I recently saw some discussion on Twitter about the potential for a historically low vote share in defeat for Donald Trump in the 2016 election. Given that all the major forecasters have pegged Hillary Clinton’s probability of winning this race between 86 and 99 percent, I don’t think we’re getting ahead of ourselves in entertaining the question of how Trump’s poor general election candidacy fits in history.
At the same, we are jumping the gun by not using current vote intention numbers in polls correctly. After a brief discussion over Twitter a few weeks ago with political science professor Hans Noel on another case of vote share calculation, I came away placing newfound importance on dealing with those who choose “undecided” when asked about their vote preference in polls. Namely, this addresses the following question: when predicting the eventual vote share on election day, should we remove the undecided vote from the denominator of the vote share calculation, or leave it in untouched?
For example, when we’re calculating Clinton’s vote share as a way of gauging what will best match the eventual outcome, we have two options:
- Clinton% / (Clinton% + Trump% + other% + undecided%)
- Clinton% / (Clinton% + Trump% + other%)
In deciding between what best models the future outcome, there was the obvious fact that people don’t actually vote “undecided” on election day. There was also a very interesting point another political science professor, Lynn Vavreck, made in an article for The Upshot (emphasis mine):
- In 2012, roughly half the people who were undecided at the end of the campaign stayed home. The other half split in similar proportions to the overall vote share. Despite the many ways that the 2016 election has been unexpected and unusual, the characteristics of undecided voters is not one of them. This may mean that we can expect half of them to stay home and the other half to split similarly to those who have decided already. They may be holdouts, but they are unlikely to be pivotal.
I believe Vavreck is making this important claim from a book she co-authored about the 2012 election. So the two main points here are 1) half of undecided’s stay home and 2) the other half broke equally for Obama and Romney. I agree with her subsequent assertion that the same dynamic should be more or less applicable to the 2016 election. If that’s the case, for the purposes of gauging the floor (minimum) support Trump will get from the American public–and understand if there’s a chance it will be historically low–removing undecided’s from the denominator is a prudent decision.
Of course, doing so rests on the assumption that current undecided voters will break evenly for Clinton and Trump. Maybe the strangeness of this election will encompass breaking this assumption, but it’s really hard to tell at this point. At the same time, it’s worth considering two points: for one, the 2012 outcome of half of undecided’s not turning out to vote doesn’t seem like it would change in this election (but on the other hand, for example, it’s conceivable that more undecided’s will break away from the most historically disliked nominee in Trump than usual). If we’re left dealing with just half of this already small group, it’s unlikely that too much of an effect or change will come about. Another more broader point: we always disproportionately focus on the irregularities observed in politics, and less on the reality that things are more likely to stay the same than change. The behavior of undecided voters may fall under this general rule.
So what happens if we do exclude undecideds from the denominator? In the Twitter discussion mentioned before, four losing presidential candidates were mentioned as having exceptionally low support: Barry Goldwater (1964), George McGovern (1972), Jimmy Carter (1980), and Walter Mondale (1984). For all four of these cases, I looked back at 2-3 presidential vote intention questions from polls around the same time we’re currently in, late October. In the below chart, I recorded the date of the polls with these questions, the support of these four candidates and that of their opponents–as a whole (“With undecideds”) and after adjusting to exclude undecided’s (“Without undecideds”)–the eventual vote share on election day, and whether adjusting to exclude undecided’s from the denominator brought the polling support closer to the actual vote share of the losing candidate.
Of the 11 cases examined for these four losing candidates (I couldn’t find a third poll in 1964), exclusion of undecided’s came closer to eventual vote share –relative to unadjusted numbers–nine times. Even when taking into account that 1) I’m examining four exceptional cases and 2) that in most cases I only used three polls in a certain time period, removing undecided percentages from the denominator of vote share calculations seems like the better option when predicting eventual vote share.
What does that mean for the 2016 election? At the bottom of the above chart, I include unadjusted numbers for Clinton and Trump–from HuffPost Pollsters two-way and three-way vote share trend lines–alongside adjusted ones that remove undecided’s. Trump’s current (unadjusted) estimated vote share of 41 and 39 percent in the two-way and three-way model, respectively, puts him squarely in the range of the Goldwater’s and Mondale’s of America’s electoral history. But removing undecided’s bumps him up 4-5 points at the moment. Considering the research and evidence shown here from the past, I think it’s very reasonable to hold off on the talk that Trump might end up with historically low numbers of support after all the votes are counted on November 8th.