Burch at the Royal Statistical Society
Burch was accorded the honour of addressing the Royal Statistical Society on 17 May 1978. His lecture was entitled Smoking and Lung Cancer: the Problem of Inferring Cause. The story can be partly told in his own words, as he wrote his own account of it for the non-technical readership of Tabak-Journal called Smoking, lung cancer and scientific debate.
The Royal Statistical Society convenes some of its ordinary meetings in London to a fairly elaborate protocol. First, the Council of the Society chooses a potential speaker and subject, and the Meetings Secretary invites the person so honoured to “submit a paper”. Submission of an invited paper does not, however, guarantee acceptance: the paper undergoes “the normal referring process” by two referees, and “acceptability depends on the reports received”. A paper that survives these hurdles is set up by the printer in galley proof form and despatched to persons known by the Society, or the author, to be interested in the subject matter.
At the Meeting, galley proofs are placed at the entrance to the lecture theatre for the benefit of those attending who have not already availed themselves of copies. The author is not expected to read the paper literally but to follow it “fairly closely” in an oral and more informal version. The speaker’s presentation is followed by contributions from the proposer and seconder. “Their function is not merely to be polite. If they disagree with the paper in any particulars, it is the tradition of the Society that they, like all other speakers, are expected to be uninhibited in saying so.” The seconder is followed by contributions from the floor, including written contributions submitted by persons unable to be present at the meeting. At the end of the discussion the author is expected to make a brief verbal reply. After the meeting remarks of discussants, together with written contributions, are sent to the author, who is expected to prepare a “considered reply” for publication with the paper and the discussion. The procedure is complicated, but provided all deeply interested persons are alerted and can be persuaded to participate, it could be thorough.
The audience included Peter Armitage, who proposed the vote of thanks, as he had done for Doll’s paper, Peter Lee observing for the Tobacco Research Council, and Carl Seltzer, who had come to London for the occasion and made a contribution from the floor. Not present were Doll, Peto or any of their associates. “I am disappointed, as I have been in the past, by the low profile maintained by the staff of the Regius Professor of Medicine at Oxford,” Burch sardonically noted.
As to content, the paper made Burch’s now standard case against the causal theory of smoking and lung cancer. He emphasised statistical inference, dose-response relationships and the statistical distributions associated with them in great detail, while saying little about his biological theories. Some of the material was new, for instance an analysis of Doll and Peto’s dose-response relationship, using data from their 1976 paper on the Doctors’ Study. Also new was a demonstration that the falling rates of lung cancer in British male doctors, which they claimed to demonstrate in that paper, were illusory. Doll and Peto had quoted relative rates with respect to the general population, and when these were converted back into absolute rates, no fall in incidence was seen. Burch also called attention to the latest results from the Swedish twins study, which continued to confirm Burch’s hypothesis (despite Doll and Peto’s insinuation to the contrary in the 1976 paper) and to Armitage’s comments on Doll’s lecture considered above. For readers with the mathematical and statistical grasp of a Burch, this paper is the best exposition of his views, while for most it will not be.
The really interesting material comes in the audience discussion and Burch’s considered responses to it. Some participants were confused and uncomprehending, and Burch easily disposed of them. Most were penetrating and raised points that go to the heart of the matter: Burch considered these at length and answered them in detail. In response to several contributions questioning Burch’s use of the parameter S as a measure of lung cancer incidence, he repeated certain crucial calculations using the more usual standardised death rate (SDR) and showed that his conclusions still stood: an important result.
The discussion was led by Peter Armitage. Armitage reaffirmed his adherence to the causal theory on the general grounds that a large number of strong positive correlations counts for more that the discrepancies between them. He identified the issues as “the reliability of sources of data, the possible effects of confounding variables and the plausibility of alternative biological hypotheses. These problems are basically for the epidemiologist, rather than the statistician.” After some remarks about correlation and causation, and about the parameter S, he finished up: “It is, in my view, right that public health authorities, and employers, should act on evidence of this sort, as they do. Does Professor Burch think they are wrong to do so?”
This looks like a polite way of saying that statistics should keep its nose out of epidemiology. As a separate science it is entitles to set its own intellectual standards, even if they are those of the soft, not the hard, sciences, or public health measures and preventive medicine would rest on no basis.
Burch’s reply began “I doubt whether Professor Armitage and I disagree deeply about approaches to the general problem of inferring cause” and ended “I do believe that unbridled enthusiasm for preventive medicine is a serious hazard to us all.”
P. D. Oldham
Dr P. D. Oldham, speaking second, agreed with Burch that the Doll-Hill theory was unproven and that the scientific standards of epidemiology were low, though he questioned Burch’s reliance on his parameter S. He nicely characterises the methods of risk factor epidemiology: “too often the authors of such studies succeed in implying that a tenable hypothesis, developed from a careful epidemiological survey, is a proven one.” He also has some interesting observations on the way the case against cigarettes took hold in the 1950s.
