1 Behavioral Economics: Past, Present, Future Colin F. Camerer Division of Humanities and Social Sciences 228-77 Caltech Pasadena, CA 91125 email@example.com George Loewenstein Department of Social and Decision Sciences Carnegie-Mellon University Pittsburgh PA 15213 firstname.lastname@example.org draft: 10/25/02 Behavioral economics increases the explanatory power of economics by providing it with more realistic psychological foundations. This book consists of representative recent articles in behavioral economics.
1 This chapter is intended to provide an introduction to the approach and methods of behavioral economics, and to some of its major findings, applications, and promising new directions. It also seeks to fill some unavoidable gaps in the chapters 9 coverage of topics. What Behavioral Economics Tries To Do At the core of behavioral economics is the conviction that increasing the realism of the psychological underpinnings of economic analysis will improve economics on its own terms -- generating theoretical insights, making better predictions of field phenomena, and suggesting better policy.
This conviction does not imply a wholesale rejection of the neoclassical approach to economics based on utility maximization, equilibrium, and efficiency. The neoclassical approach is useful because it provides economists with a theoretical framework that can be applied to almost any form of economic (and even non-economic) behavior, ... more. less.
and it makes 1 Since it is a book of advances, many of the seminal articles which influenced those collected here are not included, but are noted below and are widely reprinted elsewhere. 2 refutable predictions.<br><br> Many of these predictions are tested in the chapters of this book, and rejections of those predictions suggest new theories. Most of the papers modify one or two assumptions in standard theory in the direction of greater psychological realism. Often these departures are not radical at all because they relax simplifying assumptions that are not central to the economic approach.<br><br> For example, there is nothing in core neoclassical theory that specifies that people should not care about fairness, that they should weight risky outcomes in a linear fashion, or that they must discount the future exponentially at a constant rate. 2 Other assumptions simply acknowledge human limits on computational power, willpower, and self-interest. These assumptions can be considered 'procedurally rational' (Herbert Simon 9s term) because they posit functional heuristics for solving problems that are often so complex that they cannot be solved exactly by even modern computer algorithms.<br><br> Evaluating Behavioral Economics Stigler (1965) says economic theories should be judged by three criteria: congruence with reality, generality, and tractability. Theories in behavioral economics should be judged this way too. We share the positivist view that the ultimate test of a theory is the accuracy of its predictions.<br><br> 3 But we also believe that, ceteris paribus, better predictions are likely to result from theories with more realistic assumptions. Theories in behavioral economics also strive for generality 3 e.g., by adding only one or two parameters to standard models. Particular parameter values then often reduce the behavioral model to the standard one, and the behavioral model can be pitted against the standard model by estimating parameter values.<br><br> And once parameter values are pinned down, the behavioral model can be applied just as widely as the standard one. 2 While the papers in this book largely adhere to the basic neoclassical framework, there is nothing inherent in behavioral economics that requires one to embrace the neoclassical economic model. Indeed, we consider it likely that alternative paradigms will eventually be proposed which have greater explanatory power.<br><br> Recent developments in psychology, such as connectionist models that capture some of the essential features of neural functioning, bear little resemblance to models based on utility maximization, yet are reaching the point where they are able to predict many judgmental and behavioral phenomena. 3 Contrary to the positivistic view, however, we believe that predictions of feelings (e.g., of subjective well-being) should also be an important goal. 3 Adding behavioral assumptions often does make the models less tractable.<br><br> However, many of the papers represented in this volume show that it can be done. Moreover, despite the fact that they often add parameters to standard models, behavioral models, in some cases, can even be more precise than traditional ones which assume more rationality, when there is dynamics and strategic interaction. Thus, Lucas (1986) noted that rational expectations allows multiple inflationary and asset price paths in dynamic models, while adaptive expectations pins down one path.<br><br> The same is true in game theory: Models based on cognitive algorithms (e.g., Camerer, Ho & Chong, 2001) often generate precise predictions in those games where the mutual consistency requirement of Nash permits multiple equilibria. The realism, generality and tractability of behavioral economics can be illustrated with the example of loss-aversion. Loss-aversion is the disparity between the strong aversion to losses relative to a reference point and the weaker desire for gains of equivalent magnitude.<br><br> Loss aversion is more realistic than the standard continuous, concave, utility function over wealth, as demonstrated by hundreds of experiments. Loss aversion has proved useful in identifying where predictions of standard theories will go wrong: Loss-aversion can help account for the equity premium puzzle in finance and asymmetry in price elasticities. (We provide more examples below.) Loss aversion can also be parameterized in a general way, as the ratio of the marginal disutility of a loss relative to the marginal utility of a gain at the reference point (i.e., the ratio of the derivatives at zero); the standard model is the special case in which this "loss-aversion coefficient" is one.<br><br> As the foregoing suggests, loss-aversion has proved tractable 4although not always simple-- in several recent applications (e.g., Barberis, Huang & Santos, 2001). The Historical Context Of Behavioral Economics Most of the ideas in behavioral economics are not new; indeed, they return to the roots of neoclassical economics after a century-long detour. When economics first became identified as a distinct field of study, psychology did not exist as a discipline.<br><br> Many economists moonlighted as the psychologists of their times. Adam Smith, who is best known for the concept of the "invisible hand" and The Wealth of Nations , wrote a less well-known book The Theory of Moral Sentiments , which laid out psychological principles of individual behavior that are arguably as profound as his economic observations. The book is bursting with insights about human psychology, many of which presage current developments in behavioral economics.<br><br> For 4 example, Adam Smith commented (1759/1892, 311) that "we suffer more... when we fall from a better to a worse situation, than we ever enjoy when we rise from a worse to a better. d Loss aversion! Jeremy Bentham, whose utility concept formed the foundation of neoclassical economics, wrote extensively about the psychological underpinnings of utility, and some of his insights into the determinants of utility are only now starting to be appreciated (Loewenstein 1999).<br><br> Francis Edgeworth 9s Theory of Mathematical Psychics , which introduced his famous "box" diagram showing two-person bargaining outcomes, also included a simple model of social utility, in which one person 9s utility was affected by another person 9s payoff, which is a springboard for modern theories (see chapters 9 and 10 for two examples). The rejection of academic psychology by economists, perhaps somewhat paradoxically, began with the neoclassical revolution, which constructed an account of economic behavior built up from assumptions about the nature 4that is, the psychology 4of homo-economicus. At the turn of the 20 th century, economists hoped their discipline could be like a natural science.