[R-sig-ME] log likelihood tests

[Andrew Robinson]

Hi Katie,

I'm not sure that this is the best group to send that question to, as
it's a list serve dedicated to mixed-effects modeling questions in R,
not SAS.  I am under the impression that similar list serves exist for
SAS.  I recommend that you try those in the first instance.  That
said, there may be someone in this group with experience of SAS.  I
don't have any.

Good luck,

Andrew

On Fri, Dec 09, 2011 at 03:05:56PM -0600, Katie McGhee wrote:
> Dear mixed model list serve group,
> 
> I was wondering if anyone could advise me on the significance testing of
> random effects in a generalized linear mixed model? I am a biologist with
> basic statistical knowledge and since I am using a complex method, I want
> to make sure that I am doing things correctly.
> 
> My experiment:
> I performed a paternal half-sib breeding design to examine whether a
> variety of mating and non-mating behaviors are significantly heritable. I
> am interested in whether "sire" explains a significant amount of the
> behavioral variation.
> 
> My analysis:
> The behavioral data was non-normal and had a lot of zeros, so I decided to
> try a GLMM. I conducted my analysis in SAS GLIMMIX with sire and dam(sire)
> as random effects and no fixed effects. I specified a negative binomial
> distribution with a log link function and used the Laplace approximation to
> get a true log-likelihood.
> 
> My question:
> When I want to figure out whether my random effect of sire is significant,
> do I compare models with and without the sire effect in their -2 log
> likelihoods under the "Fit Statistics" or the -2 log likelihood (behavior |
> random effects) under the "Fit Statistics for conditional distribution"?
> Both of these things are included in the SAS output.
> 
> I am confused because I have seen some statistics papers (way over my head
> so I may be totally off-base here, e.g.Paul and Deng 2000) that suggest
> using the conditional distributions for sparse data (which mine is). I have
> also analyzed my zero-rich data as binomial (scored as 1/0, binomial
> distribution, probit link function, laplace estimation) which seems much
> more appropriate and the conclusions in terms of the significance of sire
> match better with those from the log likelihood tests using the conditional
> distribution than the regular log likelihood test from the marginal
> distribution.
> 
> I apologize if this is super obvious but I have searched SAS
> discussions,etc and as of yet, I can find no straightforward answer. I
> would appreciate any help you could offer!
> 
> Thank you!
> 
> Sincerely,
> 
> Katie
> 
> -- 
> ---- ><((((?> ----------- ><((((?> ----
> 
> Katie E. McGhee
> 
> Postdoctoral Fellow
> Integrative Biology
> University of Illinois
> 433 Morrill Hall
> 505 S. Goodwin Ave.
> Urbana, IL, 61801
> 
> kemcghee at illinois.edu
> 
> 	[[alternative HTML version deleted]]
> 

> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models


-- 
Andrew Robinson  
Deputy Director, ACERA 
Department of Mathematics and Statistics            Tel: +61-3-8344-6410
University of Melbourne, VIC 3010 Australia               (prefer email)
http://www.ms.unimelb.edu.au/~andrewpr              Fax: +61-3-8344-4599
http://www.acera.unimelb.edu.au/

Forest Analytics with R (Springer, 2011) 
http://www.ms.unimelb.edu.au/FAwR/
Introduction to Scientific Programming and Simulation using R (CRC, 2009): 
http://www.ms.unimelb.edu.au/spuRs/