[R-sig-ME] Tackling various errors in a mixed-effects Poisson model

[Ken Beath]

5 random effects is a lot to be fitting, and is probably causing the
convergence problems. I would start with just a random effect for the
intercept and then add in the others one by one and look at the effect on
AIC.

On 15 January 2015 at 07:12, Eric Lofgren <lofgrene at vbi.vt.edu> wrote:

> (Some of this content cross-posted to either StackExchange or
> CrossValidated)
>
> Hey folks,
>
> I’m currently working on an interrupted time-series project, and have run
> into some implementation issues I was hoping folks might be able to help me
> through. The data I’m working with has about ~1200 observations, with the
> observations split evenly between 17 sites for a total of 72
> observations/site.
>
> At the moment, I’m trying to estimate a single rate as a response to an
> indicator variable indicating a change in policy, the time up until the
> policy, the time after the policy, a binary cofounder and a continuous
> confounder.
>
> If I was just doing a single site, the model would look something like:
>
> mod <- glm(Outcome ~ Exposure + t_before + t_after + Var1 + Var2 +
> offset((log(PersonTime)), family=Poisson, data=data)
>
> Since we’ve got multiple sites, I’ve moved to a mixed-effects model that
> looks like so:
>
> mod <- glmer(Outcome ~ Exposure + t_before + t_after + Var1 + Var2 +
> (t_before + t_after + Var1 + Var2|Site) + offset((log(PersonTime)),
> family=Poisson, data=data)
>
> I’ve also added control=glmerControl(optimizer="bobyqa",optCtrl =
> list(maxfun = 500000)) because of some convergence issues in a test model.
> This brings me to the two major issues I’m having:
>
> 1. In some of the models, lme4 is generating this error "Error:
> (maxstephalfit) PIRLS step-halvings failed to reduce deviance in
> pwrssUpdate” in the presence of some, but not all offsets. I can’t see a
> clear difference between the offsets that are triggering this and the
> offset that isn’t (save that the triggering offsets are smaller, but
> they’re both reasonably normal when log transformed).
>
> 2. Generally, these models are having some serious convergence issues with
> lme4. I’m reluctant to simplify the models much further, but does anyone
> have suggestions in that regard? Would an MCMC based approach work better
> (if more computationally intensively) for this kind of data? If so, how
> would you compose the model above in something like MCMCglmm?
>
> Thanks very much for your help,
>
> Eric
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>



-- 

*Ken Beath*
Lecturer
Statistics Department
MACQUARIE UNIVERSITY NSW 2109, Australia

Phone: +61 (0)2 9850 8516

Building E4A, room 526
http://stat.mq.edu.au/our_staff/staff_-_alphabetical/staff/beath,_ken/

CRICOS Provider No 00002J
This message is intended for the addressee named and may...{{dropped:9}}