When Computers Were Human
Page 21
In January 1921, the United States inaugurated a new president, Warren G. Harding, and the Department of Agriculture welcomed a new secretary, Harry C. Wallace. One of Wallace’s first acts was to create a new research office called the Bureau of Agricultural Economics. This office collected together all the employees of the department, including Howard Tolley, who were engaged in studying problems of production, markets, and financing. Within this group, Tolley promoted the method of least squares as a means of analyzing agricultural data. This application of least squares was substantially different from its use in survey adjustment, even though the method of calculation was unchanged. In agricultural studies, least squares did not adjust data. Instead, it took data apart in order to identify underlying causes or forces. It could identify the effect of fertilizer on crops or the feed that increased the weight of farm animals. It was sometimes called “regression analysis” or “the analysis of variance.”
In the fall of 1922, Tolley gave a series of lectures in order to introduce the staff of the Bureau of Agricultural Economics to the ideas of statistical least squares.9 Tolley illustrated the theory with an example that concerned the damage done to cotton crops by the boll weevil. His data, gathered in the southern states by agents of the department, contained the extent of damage on each field, the typical size of the cotton plants, the time of year, the amount of rain recorded in the area, and the daytime temperature. Using least squares analysis, Tolley showed how it was possible to determine which of these factors were present in the most heavily damaged fields. According to his results, the plants were most vulnerable at a certain stage of their development.10
Though the method of least squares promised much to the agricultural researchers, it was of little use unless the department could provide a computing office. The calculations were every bit as daunting as the least squares calculations of survey adjustment. A survey calculation would be done only once. A least squares analysis of a specific problem might have to be repeated every season. Tolley believed that at least some of the calculations might be handled with punched card equipment. In agricultural statistics, as in survey adjustment, least squares computations had two distinctly different parts. The first part reduced the data to a series of normal equations. By itself, this activity was especially demanding for agricultural researchers, as they were often dealing with large collections of data that spanned counties or states or regions. The punched card equipment of 1922 could summarize data for many applications, but it needed to be used in a special way for least squares calculations. The normal equations required multiplications, and punched card tabulators could only compute sums. There was a simple way to force the machine to perform multiplications, but it required a skilled and attentive operator. The method, called progressive digiting, reduced multiplication to primitive additions. A simple product, such as 24 × 127, would require six cards. Four cards would have the number 127 punched on them. The other two would have the number 1270. To compute the product, the tabulator would sum the six cards.
When applied to real problems, progressive digiting seemed to be an awkward process, a complicated operation that should have been straightforward. It was something akin to counting the number of sheep in a field by summing the number of legs, adding the number of ears, and dividing the result by six. To handle real problems, operators had to punch multiple cards, sort them, sum them in a tabulator, shift the values, and sum again. It was difficult work, but it was faster and more accurate than the alternative of doing the computations by hand. For large collections of data, those that had been gathered from one thousand or two thousand farms, the punched card equipment provided the only practical way of preparing the normal equations.
Punched card technology offered no help with the second step of least squares analysis, the step of processing the normal equations. The only way to do it was to give the numbers from the tabulating equipment to a staff of human computers and let them complete the work. They would use a mechanical calculating machine and the mathematical method invented by Myrrick Doolittle some forty years before.11 Even with Doolittle’s method, this part could be time-consuming. Tolley advised researchers to minimize the labor by doing all calculations with only two digits after the decimal point. He defended this procedure by noting that agricultural research was imprecise and that “astronomical accuracy is really not necessary.”12
After teaching his class on least squares analysis, Howard Tolley moved to create a central computing laboratory for the Bureau of Agricultural Economics, an office that had punched card equipment and the expertise to perform least squares calculations.13 In 1922, three separate offices of the bureau had punched card equipment, but none of them was fully utilizing its equipment. “The installation of the Cost of Marketing Division is busy most of the time,” he reported, but “that in the Division of Land Economics about half the time, and that in the Division of Statistical and Historical Research something like one-third of the time.” His research suggested that none of these computing offices really understood how to prepare a problem for machine tabulation and that at least one office had started problems that it had been unable to complete.14
25. Computing room at the U.S. Department of Agriculture
