If we have two RNA-Seq libraries and run tophat on each of them, then combine the resulting bam files and run cufflinks on that, will that produce the exact same result as combining the fastq files before running tophat? I know that it wouldn't be the same to combine the results after cufflinks, since a transcript may not be able to be built from reads in a single library, but combining reads from different libraries would allow it to be assembled. I'm wondering if there is something similar with tophat.
It depends on what you mean by combining.
If, for example, you are combining different lane / run of the same sample, then it should be the same whether if you merge the bam or fastq file. However, you might need to make sure the read group setting allow you to specify them as the same samples.
However, if your are trying to combine different samples, then you should not combine them before running tophat unless your main goal is the detection of novo transcripts. The problem is that if you merge the fastq before the alignment, you will lost the information of origin, e.g. you don't know if read A is from sample 1 or sample 2.
Now if you are trying to detect novo transcripts, Trinity does suggest the merging of fastq file before the denovo assembly for the reason you've mentioned: The novo transcirpt might only be partially captured in single library. So according to my experience (which was 2 years ago, might have changed now), to detect the novo-transcripts, you will merge the fastq file and try to construct the novo-transcripts, then you align the reads of individual samples back to the novo-transcript list to get per individual alignment info