Presidential Election a Victory for Quants
If there was one lesson for political pundits from last week's presidential election, it was that basic statistical modeling techniques can be used to predict election outcomes with stunning accuracy.
Wed, November 14, 2012
Computerworld — As far back as June , Drew Linzer, an assistant professor of political science at Emory University, predicted that Obama would win reelection and secure at least 52%of the popular vote. Like Silver, Linzer also had Obama winning 332 electoral college votes and Romney taking home 206 votes.
Even as political pundits breathlessly forecast a tight race, Linzer's blog site Votamatic consistently had the president winning the election by a small but comfortable margin.
"I never saw it as being a close race" Linzer said speaking with Computerworld this week. "When I started producing my forecast in late May, the historical model that I was using showed that Obama would get about 52% of the major party vote."
Despite minor fluctuations in support for both candidates in the weeks leading up to the elections, such as immediately after the first debate, the data always showed Obama winning in the end, he said.
Linzer, like Silver, made his forecasts by aggregating state-level poll data with economic indicators and data from previous polls. He started by constructing a baseline forecast for each state by using a statistical model developed by Alan Abramowitz, a fellow Emory professor , who used the model to predict the outcome of the 2008 elections.
The model, called Time-For-Change, predicts the incumbent party candidate's national vote share by looking at factors such as the president's approval rating in June, the percentage change in gross domestic product in the first two quarters of the year, and the number of years the incumbent party has held the presidency, Linzer said.
Historical data shows that these measures are especially useful indicators of how a first-term president will fare in the elections, Linzer said. For instance, since 1948, presidents that have been popular in June have been much more likely to get reelected in November, he noted.
As the weeks progressed, Linzer began basing his forecasts increasingly on state-level opinion poll data and less on the historical data that he had used to build his baseline model. "When I started off in May and June, the forecasts were based on long-term fundamental economic and political variables," because there was little poll data available at the time.
As more poll data became available, it was thrown into the mix. "The basic idea is that on Election Day or the weeks leading up to Election Day, the polls are the best indicator," of an outcome, he said.
One of the mistakes that many pundits were making was to look at national-level poll data to predict the outcome, Linzer said. National polls often are unable to detect local trends and patterns with the same level of granularity that a state poll does. "In a national poll, I would only get a few respondents from an Ohio or a New Hampshire where the elections are being decided," Linzer said.