[GenABEL-dev] faster polygenic

[Xia Shen]

Hi Yurii,

Attached is an updated version of polygenic_hglm.

- The reference list is now in a chronological order.
- Default convergence criteria is now 1e-6. In the documentation of @param conv, we say for great precision, use 1e-8.
- An example is provided showing how one can test the fixed effects parameters and how the standard errors can be extracted.

Cheers

Xia

-------------- next part --------------
A non-text attachment was scrubbed...
Name: polygenic_hglm.R
Type: application/octet-stream
Size: 5960 bytes
Desc: not available
URL: <http://lists.r-forge.r-project.org/pipermail/genabel-devel/attachments/20110417/462a2cfa/attachment.obj>
-------------- next part --------------


On Mar 22, 2011, at 8:43 PM, Yurii Aulchenko wrote:

> Hi Xia,
> 
> Many thanks! The code looks good to me, and examples are very nice!
> 
>> - Reference list;
>> 
> 
> I suggest we change order to chronological, so your paper is 1st, then
> Lee2001, Lee1996
> 
>> Perhaps the default 'conv' can be relaxed a bit. 1e-6 is good enough to me, and also 1e-8 might be two strict for some binary trait - getting a precision of 8 digits for binary data is too difficult, sometimes it might not even converge.
>> 
> 
> I agree. This will probably make it much faster. But we should tell
> that in the documentation -- like "if you want really great
> convergence, use 1e-8..."
> 
>> We are working on updating the 'hglm' package on CRAN, where I will make extracting internal variance-covariance matrix possible.
>> 
> 
> These are great news!
> 
> One thing which can make polygenic_hglm even more apparent news: I
> understood that for the fixed effects, you get standard errors of the
> estimates, correct? And these can be used for significance testing,
> right? E.g. SNP or haplotype significance (if one puts a SNP as
> covar). This would be extremely valuable, see discussion
> 
> http://forum.genabel.org/viewtopic.php?f=6&t=134
> 
> If easy, I suggest to pull fixed effects SE more close to the user,
> and provide an example of doing test for a SNP. I really think this
> will help polygenic_hglm to make to the news-lines -- and this is
> something quite unique [not in 'polygenic', not, to my knowledge, for
> FMM] :)
> 
> Yurii