[R-sig-ME] Fwd: Re: New SIM paper by Zhang, Lu, Feng, Thurston, Xia, Zhu, and Tu

[David Atkins]

fyi, this may (or may not) be of interest to folks here, but a new paper 
is out comparing software (SAS and R, including lme4 and glmmML) for 
fitting GLMMs via simulation.  getting some chatter over on the 
multilevel listserv.

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)


-------- Original Message --------
Subject: Re: New SIM paper by Zhang, Lu, Feng, Thurston, Xia, Zhu, and Tu
Date: Tue, 14 Jun 2011 14:15:24 -0500
From: Brian R Gray <brgray at USGS.GOV>
Reply-To: Multilevel modelling discussion list <MULTILEVEL at JISCMAIL.AC.UK>
To: MULTILEVEL at JISCMAIL.AC.UK

possible that these results reflect R coding from 2009 (when the paper was
submitted).  as weak evidence, note that the authors define GLIMMIX as
being constrained to PQL only (GLIMMIX has offered Gaussian quadrature and
Laplacian estimation for some time)

Brian




From:
David Judkins <JUDKIND1 at WESTAT.COM>
To:
MULTILEVEL at JISCMAIL.AC.UK
Date:
06/14/2011 01:00 PM
Subject:
[MULTILEVEL] New SIM paper by Zhang, Lu, Feng, Thurston, Xia, Zhu, and Tu
Sent by:
Multilevel modelling discussion list <MULTILEVEL at JISCMAIL.AC.UK>



"On fitting generalize linear mixed effects models for binary responses
using different statistical packages." (pre-publication from the Wiley
website)

Not surprisingly, they got horrible results for a two-level logistic model
using GLIMMIX and some R procedures when the level 1 sample size is just 3
per level 2 unit.  More surprising is that NLMixed still gave decent
results.

They also found startling differences between NLMIXED and glmmML run with
the Gauss-Hermite option.  Wonder if this suggests problem in glmmML.

It would be interesting if someone were to take their examples and run
them under HLM, MLwin, and MPLUS.

One other oddity.  Most simulations have looked at the type I error rates
for null hypotheses that involve slope for some predictor =0 versus not 0.
  They looked at slope=1 versus not 1.  It produces very different results.
  Any theories on why?



David Judkins
Senior Scientist
Westat
1600 Research Boulevard
Rockville, MD 20850
(301) 315-5970
DavidJudkins at westat.com


-------------------------- Multilevel list --------------------------
To leave, send    leave multilevel    to jiscmail at jiscmail.ac.uk
For further info about the Multilevel list, please see
http://www.jiscmail.ac.uk/lists/multilevel.html     and
http://www.nursing.teaching.man.ac.uk/staff/mcampbell/multilevel.html



-------------------------- Multilevel list --------------------------
To leave, send    leave multilevel    to jiscmail at jiscmail.ac.uk
For further info about the Multilevel list, please see
http://www.jiscmail.ac.uk/lists/multilevel.html     and
http://www.nursing.teaching.man.ac.uk/staff/mcampbell/multilevel.html