college expansion and the marginal returns to education ...ftp.iza.org/dp8735.pdf · returns to...
TRANSCRIPT
DI
SC
US
SI
ON
P
AP
ER
S
ER
IE
S
Forschungsinstitut zur Zukunft der ArbeitInstitute for the Study of Labor
College Expansion and the Marginal Returns to Education: Evidence from Russia
IZA DP No. 8735
December 2014
Olga BelskayaKlara Sabirianova PeterChristian Posso
College Expansion and the Marginal
Returns to Education: Evidence from Russia
Olga Belskaya UNC‐Chapel Hill
Klara Sabirianova Peter
UNC‐Chapel Hill, IZA and CEPR
Christian Posso
UNC‐Chapel Hill
Discussion Paper No. 8735 December 2014
IZA
P.O. Box 7240 53072 Bonn
Germany
Phone: +49-228-3894-0 Fax: +49-228-3894-180
E-mail: [email protected]
Any opinions expressed here are those of the author(s) and not those of IZA. Research published in this series may include views on policy, but the institute itself takes no institutional policy positions. The IZA research network is committed to the IZA Guiding Principles of Research Integrity. The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center and a place of communication between science, politics and business. IZA is an independent nonprofit organization supported by Deutsche Post Foundation. The center is associated with the University of Bonn and offers a stimulating research environment through its international network, workshops and conferences, data service, project support, research visits and doctoral program. IZA engages in (i) original and internationally competitive research in all fields of labor economics, (ii) development of policy concepts, and (iii) dissemination of research results and concepts to the interested public. IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion. Citation of such a paper should account for its provisional character. A revised version may be available directly from the author.
IZA Discussion Paper No. 8735 December 2014
ABSTRACT
College Expansion and the Marginal Returns to Education: Evidence from Russia*
This paper evaluates whether the expansion of higher education is economically worthwhile based on a recent surge in the number of campuses and college graduates in Russia. Our empirical strategy relies on the marginal treatment effect method in both normal and semi‐parametric versions, and estimating policy‐influenced treatment parameters for the marginal students who are directly affected by college expansion. We use high‐quality panel data with multiple wage observations, many birth cohorts, disaggregated location information, and past economic conditions. We find that college expansion attracts individuals with lower returns to college, but the returns for marginal students vary considerably depending on the scale of expansion and the type of location where new campuses are opened. Marginal individuals in smaller cities and locations without college campuses receive the largest benefits from new campuses. The results provide important implications for the design of policies targeting the expansion of higher education. JEL Classification: J31, I21, P36 Keywords: education, returns to college, marginal treatment effect,
policy‐relevant treatment effect, local instrumental variable, selection bias, college expansion, Russia
Corresponding author: Klara Sabirianova Peter Department of Economics Carolina Population Center University of North Carolina-Chapel Hill Chapel Hill, NC 27599 USA E-mail: [email protected]
* We are thankful to Brian McManus, Saraswata Chaudhuri, Donna Gilleskie, and Tiago Pires for useful comments. We also acknowledge helpful comments received from participants of the UNC Applied Microeconomics Workshop and Southern Economic Association meetings in Atlanta, GA.
1
1.Introduction
Overthelast20years,manycountrieshaveexpandedhighereducationandincreasedthe
number of college graduates. Between 1990 and 2010, the number of students in tertiary
educationper100,000peoplehastripledinBrazil, IndiaandRussia,andincreasedtwelve‐foldin
China; thenumberof collegeshas also increased several‐fold in the largest developing countries
(Carnoy et al. 2013). Answering pertinent policy question – whether such rapid and massive
collegeexpansion iseconomicallyworthwhile–requiresproperanalytical tools. Toevaluate the
effectivenessofcollegeexpansionpolicies,analystsneedtoestimatethereturnstocollegeforthe
marginalindividualwhopreviouslywouldnothavehadanopportunitytostudyincollegeandwho
isaffectedbycollegeexpansion.
Weevaluatethemarginalreturnstocollegeinresponsetocollegeopeningsusingthecase
ofRussia.1 Between1992and2003, thenumberof students inRussianhighereducationalmost
tripled, the number of universities increased two‐fold, and the number ofmunicipalitieswith at
leastonecampusmorethandoubled.2ThelegalframeworkfortheRussiancollegeexpansionwas
provided by the 1992 Law on Education,which allowed the opening of private universities and
legalized tuition‐based programs and branches in public universities (Federal Law 1992).3 The
earlyexpansionperiodwasaccompaniedbyarapidincreaseintheaveragecollegewagepremium,
ascanbeseeninFigure1.Followingthelow‐returnSovietperiodwithitscompressedcentralized
wage structure, the OLS‐estimated averagewage premium per year of college education surged
from 2.8 percent in 1990 to about 8.7 percent in 1998 during Russia’s transition to a market
economy,butthenthepremiumstabilized.
1WeprovideabriefsummaryoftheSovietandRussiansystemsofhighereducationinthewebappendix.2SourcesareRegionsofRussia(1995,2004)andauthors’calculations.3 Russian institutions of higher education are referred to by a variety of names, including “Universities,”“Institutes,” “Academies,” and “Higher Schools”. For compatibility of terminology with internationalliteratureonthistopic,thispaperusestheterms“Universities”and“Colleges”,eventhoughthelattertermisnottechnicallyaccuratefromtheRussianlanguagepointofview.Thus,suchtermsas“College,”“University,”and“Institutionofhighereducation”areusedinterchangeablythroughoutthispaper.
2
Ourempiricalstrategyreliesonthemarginaltreatmenteffect(MTE)methodinitsnormal
and semi‐parametric versions (Heckman and Vytlacil 2001, 2005, 2007) and a construction of
policy treatment parameters (Heckman and Vytlacil 2001; Carneiro et al. 2010). Unlike the
previousMTEstudiesthatarebasedoneithercross‐sectionaldataorone‐cohortpanelssuchasthe
NationalLongitudinalSurveyofYouth(NLSY),westudymultiplecohorts,includingonesthatmade
their college decision in the pre‐expansion period (1985‐1992) as well as cohorts that entered
collegeduring theexpansionperiod(1993‐2003). Thus,wecanevaluate theoutcomesofactual,
real‐world college expansion rather than a hypothetical increase in the probability of college
attendance.
The identificationofMTErelieson thevariation invariablesdirectlyaffecting thecollege
decision. These variables capture the institutional environment and economic conditions that
prevail during individuals’ late teenage years, the time when individuals make their college
decisions.4Ourkeyinstrument,thenumberofcampusesinthemunicipalityofresidenceatage17,
measurestheextentofcollegeexpansioninRussiaovertimeandacrosslocations.Thisinstrument
is an improvement over a commonly used binary instrumental variable (IV) ‐ the presence of a
collegeintheU.S.countyofresidenceduringteenageyears(Card1995;CameronandTaber2004),
asitallowsforcomputingpolicy‐relevanttreatmenteffectsseparatelyfortheestablishmentofthe
firstcampusinthelocalitythatneverhadacollegeaswellasforthemarginalchangeinthenumber
ofcampuses.Sincetheinstrumentisnumeric,itspolynomialfunctionalformisutilizedtocapture
the strong non‐linear effects of college availability on college attainment and on the marginal
returns. Wealsoprovideseveralargumentsandstatisticalteststooffsetpotentialconcernswith
thevalidityofthisinstrument.
ConsistentwiththeMTEliterature,wefindthatthereturns tocollegeareheterogeneous,
and that individualswith thehighestpropensity togo tocollegeenjoy the largestmarginalgains
4SeeCard(2001)forareviewofthevariables(mainlycostshifters)thataffecttheschoolingdecision.
3
from college education. We also find strong evidence of the positive sorting of individuals into
college basedonunobservedgains. Themagnitudeof the sorting gain inRussia – a5.7percent
wageincreaseperyearofcollege–fallsbetweenthecorrespondingestimateof2percentforChina
(HeckmanandLi2004)and7.6percentfortheU.S.(Carneiroetal.2011b).5
Inadditiontoconventionaltreatmentparameters,weestimatethereturnstocollegeforthe
individualswhochangecollegeparticipation inresponsetoamarginal increase inthenumberof
campuses.Theestimatedmarginalpolicy‐relevanttreatmenteffectparameterof9.6percentwage
increase per year of college indicates large gains for the marginal individuals affected by the
establishmentofnewcampuses.Ourpolicysimulationsshowthattheopeningofacollegecampus
in constrained municipalities – smaller non‐capital cities or municipalities that did not have
institutionsofhighereducationpriortocollegeexpansion–attractsstudentswithhigherreturns
compared to the effect of the same policy in unconstrained municipalities with more college
choices.Inanotherpolicysimulation,weshowthatifthenumberofcampusespermunicipalitydid
not increase or remained at the 1992 level, then a considerable portion of populationwith high
potentialgainsfromcollegeeducationwouldnothavebeenabletorealizethesegains.Wealsofind
larger discounted net benefits for students from constrained municipalities by applying the
estimatedreturnsinthetraditionalcost‐benefitanalysis.
Our paper contributes to several strands of the literature. First, we contribute to the
literatureon themarginal treatmenteffectof collegeeducation.6 Mostof this literatureuses the
NationalLongitudinalSurveyofYouth(NLSY)andconsiderschangesintuitionasapolicyaffecting
the collegedecisions (Carneiroetal. 2010, 2011b;Carneiro andLee2009;Heckmanetal. 2014;
5 Our results support a “comparative advantage model” of the labor market rather than a “single abilitymodel”,which isconsistentwiththepast literatureonself‐selectionthatgoesbacktoRoy(1951). A largenumber of earlier studies established empirical evidence of the non‐random sorting of individuals intodifferent levelsofeducationandcalled for theselectivitycorrection in thereturns toschooling (e.g.,WillisandRosen1979,Garen1984,BjörklundandMoffitt1987). 6Outsidethefieldofeducation,theMTEmethodisonlybeginningtobeappliedinasystematicway;e.g.,seeBasu et al. (2007) and Evans and Basu (2011) for theMTE applications in health related topics. Moffitt(2014)usestheMTEmethodtoestimatetheeffectofatransferprogramonlaborsupply.
4
HeckmanandVytlacil2001,2005).MTEestimatesoutsidetheU.S.arerareandoftenlimitedtoa
singlecrosssection(Carneiroetal.(2011a)forIndonesia;HeckmanandLi(2004)forChina;Kyui
(2013)forRussia;Zamarro(2010)forSpain).7CollegeexpansionintheMTEliteratureiseithernot
modelled at all or simulated as an additive exogenous change in the probability of college
attendance.Thekeycontributionofourstudyistheestimationoftheexpansion‐relatedmarginal
treatmenteffects. Tothebestofourknowledge, this isthefirststudythatutilizesdataonactual
college openings to derive policy‐relevant treatment parameters and evaluate counterfactual
scenariosofcollegeexpansion.Wealsoshowthatthereturnstocollegeformarginalstudentsvary
considerably depending on the scale of college expansion and the location of new campuses.
Additionally,weextendtheMTEliteraturebyusinghigherqualitypaneldatawithmultiplewage
observations, several birth cohorts, disaggregated location information, and past economic
conditions. As shown by Heckman et al. (2006), cohort‐based models fitted on repeated cross
sectionsprovidemore reliable estimatesof the returns toeducation than theestimatesobtained
fromasinglecrosssection.
Second, our study is directly related to the literature that evaluates the effect of college
expansion on the returns to college. A number of earlier papers (Katz andMurphy 1992, Topel
1997,amongothers)linkedtheincreasedsupplyofcollegegraduateswithaloweraveragecollege
wage premium. Carneiro et al. (2011b) and Moffitt (2008) show that an increase in college
participation leads to a decline in the marginal returns to college. We also find that college
expansion draws individuals with lower marginal returns to college. However, the general
equilibriumeffectswereeitheroffsettingornotstrongenoughtoshiftconsiderablytheaggregate
7 Carneiro et al. (2011a) computes the MTE of post‐secondary education in Indonesia in 2000 using thedistance from the current village center to the nearest secondary school as an identifying restriction.HeckmanandLi (2004) estimate theMTEof collegedegree fora cross‐sectional sampleof youngChineseworkersin2000,usinglimitedexclusionrestrictionssuchasparentalincomeandparentaleducation.Kyui(2013) estimates standard college‐related treatment parameters using 2006 RLMS and provides the firstapplicationof theMTEmethodtoRussia. Zamarro(2010)developsanewmethod forestimatingMTEformultipleschoolinglevelsandappliesittotheevaluationoftheeducationreforminSpain.
5
equilibriumskillpricesandtheirdistribution.Inthispaper,wedonotdisentanglevaryinggeneral
equilibriumeffectsofcollegeexpansionandleavethisquestionforfutureresearch.8
Third, thispaper contributes to the literatureevaluating the effectofnewuniversitieson
collegeenrollment. Wefindthattheopeningof the firstcampusinRussianmunicipalitieswhere
there were previously no colleges increases the probability of receiving a college degree by 11
percentage points. Studies fromother countries also report positive, but lower effects of a new
university on college enrollment, e.g., 8 percentage points in Italy (Oppedisano 2011) and 6.4
percentagepointsinCanada(Frenette2009).
Fourth,ouruseofthenumberofcampusesinthemunicipalityofresidenceatage17asa
supply‐side instrument for college education builds upon previous applications of supply‐side
shifters in identifying the returns to schooling; some earlier examples include construction of
elementary schools in Indonesia (Duflo 2001), a dummy for living near college (Card 1995), the
distance to the nearest college (Kane and Rouse 1995) and supply disruptions caused by awar
(IchinoandWinter‐Ebmer2004).
Finally,ourpapercontributes toa large literatureon thereturns toeducationduring the
transitiontoamarketeconomy(Andrénetal.2005,Brainerd1998,Fangetal.2012,Fleisheretal.
