Model Evaluation

• It is necessary but not sufficient to write down our model assumptions: we must also test them.

• Model checking, for prediction accuracy (or "goodness of fit") is best first done visually, in the data space. Posterior predictive checks, using well-designed test statistics and discrepancy measures, can then quantify the innaccuracy.

• The next step is typically model expansion: improving the model's ability to fit the data by changing its form, perhaps adding more parameters.

• Comparing alternative models against each other can be done just with accuracy (via the check results), or with efficiency considerations as well, using the Bayesian evidence.