Mechanism-based explanation: a primer

Microfoundationalism – as Wendt and Leon refer to itFootnote ³¹ – has its roots in methodological individualism. This well-known doctrine is, as Lukes puts it, ‘a prescription for explanation’, asserting that all explanations of social phenomena should be ‘couched wholly in terms of facts about individuals’.Footnote ³² Taken to its logical extreme, it seems to demand that all theories of social phenomena be reduced to individualistic ones. This has been found to be highly problematic, however,Footnote ³³ and many if not most individualistically inclined social scientists now accept the more sober but no less interesting view that, as Wendt puts it, we endeavour to ‘identify the micro-level mechanisms by which macro-structures achieve their effects’.Footnote ³⁴ Debate now turns largely on the issue of whether mechanisms and microfoundations can be provided for structural phenomena.Footnote ³⁵

Mechanism-based explanation is the means by which microfoundations are provided, and is the core of the microfoundational approach. Also referred to as ‘mechanistic’ or ‘reductive’ explanation, mechanism-based explanation has become something of a hot topic in biology, philosophy of science, and social science, thanks largely to the spread of the mechanistic worldview, mentioned earlier. According to this view, causes are mechanisms, and explanation is a matter of uncovering them. The most elaborate accounts of mechanism-based explanation are to be found in philosophy, especially the philosophy of biology and the philosophy of mind. I draw on each in the following account.

Mechanism-based explanations come in two main forms: causal and non-causal.Footnote ³⁶ Causal mechanistic explanations, like causal explanations in general, involve specifying the prior conditions responsible for an effect. They show how effects are brought about; they explain changes. The difference between this and more traditional models of causal explanation like deductive-nomological explanation is that the mechanistic approach incorporates a mechanistic notion of causation and causal relevance, which requires looking more deeply into the nature of posited causes (among other things).

More relevant to us, however, is the non-causal form of mechanistic explanation. This is the kind most philosophical discussions focus on. (In the remainder, I refer to it simply as mechanism-based explanation.) This kind of explanation is generally understood as a process of specifying the mechanisms responsible for causal work associated with higher level phenomena. It has two especially important features. The first is that it is inter-level, or as some put it, ‘componential’. It involves looking inside systems, to their parts, for explanatory mechanisms. As Elster puts it, to explain via mechanisms is to ‘open up the black box and show the nuts and bolts, the cogs and wheels of the internal machinery’.Footnote ³⁷ Or as Kim puts it, we look to ‘explanatory resources at a different level, a “lower” level in relation to the level of the phenomenon to be explained’.Footnote ³⁸ Its other notable feature is that it is non-causal. In a causal explanation (mechanistic or otherwise), the aim is, as Cummins puts it, to ‘explain changes of state in a system as effects of previous causes’.Footnote ³⁹ In a mechanism-based explanation, however, the phenomenon to be explained is the behaviour of a macro-level system. Mechanism-based explanation involves identifying the micro-level conditions that make it possible for a structure to play a certain role. Such explanations are answers to ‘how-possible’ questions of the form: ‘How is it possible that structural factor x plays role y?’

In order to see exactly why this is not causal, we need to know more about the terms ‘role’ and ‘mechanism’ here. By role I mean disposition. Mechanism-based explanations usually target general dispositions of systems, rather than isolated behaviours. A disposition is a causal tendency – something an object tends to do, in certain conditions. Dispositions are often referred to as dispositional properties or ‘powers’. Classic examples are solubility and flexibility, and mental properties like pain and hunger are also defined this way. Dispositional properties have the special characteristic of being realised by properties at lower levels of analysis. The latter are often referred to as realisation ‘bases’. Pain for example, is something like the property of tending to wince and retract, in response to tissue damage, and it is realised by physical-chemical states of the brain.Footnote ⁴⁰ The main aim of a mechanism-based explanation is to show how a dispositional property is realised at the micro-level, or identify its realisation base. Realisation-bases are mechanisms. The literature contains a large number of definitions of mechanism, but there are important points of agreement,Footnote ⁴¹ including that mechanisms are realisers.Footnote ⁴² Craver's definition is a good starting point: a mechanism, he says, is ‘a set of entities and activities organised such that they exhibit the phenomenon to be explained’.Footnote ⁴³ The ‘phenomenon to be explained’ is a causal role played by a higher level property; the mechanism is what realises this property and its role. Realisation relations are not causal – realisation-bases and dispositions instantiate instantaneously.Footnote ⁴⁴ Pain, for example, is a brain-state usually caused by tissue damage, but realised by a person's nervous system. A causal explanation of a person's pain would cite the events that preceded and led to the tissue-damage, and the tissue-damage that in turn led to the pain. A mechanism-based explanation, in contrast, would talk about the state of the person's nervous system at the time of the pain.

