Cultural diversity, discrimination, and economic outcomes: an experimental analysis.

Ferraro, Paul J. ; Cummings, Ronald G.

I. INTRODUCTION

Economic disparities have long existed across nations and between racial and ethnic groups within nations. Recently, economists have taken a closer look at the role of "culture" in explaining global variability in economic behavior and outcomes. One path of inquiry focuses on cross-cultural differences in behavior, as in Brandts, Saijo, and Schram (2004), Henrich (2000), Croson and Buchan (1999), Ockenfels and Weimann (1999), Burlando and Hey (1997), and Roth et al. (1991). In particular, an initiative to explore the effect of culture in 15 small-scale societies across the globe found striking variability in the outcomes of economic experiments as in Henrich et al. (2001, 2004).

Others have taken the "cultural effects" inquiry in a different direction: If cultural differences affect economic behavior and outcomes (or indeed even if they do not), do intercultural relationships affect behavior and outcomes? A controversial empirical literature has developed over the role that cultural diversity may play in explaining cross-national or cross-regional differences in economic outcomes. Some authors, for example, Easterly and Levine (1997) and Alesina et al. (2003) find that there is an inverse relationship between economic growth and cultural diversity, while others like Collier (2001) and Fearon (2003) contest this conclusion. In the United States, Alesina, Baqir, and Easterly (1997) find that cultural diversity is an important determinant of local public finances. In particular, they find an inverse relationship between diversity and spending on education, roads, and sewers, which they attribute to majority of white citizens reacting to the size of minority groups. Miguel (1999) finds similar results in Kenyan primary schools: high levels of ethnic diversity are linked to lower school funding, lower student to teacher ratio, and lower parental involvement in school functions.

One possible mechanism through which cultural diversity can affect economic behavior and outcomes is discrimination. Economists have performed many empirical analyses to identify discrimination in the marketplace and to determine the nature of the discrimination, as noted by Yinger (1998), Altonji and Blank (1999), and Riach and Rich (2002). To test for discrimination, economists depend on regression-based methods and field experiments. The former technique tests for a statistical relationship between an outcome measure, such as wage or price, and a group membership indicator, as noted by Goldberg (1996). The latter includes audit studies, for example, Neumark, Bank, and van Nort (1996) and correspondence tests, for example, Bertrand and Mullainathan (2004).

Our paper provides an experimental framework that can tie together these disparate literatures and help economists move toward a synthesis of the effects that "culture" has on economic behavior and outcomes. In particular, our analysis complements existing research in three important ways.

First, empirical analyses of the economic effects of cultural diversity at the level of communities and nations suffer from the inability to control many of the factors that affect the observed outcomes and the classic problem of having only one observation of the world at time t. By virtue of the experimenter's ability to control and manipulate the cultural diversity within laboratory sessions, our experimental framework provides a path of inquiry that can yield insight into the recent debate about the role of cultural diversity and economic outcomes in societies. In the laboratory, one can reproduce existing cultural diversity patterns or create counterfactual societies. Observations of these laboratory societies offer insight about behavior and outcomes in the naturally occurring societies outside the laboratory. We know of no other experiment that is designed to determine whether the cultural diversity of the experimental session affects the manner in which subjects make decisions.

Second, economists have been successful in developing techniques to detect discrimination, but less successful at explaining observed discrimination, for example, List (2004). If rational agents have no information about the behavior of the person with whom they are interacting but have information about the average behavior of the group to which the person belongs (e.g., an ethnic group), they may condition their decision on this average behavior. Such discrimination is called "statistical discrimination" or "rational stereotyping," given in Arrow (1973) and Phelps (1972). If, in contrast, rational agents simply prefer to behave differently when interacting with an individual from a given group, such behavior is called "preference-based discrimination" or "a taste for discrimination" as given in Becker (1971). From a theoretical and policy perspective, the difference between these two types of discrimination is important, but their relative empirical importance is controversial as in Ladd (1998).

Like the recent paper by List, we demonstrate how our experimental framework can provide insight into the economics of discrimination. The paper complements List by demonstrating an alternative method for distinguishing between statistical discrimination and preference-based discrimination in situ, without requiring inferences to be drawn from behavior in other experiments.

Third, cross-cultural experiments often emphasize cross-national differences in behavior. We wish to determine if cross-cultural behavioral differences can be detected within two cultures that coexist in the same industrialized society (we are not the first to do so). Furthermore, through the control afforded by the laboratory, we ensure that we do not attribute to "culture" any differences in behavior that stem from variability in the socio-demographic attributes of our subjects.

To address these issues, we organized experimental sessions of a simple bargaining game with members of two cultural groups from New Mexico: Navajo Indians and Hispanic Americans. We varied the cultural mix of our experimental sessions in order to infer the effect of intercultural interactions on economic behavior. In the next section, We define what we mean by "culture" and describe how our study builds on previous research. In Section III, we describe the design of our experiments. Results are reported in Sections IV and V. Conclusions are offered in Section VI.

II. CULTURE, ETHNICITY, AND RACE

Our experiments were conducted in Albuquerque, New Mexico. New Mexico is arguably the most unique state in the United States, in terms of ethnic diversity, with three major ethnic groups, each accounting for a sizable proportion of the population. In 2001, New Mexico's population was 42.1% Hispanic, 45% Anglo, and 10% Native American, with blacks and Asians accounting for the remaining 2.9% (1) New Mexico has a higher Hispanic population, in terms of percentage of total population, than any other state in the United States. Other states have a higher proportion of Native Americans, but no other state has a mix of cultures comparable to New Mexico. Native American and Hispanic cultures are distinct and dominant in the state and in Albuquerque.

Economists who work with concepts like culture, ethnicity, and race often avoid defining such words. Their definitions, however, are subject to much debate in other disciplines as given in McElreath, Boyd, and Richerson (2003). (2) We use the word "culture" to refer to the statistical distribution of beliefs, values, and modes of thinking that shape behavior among a group of people (e.g., notions of fairness). "Ethnicity" is related to symbolically marked groups (e.g., marked by language, dialect, or clothing). Cultural differences may be present in a population when ethnicity is not marked (e.g., southern-born and northern-born whites in the United States, as in Nisbett and Cohen (1996). Similarly, ethnic differences may exist when no cultural differences exist (except for the ethnic marking). "Race" is like ethnicity, except the "markers" are genetically transmitted, for example, physical characteristics, as in Gil-White (2001).

Navajo and Hispanic individuals in our experiments are distinct culturally, ethnically, and racially. We test whether such distinctions make any difference in the bargaining behavior of our subjects. In our experiment, we cannot empirically differentiate the separate effects of culture, ethnicity, and race. Thus, we use the term "cultural differences" to describe any differences that result from differences in culture, ethnicity, or race. As in previous papers that find relationships between an individual's culture, ethnicity, or race and his or her behavior or economic status, we cannot be certain that what we describe as cultural determinants are not actually noncultural determinants (for which we have no data) that are correlated with culture. In this sense, what economists call "culture" in analyses of economic behavior is best viewed as a residual category. By controlling for differences in behavior that stem from variability in the socioeconomic attributes of our subjects, we attribute to "cultural differences" any remaining variability in behavior across cultural groups.

Although experimental economists have explored cross-cultural differences in behavior, they have generally ignored the question, "Do individuals interacting with others sharing the same culture behave differently than when interacting with others from a different culture?" We find only two published studies that address this question: Fershtman and Gneezy (hereafter FG, 2001) and List (2004). (3)

In experiments with two major Israeli ethnic groups, the Ashkenazic and Eastern Jews, FG address the effects of ethnic stereotyping on trust and bargaining. In their Ultimatum Game experiment, larger offers are proposed to Eastern players. (4) However, there is no significant difference between the percentage of Eastern and Ashkenazic players that reject a (low) split of 25% of the pie. FG writes, (2001, 370) that the observed discrimination "is probably an outcome of a common ethnic stereotype in Israeli society, according to which men of Eastern origin are believed to react more harshly if treated unfairly." Although FG do not explicitly refer to their Dictator Game experimental results to interpret their Ultimatum Game results, one could interpret the absence of any discrimination in their Dictator Game as indirect evidence that behavior in their Ultimatum Game stems from erroneous statistical discrimination. However, in the absence of information about players' expectations of partner responses, one has only indirect evidence for this conclusion. (5)

List studies discrimination among participants in a sportscard field experiment. He observes starting and final offers for a card and collects information on subject attributes (age, experience, gender, education, income, height, and weight) and the length of the bargaining session. Subjects are put in four categories: white males aged 20-30, white females aged 20-30, white males aged 60+, and "nonwhite" males aged 20-30. Given that race is not asked on the questionnaire, it is unclear as to how the author determines race and what race, or races, the term "nonwhite" includes.

List finds that average initial and final offers from dealers to "minority" buyers, (females, older males, and nonwhite males) are higher than those received by young white males. After controlling for experience, the differences among final offers are small for experienced buyers (but minority buyers do spend more time to obtain their final offers), and are only significantly different among inexperienced older male and young female buyers. (6) Like FG, List uses complementary experiments (Dictator Game, Decentralized Chamberlain Market, and a Vickery second-price auction) to elucidate the underlying reasons for the observed discrimination. Behavior in the complementary experiments suggests that discrimination by dealers in the field experiment stems from statistical discrimination rather than preference-based discrimination.

We develop an alternative method to distinguish between statistical discrimination and preference-based discrimination that does not require inferences to be drawn from other experiments: we elicit beliefs in the experiment itself. Note also that the ways in which intercultural effects are induced in FG and List are different from our experimental framework. FG's inquiry is based on a design, wherein players attempt to infer the ethnicity of their partners, who are in a different location, from the partners' surnames. In List, subjects can observe the race, gender, or approximate age of their partner or are told these attributes by the experimenter. While these are important contexts, we wish to explore behavioral variability in response to changes in the proportional representation of two cultural groups in an experimental session. In other words, we wish to determine if subjects behave differently in the following three contexts: (1) All players share the subject's culture, (2) the player's culture makes up a large majority of the players, and (3) the player's culture is a small minority of the players.

Finally as mentioned in the Introduction, we wish to reduce the chance that we mistakenly attribute to "culture" differences in behavior that stem from variability in subjects' socio-demographic attributes. Thus, we control for socio-demographic attributes that may affect bargaining behavior (which is also done in List, but not in FG).

III. EXPERIMENTAL DESIGN

Throughout the world, ethnic, racial, and religious conflicts persist in the face of potential settlements that plainly serve the interests of all sides. We thus conduct our analysis in the simplest of bargaining environments: the Ultimatum Bargaining Game. Two players, a Proposer and a Responder, bargain over $10. The Proposer offers $x to the Responder, leaving himself $10-x. The Responder can either take the offer, in which case each obtains the proposed split of the $10 pie or reject it and both get nothing. The Ultimatum . Game is too simple to be a good model of the complicated processes of most real-world bargaining. Yet, as noted by Camerer (2003, 8), its simplicity offers a useful environment for testing hypotheses about the factors that influence how people feel about the allocations of money between themselves and others.

Sixty Hispanic subjects were recruited by distributing flyers in Hispanic neighborhoods. All Hispanic subjects were raised in the United States. Sixty Navajo subjects were recruited primarily by distributing flyers at three Navajo organizations: the Southwest Indian Polytechnic Institute, the Public Health Service Indian Hospital, and the Albuquerque Indian Center. "Navajo neighborhoods" do not exist and these organizations serve as the closest equivalent. Overall, 45% of the subject pool is male, 59% report an annual income of less than $15,000, 47% of the sample consists of full- or part-time students, 15% is married, and the mean age is 29 years.

Session 1, followed immediately by Session 2, took place on one night. Session 3, followed immediately by Session 4, took place the next night. The experimental sessions were held in a large room rented at the Menaul School, centrally located in Albuquerque. Prior to each session, subjects were placed in a room in which food and refreshments were offered. We grouped subjects prior to entering the experimental room for two reasons: (1) to allow subjects to observe the ethnic makeup of their session (Navajo and Hispanic subjects are visually very different) and (2) to allow us to conduct back-to-back sessions without risking cross-session observation or communication. This simple and efficient approach permits one to highlight the cultural composition of the session without emphasizing it in a way that would allow subjects to infer the purpose of the experiment.

A portable experimental laboratory was used that consisted of 32 networked notebook computers with wireless connection to a laptop computer that acted as a server. The subjects' computers were situated in folding partitions to ensure private decisions. Standard rules of the Ultimatum Game were explained to subjects, and subjects were required to complete practice questions to ensure that they understood as to how their earnings would be calculated. The instructions for the experiments were conveyed orally and in writing (available upon request). A portable projector demonstrated the subject interface.

Subjects played the role of both Responder and Proposer, as was done in the original application of the Ultimatum Game by Guth, Schmittberger, and Schwarz (1982) and in later studies like Andreoni, Castillo, and Petrie (2003), Carter and Irons (1991), and Kahneman, Knetsch, and Thaler (1986). Subjects were told that they would make decisions as Responders and as Proposers. At the end of the experiment, the computer randomly assigned each subject to the role of Responder or Proposer and randomly paired the subject with another subject in the room (not known to him or her) who played the opposite role. Subjects were cautioned to take each role seriously, given the equal chance of being assigned the roles of Responder or Proposer. With the exception of the All-Navajo and All-Hispanic sessions the ethnicity of a subject's partner was uncertain, but the ethnic composition of the session was obvious: The subject's ethnic group constituted either a large majority or a small minority of the subjects. (9)

Note that our design differs from FG's in that subjects from one culture interact directly with subjects from the other culture. The only contact that an Ashkenazic subject in FG's experiment had with an Eastern subject was a visual inspection of the Eastern subject's name on a form (from which the subject had to infer the ethnicity of his or her partner).

[FIGURE 1 OMITTED]

The amount of money given to the Proposer, known by all subjects, was $10.00. Subjects first saw a screen that asked them to make the decisions of a Responder. They were asked to indicate, for each dollar amount between $0 and $10, if they were assigned the role of Responder and if that dollar amount were sent to them, whether they would accept it or reject it. Eliciting the behavior of Responders through the strategy method allowed us to collect data on all information sets of the game, not just those that were actually reached in the course of the game. Subjects were cautioned that, if assigned the role of Responder, they would be bound by the decisions that they recorded on the screen.

Subjects were then asked to play the role of a Proposer. To allow us to make inferences about discriminatory behavior that may be observed in the laboratory, subjects were first asked to predict how they believed Responders would respond to each possible amount that they might send to a Responder, from $0 to $10. Subjects predicted the percentage of Responders in the session that would accept each amount. To create incentives for subjects to think about their estimates, subjects were informed that the individual whose estimates were the closest to the actual percentage of Responders accepting each amount would win $10.00. (10)

Subjects were then asked to decide how much they would send to a Responder if they were assigned the role of a Proposer. They were notified that if assigned the role of Proposer, the amount they chose on this screen would be sent to the Responder.

Finally, subjects responded to a questionnaire about the motivations for their decisions as Responders and as Proposers (Figure 1). At the end of the session, the computer randomly assigned each subject to the role of Responder or Proposer and randomly paired the subject with another subject in the room. Demographic information was then obtained from each subject. The same person conducted all the sessions.

IV. RESULTS--SUMMARY STATISTICS

Table 1 summarizes the results from the four experimental sessions. This summary shows rough trends in the data. In the next section, we analyze the data controlling for subject characteristics.

