Financial Benefits of Rapid Student Loan Repayment: An Analytic Framework Employing Two Decades of Data

Abstract

This study analyzes the benefits of rapid student loan repayment, defined here as borrowers who repaid their cumulative undergraduate loan debt in half the time of the expected repayment cycle (10 years). Drawing from data on borrowers in two nationally representative samples, I first explain the analytic framework employed to identify rapid loan repayers, then examine whether rapid loan repayment is associated with financial benefits in terms of salary, homeownership, and non–poverty level, identifying how rapid loan repayers differ from their non–rapid loan repayer and nonborrower counterparts. Results show salary benefits associated with rapid loan repayment and indicate that among rapid loan repayers, cumulative loan debt generally did not surpass $15,000. These findings suggest that policy-makers may consider the adoption of shorter repayment plans with clear eligibility cutoff amounts as an alternative to the more common 10-year fixed plan.

Keywords

student loan debt rapid loan repayment quasi-experimental design financial aid propensity score weighting Heckman control function

On August 9, 2016, the Federal Reserve Bank of New York (FRBNY) released its Quarterly Report on Household Debt and Credit (2016). The data included in the report came from the Center for Microeconomic Data based on credit records from Equifax and covered 13 years, ranging from the first quarter of 2003 to the second quarter of 2016. This report breaks household debt balances into six categories according to a given loan type. These categories include mortgage debt, student loans, auto loans, credit cards, home equity revolving debt, and other debt. Student loan debt surpassed all other forms of debt, except mortgages, in the first quarter of 2010. As shown in Figure 1, compared to its baseline in 2003, student loan debt has more than quintupled over this 13-year period (Kiefer 2016).

Figure 1

Student Loan Debt Growth since 2003

These ever-increasing outstanding debt amounts, which continued even during and after the most recent economic crisis (Looney and Yannelis 2015), have justified the growth of a line of research focused on the effects of this debt on health, financial capability, wealth accumulation, and transitions to adulthood such as marriage and homeownership (American Student Assistance 2013; Cho, Xu, and Kiss 2015; Elliott, Grinstein-Weiss, and Nam 2013a, 2013b, 2013c; Gicheva 2011). Two common trends found among these studies are that authors have compared (1) households (or students) with student loan debt against households (or students) with no student loan debt (American Student Assistance 2013; Elliott, Grinstein-Weiss, and Nam 2013a, 2013b, 2013c; Fry 2014; Rothstein and Rouse 2007) and (2) students with differing amounts of debt (Akers 2014; Houle and Berger 2014) to estimate the effect of outstanding debt on students’ economic well-being.

The current study builds on this literature, but rather than assessing the effect of outstanding debt or its magnitude, the analytic approach here is to measure the effect of rapid full-loan repayment on students’ economic well-being. Specifically, this study assesses whether fully repaying debt shortly after college completion is associated with any discernible financial benefit compared to the outcomes realized by students who did not repay debt in full and students who did not rely on loan debt to pay for college. The use of two comparison groups is purposeful, as it allows for testing of potential benefits associated with rapid full-loan repayment that have not been examined in the loan-debt literature.

The term rapid full-loan repayment is adopted in this study to refer to a group of participants who repaid their debt in full within a median time period of 4.5 years, less than half the repayment cycle reported by Looney and Yannelis (2015). These authors indicate that the median time observed by borrowers at the end of the 1990s and the early 2000s was 10 years (see Looney and Yannelis 2015, 35). The rapid loan repayment behavior analyzed herein took place naturally; that is, there was no program or policy change that served as an incentive for participants to repay faster than the national norm. The most popular repayment plan (standard 10-year repayment plan with fixed payments) has consistently enrolled more than 50 percent of the 21 million borrowers since 2013 (National Student Loan Data System 2016), leading to times-to-repayment like those reported by Looney and Yannelis (2015).

The study of rapid full-loan repayment is timely and relevant, considering that never before has the number of students facing debt been greater (Baum 2015) and that these students, sooner or later, will also have to make important decisions about how, and when, to repay this debt in full. The current study identifies a subset of the student-borrower population that paid off debt in half of the time of the national average, giving us new knowledge about whether the decision to repay debt quickly translates into short-term financial gains.

From a policy- and decision-making perspective, one concern is that high levels of outstanding debt in the United States could lead to delinquency and default, regardless of actual loan amount (Cunningham and Kienzl 2011; Dynarski and Kreisman 2013). Accordingly, students who completely repay their student loan debt may fare better in terms of socioeconomic and financial outcomes as they are free from “payment burdens that can prevent [them] from attaining financial independence and stability … leading to financial distress and damaged credit records and rapidly rising balances” (Dynarski and Kreisman 2013, 6). In a similar vein, but moving beyond the individual level, students who repay their loans will help to bring down the outstanding stock of debt, which could yield positive social and economic outcomes for the whole of society. The current study is also important given recent claims that there is a student loan “repayment crisis” in the United States (Dynarski and Kreisman 2013, 6).

The purpose of this study is threefold. First, I propose an analytic framework that, by highlighting the complexity involved in modeling the effects of loan repayment, offers clarity in the methodological steps used to account for systematic differences among students who fully repay their loan debt rapidly and those who do not. Second, I evaluate the potential benefits of loan repayment by comparing the outcomes of students who fully repaid loan debt rapidly with the outcomes of students who did not rely on loan debt and students who relied on student loans and did not repay them within 10 years after college enrollment. Finally, considering the scarcity of empirical evidence about the characteristics associated with rapid student loan debt repayment, I compare the baseline characteristics of study participants conditional on repayment status. This comparison improves our understanding of the best predictors of rapid loan repayment. Special emphasis is placed on examining differences in amounts borrowed by students who repaid and did not repay rapidly, as well as differences in predictor variables across all comparison groups, including the individual-, institutional-, and state-level characteristics that may affect repayment status and that are typically employed in college outcomes and institutional effects literature (see Doyle 2009; Leigh and Gill 2003; Long and Kurlaender 2009; Melguizo and Dowd 2009; Reynolds and DesJardins 2009; Stephan, Rosenbaum, and Person 2009).

Background and Conceptual Framework

An important line of loan-debt inquiry has focused on its effects on financial hardship. In this line of inquiry, researchers have examined the likelihood of owning a home or having a mortgage, salary, and sector of employment (Akers 2014; American Student Assistance 2013; Cho, Xu, and Kiss 2015; Elliott, Grinstein-Weiss, and Nam 2013a, 2013b, 2013c; Field 2006; Houle and Berger 2014; Rothstein and Rouse 2007). Akers (2014) found that there is not a strong positive relationship between student debt and financial hardship. The highest rates of financial hardship are found among households with relatively little outstanding student loan debt (Akers 2014). Although the author did not control for baseline differences, this study has important implications. Akers argues that discouraging borrowing by implementing restrictive limits on federal borrowing may have negative implications across students who need this form of aid to finance college.

Somewhat similar to Akers (2014), Rothstein and Rouse (2007) found that, compared to students attending a highly selective private college who received grants and did not borrow to meet financial need, borrowers were located in higher-salary jobs. In contrast, Elliott, Grinstein-Weiss, and Nam (2013a, 2013b, 2013c), in a series of interrelated articles, found that households with a member who finished a four-year degree while incurring student loan debt are significantly less likely to have retirement savings, own a home, and accumulate assets when compared to households with a member who graduated from college without having to incur student loan debt. These apparently contradictory findings may be due to the differences in samples used. Rothstein and Rouse estimated their models using data that are perhaps not representative of the typical student population that faces debt—students enrolled in a prestigious and highly selective college. Elliott, Grinstein-Weiss, and Nam relied on the 2007–2009 Survey of Consumer Finances longitudinal data, which seeks representativeness by randomly selecting the approximate 6,500 participating families.

Elliott, Grinstein-Weiss, and Nam (2013a, 2013b, 2013c) and Akers (2014) relied on the same source of data, yet their results do not perfectly align with one another. Elliott, Grinstein-Weiss and Nam found negative effects of debt, a sign of potential financial distress, which is the same outcome variable used by Akers. Akers found little evidence of a relationship between debt and financial distress (or hardship). One possible reason for this apparent contradiction is that Elliott, Grinstein-Weiss, and Nam compared borrowers versus nonborrowers, whereas Akers compared different debt amounts among borrowers only and estimated the effect of those amounts on financial hardship.

If one considers that, as stated by Dynarski and Kreisman (2013), “payment burdens can prevent young workers from attaining financial independence and stability … leading to financial distress and damaged credit records and rapidly rising balances, all of which increase the costs of borrowing for a home, a car, and may also lead to lost employment opportunities” (p. 6), repayment should positively impact the financial well-being of students who make the effort to repay their student loans. Nonetheless, as previously mentioned, the current state of the literature indicates that this conclusion remains an empirical question yet to be addressed, thus constituting the overarching purpose of the current study.

Guiding theoretical perspectives

Considering that the median time to full repayment observed in the analytic samples is about half of the time reported by Looney and Yannelis (2015, 35), this unique group of rapid repayers may be systematically different from their nonborrower and non–full repayer counterparts. This situation requires the use of conceptual lenses that model a multidimensional set of observed and unobserved factors that may influence participants’ “rational” decisions, which in the context of this article refer to decisions about whether to repay early. Given that these decisions are affected by each participant’s economic and social sources of support and ways of seeing and experiencing the world, this study relies on social stratification theory (Bourdieu 1986) and merges it with human capital principles (G. Becker 1962; Mincer 1958). Under social stratification theory (see Marx 1887, 2000), it is clear that access to monetary, social, and cultural resources matters in that groups of people, institutions, and organizations that are able to place themselves closer to these resources will increase their overall probabilities of success in maintaining their privilege and expanding their resources. Economists have agreed that the rates of return to investment in human capital, such as education and training, may be unequal and even skewed to favor some groups over others, often based on characteristics like gender, ethnicity, and socioeconomic status (G. Becker 1962; Mincer 1958). Drawing from both theoretical perspectives acknowledges that not everybody can afford to invest in human capital as this investment requires economic, social, and cultural capital (Bourdieu 1986).

Human capital and social stratification theory suggests that students may have access to different forms of capital and have other unobserved reasons for incurring, repaying, or not relying on debt to finance college. Ignoring these differences would likely lead to biased conclusions. The analytic techniques employed herein account for these observed and unobserved factors that in addition to driving rapid loan repayment decisions may be driving the variation in the outcomes of interest as detailed in the Analytic Techniques section, below.

At the heart of the conceptual and analytic frameworks of this study lies the identification of predictors of rapid loan repayment. These theoretically and empirically relevant variables are derived from the higher education literature that deals with nonrandom assignment issues, specifically college-choice and sector effects (Doyle 2009; González Canché 2012, 2014, 2017; Hillman 2014; Leigh and Gill 2003; Long and Kurlaender 2009; Melguizo and Dowd 2009; Stephan, Rosenbaum, and Person 2009). It is worth noting that variable selection in these studies was guided by the notion that different contexts affect the choices that students made and incorporated both human and sociological perspectives, as proposed by Perna (2006). More specifically, the predictors of rapid loan repayment utilized in this study aim to capture a multidimensional spectrum of indicators that ideally dismantle baseline differences between treated (students who rapidly repaid) and comparison participants (students who did not rapidly repay in full and those who did not borrow). These variables are presented in Table 1, and their corresponding summary statistics are included in the online appendix in Tables A1 and A2.¹

Table 1

Individual-, Institution-, County-, and State-Level Characteristics Used as Predictors of Rapid Loan Repayment

Precollege Entry Indicators^a	Factors Identified during College Enrollment^b
Student Level	Institution Level	Geographic	State Level	County Level
Gender & ethnicity	Total cum. tuition cost	City	Merit aid millions	Cost of living
SES family 12th grade	Total cum. fee charges	Suburb	Need aid millions	Median income
Public K–12 education	Total cum. grant aid	Town	Loan aid millions	Crime rate
Expect BA degree	Total cum. loan aid	Rural	Prop 18–24 college	Unemployment rate
Acad. award 10–12 grades	Total cum. work-study aid	New England	Tuition agreements	Educ. attainment
Place study at home	Tuition waiver	Mideast	WICHE^c
Importance of financial aid	Selectivity level	Great Lakes	SREB^d
Importance college selectivity	Open admission policy	Plains	Midwest student exchange
Importance placed to grades	SAT math 25 & 75	Southeast	New England
Married at admission	ACT comp 25 & 75	Southwest	No tuition agreement
Children at admission	SAT verb 25 & 75	Rocky Mountains	No. IHEs agreement
Mother college support	Changed major	Far West
Father college support	Major of degree
Relatives college support	Number of majors
Teacher/counselor college sup	IHE offered grant
AP classes	IHE offered loan
Took SAT/ACT	IHE offered work-study
Took PSAT	IHE offered waiver
Extra HS class preparation	Highest degree offered
Took private class preparation	Institutional size
Used book preparation SAT	Carnegie classification
GPA 9th to 12th grades	Time to degree
Standardized math & reading scores	Transfer (2 to 4)
Number of siblings	Transfer (4 to 2)
Worked/no working 8–12	No. of institutions attended
Never worked 8–12	Distance from home
Ever dropped 8–12	Residency status

SOURCE: Integrated Postsecondary Data System (IPEDS), U.S. Census Bureau, Bureau of Labor Statistics, Bureau of Economic Analysis, U.S. Department of Commerce, U.S. Census Bureau, Geography Division; U.S. Department of Housing and Human Development; American Community Survey (ACS); Current Population Survey (CPS).

These indicators have been used in the higher education literature when dealing with nonrandom assignment issues and were used in this study in the first stage of Propensity Score Weighting (PSW) and Heckman specifications.

These indicators were also included in the prediction of repayment status given that this study departs from the literature on high school to college transition (see González Canché 2017).

Western Interstate Commission for Higher Education.

Southern Regional Education Board.

Data and Methods

Methodological conceptualization

This study relies on the counterfactual or potential outcomes framework (Lewis 1973; Rubin 2005) to account for students’ empirical systematic differences across repayment status before making any inferential claims. More specifically, the potential outcomes framework indicates that if researchers could observe the same student in two different scenarios (for example, fully repaying loan debt rapidly and not fully repaying loan debt rapidly) with some nonzero probability of occurrence, then they would simply observe the difference in outcomes between the two scenarios to evaluate which decision rendered better results (Caliendo and Kopeinig 2008; Holland 1986; Reynolds and DesJardins 2009; Rubin 2005).

This individual causal effect (ICE) estimation is impossible to obtain (Holland 1986; Rubin 2005), and researchers are constrained to comparing the outcomes of students who repaid rapidly with the outcomes of students who did not. The issue with this approach is that, as described in the Guiding Theoretical Perspectives section, above, borrowing and repayment decisions may be influenced by multiple factors. For instance, differences in access to financial resources across borrowing behaviors may be the actual driver of financial well-being rather than rapid loan repayment status itself. The inability to observe the counterfactual justifies the use of analytic techniques (propensity score weighting and Heckman control function used herein) that deal with systematic individual differences.

Analytic framework

An important reason for the scarcity of studies on the benefits of loan repayment is the complexity of modeling this phenomenon. As mentioned above, students who repaid loan debt rapidly may be systematically different from students who did not repay in full, but also from those who did not borrow. Accordingly, analytic strategies must account for and model the differences that any observed or unobserved baseline characteristics signal before fitting the final outcome models or making inferential claims. In addition to accounting for baseline differences, the modeling framework employed here ensured that outcomes were temporarily exogenous (Wooldridge 2010) from college enrollment. Accordingly, all participants included in the analytic samples were not enrolled in college for at least two years prior to gathering outcome data. The analytic plan represented in Figures 2 and 3 was purposefully designed to address these challenges.

Figure 2

Conceptual Model

Figure 3

Analytic Samples

Figure 2 illustrates three periods created based on official college-enrollment dates contained in the National Education Longitudinal Study (NELS) and Education Longitudinal Study (ELS) datasets (the two data sources used in this study; see below). The first period is college enrollment, during which students are allowed to take out loans, usually with deferred repayment. This period spans from 1991 to 1998 in NELS and from 2002 to 2010 in ELS. This restriction allows for at least two years as a postcollege transition period (second period in Figure 2) before outcomes are measured. During this second period, students must face an adjustment period during which loan amounts are due for repayment and a group of students will begin and perhaps complete repayment of their student loans. Following the restriction applied for college enrollment, the observed postcollege transition period in both the NELS and ELS samples lasted 4.5 years, on average, before measurement of financial outcomes. This last period is referred to as “the outcomes period” and took place during the last wave of data gathering in the NELS and ELS samples.

