Variation in Center Quality in a Universal Publicly Subsidized and Regulated Childcare System

ContentslistsavailableatScienceDirect

Labour Economics

journalhomepage:www.elsevier.com/locate/labeco

Variation in center quality in a universal publicly subsidized and regulated childcare system

Mari Rege

, Ingeborg Foldøy Solli

, Ingunn Størksen

, Mark Votruba

aUniversity of Stavanger, Norway

bUniversity of Stavanger, Norway

cUniversity of Stavanger, Norway

dCase Western Reserve University, United States

a r t i c le i n f o

Keywords:

Child development Childcare quality

Early childhood education and care JEL codes : I20

I24 J13

a b s t r a ct

Alargeliteraturesuggeststhathighqualitychildcareprogramscanproducepositiveandlastingeffectsbypro- motingmath,languageandsocial-emotionalskills,referredtoasschoolreadinessskills,especiallyforchildren ofparentswithloweducation.Hence,auniversalchildcaresystemwitheasyaccesshasthepotentialtomake asubstantialdifferenceinchildren’slivesandreducesocio-economicdisparitiesineducationaloutcomes.How- ever,ifchildcarequalityvariesacrosscenters,universalchildcaresystemscanalsopotentiallyincreasedisparities inschoolreadinessifthechildrenofmorehighly-educatedparentsselectintocentersofhigherquality.Using auniquedatasetwithone-to-oneassessmentsofschoolreadinessskillsamong627five-year-oldsattending67 differentchildcarecenters,weinvestigatedifferencesinchildcarequalitybytestingwhethercovariateadjusted assessmentsscoresareclusteredbycenter.Throughfixedeffectandrandomeffectanalyses,wedemonstratesig- nificantvariationinschoolreadinessacrosscenters.However,selectionintocentersofdifferentqualityappears tobelimited.

1. Introduction

Thereisalarge literaturelinkingearlychildhoodlanguage,math andsocio-emotionalskillsasfoundationalforfuturelearningandde- velopment(Duncanetal.,2007;Halletal.,2016;Rabineretal.,2016).

Moreover,studiesshowthathighqualitychildcareprogramsstimulat- ingtheseschoolreadiness skillscanproducepositive andlastingef- fectsbypromotingschoolsuccessandfosteringworkforceproductivity (GuptaandSimonsen,2016;HavnesandMogstad,2009;Heckmanand Kautz, 2013; Melhuish, 2011; Reynolds et al., 2011; Weiland and Yoshikawa,2013).Thisisespeciallytrueforchildrenofparentswith loweducation.

Basedontheevidencefromtheearly-childhood-education-and-care (ECEC)literature,publiclysubsidizedchildcarewitheasyaccessforlow- incomechildrenisoftenaccentuatedasakeypolicytoprovidechildren moreequalopportunities.Norwayisfrequentlyconsideredafrontrun- nerinthisrespect,withpubliclysubsidizeduniversalchildcareforall childrenagesone-to-fiveandfreeaccessforchildrenoffamilieswith lowincome(e.g.BennettandTayler,2006).Indeed,Norwayisamong theOECDcountrieswiththehighestpublicspendingonearlychildhood educationandcare(Engeletal.,2015).However,theNorwegianchild- caresystemhasalsobeencriticizedbecausetheregulatorystandards

∗Correspondingauthor.

E-mailaddresses:[email protected](M.Rege),[email protected](I.F.Solli),[email protected](I.Størksen),[email protected](M.Votruba).

forstructuralandprocessqualityarelenientandimprecise(Bennettand Tayler,2006;Engeletal.,2015),whichcouldcontributetolargequality differencesacrosschildcarecenters.Suchvariationisconcerningsince it suggestsmissed opportunitiesto improveschool readiness among childrenattendinglowerqualitycenters.Moreover,ifchildcarequality variesacrosscenters,universalchildcaresystemscouldincreasedispar- itiesinschoolreadinessifthechildrenofparentswithhigheducation selectintocentersofhigherquality.

Thispaperinvestigatesdifferencesinchildcarequalityacrosscen- tersintheuniversalchildcaresysteminNorway.Weutilizeaunique datasetcollectedintheAgder-project¹withone-to-oneassessmentsof literacy,mathandself-regulationof627five-year-oldsin67different childcarecentersinNorway.Theassessmentdataismatchedwithin- dicatorsforchildcarecenterandregistry dataonparentaleducation,

1TheAgder-projectisarandomizedfieldexperimentinvestigatingeffectsofa preschoolinterventionforfive-year-oldsatNorwegiandaycarecenters.Treated childcarecentersreceiveteachereducation,afocusedcurriculumwithconcrete examplesofplayfullearningactivitiescultivatingschoolreadinessskills,and resourcesprovidingteacherswithtimeforengagingthechildrenintheplayful learningactivities.Weusethepre-interventiondatawaveoftheAgder-project inthispaper.

https://doi.org/10.1016/j.labeco.2018.10.003

Received1November2017;Receivedinrevisedform28September2018;Accepted12October2018 Availableonline12October2018

0927-5371/© 2018TheAuthors.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense.

(http://creativecommons.org/licenses/by-nc-nd/4.0/)

earningsandimmigrant statusfromStatisticsNorway.Ourempirical analysestestwhethercovariate-adjustedassessmentscoresvaryacross childcarecentersbymorethan wouldbe predictedbyrandomvari- ationinchildren’sskills.Ourfixedeffectandrandomeffectanalyses demonstratesignificantvariationinschoolreadinessskillsatagefive.

Thedifferenceinschoolreadinessskillsincentersatthe90thand10th percentileinthecentereffectdistributionisestimatedtobeoverone half(0.55)ofastandarddeviation.Thus,thedifferencesinschoolreadi- nessskills(notaccountedforbycovariates)acrosschildcarecentersin Norwayappearstobe quitesubstantial.Differentialselectionofchil- drenintohigher/lowerqualitycentersonthebasisofsocio-economic backgroundappearsquitelimited.However,wecannotruleoutthatthe differencesincentereffectscaptureunobservedsimilaritiesinchildren’s priorskillsandparentalbackground.

Notably,thevarianceinschoolreadinessscoresattributabletocen- terscaptures thecombinedeffectofteacher quality,peers,childcare pedagogy, in addition tostructural qualitysuch aschild-teacher ra- tios. Moreover, children in the same center mayalso be co-located residentially,andmayexperiencecommonshocksaffectingchildde- velopment.Alimitationofourstudyisthatthemechanism(s)behind thevariationacrosscenterscanbeexploredonlypartially.Weextend ouranalysesbyinvestigatinghowmeasuresofstructuralquality– the child-teacherratio, center sizeandthetenure of thedirector– pre- dictthedevelopment ofschoolreadinessskills.Thisevidenceshould beinterpretedwithcautionasthestructuralindicatorsarenotrandom andmaybeassociated withotherunobservedindicators forquality.

Nevertheless,wefindthattheteacher-childratioisassociatedwitha largeandsignificantincrease inschoolreadiness skills,andthis sin- glecharacteristicexplainsameaningfulportion(about30percent)of thevariance in schoolreadiness acrosscenters. Theother structural characteristicsofcentersfailtosignificantlypredictschoolreadiness scores.

Our studyrelates tothelarge literature in economics investigat- ingqualitydifferencesacrossclassroomsandteachersinschools(e.g.

Chettyet al., 2011; Rivkinetal., 2005; Rockoff, 2004).This litera- turehasdemonstratedlargevariationacross classroomsandteachers inchildren’slearning,whichcannotbeexplainedbyobservablessuch asclasssize,child-teacherratios,andteacherexperienceandeducation.

Aswedonotobservepre-measuresofchildren’sskills,ourempiricalap- proachislessrigorousthanthevalue-addedapproachesintheabove- citedstudies.Nevertheless,wearenotfamiliarwithanyotherpaperin- vestigatingqualitydifferencesinlearningacrosscentersinauniversal childcaresystem.Moreover,ourrichdatasetallowsustodoarigorous selectionanalysis,whichsuggeststhecentereffectsarenotdrivenby selection.

Additionally,ouranalysescontributetotheemergingeconomiclit- eratureinvestigatinghowchildcarestructuralqualityindicatorsaffect childdevelopment (e.g.Bauchmülleretal.,2014;Blau,1999; Currie andNeidell,2007;DrangeandRønning,2017).Theliteratureprovides mixedevidenceoftheeffectofstructuralparameters,andmoreresearch isneeded.Ingeneral,theevidencefromthesestudiesseemstomimic evidencefromsimilarstudiesinschools(seereviewinJacksonetal., 2014),whichsuggestlimitedpotentialforimprovingchilddevelopment bymerely investinginstructuralcharacteristics.Nevertheless,consis- tentwithourfindings,Bauchmülleretal.(2014)andCurrieandNei- dell(2007)demonstratepositiveassociationsbetweenteacher-childra- tiosandchildoutcomes.Incontrast,Blau(1999)findnosignificantas- sociationsbetweenthesemeasures.

Ourpaperproceedsasfollows.InSection2,weprovidesomeback- groundinformation on keyschool readinessskills, andexisting evi- denceonhowchildcarecenterscanpromotetheseskills.InSection3, wedescribetheNorwegianchildcaresystemandcontext.InSection4, we present our measures of key school readiness skills, procedures fordata collectionandsample. In Section5,we present ourempir- ical analyses. Finally, in Section 6 we conclude by discussing the results.

2. Childcarequalityandkeyschoolreadinessskills

Childcarequalityisoftendefinedandmeasuredaccordingtotwoba- sicaspects—structuralqualityandprocessquality(OECD,2015).Struc- turalqualityincludeseasilyobservableaspects,suchasteachereduca- tion,child-staff ratios,teacherandmanagementexperience,andclass size.Processqualityrepresentsthedirectexperiencesofchildren,and includefactorssuchasthesensitivityandresponsivenessofcaregivers toindividualneeds,thepedagogicalapproachesandmaterialsavailable forlearning,andinteractionswithteachersandpeers.Evidencesuggest that increasingstructuralquality inisolationhasamodest orno ef- fectonchildren’sdevelopment(Bauchmülleretal.,2014;Blau,1999).

However,increasedstructuralqualityisoftenneededtoimprovepro- cessquality,whichseveralstudiesdocumentascrucialforearlychild- hoodlearning,development,andwell-being(Hamre2014;LeeandBier- man,2015;Saboletal.,2013).

Animportantaspectofprocessqualityiscompassionateandsystem- aticcultivationoflanguage,mathandsocio-emotionalskills.TheECEC literaturedemonstratesthattheseskillsarecriticalforschoolsuccess andmalleableinearlychildhood(Jensenetal.,2017;Melhuish,2016; TomineyandMcClelland,2011;Yoshikawaetal.,2013).Belowwesum- marizetheevidenceonthesekeyschoolreadinessskills:

2.1. Language

Youngchildrenwithadvancedlanguageskillsaremorelikelytobe preparedforformalschoolingcomparedtochildrenwithweaklanguage skills(Duncanetal.,2007;McGintyandJustice,2010).Inparticular, vocabularyatschoolentryexplains asubstantialportionofthevari- anceinchildren’searlyliteracy,includinggrowthinliteracy(Lervågand Aukrust,2010).Earlyvocabularyalsosignificantlypredictsfutureread- ingcomprehension.Researchindicatesthatinteractivebookreading, whencombinedwithprofessionaldevelopmentactivitiesdesignedto improvethequalityofteacher’slanguage,canbeeffectiveinpromoting richerconversationalexchangesintheclassroomandgainsinchildvo- cabularyandoralcomprehensionskills(Biermanetal.,2008).Further- more,exploringrhymes,lettersandsoundsatanearlyageisgenerally seenasanimportantpre-readingactivity(Muteretal.,2004).

2.2. Math

Theimportanceofearlymathskillsisincreasinglyevident.Awidely citedU.S.studyofchildren’sgrade-schoolachievementuseddatafrom sixdifferentlarge-scalesurveysandshowedthatthestrongestpredic- tors of later achievement were school-entry mathematical,attention (anaspectofself-regulation),andreadingskills(Duncanetal.,2007).

Amongthepossiblepredictors, mathematicalskillsseemedtobethe mostprominent.ThesefindingshavesincebeenreplicatedusingCana- dian data(Romanoetal.,2010). Thequantitative,spatial, andlogi- cal reasoning competenciesof earlymathematics mayforma cogni- tivefoundationforthinkingandlearningacrosssubjects(Clementsand Sarama,2011).Randomizedcontrolledtrialshavedemonstratedthatin- terventionsdesignedtofacilitatechildren’smathematicallearningdur- ingages3to5yearscanhaveastrongeffectonchildren’smathematical achievement(ClementsandSarama,2011).

2.3. Socio-emotionalskills

Research consistentlyconfirmsthe importanceof children’s early abilitiestocontroltheirbehavioranduse theircognitiveabilitiesto successfullynavigateschoolsettings(HeckmanandKautz,2013).Self- regulation canbe defined astheabilitytocontrolthoughts,feelings andbehaviors inorder toadapteffectivelywithdemandsandsocial standardsandexpectationsintheenvironment,andtobeabletoreach futuregoals(Berger,2011).Preschoolself-regulationhasbeenlinked toemergentliteracy, vocabulary,andmathskills(McClellandetal.,

2007) andits importancehasbeen demonstratedin variouscultures (von Suchodoletz etal., 2013; Wanless et al., 2011a).Several early childhoodinterventionshavedemonstratedsignificantimpactsonself- regulation(DiamondandLee,2011;Merrittetal.,2012;Raveretal., 2011;Schmittetal.,2018;TomineyandMcClelland,2011,2013),andit seemsliketeachersensitivity(highprocessquality)isanessentialaspect forchildren’sdevelopmentofself-regulation(BlairandRaver,2015).

Inadditiontoself-regulation,children’semotioncomprehensionor understandingabouttheir ownemotionsandtheemotions ofothers helpslaythefoundationforstrongsocio-emotionalskills(Denham&

Brown, 2010).Children’s earlyrelationships withothers aresignifi- cantlyrelatedtofutureschoolsuccess.Forexample,childrenwhoare rejectedandisolatedareatriskoffutureschoolavoidance(Buhsetal., 2006).Importantly,emotioncomprehensionandrelationshipskillscan beenhancedinearlychildhoodandpromotelastingeffects(Nixetal., 2016).Unfortunately,wedonotmeasureemotioncomprehensionand relationshipskillsinthisstudyduetolackoftestsvalidatedinaNorwe- giancontext.Still,eveniftheseskillsarehardtomeasure,theybelong inalistofkeyschoolreadinessskills.

3. TheNorwegiancontext,childcaresystemandhypotheses

Norwayhasastrongwelfarestatewithmanyfamilypoliciesfacili- tatingbothchildwell-beingandastronglabormarketattachmentfor parentsofyoungchildren.Inassociationwithchildbirthoradoption, parentshavetherightto11monthsparentalleavewithfullwagecom- pensation²andjobsecurity.Allchildrenagesonetofiveyearsoldhave therighttopubliclyregulatedandsubsidizedchildcare.Theutilization ofthechildcaresystemisveryhigh,withanuptakerateof97percent amongfive-year-olds.ChildreninNorwaystartprimaryschoolinAugust theyeartheyturnsix.Allchildrenareobligedtoattendprimaryschool inNorway,andmostchildrengotopublicschool;only3.6percentgo toprivateschools.

TheNorwegianchildcaresystemhasbeencriticizedbecauseitgives childcarecentersalargedegreeoffreedomwithrespecttopedagog- icalcontent,whichcan giverisetolargedifferencesin processqual- ityacross centers(BennettandTayler,2006).Thesystemwasorigi- nallyestablishedasaresponsetoaneedforhighqualitychildcareas mothersenteredthelabormarket.Eveniftheeducationalanddevel- opmentalpurposeisnowprevailing,theprogramisstilldominatedby thesocialpedagogicaltradition,whichhasalimitedcurricularfocus.

Theprogram’spedagogybuildsonabeliefthatpreschoolagechildren havetheirbestlearningexperiencesthroughfreeplayandactivitiesthat buildonthepreschoolersowninitiatives.Thecenters’pedagogicalcon- tentisregulatedbytheNationalFrameworkPlanforContentandTasks ofKindergartens(Ministryof EducationandResearch,2011)³ which definessevenlearningareas:1.Communication,languageandtext;2.

Body,movementandhealth;3.Art,cultureandcreativity;4.Nature, environmentandtechnology;5.Ethics,religionandphilosophy,6.Lo- calcommunityandsociety;7.Numbers,spacesandshapes.However, theselearningareasareonlylooselydescribedandarethesameforall childrenages1–5.Moreover,therearenospecificguidelinesforhow thechildcarecentersshouldimplementthelearningareasandteachers havenobenchmarksforchildren.

Additionally,theNorwegianchildcaresystemhasbeencriticizedbe- causethestandardsforstructuralqualityarelenient(Engeletal.,2015), allowingvariationwhichmightalsofacilitatequalitydifferencesacross

2 Thecompensationfromthegovernmenthasaceiling.However,thisceiling issetaboveaverageearningsandformostemployeeswithearningsabovethe ceiling,theemployerwillcompensateforthedifferential.

3 TheFrameworkPlanwasrevisedin2017.However,wedescribetheFrame- workPlanof2011,asthisistherelevantplanforthechildreninoursample.It isworthnotingthatthekeyconcernswiththePlanof2011remainsintheplan of2017:Thelearningareasareonlyvaguelydescribedandarethesameforall childrenagesone-to-five.Thereisnodetailedageappropriatecurricularfocus.

childcarecenters.InNorwayanECECteacherhasabachelordegree inearlychildhoodeducation.Theadult-childratioisregulatedsothat theyoungestchildrenhaveatleastoneECECteacherper7– 9chil- dren,whilsttheolderchildrenhaveatleastoneECECteacherper14– 18children.However,centersoftenapplyforexemptionsbecauseofa shortageofqualifiedpersonnel.Moreover,therearenomandatorycon- tinuingeducationprogramsforchildcareteachers.Inadditiontothe childcareteacher,eachchildgrouphastwoassistants.However,there arenoformalqualificationrequirementsforthesesassistants– itisnot evenrequiredthattheyhavecompletedhighschool.

Theimpreciseandlenientstandardsforprocessandstructuralqual- ityintheNorwegianchildcaresystem,discussedabove,leadustohy- pothesize:

Hypothesis1. Thereissignificantvariationincenterqualitymeasured bychildren’sschoolreadinessskills.

Iftherearesignificantdifferencesinlearningopportunitiesacross centers, we might expectchildren of parentswith low education to be especiallyvulnerable tothesedifferences. Severalstudiesdemon- stratethat parents’ socio-economic background is astrong predictor ofearlychildhooddevelopment.ThisisalsotrueinNorwaydespitea generouswelfaresystemandlimitedchildpoverty (Bøeetal.,2016; Schjølberg etal., 2008; Størksen etal., 2013; Størksen etal., 2015; StørksenandMosvold,2013).Severalstudiessuggestthatthegapsin skilldevelopmentacrossfamilybackgroundcanpartiallybeexplained byparents’abilitytocreateahomeenvironmentthatstimulateslearn- inganddevelopment(Guryanetal.,2008;Harrisetal.,1999;Kaliletal., 2012).Moreover,researchsuggeststhatthepositiveimpactofquality childcareismostpronouncedforlow-incomechildren,whogenerally havefewalternativelearningopportunities(DearingandMcCartney).

Assuch,wehypothesize:

Hypothesis2. Therelativeadvantageinschoolreadinessskillsforchil- drenfromhigherSESfamilieswilldecreaseforchildreninhigherquality centers.

4. Assessments,datacollectionandsample

InAugust2016,theAgder-projectassessedschoolreadinessskillsby invitingallthefive-year-oldsattending71childcarecentersinsouthern Norwaytolocalsciencemuseums.Thechildrenhadaccesstoallthe activitiesatthemuseum.Ata scheduledtime, eachchildcarecenter broughttheirchildrentoanassessmentstation,whereeachchildcom- pletedplayfultasksoncomputertabletsinaroomwithatrainedand certifiedtester.Assessmentstookapproximately 40 minutesforeach child.

The testers used computer tablet instruments developed for the Skoleklar-project(Størksenetal.,2013;StørksenandMosvold,2013) andfurtherrefinedfortheAgder-project.Thetabletswereloadedwitha specially-designedapplicationcontainingabatteryofsixtestsdesigned toassessmath,literacyandself-regulationskills,whichareconsidered criticalforsuccessfulschooladjustment(seeSection2).

Math skills wereassessedvia the Ani BananiMath Test(ABMT) (StørksenandMosvold, 2013).TheABMTisan18-itemdigitalmath assessmentonatabletapplication,whichincludesitemscoveringthree areasofmathematics– numeracy,geometryandproblemsolving.Chil- drenhelpthemonkeyonthescreenwithdifferenttasks,suchascount- ing bananasandsetting the tablewith enoughplates for theguests in abirthdayparty.Themeasurecorrelatesstronglywithanexisting schoolbasedmathassessmentr=0.69(Utdanningsdirektoratet,2017) andwithanothervalidatedearlynumeracytaskr=0.74(NumberSense Task)(VanLuitandVandeRijt,2009).Cronbach’salphainthisstudy is𝛼=0.72.

Twoassessmentswereconductedtomeasureliteracy,onepertain- ingtovocabularyandtheothertophonologicalawareness.Vocabulary wastestedwiththeNorwegianVocabularyTest(NVT)(Størksenetal.,

2013).TheNVTisa“namingtest” (total20words)whereanillustra- tionappearsonthecomputertabletscreenandthechildissubsequently askedtonameit.Cronbach’salphainthisstudyis𝛼=0.81.Assessments ofchildren’sphonologicalawarenesswereconstructedfromtheofficial literacyscreeningbatteryfromTheNorwegianDirectorateforEduca- tionandTraining.Themeasureconsistsofa12-itemblendingtask.For eachtask,atargetwordispresentedinitsindividualphonemesbythe experimenterandchildrenhadtoindicatethecorrespondingalterna- tivefromfourpresentedimagesonatabletscreen.Allcorrectanswers weregivenonepointandsummedup,resultinginscoresrangingfrom 0–12.Notably,thetestisdifficultforthechildrenatthisage,andwe seeconsiderableflooreffects.This wasexpectedbecausemost child- carecentersdonotworktostimulatethisskill,sinceitisnotempha- sizedintheFrameworkplan.Nevertheless,itwasincludedamongthe pre-interventionmeasuresofschoolreadinesscapturedbytheAgder- project.

Three assessments were conducted to measure children’s self- regulationskills.TheHead-Toes-Knees-Shoulderstask(HTKS)integrates attention,inhibitorycontrol,bodycontrolandworkingmemorydemands intoashorttaskofbehavioralself-regulationappropriateforchildren aged4–8years(McClellandetal.,2014).Ithasstrongreliabilityandva- lidity,issignificantlyrelatedtoothermeasuresofself-regulation,and tochildren’s academicoutcomesin diversesamples(CameronPonitz etal.,2009;Fuhsetal.,2014;McClellandetal.,2014;Wanlessetal., 2011b) including Norwegianchildren (Storksen et al., 2015). Cron- bach’salphainthisstudy𝛼=0.76.Second,intheHeartsandFlowers task(Davidsonetal.,2006),childrenhavetorespondbypressingakey onthesamesideofthestimuluswhentheyseeaheartandbypressing akeyontheoppositesidewhenthestimulusisaflower.Themeasure hasstrongreliabilityandvalidity(Davidsonetal.,2006).Thistestpro- videsamorenarrowmeasureofcognitivecontrol,asopposedtothebe- havioralself-regulationandworkingmemoryskillsalsoatworkinthe HTKS.Thethirdself-regulationassessmentusestheForward/Backward DigitSpansubtestsfromtheWechslerIntelligenceScalesforchildren- III(Wechsler,1991).Digitswerereadaloud,onedigitpersecond,and thechildrenwereaskedtorepeatthesequenceofdigits.Firsttheyhad torepeatsequencesinthesameorderastheywerereadaloud,thenin reversedorder.Thetestwasautomaticallydiscontinuedaftertwosub- sequenterrors.Itmeasuresworkingmemoryandtheabilitytofocus, consideredaspectsofself-regulation.Themeasurehasstrongreliability andvalidity(Davidsonetal.,2006).

71childcentersfrom17municipalitiesparticipatedintheproject.

There werearound855five-year olds(children in theirlast yearof ECEC)inthesecenters,rangingfrom4to29five-yearoldsineachcen- ter.Amongthese,701childrenhadparentalconsent(82percent),of which669showedupfortesting.Throughthecollectionofpersonal identifiers,we wereabletomatchchildren’s assessmentscores with registrydatafromStatisticsNorway,usedtoconstructmeasuresofthe families’socio-economicstatus.

Duetoourinterestincenter-levelvariation,weexcludedfromour sample14childrenfromfourcenterswhichhadfewerthan5children represented.Weadditionallyexcludedthreechildrenwhocouldnotbe matchedtoNorwegianregistry filesin 2015.⁴ These criteriayielded asample with648 childrenfrom67 childcarecenters,with registry dataonchildandfamilybackgroundvariables(birthmonth,gender, parentaleducationandearnings,parent’scountryofbirth).Allchildren completedthemathtest,butonly601childrencompletedallsixassess- mentsbecauseweendedtheassessmentearlyforchildrenwhobecame uncomfortablewiththetestsituation.Twenty-sixchildrenweremissing oneassessment,and21weremissingtwoormore.

Forourmainanalysis,we utilizeanaggregate“schoolreadiness” scoreasaweightedaverageofthesixindividualassessments.Weem-

4 The3childrenwithnoavailableregistrydataarelikelytobefamiliesre- centlyimmigratedtoNorway.

ployconfirmatoryfactoranalysistodeterminetheweightapplied to each assessment score (see Brown, 2014 for details).⁵ In doing so, we effectively assume the six assessments share a common compo- nent (“schoolreadiness” )whichisbestapproximatedemployingthe empirically-derivedweights.Unfortunately,producingthisindexscore isonlypossibleforchildrenthathavescoresonallsixassessments.To maximizeoursample,werunthisanalysisincludingthe26childrenwith asinglemissingscore(N=627total)bygeneratingapredictedvalue (viaOLS)forthemissingscoreasafunctionofthechild’snon-missing scores,estimatedoverthosewithallsixassessments.Thisreducedour estimatesofcenter-levelvariationslightly,butmodestlyimprovedthe poweroftherelevantstatisticaltests.⁶Whileourmainanalysisfocuses onthisaggregatescoreofschoolreadiness,resultsarealsoprovidedfor theindividualassessments.

InTable1,weseethatthesixassessmentscoresarestronglycorre- latedtooneanother,indicatingchildrenwithhighcompetencywithin onedevelopmentalareaarealsolikelytohavehighercompetencyin otherareas.Thesixassessmentscoresalso,byconstruction,correlate stronglywithour indexscoreforschoolreadiness. Thedifferencein thosecorrelationsrevealthattheindexscoreplacedsomewhathigher weightonthemathandworkingmemoryassessments,withrelatively lessweightonphonologicalawareness.

The67childcarecentersrepresentaselectedsample,whichmatters ininterpretingourresults.Thesampleincludeschildcarecentersinthe AgdercountiesofsouthernNorwaythatself-selectedforparticipation intheAgder-project.TheAgderregionisnotrepresentativeofNorway asawhole. Inthisregion,workforceparticipationislowerandwel- fareparticipationhigherthantherestofNorway.Moreover,itismore commonformotherstostayhomewithyoungchildrenthanintherest of thecountry.Inaddition,thechildcarecentersselecting topartici- pateintheAgder-projectarelikelynotrepresentativeofcentersinthe Agderregion.AstheAgder-projectwasaninterventionstudyinvolving continuingeducationandmoresystematiccurriculumuse,thecenters selectingtoparticipatemayhavebeenthosealreadyworkingmoresys- tematicallytoimproveschoolreadinessskillsthantheaveragechildcare center.Manyofthecenterswhodeclinedourinvitationtoparticipate statedthattheybelievedtheAgder-projecthadtoomuchstructuredac- tivities.Ourfocusoncentersinaparticularregionthatself-selectedinto theAdger-projectisthereforeexpectedtoproduceamorehomogeneous sampleofcentersthanarandomsamplingofNorwegiancenterswould produce,causingestimatesofcenter-levelvariationtobesmallerthan whatexistsinthebroaderuniverseofNorwegiancenters.Additionally, wehadselectionattheindividuallevel, becausenotallparentscon- sentedfortheirchildtoparticipateinthestudy.Amongthemothers ofthechildreninoursample52percenthasuniversityorcollegeedu- cation,whichisslightlyhigherthanthefullpopulationofmothersof fiveyearoldsinAgder(49percent),andslightlylowerthancomparable mothersinallofNorway(56percent).⁷

Table2presentssummarystatisticsofourcovariates.Slightlymore thanhalfofourmainanalyticsample(N=627)isfemale.⁸Birthmonth isacontinuousvariable,takingthevariable1fortheyoungest(born inDecember2011),and12fortheoldest(borninJanuary2011).An averagebirthmonthof6.8impliesthataverageageatassessmentwas 5.15years.Mothersaverage14.3yearsofeducation,whilefathersav- erage13.7years.Mothersaremorelikelytohavecompletedacollege thanfathersinoursample.Fathers’meanearningsareabout70percent higherthanmothers’meanearnings.Asizablefractionofchildrenhada non-westernimmigrantmother(13.6%)orfather(10.5%).Onaverage, centershad11.7childrenincludedinoursample,rangingfrom5to22.

5ImplementedusingStata’s“confa” command.

6Ifwealsopredictthemissingscoresofthosemissingtwoassessments,results areunchanged.

7OwncalculationfromregistrydataprovidedbyStatisticsNorway.

8Summarystatisticspertainingtothefullsample(N=648)arepresentedin AppendixTableA1.

Table1

Pairwisecorrelationbetweentestscores.

Index Math Self-regulation HTKS Vocabulary Working memory Phonological awareness

Math 0.847

Self-regulation HTKS 0.675 0.426

Vocabulary 0.662 0.466 0.364

Working memory 0.756 0.490 0.418 0.401

Phonological awareness 0.432 0.273 0.234 0.249 0.277

Self-regulation H&F 0.592 0.475 0.334 0.210 0.368 0.174

Note:Allcorrelationcoefficientsaresignificantat1percentlevel.“Index” isaweightedmeanofthesixassessmentscores, withweightsdeterminedbyconfirmatoryfactoranalysis(seetextfordetails).

Table2

Summarystatistics.

Mean Std. dev. Obs

Female 0.507 0.500 627

Birth month 6.83 3.194 627

Mother education (# years) 14.25 2.560 607 Father education (# years) 13.68 2.477 599

Mother drop out 0.188 0.391 607

Mother high school 0.283 0.451 607

Mother college 0.529 0.500 607

Father drop out 0.207 0.406 599

Father high school 0.409 0.492 599

Father college 0.384 0.487 599

Mother earning (NOK) 331,260 214,649 627 Father earning (NOK) 563,132 266,978 613

Mother immigrant 0.136 0.343 627

Father immigrant 0.105 0.307 627

# 5 yrs old in center 11.66 5.05 67

5. Analysesandresults 5.1. Empiricalstrategy

We investigate the magnitude of center-level variation in school readinessbyestimating aseriesof fixedandrandomeffectsmodels, representedasfollows:

𝑌_𝑖𝑐=𝛼_𝑐+𝛽𝑋_𝑖+𝜖_𝑖𝑐 (1)

𝑌_𝑖𝑐=𝛼+𝛽𝑋_𝑖+𝜑_𝑐+ 𝜇_𝑖𝑐 (2)

whereY_ic isthescoreforchildiinchildcenterc,andX_i isavector ofchildandparentcharacteristics.Inthefixedeffects(FE)model(1), testsofsignificantcenter-levelvariationinscoresarebasedonanFtest ofthejointsignificanceoftheestimatedfixedeffectterms(𝛼̂_𝑐).Inthe randomeffects(RE)specification(2),centerlevelseffectsareassumed tobenormallydistributedandindependentoftheincludedcovariates, butarenotexplicitlyestimated.⁹Instead,onlythevarianceincenter effectsisestimated,andthetestofsignificantcenter-levelvariationis formedfromthelikelihoodratiotestcomparingtheREmodeltoone whereVar(𝜑c)isassumedtobezero.

Inmostofouranalyses,wefocusonaparsimonioussetofcovari- atesfoundtobethestrongestindividualpredictorsofchildoutcomes – childsex,childbirthmonth,mother’seducation¹⁰andindicatorsof eachparent’simmigrantstatus.“Fullcovariate” modelsadditionallyin- cludecovariatesforfather’seducation,motherearningsandfatherearn-

9 EstimationoftheREmodelisperformedviaMLEusingStata’s“mixed” or

“xtreg/mle” command.GLSestimationoftheREmodelsyieldverysimilares- timatesofthevarianceincentereffects,butdoesn’tallowustotestthesignifi- canceofthatvariance.

10 Educationcontrolswereincludedasidentifiersforhighschoolandcollege completion.

ings.¹¹Forsomechildren,parentalearningsoreducationwasmissing intheregistrydata.Intheregressions,indicatorsfor“missingness” were includedasappropriate.

Beyondtestingforsignificantcenterleveleffects,wealsohaveanin- terestinquantifyingthemagnitudeofcenter-leveldifferencesinschool readiness. Due tothemodest numbersof childrendrawnfrom each center,theestimatedcenterfixedeffectssufferfrom“over-fitting.” The variationincenterfixedeffectsthereforepredictablyoverstatesthetrue levelofcenter-levelvariation.Ourestimatesforthemagnitudeofcenter- levelvariationarethereforebasedonestimatesof Var(𝜑c) intheRE models.SinceestimatesofVar(𝜑c)arenotparticularlyintuitive,weuse thisestimatetoinferthepredicteddifferenceinschoolreadinessscores acrosscentersatthe90thand10thpercentileofthe𝜑cdistribution.In ourresultstables,thisresultispresentedasthe“good/baddifference.” TheREmodelisalsoextendedtoexplorewhethercenterqualityhas differentialeffectsonchildrenbasedontheirfamily’ssocio-economic status,for whichweuse mother’seducationasaproxy.Specifically, Eq.(2)isextendedtoallowtheinfluenceofmother’seducationtovary acrosscentersindifferentpartsofthe𝜑cdistribution.¹²Themodelwe estimateisfrequentlycalledamixed-effectslinearregressionmodeland takestheform

𝑌_𝑖𝑐=𝛼+𝛽𝑋_𝑖+𝛾_𝑐𝑀𝑒𝑑_𝑖+𝜑_𝑐+ 𝜇_𝑖𝑐 (3) withvariancetermsestimatedforVar(𝜑c),Var(𝛾c),andCov(𝜑c,𝛾c).¹³A negativevalueonthecovariancetermwouldindicatethatthepositive gradientbetweenmaternaleducation(Med_i)andchildreadinessscores issmallerforchildrenincentersathigherpointsonthe𝜑cdistribution.

5.2. Center-levelvariationinassessmentscores

AsdocumentedinTable3,thefixedeffectsmodelssupporttheno- tionofsubstantialheterogeneityintheschoolreadinessofchildrenin differentcenters.Column1reportsresultswithnocovariates,andthe estimated fixedeffects arehighlysignificant(p<.01).Theestimated centervariationdecreaseswhentheparsimonioussetofcovariatesare included(column2),likelyreflectingselectionintoneighborhoodsand thatmostchildrenattendchildcareatacenterintheirneighborhood, buttheestimatedfixedeffectsremainstatisticallysignificant(p=.04).

Notably,thetestofincludingtherichersetofcovariateshasvirtually no effect onthevariance inestimated centereffects, though itdoes weakenthefindingofsignificantcentereffects(p=.07).Inlightofour smallsamplesize,includingextraneouscovariatesinflictsaparticularly highcostintermsofthepowerofstatisticaltests,sowefocusonmod- elscontainingtheparsimoniouscovariatesintheremainingcolumnsof Table3.¹⁴

11Covariatesforearningsareenteredasindicatorsforeachquartileinthe earningsdistributioninoursample.

