Bayesian methods can be effective for conducting meta-analyses as they allow researchers to integrate their own predictions into the analysis. However, researchers may struggle to decide which prior distribution is appropriate. To help researchers in this pursuit, we propose a data-driven method to develop prior distributions. As a proof of concept, we constructed three prior distributions based on data from 100 social psychology meta-analyses of the past 20 years. Data was collected from the journal Psychological Bulletin. The prior distributions were fit to the data using maximum likelihood estimation. We then tested the effectiveness of our prior distributions against uninformed alternatives. The data-driven prior distributions resulted in larger Bayes Factors than commonly-used default priors. We hope that our method may help future researchers who wish to conduct a Bayesian meta-analysis.
