Uncertainty in statistics relative to the reference model(?) about projpred HOT 3 CLOSED

If your reference model is the model with 10 variables, then yes, the relative difference should collapse more or less to zero at 10 variables. However, if you run the selection with default settings, then the projection uses I think 5-cluster projection by default, which introduces some discrepancy, meaning that projecting even the reference model onto itself is not exact. If you set argument ns to the number of posterior samples in your reference model, then the projection onto the reference model should be exact (this will be computationally slower though).

from projpred.

@StaffanBetner Is your problem solved, so can we close this issue?

from projpred.

Closing this now

from projpred.