Meaning and Modularity: The Multivalence of “Mechanism” in Sociological Explanation

What is Mechanistic Explanation?

To explore how mechanisms function in cultural explanations, it is helpful to first set out some terminology. In this article, we use causal system to refer to all the causal processes at play in a given phenomenon under study; we refer to mechanisms as causal structures that obtain within those systems. Mechanism models are representations of mechanisms, meaning they are models of causal structures and are almost never models of entire causal systems themselves. Finally, we consider mechanistic explanation to be achieved when an analyst provides an accurate mechanism model of an actual mechanism (or set of mechanisms).

Yet what are the mechanisms that social researchers attempt to model? Given the term’s ubiquity, it is little surprise that the concept is saddled with a diversity of not always consistent definitions. As an exhaustive analysis of these various approaches is outside the bounds of this article, we first distinguish between two general theoretical claims associated with mechanisms and limit our discussion to one of these.

There are, broadly, two main claims often made by social mechanists (Levy 2013). The first is ontological and states that causal relationships exist in virtue of mechanisms. In contrast to a Humean account of causation as constant conjunction (or more broadly, a pattern of association among events), proponents of this “realist” view hold that events are only linked insofar as underlying mechanisms give rise to them: “Mechanisms rather than regularities or necessary/sufficient conditions provide the fundamental grounding of causal relations” (Little 2012:5). ¹

The second claim is epistemological and is a claim about social explanation. Social mechanists making the epistemological claim argue that good explanations must reference mechanisms—not because the inner nature of causation itself is grounded in mechanisms but rather because mechanistic explanations are superior to other approaches at depicting causal structures (mechanisms are classically juxtaposed with statistical black-box analyses). ² For instance, the epistemological warrant for mechanistic explanation is a core principle of analytic sociology: “one should . . . aim for models which show how a proposed mechanism generates the outcome to be explained” (Hedström and Bearman 2009:6).

While these two claims are related and those advocating the ontological approach typically also argue for mechanism models’ explanatory importance, they are conceptually distinct. Realist proponents of mechanisms as the ontological basis of social causation may not always advocate for a given mechanism’s explanatory relevance (as in the case of “unmanifest” mechanisms; Steinmetz 1998:178). Similarly, scholars advocating the explanatory claim do not commit themselves to arguing for mechanisms as the essence of causal relations. ³ One can coherently advocate for mechanistic explanation while seeking a “deeper” metaphysical account of causation elsewhere (Franklin-Hall 2016; Hedström and Ylikoski 2010).

In this article, we focus on the epistemological approach to mechanistic explanation for two main reasons. First, recent invocations for meaningful mechanisms typically focus on cultural explanation rather than theories of causation, so we wish to deal with these endeavors on their own terms. For instance, Tavory and Timmermans (2013:688) advocate for mechanism models of pragmatist semiotics while conceding that mechanisms themselves do not “solve the metaphysical problem of causality.” ⁴ Second, focusing on the epistemological approach allows us to analyze mechanistic explanation while setting aside many of the thorny philosophical problems that arise in realist accounts and that are less directly relevant for scholars engaged in proposing and analyzing sociological models. ⁵

Modular Mechanism Models and Carving Up Causes

We argue that differences in how social analysts fulfill the causal, generality, and carving criteria result in incompatible approaches to mechanism models. To establish this claim, we first consider the modular approach, exemplified by Hedström and Ylikoski’s (2010:52) well-cited explication of social mechanisms:

A mechanism-based explanation describes the causal process selectively. It does not aim at an exhaustive account of all details but seeks to capture the crucial elements of the process by abstracting away the irrelevant details. The relevance of entities, their properties, and their interactions is determined by their ability to make a relevant difference to the outcome of interest. If the presence of an entity or of changes in its properties or activities truly does not make any difference to the effect to be explained it can be ignored. This counterfactual criterion of relevance [italics added] implies that mechanism-based explanation involves counterfactual reasoning about possible changes and their consequences.

Hedström and Ylikoski (2010) ground their account in a central question that faces all social analysts engaged in mechanistic explanation: Which elements of noisy social phenomena should be carved out as truly causal in modeling that phenomena? ⁷ Hedström and Ylikoski’s answer defines the modular approach: Relevant and irrelevant elements of a model are distinguished according to whether an element’s alteration would change the outcome—that is, according to the interventionist counterfactual criterion of relevance. ⁸ On this view, to provide a causal explanation for a phenomenon is to identify the preceding events on which the phenomenon depends, in the specific manipulationist sense of depends. That is, one must identify the events that could be “intervened” on to change that phenomenon, enabling researchers to answer questions regarding how things could have been otherwise. ⁹

This interventionist counterfactual account conforms with many intuitive notions of causal identification and has been explicitly adopted by mechanists in both philosophy (e.g., Craver 2006) and sociology (e.g., Demeulenaere 2011; Hedström and Ylikoski 2010; Morgan and Winship 2007). ¹⁰ In sociology, this account is often cited by scholars in the tradition of (particularly European) analytic sociology (e.g., Demeulenaere 2011; Kuorikoski and Pöyhönen 2012), but it is implicitly assumed by a much broader array of sociological scholarship. Interventionist counterfactuals are built into many statistical methods used in causal analysis (Gangl 2010; Gelman 2011), and they are especially relevant when mechanisms are modeled as mediating processes. For instance, Imai and colleagues (2011) explicitly draw on the potential outcomes approach to causal inference in developing a mediation framework that decomposes causal effects into independent mechanistic pathways (for applications, see Pager and Pedulla 2015; Pedulla 2016).

Yet even where mechanisms are not explicitly framed in the counterfactual approach, they are regularly understood as analytically independent pathways linking cause to effect. Such models are ubiquitous in sociological research and appear, for instance, in analyses of the independent processes generating wage inequality (e.g., Cheng 2014), educational outcomes (Alon 2009; Tam and Jiang 2014), and friendship networks (Schaefer, Kornienko, and Fox 2011; Wimmer and Lewis 2010). Across these studies, the causal efficacy of various mechanistic pathways is assessed by analyzing how variation in the outcome depends on variation in observed measures of the mechanisms.

The interventionist approach is most apparent in studies using statistical methods, but the general logic also appears in ethnographic, historical, and policy-related analyses. Rather than “mediation” or “decomposition,” ethnographers often rely on comparative strategies that borrow from the rationale of experimental design (Abend et al. 2013). For instance, Small (2009:23) encourages researchers to devise hypotheses from their cases of the form “when X occurs, whether Y will follow depends on W.” Through selecting positive and negative cases, process-tracing depends on identifying “whether the condition or mechanism makes a difference to the outcome” (Rohlfing 2014:617). Policy analysts, too, adopt an interventionist framework when they conceive of mechanisms as potential social levers for policy intervention (e.g., Reskin 2003).

At this point in our argument, it is important to recognize that the interventionist account is not without its detractors (e.g., Groff 2011; Porpora 2015). But for our purposes, what matters is not whether this approach is, in fact, the correct account in the sense of a final, philosophical definition of causation. Rather, our goal is to examine whether a prominent account of causality that features in and is indeed required by many methodological approaches to social mechanistic explanation can be reconciled with interpretative and semiotic mechanism models.

We argue that the difficulty arises when one considers that the counterfactual approach to causality entails a more technical requirement. Namely, when the interventionist approach is applied to causal systems, it requires the elements of a representation of that causal system to be modular. In brief, the elements of a causal model are considered modular when changes to one part of the model do not affect the relations among the other parts of the model. When applied to mechanisms, the idea is that if two mechanisms are components of a causal structure, then they should be independent such that it is possible, at least in theory, to alter one component without altering the relationship among other components. ¹¹ That is, “if two mechanisms are genuinely distinct, it ought to be possible to interfere with one without changing the other” (Hausman and Woodward 2004:158). ¹²

To see how this works in practice, consider an example. In his study on how nonstandard employment histories lead to discrimination in the labor market, Pedulla (2016) posits two potential mechanisms that could account for why employers negatively evaluate job candidates with nonstandard employment histories: (1) assessments of competence and (2) assessments of commitment to work. Modularity entails that if both these processes are included in a correctly specified model, it should be possible, at least in theory, to alter an employee’s commitment to work (or the employer’s assessment of that commitment) without necessarily altering (i.e., while holding constant) the employee’s competence (or the employer’s assessment of competence). ¹³ Importantly, modularity does not require that there be no interactions or causal complexity. Modularity only implies that such interactions or complexity must be specified features of the causal model. If, for instance, changing assessments of commitment necessarily altered assessments of competence, this would imply we were missing a causal link between these mechanisms. The key takeaway is that if a given mechanistic process is not causally connected to another process, then altering that first process should not alter the functioning of the second.

It is important to differentiate modularity, a feature of a representation of a causal system, from the distinct concept of statistical independence. Statistically independent events are ones where the joint probability of a set of events is equal to the product of their individual probabilities; in other words, statistically independent events are like coin flips, where getting heads on the first flip provides no information on the chance of getting heads on the second. ¹⁴ Relations of statistical dependence say nothing about causal relationships; causally related events will necessarily be statistically dependent, yet statistical dependence is obviously not sufficient for causation. ¹⁵

The notion of modularity specifies how the modular approach to mechanistic explanation fulfills the causal and carving criteria to show “how a system works.” Like the “cogs and wheels” to which they metaphorically refer, modular mechanism models “carve” a complex causal system into independent, component processes and identify the causal dependence relationships among these components. Crucially, however, there are plenty of instances where modularity is thought to fail (Cartwright 2007). Indeed, there are entire regions of sociological argument in which modularity does not and in fact cannot hold. We now turn to one of those regions.

