I have mentioned before that many political scientists eschew opinion polls and look more closely at election predictions made on the basis of what are called the ‘fundamentals’, that depend upon objective measures (GDP, disposable income, etc.). Of course, which model you choose can affect the results that you get.
I spent considerable time on Douglas Hibbs’ Bread and Peace Model but there are others. In this article, Larry Sabato compiles the predictions of many fundamentals models for the incumbent party’s share of the two party vote and gets an average of 50.2% in favor of Barack Obama. Hibbs comes in at the very low end of the distribution at 47.5%.
A more sophisticated analysis by Montgomery, Hollenbach, and Ward uses Bayesian model averaging. When applied to the presidential election, they get 50.3% in favor of Obama.
[Update: The EMBA authors have published their computer program code that does this and others can use this to obtain the result.]
Usually I think of models as being competing constructions, so averaging their predictions seems weird and wrong. I guess it might sort of make sense if you are treating each model as a linearly independent contribution to the popular distribution. Like a bunch of non-competing terms that you are adding on. Some time I will have to learn more about why that is even remotely a reasonable approximation.
I agree that it seems a little weird to average to results of models that use different methodologies. Each model is really a proxy for some unknown and undetermined factor that leads to voting in a particular way.
I guess the point of this kind of averaging is to say something about the variability of the quality of the proxy.
This is great, thank you very much, Mano!
It’s going to take a while to digest the Bayesian model averaging. Not to mention that ICEWS (Integrated Crisis Early Warning Systems) looks as interesting as the presidential election. For example, are the Greeks on the verge of insurrection? And if so, can the EU, ECB, and IMF impose more austerity?
I did not find a reference for the EBMA estimate-the link only covers up to 2008. It would be great if you could post it.
The authors have published their computer program code that does this and others can use this to obtain the result. I think that people like Sabato simply run the program and you can too.
Oops, sorry it is in the Sabato article.