) adds the nificance of indicatorvariables for panics could Chicago panic indicator and the NEARFAILS conceivably provide a misleading picture of the variable to the regression (5) specification. Re- effects of panic episodes on the bank failure gression (8) drops many of the insignificant process. For example, even if panics are epi- regressorsfrom regression(7). Regression (9) is sodes in which liquidity mattersa great deal for the same as regression (8), but includes only the incidence of bank failure, and in which banks in the principal 215 cities in the United indicatorsof fundamentalsolvency do not mat- States, which permits a comparison of the deter as much as during nonpanic episodes, panic terminantsof failure in rural and urban areas. indicator variables might not prove negative In regression (2), and in all other specifica- and significant in a regression that assumes re- tions, we find that the indicator variables for gression coefficients are constant. two of the four Friedman-Schwartz panics Thus, it is conceivable that our conclusions (those of late 1930 and mid-1931) are positive about the first three Friedman-Schwartzepi- and, in one case, significant.That is, contraryto sodes could change if we took account of Friedman and Schwartz, those episodes were changes in regression coefficients during those times of unusually high bank survival, after episodes. To investigate that possibility, we re- controlling for fundamentals, not episodes of laxed the assumptionthat the coefficients on our inexplicably high bank failure. We do not view fundamentalswere constant, and allowed them this result as indicative of "irrationalexhuber- to vary over time. Specifically, we allowed co- ance" on the part of depositors during those efficients to change during the first three epi- episodes, but ratherof an incomplete model of sodes identified by Friedman and Schwartz as fundamentals. The indicator variable for the panics. Our results did not supportthe view that September-November 1931 period is signifi- indicators of liquidity mattered more during cant and negative in regression (2), as are the panics, or that indicatorsof fundamentalinsol- indicator variables for January and February vency matteredless during panics. Indeed, our 1933. The indicatorfor March 1933 is insignif- failure risk model was remarkablystable. In the icant. If we had assigned all bank failures for first Friedman-Schwartzepisode (late 1930), April-December 1933 to March 1933 (which three variables out of 28 showed somewhat we argue, on balance, against doing above) the significant changes in coefficients during the only qualitative difference in our results is the episode: the state bank indicator(a 0.0159 sig- indicator variable for March 1933, which be- nificance level), the ratio of privatebills payable comes much larger in absolute value and sig- and rediscountsto total bills payable and redis- nificant.Thus, unsurprisingly,one cannot reject counts (a 0.0479 significancelevel), and interest the possibility that March bank failures resulted cost (a 0.0263 significance level). In the second from contagion if one includes many banks that Friedman-Schwartz episode, no coefficient 1634 THE AMERICANECONOMIC REVIEW DECEMBER2003 changes were significant. In the third episode, risk of Chicago banks, and only in one month of due from banks as a fraction of cash assets was the sample. significant (at the 0.0033 significance level). With respect to local contagion, the NEAR- Three facts are salient. First, randomly one FAILS variable is significant in all survival should expect that three out of 84 regressors regressions that include it, even when the would be significantat the 0.0357 level (that is, Friedman-Schwartzand Wicker indicator vari- 1/28), and we found that only three variables ables are also included. The June 1932 indicator were significant at that level. Second, different for Chicago is significanteven when we include interactionvariables were significantacross ep- the NEARFAILS variable in the regression. isodes. Third, only one of the coefficients is Since Calomiris and Mason (1997) provide ev- possibly interpretableas a special panic liquid- idence against viewing bank failures in Chicago ity effect-the negative effect for the interaction in June 1932 as the result of contagion, we view of the third episode with the due from banks that finding as illustrative of the danger of in- variable, (FSPANIC-31b) X (DFB_CashAss). terpretingindicator variables as proof of con- In other words, in the fall of 1931, banks that tagion (rather than as evidence of missing relied on deposits in other banks as a source of fundamentals). cash assets found that those assets were not In regression (9), we investigate differences perfect substitutes for other cash assets (cash, between city banks and rural banks. All vari- reserves at the Fed, and governmentsecurities). ables are defined similarly to those of the pre- To furtherinvestigate the role of the due from vious regressions with one exception; building banks variable during alleged panics, in regres- permits are defined at the level of the city in sions (5) and (7) of Table 5b we include inter- which the bank is located, rather than aggre- action effects that allow the coefficient on gated to the state level. Results are similar to DFB_CashAss to vary during all the episodes those of regression(8), apartfrom differencesin identified by Friedman and Schwartz or by significance that may be attributableto the rel- Wickeras panics. As Table 5b shows, this effect atively small sample size of city banks (which is only significantfor (FSPANIC-3lb) X (DFB_ comprise roughly one-fifth of our nationwide CashAss). Including that variable, however, sample). The main differences between regres- changes the sign of the FSPANIC-3lb indicator sions (8) and (9) are as follows. Deposits held variable to positive, and reduces the overall by banks (DTB_TD) has a smaller and insignif- combined effect of the two variables (when icant sign in the city bank failure regression. evaluatedat the mean of DFB_CashAss). Over- This result suggests that deposits held by rural all, we concluded that our survival model is banks in city banks were not a source of special quite stable over time and that there is little illiquidity risk during the Great Depression. gained from allowing coefficients to change Other differences include the insignificance of during episodes of alleged panic.5 county unemployment for city banks, the In results not reportedhere, we also experi- smaller and less significant coefficients for the mented with disaggregation of the DFB_ Wicker-3lb indicator and the indicator for the CashAss variable (which our data allow us to Chicago panic of June 1932. Perhaps not sur- divide among accounts held in Chicago, New prisingly, the latter effect indicates that in com- York, or other cities). We found that accounts parison to other cities, and using a model held in Chicago entered negatively relative to derived solely from the experience of city those of other cities, but this result disappearsin banks, there is less of an unexplained residual the presence of the indicator variable for the for Chicago bank failures in June 1932. June 1932 Chicago banking panic. In other In summary, our results provide evidence words, the illiquidity of money held in Chicago against the Friedman-Schwartzview that the banks was mainly relevant only for the failure bank failures of late 1930 and 1931 were autonomous shocks producedby nationwide contagion or panic. Our results are consistent with that 5 (but do not prove) Wicker's (1996) view We also experimentedwith varying regression coeffi- ratherthan national charac- 1933. One of 28 regional contagion cients duringJanuary and February regres- and late and sors-total bills payable and rediscounts (BPR_TD)-was terized the crises of late 1930 1931, somewhat significant (with a 0.051 significance level). they are consistent with (but do not prove) VOL. 93 NO. CALOMIRISAND MASON:BANK DISTRESSDURING THE DEPRESSION 1635

