Causal criteria and the problem of complex causation
- First Online:
- Cite this article as:
- Ward, A. Med Health Care and Philos (2009) 12: 333. doi:10.1007/s11019-009-9182-2
- 303 Views
Nancy Cartwright begins her recent book, Hunting Causes and Using Them, by noting that while a few years ago real causal claims were in dispute, nowadays “causality is back, and with a vengeance.” In the case of the social sciences, Keith Morrison writes that “Social science asks ‘why?’. Detecting causality or its corollary—prediction—is the jewel in the crown of social science research.” With respect to the health sciences, Judea Pearl writes that the “research questions that motivate most studies in the health sciences are causal in nature.” However, not all data used by people interested in making causal claims come from experiments that use random assignment to control and treatment groups. Indeed, much research in the social and health science depends on non-experimental, observational data. Thus, one of the most important problems in the social and health sciences concerns making warranted causal claims using non-experimental, observational data; viz., “Can observational data be used to make etiological inferences leading to warranted causal claims?” This paper examines one method of warranting causal claims that is especially widespread in epidemiology and the health sciences generally—the use of causal criteria. It is argued that cases of complex causation generally, and redundant causation—both causal overdetermination and causal preemption—specifically, undermine the use of such criteria to warrant causal claims.
KeywordsBradford HillCausal criteriaCausal overdeterminationCausalityLate causal preemptionRedundant causationWarranted causal claims
Nancy Cartwright begins her recent, 2007 book, Hunting Causes and Using Them, by noting that while a few years ago real causal claims were in dispute, nowadays “causality is back, and with a vengeance” (2007b, p. 1). In the case of the social sciences, Keith Morrison (2008, p. 1) writes that “Social science asks ‘why?’. Detecting causality or its corollary—prediction—is the jewel in the crown of social science research.” With respect to the health sciences, Judea Pearl (2001, p. 189) writes that the “research questions that motivate most studies in the health sciences are causal in nature.” In a more immediately practical vein, Sarah et al. (2008, pp. 340–342) argue that understanding how upstream social characteristics such as poverty and neighborhood crime are causally related to downstream characteristics such as health status is crucial in the creation of sound health policy. (Also, see Weed 2005, p. 112).
Often people base claims about causal relations on applications of statistical methods to data. However, the source and type of data place important constraints on the choice of statistical methods as well as on the warrant attributed to the causal claims based on the use of such methods. (See Rosenbaum 1989, pp. 169–170.) Even in those cases in which there are good reasons to believe that an association between a purported cause and effect is not an artifact of selection or measurement bias, an ever-present danger is that the association is the result of confounding and not a genuine cause-and-effect relation. In order to address this possibility, many users of statistical methods, especially those of the Neyman-Pearson or Fisher type (See Reiter 2000, p. 24; Suppes 1982, p. 465), claim that random assignment to the exposure of interest is a necessary condition for warranted causal claims. One of the principal virtues of randomization, according to the traditional view, is that its use, in principle, assures us that variation due to both measured and unmeasured covariates is random and, because of this, not a source of error when making causal inferences. (See Greenland 1990, pp. 421–424; Reiter 2000, pp. 26–27.) As noted by Shaffer and Johnson (2008, p. 4), ideally, randomization “reduces the chance of confounding with other variables.” (Also, see Freedman 1999, p. 244.) This contention is the centerpiece of the widely held belief that randomized clinical trials (RCTs) are, and ought to be, the “gold standard” of evaluating the causal efficacy of interventions (treatments). (See Cobb and Moore 1997; DeMets 2002.)
However, even if we assume that randomization is a “panacea for all problems”, an assumption that is clearly false (See Hernán 2004, p. 268; Jager et al. 2007, pp. 671–672; Rickles 2008; Smith 1990), not all data used by people interested in making causal claims come from experiments that use random assignment to control and treatment groups. For example, Erick Von Elm et al. (2008, p. 344) write that “Much of biomedical research is observational”, while Jan Vandenbroucke (2004, p. 1728) writes that observational studies are essential to our “knowledge about causes and pathogenesis—e.g., genetic, environmental, or infectious causes of diseases.” Thus, one of the most important problems in the social and health sciences concerns making warranted causal claims (i.e., causal claims supported by the proper application of appropriate methodologies and techniques) using non-experimental, observational data; viz., “Can observational data be used to make etiological inferences leading to warranted causal claims?” In what follows, I examine one method of warranting causal claims that is especially widespread in epidemiology and the health sciences (including health services research), and in legal cases involving toxic tort litigation (See Kever 2002; Muscat and Huncharek 1989, pp. 997–100; Thompson 1992–1993)—the use of causal criteria. I argue that cases of complex causation generally, and redundant causation (See Lewis 1986, pp. 193ff; McDermott 1995)—both causal overdetermination and causal preemption—specifically, undermine the use of such criteria to warrant causal claims.
