The aim of the quantitative analysis of social programmes is to assess the causal impact of the programme on a certain outcome variable. However, this causal impact is never observed in reality. Hence, the fundamental problem of every programme evaluation is that one would like to compare the labour market outcome of a programme participant with the labour market outcome of the same participant, if he or she had not participated in the programme. The latter, however, is never observable. The fundamental problem is solved if the participants (treatment group) of a programme and the non-participants (control group) do not differ systematically in their characteristics and behaviour before the programme starts, i.e. in the absence of a sample selection bias. Such a situation is typically built up by a social experiment. In principle, the social experiment is viewed as the best solution of the evaluation problem, however experiments are very expensive (see Heckman and Hotz, 1989) and e.g. in Germany social experiments are not allowed because of ethical reasons. Moreover, Heckman and Smith (1995) show that social experiments also face several selection problems. Since in reality a sample selection bias is present for most social programmes either due to self-selection or due to programme-selection processes a naive comparison of participants and nonparticipants would not yield the causal impact of the programme. Hence, as an alternative to the experimental method a non-experimental (econometric) procedure is applied to address the sample selection bias. In principle, the econometric procedures assemble the control group ex post out of the pool of non-participants and aim to estimate a hypothetical value of the outcome variable for the treatment group on the basis of that control group.
KeywordsFirm Level Matching Procedure Labour Market Outcome Average Treatment Effect Full Information Maximum Likelihood
Unable to display preview. Download preview PDF.
- 33.The following discussion heavily draws on Heckman and Hotz (1989).Google Scholar
- 34.See Heckman and Robb (1985), Heckman and Hotz (1989), and Heckman et al. (1999). The experimental method is not further discussed here.Google Scholar
- 35.The linear control function estimator is applied in Chapter 6.Google Scholar
- 36.As the instrumental variable estimator has no practical importance for the programme evaluation in Germany, because of a lack of adequate instruments, this procedure will not be further discussed. For an extensive overview on the instrumental variable esti-mator see Heckman and Robb (1985), Heckman et al. (1999), Imbens and Angrist (1994)or Angrist et al. (1996).Google Scholar
- 37.Another class of models allowing for selection on unobservables are panel estimators, e.g. the fixed effects estimator of the linear model (see e.g. Heckman and Hotz, 1989 or Heckman et al., 1999). The main pre-condition for the usage of these estimators is the availability of longitudinal data.Google Scholar
- 38.Recent studies mostly prefer the full information maximum likelihood (FIML) esti-mator instead of the two-step estimator. For a discussion of the properties of these two estimators see Section 8.1.Google Scholar
- 39.The simultaneous estimation is described in Section 8.1. For a derivation of the Heckman selection model see e.g. Greene (1997).Google Scholar
- 40.For a more detailed description see Ronning (1991).Google Scholar
- 41.Lalive et al. (2000) present a method of including selection on unobservables within duration models. However, their application on continuous time duration models is not appropriate to models in discrete time.Google Scholar