View - International Journal of Epidemiology

© International Epidemiological Association 2001

International Journal of Epidemiology 2001;30:707–709

Printed in Great Britain

Commentary: Positivized epidemiology and the model of sufficient and component causes Charles Poole

The dialogue form obliges the reader to consider the discussants’ relationships to each other and to the author. In Karhausen’s dialogue on Rothman’s model of sufficient and component causes (SCC) (e.g. ref. 1, pp.7–17) in this issue of the International Journal of Epidemiology2 these relationships are complicated by the fact that the participants—Socrates and Epimenide—are historical figures to whom historically questionable statements have long been attributed. Socrates’ role in Plato’s dialogues is a matter of perpetual uncertainty. Russell found it ‘very hard to judge how far Plato means to portray the historical Socrates, and how far he intends the person called “Socrates” in his dialogues to be merely the mouthpiece of his own opinions’.3 We have no such difficulty with Karhausen’s Socrates.2 He speaks for Karhausen as unambiguously as Salviati spoke for Galileo in another dialogue patterned on the Platonic form.4 Rothman has written favourably on epidemiological applications of some of Popper’s ideas (e.g. ref. 1, pp.18–19), but Karhausen abhors ‘Popperian epidemiology’.5 So Karhausen’s Socrates,2 a crafty argumentarian and his master’s faithful servant, takes every opportunity to turn Popper against Rothman. Karhausen employs Epimenide roughly as Galileo employed Salviati’s foil Simplicio: as a dimwit barely familiar enough with established dogma to give a garbled account of some of its more obvious features. This imbalance makes Karhausen’s dialogue read much as Feyerabend’s second dialogue on knowledge read to Feyerabend: ‘not really a dialogue but a diatribe directed at a helpless victim’.6 The historical Epimenide was a Cretan alleged to have said that all Cretans are liars. If a liar is one who never tells the truth, he had to be lying and at least one Cretan had to tell at least one truth. Ironically, as we shall see, the less than honest discussant in Karhausen’s dialogue2 is not Epimenide, but Socrates.

unable to distinguish among causes that help each other bring about disease, causes that have causal connections to each other, and causes that are related to each other in both ways.7,8 One easily remedied limitation that Socrates identifies in the SCC model is its lack of a requirement that an explicit alternative or ‘control’ be specified for each component cause.2 This ambiguity is shared by directed acyclic graphs (DAG),9 which traditionally trained epidemiologists recognise as elaborate formalizations and extensions of the old-fashioned ‘confounding triangle’. Directed acyclic graphs also display only variable names and do not specify particular contrasts between values or levels of those variables. This same ambiguity is also shared by common parlance: ‘Air bags kill!’ ‘Rare meat cuts cancer risk!’ ‘New treatment has no effect!’ Depending on the answer to the question, ‘Compared with what?’ each of these claims may be reasonable (rare meat cf. well-done meat) or dubious (rare meat cf. tofu). More specific labelling of the slices fixes this problem in piechart depictions of sufficient causes (Figure 1). Each chart represents a class of aetiological mechanisms. For instance, suppose the outcome is disease remission, X = 1 is an experimental therapy and X = 0 is the standard therapy. The first pie chart, with a slice labelled ‘X = 1 cf. X = 0’, represents all favourable responses to the experimental therapy, relative to the standard one. The second chart, with a slice labelled ‘X = 0 cf. X = 1’, represents favourable responses to the standard therapy, relative to the experimental one. The third chart, with no labelled slices, represents spontaneous remission, with specific regard to these two treatments. A patient whose disease would remit given either of these two therapies could be

A flawed model among many Karhausen’s characters discuss the SCC model as though it were the only causal model in epidemiology.2 They seem blissfully unaware of others, especially counterfactual models (ref. 1, pp.60–62),7–9 whose developers are well on their way to discerning exactly what is needed for association to equal causation. Like the SCC model, all these causal models have limitations. They must. They are only models. Some of the SCC model’s limitations are inherent. As a model for aetiological mechanisms, it is too ‘deep’ to serve as a model for individual or group risk.7 The model is also inherently

