Causal Inference and the Language of Experimentation - Literature review Example

?On the debates pertaining to inference of an effect from its cause David Hume held that there are three conditions for inferring cause contiguity between the effect and the cause; 2) precedence of the cause in relation with the effect; and 3) constant presence of the cause whenever the effect is obtained (Cook and Campbell 1979, p. 172). Yet, Rothmand and Greenland (2005, p. S147) reported that Hume has expressed that “proof is impossible in empirical science.” Despite this, the Hume perspective has been one of the foundations of the science of prediction and forecasting (Cook and Campbell 1979, p. 172). Taking off from Hume, John Stuart Mill “held that causal inference depends on three factors: first, the cause has to precede the effects; second, the cause and effect have to be related ; and third, other explanations of the cause-effect relationship have to be eliminated” (Cook and Campbell 1979, p. 182). In other words, the notion of causation and effect that can be found in the ideas of John Stuart Mill is that causation requires precedence of the cause from the effect, correlation, and that rival hypotheses are ruled out. For Cook and Campbell (1979), however, the most significant contribution of John Stuart Mill to the theory of causality pertains to his notions of the criteria, principles, or “methods” of agreement, differences, and concomitant variation. The principle of agreement “states that an effect will be present when the cause is present” (Cook and Campbell 1979, p. 182). The principle of difference “states that the effect will be absent when the cause is absent” (Cook and Campbell 1979, p. 182). Finally, the principle of concomitant variation “implies that when both of the above relationships are observed, causal inferences will be all stronger since certain other interpretations of the co-variation between the cause effects can be ruled out” (Cook and Campbell 1979, p. 182). According to Cook and Campbell (1979, p. 183), “at its core , the Method of Concomitant Variation involves examining a somewhat complex pattern of data in order to induce whether a treatment is associated with an effect and whether some rival causal explanations can be ruled out on the basis of when the treatment and effect are and are not related.” Cook and Campbell (1979, p. 183) pointed out that “the concept of a control group is implicit here and is clearly central in Mill’s thinking about cause.” In 1913, Bertrand Russell “looked to physics and astronomy of his day as the most mature sciences, and he noted their lack of concern with unobservables and explicitness and parsimony of the functional relationships that physicists sought to test” (Cook and Campbell 1979, p. 172-173).1 However, Russell had asked that asked whether the concept of cause continues to be relevant given that cause “is not implied by functional relationships of mathematical form” (Cook and Campbell 1979, p. 173). The Russell viewpoint is positivist “rejecting unobservables (like cause), and seeking to establish explicit functional laws between continuously measured variables in a closed system” (Cook and Campbell 1979, p. 173). Positivists like Russell believe that “causation is unnecessary because it is unobservable” (Cook and Campbell 1979, p. 175). The essentialist viewpoint “argue that the term cause should only be used to refer to variables that explain a phenomenon in the sense that these variables, when taken together, are both necessary and sufficient for the effect to occur” (Cook and Campbell 1979, p. 177). The essentialists “equates cause with a constellation of variables that necessarily, inevitably and infallibly results in the effect” (Cook and Campbell 1979, p. 177). In contrast, those “who restrict cause to observable necessary and sufficient conditions (or sufficient conditions that operate when all the necessary conditions are met) reject as causes those factors which are known to bring about effects sometimes, but not always” (Cook and Campbell 1979, p. 177). For Cook and Campbell (1979, p. 184), the view articulated by Karl Popper in 1959 is among the “more contemporary.” Popper emphasised the necessity of “basing knowledge on ruling out alternative explanations of phenomena so as to remain, the researcher hopes, with only a single conceivable explanation” (Cook and Campbell 1979, p. 184). Popper’s perspective was founded “on an acceptance of Hume’s critique of induction (e.g., to say that night has always followed day does not logically justify the inductive conclusion that night will always follow day)” (Cook and Campbell 1979, p. 184). Agreeing to the critique implies “denying the possibility of confirmatory knowledge based on generalizing from particular observations to general scientific propositions” (Cook and Campbell 1979, p. 184). Despite the situation, Popper asserted that deductive knowledge remains possible: “corroboration can never prove the theory to be true, although failures to confirm the prediction can falsify the theory under test” (Cook and Campbell 1979, p. 184). Thus, for Cook and Campbell (1979), there are two streams in logical positivism: the “confirmationist” versus the falsification school initiated by Popper. According to Cook and Campbell (1979, p. 185), “both points of view assume that experimental and observational ‘facts’ can often be generated that are relevant to the validity of a theory.” For Popper’s falsification school, “corroboration gives only the comfort that the theory has been tested and has survived the test” (Cook and Campbell 1979, p. 186). Yet, “even after the most impressive corroborations of predictions, it has only achieved the status of ‘not yet disconfirmed’” (Cook and Campbell 1979, p. 186). This by itself is an achievement because the status of being “not yet disconfirmed” is rare situation even if far from being true (Cook and Campbell 1979, p. 186). Cook and Campbell (1979) emphasised that Popper’s falsification requires putting “theories in competition with each other” (Cook and Campbell 1979, p. 199). However, Cook and Campbell (1979, p. 199) criticised that “there can be no guarantee that the researcher has conceptualised all the relevant alternative theories” because the researcher can at most only conceptualise the “plausible” alternatives. It must be pointed out, however, that the perspectives of both the confirmationist and the Popperian falsification schools have been rejected by post-positivists (Cook and Campbell 1979). Cook and Campbell (1979, p. 185) noted that “observed data never exactly match a quantitative-prediction and thus they ‘fail to confirm’ or ‘falsify’ every theory and every numerically specified causal hypothesis.” In the field of epidemiology, there is a recognition that “philosophers agree that causal propositions cannot be proved, and find flaws or practical limitations in all philosophies of causal inference” (Rothman and Greenland 2005, p. S144). Thus, “the role of logic, belief, and observation in evaluating causal propositions is not settled” (Rothman 2005, p. S144). Rothman and Greenland concluded that “inference in epidemiology is better viewed as an exercise in measurement of an effect rather than as a criterion-guided process for deciding whether an effect is present or not” (Rothman and Greenland 2005, p. S144). It follows from Rothman and Greenland (2004) that effects are assumed to be based on theory and talk of causal inference revolves on the measurement of effects that were merely assumed based on theory. In other words, diseases may be associated with a certain bacteria following a theory that a disease might have been caused by that bacteria, ignoring for example the possibility that that there may be another bacteria or another life coexisting or associated with the bacteria to which a disease is attributed. Yet, at the same time, Rothman emphasised that notion of causality in epidemiology highlights “important principles such as multi-causality, the dependence of the strength of component causes on the prevalence of complementary component causes, and interaction between component causes” (Rothman and Greenland 2005, p. S144). Rothman and Greenland (2005, p. S144) noted that our notion of causality is influenced by observation: switching on the lights can lead to a mistaken notion that it was the switch that was responsible or the cause. We know, of course, that electricity is the one really causing the light----and the switch is only a small part of a system of mechanism responsible for the light. Rothman and Greenland (2005, p. S144) argued that in epidemiology, “we can define a cause of a specific disease event as an antecedent event, condition, or characteristic that was necessary for the occurrence of the disease at the moment it occurred, given that other condition are fixed.” Therefore, “a cause of a disease event is an event, condition, or characteristics that preceded the disease event and without which the disease event would not have occurred at all or would not have occurred at all or would not have occurred until some later time” (Rothman and Greenland 2005, p. S144). Rothman and Greenland’s (2005) perspective echoed therefore the perspective of John Stuart Mill on causality or the inference of an effect from its cause. Rothman and Greenland (2005, p. S144) emphasized however that the definition focused only on “a component” of the cause and that “a ‘sufficient cause,’ which means a complete causal mechanism, can be defined as a set of minimal conditions and events that inevitably produce disease.” Rothman and Greenland (2005, p. S144) stressed that “minimal” means that “all the conditions or events are necessary to that occurrence.” In sum, the Rothman and Greenland (2005, p. S145) argued that causation and inference of effects from a cause in epidemiology requires the recognition of multi-causality, strength of specific causes in producing certain effects and the interaction of causes. For Rothman and Greenland (2005, p. S145), multi-casualty implies “that most identified causes are neither necessary nor sufficient to produce disease.” Further, “a cause need not be either necessary or sufficient for its removal to result in disease prevention” (Rothman and Greenland 2005, p. S144). At the same time, for Rothman and Greenland (2005, p. S145), another implication is that “if a component cause that is neither necessary nor sufficient is blocked, a substantial amount of disease may be prevented. The authors also elaborated “that the component cause is not sufficient implies that other component causes must interact with it to produce the disease, and that blocking any of them would result in prevention of some cases of disease” (Rothman and Greenland 2005, p. S145). Another very important implication is that “one need not identify every component cause to prevent some cases of disease” (Rothman and Greenland 2005, p. S145). Meanwhile, Rothman and Greenland (2005, p. S148) noted that Hill suggested in 1965 that strength, consistency, specificity, temporality, gradient, plausibility, coherence, and evidence are important in distinguishing causal from non-causal correlation. With regard to strength, “strong associations are more likely to be causal than weak associations” (Rothman and Greenland 2005, p. S148). Consistency refers to “the repeated observation of an association in different populations under different circumstances” (Rothman and Greenland 2005, p. S148). However, the same material said that a lack of consistency does not rule out a causal association. According to Rothman and Greenland (2005, p. S148), “the criterion of specificity requires that a cause leads to a