R.A. Fisher smoking a pipe, 1956. 

Contexts in source publication

Context 1

... crucial experiment Joseph Berkson and R.A. Fisher have become infamous for their stubborn opposition to the emerging evidence about the dangers of cigarettes (Fig. 1). Both argued that biological knowledge was essential and statistical methods limited, but their aim was ultimately to defend a particular kind of statistical expertise -that of the biostatistician who designed and interpreted randomized controlled experiments. Berkson reinforced the central role of biological knowledge in guiding ...

Context 2

... have struggled to understand the scientific and personal factors that drove prominent biostatisticians Ronald Aylmer Fisher and Joseph Berkson to the ‘wrong’ side of the debate over smoking and lung cancer in the 1950s. They both maintained a strong skeptical stance towards the mounting evidence far beyond, some suggest, the point of reason. Viewing the debate retrospectively, it is all too easy to divide the participants into those who were right and those who were wrong, or those who were against cigarettes and those who were for them. But, of course, that is too simplistic. While Berkson and Fisher stand out in the debate because of their professional stature and their strong rhetoric, a substantial number of biostatisticians and epidemiologists shared their concerns. For these scientists, it was not merely tobacco that was at stake, but the role of biostatistics and epidemiology as scientific disciplines. The debate over the epidemiological data linking smoking and lung cancer during the 1950s provides an excellent case study to understand the tensions in methodology that were emerging at this time. The epidemiological tools used in the early cigarette studies were relatively new and perhaps it is unsurprising that they were misunderstood and viewed suspiciously by pathologists, physicians and others lacking statistical training [1,2]. But the suspicions of statisticians and epidemiologists themselves cannot be put down to the ignorance of the na ̈ve laboratory researcher. ‘Now I should be the last person to attack evidence for being merely statistical,’ Fisher acknowledged [3]. The epidemiologists and biostatisticians who participated in this debate differed on what were sometimes subtle points of methodology, but these difference had substantial implications for characterizing the potential hazards of cigarette smoking. All of the participants saw the need for explicit and rigorous standards for evaluating etiological hypotheses, but they held conflicting views about what those standards should be. The differing opinions on the evidence reflect two different models of research – controlled experiment as the crucial, objective test of a causal hypothesis versus inferential judgment based on a diverse body of evidence. The first half of the 20th century saw the development of two distinct but parallel approaches to etiological research – the randomized controlled experiment and the analytic epidemiological survey. The first approach came out of biostatistics, developed and vigorously promoted by Fisher, whose Statistical Methods for Research Workers [4] and The Design of Experiments [5] became the standard textbooks. Followers of Fisher, including Major Greenwood and Austin Bradford Hill, took these tools into the field of medical research and promoted what they called ‘experimental epidemiology’, which initially meant testing epidemiological hypotheses in animal experiments [6]. The experimental approach (albeit without true randomization) was also taken up by public-health researchers for evaluating the efficacy of new vaccines. At the same time, some influential public-health investigators emphasized the role of the epidemiologist as one of assembling a diverse body of facts from different sources into a coherent explanation. American epidemiologist Wade Hampton Frost wrote that ‘epidemiology is something more than the total of its established facts. It includes their orderly arrangement into chains of inference which extend more or less beyond the bonds of direct observation.’ [7]. Instead of emphasizing a particular experimental design or statistical method, Frost stressed the importance of analytical skills and knowledge of a wide range of facts. These two approaches – experimental and inferential – are not mutually exclusive, but they do represent two distinct trends, and this subtle divide was crucial in the debate over cigarettes and lung cancer. This debate began in 1950, with the publication of five case-control studies that compared the smoking habits of lung cancer hospital patients with non-lung cancer patients [8– 12]. Case-control studies were especially useful at the early stages of the investigation, when several potential causes were being investigated. Yet the method clearly did not meet Fisher’s requirements for experimental design. A group of about a dozen biostatisticians in the Biometry Branch of the National Cancer Institute (NCI) played a leading role in advancing new methods for the analysis of case-control data and in building the case against cigarettes [13,14]. Large cohort studies, which compared cancer rates in smokers and nonsmokers over time, also provided crucial epidemiological support, although some biostatisticians also questioned these more rigorous studies. Joseph Berkson and R.A. Fisher have become infamous for their stubborn opposition to the emerging evidence about the dangers of cigarettes (Fig. 1). Both argued that biological knowledge was essential and statistical methods limited, but their aim was ultimately to defend a particular kind of statistical expertise – that of the biostatistician who designed and interpreted randomized controlled experiments. Berkson reinforced the central role of biological knowledge in guiding research: ‘if biologists permit statisticians to become arbiters of biological questions, scientific disaster is inevitable.’ However, knowledge of biological mechanisms was important not as an end in itself, but primarily because it suggested experiments. ‘The most important consideration with respect to a theory is not whether it appears plausible, but whether it suggests experiments, and what experiments are suggested.’ [15]. Fisher clung to the possibility that lung cancer and the smoking habit may both be produced by some common genetic trait. The statistical association between smoking and cancer, he insisted, did not imply that one must cause the other. Only a randomized controlled experiment could rule out the common cause hypothesis, he argued [3]. But Fisher was not alone in taking this hypothesis seriously. In fact, even Charles Cameron, president of the American Cancer Society, proposed that hormonal differences might explain both smoking habits and cancer susceptibility [16]. Moreover, studies conducted during the 1950s provided evidence that smokers did in fact differ in several ways from nonsmokers, including personality, hospitalization rates, occupation, diet and physical characteristics [17,18]. Two of these studies were conducted by epidemiologists, one by Clark Heath, with tobacco industry funding, and one by Abraham Lilienfeld, then at Roswell Park Memorial Institute. But this debate rested on more than a pet hypothesis. While Berkson and Fisher were especially outspoken and uncompromising in their views, their concerns were shared by a substantial number of senior biostatisticians at the time, including Donald Mainland at New York University, Antonio Ciocco at the University of Pittsburgh, and K.A. Brownlee at the University of Chicago. This debate was going on while the randomized controlled trial was a recent development in medical research and its place was still in doubt. Biostatisticians and clinical trialists were trying to persuade the medical establishment that their expertise was vital to research and that their scientific methods were comparable in validity to laboratory experiments [19,20]. Simple procedures like the test of statistical significance had worked their way into medical research, but statisticians objected that they were often naively applied by those without statistical training. Fisher himself probably contributed to this tendency by strongly emphasizing tests of statistical significance in his textbooks and showing an interest only in the interpret- ation of the results of single trials and not from a diverse body of evidence. In short, it was not just tobacco that was at stake in the debate, but the authority of the biostatistician in medical science. Jacob Yerushalmy, a biostatistician at Berkeley, later took a similar critical stand in response to studies showing a link between maternal smoking and infant mortality. Here he argued that the data were equally consistent with competing interpretations, that either the smoking or the constitution of smokers could explain an association between maternal smoking and low birth weight. Non- randomized studies simply were helpless to resolve the conflict. Yerushalmy made it clear that the problem was not with the statistical form of the results. ‘The evidence may not be convincing, but not because it is ‘only statistical’, rather because the evidence is nonstatistical in the sense that the method of study which produced the evidence violates the basic principles for valid statistical inference.’ [21]. Statisticians also argued that smokers and nonsmokers in epidemiological follow-up studies might differ in ways other than in their smoking habits, because smokers (or nonsmokers) might be more or less inclined to participate in a voluntary study, thus introducing selection bias [22]. Experienced public-health research methodologists were likely aware of earlier discussions over the problems of using volunteers in vaccine studies. Additionally, some statisticians criticized E. Cuyler Hammond and Daniel Horn’s use of 22 000 American Cancer Society volunteers to recruit and interview subjects for their smoking study. Mainland argued that researchers, particularly untrained volunteers, were likely to select a particular type of individual for the study; he conducted a survey of his students to demonstrate that they did indeed suffer from this procedural bias [23]. Most epidemiologists and statisticians, including those who defended the link between cigarettes and lung cancer, admitted that a controlled experiment, when it was available, offered the strongest evidence that any ...