I am analyzing RNA-seq data using a hisat2 --> htseq-count --> edgeR pipeline. To do this, I am feeding CPM values generated by htseq-count into edgeR, which produces a "toptags" table, or list of differentially expressed genes with their associated logFC, logCPM, and p-value. I have noticed that often times genes that are differentially expressed in the toptags table don't seem to correlate to htseq-count's CPM values. For example, the toptags table generated by edgeR shows that for one gene there is a logFC increase of 0.73, however it doesn't look like there are any significant differences in CPM values between conditions- control: 766, 874, 881 compared to test: 706, 766, 1095. I am not sure what to make of this, and it seems like every time I find an interesting differentially expressed gene I have this same problem.
 

Thank you in advance for your help.

For starters, edgeR requires raw read counts as input, not the CPM values - have a look at Section 2.7.6 of the edgeR user's guide. Supplying CPM values will break edgeR's attempts to model the mean-variance relationship, especially at low "counts". Applying TMM normalization to CPM values won't make much sense, either, which might explain the inconsistency with the log-fold changes.