The Sixteenth Dubrovnik Economic Conference

Transcription

1 The Sxteenth Dubrovnk Economc Conference Organzed by the Croatan Natonal Bank Ivana Herceg and Vedran Šošć The Anatomy of Household Debt Buld Up: What Are the Implcatons for the Fnancal Stablty n Croata? Hotel "Grand Vlla Argentna", Dubrovnk June 23 - June 26, 2010 Draft verson Please do not quote

2 The Anatomy of Household Debt Buld Up: What Are the Implcatons for the Fnancal Stablty n Croata? Ivana Herceg and Vedran Šošć June, 2010 Draft verson Please do not quote

3 1. Introducton Rapd growth of Croatan household debt n perod pror to the outbreak of the fnancal crses relaxed the fnancal constrant of the households, allowng them to frontload some of ther consumpton on expectatons of rapdly growng ncomes. However, at the same tme t rased concerns about the potental mplcatons of household over-ndebtedness on stablty of the fnancal system as economc outlook hugely deterorated. The goal of ths paper s to explore mplcatons of the rapd debt accumulaton by the households for fnancal stablty 1. Standard approach to ths topc observes dfferent macro-drvers of sharp household lendng wth specal attenton gven to the EU ntegraton process and real convergence. More recently, a consensus has started to form n the lterature that the level of systemc rsk n the fnancal sector depends on the actual dstrbuton of debt (and assets) amongst the households rather than the aggregate level of ndebtedness, promptng strong relance of the feld on mcro data sources (Beck et. al., 2010 and World Bank, 2009). Such an approach gnores macroeconomc rsks stemmng from buld up of external mbalances, but t s nevertheless an mportant extenson of macro-prudental tools, such as early-warnng ndcators for sudden stops n captal flows or fnancal crses, n case those rsks materalze. Ths paper expands the range of the tradtonal mcro-data analyss technques as t ams to account for the changes n the dstrbuton of household debt. Frst step n ths drecton s dentfcaton of the determnants of household debt over the observed perod. However, nstead of lookng the man determnants of the amount owned by households at the mean of the dstrbuton (estmated by the OLS), whch s standard approach n the debt determnants lterature, ths paper employs a quantle regresson analyss. The quantle regresson (QR) allows for dentfcaton of the effect of dfferent households' demographc and socoeconomc characterstcs upon the total amount of debt across the whole dstrbuton of the ndebted households. Snce dentfcaton of the debt determnants s preformed on the non-random sample of only ndebted households, the problem of the sample selecton bas s lkely to emerge. Ths problem could arse because some groups of households may be subject to fnancal constrant due to banks' lendng polcy or subjectve percepton of those polces. In order to control for the non-random bas, a modfcaton of a two-stage Heckman model for the quantle regresson proposed by Buchnsky (1998) was employed. Accordng to ths methodology unknown form of the sample selecton bas term can be approxmated followng two-step procedure: frst, the credt market partcpaton selecton parameter s estmated usng dstrbuton free semparametrc least squares estmaton of sngle ndex (Ichmura, 1993), followed by a power seres approxmaton of the bas term. As a next step, changes n the characterstcs of the ndebted households wll be used to dsentangle ther effects from the effects of changes n the estmated QR coeffcents on the rse of total household debt durng observed perod. These characterstcs are used to proxy for household credtworthness and changes of estmated QR coeffcents could be used to approxmate possble relaxaton of bank's lendng standards and/or wllngness of certan types of household to take more debt aganst ther ncomes. Decomposton method proposed by Machado and Mata (2005) wll be employed n order to separate the effects of standard 1 Conducted analyss and estmated results outlned n ths paper are based on the study "Household Credt Rsk n Croata. An Analyss Based on the Households Budget Survey" (2009) carred out n collaboraton of the Insttute of Economcs, Zagreb and Croatan Natonal Bank.

4 relaxaton vs. household characterstcs on the debt dstrbuton. Ths technque extends the tradtonal Oaxaca decomposton of these effects on mean (Oaxaca, 1973) to the entre dstrbuton. A specal attenton wll be gven to the households carryng the hghest amounts of debt. The paper s organzed as follows. Secton 2 gves an overvew of the related lterature and presents the motvaton for the analyss outlned n the rest of the paper. Secton 3 descrbes the methodology used n estmaton of household debt determnants and decomposton of ts growth from 2005 to Secton 3 descrbes the used dataset and expected effect of dfferent households' characterstcs on ther decson to take bank loan. Emprcal results of the selecton and outcome equaton of the sample selecton model together wth decomposton of household debt growth are gven n secton 5. Secton 6 concludes. 2. Lterature revew: Household lendng and fnancal (n)stablty Levels, dynamcs and qualty of household lendng have not tradtonally been subject of researchers seekng to dentfy causes of bankng system dstress. Households were consdered to be a trustworthy borrower that repad ts debts n tme, or at least had suffcent collateral to prevent excessve bank losses on household lendng. Thus, epsodes of deteroraton n qualty of loans that closely followed recessons have hstorcally mostly been confned to the corporate sector. In that ven Capro (1998) noted that "a fnancal crss usually nvolves a corporate debt problem n the non-bank fnancal sector". Such a bengn percepton of household lendng became ncreasngly scrutnzed snce the late 1990's, as the run-up n household debt accelerated, and eventually dssolved wth the advent of the sub-prme crses. Most advanced countres experenced a synchronzed growth n household ndebtedness over the prevous decade. The extent and composton of the household debt ncrease dffered from one country to the next, but very few of them managed to avert the general trend of rapd debt buld-up. In parallel wth growng ndebtedness, durng the late 1990's and especally over the prevous decade, a growng lterature on household borrowng started to emerge (Grouard et al., 2006). In that sense, some researchers fascnaton wth household borrowng prompted by the rapd growth of household ndebtedness and household leverage s a rather novel phenomenon. Most of the papers on household lendng publshed n ths perod could be classfed n several lnes of research. Many of the papers adopted a macro approach as they employed aggregate economc data n dealng wth causes of a wdespread growth n household ndebtedness. Authors usng ths approach dentfed a range of dfferent factors that accounted for the observed dynamcs of household debt, such as combnaton of favorable fnancal condtons related to bengn monetary polcy, buoyant housng markets and now notorous nnovatons n credt markets that have eased the access to credt for lower-ncome borrowers and reduced fnancal constrants for frst-tme homebuyers (see Grouard et al., 2006; Dynan and Kohn, 2007; and Dynan, 2009). Papers adoptng a macro approach also dealt wth consequences of growng household ndebtedness. However, observng aggregate household balance sheets and aggregate data on debt servce burdens provdes a very rough gudance on actual household vulnerabltes as t conceals potental major dfferences between groups of households. As ths paper explores another avenue, rather than provdng a comprehensve overvew of the ssues and methods

