[R-sig-ME] R-sig-mixed-models Digest, Vol 53, Issue 26

Ben Bolker bbolker at gmail.com
Mon May 23 05:46:30 CEST 2011


-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 11-05-24 02:53 PM, Nicholas Lewin-Koh wrote:
> Comparing non-nested models is tricky. One approach is to use the LRT,
> with model A as the null
> and then with model B as the null. The problem is that the distribution
> of the LRT is not
> known in the non-nested case. Essentially you acn simulate from each
> model to generate the null distribution.
> A nice review of approaches is in
> Fraser et al 2010. A unified approach to model selection using the
> likelihood ratio test. Methods in Ecology and Evolution 2(2) 155-162
> 
> Nicholas
> 
> 
>> Message: 3
>> Date: Tue, 24 May 2011 02:26:29 -0700 (PDT)
>> From: Iasonas Lamprianou <lamprianou at yahoo.com>
>> To: r-sig-mixed-models at r-project.org
>> Subject: Re: [R-sig-ME] compare non-nested logistic models
>> Message-ID: <990843.69746.qm at web120609.mail.ne1.yahoo.com>
>> Content-Type: text/plain
>>
>> Dear friends, I would like to compare these two non-nested  models:
>>
>> Alt1 <- glmer(contplt2 ~ 1+A+B+(1|D),family=binomial, data=partNM3)
>> Alt2 <- glmer(contplt2 ~ 1+A+C+(1|D),family=binomial, data=partNM3)
>>
>> I tried Faraway(2005) but he only deals with nested models (unless I
>> missed 
>> something). Can anyone drop a very quick line of help (or just a link
>> which I 
>> can follow?)

  What kinds of inferences do you want to make about the comparison?
Nicholas's advice above is for testing significance/model selection.  If
you just want to compare the goodness of fit, just compare the
log-likelihoods on the 'usual' log-likelihood scale -- i.e. less than 2
is a small difference, >10 is a huge difference.

>>
>> Besides glmer, I tried glmmPQL and MCMCglmm and they all give the same
>> results. 
>> I would like to thank all the authors of the packages for giving us free 
>> alternatives. Does anyone know when each of the three is mostly
>> appropriate(or 
>> they do exactly the same job?)

   glmmPQL uses penalized quasi-likelihood, glmer uses Laplace
approximation or adaptive Gauss-Hermite quadrature, MCMCglmm uses Markov
chain Monte Carlo.

  See <http://glmm.wikidot.com/faq#estmethods> (taken from Bolker et al
2009 _Trends in Ecology and Evolution_)



>>
>> Thanks
>>
>>  Dr. Iasonas Lamprianou
>> Department of Social and Political Sciences
>> University of Cyprus
>>
>>
>>
>>
>>
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAk3Z2JYACgkQc5UpGjwzenNAugCfb7azsnJeqMu2jaAN+OXph5qc
yxMAniNKky79+jCTqlMrWFp0Fref4aEQ
=12pL
-----END PGP SIGNATURE-----




More information about the R-sig-mixed-models mailing list