A Trillion Dollar Question: What Predicts Student Loan Delinquencies?

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. A Trillion Dollar Question: What Predicts Student Loan Delinquencies? Alvaro A. Mezza and Kamila Sommer 2015-098 Please cite this paper as: Mezza, Alvaro A., and Kamila Sommer (2015). “A Trillion Dollar Question: What Predicts Student Loan Delinquencies?,” Finance and Economics Discussion Series 2015-098. Washington: Board of Governors of the Federal Reserve System, http://dx.doi.org/10.17016/FEDS.2015.098. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

A Trillion Dollar Question: What Predicts Student Loan Delinquencies?∗ Alvaro Mezza† Kamila Sommer‡ October 2015 Abstract The recent significant increase in student loan delinquencies has generated interest in understanding the key factors predicting the non-performance of these loans. However, despite the large size of the student loan market, existing analyses have been limited by data. This paper studies predictors of student loan delinquencies using a nationally representative panel dataset that anonymously combines individual credit bureau records with Pell Grant and Federal student loan recipient information, records on college enrollment, graduation and major, and school characteristics. We showthatborrower-levelcreditcharacteristicsareimportantpredictorsofstudentloan delinquencies. In particular, credit scores of young borrowers are highly predictive of future student loan delinquencies, even when measured well before borrowers enter repayment. In marked contrast, our results point to only a limited power of student debt levels in predicting future student loan credit events. Our findings have potentially useful practical implications. For example, access to credit file information when borrowers exit school could help to more effectively target student loan borrowers who might benefit from enrolling in income-driven repayment or loan modification plans. ∗WewouldliketothanktoRodneyAndrews, MosheBuchinsky, RohitChopra, SusanDynarski, Jennifer Hunt, Sarena Goodman, Michael Palumbo, Karen Pence, Daniel Ringo, Shane Sherlund, and to the participantsoftheResearchSymposiumonStudentLoansatSuffolkUniversityLawSchoolandoftheAssociation forEducationFinanceandPolicy’s(AEFP)40thAnnualConferenceforhelpfulfeedback. AmandaN’guyen provided excellent research assistance. The analysis and conclusions contained in this paper are those of the authors and do not necessarily reflect the views of the Board of Governors of the Federal Reserve System, its members, or its staff. †Federal Reserve Board, email: Alvaro.A.Mezza@frb.gov ‡Federal Reserve Board, email: Kamila.Sommer@frb.gov

1 Introduction Over the past ten years, the real amount of student debt owed by American households more than doubled, from about $450 billion to more than $1.1 trillion, with average real debt per borrower increasing from about $19,000 to $27,000. As a result of this increase, student loan debt surpassed credit card debt as the largest class of non-housing consumer debt in 2010. A potential consequence of the higher reliance on student debt to finance higher education, coupled with the adverse effects of the Great Recession, is difficulty in meeting these debt obligations. As a possible reflection, the share of balances 90 or more days delinquent past increased from 6.7 percent to 11.7.1 Given the unprecedented rise in student loan debt and delinquencies, the trillion dollar questionis: “Whatpredictsstudentloandelinquencies?”Popularcommentaryhasfrequently linked greater odds of repayment difficulties to high student loan debt burdens but, given the scarcity of comprehensive data, such statements have been mostly based on press anecdotes.2 Using a new, unique data set, this paper studies which individual educational, credit, and schoolcharacteristicspredictstudentloandelinquencies. Ournationallyrepresentativepanel data anonymously combine individual credit bureau records with Pell Grant and Federal student loan recipient information, records on college enrollment, graduation and major, and school characteristics. The data are based on a sample of individuals who were 23 to 31 years old in 2004 and span the period 1997-2010. Using a probability model framework designed to estimate the likelihood with which a student loan borrower becomes 120 or more days delinquent on her student loans within five years of entering repayment, we show that borrower-level credit characteristics are important predictors of future student loan delinquencies. In particular, despite the notion that credit histories of young student loan borrowers are not necessarily well established and, consequently, less likely to be predictive of future credit behavior, we show that credit scores of young borrowers are highly predictive of future student loan delinquencies, even when measured prior to borrowers’ entering repayment. In marked contrast, and perhaps contrary to the popular narrative, student loan balances are generally not a significant predictor of 1Figures are based on authors’ calculations from the NYFed CPP/Equifax data set for 2005:Q2 and 2015:Q2. Nominal amounts are deflated by CPI-U into constant 2015:Q2 dollars. 2See, for example, “Graduating College with $120K in Student Loan Debt,” Here and Now, NPR, May 15, 2012; “The $555,000 Student-Loan Burden,” Wall Street Journal, February 13, 2010; “Finding Debt a Bigger Hurdle than Bar Exam”, New York Times, July 1, 2009. 1

student loan delinquency risk, both in statistical and economic terms. Our analysis offers additional interesting insights. In particular, other credit indicators remain highly predictive of future student loan delinquencies, even after controlling for borrower credit scores. For example, and perhaps counter to simple intuition, borrowers with credit card or mortgage debt prior to entering repayment are less likely to become delinquent ontheirstudentloansthanborrowerswithnosuchdebts, potentiallybecauseborrowerswith less risky underlying credit profiles are more likely to qualify for such debt while in school. Moreover, in our regression specifications that exclude credit variables, degree completion and for-profit attendance (see Looney and Yannelis (2015) for supporting evidence) are some of the strongest predictors of future student loan delinquencies, both in statistical and economic terms. However, once credit controls are included in the regression specification, their statistical and economic significance declines (though the dropout indicator remains highly statisticallysignificant),suggestingthatcreditvariables—andcreditscoresinparticular—are correlated with degree completion and attendance of for-profit institutions. These findings, coupled with the low predictive power of student loan debt, corroborate the notion that the observedriseinstudentloandelinquenciesingeneralmaynotbedrivenbylargelevelsofstudent loan debt, but rather by other factors that correlate with a borrower’s ability to repay it (see, for example, Dynarski and Kreisman (2013) and Hylands (2014)).3 Finally, similar to otherstudies(seeGrossetal.(2009)foracomprehensivereviewoftheexistingliteratureand Looney and Yannelis (2015) for a recent analysis of defaults on federal student loans), we confirm that other factors such as Pell Grant controls—which partly proxy for a borrower’s socio-economic background—are important predictors of student loan delinquencies. Our findings that credit characteristics are salient predictors of future student loan delinquencies, even when measured before borrowers enter repayment, have important practical implications. Coinciding with the rapid increase in student debt and delinquencies, a number of initiatives have been put forth to help borrowers manage their debt.4 For example, new plans tied to borrowers’ incomes (so-called “income-driven” repayment plans) were introduced to help borrowers lower monthly payments to manageable levels relative to their 3Some of these factors likely are future employment prospects and realized incomes relative to debt incurred to fund one’s post-secondary education. 4For an example of such initiatives, see the DoEd press release “U.S. Departments of Education and Treasury Announce Collaboration with Intuit Inc. to Raise Awareness about Income-Driven Repayment Options for Student Loans” from January 24, 2014 at www.ed.gov. 2

incomes.5,6 Whileincome-drivenrepaymentplansareapotentiallypromisingwaytoalleviate student loan burdens for borrowers who might otherwise be at risk of delinquency, a limited number of tools—such as high student loan balances and troubled status of the loans—has made it difficult for policy makers to efficiently target the at-risk population.7 While our results point to a limited explanatory power of the level of student loan debt in predicting future student loan delinquency risk, our results also suggest that borrowers’ credit scores could be used to identify at-risk borrowers. In particular, even the most rudimentary regression specification that includes credit scores and student loan balances is able to capture 60 percent of all student loan delinquencies among the 25 percent of borrowers most likely to become delinquent. Furthermore, the riskiest 15 and 30 percent of borrowers (as predicted by the model) account for 40 and 70 percent of all student loan delinquencies, respectively. The purpose of this study is not to devise a method that could be used to underwrite student loans at the time a borrower applies for college. Rather, we aim to identify variables that policy makers could use to effectively target at-risk borrowers for enrollment in programs designed to mitigate delinquency risk and potentially increase student loan debt manageability at the time when these borrowers exit school or enter repayment. To this end, we use a probability model to estimate the likelihood with which a student loan borrower becomes delinquent on their student loans using a set of predictive variables that are observed at or shortly before exiting school or entering repayment.8 This paper is organized as follows. Section 2 describes the newly constructed, core data 5TwosuchrecentlyintroducedplansaretheIncome-BasedRepayment(IBR)plan—availablesince2009— and the (ii) Pay-As-You-Earn (PAYE) repayment plan—available since 2012. While the two plans vary in someoftheeligibilityrequirements,theybothofferlowincome-basedpaymentstiedtodiscretionaryincome over a long amortization periods (from 20 to 25 years, depending on the specific plan). Additionally, the Income-ContingentRepayment(ICR)planhasbeenavailableforDirectLoanProgram(DLP)loanborrowers since the inception of the DLP in 1994. 6Income-driven repayment plans are intended to make student loan debt more manageable by reducing monthly payment amounts. While we are not able to measure debt manageability in our data per se, there is likely a link between borrowers’ ability to manage their student loan debt and delinquency risk. For an analysis of student debt manageability, see Thompson and Bricker (2014). 7As of 2015:Q2, about 19 percent of borrowers and 33.5 percent of outstanding federal Direct student loan balances are enrolled in income-driven repayment plans (https://studentaid.ed.gov/about/datacenter/student/portfolio). These figures include those enrolledin ICR, IBR, and PAYE plans. Interestingly, the enrollment figures indicate that those currently enrolled have higher balances, on average, than the average DLP loan borrower (about $50,000 versus $28,000), suggesting that a significant number of borrowers taking advantage of these plans are borrowers with high balances. As we will show, these are not the borrowers that are most frequently associated with delinquencies and defaults. 8Student loan borrowers do not have to start repaying their student loans right away after school exit, in general. The “waiting” period after school exit and before repayment begins is known as the grace period, and typically lasts six months. 3

set from which the final subsample used in the analysis is drawn. Section 3 discusses the sub-sampling criteria used for construction of the final estimation sample and provides basic summary statistics. Section 4 briefly describes the empirical design, interprets the regression output, and uses a cumulative delinquency curve framework to assess the predictive power of various model specifications. Section 5 concludes the paper. 2 Data Our data are pooled from several sources.9 Appendix A.1. discusses the details of the data, checks the representativeness of the merged data set against alternative data sources, and provides caveats relevant for the analysis. By way of summary, the data set starts with a nationally representative random sample of credit bureau records picked and provided by TransUnion, LLC, for a cohort of 34,891 young individuals who were between ages 23 and 31 in 2004, and spans the period 1997 through 2010. Individuals are followed biannually between June 1997 and June 2003, and then in December 2004, June 2007, and December 2008 and 2010. The data contain all major credit bureau variables, including credit scores, tradeline debt levels, and delinquency and severe derogatory records. In the next step, individual educational records through 2007 are merged from the DegreeVerify (for degrees) and Student Tracker (for enrollments) programs by the National Student Clearinghouse (NSC) on the TransUnion data. The NSC educational institution identifiers are then used to further merge school-level information from the Integrated Postsecondary Education Data System (IPEDS), as well as the historical school-level 2-year cohort default rates (CDR) from the Federal Student Aid (FSA).10,11 Finally, individual-level information about Pell Grants and Federal loans—sourced from the National Student Loan Data System (NSLDS)—is merged onto the data for Pell Grant and Federal student loan recipients. 9Allthemergesofindividual-levelinformationhavebeenperformedbyTransUnion, LLC,inconjunction with the National Student Clearinghouse, and the Department of Education. The merges have been done basedonacombinationofSocialSecuritynumber,dateofbirth,andindividuals’firstandlastnames. None of the variables used to merge individuals across sources is available in our data set. 10The IPEDS includes school-level information such as tuition, sector (e.g., public, private for-profit and not-for-profit, open admission), and SAT and ACT scores. 11Theschool-levelCDRiscomputedasthepercentageofborrowersatagivenschoolwhoenterrepayment on federal loans during a particular federal fiscal year and default on their student loan(s) prior to the end of the next fiscal year. 4

Table 1: Student Loan Repayment Shares by School Type: NSC vs FSA Comparison Public Private Not-for-profit For-profit Fiscal Year 4-year 2-year 4-year 2-year Obs. Panel A: NSC Sample 1995 62.5 25.0 12.5 0.0 0.0 8 1996 45.5 34.5 16.4 3.6 0.0 55 1997 48.8 28.4 21.6 0.6 0.6 162 1998 52.7 21.4 23.5 0.3 2.0 294 1999 54.5 17.9 22.5 0.7 4.4 457 2000 46.1 24.0 23.4 0.8 5.8 521 2001 47.8 18.2 27.3 1.0 5.7 671 2002 48.8 19.9 25.1 0.3 5.9 742 2003 47.6 20.0 23.5 0.7 8.2 892 2004 45.2 21.4 22.3 0.3 10.8 1,035 2005 46.1 19.6 22.1 0.4 11.9 1,088 2006 41.5 22.2 22.1 0.4 13.7 1,021 2007 37.3 23.9 21.6 0.5 16.7 1,270 2008 38.3 23.7 21.6 0.6 15.9 1,252 2009 43.3 14.5 25.6 0.5 16.0 1,053 Panel B: Borrowers Entering Repayment in the U.S. (from FSA)∗ 1995 42.0 14.0 27.0 1.8 15.3 1,864,691 1996 42.7 14.1 27.1 1.7 14.4 1,986,085 1997 43.3 14.0 26.8 1.5 14.4 2,115,860 1998 43.5 14.2 26.7 1.4 14.2 2,180,169 1999 44.0 14.1 26.6 1.3 14.0 2,266,807 2000 44.1 13.8 26.7 1.3 14.1 2,357,069 2001 43.2 13.2 26.7 1.2 15.7 2,359,820 2002 41.7 13.4 26.5 1.0 17.4 2,392,210 2003 40.3 14.0 25.7 1.0 19.0 2,569,024 2004 39.2 14.6 24.7 0.9 20.5 2,858,642 2005 38.8 14.0 25.9 0.8 20.5 3,528,391 2006 37.1 14.4 26.0 0.8 21.6 3,939,480 2007 36.8 15.2 22.5 0.8 24.7 3,363,857 2008 36.9 14.9 21.6 0.6 26.0 3,409,557 2009 36.3 14.0 21.9 0.7 27.0 3,709,845 Note∗: ForFederalstudentloans. Duetoalownumberofobservations,borrowersenteringrepaymentprior to fiscal year 1998 are dropped from the final sample, as specified in Section 3. 5

Out of the 34,891 individuals in the TransUnion sample, 54 percent (or 18,748 individuals) have existing NSC postsecondary education enrollment records, indicating that these individuals have at least some college education. While this “college-going” rate broadly matches the comparable rate of 58 percent in the nationally representative 2007 American Community Survey (ACS), it understates the rate slightly.12 In particular, besides never enrolling in college (a factor that accounts for the preponderance of the cases with missing NSC records), a NSC record might be missing because the postsecondary education institution in which an individual was/is enrolled does not participate in the NSC Student Tracker and DegreeVerify programs. The non-participating institutions being disproportionately more likely to be concentrated among the private for-profit schools (Dynarski et al. (2013)).13 To address the NSC coverage issues, we proceed in two steps. First, whenever available, we augment the NSC enrollment information with enrollment data from the NSLDS for Federal student loan recipients.14 Second, we develop yearly sample weights designed to correct for the underrepresentation of certain school sectors in the NSC data. Panels A and B in Table 1 show the yearly shares of borrowers entering repayment across school sectors in our NSC sample and the nationally representative FSA data, respectively.15 For each sector, theweightsareconstructedastheratiooftheseyearlysharesintheNSCsampleandtheFSA data by fiscal year.16 As seen in the panels, private not-for-profit and for-profit institutions are underrepresented throughout the NSC sample relative to the nationally representative FSA data, though their coverage increases significantly in later years of the NSC sample. The coverage issues are particularly severe prior to fiscal year 1998, leading us to eliminate individuals who entered repayment for the last time before fiscal year 1998 from the final sample. We discuss the final data selection next. 12The ACS estimate is based on a cohort of individuals between ages 25 to 34 in 2007. In our sample, individuals were between ages 26 to 34 in 2007. 13An additional reason for why an enrollment record may not be present in our data is due to data truncation related to the 2007 NSC data collection cutoff that is specific to our data set. However, this channel has no effect on comparability of the college-going rate in the NSC sample with the ACS estimate. 14The NSLDS contains information on enrollment spells that have been funded by Federal student loans. 15The FSA data are available for download at: http://www.ifap.ed.gov/DefaultManagement/press/. Unfortunately,theFSAdatadonotaccountforthoseborrowersenteringrepaymentwhoexclusivelyholdprivate studentloandebt. However,accordingtoourdata,lessthan2percentofborrowersholdonlyprivatestudent loan debt, and no Federal student loans, by the time they enter their last repayment spell observed in our sample. 16FiscalyearisdefinedastheperiodbetweenOctober1ofagivenyearandSeptember30ofthefollowing year. 6

3 Estimation data set and Summary Statistics In this section, we describe the final subsample used in the analysis. We focus on the population of individuals with positive student loan debt in the original “college-going” NSC sample: 11,766 individuals. The fraction of individuals with student debt in the NSC sample suggests that approximately 60 percent of college-goers use student loans to fund their postsecondaryeducation.17 Wenextdrop1,231individualswithzerostudentloanbalancesatthe time when they entered their last repayment observed in the sample, with the last repayment defined as starting 200 days after a borrower leaves school for the last time in our sample. The 200-day period between school completion and the beginning of the repayment spell reflects the six-month grace period.18 Next, consistent with our above discussion on sample weighting, we drop 3,808 borrowers who entered their most recently observed repayment spell either before fiscal year 1998 or after fiscal year 2006. Additionally, due to the very small sample and corresponding issues with weight adjustments, we drop 34 individuals who last attended a private not-for-profit 2-year school, as well as 6 individuals who last attendedaprivatefor-profitschoolinfiscalyear1998. Furthermore,wedrop5borrowerswho entered their last repayment spell before age 19 and additional 31 individuals with missing information on the CDR or school sector associated with the last institution attended. The final sample thus contains 6,651 student loan borrowers with existing educational records. All dollar-nominated variables are deflated into 2010 dollars using the CPI-U. Table 2 shows the age distribution when entering repayment in our final sample. As shown, about 8 percent of student loan borrowers were age 21 or younger when entering repayment for the last time in the sample. These borrowers generally represent those with completed one- or two-year degrees, dropouts, or—in rare instances—very young 4-year degree college graduates. The next 42 percent of student loan borrowers with at least some collegeenteredrepaymentforthelasttimebetweenages22and24. Theremaining50percent entered their final repayment spell between ages 25 and 33. These borrowers generally represent individuals with advanced degrees, as well as individuals who were either older 17Eventhoughnotperfectlycomparable,about60percentofindividualswhoearnedBachelor’sdegreesin 2011-12 from a non-profit institution at which they began their studies graduated with debt (CollegeBoard (2013)). This fraction has fluctuated only in a narrow range since 1999. 18In the credit bureau data, we cannot identify borrowers who entered deferment or forbearance after a schoolexit. Thispotentiallyreducestheobserveddelinquencyrateinoursample, andpotentiallybiasesour results if borrowers’ decisions to exercise these options are correlated with their characteristics. 7

Table 2: Age Distribution when Entering Last Repayment Age N Percent Cum. 19 85 1.28 1.28 20 177 2.66 3.94 21 267 4.01 7.95 22 771 11.59 19.55 23 1,100 16.54 36.08 24 938 14.10 50.19 25 787 11.83 62.02 26 712 10.71 72.73 27 567 8.53 81.25 28 457 6.87 88.12 29 308 4.63 92.75 30 233 3.50 96.26 31 150 2.26 98.51 32 84 1.26 99.77 33 15 0.23 100.00 Total 6,651 100.0 100.0 when entering college, or those with longer or interrupted educational histories. Given the focus on credit variables as predictors of student loan delinquencies, Table 3 shows the distribution of credit scores by age of a student loan borrower.19 At least in theory, and in particular for FICO scores, any individual with at least one credit tradeline account that is open for six months and that is also reported to the credit bureau agency should have a credit score.20 To the extent that the need for credit increases with age, one would expect that the fraction of individuals with credit scores will be increasing with age. In our final sample, only 30 percent of individuals have a credit score at age 18, but 70 percent have a credit score by age 19, and 93 percent by age 22. For those age 26 or above, close to 100 percent have a credit score. Moreover, the age profile of the average credit score for those with scores follows a U-shape. The average credit score starts at about 650 at age 18 but—perhaps somewhat counter to initial intuition—declines monotonically by about 80 points to 567 at age 26 before rising again to about 600 at age 31. Additionally, the variance 19The credit score used in this analysis is the TU TransRisk AM Score and it ranges from 270 to 900 points. 20Additional criteria for a FICO score apply. The borrower also has to have at least one undisputed account that has been reported to the credit bureau within the past six month, and there must be no indication that individual(s) associated with the credit record is/are deceased. For more information, see http://www.myfico.com/CreditEducation/questions/requirement-for-fico-score.aspx. Scores developed by other companies, such as the score used in this analysis, may differ slightly in the criteria used to score individuals. 8

Table 3: Credit Score Distribution by Age With No Credit Score With Credit Score Total Age Percent Percent Mean Std. Dev. N 18 72.4 27.6 645.5 101.6 2,559 19 29.5 70.5 629.6 111.7 3,301 20 17.5 82.5 612.3 133.0 3,765 21 11.0 89.0 596.5 149.4 4,457 22 6.9 93.1 583.4 157.7 4,790 23 5.0 95.0 576.8 166.8 6,004 24 3.8 96.2 569.8 170.5 5,226 25 3.4 96.6 570.7 176.9 5,279 26 2.4 97.6 566.8 180.8 5,187 27 2.5 97.5 578.8 186.4 5,985 28 1.8 98.2 585.5 185.9 5,146 29 1.6 98.4 589.7 190.3 4,506 30 1.5 98.5 592.0 189.8 3,922 31 1.5 98.5 603.1 191.0 3,128 32 2.3 97.7 599.6 193.0 2,218 33 1.9 98.1 606.8 194.6 1,965 34 1.7 98.3 602.0 196.2 1,355 35 1.6 98.4 606.9 201.5 910 in credit score increases with age. Combined, the initial decline in the average credit score and the increasing variance suggest that while young borrowers start their credit histories as a relatively homogeneous group in terms of their observable credit characteristics, over time the scores become increasingly heterogeneous and, consequently, increasingly predictive in explaining future credit events—a fact that we will exploit in our analysis. Table4providesbasicsummarystatisticsforthefinalestimationsample. Outofthe6,651 individuals in our sample, 43.5 percent left school without a degree, with the remainder of the sample earning at least a Certificate or an Associate’s degree. Additionally, 47.3 percent werelastassociatedwithapublicfour-yearinstitution. Moreover, 8.6percentlastattendeda private for-profit institution and an additional 5.1 percent attended a for-profit institution at some point but not as their last school. On average, individuals entered their last repayment spell with almost $22,000 in student loan debt (in constant 2010 dollars) on average, and 55.9percentreceivedaPellGrant. Interestingly, thedelinquencyrate—definedasaborrower beingatleast120+dayspastduewithin5yearsafterenteringthelastrepaymentspellinthe sample—derived solely from credit records is 19.4 percent in the sample. However, when the 9

