On Intergenerational Immobility: Evidence that Adult Credit Health Reflects the Childhood Environment

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. On Intergenerational Immobility: Evidence that Adult Credit Health Reflects the Childhood Environment Sarena Goodman, Alice Henriques, and Alvaro Mezza 2017-032 Please cite this paper as: Goodman, Sarena, Alice Henriques, and Alvaro Mezza (2017). “On Intergenerational Immobility: EvidencethatAdultCreditHealthReflectstheChildhoodEnvironment,”Finance and Economics Discussion Series 2017-032. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2017.032r1. 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.

On Intergenerational Immobility: Evidence that Adult Credit Health Reflects the Childhood Environment1 Sarena Goodman, Federal Reserve Board of Governors Alice Henriques, Federal Reserve Board of Governors Alvaro Mezza, Federal Reserve Board of Governors November 2017 Abstract Using a novel dataset that links socioeconomic background to future credit, postsecondary education, and federal student loan and grant records, we document that, even though it is not and cannot be used by credit agencies in assigning risk, background is a strong predictor of adult credit health. A relationship remains upon inclusion of achievement, attainment, and debt management metrics. These findings reveal a new dimension along which childhood circumstances persist into adulthood and imply that the many important contexts in which credit scores are relied upon to evaluate individuals (e.g., lending, insurance, employment) may be helping to preserve inherited inequities. 1 Email: sarena.f.goodman@frb.gov, alice.m.henriques@frb.gov, and alvaro.a.mezza@frb.gov. This paper does not necessarily reflect the views of the Federal Reserve Board. A previous version of this paper was circulated as “Where Credit Is Due: The Relationship between Family Background and Credit Health.” We thank Kyle Coombs, Peter Hansen, and Steve Ramos for excellent research assistance, and Bradley Hardy, Don Carmichael, Joanne Hsu, Karen Pence, Michael Palumbo, and Ronel Elul for helpful comments. We thank participants at the 2016 Association for Education Finance and Policy Annual Conference, the 2016 FRS Applied Micro Conference, the fall 2016 FRB Applied Micro Lunch seminar, the 2017 Society of Labor Economists Annual Meeting, the 2017 International Banking, Economics, and Finance Association Sessions of the Western Economic Association International Annual Conference, the 2017 Boulder Summer Conference on Consumer Financial Decision Making, and the 2017 FDIC Annual Consumer Research Symposium. All errors are our own.

I. Introduction While equal access to opportunity is the cornerstone of the American Dream, socioeconomic status (SES) is highly correlated between parents and their children (Solon, 1999, Chetty et al., 2014). This persistence raises questions of whether individuals from different backgrounds can access the same opportunities, and, if not, where policy may help level the playing field. Research progress on these questions requires the analysis of measurable economic outcomes, both to understand the evolution of gaps over time and to identify the forces generating that evolution. A particularly promising line of inquiry would examine outcomes of relatively young individuals that could cumulatively affect economic circumstances over the life-cycle. For example, findings that early academic achievement both reflects family background and can influence economic paths throughout adulthood imply that educational interventions may help address inequality of opportunity.2 This study exposes a gap in outcomes within credit files. Examining the future credit records of a random sample of college-bound individuals, we establish a link between socioeconomic background and early-career credit health. We find that, even though background is not and cannot be used by credit agencies in rating risk, a clear gap in credit scores has materialized by the time the members of our sample are about 30 years old. This gap is evident whether we consider raw credit scores or summary measures of access to credit that can be derived from scores, and when we examine these outcomes six years later. It also remains upon inclusion of a full battery of achievement, postsecondary education, and debt management metrics. Altogether, the results reveal a new dimension along which childhood circumstances persist into adulthood, 2 Chetty et al. (2011) suggests that gains in kindergarten achievement translate into improvements in adult outcomes. See Reardon (2011) for evidence of gaps in early achievement. 2

with the resilience of the relationship raising the concern that background might have predictive power in credit-scoring models. The results also imply that the credit market, as well as the many other important contexts in which credit scores are relied upon to evaluate individuals, may help preserve inequities that existed during childhood and thus serve as conduits for the transmission of economic disparities across generations. The credit score is a critical metric for understanding economic mobility. Technically, it is a dynamic statistic that reflects the likelihood an individual will default on debt within a set time frame, based on her historical interaction with credit markets at the point in time it is generated.3 Despite this narrow definition, credit scores are relied upon within a broad variety of contexts to assess the risks of contracting with individual consumers, and thus can greatly influence whether they can consumption-smooth over the lifecycle or through periods of economic hardship (Herkenhoff, 2015; Herkenhoff, Phillips, and Cohen-Cole, 2016).4 Lenders use credit scores to set prices and terms of loans, whereby sufficiently low scores can even render certain lenders or loan types completely unavailable.5 Many businesses and organizations that do not provide credit 3 There are a number of types of credit scores, varying in purpose and definition, but they are all derived from observable dimensions of credit records. In general, demographics are not observable to credit scoring agencies and, by law and regulation, cannot be included in the models that generate scores. Examples of key inputs include: the credit markets in which she has participated, the amount of debt she has outstanding, her prior repayment behavior, and the length of time she has maintained healthy credit, all of which predict delinquency to some extent. One popular variant reflects the likelihood a consumer will become seriously delinquent within 24 months of scoring. (For more information, see: https://www.fdic.gov/regulations/examinations/credit_card/pdf_version/ch8.pdf.) 4 While this discussion ignores endowments, the ability to consumption-smooth vis-à-vis credit markets is especially important for individuals with small or negligible endowments (e.g., those from economically disadvantaged backgrounds) relative to expected future income. This rationale is the policy basis for broad-based federal student lending programs, which, in theory, allow individuals with little or poor credit histories the opportunity to fund human capital investments with loans that are generally not underwritten at the student level and that, in expectation, will offer large payouts in earnings in the future. 5 We expand our understanding of how young adults from different backgrounds interact with the credit market using the Federal Reserve Board’s 2016 Survey of Consumer Finances (SCF). The SCF shows that young adults from disadvantage are 3 to 5 percentage points more likely to be denied credit and, among those who do not apply for credit, are about 6 to 7 percentages points more likely to report “fear of denial” as the explanation. Such individuals are, however, 4 to 5 percentage points more likely to apply for credit. 3

also use credit scores to evaluate risk; for example, lower scores might restrict access to insurance, rental housing, utility contracts, and employment opportunities. The paucity of research in this area primarily reflects data constraints, as few, if any, datasets include observations of both childhood socioeconomic background and adult credit outcomes. To overcome these constraints, we merge College Board (CB) records for a sample of SAT-takers who graduated high school between 1994 and 1999 (when the SAT was elective) to their future administrative credit bureau, postsecondary education, and federal student loan and Pell Grant records.6 The analysis leverages four measures of background, each derived from either parent’s educational attainment or Pell Grant receipt (a major form of need-based Federal financial aid). In general, the individuals who form our sample were, when they were in high school, planning to attend (a selective) college. Importantly, in 2008, the focal year of our analysis, they were, on average, 30 years old and thus had generally completed their schooling. Within this sample, family background unambiguously predicts credit health, approximated with either raw credit scores or a binary measure of access to prime lending. In our simplest specification, we estimate that credit scores are about 100 points lower for individuals from disadvantaged backgrounds, and that such individuals are about 20 percentage points more likely to be scored in the subprime range. Conditioning on achievement erases as much as half the gap; even then, another 100 to 200 SAT points are required to fully eliminate it. Estimates are extremely similar if 2014 credit outcomes (i.e., when the sample is in their mid-30s) are analyzed instead. 6 Our analysis draws upon a large data effort specially prepared for the Federal Reserve Board that drew randomly among individuals with credit records in 2004 who were 23 to 31 years old at the time, as described in Mezza and Sommer (2016). Our sample is the subset of these individuals for whom we can match CB records. 4

We extend the analysis in two ways. First, we allow the role of achievement to vary by background. The results indicate that although, all else equal, higher achievement reduces the gap in credit health, a gap is evident throughout the achievement distribution, even among students with very high achievement. Second, we investigate whether factors that may correlate with background and influence credit health—either directly (e.g., the take-up of, the amount of, and defaulting on federal undergraduate loans; delinquency on other types of debt; length of credit history) or indirectly (e.g., school quality; educational attainment)—mediate the relationship between background and credit.7 Within every specification we consider, differences in credit outcomes remain. Within these analyses, we also find evidence that taking a federal student loan—which, unlike other forms of debt, does not entail credit underwriting or risk-pricing—is negatively associated with adult credit health.8 Ironically, students who borrow through the federal student loan programs, which were created to equalize opportunity, may struggle with early debt obligations and foreclose on future opportunities as a result. This finding points to a new avenue through which student loans may deleteriously affect young adults’ financial health (e.g., Mezza et al., 2016; Dettling and Hsu, 2014, Bleemer et al., 2014). 7 For example, individuals from disadvantaged backgrounds attend lower-quality schools than their peers with similar SAT scores, consistent with evidence that disadvantaged students tend to “under-match” to postsecondary programs and, more broadly, have worse outcomes. In addition, such individuals are less likely to complete programs and more likely to borrow for college and default on this debt. And although they have, on average, credit histories that are two to three months longer, this difference is small (about 2 percent of the sample mean) and may simply be an artifact of increased college borrowing rates among this group. 8 Although federal undergraduate loans are not underwritten and are generally widely available, some restrictions in eligibility still apply. For instance, students who are currently in default on a student loan may not take out another. In addition, students face maxima in the amount they can borrow both in a single year and over time. Other types of student loans that are available but that we do not study (e.g., Parent PLUS loans; private student loans) are less widely used and tend to have more stringent requirements. 5

More broadly, our findings relate to the battery of studies within the Equality of Opportunity Project—a research initiative that uses federal tax records to examine intergenerational mobility in the United States—along several dimensions: Relying on linked administrative records to follow an individual’s life course, documenting a clear relationship between parents’ and children’s socioeconomic circumstances, and finally, illuminating a setting in which policy may help mitigate the transmission of disadvantage.9 Building on this work, the gap we uncover implies a material difference in how well individuals from different backgrounds can consumption-smooth and access key markets at relatively young ages, which, given how credit scores are constructed and used, is both very likely to persist and to contribute to differences in other important lifecycle outcomes.10 In the paper’s conclusion, we review the phenomena that could generate this gap (e.g., social norms, financial literacy, or household conditions; how credit scoring models assess risk), many of which may be addressed through policy, and bring forward new evidence from the Federal Reserve Board’s 2016 Survey of Consumer Finances (SCF). The rest of the paper proceeds as follows. Section II summarizes the institutional background and related literature. Section III describes the sample and construction of key variables. Section IV links family background and pre-collegiate achievement to credit outcomes. Section V extends the analysis by allowing for achievement to interact with background and by including an array of factors that could mediate credit health. Section VI discusses the possible explanations for the gap in credit health and concludes. 9 See http://www.equality-of-opportunity.org/. 10 Negative credit events remain on credit records for fairly long periods of time and the presence of such an event is a major predictor of delinquency risk. For example, a personal bankruptcy can remain on a credit record for up to 10 years (Musto, 2004). More broadly, credit scores depend on one’s historical experience managing and maintaining credit, and access to credit is reduced for those with lower scores. 6

II. Setting a. Institutional Background A credit score is a dynamic summary measure of an individual’s credit risk, derived from elements of her existing credit record. There are many variants of these scores, with each of the three major credit bureaus (i.e., Equifax, Experian, Transunion) producing several of their own measures. Some scores are intended for more general use while others denote the risk for a particular credit product, but all variants share the same basic properties, described below. Technically, a credit score reflects an individual’s relative risk of default within a fixed time period, based on the risk profiles of other individuals with credit records and the experiences of those whose records are similar. These risk profiles are derived solely from the interactions with credit markets that are reported to the credit bureaus. While all credit scores are based on proprietary formulas, according to the Fair Isaac Corporation (FICO)—the data analytics company that produces the “FICO score” (perhaps the most well-known variant of credit score)—the elements of credit records that predict credit risk fall into one of five categories, each assigned some weight within the methodology for a particular score. For the FICO score, payment history is given the most weight (0.35), then amount owed (0.3), length of credit history (0.15), new credit (0.1), and credit mix (0.1).11 (Our study relies on a distinct credit score, the TransUnion TransRisk Account Management score—TU TransRisk AM 2.0.—which is computed similarly to the FICO score but may weight these categories differently.) More generally, for each score variant, each element is weighted by the extent to which it has predictive power for the type of risk being modeled. In credit scoring models, a negative credit 11 For additional details, see: http://www.myfico.com/CreditEducation/WhatsInYourScore.aspx. 7

event, such as a prior default or a bankruptcy, can remain on a report for a long time and is often a strong predictor of delinquency. Credit scores are used to evaluate applications for credit, identify prospective borrowers, and manage existing credit accounts. They also enter into eligibility determinations for insurance, rental housing, utility contracts, and employment opportunities. Individuals are frequently then further classified as either prime or subprime according to their credit scores. While the cutoff varies by credit score variant and institution, there is usually a distinct break in the types and costs of products and services available to borrowers of each type, particularly within the lending and insurance industries. In general, prime individuals have markedly higher access to credit (and to opportunities in other markets that rely upon credit scores) than subprime individuals. In general, demographics are not observable to credit scoring agencies and cannot be included in the models that generate scores. Fair lending and credit laws mandate protections for certain groups—such as those formed from race, ethnicity, sex, or the receipt of public assistance—many of which strongly correlate with disadvantage.12 By law and regulation, scoring models cannot include an input that only appears to predict credit risk because it proxies for membership to one of these groups. Beyond those specifically protected, because of their many uses, it would clearly be very worrisome if credit scoring models penalized any demographic or socioeconomic group. Indeed, in light of these concerns, Section 215 of the Fact Act of 2003 called for a comprehensive examination of the impact that scoring models and scores have on the availability and affordability of credit and insurance with respect to “geography, income, ethnicity, race, color, religion, national origin, age, sex, marital status, and 12 For details, see the 1974 U.S. Equal Credit Opportunity Act (codified at 15 U.S.C. § 1691). 8

creed.” The analyses that followed indicated that model inputs maintain predictive power within each group and also that they do not proxy for race, ethnicity, or sex (Avery, Brevoort, and Canner, 2012; Board of Governors, 2007).13 However, systematic differences by socioeconomic background have never been ruled out. b. Related Literature The correlation of income and wealth across generations in the United States is well documented. The most recent work, which relies on administrative income data, estimates that the current level of persistence is high relative to other countries but also generally unchanged from several decades ago (Lee and Solon, 2009; Chetty et al., 2014; Mitnik et al., 2015; for a review of prior work, Solon, 1999; Black and Devereux, 2011). Mitnik et al. (2015) find that the correlation appears to be stronger when children’s outcomes are observed later in their lifecycle (their 40s instead of their 30s), suggesting that the role of family background does not diminish over time and may even grow. Studies that estimate intergenerational correlations within segments of the population offer mixed conclusions. Mazumder (2005) finds that persistence is highest among families with low net worth, while Mitnik et al. (2015) find the highest persistence in the upper-middle class. Two recent studies in this area introduce a credit dimension. First, Ghent and Kudlyak (2015) examine intergenerational persistence within the FRBNY Consumer Credit Panel, a panel dataset comprised of the full set of credit records for a random sample of the population, augmented with the records of any person sharing an address with these individuals for the duration of their 13 According to this literature, different demographic groups have substantially different credit scores, on average. For example, individuals residing in low-income census tracts have lower credit scores than individuals residing in more affluent areas. However, these differences narrow—but are not always eliminated—when personal demographic characteristics, neighborhoods, or census-tract-based estimates of income are taken into account. 9

co-residence. They identify parent-child pairs from the set of 19-year olds (“child”) that live with an older individual (“parent”) and detect a positive association between “parents’” credit scores when their “children” were 19 years old and the credit scores of these “children” 5 to 10 years later. Second, Ringo (2015), using the Survey of Consumer Payment Choice (SCPC), estimates a positive relationship between parents’ reported credit scores and both the likelihood their children attend college and the likelihood they complete a four-year degree. While both studies document strong correlations, parental credit scores may reflect other household conditions, including those that vary over time. Finally, a related strand of literature examines household conditions and college attendance. According to standard theory, the decision to attend college primarily reflects its net return.14 While existing estimates imply substantial returns to college (and college quality), particularly among disadvantaged students (e.g., Card, 1995; Black and Smith, 2006; Hoekstra, 2009; Dale and Krueger, 2002, 2011; Zimmerman, 2014), children from low-income households are far less likely to attend than their peers, with the disparity increasing over time (Bailey and Dynarski, 2011).15 Moreover, among those who aspire to attend college, children from low-income households more frequently do not apply to or attend a college commensurate with their abilities (e.g., Bowen, Chingos, and McPherson, 2009; Pallais and Turner, 2006; Spies, 2001; Hoxby and Avery, 2013; Smith, Pender, and Howell, 2013; Black, Cortes, and Lincove, 2015).16 This 14 In a credit constrained environment, family income is independently relevant, though the evidence that such constraints bind is mixed (Lovenheim and Reynolds, 2013; Hilger, 2016; Bulman et al., 2016). 15 Meanwhile, earnings have been essentially steady among the college-educated and have dropped substantially for everyone else (The College Board, 2007; Deming and Dynarski, 2010). 16 Further, many of these studies rely on application and score data from elective admissions tests; thus, while the shortage of these students attending and applying to top schools is likely larger than conventional estimates suggest, another potential takeaway is that relatively ambitious students (i.e., those who plan to attend competitive schools) are under-matching. 10

collection of findings has two important implications for our study. First, they suggest that interventions that raise college access could improve economic mobility, similar to how our study seeks to lay the groundwork for innovations in the credit space. Second, they suggest that an analysis based upon a sample of individuals who took the SAT exam—which, in the period we examine, was primarily required of students planning to attend a selective college—may disproportionately reflect the behaviors and outcomes of high-income (or high-achieving) students, which bears on the interpretation of our results. III. Data Our sample consists of person-level records that link socioeconomic background and achievement to postsecondary credit outcomes. The sample draws upon the universe of administrative credit records held by TransUnion (TU) in 2004. The “base cohort” is formed by randomly selecting a nationally representative cohort of about 35,000 individuals with credit files who were 23 to 31 years old in 2004, for whom we are able to observe an array of credit outcomes in snapshots taken periodically between 1997 and 2014. For a subset of these individuals, we identify administrative records from other institutions, allowing us to observe additional characteristics, including demographics and postsecondary schooling. (A double-blind process between TU and the other data sources was used to maintain the integrity and privacy of each party’s records. The records in our dataset are anonymous.) Specifically, we acquired CB data for SAT-takers spanning the 1994 to 1999 high school graduation cohorts, and, for about 15 percent of the base cohort, we are able to identify a matching SAT score record. The SAT is an elective competitive exam administered during students’ junior and senior years of high school that is used in admissions determinations at selective colleges (and course placement at non-selective colleges). During the period we study, 11

the SAT was fully elective and only considered to be a requirement among college admissions and placement committees, so the subsample of individuals for whom we can identify SAT records very likely planned to attend college.17 Hence, in general, the subset of students for whom we can successfully identify an SAT record are “college bound.” In addition to SAT scores, the CB records also contain student demographic characteristics (e.g., parental education; student gender; state of residence) from a survey that the CB administers to students who take the SAT. These characteristics are part of an endowment bundle that individuals inherit from their parents’ genetics, household conditions, and other circumstances beyond their control. For these same individuals, we identify any records that exist within (1) the Department of Education (DoEd) National Student Loan Data System (NSLDS) pertaining to their federal Title IV grant and borrowing behavior (e.g., whether an individual received Pell Grants, borrowed for education, the total amount she borrowed, and whether she defaulted over our period of study), and (2) the National Student Clearinghouse (NSC) pertaining to their enrollment and educational 17 Because, in general, students must register for the SAT exam with their Social Security Numbers, the matched set likely approximates the full set of SAT-takers among individuals in our base sample. Still, not all students who attend postsecondary institutions necessarily appear in the CB data. Indeed, about 75 percent of the individuals in our base credit sample for which we can identify either a National Student Clearinghouse (NSC) or National Student Loan Data System (NSLDS) record cannot be matched to a CB record. There are at least three explanations for this seemingly low match rate. First, although our base credit bureau sample is formed from nine birth cohorts, our CB data span only six graduation years. Thus, to begin with, we would expect, at most, a 67 percent match rate. Second, students may elect a competing exam, the ACT, for their college applications. Because a student’s proclivity to elect a particular exam is not necessarily randomly assigned, the omission of ACT-takers is a potential threat to the external validity of our analysis. However, our descriptive statistics are broadly in line with national statistics among all college students and are little changed when we restrict the sample to students from states where the SAT prevails. (Appendix Table 2 indicates our main estimates are not sensitive to this restriction.) Thus, for our purposes, the election of the ACT over the SAT in our sample is approximately random. Third, many postsecondary students attend schools that do not require an admissions exam. In 2000, less than 10 percent of four-year postsecondary institutions, but 80 percent of two-year institutions, fell into this category (Breland et al., 2002). Therefore, our analysis is most precisely an examination of four-year college-bound individuals, which, if anything, should be a positively selected group among the full set of postsecondary attendees who appear in the NSC or NSLDS. We will further discuss implications of such selection when we turn to our analysis. 12

attainment.18 These records each pertain to a particular postsecondary institution, which can then be linked to two DoEd external data sources to identify important characteristics of that school. The first is the Integrated Postsecondary Education Data System (IPEDS) database, which compiles responses from an annual survey of all Title IV institutions over our full period of study and contains snapshots of school characteristics (e.g., sector; level; selectivity; price) over time. The second is the “college scorecard,” which includes borrowing and later-life earnings outcomes for every cohort beginning in 1996.19 The final sample comprises the SAT survey and testing, postsecondary, and credit outcomes of 5,421 college-bound individuals. Within this sample, we generate several key variables. As a first measure of family socioeconomic circumstances, we use parents’ educational attainment, as reported by SAT-takers in the CB survey. We code the following two measures: (1) a binary measure of the mother’s B.A. status (“mom”), where a value of 0 reflects completing at least a B.A. and 1 reflects not completing a B.A.; and (2) a binary measure of the father’s B.A. status (“dad”), coded similarly to measure 1.20 Note that higher values of the parents’ education variables are associated with less education (i.e., “disadvantage”) in order to permit a consistent interpretation of the signs across all of our family background variables. As a complement to these variables, we derive a second set of measures from the NSLDS records. Specifically, Pell Grants, which are awarded to qualifying low-income financial aid applicants with the amount of the grant fully determined by financial need, enrollment status, and 18 These data sources contain the universe of their respective administrative records in a given academic year and contain records through 2008. 19 https://collegescorecard.ed.gov/data/ 20 The analysis excludes students for whom the corresponding parent’s education is either missing or reported as a zero. 13

school’s tuition level, offer a second snapshot of a student’s socioeconomic background, particularly financial well-being, around the time of the schooling decision. We code the following two measures: (1) whether the individual was ever awarded a Pell Grant (“any pell”), and (2) because the award amount is subject to a statutory limit set in each year, whether the individual was awarded the maximum Pell Grant (“max pell”).21 Note that any student who receives the maximum Pell Grant, by definition, is coded as 1 for “any pell”; thus, relative to “any pell,” “max pell” captures a more extreme measure of need (and cost). The two sets of measures together offer a fairly comprehensive snapshot of an individual’s background. Compared with parental education, the Pell measures offer the relative benefit of directly quantifying financial need; however, unlike the parental education measures, they in part reflect a student’s schooling decision (and decision to apply for financial aid) and, thus, are not fully predetermined. Because of the differences in what each measure captures, the correspondence is not exact; for instance, the correlation between “dad” and “any pell” is 0.25. Still, those with less educated fathers are about twice as likely to receive a Pell Grant as those with more educated fathers, which mirrors the association between parental education and Pell Grant receipt within the 2003–04 DoEd National Postsecondary Student Aid Study (NPSAS). The other key variables are achievement and credit health. Achievement is measured prior to college, using SAT scores (“maxsat100”), measured in hundreds and ranging from 4 to 16 in increments of 0.1.22 The SAT score reflects a combination of an individual’s cognitive ability, academic preparation, and, because of its direct use in the college admissions process, 21 Technically, the Pell Grant measures are derived from observations of whether individuals were scheduled to receive a Pell Grant and the amount that they were scheduled to receive (rather than the actual amounts). 22 For ease of interpretation of the coefficients and the constant, we transform this variable to range between 0 and 12 in the regression analysis. 14

opportunity set at the time she graduates high school. We measure credit health in two ways, both using the TU credit score observed in 2008 (“tuscore2008”), at which point the youngest individual is 27 or 28 years old. First, we use the raw score, which ranges from 270 to 900. Second, to provide a mapping between credit health and credit access, we derive a binary variable (“prime”) that approximates whether an individual would qualify for most types of credit in 2008. By this metric, members of our sample are considered prime borrowers if their credit score is above the base cohort’s median score in 2008.23 Our results are not very sensitive to the choice of this cutoff.24 We augment our analysis with an array of postsecondary variables that may influence credit health, either directly or indirectly. We use the NSLDS borrowing records to create the following: a binary measure (“borrowed ug”) that takes a value of 1 if the individual took federal student loans to fund her undergraduate studies; the cumulative undergraduate student loan borrowing through federal loans, measured in thousands of dollars (“amount borrowed”); and a binary measure that takes a value of 1 if the individual ever defaulted on a federal undergraduatelevel student loan (“defaulted”). Moreover, we make use of two school quality measures based 23 This threshold corresponds to a TU credit score of 580.5. Key distributional characteristics of the TU score differ from those of the FICO score, with which people are most familiar. Laufer and Paciorek (2016) show that there is a close relationship between the FICO score and the Equifax Risk score within the FRBNY Consumer Credit Panel (CCP), a dataset that is available to us as well. We mimic our sample restrictions within the CCP and calculate a median Equifax Risk score in December 2008 of 645. (According to the “Quarterly Report on Household Debt and Credit,” less than one-quarter of mortgages and slightly more than one-quarter of auto loans during the last quarter of 2008 were originated to individuals with scores below 645.) 24 Appendix Table 3 presents estimates using alternative thresholds, two of which relate to mortgage lending standards and a third that roughly accords with a current industry consensus definition of a prime borrower: (1) a TU score of 526, which corresponds to an Equifax Risk score of 620—a cutoff commonly used by mortgage lenders in applications for credit, especially after 2009 (Laufer and Paciorek, 2016); (2) a TU score of 351, which corresponds to an Equifax Risk score of 550—a score at which very few mortgage originations occur, even in 2008; and (3) a TU score of 620, which approximates the probability of default associated with a FICO score between 680 and 700. (We thank Ezra Becker and Transunion for helpful guidance in developing our third threshold.) Estimates are qualitatively similar under all three measures, even under the most stringent threshold. 15

on the first college an individual attends—i.e., the first enrollment spell we observe in either the NSC or the NSLDS. The first is the average income in 2007 among employed individuals who had been enrolled in that school in 1997 (“school’s mean income,” measured in thousands of dollars). The second is the average SAT score of students admitted in 2003 (“SAT school”).25 We construct four mutually exclusive measures of degree status. We primarily rely on NSC graduation records but complement this information with NSLDS records when possible. We group degrees into the following categories: (1) dropouts (i.e., those with at least some college but no degree), (2) certificate or associate degree, (3) bachelor’s degree, and (4) master’s degree or more.26 We also construct a persistence variable that counts days enrolled and expresses them in years (“years in school”), combining information from NSC and NSLDS enrollment records. Finally, we derive two variables from credit records, one that measures whether an individual has ever been delinquent on non-student loan debt (“delinquent on non-student loan debt”) and another that counts the number of months an individual has had an established credit record (“length of credit history”). Table 1 describes the final dataset. About 20 percent of the sample received the maximum Pell Grant in at least one year during our period of study, and nearly 40 percent received a Pell Grant at least once. The latter figure is higher than statistics on Pell Grant receipt among college students in 2003–04, which indicate take-up of 27.2 percent.27 Our period of study covers at least one economic downturn, so a discrepancy between our average and need during a healthier year 25 We also code a binary measure (“no SAT school”) that takes a value of 1 if an individual’s first school does not require standardized tests for admissions. 26 For some individuals, we observe a graduation date, but no degree reported. This group is labeled as “graduated but degree unknown” in Table 1, but, for practical purposes, it has been included in the group of certificate or associate degree holders in the regression analysis. 27 Statistics generated using NCES Quickstats tool for 2003–04 NPSAS. 16

for the economy is unsurprising. Turning to our other, more static measure of background, we see that about 40 percent of fathers and just over 30 percent of mothers have completed a B.A.. These statistics are roughly in line with what national estimates imply; for instance, according to the DoEd, about 40 percent of undergraduates in 2003–04 had at least one parent who earned a B.A. In addition, within our sample, the average SAT score is 1014—ranging from 430 to a perfect 1600—almost exactly corresponding to published statistics for the full population of SAT-takers around our timing.28 The average SAT score for the 1996–97 graduating cohort was 1016. The similarities between the statistics we can produce from our data and published statistics on parental education among college students and SAT-scoring among test-takers lend credence to the national representativeness of our sample in describing college students. Parental education among college students is higher than parental education among all children.29 (For instance, in 2005, 25.5 percent of mothers and 29.7 percent of fathers of children aged 6 to 18 had earned at least a B.A.30) Similarly, the average credit score in our final sample is 639, which is well above the threshold we use for credit scores in the prime range. Indeed, about 68 percent of our sample meets our definition of prime borrowers. These statistics suggest that our sample is positively selected from the population, but also imply an association between background and credit health.31 The remainder of our analysis explores this relationship. 28 U.S. Department of Education, National Center for Education Statistics (2015). Digest of Education Statistics, 2013 (NCES 2015-011), Table 226.10. 29 SAT-taking is very highly correlated with a student’s family background, as admissions test-taking generally reflects an aspiration to attend college and such aspirations are typically higher among students from high-SES families. According to a study by the DoEd of a sample of 1992 high school graduates, the admissions test-taking rate was more than two times higher when at least one of a graduate’s parents completed a B.A.. Even among graduates who indicated in 10th grade that they planned to pursue a B.A. (at least twice as common among graduates with more educated parents), the fraction of students who went on to take an exam was about 25 percent higher when they had more educated parents. See http://files.eric.ed.gov/fulltext/ED546120.pdf. 30 See http://nces.ed.gov/pubs2007/minoritytrends/tables/table_5.asp#sthash.yxGEbobp.dpuf. 31 See Appendix Figure 1 for full distribution. 17

IV. Family Background and Credit Health a. Basic Relationship We begin by examining the simple reduced-form relationship between family background and early-career credit health. Figures 1a and 1b plot the distribution of credit scores in 2008 according to our binary measures of SES. Each indicate that children from higher-SES backgrounds have higher credit scores. Next, we generate a regression-adjusted correspondence. Specifically, we estimate: , (1) 𝑐𝑐𝑖𝑖𝑖𝑖 = 𝛽𝛽0 +𝛽𝛽1∗𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑖𝑖𝑖𝑖 +𝛿𝛿𝑖𝑖 +𝜀𝜀𝑖𝑖𝑖𝑖 where is one of our two measures of 2008 credit access, is one of four 𝑐𝑐𝑖𝑖𝑖𝑖 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑖𝑖𝑖𝑖 family background indicators, i denotes a college-bound individual, and y denotes a graduation year, whereby is a high school graduation year effect that absorbs fixed differences between 𝛿𝛿𝑖𝑖 cohorts. represents the association between family background and credit health. Acro𝛽𝛽ss 1 the board, children from higher-SES backgrounds tend to have higher credit scores, though the extent varies by measure (Table 2). For example, if an individual’s father did not earn a B.A., her credit score, on average, is about 80 points lower (nearly one-half a standard deviation) (column 2), while receiving the maximum Pell Grant is associated with a 120 point lower score (two-thirds of a standard deviation) (column 4). Turning to our binary credit measure, children from better backgrounds appear to have greater access to credit: they are about 14 to 26 percentage points more likely to be prime borrowers than their peers (columns 5–8). Finally, if we examine credit outcomes in 2014 instead of 2008 (i.e., credit scores when the individuals in our sample are in their mid-30s), results are qualitatively similar on both dimensions (Appendix Table 4). 18

