Are most published research findings false in a continuous universe?
Fig 2
PPV as a function of the prevalence of true effects, statistical power and bias for (A) Dichotomous, (B) Two Peaks, (C) Overlapping Peaks, (D) Two Normals and (E) Single Normal effect size distributions.
Each point corresponds to a simulation with 5,000 published findings. True effects are defined as those above the minimum of interest (Cohen’s d > 0.5 in all simulations). In each panel, graphs correspond to 20%, 50% and 80% power from left to right. Blue, red and green curves correspond to bias (i.e. prevalence of negative results that are repeated until becoming positive and then published) of 0, 0.2 and 0.6, respectively. Prevalence ranges vary among scenarios, as for some distributions some prevalences cannot be achieved without changing other parameters: in the Two Normals model in particular, one cannot obtain a prevalence above 62% unless either the SD of the wider normal or the minimum effect of interest is changed.