2012 Elections Conference: Forecasting the Elections

Panelists: Jim Campbell (SUNY-Buffalo), Robert Erikson (Columbia University), and Drew Linzer (Emory University).

Dr. Campbell assessed the performance of some of the prominent forecasting models – including his own – of the 2012 presidential election. Dr. Campbell’s Trial-Heat and Economy model (developed in 1990) uses the trial-heat poll at Labor Day and second-quarter growth rate in the real GDP. His revised Convention Bump and Economy model (developed in 2004) substitutes the two pieces of information – the pre-convention trial-heat poll and the change produced by the party conventions – for the Labor Day trial-heat poll, and keeps the economic measure (real GDP growth rate) from the earlier model. This revised model has proven more accurate:

Click images to enlarge

The idea is that the fundamentals of the campaign are captured in these two measures in the following way:

Click image to enlarge

Dr. Campbell’s 2008 forecast (based on the Convention Bump and Economy model) called for Obama to win 51.3% of the two-party vote. This model is preferable, in his view, because it accounts for the fact that Obama had a more successful convention (received a larger bump) than did Romney. Campbell’s model tied for second-closest to the actual outcome (51.8% of the two-party vote for Obama) among thirteen 2012 presidential vote forecasts included in the forthcoming PS Symposium (note that DBE stands for “days before the election”):

Click images to enlarge

In Dr. Campbell’s view, what separated successful from unsuccessful forecasts of the 2012 presidential election was whether they included a public opinion indicator. Those that did tended to fare better, and those that did not seem to be omitting a major component of the election.

Next, Dr. Erikson discusses how economic conditions (specifically, income growth [objective] and economic perceptions [subjective]) have both generally grown more predictive of vote choice later in past presidential campaigns:

Click images to enlarge

However, according to Dr. Erikson, 2012 was one election in which objective indicators and subjective evaluations did not match. Objective indicators tended to predict a Romney victory, while subjective evaluations favored Obama:

Click images to enlarge

Finally, Dr. Linzer – who ran the VOTEAMATIC forecasting model during the 2012 campaign (which correctly predicted 332 electoral votes for Obama) – discusses his approach to modeling polling data. More details about his methodology can be found here, but the basic idea is to establishing a series of baseline state predictions (based on electoral history and national forecasts) and use state polling data to update and gradually improve forecasts within a Bayesian framework. In Dr. Linzer’s view, a problem with other poll aggregators is that their estimates have uncertainty bounds that are too large and/or that they are too volatile and responsive to swings in public opinion (which can be expected to happen over the course of the campaign, but usually do not persist until election day). A Bayesian model produces estimates that are more stable and have more realistic uncertainty bounds.

This entry was posted in 2012 Elections Conference. Bookmark the permalink.

Comments are closed.