The examples I cited in the previous section were all at least partial examples of explanation of this kind. Alliance polarisation's war-spreading role, for example, was explained in terms of unit-level conditions – alliance linkages. Polarity's stability-influencing role was explained with reference to conditions of uncertainty and utility calculations of states. In these and the other cases, the object of explanation is a causal role played by a system property, and the entities responsible constitute part of their realisation bases. Consider the explanation of polarisation. Alliance linkages do not cause alliance polarisation to spread wars. They constitute alliance polarisation, and actually fulfil this role themselves. This is an explanation of alliance polarisation's causal capacity, which is a property it possesses, rather than an effect or outcome.Footnote ⁴⁵

Perhaps surprisingly, this sort of explanation is not commonly found in reductionist IR research. Reductionist theories are generally seen as those that explain international outcomes in terms of micro-level causes, like the attributes and interactions of states, or the attributes and interactions of decision-makers within states.Footnote ⁴⁶ Some go a step further and try to explain the emergence of institutions or regimes from the interactions of actors, as noted above. These sorts of theories are reductive in the traditional sense of exhibiting a micro-level explanatory orientation – they invoke micro-level causes for macro-level effects. But they do not, for the most part, provide many examples of mechanism-based explanation proper, because they are primarily causal. They treat institutions as outcomes, and explain them by describing the processes of interaction that cause them to emerge. They are, as Cummins would say, transition theories – they explain the system's transition from lacking a certain institution, to gaining it; or from having a regime of one kind, to having one of another kind.Footnote ⁴⁷ Still, this sort of work may be regarded as being microfoundational in a less direct sense. These theories suggest that lower levels of analysis are where most of the real causal work gets done in world politics. This, in turn, seems to imply that how things are at lower levels determines how things are at higher levels. If systemic factors appear to being doing any causal work, then it is probable that this is really being fulfilled at lower levels. What are lacking in these theories, however, are explanations linking system-level causal roles and micro-level mechanisms directly.

The distinction is crucial, because each sort of explanation provides us with a fundamentally different kind of explanatory information. Causal explanations tell us how outcomes or changes are brought about; mechanism-based explanations tell us how the causes cited in causal explanations actually do the causal work assigned to them. Causal explanations attribute causal roles to things, in order to explain why outcomes or changes occur. In effect, they attribute dispositions to them. Those that explain the spread of war in certain periods with reference to alliance polarisation attribute it, in effect, a war-spreading disposition. This often prompts an additional question, as to how it is possible that the causal entity or property in question is able to do the causal work assigned to it. In virtue of what does alliance polarisation spread wars? How is it able to fulfil this role? The answer is not some prior event, but mechanisms found in its own realisation-base.

Why is it that mechanism-based explanations clarify macro-level explanations, and not the other way round? This has the appearance of being arbitrary. But from the point of view of microfoundationalism, it is quite reasonable. This ‘downward’ orientation is motivated primarily by the idea that lower levels of organisation are more causally fundamental than higher ones. All things have parts, and it is natural to think that these parts are a necessary condition for their existence and for any causal work they do. Philosophers often articulate this point with reference to ideas like physicalism, ‘supervenience’, and ontological individualism. Physicalism is the view that there is only one substance in the universe – matter – and all things and processes are essentially physical. Ontological individualism is physicalism's sociological cousin, and says that there is nothing more to social life than individual-level stuff – individuals, their properties, interactions, and so on. Both are often stated in terms of supervenience. Supervenience is a dependence relation between properties of objects – property a supervenes on property b if a could not differ without b also differing.Footnote ⁴⁸ The mind supervenes on the brain, in the sense that people in identical brain states will be in identical mental states. Society supervenes upon individuals in the sense that, as Sawyer puts it, ‘if a collection of individual properties with a given set of relations causes a certain social property to obtain on one occasion, then that same collection of individual properties in that same set of relations on another occasion will cause the same social property to obtain’.Footnote ⁴⁹ Physicalism and supervenience are both popular theses in philosophy,Footnote ⁵⁰ and ontological individualism is also well-regarded, especially when stated as a supervenience thesis.Footnote ⁵¹ Together they have the implication that how things are at lower levels determines how things are at higher levels. The individual level carries, or does the causal work of, the macro-level.Footnote ⁵² This is why, at least for the microfoundationalist, it is rational and not at all arbitrary that we should look to lower levels for clarification and explanation of macro-level phenomena, and not the other way around. These ideas are not without their detractors, however, and I consider one prominent criticism – emergence – below.

