Memo: 2020 Polling Retrospective
By Colin McAuliffe, Johannes Fischer, Charlotte Swasey, Jason Katz-Brown, Jason Ganz, Gustavo Sanchez, and Sean McElwee
Executive Summary
Since 2018 Data for Progress has helped prove the case for progressive policy with its best-in-class polling, while being on the forefront of innovative survey techniques. In the spirit of innovation, this report is the first of a series of analyses to rigorously assess polling’s successes and failures over the past cycle while also offering concrete methodological improvements to address future errors.
In the week before the 2020 general election, Data for Progress conducted polling in 15 states with best-in-class accuracy. As we analyze our performance we’ve drawn three important conclusions:
Partisan Nonresponse and Activist Overrepresentation: Conservative white voters are opting out of polling, while liberal voters are disproportionately opting in, creating an underlying bias in our respondent pools. We also have evidence that liberal partisan activists are systematically overrepresented in our surveys.
Geographic Heterogeneity in Respondents: There are substantial differences between urban and rural white voters in terms of likelihood of voting Biden in 2020 or switching from Trump to Biden. Respondents living in zip codes which display the most firm Trump support are less likely to respond to polls — even when you control for their partisanship and other demographics.
Geographic Heterogeneity in Electoral Performance: Biden overperformed House Democratic candidates in heavily white suburban and urban counties, and underperformed them in rural counties that were majority non-white.
