Tilburg University
Microeconometric essays on migration, trust and Satisfaction
Bellemare, C.
Publication date:
2004
Document Version
Publisher's PDF, also known as Version of record
Link to publication in Tilburg University Research Portal
Citation for published version (APA):
Bellemare, C. (2004). Microeconometric essays on migration, trust and Satisfaction. CentER, Center for
Economic Research.
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal
Take down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
.1.
TILBURG *1 11T * UNIVERSITY
•1 •
UNIVERSITEIT * * VAN TiLBURG *
BIBLIOTHEEK TILBURG
Microeconometric
Essays on
Migration,
Trust, and
Satisfaction
PROEFSCHRIFT
ter verkrijging vandegraad van doctor aan de Univer-siteit van Tilburg,op gezag vande rectormagnificus, prof. dn F.A. van derDuynSchouten, in het openbaar
te verdedigentenoverstaan van een door het college
voor promotiesaangewezen commissie in de aula van
de Universiteit op
vrijdag 28 mei 2004 om 10 :15 uur
door
CHARLES BELLEMARE
A France et Fran ois pour m'avoir donnd le gout de la vie et de la recherche
A lean-Fran ois,
Acknowledgements
Thisthesiscontains most of the work I pursued during my
four
years stayin Tilburg.Over thistime period,manyindividuals contributed either directly or indirectly to the writing ofthis thesis.
First of all, Ihavegreatly benefited from
my
collaborationwith Arthur
vanSoest who supervised my work from the verystart, andhasco-authored thefinalchapter ofthis thesis. I thank him for havingencouraged me topursue my ownresearchideas, andfor having given so much of his time, guidance, and words
of
encouragement.Arthur's joy
toinstructothers, and his disinteresteddedication tohis students, sets an examplewhich I will try
to emulateallthrough myacademic life.Duringthefirstsemester ofmystudies, I wasconfronted withthe technicalaspects of econometrics for which I was not reallyprepared. Making this transition
would
have beenmore difficult without the help ofBetrand Melenberg. The extra evening
lectures he gave me on thewhite board of my office willhavegreatlyimproved my understanding ofthefieldof econometrics. Bertrandco-authored the lastchapter of this thesis, and I amdelighted thatheaccepted to join
my
thesiscommittee.Special thanks gotheother members ofmy thesis committee, Claude
Montmar-quette,ChristianDustmann, and Jan van Ours.
My
discussions with them have shaped the way Ithink
about economic problems. I am especiallygratefulto ClaudeMontmar-quette
for
havingintroduced me toexperimentalandbehavioraleconomics while I was working attheCIRANObefore entering the Ph.D.programin
Tilburg. Hisenthusiasmforthesefields has beenasourceofgreatinspiration. I look forwardtocollaborate with
him and his team at theCIRANO inthe years to come.
I am grateful to CentER and to theDepartment ofEconometricsfor havinggiven me
the possibility topresent my workat numerous conferences,workshopsandsummer schools. I have gained muchfromtheseexperiences,both personallyand
profession-ally. I would also liketothank Marcel Das and all the teamatCentERdatafor having
Viii
My
stayin'Illburg hasbeenhighlighted by myjoint
venturewith
Steffan Berridge in runningtheinfamousMieredikhof Hotel. Memorable cocktail parties and backyardfires have helped us andfellow students to remain cheerfulthrough the cold winter
nights
of
111burg. I willespecially remember Steffan'smild
chilies, his barbecueblack-out, and his special way
of
havingwhiskey.My visit
toChina with him and Tu Qinhas left me withawealth
of
memories. It has beenapleasure to know him and I am honored thatheaccepted to be one ofmyparanymphs.I would also like tothank my twooffice mates, JoostDriessen and in particular
Pierre-Carl Michaud, for stimulating conversations anddebates.
My
thanks extendto all thePh.D. students and friends in
lilburg,
with
specialmention to Pierre-Carl, Antonis, David, TuQin, Youwei, Dantao, Steffan and Vera, who have allhelped mebalance mylife between workand leisure.
It does not take much reflection torealize how much ourachievements in life
de-pend on thesupport given to us by our family. Leavingmy
parents and brother has unquestionably been the most
difficult part of
mystudies abroad. I rejoicemyself in the thoughtofcoming back home, and it iswith
great emotion that I dedicatemythesis to them.Finally,Sabine hasshared mylife throughoutthe
writing of
this thesis. Her care,comfort,andpresence,
will
remain themost importantheritage of my timein Tilburg.Charles Bellemare
Contents
Acknowledgements Vii
Contents ix
1 Overview ofthethesis 1
2 IdentificationandEstimationofEconomicmodelsof Outmigration usingPanel Attrition 9
2.1 Introduction . . . , . . . , , , , . . , . . . 9
2.2 Identificationof outrnigrationparameters . . . 12
2.3 Parametricmodelandestimationmethod . . . 16
2.4
Data... 20
2.5 Results and simulations . . . . . . . . . . . . . . 22
2.6 Conclusions . . . ... 29
Appendixto Chapter 2 30
3 A Life-Cycle Model of Outmigration and EconomicAssimilation of Immi-grantsin Germany 41
3.1 Introduction . . . .4 1
3.2 Economicm o d e l. . . 45
3.3 Estimation procedure . .. .. . . ... . .. . . .. 49
3.4 Background and Data . . . 56
3.5 Estimationresults for thestructuralmodel . . . .. . . . .. 58
3.6 Conclusions . . . 66
Appendixto Chapter 3 68
X
4 OnRepresentativeTrust 81
4.1 Introduction . . . .8 1 4.2 TheExperimental Design and the
Sample ... 85
4.3 ResultsonTrust . . . . . .
. . 88
4.4 Results on
Trustworthiness . . . 95
4.5 Results onParticipation .
. . . .9 9
4.6 Discussion andConclusions . . . , . . . 100
Appendix
toChapter4 103 5 Semi-ParametricModels for Satisfaction withIncome 115 5.1 Introduction . . . , . 115 5.2 Models, EstimationTechniques,and SpecificationTests . . . 1175.3 Data and
Variables . . . 125
5.4 Results . . . .
. . . 1275.5 Conclusions...···
131Appendix
toChapter5 132Bibliography 143
List
of
Tables
2.1 Panelattrition forWest-German andImmigrantsamples1985-1999. . . 31
2.2 Descriptive statistics for West-German and Immigrant samples, 1985
and 1999 . . . 32
2.3 Unemployment Rate per Wave and Land1985-1999 . . . . . . . . . . 33
2.4 Covariancestructure of thetimevariant and timeinvariantcomponents 34
2.5 Estimation resultsfor o u t m i g r a t i o n. . . 35
2.6 Estimation resultsforincome andwork equations . . . . . . 36
2.7 Simulation
Results . . . 37
3.1 Descriptive
statistics, 1985 and 1999 . . . . . . . 69
3.2 Estimation results of thestructural
model ... 70
3.3 Panelattrition
for
West-German andImmigrantsamples1985-1999. . . . 713.4 Simulatedmigration
durations . . . 72
3.5 Estimation results of the reducedform model ... 73
4.1 Descriptive
statistics . . . . . . . . . . . . . . 108
4.2 Sender
results . . . . . . . . . . 109
4.3 Responder results
-
Tobitestimator . . . 1104.4 Responder results-STLS-estimator . . . 111
5.1 DescriptiveStatistics of thesample . . . 133
5.2 LM SpecificationTestsOrderedProbit . . . 134
5.3 Estimation Results, ExtendedSpecification . . . . . . . . . . . . . . . . . 135
5.4 EquivalenceScales . . . 136
List of
Figures
2.1 Logmonthlyearnings andworkpropensities-Germans andimmigrants 38
2.2 Simulation
Results . . . 39
3.1 Proportions
of
immigrantsworking
inGermany,notworking
and attri-tion per time period,1987-1999. . . . . . . . . 74
3.2 Goodness of fit of themodel ... 75
3.3 Simulateddistributions fortheforward looking
model . . . 76
3.4 Simulateddistributions forthemyopic looking
m o d e l. . . 77
4.1 Distribution
of
experimentaltrust. . . . 1124.2 Returnratio
of
responders for eachunitsreceived, strategymethod. . . 1124.3 Estimateddensity
of
potentialreturns oninvestments . . . 1135.1 Simulated rejection probabilities Fan and Li (1996) test . . . 137
5.2 Distribution
of
satisfactionwith
income . . . 1375.3 Distribution of loghousehold incomebyhousehold size . . . 138
5.4 Nonparametricregressionofsatisfaction
with
income onlogincome byhousehold size . . . 138
5.5 Non-parametricestimate oflink function, with95%uniformconfidence bands.Ichimura'sSLS;specification 2 . . . 139
5.6 Non-parametricpartpartiallylinearmodels . . . 139
5.7 Non-parametricpartgeneralizedpartially linearmodels . . . 140
Chapter 1
Overview of
the thesis
This thesis presentsfouressayswhich illustratetheprocessof makingmicroeconomic inferencesoneconomicallyrelevant parameters characterizing
individual
welfare and decisionmaking usingindividual
andhousehold data.Thisprocess isusually thought of asathreestepapproach, whichstartsbydrawingoneconomictheoryto focus on the parametersofinterest.Giveneconomically relevant parameters havebeendefined, the second step of theprocessconsists of obtainingan "ideal" datasetwhich is richenough to
identify
and estimatetheseparameters. Itisoften the caseineconomics andotherobservational sciences that this"ideal" data set isnot available,whichcanoften jeopardizeidentification of some or all parameters. In thiscase, researcherstypically trytocircumventtheproblembyeitherimposingassumptions onindividualbehavior,
or bygettingbetter data. Ofcourse, in theformercase,thecredibility oftheinferences depend on the realism andfalsifiability oftheassumptions made,whilethefeasibility ofthelatterapproach depends onwhether or not onecanactuallycollect richer data. Once identification hasbeenachieved, thefinalstep consistsof using existing or new estimatorsin orderto extract therelevantinformation from thedata. Becauseseveral estimators,under different set
of
assumptions,cansuccessfully extractthis informa-tion, empirical researchers havea degree offreedom to choose the model which isestimated. Fortunately, economictheoryoften restricts the setofpossible econometric models, andmodeltesting procedures and goodness of
fit
evaluationscansometimes be used to test thevalidity
of maintainedassumptions,providing a waytoincrease thereliability oftheinferences made.
The essays in thisthesis arepresented asindependent chapters, and analyze the characteristics anddecisionrulesof immigrants wholeavetheirhost country, evaluate
2 Chapter 1. Overview of the thesis
the determinants of trust andtrustworthiness for different sub-groups ofthe
popula-tion, and estimate household equivalence scales. Ineachchapter,identification of the
economic parametersofinterestisachieved either byimposingrealistic and falsifiable
assumptionson
individual
behavior, orby
collecting better data. Taken together,theseessaysdevelop a widerangeofeconometric modelsin orderto estimate the
parame-tersofinterest,fromdiscretechoicestaticand panel data models, tostructural dynamic programmingmodels, nonparametric models, and semiparametricmodels. Although
issuesof identificationandestimationarediscussed inallessays,theirrelativeweight
differsacrosschapters.
The firsttwo chapters deal
with
outmigration,defined as thedeparture ofimmi-grantsfrom their adoptivehomeland.During thelastdecade,economists have increas-inglybecome aware that outmigration is not onlya world widephenomena, but the
magnitude oftheoutflows insomecasesbe sizeable. WithinEurope, the case of
Ger-many stands outprominently, withanestimated halfa
million
immigrantsleaving thecountryevery year (OECD,2001).Understandingthemotivations
for
thesedeparturesis animportant input for policy makers whomust forecastoutflows
of
immigrants in orderto adjusttheir immigration policies to fitthefutureneedsof theirlabor markets, and to assess the economic performance of immigrants who remain in the country. Several theories havebeenproposed toexplainthesemovements, someof which pre-dict thattherelativelysuccessfulimmigrantsleavetheir adoptivecountry,whileother
theoriespredict that most ofthe departures areby unsuccessfulimmigrants. Some-what surprisingly,there hasbeen
little
structuralinference on the preferences, tastes, and other primitivescharacterizingimmigrants which would allowtovalidatethese theories, helpusunderstand the typeof immigrants whichself-select themselves out of the country,andwould
enableinteresting policyexperiments. Given thattheparam-eters ofinterestare clearlydefined, the lack ofinferences can betraced back to data
limitations. The body ofdatatypicallyavailable toconducttheseinferences usually
consists ofasampleof immigrants, reinterviewed in subsequentyears. This type of
surveydata rarely containsindicators of immigrantsdepartures. Instead, what is em-piricallyobserved isattrition fromthe panel,caused eitherby immigrantdepartures or by non-response of the nondeparting immigrants. Thus, thedata fails todirectly
identify
the parametersofinterest, and gatheringthe missing informationwould in
many instancesbeprohibitivelycostly.
im-3
posed on
individual
decisionmaking in order tononparametricallyidentifyseveralimportant economic parameters characterizing the performance
of
outmigrants, no-tably theconditional outmigration probability, theconditionalwork
probability, andtheconditional expected labormarketearnings
of
outmigrants. Theassumptionre-quired isthe existenceof immigrants whose observable characteristics are such that
their outmigration probability isclose to zero. Equally important,becausepanel
at-trition is, contraryto outmigration,always observable in surveydata, these results open up thepossibility tomake inferenceswhich werenotpreviouslypossible due to
data limitations. Toextract theparametersofinterestfrompanel attrition, the
chap-ter presentsa
limited
dependent variable model forpanel datawith
randomeffects,wherelabor marketearningsareestimatedjointly with both workand outmigration equations. An importantaspect of theempirical model is that itallows to test the key
identification assumption inasimpleway. Indeed, the results indicate thata
substan-tial number of immigrants haveanoutmigration probabilityclose to zero. The main
results fromtheempiricalanalysis show that outmigrants arenegativelyselected in terms of labor marketearnings and
work
propensities. Throughoutthe sample pe-riod, we find that the gap ineconomic performanceof immigrants who remained in Germanyand thosewho outmigrated was sizeable,with
outmigrants having work probabilitiesand expected labormarketearnings respectively 18% and30%lower than immigrantswho remain inGermany. Finally, the estimated outmigration rate in thesamplepopulationisfound toberoughly 3% per year.
Chapter3presents and estimatesa structural dynamic model whereimmigrants
simultaneouslychoosetheir workandmigration durationsover theirlife-cycle.
Con-trary to existinglife-cycle models,immigrantsareallowed tofaceuncertainty about
their future labor market outcomesall through their migration experience. Because outmigrationis rarely observed at the individual level, structural estimates of
life-cycle models
of
outmigration have yet tobeformallytested.This chapter draws on the identification results ofthepreviouschapter and extends the approach toadynamic programmingsetting. This allows to make direct structural inferencesonoutmigrationlife-cyclebehavior
without
havingto observeoutmigration and to testthevalidity of
several existinglife-cycle theories. Our findings confirmthe hypothesis recently put forward intheliteraturethatoutmigration isnotentirely driven byearnings
differen-tials. Specifically, we findthatimmigrants whofeelintegrated inthe Germansociety,
na-4 Chapter 1. Overview of the thesis tive country are less
likely
tooutmigrate. The results ofthis chapter alsohighlight
the importanceof incorporating the
work
decisionalong withthemigration durationdecisionof immigrants,afeaturepreviously ignored intheoutmigration literature. It is found that both immigrants with relatively low and highlabormarketexperience
haveagreater overallutilitytooutmigrate, whichsuggests aUshaperelationbetween
labor marketexperience and theoverall
utility
to outmigrate. The decreasein overall outmigrationutility
starting from lowlevelsofexperienceisconsistentwith
increas-ingpsychologicalcosts associatedwith outmigration,keeping earnings constant. The convex increase
in
overall outmigrationutility
predicted to occurbeyond25years oflabor marketexperienceisconsistent
with
progressivelylowerpsychological costs of outmigrationanddiminishingreturns to labormarketexperience in the host country.Theseresults are interestinggiven that most of the outmigration literature has
ana-lyzed outmigration
within
anearningsdifferential paradigm which, byconstruction, orient policy recommendations towards measures aimed atinfluencing theearningsdifferentialbetween the host and thenew destination. Theresults of this chapter do not rule outtheimportantrole playedby labormarket earningsin determining migration durations. However, they doindicate thatthe shape of themigration duration distri-bution isalsodetermined by pastworkdecisions,indicating that much canbegained
fromananalysisin whichlaborsupply decisionsareendogenously determined. The
chapteradditionallypresentsaseriesof policysimulations aimed atevaluating how
themigration duration distribution of immigrants is affectedby government policy
aimedatspeedingup economic integration of newly arrived immigrants. Itisfound
thatmeasuresaimed atimprovingaccess to thehostlabor
market
improvingacquisi-tion
of
language skills, andlump-sum taxation willhavelittle
impact ofthemigration duration distribution ofarepresentative immigrant, buthavesubstantialinfluence of migration durationdecisions oflow
incomeimmigrants.In chapter 4, the economic parameters
of
interest arethedeterminantsofindividual
trust andtrustworthinesspropensities. Compared to theprevioustwo chapters,
iden-tification oftheseparametersisachieved, notby imposinga
priori
assumptions on theindividual
decision maker, butbyusingbetter data in the form ofa randomsample oftheDutch population who playanexperimentalgamemeasuringindividual
trust andtrustworthinesspropensities, bothof which havebeenshown tobestrongly corre-latedwith
economicgrowth.Theidentificationapproachisnovel in many respects and5
experiments which uses a sample
of
subjectswhich are representative ofa nation'spopulation. Substantial heterogeneity
in
trust and trustworthinessbehaviorisfound intheDutch population. Of particularinterest are the rolesplayed bythe education level and the ageofsubjects. The chaptershows that both ofthesebackgroundchar-acteristics playan important role in determiningtrust andtrustworthiness, although
they affect trust and trustworthiness in very different ways. While the inverted U
shaperelation between trust and ageusually
found in
the literature is also presentin our data,
with
trust increasinguntil the age of 30and decreasingbeyond that, the relationbetween trustworthiness and ageisfound to beU shaped,with
trustworthi-nessdecreasing until the age of 30, andrisingagain beyondthat point. Theeducation patterns are also very different. We find an inverted-Ushaperelationbetween edu-cation and trust, and a U shaperelationshipbetween education andtrustworthiness. Suchopposite influencesofbackground characteristicsarepuzzlinginsofar as it istyp-icallyassumed that trust andtrustworthiness go hand inhand,which
would
suggestthat botharedeterminedin similar ways. Asa secondmain contribution, the chapter presents a new andsimple wayto compare theinformationalcontent ofbothrevealed (experimental) and stated (survey) trust measures. The literature has until now
as-sessedthe
validity
ofsurvey trust questionsby testingwhether or nottheypredictedwell
experimentaltrust. One of themainmessages ofthe chapter is that thismethod of validation hasbeen given too muchattention, primarilybecausethepredictive powerof surveymeasuresisintimately linked tothe sample used, the amount ofbackground
informationavailable on the subjects, and theexperimental design. Itisshown that by carefullyselecting samples and designs, experimentersincreasetheir odds of finding
either a low or high predictive power ofthesurveytrustmeasure. Thus,despite that
contrary totheexisting literaturethesurvey trustmeasure used in this chapter predicts well trust intheexperiment, this is not takenasevidencevalidating the use of survey
trust questions. Withregards to theinformationalcontentbothmeasurescarry on the
determinants oftrust,thedifferencesarecompelling. Itisshownthateducation has an
inverted-Ushaperelation
with
experimental trust while it does notcorrelate at all withanswers to the statedtrustquestion. Incontrast,religioncorrelates wellwithanswers tothestated trust question, but not at all
with
experimentaltrust.Instark contrast, chapter5 presents aexample where statedresponsesprovide a
simple solutiontoidentificationproblemswhichcannotbeovercome using data on
6 Chapter 1. Overview of the thesis
incomewhichshouldbegiven toahousehold withaspecific compositionofadults and
childrentoreach the samelevelofwelfare thanareference household.Thesequantities are important for public policyconcerningwelfarebenefits andchildallowances, and
have also been used to analyze incomeinequality
within
and between countries andforthe analysis
of
poverty.Thetraditionalapproach to estimationofequivalencescales usesconsumerdemandsystems,relyingonvariation inexpenditure on commoditiesacrossfamilieswith differentcomposition. PollakandWales (1979)andBlundell and
Lewbel(1991) demonstrated that consumer demand data alone is notrichenough to nonparametrically identifythelevel of the equivalence scales. In this chapter, the
fail-ure to identify levels using revealed consumption choices is overcome using stated
responses to questions on satisfaction with income, which directly reveal a cardinal
measureofconsumerutility. Statedanswers to such questions provide empirical
re-searchers with a discrete and ordered response variable,usually
analyzed using
or-deredprobitand linearregression models. The emphasis of the chapter is to recover
estimates ofthelevelofequivalencescalesforGerman householdsusingabroad range
ofestimation procedures, from the usualorderedprobitandlinearregression models,
to more recent nonparametric andsemiparametricmodelswhichweaken parametric
assumptionsof linearand ordered probit regression models. Given the broad range
ofpossible estimators, the chapter uses consistent waysof testing each underlying model assumption against generalforms
of
mis-specification. Themain findings of the chapter arethatorderedprobitmodelsarerejected,whileseveralsemiparametric estimators are not. Despite this, estimates of the equivalencescales arefound to bero-bustacrossmodels. Finally, itisfound thatequivalencescalesmonotonicallyincrease
with
thenumber of children for low andmiddle
income households, butarefound to be relatively flat for highincome households.Collectively, these essayspresent waystoovercomeidentification problems in
mi-gration, experimental economics, and measurementofhousehold equivalencescales. Ineach chapter,illustration ofthe approachesrequired that theybeapplied either to veryspecific countries,institutionalsettings, and choice environments. However, the usefulnessofthese essays goes beyond theseapplications, andprovidemeans for a
significant number of extensions. Theidentification strategy proposedin chapter 1
opens up thepossibility todoresearch onthe economicsof outmigration in practically any country with anongoing panel surveying immigrants. Because
outmigration is
aworldwidephenomena, this avenueoffutureresearch is
or-7
der to improve our understanding ofthe flows of human capital takingplace across
the globe. Moreover, because this strategy is relatively general, itcaneasily be ap-plied to analyze any form of migration movements, fromthe departures ofnatives, to movementsofindividualswithinacountry Theresults presented in thefirst two
chapters of this thesis show that such extensions haveapromising future. The thesis also showed that one cannowsuccessfully makepopulationinferences onbehavior in
experimentalgames. Whilethe thesishasshown that inferences based onsurvey
re-sponses can insomecasesdiffer remarkably fromthose basedon revealed preference data collectedfrom experiments, it has alsobeen shown thatsurveydata remains a useful source toidentifyeconomic parameters related to measuresofhousehold sat-isfaction. More generally, it ishoped that this thesis has alsodemonstrated that the combinationofsurvey andexperimentalmethods has thepotentialtoprovideuseful
insights in many differentgames,settings, andpopulations, which
will
undoubtedlyChapter 2
Identification
and
Estimation of
Economic
models
of Outmigration
using
Panel
Attrition
2.1 Introduction
The widespread and often sizeableflows
of
immigrants leaving theiradoptive home-lands,which wewill
refer toasoutmigration,have recently received a lotofattention. The case of Germany is a particularly revealingexample, withan estimated yearly outflow of halfamillion
immigrants over thelast decade (OECD,2001). Several the-ories have beenput forward to motivate outmigrationmovements. Theories basedon earningsdifferentialsbetween the current and newdestination (Harrisand Todaro,
1970),higher marginal
utility
of
consumption in thehome country (Djajic andMil-bourne, 1988),highreturns tohumancapitalinvestments in thehost country
(Dust-mann, 1993), information dissemination(Stark, 1995),credit market rationing in the native country (Mesnard, 2001), and several sociological factors suchas
family
uni-fication, health satisfaction, feelingsofbeing integrated insociety (Stark, 1998) , and the qualityandproductivity ofanimmigrant'ssocialnetwork(e.g. Carrington, Detra-giache and Vishwanath, 1996).
Thesetheories do not
trivially
predict a specificcomposition of departingimmi-grants. Itcould be the caseforexamplethateconomically successfulimmigrants with
arelatively higher marginal
utility
of consumption in their native country might optto leave despiterelatively lowerearnings inthatcountry,while persistently
10 Chapter2. Identification ofOutmigration
ful
immigrantsmight find aworthwhile to keepon searchingforbetter labor market prospects and move to anew destination. The lack ofaclear-cut theoreticalpredictionconcerning theeconomic performance
of
outmigrants complicates the design of im-migrationpolicieswhichareoftentailoredaround thenotion thatlabor marketequi-libriums may be affectedby outmigration flows of non-randomlyselectedworkers. Careful forecasting ofthequality ofthemigration flows isthusnecessary
if
immigra-tion policies are to meetthe futureneeds of the labor market. Additionally, recenttheoretical andempiricalevidencehassuggestedthatmeasuresofeconomic progress
of immigrants in theirhost country canbeadversely affectedbyselectiveoutmigration
(Schultz, 1998, andEdin, LaLonde, andAslund,2000).Policies aimed atimproving the
labormarket integration ofitsimmigrant population may thus alsobemisguided if it
reliesonthesepotentiallybiased measuresof immigrant assimilation.
In thischapter, weareinterestedintestingcompetingexplanationsof outmigration
decisions while at the same timeassessing therobustnessofmeasuresofeconomic
as-similationtooutmigrationflows. We do sobyestimatingalimited-dependentvariable panel data model where labor market earnings,workdecisions, andoutmigration de-cisionsarejointlydetermined and depend on earningsdifferentials, family unification, creditmarketrationing,health satisfaction, and feelingsof beingintegrated in the host
society. Unobservable characteristics such as animmigrant's ability orthequality and
productivity ofhis socialnetworkareincorporated inthemodelasindividualspecific
unobserved heterogeneity components. The generalerror structure ofthe model al-lows totestwhether outmigrants areself-selected in terms
of
labor marketearnings andin terms ofwork
status. Themodel alsoallowstoassess theimpactof outmigra-tion selecoutmigra-tiononestimates ofmeasuresofeconomic assimilationrates. Compared to existing empiricalmodelsof outmigration(seebelow),ourmodel hastheadvantageof both incorporating thedecision to work in a natural wayand characterizing the relationshipbetweenworkstatus andoutmigration.
Estimation of our modelrequires panel data onimmigrants followed over a rela-tively longperiod of time. As Dustmann(2002)recentlypointed out,interesting
iden-Section2.1.
Introduction 11
tify
the economic parameters characterizing the performanceof outmigrantswithout
observingoutmigrationdecisions. Jassoand Rosenzweig(1990)
identify
thedirection of outmigrationearningsselectivityby comparingtheskill composition ofspecificco-horts over time. Hu (2000) andLubotsky(2000)estimate the parameters of the earnings functionof immigrants who remain inthecountry,controlling for non-random outmi-gration selectivity by matchingcross-section data setsand
longitudinal
socialsecurityearnings records. Theseapproaches are not
without
their own limitations-they do notidentify
thelevelofearningsofoutmigrants,censusand earnings records often havelittle information on boththehuman capital level and sociological characteristics of migrantswhicharerequired to testoutmigrationtheories 2, they do not easilyallow migration duration decisions to depend on unobserved characteristics suchas inher-ent ability orthequality ofanimmigrant'ssocialnetwork which,at leaston theoretical grounds,areimportantdeterminantsof migration durations.
Inthis chapter, we presentconditions under whichtheoutmigration probability,
the
work
probability, andthe expected labor market earningsof
outmigrants are,con-ditionallyon observable characteristics,nonparametrically identified fromsurvey data
with
sampleattrition.The cornerstone ofour identificationapproach consistsof usingpanelattrition asaproxyvariable for outmigration,and subsequently separating
at-trition which is not duetooutmigrationmovements from realoutmigrationdecisions.
Ourapproach overcomes several of the shortcomings
of
earlier approaches used toidentifythe economic performanceof outmigrants.First,ourapproach is toour knowl-edgethefirstonewhich hasthepotentialtoprovide nonparametric identification of the economic performance and movementsof outmigrants.Second,becausethe approach
proposed uses survey panel data instead ofcensus data, unobserved heterogeneity
can easily beintroduced in themodel. Third, the approachis general enough to be applied toany
country with
an ongoing panelof immigrants (examplesofcountrieswithsuch panelsareCanada, Mexico, Germany, and the United-States.) and can be
easily extended to estimatemanydifferenttypesofeconomic modelsof outmigration.
An example ofthis flexibility is our abilityto analyze theinteractionbetween earnings, work,and outmigration decisions ina unifiedframework which wasnot previously possible
without
observing actualmigrationdecisions. Finally, the estimator proposedis simple to apply, andissimilar in
spirit
to estimators proposed to dealwith
12 Chapter 2. Identification of Outmigration
We estimate our model using thepublic use file of the GSOEP. We use data on
native Germans as a reference group to compute earningsassimilation rates for our
immigrantsample. We show how theavailability ofasampledrawn from thenative population has theadditionaladvantageof providinganatural way to test some of the identifyingassumptions ofthemodel.Weestimate the annualoutmigration rate in our
sample toberoughly 3% per year over our time horizon. Our simulation results
indi-catethataveragelog earnings
of
outmigrantsremainedroughly18%lowerthanthoseof immigrantstayers,aclearindication that outmigrantsaredrawn fromthebottom of the incomedistribution. Moreover,outmigrants areshown to have
work
probability25% to45%lower than thatofimmigrantstayers over theperiodconsidered. Finally, we do not find thatassimilationrates areparticularlysensitivetooutmigration, which
contrastswithexisting resultsfound inthe literature.
The rest of this chapter isorganizedasfollows. Section2.2presentsourapproach
to identifytheeconomic parametersofinterest. Section2.3presents the econometric model used tomodel outmigration in conjunction with the
work
decision and labor marketearnings. Section 2.4presents the data used in the chapter. Section2.5 discussestheempirical results of the model and tests for thepresenceof outmigration bias. It furtherpresents somesimulation results usedtoevaluate the fit ofthemodel and to quantifythe economic performanceof outmigrants.Section2.6concludes.
2.2 Identification of outmigration
parameters
Each immigrant ofapopulation living in thehostcountryischaracterized in agiven
time period by thevector (w, p, ru, x, z,s) wherew denoteshispotential labor mar-ketearningsconditionalon characteristics x, p isabinary variable takinga value of 1
whentheimmigrant worksandwhose outcomeisconditioned onavectorof observ-able characteristics z, ru isabinary indicator takingavalue of 1 when theimmigrant
outmigrates in the next time period andwe conditionthis outcome onavector of
char-acteristics s. Wedefine X as the vector of all distinctelements of (x, z, s). We are
interested in makinginferences on Pr (p = 1 lru = 1,
X),
conditionalwork
probabilityof an outmigrant, and on E {wlp = 1, ru = 1, X},theconditionalexpectedearnings of outmigrants. Theinferentialproblem consistof identifyingthesequantities when,
in-steadofobservingoutmigration,we observea
proxyvariablero,panelattrition, which
Out-Section2.2.Identifcation results 13 migrationand
attrition
arerelatedbecause animmigrant wholeavesthecountry mustalso leavethe panel
with
probability 1.3We illustrate theidentificationproblems for thecasewhere we want tomake
infer-enceson E{wlp = 1, ru =
1.X}.
However,allresults extenddirectlytoidentification ofPr (p = liru = 1, X).4 Theconditionalexpected labor marketearningsof immigrants
who leave thepanel in the next timeperiod can be expressed,usingiterated
expecta-tions, as
E{wlp = l.ro = 1,X} = E{wlp =l, r° =l, ru =l,X} ·Pr(ru =llro = 1,X)
+E{wlp = 1,78 = 1,ru = O. X} · Pr(ru = 0170 = 1. X) = E {wlp = 1, ru = 1, X} ·Pr (ru = 1 lro = 1, X)
+E {wlp = 1, ru = 0, X} · Pr (ru = Olro = 1,
X)
(2.1) wherethesecondequalityfollows from the fact that once we know ifanimmigrant leftor not the country, observing whether he left of not the panel doesnotcontain any addi-tional information onthe earnings ofthisimmigrantconditional on X.5 Animmediate
consequenceofequation (2.1) is thatusing panelattrition asaproxyvariablefor
outmi-gration
in itself will
give biased and inconsistent estimates of theconditionalearningsof outmigrants. This is sobecause the conditionalexpected earnings of immigrants
who leave thepanel will ingeneral beaweightedaverage of theconditionalexpected
earnings of outmigrants mixed withthe conditional earningsof immigrants who
re-main in the host country.Themixing probabilitiescontrol the size of the bias. The key
parameter is Pr (ru = Olr° = 1,
X),
whichrepresents theprobability thatanimmigrantstays in the thehostcountrygiven that heisobserved to leave the panel. Thehigher this probability, the higher will be the bias.
If
everyimmigrant wholeaves the panelalso leaves thecountry, r°would perfectlymeasureoutmigration, Pr (ru = Olr° = 1, X)
wouldbeequal to zero, and the bias would be zero.
Next, wefollow thesame stepstoderivetheconditionalexpected earnings of im-migrantswho remain inthe panel
E{wlp = 1, ro = O,X} = E{wlp = 1, ru = 1, X}·Pr(ru =llro = O,
X)
(2.2)+E{wlp = 1, ru = 0,X} ·Pr (ru = Olro = O, X)
Becauseanimmigrantcannotbeobserved to have left thecountrygiven heisobserved
to be in the panel, Pr (ru = 1 Ir° = 0, X) = 0, Pr (ru = 0 Ir° = 0, X) = l and (2.2)
simpli-fies to
14 Chapter 2. Identification of Outmigration whichindicates thattheconditionalearningsof immigrants who remain inthe panel
coincides withthe expected earningsof immigrants who remain in thehost country, andit follows that E {wlp = 1, ru = 0, X} isnonparametrically identified from data on immigrants inthe sample.Substituting (2.3) in (2.1)weobtain
E{wlp = 1,f = 1,X} = E{wip = 1, ru - 1, X}·Pr(ru =llro = 1, X)
+E{wlp = 1, ro = 0, X} ·Pr (ru = Olro = 1. X)
which canbesolved interms of E {wlp = 1, ru = 1, X},theparameter we hope to
iden-tify,
E{wlp=l, ru=l,X} = E{wlp=l, ro -1, X} ·Wi(X)-1
-E{wlp = 1,t - 0, X}· Wo (X) Wi (X)-1 (2.4)
Equation(2.4)shows that theconditionalexpected earningsofoutmigrants can be ex-pressed as a weighted average oftwo conditional expectationswhichareidentified from the data. Iftheweights canbeidentified, thentheconditionalearnings
of
outmi-grants willbeidentified. ApplyingBaye's rule, theweightsaregiven byWo (X) = Pr (ru= Olt =1, X)
= Pr (ro = 1lru = 0, X) Pr (ru = OIX)
Pr (r° = 1 IX) Wl (X) = Pr (ru = 1 lro = 1, X) - Pr (P = 1 1 r u - 1 X)Pr (ru = 1 IX) v ' , Pr ( r°= 1 IX) =1 Pr (ru = 1 IX) Pr (ro= 1IX)
Pr (r° = 1 IX) is identified fromthe attrition data. What remains to beidentified is
Pr (ru IX) and Pr (ro = 11ru = 0, X). Itisclear assumptions mustbeplaced on the data generatingprocesstoidentifytherelationshipbetweentheobservable
attrition
indica-tor r° and the unobservableoutmigrationindicator ru.To
simplify
thenotation, we willdenote a10 (X) E Pr (r° = 1 lru = O.
X)
Using the law oftotalprobability,theattritionprobability canbe expressedin general terms as
Pr (ro = 1IX) = alo (X) + 11 - alo (X)] · Pr (ru =
lIX)
(2.5)and theprobabilityofremaining inthe sample as
Pr (r° =
OIX) =1- {alo (X) t[l- trio
(X)]
Pr (ru = 1IX)}Section2.2.Identification results 15
Several assumptions canbeimposed on the data generatingprocessin orderto
identify
both alo (X)andPr (ru = 1IX).
Assumption Al a10 (X) = alo 2 0
Proposition2.1. If Al holdsandthereexists a X such that Pr (ru = 1 IX) = 0. E {wip = 1, ru = 1, X} is nonparametrically ident(Red.
Proof. Given the foregoing discussion, it suffices to establish that theconditions of
theProposition
identify
the weights Wo (X) and Wi (X). From (2.5), it follows thatPr (r° = 1IX) = 0:10whichidentifies 0:10from
limit
observations satisfying Pr (ru = 1 IX) =0. Given a lo isidentified, Pr (ru = 1IX)isidentified from(2.6),which implies that the
weights Wo (X) and Wo (X) are bothnonparametrically identified. 0
Proposition 1 shows that
if
attrition which is not duetooutmigrationisrandom inthe population (Al),
all importanteconomic parameters characterizingoutmigrationbehavior canberecovered from the data
if
there existsasampleofpermanentmigrants; i.e. immigrantswhoseoutmigrationprobabilityisclose to zero. Inpractice, this does not seem to beastrong requirement, especiallyfor
western countries wherepermanentmigrationisknownto occur at averylarge scale (OECD, 2001). Note thatinpractice,
Al needs not to hold
if
attrition which isnotrelated tooutmigration does not vary in the population (i.e. ifthevariance V (trio (X)) . 0). This canbeverifiedforexample bycomputing marginaleffectsfrom binarychoice regressions onattritionoutcomes for a sampleofindividuals who by construction donotoutmigrate, and test
if
theseeffectsaresmall. Nativesliving in thehostcountry isoneexample ofasamplenotprone to outmigration.
If attrition
which is not duetooutmigrationisbelieved toberelated to observable factors which induce sufficient variation in the attritionprocess across individuals, nonparametric identification oftheeconomic parametersof outmigrantsrequires someexclusion restriction.
Assumption A2 alo (X) = 0:10 (Xi) 2 0where X = (XH, X6)'
Proposition2.2. IfAZ holdsandthereexists a X; given Xi such that Pr (ru = 11Xl, XS) = 0,
16 Chapter 2. Identification of Outmigration proof· Fora given X = (X;, X4):a10 (X) = alo (Xl) fromA2.Using Pr (ru = 11Xi, X;) =
0, the subsampleof immigrants
(X;,Xi')'
identifies 0: 10 (Xl ) from (2.5).Given c:10 (Xl) is identified, Pr (ru = 1IX)isidentified from(2.6),which implies thattheweights Wo (X) and Wo (X) are bothnonparametricallyidentified. 0Partlybecauseattritionandoutmigration are very differentprocesses,finding
real-isticexclusion restrictions whichsatisfy the requirementsof Proposition 2 is not very
restrictive. Usually outmigrationismodelled asalife-cycle event,influenced by poor
labor marketperformance,integrationfeelings, credit rationing in thehomecountry and ageatimmigration. Whether
attrition
which is not duetooutmigrationisrelatedtoallthesefactorsseemsa
priori
unlikely, given that part ofthesurvey non-response isgenerally based on respondents refusal to continue working with thesurveyagencies.
2.3
Parametric
model
and
estimation
method
In thissection, we develop and estimateaparametric model whichallows us to extract
outmigrationbehaviorfrompanelattrition.The choice ofaparametricmodel is
moti-vated byourdesiretomodelselection into workandoutmigration asadecisionbased
on
individual
specific unobserved heterogeneity. We are not aware ofany existingsemiparametric techniquewhich would allow us toestimatethesystemofequations presented below.
We have a measure of N immigrants in period 1, whereimmigrant i remains in the panel forTi periods. Foreach
immigrant i,
we observe in period t,whether hepvorks pit, hismonthly labormarketearnings e(wit), and his attrition
status r;, in the
next period. The log ofthepotentiallabormarketearningsisassumed tobegenerated
by aloglinear earnings equation
Wit= 4,0 + Vil +
ElI (2.Dwhere Bareunknownparameters, Yii isan unobservedtimeinvariant individual spe-cificcomponentofincomewhileElIrepresentsastochasticshock. Theselabor market
earnings areonlyobservable whenan immigrant works. The
work
decision pit is as-sumed tobegenerated byalatentprocessSection2.3.Parametricmodelandestimationmethod 17 where 0 are unknown parameters, 11,2 is an unobserved
component of work and 4
representssomestochasticshock to the
work
propensity. Participationisdeterminedby
theobservation rule pit = 1 [pi > 0]. When pit = 1, earnings wit are observed.Both and Zl can
bethought of capturing immigrantsunobservedability
to generatehigherearnings and to find jobs. They can alsobethought ofasincludingunobserved
family background characteristics and preferences for workand leisure. Finally, an immigrant'sunobservableoutmigration propensity r is assumed tobedetermined by
another latentprocess
ri = s;,7 + 4,3 +
2
(2.9)where 7 are unknownparameters, 4,3 captures the
individual
specific attachment to hisnative country and 4 isa stochastic shock. Thetriplet
{,li, 111·'13} isassumed to be observed by theimmigrant who takes it intoaccountwhen makinghis decisionsbut it is notobserved by the econometrician. Let
rM = 1 [ri > 0] be
thedecision rule governing thetrueoutmigrationdecisionin
period t +1.Outmigration r;;isunobserv-able. Inour empirical application, weassume thatAlholds6andexpresstheattrition probability as
Pr (4 = lisit) = a1O + [1 - Kio]I
Pr (r# =
lIsit)
(2.10)Equation (2.10) is the samplecounter part ofequation (2.5).7
The earnings, work andoutmigrationoutcomes are notlikely tobeindependent of each other. This will notbeindependent if, forexample, immigrants who find work very easily and/or who earn a highincome are more reluctant to outmigrate. The
unobserved heterogeneity components7,1,11,2 and 4,3 canbetreated eitherasfixed
con-stants orasrandomvariables. Themainadvantage of the fixed effect approach is that
it doesnot requirethatincluded explanatoryvariablesbestrictlyexogenous to the un-observed heterogeneity components ('1 , '1 ,'1 )· However, estimationof fixedeffects
in
nonlinearmodels remains todaya sizeablecomplication, with very little guidancein
the choiceofmodels (see the recentreviewof
Arellanoand Honord, 2001).A
seconddrawback
of
fixed effectestimation is that by treatingthe unobserved heterogeneity componentsasfixed,crossequation correlationswhich drive selection into work and outmigrationbasedon unobservableindividual
characteristics arenotidentified. Asthe present chapterismainlyconcerned
with
selectionissues, fixedeffectestimation would limit our insights in the type ofselections present in the data. Wethereforein-troducethesedependenciesbyassuming that thestochastictime-invarianteffects are
covari-18 Chapter2.Identifcation of Outmigration ance matrix 4 p:.25,5, PY,35" 5,
0 = · 4 P2420,20,3
2. 993
-where
4
denotesthe variances of the unobserved heterogeneity components, and p5·denotestheircorrelations.8 Thesecorrelationsareindicativeof whether or not
immi-grantsself-selectthemselves into work andinto outmigrationbasedontheir unobserv-able
individual
characteristics.A significantandpositivep .2indicatesthat individualswho are more likely to work are also morelikely tohavehigherearnings,giveobserved
characteristics. p 1 has,- asimilar interpretation and isindicative of outmigration bias.
This coefficient willbenegative(positive)
if
immigrants who haveahigherprobabil-ity of
outmigratinghavebelow (above)averagemonthlyearnings. Finally, PL can beinterpretedasmeasuring outmigration bias in the
work
decision and whose sign has asimilar interpretation.
Finally, we assume that the
vector [el, 4,4]' is i.i.d
normally distributed withmean 0and covariancematrix
-2 E- -2 aw Pl,2"W pl,3gw
E= · 1 pb
1
where qi isthe variance oflogearnings,
while
the variances of the unobserved stochas-ticshocksentering the workandoutmigrationequations are set to 1for identificationpurposes. Contemporaneous correlations between the three stochastic components are
captured bythecorrelation coefficientsP ,2 Pf,3 and p6J·
To simplifythepresentation of thelikelihood function, wedividethe observable characteristics of immigranti into a set Yi = {Pit, 4, wit · Pit } il of dependent
vari-ables, a set Xi = {xit, zit, sit } ii ofexogenous variables, andavector vi = (qil, '1 , 'li)
containingunobservedtimeinvariantheterogeneity. Moreover, we denoteby g(·, ·. · lili)
Section2.3. Parametricmodel andestimationmethod 19 firststepdensityisgiven by
fc Cy,Ixi,qi; B. 8,7, r, x1O)
Tirff
= Il / / 4 (1- rlt) (1- trio) g(p;„riA, wit;El,li) dr;,
t=i ./Qi, ./Ci,
l
J -00CO fo
+41 f g(PA, rii, wit:ElI/i)0 J -00dr;t.+-0:10 11 g (PA, r,$, wit; Elqi) dr;, dp;; drvit
The casewhereoutmigration isperfectlyobserved follows bysettinga10 equal to 0.
The sets it anditdefine thedomainofintegration over the wage and
work
spaces andvary over timeas individualsmakedifferentchoices ineachperiod according to the
following
tableIntegrationdomainsin period t Qi: Cit
Work - [0.co)
Not Work (-co, 00) (-co,0]
Income isintegrated out in waveswhere individuals do not work. The integration
domain for the
work
propensity follows from the work decision rule. In thesecondstep, theunconditionallikelihood functionis obtainedby integrating outtherandom
individual
effects over R3f (yilxi; 4, 0,7,li, Cl,o:10) = f f (yiIX,·,1/i; 4,0,7, E, alo) h (qi; fl) dui JR3
where H denotes the trivariate normal cumulativedistribution function with mean
vector 0 and covariancematrix O.
To solvethenumerical integrationproblem, weapproximate theintegral by a sim-ulated mean:asequence of r = 1,2,..., Ri.i.d.
draws 11)) = (11,1('), 42('),119(')) is taken
fromthemultivariate normal distribution H atagivenvalue of 0.9 For each draw, the
conditional likelihood
function fc
isevaluated. ThepartialMSLestimatorconsists ofreplacing f bythesimulated mean
Ell, log R El, fc (Y,IXi,111'); 4,8,7,E,ir,o)11
Theresultingestimatoris inconsistentforfixed R but willbeconsistent ifRtends to
20 Chapter2.
Identifcation
ofOutmigration likelihood(Train,2003). Finally, wemodellabormarketearnings andwork
decisionsofGermans usingsimilarspecifications
of
equations (2.7) and (2.8) and estimate theparameters using thesimulationtechniques described above.
2.4 Data
The data used inthis chapteristaken fromthepublic use file oftheGSOEPandcovers
the 1985-1999period. Until 1990,the GSOEPconsisted of two samples, A and B.
Sam-ple A consistsofhouseholds
with
Germanheadsliving
in formerWest-Germany Sam-ple B consists ofansampleof immigrantsliving in
West-Germany coming from coun-trieswhichhad signedabilateral migrationagreementwith
Germany inthe 1950s and1960snamelyGreece,Italy, Spain, Turkey and Yugoslavia.10 Data on speaking fluency,
integrationfeelingsof immigrants, intendedlength of stay and remittances directed to
their
family living
outside Germany were givenin
consecutive waves from1984until
1987. Starting in 1987, thisinformationwas gathered every other yean
In
order to keep constant the timeperiodbetween observations, we have chosen to keep the8 wavesofthe panelwhere detailedinformation onimmigrantswas available,eachspanned
by oneyear,starting in 1985 andending in1999. Following theliteraturemeasuring
the economic assimilation rate(e.g. Borjas,1999a),werestrictour attentiontomales
between 18 and64years ofageduringthe1985-1999period. Excluded from the
sam-pleareindividuals whodiedduringtheobservationperiodandindividuals who gave incompleteinformation onany singlevariable enteringtheempiricalmodel in any of
the 8 waves.This leaves us withasample of1987nativeGermans and732immigrants
starting in 1985.
Theidentificationapproach presentedinsection2.2relies on theinformation
con-tainedinpanelattrition. Itbecomesinstructiveto contrast theattrition pattern of our immigrantsample with thatofGermanswhose
attrition
cannotobviouslybeattributedtooutmigration. Table2.1containsinformation onthenumberofindividualsobserved
along withthepercentage oftheoriginal1985samplewhoremains inagiven wave.11
41.9%ofGermans and 26.7%of immigrants havebeeninterviewedsuccessfully in all
the waves.Theattrition rate inagiven waveisdefined as the percentage ofindividuals not observed inthe given wavebutobserved in thepreceding wave. Over our
Section 2.4. Data 21
notan issue. Assuming thatthedifferencein attritionrates is due tooutmigration, a
back of the envelopecalculation implies thatwewouldexpect theoutmigration rate in
our sampleof immigrants to be6%every two years, or 3% per year,anumberwhich would be in line
with
thosereported intheliterature(seeBorjas andBratsberg, 1996).Of course, thiscalculationrelies onthe assumptionthatattrition intheimmigrant pop-ulationwhich is not dueto outmigration is ofcomparablemagnitude to that
of
natives. We will come back to thisin
section2.5whichreportsindirectevidence suggesting thatthisshouldindeed hold
in
Germany.The top panelof Figure2.1 shows theaverage monthly gross income for work-ing immigrants andGermans over the periodcovered. In 1985, the meanincome of
Germans was 3,357 DM permonthcompared to 2,690 DMpermonth for immigrants, givinganincomeratio of1.25favoringGermans. The mean wagedifferentialremaind
relativelysteadyuntil 1991, afterwhich, the meanincomedifferential widened even more between the twogroups toreach aratio of 1.34 in 1999,
with
Germansreceiv-inganaveragemonthly wage of 5,848DM while immigrantswere receiving 4,348 DM
per month. The divergingeconomic progressofGermans andimmigrantsafter 1991 is alsoreflected in theworkfrequencies. Thebottompanel of figure 2.1 shows the sample frequenciesofindividuals working inthemonthpreceding theinterview. We
can see that
until
1991, laborforceparticipation was verysimilar for
both Germansandimmigrants. After 1991, we observeasteady decline in the
work
frequencies forbothgroups. During that period, the percentage of immigrants
working
remainedsteadily below that ofGermans. The severe drop in
work
frequencies forbothgroupscoincides withthe generaldeterioration ofthelabormarket in the regionsof former
West-Germany. Table 2.3gives the unemployment rate per yearby state. With the
exceptionofBerlin,all provinces experiencedtheirlowestunemployment rate of the
1985-1999period in1991. After 1991,theunemployment ratehasprogressively risen
apart fromaslight fall in 1999 formost provinces.
