LOO error with a brmsfit model

Interfaces brms
1

Hello,

Very new to this Forum, and Bayesian analyses too, but learning a lot from the discussions here!

Im trying to do a model selection with loo, based on brmsfit models, but in some of them i’m getting these kind of errors:

loo.m1.3.3_arvoreta<-loo(m1.3.3_arvoreta,reloo=TRUE)
1 problematic observation(s) found.
The model will be refit 1 times.

Fitting model 1 out of 1 (leaving out observation 2)
Start sampling
Error in if (p.rank > ncol(object$X)) p.rank ← ncol(object$X) :
argument is of length zero

An this is the model code:

prior_1.3.3<-c(prior(normal(1,1),class=Intercept),
prior(normal(0,1), class = b),
prior(normal(0,1), class = Intercept,dpar=sigma),
prior(normal(0,1), class = sds),
prior(normal(0,.1), group = area, class = sd))

form1.3.3=bf(arvoreta~s(parcela) +(1|area))+lf(sigma~1,hu~1)

m1.3.3_arvoreta<-brm(form1.3.3,file=“m1.3.3_arvoreta”,sample_prior=“yes”,control = list(adapt_delta = 0.999, max_treedepth=15),
data=data.areabasal,family=hurdle_lognormal(),iter=10000, prior=prior_1.3.3)

Does anyone have an idea of what was my mistake?

Thanks in advance !

Operating System: Windows 10 x64 (build 17134)
Interface Version: loo_2.0.0 , brms_2.4.0

2

Could you try out brms 2.5.0 from CRAN and if that doesn’t work the dev version from github (https://github.com/paul-buerkner).

I believe this problem should be fixed already in at least the dev version.

3

It sounds from the error message like when you leave some observation out, it makes the design matrix rank deficient. If so, then almost by definition, the k-hat is going to be too large for that observation.

4

Thanks a lot for the quick answer!
It’s working normally now, with the brms 2.5.0.