How the polls helped Trump beat Clinton in the 2016 Presidential Election
If there is one thing with which almost everyone will agree about our recent elections, is that the results were surprising. To be specific, what surprised people was the discrepancy between the polls leading up to the election, and the outcome of the election.
This has been the subject of significant discourse since the election. Some point out that statistics are misinterpreted, especially the difference between probability and certainty; some argue that what people say doesn’t always correspond to what people do; others point out the flaws in polling methodologies; and yet others suggest that data-driven approaches fail to account for context. Ultimately, all of these arguments revolve around whether polls are good predictors of actual outcomes.
I would like to suggest that arguing about the validity of polls is sillier than closing the barn door after the cows have gone: it’s like arguing about the door when the barn is missing an entire wall. Why? Because what we should be asking about polls is not whether they are accurate, but whether they have a direct impact on the outcome. And the answer will not come from understanding statistical models, but from understanding human behavior, as I will now explain.
First of all, it should be clear that the outcome of elections depends entirely on the behavior of individuals, namely the act of casting a vote. Leaving aside the nuances of our electoral system, if more people actually vote for one candidate, that candidate will win, even if a greater number of people actually prefer the other candidate.
Polls, on the other hand, reflect the opinion of individuals. When we analyze poll results, we make the implicit assumption that the likelihood that an individual will actually vote is independent of which candidate they prefer. In other words, there is no a priori reason to believe that members of one party are implicitly more or less likely to vote than members of the other party.
However, what is being overlooked is that polls, while probably not influencing the opinion of individuals, can actually influence the behavior of individuals. Specifically, I believe that polls have a direct and significant impact on the likelihood that people will actually cast a vote, and that in some circumstances the polls can have a different influence on different supporters in such a way that effectively reverses the outcome of an election.
Let’s begin with some observations. There is evidence from academic research that the factors that influence our opinion about something, are distinct from the factors that lead us to act upon it. Consider for example the difference between thinking that exercise is good and actually going to the gym; or between agreeing with a Foundation’s initiative and actually donating money; or between liking a particular product and writing an online review for it. In other words, the psychological mechanisms that influence our preferences are distinct from the psychological mechanisms that lead us to take actions based on those preferences.
In the context of voting to elect the next President, our opinion reflects which candidate we prefer, while our action is whether or not we actually cast a vote. Assuming for a moment that an individual has a clear preference for one candidate, let’s think about two factors that influence whether or not this individual will actually cast a vote: (1) the strength of the individual’s preference for one candidate over the other; generally, the greater the preference, the more someone will be motivated to vote. (2) The perceived importance of casting a vote, in the sense of “how much will my vote count?”; generally speaking, if you feel that your vote will make no difference whatsoever, your likelihood of casting a vote will be lower than if you think your vote will make a difference.
And this is where the polls come into the picture. Consider an individual with a strong preference for Clinton, and a second individual with an equally strong preference for Trump. If the candidates are neck-to-neck in the polls, then, all other things equal, the two supporters will have the same degree of urgency to cast their vote. On the other hand, if the polls show that Clinton is leading, the Trump supporter will probably have a slightly greater sense of urgency to cast a vote than the Clinton supporter. This, in turn, means that more Trump supporters will vote.
How big of a factor is this, and how much of a difference could it make? In states like California and Massachusetts, where Clinton won by a wide margin, or Wyoming and West Virginia, where Trump won by an even larger margin, the impact of polls on the urge to vote may be insignificant. But in swing states like Florida and Pennsylvania, where Trump won by small margins, this may have been a determining factor.
Let’s look at Florida and Pennsylvania, because if Clinton had won these two states, she would have had 277 electoral votes to Trump’s 241. In Florida, 4,605,515 people voted for Trump, 4,485,745 for Clinton, and 295,490 voted for other candidates. This means that 119,770 Republican votes out of a total of 9,386,750 votes swung the election in Trump’s favor — that’s only 1.3% of the voting population. In Pennsylvania, the same reasoning shows that a mere 1.1% of the voting population pushed Trump to a victory.
So, if the poll results showing a strong lead for Clinton motivated about 150,000 Florida Republicans to vote who may otherwise not have voted, and led 150,000 Florida Democrats not to vote who may otherwise have voted, the polls effectively caused the result to flip. Similarly, if the polls influenced the voting behavior of about 60,000 Republicans and 60,000 Democrats in Pennsylvania, they flipped that result as well.
In fact, we can envision a situation in which one candidate is truly and measurably preferred by more people, and yet the very act of reporting that observation causes the less-preferred candidate to win, simply because his supporters are more motivated to vote. This hypothesis is very different from assuming that the polls somehow failed to reflect “reality.” Rather, I am suggesting that even if the polls are an accurate reflection of reality, making the poll results known can cause the outcome to be reversed.
While political polls have been around for a long time, in the last several years there has been an explosion both in the number of polls and how easily individuals can access the polls. Hence, even if the influence of poll results on the likelihood of voting is modest, the greater availability can tip the scales. And I suspect that this may have been a deciding factor in Trump’s somewhat surprising victory in 2016.
PS: Inspired by this post I have created a simple computer simulation in NetLogo that shows that it is in fact possible, under very reasonable circumstances, to see exactly the kind of “flipped” result that I predicted in this post. I will write a separate post with an explanation of the model and the results.