[R-sig-ME] semiparametric mixed models

[David Duffy]

a couple of recent queries about appropriate transformations of Y
variables for mixed models reminded me of the approach used in

http://www.bios.unc.edu/~lin/publications/2006/DiaoLin06b.pdf
http://www.bios.unc.edu/~lin/publications/2005/DiaoLin05.pdf

The only software I know of:

http://www.bios.unc.edu/~lin/software/SVCC/

It's a little bit different, AIUI, from the various spline based 
approaches that are available in R.  It gave pretty similar results to 
analysing Y after choosing a transformation based on a Box-Cox regression 
ignoring any clustering in my one major use of it (but it did do it 
automatically ;)).  The approach can also be used for mixed modelling of 
survival data, but we didn't get very far with their software for that. 
Maybe a little R project for someone?

Cheers, David Duffy.

PS Actually, one could probably use coxme() for this.

PPS Also some of the programs that fit a linear mixed model for genetic 
data allow one to include a power transformation (a la Box-Cox) in the 
likelihood to be maximized (PAP comes to mind).  I know of no evidence 
that this is a particularly good idea.  Folks from a biometrical genetic 
background will know that some variance components in standard models for 
pedigree data, eg genetic dominance, can be made to go to zero by such a 
transformation of the phenotype.


-- 
| David Duffy (MBBS PhD)                                         ,-_|\
| email: davidD at qimr.edu.au  ph: INT+61+7+3362-0217 fax: -0101  /     *
| Epidemiology Unit, Queensland Institute of Medical Research   \_,-._/
| 300 Herston Rd, Brisbane, Queensland 4029, Australia  GPG 4D0B994A v