Projpred: Non-Convergence of Predictive Performance with Reference Model

General
,
12

Hi @fweber144 ,

Quick update.cI have upgraded to the University’s new HPC cluster, which has R 4.3 installed. I installed the most recent version of (master branch from GitHub) Interestingly when I a the model with search_terms=NULL and I don’t set nterms_max, I get this error:

Error in eigen(sigma, symmetric = TRUE) :
  infinite or missing values in 'x'
Calls: cv_varsel ... repair_re -> repair_re.merMod -> t -> <Anonymous> -> eigen
In addition: Warning message:
In parse_args_varsel(refmodel = refmodel, method = method, refit_prj = refit_prj,  :
  In case of the Gaussian family (also in case of the latent projection) and multilevel terms, the projection onto the full model can be instable and even lead to an error, see GitHub issue #323. If you experience this and may refrain from the projection onto the full model, set `nterms_max` to the number of predictor terms in the full model minus 1 (possibly accounting for submodel sizes skipped by custom `search_terms`).

I appreciate the added warning message, it wasn’t there in 2.6.0. I then try to run and I set nterms_max to N-1 as instructed in the warning message. I keep search_terms to null. It now runs successfully, but again still the non-convergence issue:

image
image813×514 17 KB

A quick question. My deadline for this analysis is getting slightly shorter (about 2 weeks). Just to be clear, I want you to spend any extra time on this, and I am not putting any pressure! I just want to know whether I should plan to use something else, or whether you think it will be possible to fix this convergence issue soon enough?

Again thanks so much for all the help!