|Causation in Biology: Stability, Specificity, and the Choice of Levels of Explanation*
Philosophical discussion of causation has tended to focus, understandably enough, on finding criteria that distinguish causal from non-causal relationships. There is, however, another important project, also belonging to the philosophy of causation that has received less attention, at least among philosophers. This is the project of elucidating and understanding the basis for various distinctions that we (both ordinary folk and scientists) make among casual relationships. This essay attempts to contribute to this second project. In particular, I focus on certain causal concepts (used to mark distinctions among causal relationships) that are employed in biological contexts; these include the notions of non-contingency of association, appropriate choice of level of causal description or explanation, and causal specificity. These notions turn out to be interrelated in various complex ways.
In saying that attention has tended to focus on the first of the two projects distinguished above, I do not mean that the second project has received no attention at all. One does find self-conscious discussion of notions like causal specificity among researchers in many different areas of biology, including epidemiologists, geneticists, and molecular biologists. Moreover, in the philosophical literature there are discussions of closely related ideas, although the connections with causal notions of biological interest are rarely explicitly recognized. In particular, as I discuss below, the notion of non-contingency of association is closely related to the notion of the stability, insensitivity or invariance of a causal relationship, as discussed by, e.g., Mitchell (2000) and by me (Woodward, 2003, 2006), the notion choosing an appropriate level of explanation is related to Yablo’s idea (1992) that causes should be “proportional” to their effects, and the notion of causal specificity has interesting relations both to the notion of proportionality and to Lewis’ (2000) notion of influence. However, recognition of these connections is complicated by the fact that both the biological and philosophical literatures sometimes fail to distinguish between the two projects described above. More specifically, the features under discussion (non-contingency, specificity etc.) are not infrequently treated (e.g., in Susser, 1977) as conditions that can be used to distinguish between causal and non-causal relationships, rather than (as I would urge) features that should be used to distinguish among causal relationships. In particular, it is common in the biological literature (particularly in epidemiology- e.g. Hill, 1965) to refer to these features as “criteria for causation”; this has suggested both to biologists and others that the features are proposed as necessary conditions for a relationship to be causal. This in turn prompts the response that relationships can qualify as causal even if they lack some or all of the features of stability, specificity and so on. I agree, but urge that it does not follow that the features are unimportant for theorizing about causation or that they do not play important roles in particular scientific contexts
My aim in this essay is to elucidate what is meant when causal relationships are described as more or less contingent, specific, or framed at an appropriate or inappropriate level, to explore some of the interrelationships among these notions, and locate them notions within a larger framework for discussing causation and explanation. I will also try to illustrate how a concern with whether causal relationships are specific, stable and so on arises in a very natural way in many biological contexts. I will add that although I have attempted to provide biological illustrations of these causal notions, my primary interest is in the content of the notions themselves and less in the empirical details of the illustrations. For example, it is commonly claimed that the causal relationship between DNA sequence and the proteins for which it “codes” is “specific”. My concern is with what this claim means—with the empirical features that biologists believe this relationship to possess which leads them to think of it as specific -- and only secondarily with the complicated and controversial question of whether the relationship in fact possesses these features. For example, some will hold that it is more accurate to think of the causal specificity achieved in protein synthesis as not due to DNA sequence alone but instead as the result of the interaction of DNA with many other transcriptional factors1. Others may think that many causal relationships in biology – e.g., those having to do with gene action -- are less “specific” than commonly supposed. But even in these cases, we still face the questions of what features a relation must possess in order to count as specific and what the contrast between specificity and non-specificity amounts to. It is these sorts of question that I will be exploring2.
2. Causation and Explanation.
My strategy in what follows will be to introduce a very undemanding or minimalist notion of causation, based on the interventionist framework described in Woodward, 2003. I will then use this as a basis on which to explore the various other distinctions, having to do with stability, specificity and so on, that might be made among causal relationships satisfying this minimalist conception.
Consider the following characterization of what it is for X to cause Y (where “cause” here means something like “X is causally relevant to Y at the type-level”):
(M) X causes Y if and only if there are background circumstances B such that if some (single) intervention that changes the value of X (and no other variable) were to occur in B, then Y or the probability distribution of Y would change.
Here X and Y variables, which as Woodward, 2003 explains, are the natural candidates for the relata of causal claims within an interventionist framework. A variable is simply a property, quantity etc, which is capable of at least two different “values”3. Background circumstances are circumstances that are not explicitly represented in the X-Y relationship, including both circumstances that are causally relevant to Y and those that are not. An intervention on X with respect to Y as an idealized experimental manipulation of X which causes a change in Y that is of such a character that any change in Y occurs only through this change in X and not in any other way4.
As an illustration, according to M, short circuits cause fires because there are background circumstances (including, e. g., the presence of oxygen) such that in these circumstances, intervening to change whether a short circuit is present or absent will change whether a fire occurs (or the probability of whether a fire occurs) . Similarly, consider Richard Dawkins’ (1982) hypothetical example of a gene R, such that those with some abnormal variant r of this gene do not learn to read (because they have dyslexia) while those with the normal form r* do learn to read (given appropriate background conditions, including the right sort of schooling etc.) Assuming that intervening to change the normal form r to the variant r* (or vice-versa) is associated (again in appropriate background circumstances) with changes in whether its possessor learns to read, R will count as a gene that causes reading, according to M.
This last example emphasizes what I meant in saying that M characterizes a weak and undemanding notion of “cause”; undemanding in the sense that it allows a relationship to qualify as causal even if it lacks features thought by some to be characteristic of paradigmatic causal relationships. Thus a not uncommon reaction to Dawkins’ example is that, if the facts are as he describes them, it is in some way misguided or misleading or perhaps just false to describe R as causing reading – hence that M is in need of emendation since it supports this description. For those who are worried about M for this reason, I emphasize again that my strategy in what follows is to use M as a foil or baseline to which other more demanding conditions on causation (having to do with stability, specificity etc.) may be added. These additional conditions (particularly, in this case, stability—see section 2) may be used to capture what it is misleading or defective about Dawkins’ causal claim and more generally to characterize “richer” notions of causation.
Note that according to M, the claim that X causes Y in itself commits us to nothing specific about which changes in X (produced by interventions) are associated with changes in Y and also says nothing about the particular background conditions B under which this association will occur. (It is enough that there exists such B.) Within the interventionist framework, information of these latter sorts is spelled out in terms of more detailed and specific interventionist counterfactuals specifying in a more detailed way just how Y changes under various possible interventions on X and under what background conditions such changes will occur. It is this more detailed information which is related to the considerations having to do with stability, specificity, and appropriateness of level which are the focus of this essay. One way (but by no means the only way) of spelling out this more detailed information is to describe mathematical or logical relationships (e.g., equations) connecting changes in one variable or set of variables to changes in another.
So far my focus has been on causation rather than causal explanation. However, unlike some philosophers, I draw no sharp distinction between providing a casual explanation of an outcome (hereafter the explanandum-outcome) and providing information about the causes of that outcome. According to the interventionist conception, when we provide such causal information we provide information that can be used to answer a what- if –things- had- been -different question: we identify conditions under which the explanandum-outcome would have been different, that is, information about changes that (“in principle, and assuming we were able to perform them) might be used to manipulate or control the outcome. More generally, successful causal explanation consists in the exhibition of patterns of dependency (as expressed by interventionist counterfactuals) between the factors cited in the explanans and explanandum – factors that are such that changes in them produced by interventions are systematically associated with changes in the explanandum- outcome. Other things being equal, causal explanations will be better to the extent that the cited patterns of dependency are detailed, complete, and accurate in the sense of identifying the full range of changes in all those factors (and only those factors) such that, if these were to be changed by interventions, such changes would be associated with changes in the explanandum- outcome. In other words, good explanations should both include information about all factors which are such that changes in them are associated with some change in the explanandum- outcome of interest and not include factors such that no changes in them are associated with changes in the explanandum – outcome. (As we will see below, satisfaction of this feature is related to the notion of proportionality). Moreover, the patterns relating explanans and explanandum should be (in a sense to be described below) stable or invariant under changes in background conditions.
3. Stability and Non-Contingency of Association.
With this as background, I turn first to the notion of stability (also called non-contingency, insensitivity, invariance). Suppose that a relationship qualifies as causal according to M: there is a change in the value of X that when produced by an intervention in some background circumstances Bi is associated with a change in the value of Y and in this sense there is a relationship of counterfactual dependence between the effect and the cause in circumstances Bi. The stability of this relationship of counterfactual dependence has to do with whether it would continue to hold in a range of other background circumstances Bk different from the circumstances Bi. To the extent that the relationship of counterfactual dependence would continue to hold under a “large” range of changes in background circumstances or under background circumstances that are judged “important” on the basis of subject mater specific considerations (see below for more on both these notions), that relationship is relatively more stable; to the extent that the relationship would be disrupted by changes in background circumstances, it is less stable. Stability thus comes in degrees—rather than trying to identify some set of privileged set of background changes that we can use to classify relationships on one side or another of a “stable versus unstable” dichotomy, it is more plausible and better motivated to simply recognize that relationships can be more or less stable or stable under one set of background changes and not another. To the extent that the stability range of a generalization is known, we may help to spell out the content of the generalization by providing such details.
David Lewis (1986) provides an illustration of a relatively unstable (or as he calls it, “sensitive”) causal relationship which (slightly modified by me) is this: Lewis writes a letter of recommendation L that causes candidate X to get a job she would not otherwise have got. This in turn has various other effects: X meets and marries a colleague she would not have married if she had not taken the job, they have children and grandchildren that would not otherwise exist in the absence of Lewis’ letter, these grandchildren do various things A and so on.
Now consider the following claim:
(2.1) Lewis’ writing the letter L caused X’s grandchildren to exist and to do A.
Given the facts just specified, it follows from Lewis’ own theory of causation, as well as the account specified in M, that (2.1) is true. Whether or not we accept this judgment, virtually everyone will agree that there is something non-standard, or misleading about (2.1). Lewis traces this to the fact that (2.1) is highly sensitive or unstable. The counterfactual dependence associated with (2.1) may hold in the actual background circumstances but if these had been different, in a variety of “small” ways, then if Lewis had written the letter, X’s actual grandchildren would not have existed and would not have done A. This might have happened if, for example, X’s future spouse Y had not also taken a job at the same school as X, if other contingencies had led X not to marry Y and so on.
In characterizing the notion of stability, I said that what matters is whether some relationship of counterfactual dependence would continue to hold under a “large” or “important” range of background circumstances. Application of the quoted words depends on several considerations5. One straightforward possibility is that the range of background circumstances under which generalization G’ is stable is a proper subset of the circumstances under which generalization G is stable; in this case, we can at least say that G is more stable than G’ or stable under a larger range of background circumstances. In other cases, we rely on (i) subject matter specific information to tell us which sorts of changes in background circumstances are most “important” for the assessment of stability and/ or (ii) attach particular importance to stability under background circumstances that (again perhaps on the basis of subject matter considerations) are regarded as “usual” or “normal”. As an example of (i), in assessing the stability of gene phenotype relationships we may attach particular importance to whether the relationship is stable under changes in environmental conditions that are “external” to the organism. More ambitiously and demandingly, we may also ask whether the relationship is stable under various changes that might occur elsewhere in the genome.
As a biological illustration, return to Dawkins’ example of the gene R, which is such that when variant r is present, subjects have dyslexia and fail to learn to read (even if the “right” background circumstances are present) but also such that when variant r’ is present, subjects do learn to read, given the “right” background circumstances. Although M agrees with Dawkins’ assessment that R is a gene “for” (i.e., that causes) reading, the relationship of counterfactual dependence between R and whether or not subjects learn to read is itself relatively unstable under various changes in background conditions: change whether primary education is available (or even more dramatically, whether the culture is one in which there is a written language) and whether the subject learns to read will no longer be dependent on whether she possesses r or r’.
Contrast this case with claims about the genes that cause, e.g., eye color or external sexual characteristics. Of course the relationship of counterfactual dependence between possession of a Y chromosome and external sex characteristics depends upon many additional “background conditions” that are involved in sex determination. But although this relationship is not stable under all possible changes elsewhere in the genome or under suitable changes in various other processes involved in development, it is plausible that it is more stable under relevant environmental changes than the R reading relationship (“More” in the sense that to a first approximation, the range of changes in background circumstances in which the R reading relationship is stable is a proper subset of changes under which the relationship between possession of a Y chromosome and external sex characteristics.) Moreover, even though the gene eye color or gene sex characteristics relation requires the operation of many other factors that are internal to the organism and involved in development and gene expression, it is plausible that as long as these remain within some biologically “normal” range, the above relationships will hold; not so for the R reading relationship.
Some readers may balk even at the suggestion that external sex characteristics or eye color are “genetically caused”. My interest is not in arguing about these claims, but simply in observing that they at least seem more natural and less misleading than Dawkins’ claims about the genetic causation of the ability to read. I suggest that differences in the relative stability is one important consideration (but not the only consideration – see below) that leads us to have this reaction. Put slightly differently, my suggestion is that part of whatever resistance we may feel to the claim that R causes reading has the same source as our resistance to the claim that Lewis’ letter causes the existence of X’s grandchildren.
What I take to be a very similar idea is developed by the psychiatric geneticist Kenneth Kendler under the heading of “non-contingency of association”. Kendler (2005) describes a number of different “criteria” (of which non-contingency of association is one, along with “causal specificity” and choice of the appropriate level of explanation) that (he holds) should be satisfied for it to be appropriate to characterize a gene as a “gene for” a phenotypic trait or psychiatric disorder. According to Kendler,
Noncontingent association means that the relationship between gene X and disorder Y is not dependent on other factors, particularly exposure to a specific environment or on the presence of other genes. (2005, 397)
Kendler’s non- contingency condition is a stability or insensitivity condition: a gene disorder relationship is stable or “non-contingent” to the extent that its holding does not depend on the presence of some specific or special environment (with the relationships not holding in other environments) or on whether particular forms of certain other genes are present.
Kendler claims that satisfaction of this criterion of non-contingency is
… a typical (albeit not uniform) feature of genes that cause classical Mendelian disorders in humans (2005, 397)
In contrast, according to Kendler, there is considerable evidence that the effects of specific genes on psychiatric disorders is influenced both by environmental events and by other genes; hence that such relationships are less stable than “classical Mendelian” gene—> phenotype relationships. To the extent this so, it becomes less appropriate to describe these genes as genes for the disorders in question. Although (as indicated above) I find this claim plausible, my primary interest is not in defending it. Rather I put it forward as an illustration of how the notion of stability captures something of biological interest6.
To further explore this notion, consider the connection between stability of a relationship and how proximate or distal it is. Obviously, as a general rule, more distal causal relationships with many intermediate links will be less stable than the individual links themselves. Suppose that we have a chain of causal relationships X1 X2, X2X3…Xn-1 Xn which holds in the actual circumstances B in the sense that each individual link satisfies M in circumstances B and that in addition there is an overall relation of counterfactual dependence in the sense of M between Xn and X1.7 Suppose that X1—>X2 would fail to hold in some set of circumstances B1, X2 X3 would fail to hold in set of circumstances B2, and so on. Then (assuming no additional complications such as backup mechanisms) the overall dependence from X1 to Xn will be disrupted if any one of the circumstances in B1 or B2 or B n-1 holds. So unless B1...B n-1 are strongly overlapping (e.g. most members of B2 are already in B1 etc.) the overall X1 Xn relationship will be less stable than any of the individual links Xi X i+1. Thus to the extent that we value finding stable causal relationships, we will often be able to accomplish this goal by looking for more proximate causal relationships that mediate distal relationships8. It follows that a concern with stability can sometimes (but need not always9) drive us in a “reductive” direction, toward the identification of more fine-grained, “micro” relationships. Note, though, that this does not mean that stability is just another name for how proximate a causal relationship is. For one thing, it is perfectly possible for a distal relationship to be relatively stable (and even no less stable than its individual proximate links) given the right relationship between B1, B2, Bn). More generally, how proximate a causal relationship is seems to be relative to the coarseness of grain in variable description one employs10. By contrast, stability is not representation-dependent in this particular way.
One reason why the stability of a causal relationship matters biologically is that his may bear on the question of how readily the relationship can be altered, whether by processes such as natural selection or by human intervention (the latter consideration mattering for biomedicine and social policy) and on the extent to which this alteration can occur independently of changes in other processes. For example, in eukaryotes the causal relationship between a particular DNA sequence and the pre-mRNA for which it codes is more proximal and also more stable than the DNA sequence –-> mature mRNA sequence relationship since the latter is mediated by the activity of various splicing enzymes. The DNA- mRNA relationship is in turn more stable than the relationship between DNA sequence and yet more distal phenotypical features. (Arguably it is also true the DNA --> mRNA relationship in prokaryotes is more stable than this relationship in eurkaryotes since the former is not affected by splicing agents.) Of course it is true that even the DNA –-> pre-mRNA relationship is not completely stable—it depends on factors like the presence of RNA polymerase and various other cellular features. But the relation between the DNA sequence and its more distal effects in eurkaryotes is even less stable – it depends both on these factors and on more besides such as the activities of various splicing enzymes.
One consequence is that in eukaryotes it may be easier (in the sense that there are more possible changes that will produce this outcome) for natural selection or mutation to alter the relationship between DNA sequence and mRNA than for these to alter the relationship between DNA sequence and pre- mRNA. The former alteration might occur, for example, via changes in the genetic regulation of the activities of splicing enzymes which leave the DNA- -> pre-mRNA relationship unchanged. Similarly, changes in regulation of expression of structural genes can have profound phenotypic effects even though the relationship between the structural genes themselves and the proteins for which they code remains stable11.
Before leaving the notion of stability an additional remark may be helpful in placing this notion in a more general perspective. The issue of whether biology contains “laws” (and if so, which biological generalizations count as laws) has been the subject of a great deal of discussion among philosophers of biology. I won’t try to settle this question here, but two points seem uncontroversial. First, there is an obvious connection between lawfulness and stability: paradigmatic laws drawn from physics and chemistry are very stable generalizations—they hold over a wide range of background conditions. Second, many biological generalizations, including many we think of as describing causal relationships, have somewhat more restricted ranges of stability than fundamental physical and chemical laws—for many such generalizations there are not just nomologically possible but actually occurring, biologically relevant conditions under which they break down or have exceptions. It is an important point that we may ask about the conditions under which such generalizations are (or are not) stable and, as illustrated above, make assessments of their relative stability without trying to settle the difficult question of whether the generalizations are properly regarded as “laws”. In other words, at least some of the concerns that motivate discussions of the role of laws in biology can be addressed by focusing directly on the notion of stability, rather than the notion of law.