Estimation of between-trial variance in sequential meta-analyses: A simulation study

Meta-analysis of Results of Testosterone Therapy on Sexual Function Based on International Index of Erectile Function Scores

“…The present and previous [11,13] positive results are in apparent contrast to those reported by Tserstvadze et al [38], who did not document any significant effect of TTh on erection, either alone or when added to PDE5i. It should be recognized that Tserstvadze et al [38] used the DerSimonian and Laird [39] method for the calculation of the effect size, which may underestimate the true between-study variance, potentially producing overly narrow confidence intervals for the mean effect [40], especially when the between-study variance is large [41]. In addition, Tserstvadze et al [38] analyzed only nine RCTs enrolling mixed eugonadal/ hypogonadal patients, which may have resulted in a possible inclusion bias.…”

Small studies are more heterogeneous than large ones: a meta-meta-analysis

“…The high occurrence of zero estimates may be caused by the fact that many of the selected meta-analyses were based on a few small studies and that sample variances for small studies are large. When studies have equal sample variances s 2 , the t 2 estimator corresponds to the sum of (y i ey) 2 /(k-1) minus s 2 [21,25], or zero if the result is negative, where k is the number of studies. Consequently, for imprecise studies with large s 2 , the estimate of t 2 will often be zero, even when the true heterogeneity is substantial.…”

“…We used the empirical Bayes estimator for t 2 , equivalent [24] to the robust [25] Paule and Mandel [25e27] estimator and more accurate than the widely used Method of Moments estimator of DerSimonian and Laird (DL) [21,28], which is based on large sample assumptions [27]. Furthermore, the relative bias is similar for small to large heterogeneity, as opposed to the DL and restricted maximum likelihood estimators, that underestimate t 2 more as the heterogeneity increases, especially for dichotomous outcomes [25,28]. We estimated t 2 for all meta-analyses, even when the authors originally performed a fixed-effects analysis.…”

Meta-analysis approach as a gene selection method in class prediction: does it improve model performance? A case study in acute myeloid leukemia

“…In this study, we combined the corrected standardized effect size for each gene by random effects models, similar to a study conducted by Choi et al [10]. However, we estimated the between-study variance by Paule-Mandel method, which outperforms the DerSimonian-Laird method in continuous outcome data [28]. We used a broad selection of classification functions to build predictive models in order to evaluate the added value of meta-analysis in aggregating information from gene expression across studies.…”

“…(1) is the Hedges’ g correction for SMD [26]. The estimation of between-study variance was performed by Paule-Mandel (PM) method [27] as suggested by [28, 29]…”