Abstract:
|
Standard meta-analytic random effects models assume that heterogeneity is constant across studies. In practice, however, this assumption is often not supported by data resulting in poor model fits and apparent outliers. To address this, we assume a more general model in which studies can belong to one or more heterogeneity clusters. Additional variation due to heterogeneity (i.e., the variation represented by the random effects parameter in a traditional model) is assumed to be constant within, but not across, each cluster. By identifying such clusters and finding the best model fit among possible groupings, we are better able to estimate the total uncertainty associated with each study. In cases where such extra heterogeneity is present, the result is an improved understanding of the contribution of each study and a less biased pooled estimate. In this talk, we apply this method to data from a meta-analysis aimed at estimating the prevalence of postmenopausal bleeding among women with endometrial cancer.
|