<br><br> Psychology was just emerging at that time, and was not very scientific. The economists thought it provided too unsteady a foundation for economics. Their distaste for the psychology of their period, as well as dissatisfaction with the hedonistic assumptions of Benthamite utility, led to a movement to expunge the psychology from economics.<br><br> 4 Expunging psychology from economics happened slowly. In the early part of the 20th century, the writings of economists such as Irving Fisher and Vilfredo Pareto still included rich speculations about how people feel and think about economic choices. Later John Maynard Keynes very much appealed to psychological insights, but by the middle of the century discussions of psychology had largely disappeared.<br><br> Throughout the second half of the century, many criticisms of the positivistic perspective took place in both economics and psychology. In economics, researchers like George Katona, Harvey Leibenstein, Tibor Scitovsky, and Herbert Simon wrote books and articles suggesting the 4 The economists of the time had less disagreement with psychology than they realized. Prominent psychologists of the time were united with the economists in rejecting hedonism as the basis of behavior.<br><br> William James, for example, wrote that "psychologic hedonists obey a curiously narrow teleological superstition, for they assume without foundation that behavior always aims at the goal of maximum pleasure and minimum pain; but behavior is often impulsive, not goal-oriented," while William McDougall stated in 1908 that "it would be a libel, not altogether devoid of truth, to say that classical political economy was a tissue of false conclusions drawn from false psychological assumptions. d (Both quotes from Lewin (1996).) 5 importance of psychological measures and bounds on rationality. These commentators attracted attention, but did not alter the fundamental direction of economics. Many coincident developments led to the emergence of behavioral economics as represented in this book.<br><br> One development was the rapid acceptance by economists of the expected utility and discounted utility models as normative and descriptive models of decision making under uncertainty and intertemporal choice, respectively. Whereas the assumptions and implications of generic utility analysis are rather flexible, and hence tricky to refute, the expected utility and discounted utility models have numerous precise and testable implications. As a result, they provided some of the first "hard targets" for critics of the standard theory.<br><br> Seminal papers by Allais (1953), Ellsberg (1961) and Markowitz (1952) pointed out anomalous implications of expected and subjective expected utility. Strotz (1955) questioned exponential discounting. Later scientists demonstrated similar anomalies using compelling experiments that were easy to replicate (Kahneman & Tversky, 1979, on expected utility, and Thaler, 1981, and Loewenstein & Prelec, 1992, on discounted utility).<br><br> As economists began to accept anomalies as counterexamples that could not be permanently ignored, developments in psychology identified promising directions for new theory. Beginning around 1960, cognitive psychology became dominated by the metaphor of the brain as an information-processing device replacing the behaviorist conception of the brain as a stimulus-response machine. The information-processing metaphor permitted a fresh study of neglected topics like memory, problem solving and decision making.<br><br> These new topics were more obviously relevant to the neoclassical conception of utility maximization than behaviorism had appeared to be. Psychologists such as Ward Edwards, Duncan Luce, Amos Tversky and Daniel Kahneman, began to use economic models as a benchmark against which to contrast their psychological models. Perhaps the two most influential contributions were published by Tversky and Kahneman.<br><br> Their 1974 Science article argued that heuristic short-cuts created probability judgments which deviated from statistical principles. Their 1979 paper "Prospect theory: decision making under risk" documented violations of expected utility and proposed an axiomatic theory, grounded in psychophysical principles, to explain the violations. The latter was published in the technical journal Econometrica and is one of the most widely cited papers ever published in that journal.<br><br> 6 A later milestone was the 1986 conference at the University of Chicago, at which an extraordinary range of social scientists presented papers (see Hogarth & Reder, 1987). Ten years later, in 1997, a special issue of the Quarterly Journal of Economics was devoted to behavioral economics (three of those papers are reprinted in this volume). Early papers established a recipe that many lines of research in behavioral economics have followed.<br><br> First, identify normative assumptions or models that are ubiquitously used by economists, such as Bayesian updating, expected utility and discounted utility. Second, identify anomalies 4i.e., demonstrate clear violations of the assumption or model, and painstakingly rule out alternative explanations (such as subjects 9 confusion or transactions costs). And third, use the anomalies as inspiration to create alternative theories that generalize existing models.<br><br> A fourth step is to construct economic models of behavior using the behavioral assumptions from the third step, derive fresh implications, and test them. This final step has only been taken more recently but is well represented in this volume of advances. The Methods Of Behavioral Economics The methods used in behavioral economics are the same as those in other areas of economics.<br><br> At its inception, behavioral economics relied heavily on evidence generated by experiments. More recently, however, behavioral economists have moved beyond experimentation and embraced the full range of methods employed by economists. Most prominently, a number of recent contributions to behavioral economics, including several included in this book (Chapters 21, 25 and 26, and studies discussed in chapters 7 and 11) rely on field data.<br><br> Other recent papers utilize methods such as field experiments (Gneezy and Rustichini (this volume) computer simulation (Angeletos et al., 2001), and even brain scans (McCabe et al, 2001). Experiments played a large role in the initial phase of behavioral economics because experimental control is exceptionally helpful for distinguishing behavioral explanations from standard ones. For example, players in highly anonymous one-shot take-it-or-leave-it "ultimatum" bargaining experiments frequently reject substantial monetary offers, ending the game with nothing (see Camerer & Thaler, 1995).<br><br> Offers of 20% or less of a sum are rejected about half the time, even when the amount being divided is several weeks 9 wages or $400 in the 7 US (e.g., Camerer, 2002). Suppose we observed this phenomenon in the field, in the form of failures of legal cases to settle before trial, costly divorce proceedings, and labor strikes. It would be difficult to tell whether rejection of offers was the result of reputation-building in repeated games, agency problems (between clients and lawyers), confusion, or an expression of distaste for being treated unfairly.<br><br> In ultimatum game experiments, the first three of these explanations are ruled out because the experiments are played once anonymously, have no agents, and are simple enough to rule out confusion. Thus, the experimental data clearly establish that subjects are expressing concern for fairness. Other experiments have been useful for testing whether judgment errors which individuals commonly make in psychology experiments also affect prices and quantities in markets.<br><br> The lab is especially useful for these studies because individual and market-level data can be observed simultaneously (e.g., Camerer, 1987; Ganguly, Kagel & Moser, 2000). Although behavioral economists initially relied extensively on experimental data, we see behavioral economics as a very different enterprise from experimental economics (see Loewenstein, 1999). As noted, behavioral economists are methodological eclectics.<br><br> They define themselves, not on the basis of the research methods that they employ, but rather their application of psychological insights to economics. Experimental economists, on the other hand, define themselves on the basis of their endorsement and use of experimentation as a research tool. Consistent with this orientation, experimental economists have made a major investment in developing novel experimental methods that are suitable for addressing economic issues, and have achieving a virtual consensus among themselves on a number of important methodological issues.<br><br> This consensus includes features that we find appealing and worthy of emulation (see Hertwig & Ortmann, in press). For example, experimental economists often make instructions and software available for precise replication, and raw data are typically archived or generously shared for reanalysis. Experimental economists also insist on paying performance-based incentives, which reduces response noise (but does not typically improve rationality; see Camerer & Hogarth, 1999), and also have a virtual prohibition against deceiving subjects.<br><br> However, experimental economists have also developed rules that many behavioral economists are likely to find excessively restrictive. For example, experimental economists rarely collect data like demographics, self-reports, response times, and other cognitive measures 8 which behavioral economists have found useful. Descriptions of the experimental environment are usually abstract rather than evocative of a particular context in the outside world because economic theory rarely makes a prediction about how contextual labels would matter, and experimenters are concerned about losing control over incentives if choosing strategies with certain labels is appealing because of the labels themselves.<br><br> Psychological research shows that the effect of context on decision making can be powerful (see, e.g., Goldstein & Weber, 1995; Loewenstein, 2001) and some recent experimental economics studies have explored context effects too (e.g., Cooper, Kagel, Lo & Gu, 1999; Hoffman et al, 1994). Given that context is likely to matter, the question is whether to treat it as a nuisance variable or an interesting treatment variable. It is worth debating further whether helping subjects see a connection between the experiment and the naturally-occurring situations the experiments is designed to model, by using contextual cues, is helpful or not.<br><br> Economics experiments also typically use "stationary replication" 4in which the same task is repeated over and over, with fresh endowments in each period. Data from the last few periods of the experiment are typically used to draw conclusions about equilibrium behavior outside the lab. While we believe that examining behavior after it has converged is of great interest, it is also obvious that many important aspects of economic life are like the first few periods of an experiment rather than the last .<br><br> If we think of marriage, educational decisions, and saving for retirement, or the purchase of large durables like houses, sailboats, and cars, which happen just a few times in a person 9s life, a focus exclusively on cpost-convergence d behavior is clearly not warranted. 5 All said, the focus on psychological realism and economic applicability of research promoted by the behavioral-economics perspective suggests the immense usefulness of both empirical research outside the lab and of a broader range of approaches to laboratory research. 5 We call the standard approach "Groundhog Day" replication, after the Bill Murray movie in which the hero finds himself reliving exactly the same day over and over.<br><br> Murray 9s character is depressed until he realizes that he has the ideal opportunity to learn by trial-and-error, in a stationary environment, and uses the opportunity to learn how to woo his love interest. 9 Basic Concepts and Research Findings The field of Behavioral Decision Research, on which behavioral economics has drawn more than any other subfield of psychology, typically classifies research into two categories: judgment and choice. Judgment research deals with the processes people use to estimate probabilities.<br><br> Choice deals with the processes people use to select among actions, taking account of any relevant judgments they may have made. In this section, we provide a background on these two general topics to put the contributions of specific chapters into a broader context. Probability judgment Judging the likelihood of events is central to economic life.<br><br> Will you lose your job in a downturn? Will you be able to find another house you like as much as the one you must bid for right away? Will the Fed raise interest rates?<br><br> Will an AOL-TimeWarner merger increase profits? Will it rain during your vacation to London? These questions are answered by some process of judging likelihood.<br><br> The standard principles used in economics to model probability judgment in economics are concepts of statistical sampling, and Bayes 9 rule for updating probabilities in the face of new evidence. Bayes 9 rule is unlikely to be correct descriptively because it has several features that are cognitively unrealistic. First, Bayesian updating requires a prior.<br><br> 6 Second, Bayesian updating requires a separation between previously-judged probabilities and evaluations of new evidence. But many cognitive mechanisms use previous information to filter or interpret what is observed, violating this separability. For example, in perception experiments, subjects who expect to see an object in a familiar place 4such as a fire hydrant on a sidewalk 4perceive that object more accurately than subjects who see the same object in an unexpected place 4such as on a coffeeshop counter.<br><br> Third, subjective expected utility assumes separability between probability judgments of states and utilities which result from those states. Wishful thinking and other self- serving motivations violate this separation (see Babcock & Loewenstein, 1997, and this volume). Fourth, the Bayesian updating predicts no effects of the order of arrival of information.<br><br> But order effects are common in memory due to the strength of recent information in working memory 6 Because it does not specify where the prior comes from, however, it leaves room for psychological theory on the front end of the judgment process. 10 (recency effects), and increased "rehearsal" of older memories (primacy effects). These order effects mean that how information is sequenced distorts probability judgment (see Hogarth & Einhorn, 1992).<br><br> Cognitive psychologists have proposed heuristic mechanisms that will lead to judgments which sometimes violate either sampling principles or Bayes 9 rule (see Kahneman & Frederick, 2002). For example, people may judge the probabilities of future events based on how easy those events are to imagine or to retrieve from memory. This "availability heuristic" contributes to many specific further biases.<br><br> One is "hindsight bias": Because events which actually occurred are easier to imagine than counterfactual events that did not, people often overestimate the probability they previously attached to events which later happened. This bias leads to "second- guessing" or Monday-morning quarterbacking and may be partly responsible for lawsuits against stockbrokers who lost money for their clients. (The clients think the brokers cshould have known d) A more general bias is called the "curse of knowledge" 4people who know a lot find it hard to imagine how little others know.