1. Introduction

Everyone agrees there is a plurality of competing theories that model selection at different levels of biological organization. For example, some theories model evolutionary processes in terms of selection for individuals, and other theories generate models that represent the same processes in terms of selection for groups. An assumption many scientists and philosophers take for granted is that alternative model-types disagree about the facts. In levels of selection disputes, antagonists often assume that for a given selection process, there must a model that identifies the real level or levels of selection in that process. No one denies that in many cases competing models disagree about important matters of fact, but there is no a priori reason to assume this is true in every case.

In some cases, alternative theories that model the same situations differently might not disagree about matters of fact. A model couched in terms of individual selection and a different model couched in terms of group selection might both represent an evolutionary situation truthfully. Of course, many models misrepresent evolutionary processes, and biologists rightly reject them on empirical grounds. But this leaves open an important question: Are there cases in the levels dispute where models competing for attention are not disagreeing about facts? I, and a number of philosophers and biologists, have argued that the answer to this question is yes.Footnote ¹ Lisa Lloyd challenges this answer. Her spirited critique brings out important issues and raises a second question: If some situations can be individually accounted for by a plurality of true models, as pluralists contend, could scientists nevertheless have rational grounds for choosing some true models over others? I believe pragmatic reasons can exist for preferring one true model over another. Lloyd assumes the opposite.

Pluralism about levels of selection is epistemologically significant because it calls for a less strident form of scientific realism (at least with respect to evolutionary biology). The best models are, the evidence indicates, true representations. Scientific realism holds. But, a true model is not necessarily uniquely true. In some evolutionary situations, no model provides the single right way to represent the causal processes because there is no single right way; instead, there is a plurality of right ways each associated with—to use Helen Longino's (Reference Longino, Kellert, Longino and Watersforthcoming) apt phrase—a different parsing of causes. Scientists and philosophers who believe the proper aim of science is to discover the comprehensive theory that provides the single, correct way to represent the causal structure of the world (or part of the world) need to temper their realism.

In this paper, I develop and defend pluralism about evolutionary biology. In Section 2 and Section 3, I clarify the difference between pluralism and monism. Lloyd challenges my claim that a plurality of models can correctly account for situations, such as the maintenance of the sickle-cell trait, and I defend my claim in Section 4. Using the sickle-cell example, I show that competing theories don't disagree about the existence of ‘high-level’ or ‘low-level’ causes; rather, they parse the causes differently. In Section 5, I explain that pluralism acknowledges the existence of genuine empirical issues in levels debates, but holds that questions about these issues are framed differently depending on the way theorists parse the causes. In Section 6, I argue that not all parsings suit particular epistemic interests equally well, and that the merits of competing parsings can be assessed on pragmatic grounds. In Section 7, I use Jim Woodward's recent work on causation to analyze Sober's distinction between ‘selection of’ and ‘selection for’. The result concurs with Robert Wilson's (Reference Wilson2003) claim: The distinction many assumed was central to the levels debate is irrelevant. I go further and argue that it makes no sense to say the causes are at higher levels and not at lower levels. I conclude by offering a pluralist interpretation of recent work on levels of selection and making a suggestion for future philosophical research.

2. The Difference between Pluralism and Monism.Footnote ²

Lloyd refers to the position advanced in Kim Sterelny and Philip Kitcher's “Return of the Gene” (Reference Sterelny and Kitcher1988), our jointly authored response to Sober (Reference Sober1990)—called KSW—and my dissertation and subsequent paper, “Tempered Realism about the Force of Selection” (Reference Waters1991) as ‘genic pluralism’. I understand why Lloyd suspects that Sterelny and Kitcher (Reference Sterelny and Kitcher1988) wanted to have it both ways. On the one hand, they seem to argue for a strong version of pluralism according to which the genic and group theories are equivalent, while on the other hand, they seemed to argue that the genic theory is superior. I assume she uses the term genic pluralism to emphasize this contradiction. Although Lloyd acknowledges important differences between the position I advanced in “Tempered Realism” and conclusions Sterelny and Kitcher reached in “Return of the Gene,” I think her lumping our views together under the oxymoronic label ‘genic pluralism’ is misleading. Tempered realism is a form of pluralism; it is not a Trojan horse for genic selectionism.

At this point, it would help to define terms. It is useful to have separate labels for the theory that models evolutionary situations in terms of selection for individual alleles, and for the claim that this theory, and this theory alone, provides the correct account of all situations involving natural selection. I call the first, the genic theory, and the second, genic monism. In parallel fashion, I distinguish between the multilevel theory of selection and multilevel monism. The multilevel theory models evolutionary situations in terms of selection processes that can occur at one or more levels. This theory might model one situation as selection for individuals, model another situation as selection for groups, and model a third situation as selection for both individuals and groups. Multilevel monism is not an oxymoron. One might hold for a given situation, there is a uniquely correct model of any given selection situation that identifies the true level(s) of selection in that situation. According to multilevel monism, in one situation the real target of selection might be genes, in another situation the real target could be individual organisms, and in yet another situation, there might be genuine targets at multiple levels. The multilevel monist insists that for any given situation, there is a single correct answer to the question, What is the real level(s) of selection?

