Skip to content

(PR#2720)

2 messages · Jerome Asselin, Douglas Bates

Original
Jerome Asselin
# 
I apologize for cross-posting.

In the reproducable example below, I fit the same model with lme() and 
nlme(). However, I get 12 degrees of freedom with lme() and 2 df with 
nlme().

library(nlme)
set.seed(14)

a <- 2
x <- rep(rnorm(3),rep(5,3))
id <- rep(c("a","b","c"),rep(5,3))
y <- a+x+rnorm(15)
data <- data.frame(y=y,id=id)
initx <- matrix(x[c(1,6,11)],dimnames=list(c("a","b","c"),"x"))

summary(fit.lme <-
              lme(y ~ 1,data=data,random=~1|id,method="ML"))
summary(fit.nlme <-
              nlme(y ~ a + x, fixed= a~1, random=x~1|id,
              data=data, start=list(fixed=c(a=2),
              random=list(id=initx)),method="ML"))

When I try the same thing with set.seed(18), I get 12 df for both.

I have submitted a bug report on that issue on Dec 20th and I have tried 
to find the source of the problem, but no success. You can check what I 
found so far at:
http://r-bugs.biostat.ku.dk/cgi-bin/R/Add-ons?id=2384

Can anyone help find the source of this problem?

Sincerely,
Jerome Asselin

R 1.6.2 on Red Hat Linux 7.2
Package: nlme 3.1-38

______________________________________________
R-help@stat.math.ethz.ch mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help
Douglas Bates
# 
Instead of submitting the same bug report multiple times and
harrassing the developers of nlme via personal email you could perhaps
have read the section in Pinheiro and Bates (2000) on calculation of
degrees of freedom.  Had you taken a few minutes to do so you would
have discovered that the calculation of the degrees of freedom for the
constant in a linear mixed-effects model is a special case.  By
disguising the linear model as a nonlinear model you have managed to
defeat the detection of this special case and ended up with a
different number of degrees of freedom.

jerome@hivnet.ubc.ca writes: