Hi all,
as a background my study involves two groups, each with 7 samples. Groups are based on breeding values so the differences may be quite subtle. However the issue is really with biological replicates. The variance in counts that I use for analysis are very large within groups. In some cases very low counts are called as DE due to one or two 'high' counts (~30) vs very low counts (~0). This is obviously fairly easy to remove: cut all genes with x number of counts < y.
I had an idea to determine variability of counts within groups using a metric like var/mean or sd/mean ('normalised' value of variation in counts). So the groups are individually tested for this metric and any gene not meeting a certain cutoff value in either of the groups is discarded (a conservative approach would require both groups to have a maximum value). I know this is cherrypicking genes but fundamentally is it not alright to do this? I amn't picking genes based on difference in counts between groups but based on variance withing groups. So is it ok to go on and use the dataset constructed using this method?
I hope this makes sense and there isn't a very obvious reason not to do this/previous discussion. The only other option to reduce variance in count numbers is to increase sample sizes but the obvious economic arguments are there.
Your thoughts on this would be very much appreciated=)
as a background my study involves two groups, each with 7 samples. Groups are based on breeding values so the differences may be quite subtle. However the issue is really with biological replicates. The variance in counts that I use for analysis are very large within groups. In some cases very low counts are called as DE due to one or two 'high' counts (~30) vs very low counts (~0). This is obviously fairly easy to remove: cut all genes with x number of counts < y.
I had an idea to determine variability of counts within groups using a metric like var/mean or sd/mean ('normalised' value of variation in counts). So the groups are individually tested for this metric and any gene not meeting a certain cutoff value in either of the groups is discarded (a conservative approach would require both groups to have a maximum value). I know this is cherrypicking genes but fundamentally is it not alright to do this? I amn't picking genes based on difference in counts between groups but based on variance withing groups. So is it ok to go on and use the dataset constructed using this method?
I hope this makes sense and there isn't a very obvious reason not to do this/previous discussion. The only other option to reduce variance in count numbers is to increase sample sizes but the obvious economic arguments are there.
Your thoughts on this would be very much appreciated=)