Do study loans impact lifetime financial decision-making?
The influence of student debt on the saving and borrowing behavior of
households
Nikki Isabelle Epema S2208407
Thesis, MSc. Finance
Supervisor: Dr. G.T.J. Zwart January 12, 2017
Keywords: behavioral and household finance, financial decision-making, student loan debt, saving borrowing behavior
Abstract:
1. Introduction
The total study loan debt outstanding keeps on rising in the United States, as indicated in figure 1. At the start of 2016, total student debt of the United States reached the amount of $1400 billions of dollars (Federal Reserve, 2016). This link provides a clock which gives the actual study loan debt.
http://www.finaid.org/loans/studentloandebtclock.phtml
Figure 1: trend in total student loans in the United States in billions of dollars. Source: Federal Reserve, collage board, 2016
Although the total study debt in the US is tremendous, little research is done in the area of student loans and the impact on financial decision-making of postgraduates. American students obtain study loans for College and University education from a variety of sources including different (non) federal and state government. The short and long-term consequences on financial behavior of these loans are not yet fully
understood. In a recent study, Avery and Turner (2012) conclude that college students do not borrow too much, consequently they suppose that going to college is a good investment. In addition, this study clearly demonstrated that students often do not understand the conditions of student loans properly. For instance, students do not know how much money they need to pay back after their studies or against which interest rate they borrow. Moreover, they do not know or misinterpreted the future job prospects of their studies, consequently under borrowing or over borrowing could occur. However,
the study of Avery and Turner (2012) was limited to financial decision-making during the college study period. My study aims to extend this lifetime episode by looking at financial decision-making during the entire adult lifetime. Therefore, my research contributes to the existing literature by looking at the influence of study loan debt on saving and borrowing behavior of households during a long lifecycle period.
Furthermore, this study is relevant to policymakers, because it is the first study to research the influence of the study debt on the saving and borrowing behavior of American adults and therefore unique. Moreover, there are still a lot of uncertainties about the effect of the increasing student loan debts and the outcome of this study will contribute to the ongoing discussion about the impact of increasing student loan debts.
Interestingly, recently the Dutch Government (Department of Education) reformed the financial aid system for students in higher education. During the last decade, financial support consisted of a mixture of grants and study loans. All students received a similar basic grant, which in case of successful graduation would not have to be paid back and thus became a gift of the government. Since 2015, students do not get a grant anymore. Thus, the financial aid system changed from a grant system to a student loan system “a social borrowing system”. Nowadays, students in higher education can only borrow from the government causing the average study loan debt of Dutch students to rise to €24,000 on average per person and amount a total of €12 billion (CPB, 2014). These student loans have to be paid back within a period of 30 years. During this life-time cycle students, graduated or not, will get other financial obligations, e.g. mortgage, rent, car lease, etc. However, as my preliminary investigation showed, reliable information on study loans and borrowing and saving behavior after graduation are not available in the Netherlands and therefore this research was conducted on data from a country which historically has the highest study debt per capita in the world, the US, and detailed data of how household handle their finances are available (Survey of Consumer Finances Database).
The outcome of this study exploring the causes of higher study loan debt on the saving and borrowing behavior of people may be of particularly interest for the current and future situation in both the US and The Netherlands.
robustness tests by successively controlling for different factors like amount paid back of the study loan and excluding households without college degree, with low incomes, with very high incomes respectively. After analyzing all households with incomes over $20,000, my findings clearly indicate that households with study loan debt have higher additional borrowings. Contrasting the findings of the initial data examination of the full model (all respondents between 29-70) which showed, that households with study loan debt are likely to borrow less. These results may be explained by the assumption that holding study loan debt tightens borrowing constraints for instance to mortgage loans, vehicle loans etc. Intuitively, the amount of savings/borrowings could be influenced by lots of factors. However, this study aims to investigate only the causal relation between the study loan debts and later borrowing/saving behavior of households without all other unobservable factors. Therefore, I used the instrumental variable analyzes method, which is an econometrical approach to estimate causal relationships by using an instrument. In this study the variable average federal and non-federal study loan per
student is used as instrument (Gicheva, 2016).
This research will study the influence of study debt on the borrowing behavior among households. Additionally, individuals and households have to make financial decisions about their savings. Little is known about the impact of a study debt on the saving behavior of adults. Thus, this study aims to study a causal relation between study loan debt and the financial decision-making of households.
Accordingly, the following research question will be studied:
What is the influence of having a student loan on the saving and borrowing behavior during adult life?
2. Literature review
This section provides a literature review of the most relevant papers on financial decision-making. Besides, Ando and Modigliani’s (1963) life-cycle hypothesis gives guidance for the reviewed literature regarding the impact of having a student loans on different aspects in life. Furthermore, this section presents the formed hypotheses.
2.1 Saving behavior
The life-cycle hypothesis of saving is a standard economic framework developed by Modigliani and Brumberg (1954). This model defines “saving and dissaving, i.e.
spending, as the positive or negative change in the net worth of an individual during a specified time period”. Thus, if a person earns momentarily more money, he or she tends to save this for future lower earnings and vice versa, if at this point in time a person earns less money, he or she tends to save less in consideration of expected higher earnings in the future. Implying, students have on average a low income and tend to make considerable costs for, among others, tuition fees, housing, living costs etc., this with the thought of greater expected earnings after graduation as shown in figure 2. During retirement people tend to dissave, e.g. people tend to spend the savings they earned during they working life cycle (fig 2).
2.2 Borrowing behavior
The life-cycle hypotheses states consumption is primarily affected by permanent income and people tend to smooth spending’s over their life-time. Under this hypothesis students and young families would have low savings and borrow relatively more
because of higher expected earnings. However, Duca and Rosenthal (1991) found that current income also affects the level of consumption. Since consumption is either
than possible under the specific constraints. Moreover, higher study loan debt could tighten these borrowing constraints such as for example mortgage loans. This could lead to lower borrowings as a result of owning a study loan debt. On the other hand, even though there are borrowing constraints, some people attract too much debt, because they are “unaware of the consequences of their own choices” (Lusardi, 2015). This could mean, that households with study loan debts still obtain too many additional loans.
One of the possible explanations for this observation was studied earlier. In order to make adequate financial decision concerning student loan, it is useful to possess some financial knowledge. Lusardi et al. (2010) conclude financial literacy among young individuals is particularly low. They tested whether young people could answer basic questions regarding interest rates, inflation and risk diversification, only 27% could provide sufficient answers to these questions. Financial illiteracy is often linked to suboptimal financial decision-making. Implying that improvement of financial
knowledge plays an important role in “savvier savings and investment decisions, better debt management, more retirement planning, higher participation in the stock market and greater wealth accumulation” (Lusardi and Mitchell, 2014).
2.3 Student loan
Figure 2: Life-Cycle hypothesis framework (source: Wolla, 2014)
According to Davies and Lea (1995) students can be seen as a group with “relatively low-income and high debt”. However, the relatively low-income would be temporarily and debt should be made in order to perceive a higher expected permanent income. Their findings are in line with the life cycle hypothesis, as is visible in figure 2. In
addition, they found that students are averse toward student debt at first, but once they incurred debt the students became more tolerant toward their own debt level. The degree of tolerance toward debt levels is part of a student’s financial behavior. Worthy et al (2010) studied several factors that influence this financial behavior of students. During college or university students could face problematic financial behavior, due to, among others, rising tuition fees and easy accessibility to credit cards. In order to establish a relation between student financial behavior and factors like “age, gender, public assistance, adult status, sensation-seeking and risk taking” Worthy et al (2010) used Poisson regressions. This regression for count data is used to look whether there appears to be correlations between the dependent- and independent variable, rather than show causal relation.
Interestingly, other life-cycle issues seem related to student loans apart from problematic financial behavior and their high tolerance toward debt. In a recent study having a student loan was strongly associated with a smaller probability to get married after graduation compared to graduates without a student loan (Gicheva, 2016).
interesting since a similar method to establish causal relationship is used, in this case between the amount of student debt a person holds and the probability to get married. As Gicheva (2016) indicates, the problem to estimate this causal relationship is, that there could occur reverse causality, meaning that “marital status affects the schooling decisions” instead of the other way around. To overcome this bias problem, Gicheva uses
average amount borrowed per full-time student as instrument to “offer strong evidence
that student loans have a negative and significant, both statistically and econometrically, impact on the probability of marriage” (Gicheva, 2016). Figure 4 shows the instrument
average amount borrowed per student and the methodology section provides a
description of the instrumental variable regression used in this study to establish a causal relationship between the amount of study loans and saving and borrowing behavior.
The total level of study loan debt in the US continues to rise during the last 10 years (fig 1) and crossed the level of $1400 billion debt in 2016. Aker and Chingos (2014) tried to answer the question whether a student loan crisis is emerging, because of the rapid increase of total student loan debt in the United States and the inability of many borrowers to pay back their loans by comparing to the past. However, they found (Aker and Chingos, 2014) no evidence that the substantial increase of total student loan debt creates the next crisis, because students could get higher-paying jobs after graduation and therefore their financial position (of 2010) is not worse than the financial position of people in 1992. In addition, they observed financial distress to decrease for
individuals who hold a student loan over time. In contrast, Bricker and Thompson (2016) conclude that households with student loans have a higher chance of financial distress, compared to households without student loans. In addition, they showed that households with study loan debt were more likely to be denied for further loans. They analyzed a special data batch of the SCF, which is panel data for the years 2007 to 2009 (Note, the SCF data batches used in this study are not panel data). Initially, they
The reviewed literature suggests a negative relation between student loan debt and several factors mentioned before. Since, student loans are commonly paid back from the individual's total assets, savings could be lower because study loan has not been paid back yet. In addition, behavioral factors like not paying back because of financial illiteracy or just getting used to a situation of increased debt
This leads to the following hypothesis:
Hypothesis 1: Savings are expected to be lower if American adults are in possession of a study loan.
The results of Gicheva and Thompson (2015) show, that people who hold a student debt are likely to have a decreased financial stability, especially for individuals who did not complete their Bachelor's degree. Furthermore, households with study loan debt have higher probabilities to file bankruptcy and get credit constrained. Moreover, their study is particularly interesting since they use the same database (Survey of
Consumer Finances) and they establish causal relations rather than descriptive by using an instrument for the amount of study loan borrowed. As Gicheva and Thompson (2015) indicate, financial stability could be influenced by for instance job stability and starting wage, which could be influenced by (unobservable) factors such as academic success and type of education, and these two factors have influence on the amount of study loan debt. The instrumental variable approach helps to avoid that the results would be biased by these unobservable factors. Similarly, this approach is used in my study. Moreover, Mishory and O’ Sullivan (2012) suggest graduate students who would like to buy a house after college are more likely to get denied for a typical home mortgage, due to the higher debt-to-income ratio of an average graduate.
Different borrowing studies produced contradicting results. On the one hand, a study loan debt could tighten borrowing constraints, which could lead to lower
borrowing too much. Taken together these results lead to my hypothesis for the analyzes in the current study of borrowing behavior for households with study loans:
3. Data
Below a description of the Survey of Consumer Finances (SCF) data is provided. In addition, table 1 presents the main characteristics of the participants and the
summary statistics. This section identifies the key variables, as well as the questions and statements selected from the SCF.
3.1. Data
This study makes use of the SCF, the database is available via the Federal Reserve. The Federal Reserve, together with the Department of Ministry sponsors the SCF. Data collection is accomplished by National Opinion Research Center (NORC) at the University of Chicago. Participants need to answer detailed questions regarding the families’ financial situation, including their assets, liabilities, income, investment, debt, pension and demographical characteristics. In addition, detailed information concerning the respondents’ educational loans is provided. Furthermore, these questions are asked via interviews either through telephone or in person with Computer Assistant Personal Interviewing (CAPI). (SCF Codebook, 2013)
To study the influence of having a student loan on the financial decision making behavior of the participants of the SCF, data are obtained from the SCF survey batches, which ranges from 1995 until 2013. The SCF is executed every 3 years, so in total this study analyzes seven batches of data.
In order to select the between 4,000-6,000 participating US families per survey batch, the SCF uses two kinds of random sampling techniques. First, approximately 75% of the respondents were selected by standard multistage area-probability sampling, in other words, a geographically based random sampling technique. This group contains specific characteristics, which makes them a representative sample for the US
population. In other words, every American citizen has equal likelihood to be selected for the SCF, i.e. cross-sectional data. Second, in order to provide researchers the opportunity to study uncommonly held assets, SCF selected for the remaining 25% wealthy families from a list sample based on their tax returns. (SCF Codebook, 2013)
of 29 years or younger were dropped from the analyzed data, due to the inability to study their financial behavior. Although, respondents of 29 years or younger most likely already completed their higher education, they have relatively low consumer debt and savings (Gicheva and Thompson, 2015). So, it would be difficult to analyze the impact of having a study loan on saving and borrowing behavior.Additionally, seven respondents were dropped from the sample, because they studied before 1971 and data of average study loan debt per student of all American students is not available for the years before 1971. In this study, the average study loan debt per student of all American students is used as an instrument in the two stages least square method and this will be explained in the next chapter in more detail.
The SCF is carried out as a cross-sectional study, which enables to research the influences of study loan in different years of time and among different cycles of the economy. A limitation of using SCF data is the fact that SCF does not make a difference between whether the study loan was incurred by the head of the family, the partner, spouse or by the children.
3.2. Key variables
The purpose of this study is to relate saving and borrowing behavior of US adults with the amount of study loan a person has. Saving behavior and borrowing behavior are both dependent variables and the amount of study loan is the independent variable.
3.2.1. Saving behavior
or “not checked”. Enclosed in appendix B are the questions from the SCF codebook, which relate to the five ways to measure saving behavior.
Total savings are adjusted for inflation, since only deliberate acts of savings should be included, according to Nyhus and Webley (2001).
3.2.2. Borrowing behavior
The second dependent variable studied in this thesis is borrowing behavior. Federal Reserve defines in the SCF codebook that total borrowings incorporate the following forms of debt: “credit card or store debt, mortgage debt, home equity loan, other home purchase loan, home improvement loan, loan for other real estate, line of credit, business loan, vehicle loan, education loan, margin loan, insurance loan, pension loan and other instalment loan”.
To test the second hypotheses, the same approach as by saving behavior is used to measure borrowing behavior. Hence, first total borrowings excluding educational loan US dollars; second, ratio of total borrowings to net income excluding educational loan; third, willingness to borrow. Willingness to borrow is derived from a question if the respondent finds it a good idea to buy consumer goods on installment loans or on credit (Codebook, 2013).
3.2.3. Study loan
To research the formed hypotheses, the independent variable study loan is included in the model. Respondents of the SCF questionnaire specified, “the amount of educational loan their family holds” (SCD Codebook, 2013) consisting of the total
amount of study loan a family borrowed over the years, before paying off the study loan. Moreover, participants answered several questions regarding their study loan: whether they possess a study loan, the total amount of study loan borrowed, the number of study loans a person has, the total amount of the remaining study loan. Relevant questions from the survey are presented in appendix B.
Figure 3 shows the number of respondents with study loan compared to the total number of respondents that attended college. There is a slightly upward trend, caused by the enlargement of the total sample size of the dataset.
Figure 3: Number of respondents who studied with and without study loan in SCF. 3.3. Summary statistics
The summary statistics of all respondents between 29 and 70 years old of the SCF batches from 1995 until 2013 are presented in table 1. Of the nearly 27,000 analyzed respondents, 3,371 respondents possess one or more study loans, with an average of $21590.80 study loan per respondent (total amount of study loan households possess before they start paying of this study loan debt). Slightly less than half of the participants got a college degree and a large part continued studying to pass their bachelor’s degree (44%) and half of this group received a master degree or doctorate (21%).
Table 1 shows the different dependent variables. For instance, the median respondent had almost $ 5,000 in their savings account for the given year the survey was conducted, in contrast to the median borrowings of almost $50,000. Furthermore, the families’ total savings are 12% of their annual normal income. And almost one out of seven respondents spend more money than they earn and do not like to save.
Borrowings contain several loans, for example vehicle loans, mortgage, credit card debt, 0 500 1000 1500 2000 2500 3000 3500 1995 1998 2001 2004 2007 2010 2013 in n u m b er s of re sp on d en ts Year of SCF
Number of people studied within SCF data
etc. These loans have, on average, a value of almost twice the amount of a household’s yearly income.
The survey contains 78%, 11% and 8%, respectively, white-, African-American- and Hispanic people, with an average age of 49 years old. Only 19% of the respondents are female, this low percentage might presumably be due to automatic assignment of men as head of the household in a mixed-sex couple by the Federal Reserve (Codebook, 2013). For more demographical details please view the table below.
Table 1: Summary Statistics values are $, year (yr), yes no: Y/N (1,0)
Variable Observations Mean Median Min Max
Study loan ($) 3,371 21,590.80 12,000 100 400,000
Year study loan taken out (yr) 3,371 2000 2001 1971 2013
Total savings ($) 26,909 34,900.38 5,000 0 55,500,000
Total savings/normal income 26,909 0.12 0.06 0 85.77
Spending’s exceeded income
(y/n) 26,909 0.13
0
0
1
Do not save (y/n) 26,909 0.14 0 0 1
Willingness to save ($) 26,909 193,499.20 10,000 0 135,000,000
Borrowings ($) 26,909 293,978.40 49,270.41 0 276,000,000
Borrowings/normal income 26,909 1.52 0.48 0 5250
Good idea to buy things on
instalment plan (y/n) 26,909 0.28 0 0 1
Normal income ($) 29,909 705,642.10 64,600 0 328,000,000
Home owner (y/n) 26,909 0.74 1 0 1
Female (y/n) 26,909 0.19 0 0 1
College degree (y/n) 26,909 0.49 0 0 1
Bachelor’s degree (y/n) 26,909 0.44 0 0 1
Master’s degree (y/n) 26,909 0.21 0 0 1
Doctorate (y/n) 26,909 0.09 0 0 1
White (y/n) 26,909 0.78 1 0 1
Black (y/n) 26,909 0.11 0 0 1
Hispanic (y/n) 26,909 0.08 0 0 1
Age (yr) 26,909 49.28 49 29 70
College age children (yr) 26,909 0.21 0 0 10
Disabled (y/n) 26,909 0.07 0 0 1
Retired (y/n) 26,909 0.12 0 0 1
Table 2 shows the trend of study loans among respondents of the SCF. The percentage of households with one or more study loans increased from 9,3% to 17,7%. Apart from additional possession of study loans by respondents of SCF, they also borrowed a higher amount. In SCF batch of 2013, the average amount of study loan borrowed by households was more three times higher than in 1995. These observed trends in SCF data are in line with upward trends in federal and non-federal study loans nationwide in the US, as illustrated in figure 4.
