Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

Contents lists available at SciVerse ScienceDirect

Studies in History and Philosophy of Biological and Biomedical Sciences
journal homepage: www.elsevier.com/locate/shpsc

The causal structure of mechanisms
Peter Menzies
Department of Philosophy, Macquarie University, North Ryde, NSW 2109, Australia

a r t i c l e

i n f o

a b s t r a c t
Recently, a number of philosophers of science have claimed that much explanation in the sciences, especially in the biomedical and social sciences, is mechanistic explanation. I argue the account of mechanistic explanation provided in this tradition has not been entirely satisfactory, as it has neglected to describe in complete detail the crucial causal structure of mechanistic explanation. I show how the interventionist approach to causation, especially within a structural equations framework, provides a simple and elegant account of the causal structure of mechanisms. This account explains the many useful insights of traditional accounts of mechanism, such as Carl Craver’s account in his book Explaining the Brain (2007), but also helps to correct the omissions of such accounts. One of these omissions is the failure to provide an explicit formulation of a modularity constraint that plays a significant role in mechanistic explanation. One virtue of the interventionist/structural equations framework is that it allows for a simple formulation of a modularity constraint on mechanistic explanation. I illustrate the role of this constraint in the last section of the paper, which describes the form that mechanistic explanation takes in the computational, information-processing paradigm of cognitive psychology. Crown Copyright Ó 2012 Published by Elsevier Ltd. All rights reserved.

Article history: Available online xxxx Keywords: Mechanisms Mechanistic explanation Causal explanation Interventionism Structural equations Modularity

When citing this paper, please use the full journal title Studies in History and Philosophy of Biological and Biomedical Sciences

1. Introduction Recently, a number of philosophers of science have claimed that much explanation in the sciences, especially in the biomedical and social sciences, is mechanistic explanation; i.e. the kind of explanation that explains some phenomenon by revealing the mechanism that underlies it. A mechanism is, roughly speaking, a set of entities and activities that are spatially, temporally, and causally organized in such a way that they exhibit the phenomenon to be explained. Mechanistic explanation is said to be very different from covering law explanation. For instance, the aim of mechanistic explanation is not to bring a phenomenon under a covering law, but rather to reveal the mechanism underlying it. Again, mechanistic explanation does not work by showing how some phenomenon is to be expected on the basis of laws and antecedent conditions, but by showing how it is constituted by a set of entities engaging in distinctive activities and causal interactions. Moreover, while these activities and causal interactions conform to generalizations, these generalizations fall short of those invoked by covering law explanation in being narrow in scope, having exceptions, and being historically contingent.

The main works in this tradition are Simon (1969), Wimsatt (1974), Bechtel and Richardson (1993), Glennan (1996), Machamer, Darden, and Craver (2000), Craver (2007), and Bechtel (2008). These philosophers claim that recognizing this distinctive form of explanation enables us to understand many important aspects of explanatory practice in these scientific fields. I agree with them on this point and think that this recent focus on mechanistic explanation has been very fruitful in illuminating many hitherto overlooked aspects of scientific explanatory practice. Nonetheless, I shall argue in this paper that the account of mechanistic explanation provided in this tradition has not been entirely satisfactory, as it has neglected to describe in complete detail the crucial causal structure of mechanisms and of mechanistic explanation. In order to highlight this neglect I focus on the work in this tradition that offers the most detailed account of the causal structure of mechanistic explanation: Carl Craver’s influential 2007 book Explaining the Brain. Many aspects of Craver’s interesting account of mechanistic explanation in this book are, I think, correct. Even so, I argue, his account of the causal structure of mechanisms is deficient in important respects and erroneous in others.

E-mail address: peter.menzies@mq.edu.au 1369-8486/$ - see front matter Crown Copyright Ó 2012 Published by Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.shpsc.2012.05.008

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

2

P. Menzies / Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

Phenomenon

S ψ-ing
X2 φ2-ing X1 φ1-ing X3 φ3-ing X4 φ4-ing

Mechanism
Fig. 1. A phenomenon (top) and its mechanism (bottom). Modified from Craver (2007, p. 7).

This paper aims to do more than to scrutinize Craver’s account. It aims to provide a better account of the essential causal structure of mechanistic explanation by enlisting an interventionist account of causation. Craver makes a start on this project, but does not carry it as far as it needs to go. I aim to show how the interventionist approach to causation, especially within a structural equations framework, provides a simple and elegant account of the causal structure of mechanisms. This account, I go on to show, explains the many useful insights of Craver’s account, but also helps to correct the omissions of his account. One of these omissions is his failure to provide an explicit formulation of a modularity constraint that plays a significant role in mechanistic explanation. The modularity requirement states that a mechanism should be decomposable into components whose operations are independently disruptable. One virtue of the interventionist/structural equations framework is that it allows for a simple formulation of this constraint. I illustrate the role of this constraint and the general account of mechanisms in the last section of the paper, which describes the form that mechanistic explanation takes in the computational, information-processing paradigm of cognitive psychology. 2. Craver’s example of mechanistic explanation Craver’s account of mechanistic explanation is straightforward. A mechanistic explanation explains some phenomenon by revealing its underlying mechanism, which consists in ‘‘a set of entities and activities organized such that they exhibit the phenomenon to be explained’’. (2007, p. 5) Craver depicts the abstract structure of a mechanism with the following diagram (Fig. 1): In this figure, the system S’s wÀing is the phenomenon to be explained. X1, . . ., X4 are constituent parts of S and /1, . . ., /4 denote activities engaged in by these constituent parts. The arrows represent causal relations.1 The mechanism explains S’s wÀing by virtue of the fact that the spatial and hierarchical organization of S’s parts and the temporal and causal organization of their activities constitute S’s wÀing. (2007, p. 8) Craver gives a number of illuminating examples of mechanistic explanation. In one example (2007: pp. 5–6, 22–6) the explanandum phenomenon is the release of neurotransmitters by a neuron when depolarized. A much simplified mechanistic explanation of this phenomenon might go like this: the arrival of an action potential at the neuron’s axon terminal opens calcium channels in the neuron’s presynaptic membrane; the influx of calcium increases calcium concentration in the interior of the cell, which causes

