Student Loans and Homeownership

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Student Loans and Homeownership Alvaro A. Mezza, Daniel R. Ringo, Shane M. Sherlund, and Kamila Sommer 2016-010 Please cite this paper as: Mezza, Alvaro A., Daniel R. Ringo, Shane M. Sherlund, and Kamila Sommer (2016). “Student Loans and Homeownership,” Finance and Economics Discussion Series 2016-010. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2016.010r1. 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.

Student Loans and Homeownership∗ Alvaro Mezza† Daniel Ringo‡ Shane Sherlund§ Kamila Sommer¶ June 2017 Abstract We estimate the effect of student loan debt on subsequent homeownership in a uniquely constructed administrative dataset for a nationally representative cohort. We instrument for the amount of individual student debt using changes to the in-state tuition rate at public 4-year colleges in the student’s home state. A $1,000 increase in student loan debt lowers the homeownership rate by about 1.5 percentage points for public 4-year college-goers during their mid 20s, equivalent to an average delay of 2.5 months in attaining homeownership. Validity tests suggest that the results are not confounded by local economic conditions or changes in educational outcomes. (JEL D14, I22, R21) ∗WewouldliketothanktoNeilBhutta,MosheBuchinsky,AlineButikofer,LanceLochner,PaulSullivan, and Christina Wang, as well as two anonymous referees and to the participants of the 2014 Federal System Macro Conference in New Orleans, 2015 Federal System Micro Conference in Dallas, and the Spring 2015 HULM Conference at the Washington University St. Louis for helpful feedback. Taha Ahsin provided excellent research assistance. The analysis and conclusions contained in this paper are those of the authors and do not necessarily reflect the views of the Board of Governors of the Federal Reserve System, its members, or its staff. This manuscript was previously circulated as “On the Effect of Student Loans on Access to Homeownership.” †Federal Reserve Board, email: Alvaro.A.Mezza@frb.gov ‡Federal Reserve Board, email: Daniel.R.Ringo@frb.gov §Federal Reserve Board, email: Shane.M.Sherlund@frb.gov ¶Federal Reserve Board, email: Kamila.Sommer@frb.gov

1 Introduction While the overall U.S. homeownership rate has fallen markedly since the onset of the Great Recession, the decline has been particularly pronounced among young households. The homeownership rate for households headed by individuals aged 24 to 32 fell 9 percentage points (from 45 to 36 percent) between 2005 and 2014, nearly twice as large as the 5 percentage point drop in homeownership for the overall population.1 In trying to explain this rapid decline, rising student loan balances have been implicated as an important drag on homeownership for the young by an array of economists, policy makers, and by the popular press.2 Theoretically, student loan debt could depress homeownership by reducing borrowers’ ability to qualify for a mortgage or desire to take on more debt. In corroboration, recent surveys have found that many young individuals view student loan debt as a major impediment to home buying.3 Despite the attention the issue has received and the intuitive appeal of the causal claim, the evidence establishing an effect of student loans on homeownership is far from definitive. Estimation of the effect of student loan debt on homeownership is complicated by the presence of other factors that influence both student loan borrowing and homeownership decisions. Researchers have previously attempted to isolate the effect by controlling for a set of observable student characteristics (Cooper and Wang (2014) and Houle and Berger (2015)). These studies found only small negative effects of increased debt burdens on homeownership. However, the covariates recorded in available data sets may not adequately control for every important omitted factor, resulting in biased estimates. For example, students preparing for a career with a high expected income might borrow more to fund their college educations and also might be more likely to own a home in the future. To address the endogeneity of student loan debt, in their study of the effects of student loan debt on the future financial stability of student loan borrowers, Gicheva and Thompson (2014) use the national average levels of student loan borrowing as an instrument. They find a more meaningful effect size, 1Source: Current Population Survey. 2Some of the prominent figures making this claim include Nobel laureates Larry Summers and Joseph Stiglitz (“Student Debt is Slowing the U.S. Housing Recovery,” The Wall Street Journal, May 21, 2014) andSenatorElizabethWarren(“SenatorElizabethWarrenAsksFor—AndGets—Realtors’Help,”PLACE- HOLDER. See also: “CFPB Director: Student Loans Are Killing the Drive to Buy Homes,” Housing Wire, May 19, 2014; “Denied? The Impact of Student Loan Debt on the Ability to Buy a House” by J. Mishory and R. O’Sullivan at www.younginvincibles.org. 3See, for example, Stone et al. (2012) or “What Younger Renters Want and the Financial Constraints They See,” Fannie Mae, May 2014. 1

but identification in their approach may be confounded by other aggregate trends.4 In the context of the existing literature, this paper makes two key contributions. First, we use a uniquely constructed administrative data set that combines anonymized individual credit bureau records with Pell Grant and federal student loan recipient information, records on college enrollment, graduation and major, and school characteristics. The core credit bureau data—onto which the other anonymized data sources are merged—are based on a nationally representative sample of individuals who turned 18 between 1991 and 1999 and include data through 2014. The administrative nature of our data likely provides us with more accurate measures of financial variables than the self-reported data sets that are often used in the literature. Second, we use an instrumental variables approach, along with a treatment/control group framework, toidentifythecausaleffectofchangesinstudentloandebtonthehomeownership rate for individuals between the ages of 22 and 32. The instrument is generated by increases in average in-state tuition at public 4-year universities in subjects’ home states. Specifically, we instrument for the total amount of federal student loans an individual had borrowed before age 23 with the average in-state tuition at public 4-year universities from the four school years following the individual’s 18th birthday. This tuition rate directly affects the amount students at these schools may need to borrow to cover their educational expenses, but cannot be affected by any choice or unobservable characteristic of the individual. To eliminate bias from any state level shocks that could affect both the homeownership rate and public school tuition, we split the sample into a treatment and a control group. The treatment group is the set of individuals who attended a public 4-year university at any point before age 23, while the control group is all others.5 Treated individuals are directly exposed to the tuition changes and their debt balances reflect this. Control group individuals are not directly affected by the tuition at schools they did not attend, and so they absorb any variation in economic conditions at the state level that may be driving tuition rates. We show that the instrument passes several placebo tests—for example, while instrumented student loan debt has a strong negative effect on the homeownership rate of the treatment group, no such relationship between public school tuition and homeownership is apparent 4Other studies based on trend analysis include Brown et al. (2013), Akers (2014), Mezza et al. (2014); and analyses by TransUnion (Kuipers and Wise (2016)) and Zillow (http://www.zillow.com/research/ student-debt-homeownership-10563/). 5In Section 4.5 we show the results are robust to restricting the control group to other college attendees. 2

for the control group. The estimated effect of student loan debt on homeownership is also quite stable to the inclusion of various sets of controls, at both the individual and market level (including state-by-year fixed effects). A concern with this framework is that selection into the treatment group, i.e. attendance at a public 4-year university before age 23, is a choice on the part of the individual. It would seem quite plausible that the attendance choices of prospective students depend on the tuition they face, and such endogenous selection would bias our estimates. We show, however, that an individual’s probability of attending a public 4-year university is essentially uncorrelated with the average tuition charged, at least, for the relatively small increases in tuition used in this study to identify the effect of interest. In Section 4.5, we discuss the issue of endogenous selection in detail and place our findings in the context of the relevant literature. Using the aforementioned treatment/control group framework, we find a substantial negative effect of student loan debt on homeownership early in the life cycle. In particular, a $1,000 increase in student loan debt accumulated before age 23 (representing an approximately 10 percent increase in early-life borrowing among the treatment group) causes a decrease of about 1.5 percentage points in the homeownership rate of treatment group students by their mid-twenties. This is equivalent to a delay of 2.5 months in attaining homeownership, given the rapid increase in the probability of homeownership the college-going population experiences through this period in the life-cycle.6 Moreover, the estimated effect shows signs of attenuating as borrowers enter their thirties, although this change over time is imprecisely estimated. In Section 4.7, we present evidence that credit scores provide a significant channel by which student loan debt affects borrowers ability to obtain a mortgage. Higher debt balances increase borrowers’ probability of becoming delinquent on their student loans, which has a negative impact on their credit scores and makes mortgage credit more difficult to obtain. To be sure, this paper estimates the effect of a ceteris paribus change in debt levels, rather than the effect of a change in access to student loan debt, on future homeownership. Inparticular, ifstudentloansallowindividualstoaccesscollegeeducation—or, morebroadly, acquire more of it—student loan debt could have a positive effect on homeownership, as long 6 In contrast, the estimated effect from the procedure based only on observable controls is negative but very small for individuals in their twenties, similar to the results from existing studies. 3

as the return to this additional education allows individuals to sufficiently increase their future incomes. Thus, our exercise is similar in spirit to a thought experiment in which a small amount of student loan debt is forgiven at age 22, without any effect on individuals’ decisions on post-secondary education acquisition. Another caveat to keep in mind is that our estimation sample mostly covers the period prior to the Great Recession. Our findings may therefore be more relevant for times of relatively easier mortgage credit, as opposed to the immediate post-crisis period in which it was much more difficult to get a home loan. We discuss in Section 2.2 how various underwriting criteria in the mortgage market may interact with student loan debt to restrict some borrowers’ access to credit. Several recent studies have looked at the effect of student loans in different contexts, finding that greater student loan debt can cause households to delay marriage (Gicheva (2016) and Shao (2015)) and fertility decisions (Shao (2015)), lower the probability of enrollment in a graduate or professional degree program (Malcolm and Down (2012), Zhang (2013)), reduce take-up of low-paid public interest jobs (Rothstein and Rouse (2011)), or increase the probability of parental cohabitation (Dettling and Hsu (2014), and Bleemer et al. (2014)). These studies suggest credit constraints after post-secondary education may also be relevant outside the mortgage market. The rest of our paper is organized as follows. Section 2 briefly reviews the institutional background of the student loan market and examines the main theoretical channels through which student loan debt likely affects access to homeownership. Section 3 gives an overview of the data set and defines variables used in the analysis. Section 4 presents the estimator in detail, as well as the results of both the instrumental variable analysis and a naive “selection on observables” approach. The instrument is then subjected to a series of validity checks. We also extend the analysis to investigate whether student loans affect the size of the first observed mortgage balance, and whether credit scores provide a channel by which student loan debt can restrict access to homeownership. Section 5 interprets and caveats our main findings. Section 6 concludes. 4

2 Background and Mechanism 2.1 Institutional Background Student loans are a popular way for Americans to pay the cost of college, and the use of such loans has been increasing in recent years. In 2005, 30 percent of 22-year olds had accumulated some student loan debt, with an average real balance among debt holders of approximately $13,000. By 2014, these numbers had increased to 45 percent and $16,000, respectively.7 The vast majority of students have access to federal student loans, which generally do not involveunderwritingandcanchargebelowmarketrates.8 Theamountofsuchloansstudents can borrow is capped by Congress, however. Federal student loans are also not dischargeable in bankruptcy, reducing the options of borrowers in financial distress.9 Student borrowers frequently exhaust their available federal loans before moving on to generally more expensive private loans, often with a parent as co-signer.10 Historically, the typical student loan is fully amortizingovera10-yeartermwithfixedpayments. Defermentsandforbearancescanextend this term, as can enrollment in alternative repayment plans, such as the extended repayment plan (available for borrowers with high balances) and income-driven repayment plans (which have become more common in recent years and are available for borrowers with elevated debt-to-income ratios), and through loan consolidation.11 Student loan debt can impose a significant financial burden on some borrowers. Despite the inability to discharge federal loans through bankruptcy, 14 percent of recipients with outstanding federal student debt were in default as of October 2015.12 Student borrowers 7Statisticsarebasedonauthors’calculationsusingthenationallyrepresentativeFRBNYConsumerCredit Panel/Equifaxcreditbureaudata. Ouranalysisfocusesonyoungpeopleandthedebttheyhaveaccumulated before age 23. Overall debt levels are notably higher, as individuals can continue to accumulate debt past the traditional college-going age. The average outstanding loan balance for the overall borrower population was $27,000 in 2014, up from $20,000 in 2005. 8Some restrictions in eligibility apply. For instance, the post-secondary institution the student attends has to be included under Title IV to be eligible for federal student aid. Also, 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. Graduate students taking PLUS loans—as well as parents taking Parent PLUS loans—must pass a credit check. 9In2005thebankruptcycodewasamended,makingprivatestudentloansalsonotroutinelydischargeable in bankruptcy. 10The share of private loans with a co-signer increased significantly after the financial crisis, from 67 percent in 2008, to over 90 percent in 2011. Source: CFPB, Private Student Loans, August, 2012 https: //www.consumerfinance.gov/data-research/research-reports/private-student-loans-report/. 11Source: https://studentaid.ed.gov/sa/repay-loans/understand/plans. 12Source: U.S. Department of Education, Federal Student Aid Data Center, Federal Student Loan Port- 5

are often young and at a low point in their life cycle earnings profile. The financial difficulties may be more severe for students who fail to graduate. Of the federal student loan borrowers who entered repayment in 2011-12 without a degree, 24 percent defaulted within two years.13 2.2 Theoretical Mechanism Most young home buyers must borrow the money to buy their first house. We conjecture that three underwriting factors provide a channel through which student loan debt can affect the borrower’s ability to obtain a mortgage.14 First, the individual must meet a minimum down payment requirement that is proportional to the house value. While a 20 percent down payment is typical for many buyers, with mortgage insurance (whether purchased from a private company or a government agency such as the Federal Housing Administration (FHA)) the down payment can be significantly less.15 Second, the individual must satisfy a maximum debt-to-income (DTI) ratio requirement, with the ratio of all her debt payments not to exceed a percentage of her income at the time the loan is originated. Third, the individual must satisfy a minimum credit score requirement. As these underwriting factors worsen for any individual (i.e., less cash available for a down payment, higher DTI ratio and lower credit score), she will be more likely to be rejected for a loan, or face a higher interest rate or mortgage insurance premium. Itisnothardtoseehow—allelseequal—havingmorestudentloandebtcanmechanically affect one’s entry into homeownership through these three channels. First, a higher student loan debt payment affects the individual’s ability to accumulate financial wealth that can then be used as a source of down payment. Second, a higher student loan payment increases the individual’s DTI ratio, potentially making it more difficult for the borrower to qualify for a mortgage loan. Third, student loan payments can affect borrowers’ credit scores. On the one hand, the effect can be positive: timely payments of student loan debt may help borrowers to improve their credit profiles. On the other hand, potential delinquencies folio. 13Source: U.S.DepartmentofTreasurycalculationsbasedonsampledatafromtheNationalStudentLoan Data System. 14Eveninastandardlife-cyclemodelwithperfectcapitalmarketsandnopsychologicalcostofdebt(ie.,no debt aversion), student debt can affect homeownership (or, more generally, post-college decisions) through a negative wealth effect. However, for a typical individual, this effect is likely quite small, since the total student loan debt will only be a small fraction of the present discounted value of total lifetime earnings. 15The FHA requires a down payment as low as 3.5 percent of the purchase value. 6

adversely affect credit scores, thereby hampering borrowers’ access to mortgage credit. At thesametime, othernon-underwritingfactorsmighthaveeffectsaswell. Forexample, froma behavioral perspective, if individuals exhibit debt aversion and wish to repay at least some of their existing debt prior to taking on new debt in the form of a mortgage, larger student loan debt burdens can further delay their entry into homeownership. Available evidence points to the existence of debt aversion in different settings, suggesting this mechanism might play a role in reducing the probability of homeownership.16 Various factors might influence how the effect of student loan debt on homeownership changes in the years after leaving school. Since cumulative balances are generally largest immediately upon entering repayment (see Figure 15 in Looney and Yannelis (2015)), there are at least four reasons to believe that the ceteris paribus effect of higher student loan debt on homeownership access might be largest immediately upon school exit. First, given that the income profile tends to rise over the life cycle and student loan payments are fixed, the DTI constraint should ease over time, as should the budget constraint, thereby allowing the individual topotentially accumulate assets fora down payment ata faster rate. Second, once all debt is repaid, the student loan debt component of debt payments in the DTI constraint disappears entirely. Of course, the past effects of student loan payments on accumulated assets are likely to be more persistent if student loan payments significantly impaired the individual’s ability to save at a rate comparable to that of an individual with less student debt for a period of time. Third, the Fair Credit Reporting Act prohibits the credit bureaus from reporting delinquencies more than seven years old, so any difficulties the borrower had meeting payments will eventually drop off her credit report. Lastly, any effect of debt aversion induced by a higher student loan debt burden at school exit should diminish over time as the balance is paid down. A simple, two-period model illustrates the various mechanisms by which student loan 16For example, Palameta and Voyer (2010) find that some (Canadian) students are willing to accept a financial aid package with a grant but do not accept one that combines the same amount of a grant and an optional loan. Field (2009) finds evidence of debt aversion in an experiment where loan-repayment terms were randomly varied at NYU-Law school. Loewenstein and Thaler (1989) and Thaler (1992) find that payoff rates of mortgages and student loans are irrationally rapid, suggesting the existence of debt aversion. Other studies find less of an effect. In particular, focusing on students attending a highly selective university, Rothstein and Rouse (2011) find that the increase in post-graduation income and the decrease in the probability that students choose low-paid public interest jobs due to exogenous increases in student loans are more likely driven by capital market imperfections (i.e, credit constraints post-graduation) than by debt aversion. 7

debt can affect homeownership over time.17 Let hand-to-mouth consumers enter period 1 of adult life with some amount of student loan debt, L > 0. They earn income Y and can 1 choose to default on their student loans (D = 1) or pay them off (D = 0). They can also purchase a home (H = 1) at price P, which requires a down payment of fraction θ of the 1 total house price, θP. In the second period, consumers earn Y > Y . If they defaulted on 2 1 their student loans in period 1, they must pay them off in period 2 as student loans cannot be discharged. If they purchased a home in period 1, they must pay the remainder of the balance, (1−θ)P, in period 2. If they did not previously purchase a home (H = 0), they 1 have the option of paying P to purchase a home in this period (H = 1). Finally, houses 2 purchased in period 1 cannot be sold, meaning that H ≤ H . 1 2 Consumers’ utility each period is an increasing but concave function of consumption, c. Homeowners receive an additive individual idiosyncratic utility benefit γ in each period they own. However, in order to purchase a home in period 1, consumers must qualify for a mortgage. First, if consumers purchase a home in period 1, they must meet a DTI ratio constraint such that the loan-to-income ratio is less than a threshold α; i.e., L < α. Second, Y1 there is a credit history constraint. If the consumer chooses to default on their student loans in period 1 (D = 1), then they are disqualified from borrowing and cannot purchase a home in that period. Additionally, borrowers are allowed to be debt-averse. Specifically, consumers with greater student loan balances, L, experience greater disutility from taking on mortgage debt, all else equal. The utility lost to debt aversion is captured by the function (cid:15)(L) that falls with the student loan amount, L (i.e., d(cid:15)(L) < 0).18 In our stylized model, the dL debt aversion only enters consumer’s utility in period 1 when mortgage debt is required to buy a home. The consumer’s problem then becomes: max U(c )+U(c )+γ((cid:15)(L)H +H ) (1) 1 2 1 2 c1,c2,H1,H2,D subject to the budget constraints c +θPH +L(1−D) ≤ Y (2) 1 1 1 17For simplicity of exposition, we abstract from allowing households to save in deposits and, in this way, carry over funds across periods. However, it is easy to see that in an alternative set-up where savings are allowed,higherstudentloanbalanceswouldreducehouseholdabilitytosaveinperiod1andusethesesavings to partially fund housing purchases in period 2. 18Whend(cid:15)(L) =0. the debt aversion channel is not operative. dL 8

and c +((1−θ)+(1−H )θ)PH +LD ≤ Y (3) 2 1 2 2 and the credit market constraint (cid:26) (cid:27) L H ≤ 1 < α∩D = 0 . (4) 1 Y 1 Our simple model illustrates how higher student loan debt levels, L, affect the decision to purchase a home in period 1. First, due to concave utility (d2U(c) < 0) and the strictly dc2 increasing income profile (Y < Y ), the relative utility (and thus the probability) of waiting 1 2 to buy a house in period 2 rather than paying the cost of a down payment in period 1 is increasing in student loan debt, L. Second, the DTI ratio constraint (L < α) also is more Y1 likely to bind for consumers with more student loan debt, all else equal. Third, at high debt levels, defaulting also becomes a more valuable option for student loan borrowers. As L increases, consumers are more likely to default in order to shift the burden of student loan payment into a period with higher income (i.e., period 2) at the cost of restricting their access to mortgage credit in period 1 (i.e., dProb(D=1) > 0). Fourth, if borrowers are dL debt-averse (i.e., d(cid:15)(L) < 0), then the higher the student loan balance, the less utility from dL homeownership borrowers enjoy when financing a home with a mortgage in period 1 and, therefore, the less likely they will be to purchase a home in that period. Higher debt levels can also affect the decision to purchase a home in period 2. For borrowers who defaulted on their student loans (i.e., D = 1), the unpaid student loan balance, L, is due in the second period. The larger this balance is, the greater the marginal utility of consumption in period 2 and the lower the probability of choosing to purchase a home. Additionally, defaulters are less likely to buy a home in period 2 than those who did not default (and therefore have no more student loan debt to pay off), so increased student loan debt also reduces the probability of home buying in period 2 by increasing the probability of default in period 1. Among those who did not default (D = 0), however, the original student loan debt is fully paid off by the time borrowers enter period 2, and so it does not have a direct effect on their decision to purchase a home in that period.19 While our discussion thus far suggests that the effect of student loan debt on homeown- 19A formal characterization of the solution to the model is available upon request. 9

ership attenuates over time due to student loan debt repayment and rising incomes, there may be countervailing effects. In particular, the propensity for homeownership is generally relatively low among those newly out of school and increases with age. Hence, the number of marginal home buyers may peak many years after school exit, suggesting that the effect of student loan debt might be increasing as the debtor ages. Also, individuals may exhibit habit formation in their housing tenure choice. A marginal home buyer who is induced into renting by her debts may become accustomed to renting, in which case the apparent effect of student loan debt on homeownership could persist for many years. Themechanismsdiscussedinthissectionarenotspecifictostudentloandebt—autoloans andcreditcarddebtcouldimposesimilarburdensondebtorsinthehousingmarket. Student loan debt is particularly interesting to study, however, because of its ease of availability. Young people without incomes or collateral are able to take on tens of thousands of dollars of debt to pay for their education without any underwriting of the loans. In contrast, a borrower without a credit history or source of income would face very tight limits in markets forprivatelyprovidedcredit. Studentloansthereforepresentauniquechannelforindividuals to become heavily indebted at a young age. 3 Data Our data are pooled from several sources.20 Mezza and Sommer (2016) discusses the details of the data, checks the representativeness of the merged data set against alternative data sources, and provides caveats relevant for the analysis. By way of summary, the data set is built from a nationally representative random sample of credit bureau records provided by TransUnion, LLC, for a cohort of 34,891 young individuals who were between ages 23 and 31 in 2004, and spans the period 1997 through 2014. Individuals are followed biennially between June 1997 and June 2003, then in December 2004, June 2007, December 2008, and then biennially again between June 2010 and June 2014. The data contain all major credit bureau variables, including credit scores, tradeline 20Allthemergesofindividual-levelinformationhavebeenperformedbyTransUnion, LLC,inconjunction withtheNational StudentClearinghouse, the DepartmentofEducationandthe CollegeBoard. Themerges were based on a combination of Social Security number, date of birth, and individuals’ first and last names. None of this personal identifying information used to merge individuals across Sources is available in our data set. 10

debt levels, and delinquency and severe derogatory records.21 Since the credit bureau data do not contain information on individuals’ education, historical records on post-secondary enrollment spells and the institutional-level characteristics associated with each spell were merged on the TransUnion sample from the DegreeVerify and Student Tracker programs by the National Student Clearinghouse (NSC). Additionally, individual-level information on the amount of federal student loans disbursed—our main measure of student loan debt—was sourced from the National Student Loan Data System (NSLDS). The NSLDS also provides information on Pell grant receipts and enrollment spells funded by federal student loans, including the identity of each post-secondary institutions associated with the aid, which we use to augment the NSC data. Information on individuals’ state of permanent residence at the time they took the SAT standardizedtest—sourcedfromtheCollegeBoard—wasmergedforthesubsetofindividuals who took this test between 1994 and 1999, at a time when most of the individuals in our sample were exiting high school.22 Finally, we merged in institutional records, such as school sector, from the Integrated Postsecondary Education Data System (IPEDS). In what follows, we describe the construction of key variables used in our analysis: homeownership status, student loan balances, and subjects’ home state. A discussion of the remaining variables used in the analysis is available in Appendix A.1. We are not able to directly observe the individual’s homeownership status. Rather, the credit bureau data contain opening and closing dates for all mortgage tradelines that occurred prior to July 2014, which we use to infer homeownership by the presence of an open mortgage account. The obvious limitation of using mortgage tradeline information to infer the individual’s homeownership status is that we will not be able to identify homeowners who are cash-buyers. However, because our analysis is restricted to home-buying decisions made between the ages of 22 and 32, the population of cash-buyers is likely to be small, particularly among the sub-population that required student loans to fund their education. Furthermore, the credit-rationing mechanisms discussed in Section 2.2 would not bind on 21While we observe when all loan accounts have been opened and closed, as well as the complete delinquency events on these accounts, we only observe debt balances at the particular times when credit records were pulled, i.e., June 1997, June 1999, etc. 22The 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, the SAT was fully elective and, as such, not all potential college entrants took it. 11

a buyer with enough liquid assets to purchase a house outright, so there is less scope for student loan debts to affect purchase decisions for any such individuals. In our analysis, we treat the individual’s homeownership status as an absorbing state, so that if an individual is observed to be a homeowner by a given month, the individual will be treated as a homeowner at all future dates. The key explanatory variable, student loan balance, is measured as the total amount of federal student loans disbursed to an individual before they turned 23. We use disbursement of federal student loans from the NSLDS, rather than student loan balances from credit bureau data, for two reasons. First, balances in the credit bureau data are reported roughly biennially, so we do not observe student loan balances at the same ages for all individuals. Second, student loan balances from the credit bureau data are available to us for the first time in June 1997. By then, the oldest individuals in our sample were already 23 years old. A potential drawback of our approach is that the measure of total federal loans disbursed does not include accrued interest, repaid principal, or private student loans. Our instrumental variables approach relies on the imputation of the subject’s pre-college state of residence (henceforth, home state). To construct home states, we proceed in four steps. First, for individuals who took the SAT, we use these individuals’ state of legal residence at the time when they took the test, as reported in the College Board data. Fifteen percent of our sample have their home state identified in this manner. Second, for individuals for whom SAT information is not available, we use the state of residence observed in the TransUnion credit records prior to their first enrollment in college, if these data are available. An additional 20 percent have their home state identified this way. Third, for the remaining 37 percent of the sample who attended college but did not fall in either of the above two categories, we impute the home state using data on the state in which the school associated with the first enrollment spell is located.23 This last step can certainly appear problematic given that it could reflect an endogenous locationchoiceassociatedwithstate-levelcollegecostsorcollegequality. However, acasecan be made that the state of the first college attended is a reliable indicator of the individual’s 23In our data, 71 percent of individuals are identified as having attended college at some point. In the ACS,only64percentofindividualsinthecohortaged23-31in2004reportedanycollegeeducationby2015. OnepossiblesourceofdiscrepancyisthefactthatnoteverypersonintheUnitedStateshasacreditrecord. Those who did not attend college are possibly less likely to have interacted with formal credit markets, and so may be underrepresented in the TransUnion data. 12

