2. NO REVOLUTION NECESSARY

We reject the idea that neuroeconomics should be seen as a revolution or a “shift of paradigm” (See Glimcher Reference Glimcher2003: 1; CLP: 55; see also Harrison's objections in this volume.) Neuroeconomists should not promise a revolution because it is too early to bet. The question of whether neuroeconomics is revolutionary is therefore not nearly as important as the question of how the science should structure itself to maximize its potential for building new and better models and for attracting and retaining converts and graduate students. Neuroeconomics should not aim to effect a revolution, we argue, because it is more promising at this stage to pursue collaborative and integrative research projects.Footnote ⁵

To highlight the contrast between revolutionary and integrative research strategies, consider three of the core features of Kuhnian revolutions.Footnote ⁶ First, they are successional; one paradigm vanquishes and replaces another. Galileo and Copernicus did not collaborate with Aristotelians physicists and Ptolemaic astronomers to forge a new scientific perspective on the solar system; rather one global system for doing and thinking about astronomy eliminated the other (or at least that is Kuhn's narrative). Second, the rival paradigms in Kuhnian revolutions are incommensurable. They do not use the same techniques and vocabularies. They address different problems. They have different standards of evidence. They, in Kuhn's provocative phrase, inhabit different worlds. It is because the paradigms are incommensurable that revolutions must be eliminative (cf. Churchland Reference Churchland1989). Finally, and consequentially, the choice among rival paradigms cannot be grounded in any standards of evaluation mutually acceptable to members of rival paradigms. Rarely does one paradigm vanquish another by convincing the practitioners of the old to accept the new; the scientists of the old simply fail to attract students and, eventually, die. The crucial point is that revolution so conceived is in no sense a collaborative endeavour between neighbouring fields. All revolutions are victories.

Our mechanistic view of neuroeconomics differs from Kuhn's classic vision in three respects. First, we see neuroeconomics as integrative rather than successional. In this respect, a successful neuroeconomics would develop in the way Darwin and others developed the theory of evolution by natural selection, or as Watson and others worked out the genetic code. Darwin's theory integrates facts about breeding, geology, natural history, and Malthusian studies of population growth. Watson and Crick integrated diverse findings within chemistry (such as details from X-ray crystallographic findings; Chargaff's rules; basic facts about acids and bases; details about charges and chemical bonds) to elucidate the genetic code. These scientific triumphs are not victories. Second, the mechanistic model emphasizes collaboration over incommensurability. Researchers in disparate cultures build lines of communication and undertake joint projects aimed at a common collaborative goal. True incommensurability precludes collaboration; the researchers occupy different worlds and accept different standards of demonstration and evidence. Finally, in our mechanistic model, the abstract standards for evaluating causal and mechanistic explanations generally (irrespective of the details concerning any particular mechanism) are common intellectual commitments that guide the development of collaborative projects involving more or less isolated fields. The common goal of building a mechanistic model serves as the touchstone guiding the elaboration of mutually satisfactory standards for doing science and so provides the basis for communicating across the cultural divide between fields. The rhetoric of revolution emphasizes barriers to, rather than possibilities for, communication between disparate fields of science.

We sketch this mechanistic model and discuss some of its implications for the future of neuroeconomics in the final section. To see why the model is appropriate, however, it is necessary first to get clear about the goals served by economic models and, second, to see why neuroeconomists are more likely to achieve those goals if they aim at describing mechanisms.

3. GOALS OF ECONOMICS: PREDICTION, EXPLANATION, AND CONTROL

What are the goals of economics? The relevance of neuroeconomics depends on how one answers this question. And indeed economists disagree about what we should expect from economic theory. Some ways of articulating the goals of economics and the status of economic theory – in particular the revealed preference view of utility (henceforth, the RPV) and the instrumentalist conception of economic theory – run counter to our vision of interfield integration. It is not that such visions of economics are intrinsically flawed. Rather they lie at the foundation of a poor strategy for the future development of economics. In this section, we explain these interpretations and give reasons to reject them.

Our argument can be summarized as follows: Suppose that the goals of economics are prediction, explanation and control. These goals are achieved better when economics aims at the discovery of mechanisms that underlie economic phenomena than when it aims merely at producing instrumental “as if” models. These mechanisms include facts about decision-making that can be and, in some cases, have been illuminated by psychology and neuroscience. Although knowledge of neural or psychological mechanisms of decision-making is not strictly necessary for providing explanations (or making good predictions, or exercising control, for that matter), the search for mechanisms is nonetheless a good strategy for building models that achieve those goals. Interpretations of economics that impede the study of mechanisms (as do, in our view, the RPV and the instrumentalist conception of economics) should not be adopted if the goal of economics is to predict, explain and control actual phenomena.

