Woodward on "Sensitivity" and Causation
The colloquium talks that occurred here over the weekend were quite good and the conference successful on the whole. We still, of course, have Nancy Cartwright coming this week. Over the next few posts, I'll say a bit about each of the talks; at least where I think I have something to say. Let's start with Jim Woodward's talk, "Sensitive and Insensitive Causation." I was assigned to do a commentary on that talk. And what I say here will largely reflect those comments.
Most philosophers know that Woodward has developed a new and improved counterfactual theory of causation (a direct descendant of David Lewis' view, e.g., 1986). The most in-depth resource on Woodward's account is his 2003 book, Making Things Happen (which won the Lakatos Award this year). Some call Woodward's view "manipulationist" or "interventionist," which is correct, since intervention is part and parcel of the view. However, counterfactuals are central to the account and, in particular, are what provide both the advantages and disadvantages to it. So just saying that Woodward-style causation is "manipulation" hardly gets to the crux of his thesis.
At any rate, accepting some or other counterfactual account of causation requires that you recognize that the causal dependence between some cause C and some effect E is just a kind of counterfactual dependence. Thus, the occurrence of E is counterfactually dependent on the occurrence of C if and only if:
O(C) --> O(E)
-O(C) --> -O(E)
where O(C) and O(E) are propositions about the events C and E. Consider the example that Suzy's throwing the rock caused the bottle to break. Here,
If Suzy had thrown the rock, then the bottle would have broken.
If Suzy had not thrown the rock, then the bottle would not have broken.
Intervention enters Woodward's picture as a condition on which counterfactual dependence expresses causal dependence. Here's Woodward's idea informally: An intervention is an idealized experimental manipulation performed on some variable X to determine whether changes in X are causally related to changes in some other variable Y. X is a cause of Y when, on an intervention of X, a change in Y occurs only in virtue of the change in X due to the intervention. A generalization is a causal generalization when the relationship between X and Y is invariant to some degree under interventions.
Woodward's talk assumes the above. But it wasn't about the above. Rather, Woodward's talk was about the sensitivity and insensitivity of causal claims and generalizations (a topic Lewis took up first, e.g., 2000). According to Woodward, the sensitivity of a causal relation or generalization (or counterfactual) has to do with whether those relations, etc. would continue to hold under changes in other conditions, Bi, distinct from C and E. That is, sensitivity is about the extent to which causal relations continue to hold under changes in their background conditions. Take Suzy's rock throwing again. The counterfactual is quite insensitive, in particular to changes in the location of the event, timing of the event, and such similar things. Or consider, as Woodward does, a driver who hits an icy patch in the road, skids, loses control, but regains it and doesn't end up in the ditch beyond. The relation is sensitive to things like the speed of the vehicle, how much the driver had to drink, angle of attack, and so on. Sensitivity/insensitivity is a dimension of Woodward's invariance idea.
With that framework articulated, Woodward discussed a number of cases in which we're interested in the relative sensitivity of causal relations. Given my interests, I focused on Woodward's discussions of genetic causation rooted in heritability analyses and the apportioning of causal "importance" in evolutionary biology.
According to Woodward, causal claims based in heritability analyses, like "genes cause IQ," "our religiosity is genetically based", "shyness is genetic" are all sensitive causal claims, meaning that it takes very little in terms of changes in background conditions for the relation to fail. Indeed, Woodward says, it's the sensitivity of the claims that reflect our suspicions about them. It's true enough that claims like those I listed break down quickly once conditions change. But to say that such claims rooted in heritability analyses are sensitive is to give them more credence than they deserve. As I argued in my comments, such claims aren't sensitive causal claims, they are utterly mistaken causal claims about which the sensitivity claim doesn't apply. And arguing that it does takes the value out of the concept.
There are two concepts of heritability in heritability analyses, broad (H2) and narrow (h2). They are defined by way of components of phenotypic variance, VP. Commonly, phenotypic variance is partitioned into genetic and environmental components, so that VP is made up of VG and VE. VG is further partitioned into additive and non-additive genetic components. The additive component, Vg, is due to the variation in the alleles at all loci in the population when they are assumed to be acting independently of each other. The non-additive component assumed genetic interaction such as dominance, gene interaction within a locus, and epistasis, interaction between loci, Vd and Ve. The narrow sense of heritability is the variance attributable to the additive component of genetic variance, h2 = Vg/VP. The broad sense is the variance attributable to the genetic variance, H2 = VG/VP.
One of the best discussions of heritability analyses I know of is Sahotra Sarkar's (1998). My discussion is sketchy, but it follows his. Now, when heritability analyses are used to attribute genetic and environmental (or nature versus nurture) causes for some trait, two key assumptions are made: (A1) the heritability of a trait is a monotonic function (always increasing or decreasing) of the degree to which that trait can be accounted for from a genetic basis; and (A2) the heritability of a trait of an individual should only be a function of its genotype and its developmental (non-genetic) environment, that is, that VP = VG + VE. If things haven't gone wrong already, they go terribly wrong about now. There are a number of technical problems internal to heritability analyses, such as the fact that the definitions of broad and narrow heritability are lacking. But I want to set those aside for some more lurid problems.
Woodward points out the first problem, which is that A1 and A2 violate the population-level definitions of heritability. Just hearing the definitions of broad and narrow sense heritability show that A1 and A2 are misleading; there is a tacit (unexplained) reduction to individual traits from population-level concepts. A second problem stems from A2. The assumption that the heritability of a trait of an individual is a function only of its genotype and its development is simply mistaken. In one sense, it's incomplete because it ignores numerous other factors involved in inheritance, such as epigenetic (non-nuclear) factors, cytoplasmic factors, gut micro-organisms, specific social traditions, and other features of the environment such as geographic range. In another sense, A2 is flat wrong in its additivity assumption –there is no good reason to assume that VG and VE, even if they're the only factors affecting VP, contribute to VP independently, that is, tha VP = VG + VE. There is almost always an interaction between genotype and environment so that the effect of the environment on the phenotype differs between genotypes. And there are often non-random correlations between genotypes and environments. Now, we can make a repair to VP by writing it as VP = Vg + Vd + Ve + VE + VG*E + COV(G, E) (and VG = Vg + Vd + Ve). But that is still misleading if not outright wrong –there's still a reliance on an additivity assumption that's unwarranted.
It seems to me that these observations that causal claims like "genes cause IQ," when they're rooted in heritability analyses, are more senseless than sensitive. So it's not clear why one should bother to worry about sensitivity here.
Let's consider the "importance" issue. Woodward argued that differences in degree of sensitivity are connected to judgments about the relative importance of causal factors in accounting for some outcome. One way he illustrated this claim is by framing the debate between adaptationists and developmentalists over the importance of natural selection in accounting for the traits of organisms. Woodward's illustration here is fine so far as it goes. But I've never thought that debates of the "King Kong vs. the Blob" variety generate much light. Nevertheless, independently I've begun thinking about the basis of biologists' claims to the relative importance of neo-Darwinian evolutionary causes, i.e., natural selection, genetic drift, migration, mutation, and so on. And while I've not seen biologists appeal explicitly to sensitivity in apportioning importance, I think there's a role for sensitivity to play. And perhaps more importantly, I think the biologists' reasoning explains how at least some sensitivity claims relevant to the importance issue may be concretely articulated.
To make my point, I used the time-course example of mutation and drift, which I discussed in a previous post also about the relative "importance" of evolutionary causes. The idea is that an evolutionary cause is relatively more significant than another when the former affects the evolutionary trajectory of a population in a shorter time than the latter. Thus, a cause with a short time-course will swamp out one with a long time-course. Using the drift/mutation case and putting it in Woodward's terms, drift as a cause of a population's decrease in genetic variation is very sensitive to population size, a background condition, which has implications for its significance relative to other evolutionary causes, depending on the background conditions. Sensitivity, then, is relevant to time-course measures for the assignments of importance to evolutionary causes. Better yet, the time-course illustration is a specific and concrete way of explaining how biologists make judgments about the relative importance of evolutionary causes.
Ultimately, Woodward's talk was very good, and covered more ground more subtly than I have here. His paper, with the same title, is coming out in The Philosophical Review. Look for it.
References
Lewis, D. (1986), "Postscript C to 'Causation': (Insensitive Causation)", in Philosophical Papers, vol. 2. Oxford, UK: Oxford University Press. pp. 184-188.
Lewis, D. (2000), "Causation as Influence", Journal of Philosophy 97: 182-197.
Sarkar, S. (1998), Genetics and Reductionism. Cambridge, UK: Cambridge University Press.
Woodward, J. (2003), Making Things Happen. New York: Oxford University Press.
Dr. Skipper, I'm glad that you deleted my comments. It would have been embarassing to see my inexcusably boorish remarks exposed for all to see. What can I say? I'm pretty new to blogging and have discovered an unpleasant truth about myself, namely, that I'm not the kind of person in whom anonymity brings out the best. (Like some of my other GNXP co-bloggers, I have no choice about whether to blog pseudonymously or not, given the current climate in the Academy.) I hope you have accepted my apology. Needless to say, of course, I stand by my substantive points.
A word about the resolution of our differences. I doubt that any purely verbal arguments of yours could possibly persuade me to give up a paradigm that has proven so massively fruitful and productive. Why has the application of quantitative genetics in selective breeding (starting with Sewall Wright's work with cattle) proven so successful? Why does the fraction of the genome shared identical by descent between siblings (which is a random variable with expectation 0.5 and can vary in principle between 0 and 1) predict their similarity in height, and, moreover, why is the heritability estimate produced in this way so congruent with past estimates based on traditional black-box quantitative-genetic methods? Philosophical arguements are not going to move me. You will have go through a massive literature and explain why its results are more congruent with the model that you prefer and not mine.
Posted by: Darth Quixote | May 16, 2006 at 08:08 PM
The heritability debate has a tendency to get nasty very quickly. Perhaps for that reason it's a poor case to mine for insights into the conceptual foundations of genetics.
I had originally discussed heritability analyses --roughly-- in relation to other work about the causation. The strong version of my view is that the theoretical structure is too impoverished to ground specific causal claims about genetic (or environmental) causation in phenogenesis. With that said, your comments about "success," "prediction," and "results" are orthogonal that project.
There are the philosophical projects of understanding causality (after Hume) and how such an understanding is reflected in the special sciences. And there are the philosophical projects about "saving the phenomena," empirical adequacy, and so on. It seems to me that your comments here reflect the latter and not the former. No doubt we can agree on successes of heritability analyses in that context. One family of successes comes out of animal and plant breeding.
Posted by: Robert Skipper | May 17, 2006 at 10:49 AM