Assessing covariate imbalance in meta-analysis studies

The main goal of meta-analysis is to combine data across studies or data sets to obtain summary estimates. In this paper, the novelty is to propose a statistical tool to assess a possible covariate imbalance in baseline variables to investigate similarity of trials. We conducted the detection of the covariate imbalance, \ufb01rst, through some graphical comparison of the empirical cumulative distribution functions or ECDFs, which are built by putting together arms or trials according to some risk factor, and second, through some non-parametric tests such as the Kolmogorov\u2013Smirnov and the Anderson\u2013Darling tests. To overcome the huge presence of ties, we conducted the statistical tests on perturbed versions of the original data sets. The applications concern two real meta-analyses of RCTs: the \ufb01rst one, on interferon-alpha treatment of chronic hepatitis C, with 107 studies involved, and the second one, on cholesterol-lowering treatment with statins, with 14 studies involved. The applications allow for analysis of both when risk factors re\ufb02ecting demographic or clinical differences in experimental and control arms are balanced or not and when there are structural differences between the levels of some study variables, in order to proceed eventually with the pooling of the studies. We developed our suggestion, which is a quantitative way to assess combinability in meta-analysis, only with respect to RCTs, but it could be applied to a minor extent to non-RCTs. Copyright \ua9 2011 John Wiley & Sons, Ltd