home state among the sub-population that did not take the SAT or appear in credit bureau records prior to attending college. In particular, in the nationally representative 2003-04 Beginning Postsecondary Students Longitudinal Study, only 11 percent of first-time, nonforeign college entrants attended a post-secondary institution not in their state of legal residence, withthestateoflegalresidencedefinedasthestudent’strue, fixed, andpermanent home.24 Under this definition, if the student moved into a state for the sole purpose of attending college, that state does not count as the student’s legal residence. In our sample, 23 percent of students whose home state was identified by the SAT or their credit record attended an out of state post-secondary school.25 These students represent 11 percent of our total sample of college attendees, accounting for the entire expected population of out-ofstate students, and suggesting that among the remaining students the state of first college attendance is extremely likely to be their home state. We therefore do not believe that misidentification of home state is a significant issue. Finally, for the remaining 28 percent of individuals who neither attended college nor took the SAT, we impute their home states with the first state available in the credit records.26 Public 4-year university tuition rates are assigned to individuals on the basis of their home state, as imputed by the procedure outlined above.27 Several filters are applied to the baseline cohort of 34,891 individuals. First, we drop 141 observations for which TransUnion was not able to recover personal identifying information on which to perform the merge. We then drop 40 individuals who were not residing in any of the50U.S.statesortheDistrictofColumbiabeforestartingcollegeand6individualswhowe could not match to a home state. Moreover, we drop 698 individuals for whom we were not able to determine the school sectors they attended. Finally, we drop 571 individuals whose earliest enrollment record corresponds to the date a degree was obtained, rather than an actual enrollment record.28 The resulting sample used in the analysis thus contains 33,435 24Source for the definition: https://fafsa.ed.gov/fotw1415/help/fahelp46.htm. 25While the College Board data for SAT-takers is available only for a subsample of our total population, its coverage is likely skewed toward higher academically achieving individuals who are more likely to attend out-of-state selective institutions. 26The average age at which we first observe a state for this group of individuals is 22.6 27Thedataontheaveragein-statetuitionatpublic4-yearuniversitiesbystateandacademicyearareavailable on the NCES’s Digest of Education Statistics website: https://nces.ed.gov/programs/digest/. Average in-state tuition reflects the average undergraduate tuition and required fees. 28Some schools participate in the NSC DegreeVerify program, but not in the Student Tracker program. Additionally,schoolsparticipatinginbothprogramsusuallyreportgraduationdatesretroactively(frequently reporting back several years prior to their enrollment in the DegreeVerify), but report enrollment spells 13

individuals. Summary statistics for the variables we use in this analysis are presented in Table 1. 4 Estimation In this section we present our findings. First, in Section 4.1, we describe some basic correlations between student loan debt and homeownership, including how these evolve over the life cycle and vary by education level. In Section 4.2 we show the results of several naive regressions, attempting to address the endogeneity of student loan debt by controlling for observable characteristics. Our main identification strategy, using an instrumental variables approach and the treatment/control group framing, is detailed in Section 4.3. We then present the results in Section 4.4. In Sections 4.5 and 4.6 we discuss potential failures of our identifying assumptions, and run a variety of tests to validate them. Finally, in Section 4.7, we estimate the effect of student loans on individuals’ credit scores and delinquent status, and the size of their mortgage balances. 4.1 Patterns of Debt and Homeownership Student loan debt is correlated with homeownership, but this relationship is not stable over the life cycle. Figure 1 plots the probability of ever having taken on a mortgage loan against the individual’s age for different levels of student debt. In the top left panel, we compare individuals who attended college before age 23 without taking on debt to those who did borrow, as well as individuals who did not attend college by that age. Debt free college attendees have a higher homeownership rate than their indebted peers at age 22, but those with debt catch and surpass the debt free group by age 29. In the bottom left panel of Figure 1, we refine college attendees into three categories based on amount borrowed: no borrowing, less than $15,000, and more than $15,000. Students who borrow moderate amounts start off less likely to own than non-borrowers, but eventually catch up. Those who borrowed the most start with the lowest homeownership rate at age 22, but are substantially more likely to be homeowners by age 32 (the median age of first home buying, according to the National Association of Realtors). From these plots one might be tempted to conclude starting from the moment they enroll in the Student Tracker program (or just a few months prior). 14

that, at least in the medium run, higher student loan debt leads to a higher homeownership rate. Determining how student loan debt affects homeownership is not so straight forward, however. Individuals with differing amounts of student loan debt may also differ in other important ways. Notably, they may have different levels of education, which is itself highly correlatedwithhomeownership(possiblythroughaneffectonincome). Thetoprightpanelof Figure 1 restricts the sample to individuals who attained a bachelor’s degree before age 23. Within this group, those without student loan debt always have a higher homeownership rate than borrowers. In the bottom right panel, we can see that splitting the sample of borrowers further into groups by amount borrowed presents a similar picture. Students who borrowed more than $15,000 had the highest homeownership rate among the general college going population after age 27, but have the lowest rate among the subset with a bachelor’s degree at all ages. Bachelor’s degree recipients with no student loan debt have the highest homeownership rate across the range of ages. As such, simple correlations clearly do not capture the whole picture. 4.2 Selection on Observables Further factors that are correlated with both student loan debt and homeownership (and may be driving the observed relationship between these two variables of primary interest) include the type of school attended, choice of major, and local economic conditions, for example. One potential identification strategy is to attempt to absorb all these potential confounders with an extensive set of control variables. For the purpose of comparison with ourinstrumentalvariableestimates(presentedinSection4.4), werunage-specificregressions ofanindicatorforhomeownershiponstudentloandebtsandvarioussetsofcontrols. Inthese and subsequent regressions, the individual level explanatory variables (including student loans disbursed) are all measured at the end of the individual’s 22nd year. All standard errors are clustered at the state-by-cohort level. OLS and probit estimates of the effect of student loan debt on homeownership by age 26 are presented in Tables 2 and 3, respectively. Estimates are generally similar across the range of specifications in columns 1-5, which sequentially control for an increasingly rich set of covariates, including school sector, degree attained, college major, Pell grant receipt, 15

measures of local economic conditions, state and cohort fixed effects, and, finally, state by cohort fixed effects. Column 6 restricts the sample to individuals who attended any postsecondary schooling before turning 23. A $1,000 increase in student loans disbursed before age23isassociatedwithanapproximately0.1percentreducedprobabilityofhomeownership by age 26. Figure 2 plots estimates of the marginal effect of student loan debt against borrower’s age for the linear probability and probit models. These estimates are derived from the regressions using the vector of controls in columns 5 of Tables 2 and 3 for the OLS and probit specifications, respectively. Across both linear probability and probit models the estimated effect starts negative for borrowers in their early twenties and becomes positive when they reach their early thirties. Our estimates from these selection-on-observables regressions are closely in line with previous findings in the literature. Using the National Longitudinal Survey of Youth, 1997, Houle and Berger (2015) estimate that a $1,000 increase in student loan debt decreases the probability of homeownership by 0.08 percentage points among a population composed largely of 20- and 25-year olds. Similarly, using the National Education Longitudinal Study of 1988, Cooper and Wang (2014) find that a 10 percent increase in student loan debt (approximately equivalent to a $1,000 increase for our sample) reduces homeownership by 0.1 percentage points among 25- and 26-year olds who had attended college. 4.3 Instrumental Variable Estimation While the estimators used above control for some important covariates, there may still be unobservable variables biasing the results. It is not clear, a priori, in which direction the estimates are likely to be biased by such unobservable factors. For example, students with higher unobservable academic ability may borrow more, either because they choose to attend more expensive institutions or because they anticipate greater future incomes. These higher ability students would also be more likely to subsequently become homeowners, introducing a positive bias in the naive estimates. Conversely, students from wealthy backgrounds may receive financial assistance from their parents and therefore need to borrow less to pay for school than their less advantaged peers.29 Parental contributions could help these same students to later purchase a home, which would tend to introduce a negative bias. The 29For example, Lovenheim (2011) finds shocks to housing wealth affect the probability families send their children to college. 16

covariates we have may not adequately control for these or other omitted factors. Reverse causality is also a potential source of bias, if purchasing a home before leaving school affects students’ subsequent borrowing behavior. To reliably identify the causal effect of student loan debt, we need a source of variation that is exogenous to all other determinants of homeownership. We propose that the average tuition paid by in-state students at public 4-year universities in the subject’s home state during his or her prime college-going years provides quasiexperimental variation in eventual student loan balances. A large fraction of students attend public universities in their home state, so the loan amounts they require to cover costs vary directly with this price.30 Additionally, this tuition cannot be affected by the choice of any particular individual. Rather, changes in the tuition rate depend on a number of factors that are arguably exogenous to the individual homeownership decision, ranging from the level of state and local appropriations to expenditure decisions by the state universities. A short overview of the major drivers of prevailing tuition rates will help clarify the validity argument, and locate potential points of failure. One major source of tuition increases is changes to particular schools’ cost structures. According to Weeden (2015), these costs include compensation increases for faculty members, the decision to hire more administrators, benefit increases, lower teaching loads, energy prices, debt service, and efforts to improve institutional rankings, all of which have been linked to tuition increases since the 1980s. Institutions also compete for students, especially those of higher academic ability, by purchasing upgrades to amenities such as recreational facilities or residence halls. These upgrades are often associated with increased tuition to pay for construction and operation of new facilities. Finally, tuition and fees are frequently used to subsidized intercollegiate athletic ventures. In recent years, athletic expenses have increased and now may require larger subsidies from tuition and fee revenue at many colleges. Another major driver of tuition rates is the level of taxpayer support. As described in Goodman and Henriques (2015) and Weerts et al. (2012), public universities receive a large portion of their operating income from state and local appropriations. The amount of state and local revenue that public colleges receive is itself influenced by a diverse set of factors that weigh on legislators in allocating funds, including state economic health, state 30In our sample, nearly half of the students who had attended any college before age 23 had attended a public 4-year university in their home state. 17

spending priorities, and political support for affordable post-secondary education. Since public colleges can, in theory, offset the lost revenue from appropriations with increased tuition, appropriations for higher education can be crowded out by funding for other state programs. Any correlation between the tuition charged at public universities and state level economic conditions (through the effect of economic conditions on appropriations) raises a concern about the validity of tuition as an instrument. To address this potential source of bias, we split our sample into treatment and control groups, with the treatment group defined as the individuals who attended a public 4-year university before they turned 23. We then compare the outcomes among the treatment group to those of the control group, which consists of all other individuals (except in specifications show in column 6 of Tables 5and 6, where the control group is all other individuals with at least some post-secondary education before age 23). Treatment group subjects pay the tuition charged at public 4-year universities, and so their total borrowing before turning 23 is directly affected by this tuition. In contrast, the control group is not directly affected by the tuition at public 4-year universities (which they did not attend). This framework therefore allows us to control for any correlations between state level shocks and tuition rates—either by including tuition rates directly as a control variable or by using state-by-year fixed effects—with the homeownership rate of the control group absorbing unobserved variation in economic conditions.31 Specifically, we estimate the effect of student loans on homeownership via a two stage estimator that uses the interaction between tuition and an indicator for the treatment group as an instrument for student loan debt. The first stage of our instrumental variables regression is described in equation 5: X = α +α Z +α D +α Z ×D +W α +(cid:15) (5) i 0 1 i 2 i 3 i i i 4 i where X is the amount of federal student loans borrowed by individual i prior to age 23, i Z is the average tuition charged at public 4-year universities in i’s home state in the four i school years following i’s 18th birthday, and D is a dummy variable indicating i attended a i public 4-year university before i turned 23. The vector W can include a variety of controls i at the individual and state level, including fixed effects for individual’s home state, birth 31We devote further consideration to the potential endogeneity of tuition in Section 4.5 18

cohort, or for the combination of the two, i.e., state-by-year fixed effects. The interaction term, Z ×D , is the only excluded term in the second stage. We estimate the second stage i i using equation 6: Y = β +β X +β Z +β D +W β +µ (6) it 0 1 i 2 i 3 i i 4 i where Y is a dummy variable indicating i has become a homeowner by age t. The parameit ter β captures any partial correlation between tuition rates and homeownership among the 2 control group, absorbing any state level shocks that affect both tuition and the homeownership rate. Note that in specifications with state-by-year fixed effects β is not identified, 2 as the average tuition rate is collinear with the fixed effects. The parameter β captures the 3 average difference in homeownership rates between the treatment and control groups. We are left identifying β , the effect of student loan debt on homeownership, by the widening or 1 shrinking of the gap in homeownership rates between public 4-year school attendees and the general population as tuition rates change, analogous to a difference-in-differences estimator. Estimatesofβ maybeinconsistentifmembershipinthetreatmentgroupisinfluencedby 1 tuition rates. In particular, if the attendance decisions of students considering public 4-year universitiesareswayedbytheprevailingtuition, thenourestimateswouldsufferfromsample selection bias. However, we will show that the variation in tuitions exploited in this study exertnomeaningfuleffectontheprobabilityofastudentattendingapublic4-yearuniversity. Given this result, we believe it is reasonable to consider treatment group membership to be exogenous. The issue of selection into the treatment group is discussed further in Section 4.6, in which we also consider the potential endogeneity of other educational outcomes. Estimation of equation 6 produces an estimate of the local average treatment effect (LATE) of student loan debt on homeownership. That is, we are estimating the effect within the subpopulation of treatment group individuals whose debt levels are sensitive to tuition rates. The treatment group consists of traditional students—those that entered college immediately or very soon after high school, and attended a public 4-year university. Careshouldbetakenwhenextrapolatingourresultstothegeneralpopulationwhichincludes many individuals who enrolled in a private or public 2-year university, or who first attended college later in life. If such individuals respond to debt much differently than traditional students, we do not capture this heterogeneity of treatment effect in our estimates. 19

