09-17-2013 03:08 PM
I am interested in the topic of dispersion estimates for count data coming from RNAseq data. The glimmix procedure provides the option empirical=mbn to correct for small sample size,and it appears to have good statistical properties in terms of type I error rates, however there is also the possibility to adjust the denumerator degrees of freedom for the F-tests with the KenwardRoger method. Does anybody have an idea what the best choice is? In our experiments, we typically only have 3 biological replicates for each treatment.
09-24-2013 01:57 PM
I would go with the empirical= option for this, unless you have a lot of missing data. If you are on SAS/STAT 12.1 or higher, consider ddfm=KR2, the improved Kenward-Rogers adjustment.