The impact of using different random-effects models in meta-analysis: a case study on the cognitive functioning of...

Random-effects meta-analysis typically assumes a normal distribution for the underlying effects of the studies. However, under certain conditions, this assumption may not be the most appropriate. Simulation studies comparing several alternative random-effects models have found small differences in bias but larger differences in coverage probabilities and precision. To investigate the impact of using different meta-analysis models, we used a meta-analysis of 58 cohort studies comparing the cognitive functioning between preterm- and term-born children. We compared the results between seven different random-effects models based on: asymmetric distributions, mixtures of distributions and Dirichlet process (DP) prior. Sensitivity analysis using variations of the seven main models was also performed. While estimates for mean treatment effect were similar across models, the between-study variance estimates were identical. Asymmetric distribution models indicated studies with extreme effects and a left-shifted random-effects distribution while mixture models were less informative. DP models revealed clusters of studies with similar characteristics, suggesting that the observed heterogeneity may be partially explained by certain methodological and clinical characteristics of the studies. Overall, our study highlights that applying various meta-analysis models might not affect materially the summary estimates but may provide better insights into the underlying effects' distribution and the explanation of between-study variance.