5 used n ths lterature, only some major concluson of papers dealng wth Central and Eastern European countres wll be brefly addressed. Central and Eastern European countres have featured promnently n ths area of research because of rapd lendng growth drven by rapd economc development, fnancal lberalzaton and openng as well as convergence of ther fnancal systems towards structures found n Western economes. Croata had a bt of specal poston because of relatvely hgh level of ndebtedness already n md-2000 s, structure of bank borrowng that was early tlted towards household debt and slghtly slower recent growth of household ndebtedness than that found n most Central and Eastern European countres. Stll, Croatan household ndebtedness more than doubled n the perod between 2005 and 2008 and, gven the farly hgh ntal level, brought t to the top amongst Central and Eastern European countres, next only to some of the Baltc Republcs. Aggregate household debt therefore converged to the average EU level (see Fgure 1) as t reached 40.5% of GDP by the end of Ths debt growth relaxed the fnancal constrant of the households, allowng them to frontload some of ther consumpton on expectatons of rapdly growng ncomes (n lne wth lfe-cycle permanent ncome hypothess), but at the same tme t rased some concerns about the potental mplcatons of household over-ndebtedness on stablty of the fnancal system, especally after the economc outlook hugely deterorated. However, most of the papers addressng lendng expanson before the ongong crses attrbuted t to equlbratng forces of lberalzaton, convergence and algnment wth the EU structures. Mhaljek (2006) noted that for the tme beng, the expanson of prvate sector credt has been mostly a bengn phenomenon. Kss et al. (2006) agreed that large part of the credt growth n new member states can be explaned by the catchng-up process, and, n general, credt/gdp ratos are below the levels consstent wth macroeconomc fundamentals. Kraft (2006) s nterestng because of focus on household lendng n Croata. However, hs concluson seems smlar as he attrbuted t to strong performance n bankng sector reform and n mantanng low nflaton whle leavng weaknesses n enterprse reform and prvatzaton. Šonje (2009) even after the onset of the crses remarks that path of credt ntegraton looks remarkably smlar n most Central and Eastern European countres to earler epsodes of fnancal ntegraton and that credt growth n very few countres extends beyond justfable on those grounds. Zdzencka (2010) s one of a few counter-examples reportng ndcatons of "excessve" credt developments n most Central and Eastern European countres untl 2007, and even snce for some countres. Obvously very few of the researchers found some reasons for concern n general rapd lendng growth based on the "macro" approach. The second major research agenda adopted a mcro approach that was much more ntmately ntertwned wth the actual patterns of the household debt expanson. Papers n ths category predomnantly use data on ndvdual households compled from varous household surveys, allowng them to dentfy the precse profle of household vulnerabltes, n contrast to macro approach, where such nformaton remans unknown. Despte recent advances, as recently as 2004 Brown et al. noted that research nto the determnants of ndvdual debt or household level debt s surprsngly scarce. Research on the profle of household borrowng from the mcro perspectve looked for deeper causes of debt accumulaton by testng f rsng debt resulted from optmzng, welfare enhancng behavor of the households under more favorable fnancng condtons. In a way, dstngushng whether households use easng of lqudty constrant n order to algn ther spendng profles wth the optmal or market mperfectons, such as short-sghtedness,