credit data is augmented with information on Federal student loan defaults, this delinquency rate rises to 25.7 percent, suggesting that not all Federal student loan defaults recorded by the DoEd can be identified in the credit data.21 Moreover, among the 1,707 borrowers who became 120+ days past due on their student loans within five years after entering the last repayment spell, 1,094 (or 64 percent) ended up defaulting on their Federal student loans in those 5 years. For additional descriptive statistics for the final sample, see Appendix A.3. 4 Results 4.1 Bivariate Analysis This section summarizes several insightful bivariate relationships between student loan delinquencies and some of their correlates. Our delinquency measure takes a value of one if a borrower was ever 120 or more days delinquent within five years after entering her last repayment spell; zero otherwise.22 To start with, Table 5 and Figure 1 show the distribution of student loan balances for delinquent vs. non-delinquent student loan borrowers. As can be seen, delinquent borrowers are on average associated with lower (rather that higher) student loan balances than nondelinquent borrowers. SincetheresultsinTable5andFigure1donotaccountforborrowers’attainededucation, the negative coefficient on student loan debt should not be too surprising: those with lower educational attainment levels are more likely to become delinquent but, at the same time, tend to have lower levels of student loan debt. Table 6 illustrates this point by tabulating the average student loan debt and delinquency rate by the highest attained degree. The average delinquency rate and student loan balance among those who did not complete their degree are 43.5 percent and $12,524, compared to 6.8 percent and $48,260 of those with a Master’s degree or above. To further illustrate this point, Figure 2 repeats the analysis shown in Figure 1 but separates borrowers by the highest degree attained. Indeed, once 21Severe derogatory events associated with student loan debt, such as Federal student debt defaults, are included in our definition of 120+ dpd delinquency measure. Due to federal law, negative information has historically been removed from the credit file after seven years. 22Some borrowers have interrupted educational spells, meaning that they enter repayment but later reenroll and continue their education. Our dependent variable is thus defined with respect to the last time when a borrower enters repayment in our sample. 10

Table 4: Characteristics of Borrowers in the Final Sample Variable N Mean Std. Dev. Min Max Age at Last Repayment 6,651 25.4 2.8 19 33 Highest Degree Attained Dropout 2,892 0.435 Certificate/Associate’s Degree 373 0.056 Bachelor’s 2,108 0.317 Master’s or Above 604 0.091 Graduated but Degree Unknown 674 0.101 Last Attended Public 4-Year 3,143 0.473 Public 2-Year 1,375 0.207 Private Not-for-profit 4-Year 1,560 0.235 Private For-profit 573 0.086 Additional Sectors Attended (if Different from School Type Last Attended) Public 4-Year 1,169 0.176 Public 2-Year 1,597 0.240 Private Not-for-profit 4-Year 797 0.120 Private For-profit 340 0.051 Ever Had (Prior to Entering Repayment) Auto Loans 2,393 0.360 Mortgage Loans 836 0.126 Credit Card Loans 5,284 0.794 Ever Delinquent on (Prior to Entering Student Loan Repayment) Auto Loans 18 0.003 Mortgage Loans 11 0.002 Credit Card Loans 374 0.056 Any Debt (different than student loans) 398 0.060 Student Loans (in $1,000) Balance at Repayment 6,184 22.4 29.3 0.003 366.9 With Student Loans but Missing Balance 467 0.3 Fraction of Borrowers With Only Private Student Debt 129 0.019 Pell Grants Ever Had 3,720 0.559 Mean Pell Grants (>0, in $1,000) 3,720 2.3 1.0 0.4 4.31 Credit Score (Prior to Leaving School) Credit Score 6,056 599.3 156.4 271 866 Missing Credit Score Indicator 595 0.29 Days Lagged (w.r.t. Timing of School Exit) 6,056 376.7 197.9 1 731 Cohort Default Rate 6,168 4.90 3.90 0 100 Delinquency Rates by Repayment Calendar Year Cohort FY-1998 Cohort 285 0.272 FY-1999 Cohort 453 0.249 FY-2000 Cohort 511 0.296 FY-2001 Cohort 660 0.225 FY-2002 Cohort 734 0.253 FY-2003 Cohort 882 0.261 FY-2004 Cohort 1,031 0.265 FY-2005 Cohort 1,082 0.247 FY-2006 Cohort 1,013 0.266 Overall 1,707 0.257 Overall, based on TU information 1,293 0.194 Overall In Default on their Federal Student Loans 1,094 0.164 Note: Dollar-nominated variables are expressed in 2010 dollars using the CPI-U. 11

noitubirtsiD fo tnecreP 60. 40. 20. 0 Student Debt Distribution by Delinquency Status 0 10 20 30 40 50 Student Loan Debt at Repayment (in Thousands of 2010 $) Delinquent Current ® Figure 1: Distribution of Student Debt at Repayment, by Delinquency Status Note: Plotted for student loan balances at last repayment below $50,000. Table 5: Distribution of Student Loan Debt in the Final Sample, by Delinquency Status Not Delinquent Delinquent Student Loan Debt % Cum. % Obs % Cum. % Obs (0;10,000] 34.7 34.7 1,595 51.8 51.8 823 (10,000;20,000] 24.7 59.4 1,135 21.1 72.9 335 (20,000;30,000] 17.9 77.3 821 11.8 84.7 187 (30,000;50,000] 12.1 89.4 556 9.4 94.1 150 (50,000;75,000] 5.4 94.7 246 3.8 98.0 61 More than 75,000 5.3 100.0 243 2.0 100.0 32 Total 4,596 1,588 the highest degree attained is considered, the cumulative student loan balances associated with delinquent and non-delinquent borrowers are similar, with little discernible difference between the two groups. Table 7 summarizes student loan delinquency rates by degree completion as well as the school sector last attended. The table shows that, not controlling for other factors, delinquency rates are generally much higher for those who did not complete a degree, irrespective of the sector attended. Still, even among this population, student loan borrowers who last attended a private for-profit school or a public 2-year school are associated with relatively higher delinquency rates (54.3 and 46.4 percent, respectively) than borrowers who last at- 12

noitubirtsiD fo tnecreP 80.60.40.20. 0 No Degree 0 10 20 30 40 50 Student Loan Debt at Repayment (in $1,000) Delinquent Current noitubirtsiD fo tnecreP 80.60.40.20. 0 Certificate/Associate’s Degree 0 10 20 30 40 50 Student Loan Debt at Repayment (in $1,000) Delinquent Current noitubirtsiD fo tnecreP 80.60.40.20. 0 Bachelor’s Degree 0 10 20 30 40 50 Student Loan Debt at Repayment (in $1,000) Delinquent Current noitubirtsiD fo tnecreP 80.60.40.20. 0 Student Debt Distribution by Delinquency Status Master’s or Above 0 10 20 30 40 50 Student Loan Debt at Repayment (in $1,000) Delinquent Current ® Figure 2: Distribution of Student Debt at Repayment, by Delinquency Status Notes: Plotted for student loan balances at last repayment below $50,000. Table 6: Student Loan Balances and Delinquency Rate by Attained Degree Max. Degree Attained Avg. Student Loan Balances ($) Delinquency Rate No Degree 12,524 0.435 Certificate/Associate’s Degree 12,307 0.228 Bachelor’s Degree 24,133 0.111 Master’s or Above 48,260 0.068 Note: Tabulations reflect the highest reported attained degree in the sample. 13

Table 7: Delinquency Rate by School Type and Degree Completion With Degree With No Degree Sector Type Rate N Rate N Public 4-year 0.103 2,149 0.409 994 Public 2-year 0.166 361 0.464 1,014 Private 4-year not-for-profit 0.116 1,094 0.328 466 Private For-profit 0.265 155 0.543 418 Total 0.119 3,759 0.435 2,892 Note: Tabulationsarebasedonthemostrecentschoolsectoraffiliation. Individualsmostrecentlyaffiliated with private, 2-year institutions are dropped from the analysis due to limited number of observations. tendedapublic4-yearschool(40.9percent). Thedifferencesindelinquencyratesareperhaps even more striking among those who earned a degree: 26.5 and 16.6 percent at private forprofit and public 2-year schools vs. 10.3 percent at public 4-year schools. Finally, Table 8 shows the delinquency rate by borrowers’ credit score categories and their Pell Grant recipient status. To avoid the confounding effects of student loan repayment behavior on credit scores, a lagged credit score measure relative to school exit is used in the analysis. In particular, scores are lagged on average by one year relative to school exit (as reported in Table 4), depending on when we observe TransUnion records and when the school exit occurs for each individual in our sample. Not controlling for other factors, the table shows that the likelihood of a student loan delinquency following the school exit declines monotonically and steeply with borrowers’ pre-school-exit credit scores. In particular, the delinquency rate for student loan borrowers with credit scores between 500 and 599 is 21 percentage points higher than for borrowers with credit scores between 680 and 729. Moreover, the table shows that, Pell Grant recipients are significantly more likely to become delinquent on their student loans than those who never received Pell Grants: 34.1 percent vs. 14.9 percent, respectively. 4.2 Multivariate Analysis In this section, we estimate a probability model (probit) of student loan delinquency on the final estimation sample described in Section 3. The binary dependent variable corresponds to our delinquency measure defined in Section 4.1. Columns (1)–(7) in Table 9 display the final estimation output for seven baseline regression specifications that are based on the entire sample. Columns (1)–(6) are adjusted for the NSC school type underrepresentation 14

Table 8: Delinquency Rate by Borrower Credit Score and Pell Grant Recipient Indicators Credit Score Avg. Student Loan Balances ($) Delinquency Rate Obs 270–499 18,927 0.592 1,342 500–599 22,504 0.301 935 600–679 23,704 0.175 1,255 680–729 27,454 0.090 808 730–900 25,540 0.041 1360 Missing Score 11,372 0.341 484 Ever Pell Grants Avg. Student Loan Balances ($) Delinquency Rate Obs No 23,444 0.149 2,930 Yes 18,790 0.340 3,721 Note: Tabulations are based on borrowers’ credit scores that are on average lagged by one year relative to borrowers’ school exit. with FSA-based weights (previously discussed in Section 2), while column (7) is unweighted. Column (8) restricts the analysis to those with a Bachelor’s degree or less, also adjusting by the FSA-based weights. Coefficients in all regressions represent marginal effects, with the effectsbeingevaluatedatthemeansofthecontinuousvariables. Standarderrorsareclustered at the school level.23 Columns (6)–(8) capture our preferred regression specifications. Column (1) provides estimates for the model with explanatory variables restricted to borrower’s age when entering the last repayment spell, whether a borrower has received Pell Grants, the average amount of Pell Grants received, cumulative student loan balances when entering the last repayment spell, and yearly time effects (with the time dummy variables taking on a value of one for the fiscal year in which a borrower entered her last repayment spell). The omitted category is entering repayment in fiscal year 1998. In line with the bivariate relationship shown in Figure 1, the negative and highly statistically significant regression coefficient on the (logged) student loan debt in Column (1) illustrates that, not controlling for additional factors such as degree attainment, borrowers delinquent on their student loan debt tend to be associated with significantly lower cumulative student loan balances than non-delinquent borrowers.24 This stylized finding challenges the popular press narrative that frequently links borrowers with high student loan debt burdens (and often advanced degrees) to student loan debt repayment difficulties. While such anecdotes un- 23Our decision to cluster at the school level is motivated by the fact that the CDR varies by school. The significance of our regression estimates is essentially unchanged when standard errors are not clustered. 24We include an indicator variable that takes a value of one (zero otherwise) for 467 individuals who hold student loan debt when entering repayment but the balance amount is missing in the data. 15

