Importance of trends in the interpretation of an overall odds ratio in the meta-analysis of clinical trials.

This paper contains a proposition related to the publication of meta-analyses of clinical trials. We consider the situation where the results of a number of trials are summarized by a common or typical odds ratio. We show that stating such an odds ratio as the summary of evidence from a number of trials can be misleading if certain systematic differences between trials exist. In such cases the author should state not just one odds ratio but also its dependence on the relevant characteristics of the trials. In particular, we propose that those reporting a meta-analysis state in advance a (limited) number of variables to be considered for potential interaction with the exposure (risk factor or treatment) of interest. The list might include centre size and the odds in the placebo or control group if such an effect is a priori clinically plausible. The trials should be ordered according to each of these variables and a trend test for the odds ratio should be computed. Apart from a 'genuine' effect, an appreciable interaction could also be indicative of the (multiplicative) odds ratio being an inappropriate measure for the particular meta-analysis. Without any consideration as to the possibility of interaction, the meta-analysis should be considered incomplete. If such an interaction exists, the odds ratio should be stated as a function of the interacting variable, either as a formula or (preferably) in a table stating the odds ratio for a number of different values of the interacting variable, and not as a single summary statistic.