Taking errors seriously in forecasting elections



Science isn’t about predicting one-off events like election results, but that doesn’t mean the way to make election forecasts scientific (which they should be) is to build “theories of voting.” A number of people have sent me articles on statistical aspects of the recent U.S. election, but I don’t have much to say and I like to keep my blog non-political. I won’t violate this rule in making a couple of comments on Faye Flam’s Nov. 11 article: “Why Science Couldn’t Predict a Trump Presidency”[i].

For many people, Donald Trump’s surprise election victory was a jolt to very idea that humans are rational creatures. It tore away the comfort of believing that science has rendered our world predictable. The upset led two New York Times reporters to question whether data science could be trusted in medicine and business. A Guardian columnist declared that big data works for physics but breaks down in the realm of human behavior.

But the unexpected result wasn’t a failure of science. Yes, there were multiple, confident forecasts of win for Clinton, but those emerged from a process doesn’t qualify as science. And while social scientists weren’t equipped to see a Trump win coming, they have started to test theories of voting behavior that could shed light on why it happened…..

Not that these methods are pseudoscience; in fact, they employ some critical tools of science. The most prominent among those is Bayesian statistics, a way of calculating the probability that something is true or will come true.

Bayesian analysis is a core principle laid out in political forecaster Nate Silver’s book “The Signal and the Noise.” Though developed in the 1700s, Bayesian statistics had a resurgence in the science of early 21stcentury. …

Why don’t Bayesian statistics work the same sort of consistent magic for political forecasts? In science, what matters isn’t the forecast but the nature of the models. Scientists are after explicit rules, patterns and insights that explain how the world works. Those give other scientists something to build on — allowing science to self-correct in a way that other intellectual ventures can’t…..

Now that it’s over, there’s still a chance for science to explain why so many people voted for Trump. There are all kinds of guesses and judgments being thrown around about Trump voters — that they’re racist or sexist, or responding to the call of tribalism. Those aren’t the least bit scientific, but they could be turned into testable hypotheses.

You can read the rest of her article here.

Anytime a purportedly scientific method fails, a defender can always maintain the failures weren’t really scientific applications of the method. I think we did see a failure of many of the polling methods as the basis for the best-known forecasts. Methods used by Trump’s internal polling alerted them to what was happening in the “rust belt states” (according to campaign manager and pollster Kellyanne Conway), but the other polls largely missed it. They didn’t really share those internal results, and the attention Trump gave to typically blue states perplexed many [For some other activities kept under wraps, see ii]. Bill Clinton, on the other hand, “had pleaded with Robby Mook, Mrs. Clinton’s campaign manager, to do more outreach with working-class white and rural voters. But his advice fell on deaf ears.” (Link is here.)

Flam suggests, on the basis of her interviews with social scientists, that the way to turn forecasts into science is to build theories of voting. My guess is that’s the wrong way to go (I don’t claim any expertise here.) It’s an understanding of the threats to the assumptions in the particular case, with all its idiosyncrasies, that’s called for. The only thing general might be the ways you can go wildly wrong. Do their theories include tunnel vision by pollsters? Perhaps they should have asked: “If you were a person planning to vote for Trump, would you be reluctant to tell me, if I asked?”[iii]. In Trump’s internal polling, they would deliberately ask a number of related questions to ferret out the truth. Of course pollsters are well aware of the “undercover” or “shy” voter who is too worried about giving an unacceptable answer to be frank. If there was ever a case where this would be likely, it’s this—yet it was downplayed. Ironically, one might expect the more the “shy” voter should have been a concern, the less seriously a pollster would take it. (You can ponder why I say this.) It’s not enough to have a repertoire of errors if they’re not taken seriously.

As for the “consistent magic” of Bayesianism, since in this case we’re talking about an event, frequentists, error statisticians, and Bayesians can talk Bayesianly if they so choose, but my understanding is that most polling is in the form of frequentist interval estimates (perhaps with various weights attached). Maybe, as Flam suggests, some formally combine prior beliefs with the statistical data, but that’s all the more reason to have been ultra-self-critical and probe how capable the method is at disinterring fundamental flaws in the model. They should have been giving their assumptions a hard time, bending over backwards to disinter biases, and self-sealing fallacies, not baking them into the analysis.

Share your thoughts.

[i] Flam is the one who interviewed me, Gelman, Simonsohn, Senn and others for that NYT article on Bayesian and frequentist methods discussed on this post.

[ii] Apparently they also kept hidden in the “Trump bunker” a fairly extensive use of data analytics (on the order of $70 million a month, according to the Bloomberg article “Inside the Trump Bunker”), encouraging people to think it was a fledgling effort. Their polls were in sync with Nate Silver’s, they say, except for the time lag in Silver’s, owing to his reliance on other polls, but their inferences about what voters really thought differed.

Trump’s data scientists, including some from the London firm Cambridge Analytica who worked on the “Leave” side of the Brexit initiative, think they’ve identified a small, fluctuating group of people who are reluctant to admit their support for Trump and may be throwing off public polls. (Inside the Trump Bunker)

The article admits they also worked toward selectively depressing the vote. The overall data analytic project was to be the basis of an enterprise to pursue after a potential loss!

[iii] This is reminiscent of the question that permits you to get at the truth (about the correct road to town) when confronted with people who either always lie or always tell the truth.


Categories: Bayesian/frequentist, evidence-based policy

Post navigation

15 thoughts on “Taking errors seriously in forecasting elections

  1. Stan

    In the last few days Trump went to the correct states that could be won. That is a vote that his internal polling was on target.

  2. Thomas R. Dyckman

    I suggest that the voting system that is run, I believe out of Los Vegas with a monetary prize and continuously updated would offer the best results currently available, A free input with no questions asked or interpreted.

  3. Nate Silver did a pretty decent job – a lot of people were criticising him for giving Trump such a high chance (~30%) of winning.

    • OM: True, at some points though he basically gave a disjunction that was essentially tautologous. Of course he was building on other polls. I hope he will be moved to develop new methods, maybe employee some of the statistical tricks to get around a reluctance to give a frank answer.

      • Sure, it’s hard to know how to evaluate predictions like Trump has a 30% chance of winning etc. But Nate consistently warned that the race was tighter than people thought.

        Do you want him to make a more certain prediction even when the situation is uncertain or do you want him to acknowledge the uncertainty? I would prefer a ‘we don’t know/it’s hard to say’ over false certainty.

        What about those who confidently predicted a Trump victory? Should they get more ‘points’ than Nate saying it’s possible?

        More generally, I don’t think the claim ‘Bayesian inference failed’ passes a severe test. Not saying it’s the only way, of course.

  4. Polling predictions prior to the 2015 UK general election were wrong and pollsters received a lot of criticism. When you look a little closer you find that the polls correctly predicted the % voting for Labour and Conservative combined and they correctly predicted the LibDem and UKIP results. What they got wrong was the partition of votes between Conservative and Labour.

    Between 2010 and 2015 there were huge changes in voting patterns. The Libdem vote dropped from 24% to 8%, amounting to 5 million votes. Meanwhile UKIP increased from 4% to 13%, an increase of 3 million. So 8 million people (out of 30 million) changed sides. The ideologies of LibDems and UKIP are so different that it is very unlikely that 3 million LibDem votes in 2010 went to UKIP in 2015. So the additional UKIP votes in 2015 must of come from Labour and Conservative, but we didn’t notice because these lost votes were replaced by former LibDem votes. So, that’s another 3 million we can add to the 8 million million.

    So, in 2015 at least 11 million people – over 33% of voters – changed sides. Now, in my personal experience it’s not easy to change political affiliations, so we can comfortably assume (can we??) that a large percentage of those 11 million didn’t make up their minds until they were standing in the polling booth. My guess is that a lot of former LibDems swapped from Labour to Conservative at the last minute. The natural home of a lapsed LibDem would be Labour, but the Labour leader was seen as geeky and lacking leadership skills.

    If I have accurately portrayed what happened it would have been difficult to impossible for polls to predict the outcome. A theory of voting would not have helped out here. Nor would a change of methodology.

    • Peter Chapman


      I think I have double counted. It’s 8 million not 11 million. Still a lot though.

    • Peter: Our voting “booths” have become little cardboard sides–they got rid of the curtains years ago. So I don’t think much decision-making goes on while standing there.

  5. Perhaps the truth is much simpler: we all failed to update our priors

    • In other words, it was a failure of scientists, not of science

    • Enrique: No, that’s the point, Bayesian inference failed, says Faye. It’s a good example of why it would. I’ve heard people refer to a “herding” effect in polling, where pollsters adjust their polls so as to not be out of step with others or with their own beliefs. Some speculate that there was also a reluctance to report evidence that Trump might win. If they had, Clinton might have been able to adjust her focus.

  6. Steven McKinney

    We are in a time period with similarities to the circumstances of the 1940s, when the Chicago Daily Tribune printed a run of newspapers declaring “Dewey Defeats Truman” in the 1948 election.

    The trouble then was pollsters were calling people on the newfangled telephone, and it turned out that telephone ownership was biased towards more well-off urbanites who tended to be Republican.

    Now we have mountains of young people using smart phones and internet communication channels that pollsters have not figured out how to use to get at a truly random cross section of the voting population. Many smart phone users will not answer an unknown calling number, so pollsters end up talking to more older people who still just pick up the handheld home phone. They were more likely Clinton supporters.

    Until pollsters relearn how to engage a reliable random cross section of the voting public, this problem will persist. Easy and cheap on-line polls and random dialing of telephone numbers no longer provide a reasonable sample of voters with which to derive accurate estimates of voting patterns. Guesses by pollsters as to how to up or down weight various groups of respondents to compensate for this issue clearly did not work, whatever statistical modeling paradigm they chose to operate under.

  7. martin

    It seems that the biggest error in these polls is systematic rather than random. If we have biased methods then there’s little to do with statistical analysis (Bayesian or frequentist). Maybe bias analysis that can produce several scenarios.

    However, people don’t want to know about odds and scenarios they want to know who’s going to win and those who conduct and report polls have to point to that target.

    I think a 30% probability for the Republican candidate was a good bet, even more now that we know the Democrat candidate won on the popular vote.

    Thanks for the post.

Blog at WordPress.com.