Model choice is a fundamental and much discussed activity in the analy
sis of datasets. Nonnested hierarchical models introducing random effe
cts may not be handled by classical methods. Bayesian approaches using
predictive distributions can be used though the formal solution, whic
h includes Bayes factors as a special case, can be criticised. We prop
ose a predictive criterion where the goal is good prediction of a repl
icate of the observed data but tempered by fidelity to the observed va
lues. We obtain this criterion by minimising posterior loss for a give
n model and then, for-models under consideration, selecting the one wh
ich minimises this criterion. For a broad range of losses, the criteri
on emerges as a form partitioned into a goodness-of-fit term and a pen
alty term. We illustrate its performance with an application to a larg
e dataset involving residential property transactions.