To remove bias in calculating variant frequencies, we tend to remove PCR optical duplicates in Illumina exome-Seq protocols. However, in targeted sequencing, we do not remove PCR duplicates.
I am not sure about Illumina truseq rapid exome sequencing library approach. Should the BAM file deduplicated before variant calling in GATK pipeline.
I would typically expect duplicates to be removed for Exomes, but it is hard to say what is absolutely right for all variants (for all target designs).
For example, you may notice weird behavior in GATK with very high coverage (say, >1000x or >10000x) regions, and/or shifts in variant frequencies as coverage gets saturated.
I would guess this is more of an issue with more targeted panels (or even a single amplicon). While it isn't precisely shown in this paper, you can roughly see some potential value in having a customized variant calling procedure in Figure 9 (at least in terms of some samples having a noticeably higher rate of novel SNPs with the targeted panel; as a percent, not absolute counts, I would be a bit worried due to trends like Figure 7), although there is also a decrease in sensitivity in that specific example. And, perhaps more relevant to your question, the Exome example wasn't as bad for GATK :)
Also, the duplicate removal is generally for array-hybridization enrichment (for amplicon enrichment, you can't remove duplicates).