First, we explain the motivations behind the RPV and in favour of instrumentalism. Then we give reasons to think that a mechanistic interpretation is superior to the RPV if one wants to predict, explain, and control decision-making in individuals or markets. We do not claim that economic phenomena must be explained at a neural level, or any other level in particular. Rather the claim is that economics should aim at some sort of mechanistic explanation, at least in the long run, as opposed to models whose assumptions are treated as mere instruments for prediction (and derivation) and not as describing components and causal relations among them.Footnote ⁷

Consider first the RPV. This view drives some of Glenn Harrison's most pressing objections to neuroeconomics. In particular, Harrison uses the RPV to argue that the discovery that people are not motivated by the search for pleasure has no bearing on economics (Harrison Reference Harrison2008). Here's why.Footnote ⁸

The RPV is an interpretation of utility and preferences, the basic building blocks of models in microeconomics. Proposed in 1930s at the height of behaviourism and operationalism, the view sought to free economic theory from all reference to mental states such as preferences or utility in the sense of some psychological quantity of goodness. On the RPV, choices ‘reveal’ preferences – what one chooses, given certain constraints, is what one wants. In fact, as Alex Rosenberg points out, the view is a misnomer, because once preferences are identified with choices there is nothing left to reveal – talk of preferences is eliminated in favour of the talk of choices (Rosenberg Reference Rosenberg2007: 87). Of course, these choices need to satisfy certain consistency conditions – if you choose both a over b and b over a when a and b are available, then, unless you are indifferent, no respectable ranking can be constructed out of your choices. But if your choices are consistent in a particular way, then the utility function ‘revealed’ by them need not be interpreted as referring to a psychological quantity that motivates or explains your choices. Instead it is a formal representation of your choices.

Behaviourism and operationalism are no longer considered reasonable constraints on theories, neither in philosophy, nor in sciences of the mind. But in economics, the RPV has not gone away. One of its modern proponents argues for the RPV on the grounds that economists are ignorant of psychological mechanisms of motivation and decision making and it is best not to interpret economic theory as making any claims about those mechanisms (Binmore Reference Binmore2007). On the contrary, we believe this is a reason to adopt a tentative attitude toward economic models, but not a good reason to interpret them in a way that insulates them from empirical criticism.

How does the RPV shield economic models from criticism? It does so by claiming that models that purportedly describe relationships between utilities and expectations (or desires and beliefs), on the one hand, and choices, on the other, in fact describe relationships between choices (that is what utilities are) and other choices. So whatever a psychologist or a neuroscientist discovers about the human motivational system is irrelevant for the interpretation of economic models.

Some philosophers criticize the RPV on the grounds that it precludes the intuitive idea that mental states explain behaviour (for example, Hausman Reference Hausman1992; Rosenberg Reference Rosenberg2007). They also point out that the RPV cannot distinguish between irrational choices and changes in taste, such as when an agent's choices do not look consistent over time simply because the agent has reordered her preferences. Others argue that the RPV cannot accommodate the possibility of choices that violate one's preferences, as when one decides against one's best interest on the basis of morality or strategy (Sen Reference Sen1977).

Finally, others argue that the RPV robs the economic theory of its normative force. If utility is nothing but the choices one makes, then how can we make sense of the idea that it is rational to make choices that maximize utility? Both the explanation of choices and the recommendation to make these choices rational are virtually tautologous on RPV. (See Hausman Reference Hausman1992 for a full discussion of these criticisms.)

We lay aside these challenges to the RPV and argue instead that the RPV cedes the explanatory ambitions of economics. We show that those who endorse the RPV face three unattractive options. They can (1) relinquish explanation as one of the goals of model-building in economics, (2) maintain that economic models are explanatory by virtue of their predictive adequacy only, or (3) maintain that economic models are explanatory according to Woodward's difference making account (Woodward Reference Woodward2003). The first option is transparently unsavory and, further, sacrifices the ability of economic models to serve as effective bases for intervention and control. The second option relies on a faulty view of explanation. The third option, although it relies on a compelling account of explanation, forces the supporters of the RPV to embrace shallow explanations that do not admit of empirical improvement. We think none of these options is acceptable and hence encourage economists not to use RPV to shield economics from psychology and neuroscience.

Some economists endorse the second option. Milton Friedman (Reference Friedman and Hausman1953) reconciles RPV and the explanatory value of economic models by equating explanation and prediction.Footnote ⁹ For Friedman, predictive adequacy is the only goal of economic theorizing. Throughout his seminal paper he uses the term “explanation” in quotation marks to indicate that explanation means prediction. For example: “Viewed as a body of substantive hypotheses, theory is to be judged by its predictive power for the class of phenomena which it is intended to ‘explain'” (Friedman Reference Friedman and Hausman1953: 184). Friedman argues that economists should build models that predict well in the intended range of phenomena, not that describe mechanisms or otherwise allow us to understand economic phenomena.

Friedman insists further that a model's “importance” or “significance” is inversely related to the realism of its assumptions. By importance and significance, Friedman means predictive capacity or the number of true predictions a model makes: “A hypothesis is important if it ‘explains’ much by little, that is, it abstracts from the common and crucial elements from the mass of complex details and permits valid predictions on the basis of them alone” (Friedman Reference Friedman and Hausman1953: 188).

What is the intended range of phenomena? The appropriate domain of economic theory is not individual decision-making. Hence Friedman argues, “of course, businessmen do not actually and literally solve the system of simultaneous equations” (Friedman Reference Friedman and Hausman1953: 193). Indeed their actions could be “habitual reaction, random chance, or whatnot” (ibid.). The relevant prediction of the maximization of returns hypothesis is not how businessmen make decisions; rather, it is their actual behaviour in the face of external circumstances. So if psychologists or neuroeconomists demonstrate that people do not have the preferences or do not use the decision-making rules that economists assume they do, that is irrelevant for evaluating economic models. Friedman thus claims that it is inappropriate for the purposes of developing economic theory to concentrate on examining and improving upon such standard assumptions of neoclassical economics as “perfect competition”, “perfect monopoly”, and whether businessmen do indeed reach their decisions by consulting their cost and revenue curves (Friedman Reference Friedman and Hausman1953: 188).

