I am working on a model with a Beta likelihood and varying intercepts with a non-centered parameterization, which I intend to use to extrapolate back in time, rather than forward prediction.
The posterior predictives look good, relatively narrow credible intervals and the median estimates align nicely with the observed data. But, almost all of the Pareto k estimates are over 0.7 and all WAIC values are over 0.4.
If I remove the varying intercepts, the credible intervals on the predictions increase a lot, essentially filling the response scale (0 : 0.8) and the median estimates are way off for a number of the groups. But, in this case both PSIS-LOO and WAIC suggest the model fits much better.
Is this more likely a model problem or an interpretation problem?