Causality and causal inference in epidemiology: the need for a pluralistic approach (2024)

Elements of the RPOA approach

The RPOA draws inspiration from at least three distinct developments in three distinct disciplines. The first is the pragmatic attitude to causality adopted by epidemiologists studying smoking and lung cancer. The first expression of an explicitly pragmatic approach was perhaps articulated by Lilienfeld in 1957:

In medicine and public health it seems reasonable to adopt a pragmatic concept of causality. One major reason for determining etiological factors of human disease is to use this knowledge to prevent the disease. Therefore, a factor may be defined as a cause of a disease, if the incidence of the disease is diminished when exposure to this factor is likewise diminished.12

This focus on the ‘cash value’ of causal claims for epidemiologists is also evident in the attitudes of present-day advocates of the RPOA, who advocate restricting attention to those factors which we can intervene on.

The second source of inspiration is the counterfactual approach to causality. In the 1970s and 80s, philosophers like David Lewis sought to give counterfactuals clear meaning—and to use them to analyse causation.13–16 Donald Rubin’s seminal paper coining the phrase ‘potential outcomes’ was published in 1974.17 Decades later, more direct links between philosophical and statistical thinking about causation were developed by Spirtes, Glymour and Scheines18 and by Pearl.10^,19

Pearl asserts that causality has been mathematized10 although, as noted above, he adopts a more inclusive attitude towards causal claims than advocates of the RPOA in epidemiology. Pearl’s framework has direct and explicit connections with Lewis’s philosophical work. First, the ‘do’ operator can only be made sense of in a philosophical climate that permits talk about non-actual events. Second, Pearl’s directed acyclic graphs (DAGs) are similar in appearance and operation to the neuron diagrams developed by Lewis and his students, albeit governed by a more explicit set of rules and allowing variables to take numerical values. Third, Pearl explicitly seeks to harmonize his proposed framework with Lewis’s metaphysics.10^,19 Pearl’s most striking philosophical contribution is his marriage of the counterfactual and probabilistic approaches to causation. This offers the hope of deducing testable differences between competing hypotheses. The toolkit offered by Pearl has been extended by epidemiologists to remarkable effect, with certain problems that were previously intractable, or at least headache-inducing, being elegantly resolved.20^,21 In the RPOA some epidemiologists have gone one step further, to use the toolkit as a framework for expressing causal hypotheses precisely and to insist that any causal claim that cannot be expressed in this framework is not a well-defined causal claim at all, for epidemiological purposes.

Philosophers have discussed in detail and at length the semantic interpretation of counterfactual concepts.13^,22 Epidemiologists, however, are more interested in the epistemology of counterfactuals, i.e. how they lead to assessment of causality. That is perhaps a reason why the language of ‘potential outcomes’ is often preferred.23One simply imagines that one has knowledge about the outcome of two different treatments (levels of exposure) in the same individual17—and avoids the philosophical complexities of ‘possible worlds’.

The RPOA’s family tree

It is helpful to situate the RPOA within a ‘family tree’ of theories of causation (Figure 1). At the top level, we have a number of broad conceptual approaches to causation—including difference-making theories, regularity theories, probabilistic approaches, singularist approaches and dispositional analyses, among several others. Each of these approaches has received extensive treatment in the philosophical literature. A useful and up-to-date guide is the Oxford Handbook of Causation, which discusses them all.24

The RPOA is part of the family of difference-making theories of causation, which share the idea that causes are events which make a difference to their effects, in the sense that had the cause been different or absent (in some sense specified by the theory in question), the effect would also have been different or absent.

Among difference-making theories, it is possible to distinguish three broad lines of thought, at the next ‘level’ in the family tree. (These levels have no special significance; the lines in the figure are merely a heuristic aid in understanding how the various philosophical views relate to each other.) First there is the traditional counterfactual theory of causation, as advocated by Lewis, according to which a cause is something such that, had it been absent, the effect would also have been absent (for at least some individuals).15 For example: ‘Had she not been obese, she would not have developed a myocardial infarction.’ Second there is contrastivism, which holds that causation is not a relation between cause and effect alone, but a four-way relation between cause, effect, contrast for the cause and contrast for the effect.25 Third, the interventionist line of thought holds that the notion of causation is closely related to the notion of an intervention [p. 20–25]26—a cause is something upon which we can imagine intervening to alter an outcome (e.g. how would the incidence of myocardial infarction differ, due to an intervention to make obese persons lean?). The contrastivist and interventionist views share the idea that causal thinking involves explicit thinking about the contrasting states of affairs that are being considered. They differ from the counterfactual view (with which they are sometimes wrongly conflated) which lets the semantics governing counterfactual conditionals dictate what happens in the absence of the cause. The interventionist and contrastivist views differ from each other in that the contrastivist is offering a semantic thesis—that is, a theory about the meaning of causal claims,25 whereas the interventionist is more interested in explicating the use of the causal concept in causal reasoning (inference, explanation, prediction).26

The RPOA falls clearly into the interventionist camp, but this can be subdivided again. On one side there are in-principle interventionists. They believe that the notion of an intervention is not confined to what is humanly possible. Humans cannot cause an earthquake, but an earthquake could nevertheless be an intervention. The more restrictive kind of interventionist view we might call humanly feasible interventionism. This is the view that causes correspond to interventions that humans can actually make. It is here that the RPOA in epidemiology locates itself. Thus, for example, Greenland uses an earthquake as an example of an event that is not an intervention, on the basis that it cannot be brought about by humans.27 In contrast, for an in-principle interventionist such as James Woodward, it is easy to imagine an earthquake as part of a counterfactual causal history.

In fact, the general POA is not logically committed to humanly feasible interventionism, but RPOA advocates have tended to assume this commitment. Note that we have not included the more general POA in the figure since, as noted above, it could be included with the counterfactual approach or the interventionist approach (as is the RPOA), depending on which version of the POA is adopted. For example, a number of RPOA theorists have treated it as obvious that race and sex are ‘non-manipulable’.7^,28 Thus we proceed on the basis that a commitment to humanly feasible interventionism is part of the RPOA in epidemiology.

The RPOA thus represents a heavy bet on a very specific philosophical stance. This stance largely mirrors the point of view that only randomized controlled trials can assess causality, even though the history of epidemiology indicates that this narrow approach to causality has rarely, if ever, been valid or useful.29 In our view, such a restrictive approach is unwise; in epidemiology the stakes are high, and in philosophy the odds are bad for any given theory being correct, in the sense of being applicable to all instances that are judged ‘causal’.

The problem of unfeasible interventions

The problem of restricting science to studying feasible human interventions is that there are clearly causes that are not (or do not correspond to) feasible human interventions. Moving tectonic plates cause earthquakes; heat waves cause deaths; mutations cause drug resistance; being born with two X chromosomes causes lower salaries. In each case, we can specify perfectly good (counterfactual) values that the relevant variables could take: no earthquake; no heat wave; cell division without mutation; being born with an XY pair of chromosomes. At this moment, there is no humanly feasible way to bring about any of these things. However, these limits change over time: it was once not feasible to intervene on blood cholesterol, and now it is. Perhaps fracking will cause earthquakes.

These issues have been highlighted in recent papers about whether race and sex are causes,7^,28 which describe an RPOA account that is limited to humanly feasible interventions. Actually having a second X chromosome is associated with particular actual outcomes, such as higher examination marks, lesser wages, receiving tenure less often etc. However, it is not feasible to intervene on a person’s sex at the relevant life stage merely to improve examination marks or to produce more equal wages. For this reason, RPOA theorists regard antecedents of the following sort as imprecise: ‘If Jane had been a man…’ The absence of any feasible intervention that would bring this change about means, for them, that the hypothetical scenario of Jane’s non-womanhood is not well specified, and thus that causal effects are hard to conceptualize and quantify.

There are two difficulties with this stance. The first is that the boundary around the humanly feasible is not sharp; it can and does change over time. Indeed, a typical target of medical research is to bring causes that we cannot currently intervene upon under our control. Hypercholesterolaemia was held to be a cause of cardiovascular diseases, long before effective treatment by statins existed.30 Today, BRCA1 and BRCA2 are causes upon which we cannot intervene. Maybe we will one day intervene upon these. Since cholesterol can now be lowered with statins, it is straightforward to accept it as a cause. In future we may be able to intervene on BRCA1 and BRCA2 or in their biological pathways; but the cases of hypercholesterolaemia and BRCA1 and 2 are parallel, since there was a time when neither could be intervened upon. If they are parallel, then it follows that BRCA1 and BRCA2 are causes. This illustrates a logical flaw in the attempt to count only humanly feasible interventions as causes. It is only by disrespecting that boundary that medical science will be able to shift it.

The second problem arising from the RPOA’s stand on unfeasible interventions is that it represents a confusion between what we can conceptualize and what we can do. If the objection were that no possible counterfactual value could be assigned to ‘If Jane had been a man…’ then, within a framework that insists on analysing causal claims counterfactually, it would be correct to assert that the corresponding causal claim is meaningless. However, the mere fact that we cannot bring it about that Jane-like people be born a different sex does not mean that we cannot imagine a world in which Jane (more precisely, Jane’s counterpart) was born a man. In order for an intervention to be well specified (the POA prerequisite for causal claims to be well defined), it is not necessary that the intervention can be done. There is a difference between specifying and doing. We conclude that the commitment to human feasibility is something that the RPOA would be better off without.

The ‘states’ problem

A further problem with the RPOA approach is that it cannot adequately deal with ‘states’ such as obesity as causes. Hernán and Taubman suggest that obesity could never be studied as a cause under RPOA precepts:

In an observational study, we do not know the actual procedure by which each subject achieved a BMI of 20; thus, the counterfactual outcome…. when assigned to a BMI of 20 is too vague a concept. An immediate consequence of a vague counterfactual outcome is that any causal contrast involving that counterfactual outcome will be ill defined.5

When ‘states’ like obesity (or hypercholesterolaemia, hypertension, carrying BRCA1 or BRCA2, male gender) can no longer be seen as causes, this is another departure from successful historical epidemiological practice (see Lilienfeld’s quote above). The epidemiologist will no longer be a student of aetiologic factors, but only of interventions.

The specificity problem

The deeper problem with the RPOA concerns its reliance on the notion of a well-specified intervention, whether humanly feasible or not.

Hernán and Taubman point out that, if one sets about reducing the BMI of a group of obese people, one may have a different effect on mortality depending on how one intervenes5—e.g. exercise, liposuction, etc. They argue that one can only make a meaningful causal claim if one has a specific intervention in mind. Thus one cannot say that 100 000 deaths annually are attributable to obesity; one must say that 100 000 deaths among obese people could be prevented by exercise—50 000 by diet, 120 000 by a combination, none at all by liposuction and so forth. This leads Hernán and Taubman to assert that obesity is not a well-defined cause, whereas overeating or lack of exercise or a combination of the two are all well-defined causes. The difficulty is that ‘well-specified intervention’ is itself in need of definition. There are many difficulties with the claim that causation is well defined only when interventions are well specified.

The first difficulty is that the interventions specified by Hernán and Taubman are open to exactly the same kind of critique that they direct at the notion that obesity is a cause. The way that you perform one hour of strenuous physical exercise per day (e.g. running, cycling, playing rugby, boxing), or restrict calories, may have a huge impact on mortality. So the argument proposed in their paper is an instance of double standards.

A second, related difficulty is that often one does not know in advance whether an intervention is well specified. Perhaps all aerobic exercise has the same effect; perhaps all running has the same effect; perhaps all interval training; and so forth. But on the RPOA, we cannot decide whether causal questions about the effect of—say—running on obesity-related mortality are well defined until we have answered the question as to whether running is a well-specified intervention. And we cannot do that until we have answered the question as to whether—say—sprinting, as opposed to middle-distance running, is a well-specified intervention. And we cannot do that until we have found out whether 100-m sprinting as opposed to 200-m is a well-specified intervention. And so forth, ad infinitum. We are paralysed; we cannot ask a well-defined causal question without first answering an infinite series of questions concerning specification of the interventions.31

Third, there are some cases where the exact nature of the intervention does not seem very important provided it achieves its goal. In such cases, precisely specifying an intervention would be a waste of time. Hypertension arises from a number of quite diverse mechanisms (renal, cardiac, vascular), and different drug treatments exist with different mechanisms (diuretics, calcium channel blockers, angiotensin antagonists etc.). By and large, whatever drug treatment and whatever the underlying mechanism for the hypertension, lowering blood pressure has the desired beneficial consequences.32

In sum, the situation is much more complex than a simple insistence on exact specification implies. Because the RPOA refuses to acknowledge the meaningfulness or usefulness of causal claims that do not lend themselves to well-specified interventions, these theoretical problems gives rise to practical problems, to which we now turn.

The importance of ‘bad’ evidence

Our first practical criticism of the RPOA is that, in effect, it ranks evidence in a way that ignores the context-dependence of evidence. A piece of evidence which, on its own, is very poor evidence for causality, might be a keystone of a larger structure that makes a very strong case for causality.

This is illustrated by the original rejection of Fisher’s constitutional hypothesis that a tendency to smoke was ‘constitutionally’ linked to the tendency to develop lung cancer. This hypothesis can only be refuted by a long-term randomized trial of smoking starting from youth until middle and older age. The closest that analytical epidemiology might come is to study monozygotic twins discordant for smoking. Such a study was eventually conducted, and found an excess of lung cancer in the smokers, but came much too late (1996) to influence the debate.33 The actual rejection of the constitutional hypothesis in the 1950s hinged upon time trend-data.34^,35 The incidence of lung cancer had increased over only a few decades. Smoking became widespread in the first half of the 20th century. According to the constitutional hypothesis, people with a tendency to develop lung cancer would have taken up smoking, but that would not have changed the prevalence of the ‘linked’ genetic variant that increases the risk of lung cancer. Moreover, it is inconceivable that a genetic mutation would have increased that much in prevalence in a few decades—so the population incidence of lung cancer should have remained stable. The conclusion was that a new environmental factor that causes lung cancer must have been introduced.

Time-trend data are in themselves extremely weak evidence for causality. They certainly do not conform to any RCT-like emulation of a hypothetical intervention. However, in this situation, they were an essential part of the evidence for the smoking hypothesis because they were very good evidence against a competing explanation, the constitutional hypothesis. They provided an important piece of the ‘crossword’ of overlapping evidence.36 This example illustrates the dangers of crude attempts to ‘rank’ evidence: the value of evidence for assessing causality is context dependent. The RPOA makes no provision for this.

The importance of ruling out alternatives

One central scientific mode of argument for a given hypothesis is to identify the most plausible alternative hypotheses, and then find evidence which rules them out. This idea is well explored by philosophers (e.g. Karl Popper). Peter Lipton’s framework of inference to the best explanation places the ruling out of competing hypotheses at the centre of scientific inference.37 Similarly, Alex Broadbent’s model of causal inference and prediction in epidemiology emphasizes ruling out alternative hypotheses so as to arrive at ‘stable’ results.38^,39

Epidemiologists have also pointed out that ruling out alternative hypotheses is an important way to assess a hypothesis. In 1959, Cornfield stated:

If important alternative hypotheses are compatible with available evidence, then the question is unsettled, even if the evidence is experimental. But, if only one hypothesis can explain all the evidence, then the question is settled, even if the evidence is observational.40

Similar reasoning was used by Austin Bradford Hill1 in his nine causal inference considerations. According to Hill, one underlying idea was pivotal:

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. What they can do, with greater or less strength, is to help us to make up our minds on the fundamental question—is there any other way of explaining the set of facts before us, is there any other answer equally, or more, likely than cause and effect?1

The value of time-trend data in the example just discussed is precisely that it rules out alternatives. The RPOA makes no provision for this.

The importance of triangulation, negative controls and interlocking evidence

The RPOA provides no model for overcoming what is arguably the central challenge of epidemiology, namely the challenge of using different kinds of evidence to arrive at one overall verdict.

One time-honoured strategy, both within and outside epidemiology, is triangulation: one’s confidence in a finding increases if different data, investigators, theoretical approaches and methods all converge on that finding.41 For example, when the same association holds in an analysis with a propensity score and with an instrumental variable analysis that is subject to very different assumptions, the potential causality of the association is strongly bolstered.42 In contrast, the RPOA focuses on individual study design and does not account for the power of triangulation, nor guide epidemiologists seeking to implement this approach.

Negative controls can help to assess and quantify Cornfield’s ‘competing explanations’. The two main types of negative controls are: exposure controls and outcome controls.43 An interesting example was the finding of an influence of smoking habits on pregnancy outcomes. It might be argued that this might be due to other characteristics of smoking pregnant women. However, smoking by fathers had very little relation to the same pregnancy outcomes, which strongly points to an effect of smoking and not of putative other characteristics.44 Even if studies that use negative controls are constructed around RPOA principles, the idea of using negative controls, and which negative controls are most valuable, does not arise from RPOA principles but from background knowledge; furthermore, the value of also considering evidence from such negative control studies is not covered by RPOA principles.