The outcomes of interest are compared by employing models that assess the effects of repayment on students’ outcomes as comprehensively as possible. The outcomes for students who rapidly repaid their student loan debt are compared against the outcomes for (1) students who borrowed but did not repay rapidly and (2) students who did not borrow at all. This second set of methodological steps is operationalized in Figure 3.

Figure 3 builds on Figure 2 and is separated into two sections. Only the NELS and ELS participants who attended college at some point and for at least one academic year were included in the analytic samples (this was done to establish a minimum common ground in time of enrollment across participants). Figure 3 shows that not all of the students who satisfied the selection criteria necessarily requested loans. Furthermore, from the group of students who requested loans, only a subset decided or were able to repay rapidly in full the amount disbursed. This renders three mutually exclusive and exhaustive groups: (1) students who did not rely on loans to finance their education (group 1 [G1] in Figure 3), (2) students who relied on student loans and rapidly repaid this debt (G2 in Figure 3), and (3) students who relied on loans and had not fully repaid these amounts at time of measurement (G3 in Figure 3). For comparison purposes, these three groups can be classified into two different subgroups of students: debt-free and borrowers. The debt-free group comprises students in groups G1 and G2. The borrowers group contains students who belong to groups G2 and G3. G2 belongs to both comparison groups, thus creating an ideal test-case scenario for whether loan repayment has any positive differentiable effects when comparing the outcomes of repayment with those of two comparison groups (G1 and G3) that represent two ends of the spectrum of loan-debt disbursement. This strategy provides policy- and decision-makers with the strongest evidence of any potential benefit associated with rapid loan repayment in terms of nonexperimental data.

The second half of Figure 3 further stresses the complexity of measuring potential benefits associated with loan repayment. Following the counterfactual framework principles, before comparing the outcomes associated with each of the two proposed comparison groups (G1 vs. G2 and G2 vs. G3), it is necessary to identify counterfactual participants whose only observable difference is treatment status. Taking G2 versus G3 as an example—before comparing the outcomes of students who borrowed and rapidly repaid (G2) and students who borrowed but did not rapidly repay in full (G3)—we have to ensure that we are comparing students who are as similar as possible in terms of other characteristics, such as socioeconomic status and level of education attainment (S. Becker and Ichino 2002; Rosenbaum and Rubin 1983) and that the models are accounting for unobservables that may be influencing the variation of the outcomes of interest (Heckman 1979; Heckman, LaLonde, and Smith 1999). These procedures are important given that students who repaid their loans in full may have done so because they come from wealthy backgrounds, while the opposite may be true for students who failed to repay their debt in this short period of time. If this were true, any disparity in the outcomes will more likely be due to their difference in backgrounds and unobservables rather than on repayment of loan debt. This rationale also applies to G1 versus G2 comparisons.

In addition to the procedures summarized in Figures 2 and 3, a set of models is estimated for participants with four-year degrees and a different set is estimated for participants without four-year degrees. This disaggregation is important for at least two reasons. First, student loan debt amounts will most likely vary conditional on attainment of a four-year degree. Students who persisted until graduation will have longer enrollment times than, for example, students who dropped out of college during their second year. Consequently, it can be said that the longer the enrollment, the greater the resources needed to sustain this investment, and the potential greater reliance on student loans as a form of aid to finance college attendance. Second, four-year degree attainment may also affect students’ repayment behaviors and outcomes of interest. A student who did not repay and did attain a four-year degree may have different financial outcomes than a student who did not repay and was not able to finish a four-year degree. Failing to account for these differences would also represent a source of bias as the financial outcomes would be confounded by college degree attainment rather than repayment status.

Datasets

This study uses data from the NELS (1988–2000) and the ELS (2002–2012). Both of these nationally representative studies contain official loan data—gathered directly from the National Student Loan Data System (NSLDS), using participants’ social security numbers—that measured all disbursements and repayments for each loan that a student was granted over a 10-year period (August of 1990 to June of 2000 for NELS, and July 2002 to May 2013 for ELS). Given that these datasets offer access to official data on student loan debt for two national cohorts of high school graduates over time and across institutions, they offer an ideal scenario for conducting longitudinal analyses that measure loan disbursement, repayment, and potential benefits associated with early repayments. Together, the datasets offer more than two decades of evidence to assess whether there have been changes in repayment benefits across decades.

The loan debt variables analyzed in this study were restricted to undergraduate loan disbursements that took place up to December 1998 and December 2010 for the NELS and ELS samples, respectively. This restriction was feasible given that both NELS and ELS studies have a variable called “F4ELMY” and “F3A12,” respectively, that registered the “Year/month [a participant] last attended postsecondary school.” In addition, the official loan data contained in both surveys have a variable “BEGDATE” that captured the “loan period begin date.” Given the college enrollment period restriction, there were (1) no loan disbursement information after December of 1998 and 2010 available for eligible participants and (2) no participants enrolled in graduate education. Consequently, the computation of loan debt disbursed and number of loans received was naturally constrained to the undergraduate college enrollment period. Information on amounts repaid and outstanding amounts registered after December 31, 1998, and December 31, 2010, was included in the computations of loan repayment during the postcollege transition period presented in Figure 2.

Outcome variables

This study relies on three outcome variables to capture financial stability. Each of these outcomes was measured in the last wave of the NELS and ELS with a mean time after college enrollment of 4.5 years across cohorts given the restriction presented in Figure 2. The first outcome is annual income. To make the analyses and amounts comparable across decades, dollar amounts were adjusted for inflation using the Consumer Price Index Inflation Calculator provided by the Bureau of Labor Statistics (U.S. Department of Labor 2015) and are represented in 2015 dollars. The second outcome measures whether participants owned a house or had a mortgage by the last wave of data collection. Information about homeownership was provided by the National Association of Realtors, which additionally indicated that the average age of first-home purchase is the same as the average age of participants in the two surveys employed in this study, approximately 30 years old. The third outcome measures likelihood of living below the poverty threshold, operationalized as whether a given respondent had had to rely on public assistance, as defined by the U.S. Department of Health and Human Services.

Analytic techniques

The analytic techniques employed in this study were designed to minimize the risk that the variation in outcomes of interest is driven by students’ greater initial observable sources of support (by relying on Propensity Score Weighting [PSW]; Rosenbaum and Rubin 1983) and/or self-selection issues based on unobservables (by relying on Heckman control function; Heckman 1979), rather than repayment status.

Propensity Score Modeling (PSM) and the use of observables

PSM assumes that treatment assignment and selection are fundamentally based on observables (Reynolds and DesJardins 2009; Rosenbaum and Rubin 1983). These observables are conceptualized as the factors and covariates that are influential in determining participants’ probabilities of receiving treatment (i.e., having repaid debt in full rapidly). The standard procedures to obtain this probability are probit or logit estimators, which render close to identical estimates of the probability of receiving treatment. Considering that the Heckman procedure relies on the probit approach, the following lines depict this procedure in the PSM framework.

Z_{i}^{s *} = β^{s'} x_{i}^{s} + ε_{i}^{s},

where $Z_{i}^{s *}$ is a dichotomous variable with a value of 1 if student i repaid debt in full or 0 otherwise, and $x_{i}^{s}$ are the explanatory variables of repayment status. The resulting predicted probability referred to as the propensity score (e[x]) allows for the computation of the balancing score (b[x]), which ensures that the comparison between treated (Z = 1) and control (Z = 0) units are made considering similar pretreatment or baseline characteristics ( $x_{i}^{s}$ ) given b(x). Note that b(x) is a theoretical construct that is only approximated by e(x) if the researcher has access to truly influential covariates that explain treatment assignment and allow for the identification of counterfactual cases (Rosenbaum and Rubin 1983). When all observable predictors are $x_{i}^{s}$ balanced across treated (rapid in-full repayers) and control units (nonborrowers and non-in-full repayers), this approach is assumed to have statistically re-created natural treatment and control groups whose counterfactual outcomes can be compared.