Before proceeding, we make two notes regarding the interpretation of First, there are many factors that potentially correlate with both family background and cr𝛽𝛽e 1 d.i t health. Because the omission of such factors from equation (1) could introduce bias into our estimates, may not represent the causal effect of background on credit outcomes. Some of these factors𝛽𝛽 c 1 a nnot be directly observed but can be approximated in our data; for instance, the SAT score is arguably an ample proxy for student achievement. (SAT scores are particularly well-suited for our analysis because performance on the SAT explicitly affects college admissions determinations, and thus opportunity sets.) Some, however, are harder-to-quantify characteristics for which there are no good proxies in our data (e.g., grit; conscientiousness; motivation). Generally, most factors that positively correlate with childhood socioeconomic status would also positively correlate with credit, implying that our estimates probably overstate the true relationship.32 Second, the general association between family background and credit scores may be stronger than the one implied by an analysis that examines only SAT-takers. Higher-income students, as a group, take the SAT more frequently than low-income students; thus, low-income students who take the SAT may be positively selected in unobservable ways that could influence their credit scores. For example, low-income SAT-takers may exhibit more grit than high-income SAT-takers. If increased grit is associated with better credit outcomes, the association between background and credit scores that we estimate may be muted relative to the population estimate. Indeed, Appendix Table 1 indicates that, within an analysis of a broader sample that includes all 32 In the next section, we will include intermediate educational and borrowing outcomes to attempt to reduce bias in β ; however, technically, intermediate outcomes are choice variables (that is, they likely reflect qualities of 1 individuals we cannot measure) and may, thus, themselves introduce new biases that cannot easily be signed. 19

individuals that have any college record across our data sources and that measures disadvantage using only the Pell indicators, estimates of are larger than those in Table 2. b. Role of Pre-Collegiate Achie𝛽𝛽v 1 e ment As noted above, achievement potentially correlates with both background and credit scores. Figures 2a and 2b plot SAT score distributions according to family background, which indicate that the distributions are consistently bell-shaped but also are clearly left-shifted for low-SES students.33 The mean SAT score for low-SES students is about 100 points lower. (This difference suggests that, all else equal, applicants from disadvantaged backgrounds to need-blind colleges are less admissible than their peers.) Per the latter, Figure 3 displays SAT scores for prime and subprime borrowers. The distribution of scores among subprime borrowers is left-shifted relative to prime borrowers but also left-skewed. These patterns underscore the inclusion of SAT scores to reduce bias in .34 Thus, Equatio𝛽𝛽n 1 (1) becomes: (2). 𝑐𝑐𝑖𝑖𝑖𝑖 = 𝛽𝛽0 +𝛽𝛽1∗𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑖𝑖𝑖𝑖 +𝛽𝛽2∗𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 +𝛿𝛿𝑖𝑖 +𝜀𝜀𝑖𝑖𝑖𝑖 The inclusion of enables a rough estimate of the interplay between background and 𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 achievement—that is, all else equal, the amount of additional SAT points that would be needed to offset the credit effects of coming from a disadvantaged background. 33 Because, after holding background constant, test-takers appear to be drawn from similar distributions of test scores, comparisons in distributional outcomes are likely valid (since the shapes of the distributions are comparable once the level effect is removed). While for brevity, we present figures only for “dad” and the “any_pell” measures of SES throughout this section, the graphs look very similar for the measures that we exclude. 34 Also, we might expect this inclusion to absorb some other potentially important unobservables, to the extent that high-scoring students from low-SES backgrounds have more grit than those from high-SES backgrounds. 20

In each specification of equation (2), is positive and significant such that, all else equal, individuals with higher SAT scores tend t𝛽𝛽o 2 have higher credit scores (Table 3).35 Additionally, once achievement is controlled for, the association between family background and credit decreases but remains highly significant. In particular, having less educated parents is associated with a 25 to 41 point credit score reduction, and receiving a Pell Grant is associated with a credit score reduction of as much as 82 points. Combining the information in and , if a college student receives a Pell Grant, she would need an additional 200 to 300 S𝛽𝛽A 1 T po𝛽𝛽i 2 nts to ultimately have a credit score in line with her counterpart without a Pell Grant. The estimate of is quite stable across the specifications with different SES measures. 𝛽𝛽2 Estimates are qualitatively similar using the binary measure of whether an individual could qualify for most types of credit.36 Similar-achieving college students with less educated parents are about 5 to 10 percentage points more likely to be subprime. Using the Pell measures, this figure is closer to 15 to 20 percentage points. An individual generally needs at least 100 additional SAT points to compensate for her background to end up on even footing with her higher-SES peers. V. Extensions This section extends our analysis in two ways. First, we allow the role of achievement to vary by background. Then, we examine the extent to which other factors observable in our data (e.g., attainment; borrowing; length of credit history) explain the credit gap. a. Differential “Returns to Achievement” 35 These results are robust to the inclusion of higher-order polynomials in achievement (Appendix Table 5). 36 Results are again similar for 2014 credit outcomes as well (Appendix Table 4). 21

The primary goal of this exercise is to examine how meaningful achievement differences are for students from different backgrounds.37 In other words, to what extent does the credit gap narrow (or widen) when students move along the achievement distribution? We augment equation (2) to include an interaction between SES and SAT score: , (3) 𝑐𝑐𝑖𝑖𝑖𝑖 = 𝛽𝛽0 +𝛽𝛽1∗𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑖𝑖𝑖𝑖 +𝛽𝛽2∗𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 +𝛽𝛽3 ∗𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑖𝑖𝑖𝑖 ∗𝑆𝑆𝑆𝑆𝑆𝑆𝑖𝑖𝑖𝑖 +𝛿𝛿𝑖𝑖 +𝜀𝜀𝑖𝑖𝑖𝑖 whereby the estimate of reflects the differential change in credit scores for low-SES students associated with every 10𝛽𝛽0 3 point increase in SAT scores. If is large, relatively small differences in SAT scores would imply large changes in the𝛽𝛽 3 credit gap. Regression results reveal that is not particularly large but always positive (Table 4). The coefficients are largest (and statist𝛽𝛽ic 3 ally significant) under the “any pell” and “dad” specifications, and the same patterns hold whether prime status or credit scores are on the lefthand side.38 Altogether, the estimates imply that credit scores among low-SES students are a bit more sensitive to achievement—e.g., in column 3, 100 additional SAT points are associated with nearly a 35-point credit score gain among Pell Grant recipients (compared to about 25 points among those who do not receive a Pell Grant)—but also that a gap exists even in very high SAT 37 Whether SAT score differences are more or less meaningful for low-SES students is theoretically ambiguous. On one hand, high-SES students likely already have a safety net and support network and are more financially literate, so that achievement alone may have little influence on their financial health. Further, high SAT scores may be more useful in expanding opportunity sets for disadvantaged students. Finally, if high SAT scores are harder earned for low-SES students (who may have less access to SAT prep classes or increased obligations at home), we might expect SAT scores to be a better early signal of later-life successes for that group. On the other hand, the literature on under-matching finds that low-SES students are less likely to pursue the educational opportunities that higher SAT scores offer (so their actions and choices that could influence their credit health may be less likely to reflect achievement differences), which could imply that SAT score differences are less meaningful for low-SES students. 38 Behrman and Rosenzweig (2002) compare the schooling outcomes of children of twin mothers and twin fathers (with different levels of education) and find that a child’s outcomes are more strongly associated with his father’s attainment than his mother’s. Black, Devereux, and Salvanes (2005) find a similar relationship, analyzing compulsory schooling law changes in Norway, but surmise that these associations (at least in their setting) are driven by selection rather than causation. 22

ranges—the coefficient on background remains significant across all specifications and is generally at least an order of magnitude larger than . Figures 4a and 4b offer a graphical depiction of 𝛽𝛽th 3 ese results. The graphs suggest that, relative to individuals with very low SAT scores (i.e., two standard deviations below the mean), those with very high SAT scores (i.e., two standard deviations above the mean) have much more similar credit outcomes—the gap narrows by as much as 60 points (more than 60 percent) using “any pell,” and by as much as 30 points (more than 40 percent) using “dad.” Nonetheless, even within very high SAT score ranges, a gap remains. b. Potential Mediators Young adults from different backgrounds may differ along other potentially important dimensions that could correlate with credit outcomes. Our goal in this section is to examine if background maintains predictive power for credit scores, after accounting for factors that may mediate the relationship.39 In the analyses that proceed, we introduce the following (sets of) variables sequentially: (1) whether an individual took a federal student loan for college; (2) college quality/characteristics (an individual’s cumulative federal undergraduate loans, the average enrollee’s later income, the average enrollee’s SAT score, admissions test requirements); (3) educational attainment (degree, years in school); and (4) debt management (length of credit file, whether an individual defaulted on a student loan, whether an individual has been delinquent on non-student loan debt). We begin with a binary measure of whether an individual took a federal student loan for college, as such loans—unlike other forms of debt—do not entail credit underwriting or risk 39 The appendix describes the OLS relationship between background and the metrics considered in this section. 23

pricing. Upon inclusion of this metric, the credit score gap shrinks substantially, by about 40 points, as does the gap in the probability of being a prime borrower, by about 10 percentage points (Table 5). However, across the board, the coefficient on disadvantage remains statistically and economically significant. Specifically, holding ability and borrowing constant, low-SES individuals have credit scores that are 21 to 74 points lower and are 4 to 16 percentage points less likely to be prime borrowers than their high-SES peers. Interestingly, the results also point to a separate, inverse relationship between this type of borrowing and credit scores.40 We next include measures of college quality (Tables 6a and 6b, columns (1), (4), (7), and (10)).41 Unsurprisingly, since attending a higher-quality college is positively associated with credit scores, and, because low-SES students tend to go to lower-quality schools, accounting for quality decreases the credit gap, but only slightly. Then, columns (2), (5), (8), and (11) add educational attainment measures.42 As low-SES students tend to have lower attainment, the gap shrinks more, though generally remains statistically significant. (The exception is Table 6b, column (2)). As a final exercise, we add key elements from early credit histories that capture an individual’s experience managing debt (columns (3), (6), (9), and (12)). Even upon the inclusion of these elements, all of which are very likely to be inputs into the calculation of credit scores, 40 This statistical relationship might be partly driven by students from low-SES backgrounds borrowing greater sums of money or being more likely to default on such debt (Appendix Table 6), though the coefficient on undergraduate borrowing remains negative and highly significant in all of the regressions reported in Tables 6a and 6b (results available upon request). Mezza et al. (2016) find that increased student loan debt raises the probability of having poor credit. 41 In addition, we introduce an indicator variable that takes a value of 1 if the individual never pursued postsecondary education. 42 Degree attainment is entered as a series of dummy variables for various degree categories (with those who leave school without a degree serving as the omitted category). 24

the coefficient on disadvantage generally is statistically significant and continues to point to differences in scores by background. In sum, the inclusion of a full battery of education and debt management metrics is unable to fully eliminate the credit gap. There are two implications from this exercise, which together suggest that the credit market amplifies the transmission of economic well-being across generations. First, even holding many mediating factors constant, the borrowing costs (and opportunities in other markets that rely upon credit scores) faced by adults from disadvantaged backgrounds are relatively unfavorable. Second, because a clear difference in scores persists throughout these exercises—including in the most comprehensive specification that attempts to account for highly predictive elements of risk from credit scoring models—family background might also have predictive power in credit-scoring models. VI. Conclusion Prior work has documented that children’s economic opportunities and outcomes reflect their parents’ socioeconomic status. Some of this persistence invariably owes to immutable aspects of the household environment. Still, to the extent that a central policy goal is to reduce inequality of opportunity, identifying early differences in important outcomes could expose new areas for corrective policy. Our analysis estimates a gap in credit health that emerges early in the lifecycle and remains upon inclusion of an array of achievement, postsecondary education, and debt management metrics. The gap exists whether we measure credit health using raw credit scores or a summary measure derived from these scores that approximates credit access. Because of the many settings in which an individual’s credit health is a key ingredient in assessing her risk type, these early differences could be contributing to overarching socioeconomic divides; thus, our findings reveal a new area for intervention that may help level the playing field. 25

