This year's American Political Science Association (APSA) conference was set to be in New Orleans this weekend. Suffice it to say, that didn't happen.
It's a shame, not least because one of the panels was going to highlight five new studies on how to forecast American presidential and congressional elections, all of which are highly relevant to 2012. But just because those studies' authors didn't get the chance to present their work in person doesn't mean we can't take a look at it here, at least at the four papers we've obtained from the authors (the author of the fifth is in the process of updating his, and we'll post it when it's ready). So how do the models work, and what do they predict for this year's elections?
Who did it: Helmut Norpoth and Michael Bednarczuk (SUNY, Stony Brook University).
The variables: Uniquely among the studies here, Norpoth and Bednarczuk do not include any economic variables. Instead, they simply use the share each major party candidate got of the primary vote in the New Hampshire primary. To account for presidents who run unopposed, they set a maximum possible vote share of 65 percent, and a minimum possible share of 35 percent. That is, if a candidate got between 65 and 100 percent of the vote, that'd be inputted as 65 percent, and if a candidate got between 0 and 35 percent of the vote, that'd be inputted as 35 percent.
How well it does: The model has an adjusted R-squared value of 0.89. In plain English, that means that it explains about 89 percent of variation in party vote shares between presidential elections.
What it predicts: The model forecasts Obama will get 53.2 percent of the two-party vote and Romney will get 46.8 percent. This gives Obama an 88 percent chance of winning.
The variables: Lewis-Beck and Tien's "Jobs model" uses five variables: whether or not an incumbent is running, the incumbent's popularity rating according to the first July poll by Gallup, Gross National Product (GNP) growth between the second quarter of the election year and the last quarter of the previous year, growth in jobs over the first 3.5 years of a president's term, and one for the closeness of the relationship between the incumbent president and the incumbent party's candidate.
How well it does: The 2008 version of the model, the most recent one for which we have a full specification, had a standard error of 1.43 percent, and explained 94 percent of variation between president elections.
What it predicts: We don't have a copy of Lewis-Beck and Tien's latest paper, but the authors passed along their estimate that Obama will get 48.2 percent of the two-party vote share and Romney will get 51.8 percent, an estimate of which they have 77 percent certainty.
The variables: The model uses both polling data, obtained from Real Clear Politics, and the Conference Board's "leading economic indicators" (LEI) metric, which is based on a number of indicators including Treasury interest rates, unemployment claims, housing permits, manufacturing orders and hours worked, and stock prices. Specifically, they use LEI averaged through the first quarter of the election year and the incumbent's share of the two-party vote in polls in the first three quarters of the election year. At the paper's writing, polling data for the third quarter of this year was not available, so the authors only used data through July. They weighted this data less, as it did not contain voters' reactions to the conventions.
How well it does: The model's accuracy varies on how far one is from the election. One quarter out — or roughly where we are today — the model has an average error of 1.9 points and predicts 13 of the last 15 election results accurately.
What it predicts: The model predicts that Obama will get a 52.6 percent share of the two-party vote and that Romney will get 47.4 percent. It gives Obama an 80 percent chance of winning, overall.
Who did it: Alan Abramowitz, Emory University.
The variables: Abramowitz's usual "Time for Change" model uses just three variables: the president's net approval rating (approval minus disapproval) in June, the change in real GDP in the second quarter of the election year, and a "dummy variable" that equals 1 if the president is a candidate in the general election and 0 otherwise.
But Abramowitz noticed that the model was doing less well in more recent elections. He speculated that this was due to an uptick in political polarization in recent years. So he added a new "polarization variable" that equals 1 when an incumbent is running or the incumbent has a positive net approval rating, and which equals -1 when there is not an incumbent running and the incumbent's net approval rating is negative. The idea is to isolate elections where the party in power is viewed negatively by a plurality of the public.
How well it does: The model's average error in elections after World War II is only 1.1 percent, and the correlation between the model's results and the actual election results is an amazing 97 percent.
What it predicts: Abramowitz wrote the paper before last quarter's GDP revisions came in. With the most recent data, he says, the model predicts that Obama will get 50.6 percent of the two-party vote and Romney will get 49.4 percent. That's a very close election, and gives Romney a one third chance of winning, Abramowitz says. What's more, it's close enough that it's possible a candidate could win the popular vote without winning the electoral college.
Compare these to Nate Silver's model, which estimates that Obama will win, 50.8 percent to 48 percent, and has a 71.6 percent overall chance of winning, as well as to Douglas Hibbs's model, which estimates Romney will win, 53 percent to 47 percent. Overall, more models have Obama winning than not. But Hibbs and Lewis-Black/Tien should give Romney backers hope. Here are all the models together: