A shocking Iowa poll means somebody is going to be wrong
Either Ann Selzer and the New York Times, or the rest of the polling industry.
Yesterday, I complained about how so many pollsters are “herding” by publishing results that are almost an exact tie in a way that is incredibly statistically improbable given the unavoidable sampling error from surveying a small number of voters. I also noted a handful of prominent exceptions — rogue pollsters like the New York Times/Siena College that practically exist in an entirely different universe and imply a much bigger political realignment.
Another such maverick is Ann Selzer of Selzer & Co. (Selzer and NYT/Siena are our two highest-rated pollsters.) As my former colleague Clare Malone wrote in 2016, Selzer — like NYT/Siena — has a long history of bucking the conventional wisdom and being right. In a world where most pollsters have a lot of egg on their faces, she has near-oracular status.
So Selzer’s new poll of Iowa tonight was highly anticipated by polling junkies, despite Iowa being unlikely to be a decisive state. In June, Selzer’s poll for the Des Moines Register showed Donald Trump with an 18-point lead over Joe Biden in Iowa — surprisingly big, even considering how much Iowa has trended red over the years. (It voted for Trump by 8 points in 2020.)
In September, her survey had Kamala Harris just 4 points behind Trump — considered an outlier at the time.
Her new poll? It shows the state trending even bluer, with Harris leading in Iowa 47-44. 🤯
Releasing this poll took an incredible amount of guts because — let me state this as carefully as I can — if you had to play the odds, this time Selzer will probably be wrong. Harris’s chances of winning Iowa nearly doubled in our model from 9 percent to 17 percent tonight, which isn’t nothing. Polymarket shows a similar trend, moving from 6 percent to 18 percent after the survey. But that still places Harris’s odds at around 5:1 against.
The poll has a reasonable sample size: 808 likely voters. Still, the margin of error for the difference separating the candidates in a poll of that size is ±6.6 points. That means in theory, in 95 out of 100 cases, the “real” number should be somewhere between Trump +3.4 and Harris +9.6 if Selzer had surveyed every single Iowa voter instead of just an 808-person sample.
The theory, however, doesn’t capture all possible sources of polling error — especially the fact that the overwhelming majority of people who pollsters attempt to contact never complete the survey. And we ought to be good Bayesians here: even if Harris has a very good night on Tuesday, winning a state you lost last time by 8 points is a big ask.
To give us a little more perspective, there was also a second Iowa poll out tonight from Emerson College that showed Trump leading by 9 points, close to the margin from 2020. Emerson is a firm that does a lot of herding, so you ought to account for that — they virtually never publish a survey that defies the conventional wisdom.1 However, for what it’s worth, their margin of error runs from Trump +15.7 to Trump +2.5.
There is just a little bit of overlap with Selzer, then, believe it or not, given how much the toplines differ. And not entirely coincidentally, our current polling average in Iowa falls just within that range of overlap: Trump +3.4 points. (Careful: the “average” is based on only 4 polls of Iowa all cycle long.)
I have a lot to say about this survey, so let’s take the rest of this in three broad categories: what the poll means for the Electoral College, what it means for Iowa and neighboring states, and what it says about the state of the polling industry overall.