Student Loan Defaults in Texas: Yesterday, Today, and Tomorrow

Predicting which Borrowers are Most Likely to Default

Introduction

Students who obtain loans for post-secondary education do so because they have a need. Typically, they have few alternative financial resources with which to pay for their education. Loans provide a means of obtaining skills and knowledge that will enable students to earn higher salaries, which in turn will allow them to repay their loans. However, the transition from school to work is often a confusing, difficult time for borrowers, especially for those who have many life stresses (e.g. divorce, loss of job, large medical expenses, etc.). If TG can identify risky borrowers early, then preemptive prevention activities may be more successful. That is the purpose of this section: to develop an econometric model to predict which borrowers are most likely to default.

Finding the reasons for student loan defaults is no simple task. For example, it is generally known that borrowers attending four-year schools default infrequently. Many of these borrowers accumulate large levels of debt over their academic careers. Based on this limited information, one might conclude that high levels of debt are associated with low rates of default. However, statistically accounting for other factors, the opposite conclusion is correct, i.e. high levels of debt are associated with high default rates.

The objective of this section is to discuss how a logistic regression model can form a statistical picture of past borrowers using information in TG’s data files. The model is then applied to students who currently enter repayment. The model assigns the likelihood or probability of a borrower going into default.

Back to Top

Literature Review

Student loan defaults have been with us since the inception of the guaranteed student loan program in the 1960s. Despite this, only a small number of studies in the past 10 years have used multivariate techniques to examine characteristics of student defaults.

Researchers who study student loan defaulters frequently identify two categories of characteristics contributing to student loan defaults: institutional and individual. Though institutional characteristics are often viewed as less important than individual characteristics when assessing why student loan default occurs, the body of research indicates that some institutional characteristics contribute to student loan default rates. Of these, type of institution receives the most attention in the literature. In 1997, the General Accounting Office (GAO) conducted two separate studies looking at student loan defaults in 1) historically black colleges6 and 2) proprietary schools.7 Furthermore, in Wilms, Moore and Bolus’ research, the only institutional characteristic showing a positive relationship to student loan default is if a student attended a proprietary school type.8

Volkwein and Szelest’s national database analysis (NPSAS, IPEDS, and the College Board Survey) show associations between institutional characteristics, other than school type, and default rates.9 The most important associations are 1) the higher the degree offered, the lower the default rate; 2) the smaller the school, the higher the default rate; 3) the higher the rate of the school’s admission acceptances the greater the default rate; 4) in general, the less wealth of an institution, the higher the default rate; and 5) the larger the student to faculty ratio, the greater the default rate.10 In the big picture, a school’s default rate is probably influenced by a number of institutional traits rather than one particular trait.

Studies have also found that student characteristics are related to default. Wilms, Moore, and Bolus conclude that student characteristics have a greater influence on default rates than do institutional characteristics.11 A more recent study conducted by Volkwein and Szelest strongly supports these findings.12 Perhaps one of the most studied and widely accepted student characteristics which predict an individual’s default is whether or not the student graduates. Greene found a strong negative relationship between students who graduate and default.13 Thus, students who default tend to be those students who withdraw prior to graduation. In a later study, Knapp and Seaks examine two-year and four-year private schools and also find that graduation reduces default.14

Another student characteristic often associated with loan default is the race or ethnic origin of the borrower. Depending upon the study, the methodology, and sample size of the study, race/ethnicity may or may not be associated with student loan defaults. For example, Wilms, Moore and Bolus examined student characteristics at community colleges and proprietary schools using a discriminate analysis model.15 In this study, the second most important factor out of six factors in predicting student loan default was race, specifically African American. In contrast, two years later Greene’s study, which uses a Tobit regression model, found race, specifically whether a borrower was African-American, to be statistically insignificant in identifying student characteristics of defaulters.16 Additionally, Volkwein et al. found that racial/ethnic minority groups default no more than non-minority groups when grouped into categories based on degree earned, marital status, and presence of dependent children.17

Earned credit hours, or grade level, is another student characteristic influencing the predictability of default.18 Gray’s logistic regression model indicated that the number of credit hours a student earns while in college is one of six factors predicting repayment behavior. Essentially, the more hours a student earns (a proxy for grade level), the less likely that a student loan default will occur.

