As regular readers know, the meta-analysis is a hobby. I’ve only asked for support via ActBlue (nearly $45,000) or the NRSC. But now, a prize from Kevin Drum, who notes that my polling averages were extremely close to the final results. In addition to the 364 EV prediction (a bit seat-of-the-pants at the time, but now backed up by the variance minimization approach), district polls predicted a House outcome of 257 Democrats – exactly correct – and a Senate outcome of 58 seats (with the Franken-Coleman recount, it’s 59). The prize…a year’s subscription to Mother Jones. That works out to about $0.03 per hour, plus the reputation of being an MJ reader.

Was there an element of luck? A little. But more importantly, there’s a lesson about the overall reliability of Congressional polls.

The original House prediction had an error bar (68% CI) of up to +/-3 EV. Binomial statistics indicate that the odds against hitting the outcome precisely were as low as 6-1, assuming that polls are unbiased on average. But maybe the odds were better than that.

What I did was simple: I added the number of safe Democratic districts and the “toss-up” (Pollster.com definition) districts in which the Democratic candidate was ahead. For the error bar I assumed that the toss-ups were exact knife-edge races (i.e. win probabilities of 50%). This assumption leads to a very conservative estimate of the error.

But now let’s take into account the fact that polling margins always correspond to more lopsided probabilities. (This is what I did for the Presidential and Senate calculations.) The number of seats estimate would be similar, but the error bar would be substantially smaller. For an error bar of +/-2 EV the odds of a precise match are 5-1 against; for +/-1 EV, about 5-3 against.

The match of prediction with result suggests that the margin-based win probabilities can be taken at face value at all levels. In other words, polls may be essentially unbiased overall, even at the level of Congressional districts. In my view that’s a fairly remarkable finding. It complements nicely the lack of bias in Presidential and Senate polls.