bayesEfron: Fully Bayesian inference with Efron's log-spline prior in R

In meta-analyses and multisite trials, we use estimates and standard errors to infer both the distribution of effects and the individual effects. Conventional random-effects models assume normally distributed effects, whereas empirical Bayes methods treat the estimated distribution as known when obtaining intervals for individual effects. We describe bayesEfron, an R package that implements a fully Bayesian version of Efron’s log-spline prior. The model represents a flexible effect distribution by smoothly varying probabilities on a grid and accommodates unequal standard errors. Inference for individual effects includes uncertainty in this distribution. The package selects the grid, provides grid and sampler diagnostics, and computes posterior summaries and plots. The estimated distribution can also be used with the ebnm package. We illustrate the methods using racial callback gaps at 97 firms and 1,500 simulated sites with known effects.