In the winner-take-all world of politics, candidates know that even a modest lead in the polls can spell almost certain victory. Sheldon Jacobson, an operations research specialist at the University of Illinois, Urbana-Champaign, and colleagues, including a group of students, have attempted to quantify that insight for the current United States presidential election, putting their predictions for the Electoral College on a Web site, election08.cs.uiuc.edu. Using a statistical method known as Bayesian estimation, they combined an analysis of results from the 2004 Bush-versus-Kerry contest with current state-by-state polls for Obama versus McCain to produce probabilities for each candidate of carrying each state. They then converted the estimates into a probability distribution for the total number of Electoral College votes a candidate might receive. In Indiana, for example, polls as of 4 October gave McCain a slight 2.5% lead. But given that Bush carried Indiana in 2004 by 20.7%, a Bayesian calculation indicates McCain's chance of winning the state's 11 Electoral College votes at about 87%. Most states are now in the bag for one candidate or the other; only a handful are truly in Bayesian play. Current calculations give McCain no chance of victory. "However," Jacobson cautions, "if the polls move, then so will our forecasts."
Monday, October 20, 2008
Bayesian estimation on the presidential election.
From the "Random Samples" section of the Oct. 17 Science Magazine: