Online Program


Multi-treatment meta-analysis for categorical outcomes
 
 
*Christopher Schmid, Tufts Medical Center 
 

Keywords:

Meta-analytic outcomes with three or more mutually exclusive categorical responses are typically analyzed with one or more binomial models by collapsing categories or comparing different paired categories. For example, a meta-analysis comparing two breast cancer treatments might combine cancer deaths and non-cancer deaths into a total mortality outcome or it might compare death from cancer to other outcomes. Besides requiring multiple analyses, such methods may introduce bias and inefficiency by analyzing only part of the data and ignoring the correlation among responses induced by the underlying multinomial distribution. One complicating feature of the data when combining across studies is that not all of the outcomes will be reported in each study. The missing data may arise from the collapsing of some categories in some studies or from a focus on a subset of the outcomes. Further missing data are introduced when the meta-analysis is expanded to multiple treatments because different studies will now present different treatment combinations as well as different outcome combinations. This talk presents a Bayesian model for such data and applies it to analyze the effect of different statin treatments on outcomes of fatal and non-fatal stroke, myocardial infarction and other causes of mortality.