Pluralism about levels of selection is the view that monisms about the targets of selection are false. It says that in some situations, an evolutionary process can be modeled correctly as selection for units at one level, and alternatively modeled correctly as selection for units at another level (or multiple levels). This is the view I advanced in “Tempered Realism.” My basic thesis was, and still is, that the causes in evolutionary situations are sufficiently complicated that they can be described truthfully in multiple ways. I argued that the sickle-cell case provided an example of such a situation; the causal processes that maintain the sickle-cell allele can be truthfully modeled as selection for alleles (via the genic theory), and they can be truthfully modeled as selection for allele pairs (via the diploid-genotypic model). Tempered realism is not genic pluralism; it is pluralism plain and simple.Footnote ³

Although I was a consistent pluralist in “Tempered Realism,” Lloyd found passages in KSW that betray a preference for the genic view. The most incriminating passage, which Lloyd astutely identified, but graciously refrained from quoting, stated: “All selective episodes (or, perhaps, almost all) can be interpreted in terms of genic selection. That is an important fact about natural selection” (KSW Reference Kitcher, Sterelny and Waters1990, 160). I believe the first sentence quoted here is true, but I hereby retract the second sentence. In addition to confessing my careless slip into monism, I wish to point out that pluralism about levels of selection does not depend on the first of the quoted claims any more than it would depend on the claim that all selective episodes can be modeled as selection for groups. Pluralism depends only on the idea that in at least some situations at issue, the evolutionary process can be modeled accurately by alternative theories.

It is useful to have a label for the evolutionary situations that pluralists claim can be represented by a plurality of empirically reasonable models, and I call them PERM situations. As I explain in the next section, one could be a pluralist with respect to the question of whether the plurality of models for a specified PERM situation is empirically sufficient in a bookkeeping sense, and a monist with respect to whether the models get something deeper (e.g., the causality) right. The position I'm defending in this paper is pluralism on both bookkeeping and causal accounts.

One might wonder, if my real aim in “Tempered Realism” was to advocate pluralism, why did I devote so much argumentation to defending genic models of selection? Shouldn't a pluralist be evenhanded? Of course, in one sense, a pluralist should not play favorites. But, advancing pluralism to an audience of monists requires showing that some applications of models (which monists reject) can, contrary to what monists believe, accurately represent some of the situations at issue. It was my impression that among philosophers of science, genic selection models needed to be defended because William Wimsatt (Reference Wimsatt and Nickles1980), Robert Brandon (Reference Brandon1982), Elliot Sober (Reference Sober1984), Lloyd (Reference Lloyd[1988] 1994), and others had convinced potential readers that the genic theory was hopeless. If I had been writing “Tempered Realism” for Science instead of Philosophy of Science, and using Maynard Smith (Reference Maynard Smith1976) as a foil rather than Sober and Lewontin (Reference Sober and Lewontin1982), then I would have emphasized that group selection models can truthfully represent some PERM situations. Perhaps biologists would have smeared me with the label ‘group pluralist’.

I reject the label ‘genic pluralism’ and the implication that tempered realism represents some form of genic selectionism or a weaker form of pluralism than that advocated by Lloyd. In fact, it is not clear to me whether Lloyd is a pluralist at all. She seems to be defending multilevel monism.

3. A Different Look at the Basics

Although it is not clear where Lloyd stands on the epistemological issues that motivated me to join the levels debate, prominent advocates of the kind of theory she champions (the multilevel selection theory) are pluralists of sorts. Sober, for example, has consistently maintained that there is a plurality of alternative models that are empirically adequate in the sense that they can do the ‘bookkeeping’. He has acknowledged that genic models can accurately trace changing frequencies and correctly predict what multilevel selection models predict. With respect to bookkeeping, Sober has always been a pluralist.

Sober has recently taken a deeper turn towards pluralism, so it is important to distinguish between the position he advanced with Lewontin (Reference Sober and Lewontin1982) and in The Nature of Selection (Reference Sober1984), from the position he has championed with Wilson in Unto Others (Reference Sober and Wilson1998). Although Sober's position and argumentation have shifted in important ways, he has been consistent on a fundamental distinction between success in bookkeeping and a deeper kind of success (see also Wimsatt Reference Wimsatt and Nickles1980). Success in the bookkeeping sense involves correctly tracing and predicting changing frequencies. The deeper kind of success involves identifying the causes that change those frequencies. Some participants in the blossoming corner of population biology, which Lloyd favorably cites, are (like the early Sober) pluralists with respect to bookkeeping and monists with respect to something deeper. For example, John Maynard Smith made this kind of distinction when he admitted that “equivalent mathematical descriptions are often, although not always, possible” but added that he found “the gene-centered approach both mathematically simpler and causally more appropriate” (Maynard Smith Reference Maynard Smith2002, 524).

