Testing with the TestersWhat can we learn from an \audit study"?Journees Louis-Andre Gerard-Varet #8Romain Aeberhardt Denis FougereJulien Pouget Roland RathelotCREST-INSEE15 juin 2009Aeberhardt, Fougere, Pouget and Rathelot () Testing 15 juin 2009 1 / 28Plan1 Introduction2 The data3 Usual statistics { linear, parametric framework4 Conditional logit { non-linear, parametric framework5 Semi-parametric framework (NEW)6 Real interaction with the employer, statistical discriminationAeberhardt, Fougere, Pouget and Rathelot () Testing 15 juin 2009 2 / 28IntroductionPlan1 Introduction2 The data3 Usual statistics { linear, parametric framework4 Conditional logit { non-linear, parametric framework5 Semi-parametric framework (NEW)6 Real interaction with the employer, statistical discriminationAeberhardt, Fougere, Pouget and Rathelot () Testing 15 juin 2009 3 / 28Introduction2 main types of studies to measure discriminationDiscrimination on the labor market arises whenever an employer treatsdi erently, on average, individuals whose productive characteristics areequal but whose non productive characteristics such as gender or racedi er.Methods a la Blinder-Oaxaca: decomposition of a gap between anexplained part and an unexplained part.the unexplained part is not necessarily discrimination.Controlled experiments: \audit studies" / \testing"\C teris paribus ": \testers" are matched to be similarWith 2 \testers" (or more) for each job o er, it ...
Real interaction with the employer, statistical discrimination
20093/28
Introduction
1
2
The data
3
Usual statistics – linear, parametric framework
4
Ttitse51gnniujugPoanetatdRlohe
e,Pog`erandRugetebhreAF,uoratdtingTeslotathe82
Discrimination on the labor market arises whenever an employertreats differently, on average, individuals whoseproductive characteristics are equalbut whose non productive characteristics such as gender or race differ. Methodsla`aBlinder-Oaxaca of a gap between an: decomposition explained part and an unexplained part. the unexplained part is not necessarily discrimination. Controlled experiments:“audit studies”/ “testi ” ng “Cæteris paribus are matched to be similar”: “testers” With 2 “testers” (or more) for each job offer, it is possible to control for unobserved heterogeneity at the firm level.
Discrimination on the labor market arises whenever an employertreats differently, on average, individuals whoseproductive characteristics are equalbut whose non productive characteristics such as gender or race differ. Methodsa`alBlinder-Oaxaca: decomposition of a gap between an explained part and an unexplained part. the unexplained part is not necessarily discrimination. Controlled experiments:“audit studies”/ “testing” “Cæteris parib are”: “testers” matched to be similar us With 2 “testers” (or more) for each job offer, it is possible to control for unobserved heterogeneity at the firm level.
Usual statistics: focus on implicit hypotheses (strong) meaning in terms of discrimination (not so clear) Less usual results: Information carried by the observations in terms of “discrimination” – unobserved heterogeneity at the firm level– Non-linear framework –“Conditional Logit” Semi-parametric set-identificationNEW–only the sign of the effect is identified and not its magnitude Use of the specificity of the data: Testers alternate the role of the potentially discriminated candidate Taking into account tester’s unobserved heterogeneity (Heck ’ man s critique)
Usual statistics: focus on implicit hypotheses (strong) meaning in terms of discrimination (not so clear) Less usual results: Information carried by the observations in terms of “discrimination” – unobserved heterogeneity at the firm level– Non-linear framework –“Conditional Logit” Semi-parametric set-identificationNEW–only the sign of the effect is identified and not its magnitude Use of the specificity of the data: Testers alternate the role of the potentially discriminated candidate Taking into account tester’s unobserved heterogeneity (Heckman s ’ critique)
Usual statistics: focus on implicit hypotheses (strong) meaning in terms of discrimination (not so clear) Less usual results: Information carried by the observations in terms of “discrimination” – unobserved heterogeneity at the firm level– Non-linear framework –“Conditional Logit” Semi-parametric set-identificationNEW–only the sign of the effect is identified and not its magnitude Use of the specificity of the data: Testers alternate the role of the potentially discriminated candidate Taking into account tester’s unobserved heterogeneity (Heckman’s critique)
4 black “testers” and 3 white “testers” Students at the University of Milwaukee Black and White testers applied to different sets of jobs The black “testers” went to 200 interviews and the white ones to 150 Low-qualified jobs which require no more than a high-school level No competition between the “testers” Each “tester” played alternatively the role of the ex-offender or of the reference candidate This will allow us to take into account “tester effects and to answer ” Heckman’s critique Evidence of statistical discrimination using the fact that there was not always a real interview with the employer