4.4 Instrumental Variable Estimation Results First stage results from regressing student debt on the instrument and other controls are presented in Table 4. Across specifications, a $1,000 increase in the sum of average tuition across the four years after the individual turned 18 is associated with an approximately $150 increase in student loan debt for students in the treatment group. The estimates are strongly statistically significant. For reference, after controlling for state and cohort fixed effects, the residual of the four-year sum of in-state tuitions has a standard deviation of $915 across our sample. Turningnowtothesecondstage, wefindaconsiderablylargereffect, inabsoluteterms, of studentloandebtonhomeownershipthanintheearlierspecificationswithouttheinstrument. Results for the 2-Stage Least Squares (2SLS) and IV-Probit estimators are presented in Tables 5 and 6, respectively. Across both linear probability and probit models, we find a statistically significant effect at age 26, with a $1,000 increase in student loan debt leading to anapproximately1to2percentagepointdecreaseintheprobabilityofhomeownership. Since the average treatment group student in our sample had accrued, in constant 2014 dollars, approximately $10,000 of federal student loan debt before age 23, the $1,000 increase in student loan balances represents approximately a 10 percent increase in borrowing for the average person in the treatment group. Further interpretation of the magnitude of these results is presented in Section 5. The estimates from the IV specifications imply a considerably stronger effect than those from the selection-on-observables estimates in section 4.2. This difference suggests the presence of unobservable factors biasing the OLS estimates. In particular, individuals with greater levels of student loan debt are positively selected into homeownership—that is, they have a greater underlying (unobservable) propensity to become homeowners than individuals with smaller amounts of debt do. It may be, for example, that students with greater labor market ability take on more student loan debt, either due to attending more expensive schools or because they anticipate higher lifetime incomes. These high ability (and highly indebted) individuals are then also more likely to become homeowners in their mid-20s. The inclusion of educational controls in some specifications may pose a concern. Changes in tuition could affect students’ decisions about sectoral choice, completion, or which major to pursue. Failing to control for these variables could then lead to biased estimation. On 20

the other hand, these outcomes are potentially endogenous to unobserved determinants of homeownership, so their inclusion would introduce another source of bias. We show specifications with and without the controls (compare columns 1 and 2 of Tables 5 and 6) and find qualitatively similar results. In Section 4.6 we show that there is little evidence that our measured educational outcomes are affected by movements in tuition. Figure 3 plots estimates of the marginal effect of student loan debt against the borrower’s ageforthe2SLSandIV-probitmodels, respectively. Thetopleftandrightpanelsshow2SLS and IV-Probit estimates, respectively, derived from the instrumental variable regressions using the vector of controls reported in columns 2 in Tables 5 and 6. The bottom left and right panels use the vector of controls reported in columns 5 in Tables 5 and 6. Student loan borrowers seem most affected by their debt from ages 26-28. After that, the point estimates are reduced in magnitude, possibly suggesting a catch up in the homeownership rate among more indebted borrowers. The standard errors are large enough, however, that this apparent pattern is merely speculative. It is worth keeping in mind that tuition changes could affect homeownership via channels not directly measured by student loan debt. If students (or their parents) have assets they draw down to pay for college, a higher tuition leaves them with less left over for an eventual down payment on a house. This behavior would tend to bias our estimates of the effect of debt away from zero. Stripping away the assumed channel of student loan debt, we can look directly at the reducedformeffectoftuitionsonhomeownershipforthetreatmentandcontrolgroups. Table 7 presents results of regressing homeownership directly on the instrument and usual vectors of controls. Every additional thousand dollars of tuition (charged over a four year period) leads to a 0.2 to 0.3 percentage point lower homeownership rate for the treatment group at age 26, with no significant effect for the control group. It is not surprising that the reduced form effect of tuition is considerably smaller than the estimated effect of debt. Debts do not rise one-for-one with tuition hikes, as not all students attend school full time for four straight years post-high school, and not all students pay the sticker price of tuition (for example, if they receive need-based grants). Imposing an additional $1,000 cost on students would affect their homeownership rate substantially more than the 0.2 to 0.3 percentage points estimated in the reduced form specification. 21

4.5 Endogeneity of Tuition Our identifying assumption that the instrument is exogenous to unobserved determinants of homeownership is not directly testable. We can, however, test for some plausible sources of endogeneity. For example, in-state tuition rates may be correlated with local housing and labor market conditions, which in turn affect homeownership rates. To see that such omitted variables are unlikely to bias our estimates, compare the estimates across columns 3, 4, and 5 in Tables 5 and 6. Column 4 differs from column 3 by the inclusion of yearly home-state level economic controls: namely, the unemployment rate, log of average weekly wages and the CoreLogic house price index from the subject’s home state measured at age 22. The estimated coefficient on student loan debt is stable across columns 3 and 4, suggesting that these local economic conditions are not driving the results. Furthermore, column 5 includes home state-by-cohort fixed effects which should absorb the effects of all broad economic conditions at the state level. Again, the coefficient of interest is quite stable to this stricter set of controls, suggesting our findings are not substantially biased by market level factors. Further evidence that tuition affects homeownership only through the student loan channel is provided by the absence of any effect of tuition on the control group. The estimated coefficientontuition, whichmeasuresthepartialeffectonthecontrolgroup’shomeownership rate, is not significant and changes sign across specifications. This can be seen by comparing columns 1 through 4 of Tables 5 and 6. Since control group individuals do not pay tuition at public 4-year universities, their homeownership rates should not be correlated with that tuition except through omitted variable bias. We find no evidence that such omitted variables are affecting the correlations between tuition and homeownership. This is essentially a placebo test, validating the contention that we are picking up an effect of tuition rather than the influence of some unobservable factor correlated with it. Another placebo test along these lines is suggested by Belley et al. (2014), which finds that the net tuition paid by lower income students is divorced from the sticker price due to the availability of need-based grants. While we do not observe family income in our data, we do observe Pell grant receipt. We split the sample into those individuals who did and did not receive any Pell grant aid before they turned 23. The former group received need-based aid, and so their student debt burden should be much less influenced by variation in the average in-state charged tuition. We re-estimate the first and second stages of our 2SLS 22

estimator on these two subgroups (including the full vector of controls and state-by-cohort fixed effects) and present the results in Table 8. Among those who received some Pell grant aid, we do not find a significant effect of tuition at public, 4-year universities on student loan debt in the first stage, as shown in column 1. The estimated (placebo) effect on homeownership, shown in column 3, is actually positive, although not significant. In contrast, we show in columns 2 and 4 that there is a strong first and second stage effect among the population that did not receive Pell grant aid, and whose cost of college therefore varied directly with the charged tuition.32 These findingsfurthersuggestthatthecorrelationbetweenthetuitionmeasureandhomeownership is causal. Asconstructed, ourcontrolgroupincludesindividualswhoneverattendedcollege, aswell as students at private schools and public 2-year schools. A potential critique of the exclusion restriction is that tuition rates may reflect economic conditions relevant for college-goers, but not for their peers who did not receive any post-secondary education. If such were the case, our estimates may still be biased by the endogeneity of tuition to college attendeespecific economic shocks, despite the evidence discussed above. We deal with this issue by dropping all observations who had not enrolled in college before age 23 from the sample and re-estimating equations 5 and 6 on the sub-population with at least some college education. Results are presented in column 6 of Table 5 and Table 6. The estimated effect of student loan debt on homeownership is quite similar to that from previous specifications despite the redefined control group, although with a smaller sample the estimates are less precise, only reaching statistical significance in the probit model. In both the linear probability and probit models there is no significant relationship between tuition at public 4-year universities and the homeownership rate of college students that did not attend those universities, as can be seen by the estimated coefficient on the tuition measure. This test suggests that unobserved state level economic conditions specific to the college-educated population are not biasing our results. 32Similar results hold for both subsamples over different specifications or when restricting the sample to only college goers. Results not shown, available upon request. 23

4.6 Endogeneity of Educational Outcomes A further potential issue is bias from sample selection, due to the possibility that tuition rates may affect the relationship between debt and homeownership through the composition of the student population at public 4-year universities. Higher tuitions may deter some students from attending these schools. If such students have notably different propensities to become homeowners than inframarginal students, then our estimates of the effects of debt on homeownership would be biased. However, note that while the homeownership rate of the treatment group falls significantly when tuitions rise, there is no corresponding increase in the homeownership rate of the control group. The control group has a lower homeownership rate than the treatment group, so if individuals with a higher-than-average propensity to become homeowners switch out of the treatment group, then we would expect a significant increase in the control group’s homeownership rate. As can been seen in columns 1 through 4 of Table 7, the estimated effect of tuitions on the homeownership of the control group is small, statistically insignificant, and changes sign across specifications. To further address this potential source of bias, we can test whether our tuition measure affects students’ decisions to attend a public 4-year university. If variation in the average in-state tuition is not correlated with enrollment decisions, then endogenous selection into the treatment group is not a concern. In column 1 of Table 9, we show the results of regressing D —the indicator for having i attended a public 4-year university before age 23—on our tuition measure and state and cohort dummy variables. We find no evidence that changing tuition affects the probability an individual attends such a school across linear probability model and probit specifications. For completeness, in column 2 we show the estimated effect of tuition on the probability of college attendance regardless of sector, for which we find a similar null result. In column 6, we restrict the sample to only those who attended college before age 23, and again find no significant effect of tuition on the probability of attending a public 4-year university. This last test suggests that tuition at public 4-year universities does not induce switching between school sectors, at least for the relatively modest variation in the cost of schooling that our study exploits. Given the above evidence, we believe that defining our treatment group based on attendance at a public 4-year university does not meaningfully bias our estimates. Previous studies have reached mixed conclusions as to the effect of tuition on college 24

attendance. Similar to our estimates, Shao (2015) uses variation in tuition at public institutions to conclude the attendance decision is insensitive to costs. Other studies have found more significant effects. As discussed in a review paper by Deming and Dynarski (2009), this literature often focuses on low income or generally disadvantaged students, and the best identified papers find a $1,000 tuition increase (in 2003 dollars) reduces enrollment by 3 to 4 percentage points. These various findings may be reconcilable if the decision of traditional students to attend public 4-year colleges is price inelastic, while the attendance decision of marginal students considering community colleges or certificate programs is more price sensitive (Denning (2017)).33 Wecantestforthispotentialheterogeneityinpriceelasticitybyregressingtheprobability of attending a public 2-year college against the average tuition charged by such schools in the individual’s home state in the two years after they turned 18. Results of these regressions are shown in column 3 of Table 9. This test is analogous to our baseline experiment, shown in column1ofTable9. Incontrasttothenullresultfortuitionatpublic4-yearschools,wefinda significant effect of public 2-year tuition on enrollment at public 2-year colleges. Specifically, a $1,000 tuition increase (in 2014 dollars) decreases public 2-year college attendance by over 2 percentage points. This effect size is quite similar to previous estimates covered in Deming and Dynarski (2009), especially when correcting for the 28 percentage points of inflation between 2003 and 2014. Tuition may also affect other educational outcomes, such as degree completion, take up of financial aid, or the choice of major. These outcomes may in turn affect the probability of homeownership—for example, completing a college degree may boost the student’s income andallowthemtoaffordahome—whichwouldviolatetheexclusionrestriction. Wetherefore control for these outcomes in our preferred specifications. However, such outcomes may be endogenous to unobservable determinants of homeownership, in which case the estimator would still be inconsistent. Comparing columns 1 and 2 of Tables 5 and 6, we can see that 33In apparent contradiction to our results, Castleman and Long (2016) and Bettinger et al. (2016) find thatgrantaidaffectstheenrollmentofstudentsatpublic4-yearuniversities. However,asarguedinDenning (2017),grantaidmayhavestrongereffectsonthecollegeattendancechoicethanchangesinthestickerprice oftuitiondo—themarginthatwestudy. Thegrantaidprogramsstudiedinthesepaperstargetlowerincome students, which are likely more price sensitive, while changes in the sticker price affects a much larger base of students. Moreover, the size of the aid grants studied is meaningfully larger than the small year-to-year variationintuitionsweuse,whichcouldmakeforqualitativelydifferenteffects. Inparticular,theCalGrant program studied by Bettinger et al. (2016) allows qualifying students to attend public universities tuition free. 25

the estimated effect of student loan debt on homeownership is qualitatively similar regardless of whether additional educational controls are included. We can also test for whether tuition is correlated with any of these outcomes. In columns 4 and 7 of Table 9, we present estimates of the effect of tuition on the probability of completing a bachelor’s degree before age 23, for the general population and the subsample that attended college, respectively. We do not find any significant correlation between tuition and the completion of a bachelor’s degree. In columns 5 and 8, we estimate the effect of tuition on the probability of receiving any federal Pell grants for the full sample and the college-going subsample. Again, there is no significant effect. Finally, we estimate the effect of tuition on the choice of major for those attending a public 4-year school before age 23, modeled as a multinomial logit regression with majors categorized into one of 16 groups. Results are presented in Table 10. We find little evidence of an effect of tuition on major choice—the estimated relative-risk ratio is not significant at the 10-percent level for any major. 4.7 Additional Outcomes As we discuss in Section 2.2, there are multiple channels by which student loans could theoretically affect homeownership. One such channel we hypothesize is the detrimental effect of student loan debt on the borrower’s credit score.34 Increased debt balances could worsen credit scores directly if the credit score algorithm places a negative weight on higher student debt levels.35 Moreover, increased debt could lead to delinquencies which would have a further derogatory effect. The sign of the overall effect is ambiguous, however, as taking out and subsequently repaying student loans may help some borrowers establish a good credit history and thus improve their scores. We estimate the effect of student loan debt on credit scores, regressing the probability that a borrower’s credit score ever fell below one of two underwriting thresholds by a given age against their student loan debt and the usual vector of controls. The thresholds are chosen to roughly correspond to FICO scores of 620 and 680, and fall close to the 25thth 34Unfortunately, we do not have direct measures of the other hypothesized constraints—DTI ratios, down payments, and debt aversion—to test whether these additional channels play a role in explaining our main result. 35Credit scores are generally based on proprietary algorithms, however Goodman et al. (2017) find a negative effect of federal student loan debt on TU Risk Scores. 26

percentile and median credit score among our sample at age 26.36 Results from naive OLS regressions for age 26 are presented in the first and third columns of Table 11. The second and fourth columns present the results of the IV regression. In both cases the instrumented estimates are larger than those from the simple regression, suggesting that a $1,000 increase in student loan debt causes an approximately 2 percentage point increase in the probability a borrower falls below each of the thresholds. It appears that student loan delinquencies play a role in driving down borrower’s credit scores. In columns 5 and 6, we report the estimated effect of student loan debt on the probability of ever having been 30 days or more delinquent on a student loan payment for OLS and IV specifications. The IV results again are larger than the OLS estimates, and suggest that a $1,000 increase in debt increases the probability of missing a payment by 1.5 percentage points. These results suggest that borrowers are more likely to miss payments when their debt burdens are greater, and the resulting damage to their credit scores makes qualifying for a mortgage more difficult. In Figure 4 we plot the estimated effect of student loan debt on having a sub-median credit score (corresponding to a FICO score of approximately 680), and on ever having been delinquentonastudentloanpayment,byage,from22to32. Theestimatesarenotsignificant at first, but grow in magnitude and remain persistently significant after age 26. These results suggest access to homeownership could be impaired by student loan debt’s negative effect on credit scores. However, because student loan debt begins to have a significant effect on both homeownership and credit scores at about the same age, we cannot rule out the possibility of reverse causality (i.e., that mortgage debt improves credit scores). Another source of adjustment through which student loans could be affecting the housing market is by influencing the amount of mortgage debt borrowed. The direction of the effect is theoretically ambiguous. If DTI ratios or down payment constraints are binding, borrowersmaysubstitutetowardsmallermortgagesinresponsetohigherstudentdebtlevels. Alternatively, borrowers could respond to increased debt by delaying the timing of their first home purchase. If the home purchase decision is delayed to a point in the life cycle at which the borrower has a greater demand for housing (due, for example, to the presence of children), mortgage balances could conceivably rise with student debt. In the same direction, 36A FICO score of 620 is shown by Laufer and Paciorek (2016) to be a relevant underwriting threshold for mortgage lenders. We thank Ezra Becker and TransUnion for guidance in suggesting 680 as another significant threshold for underwriting. 27

student loan debt could affect the composition of the population of homeowners. That is, if marginal homeowners demand smaller mortgages than their inframarginal peers, then increased student loan debt would tend to increase the average observed mortgage as the marginal homeowners are selected out of the sample. In the first column of Table 12, we present the results from regressing the loan amount of the first mortgage we observe for each individual against their student loan debts and the usual vector of controls. Only borrowers who obtain a mortgage by age 32 are included in this regression. The estimated partial correlation is positive and statistically significant, implying a $1,000 increase in student loan debt is associated with approximately $330 higher mortgage balances. This naive estimate is likely to be biased by omitted variables similar to those that bias estimates of the effect of student loan debt on homeownership. We apply the same instrumental variable solution, and present results in the second column of Table 12. This point estimate suggests that student loan debt causes substantially lower average mortgage balances among the population of homeowners. The standard errors are very large, however, and the result is not close to statistically significant. While the point estimates suggest $1,000 in additional student loans reduces the expected first mortgage balance by almost $3,200, we cannot rule out that the additional student loan debt actually increases the first mortgage balance by almost $2,500. 5 Discussion of Findings Our baseline estimates from the IV specification in Table 5 indicate that a $1,000 increase in loans disbursed to a student before age 23 leads to a 1 to 2 percentage point reduction in homeownership by the student’s mid-twenties. To put the magnitude of the reduction into a life cycle context, Figure 5 plots the average age-profile of homeownership for the treatment group of public 4-year university attendees (the black line).37 The homeownership rate for these individuals rises sharply through young adulthood, from about 5 percent at age 22 to about 60 percent by age 32. For comparison, the red line simulates the homeownership rate under the counterfactual assumption that each individual in the treatment group is 37As a reminder, the definition of homeownership we use in this paper is an absorbing state. Individuals whoclosedtheirmortgageaccount(eitherbecausetheypaidoffthemortgageorwereforeclosedon)arestill counted as homeowners in our figures. 28

burdened with a $1,000 increase in student loan debt before age 23, using estimates from the specification of column 5 in Table 5. While the effect of the $1,000 increase in student loan debt leads to a meaningful reduction in homeownership for households in their mid- to late twenties, by age 32 the estimated 0.5 percent reduction in homeownership is a relatively small fraction of the actual homeownership rate. The apparent attenuation of the estimated effect of student debt in borrowers’ late 20s and early 30s suggests that student loan debt may cause a delay, rather than a permanent reduction, inthehomeownershiprate. Inotherwords, increasingstudentloandebtby$1,000 may induce a rightward, rather than a downward, shift in the age-profile of homeownership. Interpolating linearly between the estimated points of the counterfactual homeownership curve, we calculate that with a $1,000 increase in student loan debt, the homeownership rate of a given cohort would be delayed by 2.5 months at age 26. Due to the steepness of the homeownership-age profile during the early years of adult life, a fairly modest delay in the timing of home buying translates to a substantial decrease in the probability of homeownership at any particular age. Evenifstudentloansaffectonlythetimingofhomebuying, withnoeffectontheultimate attainment of homeownership, there are still significant aggregate implications. The overall homeownership rate would be lower than in a counterfactual world with less student loan debt, as each successive generation is delayed in becoming homeowners. Home equity is the major form of wealth holding for most households and housing services are a significant fraction of national income, so even a small change in homeownership can have wide ranging effects.38 What are the policy implications of our findings? If policymakers are interested in raising the homeownership rate among the young, our results suggest there may be additional value from promoting student loan forgiveness. Furthermore, policies directed at slowing the growth of tuition may aid student borrowers in becoming homeowners. As we show that damage to credit scores from delinquencies on student loans are a likely channel by which debts can affect homeownership, policies aimed at preventing delinquencies may also be beneficial. For example, income driven repayment plans for student loans (such as the 38In the 2013 Survey of Income and Program Participation, the median homeowner household held over $80,000 in home equity. Housing services account for 15-18% of GDP according to the Bureau of Economic Analysis. 29

Income Based Repayment and Pay As You Earn programs offered by the Department of Education) which tie debtor’s scheduled payments to their disposable income, may offer relief. Additionally, one might be tempted to interpret our findings as evidence supporting a reduction in access to federal student debt, by—for example—lowering federal student loan limits. However, our analysis does not support such a conclusion. In particular, we do not estimate the effect of access to student loans, which could directly affect students schooling choices. If access to student loans allows for increased educational attainment, the reduction in access could lead to a wide array of negative outcomes, ranging from reduced economic efficiency to increasing income inequality within and across generations (Avery and Turner (2012)). Furthermore, by lowering incomes of young individuals, reducing access to student loans could even cause lower homeownership rates.39 6 Conclusions In summary, this paper estimates the effect of student loan debt on subsequent homeownership rates. We find that a $1,000 increase in student loan debt causes a 1 to 2 percentage point drop in the homeownership rate of student loan borrowers during their mid-twenties. These results represent a larger effect than estimates attempting to deal with the endogeneity of student loan debt using a selection-on-observables approach have found. We also show that student loan debt has a negative effect on borrowers’ credit scores, potentially excluding some indebted students from the mortgage market. Our findings have implications for several recent trends and policy proposals. Tuition ratescontinuetorise, sotheamountsstudentswillneedtoborrowmayincreaseinthefuture. Increased debt levels could continue to depress homeownership rates for future cohorts of college students. Measures taken to reduce tuition—or to curb borrowing beyond what is necessarytofundattendance—couldfightthistrend. Similarly, ourresultsprovideameasure of how effective student loan forgiveness programs could be at increasing the homeownership rate of young adults. Limiting or expanding students’ access to education loans in general, however, would have ramifications that are beyond the scope of this study. In particular, if 39A large body of literature has found that returns to education remain high and indeed continues to grow—see Lochner and Monge-Naranjo (2014) and studies cited therein. 30

student loans allow individuals to access college education—or, more broadly, acquire more of it—student loan debt could have a positive effect on homeownership, as long as the return to this additional education allows individuals to sufficiently increase their future incomes. In extrapolating our results to the present day, we also have to consider some significant recent changes to the mortgage market. Individuals in our sample turned 23 years old between 1997 and 2004. Thus, the majority of our cohorts were entering their prime home-buying years in a relatively easy environment for mortgage credit. Since the housing and financial crisis, underwriting standards have tightened substantially. It is possible that student loan debt acts as an even greater drag on homeownership now that lenders are more sensitive to DTI ratios, credit scores, and low down payments. However, as the recovery continues and underwriting conditions ease, mortgage market conditions similar to the late 1990s and early 2000s may re-emerge. The growing popularity of income-driven repayment plans further complicates the picture, as it is not immediately clear how these plans moderate the link between initial student loan debt and homeownership. On the one hand, enrollment in income-driven repayment plans reduces the ratio of student loan payments relative to income, thereby relaxing the DTI constraint. On the other hand, it can extend the repayment period significantly relative to a 10-year plan, thereby potentially increasing the total interest paid by the student loan borrower over the life of the loan. We hope that further studies using even more recent data will be able to shine additional light on the issue. 31

References Akers, B., “Reconsidering the Conventional Wisdom on Student Loan Debt and Home Ownership,” The Brookings Institution Brown Center Chalkboard. Available at http://www.brookings.edu/blogs/brown-center-chalkboard/posts/2014/ 05/08-student-loan-debt-and-home-ownership-akers, 2014. Avery, C. and S. Turner, “Student Loans: Do College Students Borrow Too Much—Or Not Enough?” Journal of Economic Perspectives, 26:165–192, 2012. Belley, P., M. Frenette and L. Lochner, “Post-secondary attendance by parental income in the US and Canada: Do financial aid policies explain the differences?” Canadian Journal of Economics/Revue canadienne d’´economique, 47(2):664–696, 2014. Bettinger, E., O. Gurantz, L. Kawano and B. Sacerdote, “The Long Run Impacts of Merit Aid: Evidence from California’s Cal Grant,” NBER Working Paper No. 22347, 2016. Bleemer, Z., M. Brown, D. Lee and W. Van der Klaauw, “Debt, Jobs, or Housing: What’s Keeping Millennials at Home?” FRB of New York Staff Report, (700), 2014. Brown, M., S. Caldwell and S. Sutherland, “Young Student Loan Borrowers Remained on the Sidelines of the Housing Market in 2013,” Federal Reserve Bank of New York, 2013. Castleman, B. and B. T. Long, “Looking beyond Enrollment: The Causal Effect of Need- Based Grants on College Access, Persistence, and Graduation,” Journal of Labor Economics, 34:1023–1073, 2016. Cooper, D. and C. Wang, “Student Loan Debt and Economic Outcomes,” FRB Boston Current Policy Perspectives, (14-7), 2014. Deming, D. and S. Dynarski, “Into College, Out of Poverty? Policies to Increase the Postsecondary Attainment of the Poor,” NBER Working Paper No. 15387, 2009. Denning, J., “College on the Cheap: Consequences of Community College Tuition Reductions,” American Economic Journal: Economic Policy, 9:158–188, 2017. 32

Dettling, L. and J. Hsu, “Returning to the Nest: Debt and Parental Co-Residence Among Young Adults,” Finance and Economics Discussion Series 2014-80, 2014, Washington: Board of Governors of the Federal Reserve System. Dynarski, S., S. Hemelt and J. Hyman, “The Missing Manual: Using National Student Clearninghouse Data to Track Postsecondary Outcomes,” NBER Working Paper 19522, 2013. Field, E., “Educational Debt Burden and Career Choice: Evidence from a Financial Aid Experiment at NYU Law School,” American Economic Journal: Applied Economics, 1:1– 21, 2009. Gicheva, D., “Student Loans or Marriage? A Look at the Highly Educated,” 53:207–216, 2016. Gicheva, D. and J. Thompson, “The Effects of Student Loans on Long-Term Household Financial Stability,” Economics of Education Review, 2014. Goodman, S. and A. Henriques, “Attendance Spillovers between Public and For-Profit Colleges: Evidence from Statewide Variation in Appropriations for Higher-Education,” Finance and Economics Discussion Series 2016-098, 2015, Washington: Board of Governors of the Federal Reserve System. Goodman, S., A. Henriques and A. Mezza, “Where Credit is Due: The Relationship between Family Background and Credit Health,” Finance and Economics Discussion Series 2017- 032, 2017, Washington: Board of Governors of the Federal Reserve System. Houle, J. and L. Berger, “The End of the American Dream? Student Loan Debt and Homeownership Among Young Adults?” Third Way Discussion Paper, 2015. Kuipers, P. and C. Wise, “Are Student Loans Hurting Millenials?” The RMA Journal, 2016. Laufer, S. and A. Paciorek, “The Effects of Mortgage Credit Availability: Evidence from Minimum Credit Score Lending Rules,” Finance and Economics Discussion Series 2016- 098, 2016, Washington: Board of Governors of the Federal Reserve System. Lochner, L. and A. Monge-Naranjo, “Student Loans and Repayment: Theory, Evidence and Policy,” CIBC Working Paper Series, 2014, #2014-5. 33

Loewenstein, G. and R. Thaler, “Anomalies and intertemporal choice,” Journal of Economic Perspectives, 3:181–193, 1989. Looney, A.andC.Yannelis, “ACrisisinStudentLoans? HowChangesintheCharacteristics of Borrowers and in the Institutions They Attended Contributed to Rising Loan Defaults,” Brookings Papers on Economic Activity, 2015. Lovenheim, M. F., “The effect of liquid housing wealth on college enrollment,” Journal of Labor Economics, 29(4):741–771, 2011. Malcolm, L. and A. Down, “The Impact of Undergraduate Debt on the Graduate School Enrollment of STEM Baccalaureates,” The Review of Higher Education, 35:265–305, 2012. Mezza, A. and K. Sommer, “A Trillion-Dollar Question: What Predicts Student Loan Delinquencies?” Journal of Student Financial Aid, 46:16–54, 2016. Mezza, A., K.SommerandS.Sherlund, “StudentLoansandHomeownershipTrends,” FEDS Note, 2014. Palameta, B.andJ.Voyer, “WillingnesstoPayforPost-secondaryEducationAmongUnderrepresented Groups,” Toronto: Higher Education Quality Council of Ontario, 2010. Rothstein, J. and C. Rouse, “Constrained After College: Student Loans and Early-career Occupational Choices,” Journal of Public Economics, 95:149–163, 2011. Shao, L., “Debt, Marriage and Children: The Effects of Student Loans on Marriage and Fertility,” 2015. Stone, C., C. Van Horn and C. Zukin, “Chasing the American Dream: Recent College GraduatesandtheGreatRecession.WorkTrends.”John J. Heldrich Center for Workforce Development, 2012. Thaler, R., “The Winner’s Curse: Paradoxes and Anomalies of Economic Life,” 1992. Weeden, D., “Tuition Policy,” 2015. Weerts, D., T. Sanford and L. Reinert, “College Funding in Context: Understanding the Difference in Higher Education Appropriations across the States,” Demos.org, 2012. 34

Zhang, L., “Effects of College Educational Debt on Graduate School Attendance and Early Career and Lifestyle Choices,” Education Economics, 21:154–175, 2013. 35

Table 1: Summary Statistics Variables Obs Mean Std. Dev. Min Max Homeownership Rate Ownat22 33,435 0.068 Ownat23 33,435 0.100 Ownat24 33,435 0.143 Ownat25 33,435 0.195 Ownat26 33,435 0.243 Ownat27 33,435 0.289 Ownat28 33,435 0.332 Ownat29 33,435 0.369 Ownat30 33,435 0.401 Ownat31 33,435 0.424 Ownat32 33,435 0.445 Student Loan Debt Measures StudentLoansDisbursed(in$1,000) 33,435 4.413 9.778 0 153.527 StudentLoansDisbursed(in$1,000),ConditionalonDebt>0 9,720 15.179 12.863 0.001 153.527 Tuition(in$1,000) 33,435 17.480 5.410 5.705 38.444 School Sector Controls EverPublic4-Year 33,435 0.262 EverPublic2-Year 33,435 0.248 EverPrivate4-YearNot-for-profit 33,435 0.116 EverPrivate2-YearNot-for-profit 33,435 0.008 EverPrivateFor-profit 33,435 0.047 Degree and Pell Grant Controls NoCollege 33,435 0.458 Associate’s/Certificate 33,435 0.030 Bachelor’s 33,435 0.113 Master’sorMore 33,435 0.001 DegreeofUnknownType 33,435 0.008 EverPell 33,435 0.206 Cohort 1990-91 33,435 0.045 1991-92 33,435 0.115 1992-93 33,435 0.113 1993-94 33,435 0.109 1994-95 33,435 0.113 1995-96 33,435 0.113 1996-97 33,435 0.113 1997-98 33,435 0.118 1998-99 33,435 0.108 1999-00 33,435 0.054 Yearly State Controls AverageWeeklyWages(in$1,000,HomeState) 33,435 1.025 0.169 0.783 1.792 UnemploymentRate(HomeState) 33,435 5.0 1.1 2.3 8.8 HousePriceIndex(HomeState) 33,435 100.3 19.5 63.6 206.7 Additional Outcomes MortgageAmount(in$1,000) 10,475 152.261 112.4 0.148 2,600 EverSubprime 33,435 0.652 EverDeepSubprime 33,435 0.421 EverDelinquent 33,435 0.169 Note: Homeownership rate is measured as ever having a mortgage loan by a given age. Student loans disbursed are measured as the total amount of federal student loans disbursed to individuals before age 23. Tuition is the average in-state tuition at public 4-year colleges in the individual’s home state over the four years following his or her 18th birthday. Student loans and Tuition are in constant 2014 dollars. School sector, degree, and Pell Grant controls represent the sectors, the attained degree and whether individuals received Pell Grants before age 23. Cohorts are defined as the school-year in which individuals turn 18 years old. Yearly state controls represent local economic conditions in individuals’ home-state at age 22. Mortgage amount represents the size of the first mortgage amount observed in the dataset between ages 22 and32. Eversubprimeanddeepsubprimerepresentwhetherindividualshadscoresbelowthe50thand25th percentile, respectively, between the ages of 22 and 32. Ever delinquent represents whether individuals were delinquent on student loan debt for at least 30 days between the ages of 22 and 32. 36

Table 2: Selection on Observables: OLS Probability of Homeownership by Age 26 Variable (1) (2) (3) (4) (5) (6) Student Loans Disbursed -0.000290 -0.00120*** -0.00127*** -0.00126*** -0.00127*** -0.00129*** (0.000312) (0.000361) (0.000353) (0.000353) (0.000353) (0.000351) Tuition -0.000603 -0.00156** -0.000114 -0.000524 0.000492 (0.000672) (0.000681) (0.00213) (0.00257) (0.00381) Ever Public 4-Year 0.0745*** 0.0238*** 0.0174** 0.0173** 0.0155* 0.0142* (0.00627) (0.00832) (0.00819) (0.00820) (0.00825) (0.00825) No College -0.0621*** -0.0578*** -0.0579*** -0.0593*** (0.00957) (0.00941) (0.00942) (0.00945) Associate’s/Certificate 0.197*** 0.192*** 0.192*** 0.198*** 0.190*** (0.0338) (0.0334) (0.0334) (0.0337) (0.0336) Bachelor’s 0.221*** 0.230*** 0.229*** 0.235*** 0.230*** (0.0281) (0.0282) (0.0282) (0.0283) (0.0283) Master’s or More 0.327*** 0.350*** 0.350*** 0.349*** 0.354*** (0.0772) (0.0777) (0.0777) (0.0776) (0.0778) Degree of Unknown Type 0.298*** 0.291*** 0.290*** 0.291*** 0.285*** (0.0491) (0.0482) (0.0481) (0.0485) (0.0483) Ever Public 2-Year -0.00878 0.000226 0.000181 -0.00115 -0.000996 (0.00889) (0.00862) (0.00862) (0.00866) (0.00871) Ever Private 4-Year -0.00762 -0.00264 -0.00265 -0.00387 -0.00289 Not-for-profit (0.0101) (0.00999) (0.00997) (0.0100) (0.0101) Ever Private 2-Year 0.0649** 0.0585* 0.0585* 0.0637* 0.0573* Not-for-profit (0.0322) (0.0327) (0.0327) (0.0326) (0.0333) Ever Private For-profit -0.0279** -0.0257** -0.0257** -0.0285** -0.0260** (0.0126) (0.0122) (0.0122) (0.0122) (0.0122) Ever Pell -0.0461*** -0.0420*** -0.0420*** -0.0415*** -0.0428*** (0.00734) (0.00720) (0.00720) (0.00718) (0.00726) Avg. Weekly Wages (in 0.0158 $1,000, Home State) (0.0156) Unemployment Rate 0.00218 (Home State) (0.00482) Corelogic House Price -3.84e-05 Index (Home State) (0.000314) Constant 0.234*** 0.293*** 0.176*** 0.160 0.293*** 0.0596 (0.0111) (0.0142) (0.0517) (0.0978) (0.106) (0.144) College Major Controls NO YES YES YES YES YES Home State/Cohort FE NO NO YES YES NO YES Home State by Cohort NO NO NO NO YES NO FE Observations 33,435 33,435 33,435 33,435 33,435 18,121 R-squared 0.006 0.023 0.039 0.039 0.022 0.051 Note: This table reports OLS estimates of the effect of student loans on the probability of becoming a homeowner by age 26. Variables defined as in Table 1. Column (1) only controls for tuition and whether individualseverattendedaPublic4-yearcollegebeforeage23. Column(2)addsseveraleducationalcontrols summarized in Table 1 and 14 college major indicator variables described in the Appendix of Mezza and Sommer(2016). Omitteddegreecategoryishavingattendedcollegebeforeage23withoutgettingadegreeby that age. Column (3) adds home state and cohort fixed effects. Column (4) includes local economic controls measured at the home state level when individuals were 22 years old. Column (5) builds on column (3) by adding home state by cohort fixed effects. Column (6) restricts the sample to individuals who attended any post-secondaryschoolingbeforeturning23. Sampleisallindividualsfromanationally-representativecohort of 23-to-31-year-old individuals with credit records in 2004 after applying the filters described in Section 3. Student loans disbursed and tuitions are recorded in 1000s of 2014 dollars. Standard errors in parentheses (clustered at the home state by cohort level). ***,**, and * denote significance at 1%, 5%, and 10%. 37

Table 3: Selection on Observables: Probit Probability of Homeownership by Age 26 Variable (1) (2) (3) (4) (5) (6) Student Loans Disbursed -0.000254 -0.00113*** -0.00121*** -0.00121*** -0.00124*** -0.00132*** (0.000306) (0.000356) (0.000351) (0.000351) (0.000357) (0.000373) Tuition -0.000595 -0.00158** 6.65e-05 -0.000213 0.000822 (0.000675) (0.000693) (0.00234) (0.00275) (0.00405) Ever Public 4-Year 0.0743*** 0.0222*** 0.0160** 0.0159** 0.0143* 0.0139 (0.00627) (0.00816) (0.00807) (0.00808) (0.00821) (0.00850) Degree/Sector/Pell Grant/College Major NO YES YES YES YES YES Controls Home State Economic NO NO NO YES NO NO Controls Home State/Cohort FEs NO NO YES YES NO YES Home State by Cohort NO NO NO NO YES NO FE Observations 33,435 33,435 33,435 33,435 33,310 18,121 Pseudo R2 0.00486 0.0191 0.0334 0.0335 0.0454 0.0423 Note: This table reports Probit estimates of the effect of student loans on the probability of becoming a homeowner by age 26. Marginal probabilities reported. See Tables 1 for variable definitions and 2 for sampleselectionandspecificationdetails. Studentloansdisbursedandtuitionsarerecordedin1000sof2014 dollars. Standard errors in parentheses (clustered at the home state by cohort level). ***,**, and * denote significance at 1%, 5%, and 10%. 38

Table 4: IV Estimation: 1st Stage Total Federal Student Loans Disbursed before Age 23 Variable (1) (2) (3) (4) (5) (6) Instrument: Tuition x 0.0890*** 0.158*** 0.157*** 0.157*** 0.153*** 0.0993*** Ever Public 4-Year (0.0296) (0.0257) (0.0258) (0.0258) (0.0261) (0.0310) Tuition 0.173*** 0.0341*** 0.0532 0.0250 0.0758 (0.0152) (0.00759) (0.0427) (0.0445) (0.0779) Ever Public 4-Year 5.555*** 1.497*** 1.548*** 1.545*** 1.610*** 2.513*** (0.472) (0.449) (0.454) (0.454) (0.457) (0.521) No College -2.103*** -2.066*** -2.064*** -2.060*** (0.252) (0.252) (0.252) (0.254) Associate’s/Certificate -0.0136 -0.0669 -0.0626 0.0158 -0.134 (0.604) (0.599) (0.599) (0.608) (0.600) Bachelor’s 3.214*** 3.261*** 3.265*** 3.333*** 3.271*** (0.562) (0.557) (0.557) (0.567) (0.558) Master’s or More 4.061* 4.288** 4.282** 4.337** 4.466** (2.140) (2.135) (2.133) (2.136) (2.134) Degree of Unknown Type -0.0933 -0.166 -0.153 -0.0140 -0.193 (1.064) (1.066) (1.065) (1.076) (1.069) Ever Public 2-Year -2.580*** -2.477*** -2.473*** -2.490*** -2.385*** (0.206) (0.204) (0.203) (0.204) (0.208) Ever Private 4-Year 8.303*** 8.303*** 8.305*** 8.310*** 8.239*** Not-for-profit (0.282) (0.282) (0.282) (0.283) (0.285) Ever Private 2-Year 1.867*** 1.861*** 1.872*** 1.847*** 1.866*** Not-for-profit (0.558) (0.562) (0.563) (0.577) (0.570) Ever Private For-profit 1.871*** 1.945*** 1.944*** 1.940*** 1.987*** (0.297) (0.301) (0.301) (0.303) (0.307) Ever Pell 4.155*** 4.120*** 4.122*** 4.124*** 4.114*** (0.168) (0.171) (0.171) (0.172) (0.175) Avg. Weekly Wages (in -0.0962 $1,000, Home State) (0.223) Unemployment Rate -0.111 (Home State) (0.0767) Corelogic House Price -0.00933* Index (Home State) (0.00546) Constant -0.587*** 1.587*** 0.942 3.537** 2.076*** 0.627 (0.200) (0.258) (1.080) (1.544) (0.245) (1.832) College Major Controls NO YES YES YES YES YES Home State/Cohort FEs NO NO YES YES NO YES Home State by Cohort NO NO NO NO YES NO FEs Observations 33,435 33,435 33,435 33,435 33,435 18,121 F-stat 18.877 82.317 79.652 79.897 74.319 11.94 R-squared 0.138 0.379 0.384 0.384 0.363 0.261 Note: This table reports OLS estimates of the effect of tuition on federal student loans disbursed at the individual level. See Tables 1 for variable definitions and 2 for sample selection and specification details. Studentloansdisbursedandtuitionsarerecordedin1000sofyear2014dollars. Standarderrorsinparentheses (clustered at the home state by cohort level). ***,**, and * denote significance at 1%, 5%, and 10%. 39

Table 5: 2SLS Estimation: 2nd Stage Probability of Homeownership by Age 26 Variable (1) (2) (3) (4) (5) (6) Student Loans Disbursed -0.0295* -0.0192** -0.0151** -0.0150** -0.0161** -0.0234 (0.0164) (0.00776) (0.00746) (0.00745) (0.00788) (0.0145) Tuition 0.00516 -0.000174 0.00126 0.000478 0.00296 (0.00315) (0.000782) (0.00235) (0.00265) (0.00392) Ever Public 4-Year 0.278** 0.0970*** 0.0741** 0.0738** 0.0762** 0.106* (0.114) (0.0329) (0.0322) (0.0321) (0.0338) (0.0606) No College -0.0983*** -0.0852*** -0.0851*** -0.0885*** (0.0186) (0.0174) (0.0174) (0.0182) Associate’s/Certificate 0.196*** 0.191*** 0.191*** 0.198*** 0.187*** (0.0357) (0.0348) (0.0347) (0.0354) (0.0368) Bachelor’s 0.280*** 0.275*** 0.275*** 0.285*** 0.302*** (0.0425) (0.0410) (0.0410) (0.0428) (0.0602) Master’s or More 0.399*** 0.408*** 0.407*** 0.412*** 0.452*** (0.0952) (0.0911) (0.0910) (0.0928) (0.114) Degree of Unkown Type 0.297*** 0.289*** 0.289*** 0.291*** 0.281*** (0.0515) (0.0493) (0.0492) (0.0498) (0.0521) Ever Public 2-Year -0.0543*** -0.0334* -0.0332* -0.0374* -0.0530 (0.0209) (0.0195) (0.0194) (0.0206) (0.0350) Ever Private 4-Year 0.142** 0.112* 0.112* 0.119* 0.178 Not-for-profit (0.0652) (0.0631) (0.0630) (0.0665) (0.119) Ever Private 2-Year 0.0995*** 0.0850** 0.0850** 0.0915*** 0.0996** Not-for-profit (0.0343) (0.0351) (0.0351) (0.0354) (0.0422) Ever Private For-profit 0.00670 0.00185 0.00170 0.000849 0.0183 (0.0209) (0.0203) (0.0203) (0.0210) (0.0323) Ever Pell 0.0283 0.0147 0.0144 0.0193 0.0480 (0.0327) (0.0315) (0.0314) (0.0331) (0.0598) Average Weekly Wages 0.0143 (Home State) (0.0155) Unemployment Rate 0.000661 (Home State) (0.00481) House Price Index -0.000161 (Home State) (0.000319) Constant 0.206*** 0.309*** 0.177*** 0.196** 0.296*** 0.156* (0.0175) (0.0178) (0.0530) (0.0982) (0.0179) (0.0797) College Major Controls NO YES YES YES YES YES Home State/Cohort FEs NO NO YES YES NO YES Home State by Cohort NO NO NO NO YES NO FEs Observations 33,435 33,435 33,435 33,435 33,435 18,121 Note: This table reports 2SLS estimates of the effect of student loans on the probability of becoming a homeowner by age 26. Student loans are instrumented for using the interaction between tuition and an indicator variable for whether the individual ever attended a Public 4-year college before age 23. See Tables 1 for variable definitions and 2 for sample selection and specification details. Student loans disbursed and tuitions are recorded in 1000s of 2014 dollars. Standard errors in parentheses (clustered at the home state by cohort level). ***,**, and * denote significance at 1%, 5%, and 10%. 40

Table 6: IV-Probit Estimation: 2nd Stage Probability of Homeownership by Age 26 Variable (1) (2) (3) (4) (5) (6) Student Loans Disbursed -0.0230*** -0.0158*** -0.0125** -0.0125** -0.0134** -0.0187** (0.00875) (0.00602) (0.00624) (0.00623) (0.00637) (0.00869) Tuition 0.00403** -0.000288 0.00123 0.000665 0.00265 (0.00183) (0.000779) (0.00238) (0.00268) (0.00321) Ever Public 4-Year 0.221*** 0.0806*** 0.0615** 0.0613** 0.0634** 0.0845** (0.0496) (0.0249) (0.0265) (0.0265) (0.0271) (0.0354) Degree/Sector/Pell Grant/College Major NO YES YES YES YES YES Controls Home State Economic NO NO NO YES NO NO Controls Home State/Cohort FEs NO NO YES YES NO YES Home State by Cohort NO NO NO NO YES NO FEs Observations 33,435 33,435 33,435 33,435 33,310 18,121 Note: This table reports probit estimates of the effect of student loans on the probability of becoming a homeowner by age 26. Student loans are instrumented for using the interaction between tuition and an indicator variable for whether the individual ever attended a Public 4-year college before age 23. Marginal probabilities reported. See Tables 1 for variable definitions and 2 for sample selection and specification details. Student loans disbursed and tuitions are recorded in 1000s of 2014 dollars. Standard errors in parentheses (clustered at the home state by cohort level). ***,**, and * denote significance at 1%, 5%, and 10%. 41

Table 7: Reduced Form Effect of Instrument on Homeownership Probability of Homeownership by Age 26 Variable (1) (2) (3) (4) (5) (6) Instrument: Tuition x -0.00263** -0.00304*** -0.00236** -0.00236** -0.00246** -0.00232* Ever Public 4-Year (0.00117) (0.00112) (0.00111) (0.00111) (0.00113) (0.00130) Tuition 5.28e-05 -0.000828 0.000454 0.000102 0.00118 (0.000676) (0.000688) (0.00218) (0.00262) (0.00308) Ever Public 4-Year 0.114*** 0.0682*** 0.0507** 0.0506** 0.0504** 0.0469** (0.0198) (0.0204) (0.0203) (0.0203) (0.0206) (0.0231) Degree/Sector/Pell Grant/College Major NO YES YES YES YES YES Controls Home State Economic NO NO NO YES NO NO Controls Home State/Cohort FEs NO NO YES YES NO YES Home State by Cohort NO NO NO NO YES NO FEs Observations 33,435 33,435 33,435 33,435 33,310 18,121 Note: This table reports OLS estimates of the effect of the interaction between tuition and an indicator variable for whether the individual ever attended a Public 4-year college before age 23 on homeownership, measured at age 26. Marginal probabilities reported. See Tables 1 for variable definitions and 2 for sample selection and specification details. Tuitions are recorded in 1000s of 2014 dollars. Standard errors in parentheses (clustered at the home state by cohort level). ***,**, and * denote significance at 1%, 5%, and 10%. 42

Table 8: First and Second Stage by Pell Grant Receipt Student Loans Disbursed Homeownership by Age 26 With Pell Without Pell With Pell Without Pell Variable (1) (2) (3) (4) Instrument: Tuition x 0.0873 0.203*** Ever Public 4-Year (0.0574) (0.0318) Student Loans Disbursed 0.0233 -0.0209*** (0.0256) (0.008) Ever Public 4-Year 4.545*** -0.352 -0.129 0.0741*** (0.900) (0.593) (0.153) (0.0272) Degree/Sector/Pell Grant/College Major YES YES YES YES Controls Home State Economic NO NO NO NO Controls Home State/Cohort FEs NO NO NO NO Home State by Cohort YES YES YES YES FEs Observations 6,888 26,546 6,888 26,546 Note: Thistablereportsfirstandsecondstage2SLSestimatesoftheeffectofstudentloansontheprobability of becoming a homeowner by age 26. Student loans are instrumented for using the interaction between tuition and an indicator variable for whether the individual ever attended a Public 4-year college before age 23. Marginal probabilities reported. See Tables 1 for variable definitions and 2 for sample selection and specification details.. Columns 1 and 3 are restricted to students who received Pell Grant aid. Columns 2 and 4 are restricted to students who did not receive Pell Grant aid. Student loans disbursed and tuitions are recorded in 1000s of 2014 dollars. Standard errors in parentheses (clustered at the home state by cohort level). ***,**, and * denote significance at 1%, 5%, and 10%. 43

semoctuO noitacudE fo ytienegodnE :9 elbaT seednettAegelloC elpmaSlluF cilbuPrevE cilbuPrevE cilbuPrevE llePynA s’rolehcaB raeY-4 llePynA s’rolehcaB raeY-2 egelloCynA raeY-4 )8( )7( )6( )5( )4( )3( )2( )1( )raeY-4cilbuP(noitiuT 47100.0 57200.0 07100.0 025000.0- 12100.0 93500.0- 712000.0 SLO )35300.0( )77200.0( )92300.0( )44200.0( )08100.0( )26300.0( )83200.0( 78100.0 51300.0 12200.0 035000.0- 48100.0 49500.0- 694000.0tiborP )36300.0( )91300.0( )72300.0( )44200.0( )58100.0( )08300.0( )04200.0( )raeY-2cilbuP(noitiuT **7120.0- SLO )9010.0( **1320.0tiborP )7010.0( SEY SEY SEY SEY SEY SEY SEY SEY sEFetatSemoH SEY SEY SEY SEY SEY SEY SEY SEY sEFtrohoC 121,81 121,81 121,81 534,33 534,33 423,33 534,33 534,33 snoitavresbO semoctuo lanoitacude lareves no noitiut fo tceffe eht fo setamitse tiborp dna SLO stroper elbat sihT :etoN elihw ,setamitse SLO ot dnopserroc swor driht dna tsrfi eht ni stluseR .)sredaeh nmuloc eht yb detoned( cilbuP( noitiuT .nwohs seitilibaborp lanigraM .setamitse tiborP ot swor htruof dna dnoces eht ni stluser ruofehthguorhtetatsemohs’laudividniehtnisegellocraey-4cilbuptanoitiutetats-niegarevaehtsi)raeY-4 raey-2cilbuptanoitiutegarevaehtsi)raeY-2cilbuP(noitiuTelihw ,yadhtribht81rehrosihgniwollofsraey snaoltnedutS .yadhtribht81rehrosihgniwollofsraeyowtehthguorhtetatsemohs’laudividniehtnisegelloc )1(snmulocnielpmaS .snoitinfiedelbairavrehtorof1elbaTeeS .4102fosrallodtnatsnocnieranoitiuTdna htiw slaudividni dlo-raey-13-ot-32 fo trohoc evitatneserper-yllanoitan a morf slaudividni lla si )5( hguorht si )8( dna )6( snmuloc ni elpmas ehT .3 noitceS ni debircsed sretlfi eht gniylppa retfa 4002 ni sdrocer tiderc ni srorre dradnatS .32 ega erofeb gniloohcs yradnoces-tsop yna dednetta evah ohw slaudividni ot detcirtser dna ,%5 ,%1 ta ecnacfiingis etoned * dna ,**,*** .)level trohoc yb etats emoh eht ta deretsulc( sesehtnerap .%01 44

Table 10: Effect of Tuition on Major Choice Major Category Coef. Std. Err. 1 0.133 (0.184) 2 0.0275 (0.0535) 3 -0.0104 (0.0510) 4 0.0877 (0.0623) 5 0.147 (0.0944) 6 0.0923 (0.111) 7 -0.0379 (0.0369) 8 0.00490 (0.0656) 9 0.0599 (0.0662) 10 -0.0430 (0.0422) 11 0.122 (0.0916) 12 0.246 (0.330) 13 0.258 (0.174) 14 -0.107 (0.0795) 15 0.0409 (0.0581) Observations 8,774 Note: This table reports multinomial logit estimates for the effect of tuition on major choice. Major categories are defined as described in the Appendix of Mezza and Sommer (2016) and the omitted category is having no degree (thus, no major) before age 23. Tuition is the average in-state tuition at public 4-year collegesfromthefourschoolyearsfollowingtheindividuals’18thbirthdayandisexpressedinconstantdollars of 2014. Sample is all individuals from a nationally-representative cohort of 23-to-31-year-old individuals with credit records in 2004 after applying the filters described in Section 3 who have attended at least a public 4-year college before age 23. 45

Table 11: Probability of a Low Credit Score and Delinquency Below 25th Percentile Below 50th Percentile Ever 30 Days or More Delinquent Variable (OLS) (2SLS) (OLS) (2SLS) (OLS) (2SLS) Student Loans Disbursed 0.000270 0.0175** 0.00264*** 0.0263*** 0.00851*** 0.0149** (0.000380) (0.00829) (0.000467) (0.00906) (0.000421) (0.00752) Tuition -0.00468* -0.00649** -0.00467* -0.00716** 0.00784*** 0.00717*** (0.00253) (0.00295) (0.00257) (0.00300) (0.00219) (0.00235) Ever Public 4-Year -0.116*** -0.178*** -0.126*** -0.211*** 0.0475*** 0.0247 (0.00976) (0.0326) (0.0109) (0.0344) (0.00872) (0.0282) Constant 0.448*** 0.446*** 0.694*** 0.693*** 0.0176 0.0172 (0.0544) (0.0595) (0.0564) (0.0623) (0.0478) (0.0476) Degree/Sector/Pell Grant/College Major YES YES YES YES YES YES Controls Home State/Cohort FEs YES YES YES YES YES YES Observations 33,435 33,435 33,435 33,435 33,435 33,435 Note: This table reports OLS and 2SLS estimates of the effect of student loans on the probability an individual is ever observed with a below 25th percentile credit score between the ages of 22 and 32 in columns (1) and (2). In columns (3) and (4), the estimated effect of student loans on the probability of observing a below median credit score is reported. In columns (5) and (6), the estimated effect of student loans on the probability of being delinquent on student loan debt for at least 30 days is reported. Student loans are instrumented for using the interaction between tuition and an indicator variable for whether the individualeverattendedaPublic4-yearcollegebeforeage23. SeeTables1and2forvariabledefinitionsand sampleselectiondetails. Studentloansdisbursedandtuitionsarerecordedin1000sof2014dollars. Standard errors in parentheses (clustered at the home state by cohort level). ***,**, and * denote significance at 1%, 5%, and 10%. 46

Table 12: Dollar Value of Initial Mortgage Variable (OLS) (2SLS) Student Loans Disbursed 0.333*** -3.230 (0.112) (2.980) Tuition -4.754*** -4.289*** (1.386) (1.506) Ever Public 4-Year 9.355** 24.44* (3.661) (13.11) Degree/Sector/Pell Grant/College Major YES YES Controls Home State/Cohort FEs YES YES Observations 10,475 10,475 Note: ThistablereportsOLSand2SLSestimatesoftheeffectofstudentloansonthefirstobservedmortgage balance for individuals who opened their first mortgage tradeline between the ages of 22 and 32. Student loans are instrumented for using the interaction between tuition and an indicator variable for whether the individual ever attended a Public 4-year college before age 23. See Tables 1 and 2 for variable definitions. Student loans disbursed and tuitions are recorded in 1000s of 2014 dollars. Standard errors in parentheses (clustered at the home state by cohort level). ***,**, and * denote significance at 1%, 5%, and 10%. 47

Figure 1: Homeownership RaPteanbeyl A1ge, Debt Level and Education. Homeownership Rate With Bachelor’s Degree Homeownership Rate Homeownership Rate 0.7 1.0 W W W i i i t t t h h h N C C o o o l l l l C e e o g g l e e le a a g n n e d d F N e o d F S e t d u d S e tu n d t e L n o t a L n o s ans 0.6 With Debt 0.9 With No Debt 0.8 0.5 0.7 0.6 0.4 0.5 0.3 0.4 0.2 0.3 0.2 0.1 0.1 0.0 0.0 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age College Goers With Bachelor’s Degree Homeownership Rate Homeownership Rate 1.0 1.0 0.9 0.9 $0 Debt $0 Debt $0:$15K Debt 0.8 $0:$15K Debt 0.8 > $15K Debt > $15K Debt 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0.0 0.0 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age Note: Collegeattendanceanddegreeattainedaredefinedbasedonwhetherindividualshaveattendedcollege and obtained a degree before age 23, respectively. Student loan debt amounts reflect the amount of federal student loans disbursed before age 23. Homeownership rate at a given age is defined as ever having taken a mortgage by that age. 48

Figure 2: Estimates by Age: OLS vs Probit Marginal Effect of Student Loans on Access to Marginal Effect of Student Loans on Access to Homeownership-OLS Estimates Homeownership-Probit Estimates Percentage Points Percentage Points 0.30 0.30 0.20 0.20 0.10 0.10 0 0 -0.10 -0.10 -0.20 -0.20 -0.30 -0.30 -0.40 -0.40 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age Marginal Effect of Student Loans on Access to Marginal Effect of Student Loans on Access to NHootem:eTowhniserfisghuipr-e2SpLlSot Essetsimtiamteastes of the marginal effect oHfosmtuedoewnntelrosahnip-dIVe bPtroobnit tEhsetimparotebsability of becoming a homeowner against the borrowPeerr’csenataggee Pfooirntsthe OLS (left panel) and probit (right panPeerlc)enmtaoged Peolsin.tsThese 2.0 2.0 estimatesarederivedfromtheregressionsusingthevectorofcontrolsincolumns5ofTables2and3forthe 1.5 1.5 OLS and probit specifications, respectively. Da1s.h0ed lines represent 95 percent confidence intervals. 1.0 0.5 0.5 0 0 -0.5 -0.5 -1.0 -1.0 -1.5 -1.5 -2.0 -2.0 -2.5 -2.5 -3.0 -3.0 -3.5 -3.5 -4.0 -4.0 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age Marginal Effect of Student Loans on Access to Marginal Effect of Student Loans on Access to Homeownership-2SLS Estimates Homeownership-IV Probit Estimates Percentage Points Percentage Points 2.0 2.0 1.5 1.5 1.0 1.0 0.5 0.5 0 0 -0.5 -0.5 49 -1.0 -1.0 -1.5 -1.5 -2.0 -2.0 -2.5 -2.5 -3.0 -3.0 -3.5 -3.5 -4.0 -4.0 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age

Marginal Effect of Student Loans on Access to Marginal Effect of Student Loans on Access to Homeownership-OLS Estimates, No F.E. Homeownership-Probit Estimates Percentage Points Percentage Points 0.30 0.30 0.20 0.20 0.10 0.10 0 0 -0.10 -0.10 -0.20 -0.20 -0.30 -0.30 -0.40 -0.40 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age Figure 3: Estimates by Age: 2SLS vs IV Probit Marginal Effect of Student Loans on Access to Marginal Effect of Student Loans on Access to Homeownership-2SLS Estimates, No F.E. Homeownership-IV Probit Estimates, No F.E. Percentage Points Percentage Points 2.0 2.0 1.5 1.5 1.0 1.0 0.5 0.5 0 0 -0.5 -0.5 -1.0 -1.0 -1.5 -1.5 -2.0 -2.0 -2.5 -2.5 -3.0 -3.0 -3.5 -3.5 -4.0 -4.0 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age Marginal Effect of Student Loans on Access to Marginal Effect of Student Loans on Access to Homeownership-2SLS Estimates, State-by-Cohort F.E. Homeownership-IV Probit Estimates, State-by-Cohort F.E. Percentage Points Percentage Points 2.0 2.0 1.5 1.5 1.0 1.0 0.5 0.5 0 0 -0.5 -0.5 -1.0 -1.0 -1.5 -1.5 -2.0 -2.0 -2.5 -2.5 -3.0 -3.0 -3.5 -3.5 -4.0 -4.0 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age Note: Thisfigureplotsestimatesofthemarginaleffectofstudentloandebtontheprobabilityofbecominga homeowneragainsttheborrower’sageforthe2SLS(leftpanels)andIV-probit(rightpanels)models. These estimates are derived from the instrumental variable regressions using the vector of controls in columns 2 (top left panel) and 5 (bottom left panel) of Table 5 for the 2SLS regressions, and in columns 2 (top right panel)and5(bottomrightpanel)ofTable6fortheIV-probitregressions. Dashedanddottedlinesrepresent 95 and 90 percent confidence intervals, respectively. 50

Figure 4: Effect of Student Loans on Credit Outcomes Marginal Effect of Student Loans on Probability of Marginal Effect of Student Loans on Probability of a Below Median Credit Score Becoming Delinquent on those Loans Percentage Points Percentage Points 5.0 3.5 3.0 4.0 2.5 3.0 2.0 2.0 1.5 1.0 1.0 0.5 0 0 -1.0 -0.5 -2.0 -1.0 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age Note: This figure plots 2SLS estimates of the marginal effect of student loan debt on the probability an individual is observed with a below 50th percentile credit score (left panel) and had become delinquent on those loans (right panel) and by a certain age. The specifications corresponds to columns 2 and 6 of Tables 11, respectively. Dashed lines represent 95 percent confidence intervals. 51

Marginal Effect of Student Loans on Probability of Marginal Effect of Student Loans on Probability of a Below Median Credit Score Becoming Delinquent on those Loans Percentage Points Percentage Points 5.0 3.5 3.0 4.0 2.5 3.0 2.0 2.0 1.5 1.0 1.0 0.5 0 0 -1.0 -0.5 -2.0 -1.0 22 23 24 25 26 27 28 29 30 31 32 22 23 24 25 26 27 28 29 30 31 32 Age Age Figure 5: EffHeoctmoefowan$e1rs,0h0ip0 FIonrc TrereaasetmiennSt tGurdouenpt Loan Debt on Homeownership Percentage Points Observed 70 Simulation with $1000 More Debt 60 50 40 30 20 10 0 22 23 24 25 26 27 28 29 30 31 32 Age Note: This figure plots the average age-profile of homeownership for the treatment group of public 4-year university attendees observed in the data (the black line) and the corresponding homeownership rate the treatment group would face if their debt levels were increased by $1,000 (the red line), according to the specification presented in column5 of Table 5. 52

A Appendix A.1 Variable Definitions In this section we describe the construction and data sources for the variables not covered in Section 3. School Sector: We construct a set of five non-mutually exclusive binary indicators capturing all school sectors with which an individual was before age 23: (1) public 4-year, (2) public 2-year, (3) private 4-year not-for-profit, (4) private 2-year not-for-profit, and (5) private for-profit. To determine the school sectors in our data set, we need unique school level identifiers associated with each enrollment spell observed for a given individual in the sample. In theory, the NSC enrollment records should be sufficient to identify all enrollment spells and, consequently, allow us to observe all sectors attended. In practice, the NSC coverage is not perfect, largely due to school non-participation in the NSC Student Tracker and DegreeVerify programs (for detailed discussion, see Mezza and Sommer (2016) or Dynarski et al. (2013)). Hence, in order to supplement the NSC enrollment data, we use enrollment information from the NSLDS for enrollment spells funded by federal student loans. Highest Degree Attained: We construct a set of six mutually exclusive binary indicators for the highest degree ever attained before age 23. We group degrees into the following categories: (1) no college, (2) dropouts (i.e, those with at least some college but no attained degree), (3) associate’s or certificate degree holders, (4) bachelor’s degree holders, and (5) holders of a master’s degree or more. Moreover, for some individuals, we observe a degree has been attained, but have no information on the type of degree. In such instances, we assign individuals to the category (6) degree of unknown type.40 Major: College majors are available only for those with completed degrees. We aggregate them into 15 different categories, described in detail in Mezza and Sommer (2016). Only majors associated with degrees earned before age 23 are used. Ever Pell Grants: This binary variable indicates whether the individual received Pell grants to finance their post-secondary education before age 23. Credit scores, mortgage balance, and 30+ dpd student loan delinquencies: These variables 40TheNSCcollectsthegraduationdateanddegreeinformationfromschoolsthatreportintotheDegreeVerifyprogram. Unfortunately,somegraduationdatesarereportedwithoutthetypeofdegreeassociatedwith it. When a degree of unknown type is observed in the NSC, but borrowing from the federal government for a subsequent degree is observed in the NSLDS, we use this additional information to infer the degree. 53

are sourced from TransUnion and are defined in Section 4.7.41 Unemployment rate, average weekly wages and house price index at the state level: The unemployment rate is sourced from the yearly Local Area Unemployment Statistics series by the Bureau of Labor Statistics (BLS). The average weekly wages are sourced from the Quarterly Census of Employment and Wages by the BLS. Finally, the house price index is sourced from CoreLogic. All three variables are measured at the individual’s home state in the year when the individual turned 22. A.2 Robustness Checks In this section we check the robustness of our findings to different choices in the timing used to define our variables. We run specifications in which the explanatory variables are observed at ages 23 and 24 (instead of 22, as in our baseline results) and in which the tuition measure is taken from the first six years after the individual left high school, rather than the first four years. Results using the same set of controls as in column 5 of Table 5 are presented in Table 13. 41The credit score used in this analysis is the TU TransRisk AM Score and it ranges from 270 to 900 points 54

Table 13: Effect of Student Loan Debt on Homeownership at Age 26, Robustness Check (2SLS) ExplanatoryVariablesRecordedatAge23 ExplanatoryVariablesRecordedatAge24 First4Years First6Years First4Years First6Years ofTuitionasIV ofTuitionasIV ofTuitionasIV ofTuitionasIV Variable (1) (2) (3) (4) StudentLoansDisbursed -0.0150* -0.0132* -0.0145 -0.0120 (0.00838) (0.00783) (0.00929) (0.00851) EverPublic4-Year 0.0721* 0.0633 0.0749 0.0608 (0.0421) (0.0396) (0.0544) (0.0501) Degree/Sector/Pell Grant/CollegeMajor YES YES YES YES Controls HomeStateEconomic NO NO NO NO Controls HomeState/CohortFEs NO NO NO NO HomeStatebyCohort YES YES YES YES FEs Observations 33,435 33,435 33,435 33,435 Note: Thistablereportsfirstandsecondstage2SLSestimatesoftheeffectofstudentloansontheprobability ofbecomingahomeownerbyage26. Incolumns1and2,studentloansanallotherexplanatoryvariablesare measured when the individual was 23. In columns 3 and 4 these variables are measured when the individual was 24. Student loans are instrumented for using the interaction between tuition and an indicator variable for whether the individual ever attended a Public 4-year college. In columns 1 and 3, tuition is measured as the average in-state tuition charged are public 4-year universities in the individual’s home state, summed over the 4 years after they turned 18. In columns 2 and 4, the tuition measure is summed over the 6 years after turning 18. Marginal probabilities reported. See Tables 1 for variable definitions and 2 for sample selection and specification details. ***,**, and * denote significance at 1%, 5%, and 10%. 55