6 behavoral nerta or excessve rsk-takng, actually ncrease the volatlty of household expendtures by makng household spendng decsons ntrnscally unsustanable under constant shocks developed nto the major undertakng of ths lterature. Lfe-cycle permanent ncome hypothess 2 became the standard workhorse of the lterature as researchers strved to guess the extent to whch the relaxaton of the households fnancal constrant mproved the way they smooth ther consumpton over tme. Ths also brought focus on the actual dstrbuton of debt across households and the way t nteracts wth the household characterstcs to produce patterns of over-ndebtedness and fnancal dstress. Dfferent unantcpated shocks, such as those to ncome, nterest rates or exchange rate, have a potental to rase present value of household debt above present value of ts assets and make optmal spendng patterns unsustanable, whch s one of the ways to defne the problem of fnancal dstress. Even f the concept over ndebtedness occurred n ths lne of lterature, t was confned to a byproduct of unavodable fundamental uncertanty rather than systemc faults related to ever-easng fnancng constrant. Thus, t s not surprsng to see papers wrtten n ths sprt reportng that over ndebtedness patterns had more to do wth dosyncratc shocks households faced than any systematc ncome, age or famly structure factors, whle households n countres wth deeper fnancal systems were no more prone to over-ndebtedness (see Bett et al., 2007, Beck et al., 2010). Possblty to frontload some of the household consumpton on expectatons of growng future ncomes also rased concerns about the potental mplcatons of household over-ndebtedness. These concerns about potental consequences of household ndebtedness led to another stran of the lterature wth much more straghtforward vew of rsks stemmng from rsng household debt. Relaxaton of the borrowng constrant mtgates lqudty restrctons and allows households to ncrease spendng, whle reducng ther "savngs buffer" (and consequently ther aggregate savngs rate 3 ) as households feel they could use lendng rather than own precautonary savngs to nsure aganst shocks to ther consumpton. But a combnaton of reduced savngs buffers and hgh debt burden also makes households more susceptble to unantcpated shocks that lead to over-ndebtedness and fnancal dstress. Man and Suf (2009) clearly show the lnk between stronger growth of loans to less credtworthy customers (measured by ther FICO credt score) and subsequent rse n the number of loan defaults n US regons. Identfcaton of possble myopc behavor also opened doors to polcy advce. Brown et al. (2004) reported a major mpact of optmstc fnancal expectatons on the level of household debt n the UK and hghlghted the mportance of curbng unwarranted fnancal optmsm for avodance of excessve debt problems. Many of the feld practtoners, who flled much of the gap left by academc economsts n addressng household fnancal vulnerabltes, ddn t care whether fnancal dstress was caused by household myopa or unforeseen events. Splt n the lterature between those who saw rsng debt as a sgn of ablty of households to better algn ther spendng patterns wth optmzng behavor wthn the framework of nter-temporal budget constrant and those who emphaszed potental rsks related of excessve ndebtedness due to macroeconomc fluctuatons and myopc borrowers ddn't mpress many of those who looked for mplcatons of rsng debt on household vulnerablty. Leavng theoretcal consderatons asde allowed focus on the response of over-ndebtedness on dfferent knds of shocks. "Practtoners" 2 For more on the lfe-cycle permanent ncome hypothess see Crook (2006) and Del-Ro and Young (2005). 3 Goh (2005) uses New Zealand as an example to emphasze macroeconomc consequences as hgher household debt and lower savngs also ncrease external vulnerablty of the whole country, but such mplcatons of rsng household debt ate beyond the scope of ths paper.

7 focused on problem of over-ndebtedness from varous perspectves, such as socal mplcatons of the change n the number of persons exposed to cycle of ever-ncreasng debts (Brown et al., 2005) mplcatons of non-payments arsng from household fnancal dstress on the stablty of fnancal system (Herrala and Kauko, 2007) or the way rsng household debt nteracts wth populaton ageng to produce specfc patterns of vulnerablty. Methodologcal approaches used to tackle the problem of over ndebtedness f one of the touchng ponts between those seekng fundamental forces shapng ndebtedness patterns and practtoners lookng at mpact of shocks on fnancal dstress. Stll, avalablty of data on ndvdual households ddn t allow for resolvng n a satsfactory way the subtle lne whch denotes the transton of a household nto over-ndebtedness. There are three man methodologcal approaches to the defnton of fnancal dstress n the lterature on household ndebtedness, wth occasonal resort to some addtonal, ad hoc defntons, although dfferent varatons of those approaches make the actual number of crtera used n the lterature much larger (see Bett et al., 2007, Beck et al., 2010 for more detaled explanatons on some of those measures). Also, actual underlyng structure of the data source used, such as survey desgn, also nfluence the exact defnton of the vulnerablty crtera used and make harder comparsons over countres and dfferent. Frst approach, whch observes so-called "objectve" measures of household fnancal dstress, s based on the dea that households are vulnerable n case ther ndebtedness or debt servce ratos exceed a certan threshold. Sometmes ndcators of household consumpton relatve to ncome are used rather than debt/repayment ratos, wth hgh consumpton beng an ndcator of possble fnancal dstress (Bett et al. 2007). Ths approach s routnely used by researchers and has a natural appeal to feld practtoners nterested n fnancal dstress, such as commercal and central banks (see Uncredt, 2009 and European Central Bank, 2007). Ratos of debt to ncome n the range of 450%-600% and debt repayment to ncome of 30% are commonly used as thresholds of vulnerablty. However, ths ndcator also suffers from several serous defcences. Frst, although most emprcal studes use smlar ratos, there are no establshed thresholds that unavodably lead to household fnancal dstress. Further on, t compares ratos of debt and repayments to actual ncome rather than lfetme ncome, whle households may be nclned to run debt exactly at tmes when ther current ncome declnes below ther earnng potental. The concept of so-called "fnancal margn" s a dervatve of the "objectve" approach whch lessens some of the problems arsng from the use of arbtrary margns and has ganed much popularty among the practtoners smulatng mpacts of varous shocks on the ranks of vulnerable households. Fnancal margn refers to ncome reserve that remans after debt servce and household-specfc poverty lne have been subtracted from the household ncome. Households wth negatve fnancal margn are usually consdered to be vulnerable. However, calculaton of household specfc "fnancal margn" stll does not resolve the problem of settng an arbtrary threshold by a researcher but rather desgnates t to the nsttuton settng poverty lnes. The second approach s a "subjectve" n a sense that t reles on a subjectve evaluaton of household balance sheets and debt servcng burden. Typcally, these measures are based on the number of households reportng a degree of hardshp n servcng ts debt. One of the problems wth ths ndcator s that subjectve well-beng does not necessarly have to correlate closely wth ncome, but may be nfluenced by other factors, such as comparsons wth the reference group (Georgarakos et al., 2009).

