They can see no reasons …
—“I Don’t Like Mondays”
The Boomtown Rats.
The Fine Art of Surfacing. 1979.
“Since Tuesday night,” John Cassidy wrote in The New Yorker this week, “there has been a lot of handwringing about how the media, with all its fancy analytics, failed to foresee Donald Trump’s victory”: as the New York Times headline had it, “How Data Failed Us in Calling an Election.” The failure of Nate Silver and other statistical analysts in the lead-up to Election Day rehearses, once again, a seemingly-ancient argument between what are now known as the sciences and the humanities—an argument sometimes held to be as old as the moment when Herodotus (the “Father of History”) asserted that his object in telling the story of the Greco-Persian Wars of 2500 years ago was “to set forth the reasons why [the Greeks and Persians] wage war on each other.” In other words, Herodotus thought that, to investigate war, it was necessary to understand the motives of the people who fought it—just as Cassidy says the failure of the press to get it right about this election was, Cassidy says, “a failure of analysis, rather than of observation.” The argument both Herodotus and Cassidy are making is the seemingly unanswerable one that it is the interpretation of the evidence, rather than the evidence itself, that is significant—a position that seems inarguable so long as you aren’t in the Prussian Army, dodging Nazi bombs during the last year of the Second World War, or living in Malibu.
The reason why it seems inarguable, some might say, is because the argument both Herodotus and Cassidy are making is inescapable: obviously, given Herodotus’ participation, it is a very ancient one, and yet new versions are produced all the time. Consider for instance a debate conducted by English literature professor Michael Bérubé and philosopher John Searle some years ago, about a distinction between what Searle called “brute fact” and “social fact.” “Brute facts,” Bérubé wrote later, are “phenomena like Neptune, DNA, and the cosmic background radiation,” while the second kind are “items whose existence and meaning are obviously dependent entirely on human interpretation,” such as “touchdowns and twenty-dollar bills.” Like Searle, most people might like to say that “brute fact” is clearly more significant than “social fact,” in the sense that Neptune doesn’t care what we think about it, whereas touchdowns and twenty dollar bills are just as surely entirely dependent on what we think of them.
Not so fast, said Bérubé: “there’s a compelling sense,” the professor of literature argued, in which social facts are “prior to and even constitutive of” brute facts—if social facts are the means by which we obtain our knowledge of the outside world, then social facts could “be philosophically prior to and certainly more immediately available to us humans than the world of brute fact.” The only way we know about Neptune is because a number of human beings thought it was important enough to discover; Neptune doesn’t give a damn one way or the other.
“Is the distinction between social facts and brute facts,” Bérubé therefore asks, “a social fact or a brute fact?” (Boom! Mic drop.) That is, whatever the brute facts are, we can only interpret them in the light of social facts—which would seem to grant priority to those disciplines dealing with social facts, rather than those disciplines that deal with brute fact; Hillary Clinton, Bérubé might say, would have been better off hiring an English professor, rather than a statistician, to forecast the election. Yet, despite the smugness with which Bérubé delivers what he believes is a coup de grâce, it does not seem to occur to him that traffic between the two realms can also go the other way: while it may be possible to see how “social facts” subtly influence our ability to see “brute facts,” it’s also possible to see how “brute facts” subtly influence our ability to see “social facts.” It’s merely necessary to understand how the nineteenth-century Prussian Army treated its horses.
The book that treats that question about German military horsemanship is called The Law of Small Numbers, which was published in 1898 by one Ladislaus Bortkiewicz: a Pole who lived in the Russian Empire and yet conducted a study on data about the incidence of deaths caused by horse kicks in the nineteenth-century Prussian Army. Apparently, this was a cause of some concern to military leaders: they wanted to know whether, say, if an army corp that experienced several horse kick deaths in a year—an exceptional number of deaths from this category—was using bad techniques, or whether they happened to buy particularly ornery horses. Why, in short, did some corps have what looked like an epidemic of horse kick deaths in a given year, while others might go for years without a single death? What Bortkiewicz found answered those questions—though perhaps not in a fashion the army brass might have liked.
