s-news
[Top] [All Lists]

correlation of random effects in nlme model

To: s-news@lists.biostat.wustl.edu
Subject: correlation of random effects in nlme model
From: Buzz Burhans <wsb2@cornell.edu>
Date: Mon, 18 Nov 2002 16:22:39 -0500
I have two questions I would appreciate help with if possible. My first question:

I have used nlsList to create an object that I then update using nlme to apply a function for exponential growth to a grouped data object. The function is:

 fExpIncrease <- function(x, int, B) int*(exp( B* x))

My question involves the correlation of the random effects when both the int and B terms above are included in the model as "random = int + B ~ 1" ; they are extremely correlated (0.994), and the model is fairly unstable as the random effects are alternately dropped. However, using augPred to look at the results appears to suggest that the predictions with both in the model are better fit, and theoretically I can make a case for both random effects. Examining the model with intervals shows that both random effects are bounded away from 0, but their correlation is not, but is not, in fact it runs from -.999 to 1.

My question, is it inappropriate to leave both random effects in under these circumstances relative to their correlation ?

My second question involves a problem which arose when I tried to center the observations in an attempt to reduce this correlation, and I will post it separately.

Thanks,

Buzz Burhans
Department of Animal Science
Cornell University


<Prev in Thread] Current Thread [Next in Thread>
  • correlation of random effects in nlme model, Buzz Burhans <=