Bank lablty structure, FDIC loss, and tme to falure: A quantle regresson approach Klaus Schaeck* October 2006 * Correspondng address: Unversty of Southampton, Hghfeld, Southampton SO17 1BJ, Unted Kngdom; E-mal address: kschaeck@soton.ac.uk; Phone: ++ 44 (0) 23 8059 3557; Fax: ++ 44 (0) 23 8059 3844 Ths research was undertaken durng my stay as vstng scholar at the Department of Fnance at the Unversty of Illnos at Urbana-Champagn. Ths study has benefted from valuable gudance by George Pennacch. I am ndebted to Chrstopher James, Rosalnd Bennett, Lynn Shbut and Phlp Shvely for sharng ther extensve expertse. I also would lke to thank Roger Koenker, Masak Yamada, Martn Chak, Evangelos Benos, Smon Wolfe and conference partcpants at the 6 th Annual Bank Research Conference n Arlngton, Vrgna and semnar partcpants at the Unversty of Southampton for ther comments. All remanng errors are my own.
Bank lablty structure, FDIC loss, and tme to falure: A quantle regresson approach ABSTRACT Prevous studes that am to determne factors mpactng the depost nsurer s loss arsng from bank falures use standard econometrc technques that assume the losses are homogeneously drven by the same set of explanatory varables: However, depost nsurers are partcularly concerned about hgh-cost falures. If the factors drvng hgh-cost falures dffer systematcally from the determnants of low and moderate-cost falures, an alternatve estmaton technque s requred. Usng a sample of more than 1,000 bank falures n the US between 1984 and 1996, we present a quantle regresson approach that llustrates the senstvty of the dollar value of losses n dfferent quantles to our explanatory varables. The fndngs suggest that relance on standard econometrc technques gves rse to msleadng nferences and that losses are not homogeneously drven by the same factors across the quantles. We also fnd that lablty structure affects tme to falure and that both nsured and unnsured depostors are a source of market dscplne. Keywords: JEL Classfcaton: bank lablty structure; loss gven default; market dscplne; tme to falure; quantle regresson G21; G28; C41; C49-2 -
1. Introducton Depost nsurers need to determne the losses arsng to them from bank falures to adequately prce depost nsurance and adjust the resources of the nsurance fund accordngly. An ongong dscusson about the settng of the desgnated reserve rato and the dfferentaton of prcng schemes by bank sze motvates a recent proposal by the Federal Depost Insurance Corporaton (FDIC) to reform depost nsurance legslaton. Ths debate underscores the contnued need to nvestgate the determnants of losses caused by falures of fnancal nsttutons. Whle a consderable body of lterature exsts on the factors mpactng depost nsurers losses, these studes are lmted n two dstnct aspects: Frst, they largely focus on the faled banks asset composton and asset qualty as key drvers for the loss ncurred. However, bank lablty structure also has substantal bearng for the prcng of depost nsurance and therefore also mpacts eventually depost nsurers losses (Pennacch, 2005). In addton, Shbut (2002) underscores that the structure of deposts not only determnes whch depostors have to be compensated n case of falure, but t s furthermore an nfluental factor for an nsttuton s rsk takng behavor. 1 Ths, n turn, affects potental losses by the nsurer. Thus, lablty structure affects FDIC losses n two dstnct ways: ) drectly by determnng the FDIC s oblgatons and ts relatve poston among the faled bank s credtors, and ) ndrectly through market dscplne and the mpact on asset qualty. Second, exstng work uses standard econometrc technques such as ordnary least squares that do not suffcently account for the skewed dstrbuton of the losses and the heterogeneous populaton of the faled nsttutons. Snce depost nsurers are partcularly concerned about hgh-cost falures (n terms of absolute dollar values) due to ther possbly systemc mpact on the nsurance fund, t s pertnent to understand whether losses are homogeneously drven by the same determnants or f factors mpactng resoluton costs of expensve falures dffer systematcally from the factors observed n less expensve falures. Thus, f there are systematc dfferences between hgh-cost and low-cost falures, there may be opportuntes to develop further mprovements for the regulatory envronment. For nstance, amendments n the way troubled banks are treated by means of prompt correctve acton may be consdered snce such provsons are ntended to reduce costly falures to the depost nsurer (Shbut et al., 2003). Lkewse, mplcatons for the adequacy of the depost nsurer s reserves may be derved. Ths paper contrbutes to the lterature on losses arsng to depost nsurers n three dstnct ways: Frst, to dfferentate between the factors drvng hgh-cost and low-cost falures, we ntroduce a methodologcal advancement usng quantle regresson, also referred to as least absolute devaton regresson, for a sample of more than 1,000 bank falures n the US durng the perod 1984 1996. Ths enables us to focus on the tals of the dstrbuton of the loss varable and permts better nferences about the factors contrbutng to hgh-cost falures. Moreover, employng quantle regresson mtgates the problems assocated wth 1 Kng et al. (2006) hghlght recent changes n the envronment banks operate n. Deeper and wder fnancal markets offer new opportuntes for depostores lablty management. They stress, nter ala, that banks are relyng ncreasngly on non-core fundng such as jumbo CDs, brokered deposts, and Federal Home Loan Bank (FHLBank) advances. To the extent to whch these funds are nsured or mplctly guaranteed by the government, they gve rse to moral hazard and hence alter the rsk profle of fnancal nsttutons. - 3 -
relyng on a sngle measure of central tendency of the dstrbuton of the loss varable and permts nferences about the relatve mportance of certan regressors at dfferent ponts of the dstrbuton of the losses. Therefore, quantle regresson can be consdered superor to the prevously used estmaton technques snce t provdes more precse estmates of the mpact of the determnants of losses. Second, we test as to whether bank lablty structure plays a role n determnng the loss when banks fal. Gven the substantal evdence n the lterature that alng nsttutons tend to substtute unnsured deposts n the run-up to falure wth nsured deposts, thereby ncreasng the cost to the nsurer, t s crtcal to focus on the extent to whch dfferent types of labltes mpact upon loss. Fnally, the new Basel Captal Accord hghlghts n Pllar 3 the role of market dscplne to constran rsk-takng behavor of fnancal nsttutons. Thus, we hypothesze that depostores heavly relant on unnsured deposts are lkely to fal faster than nsttutons funded by other sources snce holders of unnsured clams can respond to mpendng falure wth wthdrawal of funds. Alternatvely, falng banks wll attempt to substtute the cash outflows wth nsured deposts, thus ncreasng the depost nsurer s rsk exposure. Our hypothess bears mportant polcy consderatons: If such banks tend to fal faster, they would have to be subject to addtonal measures of prompt correctve acton to prevent substtuton of unnsured clams wth nsured deposts. We therefore test the effect of lablty structure on tme to falure, and estmate an accelerated falure tme model wth tme-varyng covarates for a large sample of faled and non-faled depostores durng the perod 1982 1996. To our knowledge, the nexus between market dscplne and lablty structure on the one hand and tme to falure on the other has not yet been subject to extensve econometrc analyss. In addton, the relatonshp between market dscplne and depost nsurance by type of account has been wdely gnored n the extant lterature on market dscplne. We show that the evoluton of FDIC losses, defned as the log of the dollar value of the cost ncurred by the FDIC, exhbts consderable varaton across dfferent quantles of the dstrbuton. The focus of ths study s on the absolute dollar value of losses nstead of loss rates. Ths s due to the fact that the depost nsurer s concerned about costly falures n terms of the absolute dollar value as such falures can pose a systemc threat to the nsurance fund (Oshnsky, 1999; Shbut, 2002). 2 Moreover, t s well-establshed that lager banks tend to have lower loss rates (Shbut, 2002; Oshnsky, 1999). Thus, focusng on loss rates would gve rse to msleadng nferences for the purpose of ths study. Our quantle regresson results llustrate that the loss varable n dfferent quantles shows sgnfcantly dfferent senstvtes to the utlzed set of explanatory varables. In partcular, the results ndcate that bank sze, the ratos of real estate owned, C&I loans, agrcultural loans, real estate loans, and ndvdual loans to total assets exhbt a varyng mpact upon FDIC loss as we move up the dstrbuton. Smlarly, depostor preference law, bankruptcy growth and unemployment rates on the state level also exhbt non-lnear behavour. Whle our results also show an mportant effect of certan lablty varables on the loss varable, we do not detect any varyng effect of these varables between hgh-cost and low-cost falures. To ths extent, our results extend recent work by Shbut et al. (2003) that provdes crcumstantal evdence for dfferences of medans of a set of certan balance sheet and 2 The adverse repercussons of a large bank falure could be amplfed by the so-called systemc-rsk excepton, whch would further ncrease exposure of the FDIC as ths mght even ental compensaton of unnsured credtors (Shbut, 2002). - 4 -
ncome statement varables between low-cost and hgh-cost falures. 3 Algned wth theory, we fnd that the rato of Fed funds purchased to total assets s negatvely assocated wth FDIC loss for low and moderate-cost falures. Our fndngs for the sgnfcant bearng of lablty structure varables on losses underscore the mportance of consderng lablty structure when analyzng losses to depost nsurers. Lkewse, the fact that the varables that capture bank sze, the ratos of real estate owned, C&I loans, agrcultural loans, real estate loans, and ndvdual loans to total assets reveal hghly nonlnear relatonshps wth the loss varable substantate that an alternatve to standard estmaton procedures s requred when analyzng depost nsurers losses. Smlarly, depostor preference law, bankruptcy growth and unemployment rates also exhbt varyng mpact on FDIC losses. Ths furthermore suggests that relance on estmates obtaned wth standard econometrc technques gves rse to msleadng nferences wth respect to the mpact of certan factors on FDIC losses. Regardng the determnants that drve losses for costly falures, we show that these falures are partcularly nfluenced by bank sze, real estate owned, uncollected ncome, and C&I loans. Moreover, a sluggsh macroeconomy s also found to ncrease the losses arsng from expensve falures. Estmatng an accelerated falure tme model wth tme-varyng covarates, we furthermore demonstrate that the ratos of Fed funds, brokered deposts, as well as demand and tme and savngs deposts to total assets tend to shorten falure tme, whereas transactons deposts that proxy the charter value of a bank ncrease survval tme of a bank. These results are robust to controllng for the mpact of asset qualty, captalzaton, earnngs, lqudty and the macroeconomc settng whch banks operate n. Our fndngs provde a ratonale for further strengthenng dsclosure of the levels of nsured and unnsured deposts n fnancal nsttutons to enhance depostor dscplne. The plan of the paper s as follows: Secton 2 revews related work and Secton 3 presents an overvew on the methodology employed. The econometrc analyss s provded n Secton 4 and Secton 5 offers concludng remarks and avenues for future research. 2. Related Work Our survey of related studes draws from two dstnct strands n the lterature. We frst focus on work regardng the losses arsng from bank falures and then dscuss the lnk between depostor preference laws, depostor dscplne and the cost of bank falures. A number of studes model the loss on assets as a functon of the faled banks asset composton, ts asset qualty and a set of addtonal varables. Bovenz and Murton (1988) draw upon a sample of bank falures between 1985 and 1986 n the US and report an average loss rate of 33 percent of assets. Usng ordnary least squares regresson analyss, they addtonally hghlght the role of uncollected ncome, and geographc dfferences n explanng the loss on assets. Barth et al. (1990) and Blalock et al. (1991) examne resoluton costs of thrft falures. Barth et al. (1990) employ a Tobt model for the perod 1984 1988 and present evdence that tangble net worth, asset qualty and core deposts as a proxy for franchse value are sgnfcant determnants of the depost nsurer s loss. Smlarly, Blalock et al. (1991) confrm that asset mx s a major determnant of resoluton 3 Shbut et al. (2003) dvde FDIC loss by total assets and classfy falures wth resoluton cost below 12 percent of assets as low-cost falures. - 5 -
