Social Mechanisms and Explanatory Relevance

Social Mechanisms and Explanatory Relevance Petri Ylikoski Department of History and Philosophy University of Tampere & Philosophy of Science Group University of Helsinki [email protected] In this paper I will discuss mechanistic explanation in the social sciences from the point of view of the philosophical theory of explanation. My aim is to show that the current accounts of mechanistic explanation do not serve the agenda of analytical social science as well as they should. I will not challenge the idea that causal explanations in the social sciences involves mechanisms or that social scientists should seek causal explanations and a mechanistic understanding of social phenomena, but will argue that to improve explanatory practices in the social sciences analytical social scientists should employ tools more substantial than the metaphor of a mechanism. I will begin by presenting the basics of the erotetic approach to explanation and the notion of explanatory understanding. Section 2 argues that mechanisms do have many important roles related to explanation, but that they do not provide a solution to the problem of explanatory relevance. Section 3 introduces the idea of a contrastive explanandum and argues that paying attention to the identity of the explanandum is a necessary condition for proper assessment of explanatory claims. Section 4 argues that explanatory relevance should be understood as counterfactual dependence and that a manipulationist account of causation provides a fruitful framework for making sense of causal explanations. Section 5 discusses some of the consequences of these ideas for mechanistic explanation in the social sciences. 1 The purpose of the theory of explanation 1 The aim of the theory of explanation is to make sense of explanations. It addresses questions such as following: What kind of an activity is explanation and how is it related to other epistemic (and practical) activities? What kinds of explanations are there and what are their relationships to each other? By which criteria are explanations evaluated by scientists and by which criteria should they be evaluated? All of these questions circle around the central question: What constitutes an explanation? A convenient way to approach this question is to think about explanatory relevance. The problem of explanatory relevance is one of the major challenges for the covering-law account of explanation. The examples of men taking contraceptive pills explaining them not becoming pregnant, and hexing explaining why salt dissolves in water, are counterexamples of the covering-law account because they involve intuitively explanatorily irrelevant factors while fulfilling all the criteria of a satisfactory covering-law explanation. These counterexamples and other arguments have led philosophers of science to conclude that the covering-law account is a failure. It does not solve the central problem for any theory of explanation: the problem of explanatory relevance. This problem is not easy to solve. For example, Wesley Salmon’s causal theory of explanation falls victim to exactly the same counterexamples – it is not enough that we demand that an explanation only provides some information about the causal process, we want to have relevant information (Hitchcock 1995). The problem of explanatory relevance can be understood as a problem of explanatory selection (Hesslow 1983). Causal explanation provides information about causal history, but not all information about that history is regarded as explanatory. We have to pick the right aspects of the causal process to be included in the explanation. That is to say, how far in the causal history should we reach? How do we choose the events to be included in the explanation? How do we choose the right level of abstraction for describing these events? In how much detail should the events be described and which of their details should be included in the description? Apparently 2 we somehow manage to solve these problems intuitively when we are constructing explanations, but it would be much better if we could make the principles governing these judgments explicit. This is the main challenge for the theory of explanation. The theory of explanation should not be a purely theoretical enterprise; the ultimate motive is to develop conceptual tools for improving explanatory practices in the sciences (and in everyday life). The theory of explanation should have practical relevance and its success should be judged by the improvements it makes possible in explanatory practices. Such contributions are especially important in the social sciences where controversies about explanation are common. The social mechanisms movement is motivated by similar practical goals. While it has provided philosophical arguments for mechanistic explanation, the aims have been ultimately practical. The improvements in social science explanatory practice are the final evaluation criteria for philosophical arguments about explanation and causation. In this paper I will outline an account of explanation that in my judgment best advances the aims of both the philosophical theory of explanation and the social mechanisms movement. I call it the contrastive counterfactual account of explanation. The first element of this account is the erotetic approach to explanatory inquiry. Explanations as answers to questions Most philosophers of explanation agree that explanations are answers to questions. Some, like Hempel (1965), have used it as an informal starting point for their discussion, whereas others (van Fraassen 1980, Achinstein 1983) have built their theories of explanation around it. The latter approaches are commonly called erotetic approaches to explanation and are often associated with the broader question-theoretical account of scientific research. In the erotetic approach, scientific enquiry is regarded as a process of answering and elaborating questions. Empirical research is oriented towards fact-finding questions: it aims to answer what, 3 when, where, and