A simple technique investigating baseline heterogeneity helped to eliminate potential bias in meta-analyses

To perform a worked example of an approach that can be used to identify and remove potentially biased trials from meta-analyses via the analysis of baseline variables.

True randomisation produces treatment groups that differ only by chance; therefore, a meta-analysis of a baseline measurement should produce no overall difference and zero heterogeneity. A meta-analysis from the British Medical Journal, known to contain significant heterogeneity and imbalance in baseline age, was chosen. Meta-analyses of baseline variables were performed and trials systematically removed, starting with those with the largest t-statistic, until the I² measure of heterogeneity became 0%, then the outcome meta-analysis repeated with only the remaining trials as a sensitivity check.

We argue that heterogeneity in a meta-analysis of baseline variables should not exist, and therefore removing trials which contribute to heterogeneity from a meta-analysis will produce a more valid result. In our example none of the overall outcomes changed when studies contributing to heterogeneity were removed. We recommend routine use of this technique, using age and a second baseline variable predictive of outcome for the particular study chosen, to help eliminate potential bias in meta-analyses.