[R-sig-ME] Kinship vs. Pedigreemm

[Joehanes, Roby (NIH/NHLBI) [F]]

Hi Jonas and David and lmers:

Yes, I noticed that I needed to make modifications to allow one observation per individual. That was done in my last patches, which I have submitted to lme4 bug tracker. However, I noticed that the new lme4 (as of SVN revision 1703) still has bugs in the glmer function. This prevents pedigreemm to work with the new glmer (even after my patches). However, my patches allow pedigreemm to work with the new lmer.

Indeed, ZStar allows us to specify any covariance relationship matrix (provided it is symmetric positive definite), but the matrix has to be Cholesky-decomposed first. As far as I know, ZStar is undocumented. From the code, it looks very much similar to pedigreemm(..., pedigree=list(...)) except that it allows custom relationship matrix.

I will check your code, David. Optim is used even in kinship. I am pretty much against optim since I found bobyqa routine in minqa package is much more efficient. Thanks a lot.

Sincerely,
Roby


On Apr 11, 2012, at 4:53 AM, David Duffy wrote:

> On Wed, 11 Apr 2012, Jonas Klasen wrote:
> 
>> So far as I know, in the pedigreemm package, there is only the internal 
>> function pedigreemm:::ZStar which can handle covariance relationship 
>> matrices. So there is no officially supported way for implementing a kinship 
>> in pedigreemm,or I'm wrong?
> 
> I thought that pedigreemm(...pedigree=list()) was the official 
> interface for including "the (left) Cholesky factor of the relationship 
> matrix".  But yes, it sounds like it doesn't yet work with the latest 
> lme4.
> 
> Cheers, David Duffy
> 
> P.S. For the curious, you can also see some code
> 
> http://genepi.qimr.edu.au/staff/davidD/Sib-pair/Src/sib-pair.R
> 
> for the LMM and probit-normal GLMM (function do_varcomp()).  It allows to 
> you to add an empirical kinship matrix for VC linkage analysis.  Fitting 
> is direct maximization of the likelihood using optim().  It is pretty slow 
> on large kindreds (no sparse matrices etc), but a proof it is pretty easy 
> to roll your own in R.  It does have the advantage of not complaining 
> about non-full-rank NRMs.