I suspect that if empirical success in the bookkeeping sense were the only issue, we would all be pluralists. Hence, for the remainder of this paper I will use monism and pluralism to designate positions about whether a plurality of models in PERM cases represents the causality truthfully.

The issues of serious contention regarding pluralism, then, involve whether a plurality of models aimed at different levels get the causality right, not whether a plurality succeeds at bookkeeping. What is the difference between successful bookkeeping and getting the causality right? David Hull (Reference Hull1980) made an important contribution by clarifying a technical distinction between ‘interactors’ and ‘replicators’ in selection processes. But, what is the difference between identifying a pseudointeractor that merely provides a basis for bookkeeping, and a genuine interactor that is a causal target of selection? Sober offered a clear answer to this question: Good bookkeeping only requires tracing the selection of objects. But getting the causality right requires identifying the properties that are selected for. He used a toy to illustrate the distinction between selection of and selection for.

The toy is a clear plastic cylinder with several different levels and colored marbles. When the cylinder is held right side up, the marbles are all on the bottom level of the cylinder. When the toy is turned upside down and shaken, marbles descend the levels of the cylinder by falling through circular holes in the floor of each level. When marbles descend no further, some marbles have come to rest at lower levels than others. Green marbles descend to the bottom. Yellow marbles rest one level up from the bottom, and so on. The marbles are of different sizes: marbles that reach the bottom level are smallest (and are green); marbles that rest at the next level up are of the next-to-smallest size (and are yellow); and so on. The circular holes in the floors of each level are also of different sizes. The hole in the floor leading to the bottom level has the smallest circumference; the holes in the levels above have increasingly larger circumferences.

The toy is a selection device. It selects green marbles to descend to the lowest level, yellow marbles to descend to the next level up, and so on. One can model this selection in terms of color and derive conclusions about size (e.g., spherical circumference) from the correlation between the colors and sizes of marbles. This model, I'll call it M _C, gets the bookkeeping right. One can use M _C to trace descents and make predictions about where marbles of different colors (or derivatively, where marbles of different circumferences) will rest when the device is turned one way or the other. But Sober claims that M _C represents the causality wrong. Colored marbles are selected by the selection device, but color is not the causal target of selection. The causal target is circumference. I call the model based on selection for spherical circumference, M _SC. In Sober's useful terminology, there is only selection of colored objects whereas there is selection for spherical circumference. The model that gets the causality right is M _SC, not M _C. Partisans might defend M _C on the basis that one can use this model to derive information about selection of spherical circumference, but Sober could respond that the empirical success of this derivation does not show that the model gets the causality right. I agree. Both M _C and M _SC are empirically adequate in the sense that they get the bookkeeping right, but only M _SC identifies real causes.

Where I disagreed with Sober's earlier positionFootnote ⁴ is on the question of whether the difference between M _C and M _SC illustrates the difference between, say, George Williams’ (Reference Williams1966) genic account of the selective maintenance of the sickle-cell trait, and the conventional diploid-genotypic model of the same process. To clarify the point of contention, consider a third model of the toy's selection process, one based on the selection for spherical diameter (rather than spherical circumference or color). On the basis of this model, one can derive information about the selection of color (because of the correlation between spherical diameter and color), and one can derive information about the selection of spherical circumference (because of the equality $c=\pi d$ ). Consider the plurality of three models: M _C, M _SC, and M _SD. All three are empirically adequate in the sense that each gets the bookkeeping right. But M _C clearly misrepresents the causal process. What should we say about the difference between M _SC and M _SD? Does one represent the causes of selection truthfully and the other misrepresent the causes? Of course not. It is not as if the real target of selection is circumference, rather than diameter, or vice versa. M _SC and M _SD both model the causal situation accurately. In some contexts they might compete for attention, but it is not the case that one is true and the other is false. In some contexts there might be good pragmatic reasons to favor one model over the other (e.g., it might be easier to measure circumference than diameter), but this doesn't imply that one is semantically derivative of the other. I contend that the difference between the genic and diploid-genotypic models of the sickle-cell situation, to take a concrete example, is more akin to the difference between M _SC and M _SD than to the difference between M _SC and M _C.

4. How Alternative Theories of Selection Parse Causes Differently

Selection models draw conceptual divisions between environments and selected domains. On one end of two extremes, multilevel theories draw this division between conspecific groups of organisms and everything outside the groups. On the other end, genic theories of Williams’ style (Reference Williams1966) draw the division between an individual allele and everything outside the individual allele. Which entities’ properties are modeled as being selected for sometimes depends in part on where biologists draw this division. This is one way evolutionary theorists parse causes.

