Ben Goodrich writes:
We just released rstanarm 2.15.3, which fixed a major bug that was introduced back in January with the 2.14.1 release where models of the form
stan_glmer(y ~ ... + (1 | group1) + (1 | group2), family = binomial())would produce WRONG RESULTS. This only applies to Bernoulli models with multiple group-specific intercept shifts and no group-specific deviations in the coefficients. If you have estimated any such models in 2017 with rstanarm, you should re-estimate them. The other changes to rstanarm were minor.
If you are using OSX, it may be necessary to install rstanarm and / or rstan from source to work around the “c++ exception (unknown reason)” error that there are several inconclusive threads about.
The latest version of rstan is 2.15.1, released a couple weeks ago.
Another variation of a Bernoulli model that will be wrong is: y ~ … + (1 | group1/group2), family = binomial())
I’ve received several questions about this since the post, so to clarify (or really just restate what Ben wrote): this bug only affected models fit using version >= 2.14.1, and only affected Binomial models with exactly 1 trial in each row in the data (i.e. _Bernoulli_ models only). It didn’t affect other Binomial models or earlier versions of rstanarm and it’s fixed in the latest CRAN release.
I tried to install this in R 3.2.4 for MSWin; and got a result:
stanarm_2.15.3.zip’ is not available (for R version 3.2.4)
since the Depends for the package mentions only “R (≥ 3.0.2)” I was wondering if there was a problem here.
Try
install.packages(“https://cran.r-project.org/bin/windows/contrib/3.3/rstanarm_2.15.3.zip”, repos = NULL)