costs. James (1991) presents an examnaton of bank falures durng the perod 1985 1988 and reports an average loss of 30 percent of the faled bank s assets. He moreover underscores the relatve mportance of unrealzed losses, the determnants of charter value and type of resoluton procedure for the loss on assets. Brown and Epsten (1992) extend these studes and dsaggregate the loss on assets nto dfferent asset categores. Usng detaled nformaton on recevershp recoveres, they llustrate that the loss on assets vares over dfferent asset categores and over tme to reterate that portfolo composton s a key determnant of the loss rate. Osterberg and Thomson (1994) buld on prevous work and conclude that the dollar value of resoluton costs s not only a functon of asset qualty. Employng data for US bank falures between 1986 and 1992, they fnd that loss s furthermore nfluenced by bank sze, fraud and off-balance sheet tems, and that brokered deposts tend to decrease loss. Recent work by McDll (2004) drawng upon a large sample of falures between 1984 and 2002 analyses the effect of the busness cycle on resoluton costs. She contemplates that the depost nsurer s loss ncreases n a sluggsh macroeconomc envronment. Corroboratng the role of asset composton and franchse value hghlghted n prevous studes, she addtonally fnds that the pool of potental acqurers of a faled bank s an nfluental factor for the loss rate. Bennett et al. (2005) study the mpact of Federal Home Loan Bank (FHLBank) advances on expected losses to the Bank Insurance Fund (BIF) and pont out that subordnaton of FDIC clams to FHLBank advances ncreases both probablty of default and loss gven default. 4 A related body of lterature focuses on the role of depostor preference laws, desgned to reduce the cost of falures to the depost nsurer. 5 Hrschhorn and Zervos (1990) put forward that nondepost credtors mght respond wth collateralzng ther clams when depostor preference laws are enacted. The authors emprcal analyss of thrft nsttutons n the US confrms that large proportons of collateralzed clams contrbute to hgher cost of falures, gvng rse to unntended outcomes from a depost nsurer s perspectve. On the other hand, Osterberg (1996) substantates that depostor preference laws decrease resoluton costs for falures of commercal banks between 1984 and 1992. However, he also dscusses offsettng effects arsng from collateralzaton of clams by nondepost credtors. Marno and Bennett (1999) analyze falures of sx large US commercal banks between 1984 and 1992 to nvestgate f depostor preference law affects large nsttutons dfferently due to ther greater dependency on nondepost and foregn labltes. Gven that depostor preference law provdes unnsured and unsecured clamants wth an ncentve to protect themselves from losng money, an alng bank s lablty structure s lkely to change as t approaches falure. Whle the authors do not offer an econometrc analyss of the assocaton between lablty structure, depostor preference law and FDIC loss, they llustrate that lablty structure experences consderable changes pror to falure, whereby unnsured and foregn deposts decrease substantally. 4 5 Note that ther loss estmates requre knowledge of the exstng lablty structure of the bank under consderaton. The Depostor Preference Act of 1993 was desgned to shft the burden of bank falure from taxpayers to unnsured depostors. It gves depostors clams on a faled nsttuton s assets superor to those of general credtors. Several states had depostor preference laws n place pror to 1993. For detaled expostons of depostor preference see Osterberg (1996) and Marno and Bennett (1999). - 6 -
Consderable effort has gone nto the analyss of how depostors dscplne fnancal nsttutons. 6 Holders of unsecured clams have an ncentve to montor rsk-takng behavor of banks and dscplne them by demandng approprate rsk premums, collateral or by wthdrawng ther funds. Goldberg and Hudgns (1996, 2002) nvestgate the holdngs of unnsured deposts at savngs and loan assocatons over dfferent samplng perods and llustrate that falng nsttutons experence declnes n unnsured deposts. Ths result s algned wth work by Jordan (2000), who analyses lablty structure of falng banks n New England n the early 1990s. Bllet et al. (1998) study the mpact of ratngs downgrades as a proxy for ncreased rsk n fnancal nsttutons and report that downgraded banks ncreasngly rase nsured deposts. Ths not only ncreases the depost nsurer s exposure but also suggests that market dscplne nsuffcently polces banks aganst rsk takng behavor snce rsk-based captal standards and rsk based depost nsurance both fal to consder banks lablty structure. Thus, the evdence that alng nsttutons substtute unnsured deposts wth nsured deposts suggests the undermnng of market dscplne. Furthermore, ths phenomenon s bound to ncrease the depost nsurer s loss f the troubled bank eventually defaults. Park and Perstan (1998) focus on the mplcatons of rsk for prce and quantty of unnsured deposts n a sample of thrfts. Insttutons wth a hgher probablty of falure are found to offer hgher nterest rates on unnsured funds. Due to ther ncreased rsk profle, such thrfts however attract smaller amounts of unnsured deposts. These results are consstent wth the vew that unnsured depostors are a source of market dscplne. Recent work by Maechler and McDll (2006) nvestgates how banks respond to depostor dscplne. The study argues that bank behavor and depostors response s a jontly determned process and provdes evdence that depostors constran bank rsk-takng behavor. In contrast to Park and Perstan (1998), ther results ndcate that weak banks cannot rase unnsured deposts by ncreasng the nterest rates offered, whereas sound nsttutons are able to do so. Usng ndvdual bank-level data, Davenport and McDll (2006) focus on the behavor of fully nsured depostors pror to the falure of Hamlton Bank and uncover that nsured depostors are also a source of market dscplne. They present strong evdence that the total balance of nsured deposts that exted pror to the falure exceeds the amount of unnsured deposts wthdrawn. In partcular, they fnd that holders of fully nsured personal accounts wthdraw large balances n the run-up to falure, whereas certan holders of unnsured accounts vrtually exerted no dscplne. These fndngs ndcate that current regulatory practce nsuffcently recognzes the dscplnary effect arsng from protected depostors. 3. Data and Methodology Our ntal sample conssts of 1,515 faled banks that were resolved by the BIF durng the perod 1984 1996. 7 Snce falng nsttutons have been resolved by the FDIC through 6 7 We constran our revew of related studes to the drect lnk between depostor dscplne and fnancal nsttuton s response to ncreases n rsk. Some other studes nvestgate whether nvestors can dscrmnate between the rsks undertaken by US banks (Flannery and Sorescu, 1996) and how subordnate debt mpacts upon rsk-takng behavor of fnancal nsttutons (Blum, 2002). The samplng perod s constraned by the Fed funds varable. Ths varable s not avalable on the Quarterly Report of Condton and Income (Call Report) for the perod 1997 2003 and we therefore sample the faled nsttutons up to 1996 only. - 7 -
varous dfferent types of transactons, we follow the FDIC s bank falure database 8 and classfy falure as ether one of the followng nstances havng occurred: asssted merger, purchase and assumpton, transfer and assumpton of nsured deposts, re-prvatzaton, closng and reopenng, or depostor payoff. A bank s also classfed as havng faled f t was subject to the management consgnment programme. Bank specfc data are taken from the Quarterly Report of Condton and Income (Call Report) pror to falure. In nstances where no fnal report was avalable, we use the last avalable call report. 9 Informaton on the cost ncurred by the FDIC was obtaned from the FDIC s database on bank falures. Ths nformaton s an estmate of the FDIC s resoluton cost calculated as the dfference between net cash outlays and the estmated dscounted net recovery on any assets remanng n the recevershp s books. We normalze the explanatory varables by total assets to enable comparson wth prevous work (e.g. Shbut et al., 2003; McDll, 2004). Moreover, for our descrptve comparson of loss rates as detaled further below n ths secton, normalzng by total deposts would yeld unusually hgh loss rates as certan types of banks such as trust banks are not heavly relant on deposts. We apply several selecton crtera that have to be satsfed for ncluson of a faled nsttuton nto the econometrc analyss. Frst, our nferences may be msleadng f we nclude falures caused by fraud as the Call Reports may not be nformatve n such nstances (McDll, 2004). We therefore exclude publcly known falures where fraud was the man cause as mentoned by Gup (1995) to adjust the sample accordngly. For the samplng perod not covered by Gup (1995), we addtonally revew FDIC press releases and exclude those falures where fraud s mentoned as a reason for falure. Second, crossguarantee falures (e.g. Frst Republc and Frst Cty) are excluded from the analyss as they cannot be vewed as ndvdual falures (for a detaled dscusson see Ashcraft, 2003). Thrd, mult-bank holdng company falures were consoldated nto one. Fnally, mssng values for some explanatory varables further lmt the dataset to 1,066 faled bank observatons whch can be used for our econometrc analyss. In order to test the effect of lablty structure on FDIC loss, we nclude several depost and non-depost categores nto the regressons. Frst, we consder the rato of transactons deposts to total assets and antcpate an nverse relatonshp between ths rato and FDIC loss. Transactons deposts can be perceved as core deposts that proxy the charter value of a bank, whch would be lost n case of falure (James, 1991; Osterberg and Thomson, 1994). Second, the ratos of demand deposts, tme and savngs deposts, and brokered deposts to total assets are ncorporated as they capture mportant nformaton about the breakdown of the faled banks depost structure by account type. Note that these categores do not dscrmnate between the status of depost nsurance. Therefore, t s not ex-ante clear whether they ncrease or decrease FDIC loss. To the extent that they are nsured, they wll ncrease losses; to the extent that they are not nsured, they wll mtgate the depost nsurer s loss. However, as alluded to n the lterature revew, recent mcro-level 8 9 http://www.fdc.gov/bank/ndvdual/faled/ndex.html Note that relance on publcly avalable Call Report data on a quarterly bass from the Call Report mmedately precedng the falure hampers separatng out nsured and unnsured deposts. The Call Report tem contanng nformaton on depost accounts wth balances over 100,000 USD was only reported n June Call Reports pror to 1991; see also Maechler and McDll (2006). - 8 -
