Play the expert (PSYCHOLOGY)

© Copyright 2015 by Social Theory and Practice, Vol. 41, No. 4 (October 2015): 579-598 DOI: 10.5840/soctheorpract201541432

Looking for a Psychology for the Inner Rational Agent

Robert Sugden Abstract: Research in psychology and behavioral economics shows that individuals’ choices often depend on “irrelevant” contextual factors. This presents problems for nor- mative economics, which has traditionally used preference-satisfaction as its criterion. A common response is to claim that individuals have context-independent latent preferences which are “distorted” by psychological factors, and that latent preferences should be re- spected. This response implicitly uses a model of human action in which each human be- ing has an “inner rational agent.” I argue that this model is psychologically ungrounded. Although references to latent preferences appear in psychologically based explanations of context-dependent choice, latent preferences serve no explanatory purpose. Keywords: inner rational agent; behavioral welfare economics; preference purification; attention; true self A recurring finding of behavioral economics is that individuals’ choices between what might naturally be thought of as given outcomes can vary according to apparently irrelevant features of the context in which those choices are made. For example, faced with a choice between a specific amount of money and a specific consumer good, people are less likely to choose the money if the decision is framed in terms of selling something that they own than if it is framed as a straight choice.1 When choosing between alternative snacks to be delivered at a fixed time a week in the future, people are more likely to choose unhealthy but hunger-satisfying items if they are hungrier at the time they make the decision.2 In calling such contextual features “irrelevant,” I mean that they have no obvious relevance to the decision-maker’s well-being, interests, or goals; changes in these features seem therefore not to provide good reasons for a person to change her preferences. Nevertheless, there are well-grounded psycho- logical explanations of why revealed preferences are context-dependent. 1Daniel Kahneman, Jack L. Knetsch, and Richard H. Thaler, “Experimental Tests of the Endowment Effect and the Coase Theorem,” Journal of Political Economy 98 (1990): 1325-48. 2Daniel Read and Barbara van Leeuwen, “Predicting Hunger: The Effects of Appetite and Delay on Choice,” Organizational Behavior and Human Decision Processes 76 (1998): 189-205.

580 Robert Sugden These findings present a problem for normative economics, because there is a long tradition in economics of using preference-satisfaction as the criterion for evaluating alternative policy options. That public decision- makers should respect individuals’ preferences has long been an impor- tant idea in liberal political philosophy. But should we—indeed, can we—respect context-dependent preferences? Many economists and philosophers find the idea of respecting context- dependent preferences problematic, either because there seems to be no good reason for thinking that such preferences are indicators of individ- ual well-being, or more fundamentally, because the concept of respecting a person’s preferences is thought to be ill-defined unless those prefer- ences satisfy minimal properties of internal consistency.3 However, the same writers are often reluctant to conclude that there is no need to re- spect preferences at all, and that public decision-makers should simply use their own best judgments about the effects of policies on individuals’ well-being—a conclusion that seems unacceptably paternalistic. A com- mon escape route from this impasse is to argue that individuals whose choices are context-dependent are not revealing the preferences that in some meaningful sense they actually hold, and that their “true,” “under- lying,” or “latent” preferences are context-independent. The disparity be- tween latent preference and choice is attributed to psychological mecha- nisms that induce systematic biases or errors in reasoning. If these latent preferences could be recovered, it would be possible to use the traditional methods of welfare economics to work out how best to satisfy them. Fol- lowing Daniel Hausman, 4 I will call the process of recovering latent preferences preference purification. I will call the broader strategy of us- ing such preferences in welfare economics behavioral welfare economics. By using this strategy, it is thought, the principle of respect for individu- als’ preferences can be retained. In another paper, Gerardo Infante, Guilhem Lecouteux, and I have

3I have argued for an approach to normative economics that attaches value to indi- viduals’ opportunities rather than to the satisfaction of their preferences. This approach does not depend on any assumptions about the coherence of individuals’ preferences while, in a certain sense, respecting whatever preferences individuals act on. See Robert Sugden, “The Opportunity Criterion: Consumer Sovereignty Without the Assumption of Coherent Preferences,” American Economic Review 94 (2004): 1014-33, and “The Value of Opportunities Over Time When Preferences are Unstable,” Social Choice and Welfare 29 (2007): 665-82; and Ben McQuillin and Robert Sugden, “How the Market Responds to Dynamically Inconsistent Preferences,” Social Choice and Welfare 38 (2012): 617-34. The approach is therefore not vulnerable to the problems that are the topic of the current paper. However, this approach has not yet found much favor in economics or philosophy. 4Daniel Hausman, Preference, Value, Choice, and Welfare (Cambridge: Cambridge University Press, 2012), p. 102.