12Thisanalysisrequiredustocontrolformother’seducationlinearly,instead ofincategories.Thisdifferencehadnoeffectonourmainresults.

13EstimationisperformedusingStata’s“mixed” commandwiththe“covari- ance(unstructured)” option.

14CoefficientestimatesfortheincludedcovariatesarepresentedinAppendix TableA2.

Table3

Center-levelvariationinschoolreadinessscores.

(1) (2) (3) (4) (5) (6) (7)

Model FE FE FE RE RE RE RE

levels: center center center center municipality, center testing time, center testing day, center covariates none parsimonious full parsimonious parsimonious parsimonious parsimonious

Adj. R2 0.059 0.217 0.220

𝑆𝑑( ̂𝛼_𝑐) 0.431 0.365 0.362

F test 0.003 ^∗∗ 0.037 ^∗ 0.068 +

𝑆 𝑑( ̂𝜑 _𝑐) 0.168 0.106 0.168 0.115

𝑆 𝑑( ̂𝜑 _𝑚) 0.154

𝑆 𝑑( ̂𝜑 _𝑡𝑡) 0.000

𝑆 𝑑( ̂𝜑 _𝑡𝑑) 0.128

Hausman test 0.376

LR test 0.045 ^∗ 0.299 – 0.112

Good/bad difference 0.553 0.546

Obs 627 627 627 627 627 627 627

Notes:^∗∗,^∗and+denotesignificanceat1,5and10percentlevel,respectively.Parsimoniouscovariatesincludechildgenderandbirth month,mothereducationlevelandparentimmigrantstatus.Thefullsetofcovariatesadditionallyincludesfathereducationleveland parentearnings.𝑆𝑑(̂𝛼𝑐)isthestandarddeviationoftheestimatedcenterfixedeffects.Ftestreportsthesignificancelevel(p-value)ofthe fixedeffects.𝑆𝑑(̂𝜑_𝑐)istheestimatestandarddeviationincenterrandomeffects.𝑆𝑑(̂𝜑_𝑚),𝑆𝑑(̂𝜑_𝑡𝑡)and𝑆𝑑(̂𝜑_𝑡𝑑)arethecorrespondingstandard deviationinmunicipality,testingtime(hour)andtestingdaterandomeffects.Incolumns4and8,theLRtestreportsthesignificancelevel forvariationincenterrandomeffects.Incolumns5and7,theLRtestreportsthesignificanceoftheimprovedfitofferedbyinclusion oftheadditionallevelofrandomeffect(i.erelativetocolumn4).TheGood/Badvaluecalculatesthedifferenceinexpectedassessment scoresbetweenthe90^thand10^thpercentilesonthedistributionofchildcentereffects,calculatedas(+/-)1.645•𝑆𝑑(̂𝜑_𝑐).

Incolumn4ofTable3,weestimatetherandomeffectsanalogueto thefixedeffectsmodelincolumn2.TheHausmantestfailstoreject theREmodelinfavoroftheFEspecification,asignthattheestimated coefficientsarenotmuchaffectedbytheinclusionoffixed(versusran- dom)effects.Theabsenceofcenter-levelvariationcanagainberejected atconventionallevels(p=.045).Theestimatedvariationincenteref- fectswasthenusedtopredictthedifferenceinscoresexpectedacross childrenincentersatthe90thand10thpercentileinthecentereffect distribution.Thedifferenceinscoresamountstooveronehalf(0.55) of astandarddeviation.While thissuggestssizable centervariation, thespecificestimateshouldbeinterpretedwithcaution,sincewehave onlyamodestnumbersofcentersandfewchildrenpercenter.The95%

confidenceintervalaroundtheestimatedstandarddeviationis[0.083, 0.338].

Incolumns5–7weexploreanumberofconcernspertainingtothe interpretationofthesefindingsasevidenceofvariationincenterquality.

First,incolumn5,weaugmenttheREspecificationtoallowrandom effectsatthelevelofmunicipalityaswell,ofwhich17arerepresented.

ThelikelihoodratiotestfailstorejecttheREspecificationwithcenter effectsinfavorofthespecificationwithrandomeffectsatbothlevels.

Nonetheless,mostofvariationoriginallyattributedtocentersappearsto reflectcross-municipalityvariation.Whilethisisaninterestingfinding, wedonotbelieveitunderminestheplausibilityofcenterqualitydriving thevariationinschoolreadinessweobserve.Totheextentlocalpolicies anddifferentialfundinginfluencecenterquality,thesewouldemergeas variationatthemunicipality,notcenter,level.

Alargerconcernrelatestothefactthatclassmateswerealmostal- waystestedonthesamedayandatthesametime.¹⁵Ifchildrentest betteratcertaintimesofday,oriftheatmospherefortestingwasbetter onsomedaysthanothers,theseforcescouldartificiallyinflatetheim- pressionofcenter-levelvariation.Incolumn6,weevaluatethisconcern inrelationtothetimeofdaythechildrenfromaparticularcenterwere tested,augmentingtheREspecificationtoallowrandomeffectsrelated tothetimestudentsweretested(inhourlygroupings).Noevidenceof testingtimeeffectsissupported.

Column7includesrandomeffectsatthelevelof“testingday” (and center).Herewefindsizableestimatesforvariationattheleveloftest-

15 Infourcasesthechildrenweretestedonadifferentdaythantherestoftheir peersintheircenter.Droppingthesehadnoeffectonourestimates.

ingday,whichsuggestsouroriginalestimateofcentervariationmay be substantiallyinflatedbytestingdayeffects.Unfortunately,theas- sessmentdatawerenotcollectedwithaneyetowardsdifferentiating centerandtestingdayeffects.Thefactthatchildrenfromthesamecen- termostlytestedonthesameday,andonlyahandfulofcenters(6–9) weretestedonanygivenday(seeAppendixTableA3),preventsusfrom generatingameaningfuldecompositionofthecenterandtestingdayef- fects.Bychance,meancenterqualityisexpectedtovaryacrossdifferent testingdays,and,totheextentitdoes,wouldcausetheappearanceof testingdayeffectsthatareactuallyareflectionofvariationincenter quality.

Weareableshedmorelightontheexistenceofpotentialtestingday effectsbyutilizingdatafromasecondroundofassessmentscollected on thesamechildren10 monthslater (T2).Importantly, thecenters testingonthesamedayatbaseline(T1)didnotalwaystestonthesame dayforthefollowupassessments,thoughtheyoftendid(seeAppendix TableA3).

InModel1ofTable4,weshowbywayofsimpleOLSregression howassessmentscoresatbaseline(T1)varybytestingday.Iftheseesti- matesreflect“true” testingdayeffects,wewouldexpectthoseeffectsto betransitory;thedayonwhichthebaselineassessmentwasconducted wouldnotbeexpectedtoaffectsubsequent(T2)scores.If,instead,these estimatesreflectdifferences inthe(true)schoolreadinessofchildren whohappenedtotestondifferentdays,wewouldexpectT1testingday effectstobesimilaracrossthetwoperiodsofoutcomes.InModel2,we findstrongevidenceforthelatter.TestingdayatT1significantlypre- dictsT2testscores,andthepatternofestimatesareverysimilaracross thetwooutcomes.AnF-testforthejointsignificanceofthedifference intestingdayeffectsstronglyfailstorejecttheirequivalence(p-value of0.98).Incontrast,testingdayatT2hasamoremodestcorrelation withT2outcomes(seeModel3).Whenbothsetsoftesting daysare controlledfor,itisthetestingdayatT1whichmorestronglypredicts T2outcomes(seeModel4).

Together,theseresultssuggestourestimateof centervariationin Table3,Model4isnotsubstantiallycontaminated(biasedupwards)by legitimatetestingdayeffects.Instead,thevariationinassessmentscores acrosstestingdaysappearstobeanartifactofthesmallcountofcenters testingoneachdayandontheidiosyncraticgroupingsofcenterswho happenedtotestonthesamedays.Whilewecannotruleoutsomeup- wardsbiasinourestimateofcentervariationarisingthroughdifferent

Table4

Testingdayeffectsbytestingperiod.

(1) (2) (3) (4)

Test score period T1 T2 T2 T2

T1 testing day:

Monday — — —

Tuesday 0.083 0.267 − 0.013

(0.159) (0.180) (0.230)

Wednesday 0.245 0.341 ^∗ 0.046

(0.154) (0.169) (0.228) Thursday − 0.104 − 0.052 − 0.298 (0.157) (0.170) (0.236)

Friday 0.334 ^∗ 0.440 ^∗ 0.265

(0.160) (0.171) (0.194)

Monday 0.372 0.348 0.083

(0.319) (0.333) (0.362)

Tuesday 0.200 0.204 − 0.212

(0.151) (0.165) (0.958)

Wednesday 0.061 0.191 − 0.134

(0.149) (0.161) (0.970) Thursday − 0.235 − 0.190 − 0.916 (0.168) (0.185) (1.015)

Friday 0.327 ^∗ 0.335 + − 0.356

(0.164) (0.179) (1.019) T2 testing day:

Tuesday — —

Wednesday − 0.246 − 0.255

(0.158) (0.159)

Thursday − 0.184 0.289

(0.158) (0.317)

Friday − 0.131 0.219

(0.166) (0.371)

Monday − 0.333 ^∗ − 0.545

(0.157) (0.964)

Tuesday 0.083 − 0.049

(0.160) (0.953)

Wednesday − 0.183 − 0.257

(0.159) (0.953)

Thursday − 0.114 − 0.176

(0.149) (0.937)

Friday 0.160 − 0.176

(0.161) (0.965) F tests

T1 testing days 0.006 0.009 0.037

T2 testing days 0.072 0.239

T1/T2 difference 0.982

Adj R2 0.226 0.171 0.160 0.186

Obs 627 578 578 578

Note:ResultsfromOLSregressionsonaggregateschoolreadiness scoresinbaselineperiod(T1)andatfollowup(T2).Parsimo- niouscovariatesincludedinallspecifications(resultsnotshown), aswellasastudent-specific“treatment” indicatorinT2models.

Ftestsprovidep-valueforjointsignificanceofincludedtesting dayindicatorsand,acrosscolumns1and2,thep-valuetesting whether(T1)testingdayhasdifferenteffectsonT1outcomesthan onT2outcomes.

assessmentconditions,ourevidencesuggeststestingdayeffectsarenot aserioussourceofbias.

Afinalrobustnesstestofthissortwasestimatedtoverifythatthe matchingofchildrentoindividualtestersisnotbiasingourresults.In- cluding“tester” identifiershadvirtuallynoeffectonourestimatesof center-levelvariation,andconsequentlyappeartobeofnoconcern.In theREspecification,theestimatedstandarddeviationofthecenteref- fectsincreasedfrom0.168to0.170whentestereffectsareincluded.

5.3. Evidenceofselection

TherobustnessoftheFEresultstorichercovariatesandthefailure oftheHausmantesttorejectbothsuggestthatselectioneffectsareun- likelytobeimportantdriversbiasingourestimatesofcentervariation.

Toinvestigateselectionmoreformally,foreachchildweconstructeda proxyforcenterqualityastheleave-outresidualscore(i.e.conditional

Table5

Selectionintochildcarecenters.

(1) (2) (3)

Predicted test score 0.048 (0.042)

Mother immigrant 0.017 0.014

(0.051) (0.050) Father immigrant 0.090 + 0.108 + (0.052) (0.061) Mother high school degree 0.050 0.030

(0.043) (0.046) Mother college degree 0.030 − 0.017 (0.051) (0.050) Father high school degree 0.076 + (0.041)

Father college degree 0.089 ^∗

(0.038) Mother 2. quartile earning − 0.013 (0.048) Mother 3. quartile earning − 0.003 (0.038) Mother 4. quartile earning 0.039

(0.054) Father 2. quartile earning − 0.064 (0.041) Father 3. quartile earning − 0.039 (0.047) Father 4. quartile earning 0.010

(0.046)

Observations 627 627 627

Adj R2 0.004 0.010 0.030

F test 0.241 0.0849

Note:^∗and+denotesignificanceat5and10percentlevel respectively.EstimatesfromOLSregressionsontheleave- out-meanresidualscorecalculatedovereachchild’sclass- mates.Allmodelsareclusteredoncenterlevel.Ftestreports thejointsignificance(p-value)ofeachsetofincludedco- variates.Predictedtestscoresareconstructedfromthefull setofcovariates.

onthefullcovariateset)overeachchild’sclassmates.Wethentestedvia OLSregressionwhetherchildrenwithcertaincharacteristicsweremore likelytoattendacenterwith“over-performing” children.Theresultsof thisexercisearepresentedinTable5.

Incolumn1,wefindaweaklypositiveandstatisticallyinsignificant relationshipbetweenachild’sownpredictedscoreandtheLOMresidual scoreofherclassmates.Focusingonourparsimoniouscovariatesincol- umn2,highermaternaleducationisonlyweaklypredictiveofselection intoacenterwithhigherperformingclassmates.Somewhatcontraryto expectations,wefindthatfather’simmigrationstatusisapositivepre- dictorofselectionintoacenterwithhigherperformingclassmates,with anestimatethatreachesmarginalsignificance.Thisfindingisrobustto theinclusionofthefullsetofcovariatesincolumn3.Moreover,theesti- matedeffectsoffather’seducationindicatechildrenofhighereducated fatherstendtoselectintohigherqualitycenters.Thatsaid,aswedocu- mentedearlier,theinclusionoftherichercovariateshadlittleeffecton ourestimatesofcentervariation.Whilewefindsomeevidenceofselec- tion,thereislittlesupportforthenotionthatchildrenwiththestrongest SESbackgroundsaredifferentiallyselectingintobettercenterstoany greatdegree.Thismaysuggestthatthereislimiteddifferencesincenter qualityacrossdifferenttypesofneighborhoods,andthatlocationisthe primaryconcernforparentswhenchoosingchildcarecenter.Moreover, itmaybehardforparentstoobservecenterquality.

5.4. Mechanisminvestigation

Table6explorespotentialmechanismsbehindcenter-levelvariation inscoresfocusingonalimitednumberofstructuralcharacteristicswe wereabletoobtainforasubsetofthecentersinourstudy.Thisevidence

Table6

Impactofcentercharacteristicsonschoolreadiness.

