[R-sig-ME] Error: "Cannot get confidence intervals...", with lme, what does it means?

[Rafael Maia]

Hello!

Sorry to keep this going, but could you elaborate on the meaning of  
that output? I am asking this because I am trying to model my data  
set, and with the same model, I get that output if I use REML, but I  
get a normal output if I use ML - so I'm assuming it may be related to  
the model, but also in the estimating method. The intervals calculated  
using intervals() under the ML method are very similar to the ones I  
obtain when computing them by hand.

Under my limited knowledge, I believe I am using the most appropriate  
model for my data, even though it is definitely not a "good" one for  
the data (which is inherently unbalanced).

Commands follow, below. Many thanks for any help!

Best,
Rafael

 > intervals(m1)
Error in intervals.lme(m1) :
  Cannot get confidence intervals on var-cov components: Non-positive  
definite approximate variance-covariance

 > m1<-update(m1, method="ML")

 > intervals(m1)
Approximate 95% confidence intervals

Fixed effects:
                     lower      est.      upper
(Intercept)      0.2268163  1.318894  2.4109708
factor(Status)2  0.7653039  1.812026  2.8587479
Ano             -1.7888219 -1.112453 -0.4360846
attr(,"label")
[1] "Fixed effects:"

Random Effects:
  Level: ind
                                         lower       est.       upper
sd((Intercept))                   1.692475e-02  0.9802002    56.76849
sd(factor(Status)2)               2.009503e-05  0.8586423 36689.01183
cor((Intercept),factor(Status)2) -1.000000e+00 -0.8485173     1.00000

Within-group standard error:
       lower         est.        upper
3.685400e-14 3.630675e-01 3.576762e+12



On 2 Sep 2008, at 11:09, Gang Chen wrote:

> Cotter,
>
> Check the following component
>
>> lmefit1$apVar
>
> If you see something like this
>
> [1] "Non-positive definite approximate variance-covariance"
>
> it most likely indicates you have an inappropriate model for the  
> data. Try plotting out the data, and get some idea about the  
> feasible models, and then fit the data with those models.
>
> Cheers,
> Gang
>
>
> On Sep 2, 2008, at 8:29 AM, R.S. Cotter wrote:
>
>> Hello,
>>
>> In some occasions I get this error message: "Cannot get confidence
>> intervals on var-cov components: Non-positive definite approximate
>> variance-covariance".
>>
>> I have tried to figure out this by using help function, but didn't
>> find answer to the question. I address this question with describing
>> the model and the primary task that I want to solve. Sorry if the
>> question is clumsy formulated, I 'm not that experienced with R and
>> statistics.
>>
>> My model is:
>> Response= Weight(continous)
>> Explanatory variables= Time (continous) and Diet (kategorical, two  
>> groups; B&C)
>>
>> The primary question of interest is wheter the growth rates
>> (Weight/Time) differ among the two diets.
>>
>> lmefit1<-lme(Weight ~ Diet*Time,random=~1|Place,data=Total)
>>
>> Summary output is ok, so far so good. But I also wanted to get the
>> slope and confidence intervals for the growth rates for both diets
>> (B&C), so I ran intervals(). And I got the intercept, slope and
>> confidence intervals for diet B, see below.
>>
>> But I also wanted the same for the diet C, to do this I renamed  
>> diet C
>> to A in the data sheet to force C to be the dummy variable. Is this
>> the right way to do it?
>>
>> When running the intervals () once again, I got this message: "Cannot
>> get confidence intervals on var-cov components: Non-positive definite
>> approximate variance-covariance". What could be wrong..? Is there
>> other ways to get the slope and confidence intervals from a lme  
>> model?
>>
>>> intervals(lmefit1)
>> Approximate 95% confidence intervals
>>
>> Fixed effects:
>>                         lower               est.                  
>> upper
>> (Intercept)           66.040673     108.122242     150.203810
>> DietC                -175.080336   -109.638518     -44.196700
>> Time                     4.177387         5.434087       6.690788
>> DietC:Time          7.938101         11.180806     14.423511
>> attr(,"label")
>> [1] "Fixed effects:"
>>
>> Random Effects:
>> Level: Place
>>                   lower     est.    upper
>> sd((Intercept)) 0.1478599 13.50651 1233.775
>>
>> Within-group standard error:
>>  lower     est.    upper
>> 159.9128 174.8928 191.2761
>>
>> Best regards Cotter
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models