Table 2: Descriptives of SCF respondents containing a study loan
4. Methodology
This section contains the methodology and several conditions needed to process the SCF data. Furthermore, instrumental variable analyzes are conducted in order to test the hypotheses. In addition, regression equations for saving and borrowing behavior are provided.
This study uses SCF data to test the hypotheses. The used dataset is survey data and contains of multiple entities (i.e. respondents), for every respondent is a large set of variables available. Some of these variables contain nonnumeric answers or some form of rating scales; all these answers are turned into binary data in order to properly perform the statistical tests. In addition, the logarithm of normal income is used to reduce skewness of the data, because the dataset contains wealthy respondents with very high incomes (the range of highest 1% income: $11,800,000 - $328,000,000 (range) with a median of $ 23,300,000).
SCF used multiple imputation method to impute missing data into the dataset, consequently for every single respondent five observations exist in the dataset. The five observations for every respondent are combined together, so there is only one
observation per respondent. If the five ‘observations’ for one person were not equal, the average of these five observations is used.
This research aims to discover if there exists a causal relationship between
having a study loan and saving- and borrowing behavior of American adults. Specifically, whether a higher study loans cause lower savings and higher borrowings. Additionally, this research would like to provide an explanation for this phenomenon.
A standard statistical method to analyze causal relationship between study loan and saving- and borrowing behavior is ordinary least square (OLS) method. OLS is a commonly used “approach to estimate linear regression models” (Brooks, 2014). However, a drawback of using OLS is that it could be indecisive if saving (or either borrowing) behavior is influenced by study loan or by another unobserved variable, for instance income, education, etc. Moreover, the following arguments will give an
contains not constant variances in the error term, hence the dataset violates the homoscedasticity assumption and exhibits heteroscedasticity. If the error terms in the dataset suffer from heteroscedasticity, OLS is no longer Best Linear Unbiased Estimator (BLUE). To solve the heteroscedasticity problem, robust t-statistics are used. The Gauss-Markov theory implies that OLS estimator is a proper method to estimate a linear regression model, if it is BLUE (Brooks, 2014). One of the assumptions of Gauss-Markov is violated in this dataset. To be precise, the estimator is biased because there occurs to be reversed causality between study loan and the dependent variables, for instance saving, willingness to save, borrowings etc. Either savings could be lower because a person has a study loan and for example needs to pay this off or the study loan could be higher because a person does not have sufficient savings to pay for their education. When a Gauss-Markov assumption is violated, the independent variable study loan is no longer exogenous and should be treated as an endogenous variable. Endogenous
variables are determined inside the model, in contrast to exogenous variables which are “determined outside the equation” (Brooks, 2014). In this case, instrumental variable estimator could be used to “replace the endogenous variable (study loan)” by the instrument and the instrument should be correlated to study loan but not to the error term (Brooks, 2014). To look whether a variable is endogenous, one could perform a Durbin score and Wu-Hausman test to if study loan was actually exogenous (Wu’s, 1974 and Hausman, 1978). Both tests are post estimations and are performed after the
instrumental variable analysis.
Instrumental variable estimation is a tool to calculate if lower savings could partially be explained by possessing a high study loan and if higher borrowings are affected by higher study loans. Instrumental variable analysis is often referred to as two-stages least square method. The advantage of using an instrument is that it solely looks if study loan and the control variables cause either higher or lower savings/borrowings and this causal relationship would not be due to unobservable other factors for which could not be controlled for. Hence, instrumental variables analysis is an econometrical estimator, which allows for finding the specific influence of the independent variable on dependent variable. Selecting an instrument could be a difficult task because the
the United States is taken as instrument. They researched “the effects of student loans on long-term household financial stability” using SCF data from 1995 until 2010 and whether people would get married less fast if they have a study loan (Gicheva and Thompson, 2015). Since, Gicheva and Thompson use the same data source as this study, the instrument is well fitted. Further, the instrument is proper to use, because the average federal and non-federal loans per student would likely have great influence at the amount a student borrows in a specific year. This amount students borrow on average differs per year, this trend is shown in figure 4. In addition, the level of savings of an American citizen would likely not be related to the amount students borrow on average that year. Similarly, whether a person takes out a mortgage would possibly not be influenced by the trend of student debt.
Figure 4 shows the instrument, the average is taken by dividing all federal and non-federal loans by all students of that specific study year in the United States (College board, 2016). In Higher Education Act of 1965, the US Department of Education allowed banks and private lenders to provide study loans to students, in order to support
Figure 4: the instrument average federal and non-federal loans per student in 2016 US$. Source: College Board, 2016.
In addition, a first stage regression statistic test is performed in order to test if the instrument is weak. This test is called the Cragg-Donald F-test and tests whether the instrument is correlated with the endogenous variable. Thus, the test actually performs the first stage of the two stages least square method. The instrument average federal and
non-federal loans per student are in every performed instrumental variable regression
significantly correlated with the endogenous variable study loan. Hence, the used
instrument is a strong instrument in every performed analysis. The importance of using a strong instrument is that weak instruments could lead to biased outcomes and “the actual rejection rate of the null hypotheses may be larger”. (Adkins & Hill, 2011)
Furthermore, Bucciol and Veronesi (2014) state that to research influences on saving and borrowing behavior, the equation should include socio-demographic
characteristics and a dummy to control for time differences between the survey batches.
The socio-demographic characteristics of the SCF dataset include characteristics about the participants and household, like total net income, number of members, home ownership, age, gender, ethnicity, education and employment status (Bucciol and Veronesi, 2014). Moreover, in all instrumental variable regressions are survey batch
$0 $1,000 $2,000 $3,000 $4,000 $5,000 $6,000 $7,000 $8,000 $9,000 1971 1973 1975 1977 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015
Average Federal and Nonfederal Loans per student
instrumental variable analyzes use study loans in 1,000 of dollars, to enhance the impact of a study loan increase.
The following formula is established to calculate the influence of having a study loan on the saving and borrowing behavior of US adults:
Yi = + 1*study loani + 2 * socio-demographic characteristicsi + 3 * survey batch dummyi
+ i (1)
5. Results
This result section provides answers to the main research question; ‘what the influence is of having a student loan on the saving and borrowing behavior during adult life?’. In the first paragraph, outcomes of the instrumental variable analysis of the full model are analyzed, followed by several robustness tests using the instrumental variable regression. The presented robustness tests regard: 1) results limited to
respondents with college degree only; 2) a check to look if savings are lower because of respondents paying off their study loan, therefore there is controlled for paying off study loan and 3) respondents with a minimum normal income of $20,000 annually were analyzed.
5.1. Full model
This paragraph uses the instrumental variable analysis to test the influence of having a study loan on the saving and borrowing behavior of respondents between 29 and 70 years old. As indicated in tables 3 and 4, independent variable study loan is tested on dependent variables for saving behavior and borrowing behavior. As described in the ‘Data’ section, saving behavior includes the following five dependent variables: ‘how much a respondent saved; saving divided by normal income; if a particular
respondent normally spends more than their income; how much savings a respondent needs and if a respondent does not like to save’. In order to look if respondent borrowing
behavior is influenced by study loan and in particularly if respondents tend to borrow more money if they are in possession of a study loan, three dependent variables are used. These three dependent variables: ‘amount borrowed exclusive educational loan’, ‘borrowings divided by normal income’ and ‘whether respondents find it a good idea to buy
consumer goods on installment loans’, as shown in table 4.
respondent without study loan. This could be caused by lower saving expectations as a result of needing to pay off their study loan.
In addition, table 4 presents the results of the instrumental variable analyzes regarding study loan and borrowing behavior. The first column indicates; a $10,000 increase of study loan causes a decline of approximately $ 44,000 in total debt of a respondent. However, a $10,000 increase in study loan causes borrowings to increase 3% compared to normal income. Both borrowing behavior results are contradiction each other. An explanation for this contradictory observation could beborrowing
constraints imposed on low incomes. Furthermore, respondents with study loan are less likely to buy consumer goods on installment loans (table 4).
In addition, appendix A provides tables with results of the influence of having a study loan on the saving and borrowing behavior with all control variables. For instance, the control variable logarithm of normal income is highly significant for all saving
behavior variables. Log normal income is taken as covariate because naturally the amount a person earns could have impact on the amount someone saves and the amount a person could borrow. Since expected future earnings are higher when a person attends college or university, different educational achievements are included as covariates. Moreover, educational achievements could impact the amount a person saves and borrows, because higher earnings prospects could lead to attracting larger amounts of debt and could provide the opportunity to save more. In addition, several other covariates control for different demographical characteristics of a respondent.
in the model. Respondent A would save $3256.03 less than if he or she would not have no study loan. In addition, respondent A’s borrowings declined over $12,000 dollars. Moreover, respondent B saves nearly $60,000 dollars less and borrow $219,500.74 less than if he or she would not have taken out a study loan.
My first ‘intuitive’ approach to look at my data I plotted them followed by simple regression performed to look if there is a broad relation between study loan debt and saving and borrowing behavior. The outcomes of ordinary least square regression (table A.9 in appendix A) show similar to the instrumental variable approach that savings and borrowings would be lower if households are in possession of a study loan debt. The significant correlations have a less pronounced effect than the results of instrumental variable approach.
This dataset included wealthy families to research uncommonly held assets. To analyze if these wealthy families influence the outcomes, I performed an additional analyzes in which the highest 1% income group was excluded. The results did not change the results much; savings are still negatively influenced by study loan debt of households only more limited. In addition, households with study loan debt would have lower borrowings, again slightly more less pronounced than with wealthy families included.
Table 1
Instrumental variable regression of the influence of study loan debt on saving behavior using all respondents between 29 and 70 years old. Robust t-statistics are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. Appendix A.1 shows full table with the control variables. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Study loan
Saving -1172.21***
Saving to normal income -0.003*** Spending exceeded income 0.003***
Willingness to save -3828.13***
Table 4
Instrumental variable regression of the influence of study loan debt on borrowing behavior using all respondents between 29 and 70 years old. Robust t-statistics are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. Appendix A.2 shows full table with the control variables. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Study loan
Borrowings -4398.67***
Borrowings to normal income 0.003*
Good idea to purchase on installment loans -0.001**
In summary, savings and borrowings will be lower when a household has a study loan debt. However, this result is based on all respondents between 29 and 70 years old and in order to make a better comparison between the group of respondents with study loan and the group respondents without study loan, the next paragraph will solely review respondents with a college degree.
5.2. Robustness test: respondents with College degree only
Several robustness tests are performed in order to make the previous results more robust. This paragraph reviews the first robustness test, in which the instrumental variable analyzes is performed exclusively on respondents with a college degree. People with at least a college degree are likely to have higher earnings, which may lead to higher savings but also to more borrowing compared to people without college degree. Therefore, respondents without college degree are excluded from this analyzes.
