Want these election updates emailed to you right when they’re published? Sign up here.
Hillary Clinton’s chances of winning the presidential election declined by nearly 3 percentage points, to 85 percent, in FiveThirtyEight’s polls-only forecast on Wednesday. Similarly, they declined by 4 percentage points, to 75 percent, in our polls-plus forecast.
The Clinton camp doesn’t really have a lot to complain about: Her position in the polls (and in our forecasts) is much improved since the Democratic convention in late July. And she’s had strong polls this week in swing states such as Pennsylvania and Wisconsin. A Marquette University poll published on Wednesday gave her a 13-percentage-point lead over Donald Trump in Wisconsin. So why were her numbers down in our model, on a day where there wasn’t all that much polling?
It’s pretty much because of a single national poll conducted by Selzer & Company, on behalf of Bloomberg Politics, which gave Clinton a 4-point lead among likely voters. On the surface, that doesn’t seem all that bad for Clinton. Recent national polls have had Clinton ahead by around 7 or 8 percentage points. And we’ve had lots of national polls, so the average is pretty well established. One additional survey showing a result just a couple of points worse than the average shouldn’t be that big a deal, right?
So let me explain why this poll moved the needle, as far as FiveThirtyEight’s models are concerned. The short version: The trend line in the poll was awful for Clinton, since Selzer’s previous national polls had shown Clinton with leads of 12 percentage points in June and 18 (!) percentage points in March. For a more technical explanation, read on.
One of the most important features of our model is our trend-line adjustment. Basically, the model looks at trends within individual polls — for instance, that Clinton gained 3 percentage points in Quinnipiac’s latest poll of Ohio compared to its previous survey — in order to infer the overall trajectory of the race. This adjustment is then applied to the forecast in each state. Why wasn’t the model affected very much by recent polls showing Clinton up by 9 points in Pennsylvania, or by 13 points in Wisconsin? Because the model already figured she was doing well in these states, based on her improvement in national polls, and in polls of other states.
But, of course, the trend-line adjustment sometimes does get surprised. The most surprised it can get is by a very recent, high-quality poll that cuts against the prevailing direction of other surveys. The Selzer poll was one of those cases.
By “high quality,” I mostly mean a poll with a high pollster rating — which are based on a pollster’s past accuracy and methodology — since the trend line adjustment puts more weight on better-rated polls. Selzer & Company (which also conducts the prestigious Des Moines Register Poll) more than qualifies: It’s one of just six polling firms with an A-plus rating from FiveThirtyEight.
There’s another, more subtle dimension to this also, which is that Selzer hasn’t polled the general election very often this year. (Just the three national polls so far.) The trend-line adjustment is therefore designed to give a lot of weight to a Selzer poll whenever it weighs in. By contrast, it gives less weight to any given poll from a pollster that surveys the race frequently, such as by conducting a national tracking poll.1
The Selzer poll also cut against the prevailing direction of other surveys. Almost every poll over the past 10 days has shown Clinton gaining ground on Trump, or at least holding steady. This one showed a big decline for her — down from a 12-point lead in June to a 4-point lead now.2
You might raise a reasonable objection here: The 12-point lead that Selzer had for Clinton before — and certainly the 18-point lead it had for her in March — were outliers to begin with. Thus, Clinton’s decline was partly a matter of the poll reverting to the mean, instead of a true shift in the race.
This is probably true to some extent (and it’s one reason that Democrats shouldn’t worry about the poll all that much). But remember that this tends to cut in both directions. On Tuesday, for instance, Clinton benefited in the trend-line adjustment from the Quinnipiac poll that showed her going from 6 points back in Pennsylvania to 9 points ahead. Although Quinnipiac has generally had mediocre numbers for Clinton, that 6-point deficit in Pennsylvania had been especially out of line with other surveys. Thus, part of her gains in that poll were probably reversion to the mean also.
Over time, these instances tend to cancel out, and the model tends to be self-correcting. We’ll watch carefully over the next several days to see whether the latest Selzer poll is the start of a shift back to Trump, or an oddball case that the model soon works out of its system.