[R-sig-ME] lme: predictions variance collapses when one more level is added

[ian m s white]

There is another way of looking at this. The data comprise a sample of size 17 from a multivariate (normal) distribution. The sample mean vector (4x1) and covariance matrix (4x4) can be calculated, and hypothesis tests about the population mean vector  constructed (e.g. see text by Mardia, Kent and Bibby, or similar). I'm not sure whether this easily fits into the mixed model framework. Especially lme(r), which insist on having a single residual term added to everything.
  
On 25 Oct 2013, at 23:48, Ben Bolker <bbolker at gmail.com> wrote:

> Dieter Menne <dieter.menne at ...> writes:
> 
>> 
>> I have a simple mixed-model, with predictive factor treat (levels M1,M2,M3,
>> M4), continuous par, and a grouping variable subj from a cross-over 
>> experiment.
>> 
>> Everything works as expected when I only use M1, M2, M3; see subset.lme
>> below. The residuals are well distributed; 
>> resid(.,type="p")~fitted(.)|treat
>> 
>> When I add level M4 (all.lme below), the variance of the 
>> predictions shrinks
>> to almost zero. I know that level M4 adds heteroscedasticity, so I tried
>> with varPower(); this corrects for the residual, but the fitted() appear
>> nonsensical. 
> 
>  Sorry for snipping context here (I'm posting via gmane, which doesn't
> like that).  If I use  weights=varIdent(form=~1|treat)) rather than
> weights=varPower() (i.e. residual variance varies by treatment group,
> rather than as a power function of the estimated mean), I get what
> seem (at least at a quick glance) to be reasonable results.
> 
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
> 


-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.