[R-sig-ME] Variance structure in mixed models

[Ben Bolker]

 <v_coudrain at ...> writes:

> 
> Dear all,
> 
> I modelled my data with a quasi-poisson error distribution using glm:
> glm(response~variable1 + factor1 ,
> family=quasipoisson). Looking at the plot of residuals vs 
> factor 1, I noticed a large heterogeneity in residual variance
> between the different levels of my factor.
> With a linear regression, there is the option of using a 
> different variance structure for different levels using the option
> "varIdent" (funtion gls, package
> nlme). I could not find any equivalent for mixed models. Could 
> anybody explain me why and if there is another possibilty to 
> account for variance heterogeneity?
> 

  You could use MASS::glmmPQL, which allows all the same
variance structures as gls within the framework of a mixed model.
Just checking: you are looking at the Pearson residuals (i.e.
scaled by expected standard deviation), not just the raw residuals?

  Ben Bolker