On The Validity Of The Regression Discontinuity Design For Estimating Electoral E ffects: New Evidence From Over 40,000 Close Races