Tentative Answers to Questions about Causal Mechanisms James Mahoney Recently, many analysts have suggested that the study of “causal mechanisms” should constitute a central part of causal inference in the social sciences. This call comes from both leading methodologists in political science, such as George and Bennett (forthcoming) and Brady and Collier (2004), and leading social theorists in sociology, such as Hedström and Swedberg (1998), Elster (1998), and Stinchcombe (1993). These different scholars all converge on the idea that the study of causal mechanisms can offer insights that are not usually found in mainstream statistical analysis. However, this literature raises many questions about the meaning and uses of causal mechanisms that currently lack clear answers. In part, this state of affairs is a product of the newness of this literature; any new tool of analysis is likely to require refinement and clarification. However, lingering questions about causal mechanisms may also reflect some genuine confusion or perhaps disagreement about the definition and purpose of explanation via causal mechanisms. In turn, this lack of clarity and consensus may lead statistical methodologists to view less seriously arguments about the importance of causal mechanisms. In this paper, I seek to contribute to the literature on causal mechanisms by trying to address what I view as some of the most important lingering questions about causal mechanisms. I focus in particular on four questions: 1) Why do we need to study causal mechanisms? 2) What are causal mechanisms? 3) How can causal mechanisms be used in empirical research? 4) What are examples of causal mechanisms? To summarize the argument, my answers to the four questions are roughly as follows: 1) Causal mechanisms can help analysts overcome the “black box” problem that arises in all modes of causal inference that infer causality by establishing associations between variables. The black box problem specifically is the difficulty analysts have identifying the intervening processes through which an independent variable exerts an effect on a dependent variable. 2) Causal mechanisms have been defined in many different ways, and several definitions contribute important insights. I choose to define causal mechanisms as unobserved entities, processes, or structures that generate outcomes and that do not themselves require explanation. Causal mechanisms are hypothetical “ultimate causes.” 3) Causal mechanisms can be linked to empirical analysis in three ways: the derivation of testable propositions, the integration of existing statistical findings, and the explanation of particular outcomes. The first contribution is especially important to formal modelers, the second to statistical researchers, and the third to case study researchers. 4) I suggest that macro comparative work has drawn implicitly on three main causal mechanisms: the functional needs of social systems; the rationality of individuals; and the power of collective actors. These are the causal mechanisms that underpin functionalism, rational choice theory, and power theory.

Why Do We eed to Study Causal Mechanisms? The concern with causal mechanisms grows out of a distinction sometimes made in the philosophy of science between correlational analysis and causal analysis. Whereas correlational analysis involves identifying antecedents regularly conjoined with outcomes, causal analysis consists of specifying the “mechanism” that underlies and generates empirical regularities and outcomes. In contrast to correlational research, the challenge of causal analysis involves postulating “entities, properties, processes, relations . . . that are held to be causally responsible for the empirical regularities to be explained” (McMullin 1984a, p. 210; see also Blalock 1961, pp. 11-13; Keat and Urry 1982, chap. 2). These mechanisms explain why social regularities exist in the first place; knowledge of their operation allows researchers to go beyond correlations. At the base of this argument is a rejection of the Humean model of constant conjunction for understanding causation. According to Hume (1748/1988), we infer causation when we repeatedly observe putative causes followed effects; that is, two events are “constantly conjoined” in our experience. However, we can never actually confirm the existence of causation because the imagined “necessary connections” linking the two events cannot themselves be observed or known. In contrast to this Humean model, scholars concerned with causal mechanisms argue that exploration of the “black box” connecting independent and dependent variables is essential to good causal research. On this view, “we are not satisfied with merely establishing systematic covariation between variables or events; a satisfactory explanation requires that we are also able to specify the social ‘cogs and wheels’ that have brought the relationship into existence” (Hedström and Swedberg 1998, p. 7; see also Elster 1989, p. 3; Glennan 1996; Harré 1970, p. 230; Harré 1972, p. 170). The black box problem thus refers to the difficulty of explaining why a given causal variable exerts an effect on a given outcome variable. For example, while statistical methods can estimate the average net causal effects of independent variables, they do not themselves provide information about the reasons why the independent variables have the effects they do. These reasons are left as an unopened black box. The “problem” here is twofold. First, because researchers cannot meaningfully identify the connection between cause and effect, they are