doubtedly capture repayment experiences of some borrowers, the data show that they are not generally representative of the typical student loan borrower experience.25 The Federal Pell Grant program is a means-tested program that offers financial aid to low-income students. The program is large: the total Pell expenditures for the school year 2013–14 are estimated at about $34 billion, with over one-third of undergraduate students receiving a Pell Grant in that year.26 While the indicator for Pell Grant recipients is not statistically significant on its own, the average amount of Pell Grants received—itself a function of family resources for dependent students and own resources for students who are independent, as well as enrollment intensity and educational expenses—is highly significant and positively correlated with future student loan delinquency at a one percent level. While the economic significance falls when additional controls are included in the regressions in Columns (2)–(8), the statistical significance is unchanged, suggesting that borrowers from underprivileged socio-economic backgrounds are significantly more likely to become delinquent on their student loan debt. Column (2) adds information on attained degrees and college majors, with the omitted categoriesbeinganAssociate’sdegree/Certificatefordegreesandengineeringformajors.27,28 A comparison of coefficients in Columns (1) and (2) illustrates the important role of degree completionasacorrelateoffuturestudentloandelinquencies. Wheneducationalcontrolsare addedinColumn(2),thecollegedropoutindicatorbecomesthemosteconomicallysignificant predictor of future student loan delinquency.29 In contrast, having a degree beyond an Associate’s degree or a Certificate mitigates delinquency risk. Additionally, when schooling 25We are not the first to point this out; see, for example, “Student Loans and Defaults: The Facts” by Susan Dynarski, New York Times, Jun 11, 2015. For a profile overview of high-debt borrowers, see “Who Graduates College with Six-Figure Student Loan Debt?” by Mark Kantrowitz from Aug 1, 2012. 26Source: College Board (https://bigfuture.collegeboard.org/pay-for-college/grants-andscholarships/what-is-a-pell-grant and http://trends.collegeboard.org/student-aid/figures-tables/pellgrants-total-expenditures-maximum-average-grant-recipients-time). 27We group degrees into the following categories: dropouts (i.e., those with at least some college but no attained degree), Associate’s or Certificate degree holders, Bachelor’s degree holders, holders of a Master’s degree or more, and those with a completed degree for which the degree type is unknown due to NSC reporting issues. College majors are available only for those with completed degrees and we aggregate them into 15 different categories; for details, see Appendix A.2. Additionally, 138 borrowers with degree have no majorinformation. Thus,theregressionspecificationsincludeanindicatorvariablethattakesavalueofone (zero otherwise) if the individual has a major of an unknown type. 28Additionally,weconsideredanalternativespecificationwherestudentloanbalanceswereinteractedwith thedegreeindicators. Giventhelowpredictivepoweroftheseinteractionterms, weexcludedthemfromthe final analysis. 29Giventhenon-causalnatureofouranalysis,thisresultdoesnotnecessarilyimplythatpushingdropouts to finish their degrees will help them repaying their debt. 16

variables are included in the regression, the effect of the average Pell Grants received drops roughlybyathird(althoughitremainsstatisticallysignificantataonepercentlevel), largely because Pell Grant controls are positively correlated with the dropout indicator. Indeed, Pell Grant recipients are more frequently associated with socioeconomic characteristics and educational experiences that suggest statistically greater chances of dropping out of college (Wei and Horn (2002, 2009)).30 Importantly, once educational controls are added, student loan debt becomes positively correlated with student loan delinquency risk. However, the estimated partial correlation between student loan debt and delinquency is economically smallandstatisticallyinsignificant,andthisresultholdsevenasadditionalcontrolsareadded in Columns (3)–(8). This weak correlation between student loan balances and delinquency odds reflects that student loan balances—once educational backgrounds are considered—are a poor predictor of student loan delinquency risk.31 Onto majors, all the coefficients are positive, indicating that among those with an attained degree, engineering majors are less likely to become delinquent on their student debt than other majors, and several of these coefficients are significant at standard significance levels.32 Column (3) incorporates information on school sectors: public 2-year, public 4-year, private 4-year not-for-profit, and private for-profit. Individual educational histories are often complicated. While some borrowers attend only one sector, many borrowers attend multiple sectors through a series of lateral, upward or downward moves. Therefore, besides controlling for the school sector with which a borrower was most recently associated (with public 4-year schools being the omitted sector), we also control for other schools sectors previously attended.33,34 Consistent with the bivariate evidence in Table 7, having both last or ever attended a for-profit institution is positively and significantly associated with higher delin- 30Examples include delayed postsecondary enrollment, having an independent status and dependents, being a single parent, working full-time, or attending part-time. 31We further demonstrate this result in Section 4.3. 32Those with degrees in languages, literatures or visual arts; social sciences; education; legal professions; public administration; liberal arts and sciences; communications and journalism; criminal justice; and personal and culinary services are significantly more likely to become delinquent on their student debt relative to engineering majors, all else equal. 33For example, if a borrower was last associated with a private 4-year not-for-profit institution, but previously also attended a public 2-year school and a for-profit school, the total effect of her sector attendance decisions on student loan delinquency risk can be calculated by adding up the coefficients on Private 4-year not-for-profit, Ever Public 2-year, and Ever Private for-profit indicator variables in Table 9. 34Additionally,weconsideredanalternativespecificationwherestudentloanbalanceswereinteractedwith the indicators for the last school sector attended. Given the low predictive power of these interaction terms, we excluded them from the final analysis. 17

quency risk at one and five percent significance levels, respectively, even after controlling for several other factors. Similarly, attending a public 2-year institution just before entering the last repayment spell is also positively associated with future student loan delinquency risk.35 Column (4) adds information on school-level 2-year cohort default rates (CDR) for the last institution attended. The CDR is a metric constructed by the DoEd and it is mainly used to sanction schools with high student loan default rates. If the school’s CDR exceeds a given threshold, the school becomes Title IV ineligible for a period of time, thereby losing access to federal funds in the form of grants and student loans.36 To be consistent with CDR information that might be available to the DoEd at the moment when the borrower enters repayment and to avoid contaminating the CDR by the borrower’s own delinquency behavior, we lag the school-level CDR by three years with respect to the year in which the borrower enters her last repayment spell. Not surprisingly, the school CDR is highly predictive of individual future student loan delinquency. Moreover, introducing this measure takes power away from the school sector coefficients. In particular, the statistical significance of attending a public 2-year institution just before entering the last repayment spell dissipates, and the value of the coefficient drops by roughly two thirds, while the economic importance of attending a private for-profit school just before entering repayment decreases almost by a half. The reduction in the predictive power of the last school sector attended variables once the CDR is included is due to the combination of two factors. First, there is a significant correlation between school sectors and CDRs, as reflected in Figure 3. Second, even among the sectors where delinquencies are prevalent, there is significant heterogeneity in CDRs among schools within those sectors, which is why the CDR is a better predictor of future student loan delinquencies than a more generic indicator for school sector. Column (5) includes a first subset of credit variables. More specifically, the regression controls for whether the borrower had other types of debt (i.e., auto, mortgage, and credit card debt) just before entering repayment, and whether the borrower was delinquent on any debt different from student loans prior to entering repayment. Interestingly, having credit card and mortgage debt before entering repayment is associated with a lower likelihood of 35As was the case with dropouts, the positive relationship between delinquency risk and attending a forprofit institution is not necessarily causal. However, for the purpose of identifying characteristics predicting future credit risk, for-profit institution attendance is a relevant variable to consider. 36For sanctions and benefits for schools with high and low CDRs, respectively, see Section 2.4 of https://ifap.ed.gov/DefaultManagement/guide/attachments/CDRMasterFile.pdf. 18

3. 2. 1. 0 Distribution of Cohort Default Rate by Last Sector Attended 0 10 20 30 40 School Cohort Default Rate Public 4−yr Public 2−yr Private Not−for−profit 4−yr Private for−profit ® Figure 3: Distribution of School Sector CDRs future student loan delinquencies, while having auto debt is statistically insignificant. The lower incidence of student loan delinquencies among those with mortgage or credit card debt prior to entering repayment might in part reflect that borrowers with less risky underlying credit profiles might be more likely to qualify for such debts while in school. Simultaneously, the statistical insignificance of having auto debt prior to entering repayment might reflect thatautodebtisgenerallymorewidelyavailabletoborrowerswithlesspristinecreditrecords than other types of debt. Finally, being delinquent on other types of debt prior to entering repayment is highly correlated with future student loan delinquency odds. Our preferred regression specification in column (6) includes (logged) credit scores. As describedinSection4.1, creditscoresarelaggedrelativetotheschool exit. Finally, adummy variable set to unity for those with missing credit scores (zero otherwise) is included.37 Once credit variables are introduced in Columns (5) and (6), the economic effects associated with dropping out and Pell Grants received decline in value (although remain significant), suggesting that credit variables—and credit scores in particular—are correlated with degree completion and socio-economic backgrounds. At the same time, lagged credit score becomes the most economically predictive correlate of student loan delinquency odds. The highly 37Credit scores for this group are coded as zero. 19

Predicted Probability 11..0000 Prediction: No Degree 00..9900 Prediction: Bachelor’s 00..8800 00..7700 No degree and no credit score 00..6600 00..5500 Bachelor’s degree holder with no credit score 00..4400 * 00..3300 00..2200 00..1100 * 00..0000 330000 440000 550000 660000 770000 880000 990000 CCrreeddiitt SSccoorree Notes: Calculations assume that the borrower entered repayment in 2005 at the age of 25, and received a Bachelors Degree with an Engineering major. It is also assumed the borrower only studied at a Public 4-year school, received no Pell Grants, Figure 4: Probability of Student Loan Delinquency, by Credit Score had a credit score core of 700 prior to entering repayment, had no additional debt besides student loans, and the last reported cohort default rate for the school he last attended before entering repayment was 5.8%. Dashed lines represent 10% confidence intervals Notes: Calculations assume that the borrower entered repayment in fiscal year 2005 at the age of 25, only attended a public four-year school with an average (5 percent) CDR, and received a Bachelor’s Degree in Business, management, and marketing. Moreover, the borrower did not receive any Pell Grants, had a student loan debt of $22,000 prior to entering repayment and no other debt. Dashed lines represent 90 percent confidence intervals. predictive nature of lagged credit scores, coupled with the low predictive power of student loan debt, supports the notion that the observed rise in student loan delinquencies may not, in general, be driven by large levels of student loan debt, but rather by other factors that correlate with a borrower’s ability to repay it (see, for example, Dynarski and Kreisman (2013) and Hylands (2014)).38 Finally, turning to the effect of other delinquencies, not surprisingly, being delinquent on other types of debt before entering repayment becomes insignificant when introducing the credit score prior to leaving school. Although the effects of having a mortgage and credit card debt remain significant, part of the effect is absorbed by the credit score. Column (7) shows the unweighted estimates presented in Column (6), while column (8) estimates the model on a subsample of borrowers with at most a Bachelor’s degree. As can be seen, re-assuringly, the estimates and their statistical significance are largely unchanged. To illustrate the economic and statistical relevance of credit scores, Figure 4 compares 38Giventhecorrelationbetweenlaggedcreditscoreanddegreecompletion, someofthesefactorsmightbe related to borrowers’ ability to manage debt due to, for example, lack of financial literacy or income effects. 20

Credit Score: 700 Credit Score: 550 Predicted Probability Predicted Probability 0.14 0.14 0.13 0.13 0.12 0.12 0.11 0.11 0.10 0.10 0.09 0.09 0.08 0.08 0.07 0.07 0.06 0.06 0.05 0.05 0.04 0.04 0.03 0.03 0.02 0.02 0.01 0.01 0.00 0.00 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 Student Debt, in 2010 dollars, in thousands Student Debt, in 2010 dollars, in thousands Figure 5: Probability of Student Loan Delinquency, by Student Loan Balances Notes: Calculations assume that the borrower entered repayment in fiscal year 2005 at the age of 25, only attended a public four-year school with an average (5 percent) CDR, and received a Bachelor’s Degree in Business,management,andmarketing. Moreover,theborrowerdidnotreceiveanyPellGrants,hadacredit scoreof700(550)priortoenteringlastrepayment,andhadnootherdebt. Dashedlinesrepresent90percent confidence intervals. the estimated relationship between credit scores (measured prior to entering the last repayment spell) and the student loan delinquency probability for a Bachelor’s degree holder and a college dropout at a public four-year school, holding other observable characteristics constant.39 As shown, the predicted probability of becoming delinquent on student loans after entering repayment is significantly higher for those leaving school without a degree across the credit score spectrum. The differential effect ranges from about 10 percentage points for high credit score borrowers to 40 percentage points for low credit score borrowers. Turning to the effect of missing scores, notably, having no credit score prior to leaving school in our sample is not synonymous with a low score: observationally, borrowers without scores are about as likely to become delinquent on their student debt as their counterparts with a score of 575. In the same vein, Figure 5 captures the estimated ceteris paribus correlation between studentloanbalancesandthedelinquencyriskforaborrowerwitha700and550creditscore, 39For assumptions, see Notes in Figure 4. Again, this results should not be interpreted as causal. 21

respectively.40 As illustrated, and consistent with our previous discussion, the economic contribution of student loan debt plays only a relatively small role in predicting future delinquency and is only imprecisely estimated. In particular, the predicted probability of future student loan delinquency is relatively flat across the student loan debt spectrum, with sizable 90 percent confidence intervals bounding the estimated effects, indicating the limited usefulness of student loan balances in identifying borrowers in a high risk of future student loan delinquency.41 4.3 Predictive Power and Policy Implications In this section, we assess the in-sample ability of our preferred specification in Column (5) to identify borrowers in high risk of becoming delinquent on their student loans within five years after entering their last repayment spell—our dependent variable. To this end, we construct cumulative delinquency curves—an analytical metric commonly used in the mortgage industry to gauge performance of credit risk statistical models. First, we use our model to predict the probability of becoming delinquent on student loans for each borrower. Second, we calculate the cumulative delinquency curve, defined as a function L(P), where P (represented by the horizontal axis) is the cumulative portion of the population ranked by delinquency risk in a descending order, and L (represented by the vertical axis) is the cumulative portion of student loan delinquencies in the sample. The black line in Figure 6 tracks a perfect prediction for our sample and shows that roughly 25 percent of borrowers become delinquent on their student loans in our sample. Theoretically, a perfect model would assign these borrowers the highest predicted probability of student loan delinquency and would thus allot them to the bottom quartile of the population ranked by the delinquency risk in descending order, P. In practice, an estimated model is unlikely to fit the perfect prediction line exactly. However, the model’s fit relative to the perfect prediction provides a gauge for assessing how well the model separates 40To consider the maximum impact that student loan debt could have on predicting future student loan delinquency risk, unweighted results—reported in Column (7)—were used for this exercise. For additional assumptions, see Notes in Figure 5. 41The predictive power of student loans is the largest at low levels of student loan debt. For example, in the panel to the right of Figure 5, an increase in student loan debt from $1,000 to $10,000 dollars is associated with a 15 percent increase in the baseline probability of future student loan delinquency (from 6.8 to 7.9 percent). At higher student loan levels, the contribution of additional $10,000 dollars of student loan debt is much smaller. For example, an increase in student loan debt from $30,000 to $40,000 increases the baseline probability by 1.6 percent (from 8.4 to 8.5 percent). 22

borrowers in a high risk of student loan delinquency from their lower risk counterparts. Figure 6 shows that our fully specified model (represented by the red line) captures 60 percent of all student loan delinquencies among the riskiest 25 percent of student loan borrowers ranked by the model-predicted delinquency risk (relative to a 100 percent of all delinquencies under the perfect prediction). Furthermore, the bottom 15 and 30 percent of riskiest borrowers (as predicted by the model) account for 40 and 70 percent of all student loan delinquencies, respectively. While our results point to the fact that a significant amount of heterogeneity related to student loan repayment behavior exists even after controlling for the available observable characteristics, we view the ability of our model to capture an appreciable amount of student loan delinquencies in the bottom tail of the predicted risk distribution as encouraging. At this point, it is helpful to compare the predictions of our preferred model against alternative models that could be used by the DoEd to target borrowers in high risk of student loan delinquency after entering repayment. In particular, the DoEd has a number of initiatives through which it attempts to reach student loan borrowers who might benefit from enrolling in income-driven repayment plans. Unfortunately, only a limited number of tools are currently available to the DoEd to help identify these borrowers, such as high student loan balances or an existing delinquency status of these loans.42 Thus, the figure also shows the explanatory power of several (simpler) alternative models that could be potentially adopted by the DoEd to predict future delinquency risk. As illustrated, a model that uses only student loan balances—the blue line—closely tracks a 45 degree line and, as such, is associated with a minimal explanatory power. The green line shows a version of the model that complements student loan balances with additional information potentially readily available to the DoEd: the 2-year school cohort default rates and Pell Grants controls.43 Including such information greatly improves the fit of the model over the specification that exclusively relies on student loan balances; however, even such specification is far inferior in terms of its in-sample predictive power to our preferred specification. Given the statistically and economically significant effect of credit scores in our preferred specification, it is interesting to explore whether augmenting the above regressions (which 42Borrowers with loans in deferment or forbearance—other indicators that might point to a troubled status—could also be targeted. 43Asinourfinalspecification,thePellGrantcontrolsincludethebinaryindicatorforPellGrantrecipients as well as the average Pell Grants received. 23

Cumulative Percent 100 90 80 70 60 50 Full Model Student Loan Balances 40 Student Loan Balances, CDR & Pell Grants Student Loan Balances, CDR, Pell Grants & Credit Score Student Loan Balances & Credit Score 30 Perfect Prediction 20 10 0 0 10 20 30 40 50 60 70 80 90 100 Percentile Figure 6: Cumulative Delinquency Distribution, by Model-predicted Student Loan Delinquency Risk onlycontrolforriskfactorscurrentlyobservablebytheDoEd)mightappreciablyincreasethe models’ predictive power. The purple line captures the predictive power of a re-estimated model which only includes student loan balances and credit scores (measured prior to a borrower’s entering repayment). Introducing borrowers’ credit scores greatly improves the model’s performance relative to the specifications which rely exclusively on student loan balances, as well as student loan balances, CDRs and Pell Grant controls. Moreover, once credit scores are introduced, a further inclusion of CDRs and Pell Grants—in addition to student loan balances and credit scores (the orange line)—leads only to a marginal improvement in predictive power of the model and has essentially no discernible effect on the model’s ability to predict student loan delinquencies at the bottom of the risk distribution where the efficiency gains of an improvement in predictive power would be the greatest. Taken together, our results suggest that individuals’ credit scores are a potent predictor of student loan delinquency risk. Moreover, once credit scores are included, the contribution of CDRs and Pell Grant controls to the model’s explanatory power is somewhat limited. Indeed, once credit scores are included, even the most rudimentary regression specification that relies exclusively on student loan balances and credit score performs comparably to our preferred, 24

fully specified model. In particular, a simple model that includes only student loan balances and credit scores captures about 57 percent of all student loan delinquencies among the bottom quartile of the riskiest borrowers, essentially the same fraction as the fully specified model but nearly doubles the fraction of delinquencies captured by its analog that does not employ credit scores. 5 Conclusions Theincreaseinstudentloandelinquenciesanddefaultsinrecentdata, coupledwithincreased borrowing volumes, have raised concerns about students’ ability to repay their student loan debt obligations. Coinciding with these increases, new income-driven repayment plans were introduced by the DoEd to help borrowers lower monthly payments to manageable levels relative to their incomes. While enrollment in and general awareness of these programs have been on the rise, efficient targeting of borrowers who might benefit from these plans appears to have been difficult, in part due to data limitations.44 This paper contributes to the literature on predictors of student loan delinquency, with a novel application to credit bureau data. Using new and unique data, we build a series of statistical models designed to predict the probability that a borrower becomes delinquent on her student loans within the first 5 years after entering the last repayment spell. We show that, unlike student loan balances, individual credit scores are highly predictive of future student loan delinquencies, even when measured well before borrowers’ entry into repayment. In particular, even the most rudimentary probit specification that controls only for student loan balances and lagged credit scores captures about 60 percent of all student loan delinquencies in the sample among the model-predicted riskiest quartile of the student loan borrower distribution. In contrast, a specification based solely on student loan balances—a risk factor that, in absence of other viable alternatives, is readily available to the DoEd—is associated with minimal explanatory 44As of 2015:Q1, about 17.5 percent of borrowers and 31.5 percent of outstanding federal Direct student loan balances are enrolled in income-driven repayment plans (https://studentaid.ed.gov/about/datacenter/student/portfolio). These figures include those enrolledin ICR, IBR, and PAYE plans. Interestingly, the enrollment figures indicate that those currently enrolled have higher balances, on average, than the average DLP loan borrower (about $50,000 versus $28,000), suggesting that a significant number of borrowers taking advantage of these plans are borrowers with high balances. As indicated above, these are not the borrowers that are most frequently associated with delinquencies and defaults. 25

power and is, therefore, of minimal use for achieving the objective of efficient targeting. Our findings have practical implications. First, exploiting borrower credit information could vastly improve the efficacy of the borrower-targeting process. Second, our analysis used credit scores that were significantly lagged relative to the borrowers’ last repayment entryinordertoavoidconfoundingfeedbackeffectsofrepaymentbehavioronobservedcredit controls. Practically, using the most recent credit score available at the moment of targeting is likely to boost the targeting’s effectiveness relative to our results. While credit scores are generally not available to the DoEd, they could be potentially acquired by making borrowers sign a credit score disclosure form when applying for a loan. Incorporating such disclosure into the standard student loan application process could thus provide the DoEd with the ability to access borrowers’ credit score at a future time, when these borrowers exit school or enter repayment. As discussed in the Introduction, our analysis is not designed nor should be interpreted as suggesting that credit scores be used for student loan underwriting; doing so could undermine the objective of equalizing college access opportunities. 26

tuptuO noissergeR tiborP :9 elbaT )8( )7( )6( )5( )4( )3( )2( )1( elbairaV *0470.0- ***7190.0- **9980.0- 7050.0- *9650.0- *8060.0- *6260.0- ***641.0tnemyapeR tsaL ta egA )1470.0( )68200.0( )9010.0( )131.0( )8580.0( )3360.0( )7650.0( )50-e68.1( 61100.0 ***85100.0 **25100.0 709000.0 169000.0 *50100.0 *31100.0 ***26200.0 2)tnemyapeR tsaL ta egA( )241.0( )81700.0( )3320.0( )551.0( )621.0( )0090.0( )2070.0( )50-e46.5( 35400.0 40900.0 58500.0 27600.0 24700.0 68600.0 50-e88.2 ***7730.0- )000,1/tbeD tnedutS(nL )675.0( )411.0( )383.0( )513.0( )152.0( )382.0( )699.0( )90-e10.1( 8610.0 6710.0 2110.0 7910.0 4310.0 6010.0 42100.0- 0020.0stnarG lleP devieceR revE )325.0( )473.0( )906.0( )473.0( )055.0( )536.0( )759.0( )004.0( ***8930.0 ***4130.0 ***6530.0 ***9740.0 ***6950.0 ***2360.0 ***4070.0 ***2690.0 devieceR stnarG lleP .gvA )50-e30.7( )50-e23.2( )50-e66.1( )90-e16.3( )0( )0( )0( )0( ***732.0 ***102.0 ***712.0 ***332.0 ***662.0 ***872.0 ***972.0 tuoporD )441000.0( )641000.0( )931000.0( )50-e93.4( )60-e92.2( )70-e81.9( )70-e24.6( **0280.0- ***6070.0- **1070.0- ***2480.0- **5970.0- ***8380.0- ***611.0eergeD rolehcaB htiW )8030.0( )66800.0( )3130.0( )95700.0( )9310.0( )62900.0( )611000.0( **2460.0- 9060.0- ***6690.0- ***7890.0- ***701.0- ***131.0eroM ro eergeD AM htiW )4940.0( )601.0( )44600.0( )40600.0( )64200.0( )50-e33.5( **4670.0- **2380.0- ***0890.0- ***6490.0- ***0890.0- ***021.0- )nwonknU epyT( eergeD htiW )1210.0( )1030.0( )51600.0( )96800.0( )50600.0( )104000.0( 27

