[S] Varcomp for comparing multistratum models

Ray Webster (WebsterR@landcare.cri.nz)
Fri, 04 Dec 1998 14:40:26 +1200


I'm using varcomp in S-Plus 4.0 for Windows to compare unbalanced
multistratum models. In particular, I have some variable measured at
different times at various sites, but not all sites have a measurement at
each time.

I want to test the significance of various independent variables, and I'm
doing this by comparing the model likelihoods obtained using
method="reml". My problem is that sometimes the more complex model
has a lower (more negative) likelihood. An example is:

mod1_varcomp(log(rep)~site,data,method="reml",tol=1e-10)

mod3_varcomp(log(rep)~plant+site,data,method="reml",tol=1e-10)

where log(rep) is the dependent var, site is a random factor, and plant is
a quantitative independent variable (data is the name of the data frame).

I can get the likelihoods from mod1$info[7] and mod3$info[7]:

> mod1$info[7]
objective
-8.652036
> mod3$info[7]
objective
-8.867806

This makes no sense to me at all - how can the likelihood for the more
complex model be less? Is there an error in the function, or am I being
stupid here?

Thanks

Ray Webster

-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news