Causal Structure and Mechanisms
Abstract and Keywords
This chapter explicates the connection between the qualitative concept of a mechanism and the probabilistic causal concepts described in the previous chapter. The main thesis is that, for a broad range of cases of interest to the present study, it is reasonable to identify mechanisms with what is called causal structure in work on the problem of inferring causal conclusions from statistical data.
An important prerequisite for exploring the mechanisms approach to extrapolation is to explain what the qualitative concept of a mechanism has to do with probabilistic causal concepts such as causal effect and causal relevance. That is the task undertaken in this chapter and the next. In this chapter, I attempt to show that, for a broad range of cases of interest to the present study, it is reasonable to identify mechanisms with what is called causal structure in work on the problem of inferring causal conclusions from statistical data (cf. Glymour and Cooper 1999; Spirtes, Glymour, and Scheines 2000; Pearl 2000; Neopolitan 2004). Accomplishing this necessitates saying something about what causal structure is, and when and why mechanisms can be identified with it.
Explaining how this works involves reconsidering the manner in which analytic philosophers have traditionally approached the topic of causality. One of the primary activities (and perhaps the primary activity) of traditional analytic philosophy is conceptual analysis. I understand conceptual analysis to consist of providing necessary and sufficient conditions for the application of an interesting yet somewhat unclear term (e.g., “explanation,” “cause”), where these conditions satisfy the following two properties. First, the conditions are stated via concepts that can be defined independently of the target of the definition. Second, the usage of the term recommended by the analysis must agree tolerably well with the intuitions of native speakers in all conceivable circumstances. However, conceptual analysis has decidedly fallen from favor in recent years in the philosophy of science. For example, leading accounts of causality in the recent philosophy of science literature (cf. Hausman 1998; Dowe 2000; Woodward 2003) explicitly disavow any intention to provide a conceptual analysis in the sense just described. Rather than conceptual analysis, these authors endeavor to develop an account of causality that is informed by current scientific theories and methodology. Dowe, whose approach to causation owes much to Wesley Salmon (1984), strives for what he terms an empirical analysis of causality, that is, “to discover what causation is in the objective world” (Dowe 2000, 1). Dowe regards current physical theory as the most reliable source of information that would serve as a basis of an answer to this question.
But there is a simple objection to any program that would proceed with empirical analysis before conceptual analysis is complete: without prior conceptual analysis it is unclear what basis there is for asserting that the identified characteristic of the world corresponds to the term derived (p. 31 ) from ordinary language. David Lewis has posed this objection in the context of a discussion of the philosophy of mind, but it transfers easily to discussions of causation. In Lewis's words:
Arbiters of fashion proclaim that analysis is out of date. Yet without it, I see no possible way to establish that any feature of the world does or does not deserve a name drawn from our traditional mental vocabulary. (1994, 415)
After considering and rejecting Dowe's response to this objection, I propose that a better answer derives from the view that causal locutions should be treated as theoretical terms in the sense of the Ramsey‐Lewis account, according to which theoretical terms are a kind of definite description (cf. Lewis 1970). Given this perspective, an empirical analysis should be based upon a meaning postulate that specifies a particular role associated with the term in question. I will concentrate on two roles ascribed to causal structure; in particular, causal structure is that which generates probability distributions and indicates how these distributions change given interventions.
From this starting point, an empirical analysis of causal structure consists of indicating what fulfills these roles in a particular domain. Making the case for identifying mechanisms with causal structure requires some general argument for supposing that mechanisms are modular, in the sense that it is possible to alter one component without disrupting the functioning of the others. I explain how evolutionary theory can support the claim that modularity is likely to be a pervasive feature of mechanisms. However, this argument is, at present, on firmer ground in molecular biology than in social science, making the motivation for identifying causal structure with mechanisms somewhat more tentative in the latter case. An implication of this discussion is that empirical analyses of causation depend on domain‐specific scientific details and hence may differ for distinct phenomena. The question of whether social mechanisms should be identified with causal structure, and under what circumstances, will be explored in further detail in Chapter 8.
3.1 IT's NICE, BUT IS IT CAUSALITY?
An empirical analysis of causation proceeds by examining the question of what causation is in the world. For example, Dowe's conserved quantity theory advances the following two propositions as the foundation of an answer to that question:
CQ1. A causal process is a world line of an object that possesses a conserved quantity.
CQ2. A causal interaction is an intersection of world lines that involves exchange of a conserved quantity. (2000, 90)
(p. 32 ) It is striking how removed this analysis is from many ordinary discussions of causation. For instance, it is unclear what relevance exchanges of conserved quantities have to the claim that the vitamin C tablets that Bob ate did not cause him to recover from his cold. 1
Lewis's objection, then, seems quite apt: the conserved quantity theory is interesting, but why should one regard it as an account of causality? And how can this question be answered without presupposing a conceptual analysis? Dowe responds to this objection in the following way:
In drawing explicitly on scientific judgments rather than on intuitions about how we use the word, we nevertheless automatically connect to our everyday concept to some extent, because the word cause as scientists use it in those scientific situations must make some historical or genealogical connection to everyday language. (2000, 9)
Thus, basing an analysis of causation on current science connects to commonsense ideas concerning the meaning of “cause” since the usage of the term by scientists is linked to that of ordinary folk. But does this mean that empirical analysis simply amounts to a conceptual analysis of scientists' concept of causation? Dowe makes it clear that this is not his intent: “The task of empirical analysis…is not a conceptual analysis of scientists' usage of a term” (2000, 10). Rather, he maintains that the empirical analysis he pursues aims to explicate the concept of causation “implicit in scientific theories” (2000, 11).
The main difficulty I see with this response is that it is highly questionable whether there is a concept of causation implicit in current scientific theory. As Dowe observes, no physical theory contains “cause” as an explicitly defined term (2000, 9), and consequently any proposed empirical analysis of causation must inevitably be a substantive thesis over and above what is given by science (Bontly 2006, 182–83). Moreover, there are several ways that one could interpret causation in the light of current science, and it seems unavoidable that arguments for choosing one approach over another will appeal to intuitions about the proper usage of the word “cause.” To take just one issue, consider whether causation requires determinism. Dowe argues that the answer is no, on the grounds of an example concerning exposure to radioactive material.
If I bring a bucket of Pb210 into the room, and you get radiation sickness, then doubtless I am responsible for your ailment. But in this type of case, I cannot be morally responsible for an action for which I am not causally responsible. (2000, 23)
Thus, given the scientifically plausible assumption that the decay of Pb210 is a fundamentally indeterministic process, it follows that indeterministic causation exists.
