Bayesian knockoff filter selects features with lower FDR on synthetic and real data

Researchers have described the Bayesian knockoff filter (BKF), a method for selecting important features from a large set of candidates while controlling the proportion of false discoveries. The work was published in Proceedings of the National Academy of Sciences pnas.orgProceedings of the National Academy of Sciencespnas.orgOpen source ↗. The method builds the knockoff procedure into a Bayesian framework.

The BKF relies on a tight upper bound for the posterior probability of the null hypothesis. According to the authors, the bound is derived from a Bayesian "flip-sign" property implied by the procedure.

Standard knockoff methods in the frequentist paradigm usually generate the knockoff variables once, before the model is fitted. In the BKF, they are generated at every iteration of the Markov chain Monte Carlo procedure, which removes the variability that comes from a single generation. Compared with multiple knockoffs, the authors describe the BKF as more principled in using side information through prior distributions in a fully Bayesian setting.

The authors also developed an e-value version of the BKF that theoretically guarantees frequentist false discovery rate (FDR) control pnas.orgProceedings of the National Academy of Sciencespnas.orgOpen source ↗. In numerical experiments on synthetic and real data, the BKF showed comparable power, a lower FDR and more stable selected sets than existing knockoff methods and Bayesian variable selection approaches, the authors reported.