An experimental test of criminal behavior among juveniles and young adults: GARP among thieves

Michael S. Visser University of Oregon Eugene, OR 97403-1285 mvisser1@uoregon.edu Bill Harbaugh University of Oregon and NBER Eugene, OR 97403-1285 bill@harbaugh.org Naci Mocan University of Colorado at Denver and NBER Campus Box 181, P.O. Box 173364 Denver CO 80217-3364 nmocan@carbon.cudenver.edu

October, 2004 (draft, please do not cite)

JEL classifications: K42, D10, C90 Abstract: Gary Becker's (1968) model of rational criminal behavior forms the basis of how economists think about crime. In this paper we report results from economic experiments that provide a direct test of the primary hypothesis of this model: that criminal behavior responds rationally to changes in the possible rewards and in the probability and severity of punishment. We find that the data are generally consistent with revealed preference axioms, monotonic choice, and the law of demand. These results strengthen the argument that criminal behavior, and the response of criminals to changes in enforcement and penalties, can be accounted for by economic models. Acknowledgements: We would like to thank the students, teachers, and administrators of South Eugene High School for their assistance, without which this research would not have been possible. This research was supported by a grant from the NSF.

An experimental test of criminal behavior among juveniles and young adults: GARP among thieves

I. Introduction
Since the seminal paper of Becker (1968), which created the foundation for the economic analysis of criminal behavior, economists have extended the basic theoretical framework (e.g. Ehrlich 1973, Block and Heineke 1975, Schmidt and Witte 1984, Flinn 1986, Lochner 2004, Mocan et al., forthcoming). The original framework, as well as its more recent variants, postulate that participation in crime is the result of an optimizing individuals response to incentives such as the expected payoffs from criminal activity, and costs such as the probability of apprehension and the severity of punishment. Although early empirical research reported evidence suggesting that enhanced deterrence reduces crime (Ehrlich 1975, Witte 1980, Layson 1985), other papers found no significant evidence of deterrence (Myers 1983, Cornwell and Trumbull 1994). The main challenge in empirical analysis has been to tackle the simultaneity between criminal activity and deterrence. Specifically, an increase in criminal activity is expected to prompt an increase in the certainty and severity of punishment (e.g. an increase in the arrest rate and/or police effort), which makes it difficult to identify the causal impact of deterrence on crime. Three types of strategies have been used to overcome the simultaneity problem. The first solution is to find a good instrument which is correlated with deterrence measures but uncorrelated with crime. Examples

are Levitt (1997) which uses electoral cycles as an instrument for police hiring and Levitt (2002) which uses the number of per capita municipal firefighters as an instrument for police effort. The second strategy is to use high-frequency time-series data. For example, in monthly data, an increase in police effort in a given month will affect criminal activity in the same month, but an increase in crime is much less likely to alter the size of the police force in that same month, because of the much longer lag between a policy decision to increase the ranks and the actual deployment of police officers on the street. This identification strategy has been employed by Corman and Mocan (forthcoming) and Corman and Mocan (2000). The third strategy is to find a natural experiment which generates a truly exogenous variation in deterrence, as in Di Tella and Schargrodsky (2004) who use the increase in police protection around Jewish institutions in Buenos Aires after a terrorist attack to identify the impact of police presence on car thefts. Although these empirical strategies have permitted researchers to refine and improve upon earlier estimates, a convincing natural experiment is very difficult to find. The validity of any instrumental variable can always be questioned, and one can argue that if policy makers have perfect foresight about future crime, monthly data would also suffer from simultaneity. One can also argue that unobservable effort by the existing complement of police officers might respond quickly to changes in crime rates even if hiring does not. One application of Beckers model that has received a great deal of attention concerns tax compliance. By noting that tax compliance is a problem of law 2

enforcement, we can see that the regulatory agency is faced with a tradeoff between compliance and costly enforcement. Beckers model can aid these regulatory agencies in optimally solving this problem by informing decision makers about the different deterrent effects associated with penalties and the probability of detection. Theoretical models of tax compliance and enforcement follow Becker (1968) and model agents as expected utility maximizers. Early examples include Allingham and Sandmo (1972) and Srinivisan (1973). A central prediction of these models is a negative correlation between tax rates and compliance, and a positive correlation between income and compliance. Empirical tests of these predictions, however, have been sparse and somewhat inconclusive, since any data clearly suffers from the same general simultaneity problem outlined above. Feinstein (1991) pools data from the 1982 and 1985 IRS Taxpayer Compliance Measurement Program in an effort to better identify the effects of tax rate changes and income on compliance. However, the results from the pooled study suggest a positive relationship between tax rates and compliance and no relationship between income and compliance. A battery of laboratory experiments have also been applied to the problem of tax compliance. Generally speaking, experimental studies find that greater tax rates are associated with lower compliance levels (see for example Friedland, Maital, and Rutenberg 1978; Alm, Jackson, and McKee 1992; and Baldry 1987). Clark, Friesen, and Muller (2004) design an experiment to analyze the tradeoffs between two different conditional audit rules and the simple random audit rule. As expected, they find that 3

the simple audit rule maximizes compliance relative to the conditional audit rules, but is also the most costly to implement of the three audit rules. Boylan and Sprinkle (2001) conduct an experiment designed to determine whether the manner in which income is obtained affects the relation between tax rates and taxpayer compliance. They find that when income is earned (versus endowed), compliance rates increase when agents face a tax rate increase, and that the opposite is true when income is endowed. They conclude that income is not a fungible commodity, meaning that money earned is valued differently than is money endowed. Each of these studies describes behavior of optimizing agents responding rationally to the incentives put before them. While this vein of research provides some guidance for tax compliance authorities, the results have not been generalized to accommodate other types of criminal behavior, such as auto theft, embezzlement, and petty larceny. In this paper, we use a laboratory experiment to collect data on responses to unambiguously exogenous changes in the rewards and penalties pertaining to criminal behavior. The experiments involve decisions about actions that can best be described as petty larceny, and are done using high school students and real money. We use a straightforward protocol for collecting choice data that can be used to test Beckers model directly. The protocol involves collecting data on (nearly) simultaneous choices under a variety of different budget constraints. We use these to check whether peoples choices about their criminal behavior change rationally in response to changes in the probability of detection and changes in the size of the fine assessed. 4

The data are first used to check for transitivity violations. Transitivity would indicate that choices over the goods can be modeled as the result of constrained utility maximization. This provides a direct test of Beckers model, which assumes rational choice by criminals. Next, we estimate demand functions. Interestingly, Becker (1962) points out that rational choice is not necessary for choices to satisfy the laws of demand. More fundamentally, aggregate choices may obey the laws of demand even if some individual choices are inconsistent with utility maximization. Therefore, we expect to be able to provide results about these tradeoffs even if choices occasionally, or even frequently, violate the Generalized Axiom of Revealed Preference (GARP). Section II of this paper offers a discussion of the relevant literatures, including Beckers model of crime and punishment, revealed preference theory, and the experimental methods for testing it. We then describe our experiment and subjects, and analyze our data for consistency with GARP. Next we estimate demand functions for the amount of loot stolen. A discussion section concludes.

