Null hypothesis Bayes factor estimates can be biased in (some) common factorial designs: A simulation study

Bayes factor null hypothesis tests provide a viable alternative to frequentist measures of evidence quantification. Bayes factors for realistic interesting models cannot be calculated exactly, but have to be estimated, which involves approximations to complex integrals. Crucially, the accuracy of these estimates, i.e., whether an estimated Bayes factor corresponds to the true Bayes factor, is unknown, and may depend on data, prior, and likelihood. We have recently developed a novel statistical procedure, namely simulation-based calibration (SBC) for Bayes factors, to test for a given analysis, whether the computed Bayes factors are accurate. Here, we use SBC for Bayes factors to test for some common cognitive designs, whether Bayes factors are estimated accurately. We use the bridgesampling/brms packages as well as the BayesFactor package in R. We find that Bayes factor estimates are accurate and exhibit only little bias in Latin square designs with (a) random effects for subjects only and (b) for crossed random effects for subjects and items, but a single fixed-factor. However, Bayes factor estimates turn out biased and liberal in a 2x2 design with crossed random effects for subjects and items. These results suggest that researchers should test for their individual analysis, whether Bayes factor estimates are accurate. Moreover, future research is needed to determine the boundary conditions under which Bayes factor estimates are accurate or biased, as well as software development to improve estimation accuracy.