Document View

A new polling method that incorporates the quirks and biases of the electoral system could give useful answers even in the closest-run races. Researchers have developed a method that converts polling data into a probability of a candidate winning in each state, which is then weighted according to the number of electoral college votes the state has, and added to give an overall probability of winning the presidency.

Frontiers

OPINION POLLS

"TOO close to call" is often the unsatisfying answer from pollsters in the tightest electoral dogfights. Now a polling method that incorporates the quirks and biases of the electoral system could give useful answers even in the closest-run races.

Traditional polls simply measure each candidate's share of the popular vote, but that ignores the way most elections actually operate. The presidential elections in the US are particularly complex - winning the highest proportion of the vote will not necessarily propel you into office, as Al Gore found when he lost to George W. Bush in 2000. To win the presidency, a candidate must secure over 270 of the 537 electoral college votes. These are allocated to states based on population, and winning a state means you bag all its electoral votes.

"The system is peculiar," says Edward Kaplan at Yale University. It is theoretically possible for a candidate to poll just 22 per cent of the vote and still win, he says. This would never happen in reality, but it illustrates how far the popular vote and electoral college vote can differ.

Previous attempts to incorporate the idiosyncrasies of the US electoral system into polls ran into problems because there are over 2 million billion possible combinations of state wins. Simulating all of these outcomes is complicated and error-prone. But now Kaplan and his colleague Arnold Barnett at the Massachusetts Institute of Technology have developed a method that converts polling data into a probability of a candidate winning in each state. This is then weighted according to the number of electoral college votes the state has, and summed to give an overall probability of winning the presidency (Operations Research, vol 51, p 32). Because pinning down each state restricts the calculations a bit more, the method avoids becoming swamped in complexity. The maths itself is relatively simple. "Once you see it, it's trivial," says Kaplan.

In close races, the outcome predicted by Kaplan and Barnett's method can respond dramatically to even small changes in the popular vote. For example, in the American elections of 2000, a small hike in Gore's support in late summer was hardly noticeable. But the researchers' model turns the subtle change into a huge jump in the predicted probability of Gore winning (see Graph). Although the method continued to predict a Gore win to the last, Kaplan believes that no system could have called the race correctly because of the vote-- counting farce in Florida.

Kaplan says the same principle should help provide more accurate polls in any country that doesn't use proportional representation. And the method would also be useful for campaign strategists, identifying key areas where small changes in the overall vote would lead to the biggest effect on the result. For the 2000 US election, the system flagged up Florida.

However, the beauty of traditional polls is that anyone can understand them. In contrast, Kaplan and Barnett's system may be difficult to grasp.

Kathleen Frankovic, chief pollster at CBS News in New York, says she thinks that giving a prediction in terms of probability wouldn't be too confusing. "More difficult would be explaining how you got to that conclusion," she says.