Well, there's a combination of factors.
1. Unless they have been
specifically trained, most people are bad at providing an accurate probability (i.e. 95% likely) even when they have a clear view of the actual likelihood of an event.
2. Almost no-one keeps a written, objective track of their beliefs as the game progresses;
memory biases are well known.
3. The vast majority of players perform little or no post-game analysis ("lessons learned").
4. There is a significant amount of variance in Mafia which complicates any kind of analysis; the correct long-term decision can result in a local failure, and it's rare that someone plays enough games
5. It is extremely difficult to assign blame (as in causative effect) between someone acting suspicious and someone suspecting them. People who were objectively wrong to suspect someone have a pronounced tenancy to place 100% of the blame on them.
off the top of my head. What I mean is that when we do something like balance a setup, we are required to put a number on the benefit people can gain from scumhunting in that setup, and I feel this number is, in general, poorly justified at best; you may remember the statistics that towns were getting crushed in Large Games a while ago.
That said, although the majority of players probably overestimate themselves, I'd be pretty confidant most competent players do better than random over the long-term, and some players very significantly so.
Succinctness is pro-town.
Passion is inversely proportional to the amount of real information available. ~ Gregory Benford