Looking for a Psychology for the Inner Rational Agent 581

examined how this strategy has been used by behavioral economists.5 We present a critique of behavioral welfare economics from the perspective of philosophy of mind. We argue that the strategy understands human agency as if each individual human being has a “rational true self” or “inner rational agent” that has access to some mode of valid reasoning that can generate context-independent preferences. Psychological expla- nations of context-dependent preferences are then interpreted as if the in- dividual’s psychology were an external force subverting the will of the true self. The inner rational agent is not endowed with any psychology of its own, and no description is given of the mode of reasoning it is sup- posed to use. We argue that this model of agency is ungrounded and im- plausible. In section 1 of the current paper, I summarize that argument. However, my main aim in the present paper is to consider behavioral welfare economics from a different perspective, that of cognitive psy- chology. One reason for suspicion about the model of the inner rational agent is that its capacity for correct reasoning is not given any psycho- logical explanation. So one way of trying to make sense of the model is to understand decision-making, both rational and irrational, in terms of psychological mechanisms of mental processing, and to try to isolate some component or aspect of this mental processing that corresponds with rational deliberation and that is capable of generating context- independent preferences. If such a component could be isolated, and if actual behavior could be represented as the result of interaction between it and other psychological mechanisms, the isolated component might be interpreted as the psychological substrate of the inner rational agent and the other mechanisms as potential causes of error. In section 2, I consider this isolation strategy in general terms, and ar- gue that it is unlikely to succeed. A psychological explanation of context- dependent choices does not need a concept of “true” preference. In the most credible of such explanations, responses to contextual cues are an integral part of the mental processes of decision-making. The idea of re- covering latent preferences by removing the influence of these cues seems incoherent. In sections 3 and 4, I support this general claim by examining two specific models—one from behavioral economics, the other from cogni- tive psychology—which at first sight might seem to provide clues about how latent preferences can be isolated. Both models represent the role of attention in the mental processes that underlie decision-making. The first model is typical of much current work in behavioral economics in its use 5Gerardo Infante, Guilhem Lecouteux, and Robert Sugden, “Preference Purification and the Inner Rational Agent: A Critique of the Conventional Wisdom of Behavioural Welfare Economics,” Journal of Economic Methodology, forthcoming.

582 Robert Sugden of concepts of correct reasoning and latent preferences in relation to what are supposed to be models of mental processing, understood empirically. The second model has a much richer representation of mental processes and is presented with much less—but, interestingly, still with some— reference to correctness of reasoning. I will argue that in both models, concepts of correctness play no explanatory role. Thus, however success- ful these models may be in explaining decision-making, they do not pro- vide any empirical grounding for the concept of latent preferences. 1. The Model of the Inner Rational Agent6 Preference purification is at the core of “behavioral welfare economics” —a method of normative analysis that has been used by many prominent behavioral economists.7 Taking a more philosophical perspective, Haus- man gives a qualified endorsement to preference purification as a means of making judgments about individual well-being.8 In the present paper, I will focus on the particularly influential work of Cass Sunstein and Rich- ard Thaler. Sunstein and Thaler claim that the findings of behavioral economics make paternalism unavoidable. This claim is developed in relation to the now-familiar example of a cafeteria director choosing how to display food items when she knows that her customers’ choices are influenced by the prominence with which different items are displayed. Characteriz- ing their antipaternalist opponents as advocating that the director should 6This section is based on Infante, Lecouteux, and Sugden, ibid., and “‘On the Econ Within’: A Reply to Daniel Hausman,” Journal of Economic Methodology, forthcoming. 7Infante, Lecouteux, and I (ibid.) document the use or advocacy of this method by, among others, Bleichrodt et al. (Han Bleichrodt, Jose-Luis Pinto-Prades, and Peter Wak- ker, “Making Descriptive Use of Prospect Theory to Improve the Prescriptive Use of Ex- pected Utility,” Management Science 47 (2001): 1498-514); Camerer et al. (Colin Camerer, Samuel Issacharoff, George Loewenstein, Ted O’Donaghue, and Matthew Rabin, “Regulation for Conservatives: Behavioral Economics and the Case for ‘Asym- metric Paternalism’,” University of Pennsylvania Law Review 151 (2003): 1211-54); Sunstein and Thaler (Cass R. Sunstein and Richard H. Thaler, “Libertarian Paternalism Is Not an Oxymoron,” University of Chicago Law Review 70 (2003): 1159-202; henceforth “ST”); Kőszegi and Rabin (Botond Kőszegi and Matthew Rabin, “Mistakes in Choice- Based Welfare Analysis,” American Economic Review 97 (2007): 477-81); Salant and Rubinstein (Yuval Salant and Ariel Rubinstein, “(A, f): Choice with Frames,” Review of Economic Studies 75 (2008): 1287-96); Thaler and Sunstein (Richard H. Thaler and Cass R. Sunstein, Nudge: Improving Decisions about Health, Wealth, and Happiness (New Haven: Yale University Press, 2008); henceforth “TS”); and Bernheim and Rangel (Douglas Bernheim and Antonio Rangel, “Beyond Revealed Preference: Choice- Theoretic Foundations for Behavioral Welfare Economics,” Quarterly Journal of Eco- nomics 124 (2009): 51-104. 8Hausman, Preference, Value, Choice, and Welfare, pp. 100-102.

Looking for a Psychology for the Inner Rational Agent 583

“give consumers what she thinks they would choose on their own,” Sun- stein and Thaler claim that the antipaternalist position is “incoherent,” because the customers lack “well-formed” (that is, context-independent) preferences. In their 2003 paper, Sunstein and Thaler conclude that the only reasonable decision criterion for the cafeteria director is to “make the choices that she thinks would make the customers best off, all things considered.”9 In their 2008 book, they make a significant revision to this criterion, declaring that their recommendations are designed to “make choosers better off, as judged by themselves.”10 The implication is that the addressee of Sunstein and Thaler’s work—originally called the “planner,” but restyled in 2008 as the “choice architect”—tries to respect each individual’s subjective judgments about what makes him better off. But how are these judgments to be defined, and how can they be re- constructed? Sunstein and Thaler are coy about this, but they provide some clues about their thinking when, immediately after presenting the principle of trying to make choosers “better off, as judged by themselves,” they undertake to show that in many cases, individuals make pretty bad decisions—decisions that they would not have made if they had paid full attention and possessed complete information, unlimited cognitive abilities, and complete self-control.11 The implication is that what makes an individual better off “as judged by himself” is defined by the preferences he would have revealed, had his decision-making not been affected by limitations of attention, informa- tion, cognitive ability or self-control. So Sunstein and Thaler’s approach to normative economics treats context-dependent choices as the result of errors of reasoning. It requires the reconstruction of individuals’ latent preferences by simulating what they would have chosen, had their rea- soning not been subject to these errors. This is preference purification. Clearly, this approach can overcome the problem of context-dependence in actual choices only if, as Sunstein and Thaler implicitly assume is the case, the corresponding latent preferences are context-independent. Behavioral welfare economics can be characterized more precisely as having the following four properties: (1) behavioral welfare economics is intended to apply to cases in which individuals’ revealed preferences de- pend on contextual factors that have little or no apparent relevance to those individuals’ interests or well-being; (2) the normative criterion is the satisfaction of each individual’s latent preferences, defined as the

9ST, pp. 1164-65, 1182. 10 TS, p. 5 (italics in original; the italicized clause recurs with minor variations throughout TS: e.g., pp. 10, 12, 80). 11TS, p. 5.

584 Robert Sugden preferences he would reveal in the absence of any errors that might be caused by limitations of attention, information, cognitive ability or self- control; (3) latent preferences are interpreted as expressing individuals’ subjective judgments about their interests or well-being; they do not nec- essarily track objective properties of the external world (such as an indi- vidual’s monetary wealth or health status) or properties of passive ex- perience (such as happiness in the hedonic sense); (4) in the cases to which behavioral welfare economics is to be applied, latent preferences are assumed to be context-independent. Infante, Lecouteux, and I argue that this approach implicitly uses a model of an inner rational agent. By this, we mean that it treats human agency as if each human being were made up of a neoclassically rational entity encased in, and able to interact with the world only through, an error-prone psychological shell. Of course, we do not claim that behav- ioral economists think that human beings are really made up of these components. The idea of the inner rational agent is merely a way of mak- ing vivid an implication of properties (1) to (4). To be more specific, these properties imply that the human individual has a latent capacity, constant across decision environments, to form context-independent sub- jective judgments on the basis of error-free reasoning. This capacity is not always revealed in the individual’s actual decision-making behavior, but its effects can be isolated by identifying what the individual would have chosen in the absence of errors. “The inner rational agent” is our name for that capacity. In arguing that this model is problematic, Infante, Lecouteux, and I direct most of our criticism at the assumption that latent preferences, as defined by the preference purification method, are complete and context- independent. By treating context-dependent choices as revealing errors of reasoning, the preference purification approach implicitly assumes that for each individual there is some mode of latent reasoning that, if carried out correctly, would generate complete and context-independent preferences. If one interprets preferences as subjective propositions that the relevant individual holds to be true, decision theory imposes consis- tency restrictions on the set of preference propositions that an individual simultaneously holds to be true, but it provides no explanation of how she arrives at those propositions.12 The advocates of preference purifica- 12Decision theory and game theory are not theories of reasoning (that is, theories about the processes by which new propositions are inferred from existing ones); their concepts of “rationality” are consistency conditions that restrict the sets of propositions that an unmodeled process of reasoning is allowed to generate. For more on this, see John Broome, Rationality Through Reasoning (Oxford: Wiley-Blackwell, 2013); and Robin Cubitt and Robert Sugden, “Common Reasoning in Games: A Lewisian Analysis of Common Knowledge of Rationality,” Economics and Philosophy 30 (2014): 285-329.

Looking for a Psychology for the Inner Rational Agent 585

tion invoke a concept of “correct” or “undistorted” reasoning, latent in the real individual, that can somehow create complete and context- independent preferences; but they provide no account of what this rea- soning actually is. We accept that a defensible account of correct reason- ing might show that the set of preference propositions that can be de- rived by that reasoning must satisfy certain consistency conditions, per- haps including some condition of independence of “irrelevant” contex- tual features. But we challenge the implicit assumption that, for every pair of choice objects x and y, correct reasoning can lead to one of the conclusions “x is preferable to y,” “y is preferable to x,” or “x and y are equally preferable.” If one accepts that the preference relation that an in- dividual can derive by correct reasoning may be incomplete, one cannot infer errors of reasoning from context-dependent choices. Thus, contrary to a crucial implicit assumption of the preference purification approach, context-dependency in actual choice may recur in the hypothetical choices that are supposed to reveal latent preferences. 2. Trying to Make Psychological Sense of Latent Preferences Given that behavioral economists usually characterize their subdiscipline as economics with psychological foundations, it is surprising how little work has been done to explain latent preferences in psychological terms. In the remainder of the paper, I consider whether the concept of latent preference might be given empirical content by interpreting it as a com- ponent of a psychological model of mental processing. Some idea of the difficulties involved in this task can be gained by considering an example I mentioned at the beginning of the paper— people’s choices between alternative snacks to be delivered a week after that choice was made. What has been found is that a typical individual’s choices between specific food items (for example, Mars bars and apples) are influenced by his current degree of hunger, even though the date and time of delivery of the snack (and hence, the presumably predictable de- gree of hunger at the time of delivery) is held constant. This is a para- digm case in which choice is influenced by a contextual cue that seems to have no relevance for the individual’s welfare. In broad-brush terms, the psychological mechanism behind this effect is easy to understand. Mars bars and apples are goods with different mixes of attributes: the Mars bar is more energy-giving and perhaps (as viewed by the individual) tastier, the apple is more refreshing and (as an addition to the individual’s typical diet) healthier. In deliberating about which of the two snacks to choose, the individual has to bring these various attributes to mind and strike a balance between them. The hungrier he is, the more attention he

