Keywords: U.S. presidential election 2016 - performance of election polls

A curious feature of the 2016 U.S. presidential election was the pronounced impact of education on candidate preferences. Trump’s surge among voters without a college degree propelled him to the presidency. From the perspective of pre-election polling, failing to weight estimates by education might have contributed to a systemic underestimation of Trump across state polls.
The AAPOR report – “An Evaluation of 2016 Election Polls in the United States” - offers few definitive statements on the subpar performance of state pre-election polls. However, the report makes three strong assertions: 1) state polls generally underestimated Trump while national polls performed remarkably well; 2) national polls did not perform well because two large errors – underestimating Trump in red states and overestimating him in blue states – cancelled each other out; 3) as an explanation for the systemic error in state polls, the “Shy Trump” hypothesis is empirically infeasible.
The proposed paper challenges these three conclusions. First, evidence is offered that the ‘bad state polls - good national polls’ assertion rests upon questionable groundings while circumstantial evidence suggests that national polls were probably accurate because errors in the opposite directions cancelled each other out. Second, as an explanation for why most state polls underestimated Trump, the paper argues that a social desirability induced “Shy Trump” effect remains a viable hypothesis.
The AAPOR report presents a number of tests that probe the “Shy Trump” hypothesis. The proposed paper argues that most of those tests take some questionable assumptions and are largely inadequate. Most importantly, it’s been assumed that a “Shy Trump” effect would manifest in polling primarily as item non-response or misreporting pertaining to candidate preference measures. It may be a reasonable hypothesis, however, that the “Shy Trump” effect induced differential non-response, which might have been the primary source of the prediction error in state polls.
Evidence for this type of “Shy Trump” effect is inevitably circumstantial, since no data exists on individuals who do not respond to polls. Still, results from the 2016 American National Elections Studies (ANES) survey dataset offer such indirect evidence. First, the findings show that Trump supporters were significantly less likely to express their political views throughout the general election campaign than Clinton supporters. Second, Trump supporters’ public outspokenness appears to be moderated by education level: in particular, college educated Trump supporters refrained from the expression of their political views. Third, a three-way moderation effect suggests that college educated Trump voters’ reluctance to express their views was confined to less competitive states.
The paper argues that the findings are consistent with predictions derived from the Spiral of Silence framework. The normative pressures from the mainstream media and the dominant liberal segments of the college educated population could have made educated Trump supporters reluctant to reveal their political stripes. Finally, the study shows that linking the “Shy College-educated Trump supporter” phenomenon to estimated state-level differential non-response patterns could help explain the variations of prediction error across state polls.