Identifying causal mechanisms in empirical economics

In empirical economics we typically aim at evaluating the causal effect of an explanatory variable (also called "treatment") on an outcome of interest. Most studies focus on the identification of the "total" treatment effect. I.e., they do not consider the possibility that the latter may come from distinct causal mechanisms due to intermediate variables (hereafter "mediators", as the analysis of their impact is often called "mediation analysis") that lie on the causal path between the treatment and the outcome of interest. In the presence of one or more mediators, the total effect can be decomposed into a direct effect (or net effect) of the treatment on the outcome of interest and an indirect effect which operates through the mediators. Such a
decomposition of causal mechanisms often provides a better understanding of the economic problem than the total effect alone and may be crucial for deriving meaningful policy conclusions. E.g., when evaluating the employment effects of a labor market program for job seekers - such as a training - we face the problem that
the latter could induce further participation in programs which themselves affect employment. Disentangling the causal mechanisms tells us whether the initial program is effective per se (net of further participation) or only jointly with further interventions which is required for the optimal design of labor market policies.
Therefore, this project makes both methodological and empirical contributions to the identification of causal mechanisms in order to uncover the black box of total effects in policy evaluation. With respect to methodology, we develop novel nonparametric identification and estimation strategies either based on instrumental variables or on "conditional independence" assumptions (i.e., conditional exogeneity given observed variables). Our strategies invoke considerably weaker functional form and distributional assumptions than those of the (predominantly parametric) literature on mediation analysis and, thus, result in more credible causal inference. To be specific, identification will either be based on distinct instruments for the treatment and the mediators or on a sequence of conditional independence assumptions which allows controlling for both pre- and post-treatment confounders of the treatment and the mediator. As a further methodological contribution, the project demonstrates the close methodological links with the dynamic treatment literature. The latter explicitly considers the dynamic evolvement of the treatments and the control variables that is
also useful to plausibly control for treatment and mediator endogeneity in the analysis of causal mechanisms. We provide a unified framework for both strands of the literature which formally shows their overlaps and reduces ambiguity in the interpretation of the identified parameters. The empirical contribution consists of two applications in the field of labor economics where the analysis of the total effect alone appears to be incomplete. The first disentangles the effectiveness of the job counseling process provided by local employment offices on the labor market success (e.g., employment and earnings) of job seekers. It investigates whether counseling affects labor market success exclusively through placement
into programs and/or sanctions (cutbacks in unemployment benefits due to non-compliance) or also has a direct effect related to the counseling style and cooperativeness of the case worker with the job seeker (net of programs and sanctions). The second contribution identifies the labor market effects of initial programs net of further participation as mentioned above. The empirical analysis will rely on rich survey and administrative data from Switzerland and Germany and the identification strategies developed in the methodological part. By showing the relative importance of the different dimensions of the counseling process, it provides the
base for a rigorous cost-benefit analysis and the effective design of policy interventions.