Re: [R] animal models and lme

[Spencer Graves]

	  If you can solve the problem for fixed rho = (va/ve) using gls, then 
you can call gls for many values of rho, plot the log(likelihood) 
contours vs. rho, construct confidence intervals, etc.  You may even be 
able to write a function to return (-2)*log(likelihood) for a fixed rho 
and then  use "optim" to minimize that "deviance".  [I would suspect 
that the log(likelihood) might look more parabolic in terms of log(rho) 
that in terms of rho itself.  In addition, "optim" might work better 
with the minimum for log.rho = (-Inf) than with a lower bound for rho at 0.]

hope this helps.  spencer graves

Douglas Bates wrote:
<[EMAIL PROTECTED]> writes:


Not convinced that responses so far have addressed the problem. The
model is
y = mu + U + e

where e is a vector of independendent errors with variance ve, and U
is a vector of random effects with covariance matrix va*A, where A is a
known matrix (which we can assume is a correlation matrix). If we know the
ratio (va/ve), this reduces to a GLS problem, but not otherwise. Usually
we have to estimate both ve and va.


Sorry to say that I don't think lme will handle that problem gracefully.

______________________________________________
[EMAIL PROTECTED] mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help
______________________________________________
[EMAIL PROTECTED] mailing list
https://www.stat.math.ethz.ch/mailman/listinfo/r-help