A three-step strategy

Thus far I have described what mechanism-based explanations are and tend to look like. I have not yet said how they are built. A number of elaborate strategies for producing mechanism-based explanations are to be found in the literature. Some particularly interesting and useful ones are Kauffman's ‘articulation of parts explanation’, Cummins's ‘functional analysis’, Bechtel and Richardson's ‘decomposition and localization’ heuristic, and Kim's ‘functional reduction’.Footnote ⁵³ Each involves at least three steps: (i) functional analysis, (ii) mechanism search; and (iii) reductive explanation. The first step, functional analysis, involves identifying the causal role of the system or structure to-be-explained. This is to be treated as the explanandum-phenomenon. The second step involves a search for the mechanism or mechanisms that realise the causal role in question. One looks amongst the system's components, their properties, interactions, and organisation, with the aim of finding conditions sufficient to realise the causal role in question. The final step involves producing the explanation. One constructs a general theory or model to explain the correlation between the realising mechanisms and the realised causal role. The theory or model could be constructed within the apparatus of a relevant existing micro-level theory. The second step – mechanism search – is the most important, however, and the one most accounts focus on. Bechtel and Richardson, in their now-classic work on the matter, propose a strategy of decomposition and localisation, or ‘hierarchical analysis into functional components with specific functions’.Footnote ⁵⁴ Decomposition involves subdividing the overall function or causal role into a series of sub-tasks or ‘set of subordinate functions’ that together comprise ‘the behaviour we are studying’.Footnote ⁵⁵ Localisation involves associating each sub-task with a specific component of the system.Footnote ⁵⁶

This strategy is especially useful in cases of multiple realisability, as I will argue below. Strategies like it have been discussed in IR theory before. A good example is Most and Starr's ‘nice law’ strategy.Footnote ⁵⁷ They premise this strategy on the observation that certain consistent foreign policy outcomes can have ‘alternative triggers’ at the decision-making level. Self-help behaviour, for example, can be pursued on the basis of a variety of goals, even while external constraints like anarchy and the distribution of power remain constant. Thus, system-level generalisations about self-help behaviour (like Waltz's) capture necessary, but not sufficient conditions for self-help behavioural tendencies.Footnote ⁵⁸ Most and Starr argue that in order to account for alternative triggers, we break down system-level generalisations into sets of nice, ‘sometimes true’, or ‘domain-specific’ generalisations, each of which identifies a different kind of motivation for the behaviour in question. This can be interpreted as involving decomposition and localisation. In their revision of the neorealist self-help generalisation, for example, Most and Starr decompose system structure's role of inducing self-help behaviour into a set of specific sub-tasks (inducing militaristic self-help behaviour, inducing domestic-political self-help behaviour, inducing economic self-help behaviour, etc.), and localise each function to particular sorts of goals (security goals, political goals, economic goals, etc.).Footnote ⁵⁹

Mechanism-based explanation and micro-macro problems

As noted above, the primary function of a mechanism-based explanation is to account for a structural factor's causal powers. In this way mechanism-based explanations add depth to causal explanations.Footnote ⁶⁰ Another way of putting it is that mechanism-based explanations are ‘mystery-removing’ explanations.Footnote ⁶¹ Causal explanations remove mysteries of their own, concerning why things happen. But they inadvertently create new mysteries as well, concerning how causal factors are able to do their causal work in the first place. Mechanism-based explanations address the latter sort of mystery. They do this by providing microfoundations – a system's microfoundations being its realising mechanisms. This is particularly attractive from the point of view of scientific realism, for which explanation is about getting to the underlying essences of things, to fully account for their causal powers.Footnote ⁶²

Because it involves linking higher- and lower-level phenomena, some philosophers also see mechanism-based explanation as a means of resolving micro-macro problems. Some argue it might be able to resolve the mind-body problem in philosophy,Footnote ⁶³ and others have suggested it could do the same for the holism-individualism debate in social science.Footnote ⁶⁴ In my view – and whatever its results in these areas – the microfoundational strategy might help resolve IR theory's own micro-macro problems. Two problems of this kind have been discussed at length in IR theory: the level-of-analysis problem and the agent-structure problem. The former, it is said, concerns the relative explanatory value of different levels of analysis in world politics – levels usually being seen as a system for organising variables, based mainly on their scale.Footnote ⁶⁵ The microfoundational strategy appears to favour lower levels, because of its suggestion that lower levels are more causally fundamental. But the strategy is actually as conducive to macro-level analysis as it is to micro-level analysis. By removing mystery from structure – a macro-level entity – mechanism-based explanations should enhance its credibility as an explanatory tool. Indeed, this seems to be the intention behind the mechanism-based explaining that already goes on in systemic IR theory. Why would microfoundations enhance structure's credibility? For at least two reasons. First, because lower levels of organisation are more causally fundamental, at least according to microfoundationalism. Going down in level gets you closer to physical mechanisms. It helps us integrate higher level processes into an underlying ontology. The second reason is that microfoundations help bring structural theories into closer alignment with theories at lower levels. A theory's consistency with the rest of science, especially with the best-established sciences, is an important measure of its credibility.Footnote ⁶⁶ Thus, in my view, and counter to initial appearances, mechanism-based explanation should be most attractive to systemic theorists of IR.

