THE QUARTERLY JOURNAL OF ECONOMICS Vol. CXXVI November2011 Issue 4 D HOWDOESYOURKINDERGARTENCLASSROOM ow n AFFECTYOUREARNINGS?EVIDENCE lo a d FROMPROJECTSTAR∗ ed fro RAJCHETTY m h JOHNN.FRIEDMAN ttp NATHANIELHILGER ://q EMMANUELSAEZ je.o x DIANEWHITMORESCHANZENBACH fo rd DANNYYAGAN jo u rn a In Project STAR, 11,571 students in Tennessee and their teachers were ls.o randomlyassignedtoclassroomswithintheirschoolsfromkindergartentothird rg grade.Thisarticleevaluatesthelong-termimpactsofSTARbylinkingtheexper- at M/ imentaldatatoadministrativerecords.Wefirstdemonstratethatkindergarten IT testscoresarehighlycorrelatedwithoutcomessuchasearningsatage27,college L attendance, homeownership, andretirement savings. Wethendocument four ib sets of experimental impacts. First, students insmall classes aresignificantly rarie morelikelytoattendcollegeandexhibitimprovementsonotheroutcomes.Class s o sizedoes not haveasignificant effect onearnings at age27, but this effect is n J imprecisely estimated. Second, students whohad a more experienced teacher un e in kindergarten have higher earnings. Third, an analysis of variance reveals 2 5 significantclassroomeffectsonearnings.Studentswhowererandomlyassignedto , 2 0 1 3 WethankLisaBarrow,DavidCard,GaryChamberlain,ElizabethCascio, ∗ JanetCurrie, JeremyFinn, EdwardGlaeser, BryanGraham, JamesHeckman, Caroline Hoxby, GuidoImbens, Thomas Kane, Lawrence Katz, Alan Krueger, DerekNeal,JonahRockoff,DouglasStaiger,numerousseminarparticipants,and anonymousrefereesforhelpfuldiscussionsandcomments.WethankHelenBain andJayneZahariasatHEROSforaccesstotheProjectSTARdata.Thetaxdata wereaccessedthroughcontractTIRNO-09-R-00007withtheStatisticsofIncome (SOI)DivisionattheU.S.InternalRevenueService.GregoryBruich,JaneChoi, JessicaLaird,KeliLiu,LaszloSandor,andPatrickTurleyprovidedoutstanding researchassistance. FinancialsupportfromtheLabforEconomicApplications andPolicyatHarvard,theCenterforEquitableGrowthatUCBerkeley,andthe NationalScienceFoundationisgratefullyacknowledged. c TheAuthor(s)2011.PublishedbyOxfordUniversityPress,onbehalfofPresidentand (cid:13) FellowsofHarvardCollege. All rights reserved. ForPermissions, pleaseemail: journals. [email protected]. TheQuarterlyJournalofEconomics(2011)126,1593–1660.doi:10.1093/qje/qjr041. 1593 1594 QUARTERLYJOURNALOFECONOMICS higherqualityclassroomsingradesK–3—asmeasuredbyclassmates’end-of-class testscores—havehigherearnings,collegeattendancerates,andotheroutcomes. Finally,theeffectsofclassqualityfadeoutontestscoresinlatergrades,butgains innoncognitivemeasurespersist. JELCodes:I2,H52. I. INTRODUCTION What are the long-term impacts of early childhood edu- cation? Evidence on this important policy question remains scarce because of a lack of data linking childhood education D andoutcomes in adulthood. This article analyzes the long-term ow n impactsofProjectSTAR,oneofthemostwidelystudiededucation lo a experimentsintheUnitedStates.TheStudent/TeacherAchieve- de d mentRatio(STAR)experimentrandomlyassignedonecohortof fro 11,571studentsandtheirteacherstodifferentclassroomswithin m h their schools in grades K–3. Some students were assigned to ttp smallclasses(15studentsonaverage)ingradesK–3,andothers ://q je were assigned to large classes (22 students on average). The .o x experimentwasimplementedacross79schoolsinTennesseefrom fo rd 1985to1989.NumerousstudieshaveusedtheSTARexperiment jo u toshowthatclasssize,teacherquality,andpeershavesignificant rna ls causal impacts on test scores (see Schanzenbach 2006 for a re- .o rg view).Whetherthesegainsinachievementonstandardizedtests a/ translateintoimprovementsinadultoutcomessuchasearnings t M remainsanopenquestion. IT L WelinktheoriginalSTARdatatoadministrativedatafrom ib ra taxreturns, allowingus tofollow95% oftheSTARparticipants rie into adulthood.1 We use these data to analyze the impacts of s o n STARonoutcomesrangingfromcollegeattendanceandearnings Ju n to retirement savings, home ownership, and marriage. We be- e 2 5 ginbydocumentingthestrongcorrelationbetweenkindergarten , 2 0 test scores and adult outcomes. A 1 percentile increase in end- 1 3 of-kindergarten (KG) test scores is associated with a $132 in- crease in wage earnings at age 27 in the raw data, and a $94 increase after controlling for parental characteristics. Several otheradult outcomes—suchas collegeattendancerates, quality of college attended, home ownership, and 401(k) savings—are also all highly correlated with kindergarten test scores. These 1. Thedataforthisprojectwereanalyzedthroughaprogramdevelopedby theStatisticsofIncome(SOI)DivisionattheU.S. InternalRevenueServiceto supportresearchintotheeffectsoftaxpolicyoneconomicandsocialoutcomesand improvetheadministrationofthetaxsystem. KINDERGARTENCLASSROOMANDEARNINGS 1595 strongcorrelationsmotivatethemainquestionofthearticle: do classroom environments that raise test scores—such as smaller classes and better teachers—cause analogous improvements in adultoutcomes? Our analysis of the experimental impacts combines two empirical strategies. First, we study the impacts of observable classroom characteristics. We analyze the impacts of class size using the same intent-to-treat specifications as Krueger (1999), whoshowedthatstudentsinsmallclassesscoredhigheronstan- dardized tests. We find that students assigned tosmall classes D o are1.8 percentagepointsmorelikelytobeenrolledincollegeat w n age 20, a significant improvement relative to the mean college loa d attendancerateof26.4%atage20inthesample.Wedonotfind ed significantdifferencesinearningsatage27betweenstudentswho fro m wereinsmallandlargeclasses,althoughtheseearningsimpacts h areimpreciselyestimated. Studentsinsmallclassesalsoexhibit ttp://q statisticallysignificantimprovementsonasummaryindexofthe je .o other outcomes we examine (home ownership, 401(k) savings, x fo mobilityrates,percentcollegegraduateinZIPcode,andmarital rd jo status). u rn Westudyvariationacrossclassroomsalongotherobservable als dimensions, such as teacher and peer characteristics, using a .org similarapproach. Priorstudies (e.g., Krueger1999)haveshown a/ t M thatSTARstudentswithmoreexperiencedteachersscorehigher IT ontests.Wefindsimilarimpactsonearnings.Studentsrandomly L ib assignedtoaKGteacherwithmorethan10 yearsofexperience ra earnanextra$1,093(6.9%ofmeanincome)onaverageatage27 ries relativetostudentswithlessexperiencedteachers.2 Wealsotest on J whetherobservablepeercharacteristics havelong-termimpacts un e byregressingearningsonthefractionoflow-income,female,and 25 blackpeersinKG.Thesepeerimpactsarenotsignificant,butare , 20 1 veryimpreciselyestimatedbecauseofthelimitedvariationinpeer 3 characteristicsacrossclassrooms. Because we have few measures of observable classroom characteristics, weturntoasecondempirical strategythat cap- tures both observed and unobserved aspects of classrooms. We use an analysis of variance approach analogous to that in the 2. Because teacher experience is correlated with many other unobserved attributes—suchasattachmenttotheteachingprofession—wecannotconclude thatincreasingteacherexperiencewouldimprovestudentoutcomes. Thisevi- dencesimplyestablishesthatastudent’sKGteacherhaseffectsonhisorher earningsasanadult. 1596 QUARTERLYJOURNALOFECONOMICS teachereffects literaturetotest whetherearnings areclustered by kindergarten classroom. Because we observe each teacher only once in our data, we can only estimate “class effects”—the combinedeffectofteachers, peers, andanyclass-levelshock—by exploitingrandomassignmenttoKGclassroomsofbothstudents andteachers. Intuitively, we test whether earnings vary across KGclassesbymorethanwhatwouldbepredictedbyrandomvari- ationinstudentabilities.AnFtestrejectsthenullhypothesisthat KGclassroomassignmenthasnoeffectonearnings.Thestandard deviationofclasseffectsonannualearningsisapproximately10% D o ofmeanearnings, highlightingthelargestakes at playinearly w n childhoodeducation. loa d Theanalysisofvarianceshowsthatkindergartenclassroom ed assignment has significant impacts onearnings, but it does not fro m tell us whether classrooms that improve scores also generate h earnings gains. That is, are class effects on earnings correlated ttp://q withclasseffectsonscores?Toanalyzethisquestion,weproxyfor je .o eachstudent’sKG“classquality”bytheaveragetestscoresofhis x fo classmatesattheendofkindergarten.Weshowthatend-of-class rd jo peertestscoresareanomnibusmeasureofclassqualitybecause u rn theycapturepeereffects,teachereffects,andallotherclassroom als characteristicsthataffecttestscores. Usingthismeasure,wefind .org that kindergarten class quality has significant impacts on both a/ t M testscoresandearnings. Studentsrandomlyassignedtoaclass- IT roomthatis1standarddeviationhigherinqualityearn3%more L ib at age 27. Students assigned to higher quality classes are also ra significantlymorelikelytoattendcollege,enrollinhigherquality ries o colleges,andexhibitimprovementsinthesummaryindexofother n J outcomes.Theclassqualityimpactsaresimilarforstudentswho un e enteredtheexperimentingrades1–3andwererandomizedinto 25 classesatthatpoint.Hence,thefindingsofthisarticleshouldbe , 20 1 viewedas evidence on the long-term impacts of early childhood 3 educationratherthankindergarteninparticular. Ouranalysis ofclass qualitymust beinterpretedverycare- fully. The purpose of this analysis is to detect clustering in outcomesattheclassroomlevel:areachild’soutcomescorrelated with his peers’ outcomes? Although we test for such clustering by regressing own scores and earnings on peer test scores, we emphasizethat suchregressions arenot intendedtodetect peer effects.Becauseweusepostinterventionpeerscoresastheregres- sor, thesescoresincorporatetheimpactsofpeerquality, teacher quality, and any random class-level shock (such as noise from KINDERGARTENCLASSROOMANDEARNINGS 1597 constructionoutsidetheclassroom).Thecorrelationbetweenown outcomes and peer scores could be due to any of these factors. Ouranalysisshowsthattheclassroomastudentwasassignedto inearlychildhoodmattersforoutcomes20 yearslater, butdoes not shed light on which specific factors should be manipulated to improve adult outcomes. Further research on which factors contribute tohigh class quality wouldbe extremely valuable in lightoftheresultsreportedhere. The impacts of early childhood class assignment on adult outcomes may be particularly surprising because the impacts D o on test scores “fade out” rapidly. The impacts of class size on w n testscoresbecomestatisticallyinsignificantbygrade8(Krueger loa d and Whitmore 2001), as do the impacts of class quality on ed test scores. Why do the impacts of early childhood education fro m fade out on test scores but reemerge in adulthood? We