Clearly, economic theory interpreted in accordance with RPV can make predictions about economic phenomena, and so, on Friedman's view, it is capable of explaining those phenomena. In his criticisms of neuroeconomics, Harrison at times embraces the idea that economic theoryFootnote ¹⁰ need not be taken literally and instead should be read with an “as if” proviso. He also claims that economics need not discover mechanisms behind phenomena to be explanatory (Harrison Reference Harrison2008: 31). Even recognizing that sometimes standard economic theory fails to predict correctly in its intended domain of application, a Friedmanite instrumentalist rejects the need to learn about the psychological or brain mechanisms of judgment and choice. Our concern is (a) that Friedman's is not a viable view of explanation and (b) that it is not a view consistent with many of the explanations that economists routinely offer.

In the last few decades, philosophers of science have keenly debated the nature of scientific explanation. Even though there is no agreement on what explanation is, there is a lot of agreement on what it is not. When Friedman wrote his famous article, the Covering Law model of explanation dominated the philosophy of science (Hempel Reference Hempel1965a, Reference Hempel1965b). On this view, explanation is expectation; it is nothing more and nothing less than the derivation of a description of the phenomenon to be explained from premises describing laws of nature (understood as universal generalizations) and the initial and background conditions. This account is now almost universally rejected, at least as an exclusive and exhaustive account of explanation. To subsume a phenomenon under laws of nature is insufficient to explain it. It is universally true that all men who take birth control pills fail to get pregnant, and one could derive John's lack of pregnancy from statements describing his diligent consumption of birth control pills and this universal generalization, but the birth control pills are explanatorily irrelevant (Salmon Reference Salmon and Salmon1971a, Reference Salmon1971b). It is also possible to subsume a phenomenon under laws relating effects to causes (such as the length of a shadow and the height of the flagpole), while most people believe that phenomena are explained by their causes, not their effects. The height of the flagpole explains the length of the shadow, and not vice versa (Bromberger Reference Bromberger and Brody1966). It is also unnecessary to subsume a phenomenon under laws in order to explain it. We routinely accept explanations for why things happen while having no knowledge whatsoever about the relevant laws explaining how the things came about. We know that the knee's knocking on the table explained the inkspot on the floor even though few among us could list all of the relevant conditions and laws by virtue of which the stain came to be where it is (Scriven Reference Scriven, Feigl and Maxwell1962). The ability to explain without showing that the event was to be expected on the basis of laws of nature is especially evident in probabilistic explanations (which Glimcher Reference Glimcher2003 claims will be crucial for adequate explanations in neuroeconomics and neuroscience). Improbable things have causes and explanations though they cannot, by definition, be shown to have been expected on the basis of laws of nature (Scriven Reference Scriven, Feigl and Maxwell1962; Railton Reference Railton1978).

Currently the most popular alternatives to the Covering Law model all emphasize that explanations reveal causes of phenomena or the mechanisms that generate phenomena (see, e.g., Bechtel and Richardson Reference Bechtel and Richardson1993; Craver Reference Craver2007; Machamer et al. Reference Machamer, Darden and Craver2000; Salmon Reference Salmon1984; Woodward Reference Woodward, Kitcher and Salmon1989 and many others). Scientists might have an incomplete understanding of these causes and mechanisms, but to attempt to explain a phenomenon is precisely to attempt to learn what brings the phenomenon about. One variant of this view is Jim Woodward's difference-making account: roughly, x explains y if and only if had we intervened to prevent x from happening, y would not have happened (Woodward Reference Woodward2003). The advantage of a causal mechanical approach in general, and Woodward's interventionist version in particular, is that it unifies two of the most important goals of science – explanation and control.

Clearly, Friedman's idea that explanation is co-extensive with prediction falls prey to many of the problems that cast grave doubts on the tenability of the Covering Law model. If the Covering Law model is the only view that allows the RPV advocate to say that economics (interpreted in accordance with RPV) can provide explanations, that would be bad news since the Covering Law model fails many of the tests of genuine explanations. However, this is not right. RPV and Friedman have other options.

Consider now the third of the above options. On this option, one allows that economic models might explain even if they do not identify the correct underlying ontology. Models involving phlogiston or gravitation might capture important explanatory truths about burning or falling without correctly describing the underlying components and activities by which things burn or fall. In like fashion, one might interpret economic models instrumentally and in accordance with RPV, but insist that such models do support the sort of counterfactuals that are necessary to make claims about how certain variables make a difference to certain others. Take, for example, a game theoretical model that claims that rational agents with privately known valuations, under conditions of certainty, plus sealed-bid first-price rules of bidding, will bid below their true valuation. On the RPV, we could read the model as making the following claim: if these agents did not have such valuations (defined as choices) they would not make such bids. However, so long as the choices that reveal preferences are not identical to the choices that are caused by these preferences, the model can be read as making a claim about how agents would behave in a variety of counterfactual circumstances: if we intervene on one set of choices, we can change another set of choices.

