Despite certain kinds of bias, poll averages are still doing a good job predicting winners. (The Washington Post)
Before the election, I suggested that, despite lots of concerns about increasing bias in modern polling, poll averages had actually been getting better at predicting the correct winner over time. To make that point, I showed a graph of Brier scores from the last 12 election cycles. Lower scores indicate better predictions, and Brier scores have definitely been falling. Moreover, the drop seems related to the volume of polls in each cycle.
There are definitely reasons to think this trend was broken this year. The Senate polls were severely biased against Republicans in race after race (see, for example, here and here). Probably the worst example was Virginia, where Democratic incumbent Mark Warner was predicted to win by about 10 percent and actually scraped by with less than 1 percent. But there were plenty of other examples, too.
Yet for all this anti-Republican skew, it made almost no difference for predicting the winners. The graph below updates the one from my earlier post, with the numbers from 2014 added in and all polls up to Election Day included in the calculations (the earlier graph stopped at one week before). The Brier score was a little higher (worse) this year than in some other recent years, but we are still clearly in a world of better accuracy. And while the number of polls doesn’t explain all of the improvement, the correlation between the number of polls and the Brier scores is an impressive -0.65.
