Quantifying discrimination on the labour market is difficult on the sole basis of usual statistical sources, either surveys or administrative sources. The main problem stems from the impossibility of controlling for all variables that affect access to employment. Audit studies by pairs therefore appear as an interesting substitute to these standard statistical approaches. But the interpretation of their results also raises methodological difficulties. In particular, a critique made by Heckman (1998) is that it is generally impossible to take into account individual heterogeneity between testers that is observed by employers but not observed by the designer of the audit study. This paper reviews the various econometric techniques that have been proposed in the literature devoted to the interpretation of testing results. We illustrate these methods with a reexploitation of data collected in the USA by Pager (2003) concerning the consequences of a criminal record on the chances of getting a job, in interaction with skin colour. The advantage of these data is that the same testers alternatively played the role of the reference person and of the potentially discriminated person. We examine how far this allows answering Heckmans critique. In particular, we show that in a semi-parametric framework à la Manski (1987) it is still possible to take into account unobserved heterogeneity at the tester level, but then, only the sign of the effect is identifiable and not its magnitude.