calcium-sensitive proteins attached to vesicles in the interior to change shape; their changed shape enables the vesicles to dock and fuse with the membrane; and this fusion causes the vesicles to release neurotransmitters into the synaptic cleft. In this example the component entities are axon terminals, synaptic clefts, calcium ion channels, vesicles, and membranes and so on. The mechanistic explanation of a neuron’s release of neurotransmitter proceeds by describing the activities of these entities—activities such as ion channels opening, proteins attaching to vesicles, and vesicles docking and fusing with membranes—and, most importantly, by describing how these activities are causally organized in such a way that the depolarisation of the axon terminal leads to the release of the neurotransmitters. 3. Craver’s account in more detail Such examples help to reveal the general shape of Craver’s account of mechanistic explanation, but certain elements of his account still stand in need of clarification. For example, what kind of thing is the explanandum phenomenon? What kind of thing are activities and how do they relate to causal relations? How must the entities and their activities be organized in order to exhibit the explanandum phenomenon? How does a mechanism explain a phenomenon by constituting it? In fleshing out his conception of mechanistic explanation, Craver answers some of these questions. In this section I review the various causal features he ascribes to a mechanism with a view to providing a more systematic theory of the causal structure of a mechanism in the next section. I structure my discussion around three questions. 3.1. What is the explanandum phenomenon? Craver begins his characterization of the explanandum phenomenon with Cummins’ (1983) suggestion that it typically consists in a range of capacities of some system. (2007, pp. 122–8) However, he elaborates on this suggestion by noting that the relevant capacities need not be restricted to simple input-output conditions, but may include a system’s responses to various precipitating, inhibiting, modulating conditions, and non-standard conditions. For example, in the case of the release of neurotransmitters, we may be interested in the different ways in which action potentials can be formed (precipitating conditions), the different kinds of preventers of the release of neurotransmitters (inhibiting conditions), the different conditions affecting the quanta of neuro-

1 Confusingly, Craver says that the activities are represented by the /-ings (2007, p. 7) and in the same paragraph that they are represented by the arrows. He must mean that the /-ings are the activities, the arrows are causal relations, and that the activities logically imply the existence of causal relations.

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

P. Menzies / Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

3

transmitter released (modulating conditions), and conditions affecting neurotransmitter release outside the normal range. Also he notes that the manifestations of capacities can be described with greater or lesser precision. For example, the manifestation of an action potential can be described in terms of its rate of rise, peak magnitude, rate of decline, refractory period, and so on. Depending on how the output condition is specified, different explanatory mechanistic models will be required. In general, however, none of Craver’s qualifications compels us to depart from the spirit of Cummins’ proposal that the target phenomenon of a mechanistic explanation typically consists in an array of capacities of some system, specified in terms of (non-backtracking) counterfactuals that connect input and output conditions described in greater or lesser precision. 3.2. What are activities? As with many accounts of mechanistic explanation, the concept of an activity plays a crucial role in Craver’s account. He describes activities in these terms: Activities are the causal components in mechanisms. I use the term ‘activity’ as a filler-term for productive behaviours (such as opening), causal interactions (such as attracting), omissions (as occurs in cases of inhibition), preventions (such as blocking), and so on. In saying these activities are productive, I mean that they are not mere correlations, that they are not mere temporal sequences, and, most fundamentally, that they can be exploited for the purposes of manipulation and control. (2007, p. 6) Many mechanists take activities to be conceptual and ontological primitives in their scheme; and they tend to restrict activities to productive processes involving contact forces, or attraction and repulsion, or transmission of energy-momentum. (Cf. Machamer et al. (2000).) In contrast, Craver makes it clear in the passage quoted that activities for him can include negative causal relations such as those involved in inhibition and prevention. The quoted passage also makes clear that activity statements describe causal relations, and not mere correlations or temporal sequences. In order to distinguish causal sequences from these other kinds of sequences, he adopts Woodward’s (2003) strategy for drawing the distinction in terms of interventions. To illustrate Woodward’s strategy consider a familiar example: smoking causes yellow-stained fingers and also lung cancer so that we find a correlation between having stained fingers and lung cancer in the population of smokers. But this is spurious, or non-causal, correlation, as is brought out by the fact that intervening to bring about yellow-stained fingers in people would have no effect on the incidence of lung cancer among those people. In contrast, the correlation between smoking, on the one hand, and having stained fingers and getting lung cancer, on the other hand, is a genuine causal correlation. This is borne out by the fact that inducing people to smoke would induce a higher rate of stained fingers and lung cancer among them. Woodward’s strategy for drawing the distinction between correlation and causation relies on the commonsense insight that an intervention on a cause yields changes in its effects, whereas an intervention on a joint effect of a common cause does not yield changes in the other effect. More generally, Craver adopts Woodward’s interventionist approach to causation, which involves defining the concept of causal relevance in terms of an intervention. The idea behind Woodward’s approach is that causal relations are distinctively exploitable for

the purposes of manipulation and control. More specifically, he defines (2003, Chap. 2) variable X as causally relevant to variable Y just in case if an intervention were to change X it would also change Y.2 So, for example, a rise in intracellular calcium ions is causally relevant to the release of neurotransmitters because if an intervention were to raise the level of intracellular calcium ions, neurotransmitters would be released. In contrast, a rise in intracellular sodium ions is merely correlated with the release of neurotransmitters because if an intervention were to raise the level of intracellular sodium ions, neurotransmitters would not be released. The crucial notion in Woodward’s approach is that of an intervention (2003: Chap. 3). Roughly speaking, an intervention on a variable X with respect to Y is a hypothetical experimental manipulation of X that is ideal for determining its causal influence on Y. Thus, an intervention on X is an exogenous cause that completely determines the value of X in such a way that any change it brings about in X is not correlated with any other cause of Y (except for any causes that mediate the X-Y pathway). So for example, an experimental manipulation of X that sets its value on the basis of a random process such as a coin toss would count as an intervention on X with respect to Y since the coin toss would not be correlated with any cause of Y. Consequently, if X is correlated with Y under these experimental conditions, the correlation must be due to the fact that X is causally relevant to Y. Woodward’s precise definition of an intervention (2003, Chap. 3) does not restrict interventions to human manipulations. For example, a stroke that damages a brain region can count as a ‘natural intervention’ on the brain region’s functioning. The precise definition also makes it transparent that the concept of an intervention is a causal notion, so that the definition of causal relevance in terms of interventions is not presented as a reductive conceptual analysis. By invoking Woodward’s interventionist criterion of causal relevance, Craver makes substantial progress over other accounts of mechanistic explanation. For example, Machamer et al. (2000) describe mechanisms as partly constituted by ‘‘activities productive of regular changes’’, but do not say what productive activities are. They endorse Anscombe’s remark that the word ‘‘cause’’ is highly general and only becomes meaningful when filled out by other more specific causal verbs, e.g., scrape, push, carry, eat, burn. But this begs the question what all these activities have in common that makes them causal. In contrast, Craver goes beyond these platitudinous remarks by providing an informative characterization of the causal component of activities or, in other words, what makes activities causal. But in the process he has made the usual appeal to the concept of an activity otiose. If we can explain the crucial causal nature of activities in terms of an interventionist account of causation, we can employ this understanding directly in our account of the causal structure of a mechanism without the obfuscating detour through the concept of an activity. 3.3. How are the component entities and their activities organized so as to exhibit the explanandum phenomenon? In discussing the role of organization in mechanistic explanations, Craver emphasizes the spatial organization of the component entities (their sizes, shapes, locations, orientation, directions, and connections) and also the temporal organization of their activities (their order, rate, and duration). But he says very little about their causal organization. This is an unfortunate omission since it is clear that the entities and their activities have to be causally organized in a very specific way in order to exhibit the capacities of the

2 This definition corresponds to only one of a number of causal concepts Woodward defines, including the concepts of total, direct, and contributing causes. For our purposes the distinctions among these concepts do not matter. Woodward also extends the definition of causal relevance to probabilistic causal contexts by stating that X is causally relevant to Y just in case if an intervention were to change X, it would also change the probability distribution of Y. To simplify our discussion I restrict our discussion to deterministic causal contexts.

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

4

P. Menzies / Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

system under consideration. What is distinctive or essential to the causal organization of a mechanism? Despite the absence of any explicit discussion of this issue, Craver does discuss at length an issue that is indirectly related to it. This issue is: Which entities and which of their activities count as constituent parts of a mechanism? In answering this question, he formulates an account of what he calls constitutive relevance. This is one of the most original and interesting parts of his theory of mechanistic explanation. (2007, pp. 139–60) His account of constitutive relevance emphasizes the role of inter-level intervention experiments in discovering the internal components of mechanisms. He distinguishes bottom-up inter-level intervention experiments from top-down ones. In a bottom-up experiment, one intervenes to change a component of a mechanism to detect the downstream effects on the whole mechanism. Examples of such bottom-up experiments are interference experiments that lesion a portion of the brain to detect effects on task performance; and stimulation experiments that stimulate a portion of the brain to test the effect on motor behaviour. In contrast, in a top-down experiment, one intervenes to change the input to the whole mechanism to detect downstream changes in the components of the mechanism. Examples of top-down experiments include activation experiments in which a cognitive system is activated to detect markers of activity in regions of the brain. Fig. 2 below represents these two kinds of inter-level experiment. Drawing on the norms of constitutive relevance implicit in these experimental strategies, Craver develops an account of constitutive relevance. (2007, pp. 139–160) In slightly simplified form, the account states that X’s /-ing is a component in a mechanism for a system S’s wÀing if and only if (i) X is part of S; (ii) one can change S’s wÀing by intervening to change X’s /-ing; and (iii) one can change X’s /-ing by intervening to change S’s wÀing. He argues that condition (ii) (when conjoined with (i)) is not sufficient for constitutive relevance because it is satisfied by background conditions: lesioning the heart can prevent word-stem completion, but the heart is not part of the cognitive mechanism for word-stem completion. Correspondingly, condition (iii) (when conjoined with (i)) is not sufficient for constitutive relevance because it is satisfied by mere correlates: having a subject perform a word-completion task may increase blood flow, but blood flow is not part of the mechanism for word-completion. To overcome the deficiencies of conditions (ii) and (iii), Craver conjoins them in what he calls a mutual manipulability account that requires both conditions to hold in order for X’s /-ing to be constitutively relevant to S’s wÀing. There are, however, several unclarities in Craver’s account. For example, it is not immediately evident what the phrase ‘‘changing S’s wÀing’’ means, if we are to think of S’s wÀing as an exercise of a capacity. Presumably, we are to understand ‘‘change S’s wÀing’’ as meaning changing the input condition of w in condition (ii), and as meaning changing the output condition in condition (iii). It is also unDetection

clear whether X’s /-ing fails to be constitutively relevant to S’s wÀing when either condition (ii) or (iii) obtains but the other does not. Craver says that X’s /-ing counts as constitutively irrelevant to S’s wÀing when both conditions (ii) and (iii) fail to obtain, but leaves it indeterminate whether constitutive relevance or irrelevance results from the situation in which one obtains while the other does not. At any rate, Craver’s account of constitutive relevance highlights the importance of identifying the component parts of a mechanism. He is surely correct in supposing that the purpose of the different kinds of intervention experiment is to identify those entities and activities that are parts of the mechanism underlying some phenomenon. But there is much about which his account is silent. For example, the account lacks any discussion of the causal structure of a mechanism. It helps to identify the parts of a mechanism, but it does not tell us how these parts must be causally organized in order to function as a mechanism. Nor does it tell us how the causal structure of a mechanism enables it to exhibit or constitute the phenomenon to be explained. Furthermore, Craver’s account makes certain assumptions about causal structure without making them explicit. For example, his description of inter-level intervention experiments presupposes that mechanisms are modular in the sense that it is possible to intervene to change the behaviour of one component without interfering with the behaviour of others. So, for instance, a bottom-up intervention experiment that tests whether X’s /-ing is a component of the mechanism for S’s wÀing assumes that one can change X’s /-ing without changing any other component (except for its effects). If an intervention on X’s /-ing affected other components as well, one could not be sure that a change in S’s wÀing was due to the change in X’s /-ing rather than some other component. As we shall see, the modularity assumption is a significant assumption, which ought to be made explicit in a satisfactory account of the causal structure of a mechanism. Can we characterize the causal structure of a mechanism more directly and more explicitly? 4. A structural equations account of mechanism I believe we can provide a more systematic and accurate account of the causal structure of a mechanism that builds on Craver’s account by deploying the interventionist/structural equations framework for causation. Here I draw on the work of Pearl (2000), Spirtes, Glymour and Scheines (2000), and Woodward (2003), who have done much to clarify the basic ideas of this framework. In this framework, causal relations are represented within a causal model. Formally, a model consists of an ordered pair <V, E>, where V is a set of variables and E is a set of structural equations. It is best to explain these elements in terms of an example such as Craver’s example of neurotransmitter release. Simplifying drastically, we might represent the causal situation here in terms of the following variables, which can take one of two values:

Intervention

Intervention

Detection

Fig. 2. Bottom-up experiment (left) and top-down experiment (right). Modified from Craver (2007, p. 146).

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

P. Menzies / Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

5

AP

CC

P

F

NR

Fig. 3. Causal graph for neurotransmitter example with no interventions.

AP = 1 if action potential arrives at axon; 0 otherwise. CC = 1 if calcium channels open in cell membrane; 0 otherwise. P = 1 is calcium sensitive proteins attached to vesicles change shape; 0 otherwise. F = 1 if vesicles fuse with presynaptic membrane; 0 otherwise. NR = 1 if neurotransmitters in vesicles are released into synaptic cleft; 0 otherwise. Of course, a realistic model would contain many more variables— many-valued variables as well as binary variables. But a muchsimplified model has expository advantages. Apart from variables, the other element in a model is a set of the structural equations, which are of two kinds. The structural equations for the exogenous variables state the actual value of each such the variable. In this case, there is only one exogenous variable and its structural equation states: AP = 1. The structural equations for the endogenous variables state how the value of each such variable is determined on the basis of the values of other variables, both exogenous and endogenous. Such a structural equation has the form Y = f(X1, . . ., Xn), where Y is the endogenous variable and X1, . . ., Xn are all its direct causes. The structural equations for the endogenous variables in the neurotransmitter example are very simple: CC = AP P = CC F=P NR = F These particular structural equations state that the value of each endogenous variable is determined very simply by the value of one other variable. In a more realistic model, the structural equations would describe the endogenous variables as functions of many causally relevant variables. Pearl and Woodward espouse the view that in order for these equations to capture causal relations correctly they must hold invariantly under interventions, or, in other words, the equations must continue to hold not only when the variables on the right-hand side take on values in the normal course of events, but also when these variables have their values set by a range of possible interventions. As we have seen earlier, the motivation for this requirement is that what distinguishes causal relations from mere correlations is that they are robust under manipulations of the causes. If an individual structural equation holds invariantly under a range of interventions, Pearl and Woodward say that the equation describes a causal mechanism of the system in question. Their terminology is potentially confusing in the present context, as it differs from the terminology adopted here. I apply the term ‘causal mechanism’ to a system of causal relations described by a complete set of structural equations rather than a single structural equation. For this reason I adopt a different term and say that a structural equation that holds invariantly under the appropriate range of actual and possible interventions describes a causal capacity of the system. This terminology is motivated by the fact that a structural equation that is invariant under interventions implies a battery of interventionist counterfactuals. For example, suppose that a structural equation Y = f(X1, . . ., Xn) holds robustly under an intervention that sets the values of the variables at X1 = x1, . . ., Xn = xn. Then this structural equation implies a true interventionist counterfactual ‘‘If it were the case that X1 = x1, . . ., Xn = xn, then it would be that

Y = f(X1 = x1, . . ., Xn = xn)’’. Accordingly, it is appropriate to describe the structural equation as encoding the capacity of the system to respond with values of Y given actual and possible values of X1, . . ., Xn. The procedure for evaluating an interventionist counterfactual given a set of structural equations is simple. To evaluate a counterfactual of the form ‘‘If it were the case that X = x then it would be that Y = y’’, one replaces the equation in the set for X with a new equation stipulating X = x, and then solves the new set of equations to determine whether Y = y holds; if it does, the counterfactual is true and otherwise false. For example, the original set of equations for the neurotransmitter example, which includes the equation for exogenous variable AP = 1, determine that all the other variables have the actual value 1. But if we wish to consider what would happen if the variable CC took the value 0, we must replace the equation CC = AP with the equation CC = 0, and then solve the remaining equations. It will follow obviously that P, F and NR will all take the value 0 even though AP is still set at its actual value 1. This means that the counterfactual ‘‘If it were the case that CC = 0, then it would be that NR = 0’’ would be true. By replacing the old equation for CC (ie CC = AP), we imply that the intervention that sets the value of CC at 0 overrides the way in which CC is normally determined on the basis of AP. The causal relationships in a model can be depicted in a directed graph. An arrow is drawn from one variable to another when the first appears on the right-hand side of the equation for the other. A series of arrows from tip to tail constitutes a path through the graph. A graph is acyclic just in case no path loops back on itself. The directed graph for the neurotransmitter example is given in Fig. 3 below. It is also possible to represent the result of an intervention graphically. For example, the graph of the causal structure resulting from an intervention on the variable CC is represented in Fig. 4. The breaking of the arrow from AP to CC signifies that the normal causal influence of AP on CC is overridden by the intervention on CC. In order for a set of structural equations to provide an accurate description of causal relations, Pearl (2000) and Woodward (2003) impose a structural requirement on the equations they call a modularity constraint. A set of equations is modular if and only if it is possible to intervene on a variable on the left-hand side of an equation without disturbing the other equations in the set, i.e. without rendering the other equations false. Pearl and Woodward state that a set of equations is causally correct only if they are modular. They are motivated to impose this requirement by the thought that each equation should represent the operation of a distinct causal capacity; and further that a necessary condition for two causal capacities to be distinct is that it should be possible to disrupt the operation of one without disrupting the operation of the other. As we shall see in the next section, this necessary condition does seem to reflect the reasoning manifested in the explanatory practices of the cognitive sciences. It follows that a causally correct system of equations will meet the modularity constraint. Given the conceptual machinery provided by the structural equations framework, it is possible to characterize the causal structure of a mechanism in a precise way. Suppose that we seek to explain a system’s capacity O = f(I1, . . ., In) in terms of its underlying mechanism. Such a mechanism is represented by a modular set of structural equations E1, . . . Em, each of which correctly describes a causal capacity of the system and all of which compose to yield the function O = f(I1, . . ., In).3 Put more directly, the causal structure

3 The account of the causal structure of a mechanism given here depends on the assumption that the causal structure is sufficiently deterministic that it can be represented by structural equations. However, it is possible to give an alternative account that accommodates the possibility of probabilistic causation where this assumption is implausible. This alternative account would define the causal structure of a mechanism in terms of transitive relationships of direct causation lying between the input and output variables of the explanandum phenomenon, where the relationships of direct causation are explicated by probabilistic interventionist counterfactuals.

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

6

P. Menzies / Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

Intervention

AP

CC

P

F

NR

Fig. 4. Causal graph for neurotransmitter example with intervention on CC.

X

AP

Y

NR

Z
Fig. 5. Causal structure of neurotransmitter example with more complex mediating pathways.

of a mechanism is given a by a set of modular subcapacities whose sequential exercise has the input-output profile of the capacity to be explained. This account of mechanism is very similar to one given by Woodward (2002) with some slight differences of formulation due to the fact that he does not explicitly invoke the machinery of structural equations. However, both Woodward’s account of mechanism and the one proposed here emphasize the importance of interventions in identifying the internal causal structure of mechanisms and the importance of the modularity constraint as a structural requirement on mechanisms. Consider how this account applies to the example of neurotransmitter release. The mechanism underlying the release of the neurotransmitters is described by the set of structural equations CC = f(AP), P = g(CC), F = h(P) and NR = i(F). Each of these describes a causal capacity of the system and, taken sequentially, they compose to yield the explanandum capacity; i.e., these functions compose to yield NR = i(h(g(f(AP))) = AP. In this case, the mechanism consists in a simple path of causally related variables, and so the recipe for composing functions is simple. But the same account applies to mechanisms involving more complex causal structures. Suppose that a neuron’s capacity to release neurotransmitters when an action potential arrives at its axon, i.e. the capacity captured by the equation NR = AP, were mediated by three causal pathways, as represented in the following set of equations and causal graph. AP = 1 X = f(AP) Y = g(AP) Z = h(AP) NR = i(X, Y, Z) The mechanism (Fig. 5) underlying this capacity would consist of the subcapacities described by these equations provided that they compose in the right way: i.e. NR = i(f(AP), g(AP), h(AP)) = AP. One thing that this way of representing mechanism reveals is that it is important that a sequence of causal capacities described by structural equations do not by themselves constitute a mechanism. A familiar counterexample to the transitivity of causation or causal relevance makes this clear. As a person is walking along a mountain path, his enemy pushes a boulder down the mountain with the aim of killing him. However, the person walking on the path sees the boulder, ducks, and survives. If we use the variables P, D and S to represent the enemy’s pushing the boulder, the walk-

er’s ducking, and the walker’s surviving, we can write the structural equations as P = 1; D = P; S = $P v D. We can see that the enemy’s pushing the boulder causes the walker to duck and the walker’s ducking causes him to survive. But the enemy’s pushing the boulder does not cause the walker to survive. For whether or not the enemy were to push the boulder, the walker would survive. This is brought out by the fact that when we compose the structural equations, we obtain the result S = $P v P, which implies that S gets the value 1 whatever the value of P. The point that this example brings home is that a sequence of capacities constitutes a mechanism only if they compose to yield a genuine causal capacity that supports a difference-making causal relevance relation. This account of a mechanism can be inserted into the framework of mechanistic explanation as before: mechanistic explanation aims to deliver understanding of some phenomenon by revealing its underlying mechanism. The resulting account explains many features of Craver’s account, but also delivers insights that Craver’s account does not. One similarity between the accounts is that they employ an interventionist account of causation in detailing the causal structure of a mechanism. When Craver’s account is shorn of the incidental remarks about activities as embodying the causal features of a mechanism, it implies that the causal structure of a mechanism is given by relations of causal relevance between variables, as depicted in a causal graph. One respect in which the present account is illuminating in a way that Craver’s account is not concerns the question of how a mechanism accounts for or explains the explanandum phenomenon. As Craver himself acknowledges, a model of mechanistic explanation should illuminate how a mechanistic explanation does its explanatory work. In the covering law model of explanation, a set of laws and antecedent conditions explain some phenomenon by virtue of logically entailing them, or at least propositions describing them. On his alternative model, Craver says that a mechanism explains some phenomenon by constituting it. But what kind of constitution is involved here? There are many different kinds of constitution relation: a set is constituted of its members; a statue is constituted by the marble out of which it is carved; an event of writing ‘dog’ is constituted by writing the letters ‘d’, ‘o’, and ‘g’. The present account sheds some light on this question, since it states that an explanandum capacity, as expressed in a structural equation, is constituted by series of subcapacities, described by other structural equations, in the sense that latter equations compose to yield the former equation. The form of constitution is that involved in the composition of functions, or capacities.4 Another respect in which the present model of mechanisms and mechanistic explanation is superior to Craver’s is that it provides a more comprehensive account of constitutive relevance i.e. an account of which entities and activities count as parts of a mechanism. We saw in the last section that his mutual manipulability account of constitutive relevance involves some unclarities and complexities. In contrast, the present account delivers a simple account of constitutive relevance: any variable that lies on a pathway between the input variable and the output variable of the capacity to be explained counts as part of the mechanism underlying the capacity. This account subsumes Craver’s account while being more liberal about what can count as constitutively relevant to a mechanism. Any activity that is revealed by his mutual manipulability account to be constitutively relevant to a mechanism will be represented by a variable on a pathway between the input and output variables of the capacity to be explained. In these cases the intermediate variables will be at a lower level than the input and

4 It is important to distinguish an account of the way a mechanism constitutes an explanandum phenomenon from an account of the way entities and activities are constitutively relevant to a mechanism. The account of constitution in terms of composition of functions given in this paragraph is supposed to be an account of the first kind.

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

P. Menzies / Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

7

output variables. This is illustrated by the example of the neurotransmitter release where the intermediate variables represent activities of calcium channels, proteins and vesicles that are parts of the axon whose capacities are being explained. But in other cases the present account will allow constitutively relevant intermediate variables to be at the same level as the input and output variables of the capacity to be explained. An example of this might be a mechanistic explanation of a person’s mental capacity that identifies other mental states of the person as intervening variables that connect the input and output variables of the capacity.5 This account sidesteps some of the problems Craver’s mutual manipulability account is intended to overcome. First, Craver argues that condition (ii), describing the effect of a bottom-up experiment, must be supplemented by condition (iii), describing the effect of a top-down experiment, because (ii) by itself is satisfied by a background condition. For example, lesioning a subject’s heart would impair his ability to do a word-completion task, but the normal functioning of his heart is a background condition that is not part of the mechanism by which the subject executes this task. There is no need on the present account, however, to introduce a special condition to exclude such cases. In a typical informationprocessing account of word-completion that one finds in cognitive psychology, the normal functioning of the subject’s heart is not included as a relevant variable because such models typically include only those variables with values that represent ‘serious possibilities’. (See Woodward, 2003, section 2.8) Because such models do not countenance as a serious possibility the idea that the subject will fail to have a normal functioning heart, they don’t include this background condition as a variable in the model and so they don’t count it as a part of the mechanism underlying a subject’s ability to do a word-completion task.6 Another problem that Craver’s mutual manipulability account is intended to deal with is ‘the problem of mere correlates’. Craver notes that condition (iii) of his account does not, by itself, rule out that of a mere correlate counting as a part of a mechanism: for example, having a subject perform a word-completion task may increase blood flow, but blood flow is not part of the mechanism for word-completion. For this reason, Craver believes that condition (iii) must be conjoined with condition (ii) (and with condition (i)) to yield an adequate account of constitutive relevance. However, it is again unnecessary on the present account to add a special condition to rule out mere correlates as constitutively relevant to some capacity. The present account eliminates the possibility of blood flow counting as a part of the mechanism of word-completion: the usual requirements of causal relevance ensure that a variable does not count as part of the mechanism underlying the wordcompletion capacity even if it is correlated with other variables that are part of the mechanism. Finally, the present account goes beyond Craver’s account of constitutive relevance in stating not only which things count as the components of a mechanism but also how the components must be causally organized in order to count as mechanism. On

the present account, the components of a mechanism are the variables lying on the pathways mediating the input and output conditions of the capacity to be explained; and these variables must be linked by way of modular subcapacities whose sequential exercise yields the capacity to be explained. This description of the causal organization of a mechanism is very abstract. In the next section I try to make it more concrete by looking at an example taken from cognitive psychology. 5. An example from psychology The present account of mechanisms is closely related to Robert Cummins’ theory of functional analysis, which consists ‘‘in analysing a disposition into a number of less problematic dispositions such that programmed manifestation of these analysing dispositions amounts to a manifestation of the analyzed disposition.’’ (1983, p. 28) Clearly, this resembles a mechanistic explanation of some capacity that consists in analysing it into modular subcapacities whose sequential execution reproduces the input-output profile of the capacity to be explained. Cummins’ focus is on psychological dispositions. He maintains that his account of functional analysis is compatible with many different explanatory frameworks in psychology: the belief-desire psychology framework, the computational information-processing framework, the connectionist framework, and neuroscientific frameworks Cummins (2000). All these frameworks pursue the goal of functional analysis of dispositions, but differ from each other with respect to the kinds of entities and kinds of analysing dispositions they appeal to. Alternatively, as I would express things, these different explanatory frameworks all pursue mechanistic explanation but employ different kinds of models with different ranges of explanatory variables and structural equations. Nonetheless, despite its broad applicability, I focus in this section on mechanistic explanation or functional analysis as it is applied within the computational, information-processing framework. In this framework a mechanistic explanation is often represented in terms of a box-and-arrow diagram rather than a causal graph. An information-processing capacity is represented as a box with informational inputs and outputs; and specialized causal relations mapping the flow of information are represented by arrows. A mechanistic explanation consists in decomposing one information-processing capacity into a nested series of simpler information-processing subcapacities. This is sometimes expressed in terms of the metaphor of opening up a black box or module to reveal its inner working in terms of smaller black boxes or modules. The term ‘‘module’’ here is used in its minimal sense to mean a functionally individuated capacity. This kind of mechanistic explanation of a cognitive capacity is a clear illustration of the proposed account of mechanistic explanation. Each box with its input and output arrows corresponds to a structural equation; and the decomposition of a module into a se-

5 I note that Craver himself would object that this account does not represent an account of constitutive relevance, on his understanding of the notion. According to his account, constituent relevance is always an interlevel relation and so would automatically rule out, as constitutively irrelevant, variables that exist at the same level as the explanandum capacity. However, I see no reason to restrict the notion of constitutive relevance in this way. There is a perfectly good usage according to which, for example, an explanation of a mental capacity in terms of causally intervening mental states and processes at the same level is a mechanistic explanation and the intervening mental states and processes are identified as constitutively relevant to (in the sense of being part of) the mechanism underlying the capacity. Note that this liberalisation of the notion of constitutive relevance does not blur or ignore the distinction between causal and constitutive relevance that Craver draws. Intervening variables on a causal pathway between input and output conditions of an explanandum capacity are constitutively relevant to the mechanism by virtue of being parts of the pathway that is the mechanism. Notwithstanding this, they are also causally relevant to the input and output variables by virtue of being caused by the input variables and of causing the output variables. While constitutive relevance is a different relation from causal relevance, a variable can enter into both kinds of relation at the same time. 6 There is no hard-and-fast set of rules for determining whether a possibility is a serious possibility. It seems to depend on a diverse range of contextually varying considerations to do with whether the possibility is probable, technologically feasible, and even morally significant. (For more discussion of these issues, see Woodward, 2003, section 2.8) Note that there are other ways in which an interventionist framework can distinguish background conditions from variables that are regarded as more significant for an effect. For example, interventions on background conditions tend to have unspecific effects. For example, lesioning the heart will affect word-stem completion but many other capacities too. Interventions on background conditions also seem to have switch-like effects rather finely modulated effects. For example, intervening to slow the heart up to a certain threshold does not affect the capacity for word-stem completion but beyond that threshold slowing the heart will eliminate the capacity altogether. (For more discussion see Craver 2007, p. 158.)

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

8

P. Menzies / Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

quence of smaller modules corresponds to the analysis of a structural equation into intermediate structural equations that, when composed, yield the original equation. Mechanistic explanation in the computational information-processing framework also illustrates the importance of the structural requirement of modularity to such explanation. The concept of modularity and the related concept of a module are invested with considerable theoretical significance in cognitive psychology and so are much contested. Consequently, it is worthwhile separating out the different concepts of modularity that are current in cognitive psychology to see how they relate to the concept of modularity that derives from causal reasoning. The concept of modularity has been hotly debated in philosophy and cognitive psychology ever since Jerry Fodor’s book on the subject (1983), in which he argued that we should think of cognitive modules as typically having a cluster of features. In his sense, a module is typically autonomous (independently disruptable), domain specific (responsive to a distinctive class of stimuli), innately specified, informationally encapsulated (having access only to information stored in the module), mandatory (not under conscious control), fast, and hardwired (neurally specific). In later work (2000, p. 63), he claims that information encapsulation is at the heart of modularity so that by definition a module’s operation is insensitive to information stored outside the module. Given plausible assumptions, it is reasonable to think that other features such as mandatoriness and fastness follow from informational encapsulation. (See Robbins (2009).) Fodor’s sense of modularity is evidently different from the sense deriving from causal reasoning. As we have seen, the structural equations framework spells out the relevant sense of modularity in terms of independent disruptability: structural equations describe distinct capacities or modules just in case it is possible to intervene on the left-hand side of one equation without disrupting the other equations, or, in material mode, two modules are distinct just in case it is possible to disrupt the operation of one with affecting the operation of the other. Of the concepts that Fodor associates with modularity only the concept of autonomy corresponds to modularity in the sense of independent disruptability. This is also referred to as dissociability: a capacity is dissociable to the extent that it can be selectively affected, damaged or disabled, with little or no effect on other capacities in the same system. What are the logical connections between independent disruptability and the other features Fodor associates with modularity? In particular, what is the connection between independent disruptability and the crucial concept of informational encapsulation? It is reasonably clear that independent disruptability does not directly imply informational encapsulation. Thus, suppose that some cognitive capacity is modular in the sense that its operation can be disrupted without affecting the operations of other modules in the cognitive system. It does not necessarily follow from this that the capacity has access only to limited information, for example information stored in a proprietary data-base. For example, there is evidence to suggest that people have a capacity for recognizing printed words that is dissociable from other cognitive capacities. However, consistently with this, people exercise this capacity in such a way as to call on a broad range of information. This is suggested by subjects’ performance on the Stroop task in which a word printed in a particular colour is shown and the subject is required to name the ink colour. When the stimulus word is a colourrelated word, naming the colour of the ink is faster when ink colour and word agree (e.g. the colour-naming response is faster for ‘‘sky’’ printed in blue than for ‘‘cola’’ printed in blue.) This indicates that the capacity for printed word recognition accesses semantic information in an unencapsulated way, but this is consistent with fact that it is dissociable from other cognitive capacities.

Even though modularity in the sense of independent disruptability or dissociability is clearly different from Fodor’s concept of informational encapsulation, it is nonetheless an important notion in cognitive psychology. It is arguable that this sense of modularity is the one that was originally introduced into cognitive science by its founding fathers. For example, Simon’s (1969) engineering notion of ‘near-decomposability’, which is a precursor to the notion of modularity, applies to a system when it is made of components whose workings are relatively independent of each other. Similarly, Marr’s (1982) account of the visual system appealed to a principle of modular design according to which ‘‘any large computation should be split into a collection of small, nearly independent, specialized subprocesses’’ (1982, p. 325). In any case, the meaning of modularity linked with independent disruptability or dissociability is still common in cognitive psychology. For example, Saul Sternberg, who has investigated modularity over many years, assumes that this is the basic meaning of modularity: ‘‘Two (sub)processes A and B of a complex process (mental or neural) are modules if and only if each can be changed independently of the other.’’ (Sternberg 2010) To conclude this discussion, how should we understand the notion of modularity as it is used in cognitive psychology? I think that the basic and most common use of the term ‘module’ is to refer simply to a capacity that is functionally individuated in terms of its input and output conditions. In this sense a module corresponds to something that has its own proprietary box in an information flow diagram. (See Fodor, 2000, p. 56) Of the properties that Fodor lists, the important one for this sense of module is domain-specificity, since modules responding to different classes of stimuli will necessarily be functionally individuated as distinct capacities. Nonetheless, the fact that the modularity assumption, as its arises in the context of causal reasoning, is such an inevitable and natural assumption suggests that capacities that count as modules on the minimal criterion will almost invariably be dissociable or independently disruptable as well. (Of course, the interventions that dissociate capacities need not be human experimental interventions, but may be ‘natural interventions’ taking the form of cerebral accidents, strokes etc) Consequently, in my view, these two features— domain specificity and dissociability—are the essential features of modules, which should be taken to define them. This does not mean that modules cannot have the other properties that Fodor mentions. But on this present proposal, whether modules have these additional properties is an empirical question that has to be decided on a case-by-case basis. (For a similar proposal see Coltheart 1999.) So there may be cases in which modules are informationally encapsulated, mandatory, fast, innate, and neurally specific, but these features are not necessary and invariable properties of modules proper. If something like this proposal is correct, the modular architecture of information-processing models of cognition reflects deepseated features of the kind of causal reasoning central to mechanistic explanation. It is useful to illustrate these points by outlining the causal reasoning implicit in the development of a typical information-processing model. I take as an example ‘the dual route model of reading’, which is presented as a model of ‘‘the mental machinery that enables reading’’ in Coltheart, Rastle, Perry, Langdon, and Zeigler (2001) and Coltheart (2006). Fig. 6 below represents the basic structure of the model. This is a model of the information-processing activities that go on when we read aloud. It is called ‘the dual route model’ because it posits that there are two different processes by which our minds convert printed words into speech. One process involves a lexical module that accepts input from all letters and it operates by a dictionary look-up procedure. This module is needed for reading aloud real words. The other process involves a non-lexical module

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

P. Menzies / Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

9

print

Letter identification

Mental lexicon

Non-lexical procedure

speech
Fig. 6. The basic architecture of the dual-route model of reading aloud with domain-specific modules. Modified from Coltheart (2006, p. 8).

meanings; and that people with surface dyslexia can see but can no longer recognize printed words even though they are still able to speak words and still able to appreciate their meanings. (See Coltheart (2006)). Of course, the fact that these dissociations indicate the existence of three distinct modules depends on the fact that modules are assumed to be dissociable. This in turn depends on the assumption that the causal processes going on in each module can be disrupted by experimental or natural interventions independently of each other. In sum, this typical information-processing model illustrates the way in which implicit assumptions derived from causal reasoning–assumptions about domain-specificity and dissociability as essential features of modules and about how modules can be decomposed into submodules—structure mechanistic explanation in this field. 6. Conclusion

that translates letters to sounds by left-to-right application of letter-to-sound rules. It is required for reading aloud non-words. This basic architecture of the model drops out of the assumption that modules are domain-specific, since each of the modules is responsive to a special class of stimuli. Thus the non-lexical module is sensitive to the number of letters of words in a way that the lexical module is not; and the lexical module is influenced by the frequency of occurrence of words whereas the non-lexical module is not (in the case where what is being read is a word.) However, the model depicted in Fig. 6 is not correct in all details. It is known that the mental lexicon contains at least three kinds of information about words: their spellings, their pronunciations, and their meanings. Fig. 6 represents these three kinds of information as being included in a single system, the mental lexicon. However, neuropsychological data indicate that the three kinds of information are processed in three separate modules, as shown in Fig. 7. The fact that there are three separate modules is suggested by dissociations between the capacities. It is known that people with dementia do not know the meaning of words but have normal visual word recognition and normal pronunciation; that people with anomia aphasia cannot access the pronunciation of words but have normal visual word recognition and normal knowledge of word

I began this paper by examining Craver’s account of mechanistic explanation, which gives one of the most detailed descriptions of the causal structure of mechanisms. I argued that while his account makes a useful beginning with its explanation of activities in terms of an interventionist account of causal relevance, it does not capture all the details of the causal structure of mechanisms. Furthermore, his account of constitutive relevance helps to identify the components of a mechanism but, on the other hand, does not explain how they must be causally organized in order to constitute a mechanism. I then attempted to provide a more explicit and more complete account of the causal structure of a mechanism, using the conceptual resources of causal modeling within a structural equations framework. The causal structure of a mechanism was defined in terms of the composition of functional dependences, or, in material mode, in terms of the programmed exercise of modular capacities. The last section of the paper was devoted to illustrating the fruitfulness of this definition by showing how it underlies the kind of mechanistic explanation that is typical of the computational information-processing paradigm of cognitive psychology. There is much that I have not been able to discuss in this paper due to limitations of space. Two subjects stand in need of further investigation, in my view. First, some philosophers of science con-

print

Letter identification

Orthographic lexicon Grapheme-phoneme correspondence rules

Semantic system Phonological lexicon

speech
Fig. 7. The elaborated architecture of the dual-route model of reading aloud with modules analysed into dissociated submodules. Modified from Coltheart (2006, p. 9).

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

10

P. Menzies / Studies in History and Philosophy of Biological and Biomedical Sciences xxx (2012) xxx–xxx

test the modularity constraint that I have imposed on the causal structure of a mechanism. (See Bechtel (2010), Cartwright (2001, 2004); Hoover (2010).) Some of these criticisms are mistaken in my view. (For rebuttals of some of these criticisms Kuorikoski (2010), Steel (2010), Woodward (2008).) However, it is clear that the modularity assumption does not apply universally in all domains. For instance, it is not always a safe assumption in the explanation of social phenomena: participants in a social system, for example, who are not targeted by a policy intervention may nevertheless be aware of it and modify their practices in response to it (See Steel (2008, p. 52)). The fact that members of a society may become aware of the prevailing social structures and so seek to change them is a common basis for arguments against the possibility of laws of social science. But this raises the question: In what domains is it appropriate to apply the modularity assumption? More generally, it is important to have answer to such questions as: What is the status of the modularity assumption? Is it a necessary feature of causal relations that is an a priori implication of the concept of causation? Or is it a default assumption we make in order to simplify our causal models? Secondly, my discussion of mechanistic explanation has focused on the causal structure of such explanation. But as Craver puts the point, there is more to mechanistic explanation than box-and-arrow diagrams. To be sure, my discussion has neglected to describe all the significant details of the spatial and temporal organization of mechanisms. I agree with Craver and other theorists of mechanistic explanation that the spatial and temporal organization of mechanisms is often as significant as their causal organization. One of the most neglected topics of discussion in the literature on mechanistic explanation concerns how information about the causal operation of the components of a mechanism is interwoven with information about their spatial and temporal organization. Perhaps some progress can be made on this topic by considering the way in which causal thinking tends to localize causal powers in physical objects, a feature of our causal thinking that has been obscured in this paper by the characterization of causal relations as holding between variables. Acknowledgements Versions of this paper have been presented to the research seminar of the Department of Philosophy, Research School of Social Sciences, Australian National University in August 2010; and to a conference on ‘‘Causality in the Biomedical an Social Sciences’’ at Rotterdam University in October 2010; and to a workshop on ‘‘Mechanisms and Levels of Explanation in the Cognitive Science’’ at Macquarie University in December 2010. I’m grateful to the questions and comments from the participants at these workshops, especially Bill Bechtel, Brett Calcott, Nancy Cartwright, Max Coltheart, Kevin Hoover, Kim Sterelny, and Daniel Stoljar. I’m especially grateful to Max Coltheart for providing helpful references and suggestions. Research for this paper was funded by an Australian Research Council Discovery Project grant to Peter Menzies for the project ‘‘Causes that Make a Difference’’.

References
Bechtel, W. (2008). Mental mechanisms: Philosophical perspectives on cognitive neuroscience. New York: Routledge. Bechtel, W. (2010). Dynamic modules. Paper presented at workshop on ‘‘Mechanisms and levels of explanation in cognitive science’’. 16–17 December 2010, Macquarie University. Bechtel, W., & Richardson, R. (1993). Discovering complexity: Decomposition and localization as strategies in scientific research. Princeton: Princeton University Press. Cartwright, N. (2001). Modularity: It can—and generally does—fail. In Galavotti, Suppes, & Constantini (Eds.), Stochastic causality (pp. 65–84). Stanford: CSLI Publications. Cartwright, N. (2004). Causation: One word, many things. Philosophy of Science, 71, S805–819. Craver, C. (2007). Explaining the brain: Mechanisms and the mosaic unity of neuroscience. Oxford: Oxford University Press. Coltheart, M. (1999). Modularity and cognition. Trends in Cognitive Science, 3, 115–120. Coltheart, M. (2006). Dual route and connectionist models of reading: An overview. London Review of Education, 4, 5–17. Coltheart, M., Rastle, K., Perry, C., Langdon, R., & Zeigler, J. (2001). DRC: A dual route cascaded model of visual word recognition and reading aloud. Psychological Review, 108, 204–256. Cummins, R. (1983). The nature of psychological explanation. Cambridge, Mass.: MIT Press. Cummins, R. (2000). ‘Hows does it work?’ versus ‘What are the laws?’: Two conceptions of psychological explanation. In F. Keil & R. Wilson (Eds.), Explanation and cognition (pp. 117–144). Cambridge, Mass.: MIT Press. Fodor, J. (1983). The modularity of mind. Cambridge, Mass.: MIT Press. Fodor, J. (2000). The mind doesn’t work that way: The scope and limits of computational psychology. Cambridge, Mass: MIT Press. Glennan, S. (1996). Mechanisms and the nature of causation. Erkenntnis, 44, 49–71. Hoover, K. (2010). Causal structure and levels of explanation. Paper delivered at conference on ‘‘Causalities in the Biomedical and Social Sciences’’. Rotterdam University, 6–8 October 2010. Kuorikoski, J. (2010). Varieties of modularity for causal and constitutive explanations. Department of Political and Economic Studies, University of Helsinki. http:// philsci-arhive.pitt.edu/archive/00004303/01 Accessed 11.10.11. Machamer, P., Darden, L., & Craver, C. (2000). Thinking about mechanisms. Philosophy of Science, 57, 1–25. Marr, D. (1982). Vision. San Francisco: Freeman. Pearl, J. (2000). Causality: Models, reasoning, and inference. Cambridge: Cambridge University Press. Robbins, P. (2009). Modularity of mind. In E. N. Zalta (Ed.), Stanford enyclopedia of philosophy. http://plato.stanford.edu.au/entries/modularity-mind/ Accessed 11.10.11. Simon, H. (1969). The sciences of the artificial. Cambridge, Mass.: MIT Press. Spirtes, P., Glymour, C., & Scheines, R. [1993] (2000). Causation, prediction, and search. New York: Springer-Verlag. Steel, D. (2008). Across the boundaries: Extrapolation in biology and social science. New York: Oxford University Press. Steel, D. (2010). Cartwright on causality: Methods, metaphysics, and modularity. Economics and Philosophy, 26, 77–86. Sternberg, S. (2010). Modular processes in mind and brain. Department of Psychology, University of Pennsylvania. http://www.psych.upenn.edu/~saul/ Accessed 11.10.11. Wimsatt, W. (1974). Complexity and organization. In K. Schaffner & R. S. Cohen, (Eds.), PSA 1972 (Boston Studies in the Philosophy of Science, Vol. 2, pp. 67–86). Dordrecht, Holland: Reidel. Woodward, J. (2002). What is a mechanism? A counterfactual account. Philosophy of Science, 69, S366–S377. Woodward, J. (2003). Making things happen: A theory of causal explanation. Oxford: Oxford University Press. Woodward, J. (2008). Invariance, modularity, and all that: Cartwright on causation. In S. Hartmann, C. Hofer, & L. Bovens (Eds.), Nancy Cartwright’s philosophy of science (pp. 198–237). New York: Routledge.

Please cite this article in press as: Menzies, P. The causal structure of mechanisms. Studies in History and Philosophy of Biological and Biomedical Sciences (2012), http://dx.doi.org/10.1016/j.shpsc.2012.05.008

Sign up to vote on this title
UsefulNot useful