Silver also has a fascinating analysis of which polls were accurate, and which weren’t. In particular, polls based on calling landline telephones tended to seriously underestimate the Democrat vote. Voters without landlines are more likely to be young, urban, Black, Hispanic, strapped for cash, or some combination of the five, and all five categories are more likely to vote Democrat. There are lessons here for pollsters in other countries.
The following simple NetLogo simulation model (click on the image to run it) re-rolls Silver’s electoral dice, giving alternative outcomes – exactly the kind of simulation Silver actually did to support his 90% prediction of an Obama win:
Silver appears to take a Bayesian approach to statistics. A Bayesian has been described as “one who, vaguely expecting a horse, and catching a glimpse of a donkey, strongly believes he has seen a mule.” The legendary XKCD summarises the perspective quite elegantly: