Educational Loans and Attitudes towards Risk

Transcription

1 Educatonal Loans and Atttudes towards Rsk Sarah Brown, Aurora Ortz-Nuñez and Karl Taylor Department of Economcs Unversty of Sheffeld 9 Mappn Street Sheffeld S1 4DT Unted Kngdom Abstract: We explore the relatonshp between wllngness to take fnancal rsk and the probablty of takng out a loan for educatonal purposes as well as the nfluence of rsk atttudes on the sze of the loan usng data drawn from the U.S. Survey of Consumer Fnances. The fndngs suggest a postve relatonshp between ndvduals wllngness to take fnancal rsk and the probablty of takng out a loan for educatonal purposes. Smlarly, ndvduals wllngness to take fnancal rsk appears to be an mportant determnant of the sze of the educatonal loan. The fndngs suggest that non-whte ndvduals and ndvduals from less wealthy backgrounds are less lkely to fnance educaton through loans whch could potentally ncrease nequaltes n educaton and ncome f such ndvduals are deterred from nvestng n ther human captal. Key Words: Educatonal Loans; Rsk Averson. JEL Classfcaton: I22; I23. Acknowledgements: We are grateful to the U.S. Federal Reserve for supplyng the Survey of Consumer Fnances snce We are very grateful to Pam Lenton and Arne Rsa Hole and Sheffeld for valuable advce. The normal dsclamer apples. Contact Author: Aurora Ortz-Nuñez, aurora.ortz@sheffeld.ac.uk Aprl

2 I. Introducton and background Over the last forty years, there has been an ncrease n the number of ndvduals n a range of countres attendng post-secondary school educaton and, n partcular, attendng unversty (e.g. UK, France, Italy and Span). For example, accordng to Greenaway and Haynes (2007), the number of students attendng unversty n the UK has ncreased from 400,000 n the early 1960 s to more than 2 mllon n At the same tme, some OECD countres such as Australa, New Zealand, UK and the US have moved away from provdng publc hgher educaton towards prvate hgher educaton by ncreasng unversty fees. 1 The US was one of the frst countres to move away from publc provson of hgher educaton by ncreasng the enrollment fees to attend unversty. In 1992, the Hgher Educaton Reauthorzaton Act created the Stafford Unsubsdzed loan program n the US, whch gave access to educatonal loans to all students regardless of ther economc background. 2 Ths Act created a new possblty for the fnancal markets n the US to ncrease the rate of educatonal loans taken out. To be specfc, before 1992, the Federal Stafford loan program for educaton offered a subsdzed loan for those students wth fnancal needs (.e. low ncome famles). Snce 1992, those students wthout fnancal needs were enttled to apply to the unsubsdzed loan program. Both types of loans have smlar terms and condtons, but the subsdzed loan only apples nterest repayments once the student graduates from unversty. At the same tme, banks and fnancal nsttutons started to offer prvate educatonal loans wth smlar terms and condtons to those offered by the unsubsdzed Stafford loan. Educatonal loans can be classfed nto two groups accordngly wth the repayment method: mortgage type loans and ncome contngent loans. Mortgage type loans offer a fxed repayment over a set tme perod whle ncome contngent loans offer 1 Accordng to Canton and Blom (2004), ths could be due to the recent ncrease n nternatonal labour moblty makng t less attractve for governments to nvest n hgher educaton f future benefts to educaton are to be receved by other countres. 2 From 1965 to 1992, access to educatonal loans was subsdzed and restrcted to students from low ncome famles. 2

3 a repayment system whch vares accordng to the ndvdual s future ncome. Chapman (2006) revews the nternatonal reform of hgher educaton provson and argues that n general student loans do not offer total protecton to borrowers and, n partcular, to those borrowers who are not as successful n the labour market as expected and who receve low levels of ncome. In a smlar ven, Chapman and Ryan (2005) study the effects of the ntroducton of ncome contngent loans n Australa. They conclude that hgher educaton partcpaton ncreased for ndvduals from mddle ncome famles and for females, but there was no change n the partcpaton rate of ndvduals from low ncome famles. More recently, Campan and Snnng (2011) use data from the German Mcrocensus to explore the effects of a hypothetcal ncome contngent loan system n Germany. They conclude that ncome contngent loans would be fnancally feasble even for low ncome graduates. Most of the exstng studes n ths area have focused on the effects of replacng grants wth loans for educaton. For example, n a recent contrbuton, Swarthout (2006) analyses US survey data to explore whether student loans of hgh amounts nfluence students future occupatons. He concludes that takng out a loan for educaton could nfluence the career choce of the students. Smlarly, Mncozz (2005) explores US data to show how educatonal debt affects job decsons, makng students more lkely to choose jobs wth ntally hgh wages but wth lower wage growth. In ths paper, n contrast to much of the exstng lterature, we focus on the determnants, rather than the mplcatons, of takng out a loan for educatonal purposes. We focus on one partcular nfluence on the decson to take out such a loan, namely atttudes towards rsk. It s apparent that ncreasng fees for hgher educaton can have mportant consequences for students and ther famles as ths could ncrease the fnancal pressures assocated wth attendng unversty and, hence, rsk averson may play an mportant role n the decson to attend unversty (Greenway and Haynes, 2007). Furthermore, ndvduals 3

