Contribution
It offers a deflationary reading of “top-down causation” within one sense of level, levels of mechanisms. Whatever is intelligible in an interlevel causal claim is a mechanistically mediated effect, “hybrids of causal and constitutive relations, where the causal relations are exclusively intralevel” (abstract, p. 547). An interlevel claim that cannot be translated this way is “suspect”. The paper does not argue against interlevel causation from a theory of causation. It shows a translation for six worked cases and lists the assumptions about causation that interlevel causes would break.
Key insight
A mechanism and its components are not distinct existences, so a change in the whole is already a change in some parts, and a change in a part is already a change in the whole. What looks like downward influence is a change in some components, constitutive of the whole’s new condition, that then propagates through ordinary causation among components. The symmetry of interlevel experiments, which makes interlevel causation look puzzling, is “readily explained” by constitution. “The relation is symmetrical precisely because the mechanism as a whole is fully constituted by the organized activities of its parts” (p. 554).
Assumptions
- Higher-level causes exist, in the sense that (1) mechanisms are organized collections of entities and activities, (2) mechanisms affect and are affected by things, and (3) “the parts of the mechanism taken singularly cannot be so affected or have such effects” (p. 548). This is assumed, not argued.
- All higher-level causes are explained by lower-level mechanisms (p. 548). Also assumed. Strong emergence, a property with no mechanistic explanation, is set aside: its “levels” are not levels of mechanisms (p. 551).
- Causation is intralevel, between distinct relata, and typically asymmetrical. The paper rests on Lewis’s distinctness condition (p. 552) and the asymmetry of dependence (p. 553). It grants that “None of these arguments is decisive” (p. 554).
- Level is relative to a mechanism. “Levels of mechanisms are defined locally, within the context of a given type of mechanism” (p. 550). The setting is narrower than the title: it concerns levels of composition within one mechanism, not levels of size, of science or of ontological category.
Key results
The results are a definition, an argument, and translations.
- Levels of mechanisms (p. 548). “An item X is at a lower level than an item S if and only if X is a component in the mechanism for some activity ψ of S. X is a component in a mechanism if and only if it is one of the entities or activities organized such that S ψs.” The definition takes “component” as given. It gives no test for it.
- Interlevel experiments (pp. 553–554). There are four kinds:
correlation of a component’s presence with the whole’s behaviour;
lesion or inhibition; stimulation; and intervening on the whole while
recording the parts (neuroimaging). They “share a form with Mill’s
methods”. They run in both directions, which causal tests generally do
not. Two qualifications in the footnotes matter:
- n. 10, p. 553: a deficit after a lesion “is prima facie evidence that X is a component”, “It is only prima facie since if the component is not interactively integrated into the system, it could also be construed as constituting a background condition on the operation of the mechanism, not as part of it.”
- n. 11, p. 554: neuroimaging judgments “are prima facie ones and depend on discovering that the components are deeply integrated into the system.”
- Intersize is not interlevel (pp. 550, 556). A virus that kills a general is not a component of the general, so the case is ordinary causation between things of different sizes. But “there are pragmatic contexts in which a similar interlevel analysis is appropriate” (p. 557): when the question is how the virus killed her, its path through her cells is traced as part of the mechanism. What counts as inside the mechanism moves with the explanandum.
- The translations.
- Bottom-up: the general’s heart attack (p. 557). “when we reach the state of the mechanism that constitutes the state of death, we do not say … that it causes death. It just is death.”
- Top-down, parts “carried along for the ride”: Ignatius’s hotdogs (p. 558). The hotdogs are not working parts, and “Ignatius … is doing all the work” (p. 558).
- Top-down, parts “enlisted in the ride”: Hal’s glucose metabolism (p. 559), traced through ATP use and glycolysis. Kane’s memory (pp. 560–561) is the same, given a mechanistic account of intentional content, which the paper does not try to defend.
- The general strategy (p. 561). Every intelligible interlevel claim “is analyzed into a causal claim coupled with one or more constituency claims”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Mechanisms and their components cannot stand in causal relations under transmission, distinctness or asymmetry views of causation | moderate | argument (pp. 551–554); the authors grant it is not decisive |
| C2 | Every intelligible case of interlevel causation within levels of mechanisms is a mechanistically mediated effect | weak to moderate | six worked cases (pp. 556–561); no general argument, and stated assumptions do much of the work |
| C3 | Interlevel experiments are symmetrical because the whole is constituted by its parts | moderate | argument (p. 554) |
| C4 | Such experiments show only prima facie that an item is a component; one not interactively integrated may be a background condition | stated, not argued | nn. 10–11 (pp. 553–554) |
| C5 | A strongly emergent property, lacking any mechanistic explanation, does not sit at a level of mechanisms | strong, by definition | p. 551 |
Concepts
- Level of mechanisms: a level defined by the component relation within one mechanism. It is local, not a stratum of nature (pp. 548–550).