1,400

1,200-

1,000

800 *1 ? ?u 600

400- _6,699 banks /

200- WkhFriediniBA and Schwartz PanicEffects

0- I.aI . I a.. II a I I I I I I I ? ? ? ? ? ? ? m

X0 0 0 00 =- 0 e?- Ze " a" ?

FIGURE4. PREDICTEDSURVIVAL DURING FRIEDMAN AND SCHWARTZ"PANIC" EPISODES

Friedman and Schwartz's and Wicker's view relevant fundamentalsare likely omitted from that bank failures in January-February1933 our model. resulted in part from nationwide panic. Our Nevertheless, our estimates of local, regional, finding that local bank failures raise the proba- and national panic effects can be used to place bility that another local bank will fail is also an upper bound on the importanceof estimated consistent either with omitted local fundamen- panic effects for the United States as a whole. tals or local bank contagion. Figures 4-7 plot the mean estimated survival duration(in days) for banks in our sample over C. Placing an Upper Bound on the time, using different estimation equations, with Importanceof "Panic" Effects and without taking account of panic indicator variables. All the figures display a rising trend, Are the effects of panics (national, regional, which reflects the rising average conditional or local) potentially large economically as well survival probability of banks over our time as significant statistically for the United States period. as a whole? Thus far, we have arguedthat some Figure 4 corresponds to regression (2) of indicators of local, regional, or national panic Table 5b, in which only the basic model cum cannot be rejected. That is, some bank failures Friedman-Schwartzindicator variables is used are not fully explained by a stable model of the for the estimation.The dashedline in Figure4 is bank failure process for the period 1930-1933. the mean survival estimate using all the basic Late 1930 and late 1931 may have been epi- regression coefficients, and the Friedman- sodes of regional panic. Januaryand February Schwartzindicators for the late-1931 period and 1933 may have been a time of nationwidepanic. the early-1933 period (the indicators for late Local contagion effects may have been present 1930 and mid-1931 are set at zero). The solid throughoutthe sample period. As we have re- line in Figure 4 shows the average estimated peatedly noted, all of our tests of panic or con- survival durationif all the Friedman-Schwartz tagion effects are "one-sided" and likely indicator variables are set at zero. Figure overestimate the true incidence of panic, since 4 shows that even if one wanted to argue that 1636 THEAMERICAN ECONOMIC REVIEW DECEMBER2003

1,400-

1,200-

1,000-

800-

600-

400- 1,714banks 394 banks 200 (v6is ind'IistingbWh We)

WihWicker Pan Elcts u 1 I Ia I I 9 I a? ? I I 9 1 1 Ia I I ? I I 5 1 1 I r oo o 0 0 O 0 - - - _ >

FIGURE SURVIVALDURING WICKER "PANC" EPISODES 5. zPREDICTED

FIGURE5. PREDICTEDSURVIVAL DURING WICKER "PANIC" EPISODES

the late-1931 episode was a nationwide panic case of the late 1930 regional episode, 1,714 (contraryto the evidence presentedabove), it is banks are in affected regions. still not an importantepisode of unexplained Figure 6 corresponds to regression (6) of bank failure. In contrast, including indicators Table 5b. The dashed line of the figure includes for January and February 1933 substantially the effects of all coefficients, while the solid reduces the predictedaverage survival duration line excludes the June 1932 Chicago indicator in those months. We conclude that our upper variable.The estimated effect on survival dura- bound estimate of the effects of national panic tion, while largefor affected banks, is trivial for indicate potentially importanteffects from pan- the countryas a whole because only 29 Chicago ics only in early 1933. banks are affected. Figure 5 corresponds to regression (3) of The NEARFAILSvariable can be used to put Table 5b. As in Figure 4, the dashed line is the an upper bound on the potential importanceof prediction of the basic model plus the first and location-specificcontagion effects. In a survival third Wicker indicatorvariables (as before, the regression that includes the basic model and indicatorvariable with the "wrongsign" is set at NEARFAILS, the omission of NEARFAILS zero). The solid line shows the effect of omit- from the estimation of survival durationraises ting the Wicker indicatorsfor the two episodes. the average estimated survival duration of Note that these effects, while substantialfor the banks for each month in our sample period by banks in the affected regions, are confined to an average of 0.2 percent. Figure 7 shows the those regions, and thus are not large for the effect of omitting this effect from the survival nation as a whole. In the case of the 1930 model. regional episode, only 394 banks are in the We conclude that prior to January 1933, the regions affected by the indicatorvariable; in the effects of panics-whether national, regional, VOL. 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESSDURING THE DEPRESSION 1637

1,400-

1,200-

1,000

800'

(4) on -. 600.

