I think the probability of changing the election is generally underestimated.

But the analysis that follows seems to suggest a different conclusion: (1) If you are ignorant, then your subjective probability of changing the outcome is more than people often think, (2) If you aren't ignorant, then your subjective probability of changing the outcome is very small (but, I note, it is probably comparable to the probabilities estimated by Gelman in the study linked here), and (3) the high-probability case is a bit like that where someone goes into a casino thinking he has a winning system and (correctly, given that belief) estimates a substantial probability of coming out richer.

It seems to me that the "informed" probability is the relevant one here. I'm not sure how to tell whether this is generally overestimated or underestimated; among people who have followed Coscott's link to Gelman's paper, I'd guess it's estimated reasonably well unless Gelman's analysis is bad.