Microfoundations are, of course, good for reductionists too. By showing that higher level processes can be accounted for with lower level explanatory resources, more explananda-phenomena are brought within the explanatory grasp of reductionist theory, and its ‘leverage’ is increased.Footnote ⁶⁷ Mechanism-based explanations also reveal as much about micro-level mechanisms as they do structural phenomena. Primarily, they tell us more about what lower level mechanisms do. Without explanations linking mechanisms with higher level roles, we would only know about the roles mechanisms play at their own level. Structural theories, via mechanism-based explanations, help microfoundationalists see the bigger picture – see the functional forest from the mechanical trees, so to speak. The explanation of alliance polarisation in terms of alliance commitments, for example, helps us see that alliances can spread wars throughout systems, not just within dyads – despite the fact that they are normally intended to mitigate war. And Waltz's explanation of anarchy's selection-effects in terms of competition and socialisation helps us see that the latter two mechanisms play a role in generating system-wide patterns of balancing behaviour.

Microfoundationalism's solution to the level-of-analysis problem is, then, relatively pluralistic – no particular level gets priority. In itself this is not a novel solution, nor is it likely to be met with much opposition, for most that have written on the level-of-analysis problem already recognise that no single level – micro or macro – is capable of doing all the explanatory work required of IR theory.Footnote ⁶⁸ Microfoundationalism is not so neutral when it comes to the agent-structure problem, however. This problem, it is said, concerns how we conceptualise the relation between agents and structures in world politics.Footnote ⁶⁹ Mechanistic explanation is all about capturing micro-macro relations like this. More specifically, it is designed to capture relations between macro-level phenomena and their realising mechanisms. Where successful, mechanism-based explanations show that structural phenomena are wholly dependent for their causal powers on agent-level mechanisms. Microfoundationalism's proposed solution is that we close the gap between agency and structure, as it were, by means of mechanism-based explanations. The agent-structure relation, on this view, is one of asymmetric dependence – structures depend on agents, at least for their causal potentialities.

This contrasts sharply with the way most writing on the agent-structure problem have depicted the agent-structure relation. Most argue that neither structure nor agency is primary, and that each plays a part in determining the other, in a kind of recursive interplay.Footnote ⁷⁰ These authors are, as it is sometimes put, dualists – other differences in their views notwithstanding. Some adopt the criticisms I review below to undermine the microfoundational view and support their dualism. The success of mechanism-based explanation as a solution to the agent-structure problem turns in large part on these criticisms failing, a matter to which I now turn.

Obstacle 1: structure and context

The first and most apparent obstacle is structure itself. Microfoundationalism is a reductive research strategy. In IR theory, reductionism is usually associated with a domestic, ‘inside-out’ explanatory orientation on world politics. This is evident in meta-theoretical discussions of reductionist IR theory and of the level-of-analysis problem.Footnote ⁷¹ Since mechanism-based explanation involves using the explanatory resources of lower levels of analysis, it would seem to demand a domestic, inside-out view of world politics as well. To explain system-level phenomena, microfoundationalism would have us look to the attributes and interactions of states, or perhaps even further inside states, to bureaucratic structures or decision-makers. Microfoundationalism, then, seems likely to marginalise structure. This would be a problem. While systemic views of world politics are not especially popular, most concede that structure plays some role in world politics. As noted above, most that have written on the level-of-analysis problem recognise the explanatory value of multiple levels. As Levy and Thompson observe, most theories of IR incorporate variables from multiple levels of analysis.Footnote ⁷² Even those that declare a preference for lower levels of explanation, like liberal and decision-making theorists of foreign policy, still maintain that the systemic environment plays some kind of role – usually that of a distal constraint on choice or opportunity.Footnote ⁷³ Theories now generally transcend particular levels.Footnote ⁷⁴ This has lead some working on the agent-structure problem to declare that structure's role in political life is ‘uncontentious’Footnote ⁷⁵ and even a ‘truism’,Footnote ⁷⁶ and dismiss individualistic views like microfoundationalism offhand.Footnote ⁷⁷ Waltz and Singer have dismissed reductionism on similar grounds.Footnote ⁷⁸

That the systemic environment plays a causal role in the decision-making of states and leaders seems undeniable. How, then, could the microfoundational strategy possibly work? Fortunately, the microfoundational strategy is not restricted to a purely domestic orientation. The aim with this strategy, I have been stressing, is to account for the behaviours or dispositions of systems with reference to their components and properties of their components. Context itself is the explanandum-phenomenon. Furthermore, context is part of the explanans as well. Consider the examples of section one again. Alliance polarisation's war-spreading role was explained not in terms of any domestic factor, but alliances between states. Polarity's stability-influencing roles were explained partly with reference to the dispositions of states, but also with reference to environmental uncertainty. These explanations refer us to the microfoundations of context itself, to explain how its influence is realised. They exhibit a decomposition and localisation strategy. Some of these microfoundations (for example, attributes, dispositions) are domestic in nature, but some (for example, relations, conditions) are not. The latter are contextual in nature, but this does not mean they are automatically systemic. Context is not essentially a macro-social thing. Like all substantial things, context is built out of parts, and we can break it down into parts of virtually any scale. The microfoundational strategy involves doing exactly this.

Microfoundationalism is also conducive to the use of large-scale combinations of micro-level variables. Given its large scale, the systemic context probably does most of its causal work in large spatial and temporal sweeps, producing system-wide and period-spanning outcome patterns. Rather than particular foreign policy choices (that is, particular wars or cooperative ventures) it more likely affects things like overall system stability.Footnote ⁷⁹ Mechanism-based explanations of systemic tendencies will not get far if the focus is just on isolated, small-scale, domestic variables.Footnote ⁸⁰ These rarely suffice to realise whole systemic dispositions.Footnote ⁸¹ Mechanism-based explanation would instead require showing how the parts (in plural) of the system interact, work together, and are organised, so as to realise its causal roles. Many of these parts might be, individually speaking, domestic in nature. But clustered together, they turn into large scale variables, and ones that are contextual, relative to individual states.

One might wonder why, if mechanism-based explanations can reference large-scale factors like context and structure, they can be considered ‘reductive’ or ‘microfoundational’. There are actually good reasons for thinking so, and they can be brought out by looking at precedents in social science and philosophy. Consider the mind/body problem in philosophy. This problem concerns whether mental states reduce to physical-chemical states of the brain, and pits reductionists against ‘emergentists’. Reductionists do not say that mental states reduce to states of particular particles or microscopic regions of the nervous system, but to states of the whole nervous system – that is, its components in aggregate. This is a genuinely reductionist point of view, as Kim argues, because they are saying mental states are not emergent, but reduce to ‘micro-based’ properties – clusters of lower-level properties.Footnote ⁸²

Consider also the holism-individualism debate. Individualism does not demand omission of all reference to context in the explanation of social outcomes. The idea that all social outcomes can be explained with reference to factors internal to individuals, like psychological attributes, is generally regarded as utterly implausible. Individualism does not say that context has no explanatory role, just that macro-social things have no explanatory role. And context is not essentially a macro-social thing, as we have seen. Features of context that are individualistic in scale – like other agents, relations, actions, interactions, accessible physical objects, resources, and so on – can and do feature in individualistic explanations, and standard definitions of methodological individualism attest to this.Footnote ⁸³ Individualists even use variables traditionally associated with holism, like structure and culture, but define them aggregatively. The distinction here is essentially the same as Wendt's distinction between cultures of common knowledge and cultures of collective knowledge.Footnote ⁸⁴ Common knowledge, Wendt says, is merely a common set of beliefs held by individuals, and is the kind of culture microfoundationalists recognise. Neoliberals and many constructivists conceive of culture in this way. Singer defined culture in a very similar way, asserting that ‘the cultural properties of any subnational, national, or extranational system may be described in a strictly aggregative fashion, by observing the distribution and configuration of individual psychological properties’.Footnote ⁸⁵ True holists, on the other hand, view cultures as structures of collective knowledge, which while constituted by common beliefs, do not to reduce to them – at least on the holist account.Footnote ⁸⁶

Overall, two things make mechanism-based explanations of structure and culture reductive. The first is the aggregative view of structure implied. The second is the fact that, in the final analysis, the only mechanisms attributed basic causal power are micro-level ones. In mechanism-based explanations, the parts (and their properties) are doing all the real causal work. In genuinely structural explanations, on the other hand, like those Waltz and Wendt offer, structures are attributed irreducible causal or explanatory power of their own, and depicted in a non-aggregative (that is, emergent) way. The implication is that large-scale contextual variables like culture are not off-limits for the microfoundational approach, as long as the emphasis is always on the parts. This, of course, is generally recognised already. Waltz, for example, says that interactions – which are technically outside agents – are unit level, and along with attributes are part of the domain of reductionist theory.Footnote ⁸⁷ Dyad-level and region-level theories of war treat smaller segments of the systemic environment (that is, dyads and regions) as explanatory levels distinct from the system level. There is also recognition of the fact that using large-scale combinations of variables can be reductive. Waltz himself says that reductionist theories ‘explain international outcomes through elements and combinations of elements located at national or subnational levels’.Footnote ⁸⁸ Wendt also recognises it, with his distinction between common and collective knowledge.Footnote ⁸⁹

Structure, then, is not in itself a problem for mechanism-based explanation. Structure is the problem mechanism-based explanation was designed to deal with in the first place. Rather than marginalising structure, as reductionist IR theories sometimes do, mechanism-based explanations would actually bring it back into the fold, albeit in a component-emphasising way.