2005, Gorodnichenko and Sabirianova Peter 2005,Münich et al. 2005, Yang 2005, amongmany
others). We improve this literature by estimating the distribution of returns, applying a more
rigorous identification strategy, and deriving policy‐relevant parameters for individuals at the
marginofchoice. Yet,wealsoconfirmapreviouslydocumentedpatternof increasingreturns to
schoolingduringtheearlytransitionperiod,followedbytheirlevellingoutinthelaterperiod.
Ourpaperisstructuredasfollows.Inthenextsection,wepresenttheempiricalframework
used in this study. Section 3 discusses the data and identifying variables. Section 4 presents
8Intheaccompanyingpaper,weshowthattheexpansionofhighereducationinRussiaworsenedthequalityof college entrants, reduced resources per student, and increased the market returns to college quality(BelskayaandSabirianovaPeter2014).
6
estimates of the marginal treatment effect and discusses the identification validity. Section 5
performspolicysimulationsandSection6concludes.
2.EconometricFramework
Modelset‐up
To estimate the heterogeneous returns to college, we follow a semi‐structural method
developed by Carneiro etal. (2011b). The decision rule of an individual i is characterized by a
latentvariablemodelofcollegeenrollment:
∗ 0 ,where ∗ (1)
isabinaryvariableindicatingcollegeenrollment;itequalsoneforcollegegraduatesand
zeroforhighschoolgraduates.9Weapproximate byalinearformgivenby ,where isa
vectorofobservablecharacteristicsthataffectthecollegedecision.Vector includesthreetypeof
variables: (i) variables that affect both the schooling decision and wages, ; (ii) variables that
measuretheextentofcollegeexpansionatthetimewhenindividualsmaketheircollegedecisions,
;and(iii)otherinstrumentalvariables/exclusionrestrictionsthatalsoshiftthelatentvariable ∗,
:
(2)
Weassume that is anunobserved to theeconometricianerror term that is statistically
independentof .Itcapturesthemarginalcostofobtainingacollegeeducationforindividuali.
Let P denote the probability of selecting into treatment (college) given . It is
convenient to rewrite the selection equation (2) using the following innocuous transformation.
Define ,where isacumulativedistributionfunction,and isdistributeduniformlyin
9Followingthetraditionoftheliterature,thecollegedrop‐outdecisionisnotmodelledinthisstudy.Russiahasarelativelyhighcompletionrateof79percentofcollegeentrantscompletingatertiary“typeA”programcomparedtotheOECDaverageof69percentandtheU.S.collegecompletionrateof56percent(OECD2008).
7
theunitinterval[0,1].Differentvaluesof denotedifferentpercentilesof .Thus,theselection
equationbecomes
(3)
Let the outcome of interest be ln , where is an hourly wage. The potential
outcomesaredefinedaccordingtothelevelofeducationachievedsuchthat
, , for 0, 1 (4)
where , is a vector of observable characteristics that affect hourly wages, and
includeexogenouswagedeterminantsthatareexcludedfromthecollegechoiceequation(e.g.,an
individual’sageattimetandunanticipatedtransitoryshockstolocallabormarkets). Weassume
that islinear, .
Theobservedoutcome canbewritteninaswitchingregressionform:
, 1 ,
, 1 ,
, , ,
(5)
The gross returns to college education are given by∆ , ,
, , andthemodelassumesthatagentsknowtheirreturns.
Themeanoutcome , conditional on ( , ), is a sumofmeanoutcomes for
eachlevelofeducationandweightedbytheprobabilityofbeingateachlevelofeducation:
E | , E | 1, , E | 0, , 1
E , , | , 6
The vector ( , ) is observed by both the agent and econometrician, while ( , , ) is
knownbytheagentbutunobservedbytheeconometrician.Tounderstandthepossibleeffectsof
unobserved endowments in thewage equation (’s) on the returns,we consider three cases: (i)
unobservedendowmentsarehomogeneous,i.e. , , ̅ forallindividuals,suchthatthelast
term in equation (6) cancels out; (ii) unobserved endowments are heterogeneous but mean
8
independentofcollegedecisions, i.e.,E , , | , E , , ;againthelast
terminequation(6)cancelsout; and(iii)unobservedendowmentsareheterogeneous, , , ,
andcorrelatedwiththeunobservedcharacteristicsfromthecollegedecisionequation( ),inwhich
casethelastterminequation(6)cannotbeignored.
Inthefirsttwocases,thestandardIVapproachmightbesufficienttoidentifytheaverage
returns to college education. In the third case, which is more realistic, individuals, who are
observationally identical from an econometrician’s point of view, may make different college
decisions;thatis,unobservedcharacteristics influencingthecollegedecisionarecorrelatedwith
unobserved endowments ′s in the wage equation. As a result, the returns to college, for
observationally identical individuals, will depend upon a conditional mean component
E | and an individual‐specific unobserved component
E , , | , .
In the third case, the standard IV approach may not be appropriate for estimating the
parameters of interest. It is widely recognized that an IV estimator identifies the local average
treatmenteffect(LATE),whichistheexpectedgainfromthetreatment(e.g.,collegegraduation)of
thoseindividualswhoswitchfromnotreatmenttotreatmentwhenaninstrumentchanges.Inthe
caseofseveralinstruments,eachoftheIVsidentifiesadifferentmarginofthereturntocollegethat
varies across individuals. In general, there isno simple interpretationof the IVestimator in the
presenceofseveralinstrumentsthatinduceindividualstogotocollege.10
MarginalTreatmentEffectfromtheNormalSelectionModel
The marginal treatment effect (MTE) approach allows researchers to obtain the entire
distribution of the individual‐specific returns (i.e., , , | , (Heckman and
Vytlacil2005,2007).Usingequation(6),the , isdefinedasfollows
10HeckmanandVytlacil(2005,2007)presentalimitednumberofcaseswhereitispossibletodeterminetheexactaveragetreatmentparameteridentifiedbyeachoftheinstruments.
9
,E | ,
E , , | , 7
Weassumethat includesatleastoneexclusionrestriction,and and arestatistically
independentof , , , , .Wefirstestimate , fromtheparametricselectionmodelthat
assumes a multivariate normal distribution of errors, , , , , ~N 0, . Under the above
assumptions,E , , | , Cov , , Cov , , Φ ,whereΦ ∙ is
theinverseofthestandardnormalcumulativedistributionfunction,and , isgivenby
, Cov , , Cov , , Φ (8)
HeckmanandVytlacil (2005,2007)showhow,basedon theMTEdistribution, to recover
several standard treatment parameters such as the average treatment effect in the population
, the average treatment effect on the treated , the average treatment effect on the
untreated , the ordinary least squared estimator , and the IV estimator . In
short, each treatment parameter can be obtained as a weighted average of , , where
weightsaregiveninAppendixTableA3.
PolicyParameters
In addition to the standard treatment effect parameters,we also calculate the returns to
collegeforthoseindividualswhowenttocollegeduetothecollegeexpansionpolicy.Theobjective
hereistocalculatethegainsofthemarginalgraduateswhowouldhavehadlimitedaccesstohigher
educationifthecollegeexpansiondidnotoccur,buttheychosetogotocollegeinresponsetothe
openingofmorecollegesintheplaceoftheirresidence. Inotherwords,weareinterestedinthe
returns of the marginal graduates whose college decision was influenced by an increase in the
instrument inequation(2).
To illustrate our parameter of interest, we perform the following experiment using the
estimatedparametersofselectionequation(2).Supposethatanindividual getsadrawgivenby ̃
10
such that ̃ , which implies that 0. Further, suppose that the total
numberofcollegecampusesavailabletoanindividual inhisregionofresidenceattheageof17
increasessuchthatthenewnumberofcampusesisgivenby ∗ .Ourparameterofinterest
is the average of , for those individuals who switch their decision because of the
expansion,suchthat ̃ ∗ ,whichimpliesthat 1.
Heckman and Vytlacil (2001) develop a policy relevant treatment effect ( )
parameterthatmeasurestheaveragereturnstocollegeeducationforindividualsinducedtochange
their schooling decisions in response to a specific policy (for instance, ∗ ). Essentially,
istheaverage , forswitchersandisdefinedas:
, , | | , , , 0, , , 1 (9)
To identify , , the support condition needs to be met, that is, the range of
∗, , must be contained in the range of , , . PRTE is generally applied to a fixed
discretepolicychange.Examplesofspecificpoliciesevaluatedwiththismethodinthepastinclude
atuitionchangeintheUS(HeckmanandVytlacil2001;Carneiroetal.2010,2011b;Eisenhaueret
al. 2015) and a 10 percent reduction in the distance to upper secondary schools in Indonesia
(Carneiro et al. 2011a). The PRTE approach has not been applied yet with respect to college
expansion.
Alternatively,Carneiroetal.(2010)proposedthemarginalpolicyrelevanttreatmenteffect
(MPRTE ),whichcorrespondstoamarginalchangeinpolicy.Thatis,
lim → , . (10)
TheMPRTEmethodessentiallyestimatesthemeanbenefitsofcollegeforthemarginalindividuals
whoareindifferentbetweenparticipatingandnotparticipatingincollege.Theestimatorrequires
theavailabilityofacontinuousinstrument,assuminganinfinitesimalchangein.
11
MarginalPolicyEffectsUsingLocalIV
AnalternativeapproachtoestimatingMTEisbasedonthesemi‐parametricmethodoflocal
instrumental variables (Heckman and Vytlacil 2005). Under the assumption that and are
statisticallyindependentof , , , , ,equation(6)canbere‐writtenasfollows:
E | , (11)
where E , , | , for ∈ .
TheMTEestimationprocessfollowsCarneiroandLee(2009)andconsistsofthreesteps:
1) In the first step, we obtain the propensity score, ̂ , from a probit regression of on
, , .
2) In the second step, we estimate , using a partially linear regression estimator of
Robinson (1988) and then compute ̂ . The estimates
of , areprovidedinthewebappendix.
3) In the third step, we estimate a locally quadratic regression ofR on ̂ and calculate the
derivativesoftheconditionalmeanestimate ′ ̂ .
Aftercompletingthesesteps,theMTEestimatesareobtainedaccordingtoequation(7):
,| ,
′ ̂ (12)
Comparedto thenormalselectionmodel, the local IVmethod forestimatingMTE ismore
flexibleasitdoesnotassumethenormalstructureoftheerrorterm.Butitisalsomorerestrictive
in the sense that the MTE can be estimated with this method only over the common support
of ,whichrarelytakestheentirefullunitinterval[0,1]. Sinceitispracticallyimpossibleto
estimate theMTEover the fullunit interval, thestandardtreatmentparameterssuchasATE,TT,
TUT,aswellasPRTEcannotbeidentifiedusingthesemi‐parametricapproach. Nonetheless, it is
possibletoidentifytheMPRTEparametersbecauseonlythemarginalsupportof isrequired
inthiscase(Carneiroetal.2010).Inourempiricalwork,weapplythelocalIVmethodtorecover
12
the marginal returns to college for the individuals responding to a marginal change in the
probabilityofcollegeparticipationandinthenumberofcampuses.
Thefollowingsectiondiscusseshoweachofthemodelvariables , , ismeasuredin
thedata.
3.DataandIdentificationVariables
The primary data source for this study is the 1995‐2011Russia LongitudinalMonitoring
Survey–HigherSchool ofEconomics (RLMS‐HSE),which is anationally representative stratified
sampleofhouseholdsof theRussianFederation.11 We limit thesample tohighschoolgraduates,
age25andolder,assumingthatthemajorityofpeoplecompletetheirformaleducationbyage25.
Wealso restrict our sample topeoplebornbetween1968and1986. Thus, considering that the
collegedecision inRussia is typicallymadeatage17,oursample includes individualswhomade
theircollegedecisioninthepre‐expansionperiod(1985‐1992)aswellasthosewhodiditduring
the expansion period (1993‐2003). Compared to the previous studies that are based on either
cross‐sectional data or one‐cohort panels such as NLSY, our multi‐cohort panel not only has
multiplewageobservationsovertheindividuallifecycle,butalsopermitstheanalysisofschooling
decisionsforseveralcohorts,thusmakingthestudyofcollegeexpansionfeasible.
Thewage variable is the log of deflated hourlywage rate at primary job. Asmentioned
above, we restrict the period of college decision making to years 1985‐2003. Our preferred
specification employs individual wage data from the post college decision period, 2004‐2011.
However,intherobustnesschecks,weconsiderotherintervalsofwagedata.
11RLMS‐HSEisorganizedbytheNationalResearchUniversityHigherSchoolofEconomics,Moscowtogetherwith theCarolina Population Center at theUniversity ofNorthCarolina at ChapelHill and the Institute ofSociology at the Russian Academy of Sciences. The panel started in 1994. The RLMS‐HSE surveyedindividualsin32outof83regionsandallsevenfederaldistrictsoftheRussianFederation(accordingtotheofficial classification of regions as of January 1, 2010). We drop the year 1994, because the educationquestionnaireinthatyearisincompatiblewiththeoneinsubsequentsurveyrounds.
13
To estimate the college decision equation (2), we define the treatment group as college
graduatesandabove(fouryearsormoreoftertiaryeducationwithadiploma)whereasthecontrol
groupassecondaryschoolgraduates. Thelattercategoryincludesgraduatesofbothgeneraland
professionalsecondaryschools,butexcludescollegedrop‐outs.12Insomeoftherobustnesschecks,
thedefinitionofthetreatmentgroupincludescollegedrop‐outswiththreeormoreyearsofcollege
education.Figure2Ashowsthattheshareofcollegegraduatesincreasedfrom23percentin1995
to39percentin2011inthe25‐44agegroupoftheRLMSsample.
We linkedRLMSrespondentswith the local characteristicsat age17,whichareavailable
either at the level of municipality (such as the number of campuses by category) or at a more
aggregate regional level (such as the size of cohort, earnings, and unemployment rate). The
locationofrespondentsatage17isnotdirectlyreportedbytheRLMSrespondents,but itcanbe
inferredinthemajorityofcasesfromthesurveyquestionsrelatedtomigrationhistory,completion
ofhighschool,andcollegelocation.Forexample,“locationat17”isthesameastheplaceofcurrent
residenceinthreeinstances:(i) for individualswhopermanentlymovedtotheirpresent location
before they turned 18 years old (including those who never moved); (ii) for individuals who
completed high school at present location; and (iii) for individualswho are born in the place of
currentresidenceandwhotemporarilymovedtoanotherlocationafterage17.Collegelocationis
taken as “location at 17” for those respondents who reported to reside in the same community
beforegoingtocollege. Thus, locationatage17isknownwithhighconfidencefor83percentof
oursample.13Thelocationoftheremaining17percentofoursampleatage17(e.g.,thosebornina
differentmunicipalityandmovedtoacurrentresidenceaftercompletingcollege)isunknown.We
12Professional secondaryschools includevocational schools (“PTU”), technical schools (“technikums”)andthe specialized schools that train associate professionals in various fields (such as medicine, education,business,etc.). Theyoffer3‐4‐yearprogramsafter9yearsofsecondaryschoolsor2‐3‐yearprogramsafter11yearsofcompletegeneralsecondaryeducation.13OursampleexcludesrespondentswhowerebornandstudiedoutsideRussia.
14
imputetheirlocationatage17basedoncollegelocationandthetypeofbirthplacesuchasvillage,
township,andcity(seeAppendixTableA1fordetailsofimputation).14
Table 1 categorizes the variables used in the wage and schooling equations. Individual
characteristics such as gender (=1 if female), nationality (=1 if ethnicallyRussian), andmother’s
education(=1ifmotherhasacollegedegree)enterboththewageandschoolingequations.Since
currenturban residence is likely tobeanticipatedat the timeof collegedecision, it theoretically
explains both the wage and schooling outcomes. Year dummies, age, and age squared are
controlled for in the wage equation, while four birth cohort dummies with 5‐year intervals are
included in the college equation. Dummies for theMoscow city and seven federal districts are
determinedatage17inthecollegeequationandatthecurrenttimeinwageequations.
Anothervariablethatisexpectedtoaffectboththecollegedecisionandwageistheregional
cohortsize,whichisthelogofregionalpopulationofage17atthetimewhenanindividualwas17.
Individualsfromlargercohortsarelikelytofacemoreintensecompetitionforalimitednumberof
universityslotsintheirregionandthushavealowerprobabilityofgettingintocollege.Thewage
effectoflargercohortsisambiguousduetoadownwardpressurefromthelaborsupplysideandan
upward pressure from a biggermarket size and potentially higher demand for labor. The data
source for this variable is the Russian Census and other official demographic statistics (see
AppendixTableA1fordetails). Historically,thecohortsizevariedsignificantlynotonlybetween
regionsbutalsowithinregionsovertimeduetothelargedemographicchangesassociatedwiththe
long‐lasting effects ofWWII and low birth rates in the 1990s in Russia (see Figure 2B). In our 14Inthe1990s,morethanhalfoftheinternalmigrationflowswereregisteredwithinagivenregion,e.g.,fromruralareastocitiesofthesameregion(DemographyYearbookofRussia2002).Furthermore,inourRLMSsample,62percentofcollegestudentswiththeknownplaceofbirthstudiedinthesamemunicipalityastheirbirthplace,22percentstudiedinadifferentmunicipalityofthesameregion,andonly16percentstudiedinadifferentregion.Giventhesefacts,wecanrandomlyassignthemissinglocationatage17basedonthetypeof birthplace (village, township, and city), college location for students, and current residence for non‐students. For example, if a studentwasborn in a township, then an arbitrary townshipwouldbe chosenwithin the region of his college as a place of his college decision. For non‐students,we choose a randommunicipalityofthesametypeastheirbirthplacewithintheregionoftheircurrentresidence.Notethatsuchimputationsaffectonlyonevariableofinterest,whichisthenumberofcampusespermunicipality,sinceallotherlocalvariablesareconstructedattheregionallevel.
15
sampleperiodbetween1985and2003,however, the fluctuationsovertimewererelativelymild,
thoughsignificantregionalvariationremained.
NumberofCampusesperMunicipality
Thekeyidentificationvariablethroughwhichweperformpolicysimulationsisthenumber
ofcampusesavailableinamunicipalityatage17.WediscussitsidentificationvalidityinSection4.
This variable is constructed based on the opening/closing dates of campuses from the Russian
university database (Belskaya and Sabirianova Peter 2014). The number of campuses in each
municipalitymeasuresthelocalavailabilityofcollegeeducationatthetimewhenindividualsmake
theircollegeenrollmentdecisions,anditservesasaproxyfor inequation(2).Itisalsobroken
downbycategoriesofpublic‐privateandmaincampus‐branch.
Thedescriptivestatisticsforthisvariableconfirmarapidexpansionofcollegeavailabilityin
Russiafollowingthe1992LawonEducation.Between1985and2003,thenumberofcampusesin
theRussianuniversitydatabasesurgedfrom812to2245,andthenumberofmunicipalitieswithat
least one campus increased from 198 to 442. Excluding Moscow city, the average number of
campusespermunicipality in theRLMSsamplemore thandoubledover the sampleperiod,with
mean=6.9andstd.dev.=13.9(seeFigure2C).Theupsurgeofnewcampuseswasalsoparticularly
strikingintheMoscowregionwheretheirnumberquadrupled.
Despiteseeminglylargechangesinthenumberofcampusesovertime,theshareofwithin‐
municipalityvariance isonlyabout13percentof thetotalvariance. Thevariancedecomposition
reported in Table 2 indicates that most variation in the total number of campuses is between
municipalities(87percent).However,thecompositionofvariancediffersconsiderablybythetype
ofcampus. For instance, the fact that thenumberofpublicuniversitiesbarelychangedsince the
Soviettimesisreflectedintheneartozerowithin‐municipalityvariance(lessthan1percentofthe
total variance); practically all the variation in this variable comes from the differences between
municipalities.Atthesametime,thewithin‐municipalityvariationovertimeissubstantialforthe
16
numberofbranches(bothpublicandprivate)andforthenumberofprivateuniversities(32to55
percent each). These features of the main variable will be utilized in the validity tests of our
instrument.
LocalLaborMarketConditions
Besides the number of campuses per municipality at age 17, we also use labor market
conditions in the location of residence at the time of college decisions as additional exclusion
restrictions.LocalearningsandunemploymentratearefrequentlyusedasinstrumentsintheMTE
literature (e.g., Cameron and Heckman 1998, Cameron and Taber 2004, Carneiro et al. 2011b,
Zamarro2010,Eisenhaueretal.2015,amongothers).15 Inourcase, regionalearningsat17and
regionalunemploymentrateat17capturetheopportunitycostofgoingtocollege.
Tocontrol for thepotentialcorrelationbetween localwage/unemploymentatage17and
current individualwage,we include (i) the contemporaneous (transitory) component of regional
variablesinthewageequationsand(ii)thepermanent(predicted)componentofregionalvariables
inboththeschoolingandwageequations(similartoCameronandHeckman1998;Carneiroetal.
2011b);seeTable1.
The transitory shocks to regional labor markets are proxied by the deviation of
contemporaneous regional earnings and unemployment rate from the trend‐adjusted regional
average. Specifically, regional transitoryearningsaredefinedas ̂ ina regressionof the logof
regionalmonthlyearningsinregionrandyeartonthesetofregionaldummies andaquadratic
timetrend,asinln . Theregionaltransitoryunemploymentrateisdefined
similarly.
Thepermanent regionalvariablesarepredicted fromtheregressionsof regionalmonthly
earningsandunemploymentrateonregiondummiesandaquadratic time trendover theperiod
15Local tuitionasanothercommonlyused IV isnotapplicable toourcasebecausecollege tuitionwasnotcharged during the Soviet periodand only 28 percent of all public students paid a fee for their educationduringoursampleperiodofcollegeexpansion,1993‐2003(EducationinRussia,2008).
17
1995‐2011.Essentially,thesearetrend‐adjustedaveragecharacteristicsofthelocallabormarket.
Asanalternativemeasureofpermanentearnings,wecomputeaverageregionalmonthlyearnings
overthefirst10yearsfollowingthecollegedecision.Toaccountforthepossibilitythatindividuals
makingacollegedecisionundertheSovietcentralizedwagestructuremaynothaveanticipatedthe
futuremarketearningsduetoalargestructuralbreakintheeconomy,wealsocomputepermanent
regional earnings separately for the Soviet period (1980‐1991) and the market period (1992‐
2011).
In essence, with the above structure of regional variables, we assume that individuals
making their college decision are not only influenced by the observed contemporaneous
characteristicsoftheir locationatage17,butalsohaveanimperfect foresightatage17of future
locallabormarketconditions.
In addition to the regional monthly earnings and unemployment rate at age 17, the
opportunity cost of going to college is captured by the country‐level skill wage ratio at age 17,
whichistheratioofaveragewagesofmanualworkerstoaveragewagesofnon‐manualworkersin
industry(seeAppendixTableA1fordatasources).ThroughouttheSovietperiod,manualworkers
were rewardedwithhighpay,whilewagesofnon‐manualworkerswereartificially compressed.
Figure2Dshowsthatthewageratioplummetedfrom0.94to0.55duringoursampleperiod(1985‐
2003),reflectingahistorictrendoftherisingwagepremiumforskilledworkers.
SummaryStatisticsfortheEstimationSample
Thebaselineestimationsample(age25,age17[1985,2003],wage[2004,2011],and
non‐missingvariables)amountsto17,911person‐yearobservations.Conditionalontheexogenous
sample constraintsbasedonage, surveyyear, andbirth cohortsand for agivendefinitionof the
college variable, wage is missing for about 20 percent of respondents either due to non‐
employment or non‐reporting. In robustness checks,we apply the inverse propensity score re‐
weightingtodealwiththeselectionintowages(i.e.,employmentorreportingofwages).Withthe
18
exception of mother’s education, the missing values in other covariates are trivial (about 0.6
percent).16 Table A1 in the appendix details how each variable is constructed. The descriptive
statistics inTable3 is reportedseparately forcollegegraduatesandsecondaryschoolgraduates.
Asexpected,collegegraduatesaremorelikelytobe female,resideinurbanareas,haveamother
withcollegedegree, live inmunicipalitieswithmorecampusesatage17,andearnahigherwage
rate.Interestingly,whencollegegraduateswere17,theyresidedinareaswithlowerearningsand
higher unemployment rates, but their regions of residence at the time of the survey had better
permanentlabormarketcharacteristics.
4.MarginalTreatmentEffect
BaselineEstimates
In thissection,wereport theMTEestimatesof themarginalbenefitsofcollegeeducation
anddiscuss thevalidityof exclusion restrictions. Webeginbyestimating thecomponentsof the
switching regression model given in Equations (2)‐(4). The random variables , , , , are
assumed tohave amultivariatenormaldistribution, and themodel is estimatedusingmaximum
likelihood. Table 4 presents the estimates of thewage equation for secondary school graduates
(column1), thewageequationforcollegegraduates(column2), thecollegeequation(column3),
andthemarginaleffectsofvariablesinthecollegeequation(column4).Theprobabilityofhavinga
college degree,P , is significantly higher among females, ethnically non‐Russians, urban
residents,andthosewhosemotherwenttocollege.Asexpected,individualsfromsmallerregional
cohortsatage17tendtohaveahigherrateofcollegeattainment,astheyfacelesscompetitionfora
givennumberof collegeopenings in their region. Morepeople go to college in the regionswith
higher permanent earnings, but college decisions do not appear to be responsive to regional
16Becausemother’seducationisunknownforthe12percentofourestimationsample,weincludeabinaryindicator formissing values of this variable to prevent sample loss. We drop observations withmissingvaluesinothercovariates.
19
unemployment rates and to regional earnings at age 17. Predictably, better salaries of manual
workers relative to non‐manual workers ‐ observed at age 17 ‐ deter people from pursuing a
collegedegree.
Theinstrumentsarejointlystrongpredictorsofcollegedecision(p‐value=0.000).Thekey
instrumentmeasuring the college availability at the timewhen an individualwas 17 is the total
number of college campuses per municipality. This variable serves as a cost shifter of college
decisions.Inordertocapturepotentialnon‐lineareffectsofcollegeavailabilityandtoaccountfor
the fact that some of the larger cities had a wide choice of colleges even before the college
expansion (e.g., Moscow had 100 colleges in 1990), we add a quadratic term on the number of
campuses. Because the distribution of the number of campuses is clumping at zero, we also
introduceabinaryvariabletoindicatemunicipalitieswithnocampuses(mean=0.386).Thelatter
variableisanalogoustoacommonlyusedbinaryIV,suchasthepresenceofacollegeinthecounty
of residence during teenage years (Card 1995; Cameron and Taber 2004). Results in Table 4
suggest that the opening of the first campus in a municipality without a college increases the
probabilityofreceivingacollegedegreeby0.114,i.e.,by11percentagepoints,onaverage,ceteris
paribus.Theestablishmentofadditionalcampusesalsoimprovesthecollegeattainment,thoughat
adiminishing ratepernewcampus. Thus, theeffectof collegeavailabilityoncollegeattainment
appearstobemuchstrongerfor“constrained”municipalitieswithfewercampuses.
Theestimatesofthewageequationsforthetwogroupsarestandard.17Theresultssuggest
that each grouphas a comparative advantage in the labormarket, that is, the individualswith a
higherpropensitytogotocollege(lowv) tendtodowell inthelabormarketoncetheygraduate
17Theresultsworthmentioningare:averylargegenderwagegapinthecontrolgroup(0.45logpointsor56percent difference favoring males) compared to 27 percent in the treatment group; mother’s educationaffectswagesofcollegegraduates,butnotwagesofsecondaryschoolgraduates;theindividualwagerateinbothgroups ispositively influencedbyhigher transitoryandpermanent localearnings;only in thecontrolgroup, the wage rate responds positively to local unemployment (both transitory and permanent); thecohortswithlargerregionalpopulationat17receiveahigherwagepremiumtoday.