Indeed this result is consistent with Woodward's theory according to which to explain is to be able to answer what-if-things-had-been-different questions about the system that one is modelling. One can give such answers even if one does not know all the details of the underlying mechanisms, even if one does not couch the explanation in terms of entities at a deeper level of organization, and even if one is wrong about the ontology of that underlying explanation (Woodward Reference Woodward2003: 232–3). So on a difference-making account of explanation, economic models interpreted in accordance with RPV still count as explanatory, even if they are relatively shallow in the sense that they allow one to answer fewer important questions about how things would work in different circumstances than one could answer about the system being modelled (see Hitchcock and Woodward Reference Hitchcock and Woodward2003). So does this mean that RPV and instrumentalist interpretations of economics can, in fact, satisfy the requirements of explanation, prediction and control?

While such models sometimes offer dim and cloudy sketches of the structure of causal dependency that must be identified in successful explanations, instrumentalism and the RPV unnecessarily insulate economics from progress in developing less dim and less cloudy models, that is, models that capture more of that causal structure more precisely. Our view makes use of Woodward's (Reference Woodward2003) insights about the deep connection between manipulation and explanation as a crucial component in mechanistic explanation (see Craver Reference Craver2007). The search for more detailed understanding of the mechanisms underlying economic phenomena potentially allows one to answer a greater range of questions about what would happen if things were to be changed. Not only do mechanistic models include more measurable variables (by virtue of including variables describing the intermediate mechanisms), but the search for mechanisms is likely to call one's attention to testable differences between how-possibly models of a mechanism, and so to ever more subtle distinctions in predictions made about how people behave under a variety of circumstances. Dim and cloudy models of the causal structure underlying such behaviours, such as those that assume agents obey strict rationality assumptions and those that treat preferences as mere choices, do gesture at the causal structure of decision-making. This is why they are sometimes predictively and, to a first pass, explanatorily useful.Footnote ¹¹ However, it is possible to be more accurate about how decision-making in fact proceeds, and an excellent way to do so is to think about the causal structures, the mechanisms, by which agents make decisions.Footnote ¹²

So while we are willing to admit that Harrison's favoured interpretation of economics (in accordance with RPV and Friedman's instrumentalism) allows for some explanation, prediction, and control, we do not think that these are good interpretations if the goal is to improve the predictive and instrumental value of economics in the future (more on that record later). Understanding psychology and the brain is a better bet for such improvement. Option three, in sum, is best understood as a half step in the direction of more predictively adequate and instrumentally useful mechanistic explanations of economic phenomena.

Harrison, or another advocate of the RPV and instrumentalism, might reply that the dream of discovering causal and mechanistic explanations in economics is a philosopher's fantasy – economics in its current state is unable to live up to this impossible ideal. “As if” claims and predictions are all that economists can reasonably expect. But this reply flies in the face of the actual practice of economics.

Economists do provide successful causal explanations and do construct reliable mechanisms. The growing field of design economics (Roth Reference Roth2002), which currently comprises the centre stage of empirical application of microeconomics, attests to this fact. When economists are called to construct reliable incentive-compatible institutions – for example, an auction that distributes spectrum licenses to telecommunication firms – they design this auction by testing a number of different causal claims. A specific kind of auction under specific conditions, which include agents’ beliefs and desires, cause a specific distribution of goods and revenue generation. Crucially, in design economists’ accounts of how they use theoretical models, instrumentalist “as if” interpretations are conspicuously absent (Plott Reference Plott1997; Guala Reference Guala2005; Alexandrova forthcoming). Such an attitude does not fit the goal of design economics to create mechanisms whose behaviour we can more or less understand and control. Granted models are treated as very tentative and merely suggestive. But they are nevertheless suggestions as to what causal factors may be relevant to one effect or another, and some of these causal factors are preferences and expectations understood psychologically, not behaviouristically. Neither models, nor the explanations that design economists construct, treat the underlying mechanisms as complete black boxes. So contrary to Harrison's claims (Harrison Reference Harrison2008: 31), instrumentalism is not the prevailing philosophy among practicing economists.

Perhaps design economists are mistaken; they are not entitled to any causal mechanistic claims. But if this is so, then Harrison is similarly not entitled to propose or endorse perfectly sensible causal explanations, for example, an explanation of the well-documented apparent change in people's discount rates. The explanation emphasized by behavioural and neuroeconomists is hyperbolic discounting. An alternative explanation calls into question whether experimental subjects believe the experimenters’ promise to deliver the goods in the future. Harrison argues that this explanation has not been ruled out (Harrison Reference Harrison2008: 16–17). Maybe so, but this is not an instrumentalist attitude.

The biggest problem with formulating the goals of economics in formal instrumentalist terms is not so much philosophical as strategic. Instrumentalism isolates economics from contributions from other disciplines. A related problem arises for the current trend among neuroeconomists to appeal to Marr's three-level structure as a vision for the future of neuroeconomics. Marr's levels include a computational level, which specifies the problem to be solved, the algorithmic level, which specifies an algorithm for solving it, and an implementation level for the hardware on which the algorithm is run. Marr emphasized that each higher level is formally independent of the levels beneath it (Marr Reference Marr1982). Many different algorithms can solve the same computational problem, and many different hardwares can implement the same algorithms. One can continue to investigate at the computational level independently of findings about the hardware that implements it. For this reason, in the philosophy of mind, Marr's vision has been used to argue for the irrelevance of neuroscientific evidence to the theory of mind (no matter how much Marr's practice in fact violated that very assumption). In our view, it does not follow from the fact that disciplines can work independently from each other that they ought to. The value of interdisciplinary integration rather than autonomy is that it brings constraints from multiple independent perspectives to bear upon a single phenomenon and so gives it a kind of robustness or claim to reality not shared by phenomena that are not detectable from multiple independent perspectives.

