The Theory That Would Not Die

by Sharon Bertsch McGrayne

  To appreciate the audacity of NBC’s job offer to Tukey, one needs to understand how deeply he was embedded in Cold War secrets. He had done research on topology in the late 1930s and military analysis in the 1940s. As a young man during the Second World War, Tukey worked in Princeton with the Operations Research Group that computed how a B-29 bomber speeding over Europe should aim its machine-gun fire. With pencil and paper during the Cold War, he broad-brushed the aerodynamics, trajectory, and warhead for the Nike, the first antiaircraft surface-to-air missile system. He also helped persuade Eisenhower to build the U-2, the spy plane that flew from 1956 until 1960, the year a U-2 pilot Francis Gary Powers was shot down over the USSR.

  When NBC News approached Tukey in 1960, he had been a member of the CIA’s Science and Technology Advisory Panel and the National Security Agency’s Science Advisory Board for eight years. His most famous advisory role had occurred the year before, when, as a delegate to the US–USSR Conference on the Discontinuance of Nuclear Weapon Tests, he surprised the Soviet delegation by showing that seismogram data could distinguish underground nuclear explosions from earthquakes. Once both sides knew they could police each other’s compliance, they signed the Partial Test Ban Treaty in 1963 to ban nuclear tests in the atmosphere, space, and sea.

  Tukey also helped establish a super-secret cryptography think tank at Princeton University. The communications research division of the Institute of Defense Analyses (IDA) moved into Von Neumann Hall, a new campus building surrounded by an eight-foot-high brick wall. IDA, which had “the most intimate ties” to the National Security Agency, was created to solve advanced cryptographic problems.1 Although the position did not appear on his curriculum vitae, Tukey served on IDA’s board of trustees for decades. Student protests against secret research in universities forced IDA off campus in 1970, despite Tukey’s personal appeal to Princeton’s president, Robert F. Goheen.

  Many university faculty did classified work as part of their regular duties during the 1950s and 1960s. John Pratt and Stephen Fienberg, for example, were cleared for such work at the University of Chicago. Said Fienberg, “When I joined the faculty in the Department of Statistics in 1968, it had a dual contract with the Navy office of research. One part supported basic statistical research, and the other was for statistical consulting. We had a safe in the basement where they kept the classified consulting work, although I do not know which faculty had been working on it.”2

  Tukey also worked closely with members of Princeton’s physics department, which was “highly involved in the design of the atomic and later hydrogen bombs.”3 After the United States dropped atomic bombs on Japan in 1945, the director of the Manhattan Project, Gen. Leslie R. Groves, asked the Princeton physics chair Henry Smyth to write the official explanation of the bomb, Atomic Energy for Military Purposes. In 1951 Princeton launched a secret undertaking, Project Matterhorn, to design thermonuclear weapons at its nearby Forrestal Research Center. Tukey evaluated Edward Teller’s and Stanislaw Ulam’s design for the first H-bomb early that year. According to his curriculum vitae, Tukey served Forrestal Research Center as “supervisor, Military Systems Analyst” from 1951 until 1956.4 Physics professor John A. Wheeler, who headed the weapons program, stated, “I believe that the whole country—scientifically, industrially, financially—is better off because of him and bears evidence of his influence.”5

  In addition to his military research at Princeton, Tukey taught classes and supervised more than 50 graduate students. In appearance he could be “a bouncy and beefy extrovert” and “sort of cherubic looking with a pleasant manner.” But his lecture style was oblique at best. Invited to speak at Imperial College in London in 1977, Tukey looked like a great bear of a man in old baggy pants. Sitting in a cross-legged Buddha pose on the podium, he began his lecture by asking slowly and deliberately, “Comments, queries, suggestions?”6 During the long wait that ensued, he ate prunes—12 of them, one by one—until someone in the audience finally asked if he could explain something or other. Only then did Tukey begin speaking. When a graduate student asked one January for an appointment to discuss his Ph.D. thesis, Tukey checked his diary and said he was going to a meeting in two months and if the student drove him there they could talk about it in the car.

  Tukey also advised the federal government on a wide range of civilian problems: air quality, chemical pollution, ozone layer depletion, acid rain, census methodology, and educational testing.

  How did he manage all this? Stories are legion about Tukey sitting in the back row at a seminar, dozing, reading mail, scanning newspapers, or editing articles, but then rising at the end of the talk to critique it. Tukey drafted articles in pencil while listening to baroque brass recordings, topped the article with the words “By ____ and John W. Tukey,” gave the manuscript to one of his two longtime secretaries, and then searched for a collaborator to finish the piece. He put his name to about 800 publications and worked with more than 105 coauthors, including Jerome Cornfield at NIH, but most frequently with his friend Fred Mosteller at Harvard.

  As a result of Tukey’s heavy military and teaching workloads, his future father-in-law fully expected him to whip out pad and pencil while waiting at the altar to be married. His bride, Elizabeth R. Rapp, was personnel director of the three-year-old Educational Testing Service. Later, she confided that “as the wife of [a] dedicated workaholic, I understand the selfless love and devotion, accommodation and deprivation required to ‘keep them on the road.’” After Elizabeth’s death in 1998, Tukey said, “One is so much less than two.”7

  According to Elizabeth, Tukey organized and simplified his personal life like “a New Englander through and through.”8 His conversation was quiet and measured and excluded personal comments and idle chatter. His nephew Frank R. Anscombe contended that Tukey had few wants, although they included a house near the sea, a convertible, a small catamaran, classical music recordings, and mince or apple pie. Tukey traveled with his personal table tennis paddle; collected some 14,000 mystery, sci-fi, and adventure paperbacks; lunched on fistfuls of cheese and six glasses of skim milk; and drove a 1936 wood-paneled station wagon until the passenger door fell off and his papers flew onto Nassau Street in Princeton. For 40 years he wore the same style of black polo shirt, so wrinkled that students sometimes mistook him for a janitor. But he always seemed able to squeeze in one more project, provided it was sufficiently intriguing.

  So how, given Tukey’s eminence and his time commitments, could NBC convince him that the Huntley–Brinkley news program warranted his attention? First, the reputation of opinion surveys, the mainstay of social science, was abysmal. Although sampling forms the foundation of statistics, commercial pollsters were painfully slow to adopt probabilistic random sampling. Serving on the Kinsey Report study committee with Mosteller, Tukey said he would prefer a random sample of three to a Kinsey sample of 300; Kinsey’s wife said she wanted to poison him. If Tukey aimed to improve statistical practices in the polling industry, NBC News was a high-profile place to start.

  Second, NBC’s RCA computers may have been a draw. If Tukey accepted NBC’s offer, he would not need an army of students to snip pieces of adding machine paper. RCA was a major military contractor as well as a giant in communications; it manufactured highly regarded mainframe computers for the military and big business. During the 1940s the company’s large research laboratory had designed and built the Selectron memory tube for early computers, including von Neumann’s Johnniac.

  The opportunity to use RCA’s computers to analyze election data must have been tempting. Tukey had foreseen the intimate connection between computers and statistics years earlier. When von Neumann designed an electronic computer for the Institute for Advanced Studies at Princeton in late 1945, Tukey was the only Princeton University representative on the committee and helped design the computer’s architecture and electronic adding circuit. Still, Tukey’s “most striking relationship with the computer was that he didn’t use it”; his hardware consisted of pe
ncil and paper.9

  Polling reform and powerful computers, however, may have paled in importance next to the allure of NBC’s vast amounts of voting data. As an undergraduate at Brown University, Tukey had majored in chemistry with doses of physics and geology, and his Ph.D. from Princeton, earned in 1939, was in topology, among the purest branches of abstract mathematics. Military research during the Second World War turned him into a “data analyst” committed to fighting the “mental rigidity” and “ossifications” of pure mathematics and abstract statistics and to bridging the gap between mathematics and science.10 The war moved him far beyond the statistician’s early role as a passive observer.

  After the war Tukey decided he wanted to drive “the rocky road of real problems in preference to the smooth road of unreal assumptions, arbitrary criteria, and abstract results without real attachments.”11 To do so, he accepted joint positions a half hour apart at Princeton University and Bell Labs. Later, whenever he was offered professorships at other universities he would ask, “Where could I ever find another Bell Labs?”12 Like Mosteller, he preferred exploring reality, and NBC News had plenty of that.

  But of all the enticements NBC could offer—restoring polling’s reputation, fast computers, and real data—the most important must have been the thrill of the chase. To beat other networks to the draw, he would have to work at top speed under international scrutiny to make sense of vast amounts of incomplete, uncertain information. It would be, as he put it later, “the best education in real-time statistics that anybody could have.”13 So the military consultant to presidents joined NBC’s Huntley–Brinkley Report.

  Tukey’s first evening on the job, November 8, 1960, started smoothly. The race between Kennedy and Nixon was the tightest since 1916, and Kennedy would win by 120,000 votes out of 70 million cast. By 2:30 a.m., though, Tukey and his colleagues were ready to call him the winner. The pressure was too much for NBC. Network executives hustled the statisticians into a room without telephones, locked them in and refused to let them out until 8 a.m. Tukey and his team twiddled their thumbs all night, unable to release their results until morning, when it was clear Kennedy had won. Still, Tukey had prevented NBC from mistakenly declaring Nixon the winner. Relieved and impressed, the network asked him to assemble a team for the congressional election in 1962. He would work for NBC News for 18 years.

  Tukey’s handpicked group eventually included ENIAC coinventor John Mauchly; Cornfield of NIH; Richard F. Link, Tukey’s first graduate student and pollster Louis Harris’s chief statistician; Yale psychology professor Robert Abelson; and David Brillinger, later a professor of statistics at Berkeley. When David Wallace finished Mosteller’s analysis of The Federalist papers, he too joined the group.

  Wallace arrived expecting a vacation from Bayes’ rule because Tukey was thought to look down on it. Tukey is not known to have ever published anything using Bayes’ rule, and in an often-quoted remark, he said, “There are many classes of problems where Bayesian analyses are reasonable, mainly classes with which I have little acquaintance.”14 Among those were business decision making, the bailiwick of Howard Raiffa and Robert Schlaifer. The lack of methodology for quantifying Bayes’ initial prior especially irritated Tukey. Publicly, he was a data analyst who was anti-Bayesian and even antiprobability.

  Thus when Wallace joined Tukey’s NBC team in 1964 he was surprised to find Bayes’ rule securely ensconced in the computing program: “I immediately thought, this is all very Bayesian. Also, I did a lot of the coding for a lot of the models over the next decade and a half, and, as far as I’m concerned, I was using Bayesian things.”15 Those who agree include Brillinger, who later became Tukey’s biographer and the editor of his papers, Pratt from Harvard, and Fienberg from Carnegie Mellon. Said Fienberg, NBC polling used “a form of empirical Bayes, where the past results were used to construct the prior distribution.”16

  Nevertheless, in almost two decades of election forecasting Tukey never admitted to using Bayes’ rule. Why would someone who publicly disdained Bayes’ rule and seemed to look down on it use it for something as important as announcing the next president of the United States?

  Many colleagues stress that, despite appearances, Tukey was “a very private man.” His nephew called him an “elliptical and enigmatic Delphic Oracle in a black polo shirt.” Wallace agreed: “Tukey could be close-mouthed. . . . He was a man of extreme power and brilliance and in some ways enigmatic about himself. . . . He didn’t always let everyone know what his left hand was doing. He’d deny anything Bayesian in the NBC polling.”17

  His dominating personality could be intimidating. While George Box was giving a seminar at Princeton University, Tukey thought he knew what Box was going to say and kept chiming in with his own commentary. Box finally asked for a show of hands. Who wanted Tukey to continue interrupting, and who wanted him to stop? When Box won, Tukey looked surprised. “In some ways, he was a very clever eight-year-old,” Box recalled. “He didn’t seem to understand very much about interpersonal relations.” Some colleagues pointed to his early years as a child prodigy home-schooled by his mother and said that his wife, Elizabeth, helped “warm him up.” Edgar Gilbert from Bell Labs concluded, “He was a very likeable personality but he was hard to understand.” Peter McCullagh, an Irish statistician at the University of Chicago, called him a “constructive scientific anarchist, . . . a cultural phenomenon, revered by some, feared by others, understood by few.” Part of the problem, Pratt said, was that “Tukey could argue on both sides of anything, and you didn’t know where he stood.”18

