[R] lmer - error message

[Ben Bolker]

Sean Godwin <sean.godwin <at> gmail.com> writes:

> 
> Hi all,
> 
> I am fairly new to mixed effects models and lmer, so bear with me.
> 
> Here is a subset of my data, which includes a binary variable (lake (TOM or
> JAN)), one other fixed factor (Age) and a random factor (Year).
>   lake FishID Age Increment Year
> 1  TOM      1   1     0.304     2007
> 2  TOM      1   2     0.148     2008
> 3  TOM      1   3     0.119     2009
> 4  TOM      1   4     0.053     2010
> 5  JAN       2   1     0.352     2009
> 6  JAN       2   2     0.118     2010
> 
> The model I'm trying to fit is:
> m1 <- lmer(Increment ~ 0 + Age + Age*lake + (1|Year) + (1|Year:Age) +
> (1|FishID),lakedata)
> 
> The error message I get is: *"Error in mer_finalize(ans) : Downdated X'X is
> not positive definite, 27."*
> *
> *
> >From reading up on the subject, I think my problem is that I can't
> incorporate the 'lake' variable in a fixed-effect interaction because it is
> only has one binary observation.  But I don't know what to do to be able to
> fit this model.  Any help would be greatly appreciated!
> -Sean

  In principle you should be able to fit this model, but the error message
is telling you that there are numeric problems -- it may just be that
your data are a little too sparse in some direction.  A few suggestions:

 * try centering Age, or re-introducing the intercept, to see if you
can get something to work.
 * You _might_ try the development version of lme4 (lme4Eigen, on r-forge)
 * plot your data to see if you see anything odd about the data
 * perhaps try making Year a fixed effect -- 4 levels is fairly small
for a random effect
 * Ask further questions on the r-sig-mixed-models mailing list.

  Ben Bolker