I illustrated this point in “Tempered Realism” by considering the maintenance of the sickle-cell trait. The conventional population model of this situation is the diploid-genotypic model, but Williams (Reference Williams1966) argued that it is correctly accounted for in terms of genic selection. Monists countered that the correct causal account is provided by the diploid-genotypic model, not the genic model (e.g., Sober and Lewontin Reference Sober and Lewontin1982). I argued that both models described the causal processes accurately. Since Lloyd challenges my argument, it is necessary to review the case. I start with the genic model, not because it is uniquely right or better, but because it is the one that Lloyd attacks. Next, I describe the diploid-genotypic model that Lloyd associates with the multilevel theory. Then, I defend both models and show where Lloyd's critique goes astray.

4.1. The Genic Model of the Sickle-Cell Situation

Williams’ genic theory draws the environmental divide between an individual allele and everything outside the individual allele.Footnote ⁵ Hence, if we assume there are two allele-types at the locus (designated below as S and A), then selection is occurring in two different genic environments, and there are two distinct selection processes.

One selection process occurs in environment _S, where there is an S allele at the corresponding locus. In environment_S, the form of hemoglobin causing the sickle-cell condition (which I call S-hemoglobin) is synthesized because the environmental allele is S and the presence of this allele causes the synthesis of S-hemoglobin.Footnote ⁶ In environment_S, A is viable because it leads to the synthesis of regular hemoglobin molecules and this molecule participates in a host of processes in environment_S that leads to A's survival and replication. The S allele is not viable in environment_S because it causes the production of S-hemoglobin, and in environment_S the synthesis of more S-hemoglobin leads to this allele's demise. Hence, there is selection for A and against S in environment_S. In the genic model

where $W^{S}_{A}$ is the selection coefficient for allele A in environment_S, and $W^{S}_{S}$ is the selection coefficient for allele S in environment_S. In this model, A and S are selective interactors with respect to environment_S. Notice that explaining the selective advantage of A in environment_S draws upon information about causal processes occurring in environment_S. Although application of this model depends on environmental information, the causal processes occurring in the genic environment are not modeled in the genic theory as selection processes. The genic theory parses the causes so that selective interactions are between the allele and its genic environment.

Parallel considerations can be made with respect to the other genic environment, environment_A. Explaining why S is selected for, relative to A, in environment_A requires taking into account causal processes in the genic environment involving the synthesis of regular hemoglobin, malaria viruses, mosquitoes, and human living conditions. The genic model does not deny the importance of such causal processes. It simply parses the causes such that these processes are modeled as processes within the environment, not as processes between selective interactors and their environment. In the genic model

where $W^{A}_{S}$ is the selection coefficient for allele S in environment_A, and $W^{A}_{A}$ is the selection coefficient for allele A in environment_A.

Since A alleles participate in two separate selection processes, calculating the overall change in the number of alleles in a human population requires adding the changes of allele number in each environment. Williams does the math by taking a weighted average of the selection coefficients. As Williams explains:

If the gene associated with anemia and malarial resistance is designated S, its selection coefficient in the genetic environment S [environment_S] would be very different from its coefficient in environment S’ [environment_A]. Its effective (mean) coefficient would be the mean for these two environments, weighted by the frequencies of the environments. (Williams Reference Williams1966, 60)

Although this is not the conventional way to model the situation, the kind of ‘spatially-dependent’ modeling Williams is using here is part of conventional population genetics. Following Williams, we can designate the effective selection coefficients as follows:

where $P^{A}$ is the relative frequency of environment_A, and $P^{S}$ is the relative frequency of environment_S. Notice that $\overline{W}_{A}$ is not a basic causal parameter in this model; rather, it is a mathematical device for calculating the net effects of the basic parameters. The basic parameters are selection coefficients $W^{A}_{A}$ and $W^{S}_{A}$ .

4.2. The Diploid-Genotypic Model of the Sickle-Cell Situation

The diploid-genotypic theory draws the environmental divide between diploid genotypes (or perhaps organisms) and their ecological environments. Parsing the causes this way yields a model in which there is only one environment (instead of two) and three types of selective interactors (instead of two). There is a selection coefficient for each interactor type:

1. $W_{SS}$ is the selection coefficient for diploid genotype SS in the diploid environment.

2. $W_{SA}$ is the selection coefficient for diploid genotype SA in the diploid environment.

3. $W_{AA}$ is the selection coefficient for diploid genotype AA in the diploid environment.

The relative values of the selection coefficients are

Explaining why SA is selected for, relative to AA, requires taking into account causal processes in the ecological environment involving viruses, mosquitoes, and human living conditions. The diploid model does not deny the importance of such causal processes. I repeat the point I made with respect to the genic model: The diploid-genotypic model parses the causes such that these processes are modeled as processes within the environment, not as processes between an interactor and its environment.