Interlocking of evidence happens when epidemiology draws on evidence from other sciences. In the 1959 paper on smoking and lung cancer, the authors did not only discuss epidemiological data; they examined other evidence: pathology (carcinoma in situ and epithelial dysfunction in lungs of smokers), animal experiments (high doses of tobacco-tar on skin), human observations of tar as a carcinogen (chimney sweeps).34^,35 Laboratory evidence provides a useful complement to epidemiological studies, and is an important part of the overall evidence about causation. It is illustrative to discuss one biochemical explanation of why smoking causes lung cancer: it was found invitro that benzo[a]pyrenes from tobacco smoke bind to mutation hotspots of a p53 suppressor gene in cultured lung cells.45 However, this does not establish causality by itself—if we had no knowledge of the epidemiology of smoking and lung cancer, there would be no point in pursuing such biochemical studies as their outcomes would be meaningless.46 There is no logical difference between a laboratory experiment of this kind that needs epidemiology as a background and an epidemiological inference based on observational data that is strengthened by such laboratory findings: in both cases, causal inference requires going beyond the data from the particular study and integrating the messages of two different fields of science into one complementary narrative.

Considering the huge field of possible evidence that might be relevant to the assessment of a causal hypothesis, it is an illusion that one might solve the problem of causality by methods alone. The philosopher Susan Haack uses the analogy of a crossword puzzle to describe the idea of integrating particular findings from diverse disciplines with existing knowledge.36 Triangulation and interlocking of evidence are eminently practical: e.g. in the assessment of carcinogens by the IARC, combining animal studies, mechanistic reasoning and diverse kinds of epidemiology.47

Do these approaches fit with the RPOA?

We have argued that the importance of evidence depends upon context, and that ruling out alternatives is a central method of causal inference. By using triangulation (e.g. negative controls) and interlocking evidence, observational studies can help us make very strong causal inferences by ruling out alternative explanations, even where those studies do not emulate RCTs and do not clearly specify a notional intervention. The RPOA focuses on particular study designs and analyses as crucial to causal inference. It cannot explain the value of ‘bad’ evidence in the total picture; it does not provide a way of constructing triangulation or interlocking frameworks of evidence, nor an explanation of how such frameworks can be so powerful or the importance of ruling out alternatives.

Thus, the RPOA provides a view of causal inference that is inadequate to both the practice and the theory of causal inference in epidemiology. In practice, the RPOA promotes an unwarranted restriction of the type of evidence that is ‘acceptable’, and hence a restriction of the type of questions that epidemiologists may ask.48

We are indebted to extensive discussions and comments on earlier versions by Victor Gijsbers, Patrick Bossuyt, Maria Glymour, Clark Glymour, Madelyn Glymour, Sharon Schwartz, Miguel Hernán, Nancy Krieger, Sander Greenland, Tyler VanderWeele, Ben Smart, Judea Pearl and several reviewers. The Centre for Global NCDs is supported by the Wellcome Trust Institutional Strategic Support Fund, 097834/Z/11/B.

Conflicts of interest: The authors declare no financial conflict of interest.

2. Bird A.The epistemological function of Hill’s criteria. Prev Med 2011;53:242–45. [PubMed] [Google Scholar]

4. Hernan MA.Invited commentary: hypothetical interventions to define causal effects – afterthought or prerequisite? Am J Epidemiol 2005;162:618–20; discussion 621–12. [PubMed] [Google Scholar]

5. Hernan MA, Taubman SL. Does obesity shorten life? The importance of well-defined interventions to answer causal questions. Int J Obes (Lond) 2008;32Suppl 3):S8–14. [PubMed] [Google Scholar]

6. Rothman KJ, Greenland S, Lash TL. Modern Epidemiology. 3rd ednPhiladelphia, PA: Wolters Kluwer, Lippincott, Williams & Wilkins, 2008. [Google Scholar]

