Based on an initial analysis of a TCGA RNA-Seq dataset(processed like microarrays), i would like to formulate a specific design matrix with limma, in order to perform a 2-factor type analysis. In detail, a subset of my phenoData object:

head(pdat.sel)
                sample_type age_at_initial_pathologic_diagnosis  msi gender
TCGA.AA.3662 adjacentnormal                                  80 <NA>      f
TCGA.A6.4105          tumor                                  79  MSS      m
TCGA.F4.6463          tumor                                  51 <NA>      m
TCGA.F4.6806          tumor                                  59  MSS      f
TCGA.A6.6650          tumor                                  69 <NA>      f
TCGA.AZ.6600 adjacentnormal                                  64 <NA>      m

colon_type <- paste0(pData(TCGA.RNASeqV2_eset)$sample_type,"_",

pData(TCGA.RNASeqV2_eset)$msi)
table(colon_type)

colon_type
adjacentnormal_MSI adjacentnormal_MSS  adjacentnormal_NA          tumor_MSI 
                 1                  2                 11                  8 
         tumor_MSS          tumor_NA 
                30                143

My main contrast of interest, is the tumor_MSI versus tumor_MSS (30 vs 8):

however, as you can see from the above information, a lot of cancer samples are not available (NA marked) for this information, and also a very small number of normal samples has this information (which is anyway irrelevant as i would not like to utilize at all any normal samples).

So, any naive approach like the following, which is first to subset my initial expression set of 195 samples:

ind <- sampleNames(TCGA.RNASeqV2_eset)[TCGA.RNASeqV2_eset$msi%in%c("MSS","MSI")& TCGA.RNASeqV2_eset$sample_type%in%"tumor"]

eset.sel <- TCGA.RNASeqV2_eset[,ind]

dim(eset.sel)
Features  Samples 
   20502       38 

mult_con <- factor(paste0(pData(eset.sel)$sample_type,"_",
pData(eset.sel)$msi))

levels(mult_con)
[1] "tumor_MSI" "tumor_MSS"

and then proceed:

age <- pData(eset.sel)$age_at_initial_pathologic_diagnosis
sex <- factor(pData(eset.sel)$gender)

design.2 <- model.matrix(~mult_con+sex+age) # include also these covariates to adjust for age and gender

fit <- lmFit(eset.sel,design) 
fit2 <- eBayes(fit,robust=TRUE,trend=TRUE)...

Or subsetting the initial eSet, is inappropriate, and another approach is less erroneous ?

As you probably already know, you can't have NAs in the columns of the design matrix. So you can either subset the data to those samples for which MSI status is known, or you can code the NA values of MSI status to some new arbitrary value. Either approach could be ok and neither is "erroneous". Subsetting is probably the easiest. It's up to you really.