Chapter 13
BEYOND ONE-CASE STATISTICS: MATHEMATICS, MEDICINE, AND THE MANAGEMENT OF HEALTH AND DISEASE IN THE POSTWAR ERA
Jean-Paul Gaudillère
In 1961 Daniel Schwartz, a statistician studying the relationship between tobacco and cancer for the INH—the French medical research agency—wrote a popular paper on ‘The Statistical Method in Medicine’ published in a journal widely circulated within the medical profession. Schwartz scorned French physicians who thought that statistics had little to do with medical knowledge:
The reasons explaining this backwardness are diverse. France is a leading country in pure mathematics. As a consequence applied mathematics have been neglected. The statistical method emerged in the ‘Anglo-Saxon’ countries and slowly penetrated our country.
Our peculiar frame of mind as well as our education system favor the mathematics of certainty and not the mathematics of probability. The latter prevails in all aspects of life and in decision-making processes but they are never taught…
On top of cognitive patterns, cultural traits may well have played a part in the disdain for statistics. French people are highly individualistic … This behavior dominates the medical world. The French doctor is a very good clinician and an outstanding care provider. In his perception, every patient is an individual rather than a figure in a table … These values are highly recommended in medical practice but the price paid for their supremacy is a farfetched emphasis on the individual in scientific research.
This standing is not inescapable. It is true that respect is due to the special relationship between a patient and his doctor. But it is not true that this respect should result in one-case statistics. [Schwratz, 1962, p. 1920]
In his essay Schwartz offered a multi-layered interpretation of the role of the statistician in medicine. His analysis of the problematic status of ‘big numbers’ in French medical circles highlighted the ambiguities of the expertise claimed by pioneers of medical statistics in the aftermath of World War II. Being neither clinicians nor biologists many statisticians sought the role of mathematically-trained mediators. They argued that medical judgment should be based on objective procedures, i.e. aggregated data, comparative studies, controlled samples, and signification tests. Professionals ‘applying’ mathematics or translating probabilities into rules of evaluation and algorithms were barely needed. Hence the notion highlighted by Schwartz that physicians— especially French physicians trained in the ‘great clinical tradition'— quite naturally opposed the intervention of outsiders and any radical departure from accepted individual practices.
The relationship between mathematics, statistics, and medicine is hardly an uncharted territory in the history of science. Early attempts at gathering clinical cases and computing figures for medical decision making are usually traced back to the invention of the ‘numerical method’ by Pierre Louis in the early nineteenth century [Desrosières, 1993]; [Bynum, 1994]; [Matthews, 1995]. Louis’ story also is one of medical resistance. His failure to promote the numerical method has been explained with the very same factors as those listed by Schwartz: clinical tradition, one patient-one doctor relationship, organization of the medical profession, or absence of public health bodies.
Because of this failure, the main roots of contemporary medical statistics usually are ascribed to the British ‘biometric school’ of the first half of the twentieth century. The careers of Galton, Pearson, Fisher, or Major Greenwood have been the topic of several studies [Gigerenzer, 1989]; [Mackenzie, 1981]; [Matthews, 1995]; [Porter, 1986]. One shared feature of these analyses is to locate innovation—whatever its motives, i.e. cognitive, cultural, disciplinary, professional, ideological— in the science of numbers. This perception inevitably results in viewing medicine as a field for diffusion and a site of application. Statistical tools like the regression technique invented by Galton as part of his studies of human inheritance are often viewed as tools which emerged OUT OF a biological terrain to live an independent life and to occupy all sorts of scientific niches [Porter, 1986]; [MacKenzie, 1981]; [Matthews, 1995]. Explanations for this dissemination process may be intellectual—a new way of seeing the world circulated with new representations and computational techniques or social, the profession of statistician emerged in the midst of institutional conquests. In both cases however inventors are described as mathematicians having little or no interaction with specialists working in the domain of applications.
This chapter will build on more recent analyses of the part played by statistics in public life [Porter, 1995]; [Schweber, 1996] in order to propose a more balanced account focusing on the invention of medical statistics as a process of interaction between actors involved in health and disease management and building computational tools for decision making. This perspective will be illustrated with two episodes in the history of medical statistics after the Second World War when an apparent ‘trust in numbers’ started to dominate entire segments of biomedicine.
THE RANDOMIZED CLINICAL TRIAL: A DEVICE AT THE INTERSECTION OF STATISTICS AND MEDICINE
In 1948 the British Medical Journal published the results of an experiment on the therapeutic uses of a new antibiotic called streptomycin which seemed to be a potent means of curing lung tuberculosis [Medical Research Council, 1948]. This paper is often described as the ‘birth of the modern clinical trial’ because what was at stake was a new way of evaluating therapies and the statistical method as well as the efficiency of streptomycin (something many clinicians were already convinced of).1 Run by a special committee established by the British Medical Research Council, the 1947 streptomycin assay actually was a statistician's trial. It was designed by Austin Bradford Hill and tuberculosis specialists according to methodological choices which would later become the basic norms for standard objective assays of therapeutics [Meldrum, 1994]. The streptomycin trial was planned as a ‘randomized and controlled’ trial meaning that a) patients were randomly assigned to the assay group (receiving the antibiotic) or to the control group (receiving established treatments); b) when making decisions about admission to the trial, clinicians did not know this allocation; c) clinical evaluation of therapeutic progress was based on the measurement of tuberculosis spots on X-ray pictures of the patient's lungs; d) this examination was done by radiography specialists with no knowledge of the clinical status of the patient; e) final evaluation of the significance of these results was completed by statisticians.
The history of such randomized clinical trials (RCT) may be traced back to the biometricians’ work and it seems to be a good example of ‘applied statistics'. On the one hand, there was a direct lineage from Pearson to Bradford Hill via Fisher and Major Greenwood, the last being the first physician to study with biometricians and the first a statistician appointed by a medical research institution (the Lister Institute, London, where he participated in the evaluation of sera and vaccines; he later joined the National Institute for Medical Research at Hampstead) [Cox-Maximov, 1998]. On the other hand, it is not too difficult to argue for conceptual legacy, since the basic concepts grounding the choice of randomization can be traced back to R.A. Fisher's work.
Fisher's first job was at the Rothamsted Agricultural Research Station where he focused on problems of sampling and statistical significance [Fisher-Box, 1978]. He thus worked on the design of experiments [Fisher, 1935]. Fisher's main interest lay in the reliability of a given experimental result and the role of chance, for instance, in the result of trials comparing the yields of different varieties of cultivated plants. He argued that an observed difference in yield between two varieties planted in different fields might be due either to differences among seeds or to differences in soil, exposure, treatment, and other unknown factors. If so, how to decide on the results’ meaning? Randomization was an invention targeted at taming chance in experimentation without having to depend on the practitioner's experience of terrain and agricultural practices actually involved. Chance was the enemy but it was to be defeated with more chance. Fisher built on existing practices in agricultural stations and first proposed to standardize treatments. Randomization was an additional step that consisted in dividing the plots into narrow strips and assigning the place of different varieties to be tested by use of chance mechanisms. The rational was that randomization would replace replication of experiments or that the use of large numbers and chance allocation would reduce the effects of unknown variations in experimental conditions. Randomization reduced what Fisher called the objective bias.
Fisher's argument had a significant impact on agricultural research in the 1930s. In the medical world however, it was of limited influence, although some statisticians dealing with biological and clinical problems started to refer to random allocation.2 Austin Bradford Hill, a student of Greenwood at the MRC Statistical Committee was among them, and he referred to ‘alternating’ treatment assignments [Bradford Hill, 1937]. But it is only with the postwar antibiotic trials that he fully committed himself to the organization of randomized trials and chance allocation. Bradford Hill took over Fisher's procedure but developed a different rationale. Therapeutic evaluation in medicine is more complex than decision making in agriculture because two different biases are involved. The analog to Fisher's objective bias was the fact that the same disease in two different bodies could follow variable paths according to unknown differences in the terrain, i.e. body constitution, immunity, etc. But Bradford Hill and his followers focused on another— subjective—bias originating in the medical situation. Unlike agricultural researchers, physicians take care of people and they usually think that the treatment they prescribe works while patients usually trust their doctor. This conjunction of subjective hopes and trust does affect the selection of patients to be included in a study, the management of a treatment, and its final evaluation. As Bradford Hill put it in 1951 when summing up the motives for controlled clinical trials: ‘Drugs are not ordered by doctors at random, but in relation to a patient's condition when he first comes under observation and also to the subsequent progress of the disease. The two groups are therefore remotely comparable … The same objections must be made to the contrasting of volunteers for treatment with those who do not volunteer, or between those who accept and those who refuse … The contrast of one physician, or one hospital, using a particular form of treatment, with another physician, or hospital, not adopting that treatment, or adopting it to a lesser degree, is fraught with the same difficulty …’ [Bradford Hill, 1951, p. 279]. These were the most compelling reasons for introducing randomization in medicine: to regulate and undermine the clinician's preferences and idiosyncrasies. In the statistician's eyes, chance should be used to narrow down both objective and subjective sources of variability.
This plea against the traditional investigator should not be undermined. Randomization did not only oppose the clinician's role but also the very sense of a doctor's duty and experience. As recalled by many observers, clinical trials did exist before the Second World War and sometimes included control groups [Ross, 1936]. The allocation of patients however was based on other assumptions. Clinicians were aware of the variability of patients and diseases but sought to control uncertainty by accumulating experience and controlling trial parameters. The right choice of patients to participate in a study was to be based on the doctor's intimate knowledge of the individual and the intricacies of the disease. A good assay group would therefore maximize the factors enhancing a positive response to the drug, while every patient would be matched by another individual in the control group. The latter was consciously selected by the investigator so as to show strong analogies with the former, at least with respect to the clinically most important features of the disease. Chance was not to be mastered with more chance but conquered with experience and medical knowledge. This approach was also taken to be an ethically responsible one: treatment would not be denied those who would in all medical knowledge (and not in all probability) benefit from them. The statisticians’ ethos clearly was departing from these attitudes [Bradford Hill, 1963].
This foundation story has recently been refined by several historians of medicine. On the one hand, Desiree Cox-Maximov has stressed the pre-history of clinical trials and emphasized, for instance, the legacy of the Statistical Committee and the Therapeutic Trials Committee established in the 1930s by the British Medical Research Council [Cox- Maximov, 1998]. On the other hand, Harry Marks has discussed the work of American medical reformers and convincingly argued that the norms of the RCT were far from being universally accepted and applied in postwar medicine [Marks, 1997].
In-depth analyses however do not abolish the transformation typified by the reference to the MRC streptomycin trial. How is it then that methodological power shifted to the statisticians’ hands? Cognitive history has it that the postwar expansion of randomized clinical trials was triggered by the need for objective means of evaluating medical treatments and by the technical efficiency of the statisticians’ method [Himsworth, 1982]; [Lock, 1994]. One is nonetheless left with the question of knowing which factors created and made visible this ‘need’ for mathematically-certified objectivity and outside evaluation. In order to account for postwar changes, one must analyze controlled trials and their underlying notions as decision-making procedures rather than abstract methodological norms. In other words, one must look at the changes in medical practice during and after the Second World War. This leads one to stress other aspects of the MRC streptomycin trial.
In Hill's words: ‘It is the gradual development of this [scientific] attitude of mind coupled with the concurrent introduction of one antibiotic, one modern drug, after another, that has led in the past few years to the highly organized and efficiently controlled therapeutic trial of new remedies’ [Bradford Hill, 1951, p. 278]. In other words, it was of fundamental importance that the MRC trial took part in the testing of a chemical ‘wonder drug', an early product of the ‘antibiotic revolution'.3 Postwar clinical research was actually dominated by tightened relationships among biological laboratories, pharmaceutical firms, and hospital wards [Starr, 1982J. A critical change in the role of the elite physician and the medical systems of most industrialized countries was associated with the dramatic expansion of chemotherapy. Following the development of penicillin and early antibiotics, chemical laboratories of the pharmaceutical industry started to produce thousands of compounds annualy having some structural analogy with known anti-microbial agents and therefore of putative therapeutic value. The same hope of developing ‘magic chemical bullets’ soon pervaded numerous domains other than infectious diseases caused by bacteria, with cancer above all. Clinicians faced the difficult task of making decisions regarding the claims of industrial chemists and doctors, or regarding the choice of protocols for treating a given disease. This shift obviously contributed to the creation of new ‘demands’ for evaluation procedures. The production of antibiotics however does not in itself explain why these demands took the form of ‘highly organized', ‘large scale’ and ‘cooperative’ trials, and why these trials gave rise to an alliance between the medical elite and statisticians. Changes in medical organization rather than chemical breakthroughs have to be taken into account.
In his detailed analysis of American clinical reformers, Harry Marks points to the war-time medical mobilization as a turning point [Marks, 1997, chapter 4]. His perceptive description shows how the development of early antibiotics within the context of military medicine paved the way to new collaboration patterns, acculturated elite clinicians to the use of statistical tools, and contributed to sow the seeds of postwar changes. The war assays of penicillin and streptomycin are emblematic of this trend.