Although the above argument is interesting and perhaps even persuasive, it is clear that there is more to it than merely explicating a concept (p. 33 ) implicit in physical theory. Dowe's argument depends crucially on the thesis that moral responsibility (at least in some unspecified class of cases of which the present one is an example) entails causal influence. But what is the basis of any such principle linking moral responsibility and causation? Surely it is not physical theory. Rather, any grounding for it would reside in the interconnection of ordinary concepts of responsibility and causality. As a result, one who maintained that determinism is a fundamental aspect of the concept of causality (e.g., Pearl 2000, 26–27) could avoid the conclusion of Dowe's argument by rejecting the claim that moral responsibility implies causal influence. For example, I might have a moral responsibility to provide assistance to starving people in a distant land despite the fact that I am in no way causally responsible for their unfortunate situation. Thus, Dowe's use of current physics to argue for indeterministic causation requires an antecedent clarification of the relationship between causation and moral responsibility.
Physical theory certainly does have implications for the nature of causation. In the foregoing example, modern physics makes it difficult to maintain both that causation is inherently tied to determinism and that moral responsibility entails causal influence. But this does not show that there is a single account of causality implicit in physical theory, since several different accounts of causation can be made consistent with modern science, depending on what position one takes regarding the interconnections between causation and such things as responsibility, human agency, determinism, temporal priority, spatiotemporal contiguity, and so on. Yet one significant aim of conceptual analysis is to settle questions concerning such interconnections. Hence, we are led straight back to Lewis's objection: empirical analysis cannot fruitfully proceed until matters of conceptual analysis have been settled.
Let us consider a different account of how an empirical analysis of causation can proceed even in the absence of a successfully completed conceptual analysis.
3.2 CAUSALITY AND THEORETICAL TERMS
In this section, I suggest that the cogency of empirical analysis without a successfully completed conceptual analysis can be defended by considering causal locutions as theoretical terms in the sense of the Ramsey‐Lewis account (Ramsey 1954; Lewis 1970). The Ramsey‐Lewis account proposes to treat theoretical terms as a type of definite description stated via antecedently understood concepts: 2 the theoretical entity is simply that (if anything) which satisfies the description. For example, in eighteenth‐century chemistry, phlogiston is that which is present in all flammable objects and is emitted during the process of combustion. In Lavoisier's chemistry, oxygen is that which is absorbed during combustion and is necessary for the formation of acids.
(p. 34 ) Several authors have suggested that the Ramsey‐Lewis account, in addition to applying to deliberately introduced terms of scientific theories, could also be appropriate with regard to concepts falling more squarely in the province of philosophy. For example, Michael Tooley (1987) and Peter Menzies (1996) take such an approach to causation, and Dowe (2000, 49–51) sympathetically considers the idea with respect to transference theories of causation. 3 In Dowe's formulation, such an analysis of causation would consist of three components: a meaning postulate, a contingent hypothesis, and an a posteriori identity (2000, 49). The meaning postulate is the definite description that specifies some important feature of causation: causality is that which does __. For example, one plausible claim is that causation is that which underlies the possibility of predicting the consequences of interventions (cf. Menzies and Price 1993; Woodward 2003). The contingent hypothesis would then be an empirical claim about what things in the world fulfill this role in a given domain, while the a posteriori identification would assert that (in the domain in question) causation is identical to the entity or process indicated in the contingent hypothesis.
The question, then, is how to decide what the meaning postulate should be. An agreed‐upon conceptual analysis, if one were available, clearly would be one possible basis for answering this question. For example, Tooley treats his proposal regarding the meaning postulate as a conceptual analysis (cf. Tooley 1987, 25–28). If this were the only possible way to justify one's choice of meaning postulate, then Lewis's argument that empirical analysis cannot proceed until matters of conceptual analysis have been settled would be vindicated. But there is another possibility: the meaning postulate could be derived from empirical observations of the use of causal language. For example, Thomas Bontly proposes that we regard “the concept of causation as a concept defined by its place in an inferential system or network, by the inferences it licenses and those that license it” (2006, 191). Given this perspective, the meaning postulate should be based on inferences that people actually make to and from causation. A meaning postulate, then, should indicate something that is generally regarded as evidence for causal claims as well as something that is judged to be a consequence of causal claims. A meaning postulate that focuses on the connection between causation and predicting the outcomes of interventions does both of these things. The connection between causal claims and effective strategies for achieving ends has been emphasized by many authors (cf. Cartwright 1983, chap. 1; Mellor 1988, 230; Hoover 2001; Woodward 2003). Moreover, carefully controlled interventions are generally regarded as the most reliable scientific means for testing causal claims. There is also experimental evidence that preschool‐age children regard interventions as an especially effective way of learning what causes what (Kushnir and Gopnik 2005). Similarly, covariance is generally regarded as a consequence of causal relationships and as evidence for them, at least under the right circumstances (Cheng 1997). (p. 35 ) Thus, either manipulation or covariance of the right sort is a potential basis for a meaning postulate in an empirical analysis of causation. In fact, the meaning postulate that will be discussed below—according to which causal structure is that which generates probability distributions and provides information about how they change under interventions—combines both notions. Physical contiguity is a third factor that is often relevant to causal inferences, and it is presumably the guiding thought behind Dowe's conserved quantity theory. However, physical contiguity alone is rarely sufficient to infer causation, since one event might be physically adjacent to another without having caused it. Not surprisingly, in his definition of “C causes E,” Dowe combines the definitions of causal process and interaction presented above with a requirement that the cause raise the chance of the effect (2000, 167).
The link between causation and manipulation is doubtful as a conceptual analysis of causation, since specifying what a manipulation or intervention is will inevitably involve references to causation. Nevertheless, a principle linking causation to manipulation can serve as an appropriate meaning postulate for an empirical analysis that treats causation as a theoretical term in the sense of the Ramsey‐Lewis theory. If it can be shown that the feature of the world specified in the empirical analysis makes effective manipulation possible, then there is a straightforward answer to the question: Why call it causation? Whatever causation is, knowledge of it is often important for indicating effective and ineffective strategies for achieving ends. Hence, if one identified a general feature of the world that fulfilled this function, then one would have a legitimate claim to be describing causation.
It may be objected that the connection between manipulation and causation could not serve as a meaning postulate, since manipulation is a causal concept, whereas the terms in the meaning postulate are supposed to be antecedently understood. In response, I claim that manipulation and intervention are antecedently understood: they are drawn from the vocabulary of ordinary English and everyday life. (Of course, that does not preclude the usefulness of introducing a framework for discussing them more clearly, as done in section 2.1.) The key point is that antecedently understood is a criterion distinct from independently definable: we have a reasonably clear idea of what an intervention is, regardless of whether we can define the term in a manner that eschews all reference to causation. Consequently, it is legitimate to use intervention as the basis of the meaning postulate for a Ramsey‐Lewis‐style definition of “causal structure.”