<br><br> The development psychologist Jean Piaget suggested that the difficulty of teaching is caused by this curse. (Why is it so hard to explain something cobvious d like consumer indifference curves or Nash equilibrium to your undergraduate students? 7 ) Anybody who has tried to learn from a computer manual has seen the curse of knowledge in action.<br><br> Another heuristic for making probability judgments is called "representativeness": People judge conditional probabilities like P(hypothesis|data) or P(example|class) by how well the data represents the hypothesis or the example represents the class. Like most heuristics, representativeness is an economical shortcut that delivers reasonable judgments with minimal cognitive effort in many cases, but sometimes goofs badly and is undisciplined by normative principles. Prototypical exemplars of a class may be judged to be more likely than they truly are (unless the prototype 9s extremity is part of the prototype).<br><br> For example, in judging whether a certain student described in a profile is, say, a psychology major or a computer science major, people instinctively dwell on how well the profile matches the psychology or computer science 7 Here is an example from the business world: When its software engineers refused to believe that everyday folks were having trouble learning to use their opaque, buggy software, Microsoft installed a test room with a one-way mirror so that the engineers could see people struggling before their very eyes (Heath, Larrick, & Klayman, 1998). 11 major stereotype. Many studies show how this sort of feature-matching can lead people to underweigh the "base rate" 3 in this example, the overall frequency of the two majors.<br><br> 8 Another byproduct of representativeness is the "law of small numbers": Small samples are though to represent the properties of the statistical process that generated them (as if the law of large numbers, which guarantees that a large sample of independent draws does represent the process, is in a hurry to work). If a baseball player gets hits 30% of his times at bat, but is 0 for 4 so far in a particular game, then he is "due" for a hit in his next at bat in this game, so that this game 9s hitting profile will more closely represent his overall ability. The so-called "gambler's fallacy", whereby people expect a tail after a coin landed heads three times in a row, is one manifestation of the law of small numbers.<br><br> The flip side of the same misjudgment (so to speak) is surprise at the long streaks which result if the time series is random, which can lead people to conclude that the coin must be unfair when it isn't. Field and experimental studies with basketball shooting and betting on games show that people, including bettors, believe that there is positive autocorrelation 4that players experience the "hot hand" 4 when there is no empirical evidence that such an effect exists (see Camerer, 1989a; Gilovich, Vallone & Tversky, 1985). Many studies explore these heuristics and replicate their "biases" in applied domains (such as judgments of accounting auditors, consumers buying products, and students in classroom negotiations).<br><br> It is important to note that a "heuristic" is both a good thing and a bad thing. A good heuristic provides fast, close to optimal, answers when time or cognitive capabilities are limited, but it also violates logical principles and leads to errors in some situations. A lively debate has emerged over whether heuristics should be called irrational if they were well-adapted to domains of everyday judgment ( cecologically rational d).<br><br> In their early work, Kahneman, Tversky, and others viewed cognitive biases as the judgmental kin of speech errors ("I cossed the toin"), forgetting, and optical illusions: These are systematic errors which, even if rare, are useful for illuminating how cognitive mechanisms work. But these errors do not imply the mechanisms fail frequently or are not well-adapted for everyday use. But as Kahneman and Tversky (1982, p.<br><br> 494) wrote, "Although errors of judgment are but a method by 8 However, this cbase-rate fallacy d is being thoughtfully re-examined (e.g., Koehler, 1996). The fact that base rates are more clearly included when subjects are asked what fraction of 100 hypothetical cases fit the profile is an important clue about how the heuristic operates and its limits (Gigerenzer, Hell & Blank, 1988; Tversky and Kahneman, 1983). 12 which some cognitive processes are studied, the method has become a significant part of the message." The shift in emphasis from the heuristics to the biases they sometimes create happened gradually as research moved to applied areas; the revisionist view that heuristics may be near- optimal is largely a critique (a reasonable one) of the later applied research.<br><br> Progress in modeling and applying behavioral models of judgment has lagged behind other areas, such as loss aversion and hyperbolic time discounting. A promising recent modeling approach is cquasi-Bayesian d 4viz., assume that people misspecify a set of hypotheses, or encode new evidence incorrectly, but otherwise use Bayes 9 rule. For example, Rabin and Schrag (1999) model "confirmation bias" by assuming that people who believe hypothesis A is more likely than B will never encode pro-A evidence mistakenly, but will sometimes encode pro-B evidence as being supportive of A.<br><br> 9 Rabin (2002) models the "law of small numbers" in a quasi- Bayesian fashion by assuming that people mistakenly think a process generates draws from a hypothetical "urn" without replacement , although draws are actually independent (i.e., made with replacement). He shows some surprising implications of this misjudgment. For example, investors will think there is wide variation in skill of, say, mutual-fund managers, even if there is no variation at all.<br><br> (A manager who does well several years in a row is a surprise if performance is mistakenly thought to be mean-reverting due to "nonreplacement", so quasi-Bayesians conclude that the manager must be really good.) Barberis, Shleifer and Vishny (1998) adopt such a quasi-Bayesian approach to explain why the stock market under-reacts to information in the short-term and overreacts in the long- term. In their model, earnings follow a random walk but investors believe, mistakenly, that earnings have positive momentum in some regimes and regress toward the mean in others. After one or two periods of good earnings, the market can 9t be confident that momentum exists and hence expects mean-reversion; but since earnings are really a random walk, the market is too pessimistic and is underreacting to good earnings news.<br><br> After a long string of good earnings, however, the market believes momentum is building. Since it isn 9t, the market is too optimistic and overreacts. 9 This encoding asymmetry is related to "feature-positive" effects and perceptual encoding biases well documented in research on perception.<br><br> After buying a Volvo you will suddenly "see" more Volvos on the road, due purely to heightened familiarity. 13 While other approaches that find ways of formalizing some of the findings of cognitive psychology are possible, our guess is that the quasi-Bayesian view will quickly become the standard way for translating the cognitive psychology of judgment into a tractable alternative to Bayes 9 rule. The models mentioned in the two paragraphs above are parameterized in such a way that the Bayesian model is embedded as a special case, which allows theoretical insight and empirical tests about how well the Bayesian restriction fits.<br><br> Preferences: Revealed, constructed, discovered, or learned? Standard preference theory incorporates a number of strong and testable assumptions. For example, it assumes that preferences are "reference independent" 3 i.e., are not affected by the individual 9s transient asset position.<br><br> It also assumes that preferences are invariant with respect to superficial variations in the way that options are described, and that elicited preferences do not depend on the precise way that preferences are measured as long as the method used is "incentive compatible" 3 i.e., provides incentives for people to reveal their "true" preferences. All of these assumptions have been violated in significant ways (see Slovic, 1995). For example, numerous "framing effects" show that the way that choices are presented to an individual often determine the preferences that are "revealed." The classic example of a framing effect is the "Asian disease" problem in which people are informed about a disease that threatens 600 citizens and asked to choose between two undesirable options (Tversky & Kahneman, 1981).<br><br> In the "positive frame" people are given a choice between (A) saving 200 lives for sure, or (B) a 1/3 chance of saving all 600 with a 2/3 chance of saving no one. In the "negative frame" people are offered a choice between (C) 400 people dying for sure, or (D) a 2/3 chance of 600 dying and a 1/3 chance of no one dying. Despite the fact that A and C, and B and D, are equivalent in terms of lives lost or at risk, most people choose A over B but D over C.<br><br> Another phenomenon that violates standard theory is called an "anchoring effect." The classic demonstration of an anchoring effect (Tversky & Kahneman, 1974, and in this volume) was identified in the context of judgment rather than choice. Subjects were shown the spin of a wheel of fortune that could range between 0 and 100 and were asked to guess whether the number of African nations in the United Nations was greater than or less than this number. They were then asked to guess the true value.<br><br> Although the wheel of fortune was obviously random, subjects 9 guesses were strongly influenced by the spin of the wheel. As Kahneman and Tversky 14 interpreted it, subjects seemed to "anchor" on the number spun on the wheel and then adjusted for whatever else they thought or knew, but adjusted insufficiently. Of interest in this context is that anchoring effects have also been demonstrated for choices as opposed to judgments.<br><br> In one study, subjects were asked whether their certainty equivalent for a gamble was greater than or less than a number chosen at random and then were asked to specify their actual certainty equivalent for the gamble (Johnson & Schkade, 1989). Again, the stated values were correlated significantly with the random value. In a recent study of anchoring, Ariely, Loewenstein and Prelec (in press) sold valuable consumer products (a $100 wireless keyboard, a fancy computer mouse, bottles of wine, and a luxurious box of chocolate) to postgraduate (MBA) business students.<br><br> The students were presented with a product and asked whether they would buy it for a price equal to the last two digits of their own social security number (a roughly random identification number required to obtain work in the United States) converted into a dollar figure 3 e.g., if the last digits were 79 the hypothetical price was $79. After giving a yes/no response to the question cWould you pay $79?, subjects were asked to state the most they would pay (using a procedure that gives people an incentive to say what they really would pay). Although subjects were reminded that the Social Security number is essentially random, those with high numbers were willing to pay more for the products.<br><br> For example, subjects with numbers in the bottom half of the distribution priced a bottle of wine-- a 1998 Cotes du Rhone Jaboulet Parallel 845 9 3 at $11.62, while those with numbers in the top half priced the same bottle at $19.95. Many studies have also shown that the method used to elicit preferences can have dramatic consequences, sometimes producing "preference reversals"-- situations in which A is preferred to B under one method of elicitation, but A is judged as inferior to B under a different elicitation method (e.g., Grether & Plott, 1979). The best known example contrasts how people choose between two bets versus what they separately state as their selling prices for the bets.<br><br> If bet A offers a high probability of a small payoff and bet B offers a small probability of a high payoff, the standard finding is that people choose the more conservative A bet over bet B when asked to choose, but are willing to pay more for the riskier bet B when asked to price them separately. Another form of preference reversal occurs between joint and separate evaluations of pairs of goods (Hsee et al, 1999; see Hsee & LeClerc, 1998, for an application to marketing). People will often price or otherwise evaluate an item A higher than another item B when the two 15 are evaluated independently, but evaluate B more highly than A when the two items are compared and priced at the same time.<br><br> "Context effects" refer to ways in which preferences between options depend on what other options are in the set (contrary to "independence of irrelevant alternatives" assumptions). For example, people are generally attracted to options that dominate other options (Huber, Payne & Puto, 1982). They are also drawn disproportionately to "compromise" alternatives whose attribute values lie between those of other alternatives (Simonson & Tversky, 1992).<br><br> All of the above findings suggest that preferences are not the pre-defined sets of indifference curves represented in microeconomics textbooks. They are often ill-defined, highly malleable and dependent on the context in which they are elicited. Nevertheless, when required to make an economic decisions 4to choose a brand of toothpaste, a car, a job, or how to invest 4 people do make some kind of decision.<br><br> Behavioral economists refer to the process by which people make choices with ill-defined preferences as "constructing preferences" (Payne, Bettman & Johnson, 1992; Slovic, 1995). A theme emerging in recent research is that, although people often reveal inconsistent or arbitrary preferences, they typically obey normative principles of economic theory when it is transparent how to do so. Ariely, Loewenstein and Prelec (in press) refer to this pattern as "coherent arbitrariness" and illustrate the phenomenon with a series of studies in which the amount subjects demanded to listen to an annoying sound is sensitive to an arbitrary anchor, but they also demand much more to listen to the tone for a longer period of time.<br><br> Thus, while expressed valuations for one unit of a good are sensitive to an anchor which is clearly arbitrary, subjects also obey the normative principle of adjusting those valuations to the quantity 3 in this case the duration -- of the annoying sound. Most evidence that preferences are constructed comes from demonstrations that some feature that should not matter actually does. The way gambles are "framed" as gains and losses from a reference outcome, the composition of a choice set, and whether people choose among objects or value them separately, have all been shown to make a difference in expressed preference.<br><br> But admittedly, a list of a theory 9s failings is not an alternative theory. So far, a 16 parsimonious alternative theory has not emerged to deal with all of these challenges to utility maximization. 10 Overview of the Book In what follows, we review different topic areas of behavioral economics to place chapters of the book into context.<br><br> The book is organized so that early chapters discuss basic topics such as decision making under risk and intertemporal choice, while later chapters provide applications of these ideas. Reference-dependence and loss aversion In classical consumer theory, preferences among different commodity bundles are assumed to be invariant with respect to an individual 9s current endowment or consumption. Contrary to this simplifying assumption, diverse forms of evidence point to a dependence of preferences on one 9s reference point (typically the current endowment).<br><br> Specifically, people seem to dislike losing commodities from their consumption bundle much more than they like gaining other commodities. This can be expressed graphically as a kink in indifference curves at the current endowment point (Knetsch, 1992; Tversky & Kahneman, 1991). In the simplest study showing reference-dependence, Knetsch (1992) endowed some subjects randomly with a mug, while others received a pen.<br><br> 11 Both groups were allowed to switch their good for the other at a minimal transaction cost, by merely handing it to the experimenter. If preferences are independent of random endowments, the fractions of subjects swapping their mug for a pen and the fraction swapping their pen for a mug should add to roughly one. In fact, 22% of subjects traded.<br><br> The fact that so few chose to trade implies an exaggerated preference for the good in their endowment, or a distaste for losing what they have. A seminal demonstration of an "endowment effect" in buying and selling prices was conducted by Kahneman et al (1990). They endowed half of the subjects in a group with coffee mugs.<br><br> Those who had mugs were asked the lowest price at which they would sell. Those who did 10 Some specialized models have been proposed to explain particular phenomena, such as Hsee, Loewenstein, Blount & Bazerman, 1999; Prelec, Wernerfelt & Zettelmeyer, 1997; Tversky, Slovic & Kahneman, 1990. 11 Note that any possible information value from being given one good rather than the other is minimized because the endowments are random, and subjects knew that half the others received the good they didn 9t have.<br><br> 17 not get mugs were asked how much they would pay. There should be essentially no difference between selling and buying prices. In fact, the median selling price was $5.79 and the median buying price was $2.25, a ratio of more than two: one which has been repeatedly replicated.<br><br> Although calibrationally entirely implausible, some economists were concerned that the results could be driven by cwealth effects d 4those given mugs are wealthier than those not given mugs, and this might make them value mugs more and money less. But in a different study reported in the same paper, the selling prices of one group were compared to the "choosing" prices of another: For a series of money amounts, subjects chose whether they would prefer to have a mug or money. The median choosing price was half the median selling price ($3.50 versus $7.00).<br><br> Choosers are in precisely the same wealth position as sellers 4they choose between a mug or money. The only difference is that sellers are "giving up" a mug they "own," whereas choosers are merely giving up the right to have a mug. Any difference between the two groups cannot be attributed to wealth effects.<br><br> Kahneman et al's work was motivated in part by survey evidence from "contingent valuation" studies that attempt to establish the dollar value of goods which are not routinely traded. Contingent valuation is often used to do government cost-benefit analysis or establish legal penalties from environmental damage. These surveys typically show very large differences between buying prices (e.g., paying to clean up oily beaches) and selling prices (e.g., having to be paid to allow beaches to be ruined).<br><br> Sayman and Öncüler (1997) summarize 73 data sets which show selling-to-buying ratios ranging from .67 (for raspberry juice) to 20 or higher (for density of trees in a park and health risks). Loss aversion has already proved to be a useful phenomenon for making sense of field data (see Camerer, 2000, and this volume). Asymmetries in demand elasticities after price increases and decreases (Hardie, Johnson, & Fader, 1993), the tendency for New York City cab drivers to quit early after reaching a daily income target (producing surprising upward-sloping labor supply curves; see Camerer et al, 1997 and in this volume), and the large gap between stock and bond returns (the "equity premium"; see Benartzi & Thaler, 1995, in this volume) can all be explained by models in which agents have reference-dependent preferences and take a short planning horizon, so that losses are not integrated against past or future gains.<br><br> A particularly conclusive field study by Genoseve and Mayer (2001, and this volume) focuses on the real estate market. (Housing is a huge market 4worth $10 trillion at the time of 18 their study, a quarter of the wealth in the US 4and full of interesting opportunities to do behavioral economics.) They find that list prices for condominiums in Boston are strongly affected by the price at which the condominium was purchased. Motivated sellers should, of course, regard the price they paid as a sunk cost and choose a list price that anticipates what the market will pay.<br><br> But people hate selling their houses at a nominal loss from the purchase price. Sellers 9 listing prices and subsequent selling behavior reflects this aversion to nominal losses. Odean (1998) finds the same effect of previous purchase price in stock sales.<br><br> 12 At least three features of endowment effects remain open to empirical discussion. First, do people anticipat e the endowment effect? The answer seems to be "No": Loewenstein and Adler (1995) found that subjects did not anticipate how much their selling prices would increase after they were endowed with mugs.<br><br> 13 Van Boven, Dunning and Loewenstein (2000) and Van Boven, Loewensstein and Dunning (2000) found that agents for buyers also underestimated how much sellers would demand. Second, Kahneman, Knetsch and Thaler (1990:1328) note that "there are some cases in which no endowment effect would be expected, such as when goods are purchased for resale rather than for utilization." However, the boundary of commercial non-attachment has not been carefully mapped. Do art or antique dealers "fall in love" with pieces they buy to resell?<br><br> What about surrogate mothers who agree to bear a child for a price paid in advance? Evidence on the degree of commercial attachment is mixed. In their housing study, Genesove and Mayer (2001 and this volume) note that investors who don 9t live in their condos exhibit less loss-aversion than owners.<br><br> A field experiment by List (in press) found that amateur sports paraphernalia collectors who do not trade very often showed an endowment effect, but professional dealers and amateurs who trade a lot did not. 14 An example where attachment seemed important even among 12 Though it is harder to unambiguously interpret as loss aversion in the sense we are discussing here, reference points can also serve as social focal points for judging performance. Degeorge, Patel & Zeckhauser (1999) document an interesting example from corporate finance.<br><br> Managers whose firms face possible losses (or declines from a previous year 9s earnings) are very reluctant to report small losses. As a result, the distribution of actual losses and gains shows a very large spike at zero, and hardly any small reported losses (compared to the number of small gains). Wall Street hates to see a small loss.<br><br> A manager who does not have the skill to shift accounting profits to erase a potential loss (i.e., "has some earnings in his pocket") is considered a poor manager. In this example, the market 9s aversion to reported losses can serve as a signaling device which tells the markets about managerial ability. 13 Failure to anticipate the strength of later loss-aversion is one kind of "projection bias" (Loewenstein, O'Donoghue & Rabin, 1999), in which agents are make choices as if their current preferences or emotions will last longer than they actually do.