Although ‘discovered’ in the late 1920s in Fleming's laboratory at St. Mary's Hospital, London, penicillin did not become a therapeutic agent until the Second World War when its development was taken over by the US pharmaceutical industry and the Office for Scientific Research and Development (OSRD)—the federal agency established by President Roosevelt which coordinated the scientific war effort [Macfarlane, 1984]; [Hobby, 1985]. It is well known that the mobilization of American scientists for war was organized by the OSRD. From 1941 onwards its Committee for Medical Research (CMR) prompted studies on the mass production of penicillin and acted as a regulator and coordinator for a network composed of a few federal laboratories (for instance, the US Department of Agriculture Laboratory at Peoria), a dozen university teams, the War Production Board and large pharmaceutical firms (Merck and Pfizer in the leading roles). Initially CMR was also in charge of supplying the drug to civilian clinicians, and organized investigations of penicillin (and later of other antibiotics). The series of cooperative investigations launched within this context shows how organization drifted toward the randomized clinical trial.
The distribution of penicillin by CMR was handled by Chester Keefer, a Harvard professor of medicine. CMR first relied on classical ‘trials': the drug was distributed to a handful of well known clinicians who determined who would be treated and how. CMR's role was to draw a list of targeted infections and gather information about the outcomes. As production was scaled up, and as clinicians started to use massive doses of penicillin, spectacular cures of meningitis and staphylococcal infections were reported. Although there were some tensions regarding the list of authorized pathologies, CMR's monopoly on the drug, as well as its reported wonderful effects, forced compliance among participating clinicians who feared being refused access to the ‘magic bullet'. At this stage statistics played no, or a very limited, role with the presentation of ‘all or no’ figures.4
Organizational patterns started to change in 1943 as expansion of production prompted the Army, which envisioned massive uses of penicillin, to take over the drug. For the military, syphilis was a high-rank priority because of the incidence of the disease and the slow effects of arsenic treatments then in use. Syphilis was a complex and chronic disease with latent a-symptomatic phases whose diagnosis and follow-up were based on sophisticated bacteriological and immunological procedures. Systematic and organized comparison among cases was then considered necessary for evaluating the effects of penicillin on the disease. The first study organized by the military took advantage of the large number of enrolled men. Thousands of patients were given the drug and progress was monitored. Data collection however remained highly problematic, and the army officials ultimately requested the OSRD to organize a carefully controlled trial [Pillsbury, 1946]. Under CMR's guidance, syphilis investigators agreed to set up a cooperative trial based on one single treatment scheme, standardized laboratory tests, and homogenized data collection. Although the study was plagued with problems in assuring follow-ups, it contributed to make multicentric cooperative work and large numbers more visible and more normal.
As war ended and medicine returned to peacetime conditions, this style of operations was passed on to the evaluation of other antibiotics. In September 1945, when a team of researchers working at the Mayo Clinic in New York announced that streptomycin may benefit tuberculosis patients, the army and the pharmaceutical industry once again turned to governmental scientific bodies (this time the National Research Council) to resolve uncertainties about the drug and therapeutic priorities [Hinshwa, 1954]. The investigation was organized with the Veterans Administration which ran its own network of hospitals. Conducted within a bureaucratic organization by people who had all been involved or aware of penicillin studies, the VA streptomycin trial reinforced cooperative and planning ideals. Innovations focused on the need for a control group and for ‘objective’ outside analyses of X-ray pictures. The program could not be fulfilled (for instance, the assembling of a control group was dropped after a few months because investigators could not get Veterans Administration hospitals in the study to find enough patients) but the VA study clearly promoted the notion of cooperative AND controlled trials.5 It was followed by another study of streptomycin in tuberculosis launched by the US Public Health Service. This time tight control was enforced. All decisions to admit patients were reviewed by a panel of senior investigators; a central statistical unit assigned patients to treatment or control groups; patients in the control group were not told that they participated in a study while clinicians’ scruples were softened by introducing an appeal procedure for handling control patients whose disease worsened critically; finally evaluation followed the statisticians’ recommendation for X-ray measurements rather than having clinicians hold clinical case conferences.6
The growing importance of statistics and methodological rules in this series of trials reflects forms of objectification rooted in the advent of a cooperative regime in medical research which was unequivocally facilitated by mobilization for war and the context of military medicine. Its success and expansion after the war, however, is inseparable from the changing organization of medicine and particularly from the birth of a hospital-based high-tech medicine which transformed the nature and procedures for clinical decision-making.
Placed within this series, the MRC streptomycin trial thus appears as just one more instance of coordinated clinical work organized under the umbrella of a scientific regulating body. The acceptance of randomization and related notions, as well as the new role of ‘big number’ specialists, may be viewed accordingly as decision-making tools meaningful within the large medical networks emerging during and after the war. To follow liana Löwy's analysis of the case of cancer research where the randomized clinical trial became a widely used norm in the 1950s and 1960s, the RCT was an ‘organizational device’ useful in domains where the routine use and evaluation of fast-moving chemotherapeutic innovations became part of the normal caring of patients [Löwy, 1996].
