Printer Friendly

Holding fast: the persistence and dominance of gender stereotypes.

I. INTRODUCTION

In any number of situations it may be necessary for one party to make judgments as to the risk preferences of another: a lawyer negotiating a plea agreement for her client; a financial advisor developing an investment plan for a family; a doctor prescribing a course of treatment for his patient; or a real estate agent recommending how to market a seller's property. Often this judgment is based on fairly limited information. In many cases, the prediction may be based on little more than the predictor's "read" of the individual. Financial advisors frequently have their clients complete a short risk assessment survey, but even this type of instrument is likely to give only a cursory indication of a client's true risk preferences and is likely not tested for reliability. In such circumstances, predicting the risk preferences of another may constitute a guess informed by little more than visual or verbal clues provided in a brief meeting. Lacking more relevant information, the predictor may resort to stereotypes to inform his prediction.

Stereotyping is the act of assigning to a member of a particular group a characteristic or trait based solely on the individual's membership in that group. An individual is not seen as a distinct being with his own individual attributes but solely as a member of a group conforming to some pattern. In cases where individuating, judgment-relevant information is not available, drawing on stereotypes may improve one's ability to predict another's actions (assuming the stereotype contains some kernel of truth). When presented with individuating, judgment-relevant information regarding the characteristic or action being predicted, downplaying any stereotype in favor of this information should improve the accuracy of predictions.

Whether men and women differ in their attitudes toward risk and in their willingness to accept risk is the subject of much debate. Most evidence suggests that women perceive situations as inherently riskier than men perceive the same situations, women engage in less risky behavior, and they choose alternatives that involve less risk. (1) Consistent with the evidence of a gender difference in risk aversion, Ball, Eckel, and Heracleous (2010), Daruvala (2007), Eckel and Grossman (2008, EG hereafter), Grossman and Lugovskyy (2011), and Siegrist, Cvetkovich, and Gutscher (2002) report evidence suggesting that women are perceived to be more risk averse than men; when predicting the risk choices of others, experiment subjects apply the gender stereotype.

This paper tests the relative importance of individuating information and gender stereotypes when predicting the risk preferences of others. For this study, all subjects participate in a four-part experiment. First, each subject completes a risk-assessment survey. Second, they then each select a gamble to play from a set of six gambles differing in expected payoff and degree of risk. Third, each subject then attempts to predict the gamble choice of every other subject. In one treatment, subjects make initial predictions having only visual clues regarding each subject (i.e., subjects stand up one at a time). Fourth, subjects are permitted to revise their predictions after being provided individuating information (i.e., the other subjects' responses to two of the risk-assessment survey questions). In the second treatment, the order is reversed and initial predictions are first made based on survey responses and then predictions are reassessed based upon visual clues.

I find: (1) additional evidence of the existence of gender stereotyping, and more importantly, (2) the persistence of such stereotyping even when other individuating, both judgment-relevant and judgment-irrelevant, information is provided. Results indicate that in isolation both the gender stereotype and the individuating information condition initial predictions. However, and more importantly, when visual clues are provided first, the revised predictions, after a subject's survey responses are known, only marginally reduce the evidence of stereotyping. When the individuating survey responses are provided first, the provision of the visual clues leads to systematic revisions in predictions consistent with gender stereotyping. The results suggest that individuating information is dominated by the gender stereotype.

Section II reviews the relevant literature. Section III explains the experimental design and procedures. Section IV presents the results of the experiment and Section V draws conclusions from the results.

II. LITERATURE REVIEW

Psychology literature has extensively addressed how and when social stereotypes are used in judging others. Dual-process theories argue that there are two different ways of judging a person: either by relying on social stereotypes or by assessing the individual's specific attributes or qualities. (2) Gill (2004, 629) argues that the general conclusion is that "... judgment-relevant behavioral information can undercut stereotyping in judgments of individual group members." However, this conclusion is not universally supported. Nisbett, Zukier, and Lemley (1981) and Beckett and Park (1995) argue that the dominance of individuating information over stereotypes is only evident when considering weak stereotypes or when the individuating information was made salient and the stereotypes were not made salient. Studies by Sherman (1996) and Trope and Thompson (1997) find that stereotypes dominate individuating information. Kunda, Sinclair, and Griffin (1997) find that individuating information weakens the effects of stereotypes regarding assessment of a target's traits but has little effect on stereotypes when predicting the same target's trait-related behavior.

Bodenhausen, Macrae, and Sherman (1999, 280-81) note that "[W]ell-educated undergraduates may often be quite reluctant to furnish stereotypic reactions and may well be on their guard to censor such responses when their behavior is being monitored by a researcher." This highlights an important difference between many psychology experiments and economic experiments. In a psychology experiment, compensation is for showing up and not tied to decisions made; in an economic experiment, compensation is determined by the decisions made. With monetary incentives, subjects increase their earnings by correctly predicting the behavior of the targets; they have less incentive to censor their stereotypic responses.

Hsee and Weber (1997) and Siegrist, Cvetkovich, and Gutscher (2002) conducted tests of stereotyping and risk preference predictions. Hsee and Weber (1998) report that American subjects are predicted (wrongly) to be more risk-seeking than are Chinese subjects. Siegrist, Cvetkovich, and Gutscher (2002) find a pattern of predictions suggesting that "... gender stereotypes and own-based risk preferences influence predictions about other individuals." Both of these studies had hypothetical stakes.

The first study to investigate gender differences in predictions in an environment with substantial stakes is EG. In the EG study, all subjects first selected one of five 50/50 gambles to play. One gamble was a sure thing ($10); the remaining four increased linearly in expected earnings and risk (defined as standard deviation in earnings). Subjects, acting as predictors, predicted which of the five gambles each of the other subjects, acting as targets, selected. When a target's risk choice was to be predicted, that target stood so the predictors could view him/her. (3) EG find evidence of gender stereotyping; predictors predicted females to be more risk averse than males.

Grossman and Lugovskyy (2011) use EG's risk instrument (modified to include a sixth gamble choice) and then have predictors make predictions in one of three treatments. (4) Treatment 1 uses the EG procedure. In Treatment 2, predictors are provided the targets' responses to two statements from the Weber, Blais, and Betz (2002) risk preference survey (one of which is a gamble-oriented statement), but no visual clues. For Treatment 3, predictors receive both the visual clues and the survey responses. Grossman and Lugovskyy report results consistent with persistent gender stereotyping. Even when provided the individuating, judgment-relevant information, the survey responses, predictors still made predictions consistent with gender stereotyping.

Ball, Eckel, and Heracleous (2010) use EG's procedures but instruct predictors to act as financial advisors and choose gambles for each of the targets. Ball, Eckel, and Heracleous report that advisors make choices consistent with gender stereotyping. Daruvala (2007) uses a modification of the Becker-Degroot-Marschack instrument to assess subjects' risk attitudes. She finds that both male and female predictors predict significantly lower certainty equivalences for female targets than male targets.

With the exception of Grossman and Lugovskyy (2011), the studies since EG limit the information predictors have to inform their predictions of the targets' gamble choices. Predictors only see the targets, so they can either use the visual clues and apply a gender stereotype or ignore the visual clues and randomly guess their targets' gamble choices. Their results suggest that a gender stereotype will be used if visual clues are all that are available. Grossman and Lugovskyy find that individuating information will be used if that is all that is available. They also show that when both types of information are available simultaneously, predictors still use the gender stereotype. What they are unable to determine is if this is because gender is easier and quicker to assimilate and assess than the individuating information or if predictors focused more on the visually provided gender information as opposed to the computer-provided survey responses. This study addresses this issue by providing the different information sequentially, thereby avoiding the problem of two bits of information competing for the predictor's attention. It considers whether sequentially provided information alters predictions in a systematic way and if the order in which information is provided matters for gender stereotyping.

III. THE EXPERIMENT

A. Design

The experiment consists of four tasks: a psychological survey measure of risk attitudes, a gamble choice with substantial financial stakes, predictions of others' gamble choices, and the ability to revise predictions of others' gamble choices. For the first task, subjects are asked to complete the Weber, Blais, and Betz (2002) Domain Specific, Risk-Attitude Scale (DSRAS) survey. (5) They are informed they will earn $8 for completing the survey. (6) But the subjects are made aware that this money may be at risk at a later stage of the experiment.

The DSRAS survey consists of 50 statements. The survey is designed to measure risk attitudes across five domains: financial, ethical, health/safety, recreational, and social. Subjects indicate on a 5-point Likert scale both their likelihood of engaging in a risky activity (1 = extremely unlikely; 5 = extremely likely), their perception of the risk associated with that activity (1 = not at all risky; 5 = extremely risky), and the benefits they expected they would obtain from each situation (1 = no benefits at all; 5 = great benefits). Sample statements include:
 Arguing with a friend, who has a very different
 opinion on an issue (Social).
 Investing 10% of your annual income in a very
 speculative stock (Financial).
 Buying an illegal drug for your own use (Health).
 Chasing a tornado by car to take photos that you can
 sell to the press (Recreational).
 Cheating on an exam (Ethical). (7)


For the second task, we use the same risk instrument employed by Grossman and Lugovskyy (2011), which is an adaptation of the instrument introduced by Eckel and Grossman (2002, 2008). (8) Subjects are presented with six different gambles. Each gamble has two outcomes, each with equal probability of occurring. Table 1 lists the six gambles, their payoffs and risks (subjects only receive columns 1-4 of Table 1). The gambles (excluding Gamble 6) are designed so that their expected payoff and risk increase linearly. (9) Gamble 1 offers a sure payoff of $10. The expected payoff increases by $2 with each of Gambles 2 through 5. Gamble 6 has the same payoff as Gamble 5 but with higher associated risk. This gamble is included to distinguish the risk-lovers. Gambles 1-3 offer positive payoffs regardless of the outcome. Gambles 4-6 have negative payoffs should outcome B occur. All subjects in a session select one gamble to be played at a later stage in the experiment. They are informed that if they select a gamble with a negative payoff for outcome B and outcome B occurs, those payoff amounts will be deducted from their $8 payment for completing the survey. (10)

For the third task, subjects predict the gamble choices of every other subject in their session. In this task, and the subsequent task, subjects participate both as a target subject (i.e., a subject whose Task 2 gamble choice is to be predicted by the other subjects) and as a predictor subject (i.e., a subject who attempts to predict the Task 2 gamble choices of the target subjects). Each predictor makes n-1 predictions, where n = the number of subjects in the session. To motivate predictors to make their best predictions, predictors earn $1 for every correct prediction.

Two treatments are used in this task. The VISUAL/SURVEY treatment (V/S hereafter) follows the EG procedure. The EG protocol is "... designed to activate subconscious stereotypes (p. 9)." A target stands up and the predictors predict the target's gamble choice. This procedure is followed for all subjects in a session. All sessions are mixed gender and no mention of gender or visual characteristics is made so subjects are unaware of the purpose of the task. After initial predictions are made for all targets, the predictors are provided the opportunity to revise their predictions in Task 4. Targets are again asked to stand up in the same order, but this time additional information--the targets' responses to two DSRAS survey "likelihood of engaging in the activity" statements (one social and one financial)--is provided to the predictors. The social risk statement is: Arguing with a friend, who has a very different opinion on an issue. The financial risk statement is: Investing 10% of your annual income in a very speculative stock. The predictors are informed of the target's self-reported likelihood of engaging in each activity response for both statements. Predictors are asked to again predict the target's gamble based on this additional information.

The financial statement survey response is provided to give the predictors individuating, judgment-relevant information and the social statement response is provided to give individuating, judgment-irrelevant information. These two responses are used to determine if predictors recognize the type of information provided and apply it appropriately. The financial statement used was selected because it projects the air of a gamble since the statement refers to a "very speculative stock." (11) It is believed that, since both the gamble decision and the financial survey statement involve money placed at risk, a subject's response to this statement and his gamble decision will be highly correlated. (12) Even more importantly, it is believed that predictors will believe that the two correlate well. Targets' responses to the financial statement provide predictors with individuated information more relevant to the action being predicted than to gender.

The response to the social risk statement is provided to determine if predictors use or ignore information that should be irrelevant to the action being predicted. While any one of the nonfinancial risk statements might have been used I wanted a statement that was as devoid of financial implications as possible. Many of the possible alternatives involve risk that could have financial repercussions. For example, some of the health and recreational risk statements involved activities that could lead to physical injury and financial consequences. Some of the ethics statements involved illegal activities that could also have financial repercussions. The "arguing with a friend" statement was selected since it has no obvious financial dimension and the type of risk it entails seems to be as different as possible from the risk involved with a gamble. This seems to be borne out by the fact that subjects' responses to the two statements are uncorrelated (Spearman r = -0.027, p value = .785).

For the SURVEY/VISUAL treatment (S/V hereafter), the predictors are first given only the target's two survey responses. After initial predictions are made for all targets, the responses are again given while now the target is asked to stand up. Predictors are asked to again predict the target's gamble based on this additional information.

The two-treatment design allows a test of whether predictions are driven by a gender stereotype, and whether the provision of individuating, judgment-relevant information about the subject's risk-taking choices significantly diminishes the role of the stereotype.

B. Procedures

All sessions of the experiment follow a standard procedure. The experiment is conducted with pencil and paper in classroom settings. (13) Subjects are seated; consent forms are distributed, signed, and then collected. Instructions, DSRAS survey forms, and slips of paper with ID numbers are randomly distributed. The instructions are read aloud and subjects are instructed that if they have any questions they should raise their hands and one of the experimenters will answer their questions. Subjects are informed that they will be paid $8 for completing the DSRAS survey.

After all subjects have completed the DSRAS survey the forms are collected. Instructions and forms for the gamble choice task are then distributed. Subjects are presented with the six gambles and asked to choose one. After making their gamble choices, subjects deposit their forms in a box at the front of the room.

Once the gamble choice task is completed the instructions and forms for the initial prediction task are distributed. For the V/S treatment, subjects are called one-by-one by ID number to stand up. As a target subject stands, the predictors (i.e., the n - 1 other subjects in the session) are asked to record their initial predictions for the target subject's gamble choice. This process is repeated for all subjects in the session. For the S/V treatment, one-by-one an ID number is announced and that target's survey responses are written on the blackboard (the statements are also written on the board). Predictors record a prediction for the target. This process is repeated for all subjects in the session. After predictions are recorded for all subjects in a session, the forms are collected.

Instructions and forms for Task 4 are then distributed. Subjects are told they can revise their initial predictions based on the new information to be provided. For the V/S treatment, subjects are called one-by-one by ID number to stand up. In addition to having the target subject stand, the target's survey responses are written on the blackboard (along with the statements). Predictors record their revised predictions for the target. (14) This process is repeated for all subjects in the session. For the S/V treatment, one-by-one an ID number is announced with that subject's two survey responses written on the blackboard. In addition, the target is also called upon to stand. Predictors record their revised predictions for the target. This process is repeated for all subjects in the session. After everyone completes Task 4 the forms are collected. While earnings are calculated, subjects complete a survey to collect subject characteristics.

Five subjects from each session (20 in total) are selected at random for payment (subjects are informed of this fact at the onset of each session). Index cards with ID numbers on them are placed in a box and a randomly selected subject draws five cards from the box. The five chosen subjects each roll a six-sided die to determine whether Event A or B occurs for the gamble choices they made in Task 2. A roll of 1, 2, or 3 results in Event A; a roll of 4, 5, or 6 results in Event B. The chosen subjects also receive an extra $1 for every correct Task 4 revised prediction. Subjects are privately informed of their individual earnings and paid. They are then free to go.

C. Hypotheses

Based on the design of the experiment and on the findings of previous research, the following hypotheses are made (a brief justification for each follows):

H1: Gamble choices are independent of treatment.

Treatment procedures are the same through Task 2 so there should be no differences in gamble choices by treatment.

H2: Male subjects are less risk averse than female subjects.

Findings from previous studies suggest that men are less risk averse than women.

H3: Subjects will condition their initial predictions on the information available.

Prior evidence indicates that subjects will employ gender stereotyping when the only information available is visual clues and they will use individuating information when it is the only information available. I hypothesize that, in the V/S treatment, predictions for male targets will be significantly higher than predictions for female targets; in the S/V treatment, predictions will be positively and significantly correlated with the responses to the financial risk statement (the judgment-relevant information) and uncorrelated with the social risk statement (the judgment-irrelevant information).

H4: In the V/S treatment, the revised predictions will not be significantly different from the initial predictions.

The gender stereotype regarding differences in risk preferences appears, based on the prior studies, to be a strongly held stereotype. Furthermore, predictors are predicting trait-related behavior. Both factors argue against substantial revisions of the initial predictions. I hypothesize that gender stereotyping will dominate individuating, judgment-relevant information. Predictors in the V/S treatment will not systematically and significantly alter their initial predictions when provided the individuating, judgment-relevant information. The revised predictions will not be significantly different from the initial predictions nor correlated with the individuating, judgment-relevant information.

H5: In the S/V treatment, the revised predictions will be significantly different from the initial predictions.

The gender stereotype regarding differences in risk preferences appears, based on the prior studies, to be a strongly held stereotype. Furthermore, predictors are predicting trait-related behavior. Both factors argue for substantial revisions of the initial predictions. I hypothesize that gender stereotyping will dominate individuating, judgment-relevant information. Predictors in the S/V treatment will systematically and significantly alter their initial predictions when provided gender information. The revised predictions will be significantly different from the initial predictions in a manner consistent with gender stereotyping (i.e., initial predictions for male [female] targets will be revised up [down]).

H6: Additional information, regardless of its nature, should improve (or at least not decrease) the accuracy of predictions. The accuracy of the revised predictions should be equal to or higher than the accuracy of the initial predictions.

Predictors may believe that both a target's gender and the target's individuating information are correlated with the target's gamble choice and, therefore, have some predictive power. If the additional information is believed to be judgment-relevant, it can be used in place of judgment-irrelevant information to improve the accuracy of predictions. If the additional information is believed to be judgment-irrelevant, it can be ignored.

IV. RESULTS

A. Subject Characteristics

A total of 90 individuals (51 men and 39 women) made useable gamble predictions in the experiment. (15) There were four sessions with 17, 21, 25, and 27 participants. There were two sessions for each treatment. Fifty-two subjects participated in S/V sessions and 38 subjects in V/S sessions. Table 2 reports subject characteristics by treatment. There are no significant differences in characteristics across treatments. Sessions lasted approximately 90 minutes and the 20 subjects who were paid earned an average of $33.50.

B. Gamble Choices

Table 3 reports mean gamble choices for all subjects by treatment and by gender. The overall mean gamble choice is 4.16 (SD = 1.35). (16) Mean gamble choice in the S/V treatment is 4.19 (SD = 1.17); mean gamble choice in the V/S treatment is 4.13 (SD = 1.58). A Wilcoxon two-sample test cannot reject the null hypothesis that gamble choices do not differ by treatment (p-value = .96). (17) Women are only slightly more risk averse than men: the mean gamble choices for women and men are 4.05 (SD = 1.12) and 4.25 (SD = 1.51), respectively (means test and Wilcoxon two-sample test p-values [greater than or equal to].45). This result is consistent with the findings of Daruvala (2007) and Grossman and Lugovskyy (2011). (18) This gives the first two results:

Result 1: Gamble choices are independent of the experiment's treatments. HI cannot be rejected.

Result 2: Gamble choices did not differ significantly by gender. H2 is not supported.

C. Initial Gamble Predictions

Before making their initial predictions, predictors in the V/S treatment visually observe the gender of the targets; in the S/V treatment predictors are provided the targets' survey responses. Table 4 reports the mean of Task 3 initial predictions (PI) by treatment, gender of the predictor, and gender of the target. The mean gamble prediction ([P1.sup.*]) reported is calculated as follows:

(1) [P1.sup.*] = (1/n) [[summation].sup.n.sub.i=1]//m [[summation].sup.m.sub.j=1] p 1ij,

where i = the number of predictors and j = the number of targets. (19) [P1.sup.*] is calculated for each predictor gender/target gender pairing. In the V/S treatment the predictors have only visual clues on which to base their initial predictions; in the S/V treatment predictors have only the targets' survey responses on which to base their initial predictions. (20) Regardless of the treatment, both male and female predictors predict a significantly higher gamble choice for male targets than for female targets. For the S/V treatment, this is consistent with the fact that males indicated significantly less risk aversion than did women in their responses to the financial risk survey statement (means test p-value <.10; Wilcoxon two-sample test p-value <.09). (21) In both treatments and for all predictor gender/target gender pairings, means tests and Wilcoxon matched-pairs signed-ranks tests both reject the null hypothesis of no differences in initial predictions for male and female targets (p-values [less than or equal to] .08). (22

Table 6 reports regression results for ordered probit with clustered standard errors for each treatment (gamble choices are ordered from 0 [for Gamble 1] to 5 [for Gamble 6]). (23) The results suggest that predictors do condition their predictions on the information provided. The initial predictions in the S/V treatment are significantly correlated with the two survey responses:

Financial Risk Statement and Social Risk Statement. The results suggest that predictors treat both survey responses, regardless of the type of risky behavior assessed, as predictive of gamble behavior. For the V/S treatment, the results indicate that initial predictions are significantly correlated with the Male Target (male = 1). In both treatments, predictions are conditioned on the Predictor Gamble Choice and the results suggest that more risk-loving predictors assume others are likewise more risk loving. Male Predictor (male = 1) is insignificantly correlated with PI. These findings give the third result:

Result 3: Predictors in both treatments condition their initial predictions based on the target information provided. Predictions in the V/S treatment are consistent with gender stereotyping; predictions in the S/V treatment are correlated with the individuating information. H3 cannot be rejected.

D. Revised Gamble Predictions

Table 7 reports the mean of the Task 4 revised predictions (P2) by treatment, gender of the predictor, and gender of the target. The mean gamble prediction ([P2.sup.*]) reported is calculated as follows:

(2) [P2.sup.*] = (1/n) [[summation].sup.n.sub.i=1] 1/m [[summation].sup.m.sub.j=1] p2ij,

where i = the number of predictors and j = the number of targets. (24) In the V/S treatment, the predictors now have the targets' two survey responses in addition to visual clues as a basis for their revised predictions; in the S/V treatment predictors now have visual clues in addition to the targets' two survey responses as a basis for their revised predictions. As with the initial predictions, for every predictor gender/target gender pairing, in both treatments, means tests and Wilcoxon matched-pairs signed-ranks tests both reject the null hypothesis of no differences in predictions for male and female targets (p-values [less than or equal to] .01).

To assess if and how the predictors make use of the additional information provided to them prior to their making their revised predictions, I tested whether P2 was significantly different from P1. Table 8 reports the sign of the change (P2 - P1) and p-values for both paired means tests and Wilcoxon matched-pairs signed-ranks tests by treatment, gender of the predictor, and gender of the target. In the V/S treatment, providing the predictors with the targets' survey responses resulted in no significant changes in predictions for all but one predictor gender/target gender pairing. After receiving the survey responses, only females predicting for male targets significantly revised their predictions down (paired means test p-value [less than or equal to] .03; Wilcoxon matched-pairs signed-ranks test p-value [less than or equal to] .04).

In the S/V treatment, the additional information (visual clues) about the targets provided to predictors resulted in significant changes in predictions for every predictor gender/target gender pairing but one, and, more importantly, the direction of the changes is consistent with gender stereotyping. After receiving the visual clues, both male and female predictors significantly raised their predictions for male targets (paired means test p-values [less than or equal to] .06; Wilcoxon matched-pairs signed-ranks test p-values [less than or equal to] .01). Female predictors significantly lowered their predictions for female targets (paired means test p-value = .02; Wilcoxon matched-pairs signed-ranks test p-value=.03). Only males' predictions for female targets were not lowered significantly. (25)

Table 9 reports regression results for ordered probit with clustered standard errors for P2 for each treatment. (26) In each treatment regression, I control for PI and the new information provided the predictors. PI is included to control for the information provided when making the initial prediction (i.e., the predictor's own gamble choice and whether the predictor was male, as well as whether the target was male in the V/S treatment and the survey responses in the S/V treatment). In both the V/S and the S/V treatment, the revised predictions are strongly and positively determined by their respective P1s (i.e., if a predictor's P1 for a specific target is high, the predictor's P2 for that same target also tends to be high). In the V/S treatment, the newly provided Financial Risk Statement and Social Risk Statement survey responses do not significantly affect the revised predictions; predictors ignore the individuating information in favor of their initial, gender-stereotyped predictions. In the S/V treatment, the newly provided visual information (Male Target) does significantly and positively influence the revised predictions and in a manner consistent with gender stereotyping. (27) The reported results are consistent with the results reported in Table 8. These findings give the fourth and fifth results:

Result 4: In the V/S treatment, subjects hold fast to their gender-stereotyped predictions and do not systematically alter their predictions in response to the new individuating information provided. The initially provided visual clues dominate subsequently provided individuating, judgment-relevant and irrelevant information. H4 cannot be rejected.

Result 5: In the S/V treatment, subjects systematically alter their predictions in response to the new information provided. The initially provided individuating, judgment-relevant and irrelevant information is discounted by predictors in favor of gender stereotypes when the gender of targets becomes known. Individuating, judgment-relevant (or irrelevant) information does not suppress stereotyping. H5 cannot be rejected.

Results 4 and 5 may be explained by what the psychology literature calls the anchoring effect. (28) Anchoring occurs when individuals focus too much on some initial stimulus (relevant or irrelevant) resulting in a systematic bias in their responses to a subsequent question. Tversky and Kahneman (1974) define anchoring as when "... different starting points yield different estimates which are biased toward the initial values (p. 185)." Mussweiler (2001) shows that anchoring effects are lasting, Wilson et al. (1996) report that neither forewarning of the phenomenon nor providing incentives to be accurate eliminated anchoring effects, and Thorsteinson et al. (2008) find that highly applicable anchors had a greater effect than less applicable anchors.

In this experiment, gender appears to be the anchor on which subjects based their ultimate predictions. As noted earlier, there is considerable evidence (across a wide variety of risk domains) that men are less risk averse than women. In the V/S treatment, subjects may have entered the experiment anchored to the gender stereotype and perceived it as a highly applicable stereotype. The subsequently provided survey responses, however applicable, may not have been perceived as sufficiently more useful to warrant dismissing the gender stereotype. In the S/V treatment, the gender stereotype was initially irrelevant since no visual clues were provided. Subjects had only the survey responses on which to condition their initial predictions.

The responses were perceived to be sufficiently applicable and were employed in making the initial predictions. When, however, the visual clues were provided, the gender stereotype may have been perceived as sufficiently more useful to warrant discounting the survey responses.

E. Accuracy of Predictions

Predictors change their predictions when additional information is provided, but change does not necessarily imply more accuracy. In this section, I consider: (1) the change in prediction accuracy from P1 to P2 by treatment; (2) the prediction accuracy of men versus women holding the gender of the target and treatment constant; and (3) whether or not prediction accuracy differs between treatments holding the predictor gender/target gender pairings constant.

P1 versus P2. Table 10 reports mean accuracy rates of P1 and P2 by treatment and predictor gender/target gender pairings. (29) Consider first the V/S treatment results. The accuracy of predictions for male targets was virtually unchanged from P1 to P2, while the accuracy of predictions for female targets improved from P1 to P2, but the change in accuracy rates was insignificant (both means tests and Wilcoxon Matched-Pairs Signed Ranks tests p-values [greater than or equal to] .10) in every case, except for males predicting for females (p-values [less than or equal to] .10). In the S/V treatment accuracy rates declined for each predictor gender/target gender pairing except Female/Female, but in every case the change was insignificant (means tests and Wilcoxon Matched-Pairs Signed Ranks tests p-values [greater than or equal to] .10). These findings give the sixth result.

Result 6: Additional information, regardless of the type, does not significantly improve predictors' accuracy rates. H6 is not supported.

Men versus Women. Table 11 reports p-values for tests comparing the prediction accuracy of men and women, controlling for treatment and gender of the target. (30) In the V/S treatment, men are significantly more accurate than women when predicting for male targets (both means test and Wilcoxon two-sample test p-values [less than or equal to] .01 for both P1 and P2). There is no significant difference in accuracy when predicting for female targets. (31) In the S/V treatment, the only significant difference in accuracy is in P2s for female targets. Women are significantly more accurate than men in this case. These findings give Observations 1 and 2:

Observation 1: When predicting based on visual clues only, men are significantly more accurate than women when predicting for men. The addition of individuating, judgment-relevant and irrelevant information does not alter the relative accuracy of men and women predictors.