Most marked of all was the immediate adoption by health authorities (but not revenue authorities) of the view that all that was needed in response to the evidence of an association was to utter loud and clear warnings to the general public, which would then immediately cease to smoke and so cease to acquire lung cancer. Indeed, this turned out to be quite effective with the medical profession, who after a brief period of chain-smoking miniature cigars (under the illusion that these were the cigars whose association with lung cancer was not significant in the first studies, and that not significant means unassociated) were able in large numbers to give up smoking completely, I would guess with some relief. Suggestions that studies of how and why cigarettes came to be associated with lung cancer should be carried out, and that the nature of the desire of people to smoke should be investigated, with the idea that a safe form of smoking might be found, an alternative means of satisfying the desire, or at least that scientific knowledge might benefit, were received with scorn and hostility; I can remember an exceedingly uncomfortable luncheon with some professors of epidemiology and social medicine which ended in my being virtually ostracized from their company for being unwise enough to say something of the sort.
Armitage proposed, and Oldham seconded, the customary vote of thanks.
H. J. Eysenck
Hans Eysenck spoke in unqualified support of Burch, citing unpublished work of his own on twins and personality. Burch briefly thanked him.
G. A. Stern
Mr G. A. Stern of ICL spoke in a personal capacity. The Doll-Hill theory was dubious but cigarettes were scarcely harmless. “Medical statisticians seem too often to live in a Victorian melodrama.” Burch’s model involves four parameters. “If the curve has a basis in physics this would increase its acceptability.” Burch replied that the curve indeed had a basis in physics, and by no means all data would fit these particular four parameters.
(ICL is probably Imperial Cancer Laboratories, not Imperial College, London)
H. Gwynne Jones
Professor H. Gwynne Jones of the University of Leeds said that Burch had disproved “the popular causal theory… in true Popperian fashion.” Burch expressed his gratitude for the support.
Carl Seltzer spoke in support of his friend and discussed the problems of studying ex-smokers with reference to his own work on heart disease. Burch thanked him for highlighting the problem of self-selection.
Afterwards, Seltzer reported to a friend: “the statisticians were very easy on Burch with regard to his stochastic theory… many statisticians felt that the alleged effect of smoking on lung cancer was exaggerated”.
Dr Frank Hansford Miller had an interesting suggestion of his own to make about rising rates of lung cancer: “as much as 60 x 109 tonnes of carbon in the form of CO2 has been produced by the burning of fossil fuels up to 1950.” Burch replied, “To test Dr Hansford-Miller’s hypothesis, methodological difficulties not unlike those discussed here would need to be overcome.”
Ms Joy L. Townsend spoke in defence of Doll’s model of lung cancer and the linear dose-response relationship. Burch had drawn attention to a few anomalies, but these could be explained away.
Burch’s written answer is his longest and most detailed presentation of the case against Doll. He demolished Townsend’s views, showing that what she termed ‘anomalies’ are the crucial data which disprove Doll’s hypothesis.
Dr I. D. Hill (the son of Austin Bradford Hill) had personal knowledge of the original questionnaire circulated to British doctors in 1951. There was no question about inhalation because the number of questions was restricted to three to maximise the response rate. To this, Burch’s retort was “If an epidemiological study cannot be properly conducted should it be done at all?”
Written contributions were received from four individuals who were unable to be present.
R. A. Cartwright
Dr R. A. Cartwright asserted that “all cancers are the result of external agents” and expounded some ideas of his own on the role of genetics in carcinogenesis. Burch replied that it might be true that an external agent is always involved but it was for Cartwright to persuade him of this.
Professor Alvan Feinstein of Yale University recounted some past mistaken environmental explanations of disease which lent plausibility to the suggestion that Doll might be mistaken. Burch concurred.
K. V. Mardia
Professor K. V. Mardia of the University of Leeds said that Burch had “made a deep impact on us all in Leeds with his valuable contributions” but questioned the reliance on the parameter S and enquired how the parameters were derived. Burch replied that the use of S rather than SDR was not crucial to his argument.
M. E. Wise
The last, and most perceptive, comment was received from Dr M. E. Wise of Leiden University. Wise was a radiologist who understood that Burch brought a physicist’s perspective to the problem of disease and sympathised with his ambitions to make a hard science out of medicine. He agreed absolutely that Burch had disposed of simple, obvious causal theories of smoking and lung cancer, including Doll’s, but was not convinced of Burch’s positive theses. He questioned the use of five parameters to fit ten data points, the use of a constant as an estimate of the latent period λ and the lack of curve fitting. However, the cases where only the parameters S and k varied for a specific disease (lung cancer was, in fact, one of these cases) “particularly merited further study”.
Burch agreed that he had not fitted the curves to the data: because of a lack of information on the latent period λ made the problem intractable. However, he had empirical evidence that the latent period was often a constant, and illustrated the point from his ongoing work on coronary heart disease. Diseases like lung cancer with a constant modal age of onset proved that his parameters n and r were constant with respect to individual diseases and only took low integers as values.