Obstacle 2: multiple realisability

The real obstacles to microfoundationalism are the next two: multiple realisability and emergence. Each has the implication that structures cannot be explained in terms of realising mechanisms, something that, if true, would scuttle any reductive research strategy. I deal with multiple realisability first. This influential notion is largely responsible for the major backlash against reductionism in recent philosophy, and its effects have been felt all across the sciences.Footnote ⁹⁰ It was only ever a matter of time before it was brought into IR theory. Wendt was the theorist responsible, though he acknowledges that others discussed similar ideas in the past. In Social Theory, Wendt argues that system and state-level factors are multiply realisable at lower levels of analysis.Footnote ⁹¹ He acknowledges, in line with the philosophical argument, the ontologically reductive principle of supervenience, at least for the sake of this argument. As noted above, supervenience suggests that micro-level conditions determine macro-level conditions. But importantly, as Wendt notes, the ‘upward determination … is not strict or 1:1, since it allows for the possibility that the same macro state could be realised by many different micro states’ – multiple realisability.Footnote ⁹² This, he argues, means micro-level states of affairs will be ‘sufficient for the existence of a macro-state, but not necessary’.Footnote ⁹³ Because of this, micro-level inquiry ‘may generate disparate explanations for events that in fact have a common macro-level cause’.Footnote ⁹⁴ He elaborates with the example of a window that breaks when a person (John) throws a rock at it.Footnote ⁹⁵ Wendt contends that the best explanation for the broken window is that ‘John threw a rock at it’, rather than ‘an analysis of the particular combination of sub-atomic particles’ that broke it. While the exact particle-structure of the rock was causally responsible for the breaking of the glass, ‘many other combinations would have had the same effect’. In other words, the macro-cause (John throwing a rock at the window) was multiply realisable at the particle-level. By extension, the macro-effect (the window being broken) could have been brought about in many ways at the particle-level. The only factor stable enough to explain the outcome is, therefore, the macro-cause. Wendt argues that the same is true of systemic tendencies in IR, and concludes that ‘the door is open to relatively autonomous macro-level explanations’.Footnote ⁹⁶

Wendt likens his claim to ones made earlier by Waltz and Most and Starr. Waltz argued that unit-level mechanisms vary more than the international outcomes they are supposed to explain.Footnote ⁹⁷ He concluded, much like Wendt does, that these causes cannot explain these outcomes in an ‘elegant’ or general way. Instead, systemic variables are needed, because of their relative invariability. Most and Starr made a similar observation to Waltz, but drew a rather different, reductionist conclusion, which I consider shortly. Unfortunately, however, Wendt is not clear what he means by multiple realisability, and it is not immediately clear what the relation between his argument and those of the others is. Sometimes he suggests that multiple realisability is like what Waltz and Most and Starr observe – the capacity of a given macro-level outcome for different micro-level causes – as when he discusses his three cultures of anarchy.Footnote ⁹⁸ Sometimes he suggests it is the capacity of a macro-level entity for different constitutions or arrangements of parts, like when he talks about supervenience.Footnote ⁹⁹ At other points he suggests that it is something like the capacity of a macro-level causal tendency for different micro-level fulfillers, like when he talks about the Waltzian tendency toward ‘balancing and institutional isomorphism’.Footnote ¹⁰⁰

The correct meaning is actually the third, which is unfortunately the one that Wendt pursues the least. Multiple realisability is the capacity of a dispositional property for different micro-level realisers. To realise, as we saw above, means to occupy, fulfil, or implement a causal role.Footnote ¹⁰¹ Multiple realisability, in turn, is the capacity of a given causal role for different occupants – same function, different implementing mechanisms.Footnote ¹⁰² Philosophers illustrate the idea with simple examples like mousetraps. Mousetraps function to catch mice, but the exact mechanisms they use differ greatly. Traditional ones use a spring-loaded bar with a bait-laden trip for release. So-called ‘mouth-type’ mousetraps use plastic jaws and a coil-spring triggered by a bait-laden mechanism inside. Some use electricity to shock mice. Some trap them in glue. And so on. The function that defines the kind, mousetrap, can be realised in very different ways. Multiple realisability is, therefore, neither a causal nor a constitutive relation. Wendt's broken window was caused by the prior event of John throwing a rock at it; the rock's window-breaking capacity was realised by its physical substructure. The potential of a single outcome for different micro-level causes – which is what Waltz and Most and Starr observe – would be better described as something like ‘multiple causability’.

