May 7, 2021

The 2020 Polling Post-Mortem Is Beginning To Coalesce

As more and more studies on 2020 election polling come to light, it's becoming clear that very specific under-sampling issues contributed to overstating Joe Biden's lead. The most recent is an extensive analysis by Data For Progress—a staunchly "progressive" organization on the left. They found that—among other things—their own surveys in 2020 fell short on:

Ideological non-response bias. Essentially, white conservatives disproportionately opted-out of their polls, while liberal voters disproportionately opted-in;

Geographic non-response bias. Respondents living in strongly pro-Trump zip codes were less likely to respond to their polls—even when controlling for partisanship and demography;

Basically, these findings dove-tail with our own view that a widespread lack of trust (in polling, the media and other institutions) resulted in the underrepresentation of key voting groups in most 2020 presidential polling. The fact that most of this was found among white conservative voters incorrectly led to a "consensus" that Biden was ahead by 9-10 points, when in fact it ended up closer to 5-6 points. 

We have argued for assessing this issue through a battery of questions that can be expressed as a "Trust Index". If trust metrics are disproportionately askew among certain partisan or ideological groups, adjustments can be made in vote projections. Organizations like Data For Progress (and others) should be praised for examining this issue. It's an on-going exercise that the entire polling industry needs to be committed to.