Election Outcome Forecasting Models

ELECTION FORECAST MODELS attempt to predict the outcome of elections. There are generally two main streams for election forecasts. First, there are survey polls in which survey respondents from a random sample of the voting population are asked variants of the same question regarding their vote choice. Most often referred to as the horserace coverage, the election forecasts generally vary as the campaign races toward Election Day. The other stream of election outcome forecasting models, mainly from political science reser-ach, focuses on using statistical methods to predict the election outcome.

This form of election forecast generally attempts to use predictors of the popular vote—congressional or presidential—to create out-of-sample predictions of the election of interest. The first stream, which utilizes polling techniques to track the horserace, is self-explanatory and examples are plentiful, using forecasting models to generate predictions of the outcome using statistical methods.

Though different models specify the outcome as a function of independent variables, the dependent variable is usually the same in all forecasts. The dependent variable in the econometric election forecast models is generally specified as the proportion of the two-party vote for either a candidate (if examining presidential elections) or the party (if examining congressional elections). For congressional election forecasts, the out-of-sample prediction of the party's share of the two-party vote share is taken and then converted into a vote-seats translation equation. This equation will indicate how many seats will be held by either of the major two parties: Democrat or Republican. For presidential election outcome forecasts, the dependent variable, again, is a given candidate's share of the two-party, popular vote. This can lead to several problems as evinced in the 2000 presidential election. First, the models are bound to predictions of the popular vote, but the electoral system is not that simple. The popular vote, although a good indicator of which of the two major party candidates becomes president, is not how presidents are selected. The Electoral College vote is determined by the popular vote of each of the several states. As such, many of the models for the 2000 election predicted Vice President Al Gore's likely ascension to the presidency. While he may have won the popular vote, he lost the election.

Election outcome forecast models employ several independent variables to aid in prediction of the winners at the polls. While the variables differ across models and researchers, they can be grouped into several categories: approval measures, economic indicators, economic evaluations, and electoral cycle. For presidential elections, approval measures generally take two forms: a differenced measure and, more commonly, a simple approval percentage. The differenced measure is the difference between the percentage of survey respondents approving of the job the president is doing, and the percentage who disapprove. This measure of presidential approval, which is survey-based, can be taken from many different surveys, but one of the main sources has been Gallup poll data, given its consistency and frequency of tapping approval before, during, and after the campaign.

The second approval measure is the percentage who approve from similar polls, with the most common approval measure employed in presidential forecasts. Both types of measures are expected to have a positive relationship to incumbent party vote. In other words, if the Republican Party controls the presidency, higher approval ratings at the time of the election should lead to a higher proportion of the popular vote for the Republican candidate—whether it is the president seeking reelection, the vice president seeking the presidency, or an open seat presidential election.

...