Given that b(x) can take an infinite value, one method to create balance units across treatment and control statuses is to rely on matching mechanisms (Rosenbaum and Rubin 1983), where, conditional on b(x) values, the covariates xi become balanced (see S. Becker and Ichino [2002] for a survey of the most frequently used balancing mechanisms). Another use of b(x) consists of using it as a weight to create a balanced sample. The main advantage of this weighting method is that weights can be used in a similar form to survey sampling weights, thus allowing researchers to use them in different statistical approaches, including doubly robust procedures, to adjust for covariates that were not balanced or that were captured after the treatment assignment took place (Ridgeway et al. 2014). For example, if treatment is defined as two- versus four-year enrollment and the outcome is probability of four-year degree attainment, one can balance on precollege indicators to estimate the propensity to two-year enrollment, and then use college-enrollment indicators (institutional size, financial aid, major, etc.) in the outcome equation to account for indicators that may have further affected a given student’s likelihood of four-year graduation above and beyond her or his propensity to two-year enrollment. The treatment effect of interest in the current study is the ATT, or average treatment effect for the treated, which captures the effect of debt repayment and is mathematically expressed as E[Y (1) − Y (0)|Z = 1], where Y (1) is the outcome realized by repayers, Y (0) is the salary received by nonrepayers and nonborrowers, and Z = 1 is repayment status. The propensity score weights for the ATT are defined as follows: $w (x) = K \frac{f (z = 1 | x)}{f (z = 0 | x)} = K \frac{b (x)}{1 - b (x)},$ where b(x) is the propensity (or assumed balancing) score described above and K is a normalization constant, used to reduce any probability function to a probability density function with total probability of one, that will cancel out in the outcomes analysis (Ridgeway et al. 2014, 27). The most important performance tests for PSW (model specification, covariate balance, and common support) are presented in Tables A1 and A2 and Figure A1 (located in the online appendix), all of which are described in the Findings section.

Heckman two-stage selection on unobservables

PSW assumes that unobservables do not affect the probability of receiving treatment and/or outcome variation. One way to test this assumption is to utilize the Rosenbaum bounds sensitivity analysis (Keele 2010), which assesses whether the treatment effect changes as a function of an increase in the probability of receiving treatment. Another way to test this assumption is to “rely on information about the functional form of the selection and outcome processes, such as the distribution of the disturbance terms” (Toomet and Henningsen 2008, 2). The current study relies on the latter approach. Conveniently, the first stage of modeling on the unobservables is also shown in equation 1. This estimation procedure utilizes the residuals of that equation to estimate the inverse Mills ratio, which is represented with lambda (λi). Heckman describes λi as “a monotone decreasing function of the probability that an observation is selected into the sample” (1979, 156). Mathematically, Heckman shows $λ_{i} = \frac{φ (e (x))}{1 - Φ (e (x))} = \frac{φ (e (x))}{Φ (- e (x))},$ where ϕ and Φ are the probability density function and the cumulative distribution function, respectively, for a standard normal variable. Notably, e(x) is the propensity score obtained from the implementation of equation 1. From a practical standpoint, if analysts have enough information to capture probabilities of receiving treatment, then they can estimate λi using the following procedure: (1) use probit analysis for the full sample to estimate the parameters of the probability of repayment in full, (2) estimate λi using these probabilities (see Heckman 1979), and (3) use the estimated value of λi as a regressor in the final model of the outcome of interest.

The outcome equation is represented as follows:

y_{i}^{o} = β^{o}' x_{i}^{0} + E [ε^{o} | ε^{s} \geq - β^{s}' x_{i}^{s}] + η_{i} \equiv β^{o}' x_{i}^{0} + ϱ σ λ (β^{s}' x_{i}^{s}) + η_{i},

where superscripts o and s represent outcome and selection equations, respectively, and ηi is a new disturbance term. In sum, the outcome equations are similar in the PSW and Heckman specifications except that in the former, w(x) is used to balance on, whereas in the latter, λi is included as an additional predictor.

Missing data

Tables A1 and A2 in the online appendix show the variables used in propensity score estimation. A total of eighteen of the variables taken from the NELS presented problems with missing data, which resulted in the loss of several hundred cases. In the ELS sample, missing data were present in thirteen of the variables utilized. Due to the theoretical relevance of the covariates chosen, rather than dropping missing cases, methods of multiple imputation using chained equations (van Buuren and Groothuis-Oudshoorn 2011) were employed.² The PSM and Heckman models were fitted both with and without the imputed datasets.³ Due to space limitations, this study is focused on the results using the imputed datasets of the NELS and ELS. The analyses without imputations rendered similar magnitudes regarding repayment effects; however, due to sample size limitations, the standard errors associated with such coefficients were also larger in magnitude.

Findings

The samples in the NELS and ELS cohorts were comparable in size. The number of eligible participants was 3,960 for the NELS and 4,370 for the ELS.⁴ After applying the selection criteria, the numbers of NELS and ELS participants who relied on student loans during undergraduate education were 1,870 and 2,580 (47.2 percent and 59.0 percent), respectively.

Nineteen percent of NELS and 21 percent of ELS participants rapidly repaid debt in full (treated status). The number of participants who are part of the debt-free categories (G1 and G2 in Figure 3) were 2,450 and 2,340 in the NELS and ELS samples, 14.7 percent and 23.5 percent of whom rapidly repaid debt. Tables A1 and A2 in the online appendix contain the summary statistics of the covariates and factors utilized to account for baseline differences between the two comparison groups across samples (G1 vs. G2 and G2 vs. G3 shown in Figure 3). These tables also test whether the proposed set of covariates was useful in predicting the propensity to receiving treatment (rapid loan repayment) and whether the weighting procedures reduced these baseline differences when creating counterfactual scenarios across analytic samples. Each table consists of eight columns. The column “repaid” corresponds to G2 in Figure 3 and contains the raw distribution of the variables for the treated group. Since the distribution of the indicators configuring the column “repaid” is the same across comparison groups, this column is presented only once in Tables A1 and A2.

The raw distributions of the comparison groups are presented in columns “NB uwt” and “NR uwt,” where NB indicates nonborrowers, NR indicates nonrepayers, and uwt is the unweighted or raw distribution of the predictors of debt repayment. The tables show that raw baseline distributions present statistically significant differences within the comparison groups. These differences empirically validate the assumption that some form of adjustment needs to be made to compare these groups.

In accordance with the quasi-experimental designs employed, the expectation is that these significant differences disappear after weighting. To test whether these differences dissipated, the matched or weighted covariates of the comparison groups (columns “NB wt” and “NR wt”) were compared against the “repaid” column. If at least one variable is significantly different in the PSW comparisons, then a doubly robust estimation should be implemented (Emsley et al. 2008; Orsini, Bellocco, and Sjölander 2013) by including significantly different covariate(s) in the second stage of the estimation procedures for both PSW and the Heckman models. As the tables show, the weighted comparisons were not significantly different; therefore, the use of doubly robust estimations in the second stage of the models was not required (Emsley et al. 2008; Orsini, Bellocco, and Sjölander 2013).

Two graphical assessments of the balance achieved for the PSW estimations are shown in Figure A1, also located in the online appendix. Figures 4A and 4C show the absolute matched and unmatched standard mean differences of baseline covariates used in the first stage of the PSW and Heckman estimations in the NELS sample. Figures 4B and 4D show the same diagnostic information for the ELS sample. In both samples the matching mechanisms significantly reduced baseline differences between participants. Figures 4E to 4H show the common support necessary for the PSW analyses to be conducted. All subfigures contained in Figure A1(e) to A1(h) show that the models captured two important features: the location of each participant on the propensity score spectrum and the overlap between treated and control participants’ propensity scores. These characteristics empirically demonstrate the predictive power of the variables included in the model that capture counterfactual units in the outcome equations.

Optimal number and amounts of loans disbursed and repaid

Table 2 presents the summary statistics of loan disbursement and repayment in 2015 dollars across samples by comparison groups of interest. The average number of loans disbursed to the treated groups (borrowers who repaid) was 3.8 (SD = 3.1) and 3.4 (SD = 2.8) in the NELS and ELS samples, with a maximum of 15 and 17 loans, respectively. Among the group who did not repay loans, the average number increased to 5.1 (SD = 3.5) and 6.8 (SD = 4.4) in the NELS and ELS samples, reaching a maximum of 23 and 41 loans on average. The average amount in 2015 dollars disbursed among participants who rapidly repaid their loans in the NELS and ELS samples was $12,389 and $13,494 with maximum disbursements of $105,698 and $156,673, respectively. Students who had not repaid, received $17,224 and $31,775, on average, in the NELS and ELS samples. The maximum loan amount of $117,082 in the NELS sample was practically tripled in the ELS sample, reaching $309,977. Among students with outstanding debt, NELS participants repaid around $3,000 on average, while ELS participants repaid more than four times as much, reaching an average amount repaid of $12,700. The maximum amounts repaid by borrowers still in debt across samples were $85,000 (NELS) and $200,000 (ELS). These maximums are far beyond the federal maximum for undergraduate Stafford Loans, suggesting they are either from PLUS loans or private loans. The last two sections of Table 2 present the debt situation among nonrepayers disaggregated by four-year degree status.

