Hi, I have a DGEList x containing genes in rows and samples in columns (multiple samples for same patient). I have no NA in my data because I used complete.case function(). I create the design matrix in this way:

design <- model.matrix(~0 + f1  + f2  + f3 + f4 + f5 + f6 + f7 + f8 + f9 + age + gender, data=x$samples)

where f are some features (in this case I have 9 features). These are just numeric vector not factor, so in the design matrix each feature has only one single column (equal for age). Instead gender is a factor (M or F). So in the design matrix it has 2 columns.

When I Call:

v    <- voom(x, design, plot = F)

it returns:

Coefficients not estimable: f7 f8 f9 age genderF genderM
Warning message: Partial NA coefficients for 17080 probe(s) I see that it's ok until the totale value passed to model.matrix are 6 no more.

Why ??

When I call: vfit <- lmFit(v, design) it returns the same warning and the columns corresponding (f7 f8 f9 age genderF genderM) in the vift$coefficents are only with NA. I see that passing 6 parameters in the model.matrix is ok, but no more 6.

The warning is telling you that the covariates are co-linear. It is impossible to identify coefficients for co-linear variables because of ambiguity between them. The more covariates you add the more chance one of them will be co-linear with the others.

In general, it is very difficult to get unambiguous results when including so many variables in a linear model. I have never seen myself an expression dataset for so many variables would be appropriate.