Inquiries about variable selection using projpred

[kendyteng]

Thank you very much, Aki!

@avehtari

Can you show how this is defined? It’s not necessary for answering your later questions, but helps to check if there might be other problems.

Yes, here it is.

f <- as.formula(paste("obs ~", 
                       paste(names(D.spldf.st)[9:31], collapse = "+"), 
                       "+offset(log(exp))"))
M.pois.hs <- brm(formula = f, data = D.spldf.st, 
                 family = poisson(), seed = 19871002,
                 set_prior(horseshoe(df = 1.5, par_ratio = 0.1)),
                 control = list(adapt_delta = 0.999999, max_treedepth = 12))

@avehtari

Yes, these look bad. Can you post also plots with deltas = F? This would show also how bad the reference model is.

Here are the graphs. I wonder how you judge how bad the reference model is?

@avehtari

I would say that cv_varsel is the regular one. varsel can be misleading in difficult cases.

Ok, I see. In this case, should I include any of the covariates? The covariates are actually from different data source. I don’t expect it to fit very well. The purpose is more to build a BYM2 model, but of course I was hoping that they could explain something.