2.1. Explanation by Analysis and Explanation by Subsumption

For Cummins, explanation by functional analysis is merely a subclass of a broader type of explanation—explanation by analysis. Such explanations are characterized by their explananda. The explanandum of an explanation by analysis is a particular property of a system. By contrast, an explanation by subsumption under a causal law explains why a system changes from one state to another state. These two types of explanations correspond to two types of theories, “transition theories” and “property theories.” A property theory asks what it is for a system to instantiate a property, and tries to answer this question by analyzing the system that has the property. It then explains the property of the system in terms of the properties of its components and their mode of organization. The important point, for Cummins, is that an explanation of a property by analysis is not (and is not to be confused with) a causal explanation. The reason for this is simple: once we understand how a particular property is instantiated in a given system, we know what it is for that system to possess that property. We explain the property by reference to the microstructural properties of the system that realizes the property. In other words, the explanans is token-identical with the explanandum. This rules out that the explanans causally explains the explanandum.

With respect to psychology, the properties of interest are dispositional properties. Understanding what it is for a system to instantiate such properties calls for a particular type of analysis, (i.e., functional analysis). According to Cummins, psychology tries to provide such analyses, because it aims not simply at explaining individual instances of behavior, but rather at showing what it is for a particular type of system to have the disposition for particular types of behaviors. It does so by breaking down the system's behavioral disposition into the behavioral dispositions of the system's component parts. The result can be thought of as a flowchart, where each box represents a functional component of the cognitive system. For example, the human behavioral disposition to exhibit the primacy and recency effect, has been explained by a sequence of processes, which are assumed to take place in different memory stores (Baddeley Reference Baddeley1986).Footnote ¹ These memory stores and their interactions can be represented as a flowchart. In other words, the disposition for the human cognitive system to exhibit the recency effect is token identical with (and explained by) the dispositions of the different memory systems to be related by certain causal mechanisms, given certain triggering conditions.

Now, the “components” of the cognitive system, to which I have been referring, correspond to what philosophers of mind usually call “mental states” or “mental processes.” Within psychology, it is more common to call them “intervening variables,” or entities that intervene between the triggering conditions and the outward behavior, and that are characterized in terms of certain functional characteristics. Such intervening variables form an important part of the theoretical taxonomy of cognitive psychology.

2.2. Can Intervening Variables Be Identified?

Given the notion that explanations by functional analysis explain the behavioral dispositions of a system by appeal to the dispositions of variables to enter into a causal mechanism, the question is what those variables are. A behavioral disposition can be broken down into more than one flowchart. This raises the question of whether there are any criteria for favoring one functional analysis over another. One intuitive answer is that the functional analysis that describes the true causal mechanism should be favored, and that psychologists should proceed with an empirical investigation of what the true causal mechanism is. The question, then, is how to identify the intervening variables in question.

Cummins himself (Reference Cummins1983) is skeptical of the possibility of finding the true intervening variables. Thus, psychological explanations by functional analysis can at best be “how possibly” explanations.Footnote ² He hints at two arguments to this effect. According to the first, explanation by functional analysis does not involve reference to causal mechanisms, therefore the question of which mechanistic explanation is correct is nonsensical. Explanations in psychology are not causal, but “interpretive” (Cummins Reference Cummins1983, 43). However, in light of my above construal of functional analysis, this seems to be incorrect: Just because a property theory provides no causal explanations of the property in question, it does not follow that such a theory does not involve reference to the causal mechanisms that instantiate the property. Furthermore, reference to this causal mechanism explains the output of the system in particular instances when the disposition is instantiated. Cummins provides another argument against the identifiability of the true causal mechanism that instantiates a behavioral disposition. According to this argument, the true mechanism is empirically underdetermined:

It seems that the argument is the following: Psychology explains behavioral dispositions (input-output functions) by analyzing these capacities into component parts, which are themselves dispositions (i.e., little input-output functions). Any input-output function can be calculated by many different algorithms (and executed by a different sequence of intervening variables). Therefore, the correct intervening mechanism is in principle empirically underdetermined.

This second argument for the underdetermination of psychological explanation is shared by some prominent proponents of modern cognitive psychology. For example, John R. Anderson (Reference Anderson1990) distinguishes between two levels of description in psychology, (a) the implementation level, which aims at giving the physical mechanism that mediates between input and output, and (b) the algorithm level, which models behavioral input or output functions. Anderson argues that a psychology that confines itself to behavioral data will have to stop at the algorithm level. But since any mathematical function can be computed by many different algorithms (and executed by many different mechanisms), it is very unlikely that psychology will hit on the true mechanism:

If we take Anderson's argument to be what is behind Cummins's claims, it becomes clear that the conclusion (the empirical underdetermination of cognitive mechanisms) rests on a particular premise. The premise in question is that psychology is confined to behavioral data. It seems to me that this premise is closely related to the assumption of methodological autonomy, mentioned at the beginning of this paper. There are two possible replies to this premise. The first is to accept the autonomy premise, but to argue that it is in fact possible to identify the true intervening variables solely on the basis of behavioral data. The second is to question the assumption of methodological autonomy. This latter strategy is the one that will be pursued in the remainder of this paper. I will proceed in two steps: First, I will present a construal of functional analysis that does not presuppose methodological autonomy (Section 3). Then I will discuss (and reject) some arguments for methodological autonomy (Section 4).

3.1. Searching for the Physical Occupier of a Causal Role

This proposal bears a certain resemblance to some recent accounts of the discovery process in the biological sciences (Bechtel and Richardson Reference Bechtel and Richardson1993; Craver and Darden Reference Craver, Darden, Machamer, Grush and McLaughlin2001). For example, Bechtel and Richardson state that “[a] major part of developing a mechanistic explanation is…to determine what the components of the system are and what they do” (Bechtel and Richardson Reference Bechtel and Richardson1993, 18). The authors make it clear that a hypothesis as to what a component does is going to guide the search for it. They distinguish between two strategies that are in play in the process of looking for the component in question: decomposition and localization. The strategy of (task) decomposition is essentially what Cummins means by functional analysis, i.e., the subdivision of a particular type of behavior into functionally characterized components. This strategy, then, is complemented with that of localization, which consists in the identification of physical parts of the system in which particular functions can be localized. The question is how the two strategies work together. A proposal with respect to this question can be found in Craver and Darden (Reference Craver, Darden, Machamer, Grush and McLaughlin2001, 123). The authors state that a purely top-down analysis of the phenomenon to be explained “cannot itself exhaust the discovery of a mechanism. One also has to know the components of the mechanism and how they are realized” (Craver and Darden Reference Craver, Darden, Machamer, Grush and McLaughlin2001, 123). In other words, a purely functional individuation of components is not sufficient for what they call a “how-actually-explanation.” In addition, knowledge about neurological entities and mechanisms provide “componency constraints,” knowledge about the spatial organization of those components provide “spatial constraints,” and knowledge about the order, rate, duration and frequency of neural processes provides “temporal constraints.” Thus, drawing on knowledge about the entities and processes that potentially realize the causal roles of the hypothetical intervening variables posited by a Cummins-style functional analysis, narrows down the conceptual space of such an analysis. But not only that, it also makes possible two new classes of experiments designed to further narrow down the entities and mechanisms in question. These are experiments that either intervene on the neurological level (by removing certain parts of the brain) and observe effects such a manipulation has on the behavior, or intervene on the behavior level (by setting experimental subjects certain behavioral tasks) and recording nerve activity (see Craver and Darden Reference Craver, Darden, Machamer, Grush and McLaughlin2001, 19ff.).

The authors just mentioned clearly share the intuition that a full account of the workings of a system requires both functional and structural language. Functional decompositions of the behavioral capacities of a system guide the process of looking for physical components and their dispositional properties, but such a search also requires that those components can in principle be individuated independently of a purely functional decomposition.

4.1. Multiple Realizability and Some of Its Implications

The notion of multiple realizability can, in principle, be read in two different ways: (1) as an empirical claim (according to which we find that one and the same phenomenon is realized in several different ways), and (2) as a metaphysical claim (according to which a particular phenomenon, qua being individuated functionally, can be realized in an indefinite number of ways). Fodor (Reference Fodor1974) hints at both versions. In this section, I focus on the first one.

The empirical claim of multiple realizability, too, can be interpreted in at least two ways. A particular input-output function can be multiply realized (a) by virtue of the fact that it is realized by radically different intervening variables, and (b) by virtue of the fact that the same kind of intervening variables are realized by different kinds of physical bases. These two cases are interestingly different, and each of them have been addressed in the recent literature on multiple realization. For example, Keeley (Reference Keeley2000) addresses the second possibility. He presents the case of a behavior of several different species of electric fish, arguing that the functional organization that makes this behavior possible is similar in all three species, despite the fact that these components are not all realized by the same type of biological materials (in some species, the electric organ is realized by nervous tissue, in others by muscle tissue). Bechtel and Mundale (Reference Bechtel and Mundale1999), on the other hand, address the first possibility of multiple realization, arguing that with respect to systems of the brain, there is a lot less interindividual variation than suggested by the common handwaving appeal to multiple realization. If this is the case, and if brain systems serve particular functions, then it would seem that they have to be physical realizers of the causal roles individuated by psychological explanations. Consequently, knowing what brain systems are involved in the production of a particular behavior would help identifying the true intervening variables of a psychological explanation.

