Hi there, 

I ran Combat on 450k data for ~3000 samples and was pleased with the results. However, in order to get the script to run I had to omit any probes with NA values for any sample from the dataset. Due to this, if I want to use the Combat adjusted values for my downstream analysis, I will lose well over half of my data which is not ideal. 

I am wondering if someone could assist me in better understanding where I went wrong. Should Combat have run fine in the sva package with the NAs included, was this just an issue with my file? Or, is it true that if I want to use Combat I will unfortunately have to lose any probe with an NA value in any of my samples. 

Thank you in advance for your assistance. 

Unfortunately ComBat requires no NA values to run. One option is to impute the missing values using something like the impute package, and then apply ComBat. An alternative is the procedure you talked about. Finally you could fit a regression model to each probe with the batch term included in the model and use the residuals - which would allow for missing values. However, be careful doing downstream analysis with cleaned data like this  - you may just want to include the batch variable as an adjustment in your model matrix. 

Actually, ComBat can handle some NAs, but if you are missing all the values for a batch or treatment combination, ComBat will throw an error (often there is nothing to adjust for that gene in that case anyway). So the trick here is finding the genes that don't have too much missingness so ComBat will run. This might be tricky, so Jeff's suggestions are probably good ones in your case. 

Whoops definitely follow Evan's advice on that one. My mistake re: NA's in ComBat!

This is incredibly helpful, many thanks to you both.