@avehtari pointed out that projpred does not complain when using a (brms) reference model with a truncated response distribution (truncated normal in his case). At the very least, I think projpred should throw a warning for such reference models, stating that the truncation is not taken into account when performing the projection. As a longer term objective, we could check whether it is possible to take the truncation into account.
@avehtari pointed out that projpred does not complain when using a (brms) reference model with a truncated response distribution (truncated normal in his case). At the very least, I think projpred should throw a warning for such reference models, stating that the truncation is not taken into account when performing the projection. As a longer term objective, we could check whether it is possible to take the truncation into account.