If outmigration does occur ata systematic time inthe life-cycle, itislikely to
af-fect the age and years since immigration composition of oursample of immigrants.
Table2.2gives variable descriptionsand summary statistics
for
Germans andimmi-grants for the 1985 and1999waves. We see thatbothGermansandimmigrants are, on
average, a little less than40years of age in1985while the average age ofthecohort
22 Chapter2.Identification ofOutmigration
Germansandimmigrants leftthe panel. Foragiven meanage,Germans have acquired
relatively more yearsofeducation,but relatively lower labor marketexperience than
immigrants.Theaveragenumberofmonthsof labor marketexperienceof immigrants increased byalittle less than32months compared to70monthsforGermans,which is
consistent with the fact thattheproportion of working immigrants relativeto Germans
fell dramatically inthe 1990's.
Most immigrants migrated to Germany early in their productive lives, a fact
re-flected byan average age atimmigration
of
nearly24years,afigureconsistent through-out theobservation periodwhichindicates that mostmigrants wereoldenough toau-tonomouslydecide to moveto Germany. Boththeevolutionofyearssinceimmigration and immigration yearareconsistent withthe hypothesis that outmigrationoccurs 20
years after migration (OECD, 2001). Theaverage yearof immigration of ourcohort is 1969 in the 1985 wave,butraises to 1979 in the 1999 wave,indicating thattheearlier cohorts are most susceptible to havedropped out ofthepanel. Astheearlier cohorts contain themigrants withthehigher number ofyears sincemigration in 1985, it is not surprising to seethataverageyears sinceimmigrationincreasesrelatively less than the
14 year timespan, passing from 15.75 in 1985 to 19.63 in 1999, indicatingagain that earlier cohortsare those who leftthe panel. Reported feelings onintegration in the
German society and reported speakingfluency improved over time whilehealth satis-factionseemstodeteriorate,thelatter
likely
capturinganagingeffect. Finally, 73% ofimmigrantsreportedhavingaspouselivingoutside Gerrnany in1985whileaslittle as 1% still do so in 1999. This result canbeinterpreted in twoways. First, spouses may haveeventually migratedto Germanyduring the time period. Second,
it
might be thatimmigrantswhosespousewas
living
abroad were morelikely
to outmigrate.2.5 Results
and
simulations
Theregressorsincluded inthe earnings andworkequationsareeducation, labor
mar-ket experience,labor marketexperience squared, selfreportedGermanspeaking flu-ency, and the numberofyears sinceimmigrationto Germany. These arethestandard
variables thathave appeared in thisliterature(Borjas, 1999a). Theprovincial
unem-ployment rate in each waveisadded in bothequationsto capture local labor market conditions. Finally, we include time fixed effects in each wave to captureremaining
ex-Section2.5. Resultsandsimulations 23
clusionrestriction in the
work
equation. Reported health satisfaction isavalid exclu-Sion restrictionif
health problems occur mostly at a time in which anindividual is
more likely to have found astablejob whosecontinuation depends onthe worker's
choices. Theoutmigrationequation includesasregressorswhether or not the wife of
immigrantslivesinGermany,theimmigrantsself reported feelingsofbelonging to the Germans society and theunemploymentrate. Immigrant's who arrive atayoung age
presumably have the highest incentive to investin acquiring countryspecific human capital. This effectiscapturedbyincluding age atarrival
in
Germany.Immigrants investing in their native countrypresumably havedifferentincentives
to stay in thehost country. The motives oftheseinvestments
will
affecttheir dura-tion ofstay. Immigrantsmay extendtheir stay in thehostcountryin
ordertofurtherincrease theirinvestments in their nativecountry. On the other hand,
if
immigrantsinvestin starting upabusiness they wish tomanage-asproposedby Dustmann and Kirchkamp (2002), high levels of investments willbeassociated
with
shortermigra-tion duramigra-tions. Todisentanglebothhypothesis,we include intheoutmigration equa-tionthecumulativeamountofmoneyreturned tothenative countrysince 1984 as a proxy forinvestments. Reported health satisfaction and self-reported expected length
of stayinGermany arealsoincluded, thelater captures anticipatory behavior of
mi-grants,which havebeenshown to affect the acquisitionofcountry specific skills (e.g. Dustmann, 2002b). Time dummies areadded to capture remainingmacroeconomic fluctuations.
In ordertoseparate theimpact ofselection onearnings into a workand
outmigra-tion effect, we firstestimatedan earningsequation
with
randomeffects. Oursecondspecification is abivariatemodelof labormarket earnings and work. We finally es-timated the completemodelor earnings, workandoutmigration. In the latter case,
we experimented withanalternativespecification of theoutmigrationequationwhich contained education,labor market experience and its square, speaking fluency, and
years sinceimmigrationasregressors. A log-likelihoodratio test of the
null
hypothesisthatthesehumancapital variables have nojointeffectonoutmigration could not be
24 Chapter2. Identification ofOutmigration
2.5.1
Equation
results
Covariancestructure
We beginour analysis of theresults with a discussion ofthe estimates
characteriz-ingthecovariance structure ofthe unobserved componentswhichareinformative of
theselection mechanisms. Table 2.4presents estimates of the covariance structure. Focusing on the most generalmodelwhich controls for both work and outmigration selection, we findasmallbut significant positive correlationbetween theunobserved
individualheterogeneity of work and earnings(pY.2),indicatingthatindividuals with higher probabilities ofworking are also morelikely tohavehigherearnings. The
cor-relationbetween
individual
time invariantheterogeneityof outmigration and earn-ings (p ,1 is\ W/ -56%while that between outmigration and work <p24 3) is 49.8%, both.1
significant at the 1% level. Bothcorrelations suggest that individuals with a higher
propensitytooutmigratearethose with bothalower probability of finding work, and alower labor marketearnings,which pointsto negativeoutmigrationselection. When
comparing results withthebivariate model which does not correct
for
outmigration, we find thattheestimatedvalue ofP ,2remainsstable.ResultsforGermansaresimilar to that oftheimmigrant sample, withasmallbut positiveandsignificantwork
selec-tion effect(ph)
Finally, transitoryshocksbetweenearnings and work,and shocks between workandoutmigration, areall significantly negativelycorrelated,theformerat -34.2% and thelatter at30.4%while we do not find significantcorrelation between thetransitoryshocks of the earnings andoutmigrationprocesses.
Outmigration
Table 2.5presents theparameter estimates of theoutmigration equation. We find that
immigrantswhosewife lives with them
in
Germany haveasignificantly lowerprob-ability
of outmigrating,reflectingapreferencefor family unity. Immigrantssatisfiedwith
theirhealthare significantlylesslikely
tooutmigrate, a finding consistent withthe sociologicalfindingsreportedinStark(1998).Intended length ofstay captures the expectations ofimmigrantsand offersdirectinformation on their remigration inten-tions.Not surprisingly, we findthatmigrants whoexpect to remain longer