left uncertain whether a given association reflects true causation or whether the association is simply the spurious product of an unknown antecedent variable. The black box problem thus magnifies the almost inevitable selectivity and omitted variable biases that plague social research (see Lieberson 1985). Second, while researchers often discover that a heterogeneous group of independent variables are statistically associated with an outcome, they lack tools for understanding why such diverse factors are related to the phenomenon. The result is an absence of theoretical integration, which in turn contributes to fragmentation in the social sciences (see Dessler 1991; Sørensen 1998). Researchers usually respond to the black box problem by speculating about the reasons why a given association should exist, perhaps drawing on preexisting studies and general concepts. Sometimes they may attempt to identify intervening variables that link a cause and an effect through statistical procedures (e.g., structural equation and path models) or qualitative techniques (e.g., process tracing and narrative analysis). However, unless a causal mechanism is proposed, these strategies will not resolve the black box problem. Rather, they will end up explaining an association between variables by appealing to another association between variables; the new association itself will

contain a black box and require a separate explanation. Without the identification of an ultimate mechanism, then, the analyst will be forced into an infinite regress as he or she pursues deeper and deeper intervening variables (King, Keohane, and Verba 1994, p. 86; McMullin 1984, p. 206). What are Causal Mechanisms? Unfortunately, a good deal of confusion currently surrounds the precise meaning of causal mechanism. This terminological confusion is one reason why statistical methodologists have not engaged these ideas. Here I attempt to make sense of existing definitions by arranging them into four groups, each of which contributes insight toward a synthetic conception of the term (see Table 1). First, the simplest definitions treat causal mechanisms as synonymous with independent variables or causal factors that help explain outcomes (see Table 1). For example, Boudon defines a mechanism as “the well-articulated set of causes responsible for a given social phenomenon” (1998, p. 172). This definition usefully suggests that causal mechanisms are, in fact, causes, and that they help produce outcomes. However, it does not distinguish the idea of mechanism from the standard notion of an independent variable, raising the question of what the study of mechanisms adds to statistical research. Second, other definitions see mechanisms as intervening variables, events, or processes that explain how one variable influences another (see Table 1). These definitions are useful in that knowledge of intervening processes can help researchers understand why a given independent variable exerts a causal effect on a given dependent variable. In addition, knowledge of intervening processes can increase one’s confidence that a statistical association is not spurious. However, this kind of definition leaves open the possibility that causal mechanisms must be identified and analyzed as statistical associations. That is, to locate a causal mechanism, the analyst must show how intervening processes and events are themselves associated with both independent and dependent variables. In this sense, the analyst explains a correlation by appealing to another correlation, begging the question of why the new association exists (King, Keohane, and Verba 1994, p. 86; McMullin 1984a, p. 206). To explain the new association, the analyst must identify an additional mechanism, which itself will contain a black box and require explanation. Under this definition, in short, the distinction between an independent variable and a mechanism becomes somewhat arbitrary, and the analyst may be forced into an infinite regress in search of deeper and deeper mechanisms. (p.3) Third, other scholars view causal mechanisms as underspecified causal propositions that can be applied to a fairly wide range of cases (see Table 1). For example, according to Elster’s (1998) influential definition, “mechanisms are frequently occurring and easily recognizable causal patterns that are triggered under generally unknown conditions or with indeterminate consequences” (p. 45). 1 This definition assumes that a mechanism identifies a cause-effect relationship that is applicable in many social situations. Such a cause-effect relationship differs from empirical work on correlations in that it is underspecified; that is, it makes reference to analytical constructs that are not actually observed (Rueschemeyer 2001). For example, one of Elster’s mechanisms is the “spillover effect,” defined as follows: “if a person follows a certain pattern of behavior

Pin one sphere of life, X, he will also tend to follow P in sphere Y” (p. 54). This mechanism identifies a probabilistic relationship between an underspecified independent variable (a type of behavior in one sphere) and an underspecified dependent variable (a type of behavior in another sphere). Although these underspecified causal propositions do not have empirical content, they can be used to derive empirical hypotheses. For example, the spillover effect could be used to hypothesize that alienation at work will produce alienation at home, or that greater participation at work will produce greater participation in politics (Elster 1998, pp. 54-55). While clearly useful, this definition of causal mechanism nonetheless fails to explain why the underspecified cause-effect relationship embodied in mechanism itself obtains. For example, with respect to Elster’s spillover effect, the mechanism itself does not explain the process through which a given type of behavior might spillover into a new sphere; that is, it does not explain why the underspecified independent variable affects the underspecified dependent variable. As a result, scholars have no way of knowing when the spillover effect will be in operation as opposed to some other mechanism, including a mechanism that predicts exactly the opposite of the spillover effect, such as the “crowding-out” effect (i.e., if a person follows a certain pattern of behavior P in one sphere of life, X, she will not tend to follow P in sphere Y). In short, this definition of mechanism itself contains a black box, not identifying the reasons why the underspecified association holds and thereby posing problems for empirical research. Scholars associated with the realist school in the philosophy of science offer a fourth definition that, I believe, synthesizes insights from the above definitions while avoiding their shortcomings (see Table 1). This synthetic definition understands a causal mechanism as an unobserved entity, process, or structure that acts as an ultimate cause in generating outcomes. An “ultimate cause” is a cause that itself does not require explanation but nevertheless can generate outcomes (Harré 1970, pp. 101-4; see also George and Bennett 2003; Jasso 1998, pp. 5-6; Steinmetz 1998, pp. 24-25). This definition proposes three main elements that need to be clarified: a) unobservable entity, process, or structure; b) ultimate cause; and c) generates outcomes. The first two elements raise the following two questions. First, why should causal mechanisms be unobserved entities, processes, or structures? And second, what is really (p.4) meant by the phrase “ultimate cause”? My (tentative) answer to these questions is that explanation by causal mechanisms requires that the analyst posit some entity, process, or structure that is treated “as if” it exists, even though at present time scholars cannot be certain that the entity, process, or structure really does exist. Because the entity is hypothetical, it can at least temporarily serve as an ultimate cause – i.e., an unmoved mover that explains outcomes but is not itself open to explanation. If the mechanism were directly observable, it would be clear that this mechanism is not actually the final mover of outcomes in the world, but rather must itself be explained. Hence, causal mechanisms that become observable because of better measurement start to lose their status as causal mechanisms and become regular variables. Yet so long as they are unobservable, scholars appeal to them as the ultimate causes of events in the world. That is, they treat the mechanism as fully independent variable that is not open to also being treated as a dependent variable.

In the natural sciences, researchers posit causal mechanisms in this sense of unobserved ultimate causes. For example, my understanding is that scientists assume the existence of “strings” in superstring theory in this fashion. Strings are the unobservable entities that are believed to be the real mechanisms that physically constitute or otherwise generate phenomena (including statistical associations) in the natural world. String theory is only the latest in a series of previous efforts to identify a master causal mechanism in the natural sciences. And presumably one day scientists might observe strings, such that strings cease to serve this role of master mechanism. The other element of a causal mechanism is the idea that it “generates” outcomes in the world. What does this mean? First of all, “outcomes” can be either some specific event in the world (e.g., the French Revolution) or a relationship between two variables (e.g., economic growth is positively associated with democracy). Hence, causal mechanisms are designed to explain both specific events and the existence of associations between variables. The latter is especially important, because it helps researchers overcome the black box problem. Secondly, a causal mechanism is understood to be the ultimate property that makes these events and associations exist in the world, and thus that can ultimately generate events and associations. The mechanism “generates” these outcomes because it is, on current understanding, the final cause of the things that take place in the world. Three final points about this understanding of causal mechanism are in order. First, I am interested in what might be thought of as “ultimate mechanisms”; that is, mechanisms that underlie and produce much of observable reality. Hence, I do not think that there are dozens of mechanisms in the world. In fact, I believe a coherent theoretical tradition is normally associated with a single causal mechanism. Below I discuss three of these theoretical traditions and their mechanisms. I am aware that other understandings of mechanisms, including the first three definitions reviewed in this section, lead to projects in which scholars of a single theoretical tradition work with a whole “tool kits” of different mechanisms. But I think these understandings of mechanism have real problems, as discussed above. Second, my view is that mechanisms need not be specified at any particular level of analysis, and that micro mechanisms are not necessarily superior or more basic than macro mechanisms. Below I consider mechanisms at three different levels of analysis; I