Another student characteristic identified with student loan default rates is the amount of financial support, typically from parents, that a student receives. Some studies, like Volkwein and Szelest and Knapp and Seaks, have found a negative relationship between financial support from parents and/or family and student loan default. The less financial support, the greater the likelihood of default.19 Additionally, Wilms, Moore, and Bolus have found the greater the average annual family income “the more likely the student borrower will repay.20

All of the institutional and individual characteristics discussed are used in the upcoming model presented in this paper. These characteristics repeatedly have shown to be some of the most important factors that either forecast, associate, or correlate with student loan defaults.

Back to Top

Model Development

The objective of model development is to identify likely future defaulters. Using TG’s past data files, a model was developed around the historical relationships between borrower characteristics and the incidence of default. The resulting model can then be applied to borrowers who are currently entering repayment in order to predict likely defaulters who should be the target of preemptive default prevention efforts.

The model develops statistical estimates based upon a cohort of borrowers who entered repayment during one 18-month period during FY 1990-91 (See figure 1). The model tests the relationships between student and institutional characteristics and whether borrowers default within 6 1/2 years of entering repayment. An October 1991 database provided all of the borrower information, with the exception of enrollment status, which was derived from a later database.

Why was the 6 1/2 year duration for the default window chosen? A choice of FY 1990-91 uses a time period that is relevant to current default behavior. In addition, allowing a minimum of 6 years for default provides time for borrowers to enter and progress well into their repayment period. (A typical repayment period is 10 years.) A further discussion of model development choices occurs in Appendix B.

Literature on defaults and experience at TG suggest possible student characteristics, shown in Table 10, that might be related to defaults. However, some of these characteristics were not built into the model due to the lack of data.

Table 10

The factors described in Table 10 were operationalized using TG and Census Bureau data. The model describes default behavior for borrowers of Stafford, SLS, and PLUS loans and excludes consolidation loans, since the main TG guarantee database for FY 1991 did not include them. Table 11 displays the variables used in the model.

Table 11

Figure 1 represents the time line for the model. The model tracks all borrowers who entered repayment between April 1, 1990, and September 30, 1991. The model identifies a borrower as defaulted if TG had paid a default claim by May 1998.

Figure 1

Since the TG data files do not contain the race/ethnicity and income-levels of borrowers, the 1990 census provided proxies for this information. Because these variables represent approximations, the model will not be able to absolutely discern the true relationship to default of race/ethnicity and income level. Despite this imprecision, relevant relationships were found.

Through modeling we can test for the relationships between financing an education and defaulting. For example, financial aid packages are typically associated with a student’s need. If need-based aid dominates these packages, it might make sense that students with real need would correlate with more defaults because they have fewer financial resources and a relatively higher sticker price with which to contend. Another factor, educational cost, proxies the quality of obtained marketable skills (at least according to human capital theory21). If the cost of education is sometimes a measure for quality, then students attending relatively higher priced schools might default less.

The model also examines two factors — deferment and forbearance — that delay or push defaults beyond the approximate 6 1/2 year window between the repayment date and May 1998. Whether or not borrowers had forbearance could not be examined because data were not attainable.

Back to Top

Prediction Model Results

Table 12 presents the results of the model. The model correctly predicted whether or not borrowers default about 76 percent of the time. The model correctly predicted 73 percent of borrowers who default and 79 percent of borrowers who did not default.

One result taken from this model is its ability to assign each borrower a probability. For example, a borrower who is assigned a ‘0.75’ will have a seventy-five percent chance of defaulting. The model might be used to assign a probability of default to a borrower entering repayment in FY 1998. How well this model predicts whether or not a borrower will default depends on the similarity of the default patterns between the past cohort from which this model is based and the cohort in FY 1998. However, since the patterns of borrower characteristics demonstrate considerable stability from year to year, we can conclude that these cohorts are similar and that the model will predict the incidence of default with accuracy.

