While I enjoy Neil deGrasse Tyson interviews on the Colbert Report, let me echo the sentiments made by Prof Gelman and Mark "Myster Pollster" Blumenthal: Dr. Tyson should stick to astrophysics instead of attempting poltical science. Last week Dr. Tyson proposed a simplistic model for predicting the state outcomes in presidential elections: taking the median of polling 40 days before an election. Gelman took Tyson to task not accounting for electoral variability between May and November as well as his condescending tone to political scientists. Blumenthal points out that Tyson has very few May polls with which to make his inferences.
I'd like to add that Tyson validates his model in an extremely poor way. He proposes an inference method (the "median poll method") for using early polls to predict the Novemeber winner. Thus, he should test this model on early polls in modern presidential elections. Instead, he applies his model with only October polls and only for the 2004 election. Not surprisingly, he finds a "good" model; the median method only mispredicts the winner in one state (Hawaii). Heck, here's a model that misses only two states in 2004 (NM and NH): just predict the winner by who won last time. Clearly that model sucks historically (cf., predicting 1992 from 1988), and I'm guessing so does the "median poll" method (if you examine the historical May state polling).
As Gelman says, Tyson should consult political scientists (such as James Campell, who has done a lot of work on this topic) before penning an op-ed for the Times. If he does that the next time around, I promise not to write my next blog post on Bell's Inequality.
Leave a comment