Causal criteria—the case of Austin Bradford Hill
Because of the limitations of purely statistical methods, the Report goes on to state (1964, p. 20) that to “judge or evaluate the causal significance of the association between the attribute or agent and disease, or effect upon health, a number of criteria must be utilized, no one of which is an all-sufficient basis for judgment.” The criteria used in the 1964 Report, and repeated in the 1982 Surgeon General’s report The Health Consequences of Smoking (1982, pp. 17ff), were the consistency, strength, specificity, temporal relationship, and coherence of the association.
Statistical methods cannot establish proof of a causal relationship in an association. The causal significance of the association is a matter of judgment which goes beyond any statement of statistical probability.
Following the publication of the Surgeon General’s Report, Austin Bradford Hill’s (1965) Presidential Address to the Section of Occupational Medicine of the Royal Society of Medicine took these five criteria and added four others; viz., biological gradient, plausibility, experiment and analogy. Hill (1965, p. 299) denied the “criteria” were either necessary or sufficient for warranted causal claims, and described them only as “aspects of [a statistical] association” that help us “make up our minds” about whether there is a more likely explanation for the statistical association than that it is a causal relation. Thus, as noted by Alfredo Morabia (2004, p. 103), for Hill the aspects of association were not intended as necessary and sufficient conditions whose presence guaranteed that a statistical association was also a causal relation. Instead, Hill thought of the aspects of association as jointly providing a kind of “screening test” whose function was to “screen the causal statement for illogicalities or gross contradictions between what we have found and what we think we know.” If we understand Hill’s aspects of association in this way, then, at best, the presence of such aspects provides only inductive warrant for claiming that the statistical association is a causal relation. Put a bit differently, the satisfaction of the Hill criteria—the presence of the relevant aspects—inductively, but not conclusively, confirms the truth of the claim that the statistical association is a causal relation. (See Fletcher and Fletcher 2005, p. 196.) Because inductive confirmation never guarantees the truth of the conclusion for which it provides warrant (See Carnap 1995, pp. 19–28; Chalmers 1999, pp. 41–58; Hempel 1965, pp. 3–10), the presence of Hill’s aspects of association does not, on this interpretation, guarantee that a statistical association is a causal relation. At best, they provide some general guidelines for identifying statistical relations that merit further examination as plausible candidates for being causal relations. (See Kundi 2006, p. 970.)
A problem with this interpretation is that it requires two levels of metrics for the criteria. First, there need to be metrics for determining to what degree the individual criteria are satisfied (i.e., to what degree the aspects of association are present), and second, there needs to be a metric for inferring the degree of inductive support given by an application of the first metrics. To the extent to which “there is no accepted scoring system in deciding whether there is sufficient evidence of causality” (Naschitz et al. 2003, p. 13. Also, see Swaen and Amelsvoort 2008), or the metrics, at either level, are ill-defined or otherwise unsubstantiated, any inferences drawn from applications of the criteria will, at best, be questionable. Of course, it is also true that finding such metrics was not one of Hill’s goals (See Labarthe and Stallones 1988, pp. 122–125), and that he wrote (Hill’s (1965, p. 299) that no “formal test of significance can answer” questions about whether there is a more plausible explanation for the statistical association than that it is a causal relation. (Also, see Morabia 2004, pp. 72–75.) Nonetheless, apart from the historically interesting question of what Hill originally intended, the “aspects of association” identified by Hill, typically referred to as the “Bradford Hill Criteria”, are “frequently taught to students in epidemiology” as criteria whose satisfaction will guarantee that a statistical association is a causal relation (Phillips and Goodman 2004). It is for this reason that Hill’s “aspects of association” are “referred to in the literature as ‘causal criteria’ ” (Phillips and Goodman 2004. Also, see Lucas and McMichael 2005, pp. 792–794.) Indeed, Saad Shakir and Deborah Layton (2002, p. 468) write that Hill’s Presidential Address was one “of the most important papers published in the 20th century with thoughts on the epidemiological basis of disease causation”.
While scarcely mentioned in the philosophical literature, the “basic outline of the modern set of criteria has,” according to Kaufman and Poole (2000, p. 102), evolved little “since the formulations by the Surgeon General’s Advisory Committee and Bradford Hill”, and the set has “become a central tool for the epidemiologic community in grappling with the broader issues of causal reasoning.” (Also, see Egilman et al. 2003, pp. 241ff; Greenland 2004, p. 10; Karhausen 2000, p. 63; Perrio et al. 2007, p. 333; Vineis 1991, p. 174.) One can find a clear example of this in the Environmental Protection Agency’s (2005) “Guidelines for Carcinogen Risk Assessment”. The guidelines (2005, pp. 2–11ff) explicitly recommend the use of the Hill criteria to assess whether a statistical association is causal rather than spurious. Similar examples include, but are not limited to, applications of Hill’s “criteria” to determine whether chrysotile asbestos causes mesothelioma (Lemen 2004), to determine whether second generation antipsychotic drugs cause diabetes (Holt and Peveler 2006), to evaluate the (causal) effects of “environmental carcinogens” (Franco et al.2004, p. 415), and to evaluate “causal associations in pharmacovigilance as well as pharmacoepidemiology.” (Shakir and Layton 2002. Also, see Shakir 2004; Perrio et al. 2007.) Many other studies, such as an assessment of whether exposure to “Agent Orange” raises the risk for adverse reproductive outcomes (Hatch and Stein 1986, pp. 189–200) and an investigation of the connection between use of oral contraceptives and breast cancer (Schlesselman et al. 1987), use proper subsets of Hill’s “criteria” to assess causal claims.
Based on their analysis of this aspect/criterion and the others proposed by Hill, they conclude (2005, p. S149) that “the standards of epidemiologic evidence offered by Hill are saddled with reservations and exceptions.” (Also, see Rothman et al. 2008, p. 30.) In a similar vein, Karen Goodman and Carl Phillips (2004) note that there is no clear way to quantify “the degree to which each criterion is met” (which makes comparative assessments of different statistical associations very difficult at best), or how to “aggregate such results into a judgment about causation.” These kinds of criticisms have led to a large body of literature debating the meaning, application and use of specific criteria, as well as proposals for extensions and modifications of the Hill criteria. At present, though, for those who continue to support the use of causal criteria, there is no consensus about the meaning or application of the Hill criteria (See Swaen and Amelsvoort 2008), whether only some of the criteria should be used (See Weed 1988, pp. 21–23; 2004, p. 322; 2005, p. 110), or what additional criteria (if any) must be added to warrant causal claims.
Whatever insight might be derived from analogy is handicapped by the inventive imagination of scientists who can find analogies everywhere. At best, analogy provides a source of more elaborate hypotheses about the associations under study; absence of such analogies only reflects lack of imagination or experience, not falsity of the hypothesis. (Also, see Rothman et al. 2008, p. 30.)
Causal criteria and the problem of complex causation
To be fair, ignorance of appropriate causal criteria, or of the “correct” formulation of already proposed criteria, does not justify the claim that criteria warranting causal claims do not exist. Moreover, there seem to be some situations in which the satisfaction of some version of one or more causal criteria may justifiably affect, at the very least, the (subjective) credibility of a causal claim. For example, Noel Weiss (2002, p. 8) writes that because of the specificity of the association, one is warranted in believing that the “association between use of long-acting hypnotic/anxiolytic drugs by elderly persons and the incidence of hip fracture is a causal one.” Thus, rather than becoming mired in what might otherwise be an endless dialectic of criterion proposal, refutation, and modification, let us stipulate, ex hypothesi, that appropriate causal criteria do exist whose use will warrant the claim that a specific statistical association is, in reality, a causal relation. In other words, let us suppose that the joint satisfaction, by a statistical association, of the individual criteria that constitute the set of appropriate causal criteria, henceforth denoted ‘ACC’, are sufficient (though perhaps not all necessary) for making a warranted claim that the statistical association is a causal relation. Of course, a person knowing both that there is a set of appropriate causal criteria and that specific criteria constitute the set does not entail that the person either can or does know when the criteria that make up the ACC are satisfied. Thus, let us further suppose that we actually do have such knowledge about the satisfaction of the criteria that make up the ACC. Finally, given these suppositions, we can say that the claim “X causes Y” is a warranted causal claim just in case the distribution of the value(s) of X in some population is positively correlated with the distribution of the value(s) of Y in some population (See Woodward 2003, pp. 40–41), and it is known that this statistical relationship between X and Y satisfies the ACC.
To make this more concrete, it helps to consider a simple example drawn from Louis Loeb (1974, p. 526) and Martin Bunzl (1979, p. 135). Suppose that someone flips a light switch up and a light comes on in the room where the light switch is located. Here, the value of X is the flipping up of the light switch, and the value of Y is the light in the room coming on. Does flipping up the light switch cause the light in the room to come on? The answer to the question is that if the association between the flipping up of the light switch and the light coming on satisfies the ACC, then the claim that flipping the light switch caused the light to come on is a warranted causal claim. Notice that we are not saying anything about the nature of causality in using the ACC to warrant causal claims. The goal, in using the causal criteria to warrant causal claims, is not to say anything in particular about the nature of causality; the set of criteria constituting the ACC is only a tool whose (proper) use warrants the claim that a statistical association is also a causal relationship. To make this point more clear, consider the following analogy. While I may have little (or no) knowledge about the biochemical factors involved in people being sick, when I properly use a thermometer to measure a person’s body temperature, if the thermometer tells me that the person’s temperature is 104° F., and the thermometer is functioning properly, I am warranted in claiming that the person is ill. Similarly, in the case of switching the light on, it is possible that my claim that the flipping up of the light switch is the cause of the light coming on is a warranted causal claim without my knowing all the particulars of how the light switch brought this about or even how the light bulb produced the light. In both cases, the purpose of the proper use of the ACC was to warrant a causal claim about a particular statistical association, not to provide an account of causality. As John Mackie (1974, p. 1) notes, the latter, providing an account of causality, is a project in the ontology of “how the world goes on.”
Sometimes the Hill criteria are applied to singular causal claims (this is especially true in the case of toxic tort litigation), and sometimes they seem to be applied to general causal claims such as the claim that smoking causes lung cancer. Certainly, in the example about the light switch as well as the examples considered below, the focus is on singular causal statements as opposed to general causal statements. Thus, if one supposes either that the sort of causation relating tokens is different from the sort of causation relating types, or that problems with token causation are not directly relevant to type-level causal claims, then the problems of causal overdetermination and preemption considered below may seem misguided.
Singular causation is a relation between a concrete sequence of events, the individual events in the sequence are referred to as tokens … General causation signifies categories of causal relations between event types. E.g., ‘cranial factures with dural laceration leads to meningitis’ is a statement of general causality between event types (1992, p. 236). (Also, see Eells 2001; Hausman 1998, pp. 99–110; 2005, pp. 33–34; Hitchcock 1995, pp. 263–269; Hoover 2001, pp. 70–86.)
The answer to this possible objection is that it is one thing to present a theory of causation, and another thing entirely to attempt to measure a causal effect. Whereas philosophers are typically interested in providing theories of causation (including theories of type causation and token causation), epidemiologists (and others working in the health sciences) are much more interested in measuring the effects of treatments, interventions and, more generally, the distribution and determinants of disease frequencies. (See Hennekens and Buring 1987, pp. 3–4; Fletcher and Fletcher 2005, pp. 2–4.) Thus, as Rothman and Greenland (2005) write, if an “epidemiologic study sets out to assess the relation between exposure to tobacco smoke and lung cancer risk, the results can and should be framed as a measure of causal effect”. Understood in this way, singular (token-level) causal claims have a methodological priority over general (type-level) causal claims since it is the relevant set of warranted singular level (token-level) causal claims that warrant the claim that a general causal statement (a type-level causal statement such as “smoking causes lung cancer”) is true. This reflects the fact that epidemiologic studies are constituted by measurements of specific events (reflected in singular causal claims) from which general causal claims (type-level causal claims) are inferred. This approach leaves open the question of what kind of theory of causation is best (or true), as well as whether the difference between general causal claims and singular causal claims reflects a difference (such as a difference between events, properties, tropes, etc.) in types of causation. However, what the approach does suggest is that if there are cases that undermine singular causal claims (token-level causal claims), these problems will affect which general level causal claims we hold to be warranted. In the case of the Hill criteria, this means that failures at the level of singular causal claims can lead to misidentifications of warranted causal claims using the Hill criteria. Thus, in what follows it is important to keep issues (especially important to philosophical treatments of causality) of theory and ontology separate from issues of measurement (which includes the assessment that a statistical association is a causal relation).
Here X1 refers to the flipping up of the first light switch, X2 refers to the flipping up of the second light switch, and Y refers to the light in the room coming on. Given the description of the relationship of the light switches to the light, if either X1 or X2 had occurred without the other, then also Y would have occurred; that is, if either of the light switches had been flipped up without the other light switch being flipped up, then the light would still have come on. In contrast, if neither X1 nor X2 had occurred, then Y would not have occurred; that is, if neither light switch had been flipped up, then the light would not have come on.
Following David Lewis (1986, p. 193), we can call the situation above a case of redundant causation. (Also, see Mackie 1974, pp. 44–45; McDermott 1995, pp. 523–524.) In the example, both X1 and X2 are sufficient, by themselves, to cause the light to come on. However, since each is, singularly, sufficient to cause the light to come on, it follows that neither X1 nor X2 is, singularly, necessary for the light to come on. Instead, it is the disjunction of X1 and X2 (i.e., X1 or X2) that is necessary for the occurrence of Y. Thus, if the satisfaction of the ACC by the statistical association of X1 and Y warrants the claim that X1 causes Y, then the kind of causation that the satisfaction of the ACC identifies is sufficient causation (X1 is a sufficient cause of Y). Similar reasoning leads us to say that the kind of causation identified by the ACC in the case of the statistical association of X2 and Y is also sufficient causation. Notice that understanding the role of the ACC in this way does not entail an endorsement of a deterministic interpretation of sufficient cause. In particular, we can say that X1 is a sufficient cause for Y without being committed to the stronger claim that X1 is a sufficient cause of Y if and only if whenever X1 occurs Y must also occur. Instead, the claim of sufficient cause used in the example above is consistent both with a deterministic understanding of sufficient cause and with a probabilistic understanding of sufficient cause. In the latter case we can, roughly, say that given a complete description of all confounding factors (Cartwright 2007b, p. 45), and a complete specification of what Mackie (Mackie (1974, pp. 34–35) calls the “causal field”, X1 is a probabilistic (positive) sufficient cause of Y just in case the occurrence of X1 results in a greater probability of the occurrence of Y than does the non-occurrence of X1. (See, Eells and Sober 1983, pp. 35–38; Suppes 1985, p. 48.)
In order to understand the problem redundant causation poses for making warranted causal claims using causal criteria, we need to differentiate two different cases of redundant causation. If both X1 and X2 are causally on a par (e.g., if the light switches are flipped simultaneously, the resistance in the wires from the switches to the light is the same, etc.) then we have a case of symmetrical causal overdetermination. In contrast, if X1 and X2 are not causally on a par (e.g., if one light switch is flipped up before the other light switch, with all other conditions remaining the same), then we have a case of causal (late) preemption. The preemption relevant here is not “early preemption” in which the flipping of one light switch inhibits, in some fashion, the current from the other, simultaneously flipped light switch, to the light. Instead, in the example of preemption considered here, it is the fact that the effect (the light coming on) has occurred due to the flipping of the first light switch that preempts the flipping of the second light switch being the case of the light coming on. It is in this respect an example of “late preemption”. (See Ehring 1997, pp. 20–21; Hall 2004, pp. 234–241; Hitchcock 2007, pp. 516–522, 524–529.) Each kind of redundant causation poses problems for the use of causal criteria to warrant causal claims about statistical associations.
In the case of symmetrical causal overdetermination, the problem in the example is that use of the ACC does not permit us to distinguish between cases in which a known statistical association identifies the only cause of the effect from cases in which there are multiple causes of the event, some of which are unknown. In the present case, since, by hypothesis, both the X1—association—Y and the X2—association—Y satisfy the ACC, then, if we consider them separately, we are warranted in claiming that X1 is a sufficient cause of Y, and that X2 is a sufficient cause for Y. For this reason, if we knew only about the occurrence of either X1 or X2, but not both, then we would be warranted in claiming that only one of X1 or X2 is the (sufficient) cause of B (the specific claim would be relative to which of X1 or X2 it was about whose occurrence we knew). However, while either causal claim would be warranted, both would be false. The point is that in genuine cases of symmetric causal overdetermination, even if rare, no additional descriptive content of the situation would warrant the claim that it was X1 and not X2 (or, conversely, that it was X2 and not X1) that caused Y. Since the use of the ACC does not permit distinguishing between cases in which there is symmetric causal overdetermination from cases in which there is no symmetric causal overdetermination, then the most that the use of the ACC warrants when applied to a statistical association of the form X—Y is that X is a sufficient cause of Y.
