What Is The Lod Score Replication Threshold For Linkage Analysis?

Question

[K_Star]

Hi, with regards to linkage analysis of complex traits, what is the recommended threshold for significant replication of a previous result? I have seen references that suggest that an LOD = 1.4 is accepted as the threshold for significant replication, referencing Lander & Kruglyak 1995 as the source of this threshold. However, no-where in the paper is there a mention of LOD = 1.4. Any suggestions/recommended reading?

Thanks

K

[Larry_Parnell]

In which species are you interested? The value is rather different for human and in-bred mice, for example. In which trait(s) are you interested? The LOD scores for something like hypertension is vastly different than for differences in gene expression, where LOD scores of 20 and up are routine. Actually, this is a genetics question and not strictly bioinformatics. Nonetheless, if you can provide a few more details, I or others may be able to help.

[K_Star]

Thanks for the response Larry. I am interested in quantitative traits in humans, bone mineral density in particular.

[Larry_Parnell]

You're welcome. Just a reminder that if you find a response useful, you're encouraged to vote it up so that others who have a similar question are guided to potential solution(s).

[K_Star]

Thanks for the update Larry and reading suggestion. As I am focusing on complex traits in humans, the 2.5-3.0 range would be more suitable. Still getting used to the BioStar platform!

Answer 1

[David Quigley]

The 1.4 claim probably comes from the assertion made by L&K that "Accordingly, P=0.01 should bbe required to declare confirmation at the 5% level. Note that this correction is equivalent to a multiple testing (or Bonferroni) correction for 5 markers." If you check figure 1, a P value of 0.01 will approximately line up with a LOD score of 1.4.

That said, think about the conditions L&K place on that calculation: they assume that you have done a linkage analysis with sparse markers, probably in an animal model, and that the 95% confidence interval for your region is about 20 cM and includes five markers. That's where the 0.01 comes from. Using this framework, if you test only one marker, in theory, the replication P value could be higher; if you test more, the required replication P value should be lower. In practice, the required threshold for credible replication depends on your effect size and the size of the replication population. If you were doing a classical sort of experiment (e.g. scanning for a QTL for rice grain size, then doing a follow-up experiment to test your best candidate) then this advice would probably still be good. These days, though, many people are testing many phenotypes at once (e.g. eQTL studies) so in my opinion you need stronger evidence of replication because the multiple-testing challenge is greater.

For recommendations on whether an association study is credible your go-to quote usually comes from John Ioannidis. I recommend Validating, augmenting and refining genome-wide association signals (Ioannidis NRG 2009) as a place to start.

There is a great write-up of this issue on the genomesunzipped blog. They reference a review by Visscher et al which is also worth reading.

[K_Star]

Thanks for the response David and for clearing that up for me. One of the publications that I was reading referred to the p-value of 0.01 for replication, which, as you suggested, aligns with 1.4 in the figure and L&K refer to the 0.01 value to 'declare confirmation'. I thought the value of 1.4 was a bit low also, especially considering that suggestive linkage is 1.9 even if the assumptions change with a replication study. Thank you for the help and the recommended reading.

k

Answer 2

[Larry_Parnell]

An eQTL (expression QTL) study from Dixon, Cookson et al (2007) used a minimum LOD cutoff of 6.0 to assign SNP marker to mRNA level. (They observed several instances of SNP-mRNA level associaiton exceeding a LOD of 40.) For more complex phenotypes, such as blood lipids, I've seen a LOD cutoff of 2.5 to 3.0 when 400 microsatellite markers were used and no reference to animal traits/markers.