[R-sig-ME] GLMER fixed-effect estimates systematically larger than GLM/GEE?

[David Atkins]

Assaf--

The fixed-effects estimates from GLMM and GEE are not directly 
comparable as the former are conditional on the random-effects, whereas 
the latter are marginal coefficients.  See the following for some 
further chatter and citations.

https://stat.ethz.ch/pipermail/r-sig-mixed-models/2011q1/005527.html

Hope that helps.

cheers, Dave

-- 
Dave Atkins, PhD
Research Associate Professor
Department of Psychiatry and Behavioral Science
University of Washington
datkins at u.washington.edu

Center for the Study of Health and Risk Behaviors (CSHRB)		
1100 NE 45th Street, Suite 300 	
Seattle, WA  98105 	
206-616-3879 	
http://depts.washington.edu/cshrb/
(Mon-Wed)	

Center for Healthcare Improvement, for Addictions, Mental Illness,
   Medically Vulnerable Populations (CHAMMP)
325 9th Avenue, 2HH-15
Box 359911
Seattle, WA 98104
http://www.chammp.org
(Thurs)

Hi all,

I was trying logistic GLMM via glmer() on a binary outcome in a crossover
trial. One of my research questions was whether there is enough
within-person dependence that had to be accounted for, or whether I can
ignore it and pool data to a simpler analysis.

So I ran the same data on glm() and on glmer(both REML=F and REML=T). The
estimated auto-correlations are relatively mild, and the contribution of
GLMM vs. GLM is borderline significant, depending upon whether you prefer
likelihood, AIC, BIC or some other criterion.

However, I was surprised to see an unexpectedly large difference in the
*fixed* effect estimates. The glmer() estimates were all about 30%-40%
larger in magnitude.

Then I found this Q&A online:
https://stat.ethz.ch/pipermail/r-help/2008-July/166979.html

Which showed a similar discrepancy (about 15%-20% in magnitude), and also a
gee() run that agreed with glm() and not glmer().

So I tried gee() myself, and my estimates too show an gee-glm agreement with
glmer() the odd man out, having larger fixed-effect estimates.

I'm not asking you to diagnose my specific case, but rather whether you know
of this feature/problem with glmer(). Is there some hidden scaling factor in
the estimates or defaults that I am missing?

Thanks in advance, Assaf


-- 
Assaf P. Oron, Ph.D.
Research Scientist, Lead Staff Statistician

MESA Air Pollution Study and DISCOVER center, University of Washington
4225 Roosevelt Way NE Suite 303D, Seattle WA 98105
(206)616-5551

	[[alternative HTML version deleted]]