Table 2

Summary Statistics of Loan and Income Information in 2015 Dollars

	NELS Sample			ELS Sample
	Mean	SD	Max	Mean	SD	Max
Nonborrowers (G1); NELS n = 2,116, ELS n = 1,810
Income	41,151.74	29,400.61	544,000	35,247.14	29,185.93	408,000
CredOrDegre	0.62	0.49	1	0.74	0.44	1
BaDegreeOrMore	0.44	0.497	1	0.528	0.499	1
Borrowers who repaid (G2); NELS n = 360, ELS n = 550
NumberLoans	3.832	3.087	15	3.395	2.769	17
TotDisbursed	12,388.98	11,832.74	105,697.8	13,494.32	18,055.37	156,673
Income	43,314.36	26,112.86	231,200	39,947.718	37,331.56	612,000
CredOrDegre	0.763	0.426	1	0.778	0.416	1
BaDegreeOrMore	0.607	0.489	1	0.576	0.495	1
Borrowers who did not repay (G3); NELS n = 1,540, ELS n = 2,050
NumberLoans	5.064	3.502	23	6.772	4.417	41
TotDisbursed	17,223.84	13,902.70	117,082.40	31,775.31	31,832.52	309,977
OutstBalance	14,282.09	12,861.00	110,612.88	18,822.08	20,137.17	239,182
TotalRepaid	2,941.75	6,673.47	84,679.04	12,699.24	20,478.81	199,235
PctRepaid	0.208	0.25	0.96	0.371	0.287	0.98
Income	40,086.07	31,440.21	680,000.0	32,624.65	23,740.43	306,000
CredOrDegre	0.792	0.406	1	0.823	0.382	1
BaDegreeOrMore	0.62	0.486	1	0.649	0.477	1
	Mean	Med	90%	Mean	Med	90%
Outstanding debt and 4-year degree; NELS n = 950, ELS n = 1,330
NumberLoans	5.974	5	11	7.611	7	13
TotDisbursed	21,965	19,720	36,040	38,798	28,010	80,901
OutstBalance	17,874	15,664.48	32,553	22,485	16,016.04	47,334
TotalRepaid	4,091	1,955.68	12,532	16,313	7,750	44,132
PctRepaid	0.213	0.138	0.606	0.39	0.392	0.768
Income	45,230	40,800	68,000	37,283	35,496	65,790
Outstanding debt and no 4-year degree; NELS n = 590, ELS n = 720
NumberLoans	3.583	3	7	5.219	4	10
TotDisbursed	9,500	7,140	19,308	18,767	11,693	42,764
OutstBalance	8,431	5,830.32	19,058	12,037	7,712	28,696
TotalRepaid	1,069	560.32	3,343	6,730	2,528	22,018
PctRepaid	0.199	0.088	0.6	0.336	0.294	0.81
CredOrDegre	0.453	0	1	0.494	0	1
Income	31,707.1	29,920	54,400	23,996	22,440	45,900

SOURCE: Data from NSLDS:1991–2013.

Influential predictors of repayment

The findings discussed here are taken from Tables A1 and A2 in the online appendix. After using propensity score weights, all comparison samples were balanced. This section discusses differences across treatment statuses before balancing took place to assess systematic differences across treatment and comparison groups. The models across cohorts show that socioeconomic status (SES) in high school is an important indicator of borrowing behavior. Students who did not borrow came from significantly more affluent backgrounds (as measured by SES) than students who borrowed, both those who repaid rapidly and those who did not.

The borrowers group comparisons show that students who borrowed and had not repaid came from similar SES backgrounds as students who borrowed and repaid, thus indicating that repaying is not a function of SES and that other factors influence this decision, such as total amount borrowed (e.g., about $10,000 and $25,000 greater across four-year degree holders in the NELS and ELS samples, respectively). Another example of these differences indicates that students who borrowed (including those who did and did not repay) placed a greater importance on the availability of financial aid in their college-going decisions, whereas students who did not borrow were significantly less concerned about such availability.

Overall, the most common sector of first postsecondary enrollment across samples and comparison groups was four-year public institutions (around 45 percent) located in the state where students graduated from high school (roughly 80 percent of this subset). Across samples, more than 90 percent of the students enrolled full time during their first enrollment in college. The second and third most prevalent sectors were four-year private not-for-profit and community college institutions (around 25 percent in each sector across samples). In both the NELS and ELS samples, students who did not borrow were more likely to have started in the community college sector (37 percent in the NELS and 36 percent in the ELS samples). With respect to sector of last attendance, the most popular choice was once again the public four-year sector with more than 40 percent of the enrollment.

Among students who repaid, an average of 4.05 (NELS) and 5.14 (ELS) years had passed since last college enrollment. These numbers were 4.33 and 3.96 years on average for nonborrowers and nonrepayers in the NELS sample. The corresponding average numbers of years in the ELS sample were 5.02 and 4.99, respectively. Having established that all participants had been out of college for a similar number of years, another relevant indicator for repayment was degree or credential attainment. Overall, students who repaid and students with outstanding debt were equally as likely to have attained a four-year degree or more in both NELS and ELS samples (about 60 percent). Students who did not request a loan were, in both samples, less likely to have attained a four-year degree, a finding that again justifies estimating separate models for bachelor’s degree holders and non–bachelor’s degree holders.

The institutional-level variables indicate that repayers and nonborrowers had similar estimated disbursements in tuition and fees and attended college⁵ in states with similar proportions of college-age state residents enrolled in any form of college. Nonrepayers consistently had the lowest amounts disbursed to cover tuition and fees across cohorts and groups and attended institutions located in states with less access to college.

Quasi-experimental estimates

Tables 3 through 5 contain the quasi-experimental estimates for the outcome variables of interest. Although Tables A1 and A2 reflect systematic differences in participants’ baseline indicators, thus justifying the need to rely on quasi-experimental techniques, each set of models contains a “naïve regression” specification that does not account for potential endogeneity for comparison purposes. Following the analytic framework presented in Figures 2 and 3, the three financial outcomes tested correspond to the three broad categories contained in this section: (1) repayment effects on income variations (Table 3), (2) repayment effects on probabilities of owning a home (Table 4), and (3) repayment effects on probabilities of receiving public assistance (Table 5). Each table (3 through 5) contains four sets of models. The first two sets of models correspond to four-year degree holders. The first model compares repayers’ and nonrepayers’ outcomes (G2 vs. G3 in Figure 3), while the second compares repayers’ and nonborrowers’ outcomes (G2 vs. G1 in Figure 3). The second two sets of models correspond to non-four-year degree holders, with the same comparisons, G2 vs. G3 and G2 vs. G1, respectively.

Table 3

NELS and ELS Participants’ Salary Comparison

	Naive NELS	Naive ELS	PSW NELS	PSW ELS	Heckman NELS	Heckman ELS
Four-year degree holders by repayers and nonrepayers
B0	48,441.70***	57,919.64***	46,935.60***	58,899.94***	43,934.61***	59,988.94***
	–927.32	–1,599.46	–1,339.13	–2,479.52	–6,232.35	–9,154.41
Treatment	1,945.6	6,189.59	3,451.71**	5,209.29***	3,276.21*	5,697.29**
	–2,206.1	–3,239.31	–1,376.72	–1,174.95	–1,559.97	–2,385.77
Lambda					6,133.23	–2,969.15
					–8,386.52	–12,933
Num. obs.	1,130	1,260	1,130	1,260	1,130	1,260
Four-year degree holders by repayers and nonborrowers
B0	45,680.38***	50,423.05***	46,283.33***	55,677.65***	71,078.02***	101,233.99***
	–1,182.44	–1,114.65	–1,439.27	–2,002.85	–10,495.7	–9,052.85
Treatment	4,706.92	13,686.18***	4,103.97*	8,431.57**	4,226.15*	7,946.27**
	–2,814.34	–2,577.61	–1,875.1	–2,668.84	–2,117.02	–2,747.79
Lambda					–32,106.38*	–72,521.03***
					–14,826.6	–12,824.4
Num. obs.	1,160	2,340	1,160	2,340	1,160	2,340
Non-four-year degree holders by repayers and nonrepayers
B0	36,147.18***	36,263.75***	36,419.99***	38,869.63***	32,189.18***	34,169.12***
	–958.83	–1,164.01	–1,039.44	–1,717.51	–9,187.19	–7,034.34
Treatment	1,995.65	4,678.73	2,268.46	2,072.85	1,085.55	5,281.9
	–3,157.41	–2,546.42	–1,373.88	–2,182.89	–3,793.55	–3,237.35
Lambda					5,208.93	2,907.22
					–12,024.7	–9,628.53
Num. obs.	1,160	2,340	1,160	2,340	1,160	2,340
Non-four-year degree holders by repayers and nonborrowers
B0	33,906.96***	32,127.50***	32,641.23***	34,560.91***	25,027.60***	45,627.46***
	–1,007.79	–1,110.38	–1,387.67	–1,595.23	–6,834.93	–6,708.18
Treatment	244.57	8,814.97***	1,510.3	6,381.56**	1,846.92	6,083.26**
	–2,302.26	–2,277.05	–1,748.55	–2,083.36	–2,604.03	–2,438.14
Lambda					12,508.81	–19,532.50*
					–9,523.63	–9,572.35
Num. obs.	720	940	720	940	720	940

