Nate Silver is wicked smart

In which Nate creates a model to predict how undecided voters will break:

[W]hat I'm going to do instead is tie the undecided allocation to the extent to which Barack Obama overperformed or underperformed his polls in particular types of states in the Democratic primaries. If you compare the actual results in the primaries to the final RCP or Pollster.com averages, you'll notice some fairly systematic differences.

Specifically, Obama overperformed:

1. In states with high African-American populations;
2. In states that share a border with Illinois (no, Kentucky doesn't count);
3. In states with highly educated electorates;
4. To a lesser extent, in the South (as indicated by the number of evangelicals), even after accounting for the above variables.

Meanwhile, he underperformed his polls:

1. In the Appalachian states (as indicated by the number of respondents who identify their ancestry as 'American', a practice concentrated in the Appalachian region);
2. In states with low education levels;
3. And in states with a high number of Catholics.

This can all be ferreted out via regression analysis, taking the factors I describe above as the independent variables, and Obama's performance vis-à-vis his polls as the dependent variable. The R-squared on this regression is .72, which is quite high -- it means that it was rather predictable when the polls were wrong, and in which direction.

I love this sort of analysis. I'm not statistically smart enough to pull this sort of thing off myself, but I am familiar enough with stats to recognize this as a valid way to look at things. His model also currently gives Obama a likelihood of winning at 75% (!), which is a number I would dearly like to believe but simply cannot swallow. Still, I'll take some comfort in that for the moment.

Also cool: the wizards at Pollster.com dissect the Obama rebound by type of state. Interesting.