[R-sig-ME] A mixed effect regression tool based on bayesian priors like bayesglm
bakaburg1 at gmail.com
Wed Aug 9 17:43:53 CEST 2017
I actually did that after suggestion by Gelman itself!
Sorry I didn't upgrade the thread...
2017-08-09 17:41 GMT+02:00 Phillip Alday <phillip.alday at mpi.nl>:
> Sidestepping your question a bit, you might want to consider the
> rstanarm package instead of blme because it is more directly in the
> Gelman/arm tradition/school of thought. You can move your arm::bayesglm
> models to rstanarm::stan_glm and then use rstanarm:stan_glmer for the
> mixed model. But only *you* can know / justify which priors you should
> use -- default priors are just that, default and they may not always be
> sensible for every case.
> On 05/17/2017 10:46 AM, Angelo D'Ambrosio wrote:
> > I use extensively Gelman's bayesglm for the every day use due to the
> > stability of the estimates especially in the case of separation.
> > I needed an equivalent of empirical bayesian regularization for glm mixed
> > effect models. These models are strongly influenced by extreme conditions
> > (like conditional probabilities of zero and separation) and like usual
> > logistic regression model they fail in these cases.
> > I found the blme package that does exactly what I need, solving the
> > separation problem. Now the problem is to set it up in order to be work
> > exactly as bayesglm, in order to achieve consistency in my analysis.
> > Reading Gelman paper on bayesglm() I understood I should use a t
> > distribution with 1 df (eg. Cauchy) and 2.5 scale, rescaling inputs:
> > bglmer(Out ~ arm::rescale(Pred) + (1 | PatientID), family = binomial,
> > Data.events, fixef.prior = t(df = 1, scale = 2.5))
> > Is it correct? My doubt is what to do with the cov.prior parameter;
> > I leave it as default (wishart) or should I put it to NULL? Also in
> > Gelman's paper it is said that the intercept should have the same prior
> > distribution but with scale 10, and I don't know how to specify a
> > prior for it.
> > Also, I'm starting to think that bayesglm doesn't rescale the inputs
> > directly but scales the prior distribution according to the inputs. Am I
> > right?
> > Can you help me with this?
> > Thanks
> > [[alternative HTML version deleted]]
> > _______________________________________________
> > R-sig-mixed-models at r-project.org mailing list
> > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
[[alternative HTML version deleted]]
More information about the R-sig-mixed-models