Borrow or work for it? A study of individual funding methods of Dutch higher education

Borrow or work for it? A study of individual funding methods of Dutch higher education

Céline van Essen Master thesis Msc. Economics

University of Groningen Supervisor: Prof.dr. R.J.M. Alessie

June 24, 2016

Abstract

Working part-time is increasingly common among students in higher education. Borrowing in turn is less popular. Though, together they form the two most important individual funding methods for Dutch higher ed- ucation. This paper examines which factors determine the …nancing deci- sions that students make. De Studentenmonitor, a survey-based dataset, is used containing information on …nancial and demographic character- istics of Dutch students. It is found that socio-economic background strongly a¤ects …nancing decisions. First-generation students prefer work- ing part-time over borrowing. On the other hand, students who are likely to obtain high-income jobs are more willing to borrow than others. Fur- thermore, the removal of the basic grant is likely to increase the labour supply of students from lower-income families rather than their willingness to borrow.

JEL Classi…cation: I22, I28, D91

Keywords: Student Employment, Educational Investment, Government policy

e-mail: c.c.van.essen.1@student.rug.nl.

I am thankful to my supervisor, prof. dr. R.J.M. Alessie, for his valuable guidance and keen

interest during the writing process.

1 Introduction

The grounds for cost-sharing in education arise from well-accepted elements of economic and public …nance theory (Woodhall, 2002). Even though quantifying them can be di¢ cult, both public and individual bene…ts arise from education that might justify a government’s role in its provision. If a higher level of education entails a higher level of productivity, society can bene…t from the higher standard of living that increased productivity brings about. Moreover, as long as higher education is a normal good, higher-income families would obtain more education. Government intervention can therefore also be motivated by the need for income mobility. But from an economic point of view, the least controversial rationale for government intervention in education presumably is the credit market failure in education. As higher education is costly, most students do not have the means ready to …nance it. Students also generally do not possess any collateral which would make banks willing to o¤er them a loan.

That is where the government comes into play. By o¤ering …nancial aid, the accessibility of higher education can be assured. On the individual level, the basic rationale for the student bearing at least part of the costs builds on the expectation of positive returns. These re‡ect in higher lifetime earnings and in non-monetary bene…ts, such as higher status and access to jobs of greater prestige and desirability (Schleef, 2000).

Since students in general only pay for part of the actual costs of studying

, longer study periods entail higher public costs. In Dutch higher education, the nominal study period hovers around four years. However, Dutch students tend to take on average six years to obtain their diploma. Di¢ culty does not seem to be the issue. In fact, many studies show that students spend around 28 hours per week on their studies

– instead of the 40 hour norm. Factors such as lack of self-discipline, extracurricular activities and the take up of part- time jobs all play their part. Especially working part-time whilst enrolled in higher education is common in many OECD countries. About one-third of the time, the average OECD student is employed while he studies. The Netherlands stand out herein, with around 91% of its students working part-time (OECD, 2010). Although a part-time job augments …nancial resources, it abates study time. As an alternative, students can easily borrow at favourable terms. In that

1

In 2012, the costs of a Dutch university student to the government were on average e6.200, whereas the tuition fee hovered around e1.900. Source: Algemene Rekenkamer. Publieke Kosten. (2012).

2

Allen & van der Velden (2007), Arnold (2012).

case, students can graduate sooner and enter the labour market by substituting labour time for study time. But although the amount of student loans has risen over the past, the borrowing participation rates are still particularly low: only about one third of the students borrow

.

After a long standing debate, the Dutch student …nancial aid (hereafter SF) system recently shifted into a social loan system

. Basic grants were replaced by loans with favourable terms. On the policy side, any suggestion that a government contemplates concerning student …nancial aid tends to give rise to a wave of protests: varying from protests against the introduction of a basic grant in 1984 to protests against the removal of it in 2013.

A motive for grants can be the concern that lower-income individuals are debt averse. This means that they would avoid loans because of short-sighted fears about loan repayments, thereby foregoing valuable education. A natural experiment by Field (2009) illustrates how students at New York University law school who were randomly assigned grants were twice as likely to enrol in that university compared to those who were assigned loans. This suggests that moving from grants to loans can have substantial e¤ects on students’behaviour. However, in public policy theory there is no real rationale for providing grants to the higher- income families. These families might have paid for higher education anyway, such that state funds are ine¢ ciently targeted. Following this thought, shifting state resources away from direct provision and toward loans might well enhance both the micro- and macro-e¢ ciency of higher education.

Clearly, two of the main funding methods of higher education are borrowing and working part-time. They can complement or substitute each other: those who do not borrow might simply choose to work more. It seems therefore that the impact of the reform on higher education accessibility hinges on rationality assumptions. Is one’s willingness to borrow related to his or her social back- ground? If the accessibility of higher education is in question, the ideology of equal opportunities is impaired. Hence, …nding out whether the funding sources for human capital accumulation relate to …nancial- or cultural sources in one’s social background matters for optimal policy making. Careful analysis of stu- dent’s behaviour is needed to see whether the …nancial aid system expands or narrows opportunities. For that reason, this paper aims to analyse the borrow- ing and working behaviour of Dutch students. Next to that, expectations are

3

Kreetz et al. (2012).

4

In the remainder of this paper, this shift is referred to as ‘the reform’.

5

van Walsum (2014)

formed with respect to the impact of the recently implemented reform.

Using detailed Dutch survey data on student …nancial and demographic fac- tors, I sample all students who enrolled in an academic program in the Nether- lands between 2003 and 2013. Since the analysis relies on self-reported data, one must be cautious when interpreting the results. The estimates show that socio-economic background seems to play an important role in the decision for

…nancing human capital accumulation. Also, the removal of the basic grant appears to stimulate …rst-generation students to work (part-time) rather than to borrow.

The remainder of this paper is set up as follows: Section 2 opens with an overview of relevant literature. In this context, the human capital theory is related to …nancing behaviour of students in the Netherlands and other coun- tries. Subsequently, Section 3 provides background information of SF in the Netherlands in the twenty-…rst century. Section 4 discusses the methodology, after which Section 5 provides a description of the data. Section 6 contains the results and a discussion. At last, Section 7 o¤ers some concluding remarks and implications for public policies. Further information on the data and interme- diate results are presented in the Appendix.