find h some suggestive evidence that part of the explanation may be ttp://q noncognitiveskills.WefindthatKGclassqualityhassignificant je .o impacts on noncognitive measures in fourth and eighth grade x fo suchas effort, initiative, andlackof disruptivebehavior. These rd jo noncognitivemeasuresarehighlycorrelatedwithearningseven u rn conditional on test scores but are not significant predictors of als futurestandardizedtestscores. Theseresultssuggestthathigh- .org quality KG classrooms may build noncognitive skills that have a/ t M returns inthelabormarket but donot improveperformanceon IT standardizedtests. Whilethisevidenceisfarfromconclusive, it L ib highlightsthevalueoffurtherempiricalresearchonnoncognitive ra skills. ries o InadditiontotheextensiveliteratureontheimpactsofSTAR n J on test scores, our study builds on and contributes to a recent un e literature investigating selected long-term impacts of class size 25 intheSTARexperiment.Thesestudieshaveshownthatstudents , 20 1 assignedtosmallclassesaremorelikelytocompletehighschool 3 (Finn,Gerber,andBoyd-Zaharias2005)andtaketheSATorACT college entrance exams (Krueger and Whitmore 2001) and are lesslikelytobearrestedforcrime(KruegerandWhitmore2001). Mostrecently,Muennigetal.(2010)reportthatstudentsinsmall classeshavehighermortalityrates,afindingthatwedonotobtain inourdata as wediscuss later. Wecontributetothis literature byprovidingaunifiedevaluationofseveral outcomes, including the first analysis of earnings, and by examining the impacts of teachers,peers,andotherattributesoftheclassroominaddition toclasssize. 1598 QUARTERLYJOURNALOFECONOMICS Our results alsocomplement the findings of studies on the long-term impacts of other early childhood interventions, such asthePerryandAbecederianpreschooldemonstrationsandthe HeadStartprogram,whichalsofindlastingimpactsonadultout- comesdespitefade-outontestscores(seeAlmondandCurrie2010 forareview).Weshowthatabetterclassroomenvironmentfrom ages 5–8 can have substantial long-term benefits even without interventionatearlierages. The article is organized as follows. In Section II, we re- view the STAR experimental design and address potential D o threats to the validity of the experiment. Section III docu- w n ments the cross-sectional correlation between test scores and loa d adult outcomes. Section IV analyzes the impacts of observable ed characteristics of classrooms—size, teacher characteristics, and fro m peercharacteristics—onadult outcomes. InSectionV, westudy h classeffectsmorebroadly,incorporatingunobservableaspectsof ttp://q classquality.SectionVIdocumentsthefade-outandreemergence je .o effectsandthepotentialroleofnoncognitiveskillsinexplaining x fo thispattern. SectionVIIconcludes. rd jo u rn a II. EXPERIMENTALDESIGNANDDATA ls.o rg II.A. BackgroundonProjectSTAR at M/ Word et al. (1990), Krueger (1999), and Finn et al. (2007) IT L provide a comprehensive summary of Project STAR; here, we ib brieflyreviewthefeaturesoftheSTARexperimentmostrelevant rarie for our analysis. The STAR experiment was conducted at 79 s o n schoolsacrossthestateofTennesseeover4 years. Theprogram J u n oversampled lower-income schools, and thus the STAR sample e 2 exhibits lower socioeconomic characteristics than the state of 5, 2 TennesseeandtheU.S.populationasawhole. 