[R-sig-ME] zero mean fixed effects priors in a cloglog link model

Philip Harrison pharrison at uwaterloo.ca
Fri Apr 13 21:04:06 CEST 2018

Hi List,

I have a presence-absence fish telemetry dataset that has a very high number of zero's (12,732 absences and 1120 presences). So I want fit a cloglog link GLMM, which seems to work better than a logit link.

However, I also have some quasi-complete seperation- that is I have all zeros
 in one of my factor level predictions.

So ideally I would like to try and fit zero mean normal priors for my fixed
effects levels using Ben Bolker's logit link
example https://ms.mcmaster.ca/~bolker/R/misc/foxchapter/bolker_chap.html
using bglmer.

so the logit model looks like this:
cmod_blme_L2 <- bglmer(predation~ttt+(1|block),data=newdat,
                       fixef.prior = normal(cov = diag(9,4)))

where ttt is a categorical variable four levels
the priors provide 4 � 4 diagonal matrix with diagonal
elements equal to 9, for variances of 9 or
standard deviations of 3.

So with the logit link above
-2sd will be -6 and +2SD would be +6 ...so pretty weak.

 inv.logit(-6)<- 0.002472623
 inv.logit(6)<- 0.9975274

Can I use the same priors for cloglog link?
invcloglog(-6) <- 0.002475683
  invcloglog(6)<- 1

My gut says yes it should work they are both weak and the issues are with the lower end.  The model runs nicely and gives sensible estimates.
However I figured it would be more correct to have 2xsd be -6 and +2 but I dont know how to code such a prior. I would consider using MCMCglmm too if it allowed me to fit those priors.

Any help would be much appreciated

Philip Harrison PhD
Post-Doc in Cooke and Power Labs
Department of Biology
University of Waterloo
200 University Avenue West
Waterloo, Ontario, Canada
N2L 3G1

Researchgate: http://tinyurl.com/RG-PMH
Google Scholar: http://tinyurl.com/ScholarPMH
Lab Website: http://www.fecpl.ca/people/philip-harrison/
Personal Website: https://pharriso4.wixsite.com/philipmharrison

	[[alternative HTML version deleted]]

More information about the R-sig-mixed-models mailing list