Download
mechanistic evidence for causal inference n.
Skip this Video
Loading SlideShow in 5 Seconds..
Mechanistic evidence for causal inference PowerPoint Presentation
Download Presentation
Mechanistic evidence for causal inference

Mechanistic evidence for causal inference

113 Vues Download Presentation
Télécharger la présentation

Mechanistic evidence for causal inference

- - - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - - -
Presentation Transcript

  1. Mechanistic evidence for causal inference Disambiguating the RWT Phyllis Illari University of Kent

  2. RWT • Says: ‘To establish causal claims, scientists need the mutual support of mechanisms and dependencies. … The idea is that probabilistic evidence needs to be accounted for by an underlying mechanism before the causal claim can be established’ (2007 p159.) • From our previous discussions, ‘mechanistic evidence’ as discussed in that paper (and especially as it is picked up!) is multiply ambiguous. • I want to: • Make three distinctions. • Argue that several of the disambiguated meanings fed back into the RWT are worth supporting.

  3. 1: First Distinction

  4. Distinction 1 • There are two kinds of evidence • Mechanistic and difference-making evidence are different kinds of evidence. • There are two kinds of things we have evidence of • Evidence of difference-making and evidence of a mechanism are different. Without further work, we should think that you could have mechanistic evidence of both a mechanism and difference-making, AND difference-making evidence of both a mechanism and difference-making.

  5. Distinction 1

  6. 1a: Mechanistic evidence matters

  7. Russo-Williamson • ‘two different types of evidence— probabilistic and mechanistic—are at stake when deciding whether or not to accept a causal claim.’ (2007 p163.) • This is why you have been read this way.

  8. Two kinds of evidence • Quantitative vs qualitative • Generic vs single-case • Statistical vs simple manipulation • Evidence that requires large numbers of repeated trials vs evidence that lets you see pretty straightforwardly that something is the case.

  9. Two kinds of evidence • These all matter to causal inference. • The fourth in particular is interesting as an account of mechanistic evidence as a kind of evidence. • Some evidence doesn’t require the same kind of repeated replication as statistical evidence. • Sometimes you can just see that something is the case from experimental work. It requires confirmation by a different research group, at least, but only that. • In such cases we make inductive inferences based on a few cases only. • It seems that these are cases where technological breakthrough lets you see structure that was previously unavailable.

  10. Franklin’s X-ray diffraction photograph of DNA

  11. Confirmation of the existence of replication forks

  12. Two kinds of evidence • BUT: Other cases where we make inductive inferences based on a few cases – light-bending experiments. Doesn’t look like mechanistic evidence. • Might say this is different because such experiments are only decisive against a detailed theoretical background, while these are simpler. • But that is to exaggerate the difference. Good deal of theory involved in the generation of these images. • Further, even our willingness to make inductive inferences based on very few samples in such cases are probably based on the evidential context/background knowledge (where we are confident that there is only a small likelihood of confounders). • These cases are significant in their fields, but there is NO such thing as mechanistic evidence as a type of evidence always or usually required for causal inference. Best to maintain multiple distinctions among types of evidence.

  13. 1b: Evidence of a mechanism matters

  14. Evidence of • Evidence of difference-making is just evidence of one or more difference-making relations: probabilistic, counterfactual, invariance. • ‘Mechanisms are entities and activities organized so that they are responsible for one or more phenomena.’ (Illari and Williamson.) • So evidence of a mechanism is evidence of activities, entities, and their organization. • PROBLEM: evidence of these things are got using the same kinds of experimental techniques. • observing variation in different populations as in observational surveys or clinical trials; observing the results of a few cases or repeated simple or experimental manipulations; from observing the results of either physical or computer simulations of a system.

  15. Compare Erik Weber • ‘Mechanistic evidence is bottom-up evidence: it consists in using information about the behavior of the parts in order to support causal claims about the system as a whole.’ (Weber p284.) • Also looks like evidence of a mechanism. Looks congruent with our more specific claim.

  16. The job of mechanistic evidence • ‘In other words, mechanisms allow us to generalise a causal relation: while an appropriate dependence in the sample data can warrant a causal claim ‘C causes E in the sample population’, a plausible mechanism or theoretical connection is required to warrant the more general claim ‘C causes E’. Conversely, mechanisms also impose negative constraints: if there is no plausible mechanism from C to E, then any correlation is likely to be spurious. Thus mechanisms can be used to differentiate between causal models that are underdetermined by probabilistic evidence alone—see section 4.’ (2007 p159)

  17. Evidence of still needs care • The distinction in object of evidence still matters, because the different kinds of evidence address differentproblems for causal inference. • Most importantly, they each address the central problem of the other: • When you have evidence of a mechanism: • Problem is masking • Advantage is avoiding confounding • When you have evidence of difference-making: • Problem is confounding • Advantage is it eludes masking

  18. Distinction 1

  19. 2: Second distinction (Within evidence of)

  20. Distinction 2: evidence of a mechanism • Evidence of a mechanism underlying the (allegedly causal) relation between C and E. • Evidence of general mechanistic facts about the domain. • These are different: • 2 without 1 common. • Might just have 1 without 2. (Having 1 is an item of 2, so perhaps say might have 1 without any other 2.)

  21. Evidence of a mechanism • Plausible that you can’t infer ‘C causes E’ without evidence of general mechanistic facts about the domain. • Can’t move from difference-making evidence to causal claims because of problem of confounders – need evidence of mechanisms. • The kinds of simple manipulations not needing extensive repetition seem to give you evidence of mechanism and of difference-making simultaneously. • May be some cases where it is possible to infer ‘C causes E’ without evidence of a mechanism underlying the relation between C and E. • Data mining? But perhaps only establishes narrowly specified difference-making claim. • Absences? I think mechanisms can involve absences.

  22. Distinction 2

  23. 3: Third Distinction (Within evidence of a mechanism underlying the relation between C and E.)

  24. Distinction 3: How much evidence? • Evidence of what the mechanism is underlying ‘C causes E’. • Evidence that there is a mechanism underlying ‘C causes E’. • Plausible how-so story about a mechanism underlying ‘C causes E’. • No evidence that there is not a mechanism underlying ‘C causes E’. • Obviously these start at very demanding, and become less so.

  25. Steel’s objection • Steel thinks the alleged negative role of mechanisms cannot work, because can always think of a plausible mechanism. (Steel 2004.) • Looking at 1-4 undermines this worry. Steel clearly concerned about how-so stories, but much more to it. • Steel is underestimating the importance of 4. That we have positive evidence that there is no such mechanism (from evidence of general mechanistic facts about the domain) is an important reason for scientific conclusions – storks and babies example.

  26. How much evidence? • Plausible that you almost always need NO evidence that there could NOT be a mechanism. • NOT plausible that you almost always need evidence of what that mechanism is. • When you do have such evidence, that is very helpful. • Plausible that you frequently need evidence that there is some mechanism or other. • Need care. Want evidence in this case over and above the simple fact of a difference-making relationship! (This alone is prima facie evidence of the existence of a mechanism.)

  27. ‘In other words, mechanisms allow us to generalise a causal relation: while an appropriate dependence in the sample data can warrant a causal claim ‘C causes E in the sample population’, a plausible mechanism or theoretical connection is required to warrant the more general claim ‘C causes E’. Conversely, mechanisms also impose negative constraints: if there is no plausible mechanism from C to E, then any correlation is likely to be spurious. Thus mechanisms can be used to differentiate between causal models that are underdetermined by probabilistic evidence alone—see section 4.’ (RW p159)

  28. 4: Conclusion

  29. Three distinctions

  30. Objections • Howick: Well-conducted RCT a counterexample. • I just don’t agree. Our security in the conclusions of our RCTs implicitly rely at least on evidence of mechanistic facts about the domain, if not on evidence of a mechanism underlying ‘C causes E’ itself. • Broadbent: Bias towards old knowledge. • This is a legitimate requirement on scientific method. If a new finding is inconsistent with other things we know about the domain, we should be suspicious. Further work may of course alter our understanding of the domain, and vindicate the finding.

  31. Plausible RWT Claims 1 Mechanistic evidence: • There are different kinds of evidence that matter to causal inference. • Evidence of a mechanism and mechanistic evidence are not the same. • There is no one type of evidence that looks like a good candidate for mechanistic evidence. • It is better to maintain multiple distinctions in types of evidence. • Mot plausible that mechanistic evidence as a type is necessary forcausal inference.

  32. Plausible RWT Claims 2 Evidence of a mechanism: • Evidence of a mechanism underlying ‘C causes E’ and evidence of general mechanistic facts about the domain not the same. • Plausible that you can’t infer that ‘C causes E’ without general mechanistic evidence about the domain. • Possible cases where you can infer that ‘C causes E’ without evidence of a mechanism underlying the relation between C and E.

  33. Plausible RWT Claims 3 How much evidence of a mechanism: • Plausible that you almost always need NO evidence that there could NOT be a mechanism. • Plausible that you frequently need evidence that there is some mechanism or other. • NOT plausible that you almost always need evidence of what that mechanism is.

  34. Problem 1: Regress • Natural problem when first see the thesis.

  35. Problem 2: Absences • Failing to vaccinate causing death.