When understood in this way, multiple realisability becomes quite an intuitive thesis. Whether we are talking about machines, bodily organs, computer programmes, or psychological states – indeed, anything that can be defined in terms of its causal role – it is natural to think that functions can be realised in different ways. Could the multiple realisability argument-proper be applied to IR theory? I suspect it could, but only if anti-reductionists like Wendt and Waltz are willing to conceptualise structural properties as dispositional properties, realised at the micro-level. Hilary Putnam, with whom the argument originated, used it to show that mental states like pain and hunger are functional states – dispositions of a whole organism, defined in terms of their causal roles, and realised at the physical level. The argument all but entails a functional view of higher order properties like this, something Wendt misses entirely. This is a problem, because Wendt also argues that structures are ontologically emergent from their microfoundations, and this clashes with the physical realisation thesis.

However, even if the argument were successfully applied in IR, there are ways around it. One is contained within the mechanistic approach itself: localisation. In particular, Most and Starr's nice-law strategy, being an example of a localisation strategy, is able to deal with multiple realisability. As noted earlier, Most and Starr argue that, when faced with alternative triggers (that is, multiple causability), the IR theorist should retreat from always-true, system-level generalisations to nice-laws, or domain specific generalisations. From a Waltzian perspective, this seems illogical, because it would mean throwing away perfectly good macro-generalisations. However, in the following passage, Most and Starr make a very interesting point.

The strategy Most and Starr seem to propose is this: when faced with multiple causability (different micro-level causes for a single kind of outcome), we should analyse the outcome with a finer grain of analysis, to determine whether it varies in ways that correspond to the variations in causes. This is essentially what localisation is, and it is perfect for dealing with multiple realisability.Footnote ¹⁰⁴ If we try to reductively explain a multiply realised macro-level causal role, we will find multiple mechanisms sufficient to realise that role. But realisers usually differ according to context. Multiple realisation is generally found in dispositional properties that instantiate in different species or kinds of systems.Footnote ¹⁰⁵ Pain and hunger, mental states that instantiate in everything from humans to octopi, are the most commonly cited examples. Localisation involves: (i) first splitting up a multiply realisable macro-property into subtypes according to each context or species of instantiation, (ii) then linking each subtype to a realiser. This leaves us with a set of context-specific, localised mechanistic explanations, to replace the single macro-level generalisation we started with. If multiple realisation is context-dependent, localisation should be successful.

As an example, consider Wendt's ‘Lockean’ culture of anarchy, one of the three varieties of collective order he identifies in Social Theory, and the one which seems to most resemble the current world system. Wendt argues that the Lockean culture generates behavioural tendencies like balancing behaviour, respect for sovereignty, less concern for relative gains, constrained warfare, and a low death rate amongst states.Footnote ¹⁰⁶ He concedes that this culture's behaviour- and interaction-shaping role is realised at the micro-level by a common belief amongst states that they are one another's rivals. But no particular arrangement of beliefs is necessary for a Lockean culture or the tendencies mentioned. It does not really matter if some of the states are wrong in their beliefs, or believe something else entirely. It also does not matter exactly what states or people inhabit the system, or even whether there is a full turnover in membership.Footnote ¹⁰⁷ This culture will still programme for the tendencies with which it is associated. However, if the common belief set underlying this culture ever has varied in a notable, consequential sense, then this would have made some observable difference to the mentioned tendencies. For example, the exact distribution and strength of states’ rivalry-beliefs has probably changed over time. This, in turn, will have made differences to the exact prevalence of the abovementioned tendencies themselves at different times – like slightly more or less balancing behaviour, war, dissolutions of states, and so on. It may even have meant that certain of these tendencies were not present. Accordingly, we would find that these tendencies – which constitute Lockean culture's causal role – have varied in a way that corresponds with differences in the common belief set underlying this culture. What we are looking at, then, are different realisation-bases for different kinds (or grades) of Lockean culture, not multiple realisation. This is where localisation comes in. Localisation would involve: (i) splitting up (that is, decomposing) the tendency-set into a set of discrete tendency-subsets, along the lines suggested (that is, according to their prevalence or the absence of some), then (ii) localising each to the particular historical period with which it corresponds. Reductive explanation is then a real possibility.

Whether localisation will ultimately work is a matter for further investigation. Nonetheless, I think the above example highlights a basic problem with the multiple realisability argument. Philosophers of multiple realisability typically assume that micro-level properties are differentiated in a more ‘fine-grained’, variation-sensitive way than macro-level properties.Footnote ¹⁰⁸ In my view, Wendt and Waltz do the same. But when the same, fine grain of analysis is applied at both levels – as in the example above – appearances of multiple realisability tend to wash away, clearing the way for mechanism-based explanation.