WithChicago 400- PanicEffect, 29 banks (visuallyindistinguishable) 200-

0 ------5 9 9 9 0 0 0 v 9 9 I V V W 0 9 5 v n n 0 5 9 W . M M e * m *e f ^ C) f Z I I I I I e2r c Cl c l X i o 0 0 -I -I - I I I _O o 3e i ~aIi Xa z

FIGURE6. PREDICTEDSURVIVAL DURING CHICAGO "PANIC" EPISODE

or local-contributed little to the failure risk of bank failures unrelated to measures of funda- Fed member banks in the nation as a whole. A mentals may have had little to do initially with stable model of bank fundamentalscan account a "contagionof fear" about banks and more to for the bulk of regional and temporal variation do with expectations of Roosevelt's departure in bank failure risk during the period 1930- from gold, which in the event, were accurate.In 1933. other words, one could argue that the missing Our conclusion that panic effects were not "fundamental"in the failure risk model was the potentially important until January 1933 has probabilityof the government's departurefrom three importantimplications for the literatureon gold. That interpretationof the events of 1933 bank failures duringthe GreatDepression. First, would suggest even less room for "contagions it implies a limited role for nonfundamental of fear" about bank condition as a contributing causes of bank failures during the Depression. influence to bank failure duringthe Depression. January1933 was quite late in the history of the Depression (which reached its trough in March m. Conclusion 1933). Second, it implies that bank failures during the crucial period of 1930-1932, which saw We are able to identify close links between substantialdeclines in bank assets and deposits, fundamentalsand the likelihood of individual were not an autonomous shock, but rather an bank failure from 1930 through 1933. Funda- endogenous reflection of bank condition and mentals include both the attributesof individual economic circumstances.Third, the special cir- banks, and the exogenous local, regional, and cumstancesof early 1933-the origins of which national economic shocks that affected their Wigmore (1987) traces to a run on the dollar health. We also develop a set of tests for the ratherthan a loss of confidence in the solvency presence of liquidity crises or contagion. In of the banking system-suggest that the only addition to regressors that capturefundamental nationwideepisode that saw the sudden burstin determinants of bank distress, we include 1638 THE AMERICANECONOMIC REVIEW DECEMBER2003

1,400 -

1,200 -

1,000 -

;> ._ co 800 CT

C)U 600 - 4._ I. AO WithNEARFAILS 400 -

200 -

. I . I. I. . l I I 0 II I I I I I I i I I I I I I I I I I I i I I I I Is cI I I~ 00 0 .- I 5- I I enen I en "-A r^ I I I I.. -o f^cI sa rco ..-: en IIIe * in. ;> en ?-?-5 CO en e I?I s ^3 OT 0 cis 0 ?| () 0 CA2 z z 2 z

FIGURE 7. PREDICTED SURVIVAL WITH OTHER DEPOSITS IN FAILED BANKS IN THE COUNTY

indicatorvariables to captureresidual effects of presence of panics our results have two possible alleged national and regional panic episodes interpretations:either that illiquidity crises oc- identified by Friedman and Schwartz (1963) cuffrredand resulted in some unwarrantedbank and Wicker (1996). This approachis capable of failures (as Friedmanand Schwartzand Wicker rejecting, but not convincingly confirming,the have argued) or that our model of the funda- incidence of panics. We find no evidence that mental causes of bank failures is incomplete. bank failures were induced by a nationalbank- The results of our earlier study of the Chicago ing panic in the first three episodes Friedman bankingpanic of June 1932 (Calomirisand Ma- and Schwartzidentify as panics (late 1930, mid- son, 1997) indicated that banks in Chicago in 1931, and late 1931). We do find, however, that June 1932 did not fail due to an illiquidity crisis in Januaryand February1933 there is a signif- or panic, but ratheras the result of fundamental icant increase in bank failure hazardthat is not shocks that were local and sudden. That exam- explained by our model of fundamentals. ple leads us to believe that it is quite possible Indicatorvariables for the states and periods that the significance of the January-February identifiedby Wicker as sufferingregional crises 1933 panic indicators, the NEARFAILS vari- in 1930 and 1931 indicate significant region- able, and two of the Wicker panic indicators specific increases in the probability of bank result at least in part from a failure to fully failure that are not explained by our model of model changes in relevanteconomic conditions. fundamentalsin two of the three cases identified In future work, we intend to take a closer look by Wicker (late 1930 and late 1931). At the at bank failures during the regional crises of local level, we find that the failure of nearby 1930 and 1931, and at the January-February banks (NEARFAILS) is associated with an in- 1933 nationwide experience, to investigate that crease in the probabilityof bank failure. possibility. When we find evidence consistent with the Our results indicate a much smaller role for VOL. 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESSDURING THE DEPRESSION 1639 contagion and liquidity crises in explaining the policy, given the importance of fundamental bank failures of 1930-1932 (and the contrac- sources of bank failure, it is doubtful whether tion of the money stock that accompaniedthem) the FederalReserve System, state governments, than that envisioned by Friedmanand Schwartz or the national government could have done (1963). The regional panics identified by much in the way of traditionalmicroeconomic Wicker may have been importantlocal or re- liquidity assistance (either through collateral- gional events, but they had small and temporary ized lending or temporary suspension of effects on average bank failure hazard in the convertibility) to rescue failing banks during aggregate. And the effect of location-specific 1930-1932. Only a combinationof expansion- contagion seems to have been similarly small in ary monetarypolicy and bank bailouts (e.g., in its importance. To the extent that nationwide the form of subsidized bank recapitalization,as panic may explain bank failures, its role is rel- was effected by the Reconstruction Finance atively late in the sequence of events of the Corporationafter the banking crisis of 1933) Depression and brief (lasting a matterof weeks could have prevented banks from failing in in early 1933), and involved a small proportion 1930-1932 (see Mason, 2001a). of the bank failures during the period 1930- Policy options in 1933 are more ambiguous. 1933. Our findings suggest that disaggregated The fact thatthe only nationwideepisode during analysis of bank failures and deposit shrinkage which panic indicators are significant (early leads to a much smaller role for contagion in 1933) coincided with substantialexternal drain understandingbank distress during the Great (the alleged run on the dollar in early 1933) Depression. suggests that liquidity assistance targetedto in- Nevertheless, three caveats warrantmention. dividual banks might have been helpful in Jan- First, our sample consists of Fed memberbanks. uary 1933, particularlyif it had been combined It may be that nonmemberbanks had a different with expansionarymonetary policy and an im- susceptibility to panics. Second, we use failure mediate departurefrom the gold standard.6 ratherthan suspension as our measure of bank distress. We are unaware of any records of DATAAPPENDIX individualbank suspension dates, and also consider failure a more meaningful indicator of Our data set contains a wide variety of vari- distress (following the reasoningof Ali Anari et ables that differ by frequency, geographic al., 2002), but we recognize that suspensions scope, and level of disaggregation.In this Ap- that did not result in failure may also have been pendix we describe briefly the definitions and disruptive, and that we are not able to capture sources for our data, and explain the limits of those effects in our analysis. Third, there are our sample. other potentialmechanisms for the transmission The data used in our analysis can be broken of contagion that we have not been able to down into six major types: individual bank investigate. For example, ownership linkages characteristicdata; county-level bank character- among chain banks may have been importantin istic data; county economic data; state eco- promotingruns on the solvent affiliates of failed nomic data; national economic data; and other banks. It remains to be seen how ownership data constructedto measure systemic bank dis- linkages, or perhaps correspondent linkages, tress. Following are the sources and details un- may have matteredfor the transmissionof fail- derlying each type of data. ure among banks during the Depression. With respect to the policy implicationsof our I. IndividualBank Characteristics findings, it is importantto distinguish macroeconomic and microeconomic policies. There With the help and supportof the St. Louis Fed can be little doubt that expansionaryopen mar- andthe Universityof Illinois,Urbana-Champaign, ket operations in 1930 and 1931 (or departure from the gold standard in 1931, when many 6 other countriesdid so) could have avoided mac- If Wigmore is right about the run on the dollar in early roeconomic in and 1933, one could argue that lender of last resort assistance collapse 1931-1933, there- and would have expansionarymonetary policy in early 1933 would have fore, substantially mitigated bank been ineffectual in stemming the outflow from the banks distress. Without a different macroeconomic without an immediate departurefrom gold. 1640 THEAMERICAN ECONOMIC REVIEW DECEMBER2003 and subsequently, with funding from the Na- fields. These fields include not only typical con- tional Science Foundation,we assembled bank- solidated balance sheet items, but also earnings level balance sheet and income statementdata and expense information from bank income from microfilmrecords of "call reports"of Fed- statements.Additionally, we gatheredinforma- eral Reserve member banks (see Mason, 1998, tion from detailed schedules of asset, liability, for an overview of the call report data we col- earnings, and expense compositions of each lected). Individualbank data come from origi- bank. Call reports also contain informationon nal manuscript Reports of Condition (forms the numberof branchesoperated by each bank. Federal Reserve 105 and Treasury 2129) and Table Al lists the fields we entered. Reportsof Earnings and Dividends (forms Fed- We also gatheredbank structuredata (that is, eral Reserve 107 and Treasury 2130) filed by data on individual bank failures, mergers, ac- Federal Reserve member banks. In 1946, the quisitions, name changes, etc ...) linking banks Federal Reserve Board began microfilming to exit events including record dates of failures. original manuscript call reports for selected The structuredata for NationalBanks was hand- dates back to 1914. The Board authorizedthe entered from the unpublished Comptroller of destruction of extant manuscript Reports of the Currency structurerecords located in that Condition and Reports of Earnings and Divi- agency's archives. Structure data for State dends after microfilming. Though summary Member banks was hand-entered from Rand individual bank balance sheet data were some- McNally's Banker's Directory. Our structure times publishedin annualreports of the Federal data contain almost 70 different ways a bank and state bank supervisory authorities,the mi- can exit the data set, ranging from all imag- crofilm record available at the Board is the only inable types of mergers and acquisitions to known record of original manuscriptReports of relatively simple voluntaryliquidations and re- Conditionthat remain from the early twentieth ceiverships (which, together, we term failures). century,and this recordcontains far more detail For our present work we only utilize data on than any published source. Furthermore,the voluntaryliquidations and receiverships. microfilmrecord is the only source of individual Our data on failures for Fed member banks bank data from the Reports of Earnings and identify the date at which banks were placed Dividends. into receivership or were closed by voluntary Data in the microfilmcollection are not avail- liquidation.All results reportedbelow combine able equally for both State Member and Na- receivership and voluntary liquidation into a tional Banks. For State Member Banks, June single measure of bank failure. Results not re- and December Reports of Condition were mi- portedhere show slight differencesin resultsfor crofilmed for each year during the Great De- the two categories taken separately, and thus pression. For National Banks, however, only little advantageto analyzing them separately. selected dates were microfilmed. Those dates Many banking studies have had to rely on are December 1929, 1931, 1933, and annually bank suspension ratherthan liquidationas their thereafter. measure of bank failure. Suspensions are typi- Over an eight-year period we hand-entered cally employed because data on bank failures, and checked data for all available banks in this both for numbers and deposits of failed banks, collection between the years 1929 and 1935. are not readily available at the regional or na- We used existing National Bank charternum- tional level, especially for observations at bers and constructed State Bank charter num- greaterthan annual frequency. Data on suspen- bers to link each set of individualbank records sions can provide a misleading picture of bank across time. The final data set includes 66,316 failure. In some cases, suspensions were tem- bank records from 10,092 banks. Since the porary and suspended banks quickly reopened present paper is only concerned with events (Calomiris, 1992). Furthermore,suspension is prior to the Bank Holiday of March 1933, we not consistently defined in the literature. The restrict ourselves to bank data reportedon De- Federal Reserve series on bank "suspensions," cember 31, 1929 and December 31, 1931. published in Banking and Monetary Statistics Therefore the present analysis incorporates (1976), mixes suspensions (of state-chartered 14,410 bank records from 7,931 banks. banks) and liquidations(of national banks). Each bank record contains roughly 100 data The distinction between suspension and fail- VOL 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESSDURING THE DEPRESSION 1641

TABLEAl-REPORTS OF CONDITIONAND REPORTSOF EARNINGSAND EXPENSESFIELDS

General Information Acceptances of this Bank Purchasedor Discounted Date All Other Loans FR District Schedule of State Loans: Memoranda Loans Secured National Bank by U.S. GovernmentSecurities in Items E4 and E5 Central Reserve City Bank Total Loans for Rediscount Other Reserve City Bank Eligible County Schedule of Bills Payable and Rediscounts City Bills Payable from Federal Reserve Bank Name Bills Payable from ReconstructionFinance Corp CharterNumber (National Banks) Notes and Bills Rediscountedwith Federal Reserve Bank Assets Notes and Bills Rediscountedwith Reconstruction Finance Loans and Discounts Corp Overdrafts Due from Member Banks and Trust Companies in New York Due from U.S. GovernmentSecurities Owned Member Banks and Trust Companies in Chicago Due from Member Banks and Trust Elsewhere Other Bonds, Stocks, and Securities Owned Companies in United States Customers' Liability on Account of Acceptances Due from Nonmember Banks and Trust in Banking House and Furniture/Fixtures Companies New York Real Estate Owned Other Than Banking House Reserve with Fed Due from Nonmember Banks and Trust Companies in Chicago Cash and Due From Banks Due from NonmemberBanks and Trust Companies Outside Checks and Other Cash Items Elsewhere in United States Fund with U.S. Redemption Treasury Schedule of Deposits Acceptances of Other Banks, Bills of or Drafts Sold Exchange, State, County, and Municipal Demand Deposits Securities Borrowed State, County, and Municipal Time Deposits Other Assets Total Assets Income Interestand Discount on Loans Liabilities Interestand Dividends on Investments Demand Deposits Intereston Balances with Other Banks Time Deposits Domestic Exchange and Collection Charges U.S. Government Deposits Foreign Exchange Department Due to Banks Commissions and Earningsfrom Insuranceand Real CirculatingNotes Outstanding Estate Loans Bills Payable Trust Department Rediscounts Profits on Securities Sold Capital Service Charges on Deposit Accounts Surplus Other Earnings Net Undivided Profits Reserves for Dividends or Contingencies Expenses PreferredStock RetirementFund Salaries and Wages Reserve for Dividend Payable in Common Stock Interestand Discount on Borrowed Money Other Liabilities Intereston Bank Deposits Intereston Demand Deposits Other Information Intereston Time Deposits Liabilities of Officers and Directors as Payers Interestand Discount on Borrowed Money Branches and Branch Offices in CorporateLimits of Head Office City Taxes Branches and Branch Offices Elsewhere in United States Other Expenses Schedule of Loans Net Earnings Acceptances of Other Banks, Payable in the United States, Owned Recoveries This Bank by Recoveries on Loans and Discounts Notes, Bills, and Acceptances, Other InstrumentsEvidencing Loans Recoveries on Bonds and Securities in Countries Payable Foreign Recoveries on All Other Commercial Paper Bought in Open Market Loans to Banks and Trust Companies on Securities Losses All Other Loans to Banks and Trust Companies Losses on Loans and Discounts Loans on Securities to Brokers and Dealers in New York City Losses on Bonds and Securities Loans on Securities to Brokers and Dealers Outside New York City Losses on Banking House, Furniture,and Fixtures Loans on Securities to Others Losses on Foreign Exchange Real Estate Loans, Mortgages, Deeds of Trust, and Other Liens on Losses on Other Real Estate on Farm Land Real Estate Loans, Mortgages, Deeds of Trust, and Other Liens on Net OperatingEarnings Other Real Estate 1642 THEAMERICAN ECONOMIC REVIEW DECEMBER2003 ure is particularly important during early member banks in the 1930's. Exit rates were 1933-a time when virtually all banks "sus- higher for nonmemberbanks than for member pended" operations during state and national banks duringthe Depression;nonmember banks bank holidays, but during which only some of fell as a proportion of total banks from 63 those banks failed. Developing a database on percent of the numberof banks in June 1929 to bankfailures makes it possible to investigatethe 57 percent by June 1933.8 Nevertheless, our role of contagion in bank failures during early sample of memberbanks includes a large num- 1933, which is not possible using suspension ber of failed institutions.There were 7,498 Fed data. Understanding the determinants of the member banks in our sample as of the end of large numberof bank failures in early 1933 is a December 1929. 1,528 banks in our sample crucial and omitted part of the history of bank (including banks that entered after December distress during the Depression. 1929) had failed by the end of 1933 (that is, The main obstacle to collecting comprehen- were placed into receivershipor were voluntar- sive bank failure data at high frequency is the ily liquidated). absence of readily available information for Thus our sample of memberbanks comprises non-Fed member banks (hereinafterreferred to a large segment of the banking sector, but the as nonmemberbanks). Given that our data on exclusion of nonmember banks reduces the balance sheets and income statementsonly in- bank failure rate. From an aggregatestandpoint, clude member banks, this limitation was not a the variationover time in bank failures apparent problem from the standpoint of analyzing the in Figure 1 of our paperusing our definition(the failure experiences of banks in our sample. hazard rate of survival for member banks) is Even though our sample includes nearly all quite similarto the patternwhen one defines the Fed member banks, one can question whether failure process using the deposits of all sus- that sample of banks is representative,given the pended commercial banks). In particular,our large numberof excluded nonmemberbanks. It measured raw hazard rate of failure increases is worth noting that nonmember banks were markedlyduring the episodes of banking crisis much smaller on average than member banks, identified by Friedmanand Schwartz (1963). so their number exaggerates their importance. Our analysis of bank failure risk for individ- As of June 30, 1929, nonmemberbanks com- ual banks is for 1930-1933. The beginning of prised 15,797 of the 24,504 banks in existence this periodis dictatedby the startingdate for our (of which 7,530 were national banks and 1,177 balance sheet data (i.e., December 1929). The were state-charteredmember banks). But non- end of this period is dictated by the events of memberbanks only accountedfor 27 percentof early 1933, which saw the suspension of bank banking system deposits ($13.2 billion of the operations through bank holidays, first at the total $49.0 billion).7 The broad patterns of state level via a series of actions by the various growth of member and nonmember banks are state authoritiesin Februaryand March 1933, similar in loans and deposits, although non- and finally at the federal level in March 1933. member banks grew more slowly in the period March 1933 was the end of the last wave of 1921-1929 (loans grew by an average of 27 sudden bank failure during the Depression, the percentfor nonmemberbanks, as opposed to 42 time when the United States departedfrom the percent for member banks) and shrank more gold standard,and markedthe beginning of the quickly from 1929 to 1932 (nonmemberbanks' recovery of 1933-1937. loans declined by an average of 48 percent, The numberof banks that failed in the legal compared to a 35 percent decline for member sense (i.e., that were placed into receivershipor banks). These patternsreflect in large part the voluntarily liquidated) in March 1933 under- consolidation wave produced by agricultural states the true number of failures at that time. distress in the 1920's (which was reflected in Many banks that had suspended in early 1933 the greater growth of member banks) and the remainedin limbo until the regulatoryauthori- greater continuing vulnerabilityof small, non- ties and the RFC determinedwhether to assist

8 7 Data are from Federal Reserve Board (1976, pp. 22- Data are from Wicker (1996, pp. 15), derived from 23). Federal Reserve Board (1976, pp. 22-23, 72, 74). VOL 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESSDURING THE DEPRESSION 1643 them and whether to permit them to become HI. Other County-LevelCharacteristics members of the FDIC. The large number of bank failures in late 1933 and early 1934, there- County-level characteristicdata come from fore, might be best viewed as delayed reactions Historical, Demographic,Economic, and Social to the shocks of early 1933. We discuss this Data: The United States, 1790-1970 [(ICPSR) problemin more detail in our empiricalanalysis study number00003]. This source yields county of bank failures.9 data on, among other items, unemployment rates, acres of land devoted to agriculturalpro- II. County-LevelBank Characteristics duction, the value of agriculturalproduction, land devoted to and value derivedfrom different Our county-level bank data come from the agriculturalproduct types, the numberof farms Federal Deposit Insurance Corporation(FDIC) in different size categories, and the value of Data on Banks in the UnitedStates, 1920-1936 agriculturalinvestment. These data help control [Intra-universityConsortium for Political and for different asset classes, relative values of Social Research (ICPSR) study number00007]. bank portfolios, and other exogenous factors This source provides, for state and national that may affect bank distress. banks, jointly and individually, bank deposits The county characteristic data from the and the number of banks in each county an- ICPSR are coded from decadal U.S. census nually. The data set also reports the number of data. We drew upon ICPSR data from 1920, banks and deposits in banks that suspended 1930, and 1940 so we could test various spec- each year, also at the county level. For some ifications including mixtures of levels and reason, counties in Wyoming are not included growth rates. in this data set, and so any empirical work that relies on this source excludes Wyoming IV. State-LevelEconomic Data counties. The FDIC data allow us to check our distress State characteristicdata are included to in- specification more accurately against previous corporatehigher-frequency time-series data into work by using suspension rates at the county our analysis of the causes and effects of bank level as a dependent variable in addition to distress. However, higher frequency sometimes failures (at the bank level). Furthermore,the comes at the expense of more geographical FDIC data allow us to model deposit flows at aggregation. the county level for a more robust interpreta- We felt it was importantto include available tion of bank distress. The FDIC data yield a measures of local distress that control for het- measure of the size of the banking sector at erogeneity in economic conditions across the the county level that captures all banks, not United States at differenttimes duringthe Great only Federal Reserve member banks. We use Depression. Bradstreet's Weekly published a county-level measures of deposits and sus- monthly summary of building activity in 215 pensions in our county-level regression anal- major U.S. cities during the period of the Great ysis of bank distress. In those regressions, Depression, broken down by new construction explanatory variables aggregate individual and alterations.We include the total series as a bank characteristics to the county level by monthly indicatorof economic conditions in the simple averaging (rather than weighting by local community. When our survival model is size). We use simple averaging at the county restricted to banks located in the 215 major level to correct for the undersampling of cities, this variable includes building permits at small banks that results from our reliance on the city level. When the survivalmodel includes Fed member banks. all Fed member banks, building permits are aggregatedto the state level. An important indicator of local economic 9 conditions is liabilities of failed businesses at Over the course of 1933, banks would be examined and the local level. there is no either permitted to survive or forced to close. For discus- However, single sions of bank resolution policy during 1933, see Cyril B. series of liabilities of failed businesses for the Upham and Edwin Lamke (1934), Susan E. Kennedy entire period comprising the Great Depression. (1973), Wicker (1996), and Mason (2001b). We therefore collected a state-level monthly 1644 THEAMERICAN ECONOMIC REVIEW DECEMBER2003 series of the number and liabilities of failed businesses (outside the banking sector). The businesses from Bradstreet's WeeklyDecember source of this series is Dun's Review, which 1928 through June 1931 and a regional quar- published a continuous national series through- terly series from Dun's Review first quarter out our period. 1931 throughfourth quarter1932. Third, we include Treasurybond yields that To construct quarterly state-level measures control for market conditions affecting bank of these variables at the state level for the investments,particularly relating to the value of second two quartersof 1931, and for 1932, we U.S. government securities in bank portfolios. combined annual state-level figures from Our treasury yields come from Banking and Dun's Review with the quarterly region-level Monetary Statistics, and reflect yields on figures. Specifically, for the last two quarters "... non-callablebonds or callable bonds selling of 1931, we allocated the total liabilities of below call price, in other words, bonds which failed firms to each of the two quarters by are free to reflect changes in interest rates" (p. assuming that the relative proportion of fail- 428). We also experimented with corporate ures in each quarter was the same within any bonds and the Baa spread over Treasuries, region. We used the same method to allocate though these were neither significant nor each state's 1932 failures into the four quar- robust and are excluded from our present ters of that year. We use the overlapping specification. period during the first two quartersof 1931 to calibrate the estimation. We also experimentedwith annual state in- VI. Other Data for MeasuringDistress come measuresfrom John A. Slaughter(1937), but these did not prove significant or robust. We include variables in our specifications that control for several widely held sources of bank distress. These include indicators of al- V. NationalEconomic Data leged national panic episodes as reported by Friedman and Schwartz (1963), indicators of Although we were able to obtain state-level alleged regional panics reported by Wicker data on business failures at quarterly fre- (1980; 1996), an indicator for the Chicago quency, and state-level data on building per- bank panic discussed at length by James mits at monthly frequency, we felt it was (1938) and Calomiris and Mason (1997), and desirable to include additional monthly eco- a constructed measure of nearby bank nomic data in the specification to capture failures. high-frequency fundamental changes. Addi- Friedman and Schwartz national panics are tional monthly data are only available at the captured by indicator variables that take the national level. value of 1 for all banksduring panic months and We include three nationaleconomic variables 0 otherwise. The panic periods we adopt from in our specifications. The first of these is an Friedman and Schwartz and their associated index of agriculturalprices published in The dates are:November 1930-January 1931, May- Farm Real Estate Situation(U.S. Departmentof June 1931, September-November 1931, and Agriculture,various issues). We initially gath- separate indicators for January,February, and ered seven price series on individual commod- March 1933. Robustness checks confirmedthat ity groups including: grains; fruits and extending or contractingthe panic windows for vegetables; meat animals; dairy products;poul- the 1930 or 1931 episodes reduced the magni- try products; cotton and cotton seed; and all tude and significance of the indicator variables groups (30 items). We tested these both indi- in our specifications. vidually and interactedwith our county agricul- Wicker regional panics are capturedby indi- ture characteristic data. Only the national cator variablesthat take the value of 1 for banks overall index proved significant and robust in in affected regions during panic periods and 0 our specifications, and is included in our final otherwise. The panic periods we adopt from results. Wicker and their associated states and dates in The second national economic variable in- our specification are: November 1930-January cluded in our specificationis liabilities of failed 1931 for banks in Tennessee, Kentucky, Ar- VOL. 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESSDURING THE DEPRESSION 1645 kansas, North Carolina, and Mississippi, Calomiris,Charles W. and Gorton, Gary. "The April-July 1931 for banks in Illinois and Origins of Banking Panics: Models, Facts, Ohio, and September-October 1931 for banks and Bank Regulation,"in R. Glenn Hubbard, in West Virginia, Ohio, Missouri, Illinois, and ed., Financial markets and financial crises. Pennsylvania. Chicago: University of Chicago Press, 1991, The Chicago bank panic is captured by an pp. 107-73. indicator variable taking the value of 1 for Calomiris, Charles W. and Mason, Joseph banks in the city of Chicago during June 1932 R. "Contagion and Bank Failures During and 0 otherwise. the Great Depression: The June 1932 Chi- Nearby bank failures are characterized by cago Banking Panic." American Economic using our memberbank data (describedin detail Review, December 1997, 87(5), pp. 863- above) to aggregate deposits in failed banks at 83. the state level for each monthly period. We . "Causes of Bank Distress During experimentedwith various failure windows and the Great Depression." National Bureau lag specifications on deposits in nearby bank of Economic Research (Cambridge, MA) failures. Contemporaneousmonthly data (ex- Working Paper No. 7919, September cluding each failed bank's own deposits where 2000. applicable) yielded the most significant and ro- . "The Consequences of Bank Distress bust results. During the Great Depression." American Economic Review, June 2003a, 93(3), pp. 937-47. REFERENCES . "CausalLinks Among Bank Distress, Asset Values and CorporateDistress During Allen, Franklin and Gale, Douglas. "Financial the Great Depression." Working paper, Co- Contagion." Journal of Political Economy, lumbia University, 2003b. January2000, 108(1), pp. 1-33. Calomiris,Charles W. and Wheelock,David C. Anari, Ali; Kolari, James W. and Mason, Jo- "The Failures of Large SouthernBanks Dur- seph R. "Bank Asset Liquidation and the ing the Great Depression." Working paper, Propagation of the U.S. Great Depression." Columbia University, 1995. Journal of Money, Credit, and Banking Calomiris, Charles W. and Wilson, Berry. (forthcoming). This is a revised version of "Bank Capital and Portfolio Management: Wharton Financial Institutions Center The 1930s 'Capital Crunch' and Scramble (Philadelphia, PA) Working Paper No. 02- to Shed Risk." Journal of Business (forth- 35, August 2002. coming). This is a revised version of Na- Bernanke, Ben S. and James, Harold. "The tional Bureau of Economic Research Gold Standard, Deflation, and Financial (Cambridge, MA) Working Paper No. Crisis in the Great Depression: An Interna- 6649, July 1998. tional Comparison," in R. Glenn Hubbard, Carlson,Mark. "Are Branch Banks Better Sur- ed., Financial markets and financial crises. vivors? Evidence from the Depression Era." Chicago: University of Chicago Press, Federal Reserve Board Working Paper No. 1991, pp. 33-68. 2001-51, November 2001. Bradstreet'sWeekly. New York: BradstreetCom- Comptrollerof the Currency.Record book of pany, various issues. changes to nationally chartered banks [un- Calomiris, Charles W. "Do Vulnerable Econ- published]. omies Need Deposit Insurance? Lessons Diamond, Douglas W. and Dybvig, Phillip H. from U.S. Agriculture in the 1920s," in "Bank Runs, Deposit Insurance,and Liquid- Philip L. Brock, ed., If Texas were Chile: A ity." Journal of Political Economy, June primer on bank regulation. San Francisco: 1983, 91(3), pp. 401-19. Sequoia Institute, 1992, pp. 237-349, 450- Diamond, Douglas W. and Rajan, Raghuram. 58. "Liquidity Shortage and Banking Crises." . U.S. bank deregulation in historical National Bureau of Economic Research perspective. Cambridge:Cambridge Univer- (Cambridge,MA) Working Paper No. 8937, sity Press, 2000. May 2002. 1646 THEAMERICAN ECONOMIC REVIEW DECEMBER2003