RULES OF INFERENCE: TOBACCO, CANCER, AND EPIDEMIOLOGY IN THE PUBLIC ARENA
In 1964 the US Surgeon General released an official statement on the correlation between smoking and cancer [US Department of Health, 1964]. This document was seen as bringing to a close the decade-long controversy on the causes of lung cancer, which had been debated among epidemiologists, public health officials, cancer specialists, and the tobacco industry [Brandt, 1982]; [Berlivet, 1995]. This debate, which catalyzed the disciplining of epidemiology, has been viewed as another key process in the postwar history of medical statistics.
The discussion on smoking and lung cancer was initiated by the publication of a British survey of lung cancer patients conducted by Bradford Hill and a student of his; Richard Doll. This first study compared 1,500 lung cancer patients and a control population assembled by matching every cancer case with a patient of similar age, and if possible taken care of in the same hospital. The study focused on the significantly higher number of smokers among cancer patients than among the control population, but remained quite elusive on the nature of the association between smoking and lung cancer [Doll, 1950].
For reasons that have still to be properly investigated, studies multiplied in the following decade with an increasing emphasis on statistical arguments and an imputation of causality claimed by organizers of the largest surveys, beginning with Bradford Hill and the British Medical Research Council. Opponents, like statistician R.A. Fisher or geneticist C.C. Little, repeatedly argued that a causal inference was impossible to make, that the control population was biased, or, more critically, that smoking was just a marker for a population with a high incidence of another ‘third’ (causal) factor, for example genetic predisposition or socioeconomic status [Fisher 1958a and b]. In response to critics there was in an increasing sophistication of techniques. The size of populations grew. Prospective studies complemented retrospective analyses. Control populations were selected according to additional variables: age, sex, medical history. In addition, new statistical tools were invented. These concentrated on procedures for analyzing the significance of multiple associations and comparing multiple control groups [Schwartz, 1969]. For instance, fourfold table analysis [Berkson, 1946] was developed into complex strategies for analyzing retrospective data and computing relative risks by means of an extension of chi- square testing to situations in which data could be subclassified according to variable sets of factors [Mantel, 1959].
Growing mathematical sophistication did not in itself put an end to the controversy [Brandt, 1982]; [Berlivet, 1995]; [Proctor, 1995]. Public support for—and administrative endorsement of—the notion of a causal relationship between smoking and lung cancer proved necessary. They crystallized with events like the above-mentioned publication of the US Surgeon General's 1964 report building on the comparison of two dozen studies including an American Cancer Society follow-up of one million people whose aim was to urge public action. A similar role was previously played by a 1957 report of the Medical Research Council. Warning the British government of public-health consequences of the causal relationship between smoking and lung cancer, this report was based on a survey comparing the incidence of lung cancer among smoking and non-smoking British physicians.
It is noteworthy that, in both instances, public expertise was associated with strong claims regarding the rules of inference and the nature of epidemiology. In 1964, the US report argued for its own series of criteria for causality, which emphasized large numbers and the notion of relative risk [Lilienfeld 1983]. In 1965 Bradford Hill summarized the methodological lessons of the tobacco debate with a series of nine heterogeneous criteria to be taken into account for similar issues: 1) the strength of the association (for instance the fact that ‘prospective inquiries have shown that the death rate from cancer of the lung in cigarette smokers is nine to ten times the rate in non-smokers'), 2) the consistency of the observed association (accordingly the Advisory Committee to the US Surgeon-General ‘found the association of smoking with cancer in 29 retrospective and 7 prospective inquiries'), 3) the specificity of the association; 4) the temporality of the association ('which is the cart and which is the horse'), 5) the biological gradient or dose-response curve; 6) the biological plausibility; 7) the coherence ‘with known facts of the natural history and biology of the disease'; 8) experimental evidence; 9) analogy with other pathological situations [Bradford Hill, 1965]. Although Bradford Hill viewed methodological innovation as an important outcome of the lung cancer controversy, he warned his readers that ‘none of my nine viewpoints can bring indisputable evidence for or against the cause-and-effect hypothesis and none can be required as a sine qua non … Finally, in passing from association to causation I believe that in ‘real life’ we shall have to consider what flows from that decision. On scientific grounds we should do no such a thing … But in another and more practical sense we may surely ask what is involved in our decision.’ [Bradford Hill, 1965, p. 300].