Obstacle 3: emergence

Another important obstacle to the microfoundational strategy is what is often referred to as emergence. A number of IR theorists argue that structures are emergent from their agent-level foundations, in the sense of being more than mere ‘sums of their parts’.Footnote ¹⁰⁹ If true, the reductive, agent-based ontology implied in microfoundationalism is fundamentally incomplete. Wendt is at the forefront of this attack as well, and offers several arguments. In his earliest contribution on the matter, he argued that structures are sets of ‘internal relations’, which constitute properties of their relata (states).Footnote ¹¹⁰ Later, in Social Theory, Wendt argues that structures are institutions of collective knowledge instead, but which also stand in a constitutive relation with agents, and determine certain of their properties.Footnote ¹¹¹ In more recent work, Wendt argues that structures are holographic projections of individual consciousness, which stand in relations of quantum ‘entanglement’ with agents.Footnote ¹¹² All these arguments concern relations of some kind, assert that structures constitute important properties of agents, and suggest that structures do not reduce to their microfoundations.

This line of argument is much more ambitious than that of multiple realisability. The latter is normally coupled with a reductive view of ontology. Most of its philosophical proponents agree with principles like physicalism and supervenience. Their view is that there is only physical stuff in the world, but we cannot explain everything in physical terms because of multiple realisability. The implication is that structures are irreducible in an epistemological, but not ontological, sense. Some refer to this as ‘epistemological emergence’, to distinguish it from the ‘ontological emergence’ referred to in arguments like Wendt's.Footnote ¹¹³

There are, I believe, a number of problems with Wendt's arguments, but I do not have the space to provide a full critique here. Instead, I want to focus on a particular way of dealing with emergence, in which mechanism-based explanations play a central role. Wendt's arguments rely on a set of relatively abstract examples, in which he tries to demonstrate two main things: (i) that agents depend, for certain of their properties, on the existence of structures; and (ii) that structures, for their part, do not entirely depend on agents for their own existence. His arguments, like many others against ontological individualism in social science, suggest that structures do not wholly supervene, or depend, on agents.Footnote ¹¹⁴ There is an epistemological void in these arguments, however. Wendt and the other authors mentioned do not tell us how ontological emergence is to be proven. Asserting that structures do not supervene or depend on agents is one thing, proving it in practice is another. What sorts of empirical indicators should we be looking for to prove or disprove emergence? In my view, the indicators need to be causal. To prove that structural properties are more than just collections of micro-level properties, it needs to be shown that they play causal roles distinct from those of their realisation bases. In other words, it needs to be shown that structures’ parts are insufficient to realise their causal roles. To disprove supervenience specifically, emergentists would have to show that structural factors play causal roles that can differ even while their realisation bases remain unchanged. I have adapted this criterion from what is known as the ‘causal theory of properties’ in metaphysics. This theory states that if two properties bestow the same causal powers on things, we should infer that they are really the same property.Footnote ¹¹⁵ Applied to the problem at hand, it suggests that we can only infer a structure is more than its microfoundations if it plays a causal role that distinguishes it accordingly. If not, we should infer that they are really one and the same, and there is no ontological emergence. Why? We are only able to perceive and come to know of things and properties by the way they behave and impose themselves on our senses. Thus, as Armstrong argues, if ‘things act differently in the same circumstances, then we attribute to them a different nature’.Footnote ¹¹⁶ We really have nothing to go by other than causal role, when we are talking about perceptually remote things like social structures. Thus, it would be unreasonable to hold that a structural factor was distinct from its microfoundations if both had the same detectable causal role in all like-circumstances. It would also be unreasonable to hold that a structural factor was nothing more than its microfoundations, if the two played different roles in identical circumstances.

This is essentially inference to the best explanation. Wendt actually uses inference to the best explanation to justify a very similar principle. He argues that we can infer that theories of unobservable objects refer to real things and processes from the fact that they successfully predict observable behaviour.Footnote ¹¹⁷ The best explanation for why a given theory predicts observable behaviour is that the theory is true and refers to something real. Wight suggests essentially the same thing when he asserts, quoting Bhaskar, that structures’ ‘causal power establishes their reality’.Footnote ¹¹⁸

To establish that structures are emergent from their microfoundations, I believe they need to be shown to satisfy this criterion – that is, shown to play distinct causal-explanatory roles. How can this be established? Certainly not a priori, as Wendt, Wight, and others seem to be doing. It can only be established in practice, by the failure of mechanism-based explanations. If structures really are ontologically emergent, then we will not be able to explain them in terms of their component mechanisms, and microfoundationalism will fail. Only once this strategy has failed in practice will we be able to confidently dismiss its ontology. As it stands, I do not think that systemic research thus far has established that ontological individualism is wrong. Most structural factors with established or hypothesised explanatory roles, like polarity, polarisation, and norms, are already defined in unit-level terms – that is, as collections of unit-level factors. It does not seem necessary that, for any of these factors to play the roles they have been attributed, they need to be anything ‘more’ than their microfoundations. In any case, these are only things that can be established by attempting mechanism-based explanations. Microfoundationalism may ultimately fail, but in the meantime, there are plenty of good reasons to search for mechanisms.