Another possible objection is that without a conceptual analysis of causation, there will be several potential starting points for an empirical analysis of causation. I think it is quite right that there may be several reasonable choices for starting points for an empirical analysis of causation, and that different starting points might lead to separate destinations. However, this is a problem only if one supposes that there must be (p. 36 ) a monolithic concept of causation for which a unique empirical analysis must be given. In contrast, I see no reason to rule out at the start of inquiry the possibility that the notion of causation is multifaceted. 4 Given the account proposed here, empirical analyses of causation might be pluralistic in two ways. First, a single meaning postulate might be realized differently in distinct domains. For instance, that which generates probability distributions and provides information about how they change under interventions might be one kind of thing in fundamental physics and another in molecular biology and something else again in economics. Second, there may be several reasonable meaning postulates that lead to distinct empirical analyses even within the same domain of inquiry. For example, an empirical analysis based on manipulation might lead to results different from one that emphasizes physical contiguity. The potential for this second type of pluralism raises the question of whether there are common threads linking the several meaning postulates, or whether “causation” is simply an ambiguous term with several distinct meanings. My own view is that the various causal concepts are all closely linked elements of a network of concepts relating to practical reason. However, the account of extrapolation developed in this book does not depend upon the correctness of that overarching vision of causation. All that I require is that the meaning postulate I associate with causal structure be a reasonable one.
Despite the pluralistic spirit expressed in the foregoing paragraph, it is important to stress that not any old thing can be an acceptable meaning postulate. For instance, it would be absurd to say that causation is that which is located in the top drawer of my desk. Absurd proposals like this one would clearly be disqualified by the requirement that a meaning postulate indicate something that is generally regarded as both evidence for and a consequence of causation. But some things that are conceptually linked to causation also fail this criterion. Suppose one proposed this as a meaning postulate: “Causation is that which is necessary for moral responsibility.” That there is some conceptual link between moral responsibility and causation seems clear enough. In many cases, one can be morally responsible for something only if one has some influence on it. However, moral responsibility is not something that could serve as evidence for causation. Evidence for causation is something that you can actively search for or produce in order to decide whether a causal relationship obtains. If you want to know whether A causes B, you might do an experiment in which you manipulate A and check to see if B varies concomitantly. Or you might collect statistical data to see if A and B are correlated even when potential common causes are statistically controlled for. But there is no analogous way to use moral responsibility as a basis for testing causal claims. The same point would go for the suggestion that causation is that which underlies explanation. Consequently, not everything that is conceptually linked to causation can serve as a good meaning postulate in an empirical analysis of it.
(p. 37 ) 3.3 CAUSAL STRUCTURE
A lively body of work on the problem of causal inference from statistical data uses directed graphs to represent causal structures (cf. Glymour and Cooper 1999; Spirtes, Glymour, and Scheines 2000; Pearl 2000; Neopolitan 2004). For example, consider Figure 3.1.
As in section 2.1, the nodes of the graph correspond to variables and an arrow from one node to another indicates the relationship of direct causation. For instance, Y might represent whether or not a particular power strip is switched to the “on” position, while X and Z each indicate whether or not an electrical appliance plugged into the power strip is on. Using directed graphs to represent causal structures has several advantages for theories of causal inference, the most significant of which is that it enables one to draw upon mathematical results which facilitate computationally tractable methods of deriving predictions about probabilistic independence and conditional independence from alternative causal hypotheses. 5 Directed graphs in conjunction with probability distributions are sometimes referred to as Bayesian networks, or Bayes nets for short. 6 For convenience, I shall adopt the label causal Bayes nets to refer to the approach to causal inference just briefly described.
Causal structures, then, are what directed graphs are intended to represent in the causal Bayes nets literature. But that does not tell us very much about what causal structures are; after all, directed graphs like that in Figure 3.1 can just as easily be used to represent mere correlations. And of course, things other than directed graphs—such as systems of equations and wiring diagrams—can also be used to represent causal structures. What is it, then, that these diverse modes of representation depict? Introductions to treatises on the topic typically emphasize the importance of causal inference for accurately predicting the consequences of public policy decisions (cf. Glymour and Cooper 1999, xi–xii; Pearl 2000, 337; Spirtes, Glymour, and Scheines 2000, xiii–xiv). In addition, significant effort is dedicated to inquiring how knowledge of causal structure, in varying degrees of precision, can serve as the basis of predicting consequences of interventions (cf. Spirtes, Glymour, and Scheines 2000, chap. 7). Thus, causal structures provide information concerning the results of interventions. An additional role is also attributed to causal structures: causal structures are said to “generate” probability distributions (cf. Glymour 1997, 206; Spirtes, Glymour, and Scheines 2000, 29).
(CS) Causal structure is that which generates probability distributions and indicates how these distributions will change given interventions.
A good understanding of (CS) is evidently dependent on some explication of interventions and of what it is to “generate” a probability distribution. Since the notion of an ideal intervention was explained in section 2.1, let us consider the second of these two questions.
For our purposes, the concern is with physical probability rather than probabilities interpreted as personal degrees of belief or confidence. Although the concept of physical probability is nearly as disputed as that of causation, I think that it is clear enough what sort of phenomena such probabilities usefully represent, namely, processes whose outcomes exhibit what John Venn described as a combination of “individual irregularity with aggregate regularity” (1962, 4). For example, consider the simple case of a flipped coin.
So long as we confine our observation to a few throws at a time, the series seems to be simply chaotic. But when we consider the result of a long succession we find a marked distinction; a kind of order begins gradually to emerge, and at last assumes a distinct and striking aspect. We find in this case that the heads and tails occur in about equal numbers, that similar repetitions of different faces do also, and so on. In a word, notwithstanding the individual disorder, an aggregate order begins to prevail. (Venn 1962, 5)
As Venn observed, this type of behavior is found in many other circumstances: “Fires, shipwrecks, yields of harvest, births, marriages, suicides; it seems scarcely to matter what feature we single out for observation” (1962, 6).
For our concerns, it is unimportant whether one wishes to define probability as the aggregate or macro pattern itself (as frequency interpretations do), or as the causal tendencies underlying that aggregate pattern (as propensity interpretations do). Probabilities are useful for representing, or modeling, any phenomenon that displays a combination of individual irregularity and aggregate regularity. A process can be said to generate a probability distribution, then, just in case it gives rise to an aggregate pattern of this sort. This criterion is, admittedly, somewhat vague, but it will suffice for the present purposes.
Things that generate probability distributions, then, must exhibit behavior possessing the combination of individual disorder and aggregate regularity described by Venn. I maintain that these properties are possessed by mechanisms that are impinged on by disturbances that are, from the perspective of human knowledge, largely random. Moreover, mechanisms often provide information about the effects of interventions. Consequently, mechanisms are promising candidates for causal structure. (p. 39 ) Let us consider this thought in more detail with regard to a pair of cases: molecular biology and social science.
3.4 CAUSAL STRUCTURE IN MOLECULAR BIOLOGY
Given a meaning postulate, the next stage of an empirical analysis is a contingent hypothesis, which specifies a class of entities whose extension, in a particular domain, is exactly that of the meaning postulate. In this section, I argue that in molecular biology, causal structure coincides with mechanisms, yielding the following empirical analysis:
• Meaning Postulate (CS): Causal structure is that which generates probability distributions and indicates how these distributions change under interventions.