how much questions. The relevance of these questions (and the required precision of the answers) is determined by practical how to questions and by explanatory why and how questions. The insight provided by the erotetic approach is based on the analysis of interrelations between different questions and its description of the research process as an organized series of questions. For example, the research process often requires that big explanatory questions be sliced to a series of smaller ones that can be addressed through empirical inquiry. Explanation and understanding In the erotetic approach, explanation is an answer to an explanation-seeking question. The explanation is regarded as complete when it fully answers the given question. This makes the notion of explanation quite narrow. To capture the broader goal of epistemic activities, we need another notion. I suggest that this be understanding (Ylikoski 2009). We are interested in finding answers to explanation-seeking questions because we wish to have knowledge about the dependencies governing the world. In other words, the goal of explanation is to understand these dependencies. Wittgenstein argued that understanding should not be understood as a sensation, an experience, or a state of mind. Understanding is not a special moment or phase, but a more permanent attribute. It is an ability (Wittgenstein 1953, §§ 143-159, 179-184, 321-324). I agree: when a person understands something, she is able to do certain things, which does not mean that understanding is some sort of special skill. Understanding consists of knowledge about relations of dependence. When one understands something, one can make all kinds of correct inferences about it. Many of these inferences are counterfactual: What would have happened if certain things had been different? What will happen if things were to be changed in a certain manner? Thus the fundamental criterion according to which understanding is attributed is the ability to make inferences to counterfactual situations, the ability to answer contrastive what-if-things-had-been-different questions 4 (what if questions, for short) by relating possible values of the explanans variables to possible values of the explanandum variable (Ylikoski 2009). The present account ties together theoretical and practical knowledge: they are not completely different notions. For example, in the case of causal explanation, explanatory understanding is crucial to our pragmatic interests, since answers to what if questions concerning the effects of possible interventions enable us to predict the effects of manipulation. Whereas the DN model and the associated epistemic conception of explanation conceive of the possessor of understanding as a passive observer of external events, the contrastive counterfactual theory links our theoretical practices to our roles as active agents (Woodward 2003). The degree of understanding conveyed by an explanation can be defined as the number and importance of counterfactual inferences that the explanatory information makes possible. Why do many people think that understanding is a mental state or an experience? The reason is that there exists a mental experience that is closely related to understanding: the sense of understanding. It is a feeling that tells us when we have understood or grasped something. This sense of confidence (and the feeling that often comes with it) can be easily confused with what we think it indicates: understanding. Ideally these two things would go hand in hand, and assimilating them should not create any trouble. However, real life is different. The sense of understanding is a highly fallible indicator of understanding. Sometimes one has a false sense of understanding and sometimes one understands without having any associated feelings or experiences. The sense of understanding does not give us direct access to knowledge that is the basis of our understanding, so it would be highly surprising if the sense of understanding would turn out to be perfectly calibrated to our understanding. The fallibility of the sense of understanding can be demonstrated experimentally. People often overestimate the detail, coherence, and depth of their understanding (Rozenblit and Keil 2002; Ylikoski 2009). 5 The existence of the sense of understanding should not be regarded as any kind of oddity. It plays an important metacognitive role in our cognitive life. It gives us confidence to try things, and when it is lacking we can sensibly abstain from the activity in question. It also guides the search for new knowledge, and tells us when to stop the search for new information; it signals when we know enough. In addition, the sensation associated with the sense of understanding can have a motivational role. Satisfying curiosity is highly rewarding (Schwitzgebel 1999, Gopnik 2000). It provides motivation for learning and other cognitive activities, and for this reason has an important role in human cognition. The desire to satisfy one’s curiosity also provides important psychological motivation for conducting scientific research (Ylikoski 2009). The phenomenology of the sense of understanding can mislead one into thinking that understanding is an on-off phenomenon (‘Now I got it!’). This is not the case. First, the understanding can be about different aspects of the phenomenon. Second, these aspects may be understood in various degrees. Consider an ordinary object like a personal computer. Different individuals understand to varying degrees how their computer works. Some might know about the software, or some specific piece of software, and others the hardware. Most people just use the software without any understanding of a computer’s internal workings. Despite these differences, they all understand something about their PC. The crucial question is which aspects of it they understand. A comparison of their understanding is possible, but strict assessment of overall understanding is complicated: there are many dimensions to compare. However, the important point is that by asking what has been understood, the extent of understanding can always be specified (Ylikoski 2009). This notion of understanding can be used to make sense of social science explanations. For example, we can ask what kind of what if questions does a certain account of social mechanism answer. Answering this question can be difficult, but this difficulty only emphasizes the contrast 6 between our confidence that it provides explanatory insight and our inability to articulate (or to agree about) what it explains. Surely this kind of exercise would benefit the aims of analytical social science. We can also raise the more general question of whether the social theorist’s sense of understanding is well calibrated to his abilities to make correct counterfactual inferences about social phenomena he is studying. I would be highly surprised if it were to turn out that widespread explanatory overconfidence didn’t exist among social theorists. 2 Mechanisms and explanatory relevance The idea of a mechanism has many uses in the philosophy of science. Most of these ideas also figure in social science discussions about explanation (Hedström & Ylikoski 2010). Here are four different ideas about the contribution of mechanisms to explanatory understanding. The first idea is about heuristics. According to it, existing mechanistic explanations can serve as templates, schemes or recipes for the search of causes. The knowledge about a possible mechanism tells the researcher what to look for and where. This simplifies the search for causes, especially in situations where one can be confident that the menu of possible explanatory mechanisms covers all the plausible alternatives. The second idea is related to justification of causal claims. It posits that the knowledge about possible mechanisms can provide support for causal claims. Causal claims without an account of the underlying mechanisms are possible and in principle fully legitimate, but knowledge of a mechanism makes them much more secure. This idea originates from everyday thinking, but it is also accepted in scientific contexts. Of course, the idea can sometimes be misleading: an imagined causal mechanism can give false credence to spurious causal claims. The third idea is about the presentation of explanatory information. A mechanism scheme provides useful means for presenting and organizing explanatory information. A narrative form 7 makes the explanatory information more digestible for humans, and mechanism schemes can be regarded as templates for such narratives. They outline the central features of the explanatory narrative and help people to focus on the right pieces of information. The fourth idea concerns the organization of social scientific knowledge. According to an old empiricist view, general knowledge in science consists of empirical generalizations and more abstract theoretical principles from which these generalizations can (ideally) be deduced. The mechanistic account of knowledge challenges this picture on two counts. First, the locus of generality (and explanatory power) in social scientific knowledge is considered to lie in the mechanism schemes. The social sciences do not have that many valid empirical generalizations and those that they have are not very explanatory. On the contrary, they are among the things that require explanation (Cummins 2000). Explanatory power – and general applicability – comes from knowledge of possible causal mechanisms. When social scientific knowledge expands, it does not do so by formulating empirical generalizations that have broader application, but by adding or improving items in its toolbox of possible causal mechanisms. This brings us to the second challenge to the traditional picture of social science knowledge: the ideal of knowledge in no longer an axiomatic system, but a much looser idea of an expanding theoretical toolbox. The expectation is that mature social science would be more like a textbook of modern cell biology than a treatise in elementary geometry. These are important ideas – and I subscribe to all of them – but they do not address the issue of explanatory relevance. They do not tell us what is the explanatory import of mechanisms. Nor do they tell us how to construct a good or satisfactory mechanistic explanation. Furthermore, they provide no guidance for the evaluation of suggested mechanistic explanations. The insufficiency of mechanistic ideas 8 To see why the idea of mechanisms is not very helpful in dealing with the problem of explanatory relevance, we have to understand how this notion works. Let us begin by distinguishing two ways to talk about mechanisms. Examples of both are plentiful in the literature and most of the time people seem to assume that they incorporate the same notion of mechanism. The first (let us call it A-mechanism) regards mechanisms as a particular causal chain. The mechanism is whatever connects the cause and effect. No matter how long or complicated the causal process is, it can be called a mechanism if its description answers the question how did the cause bring about the effect? The second way (B-mechanism) to talk about a mechanism regards it as a theoretical building block. Here the mechanism is an object of theoretical interest, and it often applies only to simplified and idealized explananda. The above mechanistic idea about the organization of general knowledge is based on this notion of mechanism. The differences between these two notions can be seen more easily if we consider their interrelations. The first thing to observe is that a single A-mechanism can involve many B-mechanisms. A number of different B-mechanisms can work serially or simultaneously in the same causal process. Furthermore, nothing prevents B-mechanisms from working in opposite directions. For example, if we are explaining the rise in the murder rate after the collapse of a corrupt central government, some of the B-mechanisms could work towards a reduction of crime, whereas others would increase it. The explanandum of an A-mechanism is the total (or net) effect, whereas the explanandum of a B-mechanism is more like a component effect. Clearly we should not use these two notions of mechanism interchangeably. The idea of an A-mechanism builds upon the idea that the knowledge of the details of a causal process makes the explanation better. This idea feels intuitively correct. People prefer detailed explanations to mere sketches of explanation. This preference shows that detail is regarded as a virtue in explanatory contexts. The crucial question here is what do we mean by more detailed? 9 An explanation is more detailed when it omits less of the relevant information. The key word here is relevant. The details of the causal process can be quite harmful for the explanation if they are irrelevant from the point of view of the explanandum. There has been some discussion in the philosophy of science about whether the addition of irrelevant details makes the explanation completely nonexplanatory or whether it just makes an explanation worse (see Salmon 1998). We do not need to take a stance on this partly verbal issue; it is sufficient to point out that irrelevant details at least decrease the cognitive salience of the explanation. Irrelevant details can mislead: in such cases we identify wrong factors as explanatory. A more detailed causal story might also be more difficult to grasp. This is a simple fact of human cognition: our memory and ability to focus are limited and burdening them restricts our inferential performance. In both cases our ability to answer what if questions decreases. In other words, we will understand less. Although the notion of an A-mechanism is important, it does not provide much insight into the nature of explanation. It is a placeholder notion: whatever explains the fact that c caused e is the mechanism. In this sense it is of limited analytical value: it names the challenge, but does not provide tools for dealing with the problem of explanatory relevance. Nor does the notion of an Amechanism give any guidance for constructing mechanistic explanations. It does not tell us which level of organization we should focus on, which elements of the process should be incorporated in to the mechanism, or how detailed the description of these items should be. The identity criteria for B-mechanisms are stricter, but this notion faces the opposite problem. Bmechanisms are in many ways analogical to component causes. Just like component causes, they can be involved in cases of overdetermination, in which more than one mechanism works to guarantee a certain outcome. (Sometimes they could work simultaneously, sometimes one mechanism could preempt another.) Similarly, they could be involved in the cases of counteracting causes, in which the mechanisms compete by having opposite causal effects. Finally, as with 10 an individual component cause, a B-mechanism can be simply explanatorily irrelevant. In the case of an A-mechanism the connection to the explanandum is guaranteed by definition, but such a guarantee is lacking in the case of a B-mechanism. Any causal process involves many Bmechanisms at various levels of organization (physical, chemical, etc.); it is an open question whether they are in any way relevant to the explanandum we are interested in. Here again we find the problem of explanatory relevance: we need some guidance with B-mechanisms. What is the appropriate level of organization to focus on? What makes a given B-mechanism relevant to the explanandum? In how detailed a manner we should describe the mechanism? When people construct and evaluate mechanistic explanations they are intuitively making judgments about explanatory relevance. Everybody agrees that good mechanistic explanations capture the relevant aspects of the causal process. However, nothing in the idea of a mechanism helps us in making and evaluating these judgments – we have to trust our intuitions. The trouble with intuitive assessments of explanations is that they are based on unreliable metacognitive processes. The sense of understanding is highly unreliable, and this shows in the disagreements people have in their judgments about explanatory value. Given the intuitive appeal of mechanistic storytelling in everyday reasoning, the critics of mechanistic theories of explanation might be right in being skeptical about their value. Especially when we are working with highly abstract sketches of mechanisms (as is typical in the social sciences), the danger of the illusion of depth of understanding is a real one. Clearly, the notion of a mechanism is not a sufficient tool for improving explanatory practices in the social sciences. It needs to be supplemented with additional ideas from the theory of explanation. I will now turn to these ideas and try to show how they can be used to make sense of mechanistic explanation. 11 3 The importance of the explanandum The problem of explanatory relevance in causal explanation is that the causal history of an event includes a vast number of elements and aspects that are not explanatorily relevant to the explanation-seeking question we are addressing. A natural way to start sorting out this problem is to take a closer look at the explanandum. A contrastive account of explanation is helpful here. According to this view, explanations are answers to questions in the following form: Why the fact rather than the foil happened? We want to know why things are one way rather than some other way. An anecdote about the famous bank robber Willie Sutton illustrates the basic idea of contrastive explanation. When Sutton was in prison, a journalist asked him why he robbed banks. Sutton answered, “Well, that’s where the money is.” The journalist had in his mind the question: “Why do you rob banks, instead of leading an honest life?,” whereas Sutton answered the question: “Why do you rob banks, rather than gas stations or grocery stores?” This is the basic insight of the contrastive approach. We do not explain simply ‘Why f?,’ rather, our explanations are answers to the contrastive question ‘Why f rather than c?’ (Garfinkel 1981, 21-22). Instead of explaining plain facts, we are explaining contrastive facts (Ylikoski 2007). Sometimes the contrast is imagined: we ask why an object has a particular property rather than a different property we expected it to have or our theory predicted it would have. Sometimes the contrast arises from a comparison: we ask why an object has a certain property rather than being like an otherwise similar object that has a different property. In both cases we are explaining a difference: why a fact is the case rather than its exclusive alternative. I will adopt the convention of expressing the contrast in the following manner: fact [foil], which should be read as fact rather than foil. The number of foils is not limited; there can be more than one. The contrastive idea does not put strict limitations on kinds of facts that can serve as the ex- 12 plananda: they can belong to different ontological categories. They can be properties, events, quantities or qualities. The crucial thing is that the fact and its foil should be exclusive alternatives to each other. To cover the wide variety of possible ontological categories, I will here simply speak about variables. This terminology does not commit us to any specific ontology, but allows us to make our points more generally and to avoid messy philosophical debates about the nature of events and other ontological categories. The basic idea is that whatever the real relata of explanation are, they can be represented as variables. The idea of contrastive explanandum is a practical tool for analyzing explananda and for making the intended explanandum more explicit. Although it is not always apparent from the surface appearance of explanations, all explanation-seeking questions can be analyzed and further explicated by articulating the implicit contrastive structure of the explanandum. This explication makes the aims of explanation-seeking questions more clear and makes the evaluation of the adequacy of the proposed explanations easier. Articulating the contrastive structure forces one to be specific about the intended object of the explanation and about the respects in which we think the object could have been different (Ylikoski 2007). Are we explaining a singular event or a more general phenomenon or regularity? Are we addressing properties of individuals or of populations? What is the appropriate level of description: are we after micro-level details or patterns found at the macro level? Quite often the original question turns out to be a whole set of related contrastive questions. This is a good thing: smaller questions are something that we can realistically hope to answer by means of empirical enquiry. Contrastive articulation is also useful in controversies over apparently conflicting explanations. Often the apparently competing explanations turn out to be addressing complementary or completely independent questions. This is as it should be: we can be 13 pluralists about explanation-seeking questions, but whether the answers are correct is still an objective matter. The contrastive approach does not only help to clarify the explanandum, it can also be used to evaluate explanations. Rather than asking what the intended explanandum is, we can also ask what is the contrast that the given explanation (or piece of explanatory information) can explain? This is very useful approach, particularly, in social science controversies about explanation. Rather than getting entangled with difficult interpretive problems about what certain theorists are really attempting to explain, we can take a look at the explanations they provide and consider what they in fact explain (Ylikoski 2007). The contrastive idea underlies our preferred form of causal inquiry: the scientific experiment. We contrast the control group with the experimental group or the process after the intervention with the process before the intervention. In both of these cases, we are trying to account for the differences between the outcomes. The basic idea is to keep the causal background constant and to bring about changes in the outcomes by carefully controlled interventions. A similar contrastive setup motivates comparative research. In general, the idea that explanations are contrastive is natural if one thinks that the aim of explanation is to trace relations of dependency. Our goal is to understand how changes in the cause variable bring about changes in the effect variables. We want to know what makes the difference and then leave out the factors that do not. The contrastive idea is based on a very intuitive feature of our explanatory cognition. 4 Making a difference The starting point of the contrastive-counterfactual approach to explanation is realistic: explanations attempt to trace objective relations of dependence. These dependences can either be causal or constitutive. Dependencies are objective in the sense that they are independent of our 14 ways of perceiving and theorizing about them. They are also separate from our means of describing them. That is to say, the real relata of explanation are facts, not sentences. If the explanandum is contrastive, as I have suggested, it is natural to think that the explanans is also contrastive. Following this idea, philosophers of causation are increasingly thinking that causation is doubly contrastive (Woodward 2003; Schaffer 2005). The same idea applies also to causal explanation. In the simplest form, causal explanatory claims have the following structure: c [c*] explains e [e*] In plain language, the cause variable having the value c rather than c* explains why the explanandum variable has the value e rather than e*. A natural way to understand this relation is to regard it as a counterfactual dependence: if the cause variable had had the value c*, the value of the explanandum variable would have been e* rather than e. Note that although the claim is about counterfactual situation, there is nothing counter to the facts in the relation of dependence itself. Also of importance is that counterfactual dependence is a modal notion: explanation is not about subsumption under empirical regularities, but about counterfactual dependence. The idea of a counterfactual theory of causation is very old. We can avoid some of the historical problems of counterfactual theories of causation if we do not regard it as a reductive theory of causation and if we limit its application to causal explanation. The new idea is to combine it with the idea of contrastive explanandum. Together with a sophisticated manipulation account of causation this idea helps us to solve most of the problems that counterfactual theories of causation have faced (See Ylikoski 2001). An important problem for the counterfactual theories has been the specification of the truth conditions for the counterfactuals. A manipulation account of causation is helpful here. In the manipulation account of causation c [c*] causes e [e*] if we can bring about e* [e] by bringing about c* [c] 15 This cannot serve as a reductive analysis of causation, as ‘bringing about’ is already a causal notion. However, it can serve as a starting point for an adequate descriptive analysis of the notion of causation. An advantage of the manipulation account is that it provides a natural way of distinguishing between real causal relations and mere correlations. Real causal relations can be used as bases for effective interventions, whereas mere correlations do not allow this. Experiments and other causal interventions in the real world are often quite messy. To make sense of the meaning of causal claims we need the notion of an ideal intervention, developed by Woodward (2003, 94-151). An ideal intervention I changes the value of effect variable Y only via changing the value of cause variable X as shown below. (The formulation and the graph are adapted from Craver 2007.) U 1. I does not change Y directly 4 2. I does not change the value of any causal inX S Y termediate S between X and Y except by chang2 ing the value of X that is a cause of Y 4. I acts as a switch that controls the value of X C I 3. I is not correlated with some other variable C 1 3 causal relation absence of causal relation absence of correlation irrespective of X’s other causes U. The notion of ideal intervention gives us straightforward semantics for causal counterfactuals. It also makes it possible to avoid anthropocentric features of earlier manipulation accounts of causation: the definition does not refer to a human agent. Nor must we assume that the intervention is humanly possible – we are only interested in possible interventions. For us the main advantage of this account is that it provides us with a natural way of making sense of counterfactuals in 16 causal explanation, meaning that an explanatory claim is correct if ideal intervention on the explanans variable had brought about the appropriate change in the explanandum variable. An interesting feature of Woodward’s account is his view of explanatory generalizations about effects of ideal interventions (Woodward 2003, 239-314). These invariances lack many traditional features of laws of nature. They usually hold only for a limited range of possible interventions (and changes in background factors); they can refer to particular objects, places and times; and they might contain exceptions. In short, they might hold only in a certain domain and break up outside of it. However, in contrast to the traditional empiricist accounts of laws of nature, invariance is a modal notion: it makes a claim that the invariant relationship between cause and effect variables will hold under a set of possible interventions. According to Woodward, invariances capture what is essential from the point of view of explanation, whereas most of the traditional attributes are superficial from this angle. Consequently, in his account all explanatory generalizations are not laws. This is good news for the social sciences (and other sciences outside of fundamental physics): the generalizations satisfying the traditional criteria of lawhood are rare or nonexistent. It is plausible to think that our contrastive explanatory preferences stem from our nature as active interveners in natural, psychological, and social processes. We desire to know where to intervene to produce the changes we want, and this knowledge often presupposes answers to some why and how questions. Without this knowledge one would not know when the circumstances are suitable for an intervention and one would not be able to predict the results of the intervention. I do not wish to claim that the notion of explanation can be reduced to its origins in agency. After all, we are also interested in explaining things that cannot be humanly manipulated. However, our instrumental orientation might still explain why our explanatory preferences are as they are. 17 5 The explanatory import of mechanisms In the contrastive counterfactual account causal claims are based on change-relating invariances, meaning that no mechanism is needed for making a causal claim. This helps us to avoid the regress problem that an account which assumes that mechanisms are an integral part of all causal claims would face. However, it does not challenge the legitimacy of the assumption that outside of fundamental physics there will always be a causal mechanism mediating between the variables. Nor does it imply that we should not try to have knowledge about the mechanisms (Hedström & Ylikoski 2010). However, there is one important consequence: macro variables are explanatory in the same sense as micro variables. In other words, the explanatory factors (for explananda at individual or social levels) might be found at the social level. The crucial issue is whether the right kind of invariance exits between the explanans and the explanandum variables. If by changing the macro-variable we can bring about a relevant change in the explanandum variable, we have identified the cause. The intervention account does not privilege any specific level of description; it is a purely empirical matter which level provides the variables that inform us about the relevant invariances (Steel 2006). The crucial factor is the contrast on the explanandum side: depending on the contrast, a micro or macro variable might turn out to be more relevant (Ylikoski 2001, 77-100). The relevance is determined by the nature of invariances that the variables are involved in. This blocks the simple argument for methodological individualism through causal mechanisms. Causal claims do not have to incorporate mechanisms and they can refer to macro variables. Clearly, if we want to stick with methodological individualism, we need some independent arguments. If causal claims can be made without referring to mechanisms, the question arises what is their function in causal explanation? Is it a misguided idea that adds nothing to the process of causal explanation, as some critics have suggested? This would be an overreaction. It is important to recognize the limited scope of the above claims. These are claims about singular causal state18 ments that answer to simple explanation-seeking questions. However, when we turn our attention to the broader issue of understanding, the true contribution of mechanisms becomes apparent. A singular causal claim might be based on a simple claim about counterfactual dependence, but for theoretical understanding we need more. In the erotetic approach, explanation-seeking questions are not treated separately as independent entities. The questions come in clusters of closely related what if questions. The individual questions are related to each other in many ways. When we find that the value of variable Y is dependent on the value of variable X in a manner that allows (at least an ideal) intervention, this raises a whole series of questions. First, why does the counterfactual dependence hold? Second, what is the range of interventions that this invariance allows without breaking apart? Third, what are the background conditions for this invariance and how sensitive is the invariance to changes in these conditions? Fourth, are there other alternative interventions that can bring about the same changes in Y? Fifth, is it possible to account for a more precise explanandum? And sixth, is it possible to find a generalization that stays invariant in a broader range of interventions and background conditions? The knowledge of mechanisms is involved in answering all these questions. The questions are related to each other via chains of presupposition. As we can see from the first question, underlying every why question is a possible how question. Individual claims about causal influence are of limited interest. We want to know why X can cause changes in Y. Answering this question requires knowledge about causal mechanisms: we have to know how the changes in X are transmitted to changes in Y. The fourth question about alternative interventions is closely related and knowledge about possible mechanisms greatly facilitates answering it. Answers to both of these questions expand the range of what if questions we can answer about the phenomenon – they increase our understanding. 19 An important manner in which knowledge about mechanisms advances our understanding is integration. An answer to the first question integrates the causal claim with other pieces of knowledge. This is a very important consideration from the point of view of understanding. When an explanation is integrated into a larger theoretical framework, the theoretical connections can expand the range of answers to different what if questions in two ways. First, the integration allows for inferential connections to an already existing body of knowledge, and this might make it possible to find unforeseen dimensions in which what if questions concerning the explanandum phenomenon can be answered. Second, the explanation itself may bridge previous gaps within the existing theory and thus enable answers to new what if questions that do not directly concern the original explanandum phenomenon. (Ylikoski & Kuorikoski 2009.) The second and third questions arise from the fact that we want to know how broadly our causal knowledge can be exploited; in other words, we wish to know how many other what if questions can we answer with the same information. When we know more about the mechanism transmitting the causal influence, we have a better idea of what kinds of factors can disrupt the causal link and how. With this information, we can answer more what if questions concerning situations in which the background assumptions or conditions are different, for example in situations where some parts of the mechanism were altered or when the background factors not included in the explanatory generalization change. In this way, knowledge of mechanisms contributes to understanding the domain of application of the knowledge about the invariance. Without knowledge about the mechanisms we would have trouble evaluating the range of what if questions that can be answered with our piece of causal knowledge. Knowing the limits of one’s knowledge is an important ingredient of understanding. Finally, the fifth and sixth questions suggest that knowledge about the mechanisms might help us to improve the explanatory generalization. A common way to search for more general formula- 20 tions is to look at the underlying mechanisms and explore the possibility of integrating them into the explanatory generalization. This might make it possible to formulate the explanatory invariance in such a manner that it allows us to explain either a sharper explananda or a broader range of explananda than the original (Ylikoski & Kuorikoski 2009). It might also make it possible to formulate a more general statement of invariance that holds for a broader set of interventions and/ or background conditions. More general invariances are preferable, as they also make it possible to answer a broader set of what if questions. In both of these cases the road to better explanatory generalizations goes through mechanisms: they suggest variables that could be incorporated into the explanatory generalization (Ylikoski & Kuorikoski 2009). If these observations are correct, the call for mechanisms really makes sense. Knowledge about mechanisms means an ability to answer more what if questions, i.e. more understanding. When this explanatory contribution is combined with the four other functions of mechanisms mentioned earlier, we see that the social mechanisms movement has been motivated by the right kind of intuitions. However, it might be that some of the advocates of social mechanisms have had the wrong ideas about the source of legitimacy of these intuitions: the necessity of mechanisms does not derive from ontological arguments for methodological individualism or from semantics of individual causal claims. 6 Conclusion In this paper I have suggested that the current mechanistic ideas about explanation are insufficient for the purposes of analytical social science. The aim is to improve explanatory practices, but the notion of a mechanism is not very helpful. I have also suggested a number of ideas about explanation that can be used in improving social science explanations. The first is the erotetic approach to explanations combined with the idea that all explanation is contrastive. This idea is use21 ful in making explanations more explicit and by allowing one to be more precise about the intended explanandum, as it permits a much sharper evaluation of explanatory claims. It also makes it possible to think about interrelations between explanations by considering chains of presupposition. The second idea is the counterfactual criterion of explanatory relevance. When explanatory claims are understood as claims about possible (ideal) interventions, we get a notion of explanatory relevance that is both intuitive and powerful. Among other things, this approach allows for a novel way of understanding explanatory generalizations as claims about invariances. Third, I have suggested that we employ the notion of understanding to make sense of the contribution of the knowledge about mechanisms to our explanatory enterprise. A singular causal claim does not require knowledge about the mechanisms, but once we begin to consider the broader set of what if questions, the importance of mechanisms becomes apparent. The notion of understanding also helps us to see the unreliability of the intuitive way of evaluating the relevance of explanatory information. As the sense of understanding can be highly misleading, we have a further motivation for seeking a more explicit theory of explanation. Finally, I have suggested that we should decouple the idea of mechanistic explanation from the doctrine of methodological individualism. These two ideas should be argued separately, not as grounds for each other. 22 References Achinstein, Peter 1983: The Nature of Explanation. Oxford University Press. Oxford. Bechtel, William 2008: Mental Mechanism. Philosophical Perspectives on Cognitive Neuroscience. London: Routledge. Craver, Carl 2007: Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. Oxford: Clarendon Press. Cummins, Robert 2000: ”How Does It Work?” versus “What Are the Laws?”: Two Conceptions of Psychological Explanation, in Keil & Wilson (eds.) Explanation and Cognition, Cambridge: MIT Press: 117-144. Elster, Jon 1989: Nuts and Bolts for the Social Sciences. Cambridge: Cambridge University Press. Garfinkel, Alan 1981: Forms of Explanation. Yale University Press. New Haven. Gopnik, Alison 2000: Explanation as Orgasm and the Drive for Causal Knowledge: The Function, Evolution, and Phenomenology of the Theory Formation System, in Keil and Wilson (eds.) Explanation and Cognition. Cambridge: MIT Press: 299-324. Hedström, Peter & Richard Swedberg (eds.) 1998: Social mechanisms: An analytic approach to social theory. Cambridge: Cambridge University Press. Hedström, Peter 2005: Dissecting the Social. On the Principles of Analytical Sociology. Cambridge: Cambridge University Press. Hedström, Peter & Petri Ylikoski 2010: Causal Mechanisms in the Social Sciences, Annual Review of Sociology, forthcoming. Hempel, Carl 1965: Aspects of Scientific Explanation. New York: The Free Press. Hesslow, Germund 1983: Explaining differences and weighting causes, Theoria 49: 87-111. Machamer, Peter, Lindley Darden & Carl F. Craver 2000: Thinking About Mechanisms, Philosophy of Science 67: 1-25. Mayntz, Renate 2004: Mechanisms in the Analysis of Social Macro-Phenomena, Philosophy of the Social Science 34: 237-259. Pearl, Judea 2000: Causality – Models, Reasoning and Inference, Cambridge: Cambridge University Press. Rozenblit, Leonid, and Frank C. Keil 2002: The misunderstood limits of folk science: an illusion of explanatory depth, Cognitive Science 26: 521-562. Salmon, Wesley 1998: Causality and Explanation. Oxford: Oxford University Press. Schaffer, Jonathan 2005: Contrastive Causation, The Philosophical Review 114: 297-328. Steel, Daniel 2006: Methodological Individualism, Explanation, and Invariance, Philosophy of the Social Sciences 36: 440-463. Steel, Daniel 2008: Across the Boundaries. Extrapolation in Biology and Social Sciences. Oxford: Oxford University Press. Schwitzgebel, Eric 1999: Children’s Theories and the Drive to Explain, Science & Education 8: 457-488. 23 van Fraassen, Bas 1980: Scientific Image. Oxford: Oxford University Press. Wittgenstein, Ludwig 1953: Philosophical Investigations. Oxford: Basil Blackwell. Woodward, James 2003: Making Things Happen. A Theory of Causal Explanation. Oxford: Oxford University Press. Ylikoski, Petri 2001: Understanding Interests and Causal Explanation. Ph.D. thesis. University of Helsinki, May 2001. [http://ethesis.helsinki.fi/julkaisut/val/kayta/vk/ylikoski/] Ylikoski, Petri 2005: The Third Dogma Revisited, Foundations of Science 10, 395-419. Ylikoski, Petri 2007: The Idea of Contrastive Explanandum, in Persson and Ylikoski (eds.): Rethinking Explanation. Dordrecht: Springer, 27-42. Ylikoski, Petri 2009: Illusions in Scientific Understanding, in De Regt, Leonelli & Eigner (eds.) Scientific Understanding: Philosophical Perspectives, Pittsburgh University Press, 100119. Ylikoski, Petri & Jaakko Kuorikoski 2009: Dissecting Explanatory Power, Philosophical Studies, forthcoming. 24