Bayes factor null hypothesis tests are still null hypothesis tests
Many researchers are aware that null hypothesis significance tests (NHST) have significant problems. One alternative strategy for simple research questions is the use of Bayes Factor null hypothesis tests. I will argue that while Bayes Factor null hypothesis tests are an improvement over NHST, they are unsuitable as a default analytic strategy in the social sciences. Specifically, Bayes Factor null hypothesis tests are only useful for producing conclusions about the posterior probability of hypotheses if our prior knowledge takes a very specific form: a “spike and slab” distribution, with a substantial (ideally 50%) probability that the parameter is exactly zero, and the rest of the prior density spread over a wide range of values. Such a prior is both highly informative and also rather unlikely to be a reasonable representation of the actual prior knowledge held about a particular effect or relationship in the social sciences. I argue that we need to stop worrying about testing the null hypothesis, which is almost always false. Instead, Bayesian estimation is a more useful default analytic strategy—perhaps with a particular focus on the probability that a particular parameter is positive or negative. However, to allow Bayesian estimation to become widely used by everyday researchers, we as methodologists need to produce sensible informative default priors, and implement them in easy-to-use software. I will produce some suggestions for how this might be accomplished.