[R] Convergence problem in GLMM

[John Fox]

Dear Deepayan,

Thanks for the explanation (and for pointing out that the data set is also
in the lme4 package). I had tried control=lmeControl(tolerance=1e-4,
PQLmaxIt=100) and control=lmeControl(PQLmaxIt=1000), but gave up at that
point. I just tried lmeControl(tolerance=1e-4, PQLmaxIt=1000), and that
works. I guess I just was insufficiently persistent.

Regards,
 John

--------------------------------
John Fox
Department of Sociology
McMaster University
Hamilton, Ontario
Canada L8S 4M4
905-525-9140x23604
http://socserv.mcmaster.ca/jfox 
-------------------------------- 

> -----Original Message-----
> From: Deepayan Sarkar [mailto:deepayan at stat.wisc.edu] 
> Sent: Tuesday, November 23, 2004 5:10 PM
> To: r-help at stat.math.ethz.ch
> Cc: John Fox; 'R-Help'
> Subject: Re: [R] Convergence problem in GLMM
> 
> On Tuesday 23 November 2004 11:14, John Fox wrote:
> > Dear list members,
> >
> > In re-running with GLMM() from the lme4 package a 
> generalized-linear 
> > mixed model that I had previously fit with glmmPQL() from MASS, I'm 
> > getting a warning of a convergence failure, even when I set 
> the method 
> > argument of
> >
> > GLMM() to "PQL":
> > > bang.mod.1 <- glmmPQL(contraception ~ as.factor(children) 
> + cage + 
> > > urban,
> >
> > +     random=~as.factor(children) + cage + urban|district,
> > +     family=binomial, data=Bangladesh)  # works!
> >
> > > bang.mod.1 <- GLMM(contraception ~ as.factor(children) + cage + 
> > > urban,
> >
> > +     random=~as.factor(children) + cage + urban|district,
> > +     family=binomial, data=Bangladesh, method="PQL")
> > Warning message:
> > IRLS iterations for glmm did not converge in: GLMM(formula = 
> > contraception ~ as.factor(children) + cage + urban,
> 
> This dataset is also in lme4 as 'Contraception'.
> 
> > Despite the indicated convergence failure, the two sets of 
> estimates 
> > are quite close, as one would expect. I've also played around with 
> > various arguments to lmeControl(), but without success.
> >
> > I'm curious about the source of the difference, since (as I 
> understand 
> > it), essentially the same algorithm is used by the two 
> functions. I'm 
> > using R 2.0.1 under Windows XP with the current version of lme4.
> 
> Probably because the convergence criterion is different. glmmPQL has 
> 
>         if (sum((eta - etaold)^2) < 1e-06 * sum(eta^2)) 
>             break
> 
> GLMM has 
> 
>          crit <- max(abs(eta - etaold)) / (0.1 + max(abs(eta)))
>          ## use this to determine convergence
>          if (crit < controlvals$tolerance) {
>               conv <- TRUE
>               break
>          }
> 
> I'm not sure why we chose that. Setting 
> 
> control = list(tolerance = 1e-3, PQLmaxIt = 100)
> 
> converges for me, but that's probably a bit extreme.
> 
> Deepayan