re: #131 gwangung

Models are supposed to be related to reality. And that means that you know what you’re measuring.

His model is describing a reality that is extremely volatile. I simply don’t believe the voting public is that volatile. That leads me to the hypothesis that what he’s modeling is not propensity for voting behavior, but something else that’s only loosely correlated.

His modeling is limited by the volatility of public poll results, which may or may not reflect actual volatility of the electorate. It is LIKELY, but not certain yet, that the real story of this race is stability, with Clinton always 3-6% ahead, and that the apparent volatility is mostly from nonresponse bias. But Silver has limited his statistical universe to public polls, and is trying to avoid falling in love with his own assumptions (a nice way of saying ‘believing his own bullshit’). This means the only data he has to work with are those polls, and they first-level analysis of them suggests a volatile race, even if a second-level analysis suggests stability.

So, the electorate is reality. Polls sample that reality, but they reflect it only dimly, filtered through the assumptions of the pollsters, the emotional state of the electorate, and a host of other variables.

Aggregators take those highly imperfect data, and then filter it through their own assumptions to produce their predictions. They strive to make their conclusions better reflect reality than the dim reflection of individual polls, but in the end, they’re a reflection of a reflection.