egap suoiverp morf deunitnoC – 9 elbaT )8( )7( )6( )5( )4( )3( )2( )1( elbairaV 921.0 8960.0 2770.0 301.0 911.0 511.0 721.0 erutcetihcrA )1( )243.0( )315.0( )194.0( )683.0( )903.0( )223.0( )782.0( ***272.0 ***122.0 ***152.0 ***442.0 ***552.0 ***752.0 ***162.0 strA ,serutaretiL ,segaugnaL )2( )549000.0( )50100.0( )805000.0( )654000.0( )391000.0( )281000.0( )341000.0( **571.0 501.0 **251.0 *821.0 *721.0 *321.0 *921.0 secneicS larutaN/ygoloiB )3( )8330.0( )801.0( )0830.0( )2270.0( )1960.0( )8670.0( )2860.0( *961.0 *921.0 *751.0 **471.0 **981.0 **591.0 **102.0 msilanruoJ/snoitacinummoC )4( )5460.0( )6860.0( )7450.0( )9820.0( )8410.0( )3210.0( )5010.0( 051.0 5540.0 501.0 1980.0 4470.0 1970.0 6290.0 smetsyS retupmoC )5( )781.0( )865.0( )682.0( )173.0( )544.0( )414.0( )143.0( **382.0 321.0 **732.0 *402.0 *902.0 *122.0 *222.0 ecitsuJ lanimirC )6( )5720.0( )722.0( )0640.0( )0480.0( )7880.0( )5570.0( )5670.0( **361.0 *001.0 **031.0 *811.0 **421.0 **621.0 **921.0 secneicS laicoS )7( )7820.0( )7580.0( )8140.0( )4650.0( )9830.0( )0830.0( )4430.0( 901.0 *711.0 **731.0 *221.0 *911.0 *421.0 *321.0 noitacudE )8( )991.0( )1360.0( )0140.0( )6260.0( )8060.0( )0250.0( )3750.0( 5330.0 05500.0- 6220.0 3120.0 9210.0 3710.0 2620.0 ssenisuB )01( )826.0( )719.0( )896.0( )217.0( )818.0( )067.0( )946.0( *651.0 4890.0 *811.0 1580.0 9670.0 5080.0 8080.0 snoisseforP htlaeH )11( 28

egap suoiverp morf deunitnoC – 9 elbaT )8( )7( )6( )5( )4( )3( )2( )1( elbairaV )4250.0( )511.0( )1280.0( )591.0( )722.0( )012.0( )802.0( *971.0 **932.0 **612.0 **622.0 **622.0 **422.0 snoisseforP lageL )21( )9550.0( )8020.0( )5630.0( )6920.0( )0030.0( )7130.0( 951.0 *431.0 **861.0 **591.0 **702.0 ***512.0 ***212.0 noitartsinimdA cilbuP )31( )731.0( )2860.0( )3920.0( )6810.0( )2010.0( )02800.0( )21800.0( **781.0 **361.0 **471.0 **571.0 ***202.0 ***212.0 **091.0 secneicS/strA larebiL )41( )5440.0( )3730.0( )8530.0( )0820.0( )09900.0( )88600.0( )0410.0( **971.0 **311.0 **151.0 **641.0 ***751.0 ***261.0 ***481.0 secivreS yraniluC/lanosreP )51( )0210.0( )2040.0( )2310.0( )1510.0( )11700.0( )29500.0( )81200.0( 9710.0- 2710.0- 7110.0- 58100.0- 1510.0 ***9350.0 raey-2 ,cilbuP dednettA tsaL )344.0( )933.0( )465.0( )829.0( )274.0( )73700.0( 00300.0- 71200.0 973000.0 33400.0- 77200.0- 2010.0- ,etavirP dednettA tsaL )788.0( )398.0( )389.0( )908.0( )878.0( )775.0( tfiorp-rof-ton 4010.0- 34800.0- 70-e77.8 6030.0 **8360.0 ***111.0 tfiorp-rof ,etavirP dednettA tsaL )027.0( )607.0( )000.1( )122.0( )4210.0( )60-e76.6( 1520.0- 6710.0- 9310.0- 4210.0- 0910.0- 3020.0raey-4 ,cilbuP dednettA revE )803.0( )282.0( )394.0( )945.0( )353.0( )123.0( 6210.0- 999000.0 5010.0- 7010.0- 9710.0- 2810.0raey-2 ,cilbuP dednettA revE )615.0( )449.0( )715.0( )905.0( )672.0( )172.0( 29

egap suoiverp morf deunitnoC – 9 elbaT )8( )7( )6( )5( )4( )3( )2( )1( elbairaV 45900.0 53200.0 797000.0 70700.0 048000.0 84200.0- ,etavirP dednettA revE )886.0( )098.0( )869.0( )617.0( )569.0( )698.0( tfiorp-rof-ton 2840.0 0720.0 9930.0 **5560.0 ***7170.0 **4760.0 ,etavirP dednettA revE )011.0( )062.0( )431.0( )0510.0( )59600.0( )1110.0( tfiorp-rof ***70700.0 ***12700.0 ***30600.0 ***04800.0 ***79900.0 )sraeY 3 yb deggaL( RDC )89100.0( )50-e23.5( )64200.0( )50-e34.4( )60-e45.3( **2260.0- ***2250.0- ***6760.0- ***601.0tbeD egagtroM htiW )6320.0( )73300.0( )74100.0( )80-e66.1( ***4580.0- ***7590.0- ***9480.0- ***951.0tbeD draC tiderC htiW )50-e61.1( )11-e15.8( )70-e59.7( )0( 05500.0- 60500.0 23600.0 548000.0tbeD otuA htiW )637.0( )166.0( )346.0( )159.0( 35500.0- 67800.0 50900.0 ***861.0 tbeD rehtO no tneuqnileD )348.0( )496.0( )017.0( )01-e89.5( ***294.0- ***624.0- ***744.0- )erocS tiderC(nL deggaL )0( )0( )0( ***137.0- ***505.0- ***726.0erocS tiderC gnissiM )0( )0( )0( rotacidnI 7120.0- 5810.0- 0310.0- 24400.0 5720.0 3220.0 9920.0- ***032.0tbeD naoL tnedutS gnissiM 30

egap suoiverp morf deunitnoC – 9 elbaT )8( )7( )6( )5( )4( )3( )2( )1( elbairaV )967.0( )927.0( )638.0( )549.0( )666.0( )327.0( )906.0( )80-e31.1( ***812.0 ***862.0 ***872.0 ***882.0 ***892.0 ***203.0 noitamrofnI rojaM gnissiM )77500.0( )70200.0( )378000.0( )983000.0( )152000.0( )122000.0( seY seY seY seY seY seY seY seY selbairaV ymmuD emiT 263,5 156,6 156,6 156,6 156,6 156,6 156,6 156,6 snoitavresbO 732.0 072.0 162.0 402.0 371.0 861.0 161.0 9780.0 2R oduesP .1.0<p *** ,50.0<p ** ,10.0<p *** .sesehtnerap ni seulav-P .lleps tnemyaper tsal eht gniretne retfa sraey 5 nihtiw tbed naol tneduts no tneuqniled syad +021 reve sa denfied elbairav tnednepeD 31

References CollegeBoard, “Trends in Higher Education 2013,” Tech. Rep. 140108, 2013. Dynarski, S., S. Hemelt and J. Hyman, “The Missing Manual: Using National Student Clearninghouse Data to Track Postsecondary Outcomes,” NBER Working Paper 19522, 2013. Dynarski, S. and D. Kreisman, “Loans for Educational Opportunity: Making Borrowing Work for Today’s Students,” Hamilton Project Discussion Paper, 2013. Gross, J. P. K., O. Cekic, D. Hossler and N. Hillman, “What Matters in Student Loan Default: A Review of the Research Literature,” Journal of Student Financial Aid, 2009. Hylands, T., “Student Loan Trends in the Third Federal Reserve District,” Cascade Focus, 2014. Looney, A.andC.Yannelis, “ACrisisinStudentLoans? HowChangesintheCharacteristics of Borrowers and in the Institutions They Attended Contributed to Rising Loan Defaults,” Brookings Papers on Economic Activity, 2015. Thompson, J. and J. Bricker, “Does Education Loan Debt Influence Household Financial Distress? An Assessment Using the 2007-09 SCF Panel,” FEDS Working Paper Series, (90), 2014. Wei, C. and L. Horn, “Persistence and Attainment of Beginning Students with Pell Grants,” U.S. Department of Education, National Center for Education Statistics, 2002. —, “A Profile of Successful Pell Grant Recipients: Time to Bachelor’s Degree and Early Graduate School Enrollment,” U.S. Department of Education, National Center for Education Statistics, 2009. 32

A Appendix A.1 The Data Set Our data set pools data from several sources.45 The data set starts with a nationally representative random sample of credit bureau records picked and provided by TransUnion, LLC, for a cohort of 34,891 young individuals who were between ages 23 and 31 in 2004, and spans the period 1997 through 2010. Individuals are followed biannually between June 1997 and June 2003, and then in December 2004, June 2007, and December 2008 and 2010. The data set contains all major credit bureau variables, including credit scores, tradeline debt levels, and delinquency and severe derogatory records. In order to capture information on enrollment spells and the institutional-level characteristics associated with each spell, in the next step individual educational records through 2007 are sourced from the DegreeVerify (for degrees) and Student Tracker (for enrollments) programs by the National Student Clearinghouse (NSC). The NSC educational institution identifiers allow us to further merge institutional records from the Integrated Postsecondary Education Data System (IPEDS), such as tuition, sector (e.g., public, private for-profit and not-for-profit, open admission), and SAT and ACT scores that are summarized at a school level. In addition, the NSC school identifiers allow us to also merge the historical schoollevel 2-year cohort default rates (CDR) collected by the Federal Student Aid (FSA). Finally, individual-level information about Pell Grants and Federal loans—sourced from the National Student Loan Data System (NSLDS)—is merged onto the data for Pell Grant and Federal student loan recipients. Given that only a subset of individuals in the core TransUnion sample have existing postsecondary education records, there are important differences between the TransUnion sampleanditsNSCsubsample. AsillustratedinTable10, outofthe34,891individualsinthe TransUnion sample, 54 percent (or 18,748 individuals) have an existing NSC postsecondary education enrollment record. While this college-going rate broadly matches the comparable rate of 58 percent in the nationally representative American Community Survey (ACS) 2007 45All the merges of individual-level information have been performed by TransUnion, LLC in conjunction with the National Student Clearinghouse, the CollegeBoard, and the Department of Education. The merges have been done based on a combination of Social Security number, date of birth, and individuals’ first and last names. None of the variables used to merge individuals across sources is available in our data set . 33

public files, it understates the rate slightly.46 In particular, besides never enrolling in college, a NSC record might be missing because the postsecondary education institution in which an individual was/is enrolled does not participate in the NSC Student Tracker or DegreeVerify programs (see Dynarski et al. (2013)).47 To the extent that the NSC program participation issues could be systematic, it is helpful to next examine the differences in baseline variables between the two samples. Turning to the differences in credit data indicators between the two samples, individuals represented in the NSC subsample are more frequently associated with student loan balances (62.8 versus 44.5 percent in the core TransUnion sample) as well as other types of debt, including auto debt (77.7 vs 72.8 percent), mortgage debt (50.8 vs 44.4 percent), and credit card debt (95.5 vs 87.9 percent). Mostly in line with expectations, 120 or more days past due delinquency rates on auto and mortgage debt are lower in the NSC (or college-going) subsample but, perhaps surprisingly, are higher for credit card debt: 1.1 vs 1.5 percent for auto loans, 4.3 vs 5.0 percent for mortgages, and 13.3 vs 12.7 percent for credit card debt. Consistent with these findings, the average maximum credit score per borrower is 36 points greater in the NSC subsample than in the core credit bureau data (712 vs 676). Taken together, the observed differences are in line with those one would expect when comparing the population with a subsample of individuals with at least some post-secondary education. Table10: ComparisonoftheCoreSamplewiththeNSCSubsample Variable N Mean Std. Dev. Min Max Age in 2004 All 34,891 27.5 2.6 23 31 NSC 18,748 27.2 2.5 23 31 Ever Attended Public 4-year (All) 11,022 0.316 Public 2-year (All) 11,805 0.338 Private 4-year Not-for-profit (All) 5,656 0.162 Private For-profit (All) 4,025 0.115 46The ACS estimate is based on a cohort of individuals between ages 25 to 34 in 2007. In our sample, individuals were between ages 26 to 34 in 2007. 47An additional reason for why an enrollment record may not be present in our data is due to data truncation related to the 2007 NSC data collection cutoff that is specific to our data set . However, this channel has no effect on comparability of the college-going rate in the NSC sample with the ACS estimate. 34

Table 10 – Continued from previous page Variable N Mean Std. Dev. Min Max Private 2-year Not-for-profit (All) 336 0.010 Public 4-year (NSC) 9,643 0.514 Public 2-year (NSC) 10,331 0.551 Private 4-year Not-for-profit (NSC) 4,275 0.228 Private For-profit (NSC) 955 0.051 Private 2-year Not-for-profit (NSC) 120 0.006 Public 4-year (NSC/NSLDS) 10,252 0.547 Public 2-year (NSC/NSLDS) 10,765 0.574 Private 4-year Not-for-profit (NSC/NSLDS) 5,055 0.270 Private For-profit (NSC/NSLDS) 2,245 0.120 Private 2-year Not-for-profit (NSC/NSLDS) 228 0.012 Highest Degree Attained Dropout 10,337 0.551 Associate’s or Certificate 997 0.053 Bachelor’s 4,338 0.231 Master’s or More 1,217 0.065 With Degree, but Level Unknown 1,859 0.099 Ever Had Student Debt (All) 15,526 0.445 Auto Debt (All) 25,416 0.728 Mortgage Debt (All) 15,497 0.444 Credit Card Debt (All) 30,683 0.879 Student Debt (NSC) 11,766 0.628 Auto Debt (NSC) 14,558 0.777 Mortgage Debt (NSC) 9,528 0.508 Credit Card Debt (NSC) 17,897 0.955 Ever Delinquent On Student Debt (All) 4,643 0.299 Auto Debt (All) 527 0.021 Mortgage Debt (All) 1,736 0.112 Credit Card Debt (All) 4,439 0.145 Student Debt (NSC) 3,207 0.273 Auto Debt (NSC) 210 0.014 Mortgage Debt (NSC) 800 0.084 Credit Card Debt (NSC) 2,489 0.139 35

Table 10 – Continued from previous page Variable N Mean Std. Dev. Min Max Pell Grants Ever Had (All) 11,191 0.321 Mean Pell Grants (All) 11,191 2,542.2 1,017.3 400 4,310 Ever Had (NSC) 8,453 0.451 Mean Pell Grants (NSC) 8,453 2,491.3 1,013.7 400 4,310 Maximum Level of Debt in Sample (in 2010 dollars, in Thousands) Student Debt (All) 14,522 26.6 35.2 0.001 438.6 Auto Debt (All) 23,370 22.5 15.9 0.048 528.8 Mortgage Debt (All) 15,069 212.9 175.1 0.164 3976.3 Credit Card Debt (All) 27,172 6.8 9.2 0.000 253.7 Student Debt (NSC) 11,240 29.9 37.5 0.001 438.6 Auto Debt (NSC) 13,632 22.0 14.8 0.067 528.8 Mortgage Debt (NSC) 9,355 228.2 177.6 2.704 3976.3 Credit Card Debt (NSC) 16,756 7.4 9.4 0.000 253.7 Maximum Credit Score in the Sample ALL 33,950 670.7 136.9 271 897 NSC 18,641 706.6 124.5 276 897 Not in NSC, with Student Debt 3,725 628.3 141.6 271 887 Students in Delinquency Status All 4,643 0.299 Based on TransUnion Information (All) 3,486 0.225 In Default on their Federal Loans (All) 2,858 0.184 NSC 3,207 0.273 Based on TransUnion Information (NSC) 2,488 0.211 In Default on their Federal Loans (NSC) 1,912 0.163 Diving deeper into the potential systematic bias due to reporting issues, we next turn to the group of 3,760 borrowers (or fully 24 percent of all student loan borrowers in our sample) who have existing student loan debt records in the credit bureau files and/or student loan records in the NSLDS, but no NSC records. Comparing the characteristics of these borrowers (shown in Table 11) with characteristics of student loan borrowers with existing NSC records (shown in Table 10) provides the best available gauge of the type of bias posed in our data set by the NSC reporting issues. As can be seen, in our sample, student loan 36

Table 11: Characteristics of Borrowers with Student Loan Debt but Not NSC Records Variable Obs Mean Std. Dev. Min Max Age in 2004 3,760 27.9 2.6 23 31 Credit Score 3,725 628.3 141.6 271 887 Pell Grants Ever Had 1,968 0.523 Mean Pell Grants 1,968 2,679.2 1,037.7 400 4,310 Ever Had Student Debt 3,760 1 Auto Debt 2,851 0.758 Mortgage Debt 1,514 0.403 Credit Card Debt 3,340 0.888 Ever Delinquent On Student Debt 1,436 0.382 Auto Debt 86 0.023 Mortgage Debt 215 0.057 Credit Card Debt 538 0.143 Maximum Level of Debt (in 2010 dollars, in Thousands) Student Debt 3,282 15.3 22.4 0.001 371 Auto Debt 2,583 22.6 16.3 0.222 173 Mortgage Debt 1,450 199.0 163.5 0.223 1,581 Credit Card Debt 2,774 6.6 9.6 0.000 186 Students in Delinquency Status Overall 1,436 0.382 Based on TransUnion Information 999 0.266 In Default on their Federal Loans 946 0.252 borrowers without NSC records tend to be slightly older than those with existing NSC records (27.9 versus 27.5 years old), are more likely to be Pell Grant recipients (52.3 versus 45.1 percent), and are associated with greater average Pell Grant balances ($2,679 versus $2,491).48 Also, individuals with student debt but no NSC records are significantly less likely to have mortgage debt (40.3 versus 50.8 percent), credit card debt (88.8 versus 95.5 percent) or auto debt (75.8 versus 77.7 percent). With the exception of auto debt holdings (which are, on average, slightly higher for this group), debt holdings of borrowers with no NSC records are lower on balance relative to the student loan borrowers with NSC records. Additionally, student loan borrowers with no NSC records are more likely to become delinquent on their student loan debt after entering repayment (38.2 versus 27.3 percent), and are also more frequently delinquent on their other debt before they start repaying their student loans (3.0 versus 1.4 percent for auto debt, 14.2 versus 8.4 percent for mortgage debt, and 24 versus 13.9 percent for credit card debt). Finally, the average credit score for this subsample is about 80 points lower. Given the salient determinants of student loan delin- 48The average Pell Grant balance is calculated as the total amount of Pell Grants received divided by the number of disbursements. 37

quency (described in Section 4) and also consistent with the lower student loan delinquency rate in the NSC subsample (shown in Tables 10 and 11), the existing differences in these indicators point to a possible attenuation bias in our results stemming from the elimination of student loan borrowers with no NSC records from our final estimation sample. The existing differences are, in part, caused by systematic differences in NSC coverage by school type, wherein institutions associated with a riskier credit profile of student loan borrowers (such as private for-profit schools) are underrepresented in the earlier years of the NSC coverage. To address the NSC coverage issues, we proceed in two steps. First, whenever available, we augment the NSC enrollment information with enrollment data from the NSLDS for Federal student loan recipients.49 Second, we develop yearly sample weights designed to correct for the underrepresentation of certain school sectors in the NSC data. Panels A and B in Table 1 show the yearly shares of borrowers entering repayment across school sectors in our NSC sample and the nationally representative FSA data, respectively.50 For each sector, theweightsareconstructedastheratiooftheseyearlysharesintheNSCsampleandtheFSA data by fiscal year.51 As seen in the panels, private not-for-profit and for-profit institutions are underrepresented throughout the NSC sample relative to the nationally representative FSA data, though their coverage increases significantly in later years of the NSC sample. The coverage issues are particularly severe prior to fiscal year 1998, leading us to eliminate individuals who entered repayment for the last time before fiscal year 1998 from the final sample. Related to the aforementioned NSC reporting issues, Table 12 compares completion rates by degree attained between the NSC subsample and the ACS 2007. As demonstrated, the NSC records on educational attainment in our data are incomplete for a number of individuals. In particular, in the NSC sample, 55.1 percent of individuals with at least some college attendance did not attain any degree, relative to only 36 percent in the ACS sample. This discrepancy suggests that the dropout rate is artificially inflated in our data set and—to the extent that leaving school without a degree is positively associated with 49The NSLDS contains information on enrollment spells that have been funded by Federal student loans. 50The FSA data are available for download at: http://www.ifap.ed.gov/DefaultManagement/press/. Unfortunately,theFSAdatadonotaccountforthoseborrowersenteringrepaymentwhoexclusivelyholdprivate studentloandebt. However,accordingtoourdata,lessthan2percentofborrowersholdonlyprivatestudent loan debt, and no Federal student loans, by the time they enter their last repayment spell observed in our sample. 51FiscalyearisdefinedastheperiodbetweenOctober1ofagivenyearandSeptember30ofthefollowing year. 38

Table 12: Comparison of Degree Attainment between the NSC and ACS Data ACS NSC All No College 41.7 46.0 At Least Some College 58.3 54.0 Attainment Rates Among Those with At Least Some College No Degree 36.0 55.1 Associate’s 14.0 5.3 Bachelor’s 35.8 23.1 Master’s or More 14.2 6.5 With Degree of an Unknown Type 0 9.9 Note: The ACS estimate is based on a cohort of individuals between ages 25 to 34 in 2007. In our sample, individuals were between ages 26 to 34 in 2007. delinquency (shown in Section 4)—is likely to introduce a downward bias in the effect of degree on student loan delinquency. Furthermore, there are systematic differences in the reporting of the highest degree attained between the NSC and the ACS. Namely, in the NSC subsample, 5.3 percent, 23.1 percent, and 6.5 percent of college-goers attained an Associate’s degree, a Bachelor’s degree, and a Master’s degree or more, respectively—much lower than the comparable rates of 14 percent, 35.8 percent, and 14.2 percent in the ACS data. The lower rates in the NSC subsample are partly accounted for by the fact that 9.9 percent of individuals with NSC educational records in our sample have a degree but the type of the degree is unspecified/unknown. We treat such individuals in the regression specifications as a separate category. Overall, the observed inaccuracies in the NSC data coverage are likely to reduce the statistical power of variables related to college completion and degree attainment in our econometric analysis. A.2 College Major Categories 39

Major Category Obs (1) Architecture and urban planning; construction trades 35 (2) English, foreign languages and literatures; visual and performing arts; 223 philosophy, religion, and theology (3) Biological, biomedical, and nature conservation studies; natural sciences; 220 agriculture (4) Communications and journalism; communications technologies and tech- 152 nicians (5) Computer and information systems 72 (6) Criminal justice 49 (7) Economics; geography, history, political science, sociology, and social 436 sciences; psychology; area, ethnic, and gender studies; parks, recreation, and leisure studies; family and consumer sciences (8) Education 259 (9) Engineering; engineering technologies and trades; mechanic and repair 151 technologies; precision production (10) Business, management, and marketing 547 (11) Health professions and related sciences 311 (12) Legal professions and studies 75 (13) Public administration and social work 59 (14) Liberal arts and sciences 74 (15) Personal and culinary services; transportation and materials moving; 958 other Total 3,621 40