This may not be terribly troublesome in the case of light switches and lights, but it is a problem in cases where we want to affect undesirable health states. For example, suppose that some population is exposed to two agents, A1 and A2, and subsequent to the exposure, there is a higher distribution of illness in the population than before the exposure. If A1 and A2 are symmetric causal overdeterminers of the illness, then there are at least three possibilities. First, we might know about the occurrence of A1 and, because it satisfies the ACC, we work to eliminate A1. Of course, because of the occurrence of A2, about which we know nothing, even a successful elimination of A1 will not end the adverse outcome. Second, we might know about the occurrence of A2 and, because it satisfies the ACC, we work to eliminate A2. However, mirroring the first case, because of the occurrence of A1, about which we know nothing, even a successful elimination of A2 will not end the adverse outcome. Inasmuch as the statistical association between A1 and the adverse outcome and the statistical association between A2 and the adverse outcome both satisfy the ACC, then the use of the ACC warrants adoption of both ineffective strategies. There is, though, a third possibility. Knowing about symmetric causal overdetermination, the strategy might be to eliminate all the known causes (e.g., A1 if this is the known case, or A2 if that is the known cause) while at the same time seeking other statistical associations that satisfy the ACC. At best though, this significantly weakens one of the principal purposes of warranted causal claims, namely their use in identifying effective strategies for change. (See Cartwright 1983, pp. 21–22.)
In the case of causal preemption, the problem is that using the ACC permits identification of the “true” sufficient cause only if one has complete knowledge of the preemptive action. In this respect, the problem is a variant of what Douglas Ehring (1997, p. 21) calls the “pairing problem” which occurs when “there is more than one mutually exclusive candidate for the role of effect or cause”. For example, suppose that one knows that switch X1 is flipped up and that the light in the room comes on, but does not know anything about light switch X2 (including its existence and its connection to the light). Moreover, suppose that while X1 occurs, its effect is preempted by the occurrence of X2. Because X1 is a sufficient cause of Y (though a preempted one), then the statistical association between X1 and Y satisfies the ACC; after all, if it did not satisfy the ACC, then X1 would not be a sufficient cause of Y and so the situation would not count as a genuine instance of causal preemption. However, since the statistical association between X1 and Y satisfies the ACC, it follows that the causal claim X1 is a sufficient cause of Y is a warranted causal claim. The problem, of course, is that it is a false causal claim. X1 is a sufficient cause of Y only if X2 does not preempt X1 being a sufficient cause of Y. The possibility of causal preemption means that we cannot rule out the possibility that our causal claims, warranted by the use of the ACC, are false claims. Of course, the obvious way around this is to say that we need to add additional criteria to the ACC, where the satisfaction of these additional criteria will rule out cases of causal preemption. To do this, however, requires that we add the additional requirement that we have knowledge about every other statistical association that exists between the effect in whose occurrence we are interested and possible causes of that effect. This, though, is in practice an impossible requirement, and as such, undermines the use of causal criteria to warrant causal claims.
Cases of complex causation are not limited to redundant causation. Suppose that the real situation with the two light switches and the light is that neither X1 nor X2 is, by itself, a sufficient cause for Y, but that only the concurrent occurrence of X1 and X2 is a sufficient cause for Y. In this case, we would suppose that the conjunction of the association of Y with X1 and the association of Y with X2 satisfied the ACC. Now, if the conjunction of the association of Y with X1 and the association of Y with X2 satisfies the ACC, then it follows that the association between each of the constituent elements of the conjunction (X1 and X2) and Y must partially satisfy the ACC. Each element of the conjunction is what Mackie (1974, p. 62) refers to as an insufficient but non-redundant part of an unnecessary but sufficient condition (INUS) for the occurrence of Y; an idea captured graphically in Rothman’s (1976, pp. 588–591) “causal pie model” of a sufficient cause (Rothman et al.2008, p. 6). In other words, while the conjunction (X1 and X2) is a sufficient (but not necessary) condition for the occurrence of Y, neither X1 nor X2 is a sufficient condition for Y, though both are required for the conjunction to be a sufficient condition for Y. In this context, it is important to notice that it cannot be the case that an association between Y and either part of the conjunction fully satisfies the ACC. If that were to happen, then whatever part of the conjunction’s association with Y fully satisfied the ACC would be the sufficient cause of Y, while the other part of the conjunction would be a spurious cause of Y.
