Distinguishing the Components of Household Financial Wealth: the Impact of Liabilities on Assets in Euro Area Countries

Transcription

1 Dstngushng the Components of Household Fnancal Wealth: the Impact of Labltes on Assets n Euro Area Countres Merke Kukk Workng Paper Seres 2/2014

2 The Workng Paper s avalable on the Eest Pank web ste at: For nformaton about subscrpton call: ; Fax: e-mal: publcatons@eestpank.ee ISBN Eest Pank. Workng Paper Seres, ISSN ; 2

3 Dstngushng the Components of Household Fnancal Wealth: the Impact of Labltes on Assets n Euro Area Countres Merke Kukk * Abstract The paper nvestgates the nterdependence of household fnancal labltes and assets, wth specal focus on the mpact of labltes on households holdngs of fnancal assets. The paper uses the new ECB Household Fnance and Consumpton Survey from coverng euro area countres. The paper estmates a system of equatons for households fnancal labltes and assets, takng account of endogenety and selecton bas. The results ndcate that hgher household labltes are related to lower holdngs of fnancal assets. The fndngs are consstent wth the hypothess that wder use of credt leads to lower savngs. The paper hghlghts that the dstncton between the components of households wealth provdes addtonal nsghts nto households fnancal behavour. JEL Codes: D14, E21, D12 Keywords: household debt, household wealth, fnancal assets, labltes, fnancal vulnerablty Authorʼs e-mal address:merke.kukk@ttu.ee. The vews expressed are those of the author and do not necessarly represent the offcal vews of Eest Pank. * The research was done when the author was a vstng researcher at Eest Pank. The author author would lke to thank the partcpants n semnars at Eest Pank and Tallnn Unversty of Technology, EMS 2014, IISES 9 th Academc Conference, as well as Karsten Staehr, Tar Rõõm and Jaanka Merküll for ther valuable comments. The results publshed and the related observatons and analyss may not correspond to the results or analyss of the data producers.

4 Non-techncal summary A standard approach for descrbng the fnances of a household s to use net wealth, where a household s labltes are deducted from ts assets. However, the components of net wealth,.e. fnancal labltes, fnancal assets and real assets, are very heterogeneous across households. Ths means that households wth the same value of net wealth may have dfferent levels of assets and labltes. Furthermore, the components of net wealth may be nterdependent, meanng that households may consder ther ndebtedness when they make decsons about ther fnancal assets, and the balance of labltes may depend on the balance of fnancal assets. The penetraton of debt and the volumes of debt have ncreased n developed countres over the last three decades and the fastest growth has occurred snce In lght of these developments, there has been lmted dscusson about whether and how ndebtedness affects the behavour of households beyond ther borrowng decsons. There are numerous studes whch deal wth household borrowng and the fnancal vulnerablty of ndebted households but there are fewer studes whch take a holstc vew and analyse fnancal assets and labltes jontly. There s a hypothess that the avalablty of credt reduces the need for precautonary savngs as ncome shocks can be smoothed by borrowng, meanng fewer assets are needed for self-nsurance aganst consumpton rsk. On the other hand, ndebtedness ncreases the fnancal vulnerablty of households, especally durng the perod of credt tghtenng when t s dffcult for households to borrow to smooth negatve shocks. Therefore durng tmes of fnancal crss, ndebted households need to keep hgher levels of savngs for precautonary reasons. The paper uses data from the recently ntroduced Household Fnance and Consumpton Survey HFCS. The paper uses the data of 13 euro area countres from the frst wave of the HFCS, whch was carred out n The data relate to the perod when households had experenced the man economc shocks of the recesson and had presumably adjusted ther fnances to these major shocks. The paper nvestgates whether and how the labltes and fnancal assets of households are related. The estmatons show that households labltes mpact ther fnancal assets negatvely whle no sgnfcant effect was found from fnancal assets on labltes. The negatve relatonshp between labltes and fnancal assets remans after controls for other debt related varables are ncluded. The results are confrmed by a large number of robustness tests. 2

5 The results suggest that ncreasng volumes of household debt are related to lower ncentves to keep fnancal assets, whch also apples durng a recesson. The extent to whch lower buffer stocks affect the fnancal vulnerablty of households depends on the ablty of the households to nsure themselves aganst fnancal and consumpton rsks n other ways. Accordng to the HFCS, the choce by ndebted households to nsure themselves aganst negatve shocks by addtonal borrowng or by recevng fnancal assstance from relatves s not evdently dfferent from that of households wthout debt. The negatve relatonshp between labltes and fnancal assets may therefore ncrease the fnancal vulnerablty of ndebted households as they have fewer resources avalable when they are ht by a negatve shock. In addton to the drect negatve effect of labltes on fnancal assets, there s an ndrect postve effect for households that wsh to deleverage. Accordng to the HFCS, about half of ndebted households are savng to pay back debt n euro area countres. The estmatons show that ndebted households who save to pay back ther debt have fewer labltes and more fnancal assets than ndebted households that are not savng, all else beng equal. As the motvaton for deleveragng s related to the economc downturn, these results apply specfcally to the tme of the recesson. The paper provdes evdence for the nterdependence between fnancal assets and labltes. In the lght of the ncreased ndebtedness of households t s partcularly mportant to understand how households labltes affect ther other fnancal decsons. 3

