This project consists of four components which continue from last year: The first compares a fixed and a random effects model for combining data from a series of clinical trials. The study evaluates the performance of each method under various heterogeneity assumptions and with several scales of measurement, including the risk difference and the odds ratio as measures of effect. The second develops and compares parametric and non-parametric tests to assess the assumption of homogeneity of effects in data from a series of trials. This component considers data from Gaussian, binomial, and Poisson sampling distributions. The third considers meta-analytic methods for combining the evidence from a series of HIV seroprevalence studies to estimate HIV prevalence in a target population. Sentinel studies are typically investigations of incompletely defined cohorts which are convenient to survey but are often based on non-probability samples. Self-selection and similar issues inherent in sentinel studies makes generalizing results from a single study to a target population problematic. This research assesses the use of formal meta-analytic methods for HIV prevalence estimation in a target population by incorporating information from several HIV sentinel seroprevalence studies. The fourth component addresses issues that pertain to the use of meta- analysis in the design and monitoring of clinical trials. This research evaluates the role of formal incorporation of external evidence summarized from a meta-analysis of previous or concurrent results into sample size considerations and stopping rules during the conduct of a clinical trial.
