Association for Public Policy Analysis & Management

On the eve of the presidential election, randomly selected registered voters in four early voting states, namely Ohio, Florida, Nevada and North Carolina, revealed a divide was emerging between those who had cast early ballots and those heading to the polls just days before November 8^th.  A panel of voters randomly selected from the L2 registered voter database, and recruited by cell to participate in an online survey, signaled the lead held by Hillary Clinton was trending towards Donald Trump. This trend was moving in the opposite direction of the findings of many statewide polls fielded by phone.

We hypothesized that this divergence may be correlated to a “shy” Trump voter unwilling to acknowledge a preference for a candidate that evoked strong emotional responses among the electorate. We also hypothesized that this “shy voter” effect may be correlated to Clinton’s level of expressed support. Finally, we hypothesized that this “shyness” may extend to respondents’ expressed policy preferences collected when surveying questions concerning “politically” or “socially” sensitive issues.

We found among those early voters surveyed, only 50 percent reported they had told a family member who they were voting for president. Just 42 percent had shared their vote preference with a friend and only a quarter of respondents had told a co-worker about their presidential decision. Clinton supporters were decidedly more willing to share their vote preference compared to Trump supporters by a 20 point margin.

These preliminary findings suggested several possibilities: history had intervened in the final days before the election, voters were understating their likelihood to cast a ballot, respondents were unwilling to tell phone pollsters their preferences due to interviewer effect based on race and gender, or all of these factors had converged.

Our research assesses the reliability of expressed public preferences collected through random probability surveys about “politically and socially sensitive” related questions. To what extent does a respondent’s hypothesized “shyness” disappear when surveyed online? Additionally, can design effects be minimized by pre-stratifying the sample frame to ensure “shy” voters are more likely to be interviewed in proportion to the population parameters?

To examine these questions, we began by empaneling voters in five states (N=728). Two additional surveys are planned in these states during the first 100 days of the Trump Administration and at the six-month mark. Three additional panels of registered voters are also planned in Pennsylvania, Michigan and Wisconsin. We will examine the impact survey modality had on expressed presidential choice, as well as expressed policy preferences for current immigration and civil rights policies, and the Affordable Care Act (Obamacare) among registered voters in these eight states.