I'd advocate calculating the statistical margin of error [1] before the election (which might, in fact, be +/-2% if the election is small enough) and doing the coin flip if it's within that margin.

[1] Technically margin_of_error(votes_cast).