<br><br> 14 By revisiting the same traders a year later, List showed that it was trader experience which reduced endowment effects, rather than self-selection (i.e., people who are immune to such effects become dealers.) 19 experienced traders with high incentives was described by an investment banker who said his firm combats loss-aversion by forcing a trader to periodically switch his "position" (the portfolio of assets the trader bought and is blamed or credited for) with the position of another trader. Switching ensures that traders do not make bad trades because of loss-aversion and emotional attachment to their past actions (while keeping the firm 9s net position unchanged, since the firm 9s total position is unchanged). Third, it is not clear the degree to which endowment effects are based solely on the current endowment, rather than past endowments or other reference points.<br><br> Other reference points, such as social comparison (i.e., the possessions and attainments of other people) and past ownership, may be used to evaluate outcomes. How multiple reference points are integrated is an open question. Strahilevitz and Loewenstein (1998) found that the valuation of objects depended not only on whether an individual was currently endowed with an object, but on the entire past history of ownership 3 how long the object had been owned or, if it had been lost in the past, how long ago it was lost and how long it was owned before it was lost.<br><br> These "history- of-ownership effects" were sufficiently strong that choice prices of people who had owned for a long period but just lost an object were higher than the selling prices of people who had just acquired the same object. If people are sensitive to gains and losses from reference points, the way in which they combine different outcomes can make a big difference. For example, a gain of $150 and a loss of $100 will seem unattractive if they are evaluated separately, if the utility of gains is sufficiently less than the disutility of equal-sized losses, but the gain of $50 that results when the two figures are added up is obviously attractive.<br><br> Thaler (1980, 1999 and this volume) suggests that a useful metaphor for describing the rules which govern gain/loss integration is cmental accounting d 4 people set up mental accounts for outcomes which are psychologically separate, much as financial accountants lump expenses and revenues into separated accounts to guide managerial attention. Mental accounting stands in opposition to the standard view in economics that "money is fungible"; it predicts, accurately, that people will spend money coming from different sources in different ways (O'Curry, 1999), and has wide-ranging implications for such policy issues as how to promote saving (see, e.g., Thaler, 1994). A generalization of the notion of mental accounting is the concept of "choice bracketing," which refers to the fashion in which people make decisions narrowly, in either a piece-meal 20 fashion, or broadly 3 i.e., taking account of interdependencies between decisions (Read, Loewenstein & Rabin, 1999).<br><br> How people bracket choices has far-reaching consequences in diverse areas, including finance (see Bernartzi & Thaler, 1995 and chapter 22), labor supply (Camerer, Babcock, Loewenstein & Thaler, 1997, and chapter 19), and intertemporal choice (Frederick, Loewenstein & O'Donoghue, in press and chapter 6, section 5.3.4). For example, when making many separate choices between goods, people tend to choose more diversity when the choices are bracketed broadly than when they are bracketed narrowly. This was first demonstrated by Simonson (1990), who gave students their choice of one of six snacks during each of three successive weekly class meetings.<br><br> Some students chose all three snacks in the first week, although they didn't receive their chosen snack until the appointed time, and others chose each snack on the day that they were to receive it (narrow bracketing; sequential choice). Under broad bracketing, fully 64% chose a different snack for each week, as opposed to only 9% under narrow bracketing. Follow-up studies demonstrated similar phenomena in the field (e.g., in purchases of yogurt; Simonson & Winer, 1992).<br><br> Bracketing also has implications for risk-taking. When people face repeated risk decisions, evaluating those decisions in combination can make them appear less risky than if they are evaluated one at a time. Consequently, a decision maker who refuses a single gamble may nonetheless accept two or more identical ones.<br><br> By assuming that people care only about their overall level of wealth, expected utility theory implicitly assumes broad bracketing of risky decisions. However, Rabin (2000) points out the absurd implication which follows from this assumption (combined with the assumption that risk aversion stems from the curvature of the utility function): A reasonable amount of aversion toward risk in small gambles implies a dramatic aversion to reduction in overall wealth. For example, a person who will turn down a coin flip to win $11 and lose $10 at all wealth levels must also turn down a coin flip in which she can lose $100, no matter how large the possible gain is .<br><br> 15 Rabin 9s proof is a mathematical 15 The intuition behind Rabin 9s striking result is this: In expected-utility theory, rejecting a (+$11,-$10) coin flip at wealth level W implies that the utility increase from the $11 gain is smaller than the total utility decrease from the $10 loss, meaning that the marginal utility of each dollar gained is at most 10/11 of the marginal utility of each dollar lost. By concavity, this means that the marginal utility of the W+11 th dollar is at most 10/11 the marginal utility of the W-10 th dollar 4a sharp 10% drop in marginal utility for small change in overall wealth of $21. When the curvature of the utility function does not change unrealistically over ranges of wealth levels, this means the marginal utility plummets quickly as wealth increases 4the marginal utility of the W+$32 dollar (= W+11 + 21) can be at most (10/11)(10/11), which is around 5/6 of the marginal utility of the W-10 th dollar.<br><br> Every $21 decrease in wealth yields another 10% decline in marginal utility. This implies, mathematically, that implying a person 9s value for a dollar if he were $500 or $1,000 wealthier would be tiny compared to how much he values dollars he might lose in a 21 demonstration that people who are averse to small risks are probably not integrating all their wealth into one source when they think about small gambles. Preferences over risky and uncertain outcomes The expected-utility EU hypothesis posits that the utility of a risky distribution of outcomes (say, monetary payoffs) is a probability-weighted average of the outcome utilities.<br><br> This hypothesis is normatively appealing because it follows logically from apparently reasonable axioms, most notably the independence (or ccancellation d) axiom. The independence axiom says that if you are comparing two gambles, you should cancel events which lead to the same consequence with the same probability; your choice should be independent of those equally- likely common consequences. Expected utility also simplifies matters because a person 9s taste for risky money distributions can be fully captured by the shape of the utility function for money.<br><br> Many studies document predictive failures of expected utility in simple situations in which subjects can earn substantial sums of money from their choices. 16 Starmer 9s (2000) contribution to this volume reviews most of these studies, as well as the many theories that have been proposed to account for the evidence (see also Camerer, 1989b, 1992; Hey, 1997; Quiggin, 1993). Some of these new theories alter the way in which probabilities are weighted, but preserve a "betweenness" property which says that if A is preferred to B, then any probabilistic gamble between them must be preferred to B but dispreferred to A (i.e., the gambles lie cbetween d A and B in preference).<br><br> Other new theories suggest that probability weights are "rank-dependent" 4outcomes are first ranked, then their probabilities are weighted in a way which is sensitive to how they rank within the gamble that is being considered. One mathematical way to do this is transform the cumulative probabilities of outcomes (i.e., the chance that you will win X or less) nonlinearly and weight outcome utilities by the differences of bet. So if a person 9s attitude towards gambles really came from the utility-of-wealth function, even incredibly large gains in wealth would not tempt her to risk $50 or $100 losses, if she really dislikes losing $10 more than she likes gaining $11 at every level of wealth.<br><br> 16 Some of the earlier studies were done with hypothetical payoffs, leading to speculation that the rejection of EU would not persist with real stakes. Dozens of recent studies show that, in fact, paying real money instead of making outcomes hypothetical appears either fails to eliminate EU rejections, or strengthens the rejections of EU (because sharper results which come from greater incentive imply that rejections are more statistically significant; Harless & Camerer, 1994). 22 those weighted cumulative probabilities.<br><br> 17 The best known theory of this sort is cumulative prospect theory (Tversky & Kahneman, 1992). There are three clear conclusions from the experimental research (Harless & Camerer, 1994). One is that of the two new classes of theories that allow more general functional forms than expected utility, the new rank-dependent theories fit the data better than the new betweenness class theories.<br><br> A second conclusion is that the statistical evidence against EU is so overwhelming that it is pointless to run more studies testing EU against alternative theories (as opposed to comparing theories with one-another). The third conclusion is that EU fits worst when the two gambles being compared have different sets of possible outcomes (or "support"). Technically, this property occurs when one gamble has a unique outcome.<br><br> The fact that EU does most poorly for these comparisons implies that nonlinear weighting of low probabilities is probably a major source of EU violations. Put differently, EU is like Newtonian mechanics, which is useful for objects traveling at low velocities but mispredicts at high speeds. Linear probability weighting in EU works reasonably well except when outcome probabilities are very low or high.<br><br> But low-probability events are important in the economy, in the form of cgambles d with positive skewness (lottery tickets, and also risky business ventures in biotech and pharmaceuticals), and catastrophic events which require large insurance industries. Prospect theory (Kahneman & Tversky, 1979) explains experimental choices more accurately than EU because it gets the psychophysics of judgment and choice right. It consists of two main components: a probability weighting function, and a 'value function' which replaces the utility function of EU.<br><br> The weighting function À (p) combines two elements: (1) The level of probability weight is a way of expressing risk tastes (if you hate to gamble, you place low weight on any chance of winning anything); and (2) the curvature in À (p) captures how sensitive people are to differences in probabilities. If people are more sensitive in the neighborhoods of possibility and certainty 4i.e., changes in probability near zero and 1 4 than to intermediate gradations, then their À (p) curve will overweight low probabilities and underweight high ones. 17 A technical motivation for crank dependent d theories-- ranking outcomes, then weighting their probabilities-- is that when separate probabilities are weighted, it is easy to construct examples in which people will violate dominance by choosing a cdominated d gamble A which has a lower chance of winning at each possible outcome amount, compared to the higher chance of winning the same outcome amount for a dominant gamble B.<br><br> If people rarely choose such dominated gambles, they are acting as if they are weighting the differences in cumulated probabilities, which is the essence of the rank-dependent approaches. 23 The value function reflects the insight, first articulated by Markowitz (1952), that the utility of an outcome depends not on the absolute level of wealth that results but on whether the outcome is a gain or a loss. Prospect theory also assumes reflection of risk-preferences at the reference point: People are typically averse to risky spreading of possible money gains, but will take gambles where they could lose big or break even rather than accept a sure loss.<br><br> Prospect theory also assumes "loss-aversion": The disutility of a loss of x is worse than the utility of an equal-sized gain of x. Expected utility is restricted to gambles with known outcome probabilities. The more typical situation in the world is "uncertainty", or unknown (subjective, or personal) probability.<br><br> Savage (1954) proposed a subjective expected utility (SEU) theory in which choices over gambles would reveal subjective probabilities of states, as well as utilities for outcomes. Ellsberg (1961) quickly pointed out that in Savage 9s framework, subjective probabilities are slaves to two masters-- they are used as decision weights applied to utilities, and they are expressions of likelihood. As a result, there is no way to express the possibility that, because a situation may have lots of "ambiguity," one is reluctant to put much decision weight on any outcome.<br><br> Ellsberg demonstrated this problem in his famous paradox: Many people prefer to bet on black drawn from an urn with 50 black and 50 red balls, rather than bet on black drawn from an urn with 100 balls of unknown black and red composition, and similarly for red (they just don 9t want to bet on the unknown urn). There is no way for the two sets of red and black subjective probabilities from each urn to both add to one (as subjective probabilities require), and still express the distaste for betting neither color in the face of ambiguity. Many theories have been proposed to generalize SEU to allow for ambiguity-aversion (see Camerer & Weber, 1992, for a review).<br><br> One approach, first proposed by Ellsberg, is to let probabilities be sets rather than specific numbers, and assume that choices over gambles reveal whether people pessimistically believe the worst probabilities are the right ones, or the opposite. Another approach is to assume that decision weights are nonadditive. For example, the weights on red and black in the Ellsberg unknown urn could both be .4; the missing weight of .2 is a kind of "reserved belief" which expresses how much the person dislikes betting when she knows that important information is missing.<br><br> Compared to non-EU theories, relatively little empirical work and applications have been done with these uncertainty-aversion theories so far. Uncertainty-aversion might explain 24 phenomena like voting "rolloff" (when a voter, once in the voting booth, refuses to vote on obscure elections in which their vote is most likely to prove pivotal; Ghirardato & Katz, 2000), incomplete contracts (Mukherji, 1998) and "home country bias" in investing: People in every country overinvest in the country they are most familiar with-- their own. (Finnish people invest in firms closer to their own town, see Grinblatt & Keloharju, 2001.) In asset pricing, ambiguity-aversion can imply that asset prices satisfy a pair of Euler inequalities, rather than an Euler equation, which permits asset prices to be more volatile than in standard theory (Epstein & Wang, 1994).<br><br> Hansen, Sargent and Tallarini (1999) have applied related concepts of "robust control" to macroeconomic fluctuation