specifying values in correlation matrix in nlme

The only way I know of in nlme is to transform Zu to Z1 u1 so that
u1 ~ N(0, cI), which nlme can cope with. E.g. if A = PDP' is the
spectral decomposition of A, take Z1=ZPD^{1/2}, u1 = D^{-1/2}P'u.
Unfortunately Z1 is much less sparse than Z and in my experience
this only works with small problems.

The kinship package has a function lmekin which will do what you want. It
uses ML but I reckon it could be easily modified to use REML. It does not
make use of nlme, it just evaluates the log likelihood and passes it to a
general purpose optimiser.


On Thu, 26 May 2005, Keyan Zhao wrote:

Could anyone help with a linear mixed model fitting problem ?

The model is :

Y= Xp + Zu + e
where X, Z are known design matrix, p is fixed effect factor, u is
random effect,  u~ (0, G) , e~(0,R)

The main problem is , I want to fix the covariance matrix G to be a
constant times a known covariance matrix A,   G = c*A (c is positive
constant, A is a predefined matrix with values manually set by me.

I know the correlation option in lme function can specify some kind of
correlation. but only with the Construct function defined, not whatever
ever form I want.

Any good ideas of how to do this in R ?

Thanks a lot in advance,

Keyan Zhao
Computational Biology and Bioinformatics program
Univ of Southern California
Email: kzhao at usc.edu

======================================
I.White
University of Edinburgh
Ashworth Laboratories, West Mains Road
Edinburgh EH9 3JT
Tel: 0131 650 5490  Fax: 0131 650 6564
E-mail: iwhite at staffmail.ed.ac.uk
======================================