Abstract
Meta-analyses typically quantify heterogeneity of results, thus providing information about the consistency of the investigated effect across studies. Numerous heterogeneity estimators have been devised. Past evaluations of their performance typically presumed lack of bias in the set of studies being meta-analysed, which is often unrealistic. The present study used computer simulations to evaluate five heterogeneity estimators under a range of research conditions broadly representative of meta-analyses in psychology, with the aim to assess the impact of biases in sets of primary studies on estimates of both mean effect size and heterogeneity in meta-analyses of continuous outcome measures. To this end, six orthogonal design factors were manipulated: Strength of publication bias; 1-tailed vs. 2-tailed publication bias; prevalence of p-hacking; true heterogeneity of the effect studied; true average size of the studied effect; and number of studies per meta-analysis. Our results showed that biases in sets of primary studies caused much greater problems for the estimation of effect size than for the estimation of heterogeneity. For the latter, estimation bias remained small or moderate under most circumstances. Effect size estimations remained virtually unaffected by the choice of heterogeneity estimator. For heterogeneity estimates, however, relevant differences emerged. For unbiased primary studies, the REML estimator and (to a lesser extent) the Paule-Mandel performed well in terms of bias and variance. In biased sets of primary studies however, the Paule-Mandel estimator performed poorly, whereas the DerSimonian-Laird estimator and (to a slightly lesser extent) the REML estimator performed well. The complexity of results notwithstanding, we suggest that the REML estimator remains a good choice for meta-analyses of continuous outcome measures across varied circumstances.