Recall that an individual’s credit score is derived using the default outcomes of similar consumers, which credit scoring models estimate based on elements of credit records. These characteristics include: payment history, amount owed, length of history, new credit, and types of credit used. Importantly, they exclude demographic information. So, how could such a gap emerge? Below we describe five non-exclusive possibilities, with vastly different policy implications. The first three reflect demand-side differences in interactions with credit markets, each of which can be supported with newly released data from the 2016 SCF, and the final two stem from supply-side factors.43 First, individuals from disadvantaged backgrounds may face larger financial headwinds (e.g., fewer avenues through which to build healthy credit, larger shocks to their finances, fewer resources to weather financial shocks). Indeed, within the SCF, this group is approximately 20 percentage points less likely to have access to financial assistance from friends or relatives. In addition, they are less likely to receive an inheritance, and they are more likely to have taken a payday loan over the past year. Second, individuals from disadvantaged backgrounds may be less versed in the importance of healthy credit records and may even, as a result, take unadvisable credit risks (e.g., cumulating debt they will be unable to repay).44 Indeed, they score significantly lower on a set of financial literacy questions asked within the SCF. 43 For these analyses, we describe results for individuals 27-35 years old at the time of the survey (in order to match the range of ages over which we examine credit outcomes in the main analysis), but similar patterns hold when we attempt to replicate the cohort instead. (Both SCF samples are restricted to those with at least a high school degree.) We measure family background according to the educational attainment of the mother and father of the survey respondent using the same classifications as the main analysis. 44 The CFPB (2014) found that, with respect to the information credit scoring agencies rely upon, minorities and individuals from low-income households are more likely to be credit invisible or to have unscored credit records than other groups. 26

Third, individuals from disadvantaged backgrounds may have different consumption preferences or attitudes toward risk. In the SCF, they have a shorter planning horizon for their family spending and savings behavior, and are generally less willing to take financial risks. Fourth, and relatedly, elements of the formulas for credit scores may proxy for socioeconomic background as opposed to independently predicting credit performance in a demographically neutral environment (i.e., where demographics are controlled for or where the estimation sample is limited to a single demographic group).45 For example, one type of credit might highly correlate with default in credit scoring models, but the utilization of such credit may reflect a cultural norm within disadvantaged communities. Finally, if discriminatory lending practices restrict certain groups’ ability to access credit, these groups may have a more difficult time accumulating a strong credit history, which could then affect their scores.46 Future research should explore which of the above mechanisms underlie the early gaps in credit health we detect and the effectiveness of policies in ameliorating them. In particular, a key question is whether the differences in credit scores that we document by socioeconomic group stem solely from the underlying default risk of different household types or are partially an unintended artifact of how credit scores are constructed. 45 Relatedly, CFPB (2015) found that the identical treatment of medical and non-medical collections that was being employed by credit scoring agencies was not justified by subsequent debt payment patterns. 46 Recent studies that have examined the extent to which there is evidence of race-based redlining—an illegal practice whereby residents of certain geographic areas are not given the same access to credit as similarly-qualified residents of other areas and a central concern among regulators of the mortgage industry—yield mixed results (Ethan Cohen-Cole, Brevoort, 2011). Some banks—e.g., Hudson City Savings Bank, BancorpSouth Bank—have been fined millions of dollars in relation to discriminatory mortgage lending practices. See, for example, http://www.consumerfinance.gov/about-us/newsroom/cfpb-and-doj-order-hudson-city-savings-bank-to-pay-27million-to-increase-mortgage-credit-access-in-communities-illegally-redlined/ or http://www.consumerfinance.gov/about-us/newsroom/consumer-financial-protection-bureau-and-departmentjustice-action-requires-bancorpsouth-pay-106-million-address-discriminatory-mortgage-lending-practices/. 27

Work Cited Avery, R., K. Brevoort, and G. Canner (2012). Does Credit Scoring Produce a Disparate Impact? Real Estate Economics: 40(1), 65-114. Bayle, M. and S. Dynarski (2011). Gains and Gaps: Changing Inequality in U.S. College Entry and Completion. NBER Working Papers 17633. Behrman, J., and M. Rosenzweig (2002). Does increasing women’s schooling raise the schooling of the next generation? American Economic Review: 92, 323-334. Black, S., K. Cortes, and J. Lincove (2015). Academic Undermatching of High-Achieving Minority Students: Evidence from Race-Neutral and Holistic Admissions Policies. American Economic Review papers and proceedings: 105(5), 604-610. Black, S., P. Devereux, and K. Salvanes (2005). Why the Apple Doesn't Fall Far: Understanding Intergenerational Transmission of Human Capital. American Economic Review: 96(1), 437-449. Black, S., and P. Devereux (2011). Recent Developments in Intergenerational Mobility, in Handbook of Labor Economics, Orley Ashenfelter and David Card, editors, North Holland Press, Elsevier. Black, D., and J. Smith (2006). Estimating the Returns to College Quality with Multiple Proxies for Quality. Journal of Labor Economics: 24(3), 701–728. Bleemer, Z., M. Brown, D. Lee, and W. van der Klaauw (2014). Debt, Jobs, or Housing: What’s Keeping Millennials at Home? Federal Reserve Bank of New York Staff Reports, no. 700. Board of Governors (2007). Report to the Congress on Credit Scoring and Its Effect on Availability and Affordability of Credit. Federal Reserve Board: Washington, DC. 28

Bowen, W., M. Chingos, and M. McPherson (2009). Crossing the Finish Line: Completing College at America’s Public Universities. Princeton, NJ: Princeton University Press. Bowles, S., and H. Gintis (2002). The inheritance of inequality. Journal of Economic Perspectives: 16, 3-30. Breland, H., J. Maxey, R. Gernand, T., Cumming, and C. Trapani (2002). Trends in use college admission 2000: A report of a survey of undergraduate admissions policies, practices, and procedures. Sponsored by ACT, Inc., Association for Institutional Research, The College Board, Educational Testing Service, and the National Association for College Admission Counseling. Brevoort, K. (2011). Credit card redlining revisited. The Review of Economic and Statistics, 93(2), 714-724. Bulman, G., R. Fairlie, S. Goodman, and A. Isen (2016). Parental Resources and College Attendance: Evidence from Lottery Wins. NBER Working Papers 22679. Card, D. (1995). Using Geographic Variation in College Proximity to Estimate the Return to Schooling. Aspects of Labour Economics: Essays in Honour of John Vanderkamp, edited by L. Christofides, E. K. Grant, and R. Swindinsky. University of Toronto Press, 1995. CFPB (2014). Data point: Medical Debt and Credit Scores. http://files.consumerfinance.gov/f/201405_cfpb_report_data-point_medical-debt-creditscores.pdf CFPB (2015). Data Point: Credit Invisibles. http://files.consumerfinance.gov/f/201405_cfpb_report_data-point_medical-debtcredit-scores.pdf 29

Chetty R., J. Friedman, N. Hilger, E. Saez, D. Whitmore Schanzenbach, and D. Yagan (2011). How Does Your Kindergarten Classroom Affect Your Earnings? Evidence from Project Star, The Quarterly Journal of Economics 126(4):1593-1660. Chetty, R., N. Hendren, P. Kline, E. Saez, and N. Turner (2014). Is the United States Still a Land of Opportunity? Recent Trends in Intergenerational Mobility, American Economic Review 104(5): 141-47. Clark, M., J. Rothstein, and D. Schanzenbach (2009). Selection Bias in College Admissions Test Scores. Economics of Education Review: 28(3), 295-307. Cohen-Cole, E. (2011). Credit Card Redlining. The Review of Economics and Statistics, 93(2), 700-713. Cohodes S. and J. Goodman (2012, August). First Degree Earns: The Impact of College Quality on College Completion Rates. Working Paper RWP12-033. Harvard Kennedy School Faculty Research Working Paper Series. Dale, S. and A. Krueger (2002). Estimating the Payoff to Attending a More Selective College: An Application of Selection on Observables and Unobservables. Quarterly Journal of Economics: 117(4), 1491–1527. Dale, S. and A. Krueger (2011, June). Estimating the Return to College Selectivity over the Career Using Administrative Earnings Data. Working Paper 17159, NBER. Dettling, L. and J. Hsu (2014). Returning to the Nest: Debt and Parental Co-Residence Among Young Adults. Finance and Economics Discussions Series 2014-80. Washington: Board of Governors of the Federal Reserve System. 30

Ghent, A. and M. Kudlyak (2015). Intergenerational Linkages in Household Credit. Federal Reserve Bank of Richmond Working Paper Series. Herkenhoff, K. (2015). "The Impact of Consumer Credit Access on Unemployment." Unpublished manuscript. Herkenhoff, K., G. Phillips, and E. Cohen-Cole (2016). How Credit Constraints Impact Job Finding Rates, Sorting & Aggregate Output. NBER Working Paper 22274. Hilger, N. (2016). Parental Job Loss and Children’s Long-Term Outcomes: Evidence from 7 Million Fathers’ Layoffs. American Economic Journal: Applied Economics 8(3): 247-283. Hoekstra, M. (2009). The Effect of Attending the Flagship State University on Earnings: A Discontinuity-Based Approach. The Review of Economics and Statistics: 91(4), 717-724. Hoxby, C. (2009). The Changing Selectivity of American Colleges. Journal of Economic Perspectives: 23(4), 95-118. Hoxby, C. and C. Avery (2013). The Missing "One-Offs": The Hidden Supply of High- Achieving, Low-Income Students. Brookings Papers on Economic Activity 46(1), 1-65. Kane, T. (1995, July). Rising Public College Tuition and College Entry: How Well Do Public Subsidies Promote Access to College? Working Paper 5164, NBER. Laufer, S., and A. Paciorek (2016). The Effects of Mortgage Credit Availability: Evidence from Minimum Credit Score Lending Rules. Finance and Economics Discussion Series 2016-098. Washington: Board of Governors of the Federal Reserve System. 31

Lee, C., and G. Solon (2009). Trends in Intergenerational Income Mobility. Review of Economics and Statistics: 91(4), 766-772. Lovenheim, M. and C. Reynolds (2013). The Effect of Housing Wealth on College Choice: Evidence from the Housing Boom. Journal of Human Resources: 48(1), 1-35. Mazumder, B. (2005). Fortunate Sons: New Estimates of Intergenerational Mobility in the United States Using Social Security Earnings Data. The Review of Economics and Statistics: 87(2), 235-255. Mezza, A., and K. Sommer (2016). A Trillion Dollar Question: What Predicts Student Loan Delinquencies? Journal of Student Financial Aid: 46(3), 16-54. Mezza, A., D. Ringo, S. Sherlund, and K. Sommer (2016). Student Loans and Homeownership. Finance and Economics Discussions Series 2016-010r. Washington: Board of Governors of the Federal Reserve System. Mitnik, P., V. Bryant, M. Weber, and D. Grusky (2015). New Estimates of Intergenerational Mobility Using Administrative Data. SOI working paper. Musto, D. (2004). What Happens When Information Leaves a Market? Evidence From Post- Bankruptcy Consumers. The Journal of Business: 77(4): 725-748. Pallais, A., and S. Turner (2006). Opportunities for Low Income Students at Top Colleges and Universities: Policy Initiatives and the Distribution of Students. National Tax Journal: 59(2), 357-386. 32