p < .05. **p < .01. ***p < .001.

Table 4

NELS and ELS Participants’ Likelihood of Owning a Home

	Naive NELS	Naive ELS	PSW NELS	PSW ELS	Heckman NELS	Heckman ELS
Four-year degree holders by repayers and nonrepayers
B0	0.28***	0.29***	0.31***	0.31***	0.44***	0.40***
	–0.02	–0.01	–0.02	–0.02	–0.1	–0.08
Treatment	0.01	–0.03	–0.02	–0.05	–0.04	–0.06
	–0.04	–0.03	–0.03	–0.03	–0.05	–0.04
Lambda					–0.22	–0.16
					–0.14	–0.12
Num. obs.	1,130	1,260	1,130	1,260	1,130	1,260
Four-year degree holders by repayers and nonborrowers
B0	0.26***	0.23***	0.30***	0.23***	0.53***	0.08
	–0.01	–0.01	–0.02	–0.02	–0.13	–0.1
Treatment	0.02	0.04	–0.01	0.03	0	0.05
	–0.04	–0.03	–0.03	–0.02	–0.04	–0.03
Lambda					–0.38*	0.21
					–0.19	–0.13
Num. obs.	1,160	2,340	1,160	2,340	1,160	2,340
Non-four-year degree holders by repayers and nonrepayers
B0	0.41***	0.29***	0.39***	0.29***	0.28	0.37***
	–0.02	–0.02	–0.02	–0.02	–0.15	–0.09
Treatment	–0.01	–0.01	0.01	–0.01	0.02	–0.04
	–0.05	–0.03	–0.03	–0.03	–0.06	–0.04
Lambda					0.17	–0.12
					–0.2	–0.13
Num. obs.	1,160	2,340	1,160	2,340	1,160	2,340
Non-four-year degree holders by repayers and nonborrowers
B0	0.27***	0.18***	0.28***	0.24***	0.57***	0.40***
	–0.02	–0.01	–0.04	–0.02	–0.15	–0.09
Treatment	0.14**	0.09**	1.30E+00	0.03	0.08	0.05
	–0.05	–0.03	–0.04	–0.03	–0.06	–0.04
Lambda					–0.43*	–0.32*
					–0.21	–0.13
Num. obs.	720	940	720	940	720	940

p < .05. **p < .01. ***p < .001.

Table 5

NELS and ELS Participants’ Likelihood of Living below the Poverty Line

	Naive NELS	Naive ELS	PSW NELS	PSW ELS	Heckman NELS	Heckman ELS
Four-year degree holders by repayers and non-repayers
B0	0.02***	0.02***	0.02***	0.01*	0	0
	0	0	–0.01	–0.01	–0.03	–0.03
Treatment	–0.01	0	–0.01*	0.01	–0.01	0.01
	–0.01	–0.01	–0.01	–0.01	–0.01	–0.01
Lambda					0.02	0.04
					–0.04	–0.04
Num. obs.	1,130	1,260	1,130	1,260	1,130	1,260
Four-year degree holders by repayers and nonborrowers
B0	0.02***	0.04***	0.02***	0.02***	0.01	–0.10*
	0	0	–0.01	–0.01	–0.04	–0.04
Treatment	–0.02	–0.01	–0.02**	0	–0.02	0
	–0.01	–0.01	–0.01	–0.01	–0.01	–0.01
Lambda					0.02	0.19***
					–0.06	–0.06
Num. obs.	1,160	2,340	1,160	2,340	1,160	2,340
Non-four-year degree holders by repayers and nonrepayers
B0	0.04***	0.17***	0.03***	0.14***	0	0.19*
	–0.01	–0.01	–0.01	–0.02	–0.06	–0.08
Treatment	–0.01	–0.02	0	0.01	0	–0.03
	–0.02	–0.03	–0.01	–0.02	–0.03	–0.04
Lambda					0.06	–0.04
					–0.08	–0.1
Num. obs.	1,160	2,340	1,160	2,340	1,160	2,340
Non-four-year degree holders by repayers and nonborrowers
B0	0.04***	0.23***	0.03*	0.20***	–0.01	0.06
	–0.01	–0.02	–0.01	–0.02	–0.07	–0.09
Treatment	–0.01	–0.09**	0	–0.05*1	0	–0.05
	–0.02	–0.03	–0.02	–0.02	–0.03	–0.04
Lambda					0.07	0.24
					–0.09	–0.13
Num. obs.	720	940	720	940	720	940

p < .05. **p < .01. ***p < .001.

Average treatment effects of repayment on annual income in 2015 dollars

In both the NELS and ELS samples, the direction of the treatment effect was positive and statistically significant. This finding indicates that, among four-year degree holders, having rapidly repaid student loan debt is positively associated with higher salary by the age of 30 across cohorts and comparison groups. The observed magnitude of the effect of having repaid debt was larger in the comparison between repayers and nonborrowers. The Heckman models in both the NELS and ELS samples comparing nonborrowers and repayers show that the coefficient associated with λ (wherein λ captures correlation between unobservables as a function of treatment and control statuses) was statistically significant and negative, signaling that unobserved factors that affect the probability of loan repayment status tend to be associated with decreases in annual salary. This finding indicates that the treatment coefficients of these Heckman models may be more precise than the PSW estimations due to the need to control for such unobserved factors. The λ in the comparisons between nonrepayers and repayers were not significant, and the treatment coefficients in the Heckman and PSW models were practically the same.

The last two sets of models shown in Table 3 correspond to the samples in which participants did not attain a four-year degree. Although the coefficient associated with debt repayment was consistently positive, significant differences were reached in only one comparison: ELS participants who did not attain a four-year degree but repaid earned about $6,100 more than nonborrowers, and the magnitude of this coefficient was consistent in both the PSW and Heckman estimations. Nonborrowers came from higher SES backgrounds than repayers, which indicates that rapid repayment is indeed associated with positive variations in salary, as discussed in the final section.

Average treatment effects of repayment on probabilities of owning a home

The results presented in Table 4 consistently indicate that debt repayment is not associated with changes in the probability of owning a home. The quasi-experimental models reflect only one instance where results reached statistical significance. The PSW estimate shows that, among non-four-year degree holders, treated participants were 13 percentage points more likely to own a house than nonborrowers in the NELS sample. Nonetheless, considering that (1) the treatment effect utilizing the Heckman estimator had an insignificant magnitude of 8 percent in this same estimate and (2) λ was statistically significant and negative (signaling that unobserved factors that affect the probability of loan repayment status tend to be associated with decreases in likelihood of home ownership), the PSW estimate may be upwardly biased and must be treated with caution.

Average treatment effects of repayment on probabilities of living below the poverty line

Results consistently indicated that repayment status was not associated with variation in likelihood of living below the poverty line. Only the PSW estimations in the ELS sample comparing repayers and nonborrowers showed marginally significant differences indicating that repayment status was associated with a 5 percentage point decrease in the probabilities of being classified as poor. In this case, the result obtained from the Heckman procedure showed the same magnitude⁶ and λ was not significant, suggesting that the PSW coefficient estimate is not affected by unobservables and can be trusted.

Limitations

Although this study employed two quasi-experimental techniques and relied on two different datasets for model specification, it has several limitations. The directionality of the association between loan repayment and salary increase assumes repayment drives the variation in salary across comparison groups. It may be possible that, as suggested by Rothstein and Rouse (2007), students with debt burden were compelled to take high-paying jobs—assuming they were able to do so—to repay debt. This scenario would suggest that higher salaries enable faster repayment rather than repayment leading to higher salaries. Even if this were the case, the benefits of loan repayment associated with salary variations would still hold as students motivated by the desire to repay were driven to take jobs associated with a salary that allowed them to repay their loans in half the expected time.

A second limitation is that the analyses restricted the samples to two years before the last wave of data collection took place. Although more years between last enrollment and outcome measurement would be optimal, data limitations impeded this approach. Nonetheless, students in the samples analyzed had already completed college on average 4.5 years before outcomes were measured, therefore alleviating this potential time-censoring constraint to a great extent. Despite these limitations, the results of this study correspond to a subset of the population that has remained unstudied, students who repaid loan debt fast, about five years fewer than the national trend of 10 years. While this repayment time does not allow one to make inferences about the entire population of national borrowers, it does enable analysis of benefits of rapid loan repayment, a contribution in and of itself.

Discussion

This study aimed to identify differences among students who rapidly repaid loan debt and their counterparts who did not take out loans and those who had not yet repaid loans accrued during undergraduate study. In both samples, students who did not request loans came from higher socioeconomic backgrounds than students who borrowed. Nonborrowers had lower expectations of attaining a four-year degree, tended to enroll in public two-year institutions more often than borrowers, and were less likely to hold a four-year degree. These differences justified the need to fit one set of models that included only four-year degree holders and a different set that included non-four-year degree holders.

In terms of financial benefits associated with loan repayment, the most straightforward positively associated outcome is annual salary, wherein participants who attained a four-year degree and repaid quickly realized about 8 percent more in annual earnings than their nonborrower and nonrepayer counterparts. Participants who did not attain a four-year degree but repaid quickly realized about 15 percent more in annual earnings than their respective counterparts (although the latter estimates did not reach statistical significance). Expanding the discussion beyond salaries, the quasi-experimental models rendered consistent results, suggesting that rapid repayment of debt in full is not associated with a greater likelihood of homeownership. The cross-decade comparison indicated that NELS participants had lower probabilities of living below the poverty line (3 percent) than ELS students (9.5 percent). This increase in the proportion of participants living below the poverty line over time helps to explain the 5 percent decrease in likelihood of living below the poverty line for rapid repayers without four-year degrees in the ELS sample. This finding suggests that repaying seems to ameliorate the likelihood of being classified as living in poverty for a subgroup of participants, even in the absence of attainment of a four-year degree.

Given the importance of starting salary in future salary increases, the question that emerges from this research is whether students should be advised to repay loan debt as soon as possible after college or use other strategies to finance loan debt. Benefits associated with loan repayment include avoiding the distress associated with facing outstanding debt (Akers 2014; Cunningham and Kienzl 2011; Dynarski and Kreisman 2013); having a credit history in good standing, which may enable rapid repayers to qualify more easily for a mortgage or car loan; and avoiding accrued interest, which in the long run leads to more asset accumulation. As such, repaying loan debt as soon as possible would seem to have advantages. This recommendation, however, only applies to borrowers with no graduate or professional degree enrollment (i.e., participants in this study) and whose average cumulative undergraduate loan debt is below $15,000, a consistent amount found among repayers in both analytic samples. Participants with higher debt accumulation (e.g., above $24,000) may be better served by using standard 10-year repayment plans. Participants who try to repay quickly loan amounts surpassing the average amounts found across datasets may be diverted from investment in other goods or services.

The lack of effects of loan repayment on homeownership may be due to the limited average time after college enrollment of participants (4.5 years across cohorts). Although these participants’ age (30 years) represents the median age of homeownership and none of the participants were officially enrolled in college, the effects of repayment on owning a home may be reflected in later years. Indeed, the financial effort required to repay debt may be the single most important factor that did not allow these participants to invest in buying a house. Other datasets with longer follow-up periods may implement the analytic model proposed in this study to address whether allowing for more time to pass after loan repayment produces models that indicate that repayment does, in fact, significantly influence homeownership prospects.

As stated at the beginning of this study, the most common repayment plan since 2013 is a 10-year repayment plan with fixed payments (National Student Loan Data System 2016). As shown in this study, a subgroup of borrowers was able to repay in half that time. This finding in and of itself indicates that in terms of “Who repays?” there is “heterogeneity” that merits future exploration regardless of whether repaying affects other outcomes.

In closing, it is worth highlighting a recent report in which Dynarski (2016) showed that only about 18 percent of students who owe more than $100,000 default, while 34 percent of those owing less than $5,000 do so. The datasets analyzed in this study consistently showed that the group of students borrowing up to $5,000 is far more common (25 percent and 20 percent among all borrowers in the NELS and ELS, respectively) than the group of borrowers who accrued more than $100,000 (none in the NELS sample and only 3 percent in the ELS study).⁷ These statistics are important as they indicate that not only are low debts more common across decades, but also that students with lower debt amounts are also more likely to default. In considering students with no four-year degree and outstanding debt, policy-makers may contemplate the insulation of this at-risk group of students from paying interest up to a given cut-off point of $5,000 (following findings reported by Dynarski 2016) as part of their debt-repayment plan—a provision absent from currently available repayment plans. This strategy would not only provide strong incentive toward loan repayment for these students, but would also prevent them from accruing interest amounts toward their current debt that eventually translate into longer time-to-repayment and increased risk of loan default. This increased risk of default is informed by the findings of this study. Table 2 shows that, although 50 percent of non-four-year degree holders attained a degree or credential, they still received the lowest salary amounts, on average, across all comparison groups. The extra provision proposed (waiving interest accrual for non-four-year degree holders with up to $5,000 in debt) equates to providing these students with more opportunities for financial upward mobility.

Although more research on rapid loan repayment is required, a common estimate across decades indicated that rapid repayers accrued below $15,000 in cumulative loan debt, on average. This finding may serve to guide future policies or programs aimed at implementing shorter repayment plans (e.g., five-year) with specific eligibility cut-off points up to $15,000. Policy-makers may also consider offering borrowers in this debt category lower fixed interest rates. These two strategies, waiving interest accumulation given low loan amounts accrued and shorter repayment plans, include well-delineated eligibility cut-off points that are absent from standard repayment plans currently available and may advance the goal of reducing the debt burden of students who may be at greater risk of defaulting.

Footnotes

Manuel S. González Canché is an assistant professor at the Institute of Higher Education, University of Georgia. He investigates the effect of undergraduate enrollment decisions on debt accumulation. From an institutional perspective, he also studies spatial dependence and competition for the attraction of nonresident students and the resulting tuition revenue from those students.

NOTE:

This research was supported by a grant from The Spencer Foundation (grant #201500116). I am grateful for the comments and feedback provided by Drs. Laura Perna, Nicholas Hillman, and Steve DesJardins. I am also thankful for grammatical help provided at different points in time by Jeffrey Harding, Robert Stollberg, Lucia Brajkovic, Jason Lee, and Melissa Whatley.

Notes

References

Akers

Beth

. 2014. How much is too much? Evidence on financial well-being and student loan debt. Washington, DC: The American Enterprise Institute (AEI).

American Student Assistance. 2013. Life delayed: The impact of student debt on the daily lives of young Americans. Washington, DC: American Student Assistance.

Baum

Sandy

. 2015. The evolution of student debt in the united states. Student loans and the dynamics of debt. In Student loans and the dynamics of debt, eds. Hershbein

Brad

Hollenbeck

Kevin M.

, 11–36. Kalamazoo, MI: W.E. Upjohn Institute.

Becker

Gary

. 1962. Investment in human capital: A theoretical analysis. Journal of Political Economy 70 (5): 9–49.

Becker

Sascha O.

Ichino

Andrea

. 2002. Estimation of average treatment effects based on propensity scores. Stata Journal 2 (4): 358–77.

Bourdieu

1986. The forms of capital. In Handbook of theory and research for the sociology of education, ed. Richardson

John G.

, 241–53. Westport, CT: Greenwood Publishing Group.

Caliendo

Marco

Kopeinig

Sabine

. 2008. Some practical guidance for the implementation of propensity score matching. Journal of Economic Surveys 22 (1): 31–72.