Mechanisms and the Modeling of Meaning

Recently, social theorists have pointed to the notable absence of meanings, symbols, and signs from standard approaches to social mechanisms. Gross (2009) argues that although meaning is central for social action and therefore social causation, classic approaches to mechanisms have drawn on culturally impoverished accounts of strategic action predicated on individualized beliefs. Merton’s (1968) archetypal social mechanism of a “self-fulfilling prophecy,” for example, emphasizes how actors’ beliefs about a bank’s solvency can cascade into collective panic, leading to a bank run that no individual actor intended or desired to bring about. Yet, as Gross (2009:369) rightly points out, “Merton’s postulated mechanism of the self-fulfilling prophecy functions in this case only because actors are positioned in cultural systems from which they derive their assumptions and orientations, and hence have the beliefs they do.” In other words, by emphasizing individual beliefs, Merton’s mechanism ignores the very feature that makes possible the synchronous, uncoordinated, and yet collective nature of a bank run: cultural systems that organize diverse actors’ understandings about the health of their country’s financial institutions, their trust in government, their own financial security, and therefore their best course of action.

Following Gross’s (2009) argument, the question for the researcher then becomes how to represent the causal efficacy of meaning in a mechanism model. In specifying the integration of meaning in mechanistic explanations, sociologists often focus on relations among signs made present in action and interaction. For example, Norton (2014:167) portrays meaningful action as emerging from “networks of signs and signifiers understood in meaningful relation to one another.” Here, cultural causality is realized as actors constantly interpret and perform their understandings of the social situations in which they find themselves. Crucially, these sequences of interpretations and performances iterate such that what Norton (2014:173) calls “assemblage[s] of cultural codes and systems” continuously realize their causal effect within and across situations. Tavory and Timmermans (2013) similarly ground meaningful mechanisms in sequences of semiotic chains composed of objects (things in the world), signs (signifiers of objects), and interpretants (the reaction of the interpreter of a sign). Successions of meaning-making amass as “each understanding or action” produces “another round of meaning making” where “one iteration of meaning making may become the sign for another” (Tavory and Timmermans 2013:688). These accounts therefore coalesce on the view that culture realizes its causal effects through accumulating sequences of interpretation and action. Returning to the criteria we laid out earlier, these models appear to fulfill the three requirements of mechanistic explanation: They (1) depict a causal process, (2) refer to general sequences that explain varieties of behaviors across contexts, and (3) carve the social world into constitutive parts—semiotic chains actualized by actors’ interpretations and performances.

The problem for mechanistic explanation arises from the fact that although meaningful mechanism models depict causal processes, decomposing them into their constituent parts, they do so in a manner that is fundamentally at odds with the core premises of modular mechanistic explanation. Recall that a modular model decomposes complex phenomena into discrete causal pathways, providing counterfactual information that answers the question of how things could have been otherwise given an intervention on a particular pathway, holding constant the workings of the causal system elsewhere. This characteristic requires that the model’s constituent parts are independent such that it is possible, at least in theory, to intervene on one part of the model independent from the remainder of the system. Yet consider this feature of mechanistic explanation when a model instead refers to the stringing together of signs in action to create and habitually maintain meaning. As we will show, our extant theories of culture in sociology all suggest that modularity is violated by such models.

The sociological literature on culture contains extensive debates about how precisely to characterize the process whereby meanings emerge and become more or less consistent or fixed. For example, scholars debate the impact of humans’ cognitive capacities on how meanings emerge and about whether to think of culture, metaphorically speaking, as a “public document” (Geertz 1973:10) or instead as a series of subjective states molded by intersubjective interaction but ultimately grounded in individual belief or practice.

However, despite this diversity of approaches, almost all this research either states explicitly or implies that the assignment of meaning to a given term, person, thing, or situation is dependent on the assignment of meaning to other related terms, persons, things, or situations. This can be true of very small meaning “systems” that occur for only a few people or among small groups ¹⁶ or for systems drawn on by large populations. ¹⁷ This relationality implies that if one meaning in a schema of categorization changed fundamentally, then the other meanings would also change. Given a set of meaningful words, a set of emotionally charged symbols, or a set of gestures that come one after another, a change in the meaning of “sign A” would require a change in the meaning of “sign B,” and this change can occur at the level of denotation, connotation, or both. If cats are not associated with femaleness, then dogs cannot be “male” in the symbol system cited by Sahlins (2013:219). This relationality is present in both classic philosophical and linguistic accounts of meaning and in and across theoretical debates in the sociology of culture.

Perhaps most famously, the relationality of meaning forms a part of Saussure’s linguistics, which heavily influenced structuralist approaches to culture in American sociology (Alexander and Seidman 1990). Saussure’s semiotics introduces a grid-like failure of modularity due to the dependence of one side of a meaningful binary on the other and the emergence of a set of complexly interlocked binaries in the “deep meanings” that organize culture at the mythological level (Barthes 2012; Reed 2015). As Lévi-Strauss ([1963] 2008:210) wrote, “[I]f there is a meaning to be found in mythology, it cannot reside in the isolated elements which enter into the composition of a myth, but only in the way those elements are combined.”

