Opinion, Berkeley Blogs

Deciphering election polling, from algorithms and youth votes to the Electoral College

By Laura Stoker


The outcome of the 2016 general election in the United States was momentous and surprising. Yet some commentaries seem to suggest that means we need to rethink our basic understandings of voters and elections. I disagree, albeit gently, with that kind of hyperbole. Here’s why.

Let’s not exaggerate

As in 2000, it appears the winner of the election lost the popular vote. Although the polls were off (more on that below), it is wrong to exaggerate how far they were off.

There was always a lot of uncertainty in the estimates of swing state outcomes, which — as statistician and editor of the FiveThirtyEight blog Nate Silver would repeatedly point out — are correlated. If Donald Trump looked to be doing better than expected in one, he would likely do better than expected in others, which is what happened.

A sensitive Electoral College

In an election in which over 100 million votes were cast, a remarkably small number and percentage of votes can shift the electoral votes won by Hillary Clinton versus Trump. The institution of the Electoral College takes small swings in votes and magnifies them into big swings in its outcomes.

Where the polls stood as the nation went to vote. Striped areas were within the margin of error. (Wikimedia Commons)

A former UC Berkeley political science Ph.D. student, David Hopkins, now an assistant professor at Boston College, has a forthcoming book that elaborates this theme at length: Divided by Region: How Geography and Electoral Rules Polarize American Politics (Cambridge University Press).

Clinton’s vulnerability

Related to both points, above, it has long been recognized by political scientists, pollsters, pundits and journalists that Hillary Clinton’s support was not distributed in a way facilitating an Electoral College victory.

It is better if a candidate’s votes lead to many small state margins of victory than to fewer large state margins of victory. Clinton’s support (and Democratic support more generally) tends to the latter. Her Electoral College vulnerability was greater than any poll gap would naively suggest.

Polling failures

Despite the fact that the vote tallies were within polling margins of error, generally speaking, almost all predicted a Clinton victory. What was going on?

As I said at an Institute of Governmental Studies event in September, I have worried about the veracity of the polls for some time. There is a great deal of systematic bias in telephone and web surveys (more so than in face-to-face surveys, not used in election polling), which is usually (or at least often) not adequately corrected by post-hoc weighting.

Youth vote

Moreover, a seemingly under-appreciated fact about this bias is that it is not distributed equally across demographic groups.  Much more research on this question is needed, but it appears that the bias is greatest for groups with a low propensity to respond in the first place, e.g., young, urban, non-white, less educated. With the exception of the last group, this time around anyway, these tend to be Democratic constituencies.

Polling failures in Britain (2015 election, Brexit) can be tied to the polls’ erroneous estimates regarding such groups. For example, the turnout of young people was seriously over-estimated in the 2015 YouGov British election polling, which was the most important reason that YouGov overestimated support for the Labour Party. I would not be surprised to discover that the polls overestimated youth and non-white turnout this time.

The undecideds

The level of uncertainty in the polling was higher than usual this year because there were unusually high numbers of people reporting that they were undecided at all points of the general election campaign.

Exit polls suggest that late deciders, a presumably larger group than normal, went with Trump. Exit polls also suggest that partisan defection rates were higher than in the past, which was expected, though about equal across the parties (with roughly 90 percent of partisans supporting their nominee), which was not expected.

Given the asymmetry in citizens’ evaluations of the candidates’ qualities, I would have expected defection rates to be higher for Republicans than for Democrats.

Enthusiasm for the candidates

Polling predictions are always based on some sort of likely voter algorithm.  These algorithms can be wrong.  I find it intriguing that the LA Times poll, much maligned for various reasons including the fact that it continually showed more Trump support than others, was one of the few or only one to use a self-reported enthusiasm measure in making its predictions.

Some political science research has claimed that enthusiasm for one candidate is a stronger turnout motivator than fear of the alternative candidate. Trump was winning the enthusiasm battle all campaign, according to the LA Times poll.  And early reports suggest that the turnout of key Democratic groups was down — especially the young and blacks.

The exit polls also show strong support for third-party candidates among the young (along the order of 9 percent, compared to the national average of about 3 percent). They did not turn toward Trump, but the enthusiasm for Bernie Sanders never became enthusiasm for Clinton.

There are other possibilities. One speculation is that some Trump supporters were hiding their Trump support; I’m doubtful, but it is a dim possibility. There might be mode effects operating. Telephone surveys tend to have less sample bias than web surveys, but those surveyed by the web see a reasonable facsimile of the ballot whereas those interviewed by phone do not. This difference might be worth exploring.

Trump’s appeal

In general, I would not answer the question of why Trump won the nomination the same way I would answer the question of why he won the general election, though his appeal to conservatives on immigration and trade and his anti-elitism would be a part of both stories.

Party identification proved extremely important, as always, even though defection rates were up.  Low turnout among the Democratic base, probably tied to a lack of enthusiasm for Clinton, is almost certainly key, as is the modest advantage that Trump appeared to have gained among the independents, presumably tied to his conservative messaging. Whether Clinton’s gender counted as a negative is a possibility that scholars will surely scrutinize down the road.