4.2. Two Concepts of Autonomy and the Question of Reduction

The above suggests that psychology is not methodologically autonomous. However, this raises the question of whether psychology can provide genuine explanations if it does not proceed autonomously in the identification of its theoretical taxonomy. In other words, can psychology be explanatorily autonomous if it is not methodologically autonomous? I want to argue that psychology can provide genuine explanations, even if it does not arrive at those explanations in an autonomous fashion, and the kinds of explanations it offers are explanations by functional analysis. In the following subsection, I will provide a sketch of an argument for the explanatory autonomy of psychology. In the remainder of this subsection, I will argue that this issue should not be conflated with that of reductionism.

Explanatory autonomy is sometimes taken to be implied by irreducibility. For example, Fodor argues that psychological kinds are multiply realized; hence psychological theories are irreducible, hence psychological explanations are autonomous (Fodor Reference Fodor1997). According to Rosenberg (Reference Rosenberg2001), one of the main points of being an antireductionist is to argue that the special sciences can provide genuine explanations. However, I think I can have explanatory autonomy without committing myself to any stance on the issue of reductionism. In other words, explanatory autonomy may be a consequence of irreducibility, but it does not follow that irreducibility is a necessary condition for explanatory autonomy. The reason why it is desirable to divorce the issue of explanatory autonomy from that of the reductionism/antireductionism debate is that there are many problems in trying to apply traditional models of theory reduction to the relationship between psychology and neuroscience or physics (Rosenberg Reference Rosenberg2001). Two responses have been to either propose improved models of reduction (Kim Reference Kim1999; Bickle Reference Bickle1998), or to call into question the usefulness of the concept (Machamer et al. Reference Machamer, Darden and Craver2000; Machamer and Sullivan Reference Machamer and Sullivan2001). In the present context I am remaining agnostic with respect to this issue. Instead I will focus on the issue of what makes a psychological explanation explanatory.

4.3. Defending Explanatory Autonomy

The issue is put nicely by Rosenberg (Reference Rosenberg2001). One of the problems with reductionism, he says, is that it presupposes the existence of theories and laws in the reduced science. But it is not clear in what sense of the word there are theories and laws in psychology. Hence, reductionism cannot work. But if that is the case, this is also an argument against explanatory autonomy. How are psychological statements explanatory, if not by virtue of expressing a lawful regularity, which is embedded in a well-confirmed psychological theory? The answer, I suggest, is that psychological statements are explanatory by virtue of (a) providing an analysis of what it is for a given kind of system to have a particular behavioral disposition, and (b) providing the mechanism that instantiates the disposition in any given instance. In other words, psychological explanations are explanations by functional analysis.

Now, it might be objected that the explanatory work in such explanations is being done not by the psychology but the neuroscience. In response, I want to concede that perhaps a consequence of not requiring methodological autonomy is that the science we are dealing with is going to be cognitive neuropsychology. Nonetheless, I claim that the explanations are going to be psychological explanations, because they appeal to psychological categories, which are characterized functionally (e.g., memory systems). And for that reason, references to such categories enjoy a certain explanatory autonomy.

To back this up, let us remind ourselves of the second of the two senses of “multiple realizability” mentioned above. To be a function category, by definition, is to be multiply realizable, regardless of whether it is, in fact, multiply realized. This is Fodor's reply to Kim (Fodor Reference Fodor1997, Kim Reference Kim1992). This means that any given functional state can in principle be realized in indefinitely many different ways. Fodor concludes that what we want are explanations that generalize over all these cases. And the reason why we are entitled to treat such explanations as autonomous is that the entities (intervening variables) we invoke are individuated functionally. As Fodor (Reference Fodor1990) puts it, the locomotion of a sailboat can be explained by reference to its possession of a sail (in conjunction with certain environmental conditions). Knowledge of what the sail is made of does not add anything to that explanation. In a similar vein, Kitcher (Reference Kitcher1984) has argued that the details of molecular implementations are irrelevant to the explanatory power of cytological explanations.

However, I believe that in light of my argument about the nature of functional analysis, Fodor's point must be qualified in an important way: Knowledge of what the sail is made of does not add anything to the explanation of why the boat sails. But knowledge that there actually is an entity that performs this function is crucial to the validity of the explanation. In other words, appeal to functional individuated entities is explanatory even if as a matter of empirical fact they are not multiply realized. But such appeal is explanatory only if we have independent evidence for the existence of such functionally individuated entities. On the construal presented in this paper, it is the job of functional analysis to provide such evidence.