Table 12 presents a way of comparing the effectiveness of predictors of default. Standardized coefficients allow researchers to compare the effects of different variables on a common scale. The numbers presented in this table represent the magnitude and direction of the influence of each of these factors on the probability of default. The factors with the highest standardized coefficients are — whether a borrower withdraws from school and the grade level at which a borrower last took a loan. Graduation is also a strong predictor of increased default, in comparison to borrowers who later reenroll or pursue other educational goals. It is expected that borrowers who graduate would have a higher probability of default than borrowers who return to school since the latter group will not default if they receive deferments while they are in school.

Back to Top

Borrower Characteristics

Logistic regression demonstrates the effect of each characteristic on default. Table 13 displays the change in probability that each of these characteristics has on defaults.

Table 13

Back to Top

Withdrawing, Grade Level, and Graduation

As Table 13 shows, borrowers who withdraw are much more likely to default than other borrowers, other factors held constant. Borrowers who withdraw from school are almost twice as likely to default in comparison to borrowers who graduate from school. Therefore, on average, borrowers will default less if they can be kept from withdrawing during their first two years of study regardless of whether they attend a two-year, four-year, or proprietary school.

The grade level at which a borrower last took a loan is also an important indicator of default. In particular, each one-unit increase in grade level reduces the default probability by eight percentage points. Borrowers who are unable to obtain their education goals — especially through their first year of school — are highly susceptible to default.

Additionally, borrowers who graduate are more likely to default than borrowers who return to school. In most cases, enrollment statuses of graduation, withdrawal, and less than half-time force borrowers into repayment and therefore expose them to the possibility of default. In contrast, those who return to school might obtain in-school deferments that delay or prevent default. In any case, perseverance in school is a substantial factor for preventing defaults whether or not students eventually graduate.

Back to Top

Institutional Traits

Borrowers who attend schools with past high default rates are more likely to default in the future as well. For each five percentage point increase in a school’s 1988 ED cohort default rate, the borrower’s probability of default increases by two percentage points. This result presents a mixed picture of the role of cohort rates in default prediction. This finding suggests that a past cohort rate is not entirely indicative of a future probability of default. Still, to the extent that the cohort rate indicates quality of education, this finding suggests the importance of borrowers knowing the cohort default rate of the institution before they enroll.

The type of school a borrower chooses is also crucial. Borrowers who attend proprietary schools have a default probability more than 24 percentage points higher than borrowers attending four-year colleges. However, since the time of the FY 1991 cohort — seven years ago — we know that many proprietary schools with high default rates have been closed. The remaining proprietary schools have lower default rates and proprietary schools are probably not as risky an option for students as they were in 1991.

The amount of debt a borrower accumulates while at a proprietary, two-year, and four-year school affects the likelihood of default. Each $1,000 of student debt raises the chance for default at two-year and proprietary schools by about one percentage point and at four-year schools by well over one percentage point, all other factors held constant. This finding does not mean that borrowers should attend low-cost schools as a strategy for reducing debt loads in order to decrease the probability of default. When schools are categorized into five cost groups — $0-$4,000, $4,000-$7,000, $7,000-$10,000, $10,000-$14,000, and over $14,000 — attendance at a more expensive school (over $10,000) actually lowers the probability for default by nine percentage points compared to the least expensive institution (under $4,000). Since school type is controlled, this suggests that a student’s investment in a ‘pricey’ education brings some returns, one of which is a lower risk of default.

These two findings suggest that students might be receiving a better education at more expensive schools and that relatively high levels of borrowing might be justified for attendance at schools where the quality of education more than offsets the default risk associated with high borrower indebtedness. By the same token, high levels of debt represent a dangerous risk to borrowers who attend low quality schools, even if those institutions have low attendance costs. Therefore, at low cost schools, borrowers obviously still need to pursue strategies, such as working while in school or avoiding non-essential classes, that minimize debt and ease the school-to-work transition. Increased grant aid would also reduce the need to borrow and lower default rates.

Back to Top

Financial Aid Packages

Borrowers with high amounts of Estimated Financial Assistance (EFA) default at a higher rate than borrowers with relatively small assistance packages. This result coincides with the GAO report that found schools with a higher reliance on Title IV money were associated with higher student loan default rates23 and with Greene who found grant and scholarship aid associated with higher student loan default rates.24 In our model, the variable used is the financial aid determination made by financial aid offices, not necessarily the actual aid received by the student. Borrowers offered packages (work study, grants, scholarships, and other need-based loans) of more than $3,000 have default rates that are 17 percentage points higher than borrowers with packages below $1,000. Even though it is difficult to know the true relationship between the size of the financial aid package and default, this result suggests that higher amounts of financial aid packages mirror borrower need, since needy borrowers lack resources to pay for school and to later pay off their loans. However, further investigation is necessary to find the underlying reasons of the relationship between financial aid packages and default.

Back to Top

Race/Ethnicity and Income of Borrower's Community

In contrast to some previous research on defaults, this study shows the connection between race/ethnicity and defaults to be relatively minor. The model establishes that for every 10 percent increase of the Hispanic or Afro-American share in the borrower’s zip code area, default probability increases by about three percentage points. However, it is important to note that since we are using a proxy for race/ethnicity (the race/ethnic make-up of the borrower’s Zip code), the lack of precision might result in an under or over estimate of the true relationship between race/ethnicity and default.

Back to Top

Deferments

Our research provides strong evidence that the use of deferments is an effective way that borrowers can postpone a default claim, but these delaying strategies do not ultimately avoid default. For example, analysis of a 1995 cohort model with a 2 1/2 year default time frame found that borrowers with deferred loans defaulted one-third as often as borrowers without deferred loans. Thus, as expected, borrowers who obtain deferments are quite effective in reducing defaults soon after their repayment date. However, the 1991 cohort model with a 6 1/2 year timeframe indicates that borrowers with deferred loans defaulted almost as often as borrowers who never deferred. Thus, deferments almost disappeared as a factor when the timeframe for default increased.

Back to Top

Multiple Lenders

Borrowers in the 1995 cohort who had loans with more than one originating lender had about a 20 percent increase in the odds of default compared to borrowers with one lender. However, the association between defaulting and multiple lenders is diminished over a 6 1/2 year time frame in the 1991 cohort. Generally, the default problems associated with a borrower having more than one lender diminished after a long period (about 6 years) of repayment. Probably, borrowers’ use of consolidation alleviates some problems with multiple lenders.

Back to Top

How We Can Use This Information to Identify the Most Likely Defaulters

The information from this model can be used to rank the probability of default for borrowers in the TG data files. In order to identify borrowers with a consolidation loan, a consolidation loan model is also formulated for the computation of default probability. Figure 2 shows this process.

Back to Top

Conclusions

The multivariate modeling technique used in this study provides important insights into patterns of default. Like some earlier reports on defaults, this study suggests that default behavior is more closely linked to the characteristics of students than characteristics of institutions. The most effective predictors of defaulting are whether a borrower withdraws from school, especially when the student last borrowed at the first or second year grade level. This finding agrees with past studies connecting poor academic preparation and lack of persistence of borrowers with a higher occurrence of student loan defaults.

Back to Top

Additional Research

Additional research and acquisition of data would be needed to improve the accuracy of this model. Investigating factors associated with student loan default demonstrates the complexities of this issue and the need for more empirical information. Some factors that could help the model in predicting default are:

  • Employment status
  • If employed, monthly salary
  • Major field of study in post-secondary education
  • Gender
  • Marital status
  • Dependency status

Back to Top

Telephone Surveys

To provide more texture to the description of the characteristics of defaulted borrowers and to investigate whether our predictive model had important factors that were not operationalized, TG conducted in-depth telephone interviews with 42 borrowers. This sample was not intended to be statistically representative. Instead, the purpose was to talk one-on-one with borrowers to better understand their special circumstances, to provide a fuller portrait of student loan borrowers as they face the challenges of leaving school and working to pay off their loans. While TG’s econometric model does an excellent job of predicting defaults, it is limited by the kinds of data that are available for analysis. It was hoped that through these focused interviews additional dimensions to the default issue might emerge.

