What follows is a brief summary of the model and a discussion of some of the issues both Paul and Damon see in it. As Paul said, "The forecasts in the paper are really preliminary. However, if we wait a few months, till we've re-specified the model, it won't be a forecast anymore." Questions, comments and concerns can be left in the comments section. I will forward them to Paul and Damon.
There is no shortage of presidential election forecasting models, academic or otherwise. In 2008, there are at least 15 political science forecasts, the average of which shows Obama winning approximately 52% of the two-party vote. Most rely on some combination of economic factors, presidential approval and/or incumbency to explain vote shares in presidential elections. Those factors are completely national in scope and what is lost in the process are many of the relevant state-level variables that could play a role in determining the electoral outcome. To be sure, there are also forecasting models that include state-factors, but what Paul Gurian and Damon Cann have done is to draw a distinction between the long- and short-term, state-level influences. [You can view their forecasting paper here.] In much the same way that the past polls in FHQ's weighted averages serve as an anchor to the short-term fluctuations in state polling, the long-term factors included in this forecasting model allow for historical, state-level factors to serve as a baseline of sorts for their forecast.
Those same national factors, then, are included, but are buttressed by short-term, state-level impacts (state primary divisiveness, home state, home region, etc.) as well as some of the more historical, state-level influences (state partisanship and ideology, etc.) that play a role in explaining the variation in the shares of the two-party vote. [A more thorough description of the state-level factors can be found on p. 6-7 in the paper linked above.]
The beauty of this is that you get 51 different forecasts, not just one on the national level. And that is certainly more suitable to the electoral college system. Based on the included variables over the last fifteen presidential elections, a projection of the two-party vote in each state can be made. The results can be found on p. 10-11, but a map of those results is included below. [No, I can't help myself. I have to include a map.]
[Click Map to Enlarge]The result is a rather close outcome between John McCain and Barack Obama. The line between a solid and a toss up state is whether a state's division of the two-vote is within the margin of error. You'll no doubt notice that there are several states that are on opposite sides of where they may be expected given other forecasts and projections. Iowa, New Hampshire and New Mexico, for example are shaded in red while Arkansas, North Dakota and West Virginia appear in Obama's column.
Here are some caveats that Damon adds:
A few thoughts on the states:
NV: I think the "home region" variable swings the prediction for NV
toward McCain more than has actually happened in this instance.
Without that, McCain would still be in the margin of error.
AR and ND both had strong Democratic showings in House and Senate
races in 2006/2004, probably stronger than past history would suggest
for those states. Plus AR has the Democratic history from the "old
south" and our fixed effects may be picking that up a bit with the
1948-1970s elections.
I think NH is just a matter of history--while they went for Clinton in
'92 and '96, prior to that they only went to a Democrat once, Johnson
in '64. While NH has been battleground recently, our fixed effect
(based on all elections in the sample) moves NH just outside the
margin of error.
FL is probably similar. Like NH, most of the variables for 2008
suggest it ought to be perhaps R leaning but still battleground.
However, the fixed-effect for FL slides it about 2 points closer to
McCain.
I re-ran the model dropping the fixed effects, but that decreases the
general predictive power of the model by about 10%, seemingly
generating more error than it would eliminate.
Also, thinking about this statistically, since our margin of error is
based in the 95% level of confidence and we're making 50 forecasts, we
should actually expect to see 2.5 (OK, let's call that 2-3) of our
predictions that are significantly different from 50% by sampling
error alone. But since these errors are random, they should cancel
each other out in the EC tally (as long as it's not CA that is one
error and WY as the other).
I finally re-ran the model using national fatalities per 100,000
rather than state-level fatalities. The coefficient still comes out
insignificant statistically.
I want to thank both Paul and Damon for sharing this and I hope that we can get a good discussion going that will generate some helpful feedback.
Recent Posts:
The Electoral College Map (10/30/08)
Liveblog: The Obama Infomercial
Update(s): The Electoral College from a Different Angle
