[R-sig-ME] error in LME: Non-positive definite approximate variance-covariance ????

[Douglas Bates]

On Wed, Mar 9, 2011 at 3:03 PM, Federico Bonofiglio <bonoricus at gmail.com> wrote:
> Hi Dear longitudinal addicts,
> I have a quick issue to submit u that'll take u hopefully only a few
> minutes:
>
> I ran an lme model and I receive the following error mex that says a no
> positive varcov matrix was estimated..
> Now at the moment I don't exactly know the code implemented by Pinheiro and
> Bates so I'm not able to say what was the mechanism that could possibly
> generate such a problem.
>
> First thing I thinking from scratch is that te two bolded variables in the
> model below (cd4, log(rna)) could actually be higly (negatively) correlated
> so collinearity may occurr and wreak havocs there..
>
> By removing cd4 the problem disappear indeed.
>
> I would still ask u an opinion if that won't cost u much....
>
>
>
>> RATIO1C2<-lme(ratio1~rt+rt*(rt>=25)+*cd4*+cd4:rt+
> + cd4:rt*(rt>=25)+resp2+resp2:rt+resp2:rt*(rt>=25)+
> + log(rna)+*log(rna)*:rt+log(rna):rt*(rt>=25)+
> + sex+age+rf+nadir+pharmac2+pharmac2:rt+
> + hbs+hcv+haartd+hivdur+switch,
> + subset=which(ratio1<10),data=frame4b,
> + correlation=corARMA(p=4),
> + random= ~rt|id, na.action=na.omit)
>>
>> intervals(RATIO1C2)$fixed
> *
> Errore in intervals.lme(RATIO1C2) :
>  Cannot get confidence intervals on var-cov components: Non-positive
> definite approximate variance-covariance

Well, your model is somewhat complex and the approximate
variance-covariance matrix for the parameters is determined by
second-order finite differences.  It is not really surprising that in
such a model the approximate Hessian of the deviance fails to be
positive definite.  This is a warning to you that the model is
over-parameterized.

> *
> thank u indeed
>
> Freddy
>
>
>
> --
> *Little u can do against ignorance,....it will always disarm u:
> is the 2nd principle of thermodinamics made manifest, ...entropy in
> expansion.**....But setting order is the real quest 4 truth, ......and the
> mission of a (temporally) wise dude.
> *
>
>        [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>