2 Student loans, part-time jobs and the human capital theory

This part of the paper will give an overview of relevant literature on the relation- ship between student behaviour and …nancing methods. All students face the dilemma of whether to get higher education and forego immediate earnings or enter the labour market and forego further education. The optimal outcome de- pends on the perceived value of the investment. Economic theories of education such as Schultz (1968) argue that studying is a form of investing in one’s own human capital, and if the investment if worth-while, it is presumably worth bor- rowing to make it. With respect to valuing education, it is reasonable to think that …rst-generation students

will …nd it more di¢ cult to value education than students whose parents have obtained a degree. Even though it is theoretical, some literature shows that when students consider entering higher education

6

Although there is some ambiguity in the literature with respect to the precise de…nition of

…rst-generation students, for sake of simplicity I assume it concerns a student whose parents

have not obtained a degree in higher education. For an elaborate discussion on di¤erences in

such assumptions see Toutkoushian et al. (2015).

they use some kind of cost-bene…t framework. A student might weigh oppor- tunity costs of not working against the present value of a diploma, or prospec- tive debt against expected economic returns. Acumen Research Group (2008) a¢ rms a strong link between perception and participation by showing that stu- dents who perceived high costs in relation to bene…ts were less likely to enrol in university, irrespective of grades. Also, Usher (2005) …nds that lower-income individuals are more likely than higher-income individuals to overestimate the average costs and underestimate the average bene…ts of university education.

This informational aspect strongly relates to borrowing behaviour. Students are generally more informed about how much they are allowed to borrow than about the costs and bene…ts that di¤erent borrowing levels bring about. Schmeiser et al. (2015) …nd that the government can in‡uence behaviour quite easily. Their results show that students receiving more information about student debt bor- row less on average than students who do not receive extra information, without any a¤ect on their academic performance.

In light of microeconomic analysis, a relevant theory is that of the life-cycle

model (see e.g. Ando & Modigliani, 1963). It proposes that people rationally

choose how much they want to consume at each age limited only by the resources

available over their lives. Students …t relatively well in these life-cycle models

of behaviour. That is, they are fairly homogenous with respect to income and

expenditure, which makes non-economic factors more easily observable. When

individuals expect large increases in their income, they would borrow to smooth

their consumption over the life-cycle. This predicts that students are quite will-

ing to take on debt. Since, regardless of the source, a student’s current income

is likely to be low compared to his future income. Davies & Lea (1995) con…rm

this by …nding students to be a relatively low-income, high-debt group with tol-

erant attitudes towards debt. Yet, a relatively small number of Dutch students

makes use of the favourable loans. Since in the new social loan system the basic

grant no longer exists, this number will increase in all probability. Although the

social loan system resembles a social risk insurance system, the notion of debt

aversion is vital. That is, individuals might avoid loans because they fear ex-

treme debts, even when borrowing would be the most rational …nancing method

for them. It is often presumed that debt aversion is more present among low-

income families. That is why debt aversion is one of the leading arguments

for reasoning the reform will lower the access to higher education. However,

there is no necessary link between debt-aversion and obtaining a degree: these

students might simply work in order to obtain enough funds. But when debt

aversion deters students from borrowing while it would be their optimal choice, ine¢ ciencies arise. A general rationale for debt aversion is o¤ered by Carroll (2006), who shows that people with uncertain future earnings who are su¢ - ciently rational will never borrow if there is the possibility that they will not earn enough to be able to repay their debts. The same is proposed by Ooster- beek & van den Broek (2009), to explain the low borrowing participation rates in the Netherlands. Some research suggests that aversion to loans may reduce opportunities for a subset of low-income and minority students. Callender &

Jackson (2005) for instance …nd that debt aversion is especially present among low socio-economic groups. Fairley & Weitzel (2016) nuance this and …nd that students’ aversion to borrowing may be primarily driven by their aversion to ambiguity regarding repayment of the debt. They note that ambiguity aversion a¤ects the loan amounts, but has no impact on the decision to borrow. This suggests that signi…cant borrowing constraints not necessarily exist, especially when student loans are income-contingent such as in the Netherlands. Eckel et al. (2007) con…rm this by simulating debt aversion as well as risk and time preferences and …nd no evidence that debt aversion is an important barrier to investment in higher education. Jacobs & Canton (2003) propose accordingly that risk aversion and debt aversion can be overcome by the income contingent characteristic of loans. Moreover, Lanser & ter Weel (2013) predict based on available empirics from policy evaluations in the US and Australia that the re- form will decrease the number of students by a minor percentage, suggesting the reform hardly a¤ects the accessibility of higher education. Yet, these re- sults are possibly biased as the countries are incomparable in this respect. But by analysing the Dutch SF, Ebens et al. (2004) …nd that providing grants to students implicitly means subsidising their parents. They show that each addi- tional euro on supplementary grants reduces the …nancial parental contribution with 20-40 cents, suggesting a grant system is ine¤ective.

In the past, Dutch students from higher social classes took out a loan more often than those from lower social classes (Wartenbergh et al., 2010) but re- cent research has shown that precisely the latter student group tends to have a loan more often (Kreetz et al., 2012). It appears that rather than borrowing, the majority of students prefers to …nance their studies with a part-time job.

This is con…rmed by reviewing this study’s sample, illustrated in Figure 2.1.

The percentage of students that works and borrows and the percentage that

does neither have converged in the past years, for both university and college

students. The number of students that only borrow and do not work is small,

but rose slightly in the past few years. However, the majority of students clearly

Figure 2.1: Students according to work and borrowing behavior, based on this study’s sample

works without borrowing. Hence, especially for Dutch policy makers, the re- lation between the structure of the SF system and students’ labour supply is essential. The e¤ect of student employment on academic performance is am- biguous. Working part-time next to studying can have positive e¤ects. Due to improved labour market attachment, living standards of students may increase as well as chances on the labour market (Geel & Backes-Gellner, 2012; Hakki- nen, 2006; Light, 2001). But there are also potential negative e¤ects related to the high student labour supply, such as increased risks of dropping out and severe study delays (Bound et al., 2007; Stinebrickner & Stinebrickner, 2003).

A recent paper by Avdic & Gartell (2015) studies the same relationship by re- viewing a Swedish aid reform, which relaxed working constraints and tightened borrowing constraints. The authors …nd that this reform induced students to increase labour supply, thereby indirectly decreasing their relative study pace.

An important nuance on this topic is provided by Joensen (2009) who shows that the impact of student work hours on academic achievement is non-linear:

working 1-9 hours a week is complementary to academic success, while working

more than 18 hours is detrimental. This suggests that up to a certain point,

…nancial aid subsidizes leisure, rather than studies.

Several papers propose borrowing constraints as a possible explanation for why parental income matters for educational outcomes. In the context of this paper, that suggests that the current Dutch reform will increase the importance of one’s socio-economic background. An important question to the life-cycle model in this respect is whether it is reasonable to assume that students make rational, consistent and intertemporal plans. Borrowing might sometimes be optimal. But, the di¢ culty of obtaining information paired with the lack of credit experience can potentially lead students to make suboptimal decisions about the amount of student loans (Avery & Turner 2012; Lochner & Monge- Naranjo 2015).

There are notable di¤erences between US and Dutch studies on this matter.

In the former, often emphasis is placed on the magnitude and growth of student debt loads. In the latter it is increasingly discussed how the borrowing partic- ipation rates are quite low, given the favourable terms at which students can borrow. One should bear this in mind as results sometimes tend to be found in the direction of sought policy recommendations. Nevertheless, a recurring note appears to be how much emphasis to place on a¤ecting student behaviour.

Caution about borrowing might simply be a rational response to a student’s situation, especially for those who face uncertainty about future job prospects.

Borrowing and working part-time clearly are two of the main funding meth- ods of human capital accumulation for young agents. However, most of the empirical research on student loans and on work has examined these two sup- plementary …nancing choices in isolation and has focused on di¤erent issues.

Therefore, this paper aims to contribute by showing how borrowing and work- ing decisions relate to individual and family demographics.

3 The Dutch system of …nancial aid

3.1 Grants, (conditional) loans and travelling subsidies

In the following section, a brief overview is provided of the SF system before

the reform, speci…cally over the period of 2004 to 2014. Also a number of im-

portant policy changes in the SF system are shortly discussed. Dutch higher

education broadly consists of two di¤erent levels, known as HBO and WO. These

loosely translate to ‘higher professional education’and ‘scienti…c education’re-

spectively. In this paper, students in HBO are referred to as college students

whereas those in WO are referred to as university students. The SF system is listed in the law WSF 2000

. It was designed to let the government use its resources for SF so as to make education accessible to everyone, including lower-income families. The SF consisted of a basic grant and a public trans- port card for every student. Supplementary grants are provided when parental income is below a certain threshold. In Table 3.1, an overview is given of the SF amounts throughout the years. All students are allowed to combine SF with labour income up to a speci…ed earning threshold without being penalized with a proportional reduction of the student aid. The construction of this penalty is discussed in Section 3.3. The supplementary grant in the …rst study year is always a gift. The basic grant and the travelling card by contrast are dis- tributed as a conditional loan, excluding the interest-bearing loan (which must always be repaid). The conditional loan will be converted into a grant when the student obtains a …nal certi…cate in higher education within 10 years. This performance-related characteristic encompasses an implicit incentive to gradu- ate. If a student does not meet the required standards, the conditional loan will become an interest-bearing loan. The performance is also linked to the nominal duration of education (in most cases 48 months). A student is entitled to the performance grant for this nominal duration. Hereafter, a student can obtain an interest-bearing loan for 36 months.

7

Rijksbegroting 2000. SDU Uitgevers, 1999.

Table 3.1: SF amounts per month from 2003 to 2015

Basic grant Supplementary Maximum Tuition Earning

grant loan fee threshold

Living situation Living situation Year With parents Independent With parents Independent

03-04 70

210

195

214

765 1,465 9,847

04-05 74

228

199

217

771 1,476 10,200

05-06 76

233

203

221

787 1,496 10,460

06-07 89 248 207 226 796 1,519 10,530

07-08 91 253 206 225 810

1,538 10,630

08-09 92 256 209 228 819

1,565 11,640

09-10 93 260 212 231 832

1,620 12,000

10-11 96 266 219 239 853

1,672 12,000

11-12 96 266 221 241 853

1,713 12,000

12-13 96 269 225 246 853

1,771 12,060

13-14 98 272 230 250 853

1,835 12,210

14-15 100 279 237 258 895

1,906 12,388

a

The height of these amounts also depended on the kind of medical insurance. The average standard amount is shown

b

Excluding the extra ‘college credit’

c

The displayed amounts are all converted to net amounts.

3.2 Loan facilities

The major motivation for government intervention in higher education is not to produce positive externalities but rather to correct the failure in the credit mar- ket for student loans. Banks would after all hardly provide educational loans, since there is no collateral to repossess if the loan is not repaid (Gruber, 2013).

In the same line, the Dutch government does not establish any gross target values with respect to student loan amounts, since the aim is not to endeav- our that the greatest possible number of students borrow as much as possible.

A loan is rather seen as a reasonable substitute for labour income, intended

to shorten study periods and improve academic performances. very student is

eligible for the full loan amount and application is easy: a few clicks are in

between applying for a student loan via a governmental website. This website

also o¤ers calculation modules to estimate future loan repayments. During the

repayment period, interest costs on the remaining loan are still incurred. The

interest rates on the student loans equal the interest rates the government has to pay on the capital market for state loans. Because these interest rates are much lower than those o¤ered by banks, individual bank loans are rarely used by Dutch students (Kreetz et al., 2012). Two years after the student has left higher education, the repayment period starts, which can last up to 15 years.

The monthly repayment is calculated via an annuity system. Moreover, there are capacity to pay arrangements. Graduates with low incomes can request such an arrangement. This can cause the monthly repayment to decrease or can even cause a (temporary) exemption from repayment. In other words, the loan can become income-contingent. After 15 years of repayment, the remaining student debt will be abolished. This is to prevent that students, whose individual (mon- etary) bene…t from studying does not outweigh the investment, are confronted with debt for a life-long period.

3.3 Policy changes

In the study year of 2007-2008, the college credit was introduced. This allows students to borrow an extra amount on top of the regular loan amount, equal to the statutory tuition fee. It can be seen as a plain additional loan amount however, as spending purposes are not laid down by law. Alongside, in 2007 the government implemented two policies to increase the knowledge about the SF repayment system: its website was adjusted and enlarged and a speci…c folder about borrowing was issued.

Figure 3.1 shows the average borrowing amounts of students in this study’s

sample who did borrow, grouped by their birth year. After 2007, one clearly

observes (temporary) increases in the average borrowing amounts for nearly all

the birth cohorts. These sharp increases possibly indicate a shock following the

introduction of the college credit. Moreover, since the study year of 2009-2010,

the repayment conditions were made more simple and ‡exible. Most notably,

loans became automatically income-contingent. Before 2010, when a student’s

labour income exceeded the earning threshold, a penalty was given equal to

the amount of earnings that exceeded the threshold, maximized to the total

received SF. On top of this, a …ne was imposed comprising of the monthly

value of the travelling card. For instance, if one received SF throughout the

year 2009 and earned e1,- above the threshold in that year, one had to repay

12 months of traveling subsidy (e963,-) plus the e1,- earned too much. After

2010, this additional …ne was abolished: when a student earned e1,- above the

threshold, he only needed to repay e1,-. Furthermore, the most fundamental reform has been implemented recently: a social loan system. First, it was implemented in the master phase with the start of the study year 2014/2015, and in the bachelor phase in 2015/2016. Essentially, the basic grant and the earning threshold were abolished, the supplementary grant was maintained and the loan repayment period increased to 30 years. Student loans remained to be automatically income-contingent.

Figure 3.1: Average borrowing amounts by birth cohort

4 Methodology

In this section, the analytical and empirical methodology is discussed, together with the structure and possible caveats. First, the decisions to use a speci…c

…nancing method are separately tested using two-part models. Thereafter, the decisions are estimated jointly on a subsample of students. The models are estimated on a sample of repeated cross-sections.

4.1 Working and borrowing in depth

What follows from the previous section is that next to personal preferences

and experiences, other demographic factors are also likely to in‡uence …nancing

methods. In many related studies, Heckman’s two-step procedure is applied.

However, running a separate probit for sample inclusion followed by an (OLS) regression, referred to as the two-part model (Manning et al., 1987) is an es- pecially attractive alternative. Speci…cally, it models how participation in a

…nancing method is determined and, given participation, how the severity of participating is a¤ected. Other than the Heckman model, the two-part model allows the same variables to be used in each part.

This is convenient, as it is likely that variables explaining whether someone uses a speci…c …nancing method also determine how much he uses it.

Suppose that students are able to choose whether they want to engage in a (supplementary) …nancing method based on a set of variables. The two-part model is governed by two equations. The …rst equation is a probit estimator of the probability of having a positive outcome and the second equation is a linear model of the outcome among the subsample of positive observations,. Formally, the model is as follows:

F

=

x

+ "

"

s N(0; 1) (1)

F

jF

> 0 =

x

+ "

"

s N(0;

) (2) F

= F

if F

> 0

F

= 0 if F

0

where F

represents one of the …nancing methods. equals either W or B, such that F

equals one when the student works and F

equals one when the student borrows, and otherwise zero. F

represents the amount of work (measured by time spent on working

) and F

the loan amount. The i(t) subscripts indicate that the observations come from representative and indepen- dent cross sections where individuals are only available in one period. x

is a vector of explanatory variables, time dummies and cohort dummies

. These

8

Formally, this means that the probit error term is independent of the OLS error term.

9

Students also reported their average monthly labour income. In robustness check, the amount of work is measured via labour income, shown in Table A6 in the Appendix. Results are qualitatively similar.

1 0

I group students into cohorts according to birth year, and control for possible cohort

e¤ects by including cohort dummies. Hereby I follow Jappelli (1990) who …nds that if binding

borrowing restrictions are age-related, grouping individuals by year of birth is e¢ cient because

cohort members age together.

dummies can be used to control for common trends and unmeasured variables.

"

and "

are error terms.

The probit model only allows the signs of the coe¢ cients to be usefully in- terpreted. In order to gain more from the data, the marginal e¤ects for each explanatory variable, x

; of the probit estimation are calculated:

@P r(F

= 1)

@x

= (

x

)

(3)

where is the probability density function for a standardized normal variable.

The explanatory variables of interest are parental income, beta student, re- peated class, …nancial support, social class, chance of graduating and chance at the labour market. Control variables included are year of study, gender, na- tionality and living situation. The construction of these variables is explicated in Section 5. Parental income is likely to be of high importance, as those from a higher-income family su¤er less from …nancial restrictions than those from a lower-income family. Similarly, one might expect socio-economic class to be an important determinant of …nancing choices, even though there is ambiguity. As explained more elaborately in Section 5, socio-economic class is de…ned accord- ing to the parents’education level. Students from a high socio-economic back- ground may be more (…nancially) literate than those from a low socio-economic background. The former are then likely to value education higher because of the expected positive returns, assuming their parents have bene…tted from obtain- ing a degree. In this case, these type of students more easily borrow compared to low socio-economic students, since the latter might be less aware of the value of education and hence be more debt averse. If such debt aversion exists, a preference towards work arises for the low socio-economic students. Though, one might advocate that preferences work the other way around. Assuming that ability is transmitted from parents onto children

, students from a low socio-economic background might have more di¢ culty managing their studies.

This leaves them with less time available for work and a preference towards bor- rowing. Hence, it is not a priori clear how socio-economic background a¤ects

…nancing decisons.

In order to get the rationale of the upcoming analysis, imagine a stylized

1 1

A simple probit regression shows that the level of parental education signi…cantly a¤ects

the students’ level of education, supporting the view of transmission of ability. See Table A3

in the Appendix.

labour supply model in which a student optimizes its utility by dividing time between work and leisure. Figure 4.1 graphs the model. Work is time spent on a paying job, whereas leisure includes activities where one is not paid, such as education and spare time. The indi¤erence curve shows work and leisure combinations that yield the same amount of total utility. More hours of leisure implies less hours of work. Non-labour income, V, can consist of a basic grant, a supplementary grant, parental contribution and a loan. The latter is income- contingent, allowing students to borrow for study and life expenses with the condition that the repayment depends on their future income. The earning threshold is reached when the supply curve is horizontal, the area which stu- dents tend to dodge

.

Figure 4.1: Student labour supply

After a certain point however, the supply curve rises again because a student never has to repay more than the received SF. The optimal situation, point A, is found to be somewhere in between not working at all and working up to earning

1 2

An informal analysis of the sample students shows that throughout the years, the fraction

of all working students who have a yearly income close to- or above the threshold hovers

around 4 – 8%. For more details, see Figure A1 in the Appendix.

the threshold. This simple model allows us to establish theoretical expectations regarding the reform. Particularly for students with a part-time job, which as discussed concerns the majority of the Dutch students. Broadly, the reform may alter the supply curve in two ways, represented by the red- and blue line.

In either situation, the earning threshold no longer exists. First, the red line shows how a student keeps its non-labour income constant by substituting the basic grant for a loan. As the supply curve now moves straight beyond the old earning threshold, individual labour supply remains unchanged. On the other hand, the blue line depicts how the student does not borrow the entire basic grant amount, but re-optimizes his situation and borrows only part of the basic grant. That is, non-labour income decreases from V to V’, shifting the labour supply curve down. In the new optimal situation, point B, labour supply has increased. Obviously, leisure time decreases. The impact on study time is ambiguous, as it depends on what fraction of leisure time is spent on studying.

In other words, it depends on whether working replaces study time or recreation activities. If the latter is true, it might simply be that students spend their time more e¤ectively.

Important to note is that the above model assumes that studying is leisure time, even though there is a clear trade-o¤ between studying and working.

Investing more time in part-time work and less into studying might extend one’s study period, decreasing the present value of future income. Working may increase future consumption when the accumulated work experience entails higher future wages. But, working may also lower future income if academic performance is hindered by it. The larger the di¤erence between part-time labour income (often minimum wages) and one’s future job, the less rational it might be to obtain a part-time job while studying. Hence, it is of interest to see whether students behave according to these expectations. The data used does not allow for estimating the individual returns to education. However, in the past decade, beta studies are generally found to entail higher-income jobs compared to social and cultural studies.

Therefore, the analysis tries to …nd whether beta students behave di¤erently than their fellow students in adopting

…nancing methods. As might be clear, the simpli…ed model merely serves to show the importance of analysing the decision to work and to borrow together rather than exclusively. That is why next to two-part models, the following section explicates how the two decisions are estimated jointly.

1 3

Berkhout and van der Wer¤ (2015).

4.2 A joint decision

Because F

and F

are fully observable, the vectors of parameters that a¤ect the outcomes can be estimated consistently by separate estimation of two univariate probit models, as discussed in the section above. However, De Luca (2008) …nds that when a bivariate relation exists between the two binary outcomes, it is more e¢ cient to estimate the two equations jointly.

Therefore, by means of a bivariate probit model the individual decision whether to work or not and whether to borrow or not is estimated. In order to make any predictions of the basic grant removal thereafter, this model is estimated on a subsample of students. The reasoning for this is provided in Section 4.3.

Formally, the following equations are estimated jointly:

F

=

x

+ u

F

= 1 if F

> 0; (4) 0 otherwise

F

=

x

+ u

F

= 1 if F

> 0; (5) 0 otherwise

where F

and F

again denote the continuous working and borrow- ing variables respectively, whereas F

and F

denote dummy variables with observed responses 1 and 0. Again, x

is a vector of explanatory variables, time dummies and cohort dummies. The error terms, u

and u

are al- lowed to be correlated. This recognizes that there may be unobservable student characteristics that in‡uence both whether the decision to work and the deci- sion to borrow. The student adopts one of the …nancing modes that maximizes his utility. Four possible …nancing modes are considered: (0) the base mode, involving neither work nor loans; (B) only borrow; (W ) only work; and (BW ) combine loans with work. For sake of convenience, I focus on the last three modes. An advantage of the bivariate probit context over alternative formula- tions is the explicit appearance of the joint probabilities. As mentioned before, the calculation of marginal e¤ects in probit models is common for interpretation purposes. Greene (1996) constructed the quadrant probability marginal e¤ects in bivariate probit models. These are formally displayed as follows:

1 4

A likelihood ratio test con…rms that there is signi…cant correlation between the two

binary outcomes; shown in Table 6.3.

Pr(F

= 1; F

= 1);

(ax

; bx

; ) (6) Pr(F

= 1; F

= 0);

(ax

; bx

; ) (7) Pr(F

= 0; F

= 1);

( ax

; bx

; ) (8) where ax

corresponds to the linear prediction for the working equation and bx

corresponds to the linear prediction for the borrowing equation. Fur- thermore, via adjusted predictions the individual probability of each …nancing mode before and after the reform are obtained. This allows for an informal estimation of the impact of the basic grant removal. Essentially, a student is treated as though he or she does not receive a basic grant grant (regardless of whether he or she actually does). The values of the other independent variables are left as is and the probability that this person would situate themselves in one of the cases is calculated. The same is predicted for when the individual is treated as though he or she does receive a basic grant. This process is repeated for every case in the sample. The only di¤erence between these two populations is their basic grant, which ensures the basic grant to be the cause of the di¤er- ences in the probabilities of a certain state. The results are shown in Section 5.

4.3 The reform and low income groups

Whether or not the shift to the social loan system has substantial impacts is an important point for future research. Though, based on previous years it is possible to prognose what kind of behavior is likely to be induced by the reform.

Student groups of di¤erent income groups are likely to vary in their exposure to the institutional changes.

Removing the basic grant is likely to induce parents to increase their parental contribution whenever possible. However, predictions would be biased if family support itself were adjusted in response to unobserved factors in‡uencing the student’s choice. Even though parental contribution highly correlates with parental income, this contribution amount is virtually unpredictable. That is why in the bivariate probit model, parental contribution is assumed to be unchangeable. To validate this assumption, only the students who receive a supplementary grant are taken into account. Students with such a

1 5

Among others, Cameron and Taber (2004) show that students from a lower-income groups

are in general more dependent on, and more sensitive to, changes in the student aid system.

grant have parents with (relatively) low incomes, which makes it reasonable that their parents have a tight spending limit and hence cannot substantially increase their contribution. Thus, the bivariate model is estimated on a subsample of students receiving a supplementary grant.

4.4 Limitations

There exist a number of important challenges in conducting this type of research.

The research might su¤er from sample selection bias, as the students who present the greatest e¤ect are potentially excluded. These are for instance those who might have succeeded in higher education but have not entered because of debt- aversion or work-preference. This can cause an upward bias towards borrowing preferences and a downward bias towards work preferences. Moreover, students may endogenously choose the number of hours to allocate to work and study.

Higher academic performance can be correlated with personal background, such that unobserved personal characteristics may lead to extensive borrowing as well as participation in the labour market. For instance, a student of high ability might have high academic performance and accordingly expects to obtain a high income job after graduating, making borrowing an attractive …nancing method.

At the same time, the student is probably better able to combine their studies with a part-time job than low-ability students. An attempt is made to control for this by including the self-perceived traits of students. However, these and most of the other variables are based on closed-ended questions. This might cause the validity of the data to be lower compared to administrative data.

Another drawback is that individual histories are not followed. Only a snapshot of the student’s …nancing choice is available. This choice is assumed to be the preference of that student. Though, it could be that if there was access to individual histories, di¤erent preferences would be observed.

5 Data and descriptive statistics

This section provides an overview of the data source and the sample selec-

tion method. The data stems from the Studentenmonitor, organized by Re-

searchNed. It contains repeated cross-sectional surveys over the period 2001

to 2014. 2010 is excluded, because no survey was held this year. The survey

started in 2001, but this paper uses data from 2004 on because of consistency

problems in the survey. The purpose of the Studentenmonitor is to provide a

description of a broad range of determinants of study behaviour and to monitor changes in these elements (Wartenbergh et al., 2010). The surveys take place in spring, which means that each survey year contains information about the preceding academic year. Random sampling is executed based on three strata.

These are level of education, sector type and year of study. The surveys contain nationally representative samples of undergraduate and graduate students, and thus provided data suitable to a national analysis. Several subjects are covered, including background characteristics, type of education, time expenditure and types of income. In almost every year the female population is overrepresented in the sample due to a higher response rate. This has been corrected by means of a weighing procedure on the aforementioned strata.

In order to ensure that reliable inferences can be made I excluded a number of observations, particularly invalid and missing observations. Table A1 in the Appendix shows an overview of the sample selection. Missing data are com- mon in large scale surveys, arising mainly due to non-response in cross-sectional studies. Most variables included in the sample are categorical. If categorical variables in the data have high rates of missing observations, the complete case analysis is likely to provide the smallest bias opposed to any imputation method (Pigott, 2001). The disadvantages are that information is lost and that a poten- tial bias arises when missing values are correlated for some unobserved reason.

To adjust for this, I use a modi…cation of the complete-case analysis that dif- ferentially weights the complete cases, speci…cally using sampling weights. The remaining dataset is one for which each variable has the same meaning.

The sample used consists of 47,106 observations. The descriptive statistics are shown in Table 5.1, which are alienated according to each of the four …- nancing modes. The variables are de…ned as follows.

The monthly …nancial support is composed of self-reported amounts of the basic grant, the supplemen- tary grant and parental contribution. Parental contribution itself is composed of the in cash- and in kind amounts received from parents. Several dummies were constructed from the data. The foreign variable shows whether a student is foreign, the female variable equals one when the student is female. When a student is living independently, the corresponding variable equals one. The university variable equals one when the student attends university and zero if

1 6

Clearly, the analysis relies on self-reported data such that interpretations might di¤er across cohorts and survey years. Although an attempt is made to control for this in the regressions, caution is required when interpreting as the bias is possibly not fully eliminated.