In the winter of 1923, Tolley received the approval of the department’s senior statisticians to create his centralized office. Even with this approval, the task was challenging, as none of three offices was prepared to surrender its punched card equipment. The possession of an expensive piece of machinery was a sign of power and importance, even if the machine was not well used. Tolley worked with each of the offices, arguing that a centralized tabulating facility could “rearrange schedules and adjust them to cards in such a way that both the cost of tabulation and the elapsed time between the collection of data and the completion of tabulation would be materially reduced.” By April, he had convinced all three offices that they would benefit from a central computing division.15
For the new computing laboratory, Tolley had to choose between competing brands of equipment. Two of the offices leased their equipment from the CTR company, which was in the process of changing its name to International Business Machines. The third office used the tabulators of the Powers Accounting Machine Company. Powers had been founded by a former census employee. His machines could tabulate the International Business Machines cards, but they operated in a slightly different manner.16 Tolley chose to keep a unified shop and leased only International Business Machines equipment. He completed the work by the first of May and announced that the “tabulating and computing services will be available to all” engaged in economics research.17
When the office began operations, it was able to handle only the first part of the least squares computations, even though it employed thirty workers. The staff of this office were called “operatives,” not computers. For most calculations, they punched cards and ran tabulators, following instructions that had been prepared by researchers in the Department of Agriculture.18 The office was led by a statistician who provided advice to the various divisions of the department. He showed others how to organize their data and prepare it for punching, but he did nothing with advanced mathematics. The second step of least squares calculations was left to the bureau researchers and their assistants.19
A second computing laboratory was created with the assistance of the younger Wallace, Henry A., at Iowa State College in Ames. Henry A. Wallace did not move to Washington with his father but remained in Iowa to edit the family newspaper. From this position, he continued his statistical research. With the help of friends and subscribers, he would gather data from all over the state of Iowa, punch it onto cards, and summarize it with the tabulators of a Des Moines insurance company.20 He used the results to recommend farming strategies to his readers. During the early 1920s, he urged Iowa farmers to change their mix of crops, a campaign he summarized as “More Clover, Less Corn, More Money.” Though he generally used only simple numbers to support his ideas, Wallace could not res
ist the temptation to demonstrate his mathematical sophistication. By 1922, he was conducting least squares analyses of agricultural data and publishing the results in Wallace’s Practical Farmer. At the end of one article, he proclaimed, “For the benefit of our statistical friends, we may say that our predicting formula has a multiple correlation coefficient of .91, which indicates a very real degree of accuracy.”21 Of course, few of his statistical friends read Iowa farming papers, and few of the hog farmers understood the mathematics of multiple correlation.
With his father running the Department of Agriculture, Henry A. Wallace had the opportunity to meet Howard Tolley and followed the development of the new computing facility at the department.22 In the winter of 1923, Wallace followed Tolley’s example and taught his own class on the subject of least squares. His classroom was at his alma mater, Iowa State College. The students were college researchers, senior faculty, and a graduate student or two. Some already knew the basics of least squares. One had served on the Food Administration as one of the “expert swine men.”23 In ten Saturday sessions, Wallace developed the fundamental theory of least squares, demonstrated how to prepare the basic equations, and showed how Doolittle’s method could be used to compute the final answers. For most of the classes, he had a desk calculator on hand, a “cheap key-driven machine,” he later recalled. For the last class, he decided to show how to compute correlations with a punched card tabulator. He borrowed a punched card tabulator in Des Moines, loaded it into the back of a truck from the family farm, and drove north to the Iowa State campus. He spent the morning explaining how the machine worked and demonstrating statistical calculations with it.24
Assisting Wallace in the classroom was an Iowa State College mathematics professor named George Snedecor (1881–1974). Snedecor held a master’s degree in physics and had been hired in 1913 as part of an effort to improve the university’s science and engineering departments. After his arrival at the campus, he had quickly recognized that the school had a greater need for a statistician. He had no connection to Karl Pearson, who was recognized as the founder of mathematical statistics, but he had studied some statistical methodology in college and was able to teach courses in elementary methods. He spent much of his time compiling the methods of statistical practice in a form that could be used by researchers in any discipline, though his examples tended to favor the agricultural research that dominated the campus.25
In 1924, Snedecor and Wallace wrote a pamphlet entitled Correlation and Machine Calculation, a paper that would have fit nicely into Karl Pearson’s Tracts for Computers. It dealt with “practical difficulties” of using least squares techniques for computing correlation problems. The pamphlet showed how Pearson’s correlation analysis was actually a form of least squares work and then described how Doolittle’s method could be used to do the calculations. The two authors made only a passing reference to punched card technology, but they noted that “where the number of observations runs into the thousands, punched cards should be used with sorting and tabulating machines.” The “machine calculation” promised in the title was the calculation done by a human computer with a “commercial form of adding machines.”26
While he worked on the pamphlet, Snedecor experimented with punched card calculation. He leased a card punch, which he kept in his university office. This was a small device, little bigger than a standard dictionary, and consisted of a wooden base, a frame for the card, and a sliding pad with keys numbered from 0 to 9. He punched each number one digit at a time, advancing the slide as he progressed. Any mistake would ruin the card so that he would have to start afresh. Once he had punched a set of cards, he would send them to Des Moines to have them tabulated.27 As he gained skill with punched card machinery and began to appreciate what he could do with it, he began to drift away from Wallace. There was no conflict or disagreement between the two, just diverging interests. Snedecor was devoted to statistical research and to the support of scientific study. Wallace was interested in many things and was devoting more time to political causes and new business ventures. His political goals were embodied in a congressional bill that would open new markets for farmers. His new business was a company that produced and marketed hybrid corn seed.28
Snedecor’s experiments with punched card machinery slowly grew into an organized computing laboratory. He did not have the kind of resources that could be found at the Department of Agriculture in Washington but he was able to hire a human computer in 1925 and acquired a punched card tabulator two years later. When the machines arrived, he discovered that his office did not have the proper electrical wiring for the tabulator, a minor setback that forced him to find a room in another building for his laboratory.29 Once he had the equipment operating, Snedecor was temporarily overcome with the fascination of new love, the kind of fascination that researchers before and since have bestowed upon new machines. Snedecor demonstrated punched card techniques to any interested researcher and used the tabulators to solve dozens of problems, including many that had no relationship to statistics and a few that were not particularly suited to punched cards. He summarized field trials, factored numbers, built databases, and solved differential equations.30 His dean grew tired of his lengthy descriptions of punched card computations and suggested that, in the future, Snedecor might simply summarize his accomplishments.31
Snedecor named his laboratory the “Mathematical and Statistical Service,” a title that defined the role of the office within the college and suggested that Snedecor’s machines might be leased by outside clients. It was not the only university computing office in the country. The International Business Machines company was attempting to place its products at colleges and universities. In 1927, the year that Iowa State College acquired its machines, International Business Machines leased tabulators to at least four other schools: Cornell University, Columbia University, the University of Michigan, and the University of Tennessee.32 Unlike the other sites, the Iowa State facility combined tabulating machines and human computers. The machines summarized data and computed the basic values for least squares problems, just like the machines at the Department of Agriculture in Washington. The computers completed the work, solving the problems by using Snedecor’s and Wallace’s variation of Doolittle’s method.
26. Iowa State Statistical Computing Service
One computer, Mary Clem (1905–1979), rose to play a central role in Snedecor’s laboratory. Clem came from western Iowa and had only a high school education. Mathematics, she later claimed, “was my poorest subject,” a boring topic that never interested her. By contrast, she thought that computing was something entirely different. “The more I worked with Snedecor,” she said, “the more fascinated I became with figures and data. I got to the point where that was my whole life.” She became especially good at identifying patterns that could be used for detecting errors, values that she called “zero checks.” A zero check was a sum that should equal zero if all the numbers had been correctly calculated. Such sums often escaped the eyes of those with more mathematical training. The first time that she presented one of these checks to Snedecor, he was skeptical. Clem recalled that “he sat there thinking about it a little bit, and he turned around and went through algebraic fiddling around and he said ‘that’s right they do.’ ”33
Because of her insight into calculation, Clem became the planner for the laboratory, a worker occupying a middle level between the human computers and the research scientists. She was remembered as a strong presence in the lab, a leader who set the schedule, trained new computers, and enforced discipline.34 She oversaw about six computers and one machine operator. Many of these computers were graduate students or the spouses of graduate students. A few were local high school graduates like herself. The majority were women. Clem felt that men never adapted to the work, judging that “they would rather probably teach or do research work.” Male or female, most of her computers were transients, working for only “two or three years” before they left the lab.35
Eve
n with the combination of card tabulators and human computers, Clem and Snedecor found that there were problems that were too demanding or too expensive for the lab to undertake. One such problem was presented to the lab by Henry A. Wallace. In 1930, Henry A. Wallace conceived the idea of looking at long-term weather records and comparing them to astronomical ephemerides. He reasoned that the weather was a complex system driven by outside forces. He acknowledged that the radiation of the sun and the gravity of the moon were the largest of these forces, but he speculated that the planets might also exert a physical force upon the weather. Undeterred by what others might think of his plan, Wallace began to develop a framework in which he might search for a relation between the positions of the major planets and the weather on Earth.