Department of Epidemiology (CB 7400), University of North Carolina School of Public Health, Chapel Hill, NC 27599–7400, USA. E-mail: [email protected]

Figure 1 Modified pie-chart depiction of all hypothetically possible classes of sufficient causes (etiologic mechanisms) of an outcome with regard to a well-specified index condition (X = 1) and reference condition (X = 0). Each label states the specific causal contrast postulated by the hypothetical class of sufficient causes. Unlabelled slices represent known or hypothesized component causes that are unspecified in this particular analysis, as well as unknown component causes that might be discovered in future research. If X = 1 is the presence of an air bag, X = 0 is its absence, and the outcome is death in an automobile collision, the first pie chart represents mechanisms in which ‘air bags kill’, the second represents mechanisms in which ‘air bags save lives’, and the third represents fatal etiologies in which air bags, by their presence or absence, play no role

707

708

INTERNATIONAL JOURNAL OF EPIDEMIOLOGY

susceptible to a mechanism of spontaneous remission (represented by the third chart) or to the favourable effects of both therapies (the first and second charts). If the definition of X = 0 were changed (unethically) from the standard treatment to a placebo, a given population of patients might contain many more people susceptible to mechanisms represented by the first chart and few, if any, susceptible to mechanisms represented by the other two. The other remediable limitation noted by Socrates is that the SCC model does not confine component causes to alterable conditions.2 Unalterable conditions, such as genotypicallydefined gender and the cancer experience of one’s grandparents, could be declared ineligible for component-cause duty and placed into alternative service as (often crucial) elements of context or domain.10

Determinism One of the two remaining counts of Socrates’ indictment of the SCC model is that it is completely deterministic.2 Rothman does portray the model as ‘reinterpreting chance as deterministic events beyond the current limits of knowledge or observability’.1, p.16 Interestingly, this view logically leads to the adoption of a subjective interpretation of probability as ignorance or uncertainty.11,pp.205–07 Rothman allows stochastic elements into the model only for the purpose of ‘treating unmeasured component causes (i.e. those represented by unlabeled slices in the pie charts in the figure) in the model as random events’.1,p.16 In my view, this concession does not go far enough. I see no problem in defining an instance of radioactive decay or any other event that is unpredictable, in practice or even in principle, as a component cause. Socrates infers, from Rothman’s description of the SCC model as deterministic, a claim that all disease comes about solely by the operation of ‘immutable laws’.2 I find no explicit statement of such a claim in any of Rothman’s expositions of this model. The essay of mine that Socrates cites on this point12 does not so much as allude to the SCC model or to immutable laws. Socrates2 invokes Popper’s indeterminism against Rothman’s determinism as though the former were complete determin-ism’s polar opposite: the view that no events have any causes. But Popper’s indeterminism is not nearly that extreme. It is an intermediate or mixed view: that the past is completely determined, or fixed, and that the future is partly fixed and partly open.13 To apply this view to the epidemiological concept of individual risk, one might say that all past risks are either 0 or 1 (e.g. my past mononucleosis risk of 1), that some future risks are 0 or 1 (e.g. my lifetime future spina bifida risk of 0), and that the remaining future risks lie between 0 and 1 (e.g. my 10-year future colon cancer risk of 10–3 to 10–2).14 The SCC model is compatible with this mixed view. The model can be interpreted as holding merely that the occurrence of an incident case of a disease is synonymous with the falling into place of the last part (last component cause) of an aetiological mechanism (sufficient cause) for that disease. When Socrates speaks of tobacco smoking displaying a ‘tendency towards sufficiency’ in cohort studies and a ‘tendency towards necessity’ in case-control studies of an unspecified outcome (lung cancer, for instance),2 he uses the words ‘sufficiency’ and ‘necessity’ in ways far removed from their use in the SCC

model. Tobacco smoking would never be entertained in the terms of that model as a hypothetical sufficient cause or aetiology unto itself. Component causes are held by that model to be necessary for the completion of the sufficient causes in which they participate, but experiencing any component cause (other than the last one) does not necessitate that any sufficient cause involving that component will in fact be completed. As a pragmatic matter, by the time we come to analyse the data from any epidemiological study, all the individual risks are past risks and therefore measurable only as 0 or 1. Study of those past risks of 0 and 1 can illuminate future risks and help us reduce them (and uncertainty about them), even if most of those future risks lie between 0 and 1.

