Developing a Bayesian Statistical Approach to Behavioral Intervention Trials

Bayesian statistical methods permit the incorporation of existing knowledge formally into a statistical analysis to generate estimates that leverage both past and present information. Bayesian methods are particularly suitable for intervention research since treatments are generally evaluated over the course of multiple trials. Although Bayesian methods have seen increasing use in the evaluation of biomedical trials, there has been minimal application to date in the behavioral and psychosocial intervention literature. Moreover, Bayesian techniques have been used in an ad hoc manner, with some intervention studies leveraging existing data and others utilizing non-informative priors. This study focused on developing a framework for Bayesian analysis that addresses the practical considerations of accumulating data across heterogeneous studies. We used real data from a series of three randomized controlled trials conducted in New York City, Seattle, and Beijing that evaluated behavioral interventions for improving HIV antiretroviral adherence and mental health outcomes. In contrast to previous literature that has been limited to cross-sectional statistical models, this case study demonstrated the accumulation of data using multilevel regression techniques. In the first set of analyses, we evaluated medication adherence outcomes in the New York City and Seattle studies, the most similar with respect to design and interventions. In the second set of analyses, we evaluated depression outcomes in the full sequence of studies, which introduced the complexity of accommodating a study with substantial differences in methodology. The integration of data from multiple sources led to refined estimates of intervention effectiveness. However, differences arose in the estimates of intervention effectiveness across studies, raising substantive questions about when the decision to aggregate data may be appropriate. We discuss the role of contextual, implementation, and secular effects that may influence an aggregated analysis. This case study illustrated the substantive considerations necessary to support the decision to combine data across studies, and the need for careful review of findings to confirm the appropriateness of pooling using a Bayesian approach.