Abstract
The Rosiglitazone drug has been developed for type 2 diabetes mellitus, with annual sales having exceeded two billion dollars. It is however controversial, due to alleged association with heart attacks, and it is the subject of more than 10,000 lawsuits. It is also statistically controversial, as different meta-analyses have been published with partly conflicting conclusions. The problem is that of analysing a collection of paired binomials, modelled with a common odds ratio difference parameter, for which methods are fairly routine provided the binomial data stay away from the minimum and maximum values. It is however less clear how to proceed when many of the implied 2 x 2 tables contain zeroes. I will develop a strategy for optimal inference in such situatjons and apply this to the case of the 48 published studies for the Rosiglitazone drug. I will also present alternative meta analyses based on Poisson models.