argue that they should be evaluated based on their empirical usefulness. It is true that in (p.5) certain areas of physics and biology more micro mechanisms typically replace more macro ones. However, I am not yet convinced that this must also be the case for the social sciences. For example, it is still possible that something called the “social system” really does exist, and that this system has emergent properties that it cannot be reduced to constituent parts. If this is true, it is also possible that the functional needs of this system are the ultimate causal mechanism driving certain outcomes in the world. Third, I do not see causal mechanisms as intervening variables that stand temporally between an independent and dependent variable. Rather, I see causal mechanisms as ultimate causes that come before both independent and dependent variables, and that produce the relationship that exists between these variables. Causal mechanisms subsume relations between independent and dependent variables; they do not stand between these variables as some kind of intervening process. This is why causal

mechanisms can overcome the infinite regress and black box problems mentioned above. The point might become clearer as I now turn to a review the specific ways in which causal mechanisms can be linked to empirical research. How Can Causal Mechanisms Be Used in Empirical Research? Causal mechanisms understood in the sense above can help empirical analysts in three ways: the derivation of testable propositions, the integration of existing correlational findings, and the explanation of particular outcomes (see Table 2). These three modes correspond roughly with how causal mechanisms can be most effectively used by formal modelers, statistical researchers, and case study/small-N analysts respectively. All three of these contributions assume a simple distinction between propositions and postulates (see Jasso 1988). Postulates are premises made up of assertions about conditions in the world and the mathematical relationships that govern these conditions (Fararo 1989, pp. 17-22; see also Suppes 1957). Propositions are testable hypotheses or predictions about the occurrence of specific events. They are the deduced consequences of postulates, representing the conclusion of a logical chain of reasoning. Taken together, postulates and propositions follow the rules of a syllogism: if the postulates are true, the proposition must also be true. More specifically, if the postulates are true, they are jointly sufficient for the truth of the proposition; hence, the proposition follows necessarily from the truth of the postulates. 2 In this formulation, one postulate – often the first postulate – is a causal mechanism. Since a causal mechanism refers to an unobservable phenomenon, subsequent postulates often are “bridging assumptions” (see Lindenberg 1992; Kelle and Ludemann 1998; Morton 1999) that empirically specify this mechanism. Additional postulates may then entail assumptions about conditions in the world that work in conjunction with the empirically specified mechanism. Overall, the analyst moves from a (p.6) high level of abstraction (i.e., a causal mechanism) to a concrete level of abstraction (i.e., an empirical proposition) (see Lindenberg 1992 on the “method of decreasing abstraction”; also Gerring 2001:196-97 on mechanisms and a “step-by-step” approach). In a research program built around a particular causal mechanism, it is possible that certain accepted strategies will develop for moving from a mechanism to propositions. In this sense, scholars do not necessarily have to start from scratch in their model construction but can sometimes rely on the postulate formulations of previous researchers. Proposition derivation entails the use of untested postulates to logically derive hypotheses that can be tested (Cohen 1989; Merton 1949; Friedman 1953; Homans 1967; Stinchcombe 1993). As Jasso (1988) puts it, “Test the predictions [i.e., propositions], never the postulates” (p. 19). 3 As a general rule, scholars should strive to use as few postulates as possible to generate as many testable propositions as possible. The postulates themselves (including the causal mechanism) are judged based on their capacity to yield empirically supported hypotheses, especially hypotheses that are counterintuitive vis-à-vis commonsense or other theoretical expectations (Popper 1968). Propositions derived from postulates can be false for two reasons. First, the analyst may have violated a logical or mathematical rule in the process of reasoning from postulates

to a proposition. Proposition derivers must therefore be concerned with checking “the postulates for logical consistency, doing whatever repair is necessary in order to achieve internal coherence” (Jasso 1988, p. 9). Second, one or more of the postulates – including an initial postulate about a causal mechanism – may be false. Even if a testable proposition is derived in a mathematically valid manner, we have no reason to believe it will be true if the assumptions from which it was derived are not true. Thus, when a proposition derived from a causal mechanism is not empirically supported, the analyst must review whether the problem was one of a breakdown in deductive reasoning or the falsity of a postulate. If neither condition applies, doubt is cast upon the causal mechanism. 4 Knowledge integration works backwards from an existing set of propositions that have already been tested and empirically supported to a set of postulates. Theorists who use this strategy try to illustrate how diverse statistical associations can all be viewed as the product of the same basic causal mechanism. The key difference between these (p.7) knowledge integrators and the proposition derivers is that the former scholars start with pre-existing and already tested propositions and then work to a set of postulates from which these propositions can be logically deduced, whereas the latter researchers begin with the postulates and work forward to propositions that were not previously tested or known. Knowledge integration helps overcome the fragmentation that arises when researchers discover a diverse range of independent variables that are all statistically associated with a particular outcome. This strategy can show how the associations are the product of a single causal mechanism, supplementing empirical research that begins without the aid of a general theory. However, when compared to proposition deriving, the exercise of knowledge integration provides less convincing support for the existence of a causal mechanism, given that the validity of the propositions was already known in advance (e.g., Abell 1994; Cohen 1989; Hage 1994). Finally, outcome explanation refers to the theoretical practice of logically deducing particular historical outcomes or events – rather than testable hypotheses – from a set of postulates. A historical outcome might be anything from occurrence of the French Revolution to the electoral victory of George W. Bush. When using this strategy, the analyst does not test the outcome that makes up the final proposition, since this occurrence already has taken place. Rather, the analyst seeks to test as many as possible of the postulates used to logically deduce the outcome. The strategy of outcome explaining is thus fundamentally different than the strategies discussed above – in particular, outcome explaining violates the maxim “test the propositions, never the postulates.” Scholars using this strategy test the postulates because they contain the hypothesized explanation for the outcome. If these analysts cannot establish empirical support for the postulates, they are in a weak position arguing that the postulates explain the outcome at hand. In general, therefore, outcome explainers seek to formulate postulates that are highly testable and falsifiable but empirically supported. Because causal mechanisms are unobserved entities, the initial postulate about the causal mechanism cannot be tested in the outcome-explaining strategy. It is therefore imperative that outcome explainers formulate other postulates that can be empirically evaluated. If these other postulates are eventually supported, one’s confidence in the

existence of the causal mechanism increases, provided that the overall set of postulates allows for the logical derivation of the outcome of interest. In the end, nevertheless, the analyst must make a leap of faith in the existence of the causal mechanism, since this beginning postulate is never directly tested, and since the final proposition (i.e., the outcome of interest) is also not tested (because it has already taken place). 5 In important respects, the strategy of outcome explaining is best suited for case study and small-N researchers, who often seek to explain particular outcomes. By contrast, knowledge integration is more useful to statistical researchers, who often discover that many heterogeneous variables are related to an outcome, but lack a means (p.8) of understanding why this is true. Finally, the first strategy of proposition derivation is very similar to how formal modelers view the purpose of theory – that is, one of logically deducing testable proposition from premises. What are Examples of Causal Mechanisms? In the field of comparative-historical sociology, which is the field that I know best, analysts implicitly or explicitly use causal mechanisms associated with three general theories: functionalist, rational choice, and power theories. These theories differ with respect to the causal agents and causal mechanisms that comprise their core assumptions (see Table 3). In functionalist theory, the causal agent is the social system and the causal mechanism is the “needs” or “functional requisites” of that system (Aberle et al. 1950; Levy 1952, p. 151; Turner and Maryanski 1979). Functionalism is built around the assumption that social systems really exist, that these systems have certain potentially unobservable requisites that must be met for their survival, and that those requisites are an ultimate cause of outcomes in the social world. Functionalism is a truly macro theory in that it attributes final causation to system-level attributes rather than attributes of micro-level units such as individuals or meso-level units such as groups and organizations. 6 By contrast, rational choice theory is a micro-level theory. 7 It assumes that individuals are the basic agents of social analysis and that the instrumental rationality of these individuals is the causal mechanism that produces events in the social world. Instrumental rationality is defined above all by the optimization of interests (see, e.g., Hargreaves Heap 1989; Abell 1992; Coleman and Fararo 1992; Kiser and Hechter 1991). That is, individuals evaluate behavioral options in light of their costs and benefits, and they pursue that option that maximizes the differences between benefits and costs. The approach therefore assumes that individuals are purposive, goal-oriented, and intentional actors. Beyond this, however, rational choice theory does not directly identify the content of any individual interests, choice options, or potential pay-offs. 8 Finally, power theory works at a meso-level. It assumes that collective actors (e.g., social groups and organizations) are the key causal agents and that the exercised capacity (i.e., power) of these actors is the ultimate cause of social happenings. Of the (p.9)