Sociologists have argued that there are limitations to Saussurean structuralism and its Durkheimian manifestation in the sociology of culture (Gross 2005; Lamont 1998). But importantly, the relationality of the meanings in which action is embedded is not exclusive to the theories of meaning found in cultural structuralism. A variant also appears in the pragmatist tradition. For example, a relational account of meaning and action appears in Quine’s analysis of science. ¹⁸ Quine’s pragmatist account of scientific activity arrives at the key conclusion that “it is misleading to speak of the empirical content of an individual statement” (Quine 1951:40). Rather, science is best conceived as a “field” such that when experience contradicts scientific knowledge, a process of holistic reevaluation of the statements of the field takes place. Crucially, this process is relational: “re-evaluation of some statements involves re-evaluation of others, because of their logical interconnections,” and thus “no particular experiences are linked with any particular statements in the interior of the field, except indirectly through considerations of equilibrium affecting the field as a whole” (Quine 1951:40). Undoubtedly, Quine’s theoretical picture of the relationship between meaning and action—not to mention his philosophical concerns and orientations—are quite different from that of Saussure, Lévi-Strauss, or Durkheim. But a relationality of meaning is presupposed by both genres of argument, and so in both cultural structuralism (and its poststructuralist heterodoxy) and midcentury pragmatist philosophy of science, modularity is disabled for the processes being described.

We see a similar intellectual pattern in several key theoretical statements of the sociology of culture. Lamont (2001:176) understands “patterns of boundary work not as essentialized individual or national characteristics but as cultural structures, that is, institutionalized cultural repertoires or publicly available categorization systems.” These systems rely simultaneously on several crisscrossing meanings; for example, white French working men define themselves in contrast to various others, such that a change in either the salience or connotation of one “other” (e.g., an increase in the population of blacks in France or a shift in the connotation of the category “Arab”) could change the overall space or set of meanings used to construe identities.

Similarly, Swidler’s (2001) toolkit model of culture, although developing a significantly different set of metaphors for understanding culture in action, features relationality when theorizing a working “repertoire” of cultural meanings. For Swidler, relationality inheres not in a unified system of meanings but rather in the relationship between meanings and the strategies of action—themselves cultural products—on which such meanings depend: “Indeed, the meanings of particular cultural elements depend, in part, on the strategy of action in which they are embedded” (Swidler 1986:281). In Talk of Love, Swidler (2001:184) emphasizes the interdependence of meanings and institutions such that “culture conveys meanings through adherence to or deviation from locally established expectations or conventions.” Thus, for Swidler, both the relationship between aspects of a given actor’s repertoire and the relationship of repertoires to the institutional settings that repertoires are used to navigate are constituted by the interdependence of meanings on each other. ¹⁹

Crucially, this feature of relational interdependence is not incidental to meaningful mechanisms. Rather, it is at the core of their explanatory power. In describing how meanings causally shape sequences of situations, Norton (2014:32 [italics added]) discusses how an actor’s performance “fatefully reshape[s] the environment of action” such that subsequent new environments “include the [previous] action and is thus different from what came before.” Thus, the relationships between signs that enable meaning-making are not only causally antecedent to but also partially constitutive of the interpretations, interactions, and performances they produce. Similarly, drawing on a Peircian rather than Saussurean picture of meaning-making, Tavory and Timmermans’s (2013:688) account depicts mechanism models as a “spiral of meaning making, where the interpretant of one iteration of meaning making may become the sign for another.” In a longer exploration of the logic of qualitative research, Tavory and Timmermans (2014) develop Peirce’s arguments about dynamic interpretants—thus moving away from Saussurean semiotics—to characterize the nature of inference and sociological interpretation. Yet here too, such models demonstrate relationality:

The shape the interpretant ends up taking is thus defined partially by the sign-object, but is also replete with the entire history of action, interpretation, and inference that interpreters bring to the observation. . . . Signification is not an isolated act but part of an ongoing flow of action in which signs build on each other. (Tavory and Timmermans 2014:28) ²⁰

The lively debates in the sociology of culture mobilize powerful and evocative models and provide a wide spectrum of accounts of how meanings are mobilized in social life. But these models are not modular. A mechanism model of linked situations, cascading such that each new situation contains within it elements of the previous upon which its interpretation depends—or spirals of meaning-making such that each new round contains as its referent the semiotic content of the previous—is fundamentally relationally interdependent. The crux of the issue is that we cannot, even theoretically, imagine altering one component in such a model and hold fixed that which follows. To intervene on one component part of a meaningful system does not just causally alter what follows: It also necessarily and definitionally alters it. Because meanings come into being via one version or another of semiosis—whether through a semiotic net or grid (a “culture structure”) or the more granular and pragmatic process of cascading, evolving interpretations made in a series of problem-situations over time—the counterfactuals cultural theorists would be inclined to those where a shift in a meaning entails a repertoire or structure whose meanings are reconfigured at the holistic level. ²¹ The relationships among meanings, frames, and symbols themselves are not modular, and tracing precisely that relationality is what makes cultural analysis able to explain its subject matter.

We contend that the position of cultural theory—implicit or explicit—is that meaningful relationality disables modularity. At its core, the relational theory of meaning is one in which the sense and reference that are indicated via signs exist in distinctly nonmodular relationships. The relationship between signs that produce meaning cannot be considered as independent causal processes. Note that in the example of a modular mechanism (see the previous section), if nonmodularity arose, it was the result of missing some important element of our causal structure: If commitment and competence were necessarily linked such that all interventions on commitment changed competence, then this indicated we were missing a causal relationship between the two mechanisms. If we turn to the case of meanings, however, the dependence we are imagining is not based on a causal relationship that is missing from the model but rather is based on a semantic interdependence characterized by denotation or connotation. We cannot add in another causal arrow to “fix” the problem of nonmodularity when we are modeling meanings. Nonmodularity is, for cultural theorists, a feature, not a bug.

To illustrate the nonmodularity characteristic of these systems, we can consider an example of a meaningful mechanism model elaborated by Tavory and Timmermans (2013:697–703). Tavory and Timmermans draw on ethnographic evidence to investigate how parents of newborns react to information that their infant has tested positive in an initial screening for a metabolic disorder. As they note, only a small fraction of infants who initially test positive will be diagnosed as a true positive, and therefore clinicians often find themselves informing parents whose children initially screened positive that their child is likely healthy (importantly, clinicians lack a definitive test that can confirm this healthy diagnosis). The puzzle arises in how parents react to this new information. Even after being told their child likely does not have the disorder, many parents hold on to the “disease frame” in which their child is still sick, a response that causes the clinician-parent interaction to go awry. In proposing a mechanism to explain these reactions, Tavory and Timmermans (2013:700) look for the causal “culprit” at “the semiotic lower level of aggregation.” Their discussion of the mechanism is worth quoting at some length:

The proposed ethnographic explanation . . . in the study highlights parents and clinicians’ reaction to the absence of an interpretative frame. When all parties agreed that the child was sick or healthy, clinicians and parents knew how to handle the situations and disagreements were rare. . . . The interesting divergences and semiotic changes over time and over place occurred when parents and clinicians faced a situation in which a child is potentially ill, but shows no symptoms and may never have any symptoms. This situation is uncharted territory. When there is no actionable frame for diseased-but-now-basically-healthy status, parents tended to fall back on the frame of the child as diseased while professionals gravitated either to a likely false positive understanding or to a “carrier” understanding. (Tavory and Timmermans 2013:702)

In other words, interactions between parents and clinicians (i.e., whether the interaction goes well or goes awry) depend on the relationship between parents’ and clinicians’ frames (i.e., whether the separate parties adopt the same or conflicting frames). When interactions do go awry, the semiotic “causal culprit” is the structure of the semiotic system available to parents and clinicians. That is, parents and clinicians have at their disposable a “diseased” frame and a “healthy” frame, but no frame maps on to the status of a child who is “diseased-but-now-basically-healthy.” Tavory and Timmermans’s (2013) puzzle is solved when we see that the semiotic structure accessible to parents and clinicians does not map on to the child’s status, and thus parents retain their original interpretation. Early frames determine subsequent frames, and the structure of available frames carves out the space of possible interpretations and actions.

Now consider the modularity requirement. One cannot imagine altering the “absence of a frame” while simultaneously holding constant the healthy and diseased frames. To introduce a new frame into what was once a dichotomy is to effectively change the relationship among meanings of health and disease. ²² More centrally for Tavory and Timmermans’s (2013) model of dynamic interpretants, the temporal component of the meaningful mechanism model does not display modularity. We cannot imagine holding constant parents’ continued adherence to the disease frame while varying their initial uptake of this frame. That is, the content of the later disease frame—that their child is still sick—necessarily refers to and is constituted by that which came before.

Tavory and Timmermans’s (2013) model therefore points to a process we indeed wish to recognize as causal: The relationships among frames that are interpreted and acted on determine the space of possible subsequent interpretations and actions. Yet, as the authors themselves grant, this is not a causal model that depicts constituent parts that stand in relations of counterfactual dependence. Dispensing with “a discrete-event view of causality” that operates “like a dominoes game,” Tavory and Timmermans (2013:687) suggest that a pragmatist model of causality “is much more like mixing spices in a recipe where different ingredients partly define the taste.”

Crucially, the causal discrepancy between modular and meaningful mechanism models has methodological consequences for the analysts who use them. In the world of modular mechanisms, the analyst acquires evidence that a particular causal pathway obtains in his or her data by testing whether the observed relationship between cause and effect could be explained by any other alternative pathway. By contrast, accruing evidence for meaningful mechanism models requires precisely the reverse tactic. To demonstrate the plausibility of a meaningful mechanism model, the analyst assesses variance not in the mechanistic pathways but in the outcome to be explained. When the analyst is able to demonstrate that a given meaning structure can account for a variety of different behaviors, confidence in the causal power of that semiotic structure increases. Thus, Norton (2014:173) contends that “codes are a mechanism for theorizing common orientations across multiple situations and thus for explaining how actions across multiple, potentially quite different situations, can be similarly oriented.” Similarly, Tavory and Timmermans (2013:692) explain that in the case of semiotic chains, “the researcher keeps one aspect of the meaning-making process constant to examine how different situations are refracted through—or transformed by—these semiotic aspects.” The robustness of modular mechanism models is assessed in relation to alternative pathways, whereas meaningful models are assessed in relation to alternative outcomes.

Objections: Modularizing Meaning?

Might there be a way to reconcile the modeling of meaning with modular depictions of causal processes? Here, we consider two possible objections to our claim about the incompatibility of modularity and meaning.

