[R] convergence error code in mixed effects models

[Douglas Bates]

On Dec 13, 2007 4:15 PM, Ilona Leyer <ileyer at yahoo.de> wrote:
> Dear All,
> I want to analyse treatment effects with time series
> data:  I measured e.g. leaf number (five replicate
> plants) in relation to two soil pH - after 2,4,6,8
> weeks. I used mixed effects models, but some analyses
> didn´t work. It seems for me as if this is a randomly
> occurring problem since sometimes the same model works
> sometimes not.
>
> An example:
> > names(test)
> [1] "rep"    "treat"  "leaf"   "week"
> > library (lattice)
> > library (nlme)
> > test<-groupedData(leaf~week|rep,outer=~treat,test)
> > model<-lme(leaf~treat,random=~leaf|rep)
> Error in lme.formula(leaf~ treat, random = ~week|rep)

Really!? You gave lme a model with random = ~ leaf | rep (and no data
specification) and it tried to fit a model with random = ~ week | rep?
Are you sure that is an exact transcript?

> :
>         nlminb problem, convergence error code = 1;
> message = iteration limit reached without convergence
> (9)

> Has anybody an idea to solve this problem?

Oh, I have lots of ideas but without a reproducible example I can't
hope to decide what might be the problem.

It appears that the model may be over-parameterized.  Assuming that
there are 4 different values of week then ~ week | rep requires
fitting 10 variance-covariance parameters. That's a lot.
The error code indicates that the optimizer is taking