A. Responders

We first examine the behavior of Responders (i.e., compare Hispanic Responders in All-Hispanic session to Navajo Responders in the All-Navajo session, etc.). Hispanic Responders have higher minimum acceptable offers, on average, than Navajo Responders in all sessions (significant at 2-11% level, depending on the comparison and whether one uses a Mann-Whitney or t-test). In the All-Navajo and All-Hispanic sessions, 60% were willing to accept an offer of 10% of the pie ($1). These acceptance rates are substantially higher than those observed in previous Ultimatum Game experiments in industrialized, nations. Guth, Schmidt, and Sutter (2003) report that ping over 33% is much higher than the rates typically observed in Ultimatum Game experiments that use the strategy method (including experiments in which subjects played both roles). (11)

Furthermore, both Hispanics and Navajos appear to discriminate against the other group--the minimum offer that they would accept increases as the relative proportion of their ethnic group in the session decreases.

This increase is particularly notable for the Hispanics. (12) The same pattern appears in the percentage of subjects willing to accept an offer of $1. Both Hispanics and Navajo are more willing to accept $1 as the proportion of their ethnic group in the session increases. Again, the behavior on the part of Hispanics is more striking. Navajo are willing to accept low offers at much higher rates than most other subjects in previous Ultimatum Game experiments, whereas Hispanic acceptance rates are only unusually high when playing in an All-Hispanic group.

B. Proposers

Offers by both Hispanic and Navajo Proposers are in the range observed in earlier studies regardless of their proportion of the session: between 38% and about 50% of the $10.00 to be divided. When playing with members of one's own ethnic group, however, Navajos make significantly lower offers than Hispanics (significant at 1% level under both a Mann--Whitney and t-test). In addition, Hispanics appear to persistently discriminate against the Navajo-Hispanic offers appear to decline as their majority status diminishes--while Navajos appear to make higher offers when Hispanics are in the session. (13)

C. Statistical versus Preference-Based Discrimination

Any observed changes in Responder behavior as a result of changes in the ethnic mix of the session must necessarily reflect preference-based discrimination. There is no role for statistical discrimination on the part of Responders--they simply must accept or reject a given offer.

As Proposers, however, subjects may make offers that are rational responses to the average behavior of subjects from the two ethnic groups (i.e., statistical discrimination). Navajo Responders are, on average, more likely to accept low offers and thus a rational agent without complete information may choose an offer based on the likely ethnicity of the Responder. To examine this conjecture, data on subject beliefs are presented in Table 2. The table, broken down by the ethnic mix, presents subjects' mean predictions of the percentage of individuals in the session that would accept low offers ($0-3). For example, Hispanic subjects in the All-Hispanic session believed, on average, that 35% of the subjects in the session would accept $1; when Hispanic subjects were a minority, however, they believed that only 5% of the subjects in the session would accept $1.

An examination of Hispanic beliefs does not support the statistical discrimination conjecture: As the proportion of Navajos in the session increases, Hispanic subjects believe the likelihood of a low offer being accepted decreases, yet they send lower offers. (14) Navajo beliefs are roughly consistent with the data in Table 1: There appears to be a decrease in expected acceptance rates when Hispanics are present, which would lead to higher offers, but this decrease in average expectations is only weakly statistically significant (p < 0.10). (15) In summary, there is no evidence that statistical discrimination plays a role in the behavior of Hispanic Proposers, and there is weak evidence that such discrimination plays a role in the behavior of Navajo Prosposers.

V. RESULTS--REGRESSION ANALYSES

The summary statistics in the previous section do not control for demographic variability among subjects or the differences in ethnic proportions across sessions. There is a high degree of variability in our subject pool with, for example, ages ranging from 16 to 50 years old and annual incomes ranging from less than $5,000 to more than $50,000. Such variability affects the demographic composition across sessions. For example, among Hispanic subjects in the All-Hispanic session, the mean age is 32.3 years, 40% of the subjects were male and 40% report incomes less than $5,000 per year. For Hispanic subjects in the Majority-Navajo session, the mean age is 22.1 years, 25% of the subjects are male and 12.5% report incomes less than $5,000 per year. Similar variability exists among Navajos across sessions. Some studies have found that socio-demographic attributes are important determinants of behavior in the Ultimatum Game, for example, Harbaugh, Krause, and Liday (2000), Botelho et al. (2005), Eckel and Grossman (2001), Solnick (2001), Stanley and Tran (1998), Carter and Irons (1991), and Kahneman, Knetsch, and Thaler (1986). (16) To control for their effects, and to allow us to focus on cross-cultural and intercultural aspects of behavior, we conduct regression analyses of Proposers' offers and Responders' minimum acceptable offers (reservation prices) against the variables listed in Table 3.

Hispanic ethnicity is the omitted ethnicity variable in the models. Inter-ethnic effects are measured by the variables (2) and (3). The squared interaction term (3.a) between Navajo and percentage of subjects in a session from a different ethnic group is included as a result of our finding a nonlinear relationship between Navajo Proposer behavior and the ethnic composition of the session. (17) As we will note, however, this nonlinearity is largely a result of the behavior of two subjects. Such non-linearity was not observed among Hispanics.

We estimate two models for each role in the Ultimatum Game: Model 1, which includes demographic variables, and Model 2, which includes demographic and behavioral variables (i.e., responses from questions in Figure 1 and demographic questionnaire). (18) Our impression is that some experimental economists question the usefulness of asking subjects what they believe and why they made a particular decision. By presenting two specifications, we demonstrate that our conclusions are not affected by the inclusion of self-reported behavioral variables.

Variable 16 (Prob$0-3) captures a Proposer's beliefs about the likelihood that low offers would be accepted: this variable sums a subject's estimates of acceptance rates (0-100%) for each dollar amount from $0 through $3. Finally, about 10% of the sample had what we term "nonmonotonic" Responder preferences: After choosing to accept an offer, the subject chose to reject one or more offers that were higher. This pattern has been found in other studies, for example, Andreoni, Castillo, and Petrie (2003), and may derive from error, an aversion to unequal outcomes (regardless of who benefits), or some other unknown reason. We include a dummy variable (17) for these subjects but note that our results do not change by dropping these subjects or pooling them without the dummy variable.

Given evidence of heteroskedasticity, we use the Huber/White/sandwich estimator of variance, which produces robust estimates of the standard errors. (19) In order to address the potential correlation between subject beliefs and the error term, we perform an instrumental variables regression. We instrument for Prob$0-3 using the sum of absolute values of the differences between a player's estimated percentage of subjects that would accept each potential offer and the actual percentage of subjects that accept each offer (i.e., a measure of overall accuracy of a subject's beliefs). This sum is highly correlated with Prob$0-3, but unrelated to OFFER. (20)

Results from the regressions of Responder behavior (RESERV) and Proposer behavior (OFFER) are presented in Tables 4 and 5. These results will serve as a basis for responses to three questions: "Are findings of substantial cross-cultural differences in bargaining behavior limited to cultures in small, non-industrialized societies," "Can changes in the proportional representation of an ethnic group substantially affect behavior in the Ultimatum Game," and "If discrimination is observed, what are the likely causes?" We answer these questions by first examining the behavior of Responders and then focusing on the behavior of Proposers.

A. Responders

With respect to cross-cultural effects, Navajos have significantly lower reservation prices, on average, than Hispanics in both models (Table 4). For example, the ethnicity coefficients in Model 2 suggest that, depending on income, a Navajo subject will accept, on average, between $0.50 and $3.00 less than a Hispanic subject ($0.35-2.80 less in Model 1).

With respect to our second question concerning intercultural effects, the behaviors of both Hispanic and Navajo Proposers are significantly affected by the ethnic composition of the session in both models. Both Hispanics and Navajo discriminate against the other ethnic group in the sense that their mean reservation prices increase with an increase in the proportion of subjects from the other ethnic group. This effect is most pronounced with Hispanic subjects. If, for example, the subject pool was 25% Hispanic and 75% Navajo, Model 2 predicts that the average minimum acceptable offer of Hispanics would be about $1.34 more than if the pool were 100% Hispanic ($1.44 more in Model 1).

With regard to the demographic variables, married subjects have significantly lower reservation prices than single subjects by almost $1 on average. Hispanic, but not Navajo, reservation prices are positively related to income (if anything, poorer Navajos demand a little more of the pie). Evidence of gender effects on a Responder's reservation price is weak with males requiring about $0.45 less than females on average. A weakly negative effect also derives from exposure to economics courses. (21)

B. Proposers

In terms of cross-cultural effects among Proposers, we find a significant difference in the behavior of our two cultural groups in both specifications (Table 5). On average, Navajos offer less than Hispanics. For example, the ethnicity coefficients in Model 2 suggest that, depending on income levels, a Navajo subject offers, on average, between $1.35 and $2.67 less than a Hispanic subject ($1.27-2.50 less in Model 1). Thus, our observations of Proposer and Responder behavior imply that, even among cultures that coexist in a Western industrialized society, there is cross-cultural behavioral variability.

In terms of our intercultural question--does the ethnic mix of the session "matter"--we find the ethnic composition of the session has significant effects on offers. Hispanics make the highest offers to a Responder when all subjects are Hispanic and persistently lower offers as the percentage of Hispanics in the group decreases. For example, a Hispanic subject offers, on average, $1.27 less if Hispanics make up only 25% of the session rather than 100% ($1 less in Model 1).

Turning to Navajo Proposers, the nonlinear response to ethnic composition that was evident in Table 1 is also reflected in our regression results: Mean Navajo offers rise and then fall as their proportional representation of the session decreases (reflected in the positive sum of PercentOther and NavPercentOther and the negative sign on NavPercentOther (2)). However, much of this nonlinearity is driven by two influential observations.

Using Cook's (1977) distance to identify influential observations, we identified two Navajo subjects who offered $0 as the two most influential observations in Model 2 (#29 in the All-Navajo session; #33 in the Majority-Hispanic session). Deleting these subjects removes the observed nonlinearity in the data: The coefficient on NavPercentOther (2) is statistically no different from zero (p = 0.220). Removing the two influential observations and the squared variable from the regression yields the following coefficients: PercentOther = -0.019 (p < 0.001) and NavPercentOther = 0.036 (p = 0.001). This result implies that Hispanic offers decrease linearly in the proportion of Navajo subjects in the session (almost two cents for every 1% increase in the proportion of Navajos), while Navajo offers increase linearly in the proportion of Hispanic subjects in the session (almost two cents for every 1% increase in the proportion of Hispanics).

C. Predicted Proposer and Responder Behavior

In an effort to make clear these cross-cultural and intercultural effects, an example is given in Table 6. We consider two hypothetical subjects: a Navajo and a Hispanic subject, both 25-year-old single females with incomes in the $15,000$45,000 range (Fair {reserve} = 0, Fair {offer} = 0, Reject = 0, Known = 0, Payoff {reserve} = 1, Payoff {offer} = 1). Both subjects have monotonic Responder preferences and have average expectations of low offers being accepted. For various ethnic mixes, Table 6 gives the Responder reservation prices and Proposer offers that are predicted by the regressions reported in Tables 3 and 4 (Model 2). Because the nonlinearity observed in Model 2 for Proposer Offers was driven by two influential observations, we drop these two observations and use a re-estimated Offer model without the squared term "NavPercentOther (2)."

In ethnically homogeneous sessions, reservation prices and offers differ substantially between the Navajo and Hispanic "subjects." Moreover, as the percentage of Navajo subjects in a session increases, the Hispanic subject's reservation price increases and her offer decreases. For the Navajo subject, increase in the percentage of Hispanics in the session also results in increasing reservation prices. Her offer also increases as the percentage of Hispanics increases.

D. Statistical versus Preference-Based Discrimination

Navajos are more likely to accept low offers in mixed sessions. Thus statistical discrimination could explain Hispanic and Navajo Proposer behavior. The negative and significant coefficient Prob$0-3 (beliefs that low offers will be accepted) is, in fact, consistent with statistical discrimination. For example, a Proposer who believes low offers are unlikely to be accepted--only 5% of Responders will accept $0, 10% will accept $1,20% will accept $2, and 30% will accept $3--will offer $0.39 less than a person who believes the acceptance rate is double that.

If preference-based discrimination were absent among Hispanics, however, controlling for beliefs in Model 2 should render the coefficient on PercentOther statistically indistinguishable from zero (i.e., once beliefs are accounted for, varying the ethnic mix of the session should not affect the offer level). The coefficient on PercentOther, however, does not move toward zero when we control for beliefs and other behavioral variables in Model 2; in fact, it becomes more negative. This change implies that the observed discrimination against Navajos by Hispanics becomes stronger when beliefs are incorporated into the model (because, as we saw in Table 2, Hispanics believe Navajos are less likely to accept low offers, yet Hispanics are more likely to send low offers when surrounded by Navajos). This result is consistent with a Hispanic taste for discrimination against Navajos. (22) Note that although the term "taste for discrimination" has negative connotations, it does not necessarily denote "dislike." (23)

Turning to the behavior of Navajo Proposers, the combined sum of PercentOther and NavPercentOther (which reflects the mean response of Navajo Proposers to changes in the proportion of Hispanics) does move closer to zero but is still positive and substantial. This result provides evidence against the hypothesis that Navajo discrimination against Navajos is entirely a result of statistical discrimination. Removing the two influential Navajo observations (see above) and the squared interaction term does not change this conclusion (PercentOther = -0.019 (p<0.001) and NavPercentOther = 0.036 (p = 0.001)). (24)

A reader may find strange the conclusion drawn from the regressions that Navajos discriminate against Hispanics when they are Responders, but against Navajos when they are Proposers. Previous Ultimatum Game analyses, however, suggest that the framing of the Responder's decision is different from the framing of the Proposer's decision, and thus the operative decision variables are different. In the former decision, concerns of justice, fairness, and equity are operative, but in the latter decision, strategic concerns and other regarding preferences are operative. We do not pretend to understand why these observed patterns of preference-based discrimination take place, but we note that the results are consistent across alternative model specifications.

E. Demographic and Behavioral Variables in Proposer Models

Subjects reporting concern about potential rejection of their offer sent, on average, $0.63 more to the Responder (holding constant their beliefs). Subjects reporting a desire to keep as much money as possible sent, on average, $0.79 less to the Responder. There is no strong evidence that a subject's age, marital status, gender, income, experience with economics, non-monotonic preferences as a Responder, concern about potentially knowing the Responder with whom he or she would be paired, or concern about a fair division of the pie has an effect on the average offer.

Although subjects responded strategically to their beliefs, their beliefs were, on average, inaccurate; subjects were overly pessimistic about the likelihood of Responders rejecting low offers. To explore the determinants of the accuracy of subject beliefs, we select as the dependent variable the sum of absolute values of the differences between a subject's estimated percentage of subjects that would accept each potential offer and the actual percentage of subjects that accepted each offer. We regress this measure of accuracy on demographic variables (1-11 and 17). We find some evidence that Hispanic beliefs, on average, become less accurate as the percentage of Navajo subjects increases in the session and that Navajo beliefs, on average, become more accurate as the percentage of Hispanic subjects increases (PercentOther = 1.46, p = 0.069; NavPercentOther = -2.39, p = 0.029). Thus Hispanics are most accurate when surrounded by other Hispanics, and Navajos are least accurate when surrounded by other Navajos. We also observe that subjects who are older or had nonmonotonic preferences as Responders were, on average, less accurate in their beliefs (p < 0.05).

VI. CONCLUDING REMARKS

In this study, we depart from traditional empirical investigations to provide a framework to advance our understanding of the way in which culture may affect economic outcomes. Our results demonstrate that culture can matter in explaining variability in economic outcomes and in more ways than previous research has suggested. Hispanic and Navajo subjects not only behave differently in the Ultimatum Game, but they also respond differently to the ethnic composition of the session. Twenty-six years ago, Thomas Schelling (1978, 108) observed that "undoubtedly for some behaviors ... it is proportions that influence people, not absolute numbers." Our results provide empirical support for Professor Schelling's observation.

If individual behavior can be affected by cultural diversity, as well as by the subject's own culture, economists may need to reconsider the way in which they control for cultural differences in empirical analyses. Economists testing for cultural differences in behavior may find different results depending on variability of the cultural settings in which they are observing behavior. For example, were we to remove the variables that control for the ethnic composition of the sessions in our models of Proposer behavior (Table 4), the coefficient on the dummy variable for Navajo subjects would be small and not significantly different from zero, suggesting no cross-cultural differences. (25)

In addition to demonstrating how one can use experiments to explore the effects of cultural diversity on economic outcomes, we demonstrate how one can use data on behavior, beliefs, and motivations to provide insights into the cause of the observed behavioral variability that is conditional on the cultural mix of the experimental session.

We do not claim, however, that our study is a definitive study of cultural diversity or of relationships between Navajos and Hispanics. Like previous economic studies of culture, we cannot prove that we have eliminated all bias from unobservable factors that could be correlated with the ethnic mix of our sessions. However, we believe that improvements in experimental methods hold the best promise for minimizing such bias in the future.

For example, in hindsight, we should have tested Navajo subjects' ability to speak Navajo, in order to control for unobserved heterogeneity in subjects' identification with their ethnic group (similar efforts should be made with the Hispanic subjects). One must also be careful about making inferences about entire cultural groups based on experiments conducted in cities or around universities, when the majority of the cultural group does not live in such environments. We might have obtained quite different results had we conducted these experiments on or near Navajo reservations. Furthermore, with better recruiting techniques, analysts can make sessions more homogenous along a variety of non-ethnic characteristics and thus avoid depending on regression models to control for such heterogeneity.

One weakness of our design (and of most other discrimination studies) is our inability to assign motives to the observed statistical and preference-based discrimination. Future research should design ways to elucidate such motives (e.g., experimental treatments, focus groups). Finally, experimentalists must consider that subjects' perceptions of the experimentalists' cultural characteristics may influence behavior in experiments. In our experiment, the moderator was of European descent and may have been perceived as wealthy or of high status. Could such seemingly payoff-irrelevant characteristics influence subject behavior?

Our novel experimental approach to studying cross-cultural and intercultural effects on economic outcomes should be of general interest to economists. Throughout the world, policies are formulated in societies characterized by mixed ethnicity, race, and religion, in which there are clear majority and minority groups. Allocating the costs and benefits of public decisions across citizens (e.g., setting tax policy, providing public goods) is a crucial policy issue. The way in which citizens value the potential policy outcomes, however, may not only be affected by the cultural group to which they belong, but also by the group's relative size in the society.

Our experimental approach complements ongoing empirical and theoretical work on this subject, by offering control over confounding effects and clearer insights into causal relationships. With further experimentation on different subject pools and with different treatments, economists can begin to elucidate the facets of ethnic discrimination and the role that cultural diversity may play in economic outcomes. Our hope is that our experimental design and results will stimulate research designed to address these questions.

doi: 10.1111/j.1465-7295.2006.00013.x

REFERENCES

Alesina, A., R. Baqir, and W. Easterly. "Public Goods and Ethnic Divisions." Quarterly Journal of Economics, 114, 1997, 1243-84.

Alesina, A., A. Devleeschauwer, W. Easterly, S. Kurlat, and R. Wacziarg. "Fractionalization." Journal of Economic Growth, 8, 2003, 155-94.

Altonji, J. G., and R. Blank. "Race and Gender in the Labor Market," in Handbook of Labor Economics, Vol. 3C, edited by O. C. Ashenfelter and D. Card. New York: Elsevier Science B.V., 1999, 3143-59.

Andreoni, J., M. Castillo, and R. Petrie. "What Do Bargainers' Preferences Look Like? Exploring a convex Ultimatum Game." American Economic Review, 93, 2003, 672-85.

Arrow, K. "The Theory of Discrimination," in Discrimination in Labor Markets, edited by O. C. Ashenfelter and A. E. Rees. Princeton: Princeton University Press, 1973, 3-33.

Becker, G. The Economics of Discrimination. 2nd ed. Chicago: The University of Chicago Press, 1971.

Bertrand, M., and S. Mullainathan. "Are Emily and Greg More Employable than Lakisa and Jamal? A field experiment on labor market discrimination." American Economic Review, 94, 2004, 991-1013.

Botelho, A., G. W. Harrison, M. A. Hirsch, and E. E. Rutstrom. "Bargaining Behavior, Demographics and Nationality: What Can The Experimental Evidence Show?" in Field Experiments in Economies, edited by J. Carpenter, G. W. Harrison, and J. A. List. Greenwich: JAI Press, 2005, 337-72.

Brandts, J., T. Saijo, and A. Schram. "How Universal is Behavior? A Four Country Comparison of Spite, Cooperation and Errors in Voluntary Contribution Mechanisms." Public Choice, 119, 2004, 381-424.

Burlando, R., and J. D. Hey. "Do Anglo-Saxons free-ride more?" Journal of Public Economics, 64, 1997, 41-60.

Camerer, C. F. Behavioral Game Theory: Experiments in Strategic Interaction. Princeton: Princeton University Press, 2003.

Carpenter, J., S. Burks, and E. Verhoogen. "Comparing Students to Workers: The Effects of Stakes, Social Framing and Demographics on Bargaining Outcomes," in Field Experiments in Economics, edited by J. Carpenter, G. W. Harrison, and J. A. List. Greenwich: JAI Press, 2005, 261-90.

Carter, J. R., and M. D. Irons. "Are Economists Different, and if so, Why?" Journal of Economic Perspectives, 5, 1991, 171-77.

Collier. P. "Implications of Ethnic Diversity." Economic Policy, 16, 2001, 127-166.

Croson, R. T. A., and N. R. Buchan. "Gender and Culture: International Experimental Evidence from Trust Games." American Economic Review 89, 1999, 386-91.

Cook, R. D. "Detection of Influential Observations in Linear Regression." Technometrics, 19, 1977, 15-18.

Davidson, R., and J. D. MacKinnon. Estimation and Inference in Econometrics. New York: Oxford University Press, 1993.

Easterly, W., and R. Levine. "Africa's Growth Tragedy: Policies and Ethnic Divisions." The Quarterly Journal of Economics, 112, 1997, 1203-50.

Eckel, C. C., and P. J. Grossman. "Chivalry and Solidarity in Ultimatum Games." Economic Inquiry, 39, 2001, 171-88.

Fearon, J. "Ethnic and Cultural Diversity by Country." Journal of Economic Growth, 8, 2003, 195-222.

Fershtman, C., and U. Gneezy. "Discrimination in a Segmented Society: An Experimental Approach." Quarterly Journal of Economics, 116, 2001. 351-77.

Fershtman, C., Gneezy, U., and F. Verboven. "Discrimination and Nepotism: the Efficiency of the Anonymity Rule." Journal of Legal Studies, 34, 2005, 371-94.

Gil-White, F.J. "Are Ethnic Groups Biological 'Species' to the Human Brain? Essentialism in our Cognition of some Social Categories." Current Anthropology, 42, 2001, 515-54.

--. "Ultimatum Game with an Ethnicity Manipulation: Results from Khovdiin Bulgan Sum, Mongolia," in Foundations of Human Sociality: Ethnography and Experiments in 15 Small-Scale Societies, edited by J. Henrich, R. Boyd, S. Bowles, H. Gintis, E. Fehr, and C. Camerer. Oxford: Oxford University Press, 2004, 260-304.

Goeree, J. K., and C. A. Holt. "Ten Little Treasures of Game Theory and Ten Intuitive Contradictions." American Economic Review, 91, 2001, 1402-22.

Goldberg, P. K. "Dealer Price Discrimination in New Car Purchases: Evidence from the Consumer Expenditure Survey." Journal of Political Economy, 104, 1996, 622-54.

Guth, W., R. Schmittberger, and B. Schwarz. "An Experimental Analysis of Ultimatum Bargaining." Journal of Economic Behavior and Organization, 3, 1982, 367-88.

Guth, W., and R. Tietz. "Ultimatum Bargaining Behavior: A Survey and Comparison of Experimental Results." Journal of Economic Psychology, II, 1990, 417-49.

Guth, W., S. Huck, and P. Ockenfels. "Two-Level Ultimatum Bargaining with Incomplete Information: An Experimental Study." The Economic Journal, 106, 1996, 593-604.

Guth. W., and E. van Damme. "Information, Strategic Behavior and Fairness in Ultimatum Bargaining: An Experimental Study." Journal of Mathematical Psychology, 42, 1998, 227-47.

Guth, W., C. Schmidt, and M. Sutter. "Fairness in the Mail and Opportunism in the Internet: a newspaper experiment on Ultimatum bargaining." German Economic Review, 4, 2003, 243-65.

Harbaugh, W.T., K. Krause, and S. Liday. "Children's Bargaining Behavior: Differences by Age. Gender and Height." Working Paper, University of Oregon, October 2000.

Henrich, J. "Does Culture Matter in Economic Behavior? Ultimatum Game Bargaining Among the Machiguenga of the Peruvian Amazon." American Economic Review, 90, 2000, 973-79.

Henrich, J., R. Boyd, S. Bowles, C. Camerer, E. Fehr, H. Gintis, and R. McElreath. "In Search of Homo Economicus: Behavioral Experiments in 15 Small-Scale Societies." American Economic Review, 91, 2001, 73-8.

Henrich. J., R. Boyd, S. Bowles, H. Gintis, E. Fehr, and C. Camerer, eds. Foundations of Human Sociality: Ethnography and Experiments in 15 Small-Scale Societies. Oxford University Press, 2004.

Hollander, M., and D. A. Wolfe. Nonparametric Statistical Methods, 2nd ed. Etobicoke, Ontario: John Wiley and Sons, 1999.

Kahneman, D., J. L. Knetsch, and R. H. Thaler. "Fairness and Assumptions of Economics." Journal of Business, 59, 1986, 285-300.

Ladd, H. F. "Evidence of Discrimination in Mortgage Lending." Journal of Economic Perspectives, 12, 1998, 41-62.

List, J. A. "The Nature and Extent of Discrimination in the Marketplace: Evidence from the Field." Quarterly Journal of Economics, 119, 2004, 49-89.

Mallows, C. L. "Augmented partial residuals." Technometrics, 28, 1986, 313-19.

McElreath, R., R. Boyd, and P. Richerson. "Shared Norms Can Lead to Evolution of Ethnic Markers." Current Anthropology, 44, 2003, 122-30.

Miguel, T. "Ethnic Diversity, Mobility and School Funding: Theory and Evidence from Kenya." Working Paper, Harvard University, 1999.

Neumark. D., R. Bank, and K. van Nort. "Sex Discrimination in Restaurant Hiring: An audit study." Quarterly Journal of Economics. 111, 1996, 915-41.

Nisbett, R. E., and D. Cohen. Culture of Honor. Boulder: Westview Press, 1996.

Ockenfels, A., and J. Weimann. Types and Patterns: an Experimental East-West Comparison of Cooperation and Solidarity. Journal of Public Economics, 71, 1999, 275-87.

Oxoby, R. J., and K. N. McLeish. "Sequential Decision and Strategy Vector Methods in Ultimatum Bargaining: evidence on the strength of other-regarding behavior.'" Economics Letters, 404, 2004, 399-405.

Phelps, E. S. "The Statistical Theory of Racism and Sexism." American Economic Review, 62, 1972, 659-61.

Riach, P. A., and J. Rich. "Field Experiments of Discrimination in the Market Place." The Economic Journal, 112, 2002, F480-F518.

Roth, A. E., V. Prasnikar, M. Okuno-Fujiwara, and S. Zamir. "Bargaining and Market Behavior in Jerusalem, Ljubljana, Pittsburgh, and Tokyo: An Experimental Study." American Economic Review, 81, 1991, 1068-95.

Schelling, T. C. Micromotives and Macrobehavior. New York: Norton, 1978.

Solnick. S. J. "'Gender Differences in the Ultimatum Game." Economic Inquiry, 39, 2001, 189-200.

Sonnemans, J., and T. Offerman. "Is the Quadratic Scoring Rule Really Incentive Compatible?" Working Paper, December 2001.

Stanley, T. D., and U. Tran. "Economics Students Need not be Greedy: Fairness and the Ultimatum Game." Journal of Socio-Economics, 27, 1998, 657-64.

Thaler, R. H. "Anomalies: The Ultimatum Game." Journal of Economic Perspectives, 2, 1988, 195-206.

Yinger, J. "Evidence on Discrimination in Consumer Markets." Journal of Economic Perspectives, 12, 1998, 23-40.

(1.) Department of Commerce website: Statistical Abstract of the United States, 2001, Washington. DC. tables 23 and 24.

(2.) We thank anthropologist Joseph Henrich (Emory University) for directing us to the relevant literature and definitions.

(3.) We note, however, Gil-White's (2004) interesting study of the Ultimatum Game. Similar to FG, it paired members from two cultural groups (two Mongolian tribes). We do not explore this study at length because it used deception.

(4.) FG do not make clear if this discrimination was observed among both Ashkenazic (n = 24) and Eastern (n = 33) Proposers.

(5.) Using the Dictator Game to infer the source of discrimination in another game may be problematic if the framing of the two games generates different norms or behavioral strategies among subjects. As noted by Goeree and Holt (2001. 1418), an alternative approach would be to elicit beliefs directly as the game is played.

(6.) There are differences in offers made by dealers to minority sellers, but they are not statistically as meaningful.

(7.) Given our concern with offending subjects or the organizations from which we recruited them, we allowed subjects to complete the experiment even if they were unable to successfully complete practice questions or were demonstrably unable to comprehend questions. As a result, we exclude data from three subjects: one Navajo subject from the All-Navajo. session who could not respond to the practice question (even after repeated explanations by the experimenter), had trouble using the mouse, and rejected every possible offer and one Navajo subject from the All-Navajo session and one Hispanic subject from the Majority-Hispanic session, both of whom had obvious difficulty completing the practice question and who then clicked reject and accept in alternating fashion for every potential offer that could be sent to them. For these subjects, the idea of a minimum acceptable offer makes no sense, and it is unlikely that these subjects understood the main components of the experiment. We note, however, that including these subjects in the analysis by treating their first accepted offers as their Responder reservation prices does not affect our results. When estimating the percentage of Navajo and Hispanic in a session, we include these subjects because they were observable to every subject in the room (removing them from the percentage calculation does not affect our results).

(8.) Native American ethnicity is a requirement for entry into the Navajo organizations. Thus, presumably all subjects in the All-Navajo session were Navajo. However, one subject selected "Hispanic" on the post-experiment questionnaire. We are unsure if the subject was indeed Hispanic, was of mixed heritage and did not see the option for mixed ethnicity or made a mistake filling out the questionnaire, which was completed on a computer. We treat the subject as Hispanic, but note that deleting this subject or re-coding her as -Navajo" does not affect our results.

(9.) Some readers might wonder why we did not have a 50% Navajo and 50% Hispanic session. Our analysis is based on the assumption that subjects could clearly ascertain the ethnic mix of the session, if the mix were 50:50, some subjects may perceive they are in the minority, others may perceive they are in the majority, while others would correctly infer that no group is in the majority. We believe that we give up control in a 50:50 session and would not generate insights unattainable from the other four sessions (no-shows may have also made 50:50 difficult to achieve).

(10.) Subjects were told that the absolute values of the differences between their predicted percentages and the actual percentages for each potential offer would be summed. The subject with the lowest sum wins $10.00. Although this rule is not incentive compatible, it is highly transparent and can include truth telling as one best response. A best response that deviates from true beliefs under this rule requires sophisticated strategizing about the beliefs of others in the session and mathematical acumen to solve for a best response conditional on those beliefs. Moreover, a recent study by Sonnemans and Offerman (2001) found no significant difference between the beliefs elicited from a sophisticated quadratic scoring rule and beliefs elicited from a method that pays subjects a fixed (unconditional) payment: the offer of some compensation for effort was enough to induce subjects to think carefully about their beliefs.

(11.) Our anomalous results are not likely to derive from having players who play both roles. Conducting the same experiment at Georgia State University, we find only one-third willing to accept $1 or $2 (mean reservation price was $2.77). The mean offer in this session was $4.17. This session of 30 subjects had no culture in a majority or substantial minority: 14 foreign subjects from 10 different nations, five Hispanic. three African-American, and eight White. Although we limit our comparisons to other Ultimatum Game experiments that use the strategy method, Oxoby and McLeish (2004) find no difference in behavior between Proposer and Responder behavior when strategies are elicited through the strategy method or simply observed sequentially in the game.

(12.) Results from a Jonckheere-Terpstra test (with exact p-values) indicate a significant difference in Hispanic Responder behavior across sessions (p = 0.0015). No such significant difference is found among Navajo Responders (p = 0.2837). This test is a nonparametric test for ordered differences (trend) among classes and is preferable in this context to tests of more general class differences, for example, Kruskal--Wallis H test and Hollander and Wolfe (1999).

(13.) Results from a Jonckheer--Terpstra test (with exact p-values) indicate significant differences in Proposer behavior across sessions for both the Navajos (p = 0.0237) and Hispanics (p = 0.0590).

(14.) Results from a Jonckheere--Terpstra test indicate a significant difference in Hispanic beliefs across sessions at the 1% level for offers of $0-3, as well as for the sum of predicted acceptance rates from $0 to $3.

(15.) Results from a Jonckheere--Terpstra test indicate weakly significant differences in Navajo beliefs across sessions ($0: p = 0.1383: $1: p = 0.2242: $2: p = 0.0480: $3: p = 0.0311; sum of predictions $0-3: p = 0.0917). behavior of Hispanic Proposers, and there is weak evidence that such discrimination plays a role in the behavior of Navajo Proposers.

(16.) Previous Ultimatum Game studies have not studied marital status (subjects are typically college students), but 15% of our subject pool was married. We hypothesize that these subjects may behave differently in a bargaining situation.

(17.) We detected this nonlinearity using Mallows (1986) augmented component-plus-residuals plot.

(18.) We recognize that responses to some of the behavioral questions may be ambiguous (e.g., a subject may consider a 50:50 split "fair" as a Responder, but a 70:30 split "fair" as a Proposer), but we believe they are reasonable proxies for the behavioral variables that previous studies suggest are important in the Ultimatum Game as in Thaler (1988), Guth and Tietz (1990), Guth, Huck, and Ockenfels (1996), and Guth and van Damme (1998). We attempt to control for beliefs directly in the regression and thus assume that the variable Reject captures a subject's concern about rejection (e.g., two subjects may both believe a $1 offer has a 50:50 chance of being rejected, one subject may not be concerned about rejection, while the other may).

(19.) We also use Davidson and MacKinnon's more conservative heteroskedasticity-consistent (HC3) estimator without a substantial change in the standard errors. All regressions were run in Stata v.7.

(20.) Pearson correlation coefficient is -0.45 (p < 0. 0001). Using the ordinary least squares estimator does not change our inferences about the ethnicity variables, but does change the estimate and standard error of the coefficient on Prob$0-$3.

(21.) Removing the subject who reported taking 26 courses decreases the coefficient (0.02) and t-statistic (p = 0.79).

(22.) An alternative, but less plausible, explanation is that Hispanic subjects in the sessions with Navajos were less risk averse than Hispanic subjects in the All-Hispanic session and that our demographic and behavioral variables (e.g., Reject) do not fully control for these differences in risk aversion.

(23.) If Hispanics were to behave in the same manner toward an anonymous partner of unknown ethnicity, one might say that Hispanics do not discriminate against Navajos, but rather in favor of Hispanics as given in Fershtman, Gneezy. and Verboven, (2005). Exploring this distinction is beyond the scope of this paper.

(24.) See footnote 22 for an alternative explanation.

(25.) In mixed-race Ultimatum Game sessions, Carpenter, Burks, and Verhoogen (2005) infer that black students and black workers behave differently. Inspection of the racial composition of the sessions, however, shows dramatic differences. In the worker session, blacks are 25% of the session, while whites are 25%, Hispanics are 9%, and the rest are "nonwhite." In the student session, blacks are only 12% of the session, while whites are 51%, Hispanics are 9%, and the rest are nonwhites.

PAUL J. FERRARO and RONALD G. CUMMINGS *

* We thank Daniel Houser, Uri Gneezy, Joseph Henrich, Michael McKee, Ragan Petrie, and several anonymous referees for helpful comments. We also thank Dr. C. Arundale for helping to coordinate the recruiting and experiments, and Krawee Ackaramongkolrotn for software design.

Ferraro: Assistant Professor, Department of Economics, Andrew Young School of Policy Studies, Georgia State University, P. O. Box 3992, Atlanta. GA 30302-3992. Phone 1-404-651-1372, Fax 1-404-651-0425, E-mail [email protected]

Cummings: Professor Emeritus, Department of Economics, Andrew Young School of Policy Studies, Georgia State University, P. O. Box 3992, Atlanta, GA 30302-3992. Phone 1-404-651-3963, Fax 1-404-651-0425, E-mail [email protected]

We scheduled four sessions. The ethnic
composition of each session was as follows: (7)

Session 1 (All-Hispanic) 30 Hispanic subjects
Session 2 (Majority-Hispanic) 21 Hispanic and 6
 Navajo subjects
Session 3 (All-Navajo) 29 Navajo and
 I Hispanic subject (8)
Session 4 (Majority-Navajo) 23 Navajo and 7
 Hispanic subjects.

TABLE 1
Summary of Experiment Results

 Responder

 Average reservation price

Session Navajo ($) Hispanic ($)

All subjects 1.31 1.83
same ethnicity
Subject's ethnicity 1.78 2.73
is a majority
Subject's ethnicity 2.00 3.38
is a minority

 Responder

 Responder accepting $1.00 (%)

Session Navajo Hispanic

All subjects 62 60
same ethnicity
Subject's ethnicity 61 33
is a majority
Subject's ethnicity 50 13
is a minority

 Proposer

 Average offer

Session Navajo ($) Hispanic ($)

All subjects 3.83 4.90
same ethnicity
Subject's ethnicity 5.13 4.77
is a majority
Subject's ethnicity 4.17 (a) 4.50
is a minority

(a) The average offer increases from 54.17 to $5 if one influential
subject (#33) were removed. We discuss this influential observation
in the next section.

TABLE 2
Mean Estimated Percentage of Subjects Who Would Accept an Offer of $x

 Hispanic (%) Navajo (%)

Session $0 $1 $2 $3 $0 $1 $2 $3

All subjects 20 35 40 53 13 32 43 52
same ethnicity
Subject's ethnicity 7 27 33 41 19 25 30 37
is a majority
Subject's ethnicity 1 5 9 20 12 27 29 37
is a minority

TABLE 3
Variables Used in Regression Analyses

Variable Description

Dependent variables
RESERV Responder's reservation price
OFFER Proposer's offer
Independent variables
1. Navajo Dummy variable = 1 if subject is Navajo
2. PercentOther Percentage of subjects in session from an
 ethnic group different than that of the
 subject's [0, 96.9]
3. NavPercentOther Interaction between (1) and (2)
3.a (NavPercentOther) 2 (3) squared, used only in Offer equation
4. Age Subject's age
5. Male Dummy variable = 1 if subject is male
6. Econ Number of economics courses taken by subject
7. Less$15000 Dummy variable = 1 if subject's income is
 less than $15,000
8. $15-$45000 Dummy variable = 1 if subject's income is
 between $15,000 and $45,000
9. NavLess$15000 Interaction term between (1) and (7)
10. Nav$15-$45 Interaction term between (1) and (8)
11. Married Dummy variable = 1 if subject is married
12. Fair (*) Dummy variable = I if subject ranks concern
 for fairness (Figure 1) at 4 or 5 as a {*} =
 {Responder} or {Proposer}
13. Payoff (*) Dummy variable = 1 if subject ranks concern
 for earning money at 4 or 5 as a {*} =
 {Responder} or {Proposer}
14. Reject Dummy variable = 1 if Proposer ranks concern
 about Responder's rejection at 4 or 5
15. Known Dummy variable = 1 if Proposer ranks concern
 about knowing the Responder at 4 or 5
16. Prob$0-3 The sum of subject's estimates of the
 percentage of subjects who would accept $0,
 $1, $2, and $3
17. Nonmonotonic Dummy variable = 1 if Responder had
 nonmonotonic behavior (see text)

TABLE 4
Responder's Reservation Price as Dependent Variable

 Model 1

Independent variable Coefficient t-statistic
 (standard error) (p-value)

Constant 4.165 (0.801) 5.20 (<0.001)
Navajo -2.793 (0.642) -4.35 (<0.001)
PercentOther 0.019 (0.007) 2.88 (0.005)
NavPercentOther -0.010 (0.013) -0.77 (0.443)
Age 0.011 (0.017) 0.64 (0.527)
Male -0.446 (0.292) -1.53 (0.130)
Econ 0.079 (0.032) -2.31 (0.023)
Married -0.969 (0.385) -2.52 (0.013)
Less$15000 -2.355 (0.509) -4.63 (0.000)
$15-$45000 -1.793 (0.456) -3.94 (<0.001)
NavLess$15000 2.445 (1.027) 3.20 (0.002)
Nav$15-$45000 1.970 (0.838) 2.35 (0.021)
Nonmonotonic
 Fair (reserv)
 Payoff (reserv)

 Model 2

Independent variable Coefficient t-statistic
 (standard error) (p-value)

Constant 3.39 (0.875) 3.87 (<0.001)
Navajo -3.04 (0.730) -4.17 (<0.001)
PercentOther 0.018 (0.007) 2.61 (0.010)
NavPercentOther -0.009 (0.014) -0.66 (0.510)
Age 0.014 (0.017) 0.82 (0.415)
Male -0.421 (0.283) -1.48 (0.141)
Econ -0.060 (0.031) -1.94 (0.056)
Married -0.970 (0.394) -2.46 (0.016)
Less$15000 -2.057 (0.607) -3.39 (0.001)
$15-$45000 -1.568 (0.548) -2.86 (0.005)
NavLess$15000 2.564 (0.831) 3.08 (0.003)
Nav$15-$45000 2.092 (0.942) 2.22 (0.029)
Nonmonotonic -0.525 (0.403) -1.30 (0.196)
 Fair (reserv) 0.928 (0.403) 3.25 (0.002)
 Payoff (reserv) -0.021 (0.316) -0.06 (0.948)

Model 1: F(11.105) = 10.27 (p<0.001): [R.sup.2] = 0.24; Root
MSE = 1.50

Model 2: F(14,102) = 7.48 (p<0.001); [R.sup.2] = 0.32; Root
MSE = 1.44

TABLE 5
Proposer's Offer as Dependent Variable

 Model 1

Independent variable Coefficient t-statistic
 (standard error) (p-value)

Constant 6.642 (0.914) 7.27 (<0.001)
Navajo -2.502 (1.151) -2.17 (0.032)
PercentOther -0.013 (0.007) -1.98 (0.050)
NavPercentOther 0.107 (0.034) 3.18 (0.002)
NavPercentOther (2) -0.001 (0.000) -2.75 (0.007)
Age -0.029 (0.019) -1.54 (0.126)
Male 0.092 (0.314) 0.29 (0.771)
Econ 0.052 (0.032) 1.62 (0.109)
Married -0.245 (0.442) -0.55 (0.580)
Less$15,000 -1.060 (0.470) -2.26 (0.026)
$15-$45,000 -0.689 (0.472) -1.46 (0.147)
NavLess$15,000 1.232 (1.027) 1.20 (0.233)
Nav$15-$45,000 1.050 (1.170) 0.90 (0.372)
Nonmonotonic
Fair (offer)
Payoff (offer)
Reject
Known
Prob$0-3

 Model 2

Independent variable Coefficient t-statistic
 (standard error) (p-value)

Constant 6.596 (1.163) 5.67 (0.001)
Navajo -2.671 (1.005) -2.66 (0.009)
PercentOther -0.017 (0.005) -3.28 (0.001)
NavPercentOther 0.087 (0.031) 2.74 (0.007)
NavPercentOther (2) -0.0008 (0.0004) -1.99 (0.049)
Age -0.019 (0.018) -1.06 (0.292)
Male 0.548 (0.380) 1.44 (0.153)
Econ 0.039 (0.027) 1.40 (0.164)
Married 0.104 (0.435) 0.24 (0.812)
Less$15,000 -0.472 (0.675) -0.70 (0.486)
$15-$45,000 -0.427 (0.675) -0.68 (0.495)
NavLess$15,000 1.318 (0.904) 1.46 (0.148)
Nav$15-$45,000 1.440 (1.068) 1.35 (0.181)
Nonmonotonic -0.34 (0.446) -0.76 (0.448)
Fair (offer) -0.35 (0.334) -1.05 (0.297)
Payoff (offer) -0.791 (0.371) 2.00 (0.048)
Reject 0.631 (0.371) 1.70 (0.092)
Known -0.302 (0.576) -0.52 (0.602)
Prob$0-3 -0.006 (0.003) -2.24 (0.027)

Notes: Model 1: F(12,104) = 1.97 (P = 0.034); [R.sup.2] = 0.15;
Root MSE = 1.56.

Model 2: F(18,98) = 2.53 (p = 0.002); [R.sup.2] = 0.20; Root
MSE = 1.57.

TABLE 6
Comparison of Hypothetical Navajo and Hispanic Subjects with
Identical Attributes

 Navajo Hispanic

 Minimum Minimum
Percentage of "other" acceptable acceptable
ethnic group in session offer ($) Offer ($) offer ($) Offer ($)

0 1.19 3.13 2.14 4.33
20 1.37 3.46 2.50 3.95
50 1.63 3.96 3.03 3.38
80 1.89 4.45 3.57 2.81