8 The fnal approach s co called "admnstratve" where data on actual bankruptces or debt defaults s used. As most studes use household survey data, a dervaton of ths approach uses self-reported debt arrears as an ndcator of fnancal dstress. Sometmes the concept of arrears s expanded to nclude not only arrears ncurred towards fnancal nsttutons, but also late payments of certan utltes or other blls such as rent. Also, t s possble to vary thresholds for arrears from one to several months n order to make the crtera more or less strngent or algn t wth actual bankng practces. Ths defnton s the one that s most closely algned wth the concept of bank losses stemmng from household fnancal dstress, although the fact that data a collected wth the households ntroduces some dfferences. Whle "mcro" approaches greatly dffer n ther methodologes wth respect to the actual measure of "dstress" used, there s hardly any advce on the way how to dentfy fnancally dstressed households for dfferent purposes. It s therefore not surprsng to fnd that all these ndcators n countless varatons are nterchangeably used n studes of dfferent phenomena. Most "mcro" studes of rsks arsng from household lendng conclude that rsks for the fnancal system are neglgble. Conclusons from a sample of studes performed n dfferent countres look almost exactly the same. Beer and Schürz (2007) assert that "the rsks assocated wth prvate debt that could threaten fnancal stablty n Austra are mnmal." Fuenzalda and Ruz-Tagle (2009) smulate mpact of unemployment on the Chlean fnancal system and conclude that even hgher levels of unemployment do not necessarly mply that the fnancal system wll suffer a sgnfcant default shock by households. Herrala and Kauko (2007) are a bt more cautous n ther smulatons of rsks related to household lendng n Fnland as they warn that "In most states of the economy household loans bear a relatvely low credt rsk to banks. However, under extreme condtons wth a concdence of large and persstent adverse shocks to unemployment, nterest rates and housng prces, even household loans could become a threat to fnancal stablty." Keese (2009) n hs study on Germany does not partcularly emphasze macroeconomc shocks as he fnds that "trgger events such as strokes of fate (death, separaton, or dvorce), change n household composton (cohabtaton or marrage), unemployment, and chldbrth account for most changes n the debt stuaton of a household." For Central and Eastern European countres Beck et al. (2010) called for greater use of mcro data to assess household ndebtedness and overall fnancal stablty as they note that "To date, lttle s known about the ncdence of household ndebtedness and ts dstrbuton". There are, however, a few studes wth very smlar fndngs to those n the "Western" lterature. Holló and Papp (2007) conclude that the captal adequacy rato of the Hungaran bankng system would not fall below the current regulatory mnmum of 8 per cent even f the most extreme stress scenaros were to occur. Żochowsk and Zajączkowsk (2006) for Poland conclude that none of the rsks they analyzed s mportant enough to pose a threat to the fnancal system stablty. In a more recent study Daras and Tyrowcz (2009) fnd that "even small changes n the employment persstence or unemployment rsk can lead to consderable deteroraton of households' lqudty and therefore the fnancal stablty of the whole mortgage market", but ths fndng s obtaned by employng a rather soft defnton of fnancally dstressed households whch s much hgher that non-performng loans even before the applcaton of the stress scenaros.wb (2010) s also much more cautous from the rest of the mentoned studes as they that report "results of stress tests on household ndebtedness n selected countres suggest that ongong macroeconomc shocks may sgnfcantly expand the pool of households that wll be unable to meet debt servce oblgatons. Interest rate shocks n

9 Estona, Lthuana, and Hungary, for example, ncrease the share of vulnerable households or borrowers at rsk (n percent of all ndebted households) by up to 20 percentage ponts, dependng on the magntude and severty of the shock." Whle there are several potental explanatons for rather favorable results of most stresstestng exercses performed for the household sector n Central and Eastern European countres, such as mld scenaros, part of the explanaton could probably be attrbuted to the selecton of vulnerablty ndcators used. For now there are no n-debt examnatons of ther propertes, but casual observaton of dfferent vulnerablty ndcators calculated on the bass of the Croatan data reveals several nterestng features. Frst, as can be seen from the Table 1, there are wde varatons n the levels of vulnerable households on the bass of dfferent crtera. Moreover, there s very lttle overlap n presented ndcators so all of the ndcators together cover a large porton of all households, n excess of 40% for some combnatons, whle there are a few vulnerable households accordng to multple crtera. Further on, although there has not been much dynamcs durng the observed perod, ndcators have often moved out of lne wth each other. Because t s obvous that conclusons of studes on household dstress crtcally depend on the propertes of the ndcators used, of whch very lttle s known, ths paper n goes some way back to observaton of changes n debt determnants and debt dstrbuton. 3. Methodology In order to capture n full the effects of the rsng household ndebtedness on the fnancal stablty, t s mportant to account for the changes n the whole dstrbuton of household debt. Lterature on household debt determnants s n that respect a natural startng pont. However, papers prepared n ths tradton usually rely on standard OLS regresson (Del-Ro and Young, 2005) or Tobt model (Magr, 2002 and Crook, 2006) that dentfes the determnants of the amount owned by households at the mean of the dstrbuton, thereby gnorng the effects at the debt dstrbuton tals whch may be the most mportant from the fnancal stablty pont of vew. As Fgure 2 ndcates, durng observed perod the Croatan household debt dstrbuton moved n lne wth the aggregate level of household ndebtedness, as can be seen from the rght-sde shft of the whole dstrbuton of household debt. However t became asymmetrc, ndcatng a possble ncrease n the ranks of vulnerable household. For these reasons we employ the quantle regresson (QR) analyss that allows for dentfcaton of the effect of dfferent households' demographc and socoeconomc characterstcs upon the total amount of debt across the whole dstrbuton of the ndebted households. The quantle regresson model was frst ntroduced by Koenker and Bassett (1978). It can be vewed as a locaton model n whch quantles of the condtonal dstrbuton of the response varable (log credt n our applcaton) are expressed as a functon of the observed covarates (Koenker and Hallock, 2001). It s assumed that the condtonal quantle of the response varable s lnear n covarates (Buchnsky, 1998a),.e. ' Quant y x x, where Quant u / x 0 and 0,1 / (1) The coeffcent vector s estmated as a soluton to mnmzaton problem where absolute errors are asymmetrcally weghted wth weght on postve errors and weght 1 on negatve errors (Kuan, 2007):