Bortkiewicz began by assembling data about the number of fatal horse kicks in fourteen Prussian army corps over twenty years, which he then combined into “corp years”: the number of years together with the number of corps. What he found—as E.J. Gumbel pus it in the International Encyclopedia of the Social Sciences—was that for “over half the corps-year combinations there were no deaths from horse kicks,” while “for the other combinations the number of deaths ranged up to four.” In most years, in other words, no one was killed in any given corps by a horse kick, while in some years someone was—and in terrible years four were. Deaths by horse kick, then, were uncommon, which meant they were hard to study: given that they happened so rarely, it was difficult to determine what caused them—which was why Bortkiewicz had to assemble so much data about them. By doing so, the Russian Pole hoped to be able to isolate a common factor among these deaths.
In the course of studying these deaths, Bortkiewicz ended up independently re-discovering something that a French mathematician, Simeon Denis Poisson, had already, in 1837, used in connection with discussing the verdicts of juries: an arrangement of data now known as the Poisson distribution. And as the mathematics department at the University of Massachusetts is happy to tell us (https://www.umass.edu/wsp/resources/poisson/), the Poisson distribution applies when four conditions are met: first, “the event is something that can be counted in whole numbers”; second, “occurrences are independent, so that one occurrence neither diminishes nor increases the chance of another”; third, “the average frequency of occurrence for the time period in question is known”; and finally “it is possible to count how many events have occurred.” If these things are known, it seems, the Poisson distribution will tell you how often the event in question will happen in the future—a pretty useful feature for, say, predicting the results of an election. But that what wasn’t was intriguing about Bortkiewicz’ study: what made it important enough to outlast the government that commissioned it was that Bortkiewicz found that the Poisson distribution “may be used in reverse”—a discovery ended up telling us about far more than the care of Prussian horses.
What “Bortkiewicz realized,” as Aatish Bhatia of Wired wrote some years ago, was “that he could use Poisson’s formula to work out how many deaths you could expect to see” if the deaths from horse kicks in the Prussian army were random. The key to the Poisson distribution, in other words, is the second component, “occurrences are independent, so that one occurrence neither diminishes nor increases the chance of another”: a Poisson distribution describes processes that are like the flip of a coin. As everyone knows, each flip of a coin is independent of the one that came before; hence, the record of successive flips is the record of a random process—a process that will leave its mark, Bortkiewicz understood.
A Poisson distribution maps a random process; therefore, if the process in question maps a Poisson distribution, then it must be a random process. A distribution that matches the results a Poisson distribution would predict must also be a process in which each occurrence is independent of those that came before. As the UMass mathematicians say, “if the data are lumpy, we look for what might be causing the lump,” while conversely, if “the data fit the Poisson expectation closely, then there is no strong reason to believe that something other than random occurrence is at work.” Anything that follows a Poisson distribution is likely the result of a random process; hence, what Bortkiewicz had discovered was a tool to find randomness.
Take, for example, the case of German V-2 rocket attacks on London during the last years of World War II—the background, as it happens, to novelist Thomas Pynchon’s Gravity’s Rainbow. As Pynchon’s book relates, the flying missiles were falling in a pattern: some parts of London were hit multiple times, while others were spared. Some Londoners argued that this “clustering” demonstrated that the Germans must have discovered a way to guide these missiles—something that would have been highly, highly advanced for mid-twentieth century technology. (Even today, guided missiles are incredibly advanced: much less than ten percent of all the bombs dropped during the 1991 Gulf War, for instance, had “smart bomb” technology.) So what British scientist R. D. Clarke did was to look at the data for all the targets of V-2s that fell on London. What he found was that the results matched a Poisson distribution—the Germans did not possess super-advanced guidance systems. They were just lucky.
Daniel Kahneman, the Israeli psychologist, has a similar story: “‘During the Yom Kippur War, in 1973,’” Kahneman told New Yorker writer Atul Gawande, he was approached by the Israeli Air Force to investigate why, of two squads that took to the skies during the war, “‘one had lost four planes and the other had lost none.’” Kahneman told them not to waste their time, because a “difference of four lost planes could easily have occurred by chance.” Without knowing about Bortkiewicz, that is, the Israeli Air Force “would inevitably find some measurable differences between the squadrons and feel compelled to act on them”—differences that, in reality, mattered not at all. Presumably, Israel’s opponents were bound to hit some of Israel’s warplanes; it just so happened that they were clustered in one squadron and not the other.
Why though, should any of this matter in terms of the distinction between “brute” and “social” facts? Well, consider what Herodotus wrote more than two millennia ago: what matters, when studying war, is the reasons people had for fighting. After all, wars are some of the best examples of a “social fact” anywhere: wars only exist, Herodotus is claiming, because of what people think about them. But what if it could be shown that, actually, there’s a good case to be made for thinking of war as a “brute fact”—something more like DNA or Neptune than like money or a home run? As it happens, at least one person, following in Bortkiewicz’ footsteps, already has.