• Contingent Hypothesis: In molecular biology, mechanisms are what generate probability distributions and indicate how these distributions change under interventions.
• A Posteriori Identity: In molecular biology, mechanisms are causal structure.
In this section, I argue in favor of the above contingent hypothesis. As explained in earlier sections of the chapter, empirical analyses rely upon established scientific theories of the relevant domain. In this case, evolutionary biology plays an important role in motivating the claim that mechanisms in molecular biology provide information about the consequences of interventions by providing a general reason to expect that such mechanisms are modular.
3.4.1 What's a Mechanism?
Mechanisms, in a very literal sense of the term, are paradigmatic examples of causal structures. For example, in Nancy Cartwright's words:
The car engine is a good case of a stable causal structure that can be expected to give rise to a probability distribution over the events of the cooperating causal processes that make it up. That is why it can make sense to ask about the conditional expectation of the acceleration given a certain level of the throttle. (1995a, 72)
Given that several authors have proposed that mechanisms play an important role in the life sciences (cf. Bechtel and Richardson 1993; Glennan 1996; Machamer, Darden, and Craver 2000), they are a natural place to turn for an empirical analysis of causal structure in biology. However, this must be done with some care, since the application of the word “mechanism” in distinct domains might reflect only a superficial similarity of subject matter. Thus, it is important to examine just what sorts of things biological mechanisms are and why they should be thought to fulfill the roles ascribed to causal structure.
(p. 40 ) Mechanisms are generally understood as consisting of interacting components that generate a causal regularity between some specified beginning and end points. For example, according to a definition proposed by Peter Machamer, Lindley Darden, and Carl Craver, “Mechanisms are entities and activities organized such that they are productive of regular changes from start or set‐up to finish or termination conditions” (2000, 3). This general characterization is appropriate for literal examples of mechanisms, such as the car engine, and is reasonable with regard to things referred to by the term “mechanism” in biological science. Consider, for example, the mechanism involved in protein synthesis, in which the series of nucleotide bases in strands of DNA influences the chemical structure of proteins produced within cells. Nearly any introductory biology textbook describes this mechanism roughly as follows. First, a strand of DNA unwinds and the adjoining nucleotide bases separate. The next step is the transcription of the unwound DNA by messenger RNA (mRNA), the order of the bases of the mRNA being determined by the order of the complementary nucleotide bases in the DNA strand. Finally, the strand of mRNA serves as a template for transfer RNA (tRNA), which assembles a string of amino acids into a protein. In this case, the inter‐working parts give rise to more readily observed regularities, such as correlations between genes and specific traits. Some things referred to by the term “mechanism” may not involve a regular series of changes. For example, the term “mechanism” is sometimes used to refer to a unique chain of events leading to a particular effect. However, since this book is concerned with extrapolating causal generalizations, I will use the term “mechanism” to refer to regularly operating causal relationships rather than idiosyncratic and unique chains of events. Consequently, I will restrict the term “mechanism” to processes that satisfy the “regular changes” clause of the Machamer‐Darden‐Craver definition.
Other related definitions of mechanisms exist. For example, Stuart Glennan proposes a definition that is similar to Machamer, Darden, and Craver's except that it requires that the interactions among the components of the mechanism be governed by “direct causal laws” (1996, 52). The reference to laws in this definition is problematic, since it is debatable whether there are genuine laws of nature in biology and social science, where the term “mechanism” is often used. Consequently, in a subsequent revised account of mechanisms, Glennan replaces “direct causal laws” with “direct, invariant, change relating generalizations” (2002, S344). The notion of an invariant generalization is borrowed from James Woodward (2000, 2003). An invariant generalization is one that is invariant under some range of ideal interventions on the allegedly explanatory variable. For example, the generalization that barometer readings and storms are correlated is not invariant under ideal interventions on the barometer readings (as explained in section 2.1). Hence, the barometer readings do not cause or explain storms, according to Woodward's theory. In contrast, the generalization that smoking is correlated with lung (p. 41 ) cancer would be invariant under ideal interventions that target smoking. James Tabery (2004) argues that there is an important difference between Woodward's conception of causation and the notion of “productivity” invoked in the definition proposed by Machamer, Darden, and Craver. The thought is that while invariant generalizations merely point to ways in which changes brought about by an intervention lead to specific changes someplace else, productivity pertains as well to cases in which new entities are constructed (2004, 8–9). However, the “changes” covered by Woodward's account of causation should be understood to include constructing a new product out of disparate parts. For example, imagine a cellular process that generates a particular enzyme. Let E be a variable that indicates whether or not this enzyme has or has not been produced on given occasions. Then there may be invariant generalizations relating E to other variables that represent, say, the presence of necessary components in the cell or the transcription of a particular gene. If there is a real difference between Glennan's definition and that proposed by Machamer, Darden, and Craver, I think it is only that Glennan provides more detail about his preferred interpretation of causation.
Cartwright's nomological machine is another mechanism concept. Cartwright defines a nomological machine as “a fixed (enough) arrangement of components, or factors, with stable (enough) capacities that in the right sort of stable (enough) environment will, with repeated operation, give rise to the kind of regular behaviour that we represent in our scientific laws” (1999, 50). Like the definitions of mechanism considered above, Cartwright's nomological machine consists of interacting components that generate causal regularities. The concept of a nomological machine is distinctive only insofar as it is founded on Cartwright's concept of a capacity. A capacity is a stable causal power that exerts its characteristic influence in a broad range of contexts (Cartwright 1989, Chapter 4). The pure effects of a capacity can be observed only in special experimental circumstances in which all other causes have been eliminated, but the capacity nevertheless makes its contribution to the effect even when other causes are present. Since Cartwright regards physical laws merely as descriptions of the behavior of a capacity in the idealized situation in which no other forces are acting, she regards capacities as ontologically more basic or fundamental than laws of nature. Cartwright also argues that interpreting causal relationships by reference to capacities is essential for understanding how it is possible to extrapolate causal claims from one context to another (1989, 163). I argue in Chapter 5 that capacities do not in fact have this special virtue. But for the moment, let us sum up the above survey of mechanism concepts.
All of the definitions canvassed above characterize mechanisms as consisting of sets of interacting components that generate a regular series of causal interactions. To the extent that they disagree, it is with regard to how to interpret causation. For example, Glennan's original definition (1996) characterized causation by reference to “direct causal laws,” while (p. 42 ) Cartwright prefers capacities. Fortunately, pursing an empirical analysis of causal structure does not require deciding whether laws or causal powers are more fundamental or insisting that there is one correct way to interpret causation. Instead, it requires an argument that mechanisms generate probability distributions and provide information about how those distributions change under interventions. Given this, I will adopt the Machamer–Darden–Craver definition, since it is the least specific about causation, laws, and their relation to one another. The question, then, is whether mechanisms, so defined, are causal structures. I consider this question first with regard to molecular biology and then for social science.
3.4.2 Mechanisms, Modularity, and Evolvability
There is good reason to think that if there is such a thing as causal structure in molecular biology, it would have to be mechanisms. First, note what might be called the working assumption of molecular biology: all causal relationships in living organisms are mediated by molecular processes. This working assumption rests on the attractiveness of physicalism as a general ontological principle and on the success of molecular biology as a research program. Thus, if mechanisms are not causal structures in molecular biology, it is hard to see what could be. However, this conclusion is only half of the argument. It is also necessary to show that mechanisms in molecular biology do in fact perform the functions required of causal structure.
Since causal structure is that which generates probability distributions and provides information about how those distributions change given interventions, there are two parts to this argument. Let us begin with the requirement that causal structure generate probability distributions. Is this something that mechanisms in molecular biology do? Recall the features that Venn judged to be characteristic of phenomena to which the concept of probability can be usefully applied: individual disorder combined with aggregate regularity. It is obvious that mechanisms in the sense of the Machamer–Darden–Craver definition will tend to generate large sample regularities, given the requirement that mechanisms “are productive of regular changes” from the beginning and end stages of the process. Moreover, biological mechanisms are invariably subject to an array of disturbing influences, many of which are not well understood. Thus, from the perspective of human knowledge, individual cases of the operation of a given mechanism in molecular biology will inevitably display a certain amount of random variation, which is an example of the “individual disorder” that Venn described. Notice that the same sort of situation is found in the case of human‐constructed machines, which are often given as paradigm examples of causal structure. They produce regular changes, yet are impinged upon by a variety of disturbing influences that often cannot be known with any exactitude. Consequently, we have a straightforward account of why mechanisms in molecular biology (p. 43 ) should display the aggregate regularity and individual disorder that Venn cited as the characteristic features of probabilistic phenomena. Of course, these aggregate patterns may themselves change in the course of evolution, but this simply illustrates the familiar point that probability distributions themselves can change over time (cf. Venn 1962, 14–17). This point is illustrated by such social statistics as the marriage rate or average life span. Indeed, it is exemplified by Cartwright's case of the car engine; the probability of a breakdown increases as the engine ages.
However, since knowledge of causal structure also provides information about the consequences of interventions, an account of why mechanisms in molecular biology should be thought to generate probability distributions is only half of the story. It is necessary to argue that mechanisms in molecular biology generally provide information about the results of interventions. On the face of it, it is quite plausible that this is the case. Indeed, this presumption that knowledge of mechanisms can indicate the consequences of various types of interventions is often the reason for trying to discover them. But is there some general feature of biological mechanisms that justifies this presupposition? One answer to this question has been suggested by Woodward (2002a, S374–76), who maintains that mechanisms are modular in the sense that it is possible to intervene to change a feature of one component while leaving the generalizations that govern the others unaltered. This idea is reflected in the manner in which interventions are represented in directed graphs. Consider again the case of the two appliances plugged into the same power strip, represented by the graph in Figure 3.1. Recall that an ideal intervention takes complete control of the variable it targets (say, X), so as to eliminate all other influences that otherwise affect it. Such an intervention, as we saw in section 2.1, would be represented as shown in Figure 3.2.
Of course, many real‐life interventions are not ideal. In our example, switching on one of the appliances would not be an ideal intervention, since it does not sever the influence of the state of the power strip. Such an intervention might be represented as shown in Figure 3.3:
The important point with regard to modularity in figures 3.2 and 3.3 is that besides possibly eliminating or weakening the influence of Y upon X, the intervention leaves all other causal relationships unaltered. For example, modularity would be violated if the intervention eliminated the influence of Y upon Z or created a causal chain from X to Z. The interest in modularity stems from the fact that it facilitates predicting the
One way to argue that modularity is likely to be a commonly occurring feature of biological mechanisms is to maintain that modularity is favored by natural selection. Herbert Simon (1962) was one of the first to suggest a general explanation of how modularity is adaptively beneficial in environments in which disruption or interference is common. The proposal can be illustrated with a modified version of one of Simon's best‐known examples, the parable of the expert watchmakers Hora and Tempus (1962, 470). 7 Hora constructs her watches by building independently changeable modules that can be assembled into the final product. In contrast, Tempus constructs holistic watches in which no part can be modified independently of any of the others. Hora's modular production method gives her an advantage over Tempus as their ever more popular watches are used in new circumstances. For instance, mountain climbers find that the watches of Hora and Tempus fail to operate properly at high altitudes. Hora is able to trace the problem to a specific module, and through trial and error she develops a new version that operates properly under high‐altitude conditions. In contrast, Tempus must redesign an entirely new high‐altitude watch, which means searching for a solution through the space of possible watches, which is far vaster than the space of possible modifications of a specific component. By the time Tempus has finished his holistic high‐altitude chronometer, Hora has already cornered the mountain climber watch market, as well as that for scuba divers, mariners, pilots, runners, and several other specialty niches. The moral of the parable, then, is that modularity facilitates finding quick solutions to new problems, which is essential for adapting to changing environments.
(p. 45 ) The theme of this parable is nicely illustrated by the HIV replication mechanism that will be discussed in detail in the next chapter. HIV is notorious for its ability to evolve resistance to drugs designed to block its replication. Typically, such drugs interfere with one stage of the replication mechanism, for example, by binding to and disabling an enzyme required for a step in the process. In this case, a mutation in the viral genome can result in a slightly modified version of the enzyme to which the therapeutic compound no longer binds. Given that the other components of the mechanism continue to function as before, HIV has successfully evolved resistance; but if the change to the enzyme resulted in cascading alterations to the other components, it is likely that the mutant strain would no longer be viable. Thus, the HIV replication mechanism is analogous to Hora's production method: since it is modular, alterations to one component to do not compromise the functionality of the others. Consequently, evolving resistance to a single drug requires altering only one component of the replication mechanism, and hence searching through a smaller space of possibilities. In contrast, if the HIV replication mechanism were holistic like Tempus's watches, evolving resistance to the therapeutic compound would require rebuilding the mechanism from scratch, and hence searching for a solution in the space of all possible HIV replication mechanisms. Thus, modularity is an important part of what enables HIV to quickly evolve resistance.
