Donald Trump will win - Polls aren't reliable and Trump voters are hesitant to be vocal

The argument

This argument, made ahead of the 2020 election, holds that Donald Trump will be re-elected despite trailing in the polls, because the polls are unreliable — and unreliable in a specific, pro-Trump direction: his voters are hesitant to tell pollsters the truth. The foundation is 2016. Polling forecasts, expert models and the mainstream press converged on a comfortable Hillary Clinton victory, right up until the night Trump won. The misses were worst in precisely the industrial battleground states that decided the election, where polls understated Trump's support by several points. Proponents argue the systematic error was never fixed, because its cause is structural: the 'shy Trump voter'. In a media and social climate where supporting Trump invites being labelled a racist or worse, many of his voters decline polls, or tell interviewers something other than how they will vote. Social-desirability bias of this kind skews surveys against him while leaving his actual electoral strength intact — strength visible instead in proxies pollsters cannot launder: enormous rally crowds, boat parades, registration trends and an enthusiasm gap his opponent could not match. The inference follows. If polls systematically undercount Trump's vote by even a few points, then a modest published deficit is consistent with a winning position in the states that matter — exactly the configuration of 2016, repeating. Betting on the polls means betting that an industry which failed catastrophically four years earlier, for reasons it has not cured, has somehow become reliable. From this standpoint, the silent supporters who confounded the forecasts once will do so again. Because the polls aren't reliable and Trump voters are hesitant to be vocal, this argument holds, Donald Trump will be re-elected.

Premises

Counter-arguments

The prediction was tested at the ballot box and failed: Trump lost the 2020 election, 232 electoral votes to 306, trailing by roughly seven million votes nationally. An argument whose entire content is a forecast is answered decisively by the result. The mechanism it relied on has fared no better under examination. The 'shy Trump voter' is a testable hypothesis, and it has been tested directly — through list experiments, which let respondents register a preference without disclosing it, and by comparing surveys conducted by live interviewers with self-administered online surveys, where social pressure is absent. If concealment were driving the error, those methods would diverge; they largely did not. Post-election reviews of 2016 attributed the state-level miss principally to a technical failure to weight samples by education, which left survey respondents better educated than the electorate, together with late-deciding voters breaking heavily one way. Both were correctable, and the education weighting was substantially corrected before 2020. The premise also overstates what went wrong in 2016. National polls came close to the final popular-vote margin; the failure was concentrated in state-level surveys in a handful of states, which is a narrower and more specific problem than an industry-wide collapse. Describing forecasts, models and the press as a single failed instrument obscures where the error actually was. The proxies offered in place of polling measure the wrong quantity. Rally attendance, boat parades and registration enthusiasm capture intensity among people who already support a candidate; elections are decided by breadth. A campaign can hold the most fervent minority in the country and still lose, which is what the argument's own preferred indicators were unable to detect.

Rejecting the premises

[Rejecting P1] The 2016 error is misdescribed: national polls came close to the final popular-vote margin, and the miss was concentrated in state-level surveys in a few states, which is narrower than an industry-wide failure of polls, models and forecasts together. [Rejecting P2] The explanation has been tested directly — through list experiments and comparisons of interviewer-administered with self-administered surveys — and the post-election reviews found little support for it, attributing the state error principally to under-weighting by education, which was largely corrected before 2020. [Rejecting P3] Rally attendance and parades measure intensity among existing supporters rather than breadth of support, and the inference was tested and failed: 232 electoral votes to 306, with a popular-vote deficit of roughly seven million.