Dun's Review. Various issues. Federal Reserve Bank of St. Louis Review, Federal Deposit Insurance Corporation Data May/June 1998, 80(3), pp. 151-52. on Banks in the United States, 1920-1936. . "Do Lender of Last Resort Policies Intra-university Consortium for Political Matter? The Effects of Reconstruction Fi- and Social Research Study #00007. Ann nance CorporationAssistance to Banks Dur- Arbor, MI: University of Michigan and ing the Great Depression." Journal of Intra-university Consortium for Political Financial Services Research, September and Social Research. 2001a, 20(1), pp. 77-95. Federal Reserve Board. Banking and monetary . "Reconstruction Finance Corpora- statistics, 1914-1941. Washington, DC: tion Assistance to Financial Intermediar- Board of Governors, 1976. ies and Commercial and Industrial Friedman,Milton and Schwartz,Anna J. A mon- Enterprisein the U.S., 1932-1937," in Stijn etary history of the United States, 1867- Claessens, Simeon Djankov, and Oshoka 1960. Princeton, NJ: Princeton University Mody, eds., Resolving financial distress. Press, 1963. Washington, DC: World Bank, 2001b, pp. . "The Failure of the Bank of United 167-204. States: A Reappraisal:A Reply." Explora- O'Brien,Anthony P. "TheFailure of the Bank of tions in EconomicHistory, April 1986, 23(2), United States: A Defense of Joseph Lucia." pp. 199-204. Journal of Money, Credit, and Banking, Au- Grossman, Richard S. "The Shoe That Didn't gust 1992, 24(3), pp. 374-84. Drop: Explaining Banking Stability During Ramirez, Carlos D. "Did Branch Banking Re- the Great Depression." Journal of Eco- strictions Increase Bank Failures? Evidence nomic History, September 1994, 54(3), pp. from Virginia and West Virginia in the Late 654-82. 1920s." Working paper, George Mason Uni- Historical, Demographic, Economic, and Social versity, 2000. Data: The United States, 1790-1970. Intra- Rand McNally Banker's Directory. New York: university Consortium for Political and So- Rand McNally & Co., various issues. cial Research Study #00003. Ann Arbor, Reports of Condition. Washington, DC: Board of MI: University of Michigan and Intra- Governors of the Federal Reserve System. university Consortium for Political and So- Unpublished microfilm archive, various call cial Research. dates. Imbens,Guido W. "TransitionModels in a Non- Reports of Earnings and Dividends. Washington, StationaryEnvironment." Review of Econom- DC: Board of Governors of the Federal Re- ics and Statistics,November 1994, 76(4), pp. serve System. Unpublished microfilm ar- 703-20. chive, various call dates. James, F. Cyril. The growth of Chicago banks, Slaughter,John A. Income received in the vari- volume II. New York: Harper, 1938. ous states, 1929-1935. New York: National Kennedy,Susan E. The banking crisis of 1933. Industrial Conference Board Studies, No. Lexington, KY: Kentucky University Press, 234, 1937. 1973. Temin, Peter. Did monetaryforces cause the Kiefer, Nicholas M. "Economic Duration Data Great Depression?New York: W.W. Norton, and Hazard Functions." Journal of Eco- 1976. nomic Literature, June 1988, 26(2), pp. . Lessons from the Great Depression. 646-79. Cambridge,MA: MIT Press, 1989. Lancaster, Tony. The econometric analysis of Trescott,Paul B. "TheBehavior of the Currency- transition data. Cambridge:Cambridge Uni- Deposit Ratio during the Great Depression: versity Press, 1990. A Comment." Journal of Money, Credit, Lucia, Joseph. "The Failure of the Bank of and Banking, August 1984, 16(3), pp. 362- United States: A Reappraisal."Explorations 65. in Economic History, October 1985, 22(4), . "The Failure of the Bank of United pp. 402-16. States, 1930." Journal of Money, Credit, Mason,Joseph R. "AmericanBanks During the and Banking, August 1992, 24(3), pp. 384- GreatDepression: A New ResearchAgenda." 99. VOL. 93 NO. 5 CALOMIRISAND MASON:BANK DISTRESS DURING THE DEPRESSION 1647

U.S. Departmentof Agriculture. The farm real Wicker, Elmus. "A Reconsideration of the estate situation. Washington, DC: U.S. Causes of the Banking Panic of 1930." Jour- Government Printing Office, various is- nal of Economic History, September 1980, sues. 40(3), pp. 571-83. Upham,Cyril B. and Lamke,Edwin. Closed and . The bankingpanics of the Great De- distressed banks. Washington, DC: Brook- pression. Cambridge:Cambridge University ings Institution, 1934. Press, 1996. White, Eugene N. "A Reinterpretationof the Wigmore,Barrie A. "Was the Bank Holiday of Banking Crisis of 1930." Journal of Eco- 1933 a Run on the Dollar Rather than the nomic History, March 1984, 44(1), pp. 119- Banks?" Journal of Economic History, 38. September 1987, 47(3), pp. 739-56.