586 Robert Sugden gives to those attributes on which the Mars bar is superior, and so the more likely it is that his deliberation will end in the choice of that option. Viewed in this way, what might seem to be irrational context- dependence is evidence about the underlying structure of the decision- making mechanism. If one thinks in terms of the evolutionary origins of human psychology, the role played by attention in decision-making can be understood as an integral part of a general-purpose mechanism for choosing between multi-attribute options—a mechanism that is (as if) efficiently designed to make use of other mental processes that tend to distribute attention towards what is currently important. (For example, the hungrier one is, the more important it is to be alert to possible sources of nutrition.) But how, then, are we to separate the decision-making mechanism into components of “rationality” and “error,” and to be able to claim that the rational component retains the subjectivity of the real human individ- ual? The only possible way forward that I can see is to try to identify some particular distribution of attention as “correct.” But how are we to do this? Recall that it is fundamental to the preference purification ap- proach that the individual’s latent preferences represent his own subjec- tive judgments. Thus, we cannot define the correct distribution of atten- tion in terms of some objective standard of the individual’s interest, analogous with fitness in an evolutionary model. In the absence of such a standard, the idea of a “neutral” distribution of attention between differ- ent attributes of choice options is ill-defined. (To mention just one prob- lem, suppose we define neutrality as equal attention to every attribute. In the case of the snacks, is the effect of diet on weight a single attribute, or are health and slimness two separate attributes?) It would be circular to define the correct distribution of attention as that which would generate “true” latent preferences, since latent preference have already been de- fined in terms of correct reasoning. The core of the problem is that the attention-based mechanisms that explain the individual’s decisions also explain what, given the relevant choice context, he actually prefers or desires to do: he feels the desires that prompt him to choose as he does. Viewed in the perspective of em- pirical psychology, the idea that he might have “true” preferences that are different from the actual ones seems free-floating and redundant. So far, I have been arguing in very general terms. I will now try to give further support for my skeptical conclusions by looking at two con- crete examples of the use of the concept of latent preference in behav- ioral economics and psychology.

Looking for a Psychology for the Inner Rational Agent 587

3. A Behavioral Economic Model of Attention My first case study is chosen as a characteristic example of how the con- cept of latent preference is used in behavioral economic models. It is the analysis in a recent paper by Pedro Bordalo, Nicola Gennaioli, and An- drei Shleifer (hereafter “BGS”) published in the prestigious Journal of Political Economy.13 BGS develop a model that is motivated by experi- mental findings from psychology, economics, and marketing. The core idea is expressed in a quotation from a psychological paper by Shelley Taylor and Suzanne Thompson: Salience refers to the phenomenon that when one’s attention is differentially directed to one portion of the environment rather than to others, the information contained in that portion will receive disproportionate weighting in subsequent judgments.14 BGS’s core model is of a decision problem in which a consumer faces a choice set containing two or more goods, one and only one of which is to be chosen. Each good k is characterized by the pair ⟨qk, pk⟩, where qk and pk are non-negative magnitudes, respectively representing the quality and price of that good. The consumer knows the value of qk and pk for each good in the choice set. Higher-quality goods are assumed to have higher prices. In effect, BGS assume that quality is measured in its own units on a ratio scale (i.e., a scale on which the zero point is fixed but the unit of measurement is arbitrary).15 In BGS’s leading example, the reader is asked to imagine choosing between a bottle of French wine priced at $20 and a bottle of Austrian wine priced at $10 when the reader thinks the French wine “is perhaps 50 percent better.”16 I take it that in this ex- ample, the consumer is thinking of units of quality as categorically dif- ferent from units of price, and is trying to decide how to make trade-offs between the two attributes. (For example, he might be thinking of the quality scale in terms of the answer he would give to a question asking him to rate the quality of the wine on a scale from 0 to 10; he rates the Austrian wine as 6 and the French wine as 9.) The crucial assumption of the model is stated as follows:

13Pedro Bordalo, Nicola Gennaioli, and Andrei Shleifer, “Salience and Consumer Choice,” Journal of Political Economy 121 (2013): 803-43. 14Shelley E. Taylor and Suzanne C. Thompson, “Stalking the Elusive ‘Vividness’ Effect,” Psychological Review 89 (1982): 155-81, p. 175. 15BGS actually say: “Quality and price are measured in dollars and known to the con- sumer” (p. 807). Despite the literal meaning of this sentence, I think my interpretation is faithful to BGS’s intentions. It is only because quality and price are measured in different units that the consumer faces a nontrivial choice problem. 16BGS, pp. 803-4.

588 Robert Sugden Without salience distortions, a consumer values good k with a linear utility function, uk = qk – pk, which attaches equal weights to quality and price. A salient thinker departs from [this utility function] by inflating the relative weights attached to the attributes that he perceives to be more salient … [W]e say that an attribute (quality or price) is salient for good k in the choice set … if this attribute “stands out” relative to the good’s other attrib- utes.17 I take the first sentence to mean that BGS are assuming a weighted linear utility function uk = αQ qk – αP pk, and are defining the unit in which quality is measured so that αP = αQ = 1. This utility function represents the consumer’s latent preference ordering over ⟨quality, price⟩ pairs; these preferences can be described by a family of what BGS call “ra- tional indifference curves,” which are linear and parallel. That the mar- ginal rate of substitution between units of price and quality is constant is a substantive modeling assumption; that each unit of quality is worth $1 to the “rational consumer” is merely a convenient normalization. Unless the consumer is known to be “rational,” these indifference curves are not directly revealed in choices. Notice that so far, the concepts of rationality and distortion have been given no independent definition or interpreta- tion. BGS have simply stipulated that in their model, a particular family of linear indifference curve is to be called “rational.” BGS then specify “how salience distorts the valuation of a good.”18 The first step is to define a salience function, which, for any choice set, for any good k in that set, and for any attribute j, measures the degree to which the amount of attribute j “stands out” (either as particularly high or particularly low—both are treated as sources of salience) relative to the average amount of that attribute in all goods in the choice set. The sec- ond step is to identify, for each good, which of the two attributes stands out more (as measured by the salience function). Thus, unless there is a tie, each good has a salient attribute—the attribute on which it stands out more. For my purposes, it is sufficient to consider what BGS’s assumptions imply about salience when the choice set contains only two goods, with p1 > p2 and q1 > q2. These assumptions imply that if q1/q2 > p1/p2, quality is the salient attribute for both goods; if that inequality is reversed, price is the salient attribute for both goods. To get an intuitive feel for this property, think of the wine example. Good 1 is the French wine, with p1 = 20 and q1 = 9. Good 2 is the Austrian wine, with p1 = 10 and q2 = 6. Notice that q1/q2 < p1/p2. BGS’s assumptions imply that in this case, the most salient feature of the French wine is its high price relative to the av- erage price of the two wines; correspondingly, the most salient feature of 17Ibid., p. 807. 18Ibid., p. 810.

