Since I wrote on Senate elections last week, two more high-quality prediction systems have emerged: Presidential election forecaster Sam Wang is back and looking at Senate polls, and the HuffPollster gang is turning their charts into forecasts.
As with the Drew Linzer/Daily Kos model I looked at last week, both of these are poll-driven. That is, they ask what is likely to happen in November, given the polls to date. That contrasts with systems that put some or more weight on “fundamentals” of elections, on the assumption that early polling may be misleading because it ignores considerations such as district partisanship, national trends and candidate recruitment, all of which may kick in as the campaign goes along.
Right now, in aggregate, the polls-only versions are more favorable to Democrats. So Daily Kos has a 55 percent chance of at least 51 Republican seats; HuffPollster is at 48 percent; and Sam Wang gives Republicans a 35 percent chance for control. Contrast that with the mixed-input Monkey Cage/Washington Post model, which gives Republicans a 52 percent chance of having at least 51 seats, and the New York Times, which gives them a 66 percent chance at a Senate majority. The full-blown 538 model should be coming soon, but Nate Silver’s preliminary look, also with mixed input, also favored Republicans (and there are also, mainly from political scientists, fundamentals-only models, which I should write about separately).
Who to trust? Almost everyone agrees that polls are increasingly better for predicting outcomes as the election draws closer. And almost everyone agrees that well in advance, fundamentals are more useful. So one way to look at the differences is that right now Democrats are doing a bit better than the fundamentals suggest -- and that we have no way of knowing how that will play out by November.
Beyond that, the polling-only models may differ on a lot of things. Which polls should be included? Should low-quality or partisan polls be given equal weight, discounted, adjusted or dismissed? How should polling averages be estimated -- how do the models treat polls as they get older? And in projecting forward, how “linked” are the various races -- how likely is it that Kentucky, Alaska, Louisiana and Iowa will all move independently of one another because contest-specific electioneering matters, and how likely is it that they’ll move together as part of late-breaking national trends? And of course “fundamentals” or mixed models can vary, too, on many dimensions, including the precise economic variables to use, what counts as a “quality” candidate and how much any of those variables matter. Even if purely derived from past data, modelers must make choices about how they examine the past that will have important implications for how they predict the future.
I’ll continue to emphasize that the headline numbers suggest unearned precision. Most of these modelers, most of the time, don’t really make unsupported claims -- but the numbers often suggest more certainty than the prose claims, and both the headline writers and the chart designers rarely include important caveats.
So we shouldn't trust any single model, or even a single average of the various models. Instead, the best way to read all of this is to focus on the range, both in individual models when supplied by the authors, and across models. That’s going to give smart readers uncertainty, and that’s exactly what we all should be experiencing right now. If you want certainty, try the U.S. House: It's going to stay Republican. But we don’t know who is going to win the Senate, and there’s a very good chance we won’t on Election Day morning, or on Election Day night -- until Alaska comes in, and even then we may have to wait for Louisiana to hold a runoff to really know. Or not; the range is still large enough that both solid Republican or solid Democratic control are each very plausible. But for now, the right attitude isn’t to try to figure out which model to trust; it’s that we’re better off the more we know, and each reasonable model adds to our general sense of how things things are going.
To contact the writer of this article: Jonathan Bernstein at Jbernstein62@bloomberg.net.
To contact the editor responsible for this article: Max Berley at firstname.lastname@example.org.