1 7

For an overview of the de…nitions and ranges of the variables, see Table A2 in the Appen-

dix.

college. The beta variable equals one when the student has a beta type of study (e.g. mathematics and physics). The handicap variable denotes whether one has a handicap, which can vary from physical to mental disorders (e.g. dyslexia).

Two variables indicate the socioeconomic background.

Students are catego- rized as having high socio-economic background if both parents have obtained higher education. Similarly, students are de…ned as having a low socioeconomic background if neither parent has obtained higher education. Students of which one of the parents has attended higher education serve as the reference group.

The personal trait variables were constructed from questions asking students to rate their perceived chances on both graduating and the labour market. Dum- mies were constructed equalling one when the perceived chance was high and zero otherwise.

The descriptive statistics show that those who borrow are on average older and live independently more often. For those who do borrow the supplementary grant income is larger. Working students clearly receive lower parental contri- butions than those who do not work. It also shows that students who work are much more often from a low socio-economic background than from a high one. Note moreover that the loan amounts and labour income amounts ddo not di¤er substantially between students using one and students using two of the supplementary …nancing methods. This suggests that the income sources complement each other rather than substitute each other.

1 8

Following Avdic and Gartell (2015).

Table 5.1: Descriptive statistics by …nancing mode

O B W BW

Mean Mean Mean Mean

Age 21.22 (2.46) 22.25 (2.46) 20.98 (2.25) 22.41 (2.32)

Female 0.52 (0.49) 0.47 (0.49) 0.53 (0.49) 0.49 (0.50)

Foreign 0.12 (0.33) 0.15 (0.35) 0.06 (0.24) 0.09 (0.29)

Living independently 0.56 (0.49) 0.79 (0.40) 0.44 (0.49) 0.79 (0.40)

Handicap 0.31 (0.46) 0.32 (0.47) 0.25 (0.43) 0.27 (0.45)

Parents’income

Far below median 0.09 (0.29) 0.10 (0.30) 0.06 (0.23) 0.08 (0.27)

Below median 0.10 (0.30) 0.09 (0.29) 0.08 (0.28) 0.10 (0.30)

Median 0.33 (0.47) 0.38 (0.48) 0.42 (0.49) 0.39 (0.49)

Above median 0.20 (0.40) 0.19 (0.39) 0.19 (0.39) 0.18 (0.38)

Far above median 0.27 (0.45) 0.23 (0.42) 0.24 (0.42) 0.26 (0.44)

University 0.50 (0.50) 0.57 (0.49) 0.38 (0.48) 0.49 (0.49)

Beta study type 0.29 ( 0.45) 0.27 ( 0.45) 0.23 ( 0.42) 0.21 ( 0.41) Repeated class 0.17 ( 0.38) 0.24 ( 0.42) 0.17 ( 0.37) 0.24 ( 0.43) Socio-economic class

High 0.38 (0.48) 0.37 (0.48) 0.26 (0.44) 0.33 (0.47)

Low 0.37 (0.48) 0.35 (0.47) 0.46 (0.49) 0.39 (0.48)

Hours spent working 0 (.) 0 (.) 11.02 (6.61) 11.61 (6.72)

Basic grant 147.83 (104.28) 177.37 (113.12) 140.94 (95.41) 155.54 (117.88) Supplementary grant 42.46 (89.47) 95.65 (157.69) 39.50 (84.85) 80.87 (147.41) Parental contribution 512.83 (403.81) 385.65 (317.32) 388.21 (291.65) 345.71 (279.79)

Loan amount 0 (.) 376.78 (211.43) 0 (.) 364.81 (223.25)

Labour income 0 (.) 0 (.) 292.20 (284.51) 299.00 (264.68)

Chance of graduating 0.51 (0.49) 0.62 (0.48) 0.55 (0.49) 0.66 (0.47) Chances at the 0.30 ( 0.46) 0.27 (0.44) 0.25 (0.430) 0.22 (0.41) labour market

Observations 7,641 3,758 25,482 10,225

Sample averages, standard deviations in parentheses. Note: The students are divided into groups according to their …nancing mode: O is the group not borrowing nor working, group B only borrows, W only works and group BW works and borrows.

1

Likelihood-ratio test for equality of four group means; t-test for equality of two group means

6 Results

In this section, the results of each potential explanatory variable and its e¤ect on the participation decision and the amount (severity) of loans and working are discussed respectively.

All the statements below assume ceteris paribus. Table 6.1 shows the re- sults from using all explanatory variables and their (marginal) e¤ects on the frequency and the severity. The coe¢ cients of the probit do not allow for direct inferences. Hence, the estimated marginal e¤ects and the OLS estimation re- sults are shown

, depicted by equation (3) and (2) respectively. The marginal e¤ects can be interpreted as the chance in the expected value of a dependent variable associated with changing an independent variable. All time dummies were positive and signi…cant at a 0.1% level. Hence, the overall time trend of working of students is positive, con…rming that part-time work is an increasingly important …nancing method for students.

First, the control variables a¤ect the dependent variable in the expected way. With every study year, the chance one works increases, as with the last two years for the severity. Females are more likely to work, but once a student works, being a female decreases working hours by 0.33. Vice versa for foreigners:

they are less likely to work, but once they have a part-time job they work 1.19 hours more. Being in university decreases both the probability and severity of working, possibly due to the need for more self-studying; or a higher work ethic among college students. The beta coe¢ cient shows the same sign with an even stronger e¤ect, which might be because these studies tend to be relatively time-demanding. Regarding the self-perceived traits, when students think they have a high chance of graduating and a high chance at the labour market, they are more likely to work. Note that this might capture motivation in itself: when a student is highly motivated, he might be more optimistic and work-loving at the same time. Socio-economic background provides an interesting result:

the low socio-economic class is more likely to work part-time and work more.

The high socio-economic class is less likely to do so.

Recall that the former represents students with parents that have not obtained higher education. This might cause a family preference towards working part-time as a …nancing

1 9

The probit estimation coe¢ cients are shown in Table A4 in the Appendix.

2 0

Whether this e¤ect is positive or negative depends on whether working a¤ects academic performance. As discussed in Section 2, much ambiguity exists on this topic in the literature.

Hence, the …ndings here do not a priori represent a socially desirable or undesirable e¤ect.

