Fitting PCs with nPC vs CV

[JandehF]

Hi I am conducting FarmCPU GWAS and would like to understand the difference in fitting PCs using nPC vs CV. I read in the rMPV manual (cran repository) that when using nPC FarmCPU, the PCs are fitted as fixed effects.Is this correct? My understanding of GWAS models is that we want to fit PCs as covariates to ensure association between SNPs and trait are not confounded by pop structure.

My question is which argument should I use to add principle components to the model (nPC.FarmCPU or CV.FarmCPU)? I noticed associations slightly changed when performing GWAS with the 2 different arguments. Ive attached images of the resulting Manhattan and Q-Q plots.

Thank you in advance for your help.
Kindly,
Jandeh

[YinLiLin]

Thank you for the question.
In fact, it should not have significant impact on the GWAS results between these two settings, did you check if the order of individuals in phenotype, covariates, genotype is fully consistent?

[JandehF]

Hi yes I have double-checked that the order of individuals in consistent. I just noticed I am running into a problem with coding when I double checked.

If I add PC as covariates using model.matrix, I end up with an extra variable in covariates (seems to be adding an intercept variable). I think this is what led to the results in my original post and is not the correct coding for my needs.
PC<- read.delim("NAM_PC0.15het.txt",head=FALSE)
Covariates <- model.matrix(~as.factor(PC1)+as.factor(PC2)+as.numeric(PC3)+ as.numeric(PC4), data=PC)

If I add PC as covariates using bigmemory::as.matrix (as outlined on the main GitHub page). The first row of PCs is removed and I end up with 1 less row than that total number of individuals so the GWAS analysis will not run.
MVP.Data.PC("NAM_PC0.15het.txt", out='NAM', sep='\t')
covariates <- bigmemory::as.matrix(attach.big.matrix("NAM.pc.desc", head=FALSE))

If I adding PC as covariates by just attaching as a matrix, the PC matrix seem so upload with no problem. The associations are still different from using nPC but closer than my first run.
PC<- read.delim("NAM_PC0.15het.txt",head=FALSE)
cv <- as.matrix(PC)

My questions are:

1. What is the statistical difference in considering PC in covariates vs. as a fixed effect?
2. what is the best way (coding) to upload covariates to fit in the GWAS model?

[YinLiLin]

Thank you for the detailed descriptions.
In rMVP, the PCs are fitted as covariates rather than fixed effects. And for PCs, we recommend using the argument 'nPC.XXX' to fit the model, it should be safer and more stable we believe.