20
( 0.150), but the same individuals would be worse off if they don’t go to college (
0.222).
FollowingCarneiroetal.(2011b)andHeckmanetal.(2010),weperformatestforselection
ongains(i.e.,whetherreturnstocollegearecorrelatedwithS). Asimpletestinvolvesestimating
equation(6)wherethelasttermisapproximatedwithapolynomialin ̂ (obtainedfromtheprobit
of college equation) and testing whether the coefficients on higher order polynomial terms are
jointlystatisticallysignificant.Theresultsofthistestsupportthehypothesisthatindividualsinour
sampleselectoncollegegains(seeAppendixTableA2).Thetestrejectsthatthereturnstocollege
arenotcorrelatedwithSorthat , isconstantin .
Another test for selection on gains relies on the estimated parameters from the normal
switching regression model (Table 4). The null hypothesis that the slope of the MTE is zero,
i.e., : 0,isrejectedatthe1percentlevel.Weestimatethat 0.372witha
standarderrorof0.065. This findingsupports theconclusionof “selectionongains” inTableA2
thatdoesnotimposethejointnormalityassumption.
UsingtheestimatedparametersinTable4,wecomputetheMTEaccordingtoEquation(8).
Figure3plots theestimatedMTE foragridofvaluesvbetweenzeroandone,18with90‐percent
confidencebands,evaluatedatmeanvaluesofX.Weobtainannualizedestimatesofthereturnsto
collegebydividingtheMTEby4.5,whichistheaveragedifferenceinyearsofschoolingbetween
thetreatmentandcontrolgroupsinoursample.ThenegativeslopeofMTEimpliesthatindividuals
withlowvaluesofv(thosewhoaremorelikelytogotocollege)havethelargestmarginalreturns
tooneyearofcollegeeducation.Conversely,individualswithhighvaluesofvhavelowMTE.The
heterogeneity in theMTE across the distribution of v is substantial: the returns vary from ‐18.2
percent for the highest vpersonwhowould lose from attending college to 32.9 percent for the
lowestvperson,withtheaveragereturnof7.3percentperoneyearofcollegeeducation.
18 In practice,we restrict the grid of values of to be between0.001 and 0.999,with 999 equally spacedpoints.
21
Given the vast differences in labor market institutions and data characteristics between
RussiaandtheU.S.,ourMTEestimatesforRussiaturnedouttobesurprisinglyclosetothefindings
ofCarneiroetal.(2011b).Inasampleof28to34‐yearoldwhitemalesfromNLSY,theyfindthat
thecollegereturnsintheU.S.varyfrom‐15.6percentto28.8percentperyearofcollegewiththe
meanof6.7percentandtheMTEslope 0.239.Atthesametime,ourMTEestimates
forRussiaareloweratthemeanandhaveamuchlargervariancethanthecorrespondingestimates
obtained by Heckman and Li (2004) in a cross‐sectional sample of Chineseworkers (theirMTE
rangesfrom5to15percentperyearofcollege,mean=10.8).
AlternativeInstruments
InTable5,ourmaininstrument–numberofcampusespermunicipality–isbrokendown
by categories ofmain campus‐branch (column1), public‐private19 (column2),Moscow vs. other
cities(column3).Overall,theseresultsarenotdifferentfromthebaselinespecification.Wefinda
negativeMTEslopeofsimilarmagnitude,averagereturnsof7to8percentperyearofcollege,and
aclearconcaverelationshipbetweencollegeavailabilityandcollegeattainment forall indicators.
Butthecollegeprobabilityfunctionisestimatedtobeconsiderablymoreconcavewithrespectto
the number of branches. This result could be partly explained by branches rather than main
campuses being opened in less‐populated areas where the first few local branches may have a
substantial impact on local college attainment (hence, large positive linear term), andwhere the
establishmentoffurtherbranchesmayalsoquicklyovercrowdalocalmarketforhighereducation
services(hence,largenegativequadraticterm).Lowertuitionfeesandlaxeradmissioncriteriain
branches could also contribute to a high response of the probability of going to college to the
openingofthefirstfewbranchesinlocality.
19Weuseadummyindicatorforprivatecampusesinsteadofthenumberofcampusesduetoconsiderableclusteringofobservationsatvaluesofzeroandone:only10percentofallmunicipalitiesintheRLMSsamplehadmorethanoneprivatecampus.
22
Anothernoteworthyfindingisthattheestablishmentofaprivatecampusinamunicipality
withexistingpublic campusesdoesnothavea statistically significanteffecton theprobabilityof
college attainment.20 This result could be due to the fact that private education institutions in
Russiaarecharginghigher tuition,areoftensmall insizeandtendtoopen in largercitieswhere
individualshaveotheroptionsofpursuingacollegedegree.InthethirdcolumnofTable5,wetry
toisolatetheeffectofcollege‐crowdedMoscowbytheinteractionofthenumberofcampusesper
municipalityat17withMoscowresidenceat17inalinearspecification.Theestimatessuggestthat
thecollegeattainmentinMoscowisnotaffectedbyafurtherincreaseinthenumberofcampuses,
buttheopeningofanadditionalcampusinothermunicipalitieswithatleastonecampusincreases
theprobabilityof receivingacollegedegreeby0.4percentagepoints. Allparametersof interest
remainakintothebaselinespecification.
Starting with the paper of Card (1995), the literature raised two major issues with the
validityofabinaryindicatorforcollegepresenceasaninstrument, includingnon‐randomcollege
construction and the Tiebout‐type geographic sorting of individuals in response to college
presence. These are legitimate concerns with regard to the number of campuses as well. The
validityoftheinstrumentisgoingtobecompromisedif:
′ inwageequationaffect incollegeequation;thatis,anticipatedwageshocksin2004‐2011
influencewherenewcampusesareopenedin1985‐2003;
and ′ are jointly determined; for example, individualswith, let say, positivewage draws
choosetoresideinthelocationofnewcampuses.
Toisolatethepotentialcorrelationbetweenfutureexpectedearningsandcampusopenings,
bothwageandschoolingequationscontrolforthepredictedregionalearningsandunemployment
rate in the place of residence at 17. Adding district fixed effects and a dummy for the type of
20 Interestingly, in a paper on the effect of mother’s education on birth outcomes in the U.S., Currie andMoretti (2003) also found that the opening of public colleges has a larger effect onmother’s educationalattainmentthantheopeningofprivatecolleges.
23
location serves the same purpose. These controls partly address the above concerns, including
non‐randomcollegeopeningsinresponsetohigherregionalincomeandthesortingofindividuals
with higher wage draws into the locations with better labor market conditions, which also
happenedtobeplaceswithmorecampuses.
ThevariancedecompositioninTable2givesususefulhintstocheckwhetherourbaseline
results are driven by the time‐series or cross‐sectional variation in the number of campuses. A
largewithin‐municipalityvariationinthenumberofcampusesovertimeprovidesmoreroomfor
the endogenousdecisions to open campuses in response to unobserved (by the econometrician)
futurewageshocks.Atthesametime,thecross‐sectional/geographicvariationthatiscreatedina
different economic system long before the college expansion started does not allow for such
strategic behavior. Table 2 shows that most of the variation in the number of campuses is
geographic rather than over time (87percent vs. 13percent). Furthermore,we re‐estimate our
baselinespecificationwiththenumberofpubliccollegesonly. Thisnumberremainedpractically
thesamesincetheSoviettimes;thewithin‐municipalityvarianceislessthan1percentofthetotal
variance (Table 2). We can reasonably assume that the college construction decision under the
centrally‐planned industrial structure is not correlated with wage innovations in the market
economy. The alternative specificationwith the number of public colleges does not change the
valueandtheslopeofthemarginaltreatmenteffect(column4ofTable5).
We also ran several placebo tests by adding future college openings in the college
equation.21Thegoalhereistocheckwhethercollegedecisionsareinfluencedbythefuturecollege
expansion. One of such tests is shown in column 5 of Table 5 and suggests that newly opened
campuses in the same municipality between the ages of 25 and 30 do not have any significant
impactonthelikelihoodofgoingtocollegeatage17.
21Forexample,wecontrolledforthenumberofcampusesestablishedinthesamemunicipalitybetweentheages of 30 and 35; thenwe tried adding new campuses opened between 2007 and 2011. None of thesevariableshaveanysignificantimpactonthelikelihoodofgoingtocollegeatage17.
24
Therefore, provided that the anticipated local labor market conditions are sufficiently
controlled for, the number of campuses at age 17 can be used as a sensible instrument for the
marginaltreatmenteffectestimationandsubsequentpolicysimulations.
AlternativeSpecifications
Beforecalculatingthetreatmentparameters,wecheckwhetherourresultsaresensitiveto
changes in specifications. Specifically, we focus on the sensitivity of the mean component of
MTE, 1 0 , and the covariance componentofMTE, ( ). Theestimatesof additional
specifications arepresented inTable6. In all of these specifications, the instruments are jointly
strongpredictorsofcollegedecisions,andthenullhypothesisthattheMTEslopeiszero, :
0, is rejected at the one percent level. In column 1, the model is estimated using an
alternative definition of the treatment group that includes college dropouts with three ormore
years of higher education. The rationale for changing the dependent variable is that college
dropoutswithsomeyearsofschoolingmightalsohavebeen“treated”bystudyingincollege.
Incolumn2,weuseanalternativedefinitionofregionalpermanentearningscalculatedas
averageregionalearningsoverthefirsttenyearssincecollegedecisionatage17.Specificationin
column 3 applies the same definition as in column 2 but replaces permanent earningswith the
Sovietperiodaverageregionalearnings(1980‐1991)forindividualswhoturned17beforemarket
reforms. Our motivation here is that teenagers raised under central planning may not have
foreknown future earnings in the market economy. As we introduce Soviet earnings, the
association between individual wages and regional permanent earnings becomes weaker but
remainsstatisticallysignificant.22Overall,modificationsinthefirstthreecolumnsdonotcauseany
significantdeviationoftheestimatedMTEfromabaselinespecification.
To check for potential bias due to non‐randommissingwages, specification in column 4
applies the inversepropensityweighting(IPW)toourbaselinespecification,wheretheweight is
22Resultsareshowninwebappendix.
25
the inverse of the predicted probability of having non‐missing wages. The propensity score is
constructed using the model covariates and , which are available for all respondents,
includingthosewithmissingwages.23TheMTEestimatesarenotaffectedbytheIPWcorrection.
Wealsoestimatethemodelwithoutassumingthejointnormalitybetweentheerrorsofthe
wage and college equations. Figure 4 shows the MTE using the semiparametric procedure we
describedinsection2.Overall,theshapeoftheMTEandpolicyparameterswediscussinthenext
sectionareconsistentwiththefullyparametricnormalmodel.
Next,we testwhether our results change ifwe usewagedata for different survey years.
Figure5plotstheaveragetreatmenteffectandtheMTEslopealongwiththe95‐percentconfidence
interval for different periods of the wage data; the point estimates from the 1995‐2011 sample
period is also shown in column5 of Table 6. Although the average returns to college appear to
decline over time, there is no statistically significant (at the 5 percent level) difference in the
averagereturnstocollegebetweendifferentsurveyperiods. TheMTEslopeisalsoconstantover
different periods. It seems that the general equilibrium effects of college expansion are either
cancelingeachotheroutortheyarenotsufficientlystrongtoshifttheequilibriumskillpricesand
theirdistribution.Byattractingmarginalstudentswithlowerreturns,collegeexpansionmayalter
theaggregatecompositionofcollegegraduatesandthusputadownwardpressureontheaverage
returnstocollege. However,thecompositioneffectinourestimatesisintertwinedwiththeprice
effect that partly could be demand‐driven. That is, skill‐biased demand shocks along with the
positiveproductivityspilloversfromtheincreasedstockofhumancapitalmaycompensateforthe
supply‐side effects keeping the equilibrium college wage premium constant over time.
Disentangling these varying general equilibrium effects of college expansion remains an area for
23 See Table 1 for the list of covariates. The MTE estimates are not sensitive if the IPW correction isperformedseparatelyformissingwagesamongtheemployedduetonon‐reporting(15percentofallmissingwages)andformissingwagesduetonon‐employment.TheprobitmodelsandMTEestimateswithIPWfordifferentsourcesofmissingdataareshowninwebappendix.
26
futureempiricalanalysis.Thepolicytreatmentparametersdiscussedinthenextsectionshouldbe
interpretedfromapartialequilibriumperspective.
5.PolicyTreatmentParameters
ConventionalTreatmentParameters
Recallthegrossreturnstocollegeeducationaregivenby∆ , ,
, , , withE ∆ | E | . Following Heckman and Li (2004), the probability
limitoftheOLSestimatorcanbewrittenas:
∆ E | , 1 E | , 0
E , , 1 E , , 0
E ∆ | E , 1 E , 0
E ∆ | E , , | 1 E , 1 E , 0 (ATE)(Sortinggain)(SelectionBias)
E ∆ | , 1 E , 1 E , 0 , (11)(TT)(SelectionBias)
whereATE=E ∆ | is the average treatment effect of college education for a randomly chosen
individual; TT=E ∆ | , 1 is the treatment effect on the treated; the sorting gain,E ,
, | 1 TT ATE,isthemeangainoftheunobservablesforpeoplewhoselectcollege,and
the selection bias E , 1 E , 0 ] = OLS ‐ TT is the mean difference in
unobservablesbetweensecondaryschoolgraduatesandcollegegraduatesifthelatterwouldnotgo
tocollege.
