Combining event rates from clinical trials: comparison of Bayesian and classical methods.

OBJECTIVE

To compare an empirical Bayesian, a fully Bayesian, and a classical fixed-effect (Peto) method for pooling event rates from separate epidemiologic studies or clinical trials.

DESIGN

Four data sets used in meta-analyses by previous authors were evaluated. The first data set concerned death rates observed in clinical trials of beta-blockers, the second to lung cancer and smoking in 14 case-control studies, the third to drowsiness induced by the antihistamine compound chlorpheniramine, and the fourth to the use of intravenous magnesium in patients with suspected myocardial infarction. Randomly chosen data points were made more extreme to test the methods further.

MAIN OUTCOME MEASURES: Pooled estimates of effect expressed as odds ratios and their associated 95% confidence intervals.

RESULTS

All three methods gave comparable results with respect to the 95% confidence interval, although the Bayesian methods gave generally wider interval estimates. However, the point estimates for the individual studies were substantially different, particularly for small studies.

CONCLUSIONS

For the data sets considered, Bayesian methods, which are computer intensive but intuitively appealing, provided results that were consistent with the classic fixed-effect Peto method. Introduction of the more extreme data points did not alter this conclusion.