Table 6.1:Two-part model of working

(3) (2)

Study year

Second 0.0427 *** (0.008) -0.1072 (0.125)

Third 0.0660* ** (0.008) 0.2635 ** (0.101)

Fourth 0.0734 *** (0.011) 0.8444 *** (0.176) Female 0.0129 * (0.006) -0.3303 *** (0.056) Foreign -0.1269 * (0.008) 1.1902 *** (0.239) Handicap -0.0453 *** ( 0.005) -0.0915 (0.105) Living independently -0.0246 ** (0.007) 0.9635 *** (0.161) Parents’income

Far below median -0.0839 *** (0.014) -0.1763 (0.284) Below median -0.0510 *** (0.011) -0.4004 (0.207) Above median 0.0138 * (0.005) 0.1913 (0.118) Far above median 0.0179 * (0.007) 0.4864 ** (0.154) University -0.0449 *** (0.006) -1.3146 *** (0.089) Beta study type -0.0539 *** (0.005) -1.7122 *** (0.080) Repeated class -0.0036 (0.008) 0.2896 * (0.131) Socio-economic class

High -0.0312 *** (0.006) -0.5828 *** (0.091)

Low 0.0190 ** (0.006) 0.2980 ** (0.103)

Basic grant -0.0001 (0.000) -0.0071 *** (0.001)

Supplementary grant -0.0014 ** (0.001)

Parental contribution -0.0001 *** (0.000) -0.0020 *** (0.000) Chance of graduating 0.0169 *** (0.005) 0.2203 * (0.101) Chances at the 0.0210 ** (0.006) 0.1445 (0.121) labour market

Constant 13.6740 *** (0.260)

Observations 47,106 13,693

R-squared 0.0906

* p < 0.05, ** p < 0.01, *** p < 0.001; Standard errors in parentheses.

a

The …rst year as benchmark;

The median income group as benchmark Note: column (3) shows the marginal probit estimates and column (2) shows

the OLS estimates

method. Perhaps the most striking result is the e¤ect of parental income on the decision to work. For the parental income, the median income group is taken as a reference. It seems that when students’parents have income below median income, the student is less likely to work, and vice versa for high parental income students. This strengthens the importance of distinguishing family background on both education and income levels.

In Table 6.2, the loan amount is taken as dependent variable. Like the previous table, the marginal e¤ects and the OLS estimation results are shown

. Although males are often found to be more risk loving, no signi…cant e¤ect was found. Recall that students living on their own receive a higher basic grant than those living with their parents. This higher amount does not seem to cover the higher expenses, as living independently increases the probability and severity of borrowing fundamentally. University students are less likely to borrow, but once they borrow they borrow more. This might be attributed to higher indirect study costs (e.g. books) but perhaps also because university students on average look onto a higher future income than college students, which in turn decreases potential debt aversion. Having a beta study type increases the probability of borrowing. This somewhat con…rms the reasoning that students who tend to have higher incomes are more willing to borrow. Repeating a class increases both the probability and severity. These students might want to prevent more study delay by spending more hours on studying and less on other activities, making loans an attractive …nancing source. What strikes out is that participation does not seem to be in‡uenced by the basic grant, but severity does. This mildly suggests that the basic grant removal will not signi…cantly alter participation in borrowing. But given that a student borrows, a 1 euro decrease in the basic grant increases the amount he borrows by 55 cents. This is in line with literature

…nding that grants and loans are imperfect substitutes (see e.g. Linsenmeier

& Rosen, 2002). It appears that the student seeks other ways to …nance its expenses. High socio-economic class seems to positively a¤ect the participation, and vice versa for low socio-economic class. This possibly indicates some kind of …nancial illiteracy as low socio-economic students might be less able to value education and hence be more reluctant to borrow.

However, once a student borrows, his socio-economic background does not seem to matter. In this case, parental income predicts the results in the same line as socio-economic class does.

2 1

The probit estimation coe¢ cients are shown in Table A4 in the Appendix.

2 2

As was already pointed out by Callender & Jackson (2005).

Table 6.2:Two-part model of borrowing

(3) (2)

Study year

Second -0.0085 (0.010) -2.5781 (5.141)

Third -0.0151 (0.008) -5.0446 (5.711)

Fourth 0.0246 * (0.010) 0.3680 (11.151)

Female -0.0327 (0.006) -6.4138 (5.582)

Foreign 0.0456 ** (0.014) 46.5641 *** (6.611)

Handicap 0.0315 *** (0.005) 12.2826 * (4.820)

Living independently 0.2395 *** (0.014) 106.1561 *** (8.804) Parents’income

Far below median -0.0417 *** (0.008) -27.4505 (9.725) Below median -0.0494 *** (0.009) -31.0993 *** (5.015)

Above median -0.0002 (0.009 ) 5.8958 (7.731)

Far above median -0.0034 (0.009) 27.6833 *** (5.517)

University -0.0366 *** (0.006) 21.2580 ** (6.326)

Beta study type 0.0219 *** (0.005) -10.1702 ** (3.802) Repeated class 0.0397 *** (0.008) 20.9777 *** (4.428) Socio-economic class

High 0.0542 *** (0.007) 1.7394 (4.900)

Low -0.0355 *** (0.008) -10.7665 (6.784)

Basic grant -0.0001 (0.000) -0.5530 *** (0.063)

Supplementary grant 0.0006 (0.000) 0.1208 ** (0.045) Parental contribution -0.0002 (0.000) -0.0822 *** (0.009) Chance of graduating 0.0230 (0.006) 13.8224 ** (5.014) Chances at the labour market 0.0038 (0.007) -0.2930 (7.459)

Constant 377.2567 *** (24.241)

Observations 47,106 13,693

R-squared 0.2447

* p < 0.05, ** p < 0.01, *** p < 0.001; Standard errors in parentheses.

a

The …rst year as benchmark;

The median income group as benchmark Note: column (3) shows the marginal probit estimates and column (2) shows

the OLS estimates

What follows is the bivariate probit model estimation. In order to ascertain the magnitude of the e¤ects of the independent variables on working and bor- rowing, the marginal e¤ects are shown in Table 6.3

. As discussed in section 4.3, the model is estimated on a subsample of students receiving a supplemen- tary grant. measures the correlation between borrowing and working after the e¤ects of the explanatory variables have been accounted for. The estimate is -0.0902 and the chi-squared test shows that this estimate is signi…cantly dif- ferent from zero, con…rming the need to estimate the decisions jointly. It also indicates that unobservable factors that are positively related to borrowing are negatively related to working, as for instance debt aversion.

The table shows the average marginal e¤ects, which can be interpreted sim- ilarly to those of the previous models. The e¤ect is on the joint probability of the two outcomes. The three columns show how each variable a¤ects the proba- bility of one of the three …nancing modes. The estimates are broadly similar to those obtained using univariate probit models, suggesting that students in the subsample behave in a similar fashion as those in the whole sample. However, females now seem to be less likely to only borrow and to borrow and work.

This proposes that gender-related di¤erences do exist among students from a lower-income family, but do not appear to be signi…cant when looking at the entire student population. A general rationale was already provided by Japelli (1990), stating that women are sometimes claimed to have inhibitions towards borrowing, inspired either by culture or by future plans not to work. Beta stu- dents do not seem to rely less on loans than others, but study-related di¤erences are observed in the choice to combine borrowing with work.

Living independently predicts its e¤ect in line with the previous models. It now shows that living on your own also increases the joint probability of working and borrowing. Being in university however does not seem to a¤ect this proba- bility. The socio-economic variables provide the same results as before, but high socio-economic students are more likely to use both …nancing methods, whereas being a low-socio-economic student decreases this probability. The former might view the …nancing methods therefore more as supplementary opposed to their fellow students. Note that this is line with the previous results. Once stu- dents work, low socio-economic students work 0.3 hours more than their fellow students, whereas high socio-economic students work 0.58 hours less.