4 belongng to wealthy famles may fnd t easer to pay unversty fees than students from less wealthy famles, whch could ncrease future educatonal and wages nequaltes. In order to reduce potental nequaltes n access to unversty, ncome contngent loans are desgned to facltate access to hgher educaton for everyone, offerng a repayment system, whch vares accordng to the ndvdual s future ncome. However, some ndvduals are more wllng than others to take fnancal rsks and ths could be an mportant barrer when decdng to take out a loan to fnance nvestment n unversty educaton. Takng out a loan for educaton nvolves rsk snce ndvduals are uncertan about whether they wll be able to meet the future payments as well as beng uncertan about ther apttude for studyng at unversty. For ths reason, understandng how rsk atttudes affect the decson of whether to take out a loan for educaton represents an mportant contrbuton to the economcs of educaton lterature as well as beng of potental nterest to polcy-makers. The shortage of lterature n ths area s somewhat surprsng although one possble explanaton could relate to the dffculty of fndng a sutable measure of rsk atttudes. The relatonshp between debt averson and educatonal loans has been studed by Eckel et al. (2007) usng expermental technques. They analyse the role of ndvduals atttudes towards debt when decdng to take out a loan for post-secondary educaton n Canada. The fndngs suggest that debt averson s not a barrer to takng out loans for educaton; however, those ndvduals who have had any prevous experence wth debt (.e. debt use) are found to be more lkely to take out an educatonal loan. 3 Barr and Crawford (1998) argue that t s mportant to understand whether debt averse ndvduals are the group most affected by hgh unversty fees, whch could make nvestment n human captal even more dffcult for 3 The varable debt averson ncludes questons desgned to measure the person s atttude towards borrowng such as whether the ndvduals have credt cards and whether the ndvduals would borrow from a fnancal nsttuton or from credt cards to make an unexpected expendture. Debt use s measured by whether the ndvdual has ever been behnd n a bll, loan, rent or mortgage repayment, whether the ndvdual has ever sold an asset to pay a debt and whether the ndvdual s spendng s larger than the ndvdual s ncome. 4

5 them. 4 A lack of nformaton could be another mportant negatve factor when decdng to take out a loan for educaton. Booj et al. (2008) develop an experment wth Dutch students n order to understand whether havng better nformaton about educatonal loans ncreases the number of students takng out a loan to attend unversty. Usng an nstrumental varable approach, they fnd that mprovng nformaton about the condtons of educatonal loans, such as the nterest rate and repayment perod, does not have any mpact on the take-up rate. 5 The am of the emprcal analyss presented n ths paper s to explore the relatonshp between ndvduals wllngness to take fnancal rsks and the probablty of takng out an educatonal loan as well as the sze of the loan usng data from the US Survey of Consumer Fnances. The results wll help us to understand whether rsk averse ndvduals are less wllng to take out a loan for educatonal purposes than less rsk averse ndvduals, whch may ncrease ther probablty of havng lower levels of educatonal attanment and, hence, potentally nfluences ther future labour market outcomes. Indeed, our emprcal fndngs suggest that wllngness to take fnancal rsks s postvely assocated wth the probablty of takng out a loan for educatonal purposes as well as the sze of the loan. III. Data We analyse the U.S. Survey of Consumer Fnances (SCF) developed by the U.S. Federal Reserve Board snce Ths cross-secton survey contans detaled nformaton about the balance sheet, penson, ncome, demographc characterstcs and the use of fnancal 4 A lack of nformaton could be another mportant negatve factor when decdng to take out a loan for educaton. Booj et al. (2008) develop an experment wth Dutch students n order to understand whether havng better nformaton about educatonal loans ncreases the number of students takng out a loan to attend unversty. Usng an nstrumental varable approach, they fnd that mprovng nformaton about the condtons of educatonal loans, such as the nterest rate and repayment perod, does not have any mpact on the take-up rate. 5 In a smlar lne, Osterbeek and Broek use data from the same survey among Dutch students to understand the low rate of student loans n Netherlands. They conclude that factors such as subjectve dscount rate, earnngs prospects and students rsk atttudes have a lmted explanatory power on the low rate of student loans n the Netherlands. 5

