GLM_20110823

David, Ali, Ali's students, Martha, Liya, Ann

Overview of current development status, possible improvement
Continuous development of stepwise regression, snp-snp interaction, any bottleneck?
Progress of accuracy testing (single marker regression testing), plan for testing more complicated models
Getting code running on the new high-end GPU cards in the Lonestar system and exposing it via the Foundation API

The question regarding stepwise regression is how to set appropriate thresholds for getting SNPs in and out. Ann clarified that it is more important to get SNPs in.
- Ann suggested 0.002 / 1 E-5 for in /out (though Ann is still not clear on why stepwise regression is needed, we might want to just use forward selection for now); stepwise involves computing fit of the whole set of snps
- Liya suggested empirical values for getting programs to run (make sr workflow works for testing implementation)
- Liya suggested to consider R/QTL way for handling SR: specifying how many SNPs can get in and how many SNPs can be in the final model
- Ann suggested Permony for permutation to estimate 'good' threshold
Ali will send code to Liya for independent benchmarking and exposing to Foundation API
To have DE integration in mind, program needs to take model, thresholds, population structure, maximal number of SNPs, etc. as input options
Current approach is good for testing various ways to analyze data