II. Literature
II. a. Beckers model of crime and punishment Just as the typical consumer has preferences for food, entertainment, and leisure, the rational criminal has preferences for stolen loot, the probability of being caught, and the penalty associated with being caught. It is at least plausible that a criminal would be willing to forgo some of the loot in order to reduce the probability of being caught or the penalty associated with it. In a like manner, a criminal might be willing to take

great risk only if the payoff is large enough. There is nothing irrational about this kind of behavior. The agent makes a cost-benefit calculation according to her own preferences. While we cannot observe a criminals utility function directly, we can test Beckers hypothesis of rationality with tests of revealed preference, using observations of agents different choices under different budget constraints. In the next section we review the relevant theory and some tests of it. II. b. Revealed preference theory The revealed preference approach to consumer behavior involves determining what restrictions on choice behavior are imposed by assuming people chose rationally. Integrability proofs are then used to show that choice data which are consistent with the these restrictions are also consistent with maximization of a well-behaved classical utility function, and vice versa. Paul Samuelson (1938a,b) provided the first empirically useful characterization of consumer behavior based on the observation of consumer choices. This redirection of consumer theory advanced the view that, since preferences are unobservable, utility theory was not falsifiable. (Samuelson, 1938a; p. 61):

The discrediting of utility as a psychological concept robbed it of its only possible virtue as an explanation of human behaviour in other than a circular sense, revealing its emptiness as even a construction.

Samuelson proposed an alternative view of consumer behavior, treating consumer choice, as fundamental, instead of relying on innately unobservable assumptions about preferences. The behavioral restrictions that Samuelson proposed are now known as the Weak Axiom of Revealed Preference (WARP). Using the notation of Varian (1992), these restrictions can be stated as follows:

If x t R D x s and x t is not equal to x s , then it is not the case that x s R D x t .

Where x t and x s are consumption vectors, and R D is read is directly revealed preferred to. In words, WARP says that if bundle x t is selected when both x t and x s are available, then x s cannot be selected when both x t and x s are available. This does not rule out the case where the consumer does not optimize nor does it rule out the case of indifference, but it is no more restrictive than the conventional utility maximizing hypothesis. In addition, it provides economists with an empirically tractable way to test the rationality postulate without needing to know the structure of the underlying preferences. It is important to note that WARP is only applicable to directly revealed preferences; that is, transitivity of preferences cannot be extended to bundles not directly compared to each other by the consumer. For example, if A is directly revealed preferred to B when A and B are available, and also B is directly revealed preferred to C when B and C are available, one cannot conclude simply by relying on WARP that A will be revealed preferred to C when only A and C are available.

Samuelson (1948) contains a proof showing that, by varying the relative price and income values that a consumer faces, and applying WARP, we can map out an indifference curve that would be generated by the constrained maximization of the consumers utility function: utility maximization implies the satisfaction of WARP.

u ( x t ) u ( x) x t R D x

This effectively keeps utility theory from being obsolete. However, it is still possible to find consumer choice data which satisfy WARP, but are not capable of being generated by a rational preference relation. H. S. Houthakker (1950) strengthened WARP to account for indirectly revealed preferences. His Strong Axiom of Revealed Preference (SARP) requires transitivity of preference, and is stated as follows:

If x t Rx s and x t is not equal to x s , then it is not the case that x s Rx t .

R is the transitive closure of R D , and is read is revealed preferred to. SARP implies
WARP. Houthakker is also responsible for showing that choice data satisfying SARP necessarily imply the existence of a well-behaved utility function, and that these data can be generated by structurally derived demand functions corresponding to that utility function. When combined with Samuelson (1948), produces the following result:

x t Rx x t R D x u ( x t ) u ( x) .

This was an important insight, but still only a necessary condition for choice data to be consistent with utility maximization. Houthakker also takes pains to point out in his discussion that Samuelsons revealed preference approach was originally conceived as a substitute for utility theory, but that it has instead evolved to be rather complementary.1 Sidney N. Afriat (1967) presented a theorem proving the sufficiency of SARP for utility maximization, thus completing the ingredients need for a solution to the integrability problem. Now we have an empirically useful and coherent description of consumer behavior, which can be linked to the existence of a well-behaved utility function. In other words, we can observe consumer choice, and as long as it is consistent with SARP, we know that there exists some well-behaved utility function that, when maximized, could have generated the choices we observe. Varian (1982) provides a generalization of WARP and SARP. The Generalized Axiom of Revealed Preference (GARP) can be stated as:

x t Rx s implies not x s P D x t . In other words, x t Rx s implies p s x s p s x t .

P D is read is strictly directly revealed preferred to. WARP and SARP each require
that the demanded bundle be unique, while GARP allows for more than one demanded

Samuelsons treatment of traditional utility theory in his Foundations (1947) can be taken as evidence of his acceptance of this reconciliation.

bundle. The following restatement of Afriats theorem in terms of GARP is given in Varian (1992), p. 133-4.

Afriats theorem. Let p t , x t for t = 1,..., T be a finite number of observations of price vectors and consumption bundles. Then the following conditions are equivalent.

(1) There exists a locally nonsatiated utility function that rationalizes the data; (2) The data satisfy GARP; (3) There exist positive numbers (u t , t ) for t = 1,..., T that satisfy the Afriat inequalities: u s u t + t p t x s x t for all t, s; (4) There exists a locally nonsatiated, continuous, concave, monotonic utility function that rationalizes the data.

Varian provides an algorithm for use in testing choice data for consistency with revealed preference axioms. The results from the evolution of revealed preference theory are significant in that they continue Samuelsons tradition of providing empirically useful tools that can be applied to consumer theory. Prior to revealed preference theory, utility maximization was an un-testable assumption of neoclassical economics. Now we can search for evidence in support of utility maximization via the axioms of revealed preference; we simply need some empirical choice data. Once we have verified that choice data satisfy GARP, we can estimate demand functions, which need not be structurally derived from any particular utility function.

10

II. c. Empirical tests of revealed preference theory Finding suitable choice data for testing whether choices obey the revealed preference axioms is not as straightforward as it would at first appear. The application of revealed preference axioms require consistency of preferences; that is, the underlying preferences must remain unchanged. If we observe choices over time but ignore this consistency requirement, we might find that the data appear not to obey revealed preference theory and conclude that our consumer is irrational. However, we cannot rule out the possibility that preferences have simply changed over the duration of our sampling period. Since we cannot observe the underlying preferences these two explanations are observationally equivalent in any set of choice data that has been collected over time. As a result of this hurdle, most empirical research applying revealed preference axioms has consisted of either field or laboratory experiments. Koo (1963), Mossin (1972), and Mattei (1994) are exceptions. Each of these studies employs a household expenditure survey. Households participating in the surveys are asked to write down their weekly purchases of consumption goods, aggregated up into categories. In general, these data are found to be inconsistent with the axioms of revealed preference. However, as previously asserted, there is no reason to believe that the preferences of these households have not evolved over time, thereby confounding the search for evidence of revealed preference. In addition, it is possible that households may not truthfully report their actual behavior in a survey, since there is no clear incentive for them to do so. In fact, there is evidence that many households consistently 11

underreport their purchases of consumption goods, and also that recall about price information is incomplete. Experiments allow researchers to directly control and observe prices and incomes, to control the environment wherein agents interact, and to provide sufficient incentives for agents to make choices consistent with their true preferences. These experiments are likely to produce significantly improved choice data compared to the surveys above. There are a small number of experiments which examine consumption bundles for consistency with the revealed preference axioms. Battalio, et al. (1973) observe purchases in a field experiment conducted in a psychiatric hospital. The experimenters controlled a token economy where various sundries and snacks were sold to patients, who paid using tokens which they earned by performing various tasks. The experimenters imposed large price changes every few weeks, and found that around half of the subjects made choices that violated SARP. However, the authors point to errors in the data collection that might explain nearly all of these inconsistencies. In a reanalysis of these data, Cox (1997) suggests that, given the nature of the token economy, leisure is also in the consumption bundle. When including leisure as a good, Cox finds that approximately half of the subjects have at least one violation, even when accounting for the data collection errors. The experiment also lacked good control, because patients had alternate sources of their goods, including family and another store selling similar items outside the hospital, and because preferences could change during the course of the experiment.

12

Sippel (1997) is the first example of a properly controlled laboratory experiment applying revealed preference axioms to consumption data. Subjects were asked to spend one hour in the laboratory, one at a time. While in the laboratory, the subject could consume only goods purchased from the experimenter, so the opportunity cost of consumption was essentially zero. Subjects were asked to choose among several snacks and time-passing goods, and to pay for them using an endowment of an artificial currency. Choices were made almost simultaneously, and the strategy method was employed, so that one choice was chosen randomly and then implemented.2 Price vectors were structured so as not to resemble any of the real price vectors one might face in local stores, in an attempt to avoid framing effects: a price might be interpreted as a signal of value, and thus bias choices. In each of two treatments nearly half or more subjects had at least one WARP, SARP, and/or GARP violation. The number of violations per subject was quite low, however. For example, the maximum number of GARP violations possible was 45 and the median number of GARP violations was 1.3 The demand analysis also supports the view that subjects do not choose randomly, and typically behave according to neoclassical predictions.

The strategy method involves multiple rounds of the same game, where the experimenter will randomly choose only one (or perhaps a strict subset) of the rounds to calculate the subjects earnings. In this case, the strategy method provides an incentive for subjects to treat each round as a single play of the game, and maximize expected utility over the entire random lottery. This may actually mitigate against the problem of changing or evolving preferences during the experiment. The strategy method is a common procedure in many experimental settings.
2 3

This is the number of violations reported in Sippel (1997). As noted previously, it is impossible to have exactly one revealed preference violation, so it must be the case that Sippel (1997) counts the number of violations differently than we do.

13

Harbaugh, Krause, and Berry (2001) apply the revealed preference test in a series of consumption games on children of varied ages (about 7 and 11 years old) and on undergraduates (about 21 years old). Children and undergraduates were offered choices among eleven different choice sets, each bundle consisting of a number of small bags of chips and boxes of juice. Choice sets were constructed using implicit price and income vectors which form overlapping budget sets. GARP typically assumes that preferences are locally non-satiated, but since subjects in this experiment made discrete choices this requirement must be strengthened to one of strongly monotone preferences to maintain the validity of the transitivity test. The strategy method was employed. Harbaugh et al. found that nearly half to three quarters of subjects exhibited at least one violation, though the average number of violations again tended to be relatively small. On average, even for the youngest children, there were significantly fewer (4.3) transitivity violations than would be expected under random choice (uniform 8.91, bootstrapping 8.29). Under the uniform random choice, each bundle was randomly selected with equal probability. Under the bootstrap random choice, each bundle was given a weight proportional to its frequency in the actual choices made by the experiment subjects. The average number of violations decreased in older children (2.1), but there as no significant difference between 11 year olds and undergraduates (2.0). While the authors discouraged generalization from their experiment, it does provide persuasive evidence that even children may be rational utility maximizers, at least with respect to relatively familiar consumption goods.

14

Harbaugh et al. also looked at whether children with different assessed levels of mathematical ability are better at choosing. They acquired student scores for a required standardized test on the 11 year olds in their sample (administered when these students were 10 year olds). Using the choice data, Harbaugh et al. also calculated Afriats Critical Cost Efficiency Index, which measures the costliness of revealed preference violations. One minus the index is the percentage of income that is wasted on a violation. This gives us an idea of how much one would have to relax the budget constraint in order to remove revealed preference violations. An index closer to one is associated with less costly violations. They regressed the number of GARP violations and Afriats Index on the math test score. While signs of the key slope coefficients are as expected children with better scores show fewer and smaller violations and their magnitudes are large, these coefficients are not statistically significant at standard levels. However, the number of observations is relatively small (just 37), a broader examination of this hypothesis may yield different results. Andreoni and Miller (2002) offered subjects opportunities to share an endowment with another anonymous partner. Implicit prices and incomes were varied across decisions to create intersecting budget sets, allowing researchers to check for GARP consistency. Over 98% of subjects exhibited utility maximizing choices. They also find that a majority exhibit altruistic behavior, in varying degrees. This stands in stark contrast to the predictions of neoclassical rationality that agents are only selfinterested in a very narrow sense; that is, they would never give up any resource without compensation. Andreoni and Miller further estimated three classes of standard 15

CES utility functions (perfect substitutes, perfect complements, perfectly selfish) and find that almost half of the subjects exhibited behavior that was exactly consistent with at least one of these utility functions. Becker (1968) extends neoclassical consumer theory to criminal behavior. Revealed preference theory provides us with an empirically useful strategy for verifying whether the neoclassical utility model can be applied to criminal behavior. If we find that criminal behavior is consistent with revealed preference axioms, then we know that economic models are capable of explaining criminal behavior and advising policy. Even if the data are not perfectly consistent with revealed preference axioms, we can still make policy recommendations based on aggregate data, since it is still likely to obey the laws of demand.