6 nsttutons by U.S. famles. To be specfc, the SCF contans questons related to educatonal loans. In the 1992, 1995, 1998, 2001 and 2004 cross-secton surveys, the head of the household was asked the followng questons: Not countng credt cards or loans you may have told us about, do you have any loan for educatonal expenses? If so, how much was borrowed not countng the fnance charges? 6 The responses to these questons yeld detaled nformaton not only about whether the ndvdual has taken out a loan to fnance educaton but also about the sze of the loan. The SCF also contans the followng queston related to ndvdual s wllngness to take fnancal rsk answered by the head of household: whch of the followng statements comes closest to descrbng the amount of fnancal rsk that you are wllng to take when you save or make nvestments? Take substantal fnancal rsks expectng to earn substantal returns; Take above average fnancal rsks expectng to earn above average returns; Take average fnancal rsks expectng to earn average returns; Or not wllng to take any fnancal rsks. 7 We use the responses to ths queston to create a three pont rsk atttudes ndex as follows: r 0 f they are not wllng to take any fnancal rsks 1 f they are wllng to take average fnancal rsks for average returns = 2 f they are wllng to take above average/substantal fnanacal rsks for above average/substantal returns 28.28% 47.76% 23.96% Thus, the ndex s ncreasng n wllngness to take fnancal rsk such that f the ndvdual s not wllng to take any fnancal rsk, the rsk atttudes ndex takes the value of zero, whlst f the ndvdual s wllng to take above average fnancal rsk for above average returns or s 6 Unfortunately, the SCF data set does not provde nformaton about the type of loan (.e. federal or prvate loan) or the condtons of the loan. 7 Ths measure of ndvdual rsk preference has been extensvely use n the lterature. See, for example, Schooley and Worden (1996), Grabble and Lytton (2001), Fnke and Huston (2003) and Hanna and Lndamood (2004). 6

7 wllng to take substantal fnancal rsk for substantal returns, the rsk atttudes ndex takes the hghest value of 2. 8 Therefore, we explot the responses to these questons n order to explore the relatonshp between the probablty of takng out an educatonal loan and the ndvdual s wllngness to take fnancal rsk, as well as the relatonshp between the amount borrowed for educatonal expenses and wllngness to take fnancal rsk. The sample s restrcted to those ndvduals who are n hgher educaton aged between 18 and 65 years old, yeldng a total of 1,740 observatons. 9 Gven the nature of the data, reverse causalty may be a potental problem related to the analyss. Atttudes towards rsk may be assocated wth factors such as educaton, ncome and wealth. To reduce the potental for reverse causalty, we restrct the sample to those ndvduals who are currently n hgher educaton as we are nterested n the relatonshp between ndvduals wllngness to take fnancal rsk and the probablty of takng out an educatonal loan before the completon of the nvestment n educaton takes place, as completon of the nvestment could potentally nfluence ndvduals wllngness to take fnancal rsk. 10 IV. Atttudes towards Rsk and the Probablty of takng out a Loan for Educaton Methodology We explore the relatonshp between the probablty of takng out a loan for educaton, whch s measured by a dummy varable, 11 and the ndvdual s wllngness to take fnancal rsk wth a probt model as follows: Loan = * x β + ε, ~ N( 0,1 ) ε (1) 8 Due to the small sample sze we collapse the top two categores by creatng a three pont rsk averson ndex takng not wllng to take any fnancal rsk as the base category. 9 It s apparent from the queston that t could be the case that ndvduals took out a loan for educaton for ther chldren or even spouse rather than for ther own educaton, whch stll provde nformaton on the relatonshp between ncurrng debt for educatonal purposes and atttudes towards rsk. 10 In the SCF survey, the queston relatng to educatonal loans s placed before the rsk atttudes queston. Therefore, t s mportant to hghlght the potental lmtatons of the varables and potental ssues of reverse causalty. For example, some ndvduals could ndcate that they are wllng to take fnancal rsk as selfjustfcaton for havng taken out a loan for educaton. 11 The loan varable takes the value of 1 f the ndvdual has a loan (805 observatons) and takes the value of 0 f the ndvdual does not have a loan (935 observatons). 7

8 where * Loan denotes the latent varable for the propensty to take out a loan for educaton, x represents a set of explanatory varables and ε denotes the error term, whch s normally dstrbuted. To be specfc, x ncludes the rsk atttudes ndex, r, and soco-demographc characterstcs such as age, male, whte, marred, the number of chldren of the respondent and household sze. The SCF data set does not nclude detaled nformaton about famly background, such as the educatonal attanment or occupaton of the ndvdual s parents, however, we try to capture the possble effects of famly background by ncludng n the set of controls the natural logarthm of the total net wealth of the household. 12 We also nclude n the set of controls whether the respondent (.e. the student) s workng and year dummy varables. Table 1 n the Appendx presents summary statstcs of the key varables employed n our econometrc analyss. 13 Followng Greene (2003), we model the probablty of takng out a loan for educaton wth a probt model where the probablty of observng Loan = 1 s gven by: P( Loan = 1) = P( Loan > 0) = P( x β + ε > 0) = P( ε > x β ) = Φ( x β ) (2) Results Tables 2 and 3 present the margnal effects from the probt analyss of the probablty of takng out a loan for educaton. Our fndngs n Table 2 suggest that wllngness to take fnancal rsk s postvely related to the probablty of takng out a loan for educaton. As a result of one-unt ncrease n the rsk averson ndex, the probablty of takng out a loan for educaton s ncreased by around 5.5 percentage ponts. 14 Smlarly, beng male, whte or marred are all postvely related to the probablty of takng out a loan for educaton. Males have a 9 percentage ponts hgher probablty of takng out a loan for educaton than females; 12 Total net wealth of the household ncludes: the value of land, buldngs, farms or ranches owned by the household; the value of houses, holday houses or other propertes; net worth of busnesses owned by any member of the household; the value of owned cars and other vehcles; fnancal assets and nhertances; net of mortgage and loans (excludng loans for educatonal purposes). 13 All monetary varables have been deflated to 2004 prces. 14 Ths s evaluated at the mean of the rsk atttudes ndex for the sample. 8

