There are a couple of problems with the theory of causal mechanisms that will be difficult to address. Jim Mahoney raises a general concern in “Beyond Correlational Analysis” — there is no consensus about how to define a mechanism. But there are more specific problems as well.
One is the issue raised by Johannes Persson with respect to Jon Elster’s formulation, and the apparent paradox that arises from Elster’s epistemic definition: according to Elster’s definition, if we learn more about a mechanism, it ceases to be a mechanism (link). This has to do with whether we understand mechanisms epistemically or realistically. In my reply to Persson I argue that we need to take the realist approach (link).
Second, there is the problem that Andrew Abbott raises in his critique of the social mechanisms approach (link), where he highlights the continuity rather than discreteness of social processes. This reflects Abbott’s ontological preference for activity rather than actors, processes rather than discrete mechanisms. (I made a point somewhat similar to this in a post about the “discreteness” of causal mechanisms (link). I argued we should not look at mechanisms as atoms or molecules of causation. There I was inclined to push the action down to the level of the actor.)
A third problem may be even more serious. This derives from the condition of “regularity” that many definitions of a causal mechanism invoke. “A produces B” constitutes a CM if the real properties of A work to bring about B and if A is regularly associated with B. (I refer to these as pocket-sized regularities.) The basic idea here is that once the mechanism is triggered, it leads more or less inexorably to its result.
The first part concerns the idea of real causal efficacy. Here is Rom Harre, excerpted by Mahoney:
The structures, states and inner constitutions from which the phenomena of nature flow…. the permanent or enduring conditions under which a certain kind of phenomenon will occur.” “The inner constitutions, structures, powers, encompassing systems, and so on, of which natural generative mechanisms are constituted, and of which the connection between cause and effect usually consists. Harre (1970, pp. 101,102, 104)
The second criterion expresses the idea that trigger and outcome are linked by a strong regularity or law. Here is Stuart Glennan:
A mechanism underlying a behavior is a complex system which produces that behavior by the interaction of a number of parts according to direct causal laws. (Glennan, “Mechanisms and the Nature of Causation,” Erkenntnis 1996)
The problem is that these two criteria are likely to pull apart in many circumstances. There may be a real causal push from A to B, but it may result in a relatively low incidence of A followed by B. In this circumstance it is true that A is part of the cause of B, but it is not true that A->B represents a causal mechanism. And in fact we do refer to mechanisms where the likelihood of the consequent is low even when the antecedent occurs — for example, “the mechanism of George’s becoming ill is his close exposure to Alice who had the contagious illness.” It may be that the likelihood of catching the flu is still low following exposure; but there is no other way of getting the flu.
This would entail that it is not the case that all causal linkages are transmitted by causal mechanisms, and we would be forced to refer to causal powers or processes instead. Further, the question of whether a given linkage represents a mechanism or not depends only on the relatively boring facts about the world that push up or down the probability of the linkage — not some fundamental ontological distinction between mechanisms and other causal connections.
This seems to suggest that we have a difficult choice. Either we can drop the “regularity” criterion or we can acknowledge that “causal mechanism” does not capture all reasonably direct causal connections. If we take the second route, we are ill advised to define causation in terms of causal mechanisms per se; instead, we need a more comprehensive concept of “real causal structures and processes”, of which a subset are causal mechanisms.
It may be that considerations surrounding the third issue are impossible to resolve without dropping the requirement that a mechanism corresponds to some sort of regularity.
Another possibility is to weaken the regularity criterion so that it encompasses causal relevance rather than high probability of succession. The causal relevance test offered by Wesley Salmon would handle this approach for a mechanism like “contagion through exposure” along these lines: illness | exposure > illness | ~ exposure. Would this suffice as a corrective to the current definition? Might we reframe the definition along these lines:
A => B represents a causal mechanism if
(i) the real properties of A produce B
(ii) B | A > B | ~A
(Gambetta takes something like this approach.)
But if we take this approach, we lose the ability to infer from cause to effect through the mechanism. In alternative language, we move from thinking of mechanisms as sufficient causes to necessary conditions (or rather their probabilistic equivalents).