6 Contents 1. Introducton The lterature lnkng households borrowng and savng decsons The methodology A model for estmatng the nterdependence of fnancal labltes and fnancal assets A model for estmatng the nterdependence of fnancal labltes and fnancal assets The dataset and descrptve statstcs The Household Fnance and Consumpton Survey Descrptve statstcs The estmatons for fnancal assets and labltes The baselne estmatons Estmatons ncludng other debt-related varables Addtonal robustness tests Fnal comments References Appendx A: The robustness estmatons

7 1. Introducton Ths paper studes the nterdependence of households fnancal assets and labltes n European countres. Most studes treat household labltes as negatve assets and relate households decsons to ther net wealth. However, the components of net wealth,.e. fnancal labltes, fnancal assets and real assets are very heterogeneous across households. Net wealth of 10,000 euros may denote that a household has assets worth 10,000 euros and no labltes, but t may equally well denote that a household has labltes of 100,000 euros and assets of 110,000 euros. Carroll et al. (2013) hghlght n ther study of the margnal propensty to consume across dfferent net wealth balance that the assumpton of smlar consumpton behavour among households wth the same net wealth but dfferent wealth components s mplausble. Ths paper rases another queston, namely whether households consder ther ndebtedness when they make decsons about ther fnancal assets. The mplcaton of household ndebtedness for household behavour s an mportant topc as household debt volumes have ncreased markedly n developed countres over the last three decades. The largest changes occurred n the 2000s not only n the USA but also n Europe, where the household sector s debt to dsposable ncome has ncreased from 70 per cent n 1999 to 95 per cent n 2008 (Lojschova et al. (2011)). There s a long lst of research about household borrowng and the fnancal vulnerablty of ndebted households, but there has been less dscusson about whether and how ndebtedness affects the behavour of households beyond ther borrowng decsons. Moore and Palumbo (2010) document how greater debt on household balance sheets ncreased the fnancal stress n the household sector around the onset of the recesson. The ndebtedness has mplcatons for the consumpton behavour of households as well as for ther choces regardng fnancal assets. The paper nvestgates whether and how households labltes affect ther holdngs of fnancal assets. Accordng to conventonal consumpton theory, households keep buffer stocks to nsure aganst consumpton rsk. In the buffer stock models the precautonary savng ncreases wth ncome uncertanty and hgher rsk averson. The study of Carroll et al. (2012) shows analytcally that more flexble credt markets decrease the motvaton for precautonary savng. If households are able to borrow when they suffer a negatve ncome shock, they reduce the target level of ther buffer stock, meanng they decrease ther fnancal asset poston. On the other hand, consumpton smoothng through borrowng s hampered durng an economc downturn when the bankng sector s 5

8 tghtenng the credt condtons. On top of that, debt servcng may become an unsustanable burden after ncome or wealth shocks and may lead to fnancal nstablty for the ndebted households (Barba and Pvett (2009)). In ths case, ndebted households would need hgher buffer stocks n the face of credt tghtenng and consumpton rsk durng a recesson. There s a lack of emprcal evdence on the effect of ndebtedness on households fnancal asset holdngs n the mddle of a recesson. Ths paper sheds more lght on the relatonshp between labltes and fnancal asset holdngs. The paper uses data from the recently ntroduced Household Fnance and Consumpton Survey HFCS (ECB (2013a), ECB (2013b)). The paper uses the frst wave of the HFCS, whch was mplemented n n 15 euro area countres. As the survey methodology was smlar n all the countres, common estmatons can be mplemented for the whole euro area. The survey results reflect the stuaton of the households after the man economc shocks of the recesson,.e. after the adjustments of household fnances to these major shocks. The paper contrbutes to the academc lterature n several ways. Frst, the lnkage between the dfferent components of fnancal balance sheet of households s lttle explored n the lterature. Most of the studes focus on determnants of fnancal labltes, determnants of fnancal assets or determnants of net wealth, but dfferent wealth components are not nvestgated jontly. Most of the determnants of labltes and assets are the same, but the effect on borrowng and savng mght be dfferent. Ths study s the frst to present results from smultaneous estmatons. Second, the paper provdes evdence for the nterdependence between fnancal assets and labltes whch s not taken nto account n most studes. In the lght of the ncreased ndebtedness of households t s mportant to understand how ndebtedness affects decsons about the holdngs of fnancal assets. Thrd, there are few studes that nvestgate fnancal behavour n the whole set of euro area countres as there has been a lack of approprate mcro data. The recently ntroduced Household Fnance and Consumpton Survey wdens the range of fnancal behavour of European households that can be examned. The current paper s among the frst to explot ths opportunty. The rest of the paper proceeds as follows: Secton 2 provdes a bref overvew of the theoretcal and emprcal lterature. Secton 3 comprses the hypothess and the model to be estmated. Secton 4 ntroduces the dataset and delvers the man features of the varables of man nterest. Secton 5 provdes the results and Secton 6 summarses the emprcal fndngs. 6

