Random or Fixed? An Empirical Examination of Meta-Analysis Model Choices

When conducting meta-analyses, researchers must make decisions about which statistical model is most appropriate for the specific context and aims of the meta-analysis. Although there are several meta-analysis models, most researchers choose between two general models: fixed-effect (FE) and random-effects (RE). Yet, the basis on which these two general models are distinguished and of when it is appropriate to use one or the other varies in the methodological literature. Although model-to-inference inconsistencies have been previously noted, there has been little empirical investigation of whether, and to what extent, the varying conceptualizations of the distinctions between FE and RE models are reflected in published meta-analyses. The present study explores whether conceptualizations of model distinctions among psychological researchers are consistent with those in the methods literature. We also examine model choices and rationales given by psychological researchers in two samples of published meta-analyses in psychology-related journals. We identify four primary categories for distinguishing between FE and RE models, only two of which were predominant in our samples. Although model choice appears to be reported at a moderately high rate, many researchers continue not to provide explicit rationales for their model choices or do not clearly tie model choices to the specific research aims of the meta-analyses. Implications of these findings are discussed.