On Predicting Presidential Elections

Unfortunately, as elections come and go, actual predictions of these models can be compared actual results. What do such modelers do? If their prediction turns out correct, they can rest on the result. If it is wrong, they can add more input parameters or adjust coefficients in their model until the results of the last election are explained and they can claim that new model "predicts" election outcomes.

Some other political scientists have ridiculed the election prediction business, by showing that the league or conference of the winner of the last World Series or Super Bowl or the quality of the Beaujolais harvest can be used to accurately hindcast previous elections. There have been only 13 elections since 1948. Hence, there could have been only 8192 possible outcomes of the elections. Thus, one in every 8192 random time series of any sort, uncorrelated with the election cycle, would provide an absolutely accurate predictor of past elections.

Using models to understand the relative effects of different influences on presidential elections is a valid and interesting undertaking. Extrapolating them for use in predictions without adequate specification of the inherent uncertainties is at best sport and entertainment.

Dr. Michael Lewis-Beck of the University of Iowa, one of the most prestigious predictors says of the upcoming Bush-Gore contest, "It's not even going to be close." He expects Gore to win with 56.2 percent of the two-party vote. George W. Bush should not loose heart quite yet. Lewis-Beck predicted in 1992 that G.W.'s father would win re-election with 57.5 percent of the Electoral College vote.

References:

• Kaiser, Robert, "Is This Anyway to Pick a Winner?" The Washington Post, May 26 2000, page A1.

• Greene, J. P., "Forecasting Follies: Why Political Scientists Can't Predict Presidential Elections," The American Prospect, 1993.

Go To Page: 1 2