hi Xianjun, On Wed, Feb 5, 2014 at 11:02 PM, Xianjun Dong <xianjun.dong.umass@gmail.com>wrote: > Hi Mike > > I was comparing the DEseq2 output btw formula with and without some > covariance. And found that for a specific covariance, if I include it, most > of previously significant genes become insignificant, which is quite > confusing for me. I attached MAplots from such comparison. Do you have any > thought on this? Does that mean the ‘problematic’ covariance can explain > more difference between case and control then the condition itself? > > ​Yes this is one explanation. For example -- not that this has to be the explanation in your case -- in some experiments with case and controls not perfectly balanced across batches, one might see significance 'wrongly' associated with the condition when not including a batch effect, but when including the batch effect, the effect size for condition goes away. Adding covariates, in both GLM and simple linear models, can certainly change the effect sizes and test results. Variables might 'share' effect size in way that none of them are significant for a given critical level. Adding a covariate can even change the direction of the effect size for original terms. ​Mike​ > Thanks > > Xianjun > > [[alternative HTML version deleted]]