7. VanderWeele TJ, Hernán MA. Causal effects and natural laws: Towards a conceptualization of causal counterfactuals for nonmanipulable exposures, with application to the effects of race and sex. In: Berzuini C, Dawid A, Bernardinelli L (eds). Causality: Statistical Perspective and Applications. Hoboken, NJ: John Wiley, 2012, p. 101–13. [Google Scholar]

8. VanderWeele T.Explanation in causal inference: methods for mediation and interaction. New York, NY: Oxford University Press, 2015. [Google Scholar]

9. Halpern JY, Pearl J. Causes and Explanations: A Structural Models Approach. Part 1: Causes. Br J Philos Sci 2005;56:843–87. [Google Scholar]

10. Pearl J.Causality: Models, Reasoning and Inference. 2nd ednCambridge, UK: Cambridge University Press, 2009. [Google Scholar]

11. Bollen KA, Pearl J In: Morgan SL. (ed). Eight Myths about Causality and Structural Equation Models. New York, NY: Springer, 2013:301–28. [Google Scholar]

12. Lilienfeld AM.Epidemiological methods and inferences in studies of non-infectious diseases. Public Health Rep 1957;72:51–60. [PMC free article] [PubMed] [Google Scholar]

13. Lewis D.Counterfactuals. Cambridge, MA: Harvard University Press, 1973. [Google Scholar]

14. Lewis D.Counterfactuals and Comparative Possibility. J Philos Log 1973;2:418–46. [Google Scholar]

15. Lewis D.Causation. J Philos 1973;70:556–67. [Google Scholar]

16. Lewis D.Counterfactual Dependence and Time’s Arrow. Noûs 1979;13:455–76. [Google Scholar]

17. Rubin DB.Estimating causal effects of treatments in randomized and nonrandomized studies. J Educ Psychol 1974;56:688–701. [Google Scholar]

18. Spirtes P, Glymour C, Scheines R. Causation, Prediction, and Search. New York, NY: Springer, 1993. [Google Scholar]

19. Pearl J.Causality: Models, Reasoning and Inference. 1st ednCambridge, UK: Cambridge University Press, 2000. [Google Scholar]

20. Robins JM, Hernan MA, Brumback B. Marginal structural models and causal inference in epidemiology. Epidemiology 2000;11:550–60. [PubMed] [Google Scholar]

21. Hernan MA, Brumback B, Robins JM. Marginal structural models to estimate the causal effect of zidovudine on the survival of HIV-positive men. Epidemiology 2000;11:561–70. [PubMed] [Google Scholar]

22. Lewis D.On the Plurality of Worlds. Hoboken, NJ: Wiley, 2001. [Google Scholar]

23. Greenland S, Robins JM, Pearl J. Confounding and collapsibility in causal inference. Stat Sci 1999;14:29–46. [Google Scholar]

24. Beebee H, Hitchco*ck C, Menzies P (eds). The Oxford Handbook of Causation. Oxford, UK: Oxford University Press, 2009. [Google Scholar]

25. Schaffer J.Contrastive Causation. Philos Rev 2005;114:297–28. [Google Scholar]

26. Woodward J.Making Things Happen: A Theory of Causal Explanation. Oxford, UK: Oxford Unviversity Press, 2003. [Google Scholar]

27. Greenland S.Epidemiologic measures and policy formulation: lessons from potential outcomes. Emerg Themes Epidemiol 2005;2:5. [PMC free article] [PubMed] [Google Scholar]

28. VanderWeele TJ, Robinson WR. On the causal interpretation of race in regressions adjusting for confounding and mediating variables. Epidemiology 2014;25:473–84. [PMC free article] [PubMed] [Google Scholar]

30. Stamler J, Neaton JD. The Multiple Risk Factor Intervention Trial (MRFIT) – importance then and now. JAMA 2008;300:1343–45. [PubMed] [Google Scholar]

31. Broadbent A.Causation and Prediction in Epidemiology: A guide to the ‘Methodological Revolution’. Stud Hist Philos Biol Biomed Sci 2015;54:72–80. [PubMed] [Google Scholar]

32. Staessen JA, Wang JG, Thijs L. Cardiovascular protection and blood pressure reduction: a meta-analysis. Lancet 2001;358:1305–15. [PubMed] [Google Scholar]

33. Carmelli D, Page WF. Twenty-four year mortality in World War II US male veteran twins discordant for cigarette smoking. Int J Epidemiol 1996;25:554–59. [PubMed] [Google Scholar]

