[R] glmer with non integer weights

[Emmanuel Charpentier]

Addendum to my previous answer :

In that special case, the limited range of the asin(sqrt())
transformation, which is a shortcoming, turns out to be useful. The
fixed-effect doefficients seem semi-reasonable (except for stageB) :

> (sin(coef(lm(asin(sqrt(MH.Index))~0+stage, data=similarity))))^2
   stageA    stageB    stageC    stageD 
0.6164870 0.3389430 0.5083574 0.5672021 

The random effects being nested in the fixed efect, one can't afford to
be lazy in the parametrization of the corresponding random effect :

> summary(lmer(asin(sqrt(MH.Index))~stage+(stage|site),
data=similarity))
Linear mixed model fit by REML 
Formula: asin(sqrt(MH.Index)) ~ stage + (stage | site) 
   Data: similarity 
   AIC BIC logLik deviance REMLdev
 155.3 199 -62.65    111.8   125.3
Random effects:
 Groups   Name        Variance Std.Dev. Corr                 
 site     (Intercept) 0.043579 0.20876                       
          stageB      0.033423 0.18282  -0.999               
          stageC      0.043580 0.20876  -1.000  0.999        
          stageD      0.043575 0.20875  -1.000  0.999  1.000 
 Residual             0.128403 0.35833                       
Number of obs: 136, groups: site, 39

Fixed effects:
            Estimate Std. Error t value
(Intercept)  0.93036    0.08431  11.035
stageB      -0.30879    0.10079  -3.064
stageC      -0.13660    0.09981  -1.369
stageD      -0.07755    0.14620  -0.530

Correlation of Fixed Effects:
       (Intr) stageB stageC
stageB -0.836              
stageC -0.845  0.707       
stageD -0.577  0.482  0.487
> v<-fixef(lmer(asin(sqrt(MH.Index))~stage+(stage|site),
data=similarity))
> v[2:4]<-v[1]+v[2:4]
> names(v)[1]<-"stageA"
> (sin(v))^2
   stageA    stageB    stageC    stageD 
0.6429384 0.3390903 0.5083574 0.5672021 

But again, we're exploiting a shortcoming of the asin(sqrt())
transformation.

HTH,

					Emmanuel Charpentier

Le vendredi 16 avril 2010 à 00:15 -0800, Kay Cichini a écrit :
> thanks thierry,
> 
> i considered this transformations already, but variance is not stabilized
> and/or normality is neither achieved.
> i guess i'll have to look out for non-parametrics?
> 
> best regards,
> kay