9 whte ndvduals have a 10 percentage ponts hgher probablty of takng out a loan for educaton than non-whte ndvduals; and marred ndvduals have an 8 percentage ponts hgher probablty of takng out a loan for educaton than ndvduals who are not marred. Age s also an mportant determnant of the probablty of takng out an educatonal loan. The older the ndvdual s, the hgher s the probablty of takng out a loan for educaton. In order to capture any possble non lnear effect, we also nclude age squared n the set of controls. The results confrm that older ndvduals have hgher probabltes of takng out a loan for educaton than younger ndvduals and the non lnear effect s very small. 15 We also nclude net household wealth n our analyss to try to capture the effects of famly background on the probablty of takng out a loan for educaton. The results show that the effect of total net wealth s statstcally sgnfcant and negatvely related to the probablty of takng out a loan for educaton, ndcatng that students belongng to a wealthy household are less lkely to take out an educatonal loan. Fnally, f a student s workng, he/she has between a 4 and 6 percentage hgher probablty of takng out an educatonal loan than a student who s not workng. 16 In Panel B of Table 2, we repeat the probt analyss replacng the rsk atttudes ndex wth a set of dummy varables. The results show that ndvduals who are wllng to take average fnancal rsk for average fnancal returns have a 17 percentage ponts hgher probablty of takng out a loan for educaton than ndvduals who are not wllng to take any fnancal rsk. Smlarly, those ndvduals who are wllng to take above average/substantal rsk for above average/substantal returns have a 10 percentage pont hgher probablty of takng out an educatonal loan than those who are not wllng to take any fnancal rsk. 15 Although the results mght be contrary to the human captal theory, t s mportant to underlyng that, as prevously explaned, t could be the case that ndvduals take out an educatonal loan for ther chldren rather than for ther own educaton. 16 The margnal effect on the rsk averson ndex does not change sgnfcantly f the total wealth and the control for whether the student s workng or not are omtted. 9

10 The small sze of the sample used n our analyss does not allow us to splt the sample by gender n order to explore gender dfferences n the effect of rsk atttudes on the probablty of takng out a loan for educaton. To overcome ths ssue, we ntroduce an nteracton term between the rsk atttudes ndex and the male dummy varable. Table 3 presents the results of the probt model after ncludng an nteracton term, r xmale. 17 The dfferental effect for a male who s wllng to take average fnancal rsk for average returns s around 58 percentage ponts. 18 Smlarly, the dfferental effect for a male who s wllng to take above average/substantal fnancal rsk for above average/substantal fnancal returns s around 55 percentage ponts. 19 Smlarly, n Table 3 (Panels C and D) we ntroduce an addtonal term whch nteracts the rsk averson ndex wth a dummy varable whch denotes 25 years old or younger. By ncludng ths addtonal term, we am to explore whether wllngness to take fnancal rsk at dfferent ages affects the probablty of takng out a loan for educaton. We focus on those ndvduals who are 25 years old or younger as the average age at whch ndvduals fnsh the Bachelors degree n the U.S. s 25 years old (Tulp, 2007). Our fndngs show that the effect of the nteracton term between beng 25 years old or younger and the rsk atttudes ndex s statstcally nsgnfcant. After substtutng x (Age 25) by the nteracton term between beng 25 years old or younger wth the rsk atttudes dummy varables we fnd a postve and statstcally sgnfcant suggestng that dfferental rsk atttudes effect exsts on the probablty of takng out a loan for educaton across the dfferent age groups. V. Atttudes towards Rsk and the Amount of the Loan Methodology 17 Margnal effects of nteracton terms n Table 3, have been estmated followng A and Norton (2003). 18 The results from the prevous set of controls are robust to the ncluson of the nteracton term. 19 Accordng to A and Norton (2003), the statstcal sgnfcance of the nteracton effect s often stronger when the nteracton effect as postve than when negatve, wth t-statstcs as hgh as ten. 10