Observation 2: When predicting based on individuating, judgment-relevant and irrelevant information only, neither men nor women are significantly more accurate. The addition of visual clues significantly improves only the accuracy of women relative to that of men when predicting for female targets.

At first glance, these observations seem to be contradictory. Instead, they suggest a gender difference in the ability to interpret the information provided and that the order in which information is provided matters. The reported results suggest that male predictors are better able than are female predictors to interpret visual clues (in isolation) for male targets; men are better able to "read" other men than they are women. The addition of the survey responses neither significantly improves nor diminishes predictors' relative accuracy for either target gender. Regardless of gender, predictors are not easily able to interpret the survey response clues. However, female predictors are better at "reading" the combination of the survey and then the visual clues for other women. For male predictors, the addition of visual clues does not improve their ability to "read" their targets, regardless of gender. This suggests that male predictors' "readings" of male targets are diminished by the survey information.

V/S versus S/V. Finally, Table 12 reports the results of comparing accuracy rates in the V/S treatment and accuracy rates in the S/V treatment controlling for the predictor gender/target gender pairings. For P1, there is no significant difference between treatments in the accuracy of predictions for female targets, regardless of the predictors' gender (p-values [greater than or equal to] .10). (32) Where a significant difference is observed it is in the accuracy of predictions for male targets. Men predicting for other men do significantly better using visual clues than they do using individuating information (both means test and Wilcoxon two-sample test p-values [less than or equal to] .05). Women predicting for men do significantly better using individuating information than they do using visual clues (both means test and Wilcoxon Two-Sample test p-values [less than or equal to] .05).

For P2, again, a significant difference between treatments is observed only in the accuracy of predictions for male targets. Men still predict better for other men in the V/S treatment than they do in the S/V treatment. As noted in the second subsection IV.E, in the V/S treatment, men seem to ignore the new individuating information provided prior to making their P2, and as such their accuracy rate is little changed. Likewise, women still predict better for men in the S/V treatment than they do in the V/S treatment. Like men, women in the V/S treatment seem to ignore the new individuating information provided prior to making their P2. In the S/V treatment, the new visual clue information provided reduces women's accuracy, but not enough to eliminate the significant difference. These findings give Observations 3 and 4.

Observation 3: Men predicting for men are significantly more accurate when predicting based on visual clues versus survey responses; women predicting for men are significantly more accurate when predicting based on individuating information versus visual clues. Additional information, of either type, does not significantly alter the accuracy of predictors.

Observation 4: Neither visual clues nor individuating information significantly improves the accuracy of either gender when predicting for women.

V. CONCLUSION

This study reports: (1) additional evidence of the existence of gender stereotyping, and more importantly, (2) the persistence of such stereotyping even when other, both judgment-relevant and judgment-irrelevant, individuating information is provided. Subjects participate in one of two treatments of a four-task experiment. In Task 1, subjects complete a survey designed to elicit attitudes toward different types of risk. In Task 2, subjects play a gamble exercise, and in Task 3, subjects attempt to predict the Task 2 gamble choices made by their fellow subjects based on either only visual clues or only responses to two survey statements (the individuating information). In Task 4, subjects are provided additional information about their targets, either the survey responses or the visual clues, and are permitted to revise their initial predictions.

As expected, gamble choices do not differ significantly by treatment. Treatment procedures are identical through the gamble choice task. I also do not find that gamble choices differ significantly by gender; women are only marginally more risk averse than men. This result is consistent with the findings of Daruvala (2007) and Grossman and Lugovskyy (2011).

Results from Task 3 of the experiment suggest that, in general, predictors try to use whatever available information they have about the targets to inform their initial predictions. In particular, I find strong evidence consistent with the conclusion that when given only visual clues, predictors, when making their initial predictions, default to the gender stereotype that women are more risk averse than men. Predictions for male targets are significantly higher than predictions for female targets (mean predictions differ by approximately one gamble choice). I also find strong evidence consistent with the conclusion that when given only individuating information (i.e., the survey responses), predictors use this information, both judgment relevant and judgment irrelevant, to make their initial predictions. While predictions for male targets are still higher than predictions for female targets (but the mean predictions differ by less than one-half gamble choice), this reflects the fact that males indicated significantly less risk aversion than did women in their responses to the financial risk survey statement (there was little gender difference in the responses to the social risk statement).

Results from Task 4 of the experiment provide both additional evidence of gender stereotyping and evidence that gender stereotyping dominates individuating, judgment-relevant and judgment-irrelevant, information. Predictors, when provided visual clues first and individuating information second, do not systematically revise their predictions away from their initial gender stereotyped predictions. I find no evidence that suggests that the individuating information is used to supplant gender stereotyping. Predictions for all predictor gender/target gender pairings are virtually unchanged. The one exception is that female predictors significantly lower their predictions for male targets.