evdence by Davenport and McDll (2006) provdes strong evdence that nsured depostors wthdraw larger volumes than unnsured depostors. Thus, f the majorty of these types of deposts that s left n the bank at falure date s unnsured, these results pont towards a negatve relatonshp between these types of deposts and FDIC loss. We also consder a number of non-depost varables as large nsttutons tend to rely more heavly on nondepost fundng (Shbut, 2002). The rato of subordnated debt to total assets s ncluded as the use of subordnated debt has become ncreasngly popular for banks to satsfy captal requrements (Evanoff and Wall, 2002). Snce subordnated debt s unnsured, an nverse relatonshp between subordnated debt and the loss varable s antcpated. Thrd, we test for the effect of fundng through Fed funds. Fed funds are obtaned n the nterbank market and are not nsured. Therefore, we assume that relance on Fed funds wll decrease FDIC losses. The remanng non-depost lablty components of the faled banks balance sheet are grouped together n the rato of other labltes to total assets. These types of labltes are not nsured and we expect ths varable to enter the loss equaton wth a negatve sgn. Several control varables are consdered. James (1991) and Osterberg and Thomson (1994) show that asset qualty s a major determnant for the loss varable. We therefore nclude the ratos of loans past due (90 + days) and real estate owned to total assets to control for asset qualty. The latter varable provdes nformaton on the volume of real estate obtaned due to foreclosure. Numerous prevous studes also report that the level of uncollected ncome s an mportant predctor for losses (e.g. James, 1991; Osterberg and Thomson, 1994; McDll, 2004). We therefore nclude the rato of uncollected ncome to total assets as a further control varable. In addton, bank falure s often preceded by strong asset growth n the run-up to falure (McDll, 2004). Hence, a varable that captures asset growth n the 24-month perod pror to falure s ncluded. We addtonally consder the book value of equty to total assets. Equty serves as a cushon between asset value and the payments to debt holders and we antcpate an nverse relaton between the book value of equty to total assets and FDIC loss. Brown and Epsten (1992) have shown that dfferent types of assets exhbt dfferent recovery rates and Blalock et al. (1991) propose groupng certan asset types nto separate categores due to smlar credt-rsk characterstcs. We therefore addtonally consder dfferent asset categores that capture nformaton on the loan portfolo and test for the effect of the ratos of C&I loans, agrcultural loans, real estate loans, and ndvdual loans to assets on the loss varable. These varables are expected to enter the loss equaton negatvely. Furthermore, we ncorporate varables that capture nformaton about the macroeconomc envronment. Usng formaton on personal ncome growth, bankruptcy growth and unemployment on the federal state level, McDll (2004) has shown a strong lnk between such factors and losses arsng to the FDIC. These varables are obtaned from the Amercan Bankruptcy Insttute, from the Bureau of Labor Statstcs and from the Bureau for Economc Analyss. 10 We also nclude a dummy varable that takes the value one f the observaton s taken from the perod followng enactment of the Federal Depost Insurance 10 The data for personal ncome growth were obtaned from http://bea.gov/bea/regonal/statelocal.htm; the data for unemployment rates were obtaned from http://www.bls.gov/lau/home.htm; and the data for bankruptcy growth can be retreved at http://www.abworld.org. - 9 -
Corporaton Improvement Act (FDICIA). Ths act was ratfed by Congress as response to a pervasve fear that the problems experenced n the thrft ndustry n the 1980s would spread to commercal banks. FDICIA was, nter ala, desgned to reduce costs arsng from bank falures to the depost nsurer and embodes a fundamental overhaul of depost nsurance and prudental regulaton (Benston and Kaufman, 1997). Fnally, we nclude a dummy varable that takes on the value one f depostor preference law was n place at the tme of the falure or zero otherwse. As hghlghted above, depostor preference laws are ntended to shft the burden from the taxpayer to unnsured holders of credt to mtgate the losses arsng to the depost nsurer (Osterberg, 1996). Ths dummy varable takes account of the fact that some states already had depostor preference laws n place pror to the enactment of natonal depostor preference. 11 We also control for asset sze usng the log of total assets. Larger nsttutons are assumed to have a hgher loss on assets. Table 1 presents summary statstcs for our dataset. [TABLE 1] To enable comparson wth prevous studes, we also compute a loss rate, calculated as FDIC loss dvded by total assets. The average loss rate for the full sample s 34 percent of total assets, slghtly hgher than n James (1991). Our detaled breakdown llustrates a large degree of varaton across the quantles. Whle the loss rate s 3 percent of total assets for the low-cost falures (5 th quantle), falures at the upper tal of the dstrbuton cost the nsurer more than 46 percent of total assets. The most expensve falure had a loss rate of more than 133.5 percent of total assets. It s mportant to recognze that FDIC losses more than double between the 90 th and the 95 th quantle, suggestng a consderable ncrease n losses between the quantles at the upper tal of the dstrbuton. We therefore consder expensve falures as those falures where losses le at the 95 th quantle and above of the dstrbuton. Ths reflects that depost nsurers are partcularly concerned about those falures that may pose a systemc threat to the nsurance fund. Oshnsky (1999) ponts out that the solvency of the bank nsurance fund s very closely ted to the soundness of the largest nsttutons. Hs smulatons project a 98.5 percent probablty that any future nsolvency of the bank nsurance fund wll nvolve falure of one of the 25 largest bankng companes n the US. Furthermore, our sample also shows that the faled banks have a mean of total assets of 153m USD, wth the largest falures exceedng 17bn USD. A few varables stand out: Total assets durng the 24 months pror to falure declne on average 11 percent, ndcatng that troubled depostores shrnk consderably n the two years before falure. The rato of real estate owned to total deposts s on average 5 percent. Real estate owned has been found n prevous studes to be an approprate predctor for resoluton costs snce ths category contans propertes obtaned by foreclosure. Lkewse, the rato of uncollected ncome fgured promnently n prevous work because of ts ndcatve character for loans that have not been wrtten off. Ths rato has a mean of 1 percent. In terms of the fundng structure, the average rato of Fed funds purchased to total deposts s 1 percent, whereas the rato of brokered deposts s 2 percent of total assets. Faled banks have on average a rato of 25 percent of transactons deposts to total assets. Whle tme and savngs deposts represent 11 Addtonal detals for the codng of the depostor preference law dummy are provded n the Data Appendx. - 10 -
83 percent of total assets, the rato of demand deposts to total assets amounts to 15 percent. 3.1 Cost of Falure Our sample conssts of dfferent types of banks (communty banks, savngs banks, commercal banks, etc.) that pursue dfferent types of busness actvtes. Brown and Epsten (1992) pont out that a falng bank heavly concentrated n commercal loans s therefore lkely to exhbt larger losses than an nsttuton that prmarly engages n retal lendng actvtes. 12 Moreover, our sample exhbts large varaton wth respect to bank sze. Bank sze, as llustrated by Marno and Bennett (1999), n turn, nfluences bank lablty structure, whch ultmately affects the dependent varable n our analyses. Thus, numerous factors suggest that losses vary consderably across the dstrbuton and that a regresson technque s requred that helps gan detaled nsghts as to whether the factors drvng losses dffer systematcally across the dstrbuton of the loss varable. We start analyzng the lnk between the loss varable and a set of explanatory varables usng ordnary least squares regresson, smlar to the approach pursued n prevous work. We model losses as y = α + βx + u (1) whereby y denotes the loss ncurred by the depost nsurer for bank, α s the constant term and, β captures the coeffcents to be estmated for the explanatory varables x ; u s the error term. In order to account for the skewed dstrbuton of the loss varable and draw more approprate nferences about the senstvty of the losses at the tals of the dstrbuton, we use the condtonal quantle regresson estmator developed by Koenker and Bassett (1978). Gven the heterogenety of our dataset, condtonal quantle regresson not only permts drawng more precse nferences about the mpact of regressors at certan ponts of the dstrbuton of the loss varable but also offers an estmaton procedure more robust to departures from normalty because lnear estmators would more lkely produce neffcent and based estmates. Snce we are not aware of any study n the bankng lterature employng quantle regresson, we revew the key characterstcs of ths technque below. 13 Whle classcal lnear regresson estmates condtonal mean functons, quantle regresson permts estmatng condtonal quantle functons,.e. models n whch quantles of the dependent varable are expressed as functons of a set of explanatory varables (Koenker and Hallock, 2001). 14 regresson s approprate when a large degree of varaton n 12 13 14 Brown and Epsten (1992) compute a loss rate as loss dvded by total assets. regresson has been utlzed n labor economcs, demand analyss, n emprcal fnance n the lterature on value at rsk and n ecology and bostatstcs. For recent overvews of applcatons of quantle regresson we refer the nterested reader to the surveys by Koenker and Hallock (2001) and Cade and Noon (2003). s dvde the cumulatve dstrbuton functon of a random varable nto a gven number of equally szed segments. s are the general case of certan other ways of splttng a populaton nto segments. For nstance, quartles dvde a populaton nto four segments wth equal proportons of the - 11 -
the data suggests that there may be more than a sngle slope parameter descrbng the relatonshp between the dependent varable and the regressors. Thus, quantle estmaton goes beyond lnear regresson n that t gves a more complete pcture of the effect of a set of regressors on the dfferent quantles of the dependent varable. Gven that the θ th quantle of a condtonal dstrbuton of y s lnear n x and assumng ( y, x ), = 1,..., n s drawn from the populaton of faled nsttutons whereby x s a K 1 vector of explanatory varables, we wrte the condtonal quantle regresson model as y = x β + u (2) θ θ θ { y : F ( y x θ} = xβθ Quant ( y x ) nf ) (3) Quantθ ( uθ x ) = 0 (4) where Quantθ ( uθ x ) captures the θ th condtonal quantle of y on the regressor vector β θ s the vector of parameters to be estmated for dfferent quantles x. The expresson θ, lyng n the range (0;1). The error term dfferentable c.d.f. (. x) F uθ and a densty functon ). u θ s assumed to have a contnuously (. x The entre dstrbuton of y condtonal on x can be traced by movng along the (0;1) nterval of θ. To estmate we proceed as follows and mnmze n mn ρ θ (y x βθ ) (5) whereby ρ (u) s defned as follows θ θu f u 0 ρ θ ( u) =. (6) ( θ 1) u f u < 0 Ths mnmzaton problem can be solved accordng to Koenker and Bassett (1978) usng lnear programmng technques. The covarance matrx of the parameter vector can be obtaned usng bootstrap methods to calculate standard errors and confdence ntervals. We use ths quantle estmator to nvestgate as to whether our asserton of systematc dfferences of the mpact of regressors on the loss varable s correct n Secton 4.1. 3.2 Tmng of falure To test the effect of fundng structure on tme to falure, we utlze an accelerated falure tme (AFT) model wth tme-varyng covarates. Such models are called accelerated falure tme models because the effect of the ndependent varables s to rescale tme,.e. to accelerate or decelerate tme to falure. f uθ β θ reference populaton n each segment, and the medan dvdes the populaton nto two equally szed segments (Koenker and Hallock, 2001). - 12 -
We formalze tme untl falure as a probablty densty functon of tme t. A convenent way of descrbng survval of a depostory past tme t s through ts survvor functon S( t) = P( T t) (7) whch equals one mnus the cumulatve dstrbuton functon of T. We then can compute the condtonal probablty of closure wthn the tme nterval t untl t + h, gven survval untl tme t, as P{ t T t t + ht t}. (8) Ths probablty can be dvded by h, to calculate the nstantaneous rate of falure,. e. the average probablty of leavng per unt tme perod over the nterval t untl t + h such that the hazard functon can be wrtten as λ( t) P{ t T t t + ht t} d log S( t) f ( t) lm = =. (9) h 0 h dt S( t) = Accelerated falure tme models are wrtten n the form ln( t ) = β + τ (10) j x j x j where ln( t j ) s the log of tme to falure, x j denotes our explanatory varables, β x are the parameters to be estmated and τ j s a random varable that follows a dstrbuton. Thus, to estmate the model, we need to determne the dstrbuton of τ j and specfy τ j to follow the log-logstc dstrbuton. Ths dstrbuton s rather flexble snce t permts two nflexon ponts for the hazard functon. The log-logstc dstrbuton was utlzed n prevous work on bank falures and bank ext (Cole and Gunther, 1995; DeYoung, 2003). The parameters of nterest can be obtaned usng maxmum lkelhood estmaton technque. The samplng perod for ths analyss starts n 1982 and we use the same set of faled nsttutons that underle the cost equatons. The startng date 1982 s chosen to assert that we have at least eght quarterly observatons for the banks that fal durng the frst quarter n 1984. Snce supervsors cannot dscrmnate between sound and healthy banks ex-ante, we addtonally nclude non-faled nsttutons nto the duraton model. Usng quarterly data obtaned from Call Reports, we can draw upon a large dataset of more than 456,000 bankquarter observatons for more than 13,000 banks. The rchness of the dataset gves our test consderable statstcal power. The set of nsttutons s sampled untl 1996 when the last bank remanng n the dataset fals or censorng takes place. The mnmum duraton s therefore t=8 f the bank faled n the frst quarter of 1984 and the maxmum duraton s t=56 f the nsttuton faled n the last quarter 1996. The choce of a duraton model s also drven by polcy consderatons: knowledge of the factors that drve tme to falure of banks helps obtan better estmates of when the losses wll occur to the depost nsurer. Ths enables the nsurance fund to adjust resources more effectvely. 15 15 Oshnsky and Oln (2005) provde an n-depth analyss of the factors that determne whether troubled nsttutons recover, merge, contnue as a problem bank or eventually fal. They report that the Offce of - 13 -
4. Emprcal Results We report the results for the analyss of the effect of fundng structure on the loss rate n Secton 4.1 and dscuss the mpact of fundng structure on tme to falure n Secton 4.2. 4.1 Bank fundng structure and cost of falure Table 2 presents the results obtaned usng OLS regresson to enable comparson wth prevous studes. We estmate fve setups for the loss equaton. Specfcaton (1) draws upon a parsmonous set of varables prevously found to be sgnfcant determnants of the depost nsurer s loss. Due to the fact that the regulatory envronment regardng resoluton of bank falures changed consderably wth enactment of FDICIA n 1991, we also nclude a dummy varable that takes on the value one f the falure occurred n the perod after FDICIA or zero otherwse. Measures for lablty structure are ntroduced n Specfcaton (2). Addtonal control varables are used n Specfcatons (3), (4), and (5) to test for possble omtted varable bas. The varable that captures bank sze, log of total assets, s adjusted for nflaton usng the GDP deflator. Specfcaton (1) confrms fndngs by Osterberg and Thomson (1994) and McDll (2004) that hgher levels of other real estate owned ncrease FDIC losses. Smlarly, uncollected ncome, as reported n many prevous studes (e.g. James, 1991; Osterberg and Thomson, 1994; McDll, 2004) also enters wth a postve and sgnfcant sgn and so does the rato of loans past due to total assets. Unsurprsngly, larger banks measured by the log of total assets, tend to cause hgher losses to the depost nsurer. Ths s due to the fact that we model the dollar value of losses, nstead of loss rates. 16 The dummy for the perod followng enactment of FDICA enters wth a negatve sgn and assumes sgnfcance at the one percent level. Ths underscores that FDICIA, desgned to reduce losses arsng from bank falures, helped decrease the losses born by the nsurance fund (see also Benston and Kaufman, 1997). In Specfcaton (2) the rato of transactons deposts to total deposts enters wth a negatve sgn and s hghly sgnfcant. Ths fndng can be explaned by the fact that transactons deposts resemble core deposts, often used as a proxy for the franchse value of fnancal nsttutons. Ths result s algned wth Osterberg and Thomson (1994) and James (1991). Addtonally, the ratos of tme and savngs deposts, and demand deposts to total assets also enter wth a negatve and sgnfcant sgn. It s mportant to note that these depost categores also contan jumbo CDs and large money market depost accounts, whch are typcally not nsured by the FDIC. Thus, to the extent to whch these deposts are not nsured, they decrease FDIC costs. Ths could be due to the fact that holders of nsured deposts wthdraw large volumes n the run-up to a falure as reported by Davenport and 16 the Comptroller of the Currency (OCC) hghlghts relance on volatle labltes as mportant cause of bank falure. However, Oshnsky and Oln s (2005) emprcal analyss suggests that falng banks do not experence ncreases n volatle labltes. Ths result may be due to regulatory reasons. The Federal Depost Insurance Corporaton Improvement Act of 1991 (FDICIA) restrcts the use of brokered deposts of crtcally undercaptalzed depostores. Thus, explorng the nexus between lablty structure and ts mplcatons for the tmng of falure of alng banks s a frutful avenue for research. Note however, that there s an nverse relatonshp between loss rates snce t s well establshed that larger banks have lower loss rates (e.g. Shbut, 2002). - 14 -
McDll (2006), resultng n a large proporton of these types of deposts beng unnsured at the tme of falure. Insured depostors wthdrawals are due to a lqudty effect. They may be concerned about delayed redempton of ther holdngs followng falure. The rato of other labltes to total assets also enters sgnfcantly wth a negatve sgn. Ths result s fully algned wth theory as ths varable conssts of other, not nsured labltes n the faled nsttutons. None of the other regressors that capture bank fundng structure becomes sgnfcant n ths setup. In partcular, we do not fnd a sgnfcant role of brokered deposts, a fndng that contrasts wth Osterberg and Thomson (1994) who contend that brokered deposts are a source of market dscplne. Our result may be due to the longer samplng horzon n the present study: crtcally undercaptalzed nsttutons face restrctons regardng the use of brokered deposts snce FDICIA became effectve and therefore may not be able to make extensve use of ths type of fundng. In addton, Specfcaton (2) also suggests that asset growth over eght quarters pror to falure ncreases losses, whereas the level of captalzaton enters now sgnfcantly wth a negatve sgn. Both results are algned wth prevous studes. In terms of the magntude of the coeffcents, the proxy for uncollected ncome domnates the other coeffcents, ths s consstent wth the results obtaned by McDll (2004), Osterberg and Thomson (1994) and Bovenz and Murton (1988). Controllng for addtonal varables n Specfcaton (3) does not change our nferences. We fnd that C&I loans, agrcultural loans, real estate loans, and ndvdual loans all have sgnfcant bearng for the depost nsurer s loss. Specfcaton (4) furthermore ncludes a dummy varable for the effect of depostor preference law, to test whether the law meets ts objectve of decreasng resoluton costs. The dummy takes on the value one f depostor preference law was n place at the tme of falure or zero otherwse. The varable enters wth the antcpated negatve sgn, but t remans nsgnfcant. McDll (2004) has shown that a sluggsh macroeconomc envronment plays and mportant role for explanng FDIC losses. Therefore, Specfcaton (5) consders the effect of the macroeconomc settng and ncludes addtonal varables that capture nformaton on personal ncome growth (lagged by two perods), bankruptcy growth, and unemployment rates on the state level. The results suggest that bankruptcy growth and the unemployment rate sgnfcantly ncrease losses. Whle controllng for the effect of the macroeconomc envronment renders the rato of transacton deposts to total assets nsgnfcant, t hghlghts a weakly negatve assocaton of the rato of Fed funds to total assets wth the depost nsurer s loss. Ths result s ntutve: Fed funds are unnsured labltes and therefore decrease FDIC losses. The adjusted R 2 and the Akake Informaton Crteron (AIC) ndcate that Specfcaton (5) s the most approprate setup for the model. As alluded to prevously, estmates obtaned from the OLS regresson only approxmate the central tendency of the dstrbuton and are unsutable to account for heterogeneous data wth outlers. Furthermore, depost nsurers and bank supervsors are partcularly concerned about hgh-cost falures and have therefore a vested nterest n the factors drvng losses of those costly falures. We therefore employ quantle regresson models that am to obtan better estmates for the determnants of the factors for hgh-cost falures. We present the results usng quantle regresson estmators n Table 3. In order to evaluate the effect of our explanatory varables at dfferent quantles of the dstrbuton on the loss varable, we estmate quantle regresson models to obtan - 15 -
coeffcents for the 5 th, 10 th, 25 th, 50 th, 75 th, 90 th, and 95 th quantle. The estmaton s based on the regresson setup of Specfcatons (5) n Table 2. Ths regresson setup ncludes addtonal control varables for the composton of the faled banks loan portfolos, takes account of depostor preference law, and also ncludes varables that capture nformaton from the macroeconomc envronment. We report the results n Table 3 and also nclude the coeffcents obtaned wth the OLS estmator for comparablty. In order to further llustrate the use of quantle regresson, Fgures 1 a) 1 v) plot the estmated coeffcents of nterest obtaned wth the quantle estmator aganst the dfferent quantles as the sold curve. These pont estmates can be nterpreted as the mpact of a one-unt change of the regressor on the loss varable wth the other covarates held constant. The vertcal axs ndcates the effect of the regressor and the horzontal lne represents the quantle θ scale. The gray shaded area shows a 95 percent confdence band based on bootstrapped standard errors for the quantle estmates and the dashed lne represents the OLS estmator. The plots also contan a 95 percent confdence band for the OLS estmator. [TABLE 3] [FIGURE 1 a) v)] Table 3 provdes an llustraton of the dfferences n magntude, sgnfcance and change n drecton of the relatonshp between the loss varable and our regressors as we move along the dstrbuton. Bank sze, measured by the deflated log of total assets, shows a stronger effect on FDIC loss for more costly falures and remans sgnfcant across all quantles. As mentoned prevously, larger banks tend to cause larger losses to the depost nsurer, ths ndcates a sze-effect. Moreover, plot 1 a) llustrates that the coeffcent obtaned n the quantle regresson procedure clearly departs from the coeffcent resultng from the OLS estmator, hghlghtng that relance on OLS estmates can easly gve rse to napproprate nferences. Fgure 1 b) also shows a hghly nonlnear relatonshp between the rato of real estate owned to total assets and ts mpact upon FDIC loss. Exhbtng sgnfcance across all quantles, the effect s decreasng steadly. Snce real estate owned captures the level of foreclosed real estate owned by faled nsttutons, ths fndng ndcates that the varable s less mportant n explanng hgh-cost falures. Such falures may be more strongly nfluenced by other determnants, e.g. composton of the loan portfolo. The mpact of equty to total assets s sgnfcant and negatve across all quantles but the upper tal of the dstrbuton. Ths effect does not vary much n terms of magntude as the quantle regresson estmator remans wthn the confdence band of the OLS estmator. In a smlar ven to the results obtaned n Specfcaton (5) presented n Table 2, the rato of loans past due to total assets remans nsgnfcant n the quantle regressons. Whle the rato of uncollected ncome to total assets retans ts hghly sgnfcant effect on the loss varable across all quantles, t also remans wthn the confdence band of the OLS estmator n Fgure 1e), suggestng no marked dfferences between the mpact on hgh and low-cost falures. Lkewse, Fgure 1 f) shows that asset growth over the 24 months pror to falure remans sgnfcant across all quantles. However, ths varable exhbts not much varaton as we move up the dstrbuton. - 16 -