2 3

The bivariate probit estimation results are shown in Table A5 in the Appendix.

Table 6.3: Marginal bivariate probabilities

(6) (7) (8)

Study year

Second 0.0469 ** (0.014) -0.0273 ** (0.008) -0.0022 (0.007) Third 0.0796 *** (0.009) -0.0429 *** (0.006) -0.0110 (0.012) Fourth 0.0801 *** (0.016) -0.0479 *** (0.008) 0.0071 (0.011) Female 0.0385 *** (0.010) -0.0128 * (0.005) -0.0359 *** (0.009) Foreign -0.0995 *** (0.011) 0.0701 *** (0.006) -0.0579 *** (0.014) Handicap -0.0779 *** (0.008) 0.0396 *** (0.004) 0.0169 * (0.007) Living independently -0.1075 *** (0.018) 0.0290 *** (0.008) 0.1365 *** (0.014) University -0.0368 *** (0.011) 0.0230 *** (0.005) -0.0096 (0.008) Beta study type 0.0051 (0.012) 0.0087 (0.006) -0.0548 *** (0.009) Repeated class -0.0416 ** (0.016) 0.0114 (0.009) 0.0472 *** (0.013) Socio-economic class

High -0.0621 *** (0.014) 0.0272 *** (0.007) 0.0356 ** (0.014) Low 0.0972 *** (0.006) -0.0345 *** (0.004) -0.0825 *** (0.006) Basic grant -0.0006 *** (0.000) 0.0003 *** (0.000) 0.0003 *** (0.000) Parental contribution 0.0000 *** (0.000) 0.0000 *** (0.000) -0.0001 *** (0.000) Chance of graduating -0.0258 (0.012) 0.0088 (0.006) 0.0234 ** (0.008) Chances at the 0.0201 (0.019) -0.0145 (0.010) 0.0128 (0.008)

labour market

Observations 13,693

-0.0902

Likelihood ratio of rho = 0: chi2(1) = 22.7156 Prob > chi2 = 0.0000

*p < 0.05, ** p < 0.01, *** p < 0.001; Standard errors in the parentheses

Note: (6) shows the probability of only working, (7) shows the probability of only borrowing and (8) shows the joint probability of both

Hence, the low socio-economic students simply work more hours, lowering the need to supplement their income with a loan. The basic grant and parental contribution are signi…cant, however have low economic value.

The subsample is more homogenous than the complete sample with respect

to their income and expenditure, which allows for setting up expectations about

the removal of the basic grant. Table 6.4 shows the overall predicted probability

per …nancing mode, before and after the reform. That is, the predicted proba-

bility that a student receiving a supplementary grant situates himself in one of the four modes. As can be seen, the coe¢ cients sum up to zero across the four joint probabilities. The results show that before the reform, the probability that a student from a low income family decides to only work is 49%. Estimating the same model as if the student would not receive a basic grant increases this probability up to 59%. Moreover, although it is a minor change, it appears that the modes in which borrowing is a …nancing method become less likely after the reform. This is likely due to the students more often deciding to work and to work more, which decreases the need for a loan. In other words, removing the basic grant does not necessarily make students more eager to borrow. It rather will increase the labour supply of students from low income families, as expected. Note that this does not take into account that the social loan system has more favourable borrowing conditions than before, which might cause the results to be overestimated. If upcoming students are well aware of the renewed borrowing terms, borrowing as a …nancing method might well become more attractive.

Table 6.4: Adjusted predictions of the basic grant removal

O B W BW

Before reform 0.125 *** (0.002) 0.106 *** (0.002) 0.485 *** (0.002) 0.283 *** (0.002) After reform 0.121 *** (0.004) 0.061 *** (0.007) 0.599 *** (0.019) 0.219 *** (0.017) Observations 13,693

* p < 0.05, ** p < 0.01, *** p < 0.001; Delta-method standard errors in the parentheses

Note: O is the group not borrowing nor working, group B only borrows, W only works and group

BW works and borrows

7 Conclusion

In this research I have made an attempt to explicate the determinants of bor- rowing and working behaviour of students. T he results show that one’s socio- economic background matters in deciding how to …nance human capital accu- mulation. First-generation students prefer working part-time over borrowing as a …nancing method. It is also shown that students who are likely to obtain high- income jobs are more willing to borrow than others. The life-cycle theory in this respect is therefore partly supported. An important point for future research is to what extent these factors a¤ect the decision of potential students rather than of those who have already enrolled. Many other explanatory variables predict borrowing and working behavior in a similar fashion as previous literature. That is, university students work less often, females borrow less frequently and the chance one starts working increases by study year. What is more, the reform is expected to increase the labour supply of students from lower-income families, rather than making them more eager to borrow. In addition, it appears to be vital to analyse the two …nancing methods together rather then reviewing them in isolation. This is especially relevant for those countries in which working part-time is common among students.

Since the research relies on survey data, the analysis might su¤er from biased results. Another drawback is that the data only reaches 2014, leaving out the possibility to analyse any direct e¤ects of the reform. Furthermore, potential bias might arise from excluding an important group of students: those who have not entered higher education because of …nancial reasons. At last, the lack of individual histories causes the changes in individuals …nancing decisions to be unobserved. This is an especially important note when one wants to review direct responsiveness of students to …nancial incentives

It is clearly found that students from low socio-economic backgrounds tend

to choose di¤erent …nancing methods than their fellow students. The aim of the

reform was and is to enhance the e¢ ciency of the SF system while guaranteeing

accessibility of higher education. This e¢ ciency is optimal when students choose

the most suitable …nancing method based on rational considerations. Working

part-time might bene…t a student’s e¢ ciacy as long as working hours are substi-

tuted for leisure time. Once working hours replace study hours however, making

use of student loans is a more rational option. In other words, when students

take their ability into account and self-select the …nancial method that maxi-

mizes their life-time utility, no harm is found in any one of the …nancing modes.

The government can play a paternalistic role by informing (upcoming) students about the possibilities and consequences of every …nancing mode. This may improve the …nancial literacy of students with an informational disadvantage.

Nevertheless, the e¢ ciency of the system is in question when moral hazard in- centives arise. That is, students know that they will always pay a proportional amount of their income to repay their debt. This might eliminate strong in- centives to search for high-income jobs. Yet, it is likely that compared to a system dominated by grants, the social loan system also strengthens incentives to graduate sooner.

In conclusion, this study …nds that the reform is likely to change …nancing decisions of students from lower-income families. Particularly, it shows that students will substitute labour income for the basic grant. Adverse e¤ects might occur when students are not able to combine their studies with work e¤ectively.

But when the reform does turn out to induce students to borrow consciously,

study delays might be decreased in turn. Future research might therefore focus

on whether or not the reform did improve e¢ ciency in education. This can be

done by exploiting variations in academic performances, job search durations

and job match quality.