These examples suggest that modularity enhances fitness by promoting adaptability to changing environments. Moreover, environmental perturbations of various kinds—new predators, changes in supply of resources, and so on—are a pervasive fact of life. Hence, evolutionary theory suggests a basis for expecting that modularity is a typical characteristic of biological mechanisms. In fact, the importance of modularity to adaptability is a familiar point in evolutionary biology (cf. Wagner and Altenberg 1996). There is a growing body of theoretical work that attempts to clarify the general mechanisms whereby natural selection could give rise to modularity (cf. Ancel and Fontana 2000; Lipson et al. 2002; Kvasnicka and Pospichal 2002; Kashtan and Alon 2005). This work supports the intuition that natural selection favors modularity in changing environments, but with some refinements. For example, one recent study suggests that although not all varying environments lead to modularity, modularity is favored in environments with “modularly varying goals” (Kashtan and Alon 2005, 13777). Goals vary modularly when new goals share subproblems with preceding goals (ibid., 13775). The HIV example illustrates this concept. At first, the goal of the enzyme is to achieve a particular function, say, to reverse transcribe viral RNA to DNA. After the start of the drug treatment, the enzyme must still perform its original function while also avoiding being bound to the therapeutic compound. Hence, reverse transcribing the viral RNA to DNA is a subproblem shared by the first and second goals. The situation in the watchmaker parable is similar. In redesigning the malfunctioning module, Hora (p. 46 ) must preserve its original function while avoiding the disruption that occurs at high altitudes. Modularly varying goals might drop as well as add subproblems. For instance, consider a population of fish that has colonized a network of underground pools: the fish no longer need to see, but they still need to swim.
There is also a growing number of empirical studies that examine the role of modularity in the evolution of particular lineages (cf. Beldade et al. 2002; Chipman 2002; Mabee et al. 2002; Friedman and Williams 2003; Emlen et al. 2005; Fraser 2005). 8 These studies provide fascinating concrete examples of the ways in which modularity can be manifested in living beings. For example, one study documents how threshold mechanisms allow for developmental modularity in the evolution of beetle horns (Emlen et al. 2005). Empirical studies can also test hypotheses about the relationship between modularity and evolvability. For instance, mixing and matching modules, sometimes called “compositional evolution,” may often be a more efficient means of finding a solution to a problem than randomly rearranging basic elements (Watson and Pollack 2005, 456). By analogy, one is more likely to produce a sentence by randomly combining clauses and phrases than by randomly combining letters and spaces. An additional potential advantage of compositional evolution, in contrast to gradual accumulation of slight variations, is that it can avoid suboptimal local maxima traps, since a rearrangement of modules constitutes a jump to a nonadjacent point in the fitness landscape (Kashtan and Alon 2005, 13777). And in fact a recent study finds support for compositional evolution with regard to protein modules in yeast (Fraser 2005). In the HIV example discussed above, compositional evolution would suggest that the resistant variant resulted from rearranging proteins that compose the enzyme rather than from shuffling the individual amino acids that make up the proteins.
Mechanisms that are modular in the sense of these biological discussions are ipso facto a useful basis for predicting the consequences of interventions. Although several modularity concepts can be found in biology (Schlosser and Wagner 2004), the following is a fairly standard, rough definition that is appropriate for the present context:
A modular representation of two character complexes C1 and C2 is given if pleiotropic effects of the genes fall mainly among members of the same character complex, and are less frequent between members of different complexes. (Wagner and Altenberg 1996, 971)
According to this definition, modularity states that the multiple effects of genes tend to focus on discrete trait complexes. This definition makes the connection between modularity and manipulability straightforward. For if modularity in the sense just defined obtains, it is possible, by means of appropriate alternations to the genome, to intervene to alter one component of the mechanism without significantly disturbing the others. Thus, knowledge of modular mechanisms would provide information about the (p. 47 ) consequences of interventions. Of course, it would be a mistake to take the above as a general definition of modularity. Rather, it is a rough specification of the physical basis of modularity in molecular biology—in effect, an empirical analysis of modularity in that context. An empirical analysis of modularity in social science, for instance, would have to be something rather different.
In sum, given the meaning postulate that causal structure is that which generates probability distributions and indicates how such distributions change given interventions, evolutionary theory plays a central role in an empirical analysis of causal structure in molecular biology. Evolutionary theory can be invoked to support the claim that in the context of molecular biology, mechanisms can be identified with causal structure, since it provides an account of why it should be expected that biological mechanisms are typically modular. Modularity, meanwhile, was linked to the ability to predict the consequences of interventions. Of course, since empirical analysis depends on current scientific theory, it is inherently tentative. New scientific developments might result in significant revisions to the theory, and these developments might have implications for the empirical analysis. The evolution of modularity in biological systems is a young and thriving research area, which means that we should expect surprises yet to come.
3.5 CAUSAL STRUCTURE IN SOCIAL SCIENCE
In this section, I consider the possibility that an empirical analysis identifying causal structure with mechanisms in molecular biology on the basis of evolutionary theory could work similarly in social science. On its face, the argument for the adaptive benefits of modularity in variable environments seems entirely general, and hence applicable to cultural as well as to biological evolution. However, the details of these proposals are at present far less developed in social science than in biological science. In addition, one common argument against the possibility of laws of social science can be interpreted as an attempt to show that social mechanisms will often respond in nonmodular ways to interventions. Thus, I conclude that although it is likely that the evolutionary account of modularity described above can be applied to some social mechanisms, the extent to which this is so is even more of an open question than in the case of mechanisms in molecular biology.
3.5.1 What's a Social Mechanism?
In order to consider whether social mechanisms are likely to be modular, some clarification of “social mechanism” is called for. Earlier, mechanisms in general were roughly characterized as sets of entities and activities organized so as to produce a regular series of changes from a beginning state to an ending one. Social mechanisms in particular are usually thought of as complexes of interactions among agents that (p. 48 ) underlie and account for macrosocial regularities (cf. Little 1991, 13; Stinchcombe 1991, 367; Schelling 1998, 33; Gambetta 1998, 102). The paradigm example of an agent is an individual person, but coordinated groups of individuals motivated by common objectives—such as a corporation, a government bureau, or a charitable organization—may also be treated as agents for certain purposes (cf. Mayntz 2004, 248). Social mechanisms are sometimes tied to the assumption that the agents comprising them are rational, say in the sense of being utility maximizers. For instance, Tyler Cowen writes, “I interpret social mechanisms…as rational‐choice accounts of how a specified combination of preferences and constraints can give rise to more complex social outcomes” (1998, 125). I shall not adopt this perspective, and hypotheses about social mechanisms will not be restricted to rational‐choice models.