(1) (2) (3) (4) (5)

Director #yrs tenure 0.007 0.007

(0.007) (0.007)

Teacher/Child ratio 2.211 ^∗∗ 2.289 ^∗∗

(0.750) (0.650)

Center size 0.007 0.007

(0.008) (0.007) 𝑆 𝑑( ̂𝜑 _𝑐) 0.134 0.113 0.113 0.125 0.067

LR test 0.171 0.255 0.242 0.204 0.405

Observations 482 482 482 482 482

# centers 52 52 52 52 52

Notes:^∗∗denotessignificanceat1percentlevel.Estimatesfromrandom effectsregressionsanalogoustoTable3,column4.Allregressionsinclude covariatesforchildgenderandbirthmonth,mothereducationleveland parentimmigrantstatus(notreportedintable).Reportedstandardserrors arecorrectedforclusteringatthelevelofcenters.𝑆𝑑(̂𝜑_𝑐)istheestimated standarddeviationoftherandomcentereffects.LRtestreportsthesignif- icancelevel(p-value)ofcenterrandomeffects

shouldnotbeinterpretedascasualasthestructuralindicatorsarenot randomandmaybeassociated withotherunobservedindicators for quality.For52centersinourstudy,weobtaineddataonthefollowing characteristics:thetenureofthecenterdirector,theteacher-childratio andcentersize.Duetothemuchreducedsamplesize,wereplicatethe resultsofourpreferredREspecificationincolumn1,findingthevaria- tiontobesomewhatsmallerand(unsurprisingly)nolongersignificant.

Incolumns2–4,weaugmentthismodelwithcontrolsforeachofthe structuralcharacteristics,andincolumn5weincludeallthreesimulta- neously.Theteacher-childratioprovestobeastrongpositivepredictor ofschoolreadinessscores.Theresultincolumn3suggeststhat30%

oftheoriginallyestimatedvarianceincentereffectsispotentiallyat- tributabletovariationinchild-teacherratios.Column5suggestalmost 75%oftheoriginallyestimatedvarianceincentereffectsispotentially explainedbythethreestructuralcharacteristics,thoughthisshouldbe regardedcautiouslyconsideringtwoofthethreecoefficientsarestatis- ticallyinsignificant.Nonetheless,child-teacherratios(atleast)appears tobeastrongcontributortothecentervariationweestimate.

5.5. Heterogeneouscentereffectsbymaternaleducation

Mixedeffectsmodelswereestimatedtoinvestigatewhetherthein- fluenceofcenterson schoolreadinessismediatedbyfamilySES,for whichmother’seducationservesasproxy.Ifhigherqualitycentersare effectiveatreducingtheschoolreadinessgapbetweenhighandlow SESchildren,wewouldexpectthe𝛾cand𝜑ctermsinEq.(3)tocovary negatively.¹⁶

EstimationresultsarepresentedinTable7.Unfortunately,themixed effectsmodelsprovedtooweakly-poweredtoprovidesubstantiveevi- dencetowardsthishypothesis.Thebottomofcolumn2reportsthelike- lihoodratiotestforwhethertheallowancefortheadditionalvariance terms,Var(𝛾c)andCov(𝜑c,𝛾c),significantlyimprovesmodelfitoverthe originalrandomeffectsspecification,reportedtocolumn1.¹⁷Wefailto findevidenceforsignificantlyimprovedfit(p=.33).Sowhilethedirec- tionalevidencesuggestschildrenoflowerSESmightbenefitmorefrom placementinahigherqualitycenter,weareunabletodrawstrongcon- clusioninthatregard.Thesameistruewhenifweperformthisexercise overtheindividualassessments;whileestimatesofthecovarianceterm

16 Ifchildrenofhighereducatedmothershadtheirkidsinchildcarelonger, wewouldexpectanyeffectofcenterqualitytobehigheramongthose,andthat wouldbiastheestimatedmagnitudeofanegativecovariancetermdownwards.

17 Asnotedearlier,forthisspecificationmaternaleducationiscontrolledfor linearlytofacilitateestimationofthemixedmodel.

Table7

Mixedeffectsbycenterandmothereducation.

(1) (2)

random effect:

mixed effect:

center Center

mother education

covariates: parsimonious parsimonious

𝑉 ̂𝑎𝑟 ( 𝜑 _𝑐) 0.028

(0.020) 0.094

(0.193)

𝑉 ̂𝑎𝑟 ( 𝛾_𝑐) 0.001

(0.002)

𝐶𝑜𝑣 ̂( 𝜑 _𝑐, 𝛾_𝑐) − 0.011

(0.018)

LR test 0.048 0.333

Obs 627 627

Note:Estimatesfrommixedeffectsregressionsonstandardizedas- sessmentscore.Parsimonioussetofcovariatesincludechildgender andbirthmonth,mothernumberofyearseducationandparentim- migrantstatus.𝑉̂𝑎𝑟(𝜑_𝑐)istheestimatedvarianceincentereffects (andstandarderror),𝑉̂𝑎𝑟(𝛾𝑐)istheestimatedvarianceintheeffects ofmothereducation,and𝐶𝑜𝑣̂(𝜑_𝑐,𝛾_𝑐)istheestimatedcovariancebe- tweenthetwo.LRtestincolumn1reportsthesignificancelevelfor variationincenterrandomeffects.LRtestincolumn2reportsthe significanceoftheimprovedfitofferedbymodelingtheadditional varianceterms(i.e.relativetocolumn1).

aregenerallynegative,theyneverapproachlevelsofstatisticalsignifi- cance.

5.6. Estimationbysubject

Table8reportstheresultsofFEandREspecificationswithparsi- moniouscontrolsforeachofthesixindividualassessmenttests.Several importantdistinctionsarerevealedasfarastheinfluenceoffamilychar- acteristics.Maternaleducationappearstobeanespeciallystrongpre- dictorofself-regulationHTKSscores.Thechildrenofcollegeeducated mothersalsoscorestronglyonmath,vocabularyandworkingmemory, buttheeffectsofmaternaleducationaremoremutedforphonological awarenessandtheself-regulationheartsandflowerstest.Notsurpris- ingly,thechildrenofimmigrantparentsperformespeciallypoorlyon thevocabularytest.Femaleeffectsaregenerallypositiveandsignificant, butinsignificant(andnegative)forvocabularyandtheself-regulation heartsandflowerstest.Ageeffectsareconsistentlypositive,butsub- stantiallysmallerforphonologicalawareness.

Turning tothemainfindingsof interest,for eachassessment the HausmantestfailstorejecttheREspecificationinfavoroftheFEmodel.

Formath,self-regulationHTKSandvocabularywefindsignificantevi- denceofcenter-levelvariation.Whiletheresultsfortheworkingmem- orytestfailtoachievestatisticalsignificance,theestimatedvariance termsarecomparableinsizetothemathtest.Incontrast,centerlevel variationappearsweakerforphonologicalawarenessandisespecially weakfortheself-regulationheartsandflowerstest.

6. Discussion

Ourpaperinvestigatesdifferencesinchildcarequalityacrosscenters intheuniversalchildcaresystemofNorway.Thisisimportantbecause theECECliteraturesuggeststhatifchildcarequalityvariesacrosscen- ters,universalchildcaresystemscanpotentiallyincreasedisparitiesin schoolreadiness,particularlyifthechildrenofparentswithhigheduca- tionselectintocentersofhighquality.Ouranalysisdemonstrateslarge andsignificantvariationinschoolreadinessacrosscenters.Indeed,the differenceinschoolreadinessskillsincentersatthe90thand10thper- centileinthecentereffectdistributionwasestimatedtobeoveronehalf (0.55)ofastandarddeviation.Theseresultsarerobusttonumerousro- bustnesstestsanddonotseemtobedrivenbyparentalselectioninto centersofdifferentquality.