[R-sig-ME] Another case of -1.0 correlation of random effects

[Kevin E. Thorpe]

Kevin E. Thorpe wrote:
> Ben Bolker wrote:
>> Ken Knoblauch wrote:
>>> Kevin E. Thorpe <kevin.thorpe at ...> writes:
>>>
>>>> My data come from a crossover trial and are balanced.
>>>>
>>>>  > str(gluc)
>>>> 'data.frame':    96 obs. of  4 variables:
>>>>   $ Subject  : int  1 2 3 5 6 7 10 11 12 13 ...
>>>>   $ Treatment: Factor w/ 2 levels "Barley","Oat": 1 1 1 1 1 1 1 1 1 
>>>> 1 ...
>>>>   $ Dose     : int  8 8 8 8 8 8 8 8 8 8 ...
>>>>   $ iAUC     : num  110 256 129 207 244 ...
>>>>
>>>> clip>
>>> Shouldn't you make Subject into a factor?
>>>
>>> Ken
>>>
>>
>>   It would make the plot a little bit prettier but I don't think it
>> matters in this case because variable that appears as a grouping
>> variable (i.e. on the right of the | ) is automatically treated as a
>> factor?  I think?
>>
>>   Since it is really a crossover trial, it would seem reasonable in
>> principle to have the (Treatment|Subject) random effect in there as
>> well. I'm not sure what to do about the -1 correlation: it seems the
>> choices (not necessarily in order) are (1) throw up your hands and say
>> there's not enough data to estimate independently; (2) try WinBUGS,
>> possibly with slightly informative priors; (3) try using lme4a to create
>> profiles of the parameters and see if you can figure out what's 
>> happening.
> 
> Let's see.  I wish (1) was an option.  (2) would be promising if my 
> knowledge of BUGS and Bayesian methods filled more than a thimble. 
> Thanks to Jarrod for his suggestion in response to this.  I'll take a 
> look at that too.  Option (3) is probably worth a go too.
> 
> Aside from the fact that the Dose variable are the actual doses and not 
> categories, and we all know not to categorize continuous variables, what 
> are your thoughts on treating Dose as a factor (since it seems to behave)?
> 
> Thanks all for taking the time to provide your suggestions.
> 
> Kevin
> 
Okay, I now have lme4a installed and I get an error message when I do 
(note: this is the same model from my OP):

 > gluc.lmer1a <- 
lmer(iAUC~Dose+(Dose|Subject),data=gluc,subset=Treatment=="Oat",REML=FALSE)

 > gluc.lmer1a
Linear mixed model fit by maximum likelihood  ['lmer']
Formula: iAUC ~ Dose + (Dose | Subject)
    Data: gluc
  Subset: Treatment == "Oat"
      AIC      BIC   logLik deviance
    575.1    586.3   -281.6    563.1

Random effects:
  Groups   Name        Variance Std.Dev. Corr
  Subject  (Intercept) 7492.19  86.557
           Dose          14.68   3.831   -1.000
  Residual             4727.27  68.755
Number of obs: 48, groups: Subject, 12

Fixed effects:
             Estimate Std. Error t value
(Intercept)  309.352     29.338  10.544
Dose         -14.424      3.533  -4.083

Correlation of Fixed Effects:
      (Intr)
Dose -0.647

 > pr1 <- profile(gluc.lmer1a at env)  ## using @env base on other threads
Error in x[ndat + (1L:deg) - deg] :
   only 0's may be mixed with negative subscripts

Is this because I'm trying to profile a model that profile() cannot 
handle yet, or does it indicate there really are serious problems with 
my model?

I'm at a loss as to how determine what is really going on with these data.

Kevin
-- 
Kevin E. Thorpe
Biostatistician/Trialist, Knowledge Translation Program
Assistant Professor, Dalla Lana School of Public Health
University of Toronto
email: kevin.thorpe at utoronto.ca  Tel: 416.864.5776  Fax: 416.864.3016