projpred Question concerning statement in GLMM vignette

The GLMM vignette says:

For non-Gaussian models the source of the inaccuracy may come from the fact that we are running an approximate projection. In this case, increasing the number of draws for the prediction ndraws_pred should also help.

Does "approximate" here refer to the fact that the defaults of ndraws and nclusters (as well as ndraws_pred and nclusters_pred) do not imply a draw-by-draw projection, meaning that not all posterior draws are projected or that they are clustered before they are projected?

Sep 22 '21 11:09 fweber144

Ah, now that I read this preprint, could it be that you were talking of what is mentioned in section 3.2 of that preprint, i.e., the fact that the PIRLS algorithm might have convergence issues for non-Gaussian families (also shown/discussed in sections 5.4 and 5.5)?

Oct 01 '21 19:10 fweber144

Closing due to inactivity. (In #237, I made the adjustments in the vignette by guessing what was probably meant.)

May 25 '23 06:05 fweber144