Performance of statistical methods for meta-analysis when true study effects are non-normally distributed: A simulation study.

Meta-analysis (MA) is a statistical methodology that combines the results of several independent studies considered by the analyst to be 'combinable'. The simplest approach, the fixed-effects (FE) model, assumes the true effect to be the same in all studies, while the random-effects (RE) family of models allows the true effect to vary across studies. However, all methods are only correct asymptotically, while some RE models assume that the true effects are normally distributed. In practice, MA methods are frequently applied when study numbers are small and the normality of the effect distribution unknown or unlikely. In this article, we discuss the performance of the FE approach and seven frequentist RE MA methods: DerSimonian-Laird, Q-based, maximum likelihood, profile likelihood, Biggerstaff-Tweedie, Sidik-Jonkman and Follmann-Proschan. We covered numerous scenarios by varying the MA sizes (small to moderate), the degree of heterogeneity (zero to very large) and the distribution of the effect sizes (normal, skew-normal and 'extremely' non-normal). Performance was evaluated in terms of coverage (Type I error), power (Type II error) and overall effect estimation (accuracy of point estimates and error intervals).