Nate Silver on How to Interpret Recent Polling
I’ll admit, I was a tad alarmed to see the swing state polling data yesterday. Even a vapid and patently absurd candidacy like Donald Trump’s will benefit from swings in the news cycle, and he is undoubtedly getting a bounce from the FBI’s announcement that Hillary Clinton will not be prosecuted because of her emails, as well as the killing of five police officers in Dallas last week. Nate Silver puts all of that in context here.
I don’t think Hillary is taking Trump too lightly, which is good. This nation cannot afford to have him come within an eyelash of winning. He must lose in a landslide.
Ordinarily, this is the point at which I’d urge a little patience. There’s been a lot of news over the past two weeks — the conclusion to the FBI’s investigation into Clinton’s emails and the Dallas shootings of police officers, in particular — and it would be nice to see how the polls settled in after a couple of slow weeks on the campaign trail. However, we’re entering a period of rapidly moving political news. Bernie Sanders endorsed Clinton only Tuesday. Trump is expected to name his VP later this week. And then we’ll have the party conventions. The prospects definitely look better for Trump than they did a week or two ago, but the landscape also looks blurrier, and it may not be until mid-August that we have a chance to catch our breath.
So for the rest of this article, I’m going to focus mostly on the Quinnipiac polls — both to explain why our model reacted relatively strongly to them, when some of the other data wasn’t so bad for Clinton, and as an example of how you might think about “unexpected” polling results as they arise over the next few weeks. If you’re a poll junkie, this situation will seem familiar. You think you have a pretty good idea of where the race stands, but then a couple of splashy polls come out that contradict that impression. You have to figure out how much to incorporate the new polls with the data you had previously.
The FiveThirtyEight models make that calculation automatically; we just input the new data and press “go.” That’s helpful because when people (us included) rely on their intuition about how to evaluate new polls, they tend to make one of two mistakes. The more common error is to treat every new poll as a “game changer,” inventing an elaborate story about how the whole race has been upended. Often, it turns out, these interpretations don’t hold up to scrutiny, and a highly touted new poll won’t move the forecast much at all, or a poll that comes out the next day contradicts it.
But there’s also the potential mistake of dismissing a poll as an “outlier” and ignoring it when it provides important new information. This mistake is probably becoming more common because of the influence of sites like FiveThirtyEight. People have learned to trust the polling average more than individual polls — and that’s a good lesson. However, they sometimes take this a step too far, forgetting that the average is composed of individual polls. The average isn’t an excuse to ignore data you don’t like.
The fivethirtyeight.com forecast for the election is here, and it does show Hillary with a 66.2% chance of winning, and winning comfortably, but that is down from more than 80% when the model was first posted a few weeks ago.