III. Experiment design

The complete protocol for our experiment is included in Appendix A. People are randomly and anonymously matched with a partner, and they each start with a $5 endowment. Everyone makes decisions as if they are the criminal, and we then randomly determine who is a criminal and who is a victim. The choice sets are constructed from budget sets defined over three goods: stolen loot, the probability of getting away with stealing, and the amount you keep, after repayment of the stolen loot and the fine, if you are caught.4 We construct multiple choice sets, each consisting of a

Since a fine is presumably a bad, we take a base fine and subtract from it a good that might be interpreted as a fine discount. The amount of this discount varies according to the tradeoffs imposed by

16

list of bundles of these three goods. The bundles can be thought of as different crimes; that is, some will involve taking a little money, facing a good chance of getting away with it, and a modest fine if caught, while others will involve a higher amount of loot, but a lesser chance of getting away with it, and so on. Taking nothing is always an option. The list of bundles for each choice set are constructed with different implicit incomes and prices. These prices can be viewed as the rates of tradeoff between loot, the probability of getting away with the theft, and the smallness of the fine if you are caught. That is, a high implicit price for loot relative to the price of getting away with it means that, in this choice set, choosing a crime with lots of loot will cost dearly in terms of the chances of getting caught. Incomes can be thought of as the overall extent of criminal opportunities available. A higher income means that, relative to a low income choice set, there are crimes available that involve not only a lot of loot, but also a very good chance of getting away with the crime, and small fines if caught. The incomes and implicit prices are varied in such a way that the choice sets overlap frequently, with the intersections of the constraints designed in such a way as to ensure many possibilities for intransitive choices. Choice sets include both the bundles on the frontier of the budget set, and some interior bundles to check for monotonicity. Table 1A gives summary information on the choice sets and Table 1B gives the menu of bundles for a representative choice set.

the implicit prices. The base fine is set sufficiently high to make sure that criminals who are caught are not paid for committing a crime.

17

Bronars (1987) discusses the power of revealed preference tests in a two-good setting. He finds that power is maximized when two-good budgets bisect each other at arbitrarily small angles (that is, they nearly coincide). The three-good analog is that budget planes intersect each other such that the area on either side of the intersection is equal in both budgets, and that they intersect at arbitrarily small angles. In terms of experimental design, there is a clear tradeoff between Bronars power to detect revealed preference violations and the ability to detect large violations, in the sense of the Afriat Index. If we design our choice sets to maximize Bronars power, we do not present any opportunities for costly mistakes, so we have little idea about subjects likely responses to costlier mistakes. On the other hand, if we design our choice sets to maximize the power of Afriats Index, we may not be able to detect revealed preference violations. While we take maximization of Bronars power as our starting point for parameter choice, we are forced to relax it to some degree in order to accommodate the other desirable characteristics, including a balance with the Afriat Index. The subjects are told that they must choose one bundle from each choice set. After they have made their choices, we randomly determined whether their choice was to be implemented that is, whether they were the criminal or the victim. Then, one of the criminals choice sets was randomly chosen, and whatever choice they made from that set was implemented. We used a randomized procedure to determine if they got away with any crime they might have chosen, or if they were instead caught and

18

must return any loot to the victim and also pay the fine. The criminal and the victim were then paid the resulting amounts in cash. Our protocol was designed to make sure that choices are made carefully. We gave people a chance to change their minds after thinking things over, and we included a simple test of whether or not independence holds in this context. This test is explained later. We also included duplicate budget sets just to see whether people were making internally consistent choices. In Table 1A, budgets 1 and 5, 3 and 7, and 4 and 8 are the same. For each of the choice sets, we gave the subjects 30 seconds to choose a bundle. We told them not to go on to the next choice set until the 30 seconds were up. We call these 10 choices their first choices. For the second choices, we had them go through the list again, spending a further 15 seconds on each choice set, and marking any changes they would like to make by crossing out the old and circling the new. We test for independence by giving them another chance to change their minds, after the uncertainty about their role and the actual choice set is resolved. After we told them whether they are the criminal or the victim, we had them go through the choice sheets yet again. Finally, we told them which specific choice sheet had been chosen for implementation, and they could then change their choice on that sheet, if they wish. Information about choices at each stage is retained for analysis. Choice sets were ordered starting with the low income version, and going to high. The order of choices within the choice sets was blocked, with half the participants receiving forms where the

19

amount of loot to be taken in choices goes from low to high as you go down the page, and half vice versa. Participants were recruited from high school math classes in Eugene, OR. After obtaining permission from the school district and the Principal, we contacted three teachers and performed the experiments in a total of six classes. Students were matched with other students in their class, so they know the other participants quite well, although all interactions and payoffs are anonymous and secret. One might expect that total strangers would have a greater propensity to steal compared to anonymous classmates. Because school attendance rates are high, this procedure provides a fairly representative sample of the area high school age population. However, the sample is not nationally representative Eugene is a medium sized college town with a population that is richer and whiter than the US as a whole. We also recruit subjects from an upper-division undergraduate industrial organization course at the University of Oregon. These students clearly differ from the high school students in many ways. In particular, they have higher GPAs, more money, and are older.

