The Specter of Revealed Preference Theory

3.1. The Argument from Belief

The first argument points out that in order to (correctly) infer preferences from choice-behavior, we also need information on the agent’s beliefs. Consequently, it makes no sense to identify preference relations with patterns in choice. If preferences were identical with choice, information on an agent’s choice alone should be sufficient to inform us about the agent’s preferences (see Rosenberg 1992; Hausman 2000). For example, my choice for the vegetarian option on the menu is insufficiently captured by my preference for vegetarian options unless we combine it with my belief that the option I chose (let us say falafel) is indeed a vegetarian option. If I had the belief that the option is not vegetarian, my choice would have been different even though my preferences stayed the same. To put it the other way around, if we were to ascribe beliefs about the available options to the agent that do not match how the agent conceptualizes these choices, our preference ascriptions will go wrong. Consider, for instance, that we are observing Sam, who eats a full spoon of wasabi paste. Yet, Sam mistakenly believed that the wasabi cream was avocado cream (see Thoma 2021a). Hence, if we ascribed a preference for eating a spoon of wasabi to Sam, predictions about her future behavior based on this preference would likely go wrong. Because of these considerations, preferences in economics cannot be identical with choice-behavior. Information on an agent’s preferences will not provide us with all the information on an agent’s choices, and we cannot infer preference without having knowledge of the agent’s beliefs.

To add to this, consider that beliefs and preferences are often both given explicit representations in economic models in the form of utility- and credence-functions. In such cases, just considering the representation of an agent’s preferences (i.e., the utility-function) will not provide us with all the information about an agent’s choices that is contained in the model. If preferences in economics were identical with choices, utility-functions should also directly represent the agent’s choices. So, the idea that choices are identical with preference seems to conflict with those economic models that explicitly include credence-functions.⁴

3.2. The Argument from Causality

The second argument states that preferences being identical with choice-behavior would prevent them from figuring into causal explanations of choice-behavior because an event cannot cause itself. Yet, economists often refer to preferences when attempting to give causal explanations of choice-behavior (see, for example, Guala 2012). For example, the behaviorist cannot explain why I chose the vegetarian option with recourse to my preference for the vegetarian option. After all, under behaviorism, this would amount to saying that I chose the vegetarian option because I chose the vegetarian option (see Vredenburgh 2020 for a particularly clear presentation of this argument). As a result, the behavioristic conception fails to capture specific uses of the term preference in economics because it cannot make sense of the fact that economists use preferences in causal explanations of choice-behavior.

4.1. Behaviorism and Its Alleged Advantages

We have already encountered the biggest source of motivation behind revealed preference theory, namely that the main base of evidence that economists have available is choice data (Thoma 2021b; see also Dietrich & List 2016). In light of this, the behaviorist can argue that identifying preferences with anything but choice-behavior will introduce an additional inferential step into the process of assigning preferences to an agent. Yet, if preferences are choice-behavior, no further inference is required, and our ascription of preferences is, at least on a conceptual level, more secure.

However, the fact that inferences become more secure from a conceptual viewpoint is—on its own—of little help if those inferences do not help achieve our epistemic goals. For instance, if we are interested in inferring preference in order to provide causal explanations of choice-behavior, but preferences understood as patterns in choice cannot deliver such explanations, a more secure inference to preferences does not get us very far. Similarly, if we are interested in making projections to unobserved choice-behavior based on preferences, but the correctness of those projections crucially depends on having the correct information about people’s beliefs, a more secure inference to preferences would be of little worth if behaviorism would also prevent us from acquiring correct information about beliefs. All of this is just to say that whether one sees the fact that behaviorism eliminates an inferential step in economics as an advantage crucially depends on what one thinks of the two arguments outlined in the previous sections. Consequently, I hold that the possibility of drawing more direct inferences to preferences by itself offers little motivation for behaviorism.

Yet, we also find other alleged advantages of behaviorism in the literature. Thoma (2021b) mentions three of them, which she calls advantages from black-boxing. In particular, the first alleged advantage is allowing economists to retain a more transparent disciplinary boundary to neuroscience, psychology, and related disciplines. The second alleged advantage is that behaviorism looks attractive in the face of skepticism about specific psychological processes. The third alleged advantage is that behaviorism can help economic theory to achieve greater generality because, under behaviorism, the applicability of economic theory does not depend on agents’ decisions resulting from specific processes.

However, I now argue that these alleged advantages are also possessed by other conceptions of preference. One example of such a conception is functionalism, a version of the position that construes preferences in economics as mental states (see Dietrich & List 2016). Functionalism holds that (intentional) mental states should not be individuated by their intrinsic properties but by the causal role they play in a certain system of inputs, outputs, and other mental states. Any set of entities that occupies these causal roles is, therefore, said to realize a mental state. Importantly, if different sets of entities occupy the same causal role, they all realize the same type of mental state. As long as we are sure that a particular part of a system or agent fulfills the causal role that defines a certain mental state, we do not require any detailed knowledge of this part’s internal composition, location, or organization. Hence, a functionalist conception of preferences also allows for a considerable amount of black-boxing.

Consequently, I hold that functionalism can also deliver the alleged advantages Thoma attributes to behaviorism. For instance, Dietrich and List (2016) note that understanding preferences as functionally individuated mental states does not imply that economics is reducible to psychology or neuroscience. However, they note that not being reducible to a certain discipline does not imply that those disciplines cannot aid economics by providing it with further evidence (see also Craver & Alexandrova 2008 for a nuanced defense of a mechanistic, non-reductive approach to neuroeconomics). Nevertheless, functionalism allows us to leave the disciplinary borders between economics and its neighboring disciplines in place (for whatever this is worth).

Turning to Thoma’s second advantage from black-boxing, functionalism allows us to abstract away from the precise assumptions of specific psychological theories of decision making. Hence, like behaviorism, it looks attractive in the face of skepticism about specific psychological processes. Finally, functionalism can also secure the third alleged advantage mentioned by Thoma because different processes can realize the same mental states if they all occupy the relevant causal role. Hence, similar to behaviorism, functionalism can secure the generality of choice-theory as the ascription of preferences under functionalism does not depend on the presence of specific decision processes.

To sum up, the alleged advantages identified in the recent literature do not force us to adopt behaviorism. Whether we should see the fact that behaviorism would eliminate an inferential step as advantageous depends on whether we accept that behaviorism allows us to realize epistemic aims like explanations and predictions. Moreover, the other three advantages identified by Thoma can be realized by other conceptions of preferences than behaviorism. I have illustrated this with functionalism. Hence, we have so far seen little that would motivate a rehabilitation of behaviorism. Next, I will assess another argument by Thoma that aims at establishing that there are no good alternatives to behaviorism because of the fine-grainedness of preferences in choice-theoretic models.

4.2. Against Alternatives to Behaviorism

Thoma (2021a: 914) provides us with the example of an agent who “drank coffee on the first day because she prefers tasting coffee to tasting tea, and she knew that she would taste coffee if she drank coffee.” Yet, the agent “drank tea on the second day, because she wanted to keep her nerves down for her important meeting, and believed the tea would keep her less nervous than the coffee.”

She argues that to capture this type of behavior with a consistent preference relation, we have to describe the objects of choice at a very fine-grained level that includes descriptions of the various combinations of circumstances that can affect the agent’s choice (e.g., tea before an important meeting, coffee on a relaxed day). She points out that preferences and desires in folk psychological or mentalistic explanations are usually very coarse-grained. For instance, we may say that an agent prefers tasting coffee over tasting tea in a folk psychological explanation to indicate that the agent has a desire for coffee. Yet, we do not refer to such fine-grained descriptions of options in folk psychological explanations as we require them for choice-theory.

Now, Thoma takes folk psychology to fix the causal roles that mental states are meant to occupy according to functionalist accounts of mental states. In particular, in line with arguments by Dietrich and List (2016), she takes the causal roles that preferences would need to occupy to count as mental states to be fixed by our folk psychological concept of desires. Yet, as her example is meant to show, preferences in economics do not seem to play desire-like roles because their content is far more fine-grained than the content of (folk psychological) desires. She, therefore, concludes that functionalism about preferences is unpersuasive as a preference in economics “does not, like the functionalist claims, play a desire-like role in folk psychological explanation” (Thoma 2021a: 913). Hence, according to Thoma, functionalism is not an option in economics.

