Mediation, identification, and plausibility: an illustration using children's mental health services

Analyses of mediation are important for understanding the effects of mental health services and treatments. The most common approach is to add potential mediators as regressors and to estimate the direct and indirect effects of the treatment of interest. This practice makes the strong assumption that the mediator itself does not suffer from unobserved confounding--that it is as if randomly assigned. In many instances, this assumption seems rather implausible. The objective of this article is to describe the identification problem that represents the fundamental challenge of causal inference. It outlines how mediation complicates identification and considers several identification strategies. The goal of this article is not to propose a new method for handling mediation or to identify a best method for doing so. The latter, in fact, is impossible. The contribution of the article is to illustrate how one can think about possible approaches to mediation in the context of a specific empirical study.

Using data from a large evaluation of a demonstration project in children's mental health services (n = 763), the article illustrates identification strategies. That demonstration improved service delivery in several ways but primarily by offering services "intermediate" between inpatient and outpatient. These analyses focus on the impact of these intermediate services on 6-month improvement in a behavior checklist commonly used to measure psychopathology and competence among children and youths.

The results highlight how different identification strategies produce different answers to key questions. These alternative findings have to be assessed in light of substantive knowledge of the program involved.

The analyses generally support the notion that children and youths treated at the demonstration site who received intermediate services benefited from them.