Friday, September 14, 2012

Election Model: 538

I've used several good ideas from Nate Silver's blog, located here. Some of them are:

  • Some of the outcome analysis (win electoral lose popular, blowout)
  • The idea that registered and likely voter models should be treated differently. 
  • The idea to display electoral vote outcomes in a histogram. This might be my favorite idea of his, it's a great way to visualize the electoral picture.

I also do some things differently.

  • I underweight registered voter polls, I don't adjust them. Early on likely voter models are difficult to calibrate (quick, predict in may who will vote in november). As the election draws near they get better and better. Instead of keep tracking of LV vs. RV aggregate results and adjusting accordingly, I just underweight the RV polls, more and more until they have no weight right before election day.
  • I don't adjust for house effects. I understand the desire and am even willing to consider that it might be a more complete approach, but I think those effects will tend to wash mostly out as sample size increases and adjusting too much for house effects leaves you in danger of overfitting.
  • When there is enough polling in a state to yield a sufficient sample, the model only considers aggregate poll sample size, aggregate poll outcome, and time to the election. When there is insufficient poling data I fill in the blank with the Cook's PVI value (I don't know the details behind 538s state fundamentals but I think these two ideas are similar in nature). I believe 538 introduces unnecessary complexity by including economic data. I have a hard time imagining such data adding information not already captured by polling.

No comments:

Post a Comment