Comparison of prediction accuracies of p-value, m-value, and BF.

<p>We simulate 1,000 meta-analysis of 10 studies with varying sample sizes where only a subset of the studies have an effect. Given 10,000 studies, we threshold each statistic to predict the studies having an effect and the studies not having an effect, and vary the threshold to draw the ROC curves. In A, true prediction rate is the proportion of the studies actually having an effect that are correctly predicted to have an effect and false prediction rate is the proportion of the studies actually not having an effect that are incorrectly predicted to have an effect. In B, true and false prediction rates are similarly defined but in the direction of predicting studies not having an effect. For BF, we use the asymptotic BF of Wakefield <a href="http://www.plosgenetics.org/article/info:doi/10.1371/journal.pgen.1002555#pgen.1002555-Wakefield1" target="_blank">[39]</a> with prior where .