Borrowers were not selected randomly. Instead, they were picked in a way that might reveal the most about the prediction model discussed above. Borrowers were chosen based on whether their experience confirmed or contradicted the predicted repayment behavior.

By segmenting the survey population into these four equal groups, it was hoped that we could learn what was different and what was common among these groups.

The first group of borrowers interviewed were those that the model predicted would have defaulted and who did, in fact, default on their student loan. This group had the highest rate of unemployment among the four groups and they tended to be angry about the quality of education they received. If they had jobs, they were usually not related to their education. The loan counseling they received was typically unclear or not understood. Only one-half of this group considered using a deferment and even fewer thought of requesting a forebearance. Not one borrower in this group voiced a good experience with their servicer. These borrowers expressed exasperation with the process and seemed resigned to having the IRS take their refund checks. Their attitude was one of hopelessness concerning their economic future. Significantly, this group reported the highest number of combined life traumas. While some borrowers in all groups expressed some life traumas like divorce, large medical experiences, job loss, new dependents, etc., the first group expressed the most combinations of life traumas from incarceration to job loss due to donating a kidney to an uncle.25

The second group interviewed consisted of borrowers who the model predicted would default, but who had not. While these borrowers typically had low incomes, they tended to be working in jobs related to their training. While this group had a mix of life traumas, they often had some other source of support such as parents or spouses willing to help. These borrowers knew about their loan obligations. For example, most considered using forebearances — the highest percentage of any of the four groups. This group was committed to repaying their loans, despite being the least satisfied with their education. Interestingly, these borrowers viewed exit counseling more favorably than other borrowers.

The major differences between the first two groups (those that the model predicted to default) seemed to be:

  • Repayers had jobs related to their training both during school and afterwards, while defaulters did not.
  • Repayers were more knowledgeable about their loan options and were committed to repaying, while defaulters were not.

Similarities between these two groups were:

  • Both attended short-term programs and were earning relatively low wages.
  • Both experienced a good deal of life traumas.

Borrowers predicted by the model to not default, but who did, comprised the third group interviewed. Borrowers in this group were satisfied with their education and usually were working in a job related to their training. Most had attended college for at least four years and three of the borrowers had graduated from medical school. This group had the lowest unemployment rate of the four groups. However, the third group had a very high incidence of life traumas including trouble finding a job. Fortunately, these borrowers also appeared to have strong networks of support to help them through their difficult times. Most of these borrowers reported good experiences with their lenders and servicers and with their loan counseling. However, some of these borrowers had defaulted on only one of many loans suggesting a lack of awareness of their loans.

The final group of borrowers interviewed consisted of those predicted to avoid default and who, in fact, had avoided defaulting. This group overwhelmingly indicated that repayment was easy. This group had no trouble finding jobs and had little exposure to job loss. These borrowers were pleased with their education and loan experience, although they felt that their exit counseling was vague or unmemorable. While this group had more credit cards than the other groups, the balances on these cards were less than what they currently owed on their student loans. The fourth group was also the most successful in avoiding life traumas.

The two groups that were predicted to not default seemed to have different levels of success in finding and keeping jobs that would pay enough to let them repay their loans. Also, those that defaulted tended to have several life traumas (especially job loss, new dependents, and large medical expenses), while repayers seemed free of traumatic experiences. They were similar in that they attended long-term programs and were currently holding down jobs related to their education. Also, both groups seemed knowledgeable about their loan options and obligations.

From the telephone interviews, we learned that repayers typically have jobs related to their education and often had related jobs while in school. The transition from school to work was very smooth for repayers. Repayers also seemed to know more about their loan obligations than did defaulters. For defaulters, fate had dealt them significant life traumas. Defaulters often had trouble finding and keeping jobs with wages sufficient to allow them to pay back their loans.

In this section, we outlined a model by which TG can predict defaults with a reasonable level of accuracy. From the telephone interviews, we learned that certain key factors related to default couldn’t be captured by the quantitative model, e.g. the incidence of life traumas or the degree to which a borrower’s job is related to his or her education. Given this limitation, the model will always fail to predict some defaults and some repayment successes.

Back to Top