[R-sig-ME] More help with glmer!

[Joshua Wiley]

Hi,

The lack of residual variance is not a function of glmer per se, rather the distribution you used.  The poisson distribution only has parameter---the expectation and dispersion.

The variance of the random subject effect is the variability in intercepts by subject.  In your case something like the variability in expected log count when treatment is 0 for different subjects.  It is very small.

glmer does use a maximum likelihood estimator and returns maximum likelihood estimates.  The Laplace approximation is a numerical way to calculate them (optimize the likelihood function).

You can set nAGQ to some number higher than 1 to increase the number of points evaluated and obtain slightly more accurate estimates at the cost of speed.

Cheers,

Josh

On Feb 19, 2012, at 20:16, lopez toledo <llopezt2000 at yahoo.com.mx> wrote:

> Hi all:Hope you understand I'm not an statistician and hope my question is not very basic. I did have a deep look to old posts but could not find an appropriate response. So, let me ask my question!
> I'm doing some lmer and glmer (poisson) models, but I noticed that for glmer there is not residual variance for the random effects. So, how do variance values should be interpreted in a glmer? What does the 6.8578 e-06 mean in the example below? I understand the "Subject" factor is explaining 6.8 e-06 of the variance, but what is the total or the residual variance? This number does not say many to me. 
> Additionally, is it possible to fit a glmer with maximum likelihood? or is Laplace approximation the only option?
> 
> Thanks for your help and comprehension � 
> 
>> summary(glmer1)
> Generalized linear mixed model fit by the Laplace approximation 
> Formula: Response ~ Treatment + (1 | Subject) 
> �� AIC� BIC logLik deviance
> �788.8 1052 -345.4��� 690.8
> Random effects:
> �Groups Name������� Variance�� Std.Dev. 
> Subject �� (Intercept) 6.8578e-06 0.0026187
> Number of obs: 1604, groups: NNo, 555
> 
>    [[alternative HTML version deleted]]
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models