01 3 Inthe1985–86 school year, 6,323 kindergartenstudents in participatingschools wererandomlyassignedtoa small (target size13–17students)orregular-sized(20–25students)classwithin theirschools.3Studentswereintendedtoremaininthesameclass type(smallversuslarge)throughthirdgrade, atwhichpointall 3. Therewasalsoathirdtreatmentgroup:regularsizedclasswithafull-time teacher’saide.Thiswasarelativelyminorintervention,sinceallregularclasses werealreadyassignedaone-third-timeteacher’saide.PriorstudiesofSTARfind noimpactofafull-timeteacher’saideontestscores. Wefollowtheconvention intheliteratureandgrouptheregularandregularplusaideclasstreatments together. KINDERGARTENCLASSROOMANDEARNINGS 1599 studentswouldreturntoregularclassesforfourthgradeandsub- sequentyears. Astheinitialcohortofkindergartenstudentsad- vancedacrossgradelevels,therewassubstantialattritionbecause studentswhomovedawayfromaparticipatingschoolorwerere- tainedingradenolongerreceivedtreatment.Inaddition,because kindergartenwas not mandatoryandduetonormal residential mobility,manychildrenjoinedtheinitialcohortattheparticipat- ingschoolsafterkindergarten. Atotalof5,248 studentsentered theparticipatingschoolsingrades1–3.Thesenewentrantswere randomlyassignedtoclassroomswithinschoolonentry.Thusall D o studentswererandomizedtoclassroomswithinschoolonentry, w n regardlessoftheentrygrade.Asaresult,therandomizationpoolis loa d school-by-entry-grade,andweincludeschool-by-entry-gradefixed ed effectsinallexperimentalanalysesbelow. fro m Uponentryintooneofthe79 schools, thestudydesignran- h domlyassignedstudentsnotonlytoclasstype(smallversuslarge) ttp://q but alsotoaclassroomwithineachtype(ifthereweremultiple je .o classroomspertype,aswasthecasein50ofthe79schools).Teach- x fo ers were also randomly assigned to classrooms. Unfortunately, rd jo theexact protocol ofrandomizationintospecificclassrooms was u rn notclearlydocumentedinanyoftheofficialSTARreports,where als theemphasiswasinsteadtherandomassignmentintoclasstype .org ratherthanclassroom(Wordetal.1990).Wepresentstatisticalev- a/ t M idenceconfirmingthatbothstudentsandteachersindeedappear IT toberandomlyassigneddirectlytoclassroomsonentryintothe L ib STARproject,astheoriginaldesignersattest. ra Asinanyfieldexperiment,thereweresomedeviationsfrom ries o the experimental protocol. In particular, some students moved n J from large tosmall classes and vice versa. Toaccount for such un e potentiallynonrandomsorting, weadoptthestandardapproach 25 taken in the literature and assign treatment status based on , 20 1 initialrandomassignment(intent-to-treat). 3 In each year, students were administered the grade- appropriate Stanford Achievement Test, a multiple choice test that measures performance in math and reading. These tests weregivenonlytostudentsparticipatinginSTAR,astheregular statewide testing program did not extend to the early grades.4 4. TheseK–3testscorescontainconsiderablepredictivecontent.Asreported inKruegerandWhitmore(2001), thecorrelationbetweentestscoresingrades g andg+1 is 0.65 forKG and0.80 foreach grade 1–3. The values forgrades 4–7liebetween0.83and0.88,suggestingthattheK–3testscorescontainsimilar predictivecontent. 1600 QUARTERLYJOURNALOFECONOMICS Following Krueger (1999), we standardize the math and read- ing scale scores in each grade by computing the scale score’s correspondingpercentilerankinthedistributionforstudentsin largeclasses. Wethenassigntheappropriatepercentilerankto studentsinsmallclassesandtaketheaverageacrossmathand readingpercentileranks. Notethat this percentilemeasureis a rankingofstudentswithintheSTARsample. II.B. VariableDefinitionsandSummaryStatistics We measure adult outcomes of Project STAR participants D o using administrative data from U.S. tax records. Most (95.0%) wn ofSTARrecordswerelinkedtothetaxdatausinganalgorithm loa