Previously, we were calculating FDR for a given protein ids at peptide/PSM level but now a days many journals are asking to use FRD at protein level. Please share your thought/experience if you have come across these.
The trans proteomic pipeline is a great open source tool that will calculate the probability of random assignment at peptide and protein levels among many other useful things: http://tools.proteomecenter.org/wiki/index.php?title=Software:TPP. Alternatively, you can run your searches with a reversed or randomised database concatenated to your search database and use the identification of reversed sequences to calculate protein FDR ((number of reversed identifications / total identification)*100 = % FDR) FDR at protein level can make a big difference especially when there are peptide sequences that are redundant across many proteins.
Estimating error at the protein level just requires careful documentation of whatever process you use to move from initial peptide-spectrum matches to protein and protein group identifications. Everything else should be the same. Protein and protein group counts will be closely related to how you handle protein inference and rules such as requiring, for example, two peptides distinct at the level of primary sequence. Protein grouping is key because exactly how you group indistinguishable proteins has a significant impact on counts.