On Mon, 22 Nov 2004 Chuanpu.Hu@sanofi-aventis.com wrote:
I am modeling a longitudinal dataset with nlme in S-PLUS 6.2, and nlme
suggested no random effects, i.e., all random effect estimates are “tiny”,
and anova comparisons with nls fits have p-value>0.999. This seems strange
to me and I wonder whether anyone has similar experiences.
I am not sure what exactly you are doing, but believe those comparisons to
be invalid. The point being that if the random effects are absent, the
MLEs are on the boundary of the parameter space and standard ML theory is
not applicable. (There are comments on this in Pinheiro & Bates, 2000.)
You mention `longitudinal', so do you have within-subject correlation? If
you do, it is easy to specify a model that is unidentifiable and that is
one fairly common way to get zero estimates for random effects.
A further point is that nlme is only doing approximate ML (it does ML in a
local linearization) and in particular its estimates of the maximized
likelihood can be very approximate.
A few more details: I am fitting a simple model of the form a +
b*exp(-c*Time) on a dataset with 3,700+ subjects and on average 10+
observations per subject. Other than the large size, the data do not appear
particularly “unusual.” In particular, it is difficult to believe that no
random effects are present. I used nls() results as initial estimates
(perturbations did not change anything). Does this mean that nlme has
difficulty handling large data?
I've used datasets larger than that. You could always test out your idea
on a small subset.
--
Brian D. Ripley, ripley@stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595
|