If one knows only one constituent of a complex cause (where that constituent is an INUS element in the sense outlined above), then use of the ACC will lead one to make a false claim—viz., that the constituent element of a complex cause is itself a sufficient cause of the observed effect.
If one knows all the constituent parts of a complex cause (where each constituent is an INUS element in the sense outlined above), then use of the ACC will warrant two inconsistent claims. First, it will warrant the claim that the association of whichever constituent element of the complex most fully satisfies the ACC is the true causal relationship, while the other constituent elements of the complex cause are not themselves causes of the effect. Second, it will warrant the claim that the conjunctions of the associations of each of the constituent elements of the complex to the effect jointly constitute the causal relationship, and that each of the constituent elements are INUS causes.
Independent of what some claim to be the ambiguities, fallacies and vagaries inherent in the use of causal criteria to warrant causal claims, it seems that even idealized causal criteria have fatal flaws undermining their use. What conclusions ought we to draw from this? One possibility is to accept a variation of Bertrand Russell’s 1912 claim in his presidential address to the Aristotelian Society (1949, p. 180) and say that the word ‘cause’ is so “inextricably bound up with misleading associations” as to make its complete extrusion from the scientific vocabulary desirable.1 A second possibility is to say that no observational study provides data that warrant causal claims. If we want warranted causal claims, then we should restrict our attention to data from properly conducted randomized controlled experimental studies. However, each possible conclusion is, in its own way, too Draconian.
Although there is a high statistical correlation between smoking and lung cancer, taking an anti-cancer drug is not an effective strategy for quitting smoking, which suggests that the concept of cause plays a crucial role in distinguishing effective from ineffective strategies. Thus, the cost of expunging “causal talk” from the sciences would be to undermine the practical goals of science, as well as the hope of using the results of scientific inquiry to create beneficial policies (See Cartwright 2007a, pp. 192–193; Weed 2004) and help in making sound legal decisions. (See Egilman et al. 2003; James 1994; Mengersen et al. 2007.)
We all think that for the goal of avoiding lung cancer, it is beneficial to stop smoking. Intuitively the reason is that smoking is a cause of lung cancer. The reason is not simply that there is a high statistical correlation between smoking and lung cancer. For correlation is symmetric: if there is a high statistical correlation between smoking and lung cancer, there is a high statistical correlation between lung cancer and smoking.2
Regarding the second possibility, not only would this restrict causal claims to a very narrow range of data (excluding, for example, studies that made use of survey data), it also assumes that properly conducted randomized controlled experimental studies warrant causal claims. Even ignoring the methodological problems that can and do often occur in properly conducted randomized controlled experimental studies (e.g., selection bias), asserting that the causal claims warranted by a randomized controlled experiment extend beyond the study populations requires additional assumptions that, in many cases, are problematic at best. As Cartwright (2007b, p. 39) writes, the method of randomized controlled experiments may tell us something about causal relations in the very specific circumstances of the experiment, but “tells us nothing about what the cause does elsewhere.”
… while nature begins from causation and ends in experience, we must follow a contrary procedure, that is, begin from experience and with that discover the causes.4
Russell’s exact claim was that the word ‘cause’ should be expunged from the “philosophical vocabulary” since, in his view, physics had already expunged it from their vocabulary. However, regardless of whether Russell’s assessment of physics in 1912 was correct, as noted by Wesley Salmon (1998, p. 4), nowadays, causal concepts are ubiquitous in every branch of theoretical science, in the practical disciplines and in everyday life.
Christopher Hitchcock (2003, p. 4) makes the same point, but uses cancer, smoking and stained teeth as his example, writing that “[L]ung cancer is correlated with both smoking and with stained teeth, but if we wish to avoid lung cancer, it will pay to quit smoking but not to whiten our teeth.”
Developed for use in econometrics by the economist Clive Granger and named after him, “Granger causation” is a non-deterministic statistical account of causality that makes use of time-series data. (See Granger 1980; Holland 1986, pp. 957–958).