El coste de la dicotomización en meta-análisis

Abstract

Dichotomizing continuous outcomes implies a negative bias (underestimation) in the estimation of an effect size. New effect-size indices have rccently been proposed which correct such a bias. However, when these new indices are applicd on a primary study, their variability (uncertainty) around the population effect size is larger than that of the traditional uncorrected indices. Only when the sample size in the studies is extremely large or when the studies are included in a meta-analysis and the effect sizes are averaged, the unbiased indices are also more efficient than the uncorrected ones. The purpose of this work was to explore the influence of different meta-analytic conditions in the bias and efficiency of two effect-size indices applied on dichotomized outcomes and averaged through the studies in the meta-analysis: one practically unbiased ( JP,ohu ) and the other uncorrected ( cf,, ). The mínimum number of studies in the meta-analysis necessary to get that the unbiased index performs better than the uncorrected one is presented as a function of both the sample size in the studies and the magnitud e of t he population effect size.