Volume 62, Issue 1 p. 97-128

Fixed- versus random-effects models in meta-analysis: Model properties and an empirical comparison of differences in results

Dr Frank L. Schmidt

Corresponding Author

Dr Frank L. Schmidt

Department of Management and Organizations, University of Iowa, Iowa City, USA

Correspondence should be addressed to Dr Frank L. Schmidt, Department of Management and Organizations, Henry B. Tippie College of Business, University of Iowa, Iowa City, IA 52242-1994, USA (e-mail: [email protected]).Search for more papers by this author
In-Sue Oh

In-Sue Oh

Department of Management and Organizations, University of Iowa, Iowa City, USA

Search for more papers by this author
Theodore L. Hayes

Theodore L. Hayes

Gallup Organization, Washington, DC, USA

Search for more papers by this author
First published: 10 January 2011
Citations: 460

Abstract

Today most conclusions about cumulative knowledge in psychology are based on meta-analysis. We first present an examination of the important statistical differences between fixed-effects (FE) and random-effects (RE) models in meta-analysis and between two different RE procedures, due to Hedges and Vevea, and to Hunter and Schmidt. The implications of these differences for the appropriate interpretation of published meta-analyses are explored by applying the two RE procedures to 68 meta-analyses from five large meta-analytic studies previously published in Psychological Bulletin. Under the assumption that the goal of research is generalizable knowledge, results indicated that the published FE confidence intervals (CIs) around mean effect sizes were on average 52% narrower than their actual width, with similar results being produced by the two RE procedures. These nominal 95% FE CIs were found to be on average 56% CIs. Because most meta-analyses in the literature use FE models, these findings suggest that the precision of meta-analysis findings in the literature has often been substantially overstated, with important consequences for research and practice.