9 2. The lterature lnkng households borrowng and savng decsons The lterature on the fnancal decsons of households can be dvded nto several separate areas: one focuses on topcs related to household debt, such as developments n credt markets, household borrowng decsons and the fnancal vulnerablty of households due to ndebtedness, and another focuses on topcs related to households savng decsons, such as determnants of savngs and asset accumulaton. Households fnancal labltes and assets are not nvestgated jontly n the majorty of studes, although several studes use net wealth as an mportant determnant of the dfferent fnancal decsons taken by households. There are a few studes whch nvestgate the effect on household fnancal assets of relaxng credt market constrants and easer borrowng condtons. The effect can be postve, meanng that a wde choce of credt nstruments may encourage people to fnance ther equty holdngs by borrowng, whch wll result n them havng more fnancal assets. Ths hypothess has not been nvestgated thoroughly. Davs et al. (2006) note that a wedge between the cost of borrowng and the rsk-free nvestment return argues aganst leveraged equty holdngs. On the contrary, ther model shows that households do not explot ther borrowng capacty to ncrease ther fnancal poston. There s another hypothess that has receved broader scrutny. Several studes examne the nteracton between fnancal deregulaton and household savng rates. Bayoum (1993) fnds a negatve relatonshp between fnancal deregulaton and the savng rate n the UK n the s. Japell and Pagano (1994) present a model that shows how fnancal deregulaton,.e. lowerng of lqudty constrants, lowers household savng rates. They fnd emprcal evdence n favour of ths hypothess usng cross-sectonal data from the OECD countres. The topc has reganed attenton recently. Carroll et al. (2012) use a model derved by Carroll and Toche (2009) to explan how a relaxaton of credt constrants affects household savng negatvely. The mechansm works through the decrease n precautonary savngs, whch brngs the target wealth of households to a lower level. As households can nsure ther consumpton rsk usng credt markets, they can hold lower buffer stocks than when borrowng s not avalable. Carroll et al. (2012) use aggregate data for the USA to show that ncreased access to credt n the US from the 1980s untl 2007 contrbuted sgnfcantly to the declne of the savng rate. However, ths effect has not been nvestgated at the household level. 7

10 The studes of Debelle (2004), Grouard et al. (2006) and Barba and Pvett (2009) hghlght that the senstvty of households to negatve shocks has ncreased due to the ncreased leverage of ther balance sheets. Japell et al. (2013) argue that household ndebtedness s related to the ncreased fnancal fraglty of households. Man and Suf (2010) have found a negatve relatonshp between the growth of household debt n the perod before the global fnancal crss and consumpton durng the recesson that followed the crss. They suggest that ndebted households are more senstve to house prce declnes. The ndebtedness also mpacts the fnancal asset holdngs of households; snce ndebtedness ncreases the senstvty to negatve shocks, households need to keep hgher levels of savngs for precautonary reasons. The model of Challe and Ragot (2012) predcts that households facng hgher ncome rsk accumulate more precautonary wealth; however, they do not nvestgate the sources whch mght ncrease the senstvty to ncome rsk. Carroll et al. (2012) show that the aggregate household savng rate ncreased n due to ncreased ncome uncertanty, collapsng household wealth and tghter credt markets. The lterature cted above suggests that there s a lnkage between household labltes and assets. However, there s lack of studes whch nvestgate ths relatonshp n detal. Several macroeconomc models whch explan aggregate developments n household savngs and consumpton ncorporate two types of agents, borrowers and savers (see among others Nakajma (2012), Challe and Ragot (2012)). The frst do not have any savngs but only labltes and the second do not borrow but own fnancal assets. Lookng at the ndvdual balance sheets, t appears that a sngle household holds both labltes and assets at the same tme (Tudela and Young (2005), ECB (2013b)). The standard way to descrbe the fnancal poston of households s net wealth, where labltes are deducted from assets. If a household has more labltes than assets, ts net wealth s negatve. The lterature nvestgates the relatonshp between net wealth and household debt or savngs, see among others Magr (2007), Crook (2001), Arrondel et al. (2013), Costa and Farnha (2012). There are studes whch dstngush the effect of dfferent wealth components on consumpton (see the recent studes of Carroll et al. (2011), De Bons and Slvestrn (2012), Sermnska and Takhtamanova (2012), Dynan (2012)) but the nterdependence of wealth components s underexplored. The studes of Brown and Taylor (2008) and Brown et al. (2013) dsentangle net wealth and nvestgate each wealth component separately. Brown and Taylor (2008) examne the determnants of household debt and assets; 8

11 the latter ncludes the value of the house, whch s real asset, n addton to the fnancal assets. They use the Brtsh Household Panel Survey (BHPS), the German Soco-Economc Panel (GSEP) and the Panel Study of Income Dynamcs (PSID) to explore the determnants for labltes and assets. They suggest that decson makng about (fnancal) assets and labltes s nterdependent and therefore these should be modelled jontly. The nterdependence of labltes and assets s examned n Brown et al. (2013) usng PSID, as fnancal assets are ncluded n the model for labltes and the other way around. They fnd the balance of fnancal assets s negatvely related to total debt but the balance of labltes s postvely related to fnancal assets, although the effect s economcally margnal. They treat fnancal assets and debt as censored varables but do not consder the varables to be endogenous. Wth net wealth alone, mportant nformaton s lost as the composton of net wealth from labltes, fnancal assets and real assets vares substantally across households. Households wth very dfferent levels of labltes and assets may report the same level of net wealth. The effect of dfferent wealth components on households decsons can be dfferent. On top of that, the dfferent wealth components may affect the decsons related to other wealth components. Ths s a relevant topc as the structure of household wealth has changed due to wder usage of credt. It s mportant to understand how households ndebtedness affects ther decsons about fnancal asset holdngs. Ths revew ndcates a need for a better understandng of the nterlnkages between the assets and labltes of households and n partcular the effect of ndebtedness on the savngs or fnancal asset poston durng the crss. On the one hand households need fewer savngs after the deregulaton of fnancal markets; on the other hand, ndebted households need more savngs to nsure them aganst addtonal rsks durng a crss. The net effect durng the crss s ambguous, although the frst effect may be expected to be stronger than the second. 3. The methodology 3.1. A model for estmatng the nterdependence of fnancal labltes and fnancal assets As there s no structural model for examnng the nterdependence of fnancal assets and labltes, the paper reles on the emprcal models whch are used to nvestgate household borrowng and savng behavour. 9