Metaphysics In levying his charge of metaphysics against the SCC model, Karhausen’s Socrates contends that no one has seen a sufficient cause as ‘an object’ and that Popper, an enemy of metaphysics, expelled causality from science for this reason.2 Here Socrates relies on the knee-jerk repugnance that is elicited so easily from contemporary scientists by branding any notion ‘metaphysical’. This cultural bias comes not from Popper, however. It is a vestigial influence of logical positivism, which Popper opposed. The positivist programme was an attempt to show that all metaphysical statements are meaningless, that only scientific statements are meaningful, and that the only scientific statements are those that empirical observations can verify (in early positivist doctrine) or confirm (in later versions). The positivist ship ran aground on the fallibility of sense perception. It also ran afoul of realizations by working scientists that all observations are theory-laden and that scientific theories, as a rule, contain unobservable elements. The positivist leader Mach threatened to become an exphysicist rather than accept ‘the reality of atoms’, which had never been seen.15 But when Heisenberg objected to Einstein that the latter was making use of observations giving only indirect evidence of electron paths, and reminded Einstein that he had rejected the use of unobservable quantities in developing his special theory of relativity, Einstein replied, ‘Perhaps I did use such philosophy earlier, and also wrote it, but it is nonsense all the same’.16 Weinberg notes that modern physicists have no difficulty entertaining the notion that a meson is composed of one quark and one antiquark, despite the extremely dim prospect of ever splitting a meson and observing these two component parts.17 It follows that the inability to see (or to be certain that one has seen) a sufficient cause in its entirety, with all aetiologically extraneous conditions and events stripped away, is no impediment to employing the SCC model (among others) in our own branch of science. And what of Popper, the purported enemy of metaphysics? He complained bitterly, but to little avail, about being branded a positivist and he even took credit for positivism’s demise. The issue upon which he disagreed most sharply with the positivists was the demarcation between science and metaphysics.11,pp.35–38 ‘They (the positivists) are constantly trying to prove that metaphysics by its very nature is nothing but nonsensical twaddle .... In contrast to these anti-metaphysical stratagems

AGAINST POSITIVIZED EPIDEMIOLOGY

—anti-metaphysical in intention, that is—my business, as I see it, is not to bring about the overthrow of metaphysics. It is, rather, to formulate a suitable characterization of empirical science, or to define the concepts of “empirical science” and “metaphysics” in such a way that we shall be able to say of a given system of statements whether or not its closer study is the concern of empirical science ... I do not even go so far as to assert that metaphysics has no value for empirical science. For it cannot be denied that along with metaphysical ideas which have obstructed the advance of science there have been others—such as speculative atomism—which have aided it.’ The ‘“principle of causality”’ that Karhausen’s Socrates tells us Popper proposed to ‘exclude as “metaphysical” from the sphere of science’ is ‘... that any event whatsoever can be be causally explained—that it can be deductively predicted’.11,p.61 But immediately after those words Popper wrote these: ‘I shall, however, propose a methodological rule which corresponds so closely to the “principle of causality” that the latter might be regarded as its metaphysical version. It is the simple rule that we are not to abandon the search for universal laws and for a coherent theoretical system, nor ever give up our attempts to explain causally any kind of event we can describe.’ Karhausen’s Socrates2 is a crafty argumentarian indeed. He withholds evidence of how unopposed Popper was to metaphysics, of how well the SCC model serves his pragmatic proposal for scientists to agree to keep searching for causes, and of the distinction between searching for causes and searching for universal laws. The SCC model gives epidemiologists engaged in aetiological research on any disease a clear choice between two options at any point in time: 1. Consider all remaining variability in the disease’s occurrence, conditional on its known determinants, to be due to luck18 or some other source of irreducible stochastic uncertainty, and close up shop. 2. Keep searching for additional determinants. One authority on cancer epidemiology very recently took option 1, declaring the search for cancer risk factors to be over.19 For any disease or other health outcome, a way of emphasizing a working agreement on option 2 is to include unlabelled slices in pie-chart depictions of sufficient causes (Figure 1). Rothman temporarily fell into the positivist trap of interpreting ‘metaphysical’ as an epithet. He needlessly and, in my view, unsuccessfully tried to defend the SCC model as empirically falsifiable.1 The model does make the prohibitions Rothman claimed it makes, but none of them rules out any empirical observation. For example, the SCC model (like most others) prohibits effects from occurring before their causes; but it does not forbid us from observing a person who develops gastric bleeding and then takes aspirin for a headache, however inadvisedly. A better answer to the charge that the SCC model is metaphysical is this: Of course it is. And so are other causal

709

models and a great many other helpful, meaningful, guiding notions in science.

Conclusion It is long past time to move past the opposition to metaphysics that remains from positivist brainwashing. We need metaphysical notions, including causal models, to help guide our research. We should not become too devoted to any one model, for they serve different purposes and operate at different levels. We should continue to examine critically these models and all other guiding metaphysical notions, to revise them, and to develop new ones as needed. But let us dispense with the mistaken view that they are dispensable. Without metaphysics there is no physics, and no epidemiology either.

References 1 Rothman

KJ, Greenland S. Modern Epidemiology. 2nd Edn. Philadelphia: Lippincott Williams & Wilkins, 1998.

2 Karhausen L. Int J Epidemiol 2001;30:704–06. 3 Russell B. A History of Western Philosophy. New York: Simon & Schuster,

1945, p.84. 4 Galileo G. Dialogues Concerning the Two Chief World Systems, 1632. Trans.

S Drake, 2nd Edn, Berkeley: University of California Press, 1967. 5 Karhausen L. The poverty of Popperian epidemiology. Int J Epidemiol

1995;24:869–74. 6 Feyerabend P. Three Dialogues on Knowledge. Oxford: Basil Blackwell

Ltd, 1991, p.165. 7 Greenland S, Poole C. Invariants and noninvariants in the concept of

interdependent effects. Scand J Work Environ Health 1988;14:125–29. 8 Robins JM, Greenland S. Identifiability and exchangeability for direct

and indirect effects. Epidemiology 1992;3:143–55. 9 Greenland S, Pearl J, Robins JM. Causal diagrams for epidemiologic

research. Epidemiology 1999;10:37–48. 10 Miettinen OS. Theoretical Epidemiology. New York: John Wiley & Sons,

1985, pp.44–45. 11 Popper KR. The Logic of Scientific Discovery. New York: Harper

Torchbooks, 1959. 12 Poole C. Induction does not exist in epidemiology, either. In:

Rothman KJ (ed.). Causal Inference. Chestnut Hill, MA: Epidemiology Resources Inc., 1988, pp.153–62. 13 Popper KR. The Open Universe: An Argument for Indeterminism. London:

Routledge, 1988, pp.55–77. 14 Colditz GA, Atwood KA, Emmons K et al. Harvard Report on Cancer

Prevention. Vol. 4: Harvard Cancer Risk Index. Cancer Causes Control 2000;11:477–88. 15 Mach E. Physikalische zeitschrift 1910;11:603, trans. J Blackmore,

Br J Philos Sci 1989;40:524. 16 Heisenberg W. Encounters with Einstein, and Other Essays on People, Places

and Particles. Princeton: Princeton University Press, 1983, p.114. 17 Weinberg S. Dreams of a Final Theory: The Scientist’s Search for the

Ultimate Laws of Nature. New York: Vintage Books, 1994, pp.182–84. 18 Peto R. Cancer risk. New Scientist 1977;73:480–81. 19 Colditz G. Cancer culture: Epidemics, human behavior, and the

dubious search for new risk factors. Am J Public Health 2001; 91:357–64.