Wednesday, November 18, 2020

Polls, Election Predictions, Political Correctness, Bounded Cognition (2020 Edition!)

Some analysis of the crap polling and poor election prediction leading up to Nov 2020. See earlier post (and comments): Election 2020: quant analysis of new party registrations vs actual votes, where I wrote (Oct 14)
I think we should ascribe very high uncertainty to polling results in this election, for a number of reasons including the shy Trump voter effect as well as the sampling corrections applied which depend heavily on assumptions about likely turnout. ... 
This is an unusual election for a number of reasons so it's quite hard to call the outcome. There's also a good chance the results on election night will be heavily contested.
Eric Kaufmann is Professor of Politics at Birkbeck College, University of London.
UnHerd: ... Far from learning from the mistakes of 2016, the polling industry seemed to have got things worse. Whether conducted by private or public firms, at the national or local, presidential or senatorial, levels, polls were off by wide margins. The Five Thirty-Eight final poll of polls put Biden ahead by 8.4 points, but the actual difference in popular vote is likely to be closer to 3-4 points. In some close state races, the error was even greater. 
Why did they get it so wrong? Pollsters typically receive low response rates to calls, which leads them to undercount key demographics. To get around this, they typically weight for key categories like race, education or gender. If they get too few Latinos or whites without degrees, they adjust their numbers to match the actual electorate. But most attitudes vary far more within a group like university graduates, than between graduates and non-graduates. So even if you have the correct share of graduates and non-graduates, you might be selecting the more liberal-minded among them. 
For example, in the 2019 American National Election Study pilot survey, education level predicts less than 1% of the variation in whether a white person voted for Trump in 2016. By contrast, their feelings towards illegal immigrants on a 0-100 thermometer predicts over 30% of the variation. Moreover, immigration views pick out Trump from Clinton voters better within the university-educated white population than among high school-educated whites. Unless pollsters weight for attitudes and psychology – which is tricky because these positions can be caused by candidate support – they miss much of the action. 
Looking at this election’s errors — which seems to have been concentrated among white college graduates — I wonder if political correctness lies at the heart of the problem
... According to a Pew survey on October 9, Trump was leading Biden by 21 points among white non-graduates but trailing him by 26 points among white graduates. Likewise, a Politico/ABC poll on October 11 found that ‘Trump leads by 26 points among white voters without four-year college degrees, but Biden holds a 31-point lead with white college graduates.’ The exit polls, however, show that Trump ran even among white college graduates 49-49, and even had an edge among white female graduates of 50-49! This puts pre-election surveys out by a whopping 26-31 points among white graduates. By contrast, among whites without degrees, the actual tilt in the election was 64-35, a 29-point gap, which the polls basically got right.
See also this excellent podcast interview with Kaufmann: Shy Trump Voters And The Blue Wave That Wasn’t 

Bonus (if you dare): this other podcast from the Federalist: How Serious Is The 2020 Election Fraud?

Added: ‘Shy Trump Voters’ Re-Emerge as Explanation for Pollsters’ Miss
Bloomberg: ... “Shy Trump voters are only part of the equation. The other part is poll deniers,” said Neil Newhouse, a Republican pollster. “Trump spent the last four years beating the crap out of polls, telling people they were fake, and a big proportion of his supporters just said, ‘I’m not participating.’” 
In a survey conducted after Nov. 3, Newhouse found that 19% of people who voted for Trump had kept their support secret from most of their friends. And it’s not that they were on the fence: They gave Trump a 100% approval rating and most said they made up their minds before Labor Day. 
Suburbanites, moderates and college-educated voters — especially women — were more likely to report that they had been ostracized or blocked on social media for their support of Trump. ... 
... University of Arkansas economist Andy Brownback conducted experiments in 2016 that allowed respondents to hide their support for Trump in a list of statements that could be statistically reconstructed. He found people who lived in counties that voted for Clinton were less likely to explicitly state they agreed with Trump. 
“I get a little frustrated with the dismissiveness of social desirability bias among pollsters,” said “I just don’t see a reason you could say this is a total non-issue, especially when one candidate has proven so difficult to poll.”

No comments:

Blog Archive