If you map with 100% identity you can do that, otherwise it can indeed be risky, depending on what gene you have in hand (for instance if it belongs to a large and recently expanded gene family). the only problem is that you can't normalize into rpm for comparison among different libraries. to do so you need to count the number of reads mapping on the entire genome.

Thanks all! I'm relatively new to the RNA-Seq world, and yes, I'm just using published RNA-Seq data to quantify a gene's splice-variants in several tumor samples. I think I'll map the reads to the whole genome in the end so that I can get those expression values, but I've heard from one of my colleagues that RPKM and FPKM values are not so reliable and that RSEM's TPM is a better value to use. Lots of well-cited papers for all of them.

With regards to RPKM / FPKM your colleagues are right - but only under certain circumstance - namely if there is a large change in the distribution of gene/transcript length between your two conditions - else I would always recommend using FPKM/RPKM to avoid length biases in e.g. expression cutoffs.

Do not that the FPKM/RPKM measure It is however not a problem if you use tools such as cufflinks/cuffdiff (which corrects for this internally) edgeR or DESeq for count based differential expression analysis, since they rely on CPM/TPM.

With regards to mapping I would always map against the whole genome, especially if you want to do a count based differential expression since they require uniquely mapping reads count as input - which you cannot confidently provide unless you map against the whole genome.

Lastly with mapping tools such as Hisat (http://www.nature.com/nmeth/journal/vaop/ncurrent/full/nmeth.3317.html , http://www.ccb.jhu.edu/software/hisat/index.shtml ) it's probably so fast you will not notice the difference between a full genome and a single chromosome.