Cho

Soo Hyun

Yilan

Kiss

D. Elizabeth

. 2015. Understanding student loan decisions: A literature review. Family and Consumer Sciences Research Journal 43 (3): 229–43.

Cunningham

Alisa F.

Kienzl

Gregory S.

2011. Delinquency: The untold story of student loan borrowing. Washington, DC: Institute for Higher Education Policy.

10.

Doyle

William

. 2009. The effect of community college enrollment on bachelor’s degree completion. Economics of Education Review 28 (2): 199–206.

11.

Dynarski

Susan

. 2016. The trouble with student loans? Low earnings, not high debt. Brookings Institution Note. Washington, DC: Brookings Institution.

12.

Dynarski

Susan

Kreisman

Daniel

. 2013. Loans for educational opportunity: Making borrowing work for today’s students. The Hamilton Project Discussion Paper. Washington, DC: Brookings Institution.

13.

Elliott

William

Grinstein-Weiss

Michal

Nam

Ilsung

. 2013a. Does outstanding student debt reduce asset accumulation (Tech. Rep.)? CSD Working Paper 13-32, Washington University, Center for Social Development, St. Louis, MO.

14.

Elliott

William

Grinstein-Weiss

Michal

Nam

Ilsung

. 2013b. Is student debt compromising homeownership as a wealth building tool (Tech. Rep.)? CSD Working Paper 13-33, Washington University, Center for Social Development, St. Louis, MO.

15.

Elliott

William

Grinstein-Weiss

Michal

Nam

Ilsung

. 2013c. Student debt and declining retirement savings (Tech. Rep.). CSD Working Paper 13-34, Washington University, Center for Social Development, St. Louis, MO.

16.

Emsley

Richard

Lunt

Mark

Pickles

Andrew

Dunn

Graham

. 2008. Implementing double-robust estimators of causal effects. Stata Journal 8 (3): 334–53.

17.

Federal Reserve Bank of New York. 2016. Quarterly report on household debt and credit. Available from https://www.newyorkfed.org.

18.

Field

Erica

. 2006. Educational debt burden and career choice: Evidence from a financial aid experiment at NYU law school (Tech. Rep.). National Bureau of Economic Research Working Paper, Cambridge, MA.

19.

Fry

Richard

. 2014. Young adults, student debt and economic well-being. Washington, DC: Pew Research Social and Demographic Trends, Pew Charitable Trusts.

20.

Gicheva

Dora

. 2011. Does the student-loan burden weigh into the decision to start a family? Paper published online. Greensboro, NC: University of North Carolina at Greensboro.

21.

González Canché

Manuel

. 2012. Community colleges, catalysts for mobility or engines for inequality? Addressing selection bias in the estimation of their effects on educational and occupational outcomes. PhD diss., University of Arizona, Tempe, AZ.

22.

González Canché

Manuel

. 2014. Is the community college a less expensive path toward a bachelor’s degree? Public 2- and 4-year colleges’ impact on loan debt. Journal of Higher Education 85 (5): 723–59.

23.

González Canché

Manuel

. 2017. Community college scientists and salary gap: Navigating socioeconomic and academic stratification in the U.S. higher education system. Journal of Higher Education 88 (1): 1–32.

24.

Heckman

James

. 1979. Sample selection bias as a specification error. Econometrica: Journal of the Econometric Society 47 (1): 153–61.

25.

Heckman

James

LaLonde

Robert J.

Smith

Jeffrey A.

1999. The economics and econometrics of active labor market programs. In Handbook of labor economics, vol. 3, part A, eds. Ashenfelter

Orley C.

Cards

David

, 1865–2097. New York, NY: Elsevier.

26.

Hillman

Nicholas

. 2014. College on credit: A multilevel analysis of student loan default. Review of Higher Education 37 (2): 169–95.

27.

Holland

Paul

. 1986. Statistics and causal inference. Journal of the American Statistical Association 81 (396): 945–60.

28.

Houle

Jason N.

Berger

Lawrence

. 2014. Is student loan debt discouraging home buying among young adults? Washington, DC: Association for Public Policy and Management.

29.

Keele

Luke

. 2010. An overview of rbounds: An r package for Rosenbaum bounds sensitivity analysis with matched data. White Paper. Columbus, OH.

30.

Kiefer

Len

. 2016. Consumer credit trends. Available from http://lenkiefer.com/2016/08/09/trends-in-credit.

31.

Leigh

Duane E.

Gill

Andrew M.

2003. Do community colleges really divert students from earning bachelor’s degrees? Economics of Education Review 22 (1): 23–30.

32.

Lewis

David

. 1973. Causation. Journal of Philosophy 70 (17): 556–67.

33.

Little

Roderick J. A.

Rubin

Donald B.

2002. Statistical analysis with missing data. New York, NY: Wiley.

34.

Long

Bridget Terry

Kurlaender

Michal

. 2009. Do community colleges provide a viable pathway to a baccalaureate degree? Educational Evaluation and Policy Analysis 31 (1): 30–53.

35.

Looney

Adam

Yannelis

Constantine

. 2015. A crisis in student loans? How changes in the characteristics of borrowers and in the institutions they attended contributed to rising loan defaults. Brookings Papers on Economic Activity 2015 (2): 1–89.

36.

Marx

Karl

. 1887. Capital: A critical analysis of capitalist production. Vol. 2. London: S. Sonnenschein, Lowrey, & Company.

37.

Marx

Karl

. 2000. Karl Marx: Selected writings. Oxford: Oxford University Press.

38.

Melguizo

Tatiana

Dowd

Alicia

. 2009. Baccalaureate success of transfers and rising 4-year college juniors. Teachers College Record 111 (1): 55–89.

39.

Mincer

Jacob

. 1958. Investment in human capital and personal income distribution. Journal of Political Economy 66 (4): 281–302.

40.

National Student Loan Data System. 2016. Direct loan portfolio by repayment plan. Excel file. Washington, DC: National Student Loan Data System, U.S. Department of Education.

41.

Orsini

Nicola

Bellocco

Rino

Sjölander

P. C.

2013. Doubly robust estimation in generalized linear models. Stata Journal 13 (1): 185–205.

42.

Perna

Laura W.

2006. Studying college access and choice: A proposed conceptual model. In Higher education: Handbook of theory and research, ed. Smart

John C.

, 99–157. Dordrecht: Springer.

43.

Reynolds

C. Lockwood

DesJardins

Stephen L.

2009. The use of matching methods in higher education research: Answering whether attendance at a 2-year institution results in differences in educational attainment. In Higher education: Handbook of theory and research, ed. Smart

John C.

, 47–97. Dordrecht: Springer.

44.

Ridgeway

Greg

McCaffrey

Dan

Morral

Andrew

Burgette

Lane

Griffin

Beth Ann

. 2014. Toolkit for weighting and analysis of nonequivalent groups: A tutorial for the twang package. Washington, DC: RAND Corporation.

45.

Rosenbaum

Paul

Rubin

Donald

. 1983. The central role of the propensity score in observational studies for causal effects. Biometrika 70 (1): 41–55.

46.

Rothstein

Jesse

Rouse

Cecilia Elena

. 2007. Constrained after college: Student loans and early career occupational choices. National Bureau of Economic Research Working Paper No. 13117, Cambridge, MA.

47.

Royston

Patrick

. 2004. Multiple imputation of missing values. Stata Journal 4 (3): 227–41.

48.

Rubin

Donald

. 2005. Causal inference using potential outcomes. Journal of the American Statistical Association 100 (469): 322–31.

49.

Stephan

Jennifer L.

Rosenbaum

James E.

Person

Ann E.

2009. Stratification in college entry and completion. Social Science Research 38 (3): 572–93.

50.

Toomet

Ott

Henningsen

Arne

. 2008. Sample selection models in r: Package sample selection. Journal of Statistical Software 27 (7): 1–23.

51.

U.S. Department of Labor. 2015. Consumer Price Index inflation calculator. Washington, DC: Bureau of Labor Statistics, U.S. Department of Labor.

52.

van Buuren

Stef

Groothuis-Oudshoorn

Karin

. 2011. Mice: Multivariate imputation by chained equations in r. Journal of Statistical Software 45 (3): 1–67.

53.

Wooldridge

Jeffrey M.

2010. Econometric analysis of cross section and panel data. Cambridge, MA: MIT Press.