Table 5 and 6 show the results of the instrumental variable regression on saving and borrowing behavior of respondents between 29 and 70 years old that completed college. The coefficient for study loan is in every analysis highly significant. Like previous instrumental variable regression of the full model, the dependent variable
saving is still negative and is slightly larger. Implying, respondents with a study loan
the group with study loan have less savings and lower borrowing. This may be related to tighter borrowing constraints due to high study debt.
Similar to the full model presented in the paragraph above, the coefficient of borrowings to normal income is positive and the coefficient of borrowings is negative, and again this results contradict each other. However, the positive coefficient of borrowings to normal income increased greatly by excluding respondents without college degree. In the full model a respondent with $10,000 study loan debt would borrow 3% more compared to their normal income, now a college degree respondent with $10,000 study loan debt would borrow 20% more compared to normal income.
An economic interpretation of the results presented in table 5 and 6 could be provided using the example of paragraph 5.1. Respondent A’s (recall, on 10th percentile with $2820 study loan debt) savings would be slightly lower compared to the full model. Respondent B (90th percentile with $ 50,000 study loan debt) saving would decrease $5,000 compared to the full model. Similarly borrowing is affected by study debt. More study debt results in less borrowing, but in the full model borrowings were even lower, indicating stronger borrow constraints for household without college degree and household with study loan debt.
Table 5
Instrumental variable regression of the influence of study loan debt on saving behavior using all respondents with college degree between 29 and 70 years old. Robust t-statistics are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. Appendix A.3 shows full table with the control variables. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Table 6
Instrumental variable regression of the influence of study loan debt on saving behavior using all respondents with college degree between 29 and 70 years old. Robust t-statistics are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. Appendix A.4 shows full table with the control variables. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Study loan
Borrowings -2837.71***
Borrowings to normal income 0.02***
5.3. Robustness test: Nothing paid off of study loan
This robustness test is performed to look whether savings of the previous paragraphs are lower because households are paying of their study loan. Initially, I wanted to perform this robustness test by analyzing (instrumental variable analyzes) the saving and borrowing behavior of households that finished paid off their study loan. However, surprisingly none of the households finished paying off their study loan. In other words, all 3,371 respondents with study debt are still paying off or did not yet start paying off. In theory households that did not start paying off their study loan, could not have lower savings because they use that money to pay off their study loan.
Therefore, an additional variable nothing paid off of respondents’ study loan is created to control for the fact that household’s savings could be lower because they use part of their income that was supposed to be for their saving account to pay off their study loan.
Study loan
Saving -1299.65***
Saving to normal income -0.003*** Spending exceeded income 0.002***
Willingness to save -2174.92***
Table 7 shows the results of analyzing savings, willingness to save and
borrowings for respondents with and without college degree. The covariate nothing
paid off from study loan is in nearly every analysis significant. Including this covariate
makes the results more robust. Respondents’ savings would be slightly lower if there is controlled for nothing paid off of their study loan, compared to the ‘Full model’ and the robustness test of respondents with college degree only. To compare the results of college completers dataset, indicate, an increase of $1,000 study loan, leads to $1458.62 less savings, $2819.14 wanting to save less and almost $4,000 dollars less borrows. Thus, my typical respondent A (10th percentile lowest) would save $3727.72 less and respondent B (90th percentile) would save $66891.82 less, and A and B would borrow $12,831.63 and $231,263.72 less respectively. In summary, paying off the study loan had only minor but still limiting effects on both saving and borrowing capacity in both the entire group and the college group.
Financial behavior in the lowest income category may be different, because the data showed only minimal savings (median $0) and borrowing (median $911) which may be associated with limited availability of borrowing and savings. Therefore, a next step in the analysis is to exclude respondents with lower income ($20,000).
Table 7
Instrumental variable regression of the influence of study loan debt on saving, willingness to save and borrowings using all respondents with college degree between 29 and 70 years old. In this table is controlled for respondents who paid nothing off from their study loan. Furthermore, dependent variables are presented for all respondents and respondents with a college degree. Robust t-statistics are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. Appendix A.5 and A.6 shows full table with the control variables. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Study loan
Saving -1339.45***
Saving Respondents with college degree -1458.62***
Willingness to save -4910.84***
Willingness to save respondents with college
degree -2819.14***
Borrowings -4634.10***
5.4. Robustness test: Normal income over $20,000
At first, all respondents with income higher than $20,000 are analyzed. Next, will be controlled for respondents who paid nothing off from their study loan. In addition, the dataset limits to college completers. Finally, both situations were controlled for.
Interestingly, in this specific group all borrowing coefficients are positive, meaning more debts for households with a study loan (table 8). In all previous results, borrowings would be lower if a respondent increases their study loan. However, excluding respondents with low incomes from the dataset results in larger debt for respondent possessing a study loan. If a respondent’s study loan increases with $1,000, his borrowings would increase by $2391.87 as shown in table 8. The shift of lower borrowings to larger borrowings, due to possession of a study loan, could be because of behavior changes, the respondent’s tolerance toward debt changed. Moreover, in this group the savings were also lower, but this effect was limited. For instance, the
coefficient of study loan under college completion sample is -478.09 (table 8) and the coefficient of study loan is -1299.65 (table 5). Thus, by excluding low incomes
respondents still save less when they have a study loan, but savings would be more than twice that low if there was not controlled for low incomes. Therefore, low incomes have significant influence on the analyzed models. In addition, total debt of respondent A and respondent B would be respectively, $5590.17 and $101,648.91 higher than
respondents who did not take out a student loan to finance their study. Removing the lowest income group, which may bias the data, shows a spectacular change in the borrowing capacity. This changed with around $300,000 for our respondent B.
