A Monte Carlo Analysis of Alternative Meta-Analysis Estimators in the Presence of Publication Bias

A meta-analysis (MA) aggregates estimated effects from many studies to calculate a single, overall effect. There is no one, generally accepted procedure for how to do this. Several estimators are commonly used, though little is known about their relative performance. A complication arises when the sample of published studies is subject to sample selection due to “publication bias.” This study uses Monte Carlo simulations to investigate the performance of five different MA estimators in the presence of publication bias. The author considers two kinds of publication bias: publication bias directed against statistically insignificant estimates, and publication bias directed against wrong-signed estimates. The experiments simulate two data environments. In the Random Effects environment, each study produces only one estimate and the true effect differs across studies. In the Panel Random Effects environment, each study produces multiple estimates, and the true effect differs both within and across studies. The simulations produce a number of findings that challenge results from previous research.