10 1 mn n ' y ' x ' 1 y x : y x : y x ' (2) The amount of the loans s observable only for those households who are actually ndebted. Therefore the selecton of the ndebted households n the sample s not random because t s determned by household's decson to apply for a loan and bank's decson to approve the loan. Both decson processes are based on the evaluaton of households' soco-economc and demographc characterstcs. Snce dentfcaton of the debt determnants s preformed on the non-random sample of only ndebted households, the problem of the sample selecton bas s lkely to emerge. Also correlaton between the household's decson to partcpate n the credt market and the amount of the bank loan, could gve rse to ths selectvty bas. At the end such bas caused by nadequate sample selecton can nfluence the outcome results of our analyss 4. Standard sample selecton model whch allows for sample selecton bas correcton s gven by (Schafgans, 1998): Y * X, 0,1 D I( Z e 0) and * Y Y D for 1,..., n o (3) where ( Y D, X, Z I s the ndcator functon. The frst equaton n the model s usually referred to as an outcome equaton and the second * equaton s the selecton equaton. In our analyss Y represents the possble amount of debt that each household n the sample of all households would carry dependent on ts characterstcs and D s dummy varable ndcatng whether the ndvdual household s n fact ndebted or not. Therefore the observed amount of loan for household s gven byy. Characterstcs nfluencng the household's decson to apply for a loan,.e. bank's decson to grant the loan are gven by Z and the determnants of the amount of loan are gven by X, where X varables are a subset of the Z varables., ) are observed random varables and The standard approach to the estmaton of the represented sample selecton model s the Heckman's two-step procedure 5 accordng to whch frst a probt regresson s estmated on the decson to partcpate on the credt market. Probt results are then used to compute the nverse Mll's rato. Fnally, an ordnary least squares regresson s appled to the amount of loan, where n addton to the explanatory varables the nverse Mll's rato s ncluded. Ths approach assumes that, e are bvarate normally dstrbuted, ndependent of X, Z wth zero mean and unknown covarance matrx (Schafgans, 1998). However, devatons from the normalty assumpton can lead to nconsstent and based estmator (Schafagans, 1998) In order to control for the non-random bas, a modfcaton of a two-stage Heckman model for the quantle regresson proposed by Buchnsky (1998b) s 4 For more on the econometrc mplcatons of sample selecton bas see Heckman (1979). 5 For more see Heckman (1976).

11 employed. Accordng to ths method, a condtonal quantle of the observed amount of household credt depends on, apart from the household s specfc characterstcs, a bas term of the unknown form (Buchnsky, 1998b). Followng Buchnsky (1998b) and Albrecht et al. (2008) we estmate: Quant y / Z z x ˆ h z ; 0,1 (4) The vector X conssts of the observed socoeconomc and demographc characterstcs of households that carry debt. Apart from the covarates ncluded n X, vector Z contans at least one addtonal varable that nfluences the probablty of household's partcpaton n the credt market, but whch must be uncorrelated wth the amount of the debt. In our case, resdence n rural areas s used as a proxy of hgher costs related to obtanng loans due to low densty of bankng branches, whle nvestments n lfe nsurance schemes s used as sgn of fnancal sophstcaton, both of whch standard n the lterature of debt determnants (Marg, 2002 and Ruz-Tagle and Vella, 2010). The sample selecton correcton term h z s of the unknown form, quantle specfc and doesn t assume normalty. In order to approxmate the functon h z we adopt a two-step procedure proposed by Buchnsky (1998b). Frst the credt market partcpaton selecton parameter s estmated usng dstrbuton free semparametrc least squares estmaton of sngle ndex model ntroduced by Ichmura (1993) on the whole sample of households. Afterwards, the bas term h z s approxmated by a power seres expanson as suggested by Buchnsky (1998b) and Newey (2008) 6, l z ( z l0 h ˆ ) (5) l where represents the transformaton of the sngle ndex z wth the sngle ndex representaton, z ˆ. Fnally, the vector ˆ together wth the coeffcent vector covarates z 7. We contnue estmaton l s obtaned ˆ from the quantle regresson of the log credt on the x and the approxmaton of the bas term ĥ z As households dffer n ther characterstcs, t s necessary to control for changes n the credtworthness of the ndebted households n order to capture the possble negatve effects of the rsng household debt on the fnancal stablty. To acheve that, n the second step of our analyss changes n the characterstcs of the ndebted households at dfferent quantles of debt dstrbuton wll be used n order to dsentangle ther effects from the effects of changes n the estmated QR coeffcents on the rse of total household debt durng observed perod. Households' specfc socoeconomc and demographc characterstcs are used to proxy for household credtworthness and estmated QR coeffcents approxmate the possble relaxaton of bank's lendng standards and/or wllngness of certan types of household to take more debt 8. 6 In our study bas term was approxmated by polynomal of order 5. 7 Any functon of the z can be used, ncludng the sngle ndex. For more see Newey (2008) and Buchnsky (1998b). 8 However, the ntercept n the equaton (4) s not dentfed snce t s dffcult to separate the ntercept 0 ˆ from the frst term n the power seres approxmaton of the selecton equaton ˆ 0. For more on estmaton of the ntercept n sample selecton model see Andrews and Schafgans (1996).