IV. Revealed preference results

First, we check to see whether choices from the duplicate choice sets pairs 1 and 5, 3 and 7, and 4 and 8 are the same. For the first decision, 167 of the 342 pairs (3 pairs for each subject) are the same. For the final decision, 196 of 342 are the same. Consistency is slightly higher for undergraduates. A Chi-squared test rejects independence of choices (independence predicts that 342 decision pairs are the same)

20

with p-value of 0.000 for each comparison. When we regress the number of matched choices on demographics nothing is significant, including whether the subject is a high school student and GPA measures. We note that about 75% of participants change at least one choice in round 2, about 20% change in round 3, and no one changes in round 4. We take this as evidence that choice behavior under the uncertainty, which is resolved in rounds 3 and 4, generally obeys independence. This is important, since independence is a necessary requirement for our protocol to generate data that can be used to test rationality. We check for revealed preference violations using an algorithm from Varian (1995) which we modified to handle three goods and discrete bundles. Tests of SARP and WARP yield comparative results that are very similar to those from GARP, so only results for GARP are reported here. Table 2 displays the average number of GARP violations for high school subjects, college subjects, and all subjects, and provides a comparison to random (uniform and bootstrap) choice. In the bootstrap random choice, each bundle is weighted by its frequency in the overall choice distribution. Each subject group exhibits significantly fewer violations than random choice or bootstrap choice. The average number of GARP violations across all classes is 4.2. Table 2 gives frequencies for the number of GARP violations per subject. (All of the counts in Table 2 are for the final choices.) Note that, since a minimum of two choices is required to check for a transitivity violation, it is impossible to have just one violation. Overall, about 40% of the subjects have no GARP violations. The average number of

21

violations in round 1 is about 4.8, so on average the choice modifications that people make are moving them towards greater rationality. There is no obvious standard against which to compare the number of GARP violations. The revealed preference theorems described above require that choices obey the axioms without exception. In practice, this standard is not met. Sippel (1997) used a similar protocol for eight different consumption goods, using 10 different budget sets. He found that 24 of 42 participants violated GARP at least twice. Andreoni and Miller (2002) examined 142 college students decisions about how much money to keep for themselves and how much to share with another, under eight different budget constraints. They found that nine percent of the participants committed at least some violations of the revealed preference axioms. Harbaugh et al. (2001) looked at decisions over two consumption goods and 11 choice sets. Eleven-year-olds and college students had similar patterns, with about 35% displaying GARP violations. The average number of violations was about two. The task in our experiment is more difficult, in terms of the number of goods, than that in the Andreoni or Harbaugh experiments, but simpler than that of Sippel. On this basis, our results seem consistent with those generated by other experiments. As in Harbaugh, Krause, and Berry (2001) and Andreoni and Miller (2002), our revealed preference test requires that preferences are strongly monotonic. Rather than take this on faith, our experiment is designed to test this assumption by including dominated bundles in the choice set that is, bundles with simultaneously lower loot and/or higher probabilities of detection and fine than other options in the choice set. 22

Table 3 gives the frequencies for the number of monotone choices for our sample, and compares this to what would occur with random selection using a Monte Carlo simulation with 10,000 sets of uniform draws from our choice sets. The average number of monotone choices is 7.0, and 47% of our 114 participants made eight or more monotone choices. This is compared to 17% under the Monte Carlo simulation. It is important to note that non-monotonic preferences are not irrational, and can be explained by fairness models. See for example Rabin (1993), Fehr and Schmidt (1999), Bolton and Ockenfels (2000). Sippel (1997) reported that most of his participants spent their entire budget, and those who did not, came very close to spending it all. The Harbaugh et al. (2001) paper assumed monotonicity, but did not include any procedure for verifying it. Andreoni and Miller (2002) had a significantly larger percentage of monotone choices (88%) than we find. Our experiment involved a significantly greater number of alternatives, and our monotonicity test was integrated into each budget set, while the Andreoni paper constructed separate budget sets specifically for the purpose of testing for monotonicity. It seems likely that Andreonis procedure makes the non-monotonic choices more obvious and less likely to be chosen. We attempted to correlate the rationality of participants choices with some socio-demographic characteristics collected in a post-experiment survey. Definitions and descriptive statistics of the variables are provided in Table 4. Oldest Child is a

23

dichotomous variable to indicate if the subject is the oldest child in his or her family.5 Tenure is the number of years the subject has lived in Oregon. A larger value may be considered as a proxy for enhanced ties to friends and community; therefore it may be negatively correlated with the propensity to steal. Since the data are discrete and cardinal, we estimate count data models in an attempt to explain the number of GARP violations using the socio-demographic variables. Table 5 presents the results negative binomial regressions, Poisson regressions yield similar results. The first column includes the explanatory variables age, gender, GPA, the height of the individual, whether he or she is a high school student, oldest child, and the amount of money spent per week. The second column displays a more flexible specification where the explanatory variables are interacted with the High School dummy, allowing the hypothesis of differential impact of the variables by High School vs. college status can be entertained. Being older is associated with about 0.6 fewer GARP violations. While the interaction of Age and High School is positive and significant, the net difference between the coefficients on this interaction term and Age is not. This means that, among high school students, age does not contribute to the explanation of the number of GARP violations, while among college students, being older is associated with a decrease in the number of GARP violations. Gender, height and being the oldest child have no additively separable impact on the number of GARP violations. A general conclusion seems to be that the degree of

In case of no siblings, the subject is classified as the oldest child.

24

rationality of behavior is not well explained by the available socio-demographic variables. Rationality requires that choices over crimes obey GARP. Although our data are not entirely consistent with this axiom, the number of violations is in line with what other researchers have observed for choices over general types of consumption goods. Thus, the observed behavior is at least broadly consistent with utility maximization.

V. Crime and deterrence

In this section of the paper we estimate demand functions for stolen money. Specifically, we investigate the determinants of the amount of loot stolen as a function of personal characteristics of the person who steals, the price of the stolen loot, the probability of being caught, and the amount of fine. Table 6 presents the distribution of the number of thefts. During the 10 rounds of the experiment, each individual had the opportunity to steal 10 times. Thus, in Table 6, zero thefts means that the individual never stole during the experiment, and a 10 indicates that he or she stole money in every round. There is substantial variation in the number of thefts, with 49 percent of the subjects stealing money in each round. If people are choosing rationally, then we would expect them to respond to the changes in implicit prices in ways that are consistent with the laws of demand. For example, we would expect that participants will respond to an increase in the cost of choosing a crime with high loot by tending to move toward crimes with less loot but also lower probabilities of detection and/or lower fines. An increase in available

25

criminal opportunities should also be expected to increase loot, assuming it is a normal good. In these regressions, we normalize the implicit prices by dividing through by income. We expect, therefore, to find negative own price effects and positive cross price effects. Table 7 displays our analysis of the decision to steal. Columns I and II present the marginal probabilities of the explanatory variables pertaining to the decision at the extensive margin. In this analysis each individual contributes 10 observations because they make 10 decisions in 10 rounds regarding whether or not to steal. In Column II, where the potentially endogenous variables of Money, GPA and Tenure are omitted, the age and height of the individual has a positive impact on the propensity to steal. This may indicate that having stronger physical attributes may provide motivation to steal. Being the oldest child, on the other hand, lowers the propensity to steal. Columns III and IV of Table 7 present the estimated negative binomial regression results. In these specifications each individual contributes one observation, thus the impact of prices cannot be analyzed. The results are consistent with those reported in columns I and II, but only the coefficient of Age is estimated with precision in both columns, and Money in Column III (though the magnitude is small). The Poisson regression results are similar. Table 8 presents the results of the estimated demand functions. The dependent variable is the amount of loot taken in each round. Models are estimated with OLS, and standard errors are adjusted for clustering at the individual level. Columns I through 26

III contain coefficient estimates for the commodity prices, which always have the expected sign, though the fine price is not significant. Columns II and III include coefficients for additional demographic variables. These specifications both suggest that older students tend to steal more. Column II suggests that taller people steal more. Money is included in Column III, and is significant. Columns IV and V include price interactions with the High School dummy variable. As before, the price coefficients all have the expected sign, though only the price of loot is significant. Columns IV and V also indicate that the coefficients of detection price and fine price are the same for college and high school students, but suggests that these effect of the price of loot may be smaller for high school students.

VI. Discussion and conclusion

The extent to which criminals and potential criminals respond to variations in deterrence is an important issue, both theoretically and from a public policy perspective. Despite significant progress in recent empirical analyses in identifying the causal effect of deterrence on crime, objections are still raised with regard to the validity of methods proposed to eliminate the simultaneity between crime and deterrence. In this paper we designed an experiment where subjects are exposed to exogenous variations in the relative tradeoffs between three important aspects of criminal opportunities: loot, the probability of detection, and the fine. We conduct the experiment with juveniles and young adults age groups that are frequently labeled as irrational or unresponsive to deterrence.

27

We find that behavior among these groups with respect to petty criminal decisions is not entirely rational. However, behavior is approximately as consistent with the theoretical requirements of rational choice behavior as is choice behavior over consumption goods. Furthermore, we find that, in aggregate, responses to changes in criminal opportunities are consistent with the laws of demand. This serves to underscore the deterrent properties of law enforcement effort and the penalties applied to criminal acts. Experimental research in the tax compliance literature has alluded to as much by describing the effects of audit rates and marginal tax rates on tax compliance. However, this experiment generalizes the rational criminal model to a broad range of criminal behaviors, such as auto theft, embezzlement, and petty larceny. Caveats are that the participants in these experiments are not necessarily criminals outside the laboratory, and that the crimes in our experiments involve small financial gains and losses. Given these qualifications, we believe these results strengthen the argument that criminal behavior and the response of criminals to changes in enforcement and penalties can be accounted for by economic models.

28

References
Afriat, Sidney. (1967): The construction of utility functions from expenditure data, International Economic Review, 8, 67-77. _____. (1972): Efficiency estimation of production functions, International Economic Review, 13, 568-598. Andreoni, James, and John Miller. (2002): Giving according to GARP: an experimental test of the consistency of preferences for altruism, Econometrica, 70, 737-753. Becker, Gary S. (1962): Irrational behavior and economic theory, Journal of Political Economy, 70, 1-13. _____. (1968): Crime and punishment: an economic approach, Journal of Political Economy, 76, 169-217. Block, M. and M. Heineke. (1975): A labor theoretic analysis of the criminal choice, American Economic Review, 65, 314-325. Bolton, Gary E. and Axel Ockenfels. (2000): ERC a theory of equity, reciprocity, and competition, American Economic Review, 90, 166-196. Bronars, Stephen G. (1987): The power of nonparametric tests of preference maximization, Econometrica, 55, 693-698. Corman, H. and Naci Mocan. (2000): A time-series analysis of crime, deterrence and drug abuse in New York City, American Economic Review, 90, 584-604. Corman, H. and Naci Mocan. (forthcoming) Carrots, sticks, and broken windows, Journal of Law and Economics. Cornwell, C. and W. N. Trumbull. (1994): Estimating the economic model of crime with panel data, Review of Economics and Statistics, 76, 360-366. Cox, James. (1997): On testing the utility hypothesis, The Economic Journal, 107, 1054-1078. Di Tella, R. and E Schargrodsky. (2004): Do police reduce crime? Estimates using the allocation of police forces after a terrorist attack, American Economic Review, 94, 115-134.

29

Ehrlich, I. (1973): Participation in illegitimate activities: a theoretical and empirical investigation, Journal of Political Economy, 81, 521-565. _____. (1975): The deterrent effect of capital punishment: a question of life and death, American Economic Review, 65, 397-417. Fehr, Ernst and Klaus Schmidt. (1999): A theory of fairness, competition and cooperation, Quarterly Journal of Economics, 114, 817-868 Flinn, C. (1986): Dynamic models of criminal careers. In Blumstein, A. et al. (eds.), Criminal Careers and Career Criminals. Washington DC: National Academy Press. Harbaugh, William T., Kate Krause, and Timothy Berry. (2001): GARP for kids: on the development of rational choice behavior, American Economic Review, 91, 1539-1545. Harbaugh, William T., Kate Krause, and Steve Liday. (2003): Bargaining by children, working paper, University of Oregon. Houthakker, H. S. (1950): Revealed preference and the utility function, Economica, 17, 159-174. Koo, Anthony Y. C. (1963): An empirical test of revealed preference theory, Econometrica, 31, 646-664. Layson, S. K. (1985): Homicide and deterrence: a reexamination of the United States time-series evidence, Southern Economic Journal, 52, 68-89. Levitt, S. D. (1997): Using electoral cycles in police hiring to estimate the effect of police on crime, American Economic Review, 87, 270-290. _____. (2002): Using electoral cycles in police hiring to estimate the effect of police on crime: reply, American Economic Review, 92, 1244-1250. Mattei, Aurelio. (1994): La coherence des choix des consommateurs suisses, (with English summary), Schweizerische Zeitschrift fur Volkswirtschaft und Statistik/Swiss Journal of Economics and Statistics, 130, 3-20. Mossin, Axel. (1972): A mean demand function and individual demand functions confronted with the weak and strong axioms of revealed preference: an empirical test, Econometrica, 40, 177-192.

30

Myers, S. L., Jr. (1983): Estimating the economic model of crime: employment versus punishment effects, Quarterly Journal of Economics, 98, 157-166. Rabin, Matthew. (1993): Incorporating fairness into game theory and economics, American Economic Review, 83, 1281-1302. Samuelson, Paul A. (1938a): A note on the pure theory of consumers behaviour, Economica, 5, 61-71. _____. (1938b): A note on the pure theory of consumers behaviour: an addendum, Economica, 5, 353-354. _____. (1948): Consumption theory in terms of revealed preference, Economica, 15, 243-253. Schmidt, P. and A. D. Witte. (1984): The nonparametric approach to demand analysis, Econometrica, 50, 945-973. Sippel, Reinhard. (1997): An experiment on the pure theory of consumers behaviour, The Economic Journal, 107, 1431-1444. Varian, Hal R. (1982): The nonparametric approach to demand analysis, Econometrica, 50, 945-973. _____. (1991): Goodness of fit for revealed preference tests, University of Michigan CREST Working Paper Number 13. _____. Microeconomic Analysis, 3rd edition (1992). W. W. Norton and Company Inc., New York, NY. _____. (1995): Efficiency in production and consumption, working paper and Mathematica notebook. [Online]. Available: http://emlab.berkeley.edu/eml/nsf97/varian.pdf [2000, January 24] Witte, A. D. (1980): Estimating economic models for crime with individual data, Quarterly Journal of Economics, 94, 57-84.

31

Table 1A Choice Set Characteristics Budget parameters loot_p prob_p nfine_p .25 1 .5 .5 .25 1 .5 .5 1 .25 1 2 4 2 1 2 4 2 4 2 1 1 1 1 1 1 1 1 1 1

Budget 1 2 3 4 5 6 7 8 9 10

income 1.2 2.9 3.25 2.25 1.2 2.75 3.25 2.25 3.75 1.75

Table 1B Sample Bundles, from Choice Set 5

You each start with $5

Mark one choice below Dollars to take from Person B Your payment including your starting $5 if you are not discovered $5 Chance that you are discovered Dollars paid to experimenter if discovered Your payment including your starting $5 if you are discovered ---

$0 $1.00 $1.00 $1.00 $1.00 $2.00 $2.00 $3.00

--25% 50% 75% 75% 50% 75% 75%

--$1.55 $1.30 $1.05 $1.25 $1.55 $1.30 $1.55

$6.00 $6.00 $6.00 $6.00 $7.00 $7.00 $8.00

$3.45 $3.70 $3.95 $3.75 $3.45 $3.70 $3.45

32

Table 2 Frequency of GARP Violations Number of GARP violations 0 1* 2 3 4 5 6 7 8 9 10 N HS 37% 0% 6% 7% 6% 8% 5% 4% 6% 6% 14% 83 UO 48% 0% 0% 0% 0% 10% 6% 6% 6% 10% 13% 31 All 40% 0% 4% 5% 4% 9% 5% 4% 6% 7% 14% 114 Bootstrap 3% 0% 3% 5% 6% 10% 16% 20% 21% 14% 4% 10,000

*Note that it is impossible to have exactly one violation.

33

Monotone Choices 0 1 2 3 4 5 6 7 8 9 10 N

Table 3 Monotonic bundle choices Sample Bootstrap Running Running Frequency Frequency total Total 0% 0% 0% 0% 7% 7% 0% 0% 4% 11% 1% 1% 5% 16% 3% 4% 10% 26% 10% 14% 5% 32% 20% 34% 12% 44% 27% 61% 10% 54% 22% 83% 6% 60% 12% 95% 16% 76% 4% 99% 25% 100% 1% 100% 114 10,000

34

Table 4 Descriptive Statistics Variable Loot* GARP Age Height Definition The money stolen in each round Number of GARP violations Age of the individual Height of the individual in feet High school GPA if the individual is in high school; the average of high school and college GPAs if the in college How much money the individual spends on his/her own per week Dichotomous variable (=1) if the person is male Dichotomous variable (=1) if the person is oldest child The number of years the person lived in Eugene, Oregon High School 1.23 (1.14) 4.01 (3.84) 15.98 (15.98) 5.59 (0.32) 3.12 (0.54) 18.34 (17.93) 0.51 0.34 8.43 (3.59) 83 College 1.77 (1.21) 4.00 (4.16) 22.01 (0.96) 5.88 (0.35) 3.20 (0.33) 72.71 (171.93) 0.71 0.52 5.39 (4.46) 31 All 1.38 (1.19) 4.01 (3.91) 17.62 (2.86) 5.67 (0.35) 3.14 (0.49) 33.13 (94.02) 0.56 0.39 7.61 (4.08) 114

GPA

Money Male Oldest Child Tenure

N * Loot is the average of all 10 rounds

35

Table 5 Negative Binomial Estimates of the Number of GARP Violations Variable Age Age*High School Male Male*High School Height Height*High School GPA GPA*High School Money Money*High School Oldest Child Oldest Child*High School High School n Log Pseudo-Likelihood I 0.050 (0.104) 0.115 (0.284) -0.072 (0.378) -0.001 (0.235) -0.005* (0.003) -0.566*** (0.244) 0.075 (0.561) 114 -277.35 II -0.605** (0.298) 0.764** (0.318) 0.220 (0.487) -0.039 (0.588) 0.993 (0.792) -1.571* (0.896) -0.929 (0.782) 1.083 (0.822) -0.010 (0.010) 0.009 (0.010) -0.303 (0.465) -0.406 (0.550) -10.539 (6.739) 114 -273.74

Robust standard errors in parentheses *, ** and *** indicate 10, 5, and 1% significance

36

Table 6 Number of Thefts* 0 1 2 3 4 5 6 7 8 9 10 Number of Individuals 5 2 2 4 6 2 6 6 13 12 56 Percentage of Total 4% 2% 2% 4% 5% 2% 5% 5% 11% 11% 49%

*The number of thefts is the number of rounds where the individual stole money. Thus, 0 indicates that the individual did not steal money during the entire experiment, and 10 indicates that he/she stole in every round.

37

Table 7 Participation in Crime Probit Estimates of Decision to Steal Variable Loot Price Detection Price Fine Price Age Male Height Money High School GPA Oldest Child Tenure Ia -0.004 (0.416) 0.240 (0.184) -0.141* (0.082) 0.042* (0.025) -0.066 (0.073) 0.141 (0.098) 0.001 (0.001) 0.138 (0.185) -0.023 (0.057) -0.064 (0.055) -0.006 (0.007) IIa 0.003 (0.430) 0.251 (0.191) -0.145* (0.085) 0.048** (0.023) -0.048 (0.069) 0.134 (0.104) 0.130 (0.166) -0.072 (0.057) Negative Binomial Estimates of the Number of Thefts III 0.064** (0.027) -0.069 (0.101) 0.152 (0.126) 0.0002* (0.0001) 0.232 (0.172) -0.020 (0.069) -0.094 (0.072) -0.007 (0.010) IV 0.060** (0.028) -0.055 (0.099) 0.152 (0.130) 0.187 (0.165) -0.091 (0.072) -

n 1140 1140 114 114 Log-Likelihood -528.29 -533.81 -1105.91 -1106.51 Robust standard errors in parentheses *, ** and *** indicate 10, 5 and 1% significance levels respectively a) The reported coefficients are marginal probabilities

38

Table 8 Demand for Loot Variable

Loot Price Loot Price*High School Detection Price Detection Price*High School Fine Price Fine Price*High School Age Male Height Money High School GPA Ascending Oldest Child Tenure Constant n Adjusted-R2

II

III

IV

V
-6.524*** (1.397) 3.632** (1.751) 0.749 (0.560) 0.326 (0.708) 0.044 (0.265) 0.015 (0.357) 0.137** (0.066) -0.090 (0.191) 0.404 (0.247) 0.002*** (0.0003) -0.672 (1.302) -0.029 (0.138) 0.071 (0.124) -0.185 (0.131) -0.016 (0.016) -2.522 (1.970) 1140 0.24

-3.880*** -3.880*** -3.880*** -6.524*** (0.866) (0.868) (0.869) (1.394) 0.987*** (0.348) 0.055 (0.605) 1.302** (0.605) 1140 0.14 0.987*** (0.349) 0.055 (0.188) 0.144** (0.062) -0.066 (0.168) 0.426* (0.240) 0.458 (0.422) -0.154 (0.132) -0.017 (0.017) -3.758*** (1.712) 1140 0.21 0.987*** (0.349) 0.055 (0.188) 0.119* (0.066) -0.100 (0.193) 0.409 (0.247) 0.002*** (0.0003) 0.315 (0.433) -0.060 (0.142) -0.066 (0.123) -0.170 (0.135) -2.981 (1.843) 1140 0.23 3.632** (1.748) 0.749 (0.559) 0.326 (0.707) 0.044 (0.265) 0.015 (0.356) 0.129** (0.063) -0.058 (0.167) 0.423* (0.240) -0.825 (1.290) -0.147 (0.135) -2.690 (1.853) 1140 0.22

Robust standard errors, which are adjusted for clustering at the individual level, are in parentheses. *, ** and *** indicate 10, 5 and 1% significance levels respectively.

39

Appendix A

Welcome: Today we are conducting an experiment about decision-making. Your decisions are for real money, so pay careful attention to these instructions. This money comes from a research foundation. How much you earn will depend on the decisions that you make, and on chance. Secrecy: All your decisions will be secret and we will never reveal them to anyone. We will ask you to mark your decisions on paper forms using a black pen or pencil. If you are discovered looking at another persons forms, or showing your form to another person, we cannot use your decisions in our study and so you will not get paid. Please do not talk during the experiment. Payment: Stapled to this page is a card with a number on it. This is your claim check number. Each participant has a different number. Please tear off your card now and write your claim check number on the line on the first page. You are also given a packet. Write your claim check number on top of the first page of that packet, but do not turn the page until instructed to do so. Be sure to keep your claim check number. You will present this number to an assistant at the end of the experiment and you will be given your payment envelope. The Experiment: You are going to play a game today. In this game there will be two roles A and B. Everyone will be randomly assigned one role, and will be matched with another person with the other role. You will not be told who you are paired with, and they will not be told who they are paired with, even after the experiment is over. Person A will start with $5, and person B will also start with $5. Person A will have a chance to take some of person Bs money. Taking is not without a risk. After person A makes the decision to take or not to take, there will be a discovery phase. During this phase there is a chance that person A will have to return the money taken from person B, and also pay some money to the experimenter. The chance that this happens, and the amount paid to the experimenter if it does, depend on the choice made by person A. 40

Everyone received one packet, and each packet contains 10 different sheets stapled together. We will show you an example. We call these Choice Sheets. On each Choice Sheet you will make a choice as if you are person A. You will declare your choice of how much money to take from person B by putting a check mark next to one of the choices with a black pen or pencil. When we play the game the amount of money you will end up with will really be determined by the choices you make, so you want to consider your choice very carefully. We will give you 60 seconds on the first page and 30 seconds on each subsequent page. Please leave your pencil or black pen on the desk and do not mark your choice until I ask you to do so. When the time is up I will ask you to place a check mark using a pencil next to the choice you want. It is important that you wait until the time is up to mark your choice. After we go over all 10 Choice Sheets and everyone has made all 10 choices, I will give you a chance to change your mind. This time you will have 15 seconds on each page. Please leave your pencil or black pen on the desk and do not mark your choice until I ask you to do so. To change your choice, clearly cross out (do not erase) your previous choice, and place a check mark next to the choice you want. Next, we randomly assign roles of A or B by flipping a coin. If it comes up heads, then those whose claim check number is even will be assigned the role of person A, and those whose claim check number is odd will be assigned the role of person B. Should the coin come up tails, then those whose claim check number is odd will be assigned the role of person A, and those whose claim check number is even will be assigned the role of person B. Now we have to pick which one of the 10 Choice Sheets will count. We will pick a random number from 1 to 10, by having your teacher draw a card from a deck of 10 cards. The Ace will stand for 1. The number of the card will determine which Choice Sheet counts. We will have you turn your packet to that choice sheet. Then we must complete the discovery phase to see if person A will have to return the money to person B and pay some money to the experimenter. Here is how this will work: We have 5 index cards. On each index card there is a percentage written. They are: 0%, 25%, 50%, 75% and 100%. We will randomly choose one of these index cards. Everybody looks at their choice on the Choice Sheet that has just been selected. If you are person A, and if the percentage written on the selected index card is less than the chance of being discovered for the choice you made on the Choice Sheet, then you are discovered. You will have to return the money you took from person B and pay to the

41

experimenter the number of money indicated in the choice. Otherwise you keep the money. We will collect odd numbered and even numbered packets in separate stacks. Then we will mix up each stack , and take one from each stack to match people together. The choice made on person As Choice Sheet will be used to determine their payments. We will proceed through the stacks until we are done. If there is an odd number of people in the room, then at the end there will be one packet left over. This packet will be assigned the role of person A, and will be paid according to the decision on his or her choice sheet. Note that you dont know which of your 10 decisions will count, if any. This will be determined purely by chance. So the best thing for you do to is to treat every choice sheet as if it will count, and make the choice on that sheet that you most prefer.

42