WHY DOES CAUSAL REASONING WORK?

Alongside the standard theories of causation in the philosophical literature, there has been a great deal of related work in statistics, econometrics, and computer science, among other disciplines. This work has flourished over the past few decades. To a substantial extent, this non-philosophical work has focused on what might broadly be described as epistemological or methodological issues—for example, how might causal relationships be discovered from statistical information? By contrast, although philosophers have not entirely ignored broadly epistemological questions, the bulk of philosophical attention has been directed at more metaphysical issues (what is causation? what are the truth makers for causal claims?) or conceptual analysis (how is ‘our concept’ of causation to be analysed?). One consequence is that the connection between these philosophical issues and those concerning how we discover or learn about causal relationships is often left underdeveloped. For example, even if there is a connection of some kind between the claim ‘throwing the ball caused the window to break’ and the counterfactual ‘if the ball had not been thrown, the window would not have broken’, this by itself does not tell us how this analysis of causation bears on how we actually come to form correct causal judgements.

In our BJPS article, we follow a different strategy and ask questions that have been relatively unexplored in the philosophical literature: What has to be the case for causal analysis (that is, the successful description of systems in terms of causal categories) to work? What has to be the case for the discovery techniques referenced above to work? In answering these questions, we are asking about the features that need to be present in the systems in question—the pre-conditions—for causal analysis to apply. We call these features the worldly infrastructure that supports the application of causal analysis. It is these features that are exploited by the various methods we have for discovering causal relationships.

Our main focus is on identifying the relevant features of the infrastructure, but this inquiry prompts a natural follow-up question: why, or under what circumstances, are these features likely to be present? For example, one infrastructure feature that facilitates causal reasoning is the presence of a fair amount of unconditional and conditional statistical independence—the existence of variables X, Y, and Z such that X and Y are independent, or such that X and Y are independent conditional on Z. When a system exhibits such statistical independencies, one can investigate the mechanisms that give rise to such independence.

While the worldly infrastructure we elucidate reflects features of the actual world, the project we describe diverges from currently dominant metaphysical approaches. It is an important part of our project that claims about worldly infrastructure should not be understood as, for example, claims about what does or does not belong to the concept of causation, or as claims about metaphysical necessities involving causation. We think it is a virtue of our project that it does not entangle us in such questions. Our project is the entirely naturalistic one of describing in non-metaphysical terms some of the pre-conditions for the application of causal claims. As a result, the concepts necessary for characterizing the worldly infrastructure are those already employed by scientists, rather than those developed by metaphysicians.

What, more specifically, are the infrastructure features in question? Our list is not meant to be exhaustive, but we take it to include, first, the existence of statistical independencies of the sort mentioned above; second, the existence of systems for which exogenous, intervention-like processes are possible; third, the existence of systematic relationships between probability and causation (like the well-known causal Markov condition); fourth, the existence of systems that are relatively modular at some level of analysis (where modularity means roughly that the system has parts that can be changed independently of each other); fifth, the existence of systems for which it is possible to separate governing generalizations (laws) and initial conditions; and, finally, the existence of systems that exhibit ‘upper-level’ patterns of behaviour that are realization-independent in the sense that the same pattern continues to hold over variations in the system at a more fine-grained level of description.

An important implication of our analysis is that the extent to which these infrastructure features hold for any given system is an empirical question that needs to be resolved on a case-by-case basis. It is entirely possible that if enough of these features fail to be present for some systems, then these systems are resistant to causal analysis or that the theories we use to model such systems are not susceptible of a straightforward causal interpretation. We give an example of this possibility from general relativity. Additionally, it is in principle possible for a system to be causally interpretable only at particular spatiotemporal scales.

Another implication is this: Many of the infrastructure features we discuss do not have to do with whether a system is law-governed, at least not directly. For example, the presence or absence of variables that are statistically independent is not determined only by whether a system is law-governed but also reflects facts about the distribution of initial conditions. A system in which these and other infrastructure requirements are absent may not be one that supports the application of causal claims, even if the system is law governed. Thus it is mistaken to think that as long as there is a law linking X to Y, X causes Y.

Finally, a very common strategy among philosophers interested in the metaphysics of causation or in conceptual analysis is to imagine worlds very different from our own and to ask whether various causal claims would be true in such worlds. They use this methodology to answer questions about what belongs to our concept of causation or what features of causation are metaphysically necessary. For example, one might ask whether there exists a world in which the principle of the common cause fails: X and Y are correlated but X does not cause Y, Y does not cause X, and X and Y do not have a common cause. If so, following the above strategy, one concludes that the principle of the common cause is not part of our concept of causation or does not hold as a matter of metaphysical necessity. One problem with this way of proceeding, from our perspective, is that when one contemplates such worlds, one may be contemplating situations in which the worldly infrastructure needed for the application of causal notions is absent. When this happens, the appropriate conclusion should be that our concepts and strategies for making causal judgements are inapt for analysing whether the causal claims are true in those situations. We should not expect the causal concepts and strategies that we have developed to exploit the infrastructure of the actual world to be reliable guides to causal relations in worlds where that infrastructure is not present.