12 aganst ther ncomes. Decomposton method proposed by Machado and Mata (2005) wll be employed n order to separate the effects of standard relaxaton vs. household characterstcs on the debt dstrbuton. Ths technque extends the tradtonal Oaxaca-Blnder decomposton of these effects on mean (Oaxaca, 1973) to the entre dstrbuton. The Machado-Mata decomposton s wdely used n the lterature on wage nequalty for whch t s prmarly developed 9. As far as authors are aware, t has never been used to analyze changes n the patterns of household debt. In order to approxmate the mplct scorng models of the banks two counterfactual margnal credt denstes are generated: a margnal densty functon of the log credt that would preval n the 2008 f the households' characterstcs were the same as n 2005 and a margnal densty functon of the log credt that would arse n 2008 f the "returns" to households' characterstcs n 2008,.e. mplct scorng models of banks, were the same as n X X X resdual X X (6) Ths decomposton allows us to approxmate the contrbuton of the banks' evaluaton of the household's characterstcs n the process of the loan approval (the frst term n (6)) from the contrbuton of the mprovement of the household's credtworthness n perod between 2005 and 2008 (the second term n (6)) to the observed growth of the household debt. Resdual represent the part of household debt change unaccounted for by the estmaton method. In the same way the contrbuton of the ndvdual household characterstc to rsng ndebtedness can be measured. 4. Data In our analyss we use mcro data from the Households budget survey (HBS). Snce 1998 Central Bureau of Statstcs has been annually conductng HBS on the random sample of prvate households 10 n Croata. The data on ncome, wealth and most household consumpton expendtures s collected contnuously durng 12 months perod wth changng surveyed subsample of prvate households every two weeks. There s no panel part of the sample 11. Snce approprate weght s assgned to every surveyed household,.e. the number of households n the populaton that surveyed household represents, calculaton of aggregate estmates for populaton s enabled. Apart from the household-level data on ncome and expendtures, HBS gves nsght nto socoeconomc and demographc characterstcs of surveyed ndvduals, allowng for analyss of ndebtedness of households accordng to the dsposable ncome brackets and dfferent characterstcs of households' head. Ths advantage of HBS and other mcro data sources s mportant for the household ndebtedness analyss snce debt ncdent s not equally dstrbuted amongst households of dfferent age, sex, educaton or area of resdence See Machado and Mata (2005), Albrecht and Bjorklund (2003), Albrech et al. (2008), Nestć (2010) etc employ t n order to decompose gender wage dfferental. 10 Household s every famly or other communty of ndvduals who lve together of spend ther ncome together for coverng the basc exstental needs. Household s also every person who lves alone (CBS, 2010). 11 The sample frame used for selecton of dwellngs occuped by prvate households n 2008 was based on the Census 2001 data. 12 For more detals on HBS see CBS, 2010.

13 Apart from obvous advantages that data from household-level surveys have n relaton to macro data, there are also several dsadvantages that should be kept n mnd. The bggest drawback of most surveys s undervaluaton of household dsposable ncome and debt (Daras and Tyrowcz, 2009). Lower level of aggregate household ncome and bank loans compared to macro data also appears n HBS, due to sgnfcant dstrust and unwllngness of households to completely and correctly reveal the sources, values and structure of ther ncome and debt and also possbly poorly representatve sample n respect to the hgh ncome households. Compared to avalable macro data, aggregate household dsposable ncome from HBS s on average 27% lower durng the observed perod. However, the aggregate household debt s also unevaluated for some 46% so devaton of dfferent measures of relatve household ndebtedness can be tolerated. For the purpose of our households ndebtedness research we use HBS for years 2005 and 2008, farly recent perod durng whch a contnuous expanson of household debt took place. Before employng the proposed analytcal framework the sample was cleaned from dentfed errors and omssons and households that choose not to answer and/or ddn't know the answer to the questons about the level of ther dsposable ncome and/or the amount of debt owned were removed from the sample 13.In order not to further reduce the sample sze, for several dentfed households wth only one lackng data n data matrx, the mssng value was replaced wth the explanatory varable's mean 14. The estmaton of the probablty of household holdng debt (sample selecton equaton) was performed on the whole sample of households, whle dentfcaton of the determnants of the amount of debt owed (outcome quantle regresson equaton) was based on the sub-sample of ndebted households 15 regardless of the type of ther loan. The dependant varable n selecton equaton s a bnary varable that equals 1 f household has some type of bank loan and 0 f not. The second dependant varable n output equaton s natural logarthm of the observed total amount of household's loans. A specal attenton wll be gven to the households carryng the hghest amounts of debt n order to montor changes n ther determnants of ndebtedness as well as changes n the resultng concentraton of household debt durng the observed perod 16. In Table 2 some descrptve statstcs for the key varables s gven for the whole sample of surveyed households n two observed years and three sub-samples: households wth no loan, households wth some type of bank loan and households wth bank loan taken durng the last year. In both of the observed years around 32% of surveyed households had some type of bank loan, although some mld rse of the proporton of ndebted households can be observed durng ths perod. Strong growth of Croatan economy durng observed perod gave boost to the rse of households dsposable ncome and facltated satsfyng down-payment condtons as well as gvng rse to optmstc expectatons, thereby ncreasng demand for loans. At the same tme, rsng competton n the bankng sector lowered nterest rates and non-prce lendng standards of banks. From Table 2 t s evdent that households wth debt labltes n 2005 had around 21% hgher dsposable ncome per household member than households wth no bank loan. However, ths dfference between ndebted households and those wthout any debt was sgnfcantly reduced over the same perod, ndcatng that loan expanson took place 13 The whole sample sze for 2005 and 2008 s 2651 and In years 2005 and 2008 these households account for 0.26% and 0.20% of all households n the sample. In the sub-sample of ndebted households they make 0.83% and 0.60%, respectvely. 15 The sample sze of ndebted households for 2005and 2008 s 845 and 1003, respectvely. 16 In the observed years the two hghest decles account on average for 56% of the total household debt.

14 amongst less credtworthy households. At the same tme, easer access to the loan market and cheaper borrowng s reflected n the steep rse of the average value of new loans n comparson to the average household dsposable ncome durng ths perod, whch s a sgnal of rsng vulnerabltes. Apart from level of dsposable ncome, partcpaton n the credt market and the amount of debt s also postvely correlated to the level of educaton of household s head snce majorty of households wth some type of bank loan n both observed years have mddle or hgh level of educaton whle around 45% of households wth no debt labltes haven t even fnshed hgh school. Level of educaton reflects the potental for future ncome growth, but t also mples easer collecton and evaluaton of nformaton needed before decdng whether to apply for a loan or not (Magr, 2002). Age of the household head s another mportant factor n explanng credt market partcpaton. As suggested by the lfe-cycle permanent ncome hypothess, young households wth expectatons of rapdly growng ncomes and hgh margnal utlty of consumpton are more lkely to demand debt (Crook, 2005). After certan age threshold debt ncdent and the amount borrowed s expected to decrease as household need for satsfyng basc lvng condtons and expendtures dmnshes. Table 2 shows that households carryng debt are on average around 9 years younger than households wth no debt. In both observed years percentage of ndebted households grows wth age of the household's head wth slghtly more than 55% of all ndebted households aged between 40 and 59 years, after whch debt ncdent decreases. Beng marred and havng bg famly also ncreases probablty of havng a debt, as larger famles especally wth young chldren usually have hgher lvng expenses. Households whose head s workng are more lkely to have a bank loan, especally f he s employed n the publc or prvate company engaged n tertary actvty and has permanent employment contract wth full workng tme. Men are usually reported as a household head n he HBS, but those households are dsproportonably more lkely to have debt labltes. The area of resdence s another factor affectng the decson to apply for a loan. In the whole sample almost 50% of households lve n rural areas. Ths percentage s even hgher for households that don t have debt, whereas those that do are more lkely to be lvng n towns and ctes. Ths could be the consequence of densty of bank n less populated rural area but also poor educatonal profle of populaton lvng n such muncpaltes whch rses entry cost n the debt market. Households nvestng n lfe nsurance are probably more fnancally lterate and n probably better postoned to apply for a loan, but also some types of bank loans, especally resdental loans, are also lkely to be collateralzed by lfe nsurance contract, makng t useful n dentfcaton of wheatear or not household carres debt. As HBS data shows, lfe nsurance s much more present n the sub-sample of households wth some type of bank loan than n the sub-sample of households wth no debt. 5. Estmated results 5.1. Results of the selecton equaton and the outcome equaton The emprcal semparametrc least square (SLS) estmates of the selecton equaton are presented n Table 3. Most explanatory varables take the form of dummy varables rather

15 than contnuous varables. Therefore, presented estmaton s referenced to baselne household whose head s a marred male, aged between 50 and 59 years, workng permanently, full-tme hour n a prvate company dealng n tertary actvty. He owns and lves n a real-estate n urban area wthout housng loan labltes and doesn't have lfe nsurance. Estmated parameters are dentfed up to an unknown scale, so all coeffcents are normalzed relatve to the value of the coeffcent of the only contnuous explanatory varable, logarthm of household's dsposable ncome. Although most dentfyng varables have very smlar effects on the probablty of household havng bank loan n both observed years, the effect of some explanatory varables changed durng observed perod. As expected, household's dsposable ncome has postve mpact on the household's partcpaton n credt market n both years. Even though theory suggests that probablty of havng bank loan should decrease wth rsng current ncome (Magr, 2002), the obtaned estmates could ndcate that ndebted households n our sample have low or ntermedate level of ncome whch s not hgh enough to fnance all ther expenses, forcng them to turn to loan market. However, as the number of the ncome earners n the famly ncreases, household's need for borrowng n order to meet expendtures decreases. Somewhat surprsngly, havng hgher educatonal qualfcatons also lowers probablty of havng bank loan. Ths can be due to much better fnancal stuaton of hghly educated ndvduals compared to low or even mddle educated ones, especally n the later and ncome peakng stages of ther professonal career. Probablty of havng a bank loan decreases wth age for all age brackets. Havng pror bank loans also ncreases the probablty of repeated borrowng. Men are more lkely to have debt than woman and famles who have a housng loan, together wth renters n 2008, tend to have hgher probablty of havng debt than homeowners wth no mortgage. Regardng labor market status, n 2005 ndvduals employed n prvate company had hgher probablty of partcpatng n credt market than all other workng ndvduals regardless of ther employment status. Retred were also less lkely to have debt, but oddly others unemployed ndvduals had hgher probablty of beng ndebted. Hgher probablty of debt ncdent among households' whose head s unemployed can be the consequence of ther effort to overcome a shorter declne n current ncome due to job loss n order not to have to downsze ther expendtures. Ths s especally the case among the newly unemployed ndvduals who are expectng to reemploy soon. However, n 2008 only households whose head was workng n publc company, where less volatle wages reduce the future ncome uncertanty, had hgher probablty of havng debt than those workng n prvate sector. The effect of other employment characterstcs on the probablty of credt market partcpaton s somewhat nconclusve. As expected, households lvng n rural areas have lower probablty of havng debt than those lvng n more urbanzed muncpaltes, due to lmted bank supply, lower educatonal qualfcaton and more wdespread presence of nformal credt market (Magr, 2002). Even thought t was expected that debt ncdent would be hgher among households who have lfe nsurance, results suggest that ths was the case only n In the second step of analyss the amount of bank loan was estmated on the sub-sample of only ndebted households usng quantle regressons n order to capture the changes of the effect of a varous explanatory varables at dfferent ponts of condtonal debt dstrbuton. In Table 4 and Table 5 we gve the results for log credt estmaton wthout and wth sample