Looking for a Psychology for the Inner Rational Agent 589

the Austrian wine is its low price. But now suppose that the qualities of the wines are the same as before, but the prices are p1 = 50 and p2 = 40; now q1/q2 > p1/p2 (Suppose the choice is being made in a restaurant rather than a supermarket.) In this case, the most salient feature of the French wine is its high quality and the most salient feature of the Austrian wine is its low quality. BGS’s third step is to model the behavior of a “salient thinker” (i.e., a nonrational consumer) by “distort[ing] the utility weights” that the con- sumer applies when evaluating goods. For the rational consumer, both attributes have a weight of 1 in the evaluation of every good. In contrast, when valuing any given good, the salient thinker uses a weight greater than 1 for its salient attribute and a weight less than 1 for its nonsalient attribute (with the sum of the weights always equal to 2). BGS apply this model to a wide range of consumer behavior prob- lems, using the general strategy of “introducing salience-based valuation into a ‘rational’ economic model.”19 In these applications, they describe the effects of salience as “distortions” of what would otherwise be “ra- tional” choices. All of this exemplifies the dualistic modeling strategy I described in section 1. The behavior of BGS’s “salient thinker” is deter- mined by the interaction of two systems or processes—a set of context- independent latent preferences that are deemed to be rational, and a psy- chological mechanism that distorts these preferences. The choices of the salient thinker are determined by the distorted preferences, but the hypo- thetical choices of the rational consumer—that is, the consumer who acts on undistorted preferences—provide the normative benchmark. This is a model with an inner rational agent. But what is the function of this benchmark in BGS’s model? The es- sence of the model is that the relative weights of the two attributes differ according to which attribute is salient. But which attribute is salient for any given good in any given choice set depends only on the qualities and prices of the goods in that choice set, and these are defined independ- ently of the consumer’s latent preferences. Thus, any results that come about because of changes in relative attribute weights are independent of latent preferences. The concept of latent preference serves no explana- tory purpose. BGS’s example of the two wines illustrates this point. In the story, the consumer chooses the lower-quality Austrian wine when the two wines are priced at $10 and $20, but the higher-quality French wine when the prices are $40 and $50. Leaving aside the possibility of perverse income effects, this pattern of choice is inconsistent with standard economic theory; but it has long been recognized as a common feature of human 19Ibid., p. 813.

590 Robert Sugden decision-making.20 It can be explained in various ways, for example, by assuming diminishing sensitivity to changes in each attribute,21 or by as- suming that expected prices act as reference points;22 BGS provide a new explanation in terms of salience. In the example, adding $30 to the price of each wine switches the salient attribute from price to quality. (In gen- eral, adding any constant to the prices of each of two goods while keep- ing the qualities constant can cause a switch in the salient attribute; if there is a switch, it must be from price to quality. Thus any switch in choice must be from the lower-quality good to the higher-quality good.) But notice that all of this is true (in the model) irrespective of which good the consumer rationally prefers. This does not mean that in the world of the model, rational prefer- ences are unobservable. Consider a decision problem in which there is only one good in the everyday sense of the word, but the consumer can choose whether or not to buy it. BGS represent this as a choice between ⟨p1, q1⟩ and ⟨p2, q2⟩, with ⟨p2, q2⟩ = ⟨0, 0⟩ representing “not buying.” In this special case, BGS’s preferred assumptions about the salience func- tion imply that the two attributes are equally salient, and hence that the salient thinker’s choices coincide with those of the rational consumer. Thus, rational preferences are revealed in the consumer’s willingness to pay for individual goods in situations in which only one good is on offer. Remember, however, that up to this point, the concept of rationality has not been given any interpretation, except as the benchmark relative to which distortion is defined. So the only interpretation that can be given to the proposition that willingness to pay reveals rational preferences is that rational preferences are defined by willingness to pay. To put this another way, the empirical content of the model is con- tained in the idea that choices between goods are influenced by the rela- tive attention given to their attributes, and that more salient attributes are given more attention. A rational consumer is someone who always gives each attribute the right amount of attention. But what is the right amount of attention? In effect, BGS tell us that the right amount of attention to give each attribute is the attention that it is given in willingness-to-pay problems. But they do not explain what this statement means. One possible reconstruction of the missing argument runs as follows. BGS are presupposing that the consumer has well-defined latent prefer-

20See, for example, Leonard Savage, The Foundations of Statistics (New York: Wiley, 1954), p. 103. 21Amos Tversky and Daniel Kahneman, “Loss Aversion in Riskless Choice: A Refer- ence-Dependent Model,” Quarterly Journal of Economics 106 (1991): 1039-61. 22Richard H. Thaler, “Toward a Positive Theory of Consumer Choice,” Journal of Economic Behavior and Organization 1 (1980): 39-60.

Looking for a Psychology for the Inner Rational Agent 591

ences between goods, defined as ⟨quality, price⟩ pairs, and that the latent utility of any good is independent of which other goods are in the choice set. This presupposition is essential for the rest of the reconstructed ar- gument. For the cases that BGS’s model is intended to represent, it is deemed an acceptable simplification to assume a weighted linear utility function, with the implication that the consumer has a context- independent utility weight αQ /αP for quality relative to price. Thus if (as in the wine example) his choices reveal implicit weights that are context- dependent, there must be some cases in which he chooses contrary to his latent preferences. If the qualitative pattern of context-dependence is consistent with a psychological theory of salience and attention, it is rea- sonable to infer that these are cases in which his rational judgment of the utility of the chosen good is distorted by salience effects deriving from comparisons between this good and other goods in the choice set. Thus, the best way to recover the consumer’s latent preferences is to observe his choices in situations in which there is as little scope as possible for cross-good comparisons. Willingness-to-pay problems meet this re- quirement—at least in principle. The “in principle” qualification is needed because BGS extend their basic model to allow the salience of an attribute to depend not only on the content of the choice set, but also on “alternatives that the decision maker expects to find in the current choice setting,” and hence on the consumer’s expectations about prices.23 Thus, this method of eliciting la- tent preferences requires a setting in which “a good is evaluated in isola- tion and without price expectations.” BGS suggest that such settings can be created in “lab experiments.”24 In the light of decades of attempts to elicit willingness-to-pay valuations in experiments and surveys, this sug- gestion seems extraordinarily optimistic. Responses to willingness-to- pay and willingness-to-accept questions are known to be influenced by many kinds of irrelevant cues that draw attention to particular answers.25 For example, if the elicitation exercise begins with a question of the form “Would you be willing to pay $x?” final responses are pulled towards $x; if respondents are asked to pick a point on a scale of possible values, re- sponses are pulled towards the middle of the scale. These “anchoring” 23BGS, p. 820. 24Ibid., p. 828. 25Allen Parducci, “Category Judgment: A Range-Frequency Model,” Psychological Review 72 (1965): 407-18; Paul Slovic and Sarah Lichtenstein, “Relative Importance of Probabilities and Payoffs in Risk Taking,” Journal of Experimental Psychology 78 (1968): 1-18; Eric Johnson and David Schkade, “Bias in Utility Assessments: Further Evidence and Explanations,” Management Science 35 (1989): 406-24; Dan Ariely, George Loewenstein, and Drazen Prelec, “Coherent Arbitrariness: Stable Demand Curves With- out Stable Preferences,” Quarterly Journal of Economics 118 (2003): 73-105.

592 Robert Sugden and “range/frequency” effects are particularly strong when (as in stated preference studies that try to elicit valuations for nonmarketed goods, such as changes in environmental quality) there is no customary price that the respondent can use as a benchmark. A natural interpretation of this evidence is that people find it very difficult to give a monetary valuation of any good in isolation, and that when required to do so, they unconsciously search for comparators and reference points. So it is far from self-evident that individuals have well-defined context- independent latent preferences, ready to be elicited by economists. Since latent preferences play no role in BGS’s explanation of actual choices, we have been given no reason to think that the mental processes that lie behind these choices make use of any such construct. But it is only by assuming the existence of latent preferences that the concepts of “ration- ality” and “distortion” can be given any independent meaning. 4. A Psychological Model of Attention My second example is a seminal contribution to the psychology of decision-making under uncertainty—decision field theory, as proposed by Jerome Busemeyer and James Townsend (hereafter “BT”).26 BT’s aim is “to understand the motivational and cognitive mechanisms that guide the deliberation process involved in decisions under uncertainty.” They are particularly concerned with explaining two “unavoidable facts about hu- man decision making”—that the preferences of a given individual over given pairs of alternatives are subject to stochastic variation, and that the amount of time spent making a decision influences the final choice.27 Thus, they need a model in which deliberation about what to choose is a process that occurs over time and includes some random element. BT’s basic model is of an individual who has to choose between two actions in a situation of uncertainty. Uncertainty is represented by a set of alternative events, one and only one of which will occur. An action is defined by the payoff that will occur in each event if that action is chosen. Payoffs are implicitly assumed to be measured on a ratio scale and can be positive, zero, or negative. BT assume the existence of a utility func- tion that assigns a real value u(x) to every payoff x. However, the indi- vidual is not assumed to attach objective probabilities to events. BT say that they are dealing with decisions under uncertainty (as opposed to risk), defined as problems in which “the decision maker must learn and 26 Jerome R. Busemeyer and James T. Townsend, “Decision Field Theory: A Dynamic–Cognitive Approach to Decision Making in an Uncertain Environment,” Psychological Review 100 (1993): 432-59. 27Ibid., pp. 432-35.

Looking for a Psychology for the Inner Rational Agent 593

infer the event probabilities from past experience.”28 Notice the formal similarities between this problem and the one stud- ied by BGS. BT’s “actions” and “events” are respectively analogous with BGS’s “goods” and “attributes.” The “payoffs” of actions in events are analogous with the “amounts” of attributes that goods possess. BGS’s model does not have an explicit analogue of BT’s utility function for payoffs, but that is only because BGS assume that utility is linear in amounts of attributes.29 In BT’s model, the individual’s problem is to make trade-offs between payoffs that occur in different events; in BGS’s model, it is to make trade-offs between amounts of different attributes. It may help to keep in mind a concrete example of a decision problem to which BT’s model might be applied. Consider Jane, who will be work- ing in some city for a fixed period and has to decide whether to buy a house or to rent one. She knows the current purchase and rental prices of property but is uncertain about how these prices will change over the pe- riod. If property prices rise, she will gain by buying rather than renting; if they fall, she will lose. She cannot assign objective probabilities to these events. (In fact, no one can: if she consults supposed experts, she will find that their judgments differ.) In this problem, the actions are “buy” and “rent” and the events are alternative rates of change in property prices. BT present decision field theory as a succession of amendments to de- terministic subjective expected utility theory, interpreted as a decision rule that assigns a weight to every event (normalized so that the weights sum to 1) and chooses whichever action has the higher weighted average utility. Expected utility theorists normally interpret each of these weights as a subjective probability, but BT offer a different interpretation, saying: “From a cognitive view, this weight reflects the amount of attention given to [the relevant event] on each presentation of the choice prob- lem.”30 Thus, BT’s model, like BGS’s, is one in which decisions depend on the distribution of the decision-maker’s attention (between events or between attributes). In decision field theory, deliberation is a process that occurs over time. At any given moment during this process, there is a preference state measured on a real-valued scale; positive values represent strength of preference in favor of one of the actions, negative values represent strength of preference in favor of the other. Deliberation begins with an

28Ibid., p. 436. 29One disanalogy between the problems should be pointed out. BT’s utility function is defined on payoffs, independently of the events in which they occur, and so event- independent utility measures are treated as inputs to the deliberation process. On my reading, BGS implicitly assume attribute-specific utilities as the analogous inputs. 30Ibid., p. 436 (italics in original).

594 Robert Sugden initial preference state. In “neutral” versions of the theory, the initial state is zero, but BT allow the possibility that the initial state is “biased by past experience” in the direction of the individual’s decisions in simi- lar previous problems.31 This mechanism has the effect of reducing deci- sion times and increasing the stability of choice in familiar problems. Deliberation is represented as sequential sampling of events. Each time an event is sampled, the utility difference between the two actions in that event is registered, and the preference state is updated in the direc- tion of the action with the higher utility. Deliberation ends when the preference state crosses a pre-determined upper or lower threshold; the action that is preferred in this state is then chosen. Sampling an event is interpreted as attending to its payoffs: “The basic idea is that attention may switch from one event to another within a single choice trial.”32 Clearly, the probability that a given action is chosen depends on (among other things) the probabilities with which the different events are sampled. Under certain neutral assumptions (including that the sampling probability for each event remains constant during the process, that the initial preference state is zero, and that the upper and lower thresholds have the same absolute value), the action that is more likely to be chosen is the one with the higher weighted average utility when each event is weighted by the probability that it is sampled at each stage of the process. In other words, under these assumptions it is as if the individual has a subjective probability for each event and is more likely to choose the ac- tion with the higher subjective expected utility; but the as-if probability of any event is actually a measure of the individual’s propensity to attend to it in the deliberation process. On a strict reading of BT, whether this as-if probability can be interpreted as the individual’s subjective judg- ment of the likelihood of the event itself is left open. One might say that in leaving this question open, BT are working in the spirit of Leonard Savage’s subjectivist interpretation of probability as a property of an in- dividual’s preferences over actions, as revealed in her decisions.33 In its most general form, decision field theory does not impose these neutral assumptions, and so does not necessarily generate decisions that can be rationalized by a stochastic form of subjective expected utility theory. But, given the utility function, the initial preference state, the de- cision thresholds, and a full specification of the mechanism that deter- mines the distribution of the individual’s attention, BT’s model generates stochastic decisions and associated decision times. With one exception, it does so without using any concept of “correctness” in decisions. 31Ibid., p. 441. 32Ibid., p. 438. 33Savage, The Foundations of Statistics.

Looking for a Psychology for the Inner Rational Agent 595

The exception appears in BT’s discussion of the implications of alter- native values of the threshold, on the simplifying assumption that the up- per and lower thresholds have the same absolute value θ. In this discus- sion, BT define the correct action as the action that produces the higher subjective expected utility. They then say: [T]he threshold criterion θ controls speed–accuracy or cost–benefit trade-offs in decision making. On the one hand, if the cost of prolonging the decision is low or the cost of mak- ing an incorrect decision is high, then a high threshold is selected. On the other hand, if the cost of prolonging the decision is high or the cost of making an incorrect decision is low, then a low threshold is selected.34 In a footnote to this passage, BT discuss a possible amendment to their model, according to which the value of θ decreases over the deliberation period. Since the fact that the threshold has not been crossed after a long time is evidence that the difference in attention-weighted utility between the actions is relatively low, this amendment would implement what might be called a speed–accuracy trade-off by means of a simple and well-defined psychological mechanism. If this is all that BT have in mind in the quoted passage, nothing much hangs on their definition of “cor- rectness.” Nevertheless, that definition is question-begging. For BT, sub- jective expected utility is merely a construct that, under certain assump- tions, can be read off from the decisions produced by the sequential sam- pling process; the as-if probabilities used in the definition of this con- struct are determined by the distribution of attention. It is not clear why the individual’s propensities to attend to the different events should de- termine which action is deemed to be the correct choice. Think about Jane choosing between buying a house and renting one. If she deliberates in the way described by BT’s model, her attention will switch in a random fashion between thinking about a rising property market (and about the corresponding benefits of buying) and thinking about a falling property market (and about the corresponding benefits of renting). Suppose that if she deliberates for a long time and with many switches of attention, she can be expected to spend 60 per cent of the de- liberation period thinking about a rising market and 40 per cent of the pe- riod thinking about a falling market. How can that fact make the correct choice for Jane be the one that has the higher expected utility when the probabilities of the two events are set at 0.6 and 0.4? One possible answer is that BT’s concept of “correct” choice is not intended to be normative, in the sense of saying what the individual ought to choose; rather, it is an empirical concept, referring to the long- run tendency of deliberation. (It would not sound so odd to say that Jane 34BT, p. 440.

