Correcting P-values for non-perfect SNP proxies (r2<1)? Possible?
Chris Medway
Tuesday, 08 September 2009 11:03 UTC
Lets assume there IS a highly significant functional SNP in Gene A associated with disease. If you were to geneotype this SNP or a complete proxy (r2>0.8) you would see this association (say….p=1×10-4). Now lets say neither the SNP nor a perfect proxy exists on the geneotyping platform – the best you have is a SNP of modest LD (r2=0.7). In this theoretical situation can you estimate the P-value you would get from this modest proxy?
I have been trying to find publications on this…with no joy. Whay do you think?
-
Replies
-
I think the sensible answer is “yes, but why would you want to”? The p-value is determined as much by the experimental design, so you’re calculating what statistic you would have got if you had carried out an experiment with a certain design.
I blogged on why p-values are evil last year.
What’s the real question (i.e. the biological one) that you’re trying to address?
-
Hi Bob,
Hmmm…well. This question first popped into my head while performing some in silico meta-analysis on several independent datasets. These were large genome-wide association study datasets looking at case/control SNP associations. All three datasets individually lacked power to detect small genetic effects after correction for multiple testing. My idea was to detect modest (p<0.05) but consistant SNP associations across all available datasets. However, as they used different genotyping platforms, the panal of SNPs were different and a good proxy (r2>0.8) was often not available. However if i could ‘correct’ for reduced LD, then a new threshold for scoring a modest association can be determined. I guess this is how I see this being useful..comparing SNP p-values across multiple studies where only poor proxies exist in an attempt to highlight SNPs that warrent further study.
-
Sorry for the delay in replying!
I had a chat to our resident QTL expert, and he suggested it might be more robust to combine pool data over intervals, so you use data from a few SNPs. He also mentioned this article, which should help with the combining p-values.
I wonder, though, if it might be better to use the estimated QTL effects, as these are a better indication of effect size. But this is a literature I should really read up on.
-
Sorry for the late reply, moving house at the moment! Thanks for the article, gonna have a think before I post a more scientific responce…
-
