Hi,
I have been working with LMMs for a while now and convergence has proved
to be a very common problem.
Right now, your model has two random effect terms, that is, one for the
intercept and one for the coefficient of the "treat" variable. This
implies that the default variance-covariance matrix for the random
effects has three parameters. Why don't you simply try imposing a
diagonal variance-covariance structure on your random effects? In order
to do this, have the "random" argument take value
reStruct(object = ~ treat | id, pdClass="pdDiag")
It's not exactly the same model that you wanted to fit initially, but if
it converges, you probably won't regret the sacrifice (which probably
won't be that considerable anyways)!
As to why you don't get convergence in the first place, I unfortunately
can't tell. I think the R-gurus are a lot more qualified than me to
answer such questions!
Cheers,
--
*Luc Villandré*
/Biostatistician
MUHC-MCH Research Institute/
On 04/02/2010 6:42 PM, anord wrote:
> Dear all,
>
> I am working on a data set in which I have sequentially measured egg
> temperatures ("eggtemp") in birds incubating in different ambient
> temperatures ("treat", sample data set below), "id" is not replicated within
> treatment.
>
> id treat eggtemp
> 1 79 3 30.90166
> 2 42 3 34.94044
> 3 10 3 32.69945
> 4 206 3 36.64127
> 5 23 3 31.80055
> 6 5 3 29.98338
> 7 24 3 35.72992
> 8 45 3 30.49584
> 9 29 3 33.64958
> 10 78 3 31.37673
> 11 44 3 32.85873
> 12 368 3 34.44875
> 13 79 4 31.24100
> 14 42 4 34.11634
> 15 10 4 34.73407
> 16 206 4 36.20914
> 17 23 4 34.98061
> 18 5 4 34.17590
> 19 24 4 37.71468
> 20 45 4 35.34765
> 21 29 4 35.48892
> 22 78 4 33.26593
> 23 44 4 34.86981
> 24 368 4 34.44875
> 25 79 2 27.33241
> 26 42 2 30.73269
> 27 10 2 29.54986
> 28 206 2 31.78947
> 29 23 2 29.69114
> 30 24 2 36.48199
> 31 45 2 29.76454
> 32 29 2 30.56510
> 33 78 2 27.71468
>
> For this data, I want to construct a model with a random intercept and slope
> to compare with a model containing only the random intercept to check
> whether or not different individuals respond differently to treatment.
>
>> temp2.lme<-lme(eggtemp~treat,random=~treat|id,data=temp3.df)
>>
> However, I get the below error, which I suspect might be because individuals
> are not replicated within treatments.
> > nlminb problem, convergence error code = 1
> message = iteration limit reached without convergence (9)
>
> If I specify treatment as a factor (which it really is not, just different
> ambient temperatures), the model seems to converge. Similarly, the model
> converges if I exclude one of the treatment levels. However, this does not
> feel very satisfactorily.
>
> So, how do I move on from here? Any tips and hints are much appreciated. I
> have tried to include the random slope only, which also works, but the
> inference for such a model would be quite different and not really what I am
> after.
>
>> temp3.lme<-lme(eggtemp~treat,random=~treat-1|id,data=temp3.df)
>>
> Kind regards,
> Andreas Nord
> Lund University
> Sweden
>
>
[[alternative HTML version deleted]]
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
