The selection of fixed- or random-effect models in recent published meta-analyses

Article type
Authors
Kuan Y1, Tam K1
1Shaung-Ho Hospital, Taipei Medical University, Taiwan
Abstract
Background: Most meta-analyses are based on one of two statistical models. A fixed-effect meta-analysis assumes all studies are estimating the same (fixed) treatment effect, whereas a random-effects meta-analysis allows for differences in the treatment effect from study to study. The selection of a model must be based on the question of which model fits the distribution of effect sizes, and takes account of the relevant sources of error.
Objectives: The study aims to evaluate the preference and selection of statistical model in recent published meta-analyses.
Methods: The published meta-analyses were extracted from PubMed searches before 18 March 2015. We retrieved 60 studies and investigated their selection of statistical models.
Results: Six of these 60 studies did not report whether fixed- or random-effects was used for meta-analysis. Random-effects pooling model were conducted in 27 meta-analyses. Both fixed- and random-effect models were used simultaneously in five studies. In another 22 studies, a fixed- or random-effect model was chosen according to the heterogeneity. For example, studies with an I2 statistic of > 50% were considered to have substantial heterogeneity, and therefore, a random-effects model analysis was used. Otherwise, a fixed-effect model was initially employed in the analysis. Interestingly, 21 of the 60 meta-analyses were reported from China, and 15 of them selected fixed- or random-effect models according to the heterogeneity.
Conclusions: The fixed-effect model starts with the assumption that the true effect size is the same in all studies. However, in many systematic reviews this assumption is implausible. Therefore, when studies are gathered from the published literature, the random-effects model is generally a more plausible match. Unfortunately, 36.7 % of our included studies starting their analysis with a fixed-effect model and then moved to a random-effects model if the test for heterogeneity was significant. The above strategy is thought to be a mistake, and should be discouraged.