The meta-analytical random effects model with g tends to underestimate parameters: an alternative model

The Classic Random Effects Model (CREM) has some limitations when using the standardized mean difference (SMD) as effect size index. Suero et al. (2025) have reformulated CREM as a Mixture Model (MM) and have developed unbiased estimators of the main parameters µδ and τ2. They compared their performance with two classic methods widely used: Restricted Maximum Likelihood (REML; Viechtbauer, 2005) and DerSimonian & Laird estimator (DL; 1986), finding small but systematic underestimations yielded by the classic procedures. The aim of this study is to check if Suero et al.’s (2025) results can be found beyond simulation contexts. For this purpose, we created three databases with real meta-analyses (MA) from clinical, experimental and educations fields of psychology. The results found are consistent with those found by Suero et al. (2025) being the mean estimates of MM higher than those of REML and DL. We also discuss about outliers found in real MAs such as bizarre effect sizes, disproportionated sample sizes or MAs with small numbers of primary studies.