First, one might argue that meanings can in fact be modular. For instance, although meaning systems defy modularity as we have shown, certain meaning-action bundles—meanings connected to certain tendencies of action—might, for certain questions, support a modular representation. I may not be able to intervene on “cats signify feminine” without changing the entire set of relationships between animals and gender, but I could perhaps intervene on the specific relationship composed of the meaning of cats as feminine and the propensity of female children to purchase cats as pets. If I believe abortion should be legal, it raises the probability I will vote Democratic; if I am against gun control, it raises the probability I will vote Republican. In such a modular system, it is possible to ask about the causal effect on voting if gun control beliefs change from against to for. That is, meaning and action could be treated as if they were modular for the purpose of analysis.

This approach does have traction in certain sectors of cultural sociology, particularly in work where belief, values, or motivations are thought to cause a particular behavior (Vaisey 2009). This approach to culture also appears in systems where the outcome is not necessarily action but where certain beliefs play a part in the causal web. For instance, Kirk and Papachristos (2011) investigate the causal efficacy of “legal cynicism” on neighborhood-level homicide rates. The implication of the study is not that if one were to intervene on legal cynicism, the web of meanings around law, the police, and trust in legal institutions would retain its relationships, connotations, and denotations. Rather, the point is that if one were to intervene on legal cynicism, the relationships among other elements in the causal system in which those beliefs are embedded could remain. As Kirk and Papachristos (2011:1192) put it: “[C]ynicism exerts an influence on neighborhood rates of violence independent of the structural circumstances that originally produced such cynicism.”

However, this solution brings us to the kind of cultural explanation that meaningful mechanisms were explicitly theorized to overcome: a mechanistic explanation of action based on culture as individually held beliefs. Indeed, this belief-action approach to culture has been roundly and repeatedly criticized in the sociology of culture when it is presented as a general theory of culture and action—it radically disables the understanding of how action is embedded in time, space, and multiple overlapping systems of meaning. Bourdieu’s vast oeuvre is, among many other things, an attempt to show the many limitations of thinking about culture as individual belief (for sociological and anthropological consensus on this point, see Strauss and Quinn 1997; Swartz 1997) and his central concepts of field, capital, and habitus are notorious for their rejection of the kind of independence under intervention that we discussed here under the heading of modularity.

Beginning with his early attack on subjectivism of various kinds in Outline of a Theory of Practice (Bourdieu 1977) and then carried through in critique after critique of “substantialist” approaches to various practices (e.g., sport, educational attainment, art appreciation), Bourdieu’s work articulates how the relationality of meanings upends more mechanistic, means-ends explanations of action. Although not stated in these terms, Bourdieu’s mantra, the “real is relational,” can be understood as a critique of the utility of modular causal pathways for understanding the influence of culture on social action. For Bourdieu, approaches to culture that are “inclined to treat the activities and preferences specific to certain individuals as if they were substantial properties” (Bourdieu 1998:4) are insufficient precisely because they remove from view the meaningful relationships between different practices. Lizardo (2011:36) credits Bourdieu’s cultural model as asserting that cultural objects “do not have any ‘essential’ meaning in themselves, but only in relation to other works.” Thus, wherever one stands on Bourdieu’s repeated contention that there is a homology between social positions and symbolic positions in various fields, it is important to remember that for both Bourdieu and his Saussurean opponents, ²³ the relationality of meaning cannot be removed from the working theoretical model without removing the ability to analyze the power of culture and symbolic distinctions (Alexander and Smith 2006). Any model that seeks to represent culture relationally will necessarily and by design fail modularity.

At this point, a reader might agree that our best models of culture are sociological situations where modularity fails but nonetheless argue that acts of meaning-making could themselves be objects of modular models. This brings us to the second possible objection: Although semiotic nets’ denotative and connotative relations indeed defy modular representation, perhaps the act of meaning-making, taken as an object of study, can be abstracted away from such context. In other words, we could hold on to a relational theory of meaning but not necessarily have this theory feature explicitly in our models, thus fulfilling the modularity requirement and taking on board the relational model. ²⁴

In the abstract, this poses a difficult question: Namely, does the fact that a theory of meaningful action implies the existence of symbolic systems mean it cannot escape relationality (see Sewell 2005)? But when we look in empirical sociology at the kinds of explanations proposed by advocates of meaningful mechanisms, we quickly realize that such an objection is less relevant for research as it erases from the idea of meaningful mechanisms precisely what these explanations hope to achieve. From Gross’s critique of how belief functions in Merton’s self-fulfilling prophecy to Norton’s and Tavory and Timmermans’s appeal to semiotic chains, the explanatory purpose of these models is to illustrate how dependencies among signs and codes shift and re-instantiate as individuals interpret and act on the world around them. The purpose is to model process and relationality (Olick 2013), and thus the failure of modularity cannot be cordoned off without losing precisely what the models are meant to capture.