11 Our second am s to explore the relatonshp between the ndvdual s wllngness to take fnancal rsk and the amount of the loan. Ths could be partcularly nterestng as t s mportant to understand whether ndvduals are wllng to take out a loan, whch, for example, covers all the expenses of ther educatonal nvestment (fees and lvng expenses) or whether they are just wllng to take out a loan to cover ther fees only. Accordng to Tulp (2007), the average total tuton fee n 2003 n a U.S. unversty was $8,700, 20 and the average amount of lvng expenses durng the unversty perod was around $40, He argues that the fnancal constrants of students are gven by the attendance fees and the lvng expenses. In our analyss, the dependent varable s charactersed by the presence of a hgh proporton of observatons wth the value zero (see Fgure 1 n Appendx). 22 Accordng to Tobn (1958), the tobt model allows for the consderaton of all the observatons n the sample, ncludng those that are censored at zero. Therefore, followng Wooldrdge (2002), we model the relatonshp between the amount of the loan taken out and the rsk atttudes ndex wth a tobt model as follows: l = * x 2 β + ε, ε ~ N(0, σ ) (3) l 0 = * l = x β + µ f f l * l * 0 > 0 (4) where * l s the latent varable, whch represents the natural logarthm of the amount of the loan, and l s the actual observed amount of the loan. Snce the dstrbuton of the amount of the loan s hghly skewed, followng Brown et al. (2008), we specfy a logarthmc dependent varable. Note that for students reportng a zero amount for the loan, ln( l l ) s recoded to zero as there s no reported amount of the loan between zero and the unty. x represents the set of ndvdual characterstcs that explan both the probablty of takng out a loan for 20 Accordng to Tulp (2007), the fees for a prvate unversty n the U.S. n 2003 are around $22,000 per year. 21 Ths ncludes the accommodaton, transport, books, supples and mscellaneous expenses. 22 The average sze of the loan for those ndvduals who took a loan for educaton s around $11,

12 educaton and the amount of the loan. 23 The vector of parameters to be estmated s represented by β and ε represents the normally and homoskedatc dstrbuted error term. Results Tables 2 and 3 present the results of modelng the amount of the loan for educaton. The results n Table 2 show that wllngness to take fnancal rsk s postvely related to the sze of the educatonal loan. Smlarly, beng whte and older both have a postve and statstcally sgnfcant relatonshp wth the amount of the educatonal loan. The number of chldren and the number of household members appear to be statstcally nsgnfcantly related to the amount of the loan for educaton whle the number of household members s negatvely related to the amount of the loan taken out. Marred ndvduals have, on average, much larger loans than non marred ndvduals. The results suggest that the relatonshp between household wealth and the amount of the loan for educaton s statstcally nsgnfcant. In Panel B n Table 2, we replace the rsk atttudes ndex wth a set of dummy varables. The results confrm the postve and statstcally sgnfcant relatonshp between wllngness to take fnancal rsk and the sze of the loan taken out. After ntroducng the rsk atttudes-male nteracton term (see Table 3), the results show that the effect of beng male s stll postve and statstcally sgnfcant suggestng that males have larger loans for educaton. Smlarly, the results relatng to the rsk atttudes ndex ndcate that wllngness to take fnancal rsk s postvely assocated wth the sze of educatonal loans. However, the nteracton term between rsk atttudes and beng male s negatve and statstcally sgnfcant. In Panel B of Table 3, we replace the rsk atttudes ndex wth a set of dummy varables and we fnd that the dfferental effect for a male, who s wllng to take average fnancal rsk for average fnancal return, s postve and statstcally sgnfcant for the amount of the loan taken. Smlarly, the dfferental effect of a male who s 23 Ths set of characterstcs s the same as that used to explan the probablty of takng out a loan for educaton n Secton IV. 12

13 wllng to take above average/substantal fnancal rsk for above average/substantal fnancal returns, s postve and statstcally sgnfcant for the amount of the educatonal loan take out. As wth the probt analyss, we then explore the determnants of the sze of the loan for educaton ncludng an nteracton term between rsk atttudes and a dummy varable denotng 25 years old or younger (see Table 3 Panel C). After ncludng ths nteracton term, the results relatng to the other control varables do not change wth respect to our prevous results. However, whether the student s 25 years old or younger appears to have a statstcally nsgnfcant nfluence on the sze of the loan for educaton. The analyss n Panel D n Table 3 explores the robustness of the results when we replace the rsk atttudes ndex wth the set of dummy varables. Our fndngs show that the nteracton terms between beng 25 years old or younger and the rsk atttudes dummy varables have a statstcally sgnfcant effect on the probablty of takng out a loan for educaton for those ndvduals who are wllng to take average fnancal rsk only. 24 VI. Robustness Methodology One shortcomng of the tobt model t that does not allow for the nature of the zero observatons assumng that the observatons are zero due to other factors rather than the non partcpaton decson of the respondents. Hence, n order to explore the robustness of the fndngs we explore a double hurdle specfcaton. Accordng to Cragg (1971), the double hurdle model relaxes ths assumpton of the tobt model by gvng specal treatment to the partcpaton decson. The double hurdle model assumes two hurdles n order to observe postve values, each of them s determned by a dfferent set of ndependent varables. The frst hurdle refers to partcpaton and the second one refers to the ntensty of use after overcomng ms-reportng or data problems. Gven such characterstcs, the double hurdle 24 It s apparent that the results for r = 1 and r = 2 are not monotonc. Ths could be due to the large proporton of ndvduals n the r = 1 category (almost 50 per cent). 13

14 model has been wdely used n the lterature to study consumer demand models (see for example Jones, 1992, Yen and Jones, 1997 and Arste et al., 2008), labour supply models (Blundell and Meghr, 1987) and loan default analyss (Moffatt, 2005). Cragg (1971) develops the double hurdle model by modfyng the tobt model n the followng way, where loan denotes whether or not the ndvdual reported havng a loan for educatonal purposes: * * l f loan = 1 and l > 0 l = (5) 0 otherwse where * l s the latent varable, whch represents the natural logarthm of the amount of the loan, and l s the actual observed amount of the loan. l takes value zero when there s censorng at zero or due to random crcumstances. 25 Jones (1992) rewrtes equaton (5) showng the process nvolved n observng zero values: l = l * = x 2 β 2 + u > 0 f x 1 β 1 + v > 0 and x 2 β 2 + u > 0 or (6) l = 0 f x1 β or x1 β or x1 β v + v + v > and and and x x x β β β u + u + u 0 > 0 < 0 where x 1 and x2 are two dfferent sets of varables for the frst (.e. partcpaton) and second (.e. ntensty) hurdles. Both sets of varables nclude ndvduals soco-economc characterstcs as prevously descrbed n Secton IV. However, the set of controls defnng the partcpaton equaton (.e. the frst hurdle) ncludes an addtonal varable, whch 25 On the contrary, Heckman s model (1979) assumes that all the zeros are due to the decson of non partcpaton of the respondents, whch would not be approprate for our purposes as t could be that some students dd not get a loan even f they appled for t. However, we also model the probablty of takng out a loan and the amount of the loan wth a Heckman model. The results are consstent wth the analyss, suggestng a postve relatonshp between the rsk atttudes ndex and both the probablty of takng out a loan for educaton and the amount of the loan. These results are avalable on request. 14

15 ndcates whether the ndvdual has been turned down for a loan n the last fve years. 26 Hence, a postve amount of the loan s observed only f an ndvdual takes out a loan for educaton. Results Tables 4 and 5 present the results from the double hurdle analyss under the assumpton of dependent errors (see, Jones 1992). 27 In other words, we assume there s a relatonshp between the errors of the partcpaton equaton and the errors of the ntensty equaton. The results n Table 4 suggest that beng wllng to take fnancal rsk, whte, older and male all have a postve and statstcally sgnfcant relatonshp wth the decson of takng out a loan for educaton and the amount of the loan taken out. In accordance wth the prevous results, beng marred only has a postve effect on the probablty of takng out an educatonal loan. The number of chldren n the household, as well as household sze, have negatve and statstcally sgnfcant effects n the case of the sze of the loan. Household wealth has a negatve and statstcally sgnfcant relatonshp wth the decson of takng out a loan for educaton but a postve effect n the case of the sze of the loan. Ths fndng suggests that when the ndvdual belongs to a wealthy household they have a lower probablty of takng out a loan for educaton because they may have suffcent fnancal assets to fnance ther educaton, but when they decde to take out a loan for educaton, they may be able to obtan larger loans. In Panel B n Table 4, we replace the rsk atttudes ndex ( r ) wth the set of dummy varables. In ths case, beng wllng to take average fnancal rsk has a postve and statstcally sgnfcant effect on the frst decson (.e. takng out a loan for educaton). On the 26 Unfortunately, the SCF data set does not provde nformaton about the reason for whch the loan has been turned down or what type of loan t was. The valdty of the nstrument has been checked, where the nstrument s statstcally sgnfcant n modellng the probablty of takng out a loan but statstcally nsgnfcant when modelng the sze of the loan. 27 Tables 4 and 5 show that the correlaton terms ( ρ ) between u and v are statstcally sgnfcant. Ths suggests that the error terms of the two equatons that form the double hurdle model are correlated. 15

16 contrary, beng wllng to take above average/substantal fnancal rsk has a postve and statstcally sgnfcant effect on both the partcpaton decson and the amount of the loan. Table 5 presents the results after ncludng the nteracton term r xmale. The results are robust to ts ncluson suggestng that wllngness to take fnancal rsk has a statstcally sgnfcant postve effect n both decson stages. The nteracton term between the rsk atttudes ndex and the male dummy varable has a statstcally sgnfcant negatve effect on the decson to take out a loan for educaton yet a postve effect n the case of the sze of the loan. Table 5 Panel C presents the results after ncludng the nteracton term between the rsk atttudes ndex and the dummy varable for 25 years old or younger,.e. r x( Age 25). Whether the ndvdual s 25 years old or younger has a negatve and statstcally sgnfcant effect on the sze of the loan. The rsk atttudes ndex has a statstcally sgnfcant postve effect on the partcpaton decson only. However, the combned effect of x( Age 25) r s only postve and statstcally sgnfcant n the case of the amount of the loan for educaton. Ths suggests that ndvduals aged 25 years old or younger, who are wllng to take fnancal rsk and who decded to take out a loan for educaton, have larger educatonal loans. VII. Conclusons We have explored the relatonshp between atttudes towards rsk and the probablty of takng out a loan for educaton as well as the sze of the loan usng a representatve data set from the U.S., based on poolng fve cross-sectons of the SCF. To be specfc, we have explored the relatonshp between the probablty of takng out a loan to fnance educaton and ndvduals economc and demographc characterstcs, ncludng atttudes towards fnancal rsk. We have also explored how the sze of the loan for educaton vares wth socoeconomc characterstcs and the rsk atttudes of the ndvduals. 16

