Raftery argues that P-values and the tests based on them give unsatisfactory results, especially in large samples, and that standard variable-selection procedures can be very misleading when there are many candidate predictors — and, by choosing a single model, they ignore model uncertainty and so underestimate the uncertainty about quantities of interest. He presents the Bayesian approach to hypothesis testing, model selection, and model uncertainty, made practical by the simple and accurate BIC approximation computable from standard software output, with specific results for the model types common in sociology.
"It is argued that P-values and the tests based upon them give unsatisfactory results, especially in large samples … by selecting a single model, they ignore model uncertainty and so underestimate the uncertainty about quantities of interest."
The paper that carried Bayesian model selection into applied social science by making it cheap: BIC turns a Bayes factor into something you can read off any regression printout, so the argument against P-values comes with a ready replacement rather than a counsel of despair. Its three-part critique — P-values break in large samples, stepwise selection misleads, single-model inference ignores model uncertainty — is still the standard case for model averaging, and the ability to weigh evidence for a null is exactly what substantive theory (convergence, norms) needs. It is the applied companion to Kass-Raftery (1995): where that review catalogs Bayes-factor theory and computation, this one prosecutes the case and hands practitioners the BIC shortcut. The fair caveat, which the discussion of the paper pressed, is that BIC's unit-information prior is itself an assumption, and P-values retain defenders for exploratory work.