The study of the mediation of treatment effects plays a key role in advancing the knowledge and practice of treatment implementation. Efficacy or effectiveness questions address whether treatment works, while mediation questions address why it works. However, an experimenter typically cannot directly manipulate many of the plausible mediators. Thus, even in the context of a randomized trial, the mediation analyses will not be protected by the initial randomization. This will be the case even when there is an appropriate temporal order to the mediators and outcome. Traditional data analytic approaches to mediation have been based on regression, path, and structural equation models. These approaches all make the implicit assumption of sequential ignorability, which amounts to saying that patients are randomized to their observed levels of treatment, and also randomized to their levels on subsequent mediators. Hence, there is a need for a set of statistical tools that will allow for the violation of the sequential ignorability assumption, and provide a framework to assess the extent to which violations of the assumption influence the interpretation of the traditional methods. In this proposal, we will consider statistical methods that are derived from the Rubin Causal Model, and do not rely on the sequential ignorability assumption. The goal of this application is to extend the set of statistical methods that are used by alcohol treatment researchers in studies of the mediation of treatment effects. This proposal will make new statistical methods for mediation accessible to clinical researchers, and will provide software and guidelines to enable such researchers to implement, and evaluate, the methods on their own data. Using a combination of data analyses on existing NIH datasets, and simulation studies informed by those analyses and the mediation literature, we will compare the performance of traditional and causal mediation methods in the setting of a single continuous or binary outcome with a single or multiple (binary or continuous) mediators, and also when there are repeated outcomes. For each setting, we will use a combination of data analyses on a set of three case study datasets, and simulation studies to compare the implementation and performance of traditional and causal methods, and their respective sensitivities to the assumptions on which they rely. Narrative: Improving our understanding of the mechanisms by which behavioral interventions operate is important in developing new treatments. The statistical treatment of the question of mechanism uses models for mediation, where the effects of treatment on outcome are modeled using effects on intervening variables. This study investigates the performance of new statistical methods for mediation, where a more explicit causal interpretation off mediation may be possible. We compare these methods with standard approaches. ? ? ?
| Lynch, Kevin G; Van Horn, Deborah; Drapkin, Michelle et al. (2010) Moderators of response to telephone continuing care for alcoholism. Am J Health Behav 34:788-800 |
