It’s common for pundits to recite ass-covering phrases like “it all comes down to turnout” or “anything could happen” on the eve of a big election. If you’ve been following FiveThirtyEight over the years, you know it’s not our style to do that. Instead, we issue probabilistic forecasts, which can sometimes seem quite confident: We had Barack Obama as a 90.9 percent favorite to beat Mitt Romney on the eve of the 2012 general election, for example.
So let’s get a couple of things straight before the results start trickling in from Iowa tonight:
- It all comes down to turnout.
- Anything could happen.
All right, not absolutely anything could happen. Martin O’Malley is not going to win the Democratic caucuses. Donald Trump will probably not finish behind Carly Fiorina.
Could Trump slip all the way to third place? Entirely plausible. But he could also get upwards of 40 percent of the vote and double his nearest rival’s total.
Ben Carson in second place? Rand Paul in third? The odds are against it — but equally strange things have happened in Iowa before.
We say this for the same reason we can sometimes issue highly confident forecasts just before a general election: It’s what the data tells us. That data tells us that polling in general elections is pretty accurate, at least in the final few weeks before the election. The data also tells us that polling in primaries and caucuses is not very accurate. Historically, the average error of late polls in presidential general elections is about 3.5 percentage points.1 By contrast, the average polling error associated with presidential primaries is more like 8 percentage points, more than twice as high.
So imagine that we have a forecast showing Trump 4 percentage points ahead of Ted Cruz in some state. If Trump wins by 12 points instead, or Cruz wins by 4, the pollsters would be pilloried, and we’d come in for our share of flak too. But that’s what an 8-point error looks like, and 8-point errors happen fairly often in primaries and caucuses.
What makes polling these elections so difficult? There are a few major factors:
- Turnout is much lower in primaries and caucuses, and much harder to predict.
- There are often multiple candidates running. Such races increase polling error because of the potential for tactical voting.
- There are far more swing voters because most voters like several of their party’s candidates. In the recent Des Moines Register Iowa poll, for example, the average Republican respondent had a favorable impression of four of the Republican candidates. By contrast, only a small fraction of general election voters like both the Democratic and Republican candidates.
- A substantial number of voters wait until the last few days of the campaign to make up their minds in primaries and caucuses; by contrast, the vast majority of general election voters have their minds made up well ahead of Election Day.
But if primaries and caucuses are always tough for pollsters, some are even harder than others. This is something we’ve studied extensively too. Historically, the polling error has been higher when:
- A state holds a caucus instead of a primary.
- It’s early in the nomination calendar rather than later. (Perhaps because pollsters haven’t yet had a chance to learn from their mistakes.)
- There are more candidates running. (See above for why this matters.)
You’ll note that the first two circumstances apply in the Democratic caucuses tonight, and all three do for Republicans. Iowa is a caucus state, and it’s the first state to vote. And there are still a huge number of candidates on the GOP side. In our polling average, candidates other than Trump, Cruz and Rubio have a collective 28 percent of the vote, while another 3 percent or 4 percent of voters still say they’re undecided. That’s almost a third of the vote that could easily enough recirculate to one of the front-runners.
Put another way, the uncertainty associated with forecasting tonight’s Iowa Republican caucus is about as high as it gets in a major American election. Even Ann Selzer, the best pollster in the country, could have a rough night.
That doesn’t mean we’re completely in the dark. Our forecast models are designed to account for this uncertainty. Hillary Clinton, for example, leads Bernie Sanders by 4.5 percentage points in our Iowa polling average. In a general election, that would make her a rather heavy favorite, probably upwards of 90 percent. But in the Iowa caucuses, it’s not that much of an edge. Thus, our polls-only forecast still gives Sanders a 28 percent chance of winning, and our polls-plus forecast, which likes Sanders because his Iowa numbers exceed his standing in national polls, puts his chances slightly higher, at 33 percent.
There’s even more uncertainty on the Republican side. Trump leads Cruz in our polling average by about the same margin that Clinton leads Sanders, 4.7 percentage points. But the larger number of candidates involved could make for a wild finish. Here’s each candidate’s chances of finishing in first, second or third, according to our polls-only model:
|CANDIDATE||1ST||2ND||3RD||4TH OR WORSE|
Trump has a 54 percent chance to win, according to our polls-only model, compared with Cruz’s 33 percent. But you’ll notice that the model gives Rubio an outside chance too, 11 percent. Surely Rubio will finish in the top three, at least? No, that’s not certain either; the model gives him a 24 percent chance of finishing in fourth place or worse.
Then there are some of the truly wild scenarios I described earlier. Carson is given a 5 percent chance of finishing in second place. How might that happen? I can’t tell you. But by definition, the biggest surprises are the ones no one is prepared for, like Hillary Clinton beating Barack Obama in the 2008 New Hampshire primary.
Our polls-plus forecast also has Trump favored, but only narrowly. Its algorithm gives an extra percentage point or two to Cruz and Rubio because their Iowa polls exceed their standing in national polls — historically a favorable indicator — and the opposite is true for Trump.
|CANDIDATE||1ST||2ND||3RD||4TH OR WORSE|
But again, there’s a lot of uncertainty. Polls-plus assigns Trump a 10 percent chance of finishing with 15 percent of the vote or less, which could be a campaign-ending embarrassment. However, it also gives him a 10 percent chance of finishing at 38 percent or higher, in which case he’d look unstoppable.
As primary season wears on, our models won’t be hedging their bets quite so much. Candidates will drop out, and voter preferences may become more stable. As we learn more about who is voting for whom, we may also be able to add demographic information to our forecasts, which can potentially make them quite a bit more accurate.
But for tonight? Keep an open mind about the results. And don’t be shocked if the polls are way off.
Listen to the latest episode of the FiveThirtyEight elections podcast.