We have RNASeq data from 15 conditions, for which many contrasts but not all pairwise contrasts are of interest.  

When we run the dispersion estimate across all 15 conditions, perform the fit and extract one of contrasts of interest, we obtain much less significant DE genes than if we are first subset the data by the 2 condition use for the contrast and perform an dispersion estimate based solely on the data used. 

Question is the following:

1-Wouldn't it make sense to estimate the dispersion on a per contrasts (or even per condition) basis? Granted we lose samples in the dispersion estimate, but given that the conditions are so different, we don't see why we would estimate it altogether.  

Dispersion estimate using all 15 conditions:

Dispersion estimate using just 2 conditions for the contrast of interest: