Journal of Clinical Epidemiology 54 (2001) 325331

COMMENTARY

Reflections on the limitations to epidemiology

George Davey Smith*
Department of Social Medicine, University of Bristol, Canynge Hall, Whiteladies Road, Bristol BS8 2PR, UK Received 4 July 2000; received in revised form 9 August 2000; accepted 30 August 2000

1. Introduction Epidemiology provides a good example of the dictum that the smaller and more specialised a particular field is, the more ferocious are the internal battles fought by the initiates. Over the past few years the epidemiology wars[1] have intensified considerably, in ways which, I am sure, would have entertained and engaged Petr Skrabaneks interest and elicited further contributions from him. There are several separate dimensions to critiques of current epidemiological practice, with diametrically opposed views being expressed on several key issues by different commentators. As is often the case, however, the various criticisms are sometimes linked in ways that suggest they are mutually reinforcing, when on occasion they point in different directions. The apparent crisis in epidemiology is, in fact, a series of crisesalthough most commentators have been engaged with the particular aspect of epidemiologys malaise that interests them most. The failings of epidemiology that have been identified in recent critiques embrace most aspects of the epidemiological enterprise. First, the propensity of risk factor epidemiology to become the indiscriminate identification of particular aspects of daily life as dangerous to health has been widely condemned. Second, critics censure the willingness of epidemiologists to accept black box associations between exposures and disease as serious contenders for causal relationships, with no understanding of biological mechanisms. Third, many preventive medicine campaignsregarding diet, drinking, smoking, recreational drugs, forms of sexual behaviour, tanning, sloth, obesity are seen to be driven more by moralism than by science. Epidemiology is, from this perspective, not value free, but a tool used to support predetermined goals. Fourth, epidemiology as it is currently practiced is seen to be excessively

concerned with individual risks and inadequately engaged with the social production of disease. Why should problems within a small academic disciplinea subspecialty within public health, itself a minor part of the health and health care enterprisebe of interest to other than professionals earning their living within the field? Epidemiology certainly elicits emotional responses. In his recent pop-history The Rise and Fall of Modern Medicine [2], the medical journalist James Le Fanu went so far as to suggest that the solution to the fall of his title was to close all departments of epidemiology. If it would work, this certainly sounds like a cost-effective approach to problems within medicine (except to those of us who would be looking for alternative employment opportunities). Can epidemiology really be responsible for the current malaise in the medical enterprise? The suggestion that it might is certainly helped by the high profile epidemiology receives in the popular media. In Britain the broadsheet daily newspapers tend to pick one medical story out of the weekly medical journals (the BMJ and the Lancet), which are published on a Friday. More often than not the highlighted story will be an epidemiological onesome aspect of daily life has been shown by epidemiologists to be bad for people. 2. Risk factor epidemiology A major component of Petr Skrabaneks attack on epidemiology relates to what he termed risk-factorology[3]. This is the practice of performing studies that relate a myriad of potential exposures to disease risk and identifying those that are positively associated with disease as being risk factors for the condition (and those that are negatively linked as being protective factors). Although it is continuously pointed out (by the risk-factorologists among others) that association does not mean causation, the major reason for doing such studies is to identify causes of disease. There is not a great deal of excitement in showing that cabbage consumption is linked to a lowered risk of cancer if it were not thought that by encouraging people to eat more cabbage you would lower the burden of cancer in the population [3]. A particular weakness of risk factor epidemiology is that different investigators often report contradictory findings

This paper is adapted from the third Skrabanek lecture, given at Trinity College, Dublin on 26th March 1999 in honour of the late Petr Skrabanek. The discussion of the limits of epidemiology focused on Petr Skrabaneks work, given the nature of the occasion. A collection of Petr Skrabaneks papers False premises, False promises (Whithorn: Tarragon Press, 2000) contains an earlier and shorter version of the lecture. * Corresponding author. Zetkin@bristol.ac.uk

0895-4356/01/$ see front matter 2001 Elsevier Science Inc. All rights reserved. PII: S0895-4356(00)00 3 3 4 - 6

326

G.D. Smith / Journal of Clinical Epidemiology 54 (2001) 325331 Table 1 Hypothetical outcome of examining 10,000 associations, with a significance level set at 0.05 and a power of 50% How the world is Decision Accept null hypothesis Reject null hypothesis Total Null hypothesis true 8550 450 Type 1 errors 9,000 Null hypothesis false 500 Type 2 errors 500 1,000

from such studies [35]. As Petr Skrabanek points out, this apparent weakness of the method can be turned to an advantage, since once a positive association has been identified further positive studies are noteworthy (replication being the hallmark of science, perhaps a genuine cause is being identified?) while negative findings may be controversial and thus of interest [3]. The bottom line, it almost always seems, is that more research is neededa conclusion comforting to epidemiologists working in the field. Without previous positive associations having been reported, studies finding no association between a potential risk factor and a disease are unlikely to be of great interest, however. The published epidemiological literature certainly contains reports of more positive associations than null resultsan example of publication bias, whereby only apparently exciting findings reach the light of day. By this means alone a host of purely chance findings will be published, as by conventional reasoning examining 20 associations will find one statistically significant at the P .05 level finding. If only the positive findings are published then they may be mistaken for meaningful links between exposure and disease, not as the necessary chance associations which they are. Since many studies contain long questionnaires collecting information on hundreds of variables, and measure a wide range of potential outcomes, several false-positive findings are virtually guaranteed [3,5,6]. The situation is probably worse than even this reasoning suggests. The percentage of apparent findings that are false positives will depend on the proportion of associations that are examined that are meaningful associations. Table 1 represents the scenario in which the proportion of hypotheses tested in epidemiological studies that reflect real associations is 1 in 10 (1000 out of 10,000 in the table). We consider the situation where the power of studies that are carried out is 0.5, which means that there is a 50% chance of detecting real effects if they exist. This is reflected in the null hypothesis false column, where 50% of the time the null hypothesis (of no association) is rejected (i.e., it is assumed that there is a real effect). Imagine that 10,000 associations are examined and that 90% of the associations that are tested reflect hypotheses that are not valid (i.e., are cases when the null hypothesis should be accepted). At the conventional significance level of P .05, 450 out of 9000 times the statistical test will have wrongly rejected the null hypothesis (which is referred to as a type I error), since 450 divided by 9000 0.05. Thus, of the 950 cases where the null hypothesis is rejected, 500 times it is rejected correctly and 450 times it is rejected incorrectly. In other words 450/ 950 timeson nearly half of the occasionsapparently statistically significant findings are type 1 errors [7]. Remember that this example relates to a case where 1 in 10 hypotheses is meaningful. The actual situation is probably less favourable than this, since the nature of many epidemiological studies means that a great magnitude of potential associations can be examined, especially when each potential association can be looked at within many subgroups (men,

women, the young, the elderly, those with existing disease, those without existing disease, those with high blood pressure, those with normal blood pressure, etc.). The near-random nature of some epidemiological associations that have been identified in the literature [3,5,6] is partly explained by the above reasoning. These supposed findings do not reflect links in the outside world; they are constructed by the epidemiologists methods of examining data and reporting results. A second set of associations can be uncovered that does exist in the real world, but contains associations that may have a different meaning than is conventionally attached to them. To illustrate this I would like the reader to inspect Fig. 1. This demonstrates the association between cigarette smoking and the risk of dying from three causes of death in a follow-up study of a third of a million men in the U.S. The vertical axis represents the relative risk of dying, with the baseline (relative risk 1) being for nonsmokers and a relative risk of 2 indicating twice the risk of dying for the group that experiences it. Thus, for cause of death A, the men who smoked 119 cigarettes per day (less than 1 pack per day) had 1.4 times the death rate of nonsmokers (i.e., 40% higher rate); the men who smoked 2039 cigarettes per day (1 to less than 2 packs per day) had 1.9 times the risk; the men who smoked 4059 cigarettes (2 to less than 3 packs per day) had 2.3 times the risk and the men who smoked 3 packs a day or more had 3.4 times the risk. For cause of death B the equivalent relative rates were 1.5 (less than a pack); 2.3 (12 packs); 2.5 (23 packs); and 2.7 (3 or more packs). For cause of death C the relative rates were 2.0 (less than a pack); 2.3 (12 packs); 2.3 (23 packs); and 2.4 (3 or more packs). Now, with your knowledge of the health consequences of smoking, which causes of death do you think are represented by A, B and C? C, which shows an increase in risk from nonsmokers to less than a pack a day smokers, but little increase after this, is coronary heart disease. B, which shows steadily increasing risk with increasing smoking, to a final risk about 2.5 times that of nonsmokers, is stroke. Cause of death A, which also shows a steadily increasing risk with increasing smoking, but to a final risk about 3.5 times that of nonsmokers, is not (as you may have thought) lung cancer; it is suicide [8]. This is not a chance findings; it has been reported from a series of studies [9] and, unlike the P 0.05 effects commonly reported, the level of statistical significance is

G.D. Smith / Journal of Clinical Epidemiology 54 (2001) 325331

327

Fig. 1. Relative rate of three causes of death in relation to smoking (0 nonsmoker; 119 less than 1 pack per day; 2039 1 but less than 2 packs per day; 4059 2 but less than 3 packs per day; 60 3 or more packs per day).

considerably more stringent ( p 0.0001). What is likely to be happening here is a mixture of confounding and reverse causation. By confounding we mean that the potential risk factor of interestsmokingis linked to another factor that in turn is a cause of the outcome. The conventional exposition of confounding refers to the cause of death not included in Fig. 1lung cancer. Lung cancer shows a much stronger (and certainly causal) doseresponse association with cigarette smoking, with a relative rate of more than 10 among smokers of two packs of cigarettes a day [10]. Smoking causes people to acquire yellow fingers; thus, in an observational epidemiological study yellow fingers would be associated with smoking and a nave interpretation would consider yellow fingers as a potential cause of lung cancer (Fig. 2). In our example, if people who are depressed, poor, alcoholic or chronically ill are more likely to be smokers then, as these states increase the risk of suicide, smoking will be noncausally associated with suicide. Taking away the cigarettes from smokers will not, in this case, reduce their risk of suicide (indeed for depressed people who smoke to relieve stress it may increase their risk). Given the strength of antismoking feeling in public healthconsidered below in the section on epidemiological

moralismand also the justified outrage at the tactics of the tobacco companies in covering up the genuine health-damaging effects of smoking, it is perhaps not surprising that even though the paper reporting the smokingsuicide association was subtitled illustration of an artefact from observational epidemiology [8] it has been widely taken to illustrate that smoking does indeed cause suicide. Biological mechanisms through which this could act (e.g., smoking suppressing brain serotonin levels) have been advanced. To address this we looked at an even less plausible cause of deathhomicide. Smokers of two packs or more per day were more than twice as likely to be murdered as nonsmokers. Unless the provisional wing of the health education lobby has moved on to a direct action phase, during which they shoot smokers, this association is very unlikely to be causal. The strategy of epidemiologists to deal with confounding is generally statistical adjustment or control, leaving an estimate of the independent effect of the risk factor of interest. In theory this sounds an excellent solution, but unfortunately it is of limited use. In the example of smoking and suicide we controlled for a variety of other risk factors (income, ethnicity, chronic disease), all of which are potential confounders, and this had almost no effect on the finding (the relative risk for three or more packs a day changed from 3.4 to 3.3 on such adjustment). Not only are there missing variables (e.g., depression) but also the variables that are measured (e.g., income) are measured with error, leaving wide scope for residual confounding to exist after statistical adjustment [11]. 3. Black boxes Petr Skrabanek considered risk factor epidemiology to be a form of black box thinking, as the causal mechanism between exposure and disease is unknown and may even be considered irrelevant [6]. Some epidemiologists consider that epidemiology is most useful when it ventures into territories unconstrained by biological knowledge, since it is here that it may stumble upon novel findings of public health importance [12]. Many epidemiologists, conversely, feel that the future of epidemiology must lie in a rapprochement with biomedical science. Certainly to Petr Skrabanek epidemiology without biology was nonsense; he himself moved from neurotoxicology research (particularly relating to substance P) to his particular blend of natural scientist, forensic toxicologist, doctor of medicine and connoisseur of the absurd (as the footnote to one of his papers read [13]). The black box nature of risk factor epidemiology may be particularly evident where low-level risk are being investigated. The problems of chance and particularly confounding are exacerbated when the exposuredisease associations are weak. However, since a small risk applied to a large proportion of the population can have important public health implications [14], investigating these associations could be

Fig. 2. Exposure A (e.g., smoking) is associated with the outcome (lung cancer risk); exposure B (yellow fingers) is associated with exposure A and hence to the outcome.

328

G.D. Smith / Journal of Clinical Epidemiology 54 (2001) 325331

of potential importance. Thus, it is of concern that only a small degree of measurement error in a confounder can generate residually confounded effects of the magnitude under consideration for low-level radiation, electromagnetic fields around power lines or mobile phones and cancer, for example. In many cases these associations may also be considered biologically implausible [6], although it must be recognised that what is biologically plausible and what is not changes over time. An example of the malleable nature of biological plausibility is demonstrated by two epidemiological studies that appeared within a few weeks of each other in 1991. In one study, of Kenyan prostitutes, oral contraceptive use was independently related to risk of future infection with HIV [15]. The authors suggested several mechanisms through which oral contraceptives could facilitate HIV transmission, including a direct effect on the genital mucosa, making it a less successful barrier to HIV, or an immunosuppressive action that increases susceptibility to HIV. Biological knowledge of the effects of oral contraceptives support both these possibilities. The second study, however, found the exact opposite: a protective effect of oral contraceptives on HIV transmission [16]. The authors again found a biologically supported mechanism for this effect: oral contraceptives thicken the cervical mucus, which then hampers the entry of HIV into the uterine cavity. As one of his examples of moralism misinterpreted as science, Petr Skrabanek quoted a study of cervical cancer which suggested that masturbation was a risk factor [17]. The author of this study was, however, capable of producing an apparent biological mechanism: the pathways from the nervous system to cervical tissue could promote cancer if stimulated [18]. Clearly apparent biological plausibility adds little strength to epidemiological conviction, but equally epidemiology without rapprochement with biology can produce long lists of spurious associations and make no contribution to furthering our understanding of disease aetiology and strategies for disease prevention. An important recent example of how epidemiologists and biologists working together have produced important advances in understanding both epidemiological and biological aspects of disease is seen in the fetal programming field [19]. Here initial ecological associations between indices of poor early life development earlier this century and current disease rates indicated the possible importance of in utero and early childhood growth for later health. These ecological studies were followed by analytical epidemiological studies showing that at an individual level low-birthweight and other indices of suboptimal fetal development were related to cardiovascular disease risk profiles in adults. Although there was some existing basic research on fetal development and later health this was limited, and the epidemiological work has enormously stimulated this now vibrant field of biological enquiry [20]. This exampleof epidemiology and basic research feeding off each otherillustrates the weaknesses of

both black box epidemiology and of attempts to deny that epidemiology can contribute in a fundamental way to our understanding of disease. 4. Public health moralism Petr Skrabanek used to like quoting H. L. Menckens comment that hygienethe public health of its dayis the corruption of medicine by morality. It is impossible to find a hygienist who does not debase his theory of the healthful with a theory of the virtuous. The targets of a mainstream strand of epidemiology are, as mentioned, often diet, drinking, smoking, recreational drugs, forms of sexual behavior, tanning, sloth and obesity [5,13,21]and thus the focus of health improvement schemes end up endorsing the message if it tastes good, dont eat it; if it feels good, dont do it. Another moral duty emphasised by public health is the duty to be well. Strictures against smoking, drinking and sloth can also be ways to ensure a more productive workforce, a fitter army and a more fertile population: things with which, historically, most states have been much concerned [22]. The advantage of installing the habit of expecting state interference in (at worst, only self-damaging) aspects of private life and private behavioursthrough the Trojan horse of beneficenceis also not lost on those who want their leadership accepted. Health as duty can lead to unattainable aims, such as the World Health Organisation definition of health as Not merely the absence of disease or infirmity but A state of complete physical and mental and social well-being. In Petr Skrabaneks words [21], this is the sort of feeling ordinary people may achieve fleetingly during orgasm, or when high on drugs. Certainly, Health for all by the year 2000 remains a bad joke for the large majority of the population of the world, now that we have reached the new millennium and most people still live in poor and unhealthy places. Health and longevity are, of course, not the only aims of life. To quote Petr Skrabanek: There are fates worse than death. Longevity drags if you have buried your children. Poverty, loneliness, incontinence, dependence, and dementia are some of the final rewards. Not everybody hopes for a long life followed by death from boredom. Plato in The Republic recalled the gymnastic teacher Herodicus whose skills enabled him to reach old age in a prolonged death struggle. Hesiods golden race died swiftly, as though in sleep; they had no old age. Why be afraid of sudden death from coronary heart disease if you cannot regret it the day after? The function of epidemiology is to study associations and to provide hypotheses for experiments. The epidemiologists should inform; they have no right to tell people what they should do. If they start advising government on population measures, they declare themselves as agents of social control, they become

G.D. Smith / Journal of Clinical Epidemiology 54 (2001) 325331

329

preventionists for whom the interests of the state override the interest of the individual [23]. The frequent concordance of morality and what passes at any particular time for the science of public health must surely be more than coincidence. Throughout the long history of the war against smokingfrom James I pamphleteering against tobacco, through to Hitlers vigorous opposition to smokingthe combination of puritanism and self-interest seem clear [13]. Similar self-interest is, of course, shown by those who profit from the tobacco trade and deny the evident harmful effects of smoking. However, the impossibility of separating ethics or morality from science is highlighted by realising that the 100,000 Reichsmark donation Hitler made to the Institute for the Struggle against the Dangers of Tobacco supported the first methodologically strong epidemiological study showing that smoking causes lung cancer [24,25]. Indeed, the relative risk of lung cancer amongst smokers that can be calculated from the 1943 German study, of 16.6 [26], is remarkably similar to the relative risk of 16.3 among male heavy smokers that can be calculated from the classic Doll and Hill case-control study [27]. The fact that researchers working under an appallingly inhumane dictatorship, during which the most basic elements of human morality and decency were abandoned wholesale, could produce high-calibre science is a demonstration of the impossibility of reading the truth or falsity of claims from the values underlying them. As Petr wrote with respect to prevailing views in psychiatry, whether psychotherapy (known as moral treatment in the past) or somatic treatment predominates as the ruling paradigm depends less on therapeutic results than on the dominant ideology of the profession. With respect to public health he considered that the issues of preventive medicine have little to do with science, relative risks, and risk factors. They could be more profitably debated within the framework to which they belongethics, politics, and vested interests [23]. The problematic interpenetration of facts and values has recently been highlighted in the epidemological world in a debate regarding whether or not potential conflicts of interest should be declared. Kenneth Rothman, a leading epidemiological methodologist, has suggested that forcing the declaration of conflicts of interest is akin to McCarthyism in science, and risks mislabelling everyone as potentially unethical [28]. However, there is also strong evidence that financial interests influence what people say. For example, it has been shown that the expressed viewpoints of experts regarding the beneficial effects (or otherwise) of calcium-channel antagonists is strongly related to the financial relationships the authors had with manufacturers of these drugs [29]. Rothmans belief that scientists can be objective, and objectivity is not something which is brought or sold, is an opinion I feel Petr Skrabanek would have shared. My own view is that transparencythrough full declaration of potential sources of conflicts of interestsis the best policy in

such cases [30]. Both nondisclosure of interests or censorship of viewpoints by suppressing publications by those with such interests prevent readers getting an informed and rounded view of the issues. Transparency in this regard recognises the impossibility of separating facts and values.

5. Epidemiology as an asocial science There has been considerable recent debate regarding the individualistic focus of much epidemiology on the lifestyles or psychological profiles of people abstracted from their social context [3133]. These authorsin my view correctlypoint out that there are broader social determinants of the risks to health that people suffer, and that attempts to reduce these risks should recognise this fundamental social determination. Others have strongly taken issue with this view [34]. Petr Skrabanek recognised the important role of poverty in determining healthreferring to the fact that medicine brought many benefits, he considered that the fact that these achievements have had little or no bearing on the lives of all those millions of our fellows which are still nasty, poor, brutish, solitary and short is an indictment of our selfish world [22]. He was unwilling to allow this admission of the important influence of social circumstances on health to influence public policy decisions, however [21]. The many weaknesses of epidemiological approaches that fail to locate exposuredisease associations within their historical, political and social context have been convincingly elaborated on [3133,35]. Perhaps less widely acknowledged is that the abstraction of such associations from their particular context can lead to severely misleading conclusions. Take, for example, the extensive research on beta carotene consumption and the risk of cardiovascular disease. Observational studies revealed strong apparently protective effects of beta carotene, but long-term randomised controlled trials found that, if anything, beta-carotene increased cardiovascular disease risk (see Fig. 3) [9]. There are now a series of similar examples: hormone replacement therapy, vitamin E and vitamin C intake in relation to cardiovascular disease among them. What these examples have in common is that the groups of people who were apparently receiving protection from these substances in the observational studies were very different from the groups not using them, on a whole host of characteristics of their lives. Believing that these differences could be summed up, in measures of a few potential confounders and adequately adjusted for in statistical analyses fails to recognise the complexity of the reasons why people differ with regard to particular and general characteristics of their lives. It would be gratifying if the refutation of observational studies by randomised controlled trial evidence in these areas led to a critical evaluation of approaches that abstract single elementswhich are almost always behavioural, psychological or therapeuticfrom the complexity of the life and

330

G.D. Smith / Journal of Clinical Epidemiology 54 (2001) 325331

Fig. 3. Meta-analysis of results of observational cohort studies of beta-carotene intake and cardiovascular mortality and of randomised controlled trials of the same issue. Adapted from [9].

sures on smaller samples facilitates the production of meaningful findings than poorer exposure measurements on larger samplesby quoting the Irish country saying you cannot make a pig grow by weighing him [40]. Others have been similarly (but less entertainingly) dismissive of attempts to improve epidemiological methodology [41]. In my view epidemiology will only progress if it combines a detailed understanding of the ways through which the historical, economic and political constitution of how the world is influences the health of populationsand thus the individuals within these populationswith the appropriate development of methodology and concepts to deal with this complexity [32,33]. Petr Skrabaneks self-termed destructive criticism can feed into the process of formulating better questions and developing the necessary methods for answering these questions, however. He saw such criticism as a necessary cleansing act, which then allowed the building of useful knowledge. A critical appraisal of a medical theory takes so much time and effort that most people (and some of them wise) do not bother. It is like an archaeological dig the slow painstaking search for nuggets of truth by removing layer after layer of deposits left behind by fly-by-night scholars, of dust from the dry-run research, of debris left by those who would rather publish than perish in the rat race, of crumbled clay from the legs of the clossal cacademics carrying the dogma on their shoulders, . . . . [42] An illustration of the importance of such criticism is provided by the fact that the British newspapers have recently been full of reportsbased on a recent Lancet paper [43] that breast cancer screening with mammography may not be a useful preventive strategy. Fifteen years prior to this paper Petr Skrabanek provided a devastating critique of breast cancer screening, also in the Lancet [44]; a clear demonstration that the cleansing act of criticism is required as much today as ever.

times of people, and relate these to single health outcomes. It is likely, however, that as in many decaying research programmes auxiliary hypotheses will be mobilised to explain each apparent mistake, on a case-by-case basis. In my view a more promising approach is to see the development of health potential within its particular concrete historical, social and geographical location. Such a lifecourse approach [36,37] does not abstract individual life trajectories from these contexts; indeed, the long-term and sometimes irreversible outcomes of social circumstances at different stages of life are seen to become literally embodied [38]. Low birthweight, growth in childhood (and final adult height), persistent infections acquired in early life (such as H pylori) or the failure to acquire infections (leading to immunological programming increasing the risk of atopy), lung function, degree of adiposity, a habitus which embraces particular dispositional characteristics (including attitudes, health-related behaviours and mood), modes of self-presentation, and ways of dealing with misfortune, may seem to fall within different categories, but they are all essential components of life trajectories that influence health. An epidemiology that appreciates the necessary interconnections between these different domains of life will avoid the dead-ends that research strategies based on abstraction and narrowing the focus of attention can lead to. 6. Criticism and the advance of knowledge Some recent writings about the limits of epidemiology have raised ways in which we could do things better than we have in the past. Petr Skrabanek was not interested in playing this game; indeed in a Lancet editorial he dismissed a methodological paper I coauthored about improving the design of epidemiological studies [39]which made the modest proposal that more rigorous measurement of expo-

Acknowledgments Many thanks to Liz Humphries and Claire Diamond for helping in preparing this manuscript and to Nancy Krieger and John Lynch for comments on an earlier version.

References
[1] Poole C, Rothman KJ. Our conscientious objection to the epidemiology wars. J Epidemiol Community Health 1998;52:6134. [2] Le Fanu J. The rise and fall of modern medicine. New York: Little Brown, 1999. [3] Skrabanek P. Risk-factor epidemiology: science or non-science? In: Anderson D, editor. Health, lifestyle and environment. London: Social Affairs Unit, 1991;4756. [4] Mayes LC, Horwitz RI, Feinstein AR. A collection of 56 topics with

G.D. Smith / Journal of Clinical Epidemiology 54 (2001) 325331 contradictory results in case-control research. Int J Epidemiol 1989; 18:7257. Skrabanek P. The poverty of epidemiology. Perspect Biol Med 1992; 35:1825. Skrabanek P. The emptiness of the black box. Epidemiology 1994;5: 5535. Oakes M. Statistical inference. Chichester: Wiley, 1986. Davey Smith G, Phillips AN, Neaton JD. Smoking as independent risk factor for suicide: illustration of an artifact from observational epidemiology? Lancet 1992;340:70912. Egger M, Scheider M, Davey Smith G. Spurious precision? Metaanalysis of observational studies. BMJ 1998;316:1404. Davey Smith G, Phillips AN. Passive smoking and health: should we believe Philip Morriss experts? BMJ 1996;313:92933. Davey Smith G, Phillips AN. Confounding in epidemiological studies: why independent effects may not be all they seem. BMJ 1992; 305:7579. Savitz DA. In defense of black box epidemiology. Epidemiology 1994;5:5502. Skrabanek P. Smoking and society. In: Armitage A, editor. Other peoples smoke. Berkeley: Galen Press, 1991;17383. Rose G. The strategy of preventive medicine. Oxford: Oxford University Press, 1992. Plummer FA, Simonsen JN, Cameron DW, Ndinya-Achola JO, Kreiss JK, Gakinya MN, Waiyaki P, Cheang M, Piot P, Ronald AR. Cofactors in male-female sexual transmission of human immunodeficiency virus type 1. J Infect Dis 1991;163:2339. Lazzarin A, Saracco A, Musicco M, Nicolosi A. Man-to-woman sexual transmission of the human immunodeficiency virus. Arch Intern Med 1991;151:24116. Skrabanek P. Cervical cancer in nuns and prostitutes: a plea for scientific continence. J Clin Epidemiol 1988;41:57782. Rotkin ID. Sexual characteristics of a cervical cancer population. Am J Public Health 1967;57:81529. Barker DJP. Mothers, babies and health in later life. London: Churchill Livingstone, 1998. Nathanielsz PW. Life in the womb: the origin of health and disease. New York: Promethean Press, 1999. Skrabanek P. The death of humane medicine. London: The Social Affairs Unit, 1994. Skrabanek P, McCormick J. Follies and fallacies in medicine. 3rd ed. Withorn: Tarragon Press, 1998. Skrabanek P. Preventive medicine and morality. Lancet 1986;i:1434. Schairer E, Schniger E. Lungenkrebs and Tabakverbrauch. Z Krebsforschung 1943;54:2619. (translated as Lung cancer and tobacco consumption, Int J Epidemiol 2001;30, in press)

