Improving the Accuracy of Pre-election Predictions

It is hard to figure out what all those polls really show about who is likely to win the November 7 Presidential Election, but help is now available.

Yale University political science professors Donald Green and Alan Gerber, with statistics graduate student Jay Emerson, have found a way to filter out stray “noise” caused by sampling errors and inconsistencies. They have developed new software and posted it for free on the Internet.

Called Samplemiser (www.samplemiser.com), this software separates real change in opinion from meaningless fluctuation. It automatically weights new information against older surveys, dramatically boosting the accuracy of the results.

Samplemiser bases its calculations on a longer time period than most polls, which use only the most recent three days of data. By including information that goes back for a longer period, Samplemiser smoothes out the peaks and valleys that look significant, but are really meaningless. It also saves pollsters money by increasing the useful life of the data they collect and reducing the number of interviews needed for an accurate survey.

Furthermore, data from any poll can be inserted, allowing the Samplemiser user to get a more accurate reading of the CNN/USA Today/Gallup polls, or surveys done by Newsweek, ABC News, New York Times/CBS News and others.

Using Samplemiser to clarify the significance of changes in the candidates’ relative popularity as reported in recent polls, Green found, “It’s usually nine parts noise to one part change.”

Samplemiser uses a method called “Kalman filtering,” well known in electrical engineering since the 1960s, but never before used in the analysis of survey data, according to Green.

“In 10 or 15 years, everybody will be using this technique or another one like it. We’ll look back and wonder about how polls were interpreted without it,” Green said. Using this software, “You greatly reduce your uncertainty about where opinion stands right now.”

Green, Gerber and Suzanna De Boef, assistant professor of political science at Pennsylvania State University, originally proposed this method for reducing sampling error in a scholarly article, “Tracking Opinion Over Time,” published in the Public Opinion Quarterly, Volume 63, in 1999.