Such conjunction of statistical and policy debates is emblematic of the rising notion of risk in the postwar public and medical cultures. Relative risk, attributable risk, or risk factors are epidemiological constructs, which during the two decades following the war emerged at the intersection of biological laboratories, statistical bureaus, hospitals, and public health institutions. They all combine evaluation, assessment, and collective management of pathological events. They all point to the invention of quantitative procedures for handling concerns about chronic and old-age diseases (cancer, obesity, hypertension, heart diseases, etc.), which rapidly took infections’ place in the psyche of inhabitants of Western industrial countries. One major advantage of the notion of risk and associated computational techniques was that it replaced the ‘infinite regress', typical of the search for isolated and final causes, by a juxtaposition of ‘relative’ causes, i.e. a list of factors which may encompass genes, hormones, behavioral traits, lifestyles, food, socioeconomic status, age, ethnicity, etc. In a mathematically more refined way, this form of knowledge recalled old hygienic inquiries more than the bacteriology-rooted surveys of the first half of the century. The emphasis placed on the notion of risk thus highlighted the fact that postwar epidemiology was constructed as a statistical specialty producing models and evidence for public health specialists.
This brings us back to Daniel Schwartz’ career and the diverse meanings of hybridization between statistics and medicine. The rise of epidemiology was closely associated with the tobacco question in France, too. But the field took a different shape due to the specificity of local clinical and mathematical cultures. Little emphasis was put on ‘risk’ and on the public-health dimension of the controversy.
A graduate of the École Polytechnique, Schwartz began his research career as an engineer at SEITA, the French state-run tobacco company. As a mathematically-trained statistician, he knew a little about statistics in biological research, but nothing about statistics in medicine.7 In 1954, Pierre Desnoix—then head of the cancer commission of the Institut National d'Hygiène (INH), the state medical research agency— solicited his help in launching a study of lung cancer and tobacco.8 The study was later expanded to a small group of epidemiologists, first established at Desnoix’ cancer research institute, which eventually became the first unit for medical statistics research in the country.9 Ironically, this cancer study was initially supported by both the medical research agency and the state tobacco industry.10
By a not so strange twist in the story, the mathematically-trained French analysts of ‘cancer and smoking’ focused on methodological innovations. Their contribution was not to add decisive evidence by scaling up previous inquiries (something American and British investigators were actively doing) but to refine the design of control groups and significance tests [Schwartz, 1957, 1961]. Participating in the ‘experimental culture’ characterizing the French reconstruction of biological and medical research and the INH during the 1950s and 1960s, they did not try to further the public debate, but focused on developing new tools and advocating more systematic uses of statistics in clinical research [Schwartz, I960].
It is important to note that this inovation apart, very little was done by the French administration to advance the tobacco question. As stressed by social historians, one reason for this was that public health had for decades been a rather marginal domain in French medicine [Murard & Zylberman, 1997]. Moreover, when thinking of major health problems in the country, doctors and officials in the Ministry listed factors affecting the size of the population on a large scale: child mortality, tuberculosis, and alcoholism. Smoking was a matter of lifestyle, not a scourge. Financial and bureaucratic interests in the production and sale of tobacco added to these factors to make sure that there would be little administrative pressure for expertise or public debates on tobacco and cancer. It is therefore not altogether surprising that the ‘risk culture’ played little role in the discourses of the 1950s and 1960s. It is only in the early 1970s, when Desnoix became head of the French National Cancer Commission, that tobacco, conceived of as a ‘risk factor', surfaced as a problem for administrative intervention and public inquiry [Berlivet, 1998].
French statistical epidemiology therefore developed as a specialty that emphasized both the search for the (multiple) causes of diseases and research on the ‘fundamental’ problems of causal imputation and significance. In contrast to Bradford Hill, Schwartz rarely made ‘cases for action’, but repeatedly pleaded for the statisticians’ irreplaceable role on the basis of his knowledge of the measurement of uncertainty [Schwartz, 1992].
CONCLUSION
Robert Musil's ‘The Man Without Qualities’ begins with a brilliant insight into the love of statistical data characterizing the twentieth century's understanding of modernity:
‘A barometric low hung over the Atlantic. It moved eastward toward a high-pressure area over Russia without as yet showing any inclination to bypass this high in a northerly direction. The isotherms and isothers were functioning as they should. The air temperature was appropriate relative to the annual mean temperature and to the aperiodic monthly fluctuations of the temperature … In a word that charecterizes the facts fairly accurately, even if it is a bit old-fashioned: It was a fine day in August 1913’. [R. Musil, 1995].
