1. Introduction

There is a theory about how probabilistic causation works, that says this. If C is a probabilistic cause (henceforth p-cause) of E, then what C does, when it occurs, is to determine the chance that E occurs. Of course, any full theory of p-causation will have to say a good deal more—are C and E to be understood as events, facts, properties, objects or variables, is p-causation a relation between types or tokens, does the relevant relation between C and E have to hold in every causal background context, or just some, and if just some, can it be reversed in other causal background contexts (so that C is a cause of ∼E), and so on. But at the heart of any theory of p-causation is a characterization of the nondeterministic connection between cause and effect, and the idea that C determines, on some versions causes, the chances of E serves as this characterization in many. To quote Housman and Woodward, “probabilistic causation is the deterministic causation of probabilities. By this we mean that X is a probabilistic cause of Y if and only if X is a deterministic cause of the chance of Y, ch(Y), where this is identified with the objective probability of Y” (Reference Housman and Woodward1999, 570).

Call this view, that what makes a p-cause a cause is that p-causes either cause or otherwise determine the probabilities of their effects, the Causes Determine Probabilities thesis, or CDP for short. The most recent defense of CDP of which I am aware is Housman (Reference Housman1998), but Humphreys (Reference Humphreys1989) and Papineau (Reference Papineau1985 and Reference Papineau, Kitcher and Salmon1989) also offer sustained discussions. But the view is so intuitive it hardly seems in need of defense, and it is certainly implicit in, or arises quite naturally from a metaphysically minded reading of, much prior work on p-causation (see for example, Suppes Reference Suppes1970, van Fraassen Reference Fraassen1980, and Salmon Reference Salmon1984). And CDP remains popular; Housman and Woodward (Reference Housman and Woodward1999) feel the need to note the assumption and its frequency in the literature, but not to defend it. Were there an orthodoxy about how p-causation works, CDP would be it.

But for all that, I think CDP is wrong. I think this for a number of reasons, among them at least that it does not accord with much of what we know about how people think and reason about causation, it makes nonsense of a wide range of explanatory practices in the sciences, and it implies that certain perfectly standard and unobjectionable inferential practices are mistaken. Here I want to develop a fourth complaint of a more or less ontological nature.

The objection is this. CDP is consistent with the way scientists talk about p-causation, but not with the way they reason about p-causation, because CDP implies that a particular kind of property simply cannot be causally efficacious. In particular, CDP implies that emergent, population-level causes cannot exist, but we ought regard it as at least empirically possible that such causes do exist. Hence, we ought not accept CDP. The objection requires a good deal of refining, not to speak of defense, and such will be done in the following way. I shall first develop what I take to be some consequences of CDP with respect to singular causation and then describe a phenomenon I shall call “probabilistic overdetermination.” I argue that, given CDP, this phenomenon imposes certain constraints on the identification of emergent causes, which constraints cannot be satisfied by a particular kind of emergent cause, population-level causes. I then close by briefly examining an episode from the social sciences involving causes of this kind. It will be important to note that all of the scientists involved appear to think that the issue is in the end empirical, that is, even those who argue against social causes of the relevant kind seem to think that such causes are possible, but just not adequately evidenced by the data to hand. But, of course, they are mistaken if CDP is right.

2. Singular Causes and Probabilistic Overdetermination

I shall make no assumption about whether causation is, in the first instance, a type- or token-level relation. But in what follows the causal connections of interest are singular; what will be at issue is whether one particular is, or is not, a cause of another particular. So I do assume that even if type-level causal relations are somehow metaphysically primitive, they sustain token-level causal relations, and that, if CDP is right, when type-level causal relations are tokened, the token cause either produces, or is in some sense responsible for, the probability of the effect to be tokened. Much hinges on this, since the natural reply to the considerations that follow is simply to deny that CDP requires any commitment to singular causation at all.

So first, why had CDP better be true about singular causes if it is true at all? Well, on a given analysis, either there are or there are not singular causal facts. If not, I take it that type-level causal facts simply fail to do very many of the things those in the special sciences want causes to do (explain George's cancer, the origin of the Moon, why the WTC towers failed, and so on). Those are singular events, and if they have causes, those causes are singular. That kind of causing is physical, is done by objects, or tropes, or property instances, i.e., some bit of the physical world. Whatever they are, kinds, types and the like are not like that, and hence while they may serve as relata in a causal relation, they do not cause in the relevant sense. Since we clearly can give perfectly good causal explanations of facts of the above sort, and others like them, in which explanations appear to refer to singular causes, we would seem to have very good reason to reject any analysis of causation that implied singular causal claims were simply mistaken. Suppose, on the contrary, that the analysis allows that there are such singular causal facts. Then either the causes are, or they are not, singular causes that, in their occurrence, determine the probability of their effects. Suppose they are not. Then our CDP style analysis of causation mischaracterizes whatever it is about these singular causes in virtue of which they are causes; some causes, namely these, do not, contrary to the claim, cause by determining the probability of their effects. Hence, the last alternative must hold—if CDP is right, then it had better apply to singular causes.

So I take it that on CDP singular causes cause or otherwise determine the probability of their singular effects, and the doing of this is what makes them causes of those effects. This consequence invites an immediate objection. There are contexts in which what appears to be a singular p-cause of an effect cannot be said in any straightforward way to be responsible for a change in the probability of its putative effect. The difficulty arises from the overdetermination of the probability of an effect (henceforth probabilistic overdetermination). Suppose that two nonidentical putative causes, C1 and C2 are such that 1) $\mathrm{Pr}\,(\mathrm{E}\,/ \mathrm{C}\,1) =\mathrm{Pr}\,(\mathrm{E}\,/ \mathrm{C}\,2) =\mathrm{Pr}\,(\mathrm{E}\,/ \mathrm{C}\,1\& \mathrm{C}\,2) $ , and 2) there is no change in Pr(E) between C1 and C2. Then, prima facie, at most one of them can be said to cause E. For to cause E a cause must somehow influence the Pr(E), and this suggests both that Pr(E) must change in some fashion and that this change must be attributable to the presence of the cause.Footnote ¹ Suppose, given conditions 1 and 2 above, that C1 precedes C2. Then there is no change in Pr(E) for C2 to influence; hence C1 and not C2 will deterministically cause (or simply determine) Pr(E) and so will alone count as a cause of E. Matters are complicated, but I think not improved, by the introduction of background conditions that include temporal indices. But this need not concern us here, for the cases at issue will involve contemporaneous putative causes.

Matters are even worse here, for if C1 and C2 are contemporaneous, then arguably neither causes E. For each a version of the following argument is available: C2 determines that Pr(E) = p. Since Pr(E/C1) = Pr(E/C2) = Pr(E/C1 & C2) = p, and C2 determines p, C1 causes no change in Pr(E). Hence, C1 is not a p-cause of E. Mutatis mutandis for C2. One might appeal to some kind of difference in metaphysical status to induce a nontemporal ordering on C1 and C2, and then hold that whichever is the metaphysically prior is the real cause. But then certainly the other is no cause at all, for Pr(E) having been caused to be p by the first, and necessarily remaining unchanged, it cannot be caused by the second. That being unacceptable if in fact C1 and C2 really are causes of E, CDP seems consistent with the causal facts only if probabilistic overdetermination of this kind does not occur.

3. Emergent Population-Level Causes

Unfortunately for CDP, there are certain kinds of causes, taken seriously by scientists, whose existence seems to require probabilistic overdetermination on CDP. I wish to appropriate a term introduced by Elliott Sober (Reference Sober1984), namely, that of ‘population-level cause.’ I shall mean by a population-level property any aggregate property that, strictly speaking, characterizes not any individual but the population to which one belongs, and which is, in a certain sense to be made precise below, derivable from a collection of facts about individual members of a population. For example, the frequency of individuals who smoke, their mean education level, their variances of income, and other features of distributions of individual properties will count, here, as aggregate or population-level properties.

Such population-level properties are supervenient. If, for example, the distribution N(μ = 15,356, σ² = 1003) characterizes the distribution of income in a population P, it does so by virtue of the collection of facts that Igor earns $14,600 dollars, while Ivan earns $85,000 dollars, Boris $30,000, Irina a cool 100 grand, and Svetlana nada, and so on. This collection of facts too, is a distribution, and on it the aggregate population-level facts that income is normally distributed with mean $15,356 supervene, since there is more than one way to realize this normal distribution, even in this particular population (for example, Ivan could earn 30 grand, Boris nada and Svetlana 100 grand). It will be important to distinguish between the collection of facts and the mathematical summary of them. Henceforth, the mathematical entity is a distribution and the collection of facts it characterizes an arrangement.

Population-level properties do more than simply supervene on the arrangements that realize them. Given an enumeration, or any other description of the population that entails the cardinality of the set of all members of the population, and a complete description of the relevant arrangement, a sentence asserting the occurrence of the aggregate population-level property is derivable. I take this to be definitive of population-level properties. Populations may have other properties of which this is not true, but with them I am not here concerned. The properties with which I am concerned, aggregate properties, are reducible to their basis without appeal to any further empirical laws, reduction sentences or the like, and depend on nothing like the kind of metaphysical picture of emergence characteristic of nonreductive physicalism about the mind; neither functionalism nor metaphysical hocus-pocus is needed, here.

But this does not imply that population-level properties cannot be emergent in a sense that goes beyond mere supervenience, for they might, and are sometimes thought, to exhibit independent causal powers. This is not to say that it is ever thought that these powers are unmediated by the actions of individual agents in the relevant populations. But the powers are sometimes thought nonetheless to be had by the aggregate property itself—that is, the fact that the population is characterized by a certain distribution matters above and beyond the sum of the effect of each property instance in the arrangement summarized by the distribution. The phenomenon is familiar: economists talk of inflation rates, GNP, and monetary value causing and being caused; sociologists talk of crime, divorce, and abortion rates causing one another; the examples in biology are legion: densities cause migration rates, harvest rates, and rates of reproduction, the fitness of subpopulations determines rates of survival and reproductive success, differences in fitness cause frequency change, selection operates at the level of genes, and organisms, and, according to some, whole groups of organisms, and on and on. Some of the more interesting cases occur at the intersection of biology and sociology, in the field of social epidemiology.

For example, there are number of social epidemiologists who think that the inequality in income characterizing any given population is a cause of the health achievement characterizing that population (see Kawachi, Kennedy and Wilkinson Reference Kawachi1999). Indeed, the effect of inequality is understood in ways that are apparently consistent with CDP. Kawachi et al. (Reference Kawachi1999) write: “What all of these studies suggest is that income inequality causes a shift in the income/life expectancy curve…, so that almost everyone pays the costs of inequality” (Reference Kawachi1999, xv; italics in original). The shift caused by increasing values of income inequality is exactly a shift in a probability distribution. The second phrase of the quoted sentence is also illuminating. Kawachi et.al. are claiming that income inequality has some effect on the health of individual members of the population. The point is especially clear when Kawachi et. al. write, “What the relative income hypothesis [the claim that inequalities cause health outcomes] means is that an individuals' [sic] health is affected not only by their own level of income, but by the scale of inequality in society as a whole” (Reference Kawachi1999, xvi).

There are two points worth making explicit here. Note first that whatever measure of income inequality is adopted, the property of having a certain level of inequality on that measure is not identical to the arrangement of incomes on which any given instance of it supervenes: the value of the measure is, while its basis is not, multiply realizable. Second, whatever mechanisms mediate the connection between social inequality and individual health, if the claim is right, particular individuals are caused to die the deaths they do, rather than other, different, and importantly later, deaths, by that inequality. So there are people like Boris, say, whose death is the singular causal effect of the social inequality characterizing his population.Footnote ² So whatever complex property serves as the basis for that inequality, the relative income hypothesis holds not that it, but rather the population-level property it realizes, is the cause. The basis may be one among many other possible causes, but, over and above these causes, there remains the inequality itself.

The relative income hypothesis then presupposes that supervenient, emergent, properties can cause, and can cause not only other aggregate, population-level properties, but do so by way of their instances exerting a causal influence on the individual components of the populations in question. By ‘emergent population-level causes’ I will mean causes like this—population-level variables whose values have an effect on the value of other population-level variables either directly or, more interestingly, by way of affecting members of the population, but in any case where the effect of the population-level property is something over and above that of the collection of individual-level properties affecting members of the population. That raises at least two important questions: what is it for a population-level property to exert a causal influence, and how can we discover whether a given emergent property is a population-level cause?

4. Identifying Emergent Causes

As it turns out, such treatments of these questions as exist are extremely topical. But some standards are nearly axiomatic across all serious treatments. I shall locate them in discussion of the units and levels of selection in evolutionary biology and its philosophy, as this offers an explicit consideration of emergent causation in a context congenial to CDP. On the biological accounts (as on nearly everyone's), what matters is probabilistic dependence and conditional probabilistic independence. Just which dependencies and independencies matter varies, of course, among authors. But roughly there are two camps. The first, among whom Elisabeth Lloyd numbers, holds that a higher-level property (organismal or group) can be selected for only if the reproductive success of items at the lower level (genes or organisms, respectively) is not independent of the higher-level property, conditional on all the relevant lower-level properties of that item (see Lloyd Reference Lloyd1988).Footnote ³ So, roughly, the emergent property must not be screened off from the effect by the relevant lower level properties. The other group, among whom Robert Brandon numbers (see Brandon Reference Brandon1990), requires that some system of emergent properties screen off the lower-level properties from the effect. So, roughly, the effect must be independent of the lower-level properties given the emergent property. And of course there is yet a third option, namely, require both conditions.

The basic idea underlying all three proposals, i.e., that a population- level property exerts independent causal influence only if the right conditional dependencies and independencies obtain in the joint probability distribution over emergent property, its basis, and its putative effect, sits comfortably with CDP. Recollect that CDP must avoid probabilistic overdetermination, and further holds that what causes do is determine probabilities. So, the effect E must, in the first place, be dependent on the emergent population-level cause C. C, being emergent, ought to influence the probability of its effect over and above whatever effect C's basis has on that probability, else probabilistic overdetermination ensues. That is, the effect E ought not be independent of C conditional on the arrangement that realizes it. Further, if the arrangement influences E only by way of C, then C ought screen the arrangement that realizes it from E. So, under that condition anyway, E ought be independent of the arrangement realizing C when one conditionalizes on C itself. Letting arrang_pC stand for the arrangement that realizes C in population P, we then have conditions EC1: E ⨿̸ C, EC2: E ⨿̸ C/arrang_pC and EC3: E ⨿̸ arrang_pC/C.

That an emergent property satisfies the third condition, that E ⨿̸ arrang_pC/C, would seem to be less crucial. Suppose, for example, that the effect of individual income on health outcome is partly mediated by inequality, but also has some direct influence. Then health outcome will not be independent of income conditional on inequality values, as is in fact the case. But the first two conditions seem crucial; indeed, given CDP, they must be regarded as at least partly constitutive of emergent causes. If what it is for one thing to be a p-cause of another is for the first to be, in part, responsible for the probability of the second, then surely what it is for an emergent property C, or its instance, to cause some other property E, or its instance, is for the occurrence of C, as opposed to its basis, to be in part responsible for the probability of an occurrence of E, and this requires that, in some context anyway, C not be screened from E by C's basis. So I shall take it that, given CDP, we must hold that a property of objects at a given level of organization is independently efficacious only if that property satisfies EC1 and EC2.

These two conditions are consistent with standard responses to claims about the causal efficacy of emergent population-level causes in social epidemiology. So, for example, the most influential critical responses to the income inequality hypothesis all argue that either EC1 or EC2 fail. Judge (Reference Judge1995) argues that there is in fact no association between measures of inequality and health outcomes, so there is not even an initial dependence between heath and inequality. Fiscella and Franks (Reference Fiscella and Franks1997) argue that when the right measures of individual income are chosen, and Gravelle (Reference Gravelle1998) notes that when the right (nonlinear) relation between individual income and health outcome is recognized, then health outcome is conditionally independent of inequality when one conditionalizes. Hence, say the critics, inequality has no real causal influence.

5. Emergent Population-Level Causes and Overdetermination

This, however, makes the search for emergent causes simply wrong-headed. Let us exemplify. Define arrang_pσ²_I as the collection of facts summarized by the claim that in population P, income is distributed as N(μ_I = 15,356, σ²_I = 1003), and arrang_pE as the collection of facts summarized by a joint distribution over life expectancy and income, with σ²_I being used as a measure of inequality.Footnote ⁴ P then has the property σ²_I = 1003 (henceforth, simply σ²_I). Pick some random person from P, say Boris, and let H represent the state of Boris's health and S represent a collection of other contemporaneous causes of Boris' heath. Then, if the income inequality hypothesis is right, two things need to hold for some system S of variables: first, H must be conditionally dependent on σ²_I given S (H ⨿̸ σ²_I/S). And second, supposing that arrang_pσ²_I is not in S, H must be conditionally dependent on σ²_I given arrang_pσ²_I (H ⨿̸ σ²_I/ arrang_pσ²_I). If the first condition fails, then the inequality in the society is not, relative to that system of variables, a direct cause of H at all: no one is caused by their membership in an unequal society to suffer the health outcome he suffers. If the second condition fails, inequality is not actually making a difference—its basis makes such a difference, but inequality itself is simply along for an epiphenomenal ride.

But the second of the two conditions simply cannot be satisfied. Suppose an emergent property C causes E in a population P. Consider each item i in P and note for each every nonpopulation-level cause of its behavior (i.e., for each i find an S _i that is complete at the individual level). Let S be the collection of such causes, excluding C. Now either S collects all the causes of individual behavior, or not (i.e., either there are, or there are not, emergent population-level causes). Suppose there are not. Then C is not a population-level cause of E.

Suppose however that there are emergent population-level causes. Then note that, C being a population-level property, it is by definition impossible for arrang_pC to occur without C. Hence, E ∐ C/arrang_pC & S. Now either Pr(E/C & S) = Pr(E/arrang_pC & S), or not (i.e., either there is no extra information about E in arrang_pC than simply that C, or there is). If the equality fails, then C is not in fact a cause of E. If it holds, then C is a cause of E only if we reject EC2, and thereby allow the overdetermination of Pr(E), here once by C and once by arrang_pC.

So it seems that emergent population-level causes are possible on CDP only if the overdetermination of probabilities is possible. More precisely, any causal structure in which emergent population-level causes appear will be one on which probabilities are overdetermined. This results from the fact that emergent population-level causes supervene on bases, i.e., arrangements defined as the collections of facts about which members instance which properties, and these arrangements alone suffice to determine the probability of future distributions of properties.

Nor is it just that one has to give up, on CDP, the idea of population-level causes impinging on the behavior of the items that compose the populations to which they attach. If what has been said so far is right, population-level properties cannot cause other population-level properties either. Such properties supervene on lower level arrangements of other properties. Let C be a putative population-level cause, E a putative population-level effect of C realized on this occasion by arrang_pE, and let arrangE be the class of all realizers of E. By the above result, C can cause no property instance in arrang_pE, hence it cannot cause arrang_pE, or any other member of arrangE. Since the Pr(E) must be Pr(arrangE), C cannot influence the Pr(E), and so cannot be a cause of E on CDP.

6. Conclusion

But then we seem to have a problem. Given CDP, there can be emergent population-level causes only if there is probabilistic overdetermination. But for CDP, such overdetermination induces a kind of metaphysical incoherence resolvable only by denying causal efficacy to the emergent cause. At least some of the sciences for which we aim to provide an account systematically engage in practices which can only make sense if emergent causes are at least possible. As a general rule, good philosophical accounts should make out scientists as engaged in scientific silliness only when they are in fact so engaged. There are various reasons for thinking that, for example, the social epidemiologists who think that inequality is a cause of health outcomes are not silly (wrong, maybe, but not silly). So, if alternatives are or can be made available, we ought not adopt an account of causation which forces us to regard these scientists and their practices as silly. CDP commits us so to do, while there are other accounts of p-causation available on which no such commitment follows (e.g., Cartwright Reference Cartwright1989a and Reference Cartwright, Kitcher and Salmon1989b, Cheng Reference Cheng1997 and Reference Cheng, Keil and Wilson2000, and Glymour Reference Glymour1998). Hence, we ought reject CDP.