Judge mix at a time. Theres no hard and fast rule. I do it in stages anyway; first I get the mix elements levels and dynamic ranges under sufficient control that I know that I'm just going to shave off the odd couple of db here and there from a note or beat 'colouring outside the lines' as it were, and throw the limiter on and mix into it from there
Then carry on mixing forgetting about the limiter, and at some point (between twn minutes and ten days later) where I think the low end in particular (in metal invariably the loudest part of the body of a mix in shear dbfs) is all quite balanced and seperated within itself, but a bit loud in balance with the rest of the mix/spectrum (200hz or so and up), and I crank the mix into the limiter a bit to gell it up...ignore numbers here, listen for impact/slam, space/congestion/open-ness/density/cohesion/whatever) and try to forget about it for a while.
Unless levels drift and I start to hear eveything being too smashed, then revisit the limiter.
This all on Voxengo Elephant, btw, which owns the living fuck out of event horizon. I have both. I never use horizon. It distorts really easily compared to Elephant, and is very un-transparent/coloured by comparison. You can hammer into elephant like an absolute motherfucker, and it will brickwall, cutting 10db easy without distorting. But dont do that, cos its gay, and lazy, and sounds crap anyway. Final values, when I bother to look, are usually, for a dense, heavy mix, peak limit value will be around -4 to -7 or so, mostly the end behaviour of my limiter will be very similar to the when I first put it on; nominally chopping off about 2-3db, just doing it regularly rather than shaving off the occasional peak. The only thing thats certain is the peak: -0.2db. Set it there, leave it there, forget it exists. -0.1 is fine too. You really dont get any benefit from lower peaks; lower RMS is VASTLY more important for clarity/lack of distortion etc etc. I havent in the past, but since I learned of the importance of it, I do plan to do queit masters for streaming; they come out cleaner if they're quieter, after the compression. But in the actual wavs, use all the headroom for peaks, watch the RMS.
otoh, maybe I'm doing it horribly wrong. Never had any complaints so far (except blankplank saying that he has to turn down an EP I did for a local band after quieter masters; I was compelled to do that a bit loud, since the refrence material they gave me was hot and smashed as fuck. I dont like it, its gay, but it is what it is - these guys here: http://www.youtube.com/watch?v=L5b0XUmzah0
lots of things I'd do differently about that now, but the limiting is the least of it (the overall RMS could stand to be lower mind. I too can be Rick Rubin (without the money or beard). Live and learn).