Choice of the Metric for Effect Size in Meta-analysis

Abstract

Meta-analysis, as a procedure for integrating the results of empirical studies, depends on the quantification of the results of individual investigations. The standardized mean difference in performance between treatment and control conditions has been conventionally used for this purpose. There are difficulties with this technique when group standard deviations are not homogeneous, when a control condition is not included in a particular study, or when no control condition exists. There are also difficulties in expressing effect sizes on a common metric when some studies use transformed scales, such as gain scores, to express group differences, or use factorial designs or covariance adjustments to obtain a reduced error term. This paper discusses these problems, proposes a common metric on which effect sizes may be standardized, and describes procedures for computing appropriate effect sizes for all such cases.