[R-sig-ME] random factor variance

[Ken Beath]

On 20/05/2009, at 9:35 PM, João R. wrote:

> Thanks Ken, but I did not fully understood you.
>
> This means that the variance of the random effect needed to explain  
> your data is zero.
> This part I get, although the fact that the value for variance is an  
> absolute 0 makes me wonder if there is something wrong. I would be  
> happy with a low value, but not exactly 0.

Because the value can't go below zero, it may end up as zero. Even if  
the true value is small but non-zero then sampling variation may cause  
the estimate to be zero.


> The fact that two of the fixed factors are continuous variables  
> might have something to do with it?
> The clusters vary by the same amount or less than if there was a  
> random effect, that is they can all be explained by subject variation.
> This part I don't follow...
>

Think about a population and then dividing it randomly into a number  
of groups. I will assume normal distributions but the same ideas apply  
to binomial etc. We would expect that the groups will have different  
means, and we know how they will vary based on the population  
variance. When we fit a random effect the question we are asking is do  
these vary more than predicted from the population, in which case our  
random effect variance will be greater than zero.

> Basically, I am trying to predict the occurrence of reconciliation  
> after conflicts in a primate group (dependent variable: 0-no  
> occurrence; 1-reconciliation). My random variable is the victim's  
> identity of these conflicts (since not all group members are victims  
> of conflicts, and some are "more victims" than others). As fixed  
> effects I have a set of variables (describing the type of conflict  
> and the relationship between opponents; 23 variables), some  
> continuous (ex. F1, F3) and other categorical (ex. F2, F4 e F5).  
> Using a forward selection procedure based on AIC values, the best  
> fit model is this one I presented with the five fixed factors.
>

I suspect the problem is that your model is overfitted, because of the  
number of possible covariates, and the stepwise selection it has  
constructed a model that fits well without need for a random effect.

Ken


> Thanks again.
>
>
>
>
> On Wed, May 20, 2009 at 10:09 AM, Ken Beath <ken at kjbeath.com.au>  
> wrote:
> On 20/05/2009, at 11:43 AM, João R. wrote:
>
> Hello,
> I have recently used lme4 package to run a glmm, but a get 0 variance
> explained by the random effect. The model has 5 fixed effects, and I  
> have
> run each of them separately and for two of them (F1, F3) I also get 0
> variance for the random effect. Do you have any ideas of what might be
> causing this? Is this kind of result to be expected?
> thanks
>
>
> This means that the variance of the random effect needed to explain  
> your data is zero.  The clusters vary by the same amount or less  
> than if there was a random effect, that is they can all be explained  
> by subject variation.
>
> Ken
>
>
>
> Generalized linear mixed model fit by the Laplace approximation
> Formula: DV ~ F1 + F2 + F3 + F4 + F5 + (1 | R1)
> Data: JD
> AIC   BIC logLik deviance
> 203.2 225.9  -94.6    189.2
> Random effects:
> Groups Name        Variance Std.Dev.
> R1  (Intercept)  0        0
> Number of obs: 190, groups: R1, 14
> Fixed effects:
>          Estimate Std. Error z value Pr(>|z|)
> (Intercept)   1.8949     1.1869   1.596  0.11039
> F1          4.6740     2.4365   1.918  0.05507 .
> F2         -2.0657     0.7543  -2.739  0.00617 **
> F3       21.8036     8.8890   2.453  0.01417 *
> F4   1.0968     0.4874   2.250  0.02444 *
> F5      -1.7430     0.9583  -1.819  0.06894 .
>
>       [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>
>