Meta-analysis in sport and exercise research: Review, recent developments, and recommendations

Abstract

The purpose of this article is to provide a general overview of the principles and practice of conducting quantitative psychometric meta-analytic reviews in the sport and exercise sciences and highlight some of the recent developments and recommendations from researchers regarding the conduct and validity of meta-analytic methods. After outlining the historical context, the general principles involved in a quantitative cumulation of research findings across empirical studies is reviewed. Subsequently, recent controversies and issues surrounding the use of meta-analysis are reviewed with examples provided from the sport and exercise psychology literature. Specifically, the basis for and selection of meta-analytic models (use of fixed vs. random effects models), the treatment of data from theories that explicitly demand testing the effects of multiple independent variables on a dependent variable (use of multiple regression), and how to treat studies that contain multiple tests of a given effect (use of averaging and structural equation modeling methods) are covered. Recommendations are provided for researchers conducting meta-analytic studies based on these issues.

Keywords:

• Quantitative cumulative research
• research synthesis
• psychology
• effect size

Notes

¹It is important to note that while weighting the individual effect size by the sample size i s widely advocated, recent evidence suggests that using the inverse variance weight is preferable because it takes into account the variability of individual scores within the primary studies (see Lipsey & Wilson, 1996 for a more detailed discussion).

²It is important to note that the Hunter and Schmidt (1990) approach, a random effects model of meta-analysis, generally performs well in simulation studies with respect to the Type I error rate in heterogenous cases (i.e. when the population effect size varies across studies) [Field, 2001]. However, it seems that the Hunter and Schmidt method is too liberal (i.e. more null results are found to be significant) in sets of studies in which the population effect size is homogenous (i.e. the same population effect size across studies).