Pre-Election Polls

Accuracy

Pre-election polls in the United States have a generally good record of accurately forecasting the outcome of elections. According the National Council on Public Polls (NCPP), the average candidate error for national polls in the 2004 presidential election was 0.9 percentage point; in 2000, the average error was 1.1 percentage points. Polls in state-level races were also very accurate. The average candidate error in 2004 across 198 state polls reviewed by the NCPP was 1.7 percentage points.

But inaccurate polls attract a great deal of attention and tend to be remembered for a long time. Modern polling's most spectacular failure occurred in 1948, when public polls incorrectly forecast the defeat of President Harry S Truman by Thomas Dewey. The problems that led to the 1948 polling debacle were soon addressed, and it is highly unlikely that an error of that magnitude could occur again.

History

Election polling has a long history that precedes the development of modern probability sampling. Straw polls of the 19th century were a popular means by which public opinion in elections was gauged, and an early 20th-century magazine, the Literary Digest, conducted a very large and widely followed straw poll in several presidential elections through 1936. The Literary Digest poll's spectacular failure in the presidential election of 1936, in which it incorrectly forecast the defeat of President Franklin Roosevelt by the Republican candidate, Alf Landon, by a large margin, discredited straw polls. Despite having more than 2,000,000 respondents, the self-selected nature of the mail survey respondents and the fact that the sampling frame was biased toward more affluent voters led to a gross over-representation of Republican voters in the Literary Digest sample. A Gallup poll based on a sample that more closely conformed to the principles of probability sampling was quite accurate in the 1936 election and helped to affirm the legitimacy of modern polling methods.

Purposes

Conducting pre-election polls entails many decisions, each of which can affect the accuracy of the poll: (a) the timing of the poll, (b) the sampling method, (c) the determination of likely voters, (d) the trial heat question employed, and (e) the choice of other measures to be included in the poll. Pre-election polls are conducted throughout the election cycle but are considered appropriate for forecasting the election outcome only if taken very close to Election Day. But pre-election polls are conducted for many purposes other than forecasting. Some are conducted at the beginning of a campaign to provide information about the public interest in the election, awareness of the [Page 606]potential candidates, voter views on issues including the importance of different issues, and the likely receptivity of the voters to different messages. Polls conducted during the campaign can provide all of this information as well as reactions of the public to the candidates and events in the campaign. Polls in the latter stages of the campaign are used by the candidates and parties in making decisions about how to allocate campaign resources geographically or in targeting different demographic groups.

...