Overlap between treatment and control distributions as an effect size measure in experiments.

ABSTRACT

The proportion π of treatment group observations that exceed the control group mean has been proposed as an effect size measure for experiments that randomly assign independent units into 2 groups. We give the exact distribution of a simple estimator of π based on the standardized mean difference and use it to study the small sample bias of this estimator. We also give the minimum variance unbiased estimator of π under 2 models, one in which the variance of the mean difference is known and one in which the variance is unknown. We show how to use the relation between the standardized mean difference and the overlap measure to compute confidence intervals for π and show that these results can be used to obtain unbiased estimators, large sample variances, and confidence intervals for 3 related effect size measures based on the overlap. Finally, we show how the effect size π can be used in a meta-analysis.