17 We fnd that wllngness to take fnancal rsk s postvely related to the probablty of takng out a loan for educaton and wth the sze of the loan. Smlarly, our results suggest that characterstcs such as beng older, male, whte and marred are all postvely related to both the probablty of takng out a loan and the sze of the educatonal loan. The results also suggest that household sze and household wealth are negatvely related to the probablty of takng out a loan for educaton and ts sze. Our emprcal analyss contrbutes to the exstng lterature by helpng us to understand whether dfferences n rsk atttudes could lead to nequaltes n educaton. Our results suggest that dfferences n atttudes towards fnancal rsk may affect nvestment n hgher educaton by nfluencng the decson to take out a loan, whch can ultmately lead to nequaltes n labour ncome and wealth. In addton, the results suggest that non-whte ndvduals and ndvduals from less wealthy households are less lkely to fnance hgher educaton through loans. If rsk averson s concentrated among low soco-economc groups, then our fndngs predct that ndvduals from poorer backgrounds wll be unlkely to nvest n ther human captal thereby ncreasng potentally nequaltes n educaton and ncome as well as nequaltes n the next generaton. References Arste, D., F. Peral and L. Peron (2008) Cohort, Age and Tme Effects n Alcohol Consumpton by Italan Households: A Double-Hurdle Approach. Emprcal Economcs, 53 (1), A, C and E. Norton (2003) Interacton terms n logt and probt models. Economc Letters, 80, Barr, N and I. Crawford (1998) Fundng Hgher Educaton n an Age of Expanson. Educaton Economcs, 6, Blundell, R. and C. Meghr (1987) Bvarate Alternatves to the Tobt Model. Journal of Econometrcs, 34 (2), Booj, A., E. Leuven and H. Oosterbeek (2008) The Role of Informaton n the Take-Up of Student Loans. Tnbergen Insttute, Dscusson Paper No. TI2008/039/3. 17

18 Brown, S., G. Garno, P. Smmons and K. Taylor (2008). Debt and Rsk preference: A Household Level Analyss. Unversty of Sheffeld. SERPS Canton, E. and A. Blom (2004) Can Student Loans Improve Accessblty to Hgher Educaton and Student Performance? An Impact Study of the Case of SOFES, Mexco. World Bank Polcy Research, Workng Paper Chapman, B. (2006) Income Contngent Loans for Hgher Educaton: Internatonal Reform. Handbook of Economcs of Educaton, 2, Chapman B. and C. Ryan (2005) The Access Implcatons of Income Contngent Charges for Hgher Educaton: Lessons from Australa. Economcs of Educaton Revew, 24, Capman B. and M. Snnng (2011) Student Loan Reforms for German Hgher Educaton: Fnancng Tuton Fees. Insttute for the Labor Study, DP 5532 Cragg, J. (1971) Some Statstcal Models for Lmted Dependent Varables wth Applcaton to the Demand for Durable Goods. Econometrca, 39, Eckel, C., C. Johnson, C. Montmarquette and C. Rojas (2007) Debt Averson and the Demand for Loans for Postsecondary Educaton. Publc Fnance Revew, 35 (2), Fnke, M and S. Huston (2003) The Brghter Sde of Fnancal Rsk: Fnancal Rsk Tolerance and Wealth. Journal of Famly and Economc Issues, 24(3), Grabble, J. and R. Lytton (2001) Assessng the Concurrent Valdty of the SCF Rsk Tolerance Queston. Fnancal Counselng and Plannng, 12(2), Greene, W. (2003). Econometrc Analyss. 5 th Edton. Prentce Hall. Greenway D. and M. Haynes (2007) Fundng Hgher Educaton n the UK: The Role of Fees and Loans. Economc Journal, 113 (485), Hanna, S. and S. Lndamood (2004) An Improved Measure of Rsk Averson. Fnancal Counselng and Plannng, 15(2), Heckman, J. (1979) Sample Selecton Bas as a Specfcaton Error. Econometrca, 47(1), Jones, A. (1992) A Note on the Computaton of the Double Hurdle Model wth Dependence wth an Applcaton to Tobacco Expendture. Bulletn of Economc Research, 44 (1), MIncozz, A. (2005) The Short Term Effect of Educatonal Debt on Job Decsons. Economcs of Educaton Revew, 24,

19 Moffatt, P. (2005) Hurdle Models of Loan Default. Journal of the Operatonal Research Socety, 56, Oosterbeek, H. and A. van den Broek (2009) An emprcal Analyss of borrowng behavor of hgher educaton students n the Netherlands. Economcs of Educaton Revew, 28, Swarthout, L. (2006) Payng Back, Not Gvng Back: Students Debt s Negatve Impact on the Publc Servce Career Opportunty. Los Angeles: State Publc Interest Group Hgh Educaton Project. Tulp, P. (2007) Fnancng Hgher Educaton n the Unted States. OECD Economc Department Workng Papers, No Yen, S. and A. Jones (1997) Household Consumpton of Cheese: An Inverse Hyperbolc Sne Double-Hurdle Model wth Dependent Errors. Amercan Journal of Agrcultural Economcs, 79,

20 Appendx Fgure 1a: Hstogram of the Amount of the Loan (sample = ndvduals who have a loan). Percent Sze of the Loan for Educaton 20