Social mechanisms typically involve reference to some categorization of agents into relevantly similar groups defined by a salient position their members occupy vis‐à‐vis others in the society (cf. Hernes 1998; Little 1998, 17; Mayntz 2004, 250–52). In the description of the mechanism, the relevant behavior of an agent is often assumed to be a function of the group into which he or she is classified. For example, consider the anthropologist Bronislaw Malinowski's (1935) account of how having more wives was a cause of increased wealth among Trobriand chiefs. Among the Trobrianders, men were required to make substantial annual contributions of yams to the households of their married sisters. Hence, the more wives a man had, the more yams he would receive. Yams were the primary form of wealth in Trobriand society, and served to finance such chiefly endeavors as canoe building and warfare. Although individuals play a prominent role in this account, they do so as representatives of social categories: brothers‐in‐law, wives, and chiefs. The categorization of component entities into functionally defined types is not unique to social mechanisms. Biological mechanisms (e.g., that of HIV replication) are often described using such terms as “enzyme” and “co‐receptor.” The terms “enzyme” and “co‐receptor” resemble “chief” and “brother‐in‐law” in virtue of being functional: all of these terms provide some information about what role the designated thing plays in the larger system of which it is a part. In sum, social mechanisms can be characterized as follows. Social mechanisms are complexes of interacting agents—usually classified into specific social categories—that produce regularities among macrolevel variables.
This characterization of a social mechanism can be illustrated by another, better‐known example. Consider Thomas Schelling's bounded‐neighborhood model, which is intended to account for persistent patterns of segregated housing in spite of increased racial tolerance (Schelling 1978, 155–66). In this model, the residents of a given neighborhood are divided into two mutually exclusive groups (e.g., black and white). Each individual prefers to remain in the neighborhood, provided that the proportion of his or her own group does not drop below a given (p. 49 ) threshold, which may vary from person to person. Meanwhile, there is a set of individuals outside the neighborhood who may choose to move in if the proportions are to their liking. Clearly, this model divides individuals into groups with which characteristic preferences and subsequent behavioral patterns are associated, and by these means accounts for macroregularities.
On the face of it, it might seem that the empirical analysis of causal structure given in section 3.4 easily transfers to social science. As in the case of molecular biology, it is difficult to see what could constitute causal structure in social science if not social mechanisms. Moreover, it is plausible that social mechanisms often produce stable patterns, and hence generate probability distributions. Finally, just as in the case of biology, it seems that modularity is a feature that contributes to the adaptability of social systems. Indeed, the parable of Hora and Tempus illustrates the advantages of modularity for technology and is analogous to such historical cases as the IBM PC versus the Apple Macintosh, and General Motors versus Henry Ford (cf. Langlois 2002, 23–33). However, it is unclear how far the evolutionary argument for the prevalence of modularity carries over to the social realm.
3.5.2 Modularity and Social Mechanisms
Let us consider how the evolutionary argument for modularity described in section 3.4.2 might work with regard to social phenomena. As a first stab, consider the following suggestion. Modular social mechanisms contribute to the adaptability of the social groups containing them. Such groups would be able to adapt more quickly to modularly varying environments by altering one module while leaving the others the same or by rearranging modules. And, as in biology, modularly varying environments are a pervasive fact of social life: human social groups often need to develop the capacity to solve new problems while retaining most of their prior problem‐solving abilities. Thus, groups possessing modular mechanisms would be more likely to survive and produce “offspring” in the form of offshoot or copycat groups or organizations. However, there is reason for skepticism about this scenario.
The unit of selection in the scenario just described is the social group, and one important type of social group is the organization. In fact, there is a social science research program inspired by evolutionary biology in which the units of selection are organizations, namely, organizational ecology. Organizational ecology attempts to explain characteristics of various types of organizations—businesses, labor unions, advocacy groups, churches, and so on—in distinct contexts on the basis of differential mortality and founding rates (cf. Hannan and Freeman 1989; Aldrich 1999, 43–48). For example, one important thread in this literature examines the distinct environments to which generalist and specialist organizations are best suited, for instance, inquiring into the conditions in which consolidation among generalist organizations creates resource (p. 50 ) opportunities for specialists (cf. Carroll and Swaminathan 2000). Unfortunately, the scenario sketched in the foregoing paragraph contradicts one of the basic premises of organizational ecology: the structural inertia of organizations (Hannan and Freeman 1989, 70; Aldrich 1999, 45). According to this principle, the rate of change of an organization's structure is typically much slower than the rate of change in the environment. This premise is important for a model in which Darwinian selection is the driving force. Changes in populations of organizations result primarily from old organizations disbanding and being replaced by new ones that are better suited to the new environment rather than from individual organizations adapting themselves to new situations. There are a number of reasons why organizations would be expected to exhibit structural inertia (Hannan and Freeman 1989, 67–69). For example, restructuring often shifts resources away from a segment of the organization, and hence is likely to be resisted by those members who would be disadvantaged. Moreover, there is some empirical evidence in support of structural inertia (Aldrich 1999, 168). Thus, the proposal that highly modular, and therefore quickly changeable, organizations are favored by social selection processes is problematic.
Let us try a different approach. Modularity of social mechanisms need not entail that individual organizations be quick to adapt to changing circumstances. That point can be appreciated through a consideration of modular mechanisms in molecular biology. In that case, modularity is a matter of how the genome maps onto system components, not a claim that individual organisms can quickly adapt to new environments. The adaptation that modularity engenders, occurs across generations, not in the life history of a single organism. Thus, perhaps things work similarly in the social world. Consider two general ways in which this might happen.
First, consider social mechanisms that are internal to organizations. These mechanisms might include such things as a social hierarchy or an established production procedure. In this case, the argument would be that modularity facilitates evolvability because it allows mechanisms to be modified one component at a time or for solutions to new social problems to be found by rearranging mechanism components. This scenario is consistent with structural inertia, since the altered versions of the mechanism might occur in newly founded organizations rather than in transformed versions of older ones. In this scenario, nonmodular social mechanisms would be likely to go extinct in modularly varying environments, while the varied descendants of modular mechanisms would spread throughout the population of organizations. The plausibility of this scenario is enhanced by the wide prevalence of certain types of modular structures found in organizations and social groups in general, particularly hierarchies. For example, consider the hierarchical structure of a university: the university is divided into colleges or schools, which are in turn divided into departments or units. This structure is modular, since it allows alterations to be made to one unit (say, restructuring the (p. 51 ) philosophy department) while leaving other units as they were before. Likewise, although it would be difficult for an established university to, say, eliminate a number of existing departments or to restructure its colleges, a newly founded university might readily make such changes.
A second scenario concerns social mechanisms that are not internal to specific organizations, but instead are features of the broader social context in which organizations as well as individuals are embedded and interact. Forms of economic interaction, such as a market, are examples of social mechanisms of this kind. Again, the hypothesis would be that such mechanisms, if modular, are more adaptable to changing environments. As a result, such mechanisms would be expected to proliferate more widely than their nonmodular counterparts. An economic system based upon property rights and market exchange is arguably a modular mechanism, since it allows owners wide leeway to modify their properties or enterprises independently of others (cf. Langlois 2002, 26–27). Such a system also allows for rearrangement of modules in the form of consolidation or increasing specialization of industries. A more specific example is the contrast between traditional and Silicon Valley models of research and development (Aoki and Takizawa 2002). In the traditional model, R&D is carried out in an integrated manner within a particular firm, which organizes and directs coordinated R&D projects for specific goals. In this model, it is important that each of the various design teams knows what the others are doing, so that their results can be assimilated into the final product. Clearly, communication among design teams becomes increasingly cumbersome with the increasing complexity of the task of each. In the Silicon Valley model, by contrast, the product system is divided into modules developed by separate firms, often start‐ups funded by venture capitalists. The Silicon Valley model requires standardized interfaces between modules, so that improvements to the overall product system result primarily from independent improvements in the various components (Aoki and Takizawa 2002, 770–71). The advantage of the Silicon Valley model is that it avoids the onerous communication among design teams required by the traditional model, thereby facilitating quicker solutions to new problems. The Silicon Valley model, then, is an example of a modular mechanism that structures the interactions of a collection of organizations. But there is nothing in this scenario to require that individual organizations be highly adaptable.