She, thus, considers more “substantive kinds of mentalism” like the judgmentalism of Bradley (2017) and Hausman’s (2012) view that preferences are total subjective comparative evaluations.⁵ Thoma treats these views as basically stating that preferences are consciously accessible, fine-grained attitudes that we form during deliberation out of our more coarse-grained, folk psychological desires. The problem then, so Thoma argues, is that there is very little evidence that we usually form such attitudes before making decisions. This is meant to be supported by evidence generated by our best psychological theories of decision making as well as the evidence we gain from introspection.

To illustrate this position, Thoma (2021a: 923) tells us that, from an introspective perspective, an agent who does not have an important meeting on a particular day may simply conclude that she drank coffee on that day because she likes the taste of coffee. There is no need for her to form a conscious attitude of the kind: ‘I prefer coffee on a day where I have no important meetings and tea on a day where I have an important meeting.’

That is, the agent can choose coffee without forming this attitude even if the agent would have chosen differently on a day of an important meeting. Yet, if the agent had an important meeting, something would have triggered her desire to keep her nerves so that she would have become aware of it. She would then have based her decision (at least partly) on this desire and opted for tea. In light of cases like this one, Thoma also holds that more substantive forms of mentalism like judgmentalism are unsuccessful because they would restrict choice-theory’s applications to the rare cases in which we really form such fine-grained attitudes. In sum, Thoma concludes that the idea that preferences in economics are mental states cannot be sustained because of the fine-grainedness of preferences in choice-theory. This, according to Thoma, supports behaviorism.

Against this, I hold that her argument rests on an artificial distinction between functionalist views and what Thoma calls substantive views of mentalism. Thoma argues that the latter usually assumes that mental states are consciously accessible. Yet, she acknowledges that conscious accessibility is not necessary for mental states under the functionalist picture. Hence, she dismisses functionalism by saying that preferences in economics do not play the roles desires play in folk psychology and then deals with the substantive view by saying that there is no evidence that we form consciously accessible attitudes of the kind that substantive versions of mentalism would identify with preferences.

Yet, it remains unclear why functionalism forces us to submit to the view that preferences play the role of folk psychological desires. So-called psycho-functionalist (e.g., Fodor 1968; Quilty-Dunn & Mandelbaum 2018) usually identify mental states as entities defined by their role in a cognitive psychological theory. The functional role that a mental state needs to occupy to count as a particular type of mental state would, according to this view, be identified by our best psychological theories.

In light of this, why would it be illegitimate to say that preferences in economics are identified by the role they play in economic theory? An investigation of this role may lead us to the conclusion that preferences in economics occupy causal roles similar to the causal processes that make the agent in Thoma’s example choose coffee on a day at which she has no important meeting and tea on the day at which she has an important meeting. Identifying this whole mechanism with the agent’s preference would preempt Thoma’s objection from fine-grainedness precisely because the mechanism is sensitive to all the various factors that influence our choices. That only parts of this mechanism are consciously accessible would not matter for the functionalist. Finally, that not all parts of this mechanism are triggered on a day on which the agent does not have an important meeting would also not matter for the functionalist.

Of course, Thoma could insist that if we are free to define the causal role of preferences how we like, functionalism ceases to be a plausible position. Therefore, we have to import the relevant causal roles from folk psychology. However, it is pretty standard for functionalists to claim that folk psychology can at best be a start for fixing the causal roles mental states are meant to occupy and that our (best) scientific theories also have an essential role to play here (for the textbook version of this argument see Braddon-Mitchell & Jackson 2007).⁶ Moreover, it is also unclear why desires, as Thoma understands them, are the relevant folk psychological posit for economics. As Thoma notes, many authors have already argued that there is a relevant difference between folk psychological desires and choice-theoretic preferences (e.g., Hausman 2012; Bradley 2017). So why not have the causal role of preferences under functionalism fixed by the role occupied by what Hausman’s would call total subjective comparative evaluations?⁷ Furthermore, Clarke (2020) points out that folk psychological concepts like desiring are usually too vaguely specified to serve as the basis for identifying causal roles. For instance, what are the causal roles that folk psychology would associate with the desire that ‘the government’ should do more about climate change?

Hence, I hold that Thoma’s argument against understanding preferences as functionally individuated states is unsuccessful. Her claim that preferences in economics would need to play desire-like roles in order to count as functionally individuated mental states is undermotivated. Instead, preferences could, for instance, be seen as playing total-subjective-comparative-evaluation-like roles. This suffices to classify them as mental states, according to a version of functionalism that takes economic theory as its basis. This, in turn, seriously undermines Thoma’s argument in favor of behaviorism.

However, Thoma (2021a: 926) appears to anticipate this objection. She argues that there is a sense in which we can identify the whole psychological process she describes in her coffee example as the preference of the agent even if not all parts of it are consciously accessible. Yet, she thinks that this would lead to no explanatory gain at all and that, therefore, “parsimony seems to demand we do away with mentalistic preference and stick to a behavioral interpretation of preference.” The idea seems to be that if we identify preferences with the more complex causal process in her example, all we can say in the end is that an agent chooses coffee because she prefers coffee. Hence, Thoma seems to think that no explanatory value is gained from identifying the whole processes with a preference. Therefore, we should stick with the more parsimonious behaviorism, which refrains from claims about the underlying mechanisms.

However, to have any force against functionalism specifically, such an argument from parsimony must assume that an explanation that cites a consciously accessible, fine-grained substantive mental state (whatever this may be) offers explanatory power. On top of this, we need to assume that an appeal to the whole process that could be viewed as occupying the preference role offers no explanatory power. Otherwise, such an argument would amount to outright denying the explanatory power of preferences even in cases where agents would deliberately form a consciously accessible, fine-grained substantive mental state. For instance, one would need to assume that an explanation of the kind: ‘an agent chooses coffee because she deliberately formed a total subjective comparative evaluation for coffee’ offers explanatory value and an appeal to the functionally individuated causal processes in Thoma’s example offers none. In other words, Thoma’s argument from parsimony cannot serve as an argument against a particular conception of preference if it licenses skepticism about the explanatory value of the preference concept in general.

To highlight that Thoma’s position is indeed threatened to collapse into such skepticism, consider also that explanations referring to preferences as fine-grained attitudes under substantive versions of mentalism appear to be rather shallow if we just look at one particular choice of a single agent. To see this, consider Hausman’s (2012) view that preferences are total subjective comparative evaluations. What is crucial for my purposes here is that under this view, there is no counter-preferential choice as preferences are rankings to the effect that some option X is better than another option Y that take all relevant considerations into account.

Hence, for Hausman (2012), it appears to be almost true by definition (as far as economics is concerned) that an agent chooses based on total subjective comparative evaluations.⁸ Yet, if this is the case, how could it ever be more than a very shallow explanation to state that an agent chooses based on her preferences? Rather than speaking against the explanatory power of preferences in general, I take this to indicate that there is something wrong with the argument that, construed as more complex functionally individuated states, preferences offer no explanatory power.

Instead, what we need to consider is that economic models usually do not attempt to offer explanations of single choices of a single agent. In this regard, I take it that economic explanations will always lose out against more detailed psychological explanations for single choices of a single agent even if the agent in question deliberately forms a preference understood as a consciously accessible, fine-grained attitude. In other words, explaining single choices of single agents is not the strong suit of economics. Rather, economics seems to do better in explaining patterns in large sets of choices.⁹

Economic models try to come up with utility- (and credence-)functions that allow us to derive implications for sets of choices.¹⁰ I shall now argue that utility-functions (as representations of sets of preferences) in combination with credence-functions (as representations of beliefs) that account for large sets of choices already put considerable restrictions on the agent’s decision mechanism. As a result, the explanations we gain for those sets of choices are far from shallow. Let me make this point again with the help of a functionalist conception of preference. Under this reading, what we commit ourselves to when assigning preferences and credences is that the agent’s underlying decision mechanisms are such that they can generate the whole profile of choices implied by these preferences and credences. While we abstract away from details and internal organization, we nonetheless put considerable restrictions on the agent’s decision mechanism. For instance, we rule out mechanisms that could only produce some but not all of the choices.

To further illustrate this, recall that the agent in Thoma’s example does not choose coffee on the day on which she has an important meeting. Imagine now that we observe the agent only on days on which she does not have an important meeting. Hence, we assign her a utility-function that always attaches a higher utility to ‘coffee’ than to ‘tea.’ Because we never observed the agent on the day of an important meeting, her choices so far are still rationalizable by the utility-function that we came up with. Therefore, understood as a mere representation of her actual pattern in choice, this utility-function seems to be perfectly adequate. However, suppose we now understand this utility-function as a representation of a set of functionally individuated states realizable by a particular set of decision mechanisms. In that case, it would not be adequate to assign this function because it would rule out that the choices of the agent are indeed produced by the more complex mechanism, according to which a looming meeting will impact the choice between coffee and tea. To put things differently, in assigning a utility-function that always attaches a higher utility to ‘coffee’, we commit ourselves to the position that the underlying decision mechanism is not the more complex one that Thoma presupposes in her example. Hence, utility- and credence-functions put restrictions on the processes that cause our choices via narrowing down the set of possible causal histories of your choices. In doing so, they provide us with information on the causal processes underlying the agent’s choices. While the resulting type of explanations may look shallow when it comes to explaining single choices of a single agent, this is no longer true once we focus on (large) sets of choices.

My argument here is in line with what Jackson and Pettit (1990) call program explanations. The general idea behind such explanations is that the presence of a higher-level property (e.g., a preference) “programs” for the realization of one of several lower-level properties that causes the event in question. Hence, by citing the relevant higher-level property we provide information on the causal history of the event in question via narrowing down the set of its possible causal histories (for a more detailed take on the application of program explanations to preferences in economics, see Holmes 2022).

Importantly for my purposes, the behaviorist cannot avail herself of program explanations, as according to her conception of preferences, preferences refer directly to choice. By refusing to say anything about the mechanisms that underlie our choices, behaviorism forgoes our ability to offer program explanations because it is not introducing higher-level restrictions that would “program” for the realization of one of several lower-level properties. Hence, Thoma’s argument that parsimony would favor behaviorism over other conceptions of preferences seems unwarranted because program explanations are an epistemic good that behaviorism cannot deliver.

To conclude, this section aimed to show that behaviorism is neither favored by the alleged advantages identified in the literature nor do behaviorists provide any principled obstacles to alternative interpretations of preferences. I, therefore, find that the project of rehabilitating behaviorism is insufficiently motivated. In other words, the behaviorist cannot play for a win here. Yet, she may still aim at a draw. If she could dispel the arguments that speak against behaviorism, she could still try to claim that whether someone opts to be a behaviorist or goes for a different conception is a matter of philosophical taste (e.g., for or against different theories of explanations like program explanations, for or against metaphysical jargon, etc.) (cf. Clarke 2020). However, in the following two sections, I will argue that behaviorists cannot convincingly deal with these arguments.

6.1. Causal Explanations and EUT

Recall that the second argument of the received view states that economists often refer to preferences when attempting to give causal explanations of choice-behavior. Yet, being identical with choice-behavior would prevent preferences from figuring into causal explanations of choice-behavior because an event cannot cause itself (Guala 2012). As a result, the behavioristic conception would fail to capture certain uses of the term ‘preference’ in economics. However, Craver and Alexandrova (2008) and, more recently, Vredenburgh (2020) have suggested that under an interventionist account of causation, preference understood along the lines of behaviorism can explain choice-behavior. I want to give this line of thought a more thorough treatment in what follows.

Roughly speaking, the interventionist account states that X is a cause of Y iff there is a possible intervention on X that changes Y (see Woodward 2005). An intervention is a surgical manipulation of the cause that changes only the value of the cause and severs it from all other causal factors that would determine its value in the absence of the intervention. It is assumed that such interventions are possible in some very general sense. Yet, this does not imply that any human being must be able to perform the intervention.

Note that, under this account of causation, there is, of course, a trivial sense in which choices can cause choices. For example, consider the following claim: my choice to turn on the gas of my stove caused my choice to ignite the lighter. This would qualify as a causal claim under an interventionist account. If I had not chosen to turn on the gas, I would not have chosen to ignite the lighter. Consequently, an intervention that alters my choice to turn on the gas (imagine blocking my path to the stove or informing me that the gas is not connected) would have changed the event of me igniting the lighter.

However, those who want to argue that the interventionist account can help defend behaviorism need to go much further. To see this, consider that preferences in economics are sometimes assumed to depend on each other. For an example, consider the Independence Axiom of EUT (Mas-Colell et al. 1995: 171):

L ≽ L′ if and only if pL + (1-p)L′′ ≽ pL′+(1-p)L′′.

The axiom states that you prefer a lottery L over another lottery L′ iff you would also prefer getting L with probability p (and L′′ otherwise) to a lottery that gives you L′ with probability p (and L′′ otherwise). For example, you prefer coffee instead of tea just in case you also prefer getting coffee if a coin comes up heads (and nothing otherwise) to getting tea if the coin comes up heads (and nothing otherwise). The axiom needs to be satisfied for there to be a cardinal utility-function consistent with EUT.

Hausman (2012) reads the axiom as stating that preferences over simple options determine preferences over lotteries. However, Angner (2018) points out that some economists treat the axiom as stating that preferences over lotteries determine preferences over simple options. Nevertheless, under both readings the axiom postulates a causal relationship between different preferences. These causal readings of the Independence Axiom will help me to illustrate what is wrong with the interventionist defense of behaviorism.

Under Hausman’s reading, a preference for pL+(1-p)L′′ over pL′+(1-p)L′′ is determined by a preference L ≽ L′. The behaviorist may now wish to utilize the interventionist account to recast this reading and argue that what the Independence Axiom really commits us to is that a choice between L and L′ causes a choice between pL+(1-p)L′′ and pL′+(1-p)L′′. To illustrate this, consider the claim that if I had not chosen coffee instead of tea, I would not choose coffee if a coin comes up heads (and nothing otherwise) instead of tea if the coin comes up heads (and nothing otherwise).

Of course, under a non-behavioristic interpretation of preferences, we could explain this in terms of the underlying preferences that cause both choices. Yet, the behaviorist who relies on the interventionist account will want to argue that (in those cases in which the Independence Axiom holds) an intervention on the first choice will also alter the second choice, that is, making it such that I do not choose coffee will also alter my choice over the lotteries that include the chance of getting nothing. At first glance, this might tempt us to conclude that the interventionist account enables economics to give non-trivial causal explanations of choices even if we understand preferences in terms of choice-behavior.

Yet, the main problem with this idea is that it does not take seriously the notion of a surgical intervention that is central to the interventionist account (see Woodward 2015a). According to Woodward, interventionism is an account of how we should reason about causality, that is, under which conditions we should accept causal claims or how we should go about explicating the meaning of our causal claims. In this regard, one of the main functions of interventionism is to help us disambiguate our causal claims. In other words, interventionism prescribes us to spell out the details of the intervention and, thereby, forces us to commit ourselves to what Woodward calls an ideal experiment and its outcome. Thereby, it enables us to explicate the meaning of our causal claims.

Consequently, in the case of intervening on preferences, we must ask how the envisioned intervention would look in detail in order to take the interventionist account seriously. Hence, we need to disambiguate between different interventions that could be entailed by the claim that ‘if one’s choice had been different, other choices would have been different as well.’ For example, consider again the first choice between tea and coffee and a second choice between lotteries that would give us tea and coffee with the same probability (and nothing otherwise). The Independence Axiom seems to commit us to an ideal experiment in which intervening on the agent’s preference to produce a choice of tea instead of coffee also produces a change in the choice between the lotteries. Yet, the conception of preferences employed here is unlikely to be behaviorism. The intervention will not be an intervention on just the choice-behavior under any plausible construal of the ideal experiment. To illustrate this, consider the following graph, where BB stands for the Blackbox that includes all the factors that produce choice-behavior, and C1 stands for the choice-behavior in the first choice-scenario between coffee and tea, and C2 stands for the choice-behavior in the second-choice scenario between lotteries, I stands for the intervention, and the arrows denote causal connections and their directions:

Figure 1: Causal Graph 1

In contrast to this plausible (yet vaguely specified) experiment, the behaviorist would have to commit herself to an alternative intervention in which we only change the actual choice-behavior in one of the choice situations (e.g., the words she utters or her hand-movement) while severing it from all other factors that would influence it in the absence of the intervention. Only this would constitute an intervention on preferences as they are properly understood according to behaviorism. In other words, in order to make sense of the fact that economists cite preferences in causal explanations, the behaviorist would need to argue that there is a direct causal link between the choice in the first choice situation and the choice in the second choice situation. Consequently, the behaviorist would be committed to the following picture in which there is a direct causal link from one choice to another:

Figure 2: Causal Graph 2

However, the problem that arises is that C2 is now determined by BB (the Blackbox) and C1 (the first choice). Yet, BB and C1 can have countervailing implications on C2. Assuming that the Independence Axiom holds, the behaviorist has to argue that changing the choice in C1 to coffee is (ceteris paribus) sufficient to causally produce a choice for the lottery containing coffee in C2. After all, a choice for coffee in C1 entails a choice for the lottery containing coffee in C2 according to the Independence Axiom.

Yet, given that our intervention on the choice in C1 only severs C1 from all other factors that would otherwise influence it (i.e., BB), C2 is still influenced by BB. Now, because BB is unaffected by the intervention, it still impacts C2 (i.e., it has the same impact that it had before the intervention). Yet, the impact of BB on C2 is incompatible with C1 being sufficient to causally produce choice-behavior in C2. For example, if the original choice in C1 was coffee and we now intervene in order to change it to tea, then the causal connection between C1 and C2 would imply that the lottery containing tea is chosen in C2. Yet, as BB is unaffected by the intervention, the causal connection between BB and C2 would still imply that the lottery containing coffee is chosen in C2. Hence, we get contradictory results for C2. To avoid this, one of the two causal connections must go. The only available option for the behaviorist is to deny that there is an underlying mechanism that causes choice-behavior. Yet, this is obviously an untenable position. Nevertheless, it would give us the following picture in which C1 is the only variable causally influencing C2.

Figure 3: Causal Graph 3

Assuming that there is no underlying mechanism that causes choice-behavior and that the first choice is the only variable that causes the second choice obviously solves the problem of getting contradictory results for C2. However, it would commit us to a highly implausible picture of the world. Hence, I take this response to be unconvincing. Instead, we should resolve the problem by committing ourselves to the non-behavioristic picture represented in Figure 1.

Before I proceed, let me make the following clarification. One might think that what I have outlined here is simply an instance of the so-called exclusion problem in the context of interventionism that is extensively discussed in the literature (e.g., Baumgartner 2010; Woodward 2015b). The debate concerns what it is appropriate to control for in assessing causal claims in contexts that involve relations of supervenience (or other non-causal relations like definitional or mathematical connections). As Woodward (2015b) persuasively argues, it not appropriate to control for supervenience bases when assessing causal connections, nor does a coherent version of interventionism require such control. In line with this, if one reads the relationship between BB and C1 (or C2) as C1 (or C2) supervening on BB, it would not be correct to sever C1 from all other factors that would otherwise influence it (i.e., BB) when intervening on it. Instead, we should think of the intervention on C1 as also an intervention on BB (for details see Woodward 2015b: 331–32). If we do so, we will get no contradictory result for C2. However, here I assume that an intervention on C1 is possible without a change in BB and that there is a causal connection—as opposed to supervenience relation—between the two. While I think such interventions are (sometimes) possible (e.g., think of a situation akin the scene in the Marx Brothers’ movie Duck Soup where the lemonade vendor tries to reach for his hat, but repeatedly ends up with another one), my argument does not hinge on this. What is rather important is that the behaviorist would not have gained much if they committed to the view that an intervention on C1 is also an intervention on BB in virtue of a non-causal connection that holds between the two. The question would then arise what the real difference between their position and, for instance, functionalism about preferences is.¹²

6.2. Reinterpreting Causal Claims about Preferences

A more plausible route for the behaviorist would be to argue that ascribing a preference under behaviorism is to commit oneself to a set of claims of the form ‘in context K the agent chooses x.’ The behaviorist could argue that if such claims hold, an intervention on the context K will affect choice x (see Vredenburgh 2020: 153). She could then argue that causal explanations involving preferences should really be understood as claims that causally link context and choice-behavior. In other words, we need to reinterpret causal claims about preferences as causal claims about context and choice.

The claim that different contexts have different causal influences on choice is, of course, perfectly acceptable. In fact, I take it to be so uncontroversial that we do not even need to employ an interventionist framework to convince us of this fact. Yet, how does it fare in relation to causal claims that economists make about preferences?

In order to asses this, let us look at an example that could easily work in the behaviorist’s favor. When game theorists study social norms, they are interested in showing how changing people’s preferences or beliefs would affect the stability of those norms (Lewis 1969; Young 1996; Bicchieri 2016). According to Bicchieri’s highly influential analysis of social norms, three conditions must be present for the existence of a social norm (e.g., Bicchieri & Chavez 2010): (i) individuals must believe that the norm exists, (ii) individuals must have a conditional preference for following the norm (such conditional preferences are sensitive to the empirical expectation that others comply with the norm and the normative expectation that others expect them to obey and may sanction for noncompliance), and (iii) individuals must actually possess the empirical and normative expectations.

At first glance, this definition explicitly specifies two features on which the agent’s conditional preference depends, that is, her empirical and normative expectation. In this regard, it does not seem to treat preferences as mere patterns in choice, but tells us something about their determinants. Yet, the behaviorist can argue that Bicchieri’s conditions (ii) and (iii) can be easily reformulated such that they solely focus on context and choice. In particular, she can restate these conditions as ‘individuals have a preference for/choose following the norm only in contexts in which they hold the empirical and normative expectations.’ Hence, intervening on this context would alter whether the agent chooses to comply or not. This would basically collapse the two conditions into a claim about a simple input-output relation (see Clarke 2020). The input is defined in terms of context (i.e., expectations), and the output is choice. Hence, there are cases in which claims about preferences are easily restated as claims about context and choice.

However, this reinterpretation seems unmotivated. As we have already seen, by including states like expectations into the context, the behaviorist “would make use of the sort of psychological concepts that she intended to eliminate in the first place” (Guala 2019: 386). In particular, to intervene on the relevant context (i.e., the expectations) in a way that is precise enough for an interventionist account, one would need to have some grasp of what influences people’s expectations. Yet, to acquire this information, one would have to make the kind of assumptions about an agent’s psychology that most behaviorists are trying to avoid. Hence, to argue that causal claims about preferences should be recast as causal claims about context and choice, the behaviorist would have to adopt a notion of context that betrays her original motivations.¹³ Considering this, there appears to be little basis on which the behaviorist could demand that scholars like Bicchieri are required to adopt the behaviorist’s outlook on preferences and amend their definitions accordingly.¹⁴

On top of this, I hold that there are clear heuristic advantages that favor Bicchieri’s original formulation of her conditions for social norm over the behavioristic reformulation. Bicchieri (2014) herself insists that social norms cannot be analyzed just in terms of observable behavior to highlight the importance of expectations for conceptualizing them. To illustrate why behaviorism threatens to obscure this vital feature in the game-theoretic study of conventions and social norms, consider the following example by Guala (2019: 386):

Tony [prefers] not to contribute to the organization of the Christmas fair because he believes that no one in the neighbourhood will give any money, although he regrets that the fair will not take place. Vince, in contrast, [prefers] not [to] contribute because he hates Christmas fairs and wouldn’t give any money even if everyone else did.

The preferences of these two agents are very different. Even though, in the actual world, none of them would contribute to the organization of the Christmas fair, Tony is facing the kind of coordination game that underlies social norms under Bicchieri’s view, while Vince does not. The most straightforward explanation for why they are engaged in different games is that, even though they face the same context, their preferences are relevantly different. Tony’s preference are conditional preferences that depend on certain expectations where Vince’s do not. Moreover, we can be confident in this explanation because we have evidence that Tony feels regret and Vince does not. Under a non-behavioristic interpretation, it becomes immediately clear why Tony’s regret can serve as evidence for the hypothesis that only Tony is facing a type of coordination game, that is, Tony’s regret can inform us about the determinants of his preferences.

But what is the story the behaviorist can tell us for why we can connect the evidence of Tony’s regret with the conclusion that Tony is facing a coordination game? Of course, she can hypothesize, after observing a change in neighborhood contribution and a change in Tony’s behavior but not in Vince’s, that only Tony is facing a coordination game. Yet, the behaviorist appears to lack a plain explanation for why Tony’s regret is a relevant piece of evidence in virtue of which the hypothesis could have been formed before making these observations.¹⁵

Moreover, when dealing with conventions and social norms we often would like to know why compliance differs between agents that we intuitively frame to be in the same context. In this regard, it seems natural to say that two agents make different choices in the same context (e.g., a context in which they have to choose whether to contribute to the organization of the Christmas fair or not) because of their different expectations and the resulting preferences. However, the behaviorist who also admits factors like expectations into the relevant context of choice seems to be committed to the somewhat more cumbersome claim that the two agents are making different choices because they are in different contexts (e.g., in virtue of their different expectations). I do not think that this is the best way to frame the issue given the kind of questions we are interested in when studying social norms.¹⁶

All in all, I, therefore, take it that Bicchieri’s conditions are more straightforwardly construed as claims about what influences certain preferences under a non-behavioristic interpretation than as claims about context and choice. More generally, I conclude that reinterpreting causal claims involving preferences in terms of context and choice, while (sometimes) technically possible, is insufficient to reestablish the attractiveness of behaviorism.