[R] confidence intervals with glmmPQL

[Ben Bolker]

Sally_roman <sroman <at> umassd.edu> writes:

>  Hi - I am using R version 2.13.0.  I have run several GLMMs using
> the glmmPQL function to model the proportion of fish caught in one
> net to the total caught in both nets by length.  I started with a
> polynomial regression full model with three length terms: l, l^2,
> and l^3 (l=length).  The length terms and intercept were the fixed
> effects and the random effect was a paired haul (n=18).
 
> m1<-glmmPQL(fixed=Proportion~1+Length+second+third,random=~1|Pair,
> family=binomial,data=species,verbose=T,niter=2,
> weight=(Experimental+Control))

  Why did you set niter=2?  That seems like a bad idea
(the default is 10; I can imagine increasing it if there
 are warnings that the fit hasn't converged, but I don't
 see why you would decrease it).
 
> For the majority of the models, I ended up with a constant model with no
> length effect.

This isn't quite clear: did you do some kind of stepwise model
reduction or AIC-based model selection?


> The issue I am having is with the confidence intervals that
> were calculated.  For two models the CIs are not symmetrical around the mean
> proportion from the model.  The CIs for the other constant models are
> symmetrical around the mean.  I was wondering if anyone has an idea why this
> would be or if anyone has any suggestions.   
> Thanks Sally

  There's not enough detail here to answer.  How did you get your
confidence intervals?  A reproducible example would be helpful.
If m1 is the result of a glmmPQL fit, then confint(m1) gives odd
results (because it calls confint.default, which doesn't really
know what to do with the results of the fit); intervals(m1) might
be more what you're looking for.

  Since you are fitting a binomial model, you may be looking for
confidence intervals on the linear predictor (logit) scale, in
which case they should always be symmetric around the estimate,
or on the response (probability) scale, in which case they
should always be asymmetric.

  You may want to send questions about mixed models to
the r-sig-mixed-models at r-project.org mailing list.