In November of 1941, the British meteorologist and statistician Lewis Fry Richardson published, in the journal Nature, a curious article entitled “Frequency of Occurrence of Wars and Other Quarrels.” Richardson, it seems, had had enough of the endless theorizing about war’s causes: whether it be due to, say, simple material greed, or religion, or differences between various cultures or races. (Take for instance the American Civil War: according to some Southerners, the war could be ascribed to the racial differences between Southern “Celtics” versus Northern “Anglo-Saxons”; according to William Seward, Abraham Lincoln’s Secretary of State, the war was due to the differences in economic systems between the two regions—while to Lincoln himself, perhaps characteristically, it was all due to slavery.) Rather than argue with the historians, Richardson decided to instead gather data: he compiled a list of real wars going back centuries, then attempted to analyze the data he had collected.
What Richardson found was, to say the least, highly damaging to Herodotus: as Brian Hayes puts it in a recent article in American Scientist about Richardson’s work, when Richardson compared a group of wars with similar amounts of casualties to a Poisson distribution, he found that the “match is very close.” The British scientist also “performed a similar analysis of the dates on which wars ended—the ‘outbreaks of peace’—with the same result.” Finally, he checked another data set concerning wars, this one compiled by the University of Chicago’s Quincy Wright—“and again found good agreement.” “Thus,” Hayes writes, “the data offer no reason to believe that wars are anything other than randomly distributed accidents.” Although Herodotus argued that the only way to study wars is to study the motivations of those who fought them, there may in reality be no more “reason” for the existence of war than the existence of a forest fire in Southern California.
Herodotus, to be sure, could not have seen that: the mathematics of his time were nowhere near sophisticated enough to run a Poisson distribution. Therefore, the Greek historian was eminently justified in thinking that wars must have “reasons”: he literally did not have the conceptual tools necessary to think that wars may not have reasons at all. That was an unavailable option. But through the work of Bortkiewizc and his successors, that has now become an option: indeed, the innovation of these statisticians has been to show that our default assumption ought to be what statisticians call the “null hypothesis,” which is defined by the Cambridge Dictionary of Statistics to be “the ‘no difference’ or ‘no association’ hypothesis.” Unlike Herodotus, who presumed that explanations must equal causes, we now assume that we ought to be first sure that there is anything to explain before trying to explain it.
In this case, then, it may be that the “brute fact” of the press’ Herodotian commitment to discovering “reasons” that explains why nobody in the public sphere predicted Donald Trump’s victory: because the press is already committed to the supremacy of analysis over observation, it could not perform the observations necessary to think Trump could win. Or, as Cassidy put it, when a reporter saw the statistical election model of choice “registering the chances of the election going a certain way at ninety per cent, or ninety-five per cent, it’s easy to dismiss the other outcome as a live possibility—particularly if you haven’t been schooled in how to think in probabilistic terms, which many people haven’t.” Just how powerful the assumption of the force of analysis over data can be is demonstrated by the fact that—even despite noting the widespread lack of probabilistic thinking—Cassidy still thinks it possible that “F.B.I. Director James Comey’s intervention ten days before the election,” in which Comey announced his staff was still investigating Clinton’s emails, “may have proved decisive.” In other words, despite knowing something about the impact of probability, Cassidy still thinks it possible that a letter from the F.B.I. director was somehow more important to the outcome of this past election than the evidence of their own lives were to million of Americans—or, say, the effect of a system in which the answer to the question where outweighs that of how many?
Probabilistic reasoning, of course, was unavailable to Herodotus, who lived two millennia before the mathematical tools necessary were even invented—which is to say that, while some like to claim that the war between interpretation and data is eternal, it might not be. Yet John Cassidy—and Michael Bérubé—don’t live before those tools were invented, and yet they persist in writing as if they do. While that’s fine, so far as it is their choice as private citizens, it ought to be quite a different thing insofar as it is their jobs as journalist and teacher, respectively—particularly in the case, as say in the 2016 election, when it is of importance to the continued health of the nation as a whole that there be a clear public understanding of events. Some people appear to think that continuing the quarrels of people whose habits of mind, today, would barely qualify them to teach Sunday school is something noble; in reality, it may just be a measure of how far we have yet to travel.