You might be interested in this post:
The first time I compared raw reads counts to RSEM’s expected counts, I encountered an unexpected trend: the expected counts were not slightly lower than the raw counts, they were consistently lower by a factor of 2. After thinking about this a bit, I considered the possibility that RSEM treats each pair of reads as a single unit given paired-end data. To confirm this, I selected a small subset of 10 million read pairs from one of my samples and ran RSEM twice: once in paired-end mode, and once in single-end mode disregarding the read pairing information. Consistent with my hypothesis, the expected counts produced in single-end mode were almost exactly 2 times to the expected counts produced in paired-end mode.