  Adding to the confusion was Tukey’s acceptance of one of Savage’s most controversial tenets: subjectivity. Tukey called objectivity “an heirloom” and “a fallacy. . . . Economists are not expected to give identical advice in congressional committees. Engineers are not expected to design identical bridges—or aircraft. Why should statisticians be expected to reach identical results from examinations of the same set of data?”19

  If Tukey could be hard on Bayes’ rule, he was even tougher on Fisher. Tukey believed that Fisher’s frequency-based ideas dated from “the world of infancy . . . the childhood of experimental statistics, a childhood spent in the school of agronomy. . . . Almost invariably, when closely inspected, data are found to violate [the] standard assumptions” required by frequency. “Far better an approximate answer to the right question, which is often vague, than an exact answer to the wrong question, which can always be made precise.” Tukey publicized how even slight deviations from the normal model could muddle the methods of Fisher, Neyman, and Egon Pearson. He particularly scorned frequentist “techniques for assessing significance and asserting confidence. . . . By and large, the great innovations in [frequency-based] statistics have not had correspondingly great effects upon data analysis.” Hard words indeed.20

  So where did Tukey stand? Anti-Bayes and anti-frequentism? Friends contend that, like Mosteller, he opposed any monolithic philosophy. Brillinger thought Tukey was annoyed, “not with Bayesian arguments per se; . . . [but] with some of the Bayesians.” Tukey said, “Discarding Bayesian techniques would be a real mistake; trying to use them everywhere, however, would in my judgment, be a considerably greater mistake.” The issue was knowing when and where. He often complained about “a natural, but dangerous desire for a unified approach,” explaining that “the greatest danger I see from Bayesian analysis stems from the belief that everything that is important can be stuffed into a single quantitative framework.”21

  Tukey used almost the same language about Fisher’s fiducial alternative to Bayes. While courting his wife, Tukey confided that his mission in life was to emulate Fisher by developing methods for analyzing experimental science. But after writing 64 pages in a search for the logical foundations of Fisher’s fiducial probability, Tukey decided that “the belief in a unified structure for inference is a dangerous form of hubris.” When Tukey visited Fisher at his home in England and began asking questions about his methods, Fisher stalked angrily a
way, leaving the Tukeys to find their way out of his house alone. In another version of the story, Fisher threw Tukey out of his office after telling the young man that his paper was a “long screed” and that he would understand probability statements only “if you could ever get your bullheaded mind to stop and think.” In both stories, an unstoppable force met an immovable object.22

  For Tukey the only thing that mattered was the data—shorn of computerization, mathematization, probability, and theory. He named his approach exploratory data analysis (EDA). Like Bayesians, many of its proponents were ridiculed and had trouble finding jobs.

  So how did Tukey resolve his paradoxical use of Bayes’ rule without admitting it? He called it something else. While Brillinger and Wallace called their NBC polling Bayesian, Tukey said it was “borrowing strength.”23

  “Anything he did, he’d call something else,” Wallace said, even if it already had a straightforward, well-established name. New names drew attention to ideas, and a colleague counted 50 terms coined by Tukey. Among those that stuck are linear programming, ANOVA, and data analysis. In one article, Mosteller had difficulty talking him out of using musical notation—sharps, flats, and naturals. Another colleague threatened to call him J. W. Cutie for terms such as “saphe cracking,” “quefrency,” and “alanysis.” As Wallace said, “It was not always the best way to win friends and influence people. . . . But when I talked to Tukey, I essentially tried to use his terminology.”

  Still, Bayes’ rule by any other name is Bayes’ rule. And both Tukey and Mosteller were willing to use whatever statistical tool was needed, even if it was Bayesian. Beginning work long before Election Day, Wallace built a base of initial information by combining data from preelection polls; nonstatistical, expert opinion from political scientists; and the voting histories of precincts, counties, cities, and states. Preelection opinion polls did not always ask the right questions, so they often failed to elicit all the information needed. The work of sampling people, surveying them, analyzing their answers, and summarizing the results was complex.