InTable7,wereporttheabovetreatmentparameters,whichareconstructedbyintegrating
MTEwiththeappropriateweightsdevelopedbyHeckmanandVytlacil(2005)(AppendixTableA3
and Figure A1). Standard errors are bootstrapped, and all parameters are annualized. For a
randomlychosenindividual,theaveragetreatmenteffectofoneyearofcollegeeducationisabout
7.3percent.Thetreatmenteffectonthetreatedisa13percentwageincreaseforcollegegraduates
27
comparedwithwhattheywouldreceivewithoutcollegedegree. Atthesametime,thetreatment
effectontheuntreated(TUT)isonly1.9percentwageincreaseforsecondaryschoolgraduatesif
theywould go to college. TheOLS estimate is 7.7 percent. We estimate the sorting gain of 5.7
percent and the selection bias of – 5.3 percent. Positive sorting gain,E , , | 1 0,
impliesthatindividualssortintocollegeonthebasisofunobservedgains.Negativeselectionbias
means that if college graduates did not complete college education, they would be worse off in
terms of the unobserved wage component in comparison with secondary school
graduates,E , 1 E , 0 . In other words, both college and secondary school
graduateshaveacomparativeadvantageinthelabormarket,whichisconsistentwiththeanalysis
ofWillis and Rosen’s (1979) seminal paper. An IV estimate of the returns to college (with the
propensityscore ̂ usedasIV)is16.1percent,anditispredictablyupwardbiasedcomparedtoATE
duetoheterogeneity( , , )andpositivesortinggain.AsevidentfromAppendixFigureA1,the
IVestimandweighsahighervaluedsegmentof , moreheavily.
PolicyEffects
Inadditiontoconventionaltreatmentparameters,wealsoestimatethetreatmenteffectfor
individualsatthemarginofindifferencebetweengoingandnotgoingtocollege( ). As
shown by Carneiro et al. (2010), the marginal treatment effect at the indifference point is
equivalent to the effect of the marginal policy change, which is college expansion in our case.
Following Carneiro etal. (2010),we assess themarginal returns to college for three alternative
policyregimes.Thefirstpolicyexogenouslyincreasestheprobabilityofgraduatingfromcollegeby
aninfinitesimalamount,sothat . Analternativepolicychangestheprobabilityof
collegebyatinyproportion(1+), 1 ∙ .Thethirdpolicymayinvolveasmallchange
in one of the continuous components ofZ. In ourZ vector,wehave a directmeasure of college
expansion, , , which is the number of campuses in the municipality of residencem in year t
28
(correspondingtoage17),suchthat , .24Usingequation(10)andweightsprovided
in Appendix Table A3 and Figure A1, we calculatemarginal policy‐relevant treatment effects of
threepolicyregimesfromthenormalselectionmodelandreporttheminTable8,PanelA.Similar
toCarneiroetal.(2010),wefindthattheMPRTEforamarginaladditivechangeinPisestimatedto
behigher(9.9percent)thantheMPRTEforamarginalproportionalchangeinP(7.4percent),but
theestimatesforRussiaaregreaterinmagnitudeinbothcases.Thethirdpolicyregime,whichis
more explicit and not yet reported in the literature, has returns of 9.6 percent per one year of
college for amarginal personwho is indifferent between going or not going to college andwho
wouldchangecollegeparticipationinresponsetoamarginalincreaseinthenumberofcampuses.
Alternatively,we reportMPRTEparameters from theMTEdistribution estimatedusing a
semi‐parametricmethodoflocalIV.TheestimationprocessisdescribedinSection2,andtheMTE
estimates,evaluatedatmeanvaluesofX,areplottedwith90‐percentconfidencebandsinFigure4.
We find that the semi‐parametricmethodproducesMTEwith the same shape as the parametric
one,butwithsomewhatlargerstandarderrors.Similartothenormalmodel,theMTEisdeclining
in , and we reject the null hypothesis that the returns to college are not correlated with S or
that , is constant in based on the test results reported in Appendix Table A2. The
common support of the P(Z) estimated from our sample ranges from aminimum of 0.070 to a
maximumof0.938.25PanelBofTable8presentstheMPRTEparametersfromthesemi‐parametric
model for the threealternativepolicy regimesdescribedabove.26 Inparticular, theMPRTE fora
marginalchangeinthenumberofcampusesisestimatedtobe10.7percentwageincreaseforone
yearofcollege.Theseestimatesareonlyslightlyhigher(byaboutonepercentagepoint)thanthe
24Inoursample,thenumberofcampusespermunicipalityvariesfrom0to299andistreatedasacontinuousvariable.Aninfinitesimalchangein , canbeinterpretedasanewclassroomoranadditionalstudentslotinalocality.25CommonsupportisdefinedastheintersectionofthesupportofP(Z)givenS=1andthesupportofP(Z)givenS=0.76observations,or0.4percentofoursample,falloutsidethecommonsupport.26AMPRTEparameterisaweightedaverageoftheMTEvaluesestimatedsemi‐parametricallyonthebasisofequation(12).TheMPRTEweightsarereportedinAppendixTableA3andalsoplottedinFigureA1.
29
MPRTEestimates fromthenormalmodel,soweadheretoourearlierconclusionswithregardto
themarginalreturnstocollege. Regardlessofthemethod,wefindthatamarginalstudentatthe
indifferencepointenjoysrelativelyhighreturnstocollegeinRussia,eventhoughhisreturnsare,as
expected,lowerthanthoseearnedbycurrentcollegegraduates(TT).
Next, we perform a few policy experiments for a discrete change in the number of
campuses. The experiments are evaluated using the policy‐relevant treatment effect (PRTE)
estimatorproposedbyHeckmanandVytlacil(2001)andgiveninequation(9).27Inourcase,PRTE
captures theaverageMTE for the individualswhochangedcollegeparticipation in response toa
fixed increase in thenumberof campuses from the1992pre‐expansion level, ,
∆ , .Inthefirstsimulation,weaddonecampusineachmunicipalitytothe1992levelin1993‐
2003andfindthePRTEreturnstobe9.8percentwageincreaseperyearofcollege.Inthesecond
setofsimulations,weaddonecampuspermunicipality indifferent locations. Results inTable8,
Panel C show that the returns to college vary depending on the place where campuses are
established. The returns are estimated to be larger for constrained municipalities, including
smallernon‐capitalcities,ruraldistricts,andmunicipalitiesthatdidnothaveinstitutionsofhigher
educationin1992(9.9percentwageincreaseinallthreecases).Atthesametime,theopeningofa
campus in localities with previously existing campuses or in the largest regional cities‐capitals
attractsindividualswithlowerreturnstocollege(7.7and7.5percent,respectively).
Inthefinalsetofsimulations,weaskwhatwouldthereturnstocollegebefortheaffected
individuals(thosewhoshiftedtotreatment)ifthenumberofcampusesincreasedbyonlyahalfof
the actual increase, ,∗
, , , ), or by 50 percent more of the actual increase,
,∗
, , , ).ComparedtoearliersimulationsinPanelCwhereweadduniformly
27ThePRTErequiresthattheempiricalsupportofthedistributionof tobewithintheunitinterval,andthattheempiricalsupportof hastobecontainedinthesupportof .Theseconditionsaresatisfiedinoursetting.Giventhemultivariatenormalstructureoftheerrors,thesupportof istheunitintervalbyconstruction. Additionally, the empirical support of is given by (0.044, 0.949), while the empiricalsupportof is(0.045,0.949)for=1.
30
onecampusinselectedmunicipalities,herewesimulatechangeswhicharelargeinmagnitudeand
varyinginsizedependingonthescaleofactualcollegeexpansionineachmunicipality.Onaverage,
fouradditionalcampusespermunicipalitywereestablishedbetween1992and2003.Predictably,
moresizeablecollegeexpansionsattract individualswithlowerreturns,thoughthescaleeffect is
notlinear.PRTEisestimatedtobe9.8percentforaone‐campusestablishmentpolicy,8percentfor
halfexpansion,6percentforfullexpansion,and6.2percentfor1.5ofactualexpansion.
Our PRTE estimates of the returns to college imply that the present value of additional
earningsstream foramarginalstudentovera33‐yearworking life (age22‐55) isequal toabout
$41,657 gains in non‐capital cities vs. $36,261 in capital cities (see details of calculations in
AppendixTableA4).Thedifferenceinnetbenefitsbetweenthetwotypesofcitiesisevengreater,
since students in capital cities, onaverage,paymore tuition and foregohigher earningswhile in
schoolcomparedtostudentsinsmallercities.Aroughcost‐benefitcalculationinTableA4suggests
that the net present benefits for amarginal student amount to $15,357 in non‐capital cities and
$7,334incapitalcitiesaftersubtractingthepresentvalueofforegoneearnings,averagetuitionand
othercollege‐relatedexpenses.Consideringthatrelocationtoalargecityimposesadditionalcosts
(e.g., transportation, greater living expenses, etc.), new students in constrained municipalities
clearlybenefitedfromcollegeopeningsintheplaceoftheirresidence.Thestockofhumancapital
intheremotelabormarketsisalsolikelytoincrease,ascollegegraduatestendtostayinthearea
wheretheyreceivetheireducation(Groen2004;Winters2011).
5.Conclusions
This paper estimates marginal returns to college education in Russia. Despite the vast
differences in labormarket institutionsanddatacharacteristicsbetweenRussiaand theU.S.,our
resultsareconsistentwiththepreviousU.S.literatureonMTEinthatwealsofind(i)alargedegree
ofheterogeneityinreturnstocollege,varyingfrom‐18to33percentincreaseinlifetimewagesfor
31
a year of college; (ii) a negatively‐sloped MTE showing greater marginal benefits from college
among individuals with the highest propensity to go to college; (iii) the positive sorting of
individuals intocollegebasedonobservedandunobservedmarketgainsassociatedwithcollege;
(iv)lowerreturnsforamarginalstudentthanforanaveragestudent(10vs.13percentperyearof
college);(v)a largerIVestimateofthereturnstocollegecomparedtoanOLSestimate;16and8
percent, respectively. Implicitly these results support the MTE approach and the need for the
precisecharacterizationofdifferenttreatmentparameters.
Themain focusof thisstudy isontheevaluationof the large‐scalecollegeexpansionthat
occurredinRussiaandresultedinmassopeningsofnewcolleges,bothpublicandprivate,andtheir
branches inmany localitieswhere collegeeducationwasnotpreviouslyoffered. Specifically,we
areinterestedinthereturnsofthemarginalindividualwhoswitchedintotreatmentasaresultof
college expansion. Unlike previousMTE studieswhere college expansion is characterized by an
exogenous hypothetical shift in the probability of attending college by some random amount or
proportion,weevaluateactual, real‐worldcollegeexpansionusing thenumberof campuses (and
theirtypes)atahighlydisaggregatedlevelofmunicipality.
We establish that individualswith lower returns enter colleges asmore campuses open.
However, for the marginal individuals who are affected by the establishment of an additional
campusatthetimeofmakingacollegedecision,theoverallgainsfromattendingcollegearelarge
andpositive;weestimatea10percentwagegainfortheseindividuals.Furthermore,wefindthat
the opening of a new campus in constrained municipalities – smaller non‐capital cities or
municipalitiesthatdidnothaveinstitutionsofhighereducationbeforecollegeexpansion–attracts
students with higher returns compared to the effect of the same policy in unconstrained
municipalities with at least one college existing before the expansion. Our policy estimate also
indicates that if the number of campuses permunicipality did not increase, then a considerable
32
shareofpopulationwithhighpotentiallabormarketgainsfromcollegewouldnothavebeenable
torealizethesegainsinthemarket.
Other resultshighlight the importantdistinctionbetweenpublic andprivate colleges and
between main campuses and branches and show that public campuses and local branches are
estimated tobemoreeffective in influencing local collegeparticipation. Wealso find thecollege
probabilityfunctiontobeconcavewithrespecttothenumberofcampuses,withaverylargekink
point for the firstcampusopened ina locality. Theeffectofeachadditionalcampuson the local
college attainment diminishes and eventually vanishes in college‐rich locations such asMoscow
city.
Overall, our findings show the direction for future policies targeted at expanding college
educationindevelopingcountriesandidentifyinglocationsforthefuturecollegeconstructionthat
wouldattractindividualswiththehighestpotentialgainsfromadditionaleducation.
References
Andrén, Daniela, John Earle, and Dana Sapatoru, 2005. The Wage Effects of Schooling underSocialismandinTransition:EvidencefromRomania,1950‐2000,JournalofComparativeEconomics33(2):300‐323.
Basu, Anirban, James J. Heckman, Salvador Navarro‐Lozano, and Sergio Urzua, 2007. Use ofInstrumental Variables in the Presence of Heterogeneity and Self‐selection: An Application toTreatmentsofBreastCancerPatients,HealthEconomics16(11):1133‐1157.
Belskaya,Olga, andKlara SabirianovaPeter, 2014.HowDoes theExpansionofHigherEducationChange theReturns toCollegeQuality? Insights fromSixtyYearsofRussianHistory,UNCChapelHillWorkingPaper.
Björklund,Anders,andRobertMoffitt,1987.TheEstimationofWageGainsandWelfareGains inSelf‐SelectionModels,ReviewofEconomicsandStatistics69(1):42‐49.
Brainerd,Elizabeth,1998.WinnersandLosersinRussia’sEconomicTransition,AmericanEconomicReview88(5):1094‐1116.
Cameron,Stephen,andJamesJ.Heckman,1998.LifeCycleSchoolingandDynamicSelectionBias:ModelsandEvidenceforFiveCohortsofAmericanMales,JournalofPoliticalEconomy106(2):262‐333.
33
Cameron,Stephen,andChristopherTaber,2004.EstimationofEducationalBorrowingConstraintsUsingReturnstoSchooling,JournalofPoliticalEconomy112(1):132‐182.
Card, David, 1995. Using Geographic Variation in College Proximity to Estimate the Return toSchooling,inAspectsofLabourMarketBehaviour:EssaysinHonourofJohnVanderkamp,ed.byLouisN.Christofides,E.KennethGrant,andRobertSwidinsky.Toronto:UniversityofTorontoPress,201‐222.
Card,David,2001.Estimating theReturn toSchooling:ProgressonSomePersistentEconometricProblems,Econometrica69(5):1127‐1160.
Carneiro,Pedro,andSokbaeLee,2009.EstimatingDistributionsofPotentialOutcomesUsingLocalInstrumentalVariableswithanApplicationtoChangesinCollegeEnrollmentandWageInequality,JournalofEconometrics149(2):191‐208.
Carneiro,Pedro, JamesHeckman,andEdwardVytlacil, 2010.EvaluatingMarginalPolicyChangesandtheAverageEffectofTreatmentforIndividualsattheMargin,Econometrica78(1):377‐394.
Carneiro, Pedro, James Heckman, and Edward Vytlacil, 2011b. Estimating Marginal Returns toEducation,AmericanEconomicReview101(4):2754‐2781.
Carneiro,Pedro,MichaelLokshin,CristobalRidao‐Cano,andNithinUmapathi,2011a.AverageandMarginalReturnstoUpperSecondarySchoolinginIndonesia,PolicyResearchWorkingPaper5878.
Carnoy, Martin, Prashant Loyalka, Maria Dobryakova, Rafiq Dossani, Isak Froumin, KatherineKuhns, Jandhyala B. G. Tilak, and Rong Wang, 2013. University Expansion in a Changing GlobalEconomy:TriumphoftheBRICs?Stanford,California:StanfordUniversityPress.
Currie,Janet,andEnricoMoretti,2003.Mother'sEducationandtheIntergenerationalTransmissionofHumanCapital:Evidence fromCollegeOpenings,Quarterly JournalofEconomics118(4):1495‐1532.
DemographyYearbookofRussia,2002.Moscow:Goskomstat.
Duflo,Esther,2001.SchoolingandLaborMarketConsequencesofSchoolConstructioninIndonesia:EvidencefromanUnusualPolicyExperiment,AmericanEconomicReview91(4):795‐813.
EducationinRussia,2008.Moscow:Goskomstat
Eisenhauer,Philipp,JamesJ.Heckman,andEdwardVytlacil,2015.TheGeneralizedRoyModelandtheCost‐BenefitAnalysisofSocialPrograms,JournalofPoliticalEconomy,132(2),forthcoming.
Evans, Heather, and Anirban Basu, 2011. Exploring Comparative Effect Heterogeneity withInstrumentalVariables:PrehospitalIntubationandMortality,Health,EconometricsandDataGroupWorkingPapers11/26.
Fang,Hai,KarenEggleston,JohnRizzo,ScottRozelle,andRichardZeckhauser,2012.TheReturnstoEducation in China: Evidence from the 1986 Compulsory Education Law, Working Paper, No.18189,NationalBureauofEconomicResearch,Cambridge,MA.
FederalLaw,1992.Obobrazovanii[OnEducation],July10,1992.
34
Availableathttp://www.consultant.ru/popular/edu/.
Fleisher,Belton,KlaraSabirianovaPeter,andXiaojunWang,2005.ReturnstoSkillsandtheSpeedofReforms:EvidencefromCentralandEasternEurope,China,andRussia, JournalofComparativeEconomics33(2):351‐370.
Frenette, Marc, 2009. Do Universities Benefit Local Youth? Evidence from the Creation of NewUniversities,EconomicsofEducationReview28(3):318‐328.
Garen,John,1984.TheReturnstoSchooling:ASelectivityBiasApproachwithaContinuousChoiceVariable,Econometrica52(5):1199‐1218.
Gorodnichenko, Yuriy, and Klara Sabirianova Peter, 2005. Returns to Schooling in Russia andUkraine: A Semiparametric Approach to Cross‐Country Comparative Analysis, Journal ofComparativeEconomics33(2):324‐350.
Groen, Jeffrey A., 2004. The Effect of College Location on Migration of College‐Educated Labor,JournalofEconometrics121(1‐2):125‐142.
Heckman, James J., JohnEricHumphries,GregoryVeramendi, andSergioUrzua,2014.Education,HealthandWages,WorkingPaperNo.19971,NationalBureauofEconomicResearch,Cambridge,MA.
Heckman,JamesandEdwardVytlacil,2001.Policy‐RelevantTreatmentEffects,AmericanEconomicReview91(2):107‐111.
Heckman, James and Edward Vytlacil, 2005. Structural Equations, Treatment Effects, andEconometricPolicyEvaluation,Econometrica73(3):669–738.
Heckman,JamesandEdwardVytlacil,2007.EconometricEvaluationofSocialPrograms,HandbookofEconometrics,Volume6B:4877‐5143.
Heckman,JamesandXuesongLi,2004.SelectionBias,ComparativeAdvantageandHeterogeneousReturnstoEducation:EvidencefromChinain2000,PacificEconomicReview9(3):155‐171.
Heckman, James, Daniel Schmierer, and Sergio Urzua, 2010. Testing the Correlated RandomCoefficientModel,JournalofEconometrics158(2):177‐203.
Heckman, James, Lance Lochner, andPetraTodd, 2006. Earnings Functions, Rates ofReturn andTreatment Effects: The Mincer Equation and Beyond, Handbook of the Economics of Education,Volume1,Chapter7:307‐458.
Ichino,AndreaandRudolfWinter‐Ebmer,2004.TheLong‐RunEducationalCostofWorldWar II,JournalofPoliticalEconomy22(1):57‐87.Kane,ThomasandCeciliaRouse,1995.Labor ‐MarketReturnstoTwo‐andFour‐YearCollege,AmericanEconomicReview85(3):600‐614.
Kane, Thomas J. and Cecilia Elena Rouse, 1995. Labor‐Market Returns to Two‐ and Four‐YearCollege,AmericanEconomicReview85(3):600‐614.
Katz, Lawrence and Kevin Murphy, 1992. Changes in Relative Wages, 1963‐1987: Supply andDemandFactors,QuarterlyJournalofEconomics107(1):35‐78.
35
Kyui, Natalia, 2013. Expansion of Higher Education, Employment andWages: Evidence from theRussianTransition,BankofCanadaWorkingPaper2013‐45.
Moffitt, Robert, 2008. Estimating Marginal Treatment Effects in Heterogeneous Populations,Annalesd’EconomieetdeStatistique91/92:239‐261.
Moffitt, Robert (2014). Estimating Marginal Treatment Effects of Transfer Programs on LaborSupply,WorkingPaper.
Münich, Daniel, Jan Svejnar, and Katherine Terrell, 2005. Returns to Human Capital Under theCommunismWageGridandduringtheTransitiontoaMarketEconomy,ReviewofEconomicsandStatistics83(1):100‐123.
OECD,2008.EducationataGlance:OECDIndicators.OECDPublishing.
Oppedisano,Veruska,2011.The(Adverse)EffectsofExpandingHigherEducation:Evidence fromItaly,EconomicsofEducationReview30(5),997‐1008.
RegionsofRussia.Moscow:Goskomstat,variousyears.
Robinson,Peter,1988.Root‐N‐ConsistentSemiparametricRegression,Econometrica56(4):931–54.
Roy,Andrew,1951.SomeThoughtsontheDistributionofEarnings,OxfordEconomicPapers3(2):135‐146.
Topel, Robert, 1997. ‘Factor Proportions and Relative Wages: The Supply‐Side Determinants ofWageInequality,JournalofEconomicPerspectives11(2):55–74.
Willis,RobertandSherwinRosen,1979.EducationandSelf‐Selection,JournalofPoliticalEconomy87(5):S7‐S36.
Winters, John V., 2011. Why Are Smart Cities Growing?Who Moves andWho Stays? Journal ofRegionalScience51(2):253‐270.
Yang, Dennis Tao, 2005. Determinants of Schooling Returns during Transition: Evidence fromChineseCities.JournalofComparativeEconomics,33(2):244‐264.
Zamarro, Gema, 2010. Accounting for Heterogeneous Returns in Sequential Schooling Decisions,JournalofEconometrics156(2):260‐276.
36
Table1:VariablesinWageandCollegeEquations
Variable TimeframeLevelof
aggregationCommon
Wageeq
CollegeeqEandI
Female Fixed individual x Urbanresidence Fixed individual x EthnicallyRussian Fixed individual x Mother’seducation Fixed individual x Age,agesquared Currentyear individual xBirthcohorts Fixed individual xFederaldistricts Currentyear individual xFederaldistrictsat17 Age17 individual xSurveyyeardummies Currentyear individual xRegionalcohortsizeat17 Age17 region x Permanentregionalearnings Fixed region x Permanentunemploymentrate Fixed region x Transitoryregionalearnings Currentyear region xTransitoryunemploymentrate Currentyear region xRegionalearningsat17 Age17 region xUnemploymentrateat17 Age17 region xSkillwageratioat17 Age 17 country xNumberofcampusesat17 Age17 municipality x
Notes: The vector includes wage determinants; E is the number of campuses in the municipality ofresidenceatage17;theIvectorincludesotherexclusionrestrictionsinthecollegedecisionequation;andthecommonvector belongstobothequations.
Table2:VarianceDecompositionfortheNumberofCampuses
Variables MeanTotal
variance
Within‐municipality
%
Between‐municipality
%Totalnumberofcampuses 1.43 77.3 86.8 13.3Numberofcolleges‐public 0.77 22.3 99.4 0.6Numberofcolleges‐private 0.27 17.5 59.6 40.4Numberofbranches‐public 0.33 1.0 68.3 31.8Numberofbranches‐private 0.07 0.2 44.9 55.1
Notes: Thenumber of campuses, colleges, and branches is calculated for everymunicipality in theRLMSestimation sample. The panel ofmunicipalities used in the variance decomposition is balancedwith 733municipalitiesand19timeperiodsfrom1985to2003.Acampusreferstoallbuildingsofthesamecollegeinonemunicipality.Branchisacampuslocatedoutsidethemunicipalityofthemaincampus.Totalnumberofcampusesisthesumofthenumberofmaincampuses(whichisequivalenttothenumberofcolleges)andthenumberofbranches.
37
Table3:SampleStatistics
S=0(N=10,962)
S=1(N=6,949)
Meancomparison
t‐test(p‐value)Hourlywagerate(log) 3.831 4.286 0.000 (0.751) (0.721) Age 32.855 31.919 0.000 (4.808) (4.708) Regionaltransitoryearnings(log) 0.020 0.013 0.001 (0.139) (0.146) Regionaltransitoryunemploymentrate,% 0.111 0.239 0.000 (1.687) (1.641) Female 0.492 0.606 0.000 (0.500) (0.489) Urbanresidence(binary) 0.754 0.873 0.000 (0.431) (0.333) EthnicallyRussian(binary) 0.907 0.911 0.319
(0.291) (0.284) Mother’seducation(binary) 0.067 0.302 0.000 (0.249) (0.459) Mother’seducationmissing(binary) 0.117 0.115 0.648 (0.322) (0.319) Regionalcohortsizeat17(log) 10.456 10.551 0.000 (0.610) (0.674) Regionalpermanentearnings(log) 8.519 8.626 0.000
(0.387) (0.396) Regionalpermanentunemploymentrate,% 6.841 6.201 0.000
(2.615) (2.886) Nofcampusespermunicipality 11.879 26.432 0.000 (36.070) (57.131) Municipalitywithnocampuses(binary) 0.465 0.260 0.000 (0.499) (0.439) Skillwageratioat17 0.760 0.725 0.000
(0.128) (0.126) Regionalearningsat17(log) 7.875 7.844 0.000
(0.463) (0.460) Regionalunemploymentrateat17,% 4.817 5.808 0.000
(5.209) (5.164)Notes:Descriptive statistics are provided for the baseline estimation sample (age 25, age 17 [1985,2003],wage [2004,2011],andnon‐missingvariables). Thet‐testcomparesmeansofvariablesbetweencollegegraduates(S=1)andgraduatesofsecondaryschools(S=0).ThedefinitionofallvariablesisdescribedinAppendixTableA1.Standarddeviationsareinparentheses.
38
Table4:MaximumLikelihoodEstimatesoftheNormalSwitchingRegressionModel
VariablesWageequations CollegeequationS=0 S=1 Coefficients ME
Age 0.013 0.055*** … … (0.015) (0.019) Agesquared ‐0.001 ‐0.060** … … (0.023) (0.028) Regionaltransitoryearnings(log) 0.733*** 0.780*** … … (0.095) (0.119) Regionaltransitoryunemploymentrate,% 0.010** ‐0.001 … … (0.005) (0.006) Female ‐0.465*** ‐0.239*** 0.374*** 0.142*** (0.013) (0.017) (0.020) (0.008)Urbanresidence(binary) 0.212*** 0.308*** 0.183*** 0.069*** (0.015) (0.024) (0.031) (0.011)EthnicallyRussian(binary) ‐0.040** ‐0.015 ‐0.065* ‐0.027* (0.019) (0.024) (0.038) (0.015)Mother’seducation(binary) ‐0.041 0.167*** 1.053*** 0.403*** (0.034) (0.036) (0.029) (0.010)Mother’seducationmissing(binary) ‐0.021 0.110*** 0.156*** 0.059*** (0.017) (0.023) (0.031) (0.012)Regionalcohortsizeat17(log) 0.191*** 0.064*** ‐0.181*** ‐0.066*** (0.017) (0.018) (0.026) (0.010)Regionalpermanentearnings(log) 0.618*** 0.796*** 0.281*** 0.112*** (0.040) (0.051) (0.051) (0.019)Regionalpermanentunemploymentrate,% 0.020*** 0.003 ‐0.000 0.002 (0.005) (0.006) (0.009) (0.003)
Instruments
Nofcampusespermunicipality … … 0.011*** 0.005*** (0.001) (0.000)Nofcampusespermunicipalitysquared/100 … … ‐0.003*** ‐0.001*** (0.000) (0.000)Municipalitywithnocampuses(binary) … … ‐0.312*** ‐0.110*** (0.026) (0.010)Skillwageratioat17 … … ‐1.313*** ‐0.527*** (0.368) (0.145)Regionalearningsat17(log) … … 0.040 0.016 (0.043) (0.017)Regionalunemploymentrateat17,% … … ‐0.007 ‐0.003 (0.006) (0.002)
‐0.372***(0.065) 0.330***(0.057);annualized=7.3
2–testforjointsignificanceofinstruments=360.9*** 2–testforindependenceofequations=40.6*** Notes: ***Significantat1%.**Significantat5%.*Significantat10%. N=17,911(age25,age17 [1985,2003],wage[2004,2011],andnon‐missingvariables).Thistableshowsthemaximumlikelihoodestimatesofwage and college equations for the normal switching regression model. Robust standard errors are inparentheses.Thestandarderrorof( iscomputedusingtheDeltamethod.Thecollegeequationalsoincludes dummies for four cohorts,Moscow residence at age 17, and seven federal districts at age 17;wageequationsincludedummiesforsurveyyears,currentMoscowresidence,andsevenfederaldistrictsatthetimeofthesurvey.
39
Table5:CollegeEquation:AlternativeSetofInstruments
Variables (1) (2) (3) (4) (5)Nofcolleges(maincampuses) 0.007*** … … … … (0.002) Nofcollegesquared/100 ‐0.002*** … … … … (0.000) Nofbranches 0.065*** … … … … (0.013) Nofbranchessquared/100 ‐0.287*** … … … … (0.100) Municipalitywithnocampuses(dummy) ‐0.258*** ‐0.250*** ‐0.324*** ‐0.276*** ‐0.312*** (0.030) (0.029) (0.026) (0.027) (0.026)Numberofpubliccampuses … 0.024*** … … … (0.002)Nofpubliccampusessquared/100 … ‐0.018*** … … … (0.003) Municipalitywithaprivatecampus(dummy) … ‐0.008 … … … (0.031)Nofcampuses … … 0.009*** … 0.011***
(0.001) (0.001)xMoscowresidencyat17 … … ‐0.009*** … …
(0.001) Nofpubliccolleges … … … 0.025*** …
(0.002)Nofpubliccollegessquared/100 … … … ‐0.020*** …
(0.003)Nofcampusessquared/100 … … … … ‐0.003***
(0.000)ChangeinNofcampusesbetweenages25‐30 … … … … ‐0.001
(0.001) ‐0.353*** ‐0.352*** ‐0.387*** ‐0.373*** ‐0.370***
(0.066) (0.066) (0.063) (0.064) (0.066) 0.374*** 0.369*** 0.318*** 0.347*** 0.338***
(0.061) (0.060) (0.054) (0.058) (0.059)2–testforjointsignificanceofinstruments 388.4 412.2 352.3 394.4 362.1p‐value [0.000] [0.000] [0.000] [0.000] [0.000]2–testforindependenceofequations 49.6 49.2 44.0 50.3 41.6p‐value [0.000] [0.000] [0.000] [0.000] [0.000]Log‐likelihood ‐24,466 ‐24,451 ‐24,480 ‐24,460 ‐24,476Notes: N=17,911.***Significantat1%.**Significantat5%.*Significantat10%.N=17,911.TableshowsthemaximumlikelihoodestimatesofthecoefficientsoftheNofcampusespermunicipalityinthecollegeequationusing alternative definitions. The number of campuses, colleges, and branches is computed at the level ofmunicipality.Allspecificationsusethesamesetofvariablesandthesamesampleconstraintsasinthebaselinespecification reported in Table 4 (age 25, age 17 [1985, 2003], wage [2004, 2011], and non‐missingvariables).Thestandarderrorof( iscomputedusingtheDeltamethod.Robuststandarderrorsareinparentheses.
40
Table6:SwitchingRegressionModelParametersfromAlternativeSpecifications
Includesdropouts
10‐yearearningsaverage
Sovietearnings IPWWage1995‐
2011
(1) (2) (3) (4) (5) ‐0.363 ‐0.353 ‐0.365 ‐0.369 ‐0.355
(0.064) (0.070) (0.064) (0.053) (0.061) 0.345 0.367 0.355 0.397 0.434
(0.054) (0.077) (0.064) (0.047) (0.057)2–testforjointsignificance 370.6 164.0 170.1 517.3 237.8ofinstruments[p‐value] [0.000] [0.000] [0.000] [0.000] [0.000]
2–testforindependenceof 39.4 41.2 40.6 32.1 81.5equations[p‐value] [0.003] [0.003] [0.001] 0.000 [0.000]
Log‐likelihood ‐26,100 ‐24,680 ‐24,750 30,589 ‐31,592Numberofobservations 18,823 17,911 17,911 17,911 22,584Notes:Tablereportsstatisticsforalternativespecificationsofthenormalswitchingregressionmodel.Unlessnotedotherwise, all specificationsuse thesamesetofvariablesand thesamesampleconstraintsas in thebaselinespecificationreportedinTable4(age25,age17[1985,2003],wage[2004,2011],andnon‐missingvariables).Incolumn1,thetreatmentgroup,S=1,includescollegedropoutswiththreeormoreyearsofcollege.Incolumn2,weuseanalternativedefinitionofregionalpermanentearningscalculatedasaverageregional earnings over the first ten years since college decision at age 17. The specification in column 3appliesthesamedefinitionasincolumn2butreplacespermanentearningswiththeSovietperiodaverageregionalearnings(1980‐1991)forindividualswhoturned17beforemarketreforms.Incolumn4,weapplythe inverse propensityweight from the probitmodel of non‐missingwages on covariates and . Incolumn 5,wage [1995, 2011]. The standard error of ( is computed using the Deltamethod.Robuststandarderrorsareinparentheses,andp‐valuesareinbrackets.
41
Table7:TreatmentParametersTreatmentParameter Estimatedreturn
AverageTreatmentEffect(ATE) 0.073 (0.026)TreatmentontheTreated(TT) 0.130 (0.041)TreatmentontheUntreated(TUT) 0.019 (0.020)InstrumentalVariables(IV) 0.161 (0.051)OrdinaryLeastSquares(OLS) 0.077 (0.024)
Sortinggain(TT‐ATE) 0.057Selectionbias(OLS‐TT) ‐0.053Notes: The standard treatment parameters are obtained by integrating the , using weightingfunctionsofHeckmanandVytlacil(2005):
, , ,where , , , , .The weighting functions , are defined in Appendix Table A3. Linear IV estimates use P(Z) as aninstrument. Bootstrappedstandarderrorsareinparentheses(250replications). Thereportedreturnsareannualizedbydividing theestimatesby4.5,which is thedifference inaverageyearsof schoolingbetweentreatmentandcontrolgroups.Tablereportstheresultsunderthebaselinespecification(Table4).
42
Table8:PolicyParameters
PolicyExperiment PolicyParameter
PanelA:MPRTEfromthenormalselectionmodel(forinfinitesimal) 0.099
(0.011)1 ∙ 0.074
(0.013), 0.096
(0.011)
PanelB:MPRTEfromthesemi‐parametricmodel(forinfinitesimal) 0.110
(0.012)1 ∙ 0.086
(0.014), 0.107
(0.012)
PanelC:PRTEfrompolicysimulationsAddingacampusineachmunicipality 0.098
,∗
, 1 (0.011)Openingafirstcampusin1993‐2003 0.099
,∗ 1if , 0, ∈ 1993,2003 (0.011)
Openinganadditionalcampusin1993‐2003 0.077,
∗, 1if , 0, ∈ 1993,2003 (0.012)
Addingacampusincapitalcitiesin1993‐2003 0.075,
∗, 1, ∈ 1993,2003 (0.012)
Addingacampusinnon‐capitalcitiesin1993‐2003 0.099,
∗, 1, ∈ 1993,2003 (0.010)
Addingacampusinruraldistrictsin1993‐2003 0.099,
∗, 1, ∈ 1993,2003 (0.011)
50%ofactualexpansion 0.080 ,
∗, , , ) (0.012)
Actualexpansion( , , ) 0.060 ,
∗, (0.013)
150%ofactualexpansion 0.062 ,
∗, , , ) (0.013)
Notes:Tablereportsmarginalpolicy‐relevanttreatmenteffect(MPRTE)forthenormalselectionandsemi‐parametricmodelsaswellaspolicy‐relevanttreatmenteffect(PRTE)frompolicyexperiments.Thereportedreturnsareperyearofcollege. , isthenumberofcampusesinthemunicipalityofresidencematage17inyeart; ,
∗ isthesimulatednumberofcampuses.TheMTEestimatesarebasedonequation(8)inPanelsAandC;equation(12)inPanelB.TheMPRTEareobtainedbyintegratingthe , usingtheweightingfunctions reported inAppendixTableA3. The reportedPRTE estimates are computedusingQuasi‐MonteCarlosimulationwithHaltonsequences;theseestimatesareessentiallyidenticaltothePRTEestimatesbasedonthePRTEweightingfunction.Bootstrappedstandarderrorsareinparentheses(250replications).
43
Figure1:AverageReturnstoCollegeEducationinRussia
Notes:ReturnstocollegeeducationarecalculatedfromtheOLSregressionofthelogofmonthlyearningsatprimary job on college degree, gender, age, age squared, urban place of birth, a dummy for Russiannationality,andsevenfederaldistricts.Thecomparisongroupincludesgraduatesofgeneralandprofessionalsecondaryschools,butexcludescollegedrop‐outs.Earningsfor1985and1990arereportedretrospectively.Estimationisperformedforeachyearseparatelyusingthesampleof25‐to55‐yearolds.Reportedaretheestimatedcoefficientsoncollegedegreeandtheoveralltrendfittedusingnon‐parametricsmoothing(lowess;bandwidth=0.4). Returns are in percent andper yearof college (dividedby4.5). Also shown is the90%confidenceintervalcomputedusingrobuststandarderrors.Theverticallinedemarks1991asthebreakupoftheUSSR.
USSR period
0
1
2
3
4
5
6
7
8
9
10
1985 1990 1995 2000 2005 2010
44
Figure2:TrendsinKeyVariables
Notes:PanelAshowsthepercentshareofcollegegraduatesinRLMSamong25‐44‐yearoldindividuals.Thecomparison group includes graduates of general and professional secondary schools, but excludes collegedrop‐outs.PanelBdepictstheaveragesizeof17‐year‐oldpopulationinthousandspeopleacrossallRussianregions. Vertical linesdefine the sampleperiod– from1985 to2003. Also shown is the90%confidenceinterval. Panel C shows the average number of campuses permunicipality at age 17 in the RLMS sample,excludingMoscowcity. PanelDdisplaystheratioofwagesofmanualworkerstothewagesofnon‐manualworkersintheindustrialsectorduringthesampleperiod1985‐2003.
2025
3035
40
1995 1998 2001 2004 2007 2010
Age 25-44A. Share of college graduates, %
Sample period
1020
3040
1950 1960 1970 1980 1990 2000 2010
per region, 000sB. Average population age 17
03
69
1215
1985 1988 1991 1994 1997 2000 2003
per municipalityC. Average N of campuses
.5.6
.7.8
.91
1985 1988 1991 1994 1997 2000 2003
D. Manual/non-manual wage ratio
45
Figure3:MarginalTreatmentEffectEstimatedfromaNormalSelectionModel
Notes:The figureplots theMTEestimates fora gridof valuesv betweenzeroandone,with a90‐percentconfidenceinterval,evaluatedatmeanvaluesofX.Weestimateaparametricnormalselectionmodelgiveninequations (1)‐(4) by maximum likelihood. The figure is computed using equation (8) and the estimatedparametersinTable4.Thefixedcovariancecomponentof , isgivenby 0.372withastandarderrorof0.065. Thereportedreturnsareperyearofcollege. ThestandarderrorsarecomputedusingtheDeltamethod.
-0.3
-0.2
-0.1
0.0
0.1
0.2
0.3
MT
E(x
,v)
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0v
CI 90% MTE(x, v)
46
Figure4:MarginalTreatmentEffectEstimatedfromaSemi‐ParametricModelwithLocalIV
Notes:The figureplots theMTEestimates fora gridof valuesv betweenzeroandone,with a90‐percentconfidence interval, evaluated at mean values of X. Estimation steps are described in Section 2. Theestimation is performed using a Robinson's (1988) double residual estimator to obtain the non‐linearrelationbetween the log ofwages and ,where ̂ is thepredictedprobability of graduating from college.Then, we use a Kernel quadratic local polynomial regression to evaluate the derivatives of equation(11), ′ ̂ ,withabandwidthof0.32. The reported returnsareperyearof college. 90percent standarderrorbandsarecalculatedusingthebootstrap(250replications).
-0.3
-0.2
-0.1
0.0
0.1
0.2
0.3
MT
E(x
,v)
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0v
CI 90% MTE_semi(x, v)
47
Figure5:AverageTreatmentEffectEstimatesforDifferentSamplePeriods
Notes: ; 1 0 .Figurereportstheaveragereturnstocollege(ATE)andthe MTE slope along with the 95‐percent confidence interval under the baseline specification byadding/subtractingoneadditionalsurveyyeartopreservethepanelfeaturesofwageseries.Thelastyearofeach survey period is always 2011 to keep the college decision period (1985‐2003) constant forcomparabilitypurposes.Yearindicatesthebeginningofthesurveyperiod.Theredlinecorrespondstothebaselineestimatesforthe2004‐2011surveyperiod(Table4).
