The ASA Statement on P-values and Statistical Significance was an important step forward in clarifying the issues involved in using P-values, what they can and cannot do, and situations in which their use may be inappropriate. However, the statement was fairly vague on what researchers *should* do in these latter settings. Bayesian methods for hypothesis testing, including summaries like posterior odds, Bayes factors, and penalized likelihood criteria like the Deviance Information Criterion (DIC) and its variants, have long been recommended as possible alternatives here, but this approach has its own baggage and problems to address. Even in settings where noninformative priors can be identified, various approaches may still yield different answers, and the approach does not comport well with the point-null hypothesis setting with which most investigators are most familiar and comfortable. The availability of standardized software and problems wrought through uncareful use of (often MCMC-driven) Bayesian software raise further practical challenges. In this session, we will hear from 3 experts who have worked to develop properly-calibrated Bayesian methods that could be acceptable for routine use by statisticians interested in the long-run frequency behavior of any testing procedure.