single effect not a multiple effects.” Although Rothman and Greenland (2005, p. S148) conceded that the criterion is not valid as a general rule, “specificity can be used to distinguish some causal hypotheses from non-causal hypotheses, when the causal hypothesis predicts a relation with one outcome but no relation with another outcome.” In other words, “specificity can come into play when it can be logically deduced from the causal hypothesis in question” (Rothman and Greenland 2005, p. S148). Temporality refers to precedence of the cause relative to the effect (Rothman and Greenland 2005). Gradient requires that the cause will have one direction as an effect or that it will either increase or decrease a characteristic but “the existence of a monotonic association is neither necessary nor sufficient for a causal relation” (Rothman and Greenland 2005, p. S149). The exact word used by Rothman and Greenland (2005) was “biological gradient” but, in this discussion, the word “biological” was removed because the intension was to cover topics other than epidemiology that was the subject of the Rothman and Greenland (2005) discussion. Plausibility refers to being logical or compatible with sound reasoning or being sensible even as its possible drawback is that plausibility can be associated with prior knowledge (Rothman and Greenland 2005). Coherence “implies that a cause-and-effect interpretation for an association does not conflict with what is known” (Rothman and Greenland 2005, p. S149). Rothman and Greenland’s discussion of what can be considered evidence is ambivalent or unclear but this did not stop Rothman and Greenland (2005) from emphasizing the importance of evidence with regard to the inference of an effect from a cause. It must be emphasised though that the assertions forwarded by Rothman and Greenland (2005) had pertained to epidemiology and not to causation in other disciplines. Nevertheless, at least some of their notions should be applicable to fields other than epidemiology. The work of Gerring (2005) offered four main arguments on the nature of causation in the social sciences. First, a cause raises the probability of an event. Second, causation involves a set of logical criteria. Third, it is useful to distinguish between the formal properties of causation and the methods through which causation is tested. Finally or fourth, sixteen criteria apply to the formal properties and seven criteria apply to methodology. However, more important, Gerring outlined the types of causal relationships that he described as an “ever-expanding menu of causal relationship” (2005, p. 164-165): “…These include the conjectural cause (also known as compound cause, configurative cause, combinatorial cause, conjunctive plurality of a given effect; causal equifinality (also known as disjunctive plurality of causes, redundancy, or overdetermination), where several causes act independently of each other to produce, each on its own, a particular effect; the cause in fact (also known as the actual cause) which explains a specific event as opposed to a class of events; the non-linear cause (e.g., a cause with a takeoff or threshold level); the irreversible cause (e.g., a cause with a ratchet effect); the constant cause, a cause operating continually on a given effect over a period of time; the causal chain (also known as causal path), where many causes lie between a structural Y and an ultimate Y; and the critical-juncture or path-independent cause, a cause or a set of causes at particular moment in time that has enduring effects.” Thus, for Gerring, causation has many types and causal relationships are “infinite in their diversity” (2005, p. 165). If Gerring’s assertions are true or will turn out to be true, then the possibility is raised that not all of the principles related to causation applies to all types of causation. Nevertheless, it can be argued that even if this is the case (many types of causation exist), the point is to identify the key elements of causation that are applicable or generally applicable to all the types of causation: in particular, on how effects can be inferred from causes. Gerring (2005, p. 167) did identify one key element and that is, “the core, or minimal, definition of causation held implicitly with the social sciences is that a cause raises the probability of an event occurring.” Unfortunately, however, this core or minimum definition of causation may only be expressing correlation and not causation and we may have to emphasise the important of precedence of a cause from its effect. Nowadays, statistics are available to test correlations and their significance as well as precedence and their significance. Unfortunately, eliminating rival hypotheses to prove causation conclusively is difficult and we may have to retreat to the view that causation cannot be conclusively proven; evidence can only be provided to support a specific theory of causation but the evidence will probably be never enough to conclusively prove causation. The role of research may be limited to indicating what theory of causation has greater evidence rather than conclusively proving causation. There may be a case for proving causation “beyond reasonable doubt” but perhaps that will have to be discussed in another paper. References Chignell, A. and Pereboom, D., 2010. Kant’s theory of causation and its eighteen-century background. Philosophical Review 119 (4), 564-591. Cook, T. and Campbell, D., 1979. Causal inference and the language of experimentation. In: Vaus, D., ed. 2006. Research design. London: Sage Publications Ltd., 171-208. Gerring, J., 2005. Causation: A unified framework for the social sciences. Journal of Theoretical Politics, 17 (2), 163-198. Rothman, K. and Greenland, S., 2005. Causation and causal inference in epidemiology. Public Health Matters, 95 (S1), S144-S150. Read More