d e based on standard identifiers (SSN, date of birth, gender, and d names)describedinOnlineAppendixA.5 fro m Weobtaindataonstudentsandtheirparentsfromfederaltax h ttp forms such as 1040 individual income tax returns. Information ://q from 1040s is available from 1996 to2008. Approximately 10% je .o of adults do not file individual income tax returns in a given x fo year. We use third-party reports to obtain information such as rd jo wageearnings (formW-2)andcollegeattendance(form1098-T) u rn for all individuals, including those who do not file 1040s. Data als fromthesethird-partyreportsareavailablesince1999.Theyear .org always refers to the tax year (i.e., the calendar year in which a/ t M the income is earned or the college expense incurred). In most IT cases, tax returns for tax year t are filed during the calendar L ib yeart+1.Theanalysisdatasetcombinesselectedvariablesfrom ra individualtaxreturns,third-partyreports,andinformationfrom rie s theSTARdatabase,withindividualidentifiersremovedtoprotect on J confidentiality. un e Wenowdescribehoweachoftheadultoutcomemeasuresand 2 5 control variables used in the empirical analysis is constructed. , 2 0 1 Table I reports summary statistics for these variables for the 3 STAR sample as well as a random 0.25% sample of the U.S. populationborninthesameyears(1979–1980). Earnings. TheindividualearningsdatacomefromW-2forms, yielding information on earnings for both filers and nonfilers.6 5. AllappendixmaterialisavailableintheOnlineAppendix.Notethatthe matching algorithm was sufficiently precise that it uncovered28 cases in the originalSTARdatasetthatwereasinglesplitobservationorduplicaterecords. Afterconsolidatingtheserecords,weareleftwith11,571students. 6. Weobtainsimilarresultsusinghouseholdadjustedgrossincomereported onindividualtaxreturns. WefocusontheW-2 measurebecauseitprovidesa KINDERGARTENCLASSROOMANDEARNINGS 1601 We define earnings in each year as the sum of earnings on all W-2 forms filed on an individual’s behalf. We express all monetaryvariablesin2009 dollars, adjustingforinflationusing the Consumer Price Index. We cap earnings in each year at $100,000 to reduce the influence of outliers; fewer than 1% of individualsintheSTARsamplereportearningsabove$100,000 in a given year. Toincrease precision, we typically use average (inflation indexed) earnings from 2005 to 2007 as an outcome measure. The mean individual earnings for the STAR sample in2005–2007 (whentheSTARstudents are25–27 years old) is D o $15,912. This earnings measure includes zeros for the 13.9% of w n STARstudentswhoreportnoincome2005–2007.Themeanlevel loa d ofearningsintheSTARsampleislowerthaninthesamecohort ed in the U.S. population, as expected given that Project STAR fro m targetedmoredisadvantagedschools. h ttp College Attendance. Higher education institutions eligible ://qje for federal financial aid—Title IV institutions—are required to .ox file 1098-T forms that report tuition payments or scholarships ford jo u rn a ls TABLEI .o rg SUMMARYSTATISTICS a/ t M (1) (2) (3) (4) IT STARsample U.S.1979–80cohort L ib Variable Mean Std.Dev. Mean Std.Dev. ra Adultoutcomes rie s Averagewageearnings(2005–2007) $15,912 $15,558 $20,500 $19,541 o n Zerowageearnings(2005–2007)(%) 13.9 34.5 15.6 36.3 Ju n Attendedcollegein2000 26.4 44.1 34.7 47.6 e 2 (age20)(%) 5 Collegequalityin2000 $27,115 $4,337 $29,070 $7,252 , 20 1 Attendedcollegebyage27(%) 45.5 49.8 57.1 49.5 3 Ownedahousebyage27(%) 30.8 46.2 28.4 45.1 Made401(k)contributionby 28.2 45.0 31.0 46.2 age27(%) Marriedbyage27(%) 43.2 49.5 39.8 48.9 MovedoutofTNbyage27(%) 27.5 44.7 Percentcollegegraduatesin2007 17.6 11.7 24.2 15.1 ZIPcode(%) Deceasedbefore2010(%) 1.70 12.9 1.02 10.1 consistentdefinitionofindividualwageearningsforbothfilersandnonfilers.One limitationoftheW-2easureisthatitdoesnotincludeself-employmentincome. 