R help - Jan 2010 - glmer (lme4), glmmPQL (MASS) and xtmepoisson (Stata)

Dear R users,
 
I'm trying to specify a generalized linear mixed model in R, basically a
Poisson model to describe monthly series of counts in different regions.
My aim is to fit subject-specific curves, modelling a non-linear trend for each
region through random effects for linear splines components (see Durban et al, 
Stat Med 2005, or " Semiparametric regression" by Ruppert et al,
2003).
 
I use the command 'glmmPQL' in the MASS package and replicated the
analysis with Stata's 'xtmepoisson'.
I obtained very different results, so I would like to try 'glmer' in the
lme4 package.
I guess the default correlation for the random effects in  'glmer' is
unstructured, but this choice is absolutely unfeasible for this complex random
effect nesting structure.
Unfortunately, I couldn't find a way to input simpler correlation structures
(namely diagonal or identity), in the same way as the using the functions pdDiag
or pdIdent with 'glmmPQL'.
 
I wonder if this option is still to be implemented in lme4.
In this case, any suggestion/comment?
 
Thanks for your time

Antonio Gasparrini
Public and Environmental Health Research Unit (PEHRU)
London School of Hygiene & Tropical Medicine
Keppel Street, London WC1E 7HT, UK
Office: 0044 (0)20 79272406 - Mobile: 0044 (0)79 64925523
Skype contact: a.gasparrini
http://www.lshtm.ac.uk/people/gasparrini.antonio ( http://www.lshtm.ac.uk/pehru/
)

<Antonio.Gasparrini <at> lshtm.ac.uk> writes:
> I'm trying to specify a generalized linear mixed model in R, 
> basically a Poisson model to describe monthly
> series of counts in different regions.
> My aim is to fit subject-specific curves, 
> modelling a non-linear trend for each region through random
> effects for linear splines components (see Durban et al,  
> Stat Med 2005, or " Semiparametric regression"
> by Ruppert et al, 2003).
> 
> I use the command 'glmmPQL' in the MASS package and 
> replicated the analysis with Stata's 'xtmepoisson'.
> I obtained very different results, 
> so I would like to try 'glmer' in the lme4 package.
> I guess the default correlation for the random effects 
> in  'glmer' is unstructured, but this choice is
> absolutely unfeasible for this complex random effect nesting structure.
> Unfortunately, I couldn't find a way to input simpler correlation
> structures (namely diagonal or
> identity), in the same way as the using the functions
> pdDiag or pdIdent with 'glmmPQL'.
> 
  1. I would suggest continuing this conversation on the
r-sig-mixed-models at lists.r-project.org , which is specialized
for this kind of question.

  2. I don't think that a specific replacements for pdDiag/pdIdent
are on their way any time soon, but you can 
get a diagonal structure for the random effects: see p. 16 of

 (broken URL so gmane doesn't complain about long lines ...)

http://lme4.r-forge.r-project.org/slides/
  2009-07-21-Seewiesen/5LongitudinalD.pdf
 
for an example.

  Ben Bolker