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
1995 1996 1998 2000 2001 2002 2003 2004 2005 2006Year
CI 95% ATE
A. ATE
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
1995 1996 1998 2000 2001 2002 2003 2004 2005 2006Year
CI 95% MTE slope
B. MTE slope
48
AppendixTableA1:DescriptionofVariables
Generalnotes:1. Thesourceforallindividual‐levelvariablesisRLMS2. Estimationsamplecovers1995‐1996,1998,2000‐2011timeperiods;variablesareavailableforallyears,unlessnoted
otherwise3. Thesampleisrestrictedtoindividualsbornbetween1968and1986andolderthan25atthetimeofthesurvey.4. Russianregionsincludetwofederalcities(MoscowcityandSt.Petersburg)and81territories,whichareaggregatedinto
sevenfederaldistricts.5. Duetomultiplechangesintheadministrative‐territorialstructureofRussia,allpastregionaldataarecollectedbasedon
theclassificationofregionsaccordingtothe2008amendmenttotheConstitutionoftheRussianFederation.6. Atthetimeofthesurvey,respondentsresidedin32regions,buttheygraduatedfromuniversitieslocatedin73regions7. MunicipalityinRLMSisdefinedaccordingtothe5‐digitmunicipalitycodetakenfromtheNationalClassificationof
AdministrativeandTerritorialDivision(OKATOinRussian).OKATOcodeconsistsoftwodigitsforaregionandthreedigitsforeithercityordistrict(county‐equivalent)withinaregion.
8. Municipalityatage17isthesameascurrentmunicipalityforindividualswhomovedtocurrentmunicipalitybefore17,finishedsecondaryschoolincurrentlocation,orcurrentlyresideintheplaceoftheirbirth.Municipalityatage17isthesameascollegelocationifindividualsresidedinthesameplacebeforecollege.Municipalityatage17isimputedbyrandomlyselectingamunicipalityforagiventypeofbirthplace(city,township,andvillage)withintheregionofcurrentresidenceforindividualswhodidnotattendcollegeandmovedtoanewlocationafterage17.Municipalityatage17isimputedbyrandomlyselectingamunicipalityforagiventypeofbirthplace(city,township,andvillage)withintheregionofcollegeforindividualswhoattendedcollegeandmovedtoanewlocationafterage17.
DependentVariables
Hourlywagerate(log) =Laborearningspermonthatprimaryjob/Hoursofworkpermonthatprimaryjob
Laborearningspermontharedefinedas:
monthlyaverage(overthelast12months)after‐taxlaborearningsforanemployee[1998‐2011];
totalaccumulatedwagedebtdividedbythenumberofmonthsofoverduewagesforemployeeswithwagearrears[1995‐1996];
monetary portion of lastmonth earnings for employees with nowagearrears[1995‐1996];
monetary portion of lastmonth earnings for self‐employed; self‐employed include individuals reportingplace ofworkother thanan organization as well as those involved in regular individualeconomicactivities[allyears].
Hoursofworkpermontharedefinedasfollows:
Usualhoursofworkinatypicalweektimesfour[1998‐2011]; Actualhoursofworklastmonth[1995‐1996]; Unusually highhours are top codedat 480hourspermonth (16
hoursperdaytimes30days).
Collegedegree(binary) Twoalternativedefinitions:
(i) = 1 if an individual has a college degree or higher, = 0 if anindividual graduated from a general and/or professionalsecondaryschoolwithcredentials;
(ii) =1 if an individual has three ormore years of college educationwithorwithoutcollegedegree,=0ifanindividualgraduatedfromageneraland/orprofessionalsecondaryschool.
49
CommonvectorXS
Female =1iffemale
Urbanresidence(binary) =1ifresidesinurbanlocationatthetimeofthesurvey
EthnicallyRussian(binary) =1ifethnicityisRussian
Mother’seducation(binary) =1ifmotherhasacollegedegree
Note:Availablein2006and2011surveys;extrapolatedtootheryearsbasedonindividualpanelidforcasesofconsistentreportingofmother’seducation.
Mother’seducationismissing =1ifmother’seducationismissing
Regionalcohortsizeat17(log) Logof17‐yearoldpopulationintheregionofresidenceattheageof17,1985‐2003.
Sources:Census1979,1989,2002;DemoscopeWeekly(www.demoscope.ru/);GoskomstatCentralStatisticalDatabase.
Regionalpermanentearnings(log) Threealternativedefinitions:
(i) Predictedearnings=predictedvaluefromtheregressionofthelogofregionalearningson83regionaldummiesandaquadratictrend,1994‐2011;
(ii) 10‐yearaverageearnings=thelogoftheaverageofregionalearningsoverthefirsttenyearssincecollegedecisionatage17;
(iii) 10‐yearaveragewithastructuralbreak
=10‐yearaverageearningsasin(ii)for17‐yearoldsinthemarketeconomyafter1991;
=thelogoftheaverageofregionalearningsoverthe1980‐1991periodfor17‐yearoldsintheSovietperiodbefore1992.
Note:Regionalearningsaretheaveragemonthlyearningsinagivenyeardeflatedin2000prices.
Sources:GoskomstatCentralStatisticalDatabase;LaborandEmployment(variousyears).
Regionalpermanentunemploymentrate,%
Predictedvaluefromtheregressionofregionalunemploymentrateon83regionaldummiesandaquadratictrend,1994‐2011.
Note:AlternativedefinitionswerenotusedbecauseofthelackofvariationintheunemploymentrateduringtheSovietperiod,whenunemploymentwasconsideredtobenon‐existent.
Sources:GoskomstatCentralStatisticalDatabase.
WageequationonlyXW
Age,agesquared Yearofsurveyminusyearofbirth;themodeofbirthyeariscomputedincasesofinconsistenciesacrossrounds.
Regionaltransitoryearnings(log) Residualfromtheregressionofthelogofregionalearningson83regionaldummiesandaquadratictrend,1994‐2011.
Sources:GoskomstatCentralStatisticalDatabase;LaborandEmployment(variousyears).
50
Regionaltransitoryunemploymentrate,%
Residualfromtheregressionofregionalunemploymentrateon83regionaldummiesandaquadratictrend,1994‐2011.
Sources:GoskomstatCentralStatisticalDatabase
Federaldistricts(dummies) SetofdummiesforresidinginMoscowandsevenfederaldistrictsatthetimeoftheinterview.
Collegeequation,EandI
Nofcampusespermunicipality Totalnumberofcollegecampusesinthemunicipalityofresidenceatage17,1985‐2003.
Acampusreferstoallbuildingsofthesamecollegeinonemunicipality.Eachcollegehasamaincampus;somecollegesmayalsohavebrancheswithcampuseslocatedinothermunicipalities.
Subcategoriesofcampusesincludemaincampuses;branches;publiccampuses;privatecampuses.
Sources:RussianUniversityDatabase(BelskayaandSabirianovaPeter,2014).
Skillwageratioat17 Ratioofaveragewagesofmanualworkerstoaveragewagesofnon‐manualworkersintheindustrialsector(manufacturing+mining+electricity+selectedindustrialservices)atage17,1985‐2003.
Sources:Russianyearbooks(annualissuesfrom1985to2003)
Regionalearningsat17(log) Logofregionalearningsintheregionofresidenceatage17,1985‐2003.
Note:Regionalearningsaretheaveragemonthlyearningsinagivenyeardeflatedin2000prices.
Sources:GoskomstatCentralStatisticalDatabase;LaborandEmployment(variousyears).
Regionalunemploymentat17,% Unemploymentrateintheregionofresidenceatage17,1985‐2003.
Note:Availablefor1992‐2003;assumedzeroduringtheSovietperiod,whenunemploymentwasconsideredtobenon‐existent.
Sources:GoskomstatCentralStatisticalDatabase.
Birthcohorts(dummies) Setofdummiesforindividualsbornin19851988,19891993,19941998,and19992003.
Federaldistrictsat17(dummies) SetofdummiesforlivinginMoscowandsevenfederaldistrictsatage17.
51
TableA2:TestsforSelectiononGains
PanelA.Testoflinearityof | , usingpolynomialsinP
DegreeofPolynomial 2 3 4 5
p‐valueofjointtestofnonlinearterms 0.000 0.000 0.000 0.000
PanelB.TestofequalityofLATEs(H0: 0)
RangesofvforLATE DifferenceinLATEs
p‐value Jointp‐valuej j+1
(0.00,0.04) (0.08,0.12) 0.035 0.000 0.0000(0.08,0.12) (0.16,0.20) 0.029 0.000 (0.16,0.20) (0.24,0.28) 0.025 0.000 (0.24,0.28) (0.32,0.36) 0.021 0.000 (0.32,0.36) (0.40,0.44) 0.019 0.000 (0.40,0.44) (0.48,0.52) 0.017 0.000 (0.48,0.52) (0.56,0.60) 0.016 0.000 (0.56,0.60) (0.64,0.68) 0.016 0.000 (0.64,0.68) (0.72,0.76) 0.017 0.000 (0.72,0.76) (0.80,0.84) 0.018 0.020 (0.80,0.84) (0.88,0.92) 0.019 0.044 (0.88,0.92) (0.96,1.00) 0.019 0.060
Notes toPanelA: Standard errors are bootstrapped (500 replications) to account for the fact that P isestimated. The computation of the test includes two steps. First, we estimate a probit model of collegeequationandpredictthepropensityscore . Second,weestimatealinearregressionof on , ∙ and ,where isapolynomialofdegree j=2,3,4,5. Eachcolumnpresentsthep‐valueofthetestassociated with the null hypothesis that neither nonlinear term has an explanatory power for eachpolynomial.NotestoPanelB:Wecomputethesemi‐parametricMTEasdescribedinSection2andpresentedinFigure4using250bootstrapreplications.LATEisthelocalaveragetreatmenteffect,anditisdefinedastheaverageofMTEineachoftheequally‐spacedintervalsofv,withthetotalof13non‐overlappingintervalsjseparatedbyadistanceof0.04. Thenullhypothesis is that theLATEsof twocontiguous intervalsareequal. Thep‐valueforthisnullhypothesisisshownincolumn4. Thep‐valueistheproportionofbootstrapb forwhich
inintervalj,where and for 1, … 250.Thelastcolumnreportsthep‐valueassociatedwiththenullhypothesisthatthedifferencesacross all adjacent LATEs are different from zero. The joint p‐value is the proportion of bootstrap b forwhich , where ∑ for all j intervals and ∑
for 1, … 250andalljandj+1intervals.
52
TableA3:WeightsforTreatmentEffects
__________________________________________________________________________________________________________________
, 1
, |1
P|
, |1
1 P |
,∗, ,
∆, ∆ P∗ P
,| | |
E | | |
, | |1|
for astheinstrument
| , , | , ,,
, 0
, |1|
, |1
1 |
__________________________________________________________________________________________________________________
Source:HeckmanandVytlacil(2001,2005),Carneiroetal.(2010)
Notes:ATE=averagetreatmenteffect;TT=treatmenteffectonthetreated;TUT=treatmenteffectontheuntreated;PRTE=policyrelevanttreatmenteffect;MPRTE=marginalpolicyrelevanttreatmenteffect;IV=instrumentalvariables;OLS=ordinaryleastsquare.
53
TableA4:Cost‐BenefitAnalysis
Regionalcapitals
Non‐capitalcities
COSTSPERCOLLEGESTUDENTA.Directcostsofcollegeattendance
Tuitionexpendituresforfee‐payingstudentsperyear $1,376 $1,059Shareofstudentspayingtuition 0.576 0.680Additionaldirectcosts(books,supplies,tutoring,etc.)peryear $72 $72Totaldirectcostsofcollegeattendanceperyear $865 $792
B.ForegonelabormarketearningsAverageannualnetearningsofhighschoolgraduates,age18‐30 $5,763 $4,706Probabilityofbeingemployedofhighschoolgraduates,age18‐30 0.831 0.854
Totalcostsofcollegeattendanceperyear $5,656 $4,809Presentvalueofcollegecosts $30,927 $26,300BENEFITSPERCOLLEGESTUDENTC.Improvedlabormarketincome
Rateofreturnstocollege(%) 7.5 9.9Probabilityofbeingemployedofcollegegraduates,age22‐55 0.920 0.929Additionalannualnetearningsofcollegegraduate,age22‐55 $1,789 $1,948
Presentvalueofadditionalearningsstream,32‐yearworklife $38,261 $41,657
Netpresentbenefitsofcollegeeducation $7,334 $15,357
Notes: Regional capitals are most populated cities in the region and have the highest concentration ofcolleges.Allmonetaryvaluesaredeflatedin2010pricesusingnationalCPIandthenconvertedtoUSdollars($1 = 30.6 rubles in December 2010). Tuition expenditures and the share of students paying tuition arecalculated from the RLMS questions available in 2006‐2008 surveys. Additional direct costs are fromGoskomstat(SocialConditionsandWelfareoftheRussianPopulation,2011)andassumedthesameforbothtypes of cities. Real earnings are net of tax, averaged over the 2004‐2011 period, and adjusted for theprobabilityofbeingemployed. Theratesofreturnsaretakenfromtable8(PRTEfromaddingacampusincapital and non‐capital cities). Additional annual net earnings of college graduate, age 22‐55 = (averageannual net earnings of high school graduates, age18‐30) x (percent rate of returns to college/100) x (4.5years of college) x (probability of being employed of college graduates, age 22‐55). The present value iscalculatedatage22over4.5pastyearsofcollegeand33futureyearsofworkinglifeat3percentrealinterestper annum. Calculations of net benefits do not account for earnings during school years, stipends andfinancial support, additional living expenses if a studentdecides tomove toa regional capital, and collegebenefitsbesidesearnings.
54
FigureA1:MTEWeights
Notes: PanelAplotstheMTEweightsusedincomputingATE,TT,TUT,andIVwithestimatesprovidedinTable7;seeAppendixTableA3forformulas.PanelBplotstheweightsonMTEforthreedifferentversionsofMPRTE: 1) ; 2) 1 ∙ ; and 3) , . The MPRTE estimates arereportedinTable8. Thescaleofthey‐axis inPanelBisthescaleoftheMTE,notthescaleoftheweights,whicharescaledtofitthepicture.
0.0
0.5
1.0
1.5
2.0
2.5
We
igh
ts
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0v
ATE TT TUT IV
A. Standard treatment parameters
-0.2
-0.1
0.0
0.1
0.2
0.3
0.4
MT
E a
nd W
eig
hts
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0v
MTE MPRTE1 MPRTE2 MPRTE3
B. MPRTE weights