Of course Musil is wrong. Isotherms, annual means, and weather correlations are not equivalent to a nice summer day. His comment however reminds us that when dealing with statistical entities observers should not forget ‘the fine day in August'. In the context of the historiography of the field, the warning may be translated into: one must not forget that statistics is a form of government. More precisely, statistics is a boundary field continuously creating decision-making procedures. This chapter argues this point by focusing on examples taken from the history of medical statistics. The origins and fate of randomized clinical trials as well as epidemiological debates about smoking and lung cancer show that major changes in the nature and images of statistics did not stem from mathematical breakthroughs, but emerged out of the work of statisticians operating within a rapidly changing medical arena. The rise of a new statistical culture, typified by randomization and relative risk was therefore rooted in the birth, after the Second World War, of a new form of medical organization focusing on cooperative work, chemotherapy, and state intervention. In other words, looking at statistical creativity from the viewpoint of the ‘dead-end users’ may occasionally shed new light on the dynamics of the domain.
NOTES
1. The introduction of the paper thus explained: ‘The history of chemotherapeutic trials in tuberculosis is filled with errors due to empirical evaluation of drugs; the exaggerated claims made for gold treatment, persisting over 15 years, provide a spectacular example. It had become obvious that, in future, conclusions regarding the clinical effect of a new chemotherapeutic agent in tuberculosis could be considered valid only if based on adequately controlled clinical trials.’
2. Practitioner's historiography has merely focused on this point. For example see (Armitage, 1992).
3. On this conjunction, see the forthcoming work of Alan Yoshioka, The British Clinical Trials of Streptomycin, Ph.D. thesis, Imperial College, London.
4. As Bradford Hill later put it: ‘No controls are essential to prove the value of a drug such as penicillin which quickly reveals dramatic effect in the treatment of the disease. Such dramatic effects occurring on a large scale and in many hands cannot be long overlooked. Unfortunately these undeniable producers of dramatic effect are the exception rather than the rule even in the halcyon days of the antibiotics.’ (Bradford trial, 1951, p. 281).
5. W.B. Tucker, ‘The Evolution of the Cooperative Studies in the Chemotherapy of Tuberculosis of the Veterans Administration and Armed Forces of the USA’ Advances in Tuberculosis Research 10 (1960), 3–4. W.B. Tucker, ‘Evaluation of Streptomycin Regimens in the Treatment of Tuberculosis: An Account of the Study of the Veterans Administration, Army and Navy’ American Review of Tuberculosis 60 (1949), 745–746.
6. H. Marks, op. cit. for a detailed analysis of this trial.
7. Schwartz eventually started some work on the tobacco mosaic virus.
8. Daniel Schwartz, ‘A quoi sert l'épidémiologie’ seminar EHESS, December 11, 1996.
9. D. Schwartz, Note sur le fonctionnement de l'unite de recherche statistique de I'lnstitut Gustave Roussy, undated manuscript, presumably 1960, Archives INSERM.
10. According to Schwartz the latter was interested at finding the chemicals inducing lung cancer and eliminating these side-products through the cigarette production process.
REFERENCES
Armitage, P., [1992] ‘Bradford Hill and the Randomized Controlled Trial’, Pharmaceutical Medicine, 6, 23–67.
Berlivet, L., [19951 Controverses en épidémiologìe: Production et circulation de statistiques médicales, Rapport de recherche, Mission Interministérielle à la Recherche, Paris, Mire.
Berlivet, L., [1998] Une santé à risque, faction publique de lutte contre Valcoolisme et le tabagisme en France, 1954–1997, these de Sciences Polítiques, Université de Rennes 1.
Berkson, J., [1946] ‘Limitations of the application of fourfold table analysis to hospital data’, Biometrics Bulletin, 2, 47–53.
Bradford Hill, A., [1937] Principles of Medical Statistics, London, Lancet.
[1951] The Clinical Trial', British Medical Bulletin, 1951, 7, 278–282.
[1963] ‘Medical Ethics and Controlled Trials’, British Medical Journal, 20 April 1963.
[1965 ] ‘The Environment and the Disease: Association or Causation?’, Proceedings of the Royal Society of Medicine, 58, 295–300.
Brandt, A., [1982] ‘The Cigarette, Risk and American Culture’, Daedalus Fall.
Bynum, W., [1994] Science and the Practice of Medicine in the Nineteenth Century, Cambridge, Cambridge University Press.
Cox-Maximov, D., [1998] The Making of the Clinical Trial in Britain, 1900–1950, Ph.D. thesis, Cambridge University.
Desrosières, A., [1993] La Politique des grands nombres, Paris, La Découverte.
Doll, R., Bradford Hill, A., [1950] ‘Smoking and Carcinoma of the lung’, British Medical Journal, 73, 739–748.
Fisher Box, J., [1978] The Life of Scientist, New York, John Wiley and Sons.
Fisher, R.A., [1935] The Design of Experiments, London, Oliver and Boyd.
[1958] ‘Lung Cancer and Smoking’, Nature, 12 July 1958.
[1958] ‘Cancer and Smoking’, Nature, 30 August 1958.
Gigerenzer, G. et ai, [1989] The Empire of Chance: How Probability Changed Science and Everyday Life, Cambridge, Cambridge University Press.
Hinshwa, H.C., [1954] ‘Historical Notes on Earliest Use of Streptomycin in Clinical Tuberculosis’, American Review of Tuberculosis, 70, 9–14.
Himsworth, Sir H., [1982] ‘Bradford Hill and Statistics in Medicine’, Statistics in Medicine, 1, 301–303. Hobby, G.L., [1985] Penicillin: Meeting the Challenge, New Haven, Yale University Press.
Lilienfeld, A.M., [1983] ‘The Surgeon's General Epidemiologic Criteria for Causality’, Journal of Chronic Diseases, 36, 837–845.
Lock, S., [1994] ‘The Randomised Clinical Trial: A British Invention’ in Lawrence, G., (ed.) Technologies of Modern Medicine, London, Science Museum.
Löwy, I., [1996] Between Bench and Bedside, Cambridge, Harvard University Press.
Macfarlane, G., [1984] Alexander Fleming: The Man and the Myth, Cambridge, Harvard University Press.
Mackenzie, D., [1981] Statistics in Britain, 1865–1930. The Social Construction of Scientific Knowledge, Edinburgh, Edinburgh University Press.
Mantel, N., Haenzl, W., [1959] ‘Statistical Aspects of the Analysis of Data from Retrospective Studies of Disease’, Journal of the National Cancer Institute, 22, 719–748.
Marks, H., [1997] The Progress of Experiment: Science and Therapeutic Reform in the United States, 1900–1990, Cambridge, Cambridge University Press.
Matthews, G., [1995] The Quest for Certainty, Princeton, Princeton University Press.
Medical Research Council Streptomycin in Tuberculosis Trials Committee, [1948] ‘Streptomycin treatment of pulmonary tuberculosis’, British Medical Journal, 30 October 1948.
Meldrum, M.L., [1994] Departures from Design: The Randomized Clinical Trial in Historical Context, Ph.D. thesis, State University of New York at Stony Brook.
Murard, L., Zylberman, P., [1997] L'hygiène dans la République, Paris, Fayard.
Musil, R., [1995] The Man Without Qualities (English translation by S. Wilkins), New York, Alfred A. Knopf.
Pillsbury, D.M., [1946] ‘Penicillin Therapy of Early Syphilis in 14,000 patients: Follow-Up Examination of 792 Patients Six or More Months after Treatment’, American Journal of Syphilology and Dermatology, 30, 134–135.
Porter, T., [1986] The Rise of Statistical Thinking, 1820–1900, Princeton, Princeton University Press.
[1995] Trust in Numbers: The Pursuit of Objectivity in Science and Public Life, Princeton, Princeton University Press.
Proctor, R., [1995] Cancer Wars, New York, Basic Books.
Ross, O.B., [1936] ‘Use of Controls in Medical Research’, JAMA, 145, 20–21.
Schwartz, D., Desnoix, P., [[1957] ] ‘L'enquête française sur Pétiologie du cancer broncho-pulmonaire. Role du tabac’, Semaine des hôpitaux de Paris, 62, 424–437.
Schwartz, D., Flamant, R., Lellouch, J., Desnoix, P., [1961] ‘Results of a French Survey on the Role of Tobacco, Particularly Inhalation, in Different Cancer Sites’, Journal of the National Cancer Institute, 26, 1085–1108.
Schwartz, D., Flamant, R., Lellouch, J., Rouquette, C, [1960] Les essais thérapeutiques cliniques, Paris, Masson.
Schwartz, D., [1962] ‘La méthode statistique en médecine: mode ou nécessité ?’, Gazette Médicale de France, 68, 1919–1922.
[1969] Méthodes staüstiques à Vusage des médecins et des biologistes, Paris, Flammarion.
[1992] he feu de la science et du hasard: la statistique et le vivant, Paris, Flammarion.
Schweber, L., [1996] The Assertion of Disciplinary Claims in Demography and Vital Statistics: France and England, Ph.D. thesis, Princeton University.
[1997] ‘L'échec de la démographie en France au XIXème siècle’, Genèses, 29 décembre 1997.
Starr, P., [1982] The Social Transformation of American Medicine, New York, Basic Books.
US Department of Health, Education and Welfare, [1964] Surgeon's General Report on Smoking and Health, Washington D.C., Government Printing Office.