In my experience, the length of the contest affects the calibration of the problem rating (at least in terms of implementation difficulty). I think it's because in a longer contest more people will have time to solve it so the user rating of the average solver will be lower