Sincerest thanks to Charles Berry, David Smith and Douglas Bates for
pointing out that my model is severely overdetermined, by some 11,000
degrees of freedom, and that a design matrix for it would occupy over a
gigabyte of storage! Thanks also to Peter Sherer for steering me to "S
Programming" by Venables & Ripley which he says has two function to handle
regression of larger data sets.
Douglas asks if I really want to fit a fixed-effects model involving a
factor with over 1000 levels (my G factor). I think the answer is
'yes'. That factor represents individual spots on an array, each itself
representing a gene. I'm basing my analysis on that of Kerr & Churchill
(2001) PNAS 98:8961 (though the overdetermined model was entirely
doing!). Regression is their first step toward a cluster analysis that
includes bootstrapping to assess reliability.
-John Thaden
### My Original Message #####
I'm trying to fit 26046 observations with a linear model
y = mu + A + G + AG + CG + error
where A, G and C are factors of 22, 1185, and 10 levels,
respectively. Both lm() and glm() choke on this problem ...
************************************************************
John J. Thaden, Ph.D., Research Assistant Professor
Department of Geriatrics (501) 257-5583
U. Arkansas for Medical Sciences fax: (501) 257-4822
************************************************************
|