The empirical Bayes normal means (EBNM) model plays an important role in both theoretical and applied statistics. Applications include meta-analysis and shrinkage estimation; wavelet denoising; multiple testing and false discovery rate estimation; and empirical Bayes matrix factorization. As such, several software packages have been developed that fit this model under different prior assumptions. Each package naturally has a different interface and outputs, which complicates comparison of results for different prior families. Further, there are some notable gaps in the software - for example, implementations for simple normal and point-normal priors are absent. Motivated by these issues, we developed the R package ebnm, which provides a unified interface for efficiently solving the EBNM problem using a wide variety of prior families, both parametric and non-parametric. Where practical we leverage core fitting procedures from existing packages, writing wrappers to create a unified interface; in other cases, we implement new core fitting procedures ourselves, with a careful focus on both speed and robustness. The result is a convenient and comprehensive package for solving the EBNM problem under a wide range of prior assumptions.
View on arXiv