Hierarchical Multivariate Ordinal Model Diagnostics (and tuning)

[avehtari]

PPC looks marginal distribution, while LOO-PIT looks at the conditional distributions and thus there can be a big difference in the results.

@JLC

My understanding is that the LOO PITs should be uniform for continuous data; does that apply if the the model is based on a continuous latent variable rather than a true continuous response?

For discrete data see Predictive Model Assessment for Count Data. If there are many unique values in discrete space the issue is smaller. This is something we should fix for LOO-PIT plots, but the progress has been delayed by thinking more about how to present the deviation from uniformity better than the current approach.