On the posterior predictive in Bayesian failure analysis