
155Extensions
This example illustrates that there is a trade-off between having an additional param-
eter, and therefore possibly more accurate predictions, and the loss of parsimony asso-
ciated with the addition of each new parameter. If there is strong theory or mounting
empirical evidence that suggests a different pattern for a cumulative logit, then relaxing
the parallel assumption adds utility to our analyses. Otherwise, there is no real harm in
disregarding a modest improvement in model t, which may well reect peculiarities
in the sample. In adjudicating between the parallel and nonparallel Bayesian models,
the popular Bayesian model t ...