21 Table 1: Summary Statstcs of the SCF data set (Sample= Current Student Head of Household) ALL STUDENTS STUDENTS WITH LOAN MEAN STDEV MAX MIN MEAN STDEV MAX MIN Havng a Loan Log Amount of Loan Age Age Squared Age Whte Male Number of Chldren Marred Household Sze Rsk Atttudes Index( ) x Male x (Age 25) Log of Household Wealth Student Workng Turned Down for a Loan Observatons 1,

22 Table 2: The Determnants of the Probablty of Havng a Loan for Educaton and Sze of the Educatonal Loan. Probablty of Havng a Loan Sze of the Loan ME tstat ME tstat Panel A: Rsk Atttudes Index Age (2.75) (2.87) Age Squared (3.30) (3.47) Whte (3.84) (3.99) Male (2.97) (2.88) No. of Chldren (1.10) (1.14) Marred (2.31) (2.49) HH Sze (1.53) (1.51) r (3.23) (3.44) Year (4.68) (5.25) Year (1.68) (2.09) Year (4.00) (4.48) Year (3.04) (3.51) Log Wealth (1.80) (1.72) Student Work (1.53) (1.63) Ch- Squared p=[0.000] p=[0.000] Pseudo R-squared Panel B: Rsk Atttudes Dummy Varables r = (5.78) (5.88) r = (3.14) (2.97) Ch-Squared p=[0.000] p=[0.000] Pseudo R-squared Observatons 1,740 22

23 Table 3: The Determnants of the Probablty of Havng a Loan for Educaton and the Sze of the Educatonal Loan wth r xmale and x( Age 25) Interacton Terms r Probablty of Havng a Loan Sze of the Loan Panel A: Rsk Atttudes Index Male Interacton Term ME tstat ME tstat Male (3.53) (3.58) r (3.73) (3.81) r xmale (2.05) (1.93) Ch- Squared p=[0.000] p=[0.000] Pseudo R-squared Panel B: Rsk Atttudes Dummy Varables Male Interacton Term Male (3.01) (1.23) r = (5.80) (2.56) r = (3.05) (3.73) ( r = 1)xMale (22.75) (1.79) ( r = 2)xMale (12.35) (3.22) Ch-Squared p=[0.000] p=[0.000] Pseudo R-squared Panel C: Rsk Atttudes Index Age 25 Interacton Term Age (0.31) (0.21) r (2.96) (3.06) r x( Age 25) (0.64) (0.48) Ch- Squared p=[0.000] p=[0.000] Pseudo R-squared Panel D: Rsk Atttudes Dummy Varables Age 25 Interacton Term Age (0.46) (0.33) r = (5.41) (4.81) r = (2.60) (2.71) ( r = 1) x( Age 25) (12.50) (2.31) ( r = 2) x( Age 25) (9.88) (0.46) Ch-Squared p=[0.000] p=[0.000] Pseudo R-squared Observatons 1,740 23

24 Table 4: Robustness: Double Hurdle Model Partcpaton Intensty Coef. tstat Coef. tstat Age (2.31) (4.34) Age Squared (2.83) (4.15) Whte (8.23) (4.04) Male (3.68) (3.79) N. Kds (0.52) (2.16) Marred (2.28) (0.79) HH Sze (1.67) (3.94) r (2.73) (3.57) Year (1.94) (7.37) Year (0.06) (6.86) Year (1.70) (7.06) Year (0.61) (4.72) Log Wealth (2.31) (4.83) Student Work (0.55) (3.20) Turned Down for a Loan (2.45) - - ρ (2.60) L-Lkelhood Panel B: Rsk Atttudes Dummy Varables r = (5.09) (1.11) r = (2.49) (3.48) ρ (3.27) L-Lkelhood Observatons 1,740 24

25 Table 5: Robustness: Double Hurdle Model wth Terms r xmale and x( Age 25) Interacton r Partcpaton Intensty Coef. tstat Coef. tstat Panel A:Rsk Atttudes Index Male Interacton Term Male (3.40) (0.63) r (3.61) (1.57) r xmale (1.94) (1.71) Turned Down for a Loan (2.76) - - ρ (2.68) L-Lkelhood Panel B: Rsk Atttudes Dummy Varables Male Interacton Term Male (0.96) (1.21) r = (2.60) (0.80) r = (3.37) (1.87) ( r = 1)xMale (2.38) (0.47) ( r = 2)xMale (1.44) (1.28) ρ (4.20) L-Lkelhood Panel C: Rsk Atttudes Index Age 25 Interacton Term Age (0.79) (1.93) r (2.73) (0.74) r x( Age 25) (0.30) (4.21) Turned Down for a Loan (3.03) - - ρ (2.04) L-Lkelhood Panel D: Rsk Atttudes Dummy Varables- Age 25 Interacton Term Age (0.64) (1.63) r = (5.81) (0.08) r = (3.08) (2.61) ( r = 1) x( Age 25) (1.84) (2.10) ( r = 2) x( Age 25) (1.25) (0.82) ρ (5.85) L-Lkelhood Observatons 1,740 25