Department of Biostatistics Seminar/Workshop Series

A Bayesian Missing Data Framework for Combining Multiple Outcomes in Mixed Treatment Comparisons using Aggregate and Individual Patient Data

Bradley Carlin, PhD

Professor, Division Head, Division of Biostatistics, University of Minnesota School of Public Health

Bayesian methods have been shown to be useful in network meta-analysis (NMA), as they facilitate borrowing of strength across treatments, trials, and outcomes, as well as provide a natural framework for filling in missing data values that respect the underlying correlation structure in the data. In this talk, we first describe the aggregate data framework for multiple bivariate outcomes (say, continuous efficacy and safety measures), indicating how such models enable ranking of the treatments following their implementation in BUGS, the most popular Bayesian software package. We then move on to models that incorporate individual patient data, perhaps available only on a subset of the trials in the NMA. Both contrast- and arm-based models will be considered, and both in the presence of covariate adjustment. We indicate the improved performance of our methods via simulation, and illustrate their application with real datasets obtained from the literature and private industry studies. We close by indicating areas still in need of further research, including detection of and remedies for evidence inconsistency, and the further broadening of the range of external information that may one day be incorporated into NMAs, including expert opinion, user-reported observations (say, from handheld devices), and other unstructured 'big data' historically thought of as unsuitable for inclusion in rigorous scientific investigations.

This work is joint with Drs. Hwanhee Hong and Haitao Chu of the University of Minnesota, and Drs. Karen Lynn Price and Haoda Fu of Eli Lilly and Company