I am trying to run projpred to reduce the number of terms in my model, but when I plot the resulting features, feature #2 results in an increase in rmsd and decrease in elpd, whilst feature #3 has a larger impact on these measures. Horseshoe prior with five suggested variables.

(stan_glm(as.formula(formula), data = d.1, prior=rhs_prior, QR=F, seed=123, refresh=0,family=binomial)