- Component: an entity or activity “organized such that S ψs” (p. 548). Contrasted, in n. 10, with a background condition, which is a part not “interactively integrated” into the system.
- Mechanistically mediated effect: a hybrid of intralevel causation and interlevel constitution, which stands in for an apparent interlevel cause (p. 547).
- Carried along for the ride / enlisted in the ride: components that change only because the whole moves, or components whose activity a change of the whole’s condition recruits (pp. 558–559).
Connections
It applies Craver’s account of levels (Craver 2001, and Explaining the Brain, LIT-832, then forthcoming). The account of interlevel experiments is Craver 2002, the source of the mutual-manipulability criterion. That criterion is stated in Explaining the Brain and in “Constitutive explanatory relevance” (LIT-1930), neither of which was reached. The paper answers Kim’s (2000) argument that simultaneous upward and downward influence makes causal circles. It anticipates the later charge (Baumgartner and Gebharter 2016, reported in LIT-1931) that an interventionist reading of these experiments confuses constitution with causation. Here the symmetry of the experiments is the evidence for constitution. Bechtel’s later paper on top-down causation through networks and dynamics (PhilSci-Archive preprint 2016, not held) reports critics’ objection that this account, like Kim’s, allows causation only at the lowest level. That is known here from a search snippet, not from reading it.
Limitations
- No test of componency. The definition of level takes “component” as given. The experiments supply only prima facie evidence, and the difference between a component and a background condition is left to an undefined “interactively integrated”. Constitutive relevance (LIT-832, LIT-1930) is meant to supply the test, and this paper does not.
- The conclusion is mostly assumed. That all higher-level causes have mechanistic explanations is an assumption (p. 548). So the translation’s success shows that the view is coherent, not that it is true.
- Feedback is waved through. “Even if this is made possible by the incorporation of complex interactions, including feedback loops, inside the mechanism, there is nothing mysterious” (p. 560). The paper does not ask whether a feedback loop that crosses the mechanism’s boundary changes the analysis. That is the case the record needs.
- Kane’s case rests on a mechanistic account of intentional content, which the paper explicitly does not defend (p. 560).
Open questions
- What makes a part “interactively integrated” (n. 10)? Is it a matter of causal structure, of spatial inclusion, or of the explanandum chosen? The paper’s virus example suggests the explanandum (p. 557).
- Does an item that is causally coupled to a system in both directions, but located outside it, pass the paper’s four experiments? For a state-feedback controller it does (CLAIM-530). If so, what rules it out as a component, other than a boundary drawn beforehand?
Bearing on the record’s own argument
-
CLAIM-522: the boundary it requires is the paper’s undefined “component”. CLAIM-522 says that only interventions on mechanisms, relative to a declared boundary, separate an internal producer from an outside controller that supplies the same law. Run this paper’s four experiments (pp. 553–554) on both systems, with maintenance of c as the phenomenon.
- Lesion: disabling the internal producer in A, or the controller in B, ends maintenance.
- Stimulation: raising either’s output raises c.
- Intervening on the whole: perturbing c changes the activity of either, since each reads c.
- Correlation: either co-varies with c.
Every test comes out the same. By n. 10 the controller is no “background condition”, because it is interactively integrated, reading c and acting on it. So by the paper’s own markers the controller is a component of the mechanism for “c is maintained”. What puts it outside is the declared boundary, which the phenomenon “the system maintains c” already names. That is filed as CLAIM-530. It is CLAIM-522’s own closing proviso, that the boundary is not given, now traced into the mechanist’s account of constitution.
-
CLAIM-194: a second instance of the same gap. CLAIM-194 says a coupling criterion has “no term for who made the coupling”. This paper’s experiments have no term for which side of a boundary a component sits on. Interactive integration, the paper’s only discriminator, is a coupling notion. So it inherits CLAIM-194’s blindness to who made the coupling. Same structure, not shown same mechanism: CLAIM-194 is about observational statistics, and this is about interventions.
-
CLAIM-414 and QUESTION-086: what the experiments do separate. Within a fixed boundary the lesion experiment is just CLAIM-414’s separating intervention. Disabling the inside pathway ends A’s maintenance and leaves B’s untouched. So the paper supports the record’s view that an intervention on mechanism, not state, is what separates them. It supplies no intervention that finds the boundary. QUESTION-086’s “by what intervention” is answered by this paper only relative to a boundary already drawn.
-
QUESTION-093 and the “higher-level agent”. On this paper’s account, a coalition’s higher-level “influence” on its members is a mechanistically mediated effect: members enlisted in the ride (p. 559). Nothing in QUESTION-093’s four criteria needs top-down causation in a stronger sense. The paper also gives no reason to think closure or self-individuation could be read from interlevel experiments without a boundary.
-
THEORY-072 and production. THEORY-072 draws the line at “constraints that produce one another”. In this paper’s vocabulary, mutual production is intralevel causation among components, and the organism’s persistence is constituted by it. That is consistent with THEORY-072. It adds nothing on whether production happens inside or outside the boundary.