As put by Gelman et al. (2013, page 148): 'because a probability model can fail to reflect the process that generated the data in any number of ways, posterior predictive *p*-values can be computed for a variety of test quantities in order to evaluate more than one possible model failure'.