Explaining why SA is selected for, relative to SS, in the diploid environment also invokes information about alleles and the syntheses of different forms of hemoglobin molecules. The diploid model does not deny the importance of these causal processes. It just parses the causes such that these processes are modeled as occurring within an interactor, rather than between an interactor and its environment.

Calculating the change in the relative number of diploid genotypes in the environment also depends on a mathematical device that takes a weighted average of the diploid selection coefficients. This mathematical device does not represent a basic parameter in the model. The basic parameters are the selection coefficients $W_{SS}$ , $W_{SA}$ , and $W_{AA}$ .

5. Two Different Kinds of Issues in the Units of Selection Debate

Lloyd gets lots of mileage quoting rhetorical flourishes in KSW, to the effect that “quibbling about the real unit of selection” (quoted in Lloyd Reference Lloyd2005, 288) is a waste of time because it is “an exercise in muddled metaphysics” (quoted in Lloyd Reference Lloyd2005, 290). Such remarks might make it sound as if we thought there were no empirical issues in the levels debates. This is misleading. Pluralism does not deny the existence of genuine, empirical issues. What it denies is the metaphysical interpretation that monists read into them.

The t-allele situation provides a clear example of a PERM situation in which important issues are at stake. (See Waters Reference Waters1991 for details.) For genic theorists, the issue is whether the genic environment of the t-allele is heterogeneous with one selection process occurring in spaces where t-alleles are contained within female mice in groups with all sterile males, and another selection process occurring in spaces where t-alleles are not contained within such females. For multilevel theorists, the issue is whether there is selection against groups of females in demes containing all sterile males. In a way, it seems to be the same issue. This makes sense because the questions concern the same causal situation. When the causes in this situation are parsed one way, the issue emerges as a question about whether there are multiple selection processes in a heterogeneous environment. When the causes are parsed differently, the issue emerges as a question about whether group selection is occurring. In “Tempered Realism,” I stressed that the empirical issues are as important to the genic theorist as they are to the multilevel theorist. The rhetorical flourishes were not intended to impugn these substantive questions.

Sober and Wilson insist that the following is a genuine question: “Can traits evolve by benefiting whole groups, despite being selectively neutral or disadvantageous within groups?” (Reference Sober and Wilson2002, 530). I agree. But, we should keep in mind (as Sober and Wilson point out) that this question is framed within the multilevel theory. It is a question that seems to fall between the cracks of the genic framework. The genic theorist does not ask about the causal situation by framing her question in terms of group selection; instead she asks her questions in terms of whether there are multiple selection processes in a heterogeneous genic environment. The exercise of answering such questions framed within the multilevel (or genic) framework is empirically meaningful. The exercise becomes muddled only when monists ask, with the exclusive sense of ‘or’: But is it really group selection in the ecological environment or multiple selection processes in a heterogeneous genic environment?

Some cases in dispute (e.g., the t-allele case) do involve important empirical issues, but others, such as the sickle-cell case, do not. Why do monists care about the sickle-cell case? Monists seem wedded to the Platonic ideal that science aims to cut nature at its joints. It appears that some philosophers joined the levels debate to clarify conceptual issues with the expectation that biologists could then use empirical methods to identify the real joints of natural selection. The pluralist view, on the other hand, is that there are multiple ways to divide nature, and that some questions are empirically meaningful only with respect to particular causal parsings.

6. The Grounds for Choosing among Competing Models in PERM Cases Are Pragmatic

Scientific models provide partial descriptions. They highlight some features and obscure others. In simple situations, models may highlight all causally relevant features and obscure the rest. In complicated situations, however, such as biological development and evolution, models typically make some causal features salient by obscuring others. This doesn't render models false, but depending upon scientists’ interests, it makes them more or less suitable, pragmatically speaking. In some cases, different interests might favor different model-types; hence, there could be pragmatic reasons for retaining a plurality (see Kellert, Longino, and Waters Reference Kellert, Longino, Waters, Kellert, Longino and Watersforthcoming).

