Measuring and incorporating resentment origins of non-response bias in polling data

Political polls reached new lows in accuracy during the 2020 American Presidential election. Polls systematically over-estimated Democratic candidate support across multiple offices by several percentage points within states. The AAPOR post-mortem concludes the error likely arising from non-response bias. However, these problems did not arise during midterm elections. I argue that the reason for the inconsistent polling accuracy arises from a polling non-receptive (PNR) population that favors right-wing nationalist candidates. I hypothesize a link between an area's Census response rate and polling error, voter turnout, and Trump support. I employ a mixed effects model of state level polling error from 2020 -- 2024. I additionally employ a two-stage mixed effect panel linear model of county level census response, turnout, and Trump support. I find supportive evidence, with an eight point increase in poll Democratic bias across the IQR. Additionally, census response is associated with increased voter turnout and two-party vote.