The Uncomfortable Implications for the Sociology of Culture

So far, we have made two claims: (1) A great deal of sociological work takes mechanism to imply modularity that, although not uncontroversial as an approach to causality, is a reasonably good specification of how many analysts assess independent causal pathways, and (2) modularity is incompatible with what is implicitly or explicitly taken to be an irreducible property of cultural meaning-systems by a wide variety of theories of culture in action. We now consider the implications of possible routes to reconciling this disjuncture. ²⁵

A first possibility is to give up on modeling culture in the supra-individual sense of the term and thereby let go of the relational theory of signs, symbols, and codes needed to model a web of meaning. Instead, we could insist that our models only refer to individual action as a result of beliefs, preferences, and desires held with more or less strength. In these redefinitions, because culture is no longer modeled relationally—but rather in terms of beliefs or values that can be held, not held, or held with variable strength—the possibility that culture, radically redefined, participates in modular causal systems remains open.

The reader knows well the debates that follow such a move. Such an approach to culture has been criticized from many different angles: as “atomistic,” as too dependent on a means-ends model of action, and as insufficiently “thick” or “deep” (Alexander and Smith 2006; Desmond 2014; Geertz 1973). One way to interpret this long lineage of critique is as an extensive argument that the possibility of modularity as a clean way to model culture may only be obtained at the cost of external validity. For the differences that many sociologists want to explain—for example, between different church groups fighting welfare retrenchment (Lichterman 2005) or the ways religion, although understood as a universal human phenomenon, is practiced and lived vastly differently in different historical moments (Riesebrodt 2009)—modular models of culture are unable to account for context or historicity. This is presumably why work on meaningful mechanisms that accommodates a semiotic theory of culture had to arise in the first place. We thus reject this solution, although we have no doubt that some of our colleagues will not only accept it but defend it with agility and aplomb.

A second possibility is to reject the troubling modularity requirement by rejecting its progenitor: the interventionist approach to causality. One might point to the successful explanations arrived at by the meaningful mechanism approach and argue that if an interventionist approach to causality excludes such models, we should discard it as flawed or inappropriate to social science. Instead, we should find an alternative causal account to ground our models, perhaps one arrived at by sociologists “on their own” (e.g., without engagement with philosophers of biology). ²⁶

This solution has its appeal, particularly in its unwillingness to simply model social science on natural science, but we reject it as well. As the volume of empirical work on mechanisms that we discussed earlier illustrates, the specification of causality as relations of counterfactual dependence has already demonstrated its conceptual and empirical effectiveness to sociological research. Moreover, as Hirschman and Reed (2014) point out in a different context, including “formation stories” (which are surely nonmodular in the terms of this article) as part of causal sociology does not require rejecting outright that in certain times and places, social kinds are stable enough that a modular-mechanist account of causation can work empirically so long as scope conditions are rigorously specified. In particular, if such situations can be clearly delimited, then sociology should “make intelligent interventions possible” (Sewell 2005:355). After all, even Sewell admits that the world is not reducible to semiotics.

A third possibility—and the one we pursue here—is to recognize that most uses of the term mechanism in the sociology of culture are metaphorical. The denotations and connotations of the term mechanism in the context of cultural analysis in sociology are such that it provides a useful way to describe certain social processes. Understood in this sense, “mechanism” is primarily a sensitizing concept that admits some degree of generalization and points to causal processes that unfold in a regular manner (Aviles and Reed 2017). Precisely because its utility is metaphorical rather than definitional, the term “mechanism” then enables a very different approach to the epistemology of social causation.

In this view, and accepting mechanism as metaphorical, we would consider modular mechanisms a sufficient but not necessary route to sociological explanation. Meaningful mechanisms built on fundamentally different explanatory principles would also be sufficient and not necessary. Both explanation types, modular and meaningful, would provide information about regularly unfolding causal structures: the former carving the social world at its causal joints, the latter demonstrating the productive capacity of its symbolic, interdependent junctures.

This approach is consistent with its own theory of meaning. That is, it views the meaning of causation that inheres in the concept mechanism to be specified by its relationship to surrounding signs in scientific discourse. However, what might first be taken as a laissez-faire epistemological pluralism (or anarchism, see Feyerabend 1988)—each to his or her own intellectual context—is not implied by this third possibility. Rather, it implies something much more unsettling for sociological theory and research.

By abandoning the epistemological unity of the mechanism label—by allowing different explanatory approaches to use the same term—we abandon any epistemic capital that meaningful mechanisms may gain from their nominal proximity to a modeling strategy with a well-developed statistical apparatus. In other words, having identified a mechanism will not be sufficient in and of itself to answer skeptical demands for causal clarity of how things could have been otherwise. Because production and procession, rather than dependence and intervention, characterize meaningful mechanisms, they will illustrate the constellation of semiotic understandings that make action possible, but they will not—and cannot—answer the causal questions of interventionist dependence that modular mechanism models make explicit.

This recognition—that social mechanism is a useful causal metaphor deeply riven in its approach to causality—means sociologists of culture—and perhaps sociologists more generally—must delineate strict limitations on the use of mechanistic language as a speech genre for their science. Scholars must recognize that social mechanism cannot serve as the default stand-in for explanation. For, if mechanism is primarily an explanatory metaphor characterized by a contested notion of causation, then the question arises: Which other causal metaphors should sociologists of culture draw on to build their explanations? We do not believe this question has received adequate attention in the field.

