After eight years of a Democratic administration, any successor candidate to Barack Obama - in this case Hillary Clinton - would have been at a disadvantage with a public ready for change.
Two weeks out from Election Day and it looks like the race for the White House is all but over. However, if it looks like pollsters are increasingly on the wrong side of history (Colombia’s referendum, Brexit, the 2015 British election and the Scottish referendum) it could be because they need to triple and quadruple check their assumptions before making blithe pronouncements. This is that type of thought exercise.
Regardless of who won the nomination, Campaign 2016 was always going to be close. After eight years of a Democratic administration, any successor candidate to Barack Obama – in this case Hillary Clinton – would have been at a disadvantage with a public ready for change. However, Obama’s more than 50 percent approval rating gives Republicans only a narrow, but not insurmountable, advantage. Models like this, which look to fundamentals, have a very good predictive value – about 80 percent – when it comes to election outcomes. But they’re not infallible. Perhaps most importantly in this case, they don’t take into account the individual candidates.
Republican nominee Donald Trump has run a unique campaign and has apparently altered many of the models. Over the past few months, Hillary Clinton has maintained a relatively steady 4-6 point lead in the polling, reversing what our “fundamentals model” would lead us to expect (a Republican victory). But will this remain the case – especially if voters are influenced by Friday’s announcement that the FBI is investigating additional emails relating to Clinton’s use of a private email server.
In short, what the fundamentals point to and what the polling currently suggests are materially at odds. One will be wrong, and our job is to validate and verify all our data-driven predictions.
Are polls accurately depicting the electorate, or are we systematically missing some bloc of voters that may sway the election? There are two main areas for investigation. First, are our surveys systematically under-representing Trump voters in our samples, leading us to have more Clinton voters than we should? This “shy Trumpers” thesis assumes Trump supporters are just not responding to our surveys. Second, we could have a good sample, but some Trump voters could be reluctant to volunteer their true support when asked. We’ll call this the “social desirability” thesis.
Pollsters overseas have pointed to non-response (meaning that certain groups – like Brexit voters – are less likely to respond to surveys) as a contributing factor to poll inaccuracy. The “shy Trumper” thesis asks if we are seeing that in the 2016 White House race. This potential problem is compounded by relatively low turnout in American elections. We need not only to get a correct representation of the U.S. population, but we also need to accurately anticipate which half of the population will be voters. A number of approaches have tried to answer this question, but we undertake a relatively simple one. For this “shy Trumper” thesis to be true, the electorate in our 2016 polls should be different from what we saw in previous election in a way that suggests Donald Trump supporters are being left out.
That’s not the case. Our Reuters/Ipsos polling in the 2012 presidential election was very accurate in predicting the final election results and therefore serves as a useful benchmark of the likely voter population. Using this polling, we compare the composition of our electorate (i.e. likely voters) in October 2012 to what we see in October 2016 across several demographic categories that have strong correlations to voter support.
This data indicates that the poll electorate is largely unchanged from 2012. The composition by race/ethnicity is stable; education level is mostly consistent; and age correlates well across the two election polls. Taken together, this suggests that our current polling is substantially capturing the same population we saw in our accurate 2012 polls. If anything, the population favors Trump slightly with older voters this year. Non-response does not appear to be playing a major factor in Trump’s current poll deficit.
Is Trump the victim of political correctness?
Trump and his supporters, in addition to several election observers, say that political correctness or “social desirability bias” play a part in Trump’s poll deficit. This is the notion that some people are unwilling to admit their true support for Trump because they are embarrassed or otherwise unwilling to disclose it in polls. Several studies have looked at this with mixed results from “yes, online polls show Trump stronger because there are no live interviewers” to “no, when asked in different ways, people still exhibit the same level of support for Trump.” Our sense is that social desirability is a minor factor, if one at all, particularly as Reuters/Ipsos polling is conducted online and therefore does not suffer from interviewer effects.
It seems more likely that there is a base of voters who ordinarily would vote Republican but cannot bring themselves to support Trump. After all, Trump’s favorability rating is the lowest in modern presidential polling and hostility towards him is not confined solely to Democrats.
If we take the “If the 2016 presidential election were held today, would you vote for Donald Trump or Hillary Clinton?” question at face value as our starting point, Clinton has a 6 percentage point lead among likely voters in the October Reuters/Ipsos poll. However, our presidential vote question has almost 20 percent of likely voters saying they are undecided, would vote third party or would not vote. This is a substantial block of probable voters who will most likely show up on Election Day but are not making known their presidential preference (if they have one.)
We use a few different tools to understand the potential voting behavior of this group. First, we can ask them our push question, “if you had to choose, which way would you go?” This prompts about a fifth of these undecided voters to pick a side, and these people seem to break evenly between Clinton and Trump. That still leaves us with over 15 percent of likely voters’ voting decision-making opaque to us.
From here, we have to start deriving their behavior. We know from other research that two main factors drive vote preference: the party identification of the respondent (Democrat, Republican, or independent) and how the respondent feels about the current office holder. Since we also ask about party ID and Obama approval in our Reuters/Ipsos polling, we have this information to hand. Combining this with our “push” question leaves less than one percent of likely voters still unknown.
Our data shows that when we this combine the results of both our basic vote question and the “push” question, Clinton’s lead of about six points holds solid. A second data point shows the vote totals from the combined questions and the derived support of the remaining undecided voters. Here Clinton’s lead drops to three percentage points.
Clearly, the undecideds can make a significant impact on this election. Clinton’s lead is large enough that she’ll win the popular vote even if – a big if –this bloc all shows up to vote and breaks toward Trump by almost 2 to 1. However, this potential vote could make a larger difference in several swing states – like Florida, Iowa or Ohio – where Trump and Clinton are polling much closer.
Far from a blowout, this election is closer than it seems.