In terms of the lablty structure, the quantle regressons generally do not suggest that varables that capture lablty structure have a varyng effect (n terms of magntude) upon hgh-cost and low-cost falures. Accordng to the quantle regresson estmator, the rato of Fed funds to total assets s weakly and negatvely assocated wth losses for low-cost and moderate-cost falures, whereas the effect s rendered nsgnfcant as we move up the dstrbuton. The negatve effect of Fed funds on losses s algned wth the statement by Osterberg (1996), who argues that Fed funds are hghly lqud and that falng banks able to borrow such funds wll have lower resoluton costs. The ratos of tme and savngs as well as demand deposts to total assets also sgnfcantly decrease FDIC loss for moderate-cost falures as the large proporton of these types of deposts s unlkely to be nsured at the tme of falure (Davenport and McDll, 2006). Lkewse, the rato of other labltes to total assets has a weakly decreasng effect on moderate-cost falures. Ths suggests that these varables are only of moderate mportance for explanng hgh-cost falures. Moreover, gven the wdth of the confdence ntervals for the quantle regresson estmators, cauton has to be exercsed when drawng nferences. The rato C&I loans to total assets s postve and sgnfcant across all quantles. Fgure 1 n) llustrates a consderable departure of the quantle estmator from the OLS estmator at the lower and upper tal of the dstrbuton. Ths suggests that C&I loans have a much stronger effect on low-cost falures than on hgh-cost falures, ndcatng marked dfferences n the loan portfolos between hgh-cost and low-cost falures. The rato of agrcultural loans to total assets n Fgure 1 o) exhbts a smlar pattern, ndcatng a stronger effect of ths type of loan on losses caused by low-cost falures than on moderate and hgh-cost falures. Fgure 1 p) challenges the result obtaned wth the OLS estmator n Specfcaton (5) n Table 2. The quantle regresson estmator underscores that ths varable markedly ncreases losses for low and moderate-cost falures, whereas t s only nsgnfcant at the upper tal of the dstrbuton. Fgure 1 q) plots the rato of real estate loans to total assets and also resembles the behavour of the other varables that capture the loan portfolo. Real estate loans also play an mportant role for low-cost falures whereas they do not sgnfcantly affect hgh-cost falures. Whle the lagged personal ncome growth varable s only sgnfcant at the medan and at the upper tal of the loss varable, the unemployment rate s sgnfcant across all quantles. Bankruptcy growth, however, s only of relevance for the costly falures. The ncreasng effect of bankruptcy growth and unemployment rates on costly-falures s not surprsng. In states where the economy s performng poorly, defaults of ndvdual and corporate borrowers wll also adversely affect other banks. McDll (2004) has shown that the pool of potental buyers of faled nsttutons s a further key determnant for the depost nsurer s loss. Thus, f the pool of potental buyers s operatng n the same economc envronment than the faled nsttuton, these nsttutons may be restraned n ther ablty to pay hgh prces for the faled bank s assets. Ths wll not only affect recovery rates but t wll ultmately adversely affect FDIC s loss as well. Fgure 1 u) hghlghts a nonlnear relatonshp between depostor preference law and the loss varable. Whle the OLS model ndcates no ndependent effect of depostor preference law on FDIC loss, our quantle regressons underscore that depostor preference law sgnfcantly decreases falure cost only at the lower tal of the dstrbuton. Ths mples that the law meets the objectve of decreasng FDIC loss exclusvely for lowcost falures. Ths fndng may be affected by the way a faled nsttuton s resolved. In - 17 -
nstances where an asssted merger (or purchase and assumpton transacton) took place, all depostors may have been treated as f they were nsured so that the effect of the law was lmted. By contrast, f the FDIC lqudated the faled bank and pad off depostors, the law mght have lved up to ts expectatons. 17 Fnally, the FDICIA dummy enters across all quantles wth a negatve sgn. Ths coeffcent remans wthn the OLS confdence band and does not exhbt much varaton as we move up the dstrbuton. W hle vsual nspecton of the ndvdual plots n Fgure 1 already suggests mportant nonlnear relatonshps between FDIC losses and several explanatory varables, the followng secton presents addtonal tests to valdate these nferences. Frst, we run F-tests to nvestgate f the coeffcents are jontly statstcally dfferent from zero across all quantles for each varable. Second, we test f the coeffcents at the medan and at the tals are jontly sgnfcantly dfferent from zero to evaluate f medan FDIC losses are affected dfferently by the varables from the losses n the tals of the dstrbuton of the dependent varable. Thrd, F-tests are also utlzed to nvestgate f there are sgnfcant dfferences for each coeffcent n the tals of the dstrbuton. The latter test helps ascertan f there are dfferences n the factors that drve n partcular hgh-cost and low-cost falures. [TABLE 4] The nferences from our vsual nspecton are corroborated by the addtonal tests. The F- tests reject the null hypothess for the equalty of the coeffcents across all quantles n sx nstances at the one and at the fve percent level. The results further mprove when testng the null hypothess that the coeffcents are jontly dfferent from zero between the medan and the tals (5 th and 95 th quantle) of the dstrbuton. The F-tests ndcate n seven nstances statstcal sgnfcance. Fnally, the results are even stronger when we focus on the tals of the dstrbuton (5 th and 95 th quantle) to establsh whether the explanatory varables dfferently mpact hgh-cost and low-cost falures. The F-tests suggest a statstcally sgnfcantly dfferent mpact for eght varables. In partcular, bank sze, and the rato of real estate owned to total assets exhbt varyng mpact on hgh-cost and low-cost falures. Whereas the effect of lablty structure does not appear to have varyng effects on the depost nsurer s loss, varables that capture composton of the loan portfolo do. Ths may be explaned as follows: Whle lablty structure determnes whch credtors have to be compensated and cannot be affected ex-post the falure date, the dependent varable n the cost equatons s affected after the falure date snce t s also nfluenced by the recovery value of the assets n the recevershp books. Our fndng may reflect that alng nsttutons overstate asset values n the run-up to falure, whch would explan ths varyng effect. In partcular, expensve falures may be only uncovered ex-post. In summary, our results provde clear emprcal evdence for the mportant bearng of certan types of labltes on the depost nsurer s loss. Ths fndng s consstent wth the asserton by Shbut (2002) that lablty structure nfluences the depost nsurer s loss snce t determnes whch clamants have to be compensated n case of bank falure. For nstance, Fed funds sgnfcantly decrease low and moderate-cost falures, a fndng that s only observable through the use of quantle regresson estmators. Moreover, the fndngs llustrate that relance on standard econometrc technques to assess the determnants of the depost nsurer s loss can gve rse to msleadng nferences as several explanatory varables 17 See Osterberg (1996) for the lnk between bank resoluton and depostor preference law. - 18 -
exhbt a varyng effect on FDIC loss. The observed non-lneartes are not surprsng: faled depostores exhbt dfferent characterstcs regardng bank type, busness actvtes and sze that all affect the loss varable. The proposed quantle regresson estmators accommodate the heterogenety of the data and offer more detaled nsghts nto the factors drvng FDIC losses across the dstrbuton. Ths s of partcular mportance for determnng how the explanatory varables nfluence hgh-cost falures. Whle we fnd that varables that capture loan portfolo composton have a strong dscrmnatory effect for low-cost and hgh-cost falures, no such effect can be establshed for the varables that capture lablty structure. In partcular, costly falures are largely determned by C&I loans, uncollected ncome and asset growth. In addton, costly falures are furthermore affected by a weak macroeconomc envronment. 4.2 Bank fundng structure and tme to falure We employ the AFT model n ths secton to test the effect of bank fundng structure on tme to falure. Whle prevous studes nvestgate the prce and quantty effects of rsk on bank fundng structure (e.g. Park and Perstan, 1998; Maechler and McDll, 2006), the nexus between bank fundng structure and tme to falure s an alternatve way of assessng the role of market dscplne. In addton, wth the excepton of the work by Davenport and McDll (2006), the relatonshp between market dscplne and depost nsurance by type of account has been left largely untouched n the lterature. These questons have ganed ncreasng promnence wth the advent of the new Basel Captal Accord. For nstance, Maechler and McDll (2006) argue that very rsky nsttutons cannot ncrease the volume of nsured deposts by offerng hgher nterest rates to compensate outflows of unnsured deposts. Thus, troubled banks that rely heavly on unnsured deposts mght fal faster due to ther nablty to substtute such cash outflows wth other types of funds. Ths may be nterpreted as a sgnal for the presence of market dscplne and underscores the mportance of Pllar 3 n the new Basel Captal Accord. Furthermore, evdence that certan types of account holders wthdraw deposts n the perod pror to falure would present evdence that such account holders are also senstve to the bank s fnancal condton. Table 5 presents the results of our duraton analyss whereby we use data for the faled nsttutons that also underle the cost equatons. 18 As hghlghted n Secton 3.2, ths dataset also ncludes non-faled nsttutons to avod problems arsng from sample selectvty. Ths s due to the fact that regulators cannot dscrmnate ex-ante between faled and non-faled nsttutons. The set of explanatory varables ncludes regressors that capture lablty structure, composton of the loan portfolo, varables that provde nformaton on the macroeconomc envronment on the federal state level, and a dummy that takes on the value one f the observaton s from the perod followng enactment of FDICIA or zero otherwse. We also nclude a dummy that takes on the value one f the federal state had depostor preference law n place at the tme of the observaton or zero otherwse. The log of total assets s ncluded to control for bank sze. We augment these specfcatons by 18 Note that avalablty of the explanatory varables for the AFT model slghtly reduces the number of faled nsttutons for our analyss of depostor dscplne n comparson to the cost-equatons. - 19 -
addtonal varables to capture nformaton that s commonly used by bank supervsors to predct falures of depostores. Bank supervsors refer to these varables as CAMEL 19 varables. We use the rato captal to total assets as a proxy for the level of captalzaton. The varable troubled assets s calculated as the sum of real estate owned and loans past due over total assets to measure asset qualty. The rato of operatng ncome to total assets s used as proxy for earnngs. We capture the effect of lqudty wth a varable that s calculated as the sum of securtes and cash over total assets. We do not nclude a proxy for management qualty as ths would requre detaled qualtatve nformaton not contaned n publcly avalable Call Reports. The results of the AFT models are to be nterpreted as follows: A postve coeffcent ndcates a deceleratng effect of the varable on tme to falure whereas a negatve coeffcent ndcates shortened survval tme. Specfcaton (1) n Table 5 s our canoncal model that only uses a parsmonous set of varables based on prevous studes of bank falure, augmented by the FDICIA dummy. Four of the sx regressors are sgnfcant at the one percent level and show the antcpated sgn. Unsurprsngly, banks wth a large proporton of troubled assets tend to fal faster. The proxes for captalzaton and lqudty enter the equaton wth a postve sgn, ndcatng ncreased survval tme. Clearly, better captalzed banks are better able to absorb shocks. Smlarly, more lqud nsttutons are n a better poston to accommodate sudden cash outflows than less lqud depostores. The postve coeffcent of the proxy for bank sze ndcates that larger banks exhbt ncreased survval tme. Ths may be due to ther ablty to reap benefts from dversfcaton of rsk. We ntroduce varables that capture fundng structure n Specfcaton (2). Ths augmented specfcaton not only underscores a consderable mpact of fundng structure on tme to falure, but also hghlghts the presence of omtted varable bas n Specfcaton (1). The magntude of our measure for captalzaton used n the canoncal model declnes approxmately 20 percent upon controllng for lablty structure n Specfcaton (2). The rato of Fed funds to total assets enters the equaton at the one percent level wth a negatve sgn. Ths result emprcally substantates that banks funded by such deposts tend to fal faster. Ths s algned wth research on market dscplne: Fed funds are not nsured and holders of unnsured clams tend to wthdraw ther funds from alng nsttutons as documented n several studes (e.g. Goldberg and Hudgns, 1996, 2002; Jordan, 2000; Davenport and McDll, 2006). In ths respect, our fndngs can be nterpreted as emprcal evdence for the presence of market dscplne. Cash outflows n serously troubled banks may not longer be offset by ether substtutng unnsured deposts or by offerng hgher nterest rates. Indeed, Maechler and McDll (2006) show that very weak banks,.e. banks wth CAMEL ratngs 4 or 5, face severe constrants n offsettng declnes n unnsured deposts by offerng hgher nterest rates. Ths ndcates a potentally non-lnear relatonshp between bank rsk and the cost of unnsured funds. Furthermore, banks obvously tryng to crcumvent market dscplne mght attract addtonal regulatory scrutny and regulators may be ultmately forced to act and close these nsttutons faster. 19 CAMEL s an acronym for components of the regulatory ratng system employed to assess soundness of fnancal nsttutons: Captal adequacy, Asset qualty, Management, Earnngs, and Lqudty. The ratng system has been augmented n 1997 by addng a component that captures Senstvty to market rsk. The system s therefore now referred to as CAMELS ratng system. The ratngs assgned to banks range from 1 5, whereby 1 denotes a sound nsttuton and banks rated 5 are consdered extremely rsky and unsound. - 20 -
The rato of transactons deposts to total assets enters postvely at the one percent level, suggestng that more transactons deposts tend to ncrease tme to falure. Ths result can be agan explaned wth the fact that transactons deposts are a proxy for a bank s charter value. Hgher bank charter values are lkely to encourage prudent behavor by bank managers, thus curtalng rsk-takng behavor and ncreasng tme to falure. The rato of brokered deposts to total assets enters the equaton negatvely at the one percent level. It s well documented that lablty shftng occurs pror to the falure of depostores (e.g. Marno and Bennett, 1999). Although FDICIA lmts the use of brokered deposts by crtcally undercaptalzed banks, nsttutons not subject to ths classfcaton may be nevertheless able to turn to brokered deposts to compensate outflows of other types of deposts. To the extent to whch these deposts are not nsured, they are not prced accordng to the borrower s default rsk. Thus, use of brokered deposts can be nterpreted as evdence of dstress such that the regulator s propensty to close a troubled bank faster ncreases. Smlarly, the ratos tme and savngs deposts, and demand deposts to total assets adversely mpact upon tme to falure and assume statstcal sgnfcance at the one percent level. These varables capture both nsured and unnsured deposts. Thus, to the extent they capture unnsured deposts such as jumbo CDs, the results ndcate that unnsured depostors wthdraw ther funds n the run-up to falure. Whle ths may appear counterntutve at frst glance, there are arguments that can explan ths result. Frst, recent evdence by Davenport and McDll (2006) suggests that the majorty of deposts wthdrawn n from alng banks are fully nsured. Second, Park and Perstan (1998) underscore that nsured depostors may be unwllng to supply funds to troubled banks f they become aware of an mpendng falure. For nstance, they argue that even nsured depostors may be reluctant to supply funds to alng nsttutons, whch, n turn, could accelerate tme to falure. Park and Perstan (1998) also fnd adverse effects of bank rsk on the prcng and growth of nsured deposts and propose that nsured depostors may be concerned about the nsurer s solvency or try to avod other ndrect costs arsng from the delay n depost redempton after falure. Thrd, Jordan (2000) reports that declnes n large CDs n falng banks n New England n the 1990s were more than offset by ncreases n small CDs. Thus, lablty shftng from unnsured to nsured deposts also plays a crucal role n explanng the nverse relatonshp between tme and savngs deposts and tme to falure. The rato of subordnated debt to total assets exhbts a postve and sgnfcant sgn suggestng that relance on subordnated debt ncreases survval tme. Ths may be explaned wth a sgnalng effect: an nsttutons ablty to attract subordnated debt could ndcate that they are less rsky and that large and sophstcated debt holders whch advanced montorng abltes are wllng to lend to these nsttutons. Moreover, the typcally longer maturty of these labltes means that these labltes cannot ext the bank at short notce. The remanng category that captures other labltes to total assets enters negatvely and sgnfcantly at the one percent level. These labltes are not nsured and credtors have therefore an ncentve to obtan ther funds pror to falure. All our control varables found to be sgnfcant n Specfcaton (1) reman sgnfcant n Specfcaton (2). To test robustness of these results, we nclude addtonal control varables n Specfcaton (3), (4) and (5). In Specfcaton (3), we addtonally employ several varables that capture - 21 -
composton of the loan portfolo. Whle controllng for addtonal varables decreases the magntude of several coeffcents, our results regardng the fundng structure are robust. Among the addtonal control varables, the ratos of C&I loans and agrcultural loans to total assets enter wth a sgnfcant and negatve coeffcent, suggestng lendng n these areas shortens survval tme. Depostor preference law, ncluded n Specfcaton (4), ncreases tme to falure sgnfcantly. Ths may be due to depostor s lower propensty to run when such law s n place. Fnally, Specfcaton (5) ndcates that the macroeconomc envronment has ndeed some bearng on falure tme. As antcpated, a weaker macroeconomc settng, reflected n hgher bankruptcy growth rates and hgher rates of unemployment shortens tme to falure of banks whereas an economc upswng, proxed by personal ncome growth, wll ncrease tme to falure. Both the log lkelhood functon and the Akake Informaton Crteron ndcate that Specfcaton (5) s the most approprate setup for our AFT model. To verfy our results, we perform an addtonal robustness test by examnng as to whether the tmng of the onset of rsk mpacts our nferences. Ths addtonal test redefnes the onset of rsk for each nsttuton to be the perod when the rato of equty captal to total assets falls below eght percent. 20 The results are vrtually dentcal to those reported n Table 5 and we therefore do not report them. 21 In sum, the fndngs from our AFT model provde emprcal evdence that controllng for the lablty structure when estmatng tme to falure ncreases the explanatory power of the presented model. Our results ndcate the presence of market dscplne: unnsured labltes such as Fed funds decrease tme to falure. In addton, our fndngs are suggestve for a substtuton effect of unnsured deposts wth nsured labltes such as brokered deposts. Smlarly, tme and savngs deposts, and demand deposts are found to adversely mpact survval tme of fnancal nsttutons. These results suggest that not only holders of unnsured credts such as Fed funds but also nsured depostors are a source of market dscplne. In terms of polcy mplcatons, the fndngs suggest that lablty structure deserves more attenton by regulatory bodes. Montorng of the behavor of certan types of deposts can provde better nsghts nto tme to falure of fnancal nsttutons. Moreover, applyng captal charges to labltes that tend to leave a bank faster mght curb depostores rsktakng behavor. Pllar 3 of the new Basel Captal Accord currently neglects dsclosure of nsured and unnsured deposts. 22 In lght of our fndngs, dsclosng the levels of nsured and unnsured deposts to the publc may further enhance market dscplne. 20 21 22 The eght percent rato s chosen due to reflect that prompt correctve acton captal gudelnes n FDICIA necesstate regulatory acton such as ncreased montorng or restrctons on asset growth when the (rsk-based) captal rato falls below eght percent. The addtonal results may be obtaned upon request. Nether the Consultatve Document Pllar 3 (Market Dscplne), (Basel Commttee on Bankng Supervson, 2001a), nor the Workng Paper on Pllar 3 Market dscplne, (Basel Commttee on Bankng Supervson, 2001b) menton dsclosure rules wth respect to fnancal nsttutons lablty/depost structure regardng ther status of depost nsurance. Ths nsuffcent consderaton of bank lablty structure n the context of market dscplne n general and depost nsurance n partcular s also documented n Pennacch (2005), who underscores that the Thrd Consultatve Paper on the New Basel Captal Accord (Basel Commttee on Bankng Supervson, 2003) contans no reference to depost nsurance. - 22 -
5. Concludng Remarks Ths paper analyses the extent to whch bank lablty structure mpacts on the depost nsurer s loss n case of falure of ndvdual fnancal nsttutons and how bank lablty structure affects tme to falure. These questons are pertnent to the estmaton of loss gven default snce depostores lablty structure not only determnes whch depostors have to be compensated n case of falure but also mpacts upon fnancal nsttutons rsktakng behavor. Usng quantle regresson analyss that permts takng account of the non-normal dstrbuton of the depost nsurer s losses ncurred from bank falures, we explore how the depost nsurer s loss vares across the dstrbuton and llustrate ts senstvty towards several explanatory varables across dfferent quantles. Ths examnaton s benefcal for bank regulators, supervsory agences and depost nsurers as they are partcularly concerned about hgh-cost falures. Our analyss extends prevous work n that t presents emprcal evdence for non-lnear relatonshps between losses and a number of explanatory varables. To that extent, our fndngs hghlght the shortcomngs assocated wth standard econometrc technques due to the better use of the nformaton n the sample dstrbuton. The dscovered non-lneartes are not surprsng: faled depostores exhbt dfferent characterstcs regardng bank type, busness actvtes and sze that all mpact upon the depost nsurer s loss. In partcular, we present evdence that losses are not homogeneously drven by the same set of determnants. C&I loans, uncollected ncome, and a weak macroeconomc envronment are man determnants for very costly bank falures. Investgatng the nexus between lablty structure and tme to falure, we offer evdence for the presence of depostor dscplne: unnsured labltes such as Fed funds decrease tme to falure. Brokered deposts, tme and savngs, and demand deposts are also found to adversely mpact survval tme of fnancal nsttutons. To the extent to whch nsured deposts decrease survval tme, we assgn ths fndng to market dscplne arsng from nsured depostors and to lablty shftng of troubled banks. These results are robust to controllng for numerous covarates that capture bank asset qualty and the composton of the faled nsttutons loan portfolo. Furthermore, performng a senstvty check that redefnes the onset of rsk for the banks n the sample yelds vrtually dentcal results. Fnally, the results from our AFT model provde emprcal evdence that consderaton of bank lablty structure when estmatng tme to falure of fnancal nsttutons ncreases the explanatory power of the presented model. The fndngs regardng tme to falure bear mportant polcy mplcatons. If banks that are heavly relant on short-term and unnsured funds tend to fal faster, there s a case to make them subject to addtonal measures of prompt correctve acton to lmt ther ablty to substtute unnsured deposts wth nsured deposts, thereby ncreasng the depost nsurer s loss gven default. The montorng of alng fnancal nsttutons should therefore be extended to ther use of certan types of deposts. Moreover, whle Pllar 3 n the Basel II framework underscores dsclosure as an ntegral component to enhance market dscplne, t wdely gnores fnancal nsttutons lablty structure. Thus, our fndngs ndcate that dsclosure of the levels of nsured and unnsured deposts could further strengthen - 23 -
depostor dscplne. In addton, captal charges may be approprate for certan types of labltes to polce nsttutons aganst rsk takng behavor. One caveat remans. Snce quantle regresson condtons on the dependent varable ts use as a predctve tool s lmted. Nonetheless, ths study ponts out that there exst systematc dfferences between the factors that drve hgh-cost and low-cost falures. Thus, to that extent, our results suggest closer montorng of certan categores of the loan portfolo of weak depostores to mtgate the losses that wll arse to the depost nsurer when these alng nsttutons eventually fal. Our analyss focuses on the non-lnear effect of certan varables on the depost nsurer s loss and on the mpact of lablty structure on tme to falure. Future research could buld on these results and examne the lnk between tme to falure and the loss varable and evaluate the mplcatons for the regulatory envronment n greater detal. - 24 -
References Ashcraft, A. (2003) Are banks really specal? New evdence from the FDIC-nduced falure of healthy banks. Federal Reserve Bank of New York Staff Reports No. 176, December 2003 Barth, J. R., Phlp F. Bartholomew, and Mchael G. Bradley (1990) Determnants of thrft nsttuton resoluton costs. Journal of Fnance, Vol. 45, pp. 731-754 Basel Commttee on Bankng Supervson (2003) The New Basel Captal Accord. Techncal report thrd consultatve paper. Basel: Bank for Internatonal Settlements Basel Commttee on Bankng Supervson (2001a) Consultatve document: Pllar 3 (Market Dscplne). Basel: Bank for Internatonal Settlements Basel Commttee on Bankng Supervson (2001b) Workng Paper on Pllar 3 - Market Dscplne. Basel: Bank for Internatonal Settlements Bennett, R. L., Mark D. Vaughan and Tmothy. J. Yeager (2005) Should the FDIC worry about the FHLB? The mpact of Federal Home Loan Board Advances on the Bank Insurance Fund. FDIC CFR Workng Paper 2005-10 Benston, G., and George G. Kaufman (1997) FDICIA after fve years. Journal of Economc Perspectves, Vol. 11, pp. 139-158 Bllet, M. T., Jon A. Garfnkel and Edward S. O Neal (1998) The cost of market versus regulatory dscplne n bankng. Journal of Fnancal Economcs, Vol. 48, pp. 333-358 Blalock, J. B., Tmothy J. Curry and Peter J. Elmer (1991) Resoluton costs of thrft falures. FDIC Bankng Revew, Vol. 4, pp. 15-26 Blum, J. (2002) Subordnated debt, market dscplne, and banks rsk takng. Journal of Bankng and Fnance, Vol. 26, pp. 1427-1441 Bovenz, J. F. and Arthur J. Murton (1988) Resoluton costs of bank falures. FDIC Bankng Revew, Vol. 1, pp. 1-13 Brown, R. A. and Seth Epsten (1992) Resoluton costs of bank falures: An update of the FDIC hstorcal loss model. FDIC Bankng Revew, Vol. 5, pp. 1-16 Cade, B. and Barry R. Noon, (2003) A gentle ntroducton to quantle regresson for ecologsts. Fronters n Ecology and the Envronment, Vol. 1, pp. 412-420 Cole, R. A. and Jeffery W. Gunther (1995) Separatng the Lkelhood and Tmng of Bank Falure. Journal of Bankng and Fnance, Vol. 19, pp. 1073 1089 Davenport, A. M. and Kathleen M. McDll (2006) The depostor behnd the dscplne: A mcro-level case study of Hamlton Bank. Journal of Fnancal Servces Research, Vol. 30, pp. 93-109 DeYoung, R. (2003) The falure of new entrants n commercal bankng markets: a spltpopulaton duraton analyss. Revew of Fnancal Economcs, Vol. 12, pp. 7-33 - 25 -
Flannery, M. J. and Sorn M. Sorescu (1996) Evdence of bank market dscplne of subordnated debenture yelds: 1983 1991, Journal of Fnance, Vol. 51, pp. 1347-1377 Goldberg, L. G. and Sylva C. Hudgns (2002) Depostor dscplne and changng strateges for regulatng thrft nsttutons. Journal of Fnancal Economcs, Vol. 63, pp. 263 274 Goldberg, L. G. and Sylva C. Hudgns (1996) Response of unnsured depostors to mpendng S&L falures: evdence of depostor dscplne. Quarterly Revew of Economcs and Fnance, Vol. 36 (3), pp. 311 325 Gup, B. E. (1995) Targetng Fraud: Uncoverng and deterrng fraud n fnancal nsttutons. Chcago: Probus Hrschhorn, E. and Davd Zervos (1990) Polces to change the prorty of clamants: the case of depostor preference laws. Journal of Fnancal Servces Research, Vol. 4, pp. 111 125 James, C. (1991) The losses realzed n bank falures. Journal of Fnance, Vol. 46, pp. 1223-1242 Jordan, J. S. (2000) Depostor dscplne at falng banks. New England Economc Revew, March/Aprl 2000, pp. 15 28 Kng, T. B., Danel A. Nuxoll, and Tmothy J. Yeager (2006) Are the causes of bank dstress changng? Can researchers keep up? Federal Reserve Bank of St. Lous Revew, January/February 2006, pp. 57-80 Koenker, R. and Glbert Bassett (1978) Regresson quantles. Econometrca, Vol. 46, pp. 33 50 Koenker, R. and Kevn F. Hallock (2001) regresson. Journal of Economc Perspectves, Vol. 15, pp. 143-156 Lane, W. R., Stephen W. Looney, and James W. Wansley (1986) An Applcaton of the Cox Model to Bank Falure. Journal of Bankng and Fnance, Vol. 10, pp. 511-531 Maechler, A. M. and Kathleen M. McDll (2006) Dynamc depostor dscplne n US banks. Journal of Bankng and Fnance, Vol. 30, pp. 1871-1891 Marno, J. A. and Rosalnd L. Bennett (1999) The consequences of natonal depostor preference. FDIC Bankng Revew, Vol. 12, pp. 19-38 McDll, K. M. (2004) Resoluton costs and the busness cycle. FDIC W orkng Paper 2004-01 Oshnsky, R. (1999) Effects of bank consoldaton on the bank nsurance fund. FDIC Workng Paper 1999-03 Oshnsky, R. and Vrgna Oln (2005) Troubled banks: Why don t they all fal? FDIC Workng Paper 2005-03 Osterberg, W. P. (1996) The mpact of depostor preference laws. Federal Reserve Bank of Cleveland Economc Revew, Vol. 12, pp. 1 12-26 -
Osterberg, W. P. and James B. Thomson (1994) Underlyng determnants of closed-bank resoluton costs. Federal Reserve Bank of Cleveland, Workng Paper 9403 Park, S., and Stavros Perstan (1998) Market dscplne by thrft depostors. Journal of Money, Credt, and Bankng, Vol. 30, pp. 347 364 Pennacch, G. G. (2005) Rsk-based captal standards, depost nsurance, and procyclcalty. Journal of Fnancal Intermedaton, Vol. 14, pp. 432 465 Powell, J. L. (1986) Censored regresson quantles. Journal of Econometrcs, Vol. 32, pp. 143 155 Powell, J. L. (1984) Least absolute devatons estmaton for the censored regresson model. Journal of Econometrcs, Vol. 25, pp. 303 325 Shbut, L. (2002) Should bank lablty structure nfluence depost nsurance prcng? FDIC Workng Paper 2002-01 Shbut, L., Tm Crtchfeld and Sarah Bohn (2003) Dfferentatng among crtcally undercaptalzed banks and thrfts. FDIC Bankng Revew, Vol. 15 (2), pp. 1 38 Whalen, G. (1991) A proportonal hazards model of bank falure: An examnaton of ts usefulness as an early warnng tool. Federal Reserve Bank of Cleveland Economc Revew, Vol. 27, pp. 21-31 - 27 -
Data Appendx Depostor Preference Laws State Date effectve Alaska October 15, 1978 Arzona September 21, 1991 Calforna June 27, 1986 Colorado May 1, 1987 Connectcut May 22, 1991 Florda July 3, 1992 Georga 1974 a Hawa June 24, 1987 Idaho 1979 b Iowa January 1, 1970 Kansas July 1, 1985 Lousana January 1, 1985 Mane Aprl 16, 1991 Mnnesota Aprl 24, 1990 Mssour September 1, 1993 Montana 1927 c Nebraska 1909 c New Hampshre June 10, 1991 New Mexco June 30, 1963 North Dakota July 1, 1987 Oklahoma May 26, 1965 Oregon January 1, 1974 Rhode Island February 8, 1991 South Dakota July 1, 1969 Tennessee 1969 c Utah 1983 c Vrgna July 1, 1983 West Vrgna May 11, 1981 a. Legslaton became effectve on ether January 1 or July 1. b. Passed by both houses of the state legslature on July 1; enactment date s unclear. c. Nether the month nor the day of enactment s avalable. SOURCE: Osterberg (1996) - 28 -
Table 1: Descrptve statstcs Varable N Mean Max Mn S.D. p. 5 p. 10 p. 25 Medan p. 75 p. 90 p. 95 Loss on assets 1074 20778.8 2017459.0 0.00 87937.1 566.0 997.0 2212.0 5107.0 12327.0 32780.0 78778.0 Loss/Total assets 1074 0.34 133.15 0.00 4.06 0.03 0.05 0.13 0.21 0.30 0.40 0.46 Total assets 1092 153239.8 17100000.0 1731.0 924287.0 5915.0 8054.0 13778.5 26595.0 60733.5 171578.0 400540.0 Real estate owned/total assets 1092 0.05 0.53 0.00 0.05 0.00 0.00 0.01 0.04 0.07 0.11 0.14 Equty/Total assets 1092-0.01 0.93-0.58 0.07-0.12-0.08-0.03 0.00 0.03 0.06 0.08 Loans past due/total assets 1091 0.02 0.28 0.00 0.03 0.00 0.00 0.00 0.01 0.03 0.06 0.09 Uncollected ncome/total assets 1092 0.01 0.06 0.00 0.01 0.00 0.00 0.01 0.01 0.02 0.02 0.03 Asset growth 1092-0.11 4.33-0.76 0.33-0.45-0.40-0.29-0.17-0.01 0.22 0.40 Fed funds/total assets 1092 0.01 0.34 0.00 0.02 0.00 0.00 0.00 0.00 0.00 0.02 0.05 Demand deposts/total assets 1092 0.15 0.65 0.00 0.07 0.05 0.07 0.10 0.14 0.18 0.24 0.28 Brokered deposts/total assets 1092 0.02 0.85 0.00 0.08 0.00 0.00 0.00 0.00 0.00 0.07 0.15 Transactons deposts/total assets 1088 0.25 0.92 0.00 0.11 0.10 0.13 0.18 0.24 0.31 0.39 0.43 Tme and savngs deposts/total assets 1049 0.83 1.35 0.24 0.10 0.66 0.72 0.78 0.84 0.88 0.93 0.97 Subordnated debt/total assets 1092 0.00 0.04 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Other labltes/total assets 1090 0.02 0.39 0.00 0.04 0.00 0.01 0.01 0.01 0.01 0.04 0.09 C&I loans/total assets 1091 0.17 1.39 0.00 0.12 0.03 0.05 0.09 0.15 0.24 0.33 0.40 Agrcultural loans/total assets 1091 0.07 0.61 0.00 0.12 0.00 0.00 0.00 0.00 0.08 0.23 0.34 Real estate loans/total assets 1092 0.27 0.86 0.00 0.15 0.06 0.09 0.15 0.25 0.35 0.48 0.55 Indvdual loans/total assets 1091 0.12 1.14 0.00 0.10 0.02 0.03 0.05 0.10 0.16 0.24 0.30 Personal ncome growth (lagged) 1092 0.05 0.18-0.08 0.03 0.00 0.01 0.02 0.06 0.08 0.10 0.11 Bankruptcy growth rate 1092 0.16 0.94-0.27 0.19-0.07-0.01 0.04 0.11 0.27 0.43 0.56 Unemployment rate 1092 7.03 17.40 2.80 1.61 4.70 5.10 6.00 6.80 7.80 8.90 9.50 Depostor preference law dummy 1092 0.44 1.00 0.00 0.50 0.00 0.00 0.00 0.00 1.00 1.00 1.00 FDICIA dummy 1092 0.12 1.00 0.00 0.32 0.00 0.00 0.00 0.00 0.00 1.00 1.00
Table 2: Ordnary Least Squares Regressons (1) (2) (3) (4) (5) Cost (log) Cost (log) Cost (log) Cost (log) Cost (log) Total assets (deflated) 0.8968 0.8376 0.8215 0.8191 0.8340 (0.0329)*** (0.0313)*** (0.0338)*** (0.0337)*** (0.0320)*** Real estate owned/total assets 4.4190 3.5666 4.0795 4.0322 4.0092 (0.6888)*** (0.5030)*** (0.4841)*** (0.4883)*** (0.4678)*** Equty/Total assets -1.4808-8.3374-7.3869-7.3147-7.7015 (1.5465) (1.7583)*** (1.6475)*** (1.6598)*** (1.5635)*** Loans past due/total assets 2.6701 1.9981 1.0039 1.0091 0.6931 (1.0608)** (1.0187)* (0.8654) (0.8608) (0.8021) Income earned, not collected/total assets 32.4485 33.6278 25.9984 25.8933 23.8104 (5.0442)*** (3.9075)*** (4.3290)*** (4.3262)*** (4.2849)*** Asset growth 2 years pror to falure 0.1623 0.2821 0.2533 0.2395 0.2663 (0.1019) (0.0790)*** (0.0689)*** (0.0685)*** (0.0643)*** FDICIA dummy -1.4311-0.9515-0.9667-0.9531-0.9427 (0.1467)*** (0.1596)*** (0.1569)*** (0.1599)*** (0.1630)*** Fed funds purchased/total assets -2.3840-3.2451-3.1941-3.5057 (2.1937) (2.0853) (2.0866) (1.9566)* Brokered deposts/total assets 0.7000 0.2963 0.3078 0.4287 (0.6008) (0.5188) (0.5153) (0.4237) Transactons deposts/total assets -2.0234-1.0983-1.0734-0.5216 (0.5240)*** (0.5070)** (0.5014)** (0.4524) Tme and savngs deposts/total assets -5.9227-5.6922-5.6822-5.8544 (1.6514)*** (1.5456)*** (1.5581)*** (1.4452)*** Demand deposts/total assets -4.1961-5.6017-5.6453-6.1340 (1.8182)** (1.6974)*** (1.7101)*** (1.5921)*** Subordnated debt/total assets -5.0942-4.8766-4.5684-2.6518 (15.9163) (15.7139) (15.5896) (13.6435) Other labltes/total assets -3.5719-3.3345-3.2954-4.3538 (1.9382)* (1.7924)* (1.8055)* (1.7119)** C&I loans/total assets 3.1892 3.2256 2.9026 (0.3509)*** (0.3508)*** (0.3301)*** Agrcultural loans/total assets 0.9254 0.9793 1.3805 (0.4434)** (0.4415)** (0.4420)*** Real estate loans/total assets 1.0856 1.0954 0.8580 (0.3537)*** (0.3526)*** (0.3473)** Indvdual loans/total assets 0.8364 0.8201 0.5868 (0.4054)** (0.4080)** (0.4045) Depostor preference law -0.0745-0.0750 (0.0637) (0.0640) Personal Income Growth (lagged) -0.3457 (1.0125) Bankruptcy Growth Rate 0.7166 (0.2196)*** Unemployment Rate 0.1405 (0.0231)*** Observatons/Number of falures 1066 1024 1023 1023 1023 Adjusted R square 0.5302 0.6002 0.6390 0.6392 0.6704 AIC 3081.723 2781.438 2672.087 2672.444 2582.824 We estmate OLS regressons n column (1) (5) for the perod 1984-1996. The dependent varable s the log of the dollar value of losses n the quarter pror to falure. Specfcaton (1) s the baselne model that ncludes covarates used n prevous studes. We addtonally ncorporate a dummy varable that takes on the value one f the falure occurred n the perod followng enactment of the Federal Depost Insurance Corporaton Improvement Act n 1991. Specfcaton (2) ncludes varables that capture lablty structure. We nclude addtonal control varables n Specfcaton (3) to capture asset composton. Specfcaton (4) ncludes a dummy varable that takes on the value one f depostor preference law was n place n the state n whch the bank s located or zero otherwse. In Specfcaton (5) we account for the macroeconomc envronment on the federal state level and nclude varables that capture personal ncome growth, bankruptcy growth and unemployment. Robust standard errors are reported n parentheses. Sgnfcance levels of 1, 5 and ten percent are ndcated by *, **, and ***.
Table 3: Ordnary least squares and quantle regressons regressons (1) (2) (3) (4) (5) (6) (7) (8) OLS q. 05 q. 10 q.25 q.50 q.75 q.90 q.95 Total assets (deflated) 0.8340 0.7403 0.8015 0.8063 0.8579 0.8918 0.9209 1.0129 (0.0320)*** (0.0847)*** (0.0640)*** (0.0402)*** (0.0239)*** (0.0351)*** (0.0456)*** (0.0515)*** Real estate owned/total assets 4.0092 6.7291 6.5193 4.5904 2.9859 1.8738 1.8825 1.5987 (0.4678)*** (0.9792)*** (0.7745)*** (0.5553)*** (0.4281)*** (0.4966)*** (0.6697)*** (0.8276)* Equty/Total assets -7.7015-7.2358-6.6059-8.5075-7.8334-6.8643-3.5968-4.2023 (1.5635)*** (2.8537)** (2.3610)*** (2.5157)*** (2.2400)*** (1.7194)*** (2.9675) (4.4366) Loans past due/total assets 0.6931 0.2315 1.5355 1.3033 1.0914 0.5320 0.2952 0.6637 (0.8021) (1.9248) (1.3278) (0.8923) (0.8109) (0.6891) (0.7305) (1.1217) Income earned, not collected/total assets 23.8104 36.0187 29.4740 24.1902 22.1033 19.6424 22.1703 25.9112 (4.2849)*** (9.1994)*** (6.5363)*** (6.8902)*** (4.7023)*** (4.0406)*** (6.5240)*** (7.8535)*** Asset growth 2 years pror to falure 0.2663 0.4111 0.4538 0.3671 0.3298 0.2178 0.2938 0.2801 (0.0643)*** (0.2347)* (0.1460)*** (0.0959)*** (0.0728)*** (0.0746)*** (0.1174)** (0.1360)** FDICIA dummy -0.9427-1.1659-1.0773-0.7134-0.8173-0.9233-0.9165-0.9617 (0.1630)*** (0.8986) (0.3769)*** (0.2409)*** (0.0956)*** (0.1270)*** (0.1715)*** (0.1766)*** Fed funds purchased/total assets -3.5057-8.1590-2.1938-5.4359-4.8779-2.1017 1.3986 1.0921 (1.9566)* (4.2730)* (3.2500) (2.9549)* (2.8823)* (2.5713) (3.6287) (4.7254) Brokered deposts/total assets 0.4287 0.0911 0.2265-0.0146 0.3525 0.5377 0.5904 0.4716 (0.4237) (0.8823) (0.7512) (0.3891) (0.3365) (0.3353) (0.3271)* (0.5471) Transactons deposts/total assets -0.5216-0.8848-0.7037 0.2218-0.4324-0.4575 0.1078-0.0317 (0.4524) (1.1475) (0.9034) (0.5554) (0.3675) (0.4055) (0.4773) (0.4769) Tme and savngs deposts/total assets -5.8544-4.3576-3.5360-6.0386-5.8508-5.3294-2.4002-3.2440 (1.4452)*** (2.7063) (2.3136) (2.4556)** (2.2177)*** (1.6621)*** (2.9580) (4.4287) Demand deposts/total assets -6.1340-5.0309-4.1395-6.9371-6.2356-5.7975-3.5546-3.4712 (1.5921)*** (3.6641) (2.7069) (2.5368)*** (2.2349)*** (1.7827)*** (3.1100) (4.6504) Subordnated debt/total assets -2.6518 3.5393-4.7926-13.8004 6.9646 12.2365 20.4744-8.1534 (13.6435) (97.3029) (15.1688) (23.0044) (15.6175) (16.8075) (13.3333) (17.2878) Other labltes/total assets -4.3538-3.2741-3.6733-5.4101-4.5460-4.4378-1.5226-1.9883 (1.7119)** (3.2161) (2.6510) (2.7791)* (2.3535)* (1.8795)** (3.0351) (4.5531) C&I loans/total assets 2.9026 5.3493 4.3623 3.4037 2.4651 1.7890 1.6688 1.1402 (0.3301)*** (0.7568)*** (0.5050)*** (0.4247)*** (0.2882)*** (0.2900)*** (0.3454)*** (0.4462)** Agrcultural loans/total assets 1.3805 2.9269 2.7507 1.8543 0.9489 1.0318 0.6080 0.2667 (0.4420)*** (1.0947)*** (0.7062)*** (0.6918)*** (0.4197)** (0.3883)*** (0.4515) (0.4837) Real estate loans/total assets 0.8580 2.7866 1.9997 1.3929 0.3933 0.1232 0.0554-0.1137 (0.3473)** (0.8140)*** (0.5359)*** (0.4092)*** (0.2574) (0.2824) (0.3700) (0.4528) Indvdual loans/total assets 0.5868 2.3054 1.4029 1.3813 0.7519 0.2609 0.1795-0.0232 (0.4045) (1.3410)* (0.7012)** (0.5262)*** (0.3188)** (0.2893) (0.3765) (0.4649) Depostor preference law -0.0750-0.3129-0.2317-0.1348-0.0167 0.0143 0.0490-0.0321 (0.0640) (0.1892)* (0.1226)* (0.0754)* (0.0499) (0.0590) (0.0545) (0.0595) Personal Income Growth (lagged) -0.3457 4.2330 1.3258-0.5994-3.0264-1.2976 1.3482 1.2370 (1.0125) (2.5839) (2.2323) (1.6325) (0.9345)*** (0.8257) (0.7840)* (0.7426)* Bankruptcy Growth Rate 0.7166-0.2030 0.2878 0.4985 1.0041 0.8800 0.6011 0.7352 (0.2196)*** (0.4309) (0.3384) (0.3051) (0.2088)*** (0.1872)*** (0.2100)*** (0.2653)*** Unemployment Rate 0.1405 0.2299 0.1928 0.1717 0.1283 0.1520 0.1459 0.1048 (0.0231)*** (0.0655)*** (0.0358)*** (0.0256)*** (0.0184)*** (0.0212)*** (0.0237)*** (0.0245)*** Observatons 1023 1023 1023 1023 1023 1023 1023 1023 R square/pseudo R square 0.6704 0.3384 0.3727 0.4251 0.5043 0.5702 0.6353 0.6661 We report OLS regressons n column (1) and quantle regresson estmates n column (2) (8). The dependent varable s the log of the loss on assets. Robust standard errors are reported n parentheses for OLS regressons and bootstrapped standard errors based on 500 replcatons are reported n parentheses for the quantle regressons. Pseudo R square reported for quantle regressons. The pseudo R square s calculated as 1-(sum of the weghted devatons about estmated quantle/sum of weghted devatons about raw quantle). Sgnfcance levels of 1, 5 and ten percent are ndcated by *, **, and ***.
Table 4: F-Tests for the equalty of coeffcents across quantles Varable (1) (2) (3) F-test (equalty across all quantles) F-test (equalty between 5 th, 50 th, and 95 th quantle) F-test (equalty across tals) Total assets (log) (deflated) 2.58** 6.12*** 8.97*** Real estate owned/total assets 5.25*** 8.57*** 16.34*** Equty/Total assets 0.33 0.28 0.31 Loans past due/total assets 0.30 0.14 0.04 Uncollected ncome/total assets 0.61 1.03 0.76 Asset growth 0.59 0.12 0.23 Fed funds/total assets 1.05 1.04 2.04 Transactons deposts/total assets 0.85 0.44 0.45 Brokered deposts/total assets 0.45 0.06 0.13 Demand deposts/total assets 0.23 0.19 0.07 Tme and savngs deposts/total assets 0.30 0.22 0.04 Subordnated debt/total assets 0.92 0.25 0.02 Other labltes/total assets 0.22 0.17 0.05 C&I loans/total assets 5.57*** 12.87*** 25.74*** Agrcultural loans/total assets 1.75 2.67* 5.27** Real estate loans/total assets 3.01*** 6.22*** 11.98*** Indvdual loans/total assets 1.04 2.00 3.01* Personal ncome growth (lagged) 4.39*** 11.36*** 1.15 Bankruptcy growth rate 2.23** 4.67*** 3.23* Unemployment rate 1.77 1.71 3.41* Depostor preference law dummy 1.59 1.52 2.43 FDICIA dummy 0.36 0.42 0.06 Ths table presents F-tests for the equalty of the slope coeffcents for the explanatory varables used n the cost equatons. The F-tests are based on the coeffcents reported n Table 3. Column (1) reports F-tests for the equalty of coeffcents across all quantles from the 5 th 95 th quantle; column (2) presents F-tests for the equalty of the coeffcents for the 5 th, 50 th and the 95 th quantle and column (3) presents F-tests for the equalty of the coeffcents for low-cost (5 th quantle) and hgh-cost (95 th quantle) falures.
Table 5: Duraton analyss (1) (2) (3) (4) (5) Equty/Total assets 37.1454 29.9544 30.8770 30.6854 29.5171 (20.7265)*** (12.6300)*** (14.3502)*** (14.2993)*** (13.5246)*** Troubled assets/total assets -12.6529-11.9390-11.5690-11.6190-10.6854 (9.8235)*** (9.4432)*** (9.1905)*** (9.1827)*** (8.7001)*** Operatng ncome/total assets 0.9341 1.1782 0.9808 0.9426 0.9921 (1.0061) (1.3997) (1.1754) (1.1314) (1.2236) Total assets (deflated) 0.4373 0.4457 0.3228 0.3217 0.3112 (11.8921)*** (13.9652)*** (9.8552)*** (9.9368)*** (9.7646)*** FDICIA dummy -0.0289-0.0461-0.0007-0.0081-0.0920 (0.2786) (0.5087) (0.0082) (0.0924) (1.0014) Lqudty/Total assets 2.6032 2.0650 1.7574 1.7235 1.5961 (10.4516)*** (8.7665)*** (4.9666)*** (4.8987)*** (4.6106)*** Fed funds purchased/total assets -6.1681-4.6112-4.6097-4.3518 (2.9673)*** (3.3117)*** (3.3486)*** (2.9097)*** Brokered deposts/total assets -3.1782-3.5328-3.5055-3.7120 (5.7138)*** (4.8869)*** (5.1768)*** (5.1444)*** Transactons deposts/total assets 2.0229 1.9269 1.8574 1.5842 (4.2117)*** (3.7594)*** (3.6785)*** (3.2263)*** Tme and savngs deposts/total assets -5.7039-4.5549-4.5582-4.2282 (2.8552)*** (3.5269)*** (3.5712)*** (3.0154)*** Demand deposts/total assets -6.7986-5.9254-5.8616-5.4670 (3.1883)*** (4.0635)*** (4.0685)*** (3.4914)*** Subordnated debt/total assets 23.6120 24.1684 23.5373 22.0041 (3.1029)*** (3.3877)*** (3.3287)*** (3.2729)*** Other labltes/total assets -7.3790-5.9387-5.9451-5.6141 (3.4404)*** (4.0911)*** (4.1714)*** (3.6829)*** C&I loans/total assets -1.6225-1.6529-1.5139 (4.0377)*** (4.1485)*** (3.8332)*** Agrcultural loans/total assets -1.8482-1.9151-1.9812 (4.2057)*** (4.3598)*** (4.5789)*** Indvdual loans/total assets 0.0646 0.0790 0.0588 (0.1607) (0.1978) (0.1506) Real estate loans/total assets 0.6243 0.6161 0.6214 (1.5734) (1.5664) (1.6149) Depostor preference law 0.0895 0.0873 (1.7933)* (1.7491)* Personal Income Growth (lagged) 2.4948 (2.5798)*** Bankruptcy Growth Rate -0.7044 (4.3396)*** Unemployment Rate -0.0238 (1.7039)* Observatons 456857 444315 444315 444315 443823 Number of banks 13884 13884 13884 13884 13870 Number of falures 922 903 903 903 903 AIC 2253.704 2037.574 1904.514 1903.011 1876.775 Log lkelhood functon -1118.852-1003.786-933.256-931.505-915.387 We report log-logstc duraton models wth tme-varyng covarates based on the log-logstc dstrbuton n column (1) - (5) for the perod 1982-1996. The dependent varable s the log of tme to falure. Specfcaton (1) contans varables used n prevous studes and a dummy that takes on the value one f the observaton s from the perod followng enactment of the Federal Depost Insurance Corporaton Improvement Act n 1991. Specfcaton (2) ncludes covarates that capture the fundng structure. Addtonal control varables are ncluded n Specfcaton (3) to capture composton of the loan portfolo. We ncorporate a dummy varable for depostor preference n Specfcaton (4) that takes the value one f depostor preference law s n place or zero otherwse. Specfcaton (5) addtonally ncludes varables that capture the macroeconomc settng on the federal state level. Robust standard errors are reported n parentheses. Sgnfcance levels of 1, 5 and ten percent are ndcated by *, **, and ***. - 33 -
Fgure 1: regresson estmators a) Total assets (log), deflated b) Real estate owned/total assets c) Total equty captal/total assets d) Loans past due (90 days+)/total assets 0.60 0.70 0.80 0.90 1.00 1.10 Total assets (deflated) 0.00 2.00 4.00 6.00 8.00 Real estate owned/total assets 15.00 10.00 5.00 0.00 5.00 Equty/Total assets 4.00 2.00 0.00 2.00 4.00 Loans past due/total assets e) Income earned, not collected/total assets f) Total asset growth, 8 quarters pror to falure g) Fed Funds purchased/total assets h) Brokered deposts/total assets 10.00 20.00 30.00 40.00 50.00 60.00 Income earned_ not collected/total assets 0.50 0.00 0.50 1.00 Asset growth 2 years pror to falure 20.00 10.00 0.00 10.00 Fed funds purchased/total assets 2.00 1.00 0.00 1.00 2.00 Brokered deposts/total assets ) Tme and savngs deposts/total assets j) Transactons deposts/total assets k) Demand deposts/total assets l) Subordnated debt/total assets 10.00 5.00 0.00 5.00 Tme and savngs deposts/total assets 3.00 2.00 1.00 0.00 1.00 2.00 Transactons deposts/total assets 15.00 10.00 5.00 0.00 5.00 Demand deposts/total assets 200.00 100.00 0.00 100.00 200.00 Subordnated debt/total assets - 34 -
m) Other labltes/total assets n) C&I loans /Total assets o) Agrcultural loans/total assets p) Indvdual loans/total assets 10.00 5.00 0.00 5.00 Other labltes/total assets 0.00 2.00 4.00 6.00 8.00 C&I loans/total assets 2.00 0.00 2.00 4.00 6.00 Agrcultural loans/total assets 2.00 0.00 2.00 4.00 6.00 Indvdual loans/total assets q) Real estate loans/total assets r) Personal ncome growth (2 year lag s) Unemployment rate t) Bankruptcy growth rate 1.00 0.00 1.00 2.00 3.00 4.00 Real estate loans/total assets 5.00 0.00 5.00 10.00 Personal Income Growth (lagged 2) 0.00 0.10 0.20 0.30 0.40 Unemployment Rate 2.00 1.00 0.00 1.00 2.00 Bankruptcy Growth Rate u) Depostor preference law dummy v) FDICIA dummy 0.60 0.40 0.20 0.00 0.20 Depostor preference law 2.50 2.00 1.50 1.00 0.50 0.00 FDICIAdummy - 35 -