The two scenarios described above illustrate ways in which the hypothesis about the advantages of modularity with regard to evolvability might be extended to social mechanisms. But the quantity of both theoretical and empirical research on these questions in social science is minuscule in comparison to the body of work on modularity and evolvability in biology. Robert Boyd and Peter Richerson (Richerson and Boyd 2005; Boyd and Richerson 2005) are the only authors I know of who have offered anything like a detailed evolutionary explanation of modularity in social science. Boyd and Richerson argue against the image of culture (p. 52 ) as a tightly integrated, holistic system (Richerson and Boyd 2005, 91–93), and they hypothesize that culture evolved as an adaptation to rapidly changing climates in the Pleistocene (ibid., 131–39). They develop models that illustrate how the cumulative social learning indicative of culture can be favored by natural selection in changing environments (Boyd and Richerson 2005, pt. I). The main theme of this account is that culture enhances adaptability by facilitating quick, though not necessarily optimal, solutions to new problems. Hence, Boyd and Richerson's hypothesis is very similar to the evolutionary account of modularity described in section 3.4.2. Nevertheless, the focus of Boyd and Richerson's work is explaining the origin of culture rather than modularity per se, and it is unclear to what extent their proposals could be developed to support the claim that specific types of social mechanisms are modular.
In the remainder of this section, I consider some possible reasons for thinking that social mechanisms may often be nonmodular. The first concern is based on the point that modularity is adaptively beneficial only in changing environments. Consequently, nonmodular designs may be preferable to modular ones in environments that exhibit a high degree of stability over time. Thus, there would appear to be no particular reason to expect modular social mechanisms in social contexts that have persisted without much change for a significant period. Richard Langlois suggests that certain nonmodular features of medieval European social structures were well suited to the stable social environment of this period, but eventually disappeared in the face of changing circumstances (2002, 28–29). Of course, the analogous point holds with regard to biology as well. Thus, the question here is to what extent past social and biological environments have been modularly variable rather than stable or simply chaotic. The next concern, however, is more specifically focused on characteristic features of human society.
A common challenge for social policy is that changes in one feature of a society may produce unpredictable changes elsewhere in the system, thus making it extremely difficult to anticipate the consequences of the policy intervention. One source of this difficulty is that participants in the system who are not directly targeted by the policy intervention may nevertheless be aware of it, and may perceive opportunities to advance their interests by modifying their practices in response to it. Indeed, the complex interrelation between social structures and awareness of those structures by members of the society is a common basis for arguments against the possibility of laws of social science (cf. Searle 1984; Taylor 1971). Although such arguments rarely use the term “modularity,” the modularity of social mechanisms is precisely what they aim to call into question. For if the objection is correct, it will typically not be possible to change one component of a social mechanism without producing unpredictable changes in the others.
This objection to the modularity of social mechanisms will be discussed in detail in Chapter 8. For the moment, I would like to indicate two points (p. 53 ) that would be relevant to any response to it. Whether a mechanism is modular with regard to an intervention depends on the intervention itself and on the manner in which the causal system is represented. For a given mechanism, some interventions may be modular while others are not. In Chapter 8, I call interventions that affect mechanisms in nonmodular ways structure‐altering. The second point is that even if an intervention is structure‐altering with regard to a mechanism, it might not be such with regard to other, more fundamental mechanisms that can explain why and how the first was altered. Thus, one natural response to the objection described in the foregoing paragraph is that the unintended effects of the policy intervention could be explained, and perhaps even anticipated, by individual‐level mechanisms. For example, a rational choice model might explain why an intervention that inadvertently creates new incentives leads to systematic but unintended changes of behavior. The thought that more fundamental, modular mechanisms can be described at finer‐grained levels of description is an underlying motivation of the mechanisms approach to extrapolation. It is also the central theme of Chapter 7, which discusses the relationship between mechanisms‐based extrapolation and reductionism.
This chapter began with the question of the relationship between mechanisms and the probabilistic causal concepts elaborated in Chapter 2, and it proposed the first part of an answer to this question. To the extent possible, mechanisms are to be identified with causal structure on the basis of domain‐specific empirical analyses. Since causal structure is that which generates probability distributions and provides information about how they change under interventions, this identification is a basis for linking mechanisms to probabilistic causal concepts. An important part of these empirical analyses consists of providing some general reason to think that mechanisms are modular, and evolutionary theory suggests a means of doing just this. However, this evolutionary argument is, at present, on firmer ground in molecular biology than in social science.
Yet the identification of mechanisms with causal structure alone indicates only that there is some connection between mechanisms and probabilistic causal concepts such as causal effect and positive causal relevance. It provides no indication of what the nature of that relationship is. Chapter 4 discusses a proposition, which I call the disruption principle, which says something specific about the link between probability and mechanisms identified with causal structure.
(1.) See Hausman (1998, 13–17) for a critique of the conserved quantity theory on the grounds that it fails to distinguish causally relevant and irrelevant interactions and fails to account for causal asymmetries.
(2.) In Carnap's (1936) version of this account, the antecedently understood terms were presumed to be drawn from an “observation language.” I follow Lewis (1970) and others (cf. Papineau 1996) in rejecting this requirement.
(3.) The transfer theory is, together with Salmon's (1984) proposal, one of the ancestral sources of Dowe's position. It is presented in Aronson (1971) and in Fair (1979). The transfer theory differs in some important respects from the conserved quantity theory, and unfortunately Dowe does not elaborate on how the Ramsey‐Lewis approach would work in the case of his own theory.
(6.) The term “Bayesian network” derives from the original (and continuing) use of directed graphs and probability distributions to implement expert learning and judgment in artificial intelligence (cf. Pearl 1988). In the context of causal inference, the name does not indicate a commitment to a Bayesian methodology.
(7.) In Simon's original parable, Hora and Tempus are interrupted by telephone calls, so that Tempus must continually restart construction from scratch, while Hora need only restart the last module. As Watson and Pollack (2005, 448) point out, it is difficult to interpret the original parable as an example of how modularity enhances (p. 213 ) evolvability, since it does not involve a search through a space of possibilities for a solution to a problem.