[R-sig-ME] different aic and LL in glmer(lme4) and glimmix(SAS)?

[Adam D. I. Kramer]

Also...R is giving you a model with LESS deviance. So, it's doing a better
job...why would you want to reproduce SAS? :)

--Adam


On Thu, 1 Jul 2010, Jeffrey Evans wrote:

> Good question.
>
> They are similar
>
> Compare models with nested fixed effects structures
> Full model = lnsdlmaxd + lnadultssdld + lnsdlmaxd:lnadultssdld
> Reduced model = lnsdlmaxd + lnadultssdld
>
> AIC		R		SAS
> Full		1150		1663.9
> Reduced	1159		1673.4
> -------------------------------
> deltaAIC	9		9.5
>
>
> ________________________________
>
> From: almost at gmail.com [mailto:almost at gmail.com] On Behalf Of Andy Fugard
> Sent: Thursday, July 01, 2010 12:16 PM
> To: Jeffrey Evans
> Cc: r-sig-mixed-models at r-project.org
> Subject: Re: [R-sig-ME] different aic and LL in glmer(lme4) and
> glimmix(SAS)?
>
>
> Hi Jeff,
>
> Can't answer the question, but out of interest, what happens when you
> compare nested models in R and SAS, e.g., models with and without the
> lnsdlmaxd:lnadultssdld interaction?  Would be interesting to see the
> log-likehood ratio (and maybe /change/ in AIC and BIC between the models).
>
> Cheers,
>
> Andy
>
>
> On Thu, Jul 1, 2010 at 18:03, Jeffrey Evans <Jeffrey.Evans at dartmouth.edu>
> wrote:
>
>
> 	Hello All,
>
> 	I have read several posts related to this previously, but haven't
> found any
> 	resolution yet. When running the same GLMM in glmer and in SAS PROC
> GLIMMIX,
> 	both programs return comparable parameter estimates, but wildly
> different
> 	likelihoods and AIC values.
>
> 	In SAS I specify use of the Laplace approximation. In R, I believe
> this is
> 	the default (no?).
>
> 	What's the difference, and [how] can I reproduce the SAS -2ll in
> glmer?
>
> 	Thanks,
> 	Jeff
>
> 	\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/
> 	> R_GLMM = glmer(cbind(SdlFinal, SdlMax-SdlFinal) ~
> lnsdlmaxd*lnadultssdld +
> 	 (1|ID),data=sdlPCAdat,family="binomial")
> 	> R_GLMM
> 	Generalized linear mixed model fit by the Laplace approximation
> 	Formula: cbind(SdlFinal, SdlMax - SdlFinal) ~ lnsdlmaxd *
> lnadultssdld +
> 	(1 | ID)
> 	  Data: sdlPCAdat
> 	 AIC  BIC logLik deviance
> 	 1150 1165   -570     1140        <------------------ this line!!
> 	Random effects:
> 	 Groups Name        Variance Std.Dev.
> 	 ID     (Intercept) 1.2491   1.1176
> 	Number of obs: 144, groups: ID, 48
>
> 	Fixed effects:
> 	                      Estimate Std. Error z value Pr(>|z|)
> 	(Intercept)             4.56964    0.43148  10.591  < 2e-16 ***
> 	lnsdlmaxd              -0.65936    0.05686 -11.595  < 2e-16 ***
> 	lnadultssdld           -0.64534    0.15861  -4.069 4.73e-05 ***
> 	lnsdlmaxd:lnadultssdld  0.07393    0.02166   3.414  0.00064 ***
> 	---
> 	Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
>
> 	Correlation of Fixed Effects:
> 	           (Intr) lnsdlm lndlts
> 	lnsdlmaxd   -0.923
> 	lnadltssdld -0.461  0.479
> 	lnsdlmxd:ln  0.482 -0.508 -0.994
>
>
> 	 \/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/
> 	title 'SAS GLMM';
> 	proc glimmix data=sdlPCAdat ic=pq noitprint method=laplace;
> 	class site id;
> 	model sdlfinal/sdlmax = lnsdlmaxd|lnadultssdld/ solution
> dist=binomial;
> 	random ID /;
> 	covtest glm/wald;
> 	run;
>
>
> //////////////////////////////////////////////////////////////////////
>
> 	                     SAS GLMM      19:36 Wednesday, June 30, 2010 88
>
> 	                  The GLIMMIX Procedure
>
> 	           Data Set           WORK.SDLPCADAT
> 	           Response Variable (Events)  SdlFinal
> 	           Response Variable (Trials)  SdlMax
> 	           Response Distribution     Binomial
> 	           Link Function         Logit
> 	           Variance Function       Default
> 	           Variance Matrix        Not blocked
> 	           Estimation Technique     Maximum Likelihood
> 	           Likelihood Approximation   Laplace
> 	           Degrees of Freedom Method   Containment
>
>
>
> 	                 Optimization Information
>
> 	            Optimization Technique    Dual Quasi-Newton
> 	            Parameters in Optimization  5
> 	            Lower Boundaries       1
> 	            Upper Boundaries       0
> 	            Fixed Effects         Not Profiled
> 	            Starting From         GLM estimates
>
> 	            Convergence criterion (GCONV=1E-8) satisfied.
>
> 	                    Fit Statistics
>
> 	              -2 Log Likelihood        1653.90  <------------------
> this
> 	line!!
> 	              AIC (smaller is better)  1663.90
> 	              AICC (smaller is better) 1664.33
> 	              BIC (smaller is better)  1673.25
> 	              CAIC (smaller is better) 1678.25
> 	              HQIC (smaller is better) 1667.43
>
>
> 	             Fit Statistics for Conditional Distribution
>
> 	             -2 log L(SdlFinal | r. effects)   1436.44
> 	             Pearson Chi-Square          908.07
> 	             Pearson Chi-Square / DF        6.31
>
>
> 	                Covariance Parameter Estimates
>
> 	           Cov         Standard     Z
> 	           Parm  Estimate    Error   Value   Pr > Z
>
> 	           ID    1.2491   0.2746   4.55   <.0001
>
>
> 	                Solutions for Fixed Effects
>
> 	                             Standard
> 	    Effect         Estimate    Error    DF  t Value  Pr > |t|
>
> 	    Intercept           4.5696   0.4333    47    10.55  <.0001
> 	    lnsdlmaxd          -0.6594   0.05717   93   -11.53  <.0001
> 	    lnadultssdld       -0.6453   0.1593    93   -4.05   0.0001
> 	    lnsdlmaxd*lnadultssd0.07394  0.02174   93    3.40   0.0010
>
> 	_______________________________________________
> 	R-sig-mixed-models at r-project.org mailing list
> 	https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>