A bivariate likelihood approach for estimation of a pooled continuous effect size from a heteroscedastic meta-analysis study

TL;DR

This paper introduces a bivariate likelihood method for meta-analysis that jointly estimates effect size, heteroscedastic within-study variances, and between-study variance, improving accuracy over traditional methods especially with heteroscedastic data.

Contribution

It proposes a novel bivariate likelihood approach that accounts for heteroscedastic within-study variances in meta-analysis, enhancing estimation accuracy and robustness.

Findings

Less sensitive to the number of studies

Less biased under heteroscedasticity

Performs better than DL, HT, and higher-order likelihood methods

Abstract

The DerSimonian-Laird (DL) weighted average method has been widely used for estimation of a pooled effect size from an aggregated data meta-analysis study. It is mainly criticized for its underestimation of the standard error of the pooled effect size in the presence of heterogeneous study effect sizes. The uncertainty in the estimation of the between-study variance is not accounted for in the calculation of this standard error. Due to this negative property, many alternative estimation approaches have been proposed in literature. One approach was developed by Hardy and Thompson (HT), who implemented a profile likelihood approach instead of the moment-based approach of DL. Others have further extended the likelihood approach and proposed higher-order likelihood inferences (e.g., Bartlett-type corrections). Likelihood-based methods better address the uncertainty in estimating the between-study variance than the DL method, but all these methods assume that the within-study standard deviations are known and equal to the observed standard error of the study effect sizes. Here we treat the observed standard errors as estimators for the within-study variability and we propose a bivariate likelihood approach that jointly estimates the pooled effect size, the between-study variance, and the potentially heteroscedastic within-study variances. We study the performance of the proposed method by means of simulation, and compare it to DL, HT, and the higher-order likelihood methods. Our proposed approach appears to be less sensitive to the number of studies, and less biased in case of heteroscedasticity.