34. Cornfield J, Haenszel W, Hammond EC, Lilienfeld AM, Shimkin MB, Wynder EL. Smoking and lung cancer: recent evidence and a discussion of some questions. J Natl Cancer Inst 1959;22:173–203. [PubMed] [Google Scholar]

35. Vandenbroucke JP.Commentary: ‘Smoking and lung cancer’ – the embryogenesis of modern epidemiology. Int J Epidemiol 2009;38:1193–96. [PubMed] [Google Scholar]

36. Haack S.Manifesto of a passionate moderate. Chicago, IL: Chicago University Press, 1998, p. 95. [Google Scholar]

37. Lipton P.Evidence to the Best Explanation. 2nd ednLondon and New York: Routledge, 2004. [Google Scholar]

38. Broadbent A.Philosophy of Epidemiology. Basingstoke, UK: Palgrave Macmillan, 2013. [Google Scholar]

39. Broadbent A.Stability as a Stable Category in Medical Epistemology. J Appl Philos 2015. In press. [Google Scholar]

40. Cornfield J.Principles of research. Am J Ment Defic 1959;64:240–52. [PubMed] [Google Scholar]

42. Schneeweiss S, Seeger JD, Landon J, Walker AM. Aprotinin during coronary-artery bypass grafting and risk of death. N Engl J Med 2008;358:771–83. [PubMed] [Google Scholar]

43. Lipsitch M, Tchetgen Tchetgen E, Cohen T. Negative controls: a tool for detecting confounding and bias in observational studies. Epidemiology 2010;21:383–88. [PMC free article] [PubMed] [Google Scholar]

44. Howe LD, Matijasevich A, Tilling K et al.. Maternal smoking during pregnancy and offspring trajectories of height and adiposity: comparing maternal and paternal associations. Int J Epidemiol 2012;41:722–32. [PMC free article] [PubMed] [Google Scholar]

45. Denissenko MF, Pao A, Tang M, Pfeifer GP. Preferential formation of benzo[a]pyrene adducts at lung cancer mutational hotspots in P53. Science 1996;274:430–32. [PubMed] [Google Scholar]

46. Vandenbroucke JP.175th anniversary lecture. Medical journals and the shaping of medical knowledge. Lancet 1998;352:2001–06. [PubMed] [Google Scholar]

47. Pearce NE, Blair A, Vineis P et al.. IARC Monographs: 40 Years of Evaluating Carcinogenic Hazards to Humans. Environ Health Perspect 2015. doi: 10.1289/ehp.1409149 [PMC free article] [PubMed] [Google Scholar]

48. Pearce N.Traditional epidemiology, modern epidemiology, and public health. Am J Public Health 1996;86:678–83. [PMC free article] [PubMed] [Google Scholar]

49. Glymour C, Glymour MR. Commentary: race and sex are causes. Epidemiology 2014;25:488–90. [PubMed] [Google Scholar]

50. Hall N.Two concepts of causation. In: Collins J, Hall N, Paul LA (eds). Causation and Counterfactuals. Cambridge, MA: MIT Press, 2004, p. 225–76. [Google Scholar]

51. Godfrey-Smith P.Causal Pluralism. In: Beebee H, Hitchco*ck C, Menzies P (eds). The Oxford Handbook of Causation. Oxford, UK: Oxford University Press, 2009, p. 326–40. [Google Scholar]

52. Pearl J.Causal inference in statistics: An overview. Stat Surv 2009;3:96–146. [Google Scholar]

53. Dawid P.The Decision Theoretic Approach to Causal Inference. In: Berzuini C, Dawid P, Bernardinelli L (eds). Causality: Statistical Perspectives and Applications. Hoboken, NJ: Wiley, 2012, p. 25–42. [Google Scholar]

54. Greenland S.Causal inference as a prediction problem. In: Berzuini C, Dawid P, Bernardinelli L (eds). Causality, Statistical Perspectives and Applications. Hoboken, NJ: Wiley, 2012, p. 44–52. [Google Scholar]

55. Vandenbroucke JP, Pearce N. POINT: Incident exposures, prevalent exposures, and causal inference: Limiting studies to persons who are followed from first exposure onwards may damage epidemiology. Am J Epidemiol 2015;182:826–33. [PMC free article] [PubMed] [Google Scholar]