Consider, for example, how Harrison uses RPV to argue that neuroeconomists’ bold claims reflect a fundamental misunderstanding of the nature of economic theory, which, when corrected, undercuts the significance of neuroeconomists’ claims. Findings of neuroscience and psychology, neuroeconomists emphasize, can be interpreted to suggest that the brain uses two separate though overlapping systems: one for “hedonic evaluation of stimuli” and a separate one for “motivation” (CLP: 37). If so, it is possible to be motivated to seek an object without judging that it is a greater source of pleasure than another. What bearing does this have on the central assumption of economic theory that agents choose courses of action that maximize their expected utility? Neuroeconomists claim it shows that that motivation is not grounded in pleasure seeking. However, whether or not this is relevant to economic theory depends on how utility is interpreted: interpreted as pleasure, the finding undermines the central assumption; interpreted as satisfaction of desires that may or may not be pleasurable, the finding does not bear on it; interpreted as choices, as the RPV teaches and as Harrison appears to prefer, the finding could never bear on it. We agree with Harrison that the finding does not necessarily undermine standard economic theory. But we disagree that one should choose the interpretation that precludes revision of this theory in the light of empirical findings about mechanisms that implement choice. In service of what end would one close off the possibility that one might improve the predictive or instrumental utility of one's models by studying the mechanisms by which human beings make decisions and the environmental factors to which those decision-making mechanisms do and do not respond?

It is no news that standard economic theory makes inaccurate predictions in a number of contexts. For example, game theoretical models of pretty much anything, but auctions in particular, fail to predict the bidding behaviour of real flesh and blood bidders in one-shot interactions, i.e. when the bidders are inexperienced with the interaction. The endowment effect, whereby people begin to value an object more merely because they own it, not because they derive more utility out of it, is extensively documented (Thaler Reference Thaler1980). So is its inconsistency with the central assumption of consumer theory that willingness to pay for an object should be equal to willingness to accept compensation if deprived of that object. Examples ranging from consumer behaviour, to strategic interactions, to many other areas of economics, abound. Given the predictive failures of such standard economic models, economists cannot accept Friedman's permission to ignore the falsity of their models’ assumptions. This permission is granted only to predictively successful models. How better to improve these models than through careful investigation of the psychological and brain mechanisms underlying decision-making?Footnote ¹³

Consider another example. Some neuroeconomists argue that the brain is unlikely to contain a single “reasoning” module; rather, it is likely that the psychological category, “reasoning”, is a dim and cloudy notion that in fact describes the behaviour of several distinct and more-or-less domain specific modules. Expected utility theory is naturally interpreted as a theory of decision-making according to which the agent first forms preferences about different outcomes represented by a cardinal utility function, then forms beliefs about probabilities of these outcomes, and then picks an act that maximizes expected utility. Psychologists argue that this account of decision-making is probably false as a descriptive theory because it fails to account for many biases, heuristics and automatic processes that underlie choice and decision-making.

Is this relevant to economic theory, which relies heavily on expected utility theory for its various models? Harrison does not think so. He gives two reasons. First, economics is about “human capital and compensating wage differentials”, rather than about reasoning. And secondly, the domain-generality of reasoning is assumed merely for convenience and is not intrinsic to economics (Harrison Reference Harrison2008). Both reasons appeal to Friedman's idea that not all predictions of economics are on a par when it comes to testing it because not all predictions are within the intended range of the theory. We agree that sometimes it can be acceptable to ignore complexities,Footnote ¹⁴ but we urge that this should not be a general strategy of economics. The fact that some presuppositions of economics are not central to its subject matter does not change the fact that these presuppositions are problematic if the goal of the discipline is to give causal explanations and to progressively improve on the quality of those explanations.

One can grant Harrison and other critics of neuroeconomics and behavioural economics that there exists an interpretation of economic theory under which the findings of these fields (or any other fields for that matter) are irrelevant to economics and hence their significance is overblown. The question is whether this is the right interpretation. We doubt it for a number of reasons: some philosophical, having to do with the nature of explanation; others strategic, having to do with how economics should best position itself in the light of the evident progress in sciences of mind and brain. The RPV and Friedman-style instrumentalism work jointly to insulate economics from possible contributions from other disciplines. In doing so, these views erect a wall of Kuhnian incommensurability between different fields. But this can only be a reasonable long-term strategy if one believed, as Gul and Pesendorfer do, that “rationality in economics is not tied to physiological causes of behavior” (Gul and Pesendorfer Reference Gul and Pesendorfer2005: 24), that is, sciences of mind and brain can never bear on social sciences. But this is evidently false (how, otherwise, could advertisers be successful at influencing consumer behaviour with various psychological and physiological tricks?) or at least presumptuous. So it is best to keep economics open to contributions from other disciplines, especially sciences of mind and brain.

4. NEUROECONOMICS AS A MECHANISTIC SCIENCE

In our view, neuroeconomics should be conceived as an attempt to discover multilevel mechanisms to explain regularities in decision-making. The goal is to discover the variables (internal and external) that are relevant to explaining how people behave in different sorts of tasks and to understand how the values of those variables depend upon one another and are organized together such that they give rise to the regularities in decision-making behaviour. In neuroscience specifically, the goal is to identify the entities that are involved in these decision-making processes (the brain systems, brain regions, neurons, neurotransmitters, receptors, and so on), and to identify the activities in which those entities engage (manipulation of representations, release of neurotransmitters, binding of agonists) and how they are organized together.Footnote ¹⁵ A mechanistic neuroeconomics would proceed by searching for the mechanisms in the brain (body and world) that give rise to, sustain, modify and regulate, etc. decision-making behaviour.

Much of what is called neuroeconomics at the moment is dedicated to a very early-stage effort to sketch hypotheses about how certain decision-making mechanisms work, about which areas of the brain are involved in which decision-making tasks, and about how variations in tasks produce different patterns of brain activation. Ideally the goal is to learn which environmental variables are causally relevant to which decision-making mechanisms, and one wants to know enough about how the decision-making mechanisms work to be able to say how the mechanism will behave under a wide variety of conditions, including interventions onto the system (typically, but not always, on environmental conditions). To know how the variables depend upon one another, one needs to know how interventions to change the values of the variables in the mechanism change and do not change the values of other variables in the mechanism (see Pearl Reference Pearl2000; Woodward Reference Woodward2003). Controlled causal experiments are designed to test such relations. Our view appears to fit nicely with McCabe's idea that economists study agents’ strategies while neuroscientists try to infer neural mechanisms (McCabe Reference McCabe2008: 6). The strategy is just the “function that the mechanisms are performing” (McCabe Reference McCabe2008: 7), and that can be understood as the phenomenon to be explained by the mechanism.

Mechanisms frequently span multiple levels. That is, components and activities at one level can often themselves be decomposed into organized entities and activities at a lower level, which can themselves be decomposed into organized entities and activities at still lower levels. (Relations among one set of variables can be redescribed to include more or fewer intermediate variables; the redescriptions thereby descend and ascend through levels). Researchers collaborate with one another in building such models by placing constraints on mechanisms at different levels. Researchers in one field (e.g. those who study brain systems by studying brain-damaged patients) study the gross organization of brain systems; others (such as those who use functional imaging techniques) have something to say about which locations in the brain are active during different tasks and the conditions under which those brain regions become active; others investigate neurotransmitter systems and the molecular receptors on which they act. Researchers at different levels use different techniques to place constraints on the space of plausible mechanisms for a given phenomenon. The effort to build such a multilevel model enjoins researchers to coordinate their research with work done at other levels in the same mechanisms, and this places additional constraints on any successful model. As a general sketch: the behaviour of populations might be explained in terms of the aggregation of or interaction among individual agents, whose behaviour can be explained in terms of cognitive mechanisms, which are in turn explained in terms of underlying interactions among brain regions, cells, molecules, and so on.Footnote ¹⁶ The search for multilevel mechanisms scaffolds the integration of fields.

There are three reasons neuroeconomics should be a mechanistic science. First, as noted above, the rest of neuroscience, cognitive science, and biology have adopted a largely mechanistic stance (see Darden Reference Darden2006; Craver Reference Craver2007; Bechtel Reference Bechtel2008). The search for mechanisms provides a common goal toward which researchers in different fields can contribute. Anatomy, biochemistry, cytology, developmental biology, electrophysiology, evolutionary theory, molecular biology, systems neuroscience all have something to add to an understanding of mechanisms operating at multiple levels. Economics might usefully be viewed as another perspective among these,Footnote ¹⁷ adding one more kind of tile to the mosaic unity of neuroscience.

Second, the search for mechanisms is an attempt to understand the causal structures that give rise to more or less stable regularities in decision-making. While it is possible to have predictively adequate theories that do not capture accurately the causal structure of the world (Reiss Reference Reiss2007), generally models that allow one to infer how interventions will change the behaviour of the system will include facts about the system's causal structure: about what the relevant variables are, and about how they are causally related to one another. Mechanisms are how things work. If one knows how things work, one is potentially in a position to intervene into the system to change it for good or for ill (see Woodward Reference Woodward2003). In other words, if one wants instrumentally useful models, one should search for models that describe mechanisms (see Craver Reference Craver2007).

Finally, even if it is possible to build predictively adequate theories that make no reference to underlying mechanisms (think for example of Snell's law, Hooke's law and Balmer's formula), one way to build more predictively adequate models is to posit underlying mechanisms. The search for mechanisms requires one to search for subtle differences in the predictions made by different models and for new experiments to bring those subtle differences out. One strategy for building more predictively adequate theories is to conceive and evaluate different models of underlying mechanisms and to test those models against the predictions of the model. This is not a failsafe strategy for generating more predictively adequate models. It might be, for example, that the hidden variables are too numerous and too varied to be included in models while still delivering models that are predictively adequate and/or manageable for predictive purposes. All we can say in response is that if there are hidden variables that are usefully included in the models or useful for clarifying when those models apply (and when they do not), then searching for mechanisms is one way to find them. It seems likely to us that there are such variables for economic theory, and so we think it would be fruitful to look for mechanisms. (To see some ways that this might be worked out in varying detail, see Woodward Reference Woodward2002; Craver Reference Craver2007; Steel Reference Steel2008.)

In short, if neuroeconomists would like to follow in the footsteps of other successful neuroscientists, if they want to build instrumentally useful models, and if they want to formulate predictively adequate models, then they have good reason to pursue a mechanistic research programme.

5. CO-EVOLUTION THROUGH INTEGRATION

While it is inappropriate to think of neuroeconomics as a revolution, we nonetheless believe that the effort to construct multilevel mechanisms in neuroeconomics will almost surely require some adjustments in the models and techniques of the parent fields if their models are to be integrated usefully. Patricia Churchland (Reference Churchland1986) refers to this process as coevolution (see also Bechtel Reference Bechtel and Bechtel1986, Reference Bechtel1988; Craver Reference Craver2007: Ch. 7). In the most radical cases, coevolution is Kuhnian elimination. But there are several varieties of coevolution short of revolution.Footnote ¹⁸ A review of recent literature in neuroeconomics shows evidence for the possibility of coevolution in neuroeconomics that is driven both from the top down (in which economics has the power to transform our neuroscientific models) and from the bottom up (in which neuroscience forces one to revise economic models).

From the top down, economics offers neuroscience fresh perspectives on the regularities in decision-making behaviour. More specifically, it offers neuroscience a library for taxonomizing economic decision-making behaviour, a set of tasks for manipulating and measuring aspects of that behaviour (from behavioural economics), and a set of representational and theoretical tools for describing and interpreting the results of such experiments. The cognitive faculties involved in these tasks are “central” faculties (such as evaluation, ranking and choice) rather than “peripheral” faculties (such as visual motion processing or control of saccadic eye movements). It is no trivial accomplishment to develop a set of tasks for measuring central cognitive phenomena precisely, a theoretical vocabulary for describing what those tasks assess, and sets of equations for describing the results of those measurements and allowing one to draw inferences from them. The tasks that economists have developed over the years to assess bargaining, price formation, cooperation, and preference reversals, for example, can be used as tasks in neuropsychological and neuropsychiatric experiments to determine which brain mechanisms are involved in such decision-making tasks.

For example, consider how economists determine a subject's preferences. Experimental economists have developed several elicitation procedures, the most famous being the Becker–deGroot–Marschak (BDM) mechanism. This procedure asks subjects to name their reservation price for a lottery and then the lottery is auctioned off, with the proceedings going to the subject (if the auction price is higher than the subject's reservation price), or else played. BDM creates a simple task that makes it beneficial and easy for the subject to state his correct preference. Experimental economics has similar procedures for operationalizing other theoretical concepts. This library of phenomena, to use Francesco Guala's expression (Guala Reference Guala2005), and the measurement procedures that go with it, should be a welcome addition to the neuroscientist's toolbox. Moreover, these phenomena are often described in terms of intentional categories, that is beliefs and preferences, which can function as constraints on neuroscientific explanations.Footnote ¹⁹ Clearly neuroscience potentially stands to benefit from the influx of well-characterized and quantifiable tasks for testing central decision-making mechanisms.

Viewed in the most radical light, one might see this top-down perspective as revolutionary. The thought is that economics provides a unique view of the nature of animal behaviour and so how one should break the mechanisms underlying such behaviour into components. Consider an analogy. Franz Josef Gall argued that the mind was composed of distinct faculties located in specific regions of the brain. His list of faculties is composed mostly of skills on which two individual subjects might vary independently. His faculty psychology includes artistic talent, amorousness, love for one's offspring, and other faculties that sound funny to the contemporary ear. He argued for this faculty psychology on the grounds that standard philosophical categories of mind (memory, will, cogitation) are abstractions failing to reflect the underlying modular architecture. Evolutionary psychologists argue, on like grounds, that the taxonomy of faculties implemented in the brain are likely to cross-cut the taxonomies posited by contemporary cognitive science. Economics might similarly transform how neuroscientists break the mind into nearly decomposable sub-systems and so influence what we take different regions of the brain to be doing when they perform a task. Describing the phenomenon to be explained and distinguishing it from others in a taxonomy of phenomena (such as decision-making phenomena) guide one to decompose the organism in particular ways (compare the decomposition of the body by the researcher interested in the circulatory system and the researcher interested in reaching behaviour). How one characterizes the phenomena to be explained influences how one divides the mechanism into nearly decomposable mechanistic components (Simon Reference Simon1969; Kauffman Reference Kauffman1970). We must also acknowledge the possibility, however, that the economic perspective on human behaviour enshrined in such behavioural-economic task might not correspond in any tidy way to the mechanisms that drive human behaviour. Perhaps the brain's causal structure does not fit the economist's taxonomy just as the brain failed to accommodate Gall's faculty psychology (or so we now think). Perhaps the neural structures with which we make decisions are too diverse and too varied to be studied usefully from a neural perspective.

Between these extremes lies the possibility that economics offers a perspective on some new behavioural tasks that can be integrated into the way that cognitive scientists think about the brain and its wiring rather than requiring a radical revision of contemporary cognitive science. One might wonder, as neuroeconomists do, how decision-making is related to well-known reward systems, emotional systems, and memory systems in the brain, to name a few. Most of the neuroeconomics with which we are familiar are engaged in the effort to explore how economic phenomena are related to cognitive skills that are recognized in the current taxonomy of cognitive science. This is precisely what one would expect from a mechanistic science: the effort to show how components are organized together in such a way that they explain the phenomenon. In short, the import of economic tasks and perspectives exerts a top-down influence on neuroscience to explain how regularities in decision-making behaviour are implemented in the architecture of the brain. A prudent strategy, and one that appears to describe the practice (if not the rhetoric) of neuroeconomists, would be to try to integrate the economic understanding of behaviour into the best of ever-evolving contemporary cognitive science and to call for radical revision only when the attempted integration with what is known faces threatening anomalies.

What does neuroscience offer to drive bottom-up revision? In section 3, we argued that the search for mechanisms could lead neuroeconomists to build more predictively, explanatorily, and instrumentally useful models. This is precisely what we mean by bottom-up revision. Some types of revision involve recharacterizing the phenomenon to be explained, eliminating the phenomenon, splitting the phenomenon, lumping the phenomenon with others, and revising the scope over which a model is supposed to range.

We find very few examples of such revision in the target articles for this volume, but many examples can be found in CLP's recent review article. Again, we make no bets as to the staying power of these revisions; time will tell. Our point is that successful coevolution and revision is the benchmark of success for neuroeconomics. If neuroeconomics does not eventually produce this kind of revision, it is not worth the investment. For our purposes, CLP's arguments exhibit the kind of thinking that will lead to such a productive and transformative interaction between economics and neuroscience, even if their particular examples of such coevolution do not work out.

CLP begin with a very rough taxonomy of types of processing in the brain (one which we grant for present purposes). The types differ from one another depending on whether they are controlled or automatic, on the one hand, and on whether they are cognitive or affective, on the other. Together these dimensions form four quadrants of cognitive function: controlled cognitive, controlled affective, automatic cognitive and automatic affective (CLP: 16). Many decision-making tasks, such as evaluation of utility, judgments of probability, choice, etc., straddle these quadrants. Moreover, automatic and affective dimensions, scientists are finding, influence judgment and behaviour much more than one would expect given the standard economic theory's emphasis on controlled and cognitive processes. For example, judgments about whether or not one agrees with a given editorial, how probable an event is, whether a given stimulus is positive or negative, whether consumption should be delayed or not, and many others, all take place at the intersection of several of the four quadrants (CLP). Yet only one of these processes, the controlled cognitive one, is captured by the expected utility theory.

This argument exemplifies some of the ways neuroeconomics might be expected to alter economics. First, it is an example of ‘splitting’ – in which what is thought to be a unitary phenomenon is discovered to be several different phenomena explained by different mechanisms. For example, based on the distinction between automatic and controlled processes plus experimental data, neuroeconomists propose that there are two, not one, mechanisms that bring about judgments of probabilities – some explicit, driven by controlled processes, others implicit driven by automatic processes (CLP: 45). Conflict between these systems is invoked to explain psychologists’ numerous observations of violation of probability axioms as, for example, in the famous Linda problem. Another possible example of splitting includes the use of different cognitive systems to explain different aspects of intertemporal choice in hyperbolic discounting (CLP: 39–40). The search for mechanisms, in short, sometimes requires one to recognize that what was previously thought to be a single phenomenon is in fact two or more phenomena.

CLP's discussion also illustrates how the search for mechanisms can lead one to revise the intended scope of one's model. That is, one might be forced to recognize that one's model of behaviour adequately describes the phenomenon only under a limited range of conditions. Different mechanisms are required to explain decision making outside of those conditions. Given the generality of the expected utility theory, it is tempting to view it as the laws of decision-making whenever an agent faces any trade-off.Footnote ²⁰ However, if CLP are right, only some economic decisions can accurately be described by standard economic theory. These decisions are controlled and cognitive, that is, only one of the four quadrants. Outside of that quadrant (where most decisions lie), different models are required to predict, explain and control the phenomena.

The example also illustrates how scope revision might prompt the elimination of certain descriptions of mechanisms. For example, a natural interpretation of some models in microeconomics is as explaining behaviour in terms of beliefs and desires (an agent does x because she desires y and believes x is a means to achieve y). But neural processes that take place in the “hedonic” quadrant (i.e. affective automatic) produce behaviour that is not easily assimilated within the standard folk psychological framework, because the affective automatic processes are cognitively and introspectively inaccessible. Indeed, people seem to confabulate reasons for their behaviour and fail to recognize factors that demonstrably influence their behaviour as having any influence on their behaviour (CLP: 31 and 38; Doris forthcoming). Of course, with elimination of incorrectly described mechanisms come proposals for better models of mechanisms. For example, racial discrimination might be seen as resulting not from conscious preferences or beliefs but from introspectively inaccessible automatic processes.

One final form of bottom-up co-evolutionary revision is lumping, the opposite of splitting. One might discover that two processes hitherto thought separate are in fact accomplished by the same or nearly the same mechanism. CLP, for example, argue that utility for money is no different neurologically than utility for other goods (CLP: 35), and that two different rules for learning in game theory, reinforcement, and learning by updating beliefs about other players, in fact employ the same neural mechanism (CLP: 50). Such a discovery can prompt researchers to see unrecognized connections between behaviours previously treated as distinct.

In short, the effort to integrate economics and neuroscience is likely to transform each in the process. We think it would be best at the moment to work to integrate, to the extent possible, the experimental approaches from economics with the best cognitive neuroscience. What counts as textbook cognitive neuroscience, especially concerning central processes, is hardly fixed in stone and exhibits considerable flexibility on even relatively short time scales. As such, one can expect that the effort to fit economics to the brain will exert downward pressure to revise how we think about the functions performed by different brain systems and regions. At the same time, the search for lower-level mechanisms forces one to ask whether one has characterized the phenomenon to be explained correctly, and so can exert bottom-up pressure to refine the higher-level taxonomy. These are the selective pressures that drive the coevolution of models and will drive the coevolution of economics and neuroscience in a mechanistic neuroeconomics.