Shooting a rabbit with a cannon

There’s some dude who goes around with a method for deterministically forecasting presidential elections. It’s all pretty silly given that he gives his model credit for predicting the winner of every presidential election, including tossups such as 1960, 1968, and 2000 which it’s pretty meaningless to imagine you could predict more than a day or so ahead of time.

This stuff must have hit the newspapers again recently, because Nate went to the trouble of demolishing it. This has got to be one of the less fun parts of Nate’s job—to spend your time shooting down misconceptions that never should’ve been taken seriously in the first place.

3 Responses to Shooting a rabbit with a cannon

  1. Justin September 1, 2011 at 2:47 pm #

    My question is why newspapers (and the blogosphere!) continue to make a big deal out of such qualitative models. I personally don’t find them much easier to understand, as the choosing and weighting of the “keys” seems sort of arbitrary to me. Like why 13 keys? Are they all equal? What constitutes significant policy change? etc. etc. A simple number, like a percent vote or win probability, would be much more attractive and easier to understand for the masses, regardless of how complex the method used to arrive at it.

  2. Mike September 2, 2011 at 7:45 am #

    I think there are quite a few valid reasons for being skeptical towards Lichtman’s claims, but Silver’s objection is hardly what I’d call “demolishing”. In fact, given that Lichtman’s model is designed to predict a binary outcome and nothing else it is not clear to me that it demolishes anything at all. The model makes no claims towards the categories (all of which have quite specific membership conditions, by the way) being additive or linearly related and does not purport to predict the margin of victory. So what exactly is the point of chiding it for something it was never designed to do in the first place? Moreover, using historical data (31 elections) to build a model that predicts future states (7 elections so far if you include Gore’s popular vote) sounds like perfectly normal (and valid) scientific practice to me, but perhaps things work differently in Gelman’s world.

    • Andrew Gelman September 2, 2011 at 4:52 pm #

      Mike:

      See the link in my post above for how things work in Gelman’s world. The short answer is that, as noted above, it’s basically meaningless to give any model credit for predicting the winner in 1960, 1968, or 2000.

      As Bill James put it, if you want to predict a pitcher’s won-loss record next year, you’re better off using this year’s ERA than this year’s won-lost record. To put it in statistical terms, just because you’re predicting a binary outcome, it doesn’t mean you should throw away available information.