1602 QUARTERLYJOURNALOFECONOMICS TABLEI (CONTINUED) (1) (2) (3) (4) STARsample U.S.1979–80cohort Variable Mean Std.Dev. Mean Std.Dev. Parentcharacteristics Averagehouseholdincome(1996–98) $48,014 $41,622 $65,661 $53,844 Mother’sageatchild’sbirth(years) 25.0 6.53 26.3 6.17 Marriedbetween1996and2008(%) 64.8 47.8 75.7 42.9 D o Ownedahousebetween 64.5 47.8 53.7 49.9 w n 1996and2008(%) lo a Madea401(k)contribution 45.9 49.8 50.5 50.0 d e d Mibsestiwngee(nno1p9a9r6eanntdfo2u0n0d8)((%%)) 13.9 34.6 23.9 42.6 from Studentbackgroundvariables http Female(%) 47.2 49.9 48.7 50.0 ://q Black(%) 35.9 48.0 je .o Eligibleforfreeorreduced-price 60.3 48.9 x fo lunch(%) rd Ageatkindergartenentry(years) 5.65 0.56 jo u rn Teachercharacteristics(entry-grade) a ls Experience(years) 10.8 7.7 .org Post-BAdegree(%) 36.1 48.0 a/ Black(%) 19.5 39.6 t M Numberofobservations 10,992 22,568 IT L ib Notes.Adultoutcomes,parentcharacteristics,andstudentageatKGentryarefrom1996–2008tax ra data;otherstudentbackgroundvariablesandteachercharacteristicsarefromSTARdatabase.Columns(1) rie a(3n)da(n2d)a(r4e)baaresebdaosnedthoensaam0p.2le5%ofrSaTnAdRomstsuadmenptlsewofhothweeUre.Ss.ucpcoepsusfluatlliyonlinbkorendtinoUth.Se.staamxedayteaa.rCsoalusmthnes s on STARcohort(1979–80). AllavailablevariablesaredefinedidenticallyintheSTARandU.S. samples. J Earningsareaverageindividualearningsinyears2005–2007,measuredbywageearningsonW-2forms; un thosewithnoW-2earningsarecodedas0s.Collegeattendanceismeasuredbyreceiptofa1098-Tform, e 2 issuedbyhighereducationinstitutions toreport tuitionpayments orscholarships. Collegequalityis 5 definedasthemeanearningsofallformerattendeesofeachcollegeintheU.S.populationatage28.For , 2 individualswhodidnotattendcollege,collegequalityisdefinedbymeanearningsatage28ofthosewho 01 didnotattendcollegeintheU.S.population.Homeownershipismeasuredasthosewhoreportmortgage 3 interestpaymentsona1040or1098taxform.401(k)contributionsarereportedonW-2forms.Marital statusismeasuredbywhetheranindividualfilesajointtaxreturn.StateandZIPcodeofresidenceare takenfromthemostrecent1040formorW-2form.Percentcollegegraduatesinthestudent’s2007ZIP codeisbasedondataonpercentcollegegraduatesbyZIPcodefromthe2000Census.Birthanddeath informationareasrecordedbytheSocialSecurityAdministration.WelinkSTARparticipantstotheir parentsbyfindingtheearliest1040forminyears1996–2008onwhichtheSTARstudentisclaimedas adependent.Weareunabletolink13.9%oftheSTARchildren(and23.9%oftheU.S.cohort)totheir parents;thesummarystatisticsreportedforparentsexcludetheseobservations.Parenthouseholdincome isaverageadjustedgrossincome(AGI)inyears1996–1998,whenSTARparticipantsareaged16–18.For yearsinwhichparentsdidnotfile,householdincomeisdefinedas0.Forjoint-filingparents,mother’s ageatchild’sbirthusesthebirthdateofthefemaleparent;forsingle-filingparents,thevariableuses thebirthdateofthesingleparent, whoisusuallyfemale. Otherparentvariablesaredefinedinthe samemannerasstudentvariables. Freeorreduced-priceluncheligibilityisanindicatorforwhether thestudentwasevereligibleduringtheexperiment.Student’sageatkindergartenentryisdefinedasage (indays,dividedby365.25)onSept.1,1985.Teacherexperienceisthenumberofyearstaughtatany schoolbeforethestudent’syearofentryintoaSTARschool.Allmonetaryvaluesareexpressedinreal 2009dollars.