Page 1 of 1

Mixed-models for GWA

Posted: Wed Jan 30, 2013 1:32 pm
by MariaG
Nat Genet. 2012 Jun 17;44(7):821-4. doi: 10.1038/ng.2310.

Genome-wide efficient mixed-model analysis for association studies. (GEMMA)

Zhou X, Stephens M.
Department of Human Genetics, University of Chicago, Chicago, Illinois, USA. xz7@uchicago.edu

http://www.ncbi.nlm.nih.gov/pubmed/22706312

"In summary, we have presented an efficient method for computing exact values of standard test statistics in linear mixed models (GEMMA). This method is comparable in speed to approximation methods such as EMMAX but yields exact test statistics. By analyzing two example data sets, we demonstrate the use of our method and show that the approximation methods can yield inaccurate P values when the sample structure is strong and/or when the marker effect size is large. We also find that approximation by EMMAX is more accurate than approximation by GRAMMAR (GenABEL) across the genome (a comparison made possible only by the availability of an efficient exact method)."
"...While this paper was in review, Lippert et al. also published an efficient algorithm for this model, implemented in the FaST-LMM software; the relationship between this algorithm and ours is discussed..."

Personally, I'm happy to see an affective exact method (GEMMA) applicable in case of population stratification. However, I'm surprised that approximate method implemented in GRAMMAR (GenABEL) shows much worse results in case of strong stratification comparing to GEMMA, and demonstrates more limitations comparing to another approximate approach realized in EMMAX. I would highly appreciate any comments about GRAMMAR's (mmscore() function) capacities and limitations, especially related to the presence of population stratification and relatedness.

I work on the same kind of complex sample already discussed here on Forum (mixture of family-based individuals, unrelated individuals, within the cohorts from 5 several European countries) :
http://forum.genabel.org/viewtopic.php?f=9&t=642
So I also wonder, which strategy is better to implement, and if I have to be careful with interpretation of GRAMMAR (mmscore) results?..

Maria

Re: Mixed-models for GWA

Posted: Wed Jan 30, 2013 1:50 pm
by MariaG
PS. Unfortunatelly I was not able to upload the corresponding .pdf file.
Do we have any limitations for uploading documents on Forum?
Is it possible to allow uploading .pdf at least here, in Journal Club?

Maria

Re: Mixed-models for GWA

Posted: Sat Feb 23, 2013 9:44 pm
by yurii
GRAMMAR != mmscore != FMM

As far as I understand GEMMA is similar to MixABEL::FMM (though they compute the genomic kinship differently from 'ibs'), and FaST-LMM when you do the Maximum Likelihood, and also to the EMMA (not EMMAX!).

mmscore is similar to FASTA, EMMAX, P3D, MixABEL::GWFGLS etc. All these are approximations of above methods.

GRAMMAR-Gamma is approximation to FASTA-like tests mentioned above

We also clearly show and say in the Nat Gen paper that GRAMMAR-Gamma does excellent work on human populations, but works worse for Arabidopsis (very strong structure). Our suggestion is even when using on humans check the accuracy of approximation in your particular sample before start using it.

best, Yurii