16 selecton bas correcton for both observed years at nne condtonal quantles. The ndependent varables used are the subset of the explanatory varables used n the selecton equaton. Results presented n Table 4 suggest that among the dentfyng varables statstcally sgnfcant on almost the entre debt dstrbuton and wth expected sgn are varables on the demand sde: current dsposable ncome, age of the household's head, the type of actvty the head s workng n, part-tme work dummy and tenure status whch s also connected wth varables that reflect households debt supply: total number of household's bank loans and dummy dentfyng household wth housng loan taken durng the last 12 months. Household's dsposable ncome has postve effect on the amount of debt held n both observed years at all condtonal quantles. It s the man varable explanng the dfference n holdngs of bank loans. Comparng the results for 2005 and 2008 we can observe a rse n the amount of debt supported by a gven sze of ncome, accountng for most of the household debt accumulaton, and especally pronounced at the tals of the debt dstrbuton. In 2005 households whose head s employed n company dealng n prmary or secondary actvty on average had lower amount of debt than households whose head works n tertary actvty, especally at lower or the hghest level of ndebtedness. However, n 2008 these varables showed no sgnfcance n explanng the amount of debt. Dummy varable ndcatng shorter workng hours than usual also proved to be sgnfcant wth negatve sgn n explanng amount of debt owed, partcularly n The mpact of the age of households' head on the amount of debt depends on the ndvdual's poston n the lfe-cycle. The effect of the head's age on the debt s consstent wth the theoretcal lfe-cycle model of consumpton; t s postve untl person reaches hs fftes, afterwards t becomes negatve. As expected there s a strong postve relatonshp between the amount of debt held and the number of loans, especally f on of them s resdental loan whether new (.e. taken durng the 12 months perod pror to the survey) or old one (homeowner wth housng loan). Ths effect s observed n both years suggestng that households who already have some type of loan wll have greater propensty for new borrowng. Also banks can be more wllng to borrow to households who are already ther clents snce all relevant nformaton about households' characterstcs and regularty of servcng pror credts are already avalable. However, the mpact of havng a resdental loan on the ndebtedness decreased n 2008 compared to However, when corrected for sample selectvty bas the sgnfcance of explanatory varables, together wth the sze of ther estmated coeffcents, changes sgnfcantly as can be seen from Table 5. Household's dsposable ncome and dummy varable dentfyng homeowners that repay housng loan reman the most mportant varables n explanng the amount of debt hold n both years wth unchanged sgn of ther effect on debt sze across all condtonal quantles. In 2005 sample selecton bas on the estmated coeffcents of these two explanatory varables was generally downward and n 2008 mostly upward. The bas s the most pronounced for debt dstrbuton tals. After the correcton, dummy varable ndcatng households wth new housng loan was also mportant n determnng the sze of debt along the entre debt dstrbuton, especally n Dfferent age brackets have statstcally more sgnfcant explanatory power at dfferent condtonal quantle n 2008, together wth parttme work whch negatvely nfluences the amount of debt at the hgher quantles.

17 4.2. Machado-Mata decomposton of household debt growth Fnal step n household ndebtedness analyss ncludes assessng the mplcatons of changes n mpacts of varous explanatory varables whch proxy for the evoluton of "mplct" scorng models and credt polces of banks durng observed pre-crses perod. In order to decompose the rse of households ndebtedness between 2005 and 2008 nto part attrbutable to changes n households' credtworthness (observed households' characterstcs) and changes n banks' credt standards we follow earler descrbed Machado-Mata decomposton 17. The results of MM decomposton employed on sample uncorrected for sample selectvty bas are presented n Fgure 3. Growth of ndebtedness can be observed throughout the whole dstrbuton of household debt wth average rse of 27%. Rse of debt was the strongest among the hghly ndebted households, reachng 39% at the 99-th percentle. At the same tme, characterstcs of ndebted households mproved, whch was postvely reflected on the households' ndebtedness dynamcs. However, effect of mproved households' credtworthness can explan only around 7 percentage ponts of the rse of accumulated household debt across the whole ndebtedness dstrbuton. Snce the effect of estmated coeffcents s quanttatvely more mportant then effect of mproved households characterstcs at each estmated quantle, relaxaton of banks' lendng standards and/or greater appette of some households to take more credt were the man drvers of households debt growth between 2005 and On average ther contrbuton to overall debt rse was about 18 percentage pnts across whole dstrbuton. The mpact of banks' loosened mplct credt polces and growth of households' tendency to borrow was the strongest among the hghly ndebted households, as t explans around tree quarters of ther debt ncrease durng the observed perod. At the same tme the mprovement of ther credtworthness was the slowest. Overall effect of mproved households' credtworthness s decomposed nto ndvdual contrbutons of several characterstcs that showed to be mportant n determnng the sze of debt and whch represent both the demand and supply sde of household credt market. Fgure 4 suggests that growth of household current dsposable ncome had postve effect upon household debt across almost the whole dstrbuton, accountng for around 11 percentage pnts ncrease of ndebtedness. However, at the hghest condtonal quantles the rse of dsposable ncome between 2005 and 2008 was much slower. Number of bank loans that household carres s another mportant varable. Its mpact on the debt rse s negatve n the frst half of dstrbuton. However, after medan t rses exponentally, postvely nfluencng the debt dynamcs.. Rsng number of bank loans had postve effect on the amounts of household debt, but hgher number of loans cannot be consdered as an mprovement of households' credtworthness. So f the mpact of the number of loans household carres s taken nto account, households' credtworthness observed at the hghest condtonal quantles would actually deterorate durng observed perod (see Fgure 3). Age, new housng loan dummy and educaton, whch are expected to mprove credtworthness of ndebted households, ddn't have a notceable contrbuton to the households' debt dynamcs durng observed perod, except among the hghly ndebted households where ther mldly postve contrbuton to ncreasng ndebtedness can be observed. Presence of the sample selecton bas can have an mpact on the estmated changes to households' credtworthness and banks' credt polces. Decomposton of households' debt 17 MM decomposton was repeated 100 tmes.