Reardon, S. (2011). The Widening Academic Achievement Gap between the Rich and the Poor: New Evidence and Possible Explanations. In R. Murnane & G. Duncan (Eds.) Whither Opportunity? Rising Inequality, Schools, and Children’s Life Chances. New York: Russell Sage Foundation Press. Ringo, D. (2015). Parental Credit Constraints and Child College Attendance. Unpublished manuscript. Smith, J., M. Pender, and J. Howell (2013). The Full Extent of Academic Undermatch. Economics of Education Review: 32, 247-261. Solon, G. (1999). “Intergenerational Mobility in the Labor Market,” in O. Ashenfelter and D. Card, eds., Handbook of Labor Economics: 3A, 1761-800. Spies, R. (2001). The Effect of Rising Costs on College Choice: The Fourth and Final in a Series of Studies on This Subject. Princeton University Research Report Series, No. 117. Zimmerman, S. (2014). The Returns to College Admission for Academically Marginal Students. Journal of Labor Economics: 32(4), 711-754. 33

Note: Displayed are the density functions for 2008 credit scores grouped by disadvantage (measured by dad’s education on the left panel and mom’s on the right). The solid lines reflect the densities for individuals with parents with at least a BA degree, while the dashed lines for parents with less than a college degree. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. 34

Note: Displayed are the density functions for 2008 credit scores grouped by disadvantage (measured by Max Pell on the left panel and Any Pell on the right). The solid lines reflect the densities for individuals who were ever awarded the maximum Pell Grant available in a given year (left) or ever awarded a Pell Grant (right), while the dashed lines for individuals who were never awarded the maximum Pell Grant level (left) or were never awarded a Pell Grant (right). Sample is SAT test-takers matched to the nationally-representative cohort of 23to 31-year-old individuals with credit records in 2004 as described in the main text. 35

Note: Displayed are the density functions for SAT scores grouped by disadvantage (measured by dad’s educational attainment). The solid line reflects the density for individuals with a dad with at least a BA degree, while the dashed line for a dad with less than a college degree. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. 36

Note: Displayed are the density functions for SAT scores grouped by disadvantage (measured by Any Pell). The solid line reflects the density for individuals who were ever awarded a Pell Grant, while the dashed line for individuals who were never awarded a Pell Grant. Sample is SAT test-takers matched to the nationallyrepresentative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. 37

Note: Displayed are the density functions for SAT scores grouped by 2008 credit access. The solid line reflects the density for individuals who are classified as prime, while the dashed line as subprime. Threshold for prime borrower is the median TransUnion credit score in 2008 over the base cohort sample as described in the main text. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. 38

39 002 001 0 001- 002- Figure 4a: Relationship between SAT and Credit Scores by Dad's Ed. residualized credit score -500 0 500 residualized SAT score BA Plus Less than BA Note: Graph plots residuals after netting out cohort effects and a constant.

Table 1: Summary Statistics Std. Variables Obs Mean Dev. Min Max dad 4,790 0.563 0.496 0 1 mom 4,867 0.650 0.477 0 1 any_pell 5,421 0.368 0.482 0 1 max_pell 5,421 0.193 0.395 0 1 maxsat100 5,421 10.1 2.1 4.3 16 tuscore2008 5,421 639.3 183.0 271 894 prime 5,421 0.679 0.467 0 1 no school 5,421 0.056 0.230 0 1 borrowed ug 5,421 0.578 0.494 0 1 amount borrowed for ug 5,421 12.602 17.147 0 154.064 defaulted 5,421 0.086 0.280 0 1 school's mean income 4,945 36.94 14.45 7.6 134 sat school 5,421 7.2 5.4 0 14.88 no sat school 5,421 0.352 0.478 0 1 dropout 5,421 0.478 0.500 0 1 graduated but degree unknown 5,421 0.218 0.413 0 1 certificate/associate's 5,421 0.031 0.173 0 1 bachelor's 5,421 0.235 0.424 0 1 master's or more 5,421 0.037 0.189 0 1 years in school 4,736 4.282 2.136 0.01 13.35 delinquent on nonstudent loan debt 5,421 0.496 0.500 0 1 length of credit history 5,421 132.9 25.9 16 421 graduation year: 1995 5,421 0.14 0.35 0 1 graduation year: 1996 5,421 0.17 0.38 0 1 graduation year: 1997 5,421 0.17 0.38 0 1 graduation year: 1998 5,421 0.18 0.39 0 1 graduation year: 1999 5,421 0.18 0.38 0 1 40

Table 2. Family Background and Credit Health Credit Score Prime Borrower? mom dad any_pell max_pell mom dad any_pell max_pell Disadvantage -63.92*** -78.59*** -93.60*** -119.1*** -0.138*** -0.177*** -0.210*** -0.259*** (5.390) (5.150) (4.991) (6.085) (0.0138) (0.0133) (0.0128) (0.0157) Constant 692.5*** 697.8*** 682.0*** 671.8*** 0.791*** 0.810*** 0.771*** 0.747*** (7.505) (7.151) (6.442) (6.274) (0.0193) (0.0184) (0.0166) (0.0162) Observations 4,867 4,790 5,421 5,421 4,867 4,790 5,421 5,421 R-squared 0.031 0.050 0.064 0.069 0.022 0.038 0.049 0.050 Note: Displayed are coefficients from a simple regression of a measure of 2008 credit health (denoted by the super column header) on a measure of disadvantage (denoted by the subcolumn header). Threshold for prime borrower is median TU credit score in 2008 (i.e., 580.5).Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. Regressions include graduation year effects. *** denotes significance at 1%. 41

Table 3: Family Background, Achievement, and Credit Health Credit Score Prime Borrower? mom dad any_pell max_pell mom dad any_pell max_pell Disadvantage -25.25*** -41.15*** -65.72*** -82.05*** -0.0497*** -0.0927*** -0.147*** -0.176*** (5.350) (5.189) (4.846) (5.956) (0.0139) (0.0135) (0.0126) (0.0156) maxsat100 29.90*** 28.09*** 28.82*** 28.38*** 0.0682*** 0.0636*** 0.0643*** 0.0637*** (1.252) (1.270) (1.138) (1.144) (0.00326) (0.00331) (0.00297) (0.00299) Constant 480.2*** 500.8*** 494.3*** 489.5*** 0.307*** 0.364*** 0.352*** 0.337*** (11.37) (11.21) (9.596) (9.455) (0.0296) (0.0292) (0.0250) (0.0247) Observations 4,867 4,790 5,421 5,421 4,867 4,790 5,421 5,421 R-squared 0.133 0.138 0.163 0.164 0.103 0.107 0.125 0.124 Note: Displayed are coefficients from a regression of a measure of 2008 credit health (denoted by the super column header) on a measure of disadvantage (denoted by the subcolumn header) and a student’s SAT score. Threshold for prime borrower is median TU credit score in 2008 (i.e., 580.5). Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Sample is SAT testtakers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. Regressions include graduation year effects. *** denotes significance at 1%. 42

Table 4: Family Background, Achievement, and Credit Health (Interacting Background with Achievement) Credit Score Prime Borrower? mom dad any_pell max_pell mom dad any_pell max_pell Disadvantage -33.50* -68.35*** -108.6*** -90.00*** -0.119** -0.174*** -0.254*** -0.191*** (17.78) (17.04) (14.71) (16.80) (0.0462) (0.0444) (0.0384) (0.0439) maxsat100 29.11*** 25.72*** 26.16*** 28.10*** 0.0615*** 0.0565*** 0.0577*** 0.0632*** (2.059) (1.899) (1.427) (1.267) (0.00535) (0.00495) (0.00372) (0.00331) Disadvantage × 1.262 4.277* 7.300*** 1.493 0.0105 0.0127* 0.0181*** 0.00285 maxsat100 (2.593) (2.553) (2.363) (2.950) (0.00674) (0.00666) (0.00616) (0.00770) Constant 485.9*** 517.5*** 511.5*** 491.2*** 0.354*** 0.414*** 0.395*** 0.341*** (16.23) (14.98) (11.10) (10.09) (0.0422) (0.0391) (0.0290) (0.0264) Observations 4,867 4,790 5,421 5,421 4,867 4,790 5,421 5,421 R-squared 0.133 0.139 0.165 0.164 0.104 0.108 0.126 0.124 Note: Displayed are coefficients from a regression of a measure of 2008 credit health (denoted by the super column header) on a measure of disadvantage (denoted by the subcolumn header), a student’s SAT score, and the interaction of the two. Threshold for prime borrower is median TU credit score in 2008 (i.e., 580.5). Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. Regressions include graduation year effects. ***, **, * denote significance at 1%, 5%, and 10%. 43

Table 5: Family Background, Achievement, and Credit Health (Including “Whether Individual Took a Federal Undergraduate Student Loan”) Credit Score Prime Borrower? mom dad any_pell max_pell mom dad any_pell max_pell Disadvantage -20.63*** -34.19*** -55.77*** -73.90*** -0.0397*** -0.0778*** -0.128*** -0.159*** (5.290) (5.165) (5.154) (6.038) (0.0138) (0.0135) (0.0135) (0.0158) maxsat100 29.88*** 28.34*** 28.74*** 28.16*** 0.0680*** 0.0640*** 0.0640*** 0.0632*** (1.237) (1.257) (1.138) (1.139) (0.00323) (0.00329) (0.00298) (0.00299) Federal Student Loan -60.12*** -56.19*** -43.97*** -50.48*** -0.130*** -0.120*** -0.0916*** -0.108*** (5.082) (5.112) (5.000) (4.822) (0.0133) (0.0134) (0.0131) (0.0126) Constant 515.6*** 531.0*** 521.0*** 522.6*** 0.385*** 0.430*** 0.410*** 0.409*** (11.67) (11.47) (9.996) (9.910) (0.0305) (0.0300) (0.0262) (0.0260) Observations 4,867 4,790 5,421 5,421 4,867 4,790 5,421 5,421 R-squared 0.158 0.159 0.177 0.182 0.121 0.122 0.134 0.136 Note: Displayed are coefficients from a regression of a measure of 2008 credit health (denoted by the super column header) on a measure of disadvantage (denoted by the subcolumn header), a student’s SAT score, and whether the individual borrowed from the Federal Government for undergraduate studies. Threshold for prime borrower is median TU credit score in 2008 (i.e., 580.5). Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. Regressions include graduation year effects and a dummy variable taking a value of 1 if an individual did not attend postsecondary school. *** denotes significance at 1%. 44

Table 6a: Family Background, Achievement, and Credit Scores (Sequentially Adding Covariates) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) mom dad any_pell max_pell Disadvantage -16.39*** -10.09* -5.815 -29.81*** -20.64*** -11.05** -52.47*** -42.14*** -25.97*** -71.50*** -62.22*** -41.85*** (5.316) (5.200) (4.388) (5.214) (5.126) (4.323) (5.204) (5.242) (4.444) (6.105) (6.174) (5.243) maxsat100 24.80*** 20.00*** 13.26*** 23.76*** 19.25*** 13.02*** 23.65*** 19.42*** 13.03*** 22.89*** 18.58*** 12.40*** (1.414) (1.411) (1.200) (1.429) (1.430) (1.213) (1.318) (1.324) (1.128) (1.320) (1.327) (1.129) Federal Student Loan x x x x x x x x x x x x College Quality x x x x x x x x x x x x Attainment x x x x x x x x Credit History x x x x Constant 501.4*** 482.9*** 608.9*** 518.9*** 497.6*** 616.9*** 510.4*** 486.6*** 607.8*** 511.7*** 488.8*** 608.2*** (12.50) (12.78) (19.61) (12.36) (12.63) (19.55) (10.85) (11.09) (18.38) (10.73) (10.97) (18.29) Observations 4,723 4,439 4,439 4,649 4,371 4,371 5,249 4,939 4,939 5,249 4,939 4,939 R-squared 0.168 0.228 0.452 0.168 0.226 0.452 0.185 0.237 0.457 0.191 0.243 0.460 Note: Displayed are coefficients from a regression of 2008 credit score on a measure of disadvantage (denoted by the column header) and a student’s SAT score, upon sequential inclusion of other factors that might mediate the relationship between background and credit as described in the main text. Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Debt measured in thousands and real adjusted to 2008 dollars. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. Regressions include graduation year effects. ***, **, and * denote significance at 1%, 5%, and 10%. 45

Table 6b. Family Background, Achievement, and Prime Borrower Status (Sequentially Adding Covariates) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) mom dad any_pell max_pell Disadvantage -0.0294** -0.0127 -0.00383 -0.0670*** -0.0433*** -0.0224* -0.119*** -0.0905*** -0.0548*** -0.153*** -0.127*** -0.0823*** (0.0139) (0.0138) (0.0123) (0.0137) (0.0136) (0.0122) (0.0136) (0.0139) (0.0125) (0.0160) (0.0164) (0.0148) maxsat100 0.0564*** 0.0458*** 0.0310*** 0.0536*** 0.0438*** 0.0301*** 0.0519*** 0.0429*** 0.0290*** 0.0506*** 0.0413*** 0.0279*** (0.00370) (0.00374) (0.00338) (0.00374) (0.00380) (0.00342) (0.00345) (0.00351) (0.00317) (0.00346) (0.00352) (0.00319) Federal Student Loan x x x x x x x x x x x x College Quality x x x x x x x x x x x x Attainment x x x x x x x x Credit History x x x x Constant 0.353*** 0.308*** 0.560*** 0.403*** 0.352*** 0.588*** 0.385*** 0.328*** 0.567*** 0.383*** 0.330*** 0.567*** (0.0327) (0.0339) (0.0552) (0.0324) (0.0336) (0.0552) (0.0284) (0.0294) (0.0517) (0.0281) (0.0291) (0.0516) Observations 4,723 4,439 4,439 4,649 4,371 4,371 5,249 4,939 4,939 5,249 4,939 4,939 R-squared 0.128 0.175 0.340 0.128 0.172 0.339 0.141 0.179 0.340 0.143 0.182 0.342 Note: Displayed are coefficients from a regression of 2008 prime status on a measure of disadvantage (denoted by the column header) and a student’s SAT score, upon sequential inclusion of other factors that might mediate the relationship between background and credit as described in the main text. l. Threshold for prime borrower is median TU credit score in 2008 (i.e., 580.5). Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Debt measured in thousands and real adjusted to 2008 dollars. Sample is SAT testtakers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. Regressions include graduation year effects. ***, **, and * denote significance at 1%, 5%, and 10%. 46

Appendix Appendix Table 6 presents regression results of alternative specifications of the dependent variable— —in equation (2), relating to postsecondary education and debt management. 𝑐𝑐𝑖𝑖𝑖𝑖 Panels (a) to (d) present results for each background measure in succession. Columns (1) to (3) indicate that, holding achievement constant, students from low-SES backgrounds are more likely to borrow from the federal government to fund their undergraduate studies, borrow more federal money, and are more likely to default on this debt.1 Students with less educated parents are about 10 percentage points more likely to borrow and borrow almost $3,000 more. (Using the Pell measures, these figures are closer to 15 percentage points and $5,000.2) Students with less educated parents are about 1.5 to 5 percentage points more likely to default, while those receiving Pell Grants are nearly 10 percentage points more likely to.3 Columns (4) to (6) present evidence consistent with under-matching. Column (4) indicates that students from low-SES backgrounds are more likely to attend colleges associated with lower earnings. For example, children with mothers with less than a B.A. attend schools where students make, on average, $1,770 less, compared to equally able children with more educated mothers. Additionally, children from low-SES backgrounds are more likely to attend less selective schools (columns (5) and (6)).4 1 The exception is students with less educated mothers, for whom the estimate on amount borrowed is not statistically significant. 2 The tighter link between background and borrowing when we measure background with the Pell Grant measures (compared with our parental education measures) is almost tautological, as students who qualify for Pell Grants are eligible for more subsidized loans. If borrowing is correlated with lower credit scores, this relationship could help explain why the effects we detect throughout the paper are stronger for the Pell Grant measure than the parental education measure. 3 Results in columns (1) through (3) hold if we replace the borrowing and default measures of federal student loans for undergraduate studies by total student loan borrowing and higher order delinquency measures for all postsecondary education. 4 Only students who attend schools that report (and thus tend to require) SAT scores are included in this model. Thus, this subsample is a subset of college students who attended above-average quality schools. 47

Columns (7) and (8) investigate differences in attainment. Column (7) shows that students from less affluent backgrounds are also significantly more likely to leave school without a degree. In particular, students with less educated parents are about 9 percentage points more likely to drop out, while students receiving Pell Grants are 6 percentage points more likely to drop out. Column (8) presents mixed evidence of years spent in school by background. Although children with less educated mothers are in school about one-quarter of a year less than those with more educated mothers, there is no difference by fathers’ education. Additionally, children who receive Pell Grants spend more than half a year more in school, which could be an artifact of the subsidy they are receiving to attend. Finally, columns (9) and (10) examine whether an individual’s background correlates with the likelihood she has been delinquent on (non-student loan) debt and how early her credit record began (measured in months). Those from disadvantaged backgrounds are significantly more likely to have experienced a delinquency, with the difference ranging from about 5 to 10 percentage points. The relationship between background and the length of one’s credit history is theoretically ambiguous. On the one hand, more affluent parents may be more likely to help their children build their credit files (by opening credit accounts for them). On the other hand, low- SES undergraduate students are more likely to take student loans, which might establish their credit records relatively early. The estimates are highly significant and imply that low-SES individuals have slightly longer credit histories (i.e., about two months, on average). 48

Note: Displayed are the density functions for a measure of 2008 credit health (TU credit score) for SAT test-takers (solid line) and for the full base cohort, including non-SAT test-takers (dashed line). 49

Appendix Table 1: Estimated Relationship between Family Background and 2008 Credit Health, All College-Goers Credit Access Score Prime any_pell -113.2*** -0.255*** (2.551) (0.00645) Constant 621.1*** 0.642*** (4.012) (0.0102) Observations 22,162 22,162 R-squared 0.085 0.067 Note: Displayed are coefficients from a simple regression of a measure of 2008 credit health (denoted by super column header) on Pell Grant receipt. Threshold for prime borrower is median TU credit score in 2008 (i.e., 580.5). Sample is all individuals with 2004 credit records in our sample that can be matched to college-going activity (i.e., any individual with a record in the National Student Loan Data System, the National Student Clearinghouse, or information on student loan originations to fund postsecondary education available in TU) as described in the main text. Regressions include year of birth effects. *** denotes significance at the 1%. 50

Appendix Table 2. Family Background and Credit Health, SAT States Credit Score Prime Borrower? mom dad any_pell max_pell mom dad any_pell max_pell Disadvantage -66.45*** -79.72*** -91.11*** -117.5*** -0.141*** -0.174*** -0.206*** -0.260*** (6.237) (5.923) (5.669) (6.874) (0.0161) (0.0153) (0.0146) (0.0177) Constant 686.2*** 691.4*** 672.8*** 662.8*** 0.776*** 0.792*** 0.752*** 0.729*** (8.613) (8.211) (7.316) (7.110) (0.0222) (0.0212) (0.0188) (0.0183) Observations 3,815 3,752 4,274 4,274 3,815 3,752 4,274 4,274 R-squared 0.032 0.050 0.060 0.067 0.023 0.036 0.046 0.050 Note: Displayed are coefficients from a simple regression of a measure of 2008 credit health (denoted by the super column header) on a measure of disadvantage (denoted by the subcolumn header). Threshold for prime borrower is median TU credit score in 2008 (i.e., 580.5). Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text, restricted further to test-takers from states where the SAT was the dominant exam (Clark, Rothstein, and Schanzenbach, 2009). Regressions include graduation year effects. *** denotes significance at 1%. 51

Appendix Table 3. Family Background and Prime Borrower Status, Alternative Prime Thresholds A. Mortgage Market Definitions Prime Cutoff from Laufer and Paciorek (2016) Bottom of Mortgage Market TU credit score > 526 TU credit score > 351 mom dad any_pell max_pell mom dad any_pell max_pell Disadvantage -0.111*** -0.149*** -0.194*** -0.253*** -0.0701*** -0.0887*** -0.121*** -0.156*** (0.0130) (0.0124) (0.0120) (0.0147) (0.00975) (0.00932) (0.00915) (0.0112) Constant 0.840*** 0.857*** 0.830*** 0.811*** 0.936*** 0.942*** 0.933*** 0.920*** (0.0180) (0.0172) (0.0155) (0.0151) (0.0136) (0.0129) (0.0118) (0.0115) Observations 4,867 4,790 5,421 5,421 4,867 4,790 5,421 5,421 R-squared 0.018 0.032 0.048 0.055 0.014 0.022 0.034 0.038 B. Current Industry Consensus Definition TU credit score > 620 mom dad any_pell max_pell Disadvantage -0.146*** -0.185*** -0.215*** -0.266*** (0.0144) (0.0138) (0.0134) (0.0163) Constant 0.735*** 0.751*** 0.709*** 0.684*** (0.0201) (0.0192) (0.0172) (0.0168) Observations 4,867 4,790 5,421 5,421 R-squared 0.022 0.038 0.047 0.048 Note: Displayed are coefficients from a simple regression of a measure of 2008 credit health (denoted by the super column header) on a measure of disadvantage (denoted by the subcolumn header). A TU score of 526 corresponds to an Equifax Risk score of 620—a cutoff commonly used by mortgage lenders in applications for credit, especially after 2009 (Laufer and Paciorek, 2016). A TU score of 351 corresponds to an Equifax Risk score of 550—a score at which very few mortgage originations occur, even in 2008. A TU score of 620 approximates the probability of default associated with a FICO score between 680 and 700, a current industry consensus definition of a prime borrower. Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. Regressions include graduation year effects. *** denotes significance at 1%. 52

Appendix Table 4. Family Background and 2014 Credit Health Credit Score Prime Borrower? mom dad any_pell max_pell mom dad any_pell max_pell A. Simple Relationship Disadvantage -62.36*** -76.53*** -85.04*** -108.7*** -0.144*** -0.171*** -0.198*** -0.259*** (5.232) (5.010) (4.880) (5.952) (0.0138) (0.0132) (0.0129) (0.0157) Constant 706.7*** 710.2*** 693.2*** 684.2*** 0.800*** 0.804*** 0.763*** 0.744*** (7.273) (6.943) (6.282) (6.121) (0.0191) (0.0183) (0.0166) (0.0161) Observations 4,806 4,729 5,351 5,351 4,806 4,729 5,351 5,351 R-squared 0.031 0.049 0.055 0.060 0.024 0.036 0.043 0.050 B. Including SAT Score Disadvantage -25.27*** -41.09*** -58.16*** -73.15*** -0.0570*** -0.0876*** -0.135*** -0.176*** (5.205) (5.061) (4.747) (5.835) (0.0138) (0.0135) (0.0127) (0.0156) maxsat100 28.50*** 26.56*** 27.63*** 27.22*** 0.0672*** 0.0629*** 0.0648*** 0.0636*** (1.218) (1.239) (1.115) (1.120) (0.00324) (0.00329) (0.00298) (0.00299) Constant 504.2*** 523.9*** 513.1*** 509.1*** 0.323*** 0.364*** 0.341*** 0.335*** (11.06) (10.93) (9.394) (9.255) (0.0294) (0.0291) (0.0251) (0.0247) Observations 4,806 4,729 5,351 5,351 4,806 4,729 5,351 5,351 R-squared 0.130 0.134 0.153 0.154 0.105 0.105 0.121 0.124 Note: The top panel displays coefficients from a simple regression of a measure of 2008 credit health (denoted by the super column header) on a measure of disadvantage (denoted by the subcolumn header) and the bottom panel from a similar set of regressions that include a student’s SAT score. Threshold for prime borrower is median TU credit score in 2014 (i.e., 597). Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. Regressions include graduation year effects. *** denotes significance at 1%. and a student’s SAT score. 53

Appendix Table 5: Family Background, Achievement, and Credit Scores (with Higher-Order Polynomials in Achievement) Credit Score mom dad any_pell max_pell Disadvantage -26.48*** -26.41*** -41.58*** -41.47*** -65.31*** -65.33*** -80.90*** -80.79*** (5.355) (5.354) (5.185) (5.186) (4.844) (4.842) (5.968) (5.968) maxsat100 48.98*** 19.9 47.04*** 29.32* 45.74*** 18.09 42.12*** 17.38 (5.622) (16.500) (5.705) (16.780) (5.238) (15.560) (5.263) (15.560) maxsat1002 -1.541*** 3.648 -1.519*** 1.627 -1.364*** 3.571 -1.105*** 3.313 (0.443) (2.803) (0.446) (2.837) (0.412) (2.647) (0.413) (2.647) maxsat1003 -0.278* -0.168 -0.265* -0.237* (0.148) (0.150) (0.140) (0.140) Constant 428.4*** 475.4*** 448.2*** 477.1*** 447.3*** 492.1*** 451.1*** 491.1*** (18.730) (31.300) (19.070) (31.980) (17.110) (29.260) (17.170) (29.240) Observations 4,867 4,867 4,790 4,790 5,421 5,421 5,421 5,421 R-squared 0.135 0.136 0.14 0.14 0.165 0.166 0.165 0.166 Note: Displayed are coefficients from a regression of a measure of 2008 credit health (denoted by the super column header) on a measure of disadvantage (denoted by the subcolumn header), a student’s SAT score, and a second- and third-order polynomial in a student’s SAT score, respectively. Threshold for prime borrower is median TU credit score in 2008 (i.e., 580.5). Mom and dad are binary measures of the corresponding parent’s educational attainment (where a value of 0 reflects having a BA), any_pell denotes whether the individual was ever awarded a Pell Grant, and max_pell denotes whether the individual was ever awarded the maximum Pell Grant available in a given year; therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text. Regressions include graduation year effects. ***, **, * denote significance at 1%, 5%, and 10%. 54

Appendix Table 6a: Association between Family Background and Other Covariates, Using Mom’s Educational Attainment to Measure Family Background (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Borrowed Defaulted Avg SAT for Attending a Delinquent Federal UG Cumul Ug on Federal Mean schools school that on (Non- Length of Student Bor--incl. UG Student School reporting doesn’t Years in Student Credit VARIABLES Loan non-bor Loan Income SAT report SAT No Degree School Loan) Debt History mom 0.0829*** 0.919 0.0153* -1.770*** -0.204*** 0.0375*** 0.0887*** -0.255*** 0.0544*** 1.851*** (0.0157) (0.563) (0.00921) (0.408) (0.0388) (0.0140) (0.0153) (0.0689) (0.0158) (0.617) maxsat100 0.000310 0.243* -0.0174*** 3.303*** 0.344*** -0.0728*** -0.0668*** 0.214*** -0.0520*** 0.0458 (0.00370) (0.133) (0.00217) (0.0969) (0.0101) (0.00329) (0.00360) (0.0165) (0.00371) (0.145) Constant 0.573*** 12.15*** 0.208*** 19.29*** 8.958*** 0.706*** 0.771*** 3.023*** 0.829*** 159.3*** (0.0339) (1.213) (0.0198) (0.884) (0.0909) (0.0301) (0.0330) (0.151) (0.0340) (1.329) Observations 4,562 4,562 4,562 4,456 3,185 4,562 4,562 4,259 4,562 4,562 R-squared 0.007 0.004 0.019 0.245 0.310 0.121 0.099 0.055 0.058 0.465 Note: Displayed are coefficients from regressions of various credit-related outcomes (denoted by the column headers) on mom’s education and a student’s SAT score. Mom is a binary measure of the mother’s educational attainment (where a value of 0 reflects having a BA); therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Debt measured in thousands and real adjusted to 2008 dollars. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text, restricted to those who attended post-secondary school. Regressions include graduation year effects. ***, **, * denote significance at 1%, 5%, and 10%. 55

Appendix Table 6b: Association between Family Background and Other Covariates, Using Dad’s Educational Attainment to Measure Family Background (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Defaulted Attending Borrowed Avg SAT Delinquent Cumul Ug on Federal Mean a school Length of Federal UG for schools Years in on (Non- Bor--incl. UG School that No Degree Credit Student reporting School Student non-bor Student Income doesn’t History Loan SAT Loan) Debt VARIABLES Loan report SAT dad 0.130*** 2.774*** 0.0487*** -2.529*** -0.256*** 0.0518*** 0.0973*** -0.0985 0.0871*** 1.849*** (0.0153) (0.547) (0.00888) (0.399) (0.0382) (0.0136) (0.0149) (0.0677) (0.0154) (0.604) maxsat100 0.00509 0.373*** -0.0139*** 3.258*** 0.339*** -0.0719*** -0.0654*** 0.224*** -0.0484*** 0.0602 (0.00377) (0.135) (0.00218) (0.0989) (0.0102) (0.00335) (0.00367) (0.0169) (0.00378) (0.149) Constant 0.523*** 10.23*** 0.166*** 19.85*** 8.991*** 0.693*** 0.762*** 2.837*** 0.792*** 159.5*** (0.0335) (1.199) (0.0194) (0.879) (0.0898) (0.0298) (0.0327) (0.151) (0.0337) (1.322) Observations 4,487 4,487 4,487 4,383 3,144 4,487 4,487 4,190 4,487 4,487 R-squared 0.017 0.009 0.024 0.248 0.311 0.123 0.099 0.052 0.061 0.465 N ote: Displayed are coefficients from regressions of various credit-related outcomes (denoted by the column headers) on dad’s education and a student’s SAT score. Dad is a binary measure of father’s educational attainment (where a value of 0 reflects having a BA); therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Debt measured in thousands and real adjusted to 2008 dollars. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text, restricted to those who attended postsecondary school. Regressions include graduation year effects. *** denotes significance at 1%. 56

Appendix Table 6c: Association between Family Background and Other Covariates, Using any_pell to Measure Family Background (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Borrowed Defaulted Attending a Delinquent Federal UG Cumul Ug on Federal Mean Avg SAT for school that on (Non- Length of Student Bor--incl. UG Student School schools doesn’t Years in Student Credit VARIABLES Loan non-bor Loan Income reporting SAT report SAT No Degree School Loan) Debt History any_pell 0.307*** 6.567*** 0.0919*** -0.991*** -0.0966*** -0.00122 0.0551*** 0.584*** 0.110*** 2.502*** (0.0138) (0.509) (0.00841) (0.378) (0.0374) (0.0129) (0.0141) (0.0640) (0.0144) (0.559) maxsat100 0.00946*** 0.559*** -0.0143*** 3.415*** 0.356*** -0.0774*** 0.0683*** 0.253*** -0.0484*** 0.110 (0.00330) (0.121) (0.00200) (0.0908) (0.00950) (0.00307) (0.00336) (0.0153) (0.00343) (0.133) Constant 0.442*** 7.842*** 0.161*** 18.04*** 8.817*** 0.764*** 0.805*** 2.435*** 0.798*** 159.1*** (0.0283) (1.042) (0.0172) (0.779) (0.0814) (0.0264) (0.0289) (0.132) (0.0295) (1.144) Observations 5,072 5,072 5,072 4,945 3,511 5,072 5,072 4,736 5,072 5,072 R-squared 0.090 0.035 0.044 0.246 0.309 0.123 0.093 0.063 0.067 0.475 Note: Displayed are coefficients from regressions of various credit-related outcomes (denoted by the column headers) on any_pell and a student’s SAT score. any_pell denotes whether the individual was ever awarded a Pell Grant; therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Debt measured in thousands and real adjusted to 2008 dollars. Sample is SAT test-takers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text, restricted to those who attended post-secondary school. Regressions include graduation year effects. *** denotes significance at 1%. 57

Appendix Table 6d: Association between Family Background and Other Covariates, Using max_pell to Measure Family Background (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Borrowed Defaulted Attending a Delinquent Federal UG Cumul Ug on Federal Mean Avg SAT for school that on (Non- Length of Student Bor--incl. UG Student School schools doesn’t Years in Student Credit VARIABLES Loan non-bor Loan Income reporting SAT report SAT No Degree School Loan) Debt History max_pell 0.202*** 4.685*** 0.0897*** -0.429 -0.0988** 0.0167 0.0598*** 0.642*** 0.118*** 2.577*** (0.0173) (0.624) (0.0102) (0.461) (0.0471) (0.0156) (0.0171) (0.0785) (0.0175) (0.678) maxsat100 0.00152 0.409*** -0.0151*** 3.452*** 0.357*** -0.0764*** 0.0685*** 0.252*** -0.0488*** 0.0953 (0.00342) (0.123) (0.00202) (0.0913) (0.00950) (0.00308) (0.00338) (0.0154) (0.00345) (0.134) Constant 0.572*** 10.43*** 0.183*** 17.50*** 8.792*** 0.754*** 0.815*** 2.536*** 0.819*** 159.6*** (0.0287) (1.033) (0.0169) (0.764) (0.0797) (0.0258) (0.0283) (0.130) (0.0289) (1.122) Observations 5,072 5,072 5,072 4,945 3,511 5,072 5,072 4,736 5,072 5,072 R-squared 0.028 0.014 0.036 0.245 0.308 0.124 0.092 0.060 0.065 0.474 Note: Displayed are coefficients from regressions of various credit-related outcomes (denoted by the column headers) on max_pell and a student’s SAT score. max_pell denotes whether the individual was ever awarded the maximum Pell Grant for a given year; therefore, higher values are associated with higher SES. maxSAT100 is SAT score, measured in hundreds and ranging from 0 to 12. Debt measured in thousands and real adjusted to 2008 dollars. Sample is SAT testtakers matched to the nationally-representative cohort of 23- to 31-year-old individuals with credit records in 2004 as described in the main text, restricted to those who attended post-secondary school. Regressions include graduation year effects. ***, ** denote significance at 1% and 5%. 58