Table 8
Instrumental variable regression of the influence of study loan debt on savings and borrowings for respondents (between 29-70 years old) with earning higher than $20,000 annually. A distinction is made between all respondents, all respondents controlled for nothing paid off from their study loan, respondents with college degree and
respondents with college degree and controlled for nothing paid off from their study loan. Robust t-statistics are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. Appendix A.7 and A.8 shows full table with the control variables. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Saving Borrowing
All respondents -592.55*** 663.03
Nothing paid back from study
loan -784.68*** 183.14
With college degree -478.09** 2029.21***
With college degree and nothing paid back from study loan
6. Conclusion
The main questions in this study were studied using data from the Survey of the Consumer Finances. The subsequent analysis clearly demonstrated that having a study loan debt from college means that households with income over $20,000 have higher borrowings compared to the similar group without study loans. In addition, the households with study loan debts had lower savings. This negative effect on savings, more study loan lower savings, was also present after all different robustness test I performed to correct for college degree, nothing paid back of the study loan and low incomes. Therefore, this study shows that during adult life savings are lower if American adults are in the possession of a study loan. In addition, the study reveals that
borrowings are higher if American adults are in the possession of a study loan, except for the group with the lowest income. Due the use of the instrumental variables analysis I was able to demonstrate this result and finally could confirm the causal relationships of both hypothesis of this study.
The distinctiveness of this study is robust by using an econometrical approach which discovered causal relations between the studied variables. By using the
instrumental variable method, I was able to approach my formula to estimate saving and borrowing behavior of American adults using their study loan debt and additional
control variables associated with expected earnings (higher labor market earnings, Gicheva and Thompson, 2015).
The results of the borrowing part of the analyzes showed initially contradicting results, e.g. the full model analysis showed lower borrowings for households with study loan but the ratio borrowings to income was higher, meaning higher borrowings.
mortgage, which showed that half of SCF respondents would like to have more debts than allowed under the constraint (Duca and Rosenthall, 1991).
Another interesting finding emerged from sub analysis of confounding factors like how did paying off the study loan affect the saving and borrowing behavior. To my surprise the data showed that in the entire database no one finished paying off their study loan debt despite the fact that this study has a maximum timespan of 42 years. Initially, my idea was to control for the idea that savings could be lower due to
households paying off the study debts. However, this proved impossible because I had no group to compare with because no one finished paying off. Therefore, I used households that did not start paying back their study loan as control variable. The results of these analyzes establish that having a study loan is associated with less saving than comparable households without a study loan.
A limitation of this study is that the household is the smallest unit which could be studied, because of the organization of the SCF database and not the actual individual with a study loan.
One of the implications of this study is that study debt will probably rise. A problem which may have serious implications not only for the individual student, but eventually affecting the entire household. Since study loan debts are rising a larger part of these households will be involved and have to bear this financial burden. Although, a higher individual income, economic growth or inflation may stabilize or even improve this situation.
References
Ackert, L. F. and Deaves, R., 2010. Behavioral finance psychology, decision-making and markets. South-Wester Cengage Learning, Mason.
Adkins, L.C. and Carter Hill, R., 2011. Using STATA for principles of econometrics. Fourth edition. John Wiley & Sons Inc, New York.
Akers, B. and Chingos, M.M., 2014. Is a student loan crisis on the horizon? Brookings Institution, Washington D.C.
Ando, A. and Modigliani, F., 1963. The “life cycle” hypothesis of savings: aggregate implications and tests. The American Economic Review 53, 55-84.
Avery, C. and Turner, S., 2012. Student loans: Do college students borrow too Much—Or not enough? Journal of Economic Perspectives 26(1), 165–192.
Bricker, J., Dettling, L., et al., 2013. Changes in U.S. family finances from 2010 to 2013: Evidence from the survey of consumer finances. Federal Reserve Bulletin 100(4), 1-41. Bricker, J. and Thompson, J., 2016. Does education loan debt influence household financial distress? An assessment using the 2007-2009 SCF panel. Contemporary Economic Policy 34, 660-677.
Brooks, C., 2014. Introductory econometrics for finance. Third edition. Cambridge University Press, Cambridge.
Bucciol, A. and Veronesi, M., 2014. Teaching children to save: what is the best strategy for lifetime savings? Journal of Economic Psychology 45, 1-17.
Campbell, J.Y., 2006. Household Finance. Journal of Finance 61(4), 1553-1604. College Board., 2016. Trends in college pricing, 2016. College Board, New York. College Board., 2016. Trends in student aid, 2016. College Board, New York. College Board., 1998. Trends in student aid, 2016. College Board, New York. Centraal Plan Bureau., 2014. Gemiddelde aflossing en inkomenseffecten sociaal
leenstelsel. http://www.cpb.nl/publicatie/gemiddelde-aflossing-en-inkomenseffecten-sociaal-leenstelsel
Davies, E. and Lea, S., 1995. Student attitudes to student debt. Journal of Economic Psychology 16(4), 633-679.
Gicheva, D., 2016. Student loans or marriage? A look at the highly educated. Economics of Education Review 53, 207-216.
Gicheva, D. and Thompson, J., 2015. The effects of student loans on long-term household financial stability. Student Loans and the Dynamics of Debt, 287-316.
Hausman, J.A., 1978. Specification tests in econometrics. Econometrica 46(6), 1251-1271.
Lusardi, A. and Mitchell, O.S., 2014. The economic importance of financial literacy: theory and evidence. Journal of Economic Literature 51(1), 5-44.
Lusardi, A., Mitchell, O.S. and Curto, V., 2010. Financial literacy among the young. The journal of consumer affairs 44(2), 358-380.
Lusardi, A., Tufano, P., 2015. Debt literacy, financial experiences, and overindebtedness. Journal of Pension Economics &finance 14(4), 332-368.
Mishory, J. O’Sullivan, R., 2012. Denied? The impact of student debt on the ability to buy a house. Young Invincible, Washington.
Modigliani, F. and Brumber, R.H., 1954. Utility analysis and the consumption functions: an interpretation of cross-section data. Post-Keynesian Economics
Nyhus, E. K. and Webley, P., 2001. The role of personality in household saving and borrowing behavior. European Journal of Personality 15, 85-103.
Oosterbeek, H. and Van der Broek, A., 2009. An empirical analysis of borrowing behavior of higher education students in the Netherlands. Economics of Education Review 28, 170-177.
Shefrin, H.M. and Thaler, R.H., 1988. The behavioural life-cycle hypothesis. Economic Inquiry 26, 609-643.