We argue that accepting the metaphorical approach requires accepting that there is no coherent intellectual reason to consider mechanistic explanations as inherently more scientific or rigorous than sociology of culture explanations that adopt other epistemic strategies or that use other metaphors, particularly those that are similarly productive, processual, and interpretive. Consider, for instance, narrative. Griffin (1993:1097) describes narrative explanation as “analytic constructions (or ‘colligations’) that unify a number of past or contemporaneous actions and happenings, which might otherwise have been viewed as discrete or disparate, into a coherent relational whole that gives meaning to and explains each of its elements, and is, at the same time, constituted by them.” Like meaningful mechanisms, the causal relations that narratives illustrate are based in temporal sequencing, in demonstrating the actual ideas, happenings, networks, and actions whose assemblage forms a historical event (Hirschman and Reed 2014). Also, like meaningful mechanism models, narratives are simultaneously interpretative models: The meaning of their component parts arises in relation to their holistic arrangement. Narrative explanation, however, does not model relations of counterfactual dependence. As Griffin (1993:1101) asserts, narratives—the “artful blend of explanation and interpretation”—fail to meet the requirement of a “causal” explanation precisely because “true” causal explanation “conceptually isolates and abstracts facts from their historical concatenations and asks whether their absence or modification would have altered the course of the event as it was recorded.”

Our point here is not to mount a defense of narrative explanation so much as it is to argue that should we accept meaningful mechanisms as sufficient for explanation, we must likewise accept a causal pluralist position (Cartwright 1999; also see Godfrey-Smith 2009). Epistemic value, and indeed causal explanation, might be achieved by models that do not demonstrate whether their abstracted component parts’ “absence or modification” alters the outcome of interest. Indeed, beyond narrative, we can identify a vast and varied field of nonmodular explanations in sociology: explanations that focus not on abstract interventions but on historical specificity, flow, sequence, and production. The theoretical terms and bounded explanations developed in these accounts will not carve the world into abstract, independent processes linked by counterfactual dependence. Is it possible that in the first flush of mechanisms’ celebrity, we have failed to carefully evaluate the epistemic value of various other routes to sociological explanation?

In this article, we focused on how cultural explanations that model meaning are often nonmodular. In concluding, we want to point toward the possibility that much of the social world and the kinds of models we wish to use to comprehend it may also fail to exhibit modular relations. The relationships among institutional rules and the practices they enable is a good candidate for the kind of “relationality” that disables modularity (see Abend 2018). Consider political questions, for instance, that disentangle how the existence of legislative practices like the Senate filibuster and how particular uses of the filibuster in a given administration determine gridlock. ²⁷ It is sensible to try to disentangle the causal contribution of an institutional rule from the causal contribution of a given use of that rule. However, because rules and the actions they make possible are not only causally dependent but also logically and constitutively dependent, models depicting these systems will not be modular. Future work could further examine what sorts of causal systems exhibit this kind of nonmodularity and therefore what kinds of epistemic commitments we make when referring to causal relationships among such processes.

Conclusion

We see two implications of the argument pursued here for further thinking about sociological concepts. The first is the synthesis of the sociology of sociological knowledge with epistemology. What we have in mind here is a combination of studies of the use of sociological concepts in various settings (e.g., conferences, experimental interactions with subjects, the writing of ethnographic field notes) with theoretical reflection on how to better imagine the basis of the use of these concepts, including thinking about how sociologists should use concepts. The model for this latter aspect—a philosophy of social science that is compatible with a sociology of knowledge—can be found in the extension of the Stanford school of the philosophy of science to the social science of economics, conducted by Nancy Cartwright. Greater engagement with these questions may help develop sociological accounts of mechanistic explanation that are neither modular nor meaningful. For instance, explanations of causal accounts based in production rather than intervention have provided the basis for a causal approach outside of the experimental logic of dependence (Hall 2004).

The second implication is that beyond the question of meaning, careful parsing of what we mean by the term mechanism may have salutary effects for sociological theory more generally. Completely outside of questions of culture, “social mechanisms” are repeatedly offered as a solution to the problem of how to theorize social causality. But here, too, we find a multiplicity of meanings of the term. Despite its ubiquity, the concept of mechanism takes on different meanings when it is used to explain different targets: It reproduces in fractal fashion the same divides (Abbott 2001, 2007). These differences in meaning reproduce disagreements about causation and even reproduce the familiar opposition between intervening variables, on the one hand, and more holistic and relational analysis, on the other.

Our overall message is straightforward, although perhaps ironic. The concept of mechanism was designed to open the black box of causality for sociology (Hedström and Swedberg 1998). But we have, perhaps, black boxed just how mechanisms explain—how they are part of a verifiable explanatory account delivered from one (group of) sociologist(s) to another. This article has been an effort to articulate with greater clarity what we talk about when we talk about mechanisms. By doing so, we hope to avoid falling into that particularly detrimental trap for any epistemic community: sacrificing the construction of useful models of reality to reify the encompassing reality of a single model. Looking closely at mechanism, we can see it is neither panacea nor placebo but rather a sociological tool—one among many, whose value must be approached with care, in the company of other metaphors, pursuing other explanations.