A.3 Supplemental Graphs Summary statistics calculated from data based on less than 30 observations are reported as NA. Table 13: Distribution of Student Loan Debt in the NSC Sample, All College-Goers Student Loan Debt % Cum. % Obs No Debt 43.8 43.8 8,213 (0;10,000] 18.2 62.0 3,405 (10,000;20,000] 11.7 73.7 2,194 (20,000;30,000] 8.3 81.9 1,551 (30,000;50,000] 7.3 89.2 1,362 (50,000;75,000] 3.4 92.6 642 More than 75,000 3.6 96.2 667 Missing Debt 3.8 100.0 714 Total 18,748 Table 14: Distribution of Student Loan Debt in the Final Sample Student Loan Debt % Cum. % Obs (0;10,000] 36.4 36.4 2,418 (10,000;20,000] 22.1 58.5 1,470 (20,000;30,000] 15.2 73.6 1,008 (30,000;50,000] 10.6 84.2 706 (50,000;75,000] 4.6 88.8 307 More than 75,000 4.1 93.0 275 Missing Debt 7.0 100.0 467 Total 6,651 Table 15: Distribution of Student Loan Debt in the Final Sample, by Delinquency Status Not Delinquent Delinquent Student Loan Debt % Cum. % Obs % Cum. % Obs (0;10,000] 34.7 34.7 1,595 51.8 51.8 823 (10,000;20,000] 24.7 59.4 1,135 21.1 72.9 335 (20,000;30,000] 17.9 77.3 821 11.8 84.7 187 (30,000;50,000] 12.1 89.4 556 9.4 94.1 150 (50,000;75,000] 5.4 94.7 246 3.8 98.0 61 More than 75,000 5.3 100.0 243 2.0 100.0 32 Total 4,596 1,588 41

Table 16: Distribution of Student Loan Debt, by Highest Degree Attended* Undergraduate Graduate No Debt 51.0 51.0 5,307 20.8 20.8 321 (0;10,000] 20.6 71.6 2,146 8.8 29.6 135 (10,000;20,000] 11.7 83.3 1,220 11.0 40.6 170 (20,000;30,000] 7.6 90.8 788 11.0 51.6 169 (30,000;50,000] 4.2 95.0 437 15.8 67.5 244 (50,000;75,000] 1.0 96.0 104 12.3 79.7 189 More than 75,000 NA NA NA 16.6 96.3 255 Missing Debt 3.8 100.0 394 3.7 100.0 57 Total 10,415 1,540 Note*: Undergrad: LessorExactlyABachelor’sDegree;Grad: AtleastsomeeducationbeyondaBachelor’s degree. 797 Observations missing because it was not possible to determine the degree. Table 17: Distribution of Student Loan Debt in Final Sample, by Highest Degree Attended and Delinquency Status* Not Delinquent Delinquent Student Loan Debt % Cum. % Obs % Cum. % Obs (0;10,000] 41.2 41.2 1,353 55.5 55.5 793 (10,000;20,000] 27.5 68.7 905 22.1 77.6 315 (20,000;30,000] 18.9 87.6 620 11.8 89.4 168 Undergraduate (30,000;50,000] 9.7 97.3 320 8.2 97.5 117 (50,000;75,000] 2.2 99.5 71 2.3 99.9 33 More than 75,000 NA NA NA NA NA NA Total 3,286 1,428 (0;10,000] 11.8 11.8 124 NA NA NA (10,000;20,000] 15.4 27.2 161 NA NA NA (20,000;30,000] 15.2 42.4 159 NA NA NA Graduate (30,000;50,000] 20.4 62.8 214 26.3 52.6 30 (50,000;75,000] 15.7 78.5 165 NA NA NA More than 75,000 21.5 100.0 225 26.3 100.0 30 Total 1,048 114 Note*: Undergrad: LessorExactlyABachelor’sDegree;Grad: AtleastsomeeducationbeyondaBachelor’s degree. 262 non-delinquent borrowers and 46 delinquent borrowers not reported because of missing Highest Degree Attended. Table 18: Debt Levels and Delinquency Rates, by Highest Degree Attended Mean Median Delinquency Rate (%) Obs Undergraduate 15.2 11.7 30.3 4,714 Graduate 53.3 39.1 9.8 1,162 Total 5,876 Note*: 308 borrowers not included because the highest degree attended could not be determined. 42

Table 19: Distribution of Student Loan Debt, by Highest Degree Obtained* With College, No Bachelor’s With At Least Bachelor’s No Debt 56.6 56.6 4,250 31.3 31.3 1,401 (0;10,000] 22.8 79.4 1,716 12.7 44.0 567 (10,000;20,000] 9.0 88.4 679 16.0 60.0 716 (20,000;30,000] 4.2 92.7 318 14.3 74.3 641 (30,000;50,000] 2.6 95.2 194 10.9 85.2 488 (50,000;75,000] 0.8 96.0 57 5.3 90.5 237 More than 75,000 NA NA NA 5.9 96.4 263 Missing Debt 3.9 100.0 290 3.6 100.0 162 Total 7,515 4,475 Note*: With college, No B.A.: dropouts, Certificates, Associate’s Degree holders; With at least B.A.: B.A. and advanced degrees. 762 observations not reported because highest degree obtained could not be determined. Table 20: Distribution of Student Loan Debt in Final Sample, by Highest Degree Obtained and Delinquency Status* Not Delinquent Delinquent Student Loan Debt % Cum. % Obs % Cum. % Obs (0;10,000] 54.6 54.6 952 62.0 62.0 764 (10,000;20,000] 23.5 78.2 410 21.8 83.8 269 (20,000;30,000] 11.9 90.1 208 8.9 92.7 110 With College, No BA (30,000;50,000] 7.5 97.6 130 5.2 97.9 64 (50,000;75,000] 1.8 99.4 32 NA NA NA More than 75,000 NA NA NA NA NA NA Total 1,742 1,233 (0;10,000] 20.2 20.2 527 12.9 12.9 40 (10,000;20,000] 25.4 45.6 661 17.8 30.7 55 (20,000;30,000] 22.0 67.7 573 22.0 52.8 68 With At Least BA (30,000;50,000] 15.6 83.2 405 26.9 79.6 83 (50,000;75,000] 7.9 91.1 205 10.4 90.0 32 More than 75,000 8.9 100.0 232 10.0 100.0 31 Total 2,603 309 Note*: With college, No B.A.: dropouts, Certificates, Associate’s Degree holders; With at least B.A.: B.A. and advanced degrees. 251 non-delinquent borrowers and 46 delinquent borrowers because highest degree obtained could not be determined. Table 21: Debt Levels and Delinquency Rates in the Final Sample, by Highest Degree Obtained Mean Median Delinquency Rate (%) Obs With College, No BA 12.5 8.0 41.4 2,975 With At Least BA 33.2 22.0 10.6 2,912 Total 5,887 Note*: 297 borrowers not included because highest degree obtained could not be determined. 43

Table 22: Distribution of Student Loan Debt in Final Sample (Most Disaggregated), by Highest Degree Obtained and Delinquency Status* Not Delinquent Delinquent Student Loan Debt % Cum. % Obs % Cum. % Obs (0;10,000] 54.9 54.9 811 63.0 63.0 725 (10,000;20,000] 22.5 77.4 333 20.4 83.4 235 (20,000;30,000] 12.3 89.7 182 9.3 92.7 107 Dropouts (30,000;50,000] 7.6 97.3 112 5.2 97.9 60 (50,000;75,000] 2.1 99.4 31 NA NA NA More than 75,000 NA NA NA NA NA NA Total 1,478 1,151 (0;10,000] 53.4 53.4 141 47.6 47.6 39 (10,000;20,000] 29.2 82.6 77 41.5 89.0 34 (20,000;30,000] NA NA NA NA NA NA Associate’s/Certificate (30,000;50,000] NA NA NA NA NA NA (50,000;75,000] NA NA NA NA NA NA More than 75,000 NA NA NA NA NA NA Total 264 82 (0;10,000] 24.1 24.1 426 14.5 14.5 33 (10,000;20,000] 29.8 53.9 526 20.7 35.2 47 (20,000;30,000] 25.3 79.2 447 26.9 62.1 61 Bachelor’s (30,000;50,000] 13.4 92.6 236 28.6 90.7 65 (50,000;75,000] 4.8 97.3 84 NA NA NA More than 75,000 2.7 100.0 47 NA NA NA Total 1,766 227 (0;10,000] 15.1 15.1 80 NA NA NA (10,000;20,000] 16.0 31.1 85 NA NA NA (20,000;30,000] 16.8 47.8 89 NA NA NA More than Bachelor’s (30,000;50,000] 19.6 67.4 104 NA NA NA (50,000;75,000] 14.3 81.7 76 NA NA NA More than 75,000 18.3 100.0 97 NA NA NA Total 531 41 Note*: 557 non-delinquent and 87 delinquent borrowers not reported because not able to determine the exact degree obtained. Table 23: Debt Levels and Delinquency Rates in the Final Sample(Most Disaggregated), by Highest Degree Obtained Mean Median Delinquency Rate (%) Obs Dropouts 12.5 7.9 43.8 2,629 Associate’s/Certificate 12.3 9.2 23.7 346 Bachelor’s 24.1 19.6 11.4 1,993 More than Bachelor’s 48.3 34.2 7.2 572 Total 5,540 Note*: 664 borrowers not reported because the exact highest degree obtained could not be determined. 44

60. 40. 20. 0 PDF 0 20 40 60 80 100 Student Loan Debt All Delinquent Current 1 8. 6. 4. 2. 0 Student Loan Debt Distribution Debt in Thousands of 2010 Dollars CDF 0 50 100 150 200 debt All Delinquent Current ® Figure 7: Student Loan Debt Distribution: Final Sample 50. 40. 30. 20. 10. 0 PDF 0 50 100 150 200 Student Loan Debt Undergraduate Graduate 1 8. 6. 4. 2. 0 Student Loan Debt Distribution by Highest Degree Attended Debt in Thousands of 2010 Dollars CDF 0 50 100 150 200 250 debt Undergraduate Graduate ® Figure 8: Student Loan Debt Distribution by Highest Degree Attended: Final Sample 45

60. 40. 20. 0 PDF−Undergraduates 0 20 40 60 80 100 Student Loan Debt Delinquent Current 510.10.500. 0 PDF−Graduates 0 50 100 150 200 250 Student Loan Debt Delinquent Current 1 8.6.4.2. 0 CDF−Undergradutes 0 20 40 60 80 100 debt Delinquent Current 1 8.6.4.2. 0 St. Loan Debt Dist. by Delinq. Status & Highest Degree Attended Debt in Thousands of 2010 Dollars CDF−Gradutes 0 50 100 150 200 250 debt Delinquent Current ® Figure 9: Student Loan Debt Distribution by Delinquency Status and Highest Degree Attended: Final Sample 80. 60. 40. 20. 0 PDF 0 50 100 150 200 Student Loan Debt College, No BA At Least BA 1 8. 6. 4. 2. 0 St. Loan Debt Dist. by Highest Degree Obtained Debt in Thousands of 2010 Dollars CDF 0 50 100 150 200 250 debt College, No BA At Least BA ® Figure 10: Student Loan Debt Distribution by Highest Degree Obtained: Final Sample 46

80.60.40.20. 0 PDF−College, No BA 0 20 40 60 80 100 Student Loan Debt Delinquent Current 30. 20. 10. 0 PDF−At Least BA 0 50 100 150 200 250 Student Loan Debt Delinquent Current 1 8.6.4.2. 0 CDF−College, No BA 0 20 40 60 80 100 debt Delinquent Current 1 8.6.4.2. 0 St. Loan Debt Dist. by Delinq. Status & Highest Degree Obtained Debt in Thousands of 2010 Dollars CDF−At Least BA 0 50 100 150 200 250 debt Delinquent Current ® Figure 11: Student Loan Debt Distribution by Highest Degree Obtained: Final Sample 47