Skip to content

The Patron Saint of Superheroes

Chris Gavaler Explores the Multiverse of Comics, Pop Culture, and Politics

A week before the election, I gathered and posted conglomerate polling numbers from FiveThirtyEight (updated October 26) for the ten battleground states. Here they are again, now with the actual outcomes:


Polls: Biden up 8.

Results: Biden by 3.

Difference: 5 for Trump


Polls: Biden up 6.7.

Results: Biden by .6. 

Difference: 6.1 for Trump


Polls: Biden up 5.6.

Results: Biden by 1.

Difference: 4.6 for Trump.


Polls: Biden up 3.0.

Results: Biden by .3.

Difference: 2.7 for Trump.

North Carolina

Polls: Biden up 2.5.

Results: Trump by 1.3.

Difference: 3.8 for Trump.


Polls: Biden up 2.4.

Results: Trump by 3. 

Difference: 5.4 for Trump.


Polls: Biden up 1.2.

Results: Trump by 8.2.

Difference: 9.4 for Trump.


Polls: Biden up .4.

Results: Biden by .3.

Difference: .1 for Trump. 


Polls: Biden up .1.

Results: Trump by 5.7.

Difference: 5.8 for Trump. 


Polls: Trump up 1.5.

Results: Trump by 8.

Difference: 9.5 for Trump. 


So the one good moment for the state polls:

threading the Georgia needle by .1.

The worst moments:

missing Iowa and Ohio by almost double digits.

Even worse:

missing seven of the ten states by more than four points. 

Worst overall:

under polling Trump every time. 

Okay, but what about the national polls? 

On the day before the election, Biden was up by 8.4 at FiveThirtyEight, and by 7.2 at RealClearPolitics. The national vote results are still dribbling in, but Biden has at least 3.7, and FiveThirtyEight projects an eventual 4.3. So the least possible mistake is 2.5 for Trump, falling barely within a 3-point margin of error. At worst, the mistake is 4.7 for Trump. 

That’s a lot worse than in 2016, when the RealClearPolitics final average put Clinton up by 3.2, and then she won the popular vote the next day by 2.1. That difference of 1.1 is pretty good. The mistakes in 2016 were only in some states polls, and, like with the popular vote, those mistakes only grew in 2020. 

Pollsters have at least one line of defense: their margins of error were often 5 points. I wrote here two weeks before the election:

“Placing states with Biden or Trump leads under 5% as toss-ups, on Tuesday October 20, two weeks before Election Day, the state polls put Biden at 259 electoral votes.”

Only Michigan and Wisconsin were outside a 5-point margin of error. So according to the October 20 polls, the Electoral College was a toss-up–though with a massive Biden advantage. But a week later I wrote: 

“As of today, Monday October 26, eight days before the Election, the polls would put Biden at 279 by including Pennsylvania. All of the other battleground states remain in the toss-up column’s margin of error.”

And that’s what happened. Biden won Wisconsin, Michigan, and Pennsylvania, with contested leads in Arizona and Georgia.

Of course that doesn’t explain the mistakes always falling in the same direction. If these were just garden-variety sampling errors, at least some should have favored Biden. None did.  

It’s not clear why the polls were wrong. Michael Moore offered one rationale a week before the election:

“Don’t believe these polls … the Trump vote is always being undercounted. Pollsters, when they actually call a real Trump voter, the Trump voter’s very suspicious of the ‘Deep State’ calling them and asking them who they’re voting for. It’s all fake news to them, remember. It’s not an accurate count. I think the safe thing to do, this is not scientific … whatever they’re saying the Biden lead is, cut it in half, right now, in your head. Cut it in half, and now you’re within the four-point margin of error.”

Is Moore right? Do Trump voters lie to deep-state pollsters?

Maybe. His “not scientific” cut-it-in-half methodology proved about right, but there’s no test for evaluating the accuracy for his explanation. And yet Nate Cohn (the Nate who replaced Nate Silver at the New York Times) now echoes Moore:

“It’s hard not to wonder whether the president’s supporters became less likely to respond to surveys as their skepticism of institutions mounted, leaving the polls in a worse spot than they were four years ago.”

Cohn also wonders about a mirror effect from progressives:

“the resistance became likelier to respond to political surveys, controlling for their demographic characteristics… Like most of the other theories presented here, there’s no hard evidence for it — but it does fit with some well-established facts about propensity to respond to surveys.”

The pandemic may have influenced that even further with a “post-Covid jump in response rates among Dems” because they “just started taking surveys, because they were locked at home and didn’t have anything else to do.” Cohn also wonders whether the high turn-out helped Trump, with the polls’ focus on “likely voters” (where Biden was up much higher) missing the actual votes of the less likely but registered voters who did show up on election day after all. 

So bottom line: the 2020 state polls were as bad as the 2016 state polls, the 2020 national polls were worse than the 2016 national polls, and we don’t know why.

If you ignore the margins of victory in each battleground state, the 2020 forecasts did much better. They all predicted Biden, though with a wide range of differing Electoral College totals. 

First, to their credit, all of the forecasters correctly placed Wisconsin and Michigan in the Biden column, giving him a minimum of 259. 

Politico said Biden would win with 279 electoral votes by taking Pennsylvania, but they left Arizona in the toss-up category and incorrectly predicted Georgia for Trump. So they missed two states, one for Biden, one for Trump, and dodged calling five states, including three (Florida, Ohio, Iowa) that Trump won easily. 

Crystal Ball, Cook Political Report, CNN, and NPR all said Biden would win with 290 votes, correctly predicting that he would take not only Pennsylvania but also Arizona, leaving five states in the toss-ups.

U.S. News predicted Biden with 290 too, but they also correctly placed Iowa and incorrectly placed Georgia in the Trump column. 

Things go down hill from there:

Inside Elections gave Biden 319 by adding Florida to his column.

The Economist gave him 323 by including Florida and North Carolina but not the toss-up Arizona.

Decision Desk, PredictIt Market Probabilities, and FiveThirtyEight all gave Biden 334, giving him Florida and North Carolina.

JHK Forecasts and Princeton Election Consortium predicted 335, the extra one point coming from Maine’s second district.

Since Biden’s actual count is now 306 (the same as Trump’s self-described 2016 “historic landslide”), just over half, seven out of the thirteen forecasters, overshot.

That’s mainly because they all incorrectly predicted Florida for Trump. Since Trump won Florida by a safe three points, Florida is both the polls’ and the forecasters’ biggest 2020 failure.

Five of the forecasts also gave North Carolina to Biden, but since Trump won it by only 1.3, North Carolina seems legitimately purple. Though not the incandescent purple of Georgia, Pennsylvania, Wisconsin, and Arizona where Biden’s margin was under 1%, very similar to Trump’s 2016 margins. 

So, like in 2016, the final map hides how close the Electoral College race was. 

Of course if we elected our presidents by national popular vote, a Biden 4-point margin of victory isn’t close at all. 

%d bloggers like this: