Jay DeSart and I have done some work over the years on using September state-wide trial-heat polls to predict presidential election outcomes in the states. We pool state-level data from 1992-2008 and use the Democratic candidate's share of the two-party vote in state-level September polls (averaged across publicly available polls), a lagged (four elections) vote variable, and the Democratic candidate's share of the vote in national trial-heat polls in the last half of September to predict the election outcomes. While the other variables are important to our model, the September polling average is the strongest predictor. Jay has been good enough to update our model and provide forecasts for the 2012 election. Here's what it looks like:
 |
| National Two-Party Vote : Obama 53.1% Electoral College: Obama 332, Romney 206 |
Even if you don't use a complicated statistical model, September polls are a useful guide to who will win in the states. A couple of interesting tidbits:
- In our data set, from 1992-2008, there were 164 states in which one of the candidates had a lead outside the margin of error (based on the average statewide sample size) in the state polling average. The leader went on to win in 160 of those cases (98%).
- If you throw caution to the wind (ignore the margin of error) and make predictions just based on which candidate is ahead in the September poll average, the polls predict correctly in 220 of the 250 cases (87.5%).
