Can we anticipate polling error?

G. Elliott Morris send out a recent email covering one of the issues coming to the fore as Biden’s lead over Trump spreads to more states and widens nationally. While poll skepticism has reached all time highs after 2016 (although many national polls got the national vote right, and few polls could have picked up the razor-thin margins in PA, WI and MI, it’s worth looking at this question:

Can we anticipate polling error?

With 100 days to go, it’s hard to see how Trump stages a comeback

This post is a continuation of sorts of the one I sent out to subscribers last Saturday. In it, we covered the topic of uncertainty in election campaigns, specifically how to think through what information we get from different data—and, crucially, what information we’re missing.

A related topic (and one I’ve been receiving lots of tweets and emails about) is what happens if we see a 1988-type polling error again this year. I have rebutted that such a shift is unlikely now, for multiple reasons. The first is that political polarization has caused election campaigns to be less volatile and easier to predict; fewer swing voters means, well, fewer swings. But the second reason is that when polls miss the outcome of the election they usually miss in the direction of what we’ll expect from other information. In other words, we can often (though not always) anticipate the shift beforehand.

In 1988, the economy was growing steadily and the president for the incumbent party, Ronald Reagan, was also relatively popular with an approval rating around 55%. So when Michael Dukakis’s 17-point polling lead disappeared it shouldn’t have been that surprising.

But this year, the polls and fundamentals agree: Donald Trump is in trouble. The various polling averages have Joe Biden up 8-9 points in the polls and have the president’s approval rating 16 points underwater. The economy is also in tatters (or haven’t you noticed) and the rapid spread and intensification of covid-19 around the country threatens the government’s ability to stage a timely comeback. So, if we think the polls will miss, our best guess is still that Biden will win. That wasn’t the case in 1988, 1992 or 2000, when the polls moved rapidly toward the (more accurate) fundamentals over the course of the election cycle.

Point being: over the next hundred days, a lot can change. But if you want to be empirical and rigorous about forecasting that change, your best bet is still on a healthy (though not assured) Biden victory.

I’m intentionally little about the potential uncertainty in this exercise, of which there is a healthy amount. The point is instead to calibrate our expectation of movement over the remaining election cycle, perhaps so that people realize saying “But but but Dukakis was ahead by 17 points in July!” is actually a point for the forecasters, not against them.