Poll: Doug Jones leads Roy Moore by 10 in Alabama. Other polls: actually, Moore’s winning.

Polling in the final week of the Alabama Senate special election has been all over the place. Four polls have shown Roy Moore leading by between 4 and 9 points. One new one, from Fox News, shows Democrat Doug Jones with a 10-point lead. And another, from SurveyMonkey, just posts a range of possibilities, arguing that results will largely depend on how a pollster chooses to model turnout.

Put it all together, and RealClearPolitics’s polling average shows Moore ahead by 2.5 percentage points. That would make him the favorite — but not much of one, since polls frequently “miss” by more than that amount. For instance, RCP’s polling average underestimated Ralph Northam’s margin of victory by 5.6 points in this year’s Virginia governor’s race.

The Alabama race is even more of a muddle. Even in the most normal of times, special elections are extremely challenging to poll. And as Matt Glassman, a senior fellow at Georgetown’s Government Affairs Institute, has pointed out, very little about this election is normal:

To tease this out a little more: Alabama is a deeply conservative, and deeply pro-Trump state, meaning we’d usually expect a Republican to win this race. But Roy Moore was a controversial nominee who some Republicans were wary of even before he was accused of sexually assaulting two teenage girls, meaning many GOP voters might not be so enthusiastic about turning out for him. So it’s difficult to say how many voters will view the race as a referendum on Moore himself, as one on Donald Trump, or as a choice between Republican versus Democrat.

It’s also difficult to say how all those various factors put together might make different voters more or less willing to actually turn out. For instance, the nationalization of the race could motivate Republicans, or the allegations against Moore could depress them and lead many to choose to stay home or cast write-in votes.

We’ve also had a flurry of late-breaking developments that could shake up the race. In Moore’s favor, you’ve had President Trump putting out the word that he fully backs him, some national Republican Party organs backing him again, and one accuser’s admission that she added a note to her high school yearbook that she says Moore signed.

But potentially in Jones’s favor, you’ve had Republican senior Sen. Richard Shelby more openly speaking out against Moore, Democrats forcing Sen. Al Franken (D-MN) to agree to resign due to sexual misconduct, and Moore himself skipping the final weekend of campaigning.

All polls have to make semi-informed guesses about what the electorate will look like

To understand why modeling turnout in the Alabama race is so difficult, let’s first remind ourselves of two major steps involved in any poll:

1. Gathering your raw data — basically, asking lots of people in the state who they plan to vote for
2. Weighting that raw data in an attempt to model the electorate — that is, the people who will actually vote

Step two is crucial because many people who are registered to vote don’t actually turn out — and the demographics and political views of non-voters often differ from the people motivated to turn out.

But the tricky part is that step two also involves making a whole bunch of assumptions about who will turn out. Often, those assumptions are basically semi-informed guesses.

And while for presidential and midterm elections there’s a wealth of historical turnout data to help inform those guesses about the electorate, there’s much less to go on for special elections held in particular states at odd times.

Alabama polling results really vary based on turnout assumptions

For this year’s Alabama race in particular, Mark Blumenthal of SurveyMonkey argued that his firm’s topline results from their most recent poll vary hugely depending on “Minor differences in the methods used to model or select the likely electorate.” Blumenthal writes. They “show everything between an 8 eight percentage point margin favoring Jones and a 9 percentage point margin favoring Moore,” he adds.

For instance:

• One way to model “likely voters” is simply to ask people if they plan to vote, and count everyone who says “yes” as a likely voter. A problem there, though, is that many people tend to overstate their likelihood to turn out, and history shows turnout in off-year special elections tends to be particularly low.
• So perhaps you’d want a bit more skepticism about who’d actually show up this year in Alabama. One way to go about that, Blumenthal suggests, is by counting everyone who says they voted in a post lower-turnout race — like Alabama’s 2014 midterms — as a likely voter. There’s a problem there too, though: 2014 was a national Republican wave year, so basing your modeling on that might not account for this different situation.
• You could also try to make your “likely voter” electorate resemble the 2016 presidential race. For instance, Blumenthal points out that Trump won Alabama by 28 percentage points, but when SurveyMonkey’s Alabama respondents are asked how they voted in 2016, Trump only has an 11-point margin of victory among them. That split seems a bit worrying and might suggest that the polls are “missing” a fair amount of Trump voters. Then again, perhaps many of those Trump voters really aren’t motivated to turn out for this race.

As Blumenthal writes, there are convincing arguments both for and against using any of these, or some combination of them, to model turnout. But all pollsters are making decisions about these questions, and their topline results are highly dependent on those decisions. And that helps explain why the Alabama polls are all over the place.

We have a request

Vox's journalism is free because we believe that everyone deserves to understand the world that they live in. That kind of knowledge helps create better citizens, neighbors, friends, parents, consumers and stewards of this planet. In short, understanding benefits everyone. You can join in on this mission by making a financial gift to Vox today. Reader support helps keep our work free, for everyone. Will you join us?

One-Time Monthly Annual

$95/year

$120/year

$250/year

$350/year

Other

$

Yes, I'll give $120/year

Yes, I'll give $120/year

We accept credit card, Apple Pay, and Google Pay. You can also contribute via