Abstract:
This paper studies the performance of the FAT-PET-PEESE (FPP) procedure, a commonly employed approach for addressing publication bias in the economics and business meta-analysis literature. The FPP procedure is generally used for three purposes: (i) to test whether a sample of estimates suffers from publication bias, (ii) to test whether the estimates indicate that the effect of interest is statistically different from zero, and (iii) to obtain an estimate of the mean true effect. Our findings indicate that the FPP procedure performs well in the basic but unrealistic environment of “fixed effects,” where all estimates are assumed to derive from a single population value and sampling error is the only reason for why studies produce different estimates. However, when we study its performance in more realistic data environments, where there is heterogeneity in the population effects across and within studies, the FPP procedure becomes unreliable for the first two purposes, and is less efficient than other estimators when estimating overall mean effect. Further, hypothesis tests about the mean true effect are frequently unreliable. We corroborate our findings by recreating the simulation framework of Stanley and Doucouliagos (2017) and repeat our tests using their framework.
