Projpred interpretation

[avehtari]

@henrik.eckermann

I think it probably has to do with multiple imputation.

Very likely. In your first message, you didn’t mention use of imputation.

@henrik.eckermann

I used the function brm_multiple which takes as input a list of imputed datasets.

We have not tested projpred with this. It would be useful to create a new issue at Issues · stan-dev/projpred · GitHub, so that a) we can add informational message that this is not yet supported and b) think if it is feasible to support thus.

@henrik.eckermann

But to identify relevant variables with projpred I need to run projpred multiple times using the brm function and then I will consider those variables as important that are identified by projpred with a certain consistency across the imputed datasets (not sure yet how to define that but l could report the consistency for each variable).

This is the correct solution. When running with one data, you can also get different search paths depending on the cross-validation fold and it is possible to report the probabilities of including certain variable at certain model size.