Rstanarm, QR and priors

I think it’s not clear what happens if horseshoe prior is used with QR decomposition. The users don’t necessarily notice that the prior is set on QR-space. Maybe rstanarm should give a warning if QR=TRUE and a prior is defined?

Aki

1 Like

That is why we can’t default QR to TRUE. There is a default prior, even though it is usually not specified explicitly. You mean warn if prior is not NULL? I agree that it doesn’t necessarily make sense to shrink the coefficients on Q to zero so strongly.

Yes. Just a gentle reminder that the prior is on Q-space and not on the original covariates.

I have one example where HS in QR-space seems to produce something useful, but it’s not clear whether it would work in general, and thus I think the warning would be useful.

Aki