331

[5] [6] [7] [8]

[9] [10] [11]

[12] [13] [14] [15]

[16]

[17] [18] [19] [20] [21] [22] [23] [24]

[25] Davey Smith G, Strbele SA, Egger M. Smoking and health promotion in Nazi Germany. J Epidemiol Community Health 1994;48:2203. [26] Davey Smith G, Strbele S, Egger M. Smoking and death. BMJ 1995; 310:396. [27] Doll R, Hill AB. A study of the aetiology of carcinoma of the lung. BMJ 1952;II:127187. [28] Rothman JK. Conflict of interest. The new McCarthyism in science. JAMA 1993;269:27824. [29] Stelfox HT, Chua G, ORourke K, Detsky AS. Conflict of interest in the debate over calcium-channel antagonists. N Engl J Med 1998; 338:1016. [30] Smith R. Beyond conflict of interest. Transparency is the key. BMJ 1998;317:2912. [31] Diez Roux AV. Bringing context back into epidemiology: variables and fallacies in multilevel analysis. Am J Public Health 1998;88:21622. [32] Koopman JS, Lynch JW. Individual causal models and population system models in epidemiology. Am J Public Health 1999;89:11704. [33] Krieger N. Epidemiology and the web of causation: has anyone seen the spider? Soc Sci Med 1994;39:887903. [34] Rothman KJ, Adami HO, Trichopoulos D. Should the mission of epidemiology include the eradication of poverty? Lancet 1998;352:8103. [35] Schwartz S, Carpenter KM. The right answer for the wrong question: consequences of type III error for public health research. Am J Public Health 1999;89:117580. [36] Kuh D, Ben-Shlomo Y, editors. A life course approach to chronic disease epidemiology. Oxford: Oxford University Press, 1997. p. 1541. [37] Davey Smith G, Gunnell D, Ben Shlomo Y. Lifecourse approaches to socio-economic differentials in cause-specific adult mortality. In: Leon D, Walt G, editors. Health inequalities and public policy. Oxford: Oxford University Press, 2000. [38] Najman JM, Davey Smith G. The embodiment of class-related and health inequalities: Australian policies. Aust N Z J Public Health 2000;24:34. [39] Phillips AN, Davey Smith G. The design of prospective epidemiological studies: more subjects or better measurements? J Clin Epidemiol 1993;46:120311. [40] Skrabanek P. The epidemiology of errors. Lancet 1993;342:1502. [41] Pearce N, McKinlay JB. Back to the future in epidemiology and public health: response to Dr Gori. J Clin Epidemiol 1998;51:6436. [42] Skrabanek P. In defence of destructive criticism. Perspect Biol Med 1986;30:1926. [43] Gtzsche PC, Olsen O. Is screening for breast cancer with mammography justifiable? Lancet 2000;355:12934. [44] Skrabanek P. False promises and false premises of breast cancer screening. Lancet 1985;ii:31620.