Posterior predictive p-values: what they are and what they are not