596 Robert Sugden has a latent preference for the action that she would be more likely to choose after long deliberation.) BT may also be thinking of possible extensions of their theory that could close the gap between the normative and empirical concepts of correctness. Recall that in their definition of “uncertainty,” they refer to the event probabilities that the decision-maker has to learn from experi- ence. It would not be inconsistent with this definition to assume the exis- tence of event probabilities, perhaps defined as relative frequencies in a (possibly hypothetical) series of exactly repeated trials. Perhaps BT are entertaining the hypothesis that if an individual faces exactly the same decision problem many times, the distribution of her attention between events converges to the corresponding distribution of event probabilities, irrespective of contextual factors. If this attention hypothesis were true (and given other neutral assumptions), the long-run tendency of repeated decision-making would be towards a state in which the action that was more likely to be chosen was the one with the higher weighted average utility when each event is weighted by its “objective” probability. One might call this action “latently preferred” (or the “correct” choice) in an empirical sense. If one believed that expected utility theory was grounded on compelling principles of rationality, one might also call that action “correct” in a normative sense. And so, if the attention hypothesis were confirmed, one might claim that decision field theory isolates the psychological substrate of context-independent latent preferences of just the kind that behavioral welfare economics needs. But there are some very big “if”s here. Notice in particular that the argument sketched in the previous paragraph depends on assumptions that imply that if the same decision problem is repeated many times, any systematic context-dependence effects gradually disappear. If it really were the case that the choices of experienced decision-makers reliably revealed context-independent preferences, most behavioral economists would probably agree that the preferences to be used in normative analy- sis should be those that individuals reveal after having sufficient experi- ence of relevant choice problems. But the truth is that after a quarter of a century of experimental investigation of the influence of experience on decision “anomalies,” the only general conclusion that can be drawn is that some but not all anomalies seem to decay with some but not all kinds of experience.35 I think we have to accept that context-dependent

35This literature is too large and diverse to be usefully reviewed in a philosophically oriented paper. Graham Loomes, Chris Starmer, and Robert Sugden, “Preference Rever- sals and Disparities Between Willingness to Pay and Willingness to Accept in Repeated Markets,” Journal of Economic Psychology 31 (2010): 374-87, report one experiment that found mixed results, and refer to other relevant papers.

Looking for a Psychology for the Inner Rational Agent 597

choice is not just a symptom of inexperience. Indeed, one might think that the fundamental principles of decision field theory provide reasons for expecting context-dependence to be a pervasive and persistent feature of human decision-making. If the distri- bution of an individual’s attention between alternative events or different attributes is a crucial determinant of her decisions, any “irrelevant” factor that influences the distribution of attention will be capable of inducing context-dependent choices. It does not seem at all self-evident that these influences will become less powerful or less effective as a decision- maker gains experience. If context-dependence is a systematic and persistent consequence of psychological mechanisms that control the distribution of attention, be- havioral welfare economics has to face the question to which BGS’s model provided no satisfactory answer: How can we identify context- independent latent preferences? As an explanation of how attention in- fluences choice, decision field theory is much deeper and more convinc- ing than BGS’s economic model; but it does not answer that question. What it does do is to help us understand why the presupposition of the question is mistaken. If context-independent latent preferences play no role in psychological explanations of actual deliberation or actual choice, we should not expect psychology to tell us how to identify them. 5. Discussion The aim of behavioral welfare economics, as I understand it, is to show how welfare judgments or public policy decisions can respect each indi- vidual’s own subjective preferences, as revealed in her choices after the effects of psychologically induced errors have been controlled for. My purpose in discussing these two models of attention was to explore whether a psychological analysis of decision-making as mental process- ing might allow an empirical distinction to be made between latent pref- erences and error. These models are interesting because they represent attention-based mental processes that can induce context-dependent choices, and because their authors—particularly the authors of the model that belongs to behavioral economics—make use of concepts of “rational” or “correct” latent preference. I have argued that these concepts serve no explanatory purpose. In these models, individuals’ decisions depend on the relative attention given to different attributes or events, allowing context-dependent choices to be explained by causal factors that impact on the mental proc- esses that control the distribution of attention. Of course, if one chooses to define any particular preference as “correct,” there is a correspond-

598 Robert Sugden ingly “correct” distribution of attention. And given any such definition of correctness, there is a corresponding definition of “error,” namely, that an error occurs when an incorrect choice is made; the cause of the error is an incorrect distribution of attention. One might choose to call this causal mechanism a “bias” or “distortion” of correct reasoning. But none of this is any help in determining which preferences are latent in the in- dividual and which are not. If behavioral welfare economics is to succeed in its aim, it has to be able to identify some mode of reasoning or mental processing that, in some well-defined hypothetical situation, would lead the individual to reveal context-independent latent preferences; it must have some defen- sible criterion for defining errors in reasoning; and it must provide good reasons for thinking that this situation is one in which such errors are particularly unlikely to occur. I submit that it has not found any way of doing this, and that the prospects of success are poor. The root of the problem, I believe, is that when economists (and indeed many philoso- phers, and perhaps even some psychologists) think about human agency, they find it hard to avoid using a mental model in which humans are ul- timately rational beings. This model may recognize that humans can hold irrational beliefs and make irrational decisions, but at some deep level, irrationality is understood as the product of mistakes. These mistakes must be defined relative to some “true” preferences—the preferences of the human individual’s “true self.” This is the model of the inner rational agent. We need to recognize that this model is pre-scientific. To ques- tions about the role of latent preferences, the best answer is the one that, in another context, Laplace gave to Napoleon: I had no need of that hy- pothesis.36 Centre for Behavioural and Experimental Social Science, and School of Economics, University of East Anglia [email protected]

36A previous version of this paper was presented at a workshop on the topic “Re- specting Context-Dependent Preferences” at Umeå University in 2014. I thank partici- pants at that workshop, Sudeep Bhatia, Kalle Grill, Graham Loomes, Danny Scoccia, and an anonymous referee for valuable comments. My work was supported by the Economic and Social Research Council through the Network for Integrated Behavioural Science (grant reference ES/K002201/1).

Copyright of Social Theory & Practice is the property of Florida State University, Dept. of Philosophy and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.