12 Decsons about the holdngs of fnancal assets and labltes are made n a household at the same tme but the drvers mght be dfferent. There s some emprcal evdence that households treat debt dfferently than savngs, therefore the determnants of fnancal assets and labltes mght dverge (Messner (2013)). In the paper dfferent determnants are allowed for fnancal assets and labltes and the nterdependence between dfferent assets and labltes s taken nto account. Although most of the households own fnancal assets, there s a substantal number of households who do not own any labltes. Labltes can be handled as a censored varable as has been the approach of Brown et al. (2013). However, the lterature on household debt suggests that there s a selecton ssue (see Magr (2007), Duca and Rosenthal (1993) and Cox and Japell (1993)). Gven the assumptons about the selecton ssue, the nterdependence and the endogenety of fnancal assets and labltes, and usng cross-sectonal data, the holdngs of fnancal assets and labltes are modelled as a system of equatons. The system of equatons s gven as: F = α + γ L + X ' β + Z ' φ + ε L = α + γ F + X ' β + Z ' φ + ε 2, (1) where varable F s the household s holdngs of fnancal assets, L s the household s holdngs of labltes, and X s a column vector of exogenous varables that affect both the volume of fnancal assets and labltes. Z 1 s a column vector of exogenous or predetermned varables that affect the volume of fnancal assets, whle the predetermned varables n the column vector of Z 2 affect the holdngs of labltes. The error terms ε 1 and ε 2 reflect the mpact of varous unmeasured factors on fnancal assets and debt. As the unobserved factors may affect a household s decson about fnancal assets and labltes, the errors may be correlated across the two regressons. The exogenous varables are determned outsde the system and they are uncorrelated wth the error terms. The approach s smlar to Daves (2011) and Znn (2013) who estmate a SUR model usng aggregate household data for 34 and 40 countres respectvely, whch enables them to explan cross-country dfferences usng aggregate varables. However, both studes treat fnancal assets and labltes as exogenous varables. The eq. (2) can be estmated by 3SLS where all coeffcents of the full system are estmated smultaneously, takng nto account that the error terms of the two equatons are correlated across the observatons. It combnes the 10

13 system estmaton of seemngly unrelated regressons ntroduced by Zellner and Thel (1962) wth the nstrumental varables method of 2SLS. Frst, the endogenous varables F and L have to be nstrumented to obtan consstent estmators (Greene (2012)). The nstruments for L should be orthogonal to the error term ε 1 and F whle correlatng wth L. The varables n the vector of Z 2,.e. explanatory varables that appear only n the regresson of labltes, can be used as nstruments for L. And the varables n the vector of Z 1,.e. explanatory varables that appear only n the regresson of assets, can be used as nstruments for F. Both regressons of eq. (1) contan one endogenous varable on the RHS. Hence, f each regresson contans at least one exogenous varable whch s not n the other regresson, the model satsfes the order condton for the dentfcaton of the system. The IV estmators are obtaned by estmatng the reduced form equatons: Fˆ = α1+ X ' β1+ Z1 ' φ1 + Z 2 ' φ3 + ε1 Lˆ. (2) = α + X ' β + Z ' φ + Z ' φ + ε 2 2 The reduced form s estmated by OLS and not by SUR as all exogenous varables n the system appear n both equatons. In the second step, eq. (1) s estmated usng ftted values for the nstrumented varables from eq. (2). The rank condton for dentfcaton s met f the set of exogenous varables n the vector of Z 1 truly enters the regresson of fnancal assets and the explanatory varables n the vector of Z 2 enter the regresson of labltes n the reduced form model. Ths happens f the coeffcents of φ 1 are nonzero for the holdngs of fnancal assets and the coeffcents of φ 2 are nonzero for the holdngs of labltes. The equatons are frst estmated separately to obtan the resdual vectors 1 ˆε and ˆε 2. The resduals are used for estmatng the covarance cov( ε1, ε 2 j ) =σ 12 0 where and j denote the observatons and j. The estmators are assembled nto ˆ ˆ σ ˆ 11σ12 ε = ˆ σ ˆ 12σ 22 and the weghtng matrx s computed as Ωˆ =Σˆ ε I n where I n s the n n dentty matrx. In the thrd step the system of equatons s estmated jontly by feasble GLS. The coeffcents are estmated as ˆ C ˆ FGLS ( ' C) C' ˆ β = Ω Ω Y where the vector C contans the varables n the vectors X and Z 1 of the regresson

14 on fnancal assets and the varables n the vectors X and Z 2 of the regresson on labltes n eq. (1) (Greene (2012), Ch. 10.). Only a fracton of households declare labltes, and the choce of the volume of labltes s preceded by the decson whether or not to take on the labltes. In other words, there s a latent varable for the sze of labltes L * whch s not observed. We observe the volume of the labltes f the decson s taken to hold debt L = 1, otherwse f L = 0, the latent varable L * s not observed. There s a large number of households whch do not have any labltes for varous reasons such as preferences or other unobserved characterstcs. The effect of labltes on fnancal assets should be dstngushed from the effect of unobserved characterstcs whch correspondngly nfluence the decson about debt ownershp. If the same unobserved characterstcs affect the holdngs of fnancal assets, the estmated coeffcent for labltes would suffer from selecton bas f no correcton s made. Dfferent possble ways to correct for selecton bas are provded by Basen (2011), Nchols (2007) and Vella (1998) among others. The current crosssectonal model wth a system of equatons s complemented wth a selecton equaton n order to correct for the selecton bas: S * * = W ' δ + u, (3) S *, the household where S * > 0 means that the household owns debt. If 0 does not have any labltes. The column vector of explanatory varables n the selecton equaton W contans the explanatory varables n eq. (1) and addtonal varables that affect the probablty of ownng labltes but do not mpact the volume of labltes have to be found. The probt model for the probablty of ownng labltes s estmated and the correcton factor or nverse Mlls rato s calculated as: ' φ( Wδ ) λ = '. (4) Φ( Wδ ) The term λ s the nverse Mlls rato, W denotes the vector of explanatory varables n the debt selecton equaton (4), φ( W ' δ ) s the probablty densty functon and Φ ( W ' δ ) s the cumulatve densty functon. The correcton factor s ncluded as addtonal varables n both regressons of equaton (1) and the model s estmated for the sub-sample of households wth labltes: F = α + γ Lˆ + X ' β + Z ' φ + θ λ + ε L = α + γ Fˆ + X ' β + Z ' φ + θ λ + ε (5) 12

15 where Lˆ denotes the nstrumented varable of labltes, Fˆ s the nstrumented varable of fnancal assets and λ s the nverse Mlls rato estmated by eq. (4). The current model specfcaton uses the two-stage Heckman estmator, meanng that the results of the system are corrected for selecton bas (Heckman (1979)). The man nterest of the paper concerns the effect of labltes on fnancal assets gven by γ 1. As argued by Carroll et al. (2012), ths s expected to be negatve whle the economy s growng but durng a recesson there are some factors that tlt the relatonshp n the opposte drecton. Overall, the net effect s expected to be negatve. Also of nterest s the effect of fnancal assets on household labltes, whch s captured by γ 2. One lne of argument s that the stronger the fnancal poston of the household, the lower the demand for credt (Crook (2006)). Another lne of argument derves from the ablty to use dfferent fnancal products. Households wth a hgher level of fnancal assets can use a wder range of fnancal nstruments to dversfy ther portfolos and these nstruments may nclude both labltes and assets. Emprcal evdence reveals a postve relatonshp between labltes and net wealth (Cox and Japell (1993), Magr (2007)). It s mportant to recognse that ths estmaton framework does not capture fully the dynamc feedback effects between changes n fnancal assets and changes n labltes. Ths can be estmated wth panel data, whle data allow only the search for a sutable cross-sectonal model, meanng the queston about dynamc nteracton s not addressed at present A model for estmatng the nterdependence of fnancal labltes and fnancal assets The choce of the exogenous varables for the models of fnancal assets and labltes n eq. (5) s based on the studes of the determnants of household savngs or household debt. Determnants of fnancal assets Brownng and Lusard (1996) and Attanaso and Weber (2010) gve an overvew of those determnants of households savng behavour whch also determne ther fnancal asset holdngs. Carroll and Toche (2009) propose a structural buffer stock model of consumpton whch produces a target level for fnancal assets. Household ncome, ncome expectatons, ncome uncer- 13

16 tanty, nterest rates, mpatence and relatve rsk averson are the man determnants of fnancal assets. Several emprcal studes suggest that age, famly composton, educaton and self-employment are mportant for savngs, see among others Brownng and Lusard (1996), Tudela and Young (2005), and Kulkov et al. (2009). Addtonally, a bequest motve s an mportant factor n wealth accumulaton (Gale et al. (1994), Modglan (1988)). As t s assumed that wealth components are nterdependent, real assets and labltes are added as possble determnants of fnancal assets. 1 The ncome and wealth shocks durng the recesson had profound mplcatons for households balance sheets, see the analyss of UK households by Crossley et al. (2013), of US households by Chakrabart et al. (2011) and a model explanng the emprcal evdence by Alan et al. (2012). The lterature fnds that households whch experenced ncome declnes cut back ther consumpton, but also used lqud funds to smooth ther consumpton profle. Therefore t s necessary to control for transtory changes n fnancal assets, whch mght occur due to unemployment or extraordnary low ncome. For the estmatons of eq. (5) the explanatory varables for the fnancal assets whch do not affect the volume of labltes are needed,.e. the set of varables n the vector Z 1 n eq. (5). Ths s challengng as the balance of fnancal assets and labltes s mostly determned by the same socodemographc and economc characterstcs. There s lack of theores to explan the dfferent determnants for household fnancal labltes and fnancal assets. Therefore the emprcal lterature and sample nformaton about the relatonshp between labltes or assets and the household s economc and soco-demographc characterstcs s used for determnng the set of varables n the vector Z 1. The emprcal lterature does not gve bequests as a determnant of labltes but t s an mportant determnant of fnancal assets as pnponted by Gale et al. (1994) and Modglan (1988). Sample nformaton confrms that bequests are not mportant for the balance of labltes. Evdently the bequests ncrease the balance of fnancal assets and when ths has been taken nto account, the bequests per se do not play any role n determnng the 1 The volume of real assets may be consdered endogenous as t s one of the wealth components. However, as ponted out by Flavn and Yamashta (2002), because of the transacton costs, endogenous changes n real assets are extremely nfrequent and the balance sheet adjustment s manly made va changes n lqud assets and labltes. As the current paper does not focus on the effect of real estate assets on fnancal assets, the varable can be treated as predetermned or weakly exogenous and left unnstrumented. 14

17 balance of labltes. Hence, t can be used as an explanatory varable only for the balance of fnancal assets. Determnants of fnancal labltes Another strand of lterature nvestgates the determnants of household debt or fnancal labltes. The demand for household debt can be dvded nto the partcpaton decson and, n case of partcpaton, the sze of the debt stock wanted by households. The volume of labltes also depends on the credt supply, as some households would lke to borrow more but are constraned by bank requrements. In the current specfcaton the demand and supply of credt are not modelled separately, and eq. (5) provdes the equlbrum result,.e. the holdng of labltes that s the outcome of the supply and demand for credt. The man determnants of household debt are derved from the lfe cycle model and are the same as for household savngs: ncome dynamcs, preferences and nterest rates (Crook (2006)). The emprcal studes whch nvestgate the determnants of the supply and demand for debt ndcate that the volume of household labltes also depends on the age and educaton of household members, ther occupatonal status and ther net wealth, see Magr (2007), Crook (2001), Ylmazer and DeVaney (2005) and Costa and Farnha (2012) among others. Fnancal assets and real assets are dstngushed from net wealth as the dfferent wealth components may have dfferent mpacts on labltes. Addtonally, extraordnary low ncome and unemployment may affect the volume of debt as households may smooth ther consumpton by borrowng. When we compare the explanatory varables for fnancal labltes and fnancal assets, the man economc varables and man household characterstcs are common varables for both labltes and assets. However, there are some varables whch have been found to be mportant for credt ratonng n the studes of Magr (2007), Crook (2001) and Cox and Japell (1993) and whch are not expected to be determnants of fnancal assets. Homeownershp contans addtonal nformaton beyond the data on the volume of real assets whch s mportant to the credt supply. Therefore the varables should be ncluded n the regresson for labltes, as s argued by Albuquerque et al. (2014) who use homeownershp n addton to total wealth as an explanatory varable when estmatng the determnants of aggregate household debt. Homeownershp can be used as an addtonal explanatory varable for labltes n the vector of Z 2. 15

18 Varables n the selecton model For the selecton equaton, addtonal varables to the determnants of the volume of labltes are needed for excluson restrctons,.e. the varables are consdered to nfluence the decson to borrow but they do not have any effect on the volume of labltes. Duca and Rosenthal (1993) and Crook (2001) fnd usng US data that beng black decreases the probablty of debt ownershp but t does not ncrease a household s demand for debt. Cox and Japell (1993) fnd that gender, and martal status affect partcpaton n the credt markets but are nsgnfcant n estmatng the volumes of debt. Duca and Rosenthal (1993) fnd that martal status s mportant for ownng debt but not for the volume of debt. Gven the results of the studes on partcpaton n credt markets, foregn orgn (mmgrant), gender and martal status (couple, sngle, dvorced or wdowed) are used as addtonal varables n the selecton model gven n eq. (3). Evdently these socal characterstcs contan nformaton about the atttude towards debt or wllngness to borrow whch may hnder the partcpaton n the credt market, as dscussed by Chen and Devaney (2001). These varables are not related to the volumes of labltes when households are partcpatng n the credt markets. These varables are also added to the vector of Z 1, meanng they are explanatory varables for fnancal assets and do not appear n the regresson for fnancal labltes. There s emprcal lterature that notes the dfferent savng behavour of mmgrants (see Pracha and Zhu (2012) among others) and sample nformaton shows that gender and martal status are sgnfcant n the regresson for fnancal assets, as also noted by Arrondel et al. (2013). Lkewse, Brown et al. (2013) show that gender and foregn orgn determne the balance of fnancal assets but are not related to the balance of total debt. The bequest varable s added to the selecton equaton to ensure that all the varables used n eq. (5) are also ncluded n the selecton equaton (4). Even so, the excluson restrcton may be a concern as t s dffcult to fnd a good excluson restrcton n most cases. When excluson restrctons are not avalable then the model can be dentfed from the assumpton of the jont normalty of regresson resduals. However, as ponted out by Rõõm and Dabušnskas (2011), ths may result n poor dentfcaton and hgh multcollnearty n the structural equaton. Therefore, to evaluate the goodness of excluson restrctons, dfferent sets of excluson restrctons have been tested. 16

19 4. The dataset and descrptve statstcs 4.1. The Household Fnance and Consumpton Survey The paper uses data from the frst wave of the Eurosystem Household Fnance and Consumpton Survey (HFCS). Ths s a harmonsed mcro database coverng 15 euro area countres. The survey s coordnated by the European Central Bank and carred out by the natonal central bank of each country. The survey desgn and the questonnare were harmonsed across the countres and the survey was made wthn the same tme span n all countres n The dataset covers more than 62,000 households wth sample szes n each country rangng from 340 households n Slovena to 15,000 households n Fnland. It provdes detaled household-level data on household balance sheets accompaned by related economc and demographc varables. A detaled descrpton of the methodology and the man results of the survey are provded n ECB (2013b). The database contans very detaled nformaton about the dfferent wealth components of households. The man components of households wealth are aggregated nto labltes, fnancal assets and real assets. Labltes nclude mortgages, and non-mortgage debt nstruments such as credt lnes or overdrafts, credt cards, and loans not collateralsed by real estate. The survey collects nformaton about debt repayments, so the debt repayment burden can be computed as the share of annual repayments n the annual total ncome of the household. The HFCS ncludes questons about reasons for savng, ncludng the reason of payng back debt. We can use the varables related to labltes n the robustness analyss. The fnancal assets cover sght accounts, savngs accounts, mutual funds, bonds, publcly traded shares, assets n managed accounts, nformal loans to relatves or frends, and other fnancal assets. The HFCS contans nformaton about prvate penson plans but n the baselne model ths asset type s not ncluded n the fnancal assets. Real assets nclude home equty, vehcles, valuables and self-employment busnesses. The queston about bequests and the form of bequests s used for complng a dummy for recevng a bequest n the form of fnancal assets. The total ncome of the household ncludes employee ncome, selfemployment ncome, ncome from pensons, unemployment and other socal benefts, prvate transfers receved, rental ncome, fnancal nvestment ncome, prvate busness ncome and other sources of ncome. We compute a dummy for self-employment by usng the nformaton about self-employment 17