Thaler, R.H., 1999. Mental accounting matters. Journal of Behavioral Decision Making 12(3), 183-206.
Appendix A
Table A.1
Instrumental variable regression of the influence of study loan debt on saving behavior using all respondents between 29 and 70 years old. Robust t-statistics (in parentheses) are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Dependent Variables Saving Saving to normal income Spending exceeded income Willingness to
save Don't save
Table A.2
Instrumental variable regression of the influence of study loan debt on borrowing behavior using all respondents between 29 and 70 years old. Robust t-statistics (in parentheses) are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Table A.3
Instrumental variable regression of the influence of study loan debt on saving behavior using all respondents with college degree between 29 and 70 years old. Robust t-statistics (in parentheses) are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars.
*Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Dependent Variables Saving Saving to normal income Spending exceeded income Willingness to
save Don't save
Table A.4
Instrumental variable regression of the influence of study loan debt on saving behavior using all respondents with college degree between 29 and 70 years old. Robust t-statistics (in parentheses) are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars.
*Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Table A.5
Instrumental variable regression of the influence of study loan debt on saving and willingness to save using all respondents with college degree between 29 and 70 years old. In this table is controlled for respondents who paid nothing off from their study loan. Furthermore, dependent variables are presented for all respondents and respondents with a college degree. Robust t-statistics (in parentheses) are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. Appendix A.5 and A.6 shows full table with the control variables. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Dependent Variables Saving
Saving Respondents with college degree Willingness to save Willingness to save respondents with college degree Study loan -1339.45*** -1458.62*** -4910.84*** -2819.14*** (-4.81) (-3.79) (-6.67) (-2.91)
Nothing paid off from
Table A.6
Instrumental variable regression of the influence of study loan debt on borrowings using all respondents with college degree between 29 and 70 years old. In this table is controlled for respondents who paid nothing off from their study loan. Furthermore, dependent variables are presented for all respondents and respondents with a college degree. Robust t-statistics (in parentheses) are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. Appendix A.5 and A.6 shows full table with the control variables. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Dependent Variables Borrowings
Borrowings Respondents with college degree Study loan -4634.103*** -3013.902*** (-4.05) (-3.79)
Nothing paid off from
Table A.7
Instrumental variable regression of the influence of study loan debt on savings for respondents (between 29-70 years old) with earning higher than $20,000 annually. A distinction is made between all respondents, all respondents controlled for nothing paid off from their study loan, respondents with college degree and respondents with college degree and controlled for nothing paid off from their study loan. Robust t-statistics (in parentheses) are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Dependent Variables Saving all
respondents Saving Nothing paid back from study loan Saving with college degree Saving With college degree and
nothing paid back from study loan
Study loan -592.55*** -784.68*** -478.09** -672.43**
(-3.48) (-3.64) (-2.01) (-2.28)
Nothing paid off from study
Table A.8
Instrumental variable regression of the influence of study loan debt on borrowings for respondents (between 29-70 years old) with earning higher than $20,000 annually. A distinction is made between all respondents, all respondents controlled for nothing paid off from their study loan, respondents with college degree and respondents with college degree and controlled for nothing paid off from their study loan. Robust t-statistics (in parentheses) are used, due to heteroscedasticity in the residual. Post estimations are performed in order to test whether the variable study loan is indeed endogenous (Durbin score and Wu-Hausman test) and whether the instrument is not weak (first-stage regression, Cragg-Donald F-test). In every instrumental variable regression study loan can be treated as endogenous variable. Furthermore, the null hypothesis that the instrument is weak, could be rejected in every performed instrumental variable regression at p<0.01. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Dependent Variables Borrowing all
respondents Borrowing Nothing paid back from study loan Borrowing with college degree Borrowing With college degree and nothing paid back from
study loan
Study loan 663.03 183.14 2029.21*** 1718.75**
(1.46) (0.36) (2.85) (2.24)
Nothing paid off from
Table A.9
Ordinary least square method to estimate broad correlation between study loan debt and savings and borrowings. Robust t-statistics in parentheses. In addition, dummy variables for SCF year are included in the model. Study loan in 1,000 of dollars. *Significant at p<0.1, ** significant at p<0.05 and *** significant at p<0.01.
Appendix B
This section contains all questions from the SCF questionnaire used to obtain the variables for study loan, saving behavior, borrowing behavior and control variables. The bold letters is the specific variable name, with behind the original question asked to the respondents. Source for all questions: Codebook (2013) and SCF combined extract data at
http://sda.berkeley.edu/data/scfcomb2013/Doc/hcbk.htm.
Study loan questions
• Study loan: “Total value of educational loans held by household, 2013 dollars”. • “Year study loan was taken out”.
• “Do you owe any money or have any loans for educational expenses? ‘Yes’ or ‘No’”.
Saving behavior questions
• Saving: “Total value of savings accounts held by household, in 2013 dollars”. • Spending exceeded income: “Spent more/same/less than income in past year”. • Willingness to save: “About how much do you think you (and your family) need to
have in savings for emergencies and other unexpected things that may come up? Amount in dollars”.
• Don’t save: “Don’t save. ‘Checked’ or ‘Not checked’”.
Borrowing behavior questions
• Borrowings: “Total value of debt held by household, 2013 dollars. Including, mortgage, HELOCs, other lines of credit, debt for other residential properties, credit card debt, installment loans and other debt”.
• Willingness to borrow: “In general, do you think it is a good idea or a bad idea for people to buy things by borrowing or on credit? If r says they do not need to borrow for anything, say: What do you think in general? ‘Good idea’, ‘Good idea in some ways, bad in others’ or ‘Bad idea’”.
Control variable questions
• College degree: “Did you get a college degree? ‘Yes’ or ‘No’”.
• Bachelor’s degree, master degree and doctorate: “What is the highest degree you earned? ‘Associate’s’, ‘Bachelor’s’, ‘MA/MS’, ‘PhD’, ‘MD’, ‘Law, JD’, ‘MBA’, ‘nursing degree’, ‘other certificate’ or ‘other doctorate’”.
