Hi,

I am trying to find a list of genes significantly associated with a continuous variable (e.g. glucose) so I used edgeR to create a design matrix as follows:

design <- model.matrix(~glucose)

​Followed by: 

disp <- estimateDisp(y, design, robust = TRUE)
fit <- glmFit(disp, design, robust = TRUE)
lrt <- glmLRT(fit, coef = 2)
topTags(lrt, n=Inf, adjust.method = "BH", p.value = 0.05)

to give a list of genes associated with glucose.

In addition, I wanted to adjust for gender, so I included gender as a factor.

design <- model.matrix(~glucose+gender)

However, doing so resulted in gender-related genes (coming from Y chromosome) coming up as significantly associated. At this point, I am not sure what's going on.

Separately, I tried out limma-voom to replicate what I have done above.

limma.design <- model.matrix(~glucose+gender)
limvm <- voom(y, limma.design, plot = TRUE)
fit <- lmFit(limvm, limma.design)
fit <- eBayes(fit)
topTable(fit, coef=2, adjust.method = "BH", p.value = 0.05, number = Inf)

which I think should have done the same thing, but instead there are no gender-related genes, and the number of genes are much less. To find glucose associated genes while adjusting for gender, am I doing this correctly? Thank you.

To make the limma-voom analysis similar to the edgeR analysis, you would use

fit <- eBayes(fit, robust=TRUE)

since you've done the equivalent for edgeR.

It is very hard for a linear model to fully adjust for gender effects, because the Y chromosome genes don't have any counts for females, so making it hard to estimate the baseline. In this situation, I find it better to explicitly remove the sex-linked genes (XIST and the Y chromosome) from the analysis, rather than trying to do the adjustment in the linear model.

Alternatively you can just trust that the robust=TRUE method will down-weight any sex-linked genes automatically for you (when gender is not included in the model). It seems it has already done this. To understand why this is the case, have a read of the case study in this preprint:

  http://arxiv.org/abs/1602.08678

Note, this assumes that you remove gender from the design matrix.