20 ncome rather than the employment status of a reference person, n lne wth Pssardes and Weber (1989) and Kukk and Staehr (2013). Households are defned as self-employed f ther self-employment ncome exceeds 25 per cent of the household s total ncome. Other ncome-related varables are used to ndcate ncome uncertanty, such as the unemployment of a reference person, nformaton about the ncome n a reference perod compared to average ncome, or ncome expectatons for the next perod. Addtonal soco-demographc varables are age, educaton, gender, martal status (couple, sngle, dvorced or wdowed) and foregn orgn (mmgrant) of the reference person, and number of chldren n the household. Rsk preferences are captured by a dummy f the household states that t s rsk averse. Labltes, assets and ncome are expressed n logarthms. In the pooled dataset t should be ensured that the fnancal and ncome fgures are comparable across countres. As the ntervews were carred out n dfferent tme perods n dfferent countres durng , real values could be used. However, as ntervews were conducted durng several quarters wthn a country, usng the same annual prce ndces n a country for estmatng real values does not mprove the precseness of the varables. The exact tme span of the ntervews n each country s not avalable, so quarterly or monthly consumer prce ndces cannot be used. As the annual nflaton dfferences are very small across countres, as also ponted out by ECB (2013b), the nomnal values of the varables are used n the paper. The dfferences n the cost of lvng are remarkable across the countres and therefore ncome quntles are computed for each country separately. Also the real assets are dvded nto quntles, to take account of country level dfferences n llqud assets such as real estate when the pooled dataset s used. A smlar approach has been used by Teppa et al. (2013) and Arrondel et al. (2013). A lst of the varables that have been used n the models gven n Secton 3 s presented n Table 1. The survey weghts are used when calculatng the man statstcs for the whole populaton and when estmatng the probt model for calculatng the nverse Mlls rato of eq. (4). In these cases the whole sample s used n the estmaton sample. The weghts are not used n eq. (5) as t s estmated for the sub-sample of households wth labltes and census parameters are not provded. Moreover, as standard errors are bootstrapped n eq. (5), weghts are unnecessary (Cameron et al. (2009)). To see the effects of oversamplng whch are not off-set by the use of survey weghts, addtonal estmatons are mplemented by excludng outlers. 18

21 Table 1: Defntons of the varables used n the model Varable Defnton Labltes Log of fnancal labltes (collateral and non-collateral debt) FnAsset Log of fnancal assets (sght and savng accounts, bonds, funds, shares, nvestment accounts, and other fnancal assets). Penson assets are not ncluded. Inc Fve country-specfc quntles of total ncome of the household (salares, busness ncome, captal ncome and socal benefts) RealAsset Fve country-specfc quntles of real assets (housng equty, busness equty, vehcles and valuables) Age_cat Age category wth a value between 1 and 5. It takes the value 1 f the age of the reference person s < 35; value 2 f the age s between 35 44; value 3 f the age s between 45 54; value 4 f the age s between 55 64; value 5 f the age s over 65 Educ Categorcal varable for levels of educaton, takes the value 1 f the reference person has only prmary educaton, value 2 for secondary educaton, value 3 for tertary educaton Chld Number of chldren n the household IncIncrease Dummy = 1 f the household expects ncome to ncrease n the followng perod, otherwse = 0 IncLow Dummy = 1 f the household had lower ncome durng the reference perod than usual, otherwse = 0 Unempl Dummy = 1 f the reference person of the household s unemployed, otherwse = 0 Selfempl Dummy = 1 f the busness-related ncome of the household s hgher than 25%, otherwse = 0 Bequest Dummy = 1 f the household has receved a bequest n the form of fnancal assets (.e. money, deposts or bonds), otherwse = 0 Immgrant Dummy = 1 f the reference person reports beng born abroad, otherwse = 0 Gender Dummy = 1 f the reference person s female, otherwse = 0 Martal status Categorcal varable for dfferent martal status, takes the value 1 f the reference person s marred or cohabtng, value 2 f the reference person s sngle, value 3 f the reference person s dvorced and value 4 f the reference person s wdowed Homeowner Dummy = 1 f the household owns the man resdence Dsr Annual debt payments as a share of the annual ncome of the household Paydebt Dummy = 1 f the household s savng to pay back debt Rsk0 Dummy = 1 f the household does not want to take any rsk n nvestments, otherwse = 0 Source: Household Fnance and Consumpton Survey, 1 st wave. 19