Readers might wonder whether this is consistent with the claim I made in KSW to the effect that alternative models in PERM situations are representationally equivalent. Lloyd rightly criticizes KSW for making such claims. Kerr and Godfrey-Smith (Reference Kerr and Godfrey-Smith2002) have offered the clearest account of the relationship between contextual models (e.g., the genic model of heterozygotic superiority) and group models (e.g., the diploid-genotypic model of heterozygotic superiority), and they prove that the models are mathematically equivalent. But Lloyd is absolutely correct in saying that mathematically equivalent models are not necessarily representationally equivalent.Footnote ⁷ The mathematics does not exhaust the information of applications of these models (as I read Kerr and Godfrey-Smith, they wouldn't disagree). I retract all claims about representational equivalence in KSW.Footnote ⁸

Lloyd suggests that pluralism implies that there can be no rational grounds for choosing among genic and higher-level models. I disagree. There can be pragmatic reasons for preferring one true model over another. I remained neutral on pragmatic issues in “Tempered Realism” because drawing conclusions about pragmatic advantages and disadvantages would require a different kind of analysis. But, since Lloyd's critique brings up pragmatic issues as well as semantic ones, I admit my bias: The genic theory has a number of pragmatic disadvantages. It often requires cumbersome conceptions of environments, obscures causal details, and conceals issues concerning adaptedness, which are of interest to biologists. These are pragmatic disadvantages. They do not indicate that genic models are semantically derivative of higher-level models, that genic models misrepresent causal processes, or that genic models fail to identify the true targets of selection. But depending upon biologists’ goals, these pragmatic disadvantages could provide rational grounds for preferring higher-level models.

7. Why the Claim That Selection Is for Higher-Level Traits and Not for Lower-Level Traits Is Nonsense

Arguments about levels typically invoke causal claims. It's time to take a closer look at what it means to identify a cause. Sober argued that models of selection should identify genuine causal targets. In “Tempered Realism,” I showed that the criterion that Sober and Lewontin (Reference Sober and Lewontin1982) applied to determine whether a model correctly pinpointed ‘real’ targets of selection could not establish their conclusions (Sober has subsequently abandoned his former line of argumentation). I also argued that the conception of probabilistic causation upon which their criterion rested didn't capture the sense of causation appropriate for understanding natural selection. At the time, I didn't have an alternative proposal. Now I do.

Over the past 15 years, philosophers have made significant progress developing a theory of causation that can help elucidate the idea that some selection models identify genuine causes and others identify pseudocauses. The theory I have in mind is the manipulability theory that has been set out in a wonderfully clear and careful manner in Woodward's recent book, Making Things Happen (Reference Woodward2003). According to this theory, the distinguishing feature of causal claims is that they provide the basis for answering ‘what if’ questions about what would happen if certain properties were manipulated in specified ways. Woodward formulates his account, not directly in terms of properties, but in terms of variables that take different values. He writes,

According to this theory, causal relationships involve patterns of counterfactual dependencies.

The manipulability theory involves a number of subtleties that I will not discuss here. But, one subtlety is crucial for our purposes. It is not the case that any manipulation of the value of X that subsequently changes the value of Y implies that X causes Y. For example, a manipulation of the thermometer reading brought about by decreasing temperature would change the phase state of water, but this does not mean that thermometer readings cause water to freeze. To say that the thermometer reading causes water to take particular phase states is to say that a special kind of manipulation of the thermometer reading, a manipulation that didn't independently affect other causal variables such as temperature, would change the phase state. The technical term for the special kind of manipulation is ‘intervention’. Woodward carefully spells out what counts as an intervention, but for our purposes we just need to keep in mind that the manipulability theory says that ‘X, rather than V, causes Y' roughly means that an intervention in the value of X would change the value of Y, whereas an intervention in the value of V would not change the value of Y.

Although Woodward's theory is not reductive, and although there might be different senses of causation underlying different instances of reasoning that are typically called ‘causal’, Woodward's theory helps elucidate the difference between ‘selection of’ and ‘selection for’.

Let's take a closer look at the toy illustration. I identified three different models of the selection process, each of which, modeled the process as selection for a different property: M _C (for color); M _SC (for spherical circumference); and M _SD (for spherical diameter). Following Sober, I appealed to a basic intuitive judgment when I claimed that M _SC represents the causes correctly, while M _C misrepresents the causes. But, what is the basis for our judgment? The answer, according to the manipulability theory of causation, is that the causal models disagree about what would happen if the system were manipulated in certain ways. M _C's answers to the corresponding ‘what if’ questions are wrong, and M _SC's answers are correct. Consider the following ‘what if’ question:

According to M _C, color is the difference-making property, and the intervention on color (changing it from green to yellow) would result in the marbles descending to the level of the other yellow marbles. Changing the color would count as an intervention only if the change did not independently affect other causal factors. According to M _SC, however, a genuine intervention on color (i.e., a manipulation that did not also change spherical circumference) would not affect how far the marbles descend. This claim could be tested by painting the green marbles yellow (perhaps the procedure would include grinding down the marbles before painting them so the manipulation wouldn't change the marbles’ circumference). Of course, the painted marbles would still descend to the bottom level. According to the manipulability theory of causation, this captures what it means to say that M _C provides a false account of the causal process.

The reason why we think M _SC represents the causes correctly is that we believe the level to which the green marbles descend by intervening on the circumferences of the marbles. This claim could be tested by manipulating the circumference of the green marbles (partisans of M _C might insist that we do so without changing their color). For example, we could increase the circumference of green balls by adding green clay to their surfaces. Now consider the following question:

According to M _SC, circumference is the difference-making property and this intervention on circumference would result in green marbles coming to rest at the same level as the yellow marbles. According to M _C, the change in circumference will have no affect on descent and the marbles will descend to the bottom level. Of course, M _SC answers the question correctly.

The pair of experimental results relating to the pair of ‘what if’ questions would show that the selection process depends on circumference, not color. This illustrates how, according to Woodward's manipulability theory, the idea that M _SC gets the causal process right and M _C gets it wrong involves beliefs about counterfactual dependencies concerning what would happen if particular properties were manipulated in special ways.

The manipulability theory shows why it would not even make sense to claim that M _SC identifies the true causes and M _SD does not. For such a claim would mean that an intervention on circumference would affect descent, but an intervention on diameter would not. But, it is impossible to intervene on circumference without also intervening on diameter. It is not that we lack the technological means to intervene on one without intervening on the other. The problem is that it is impossible to change one without changing the other, because of the geometrical relationship between circumference and diameter (c = πd). Hence, according to the manipulability theory, it simply makes no sense to claim that M _SC identifies the true cause and M _SD does not.

I claim that in PERM situations, the difference between individual models and multilevel models is more like the difference between M _SC and M _SD than the difference between M _C and M _SC. I explain why by returning to the case involving the sickle-cell trait. Recall that Williams claimed that the genic model M _G identifies the real cause of maintenance of the trait, while Sober and Lewontin (Reference Sober and Lewontin1982) argued that the diploid-genotypic model M _DG identifies the real cause (Sober has dropped this argument, but Lloyd persists). A monist would assume that one or the other partisan is correct. A pluralist would hold that both partisans might be correct and look for an argument to decide whether one, the other, or both models represent the causal situation truthfully.

Let's assume, for the sake of a reductio argument, that the monist is correct and either M _G or M _DG (but not both) gets the causality right. According to the manipulability theory, the meaning of the competing causal claims involves the following ‘what if’ questions:

The monist interpretation requires that the answer to first question be the outcome of the selection process would be different and the answer to the second be the outcome would not be altered (or vice versa). But, there is a problem. It is impossible to change fitnesses of the diploid allele pairs without also changing contextual fitnesses of the individual alleles (as shown by Equations 1–4 in Kerr and Godfrey-Smith Reference Kerr and Godfrey-Smith2002, 482–483). Likewise, contextual fitnesses of the individual alleles cannot be changed without changing fitnesses of the diploid allele pairs. These impossibilities do not arise because of contingent biological facts. They arise from the mathematical structures of the models. It is similar to the situation described above where spherical circumference cannot be changed without altering spherical diameter. The difference between M _G and M _DG is akin to the difference between M _SC and M _SD. (I say akin rather than identical, because applying M _G requires making explicit information that is not necessarily made explicit in applying M _DG and vice versa. Applying these different models requires bringing in different information, whereas applying M _SC and M _SD involves bringing in the same information.) Hence, claiming that M _G identifies properties that are selected for, while M _DG does not, (or vice versa) is, literally speaking, an exercise in muddled metaphysics.

Monists about levels of selection might resist this argument by denying that causation is the issue, or by claiming that Woodward's theory of causation is mistaken. To those who would deny that causation is the issue, I ask: What is the issue, if not causation? If the issue is just a matter of determining which models are empirically adequate and identifying practical advantages of different bookkeeping systems, then what Lloyd takes to be the key issue, the ‘interactor issue’, is a pseudoissue. Denying the importance of causal claims in this debate is a strongly antirealist position. As a causal realist, I wonder what such antirealists would say about Sober's selection device. Is M _C as true a description as M _SC? If not, why not?

Monists might acknowledge that causation is the issue, but claim that the manipulability theory is mistaken, and that my argument in this section is based on a false premise. (I have structured this paper so my earlier arguments do not depend on the manipulability theory.) One problem with the levels debate is that many disputants have been unclear about what they mean by cause, causal interactor, or causally appropriate (on this point, Lloyd and I agree). If monists seek to protect their epistemology by claiming the manipulability theory is mistaken, then the burden is now on monists to clarify causal notions such as ‘interactor’, ‘force of selection’, and ‘selection of’ versus ‘selection for’ in a way that doesn't undermine their position.

Wilson (Reference Wilson2003) also draws on the linkage between properties at different levels to argue that Sober's distinction between ‘selection of’ and ‘selection for’ is irrelevant to the levels debate. Wilson's interests are mainly ontological, concerning whether causes in the world are entwined, fused, or leveled. My interests are epistemological, concerning whether we should interpret the best scientific theories, as (a) potentially capturing the one true way to represent a domain, or (b) potentially capturing one of perhaps a plurality of true ways. As I read Wilson, he is claiming that the selective causes are so entwined in PERM situations that they cannot be disentangled at distinct levels, as multilevel theorists seek to do. This may be true, and if so, it is a significant point. But it is not my point. My point is that PERM situations are sufficiently complicated (or entwined) that there are multiple ways to disentangle them. I think it is an open question whether multilevel theories can be developed to sort out causes at distinct levels. I believe that just as there is a plurality of (correct) ways to sort out selective forces at single levels for particular situations (e.g., at the genic level or at the diploid-genotype level for the sickle-cell situation), there is a plurality of ways to sort causes at multiple levels. Different sortings may have different pragmatic advantages and disadvantages.

The manipulability theory provides a clear way to understand why the distinction between ‘selection of’ and ‘selection for’ is valid for distinguishing between selection models that identify true causes from those that do the bookkeeping by identifying pseudocauses. The theory also shows that the notion that selection can be for properties of groups—and not be for properties of the individuals making up the groups—does not make sense in situations where one could not, in principle, change a group property without changing an individual property (a basic condition of supervenience that all or nearly all philosophers of biology accept).

Sober's distinction between ‘selection of’ and ‘selection for’ is irrelevant to considerations about levels of selection. If the argument in this section is correct, notions widespread in philosophy of science (and especially philosophy of psychology) about identifying the real levels of causation are inherently confused.Footnote ⁹

8. Conclusion: Guidance from the Past for Future Research

Pluralism denies neither the existence of empirical issues in the levels debate, nor the idea that there can be rational grounds for making decisions among a plurality of truthful models. What it denies is what monists read into those issues and decisions. Multilevel monists assume the debate hinges on identifying the uniquely correct model of any given situation. There is a plurality of models about some evolutionary situations, they believe, because biologists disagree about facts of the matter. Decisions among the different models of a given situation are decisions aimed at choosing the one true model and rejecting the false ones.

But not all decisions affecting the content of science reduce to decisions about choosing the true theory and rejecting the false ones. Hans Reichenbach put the point this way:

Reichenbach said that an important task of epistemology is to identify volitional stipulations. He distinguished between two classes of volitional decisions. He said ‘conventional decisions’ are between ‘equivalent’ conceptions, and do not influence the content of knowledge. He cited decisions to choose a particular system of weights and measures as conventional: The choice concerns “different ways leading to the same place” (Reference Reichenbach1938, 10). Reichenbach said the other class of volitional decisions involve choices that can lead to divergent systems of knowledge. He called these ‘volitional bifurcations’.

Some writers call pluralism about levels of selection ‘conventionalism’. This label is misleading because it suggests that decisions between genic and multilevel theories have trivial consequences for the practice of evolutionary biology. The decisions between competing theories of selection actually fall under Reichenbach's category of volitional bifurcation. In cases of bifurcations, Reichenbach said philosophers should investigate the consequences of volitional decisions and elucidate entailments such as “If you choose this decision, then you are obliged to agree to this statement, or to this other decision” (Reference Reichenbach1938, 16). It is time to stop quibbling about how to identify the real level of selection, and start investigating the pragmatic consequences of different volitional decisions about where to draw the environmental divide and how to parse and disentangle the causes.

Some of the most interesting current philosophical research on levels of selection issues can be understood in this way. Samir Okasha (Reference Okasha2004), for example, has been exploring different statistical approaches for assessing models of PERM situations. His investigation, which presupposes a decision to draw the environmental divide where the multilevel selection theory does, explores the consequences of different volitional decisions about how to disentangle the causes. Sober and Wilson (Reference Sober and Wilson1998) can be understood in a similar way. Okasha argues that one approach is ‘theoretically preferable’. As a tempered realist, I interpret this conclusion in terms of pragmatics.

The manipulability theory suggests another way to assess alternative parsings. Kerr and Godfrey-Smith have started exploring the idea that some theories (or ‘parameterizations’) might turn out to be better than others in terms of accounting for perturbations.Footnote ¹⁰ As Kerr is developing this idea, some parameterizations might localize causation in certain perturbations more narrowly than others. Kerr's intuitions correspond to ideas of ecometricians who believe that causal models are better insofar as they identify autonomous parameters. Roughly speaking, a parameter is said to ‘autonomous’, if it can be manipulated without changing the value of other parameters in the model (Woodward Reference Woodward and Little1995). One might compare different selection models by examining whether the parameters within each model are autonomous in various situations. It is not clear where such comparisons will lead, but this approach (which is aligned with the manipulability theory of causation) might provide a practical basis for assessing models that parse causes differently, while still avoiding metaphysical muddles such as questioning whether maintenance of the sickle-cell trait is caused by selection for alleles in genic environments, or by selection for diploid genotypes in ecological environments.