three general theories considered here, power theory has been the least developed in any formal way. But elsewhere I have tried to flesh it (Mahoney 2003). One might be tempted to evaluate functionalist, rational choice, and power theories based on the “realism” of their causal agents and causal mechanisms. For example, individuals might seem to be a more realistic entity than social systems, and therefore rational choice theory might seem to have an inherent advantage over functionalism. However, caution should be exercised when basing critiques of general theories upon the empirical plausibility of their key assumptions (Bhaskar 1975; Churchland and Hooker 1985; Hacking 1983). Just as useful general theories in the natural sciences often make reference to hypothetical and unobservable entities (consider gravity and quarks), one should not dismiss out of hand social science theories that posit abstract entities and mechanisms. The importance of assessing general theories on the basis of their empirical contributions can be obscured if scholars are too focused on the normative implications of general theories. For example, scholars who believe that rational choice theory encourages egoistic and self-regarding behavior may be inclined to dismiss the theory independent of its empirical merits. In fact, however, the normative implications of general theories are rarely straightforward. 9 While scholars must be accountable for both the empirical and normative implications of their general theories, separate debates can and should be held over these issues. Table 1. Glossary of Definitions of “Mechanism” (adapted from Mahoney 2001; for references to all works before 2001 see that article) I. Mechanism as a cause of an outcome Boudon (1998, p. 172): “A SM [social mechanism] is, in other words, the wellarticulated set of causes responsible for a given social phenomenon. With the exception of typical simple ones, SMs tend to be idiosyncratic and singular.” Cowen (1998, p. 125): “I interpret social mechanisms (defined in greater detail below) as rational-choice accounts of how a specified combination of preferences and constraints can give rise to more complex social outcomes.” Elster (1989, p. 3): “nots and bolts, cogs and wheels – that can be used to explain quite complex social phenomena.” Tilly (forthcoming): “Mechanisms are events that alter relations among some specified set of elements.” II. Mechanism as an intervening process, event, or variable Bennett and George (1997, p. 1): “the processes and intervening variables through which causal or explanatory variables produce causal effects.” Hedström and Swedberg (1998, p. 11): “Mechanism-based explanations usually invoke

some form of ‘causal agent’ that is assumed to have generated the relationship between the entities observed.” Hedström and Swedberg (1998, p. 13): “Mechanisms . . . are analytical constructs that provide hypothetical links between observable events.” Keat and Urry (1982, p. 30): “causal explanations require the discovery both of regular relations between phenomena, and of some kind of mechanism that links them. . . . In describing these mechanisms and structures we will often, in effect, be characterizing the ‘nature’, ‘essense’, or ‘inner constitution’ of various types of entity.” King, Keohane, and Verba (1994, p. 85): “Some scholars argue the central idea of causality is that of a set of ‘causal mechanisms’ posited to exist between cause and effect. This view makes intuitive sense: any coherent account of causality needs to specify how the effects are exerted.” Kiser and Hechter (1991, p. 5): “A complete explanation also must specify a mechanism that describes the process by which one variable influences the other, in other words, how it is that X produces Y.” Koslowski (1996, p. 6): “A causal mechanism is the process by which a cause brings about an effect. A mechanism is a theory or an explanation, and what it explains is how one event causes another.” Little (1991, p. 15): “A causal mechanism, then, is a series of events governed by lawlike regularities that lead from the explanans to the explanandum.” Mahoney (2000, p. 531): “Causal mechanisms are the intervening processes through which one variable exerts a causal effect on another variable.” Somers (1998, p. 726, citing Coleman [1986, p. 1328]): “meaningful connection between events as the basic tool of description and analysis.” Sørensen (1998, p. 240): “My definition of ‘mechanism’ is simple: It is an account of how change in some variable is brought about – a conceptualization of what ‘goes into’ a process. III. Mechanism as an underspecified causal process Elster (1998, 45, his emphasis): “Roughly speaking, mechanisms are frequently occurring and easily recognizable causal patterns that are triggered under generally unknown conditions or with indeterminate consequences.” Gambetta (1998, p. 102): “I take ‘mechanisms’ to be hypothetical causal models that make sense of individual behavior. They have the form, ‘Given certain conditions K, an agent will do x because of M with probability p.’ M refers either to forms of reasoning governing decision making (of which rational choice models are a subset) or to subintentional processes that affect action both directly (as impulsiveness) or by shaping preferences or beliefs.”

McAdam, Tarrow, and Tilly would go here. Schelling (1998, pp. 32-33): “I propose . . . that a social mechanism is a plausible hypothesis, or set of plausible hypotheses, that could be the explanation of some social phenomenon, the explanation being in terms of interactions between individuals and other individuals, or between individuals and some social aggregate.” Stinchcombe (1998, p. 267): “I have defined mechanisms before as bits of ‘sometimes true theory’ or ‘model’ that represent a causal process, that have some actual or possible empirical support separate from the larger theory in which it is a mechanism, and that generate increased precision, power, or elegance in the large-scale theories.” Rueschemeyer (2001, p. 31): “Incomplete theoretical propositions . . . a causal hypotheses, but one whose conditions are insufficiently specified.” IV. Mechanism as an unobserved entity that generates outcome Bennett and George (2003): “we define causal mechanisms as ultimately unobservable physical, social, or psychological processes through which agents with causal capacities operate in specific context to transfer energy, information, or matter to other entities.” Bhaskar (1979, p. 15): “the construction of an explanation for . . . some identified phenomenon will involve the building of a model, utilizing such cognitive materials and operating under the control of something like a logic of analogy and metaphor, of a mechanism, which if it were to exist and act in the postulated way would account for the phenomenon in question.” Goldthorpe (2000, p. 149): “some process existing in time and space, even if not perhaps directly observable, that actually generates the causal effect of X on Y and, in so doing, produces the statistical relationship that is empirically in evidence.” Harré (1970, pp. 101, 102, 104): “the structures, states and inner constitutions from which the phenomena of nature flow . . . . the permanent or enduring conditions under which a certain kind of phenomenon will occur.” “The inner constitutions, structures, powers, encompassing systems, and so on, of which natural generative mechanisms are constituted, and of which the connection between cause and effect usually consists.” Steinmetz (1998, pp. 177-178): “Within critical realism, a law is not a constant conjunction of events but the characteristic pattern of activity, or tendency, of a mechanism. More specifically, real structures possess causal powers which, when triggered or realized, act with natural necessity and universality as generative mechanisms.” [Accompanying footnote: “Generative mechanisms are ‘tendencies’ rather than ‘powers’ because they are not just potentialities but potentialities that may be exercised without being manifested.”] V. Other definitions

Hernes (1998, p. 74): “A mechanism is a set of interacting parts – an assembly of elements producing an effect not inherent in any one of them. A mechanism is not so much about ‘nuts and bolts’ as about ‘cogs and wheels’ – the wheelwork or agency by which an effect is produced. But a mechanism or inner workings is an abstract, dynamic logic by which social scientists render understandable the reality they depict.” Stinchcombe (1993, pp. 24-25): “As I use the word, mechanism means (1) a piece of scientific reasoning which is independently verifiable and independently gives rise to theoretical reasoning, which (2) gives knowledge about a component process (generally one with units of analysis at a ‘lower level’) of another theory (ordinarily a theory with units at a different ‘higher’ level), thereby (3) increasing the suppleness, precision, complexity, elegance, or believability of the higher-level theory without excessive ‘multiplication of entities’ in it, (4) without doing too much violence (in the necessary simplification at the lower level to the make the higher-level theory go) to what we know as the facts at the lower level.”