It doesn't seem very easy to have access to an old RefSeq release. Why do you want to work with this particular release and not the current one? You can have access to a gtf (or gff) file via the RefSeq website (https://www.ncbi.nlm.nih.gov/refseq/) or through the UCSC table browser, choose "RefSeq genes" as table choice and "gff" in the choice of file format. If you really want to have access to RefSeq 73, you can maybe send an e-mail to the RefSeq team. Good lcuk!

Thanks for the response. I needed specific number for the following reason.

Our sequence provider did the initial mapping of RNA-Seq reads to refseq-73 transcriptome and converted it to hg-19. As you may imagine this mapping is not good for several reasons including untrimmed reads, duplicated reads etc. 

Hence I did the mapping to hg-38 after all the needed pre-processing.

Now I am trying to compare gene counts between my mapping and old mapping. 

To get the counts I would like to have refseq-73 GTF. Will you suggest using hg-19 ensembl annotation?

Yes I tried getting refseq GTF from the original provider which they somehow don't have.

First, I think you make a confusion between RefSeq gene versions and genome versions. A gff file from RefSeq could have gene coordinates either in hg19 or hg38, independently of the RefSeq version; Now if you want to compare with your primary alignment, you don't have other choice than to retrieve a RefSeq 73 gff file from your sequence provider or from RefSeq, if it's still available.

Then, for RNA-seq, it is better to first map reads to a whole genome and not to a transcriptome. If a read matches perfectly on a intergenic region and with one mismatch on an exon, you'll discard it when aligning to a whole genome, but keep it when restricting alignment to a transcriptome (in most aligner settings), then over-estimating gene expression.

Finally, if you map your reads to hg38, you have to use a gff gene reference file with hg38 coordinates to perform reads counts per gene / transcript or whatever you want to use your gff for. Once you chosen a reference genome to map your reads to (hg19 or hg38) you have to stick with it along your entire experiment. But then again, on the UCSC talbe browser (http://genome.ucsc.edu/cgi-bin/hgTables), you'll find RefSeq gff files in both hg19 and hg38 coordinates. Good luck!

Yes I understand transcript coordinates are different from genome coordinates. Our Sequencing provider mapped to RefSeq-73 transcriptome yielding mapping in transcript coordinates. Then they used "rsem-tbam2gam" to transform it to "hg19-genome" coordinates. As you stated that their mapping is not reliable. It is my understanding that hg19 genome for different refseq release will be slightly different from each other. Hence using latest hg19 gene coordinates might not help I think?. Moreover, Refseq doesnot seem to archive gff files. They said "We only retain the 'text' report files from previous releases (but not the sequence records. The text reports are available as RefSeq-release##.catalog.gz (where ## is the release number) files at this location: ftp://ftp.ncbi.nlm.nih.gov/refseq/release/release-catalog/archive/ You can find the report for release 73 in this list." 

You're right, if you map your reads on the current hg19 and use the gff of the current Refseq (version 78, I think), you won't have the exact same coordinates than your previous mapping. This said, you're working with human, which is pretty well annotated from several years ago, so you might be able to compare your mappings with the same gff, or compare read counts (if you got one from your sequence provider) even if some genes won't have the same coordinates in both versions (although 73 is not so old). It's worth a try anyway!