Read duplication may be natural (the same DNA fragment occurs and is sequenced twice) or artificial (during the sequencing procedure a copy of the same read is created and sequenced).
Some approaches are more sensitive to read duplication than others. I have also noticed that samples coming from labs with less experience with NGS library preparation typically produce very large rates of read duplications (80% or more!). Perhaps this is due to producing insufficient DNA that later needs to be amplified for the protocol.
My personal opinion is to investigate the duplication rates and remove them if there is indication that these are artificial ones (rates are way above what a natural duplication level would be). That being said very accurate ChIP-Seq type technologies (like ChIP-Exo) could produce very high rates of natural duplicates, often undistinguishable from artificial ones.
Looking at the read distribution around high duplication sites are a way to evaluate wether that location is naturally or artificially enriched. A natural site would exhibit a smoother distribution at the site, with roughly equal number of reads on both strands. An artificial site tends to show heavy imbalances by strand, with most reads being exactly the same rather than showing a distribution around the site.
I'd refer you to this reply from seqanswers by lh3. In short, for DGE analysis, I wouldn't remove PCR duplicates. There is no way of knowing if it is a PCR duplicate or it is just because of the number of fragments that were identical. Of course paired-end helps resolve this up to a certain extent. It would be even more unlikely to have a fragment that has identical start and end.
However, most of the pipelines constructed so far deal with removal of duplicates for SNP calling and not for DGE. And I think this is the way to go. But then, I also understand this could be very subjective.
Thanks Arun and Istvan Albert! Both paired-end and read distribution are helpful to sovle this issue at certain level. However, for meta-sequencing without references, read distribution is not possible. I spoke several guys in the field and there is no clear answer. I think the final asnwer may depend on a series of model experiments, including the estimation of several parameters like coverage depth, initial template amount and many others.