I think the standard is at least correcting for the exact number of tests you did for each SNP marker, for each trait. Any observed significant associations would be more reliable if they replicate in independently enrolled sample collections. Finally, the genome-wide significance threshold (P < 5 x 10e-8) provides a good guide to assessing the strength of your association. 

Do find attached some papers which may be of help.

Dear Jacob,

There are multiple points to be addressed in your question.

Before starting, what do you mean by "pool all analyses above into one"??

First of all, what you are calling "Recessive" and "Dominant" tests are not correct. For a recessive test, you contrast Aa and aa against AA (i.e does one copy of the 'a' allele result in different phenotype compared to two copies of "A"?). For the "Dominant" test, you should contrast Aa against AA and aa. Of course, in these contrasts, you have to use the average of the two genotypes (e.g [Aa + aa]/2 for the "recessive" test).

Second, you could summarize these 3 tests into 2: "Additive" and "Dominance". The "dominance" test was defined above. For the additive, you contrast AA against aa (just like one of your 3 tests). Therefore, these would be the coefficients used in your contrasts, given the order AA, Aa, aa:

Additive: -0.5 0 0.5

Dominance: -0.5 1 -0.5

Some researchers can be more stringent and say that you cannot use 3 contrasts if you have only 2 degrees-of-freedom (df), such as in your case. Therefore, this traditional approach using "Additive" and "Dominance" contrasts can be a way to get around it. In addition, the "Dominance" contrast tests both questions about "Recessive" and "Dominant". If There is significant dominance, you can look at the direction of the effect. Then, perhaps, test the contrast between the closest homozygous genotype.

Now, moving to your cohorts... Are you using data on ALL cohorts simultaneously in the model? I believe you are splitting the data, is that the case? Stratification used to be a problem because we used not to be able to efficiently identify the sub-populations within a population (in Humans - in livestock animals that is easy since we use pedigree). Since you have information on these 4-5 sub-populations, you can easily fit the effect of "sub-population" in the model. This would make your model account for this source of variation. In addition, if you believe that SNPs can have different effects depending on the sub-population, you can add an interaction term between SNP and sub-population. For exemple, you model could be written as:

Phenotype = µ + Sub-Population + additive + dominance + Sub-Population*Additive + Sub-Population*Dominance + error

Given homogeneity of variance, this model you increase the power to detect associations between SNPs and phenotype.

Now, trying to answer your question. By definition, you should use FDR correction for ALL tests you perform. Most people test each SNP once (for each trait). In this manner, you would have as many tests as SNPs. In your case, you could potentially have two tests per SNP: The 2-df test that includes "Additive" and "Dominance" and the interaction term that has 2 x 4 = 8 df. Therefore, you would have twice as many tests and the number of SNPs (2 x 20 = 40 tests). Then your FDR could be based on that. Also, you should use FDR within phenotype.

Does this make sense to you? Nick

You will not get three different p values for one SNP. You will get only one p-value and by default it will be for the minor allele (unless you opt for otherwise). This p value will be either for additive or dominance effect. You can imagine dividing your phenotypic data into three genotypes for each SNP and then fitting linear model for additive or fitting bi-modal model for dominance, and getting p value for the fitment curve. From practical point of view the dominance model may not have applicable value for selections, apart from adding it to your discussion in a publication. So, if at all you want to handle both these effects then it should be done separately as these effects are calculated separately and not pooled. Statistically they are exclusive of each other. Also as you have 20 SNPs, you will correct the p value for 20 ranks with p-value of 0.05 or 0.01 as per your preference to get q value.

Thank you all for taking time answering this question.

I’ve also talked to a statistician / epidemiologist, since its difficult a include all aspects in a short text on researchgate.

I believe a central question is: which analyses are primary to what I want to answer?  In other words:

a) What are my primary hypothesis I want to report? (central)

b) Which analyses are performed to elaborate on the primary hypothesis? (nice to know)

c) Are analyses performed solely to ease comparison with other studies (not central to what I report)

My analyses fall into these three scenarios, and we’ve believe it’s overly strict to pool all analyses when correcting for multiple testing.

If a researcher was forced to do that, he would necessarily tend to avoid examining his/her data for different hypotheses, with limited use of the data as a consequence.

I have in fact examined genotype data by comparing homozyg WT with heterozyg, homozyg variants and variant allele carriers, respectively, giving me 3 p-values per SNP. I may try to analyse for additive effect only, in future studies.

Nick: thanks for correcting me on the recessive inheritance.

Chiea CK: thanks for the attached publications. Useful reading!

Prasad: I tend to agree that dominant and recessive effect analyses exclude each other. However, I did choose to include both of these analyses when calculating FDR q-value.

Any thoughts on this are welcome.

Thank again,

Jacob