Polls and elections

Richard Morey points us to this article. We posted Yair’s take on the polling problems here.

But what really amused, or upset, me, was what I encountered when following the link on that page which read, “Smarter poll could call the closest races.” That headline set off some warning bells–the closest races are the ones that can’t be forecasted! I followed the link to an article that I couldn’t read without a subscription, so I found it through our library and tracked down the original research article, “A new approach to estimating the probability of winning the presidency,” by Edward Kaplan and Arnold Barnett, professors of management at Yale and MIT. The article appeared in 2003 in the journal Operations Research and is pretty misinformed. It’s bad in so many ways, and the also, annoyingly, call their method Bayesian. Huge amounts of detail on essentially trivial algebra and a complete misunderstanding of elections. The sad thing is, there are excellent quantitative political scientists at both Yale and MIT–if these guys had just walked over a few buildings and asked for help, they could’ve been spared this embarrassment. Seeing this stuff just makes me want to barf: it’s just not that hard to do something reasonable, and I hate the way they put in all this algebra for what are straightforward simulations of a probability distribution. (But ya gotta give the publicity office of Yale or MIT credit for getting this mentioned in the popular press.)

I hope the paper that Kari and I are writing will clear things up.

P.S. I have nothing against these guys personally. It’s the kind of thing that can happen when you come into a field from the outside and don’t know who are the right people to talk to. I’m sure if I tried to write a paper about business management, it would be equally silly.

3 thoughts on “Polls and elections

  1. I'm sure if I tried to write a paper about business management, it would be equally silly.

    Yes, but that extreme level of silliness may blend right in in the management literature.

  2. I have a feeling the blame for this one goes to both the authors and their editors; the former for underestimating the likelihood of polling to be biased (assumed away with "random sampling"), and the latter for either ignoring the unnecessary probability exposition business or for some reason insisting on it.

  3. AT,

    Yes, exactly. Without the unnecessary exposition the paper would've been 2 pages long, and then maybe the inappropriateness of the model would've been clear.

Comments are closed.