22 The HFCS contans fve mputed datasets n whch the values of fnancal and ncome varables that are mssng n the dataset are mputed (ECB (2013a)). All fve mputed datasets are used n the fnal estmatons n order to avod bas from mssng observatons. The multple mputaton (MI) pont estmate of a coeffcent γˆ s the average of the fve complete data estmates gven as 1 γ =, (6) 5 5 ˆ γ IM IM= 1 The varance var(γ ) of a completed data estmate contans two components: 1) The wthn mputaton samplng varance, whch s the average of the fve complete-data varance estmates var(γ ˆ). 2) The between mputatons varance, whch s the varance of the complete data pont estmates. Ths gves the varablty due to mputaton uncertanty. The total varance var(γ ) of an estmated coeffcent s calculated as var( γ ) = IM= var( ˆ) γ 6 / 5 ( ˆ IM + γ IM γ ). (7) 4 IM= Descrptve statstcs Table 2 shows that n the HFCS dataset 43.7 per cent of households are ndebted and most of them prefer to hold both fnancal assets and labltes. About 26 per cent of the households report that they hold both labltes and assets of over 5,000 euros. Only 4 per cent of the households have labltes of more than 5,000 euros but fnancal assets of less than 1,000 euros. The statstcs ndcate that a substantal share of households tend to have both assets and labltes. Ths paper examnes whether the holdngs of labltes and fnancal assets are dependent on each other. 20

23 Table 2: Penetraton of ndebted households wth dfferent balance sheets (1) (2) (3) (4) Total no of observatons Share of households wth labltes (%) Share of households wth labltes and fnancal assets EUR (%) Share of households wth labltes EUR and fnancal assets < EUR (%) TOTAL Austra (2010) Belgum (2010) Cyprus (2010) Span (2008) Fnland (2009) France (2010) Germany (2010) Greece (2009) Italy (2010) Luxembourg (2010) Malta (2010) Netherlands (2009) Portugal (2010) Slovena (2010) Slovaka (2010) Notes: Fgures are calculated from the HFCS database usng survey weghts and fve mputed datasets. When the fnancal assets of households wth labltes are compared wth those of households wthout labltes, some dfferences emerge. Table 3 shows that ndebted households (column 2) n most countres hold slghtly fewer fnancal assets than households wthout labltes (column 1), though the dfferences are small. If ncome s taken nto account and the rato of fnancal assets to ncome s compared (columns 3 and 4), the dfferences are more clear than when the volume of fnancal assets s consdered. The statstcs mply that ndebted households keep less n fnancal assets than households wthout any debt. However, there may be several other reasons for the dfference n the stock of fnancal assets of households wth and wthout labltes, e.g. dfferent preferences, household characterstcs or economc condtons. In order to shed lght on the dfference whch s related to labltes, the model developed n Subsecton 3.1 s estmated. 21

24 Table 3: Mean values of fnancal assets for households wth and wthout fnancal labltes (n EUR) (1) (2) (3) (4) Fnancal assets (EUR) Fn asset to ncome rato HH wth labltes HH w/o labltes HH wth labltes HH w/o labltes Total AT BE CY DE ES FI FR GR IT LU MT NL PT SI SK Notes: Frst wave of HFCS. For estmatons, fve mputatons and survey weghts have been used. Fnancal assets nclude deposts, money market funds and bonds, nvestment funds, shares, managed accounts, and money lent out. The tme of the survey has to be taken nto account when nterpretng the results. The survey was carred out n ,.e. n the mddle of a crss when households faced adverse economc shocks. As the frst wave of HFCS s used, t s not possble to track the dynamcs of household wealth n euro area countres. Changes n household fnances are comprehensvely analysed for US households usng the bannual Household Fnancal Survey (HFS) by Moore and Palumbo (2010) and Brcker et al. (2011). Some developments whch are relevant for the current estmatons are provded next. Frst, there were substantal and wdespread declnes n values of homes n all European countres, resultng n szeable eroson of home equty, although of dfferent magntudes across the countres. The values of busness equty lkewse declned. The relatonshp between real assets and other wealth components mght have been dfferent when equty prces were at ther peak. As we use the country-specfc quntles of the real assets, the relatonshp 22

25 between the real asset quntles and the other wealth components s expected to be more stable than when absolute real estate value s used. Second, there was a sharp declne n share prces, leadng to a declne n the value of shareholdngs and the total value of fnancal assets. Challe and Ragot (2012) and Alan et al. (2012) pont out that precautonary wealth accumulaton s countercyclcal. Uncertanty n the housng and equty markets leads households to ple up buffer stocks. Consequently, the composton of fnancal assets s lkely to be dfferent durng the recesson from what t was n the pre-crss perod. However, as the paper nvestgates the balance of total fnancal assets, these developments have mnor mportance n nterpretng the results. The survey was conducted n when households had experenced the man shocks of the recesson. Ths suggests that at the tme of the survey households were more aware of the dfferent rsks they mght face, ncludng rsks related to ther ndebtedness. Consequently, households were more lkely to adjust ther balance sheets accordng to the perceved rsks and so the relatonshp between labltes and fnancal assets s expected to be explct durng ths tme. 5. The estmatons for fnancal assets and labltes 5.1. The baselne estmatons We nvestgate the relatonshp between ndebtedness and the fnancal assets of households n the euro area countres by estmatng the system of equatons derved n Subsecton 3.1: F = α1+ γ Lˆ 1 + X ' β1+ Z1 ' φ1+ θλ 1 + τ1 c+ ε1. (8) L = α + γ Fˆ + X ' β + Z ' φ + θ λ + τ + ε Followng the dscusson n Subsecton 3.2 on the determnants of fnancal assets and labltes the varables that appear n the vectors of X, Z 1 and Z 2 can be lsted. Income, real assets, number of chldren, age and educaton of the reference person, dummes for future ncome ncrease, extraordnary low ncome, unemployment and self-employment of the reference person are n the vector of X as they are common varables n both equatons. A dummy for recept of bequest n the form of fnancal assets, dummes for the reference person beng an mmgrant or female and for martal status appear n the vector Z 1, whle a dummy for homeownershp appears n the vector Z c 2 23