Predictors, when provided the individuating information first and the visual clues second, do systematically and significantly revise their predictions to reflect the gender stereotype. The evidence reported suggests that gender stereotyping supplants, or at least supplements, the individuating information. Predictions for male targets are revised upward by both male and female predictors; predictions for female targets are revised downward by both male and female predictors.

The findings reported offer a caution to those who might rely upon their ability to "read" an individual or to acquire useful information from survey instruments. While the information garnered from such cursory interactions and surveys may be used to inform their assessment of the risk attitudes of others, my results suggest that their assessments are likely to be colored, and possibly dominated, by the gender stereotype. The gender stereotype that women are more risk averse than men appears to be a salient anchor, regarded by predictors as relevant, that colors assessments of risk attitude. Furthermore, my findings suggest that as an anchor, the gender stereotype is sufficiently strong such that additional, individual specific, information is largely dismissed.

The implications of this strongly and widely held gender stereotype are potentially life altering. An individual's risk preferences can influence educational, financial, legal, and medical decisions, to name a few, that are made over the course of a life. In all of these, the advice of experts is often sought and that advice may be skewed by the expert's potentially biased assessment of the individual's risk preference. Students may be steered toward career paths (i.e., a low-risk stable career in nursing as opposed to a higher-risk career as a doctor), into financial instruments (i.e., investment in blue chip stocks as opposed to growth stocks) or legal settlements (i.e., a sure thing settlement as opposed to taking the case to court with the risk of winning a lot or nothing), and medical procedures (i.e., a safe, noninvasive procedure versus a high-risk, high-return invasive procedure) that encompass too much or too little risk relative to what the individual might prefer.

Gender stereotyping does not stop with risk. Stereotypes regarding differences between men and women may exist along many dimensions. Lavy (2008) reports that boys are assumed to excel at math and science while girls excel in other subjects. (33) Males are assumed to be more competitive than females. (34) Women are assumed to be more cooperative than men. (35) Croson and Gneezy (2009) argue that women are also assumed to be more generous. Teachers, advisors, or bosses acting on these stereotypes, as opposed to actual individual preferences or behavior, may channel individuals into nonoptimal career paths, for both the individual and the employee, and other lifestyle choices.

Finally, the results reported also suggest that men may be better at interpreting the visual clues provided by other men, while women may be better at interpreting for males any individuating information that might be provided by assessment tools. The results suggest that female clients should be wary of any advice coming from either male or female advisors, regardless of the basis for that advice. Women, it would appear, are simply harder to "read" than men. This is an area that merits greater exploration.

ABBREVIATIONS

DSRAS: Domain Specific, Risk-Attitude Scale

EG: Eckel and Grossman

S/V: SURVEY/VISUAL

SCSU: Saint Cloud State University

V/S: VISUAL/SURVEY

doi: 10.1111/j.1465-7295.2012.00479.x

REFERENCES

Andersen, S., G. W. Harrison, M. I. Lau, and E. Rutstrom. "Eliciting Risk and Time Preferences." Econometrica, 76, 2008, 583-618.

Ball, S., C. C. Eckel, and M. Heracleous. "Risk Aversion and Physical Prowess: Prediction, Choice, and Bias." Journal of Risk and Uncertainty, 41, 2010, 167-93.

Barasinska, N. "Does Gender Affect Investors' Appetite for Risk?" DIW Berlin Discussion Paper 1125, 2011.

Barsky, R. B., T. J. Juster, M. S. Kimball, and M. D. Shapiro. "Preference Parameters and Behavioral Heterogeneity: An Experimental Approach in the Health and Retirement Study." Quarterly Journal of Economics, 112. 1997, 537-79.

Barber, B. M., and T. Odean. "Boys Will Be Boys: Gender, Overconfidence, and Common Stock Investment." Quarterly Journal of Economics, 116, 2001, 261-92.

Beckett, N. E., and B. Park. "Use of Category versus Individuating Information: Making Base Rates Salient." Personality and Social Psychology Bulletin, 21, 1995, 21-31.

Bodenhausen, G. V., C.N. Macrae, and J. W. Sherman. "On the Dialectics of Discrimination: Dual Processes in Social Stereotyping," in Dual-Process Theories in Social Psychology, edited by S. Chaiken and Y. Trope. New York: Guilford Press, 1999, 271-90.

Cardenas, J. C. and J. Carpenter. "Risk Attitudes and Well-Being in Latin-America." IZA Discussion Paper No. 5279, 2010.

Carpenter, J., J. R. Garcia, and J. K. Lum. "Can Dopamine Receptor Genes Explain Economic Preferences and Outcomes?" Middlebury College Working Paper, 2009.

Charness, G. B., and U. Gneezy. "Strong Evidence of Gender Differences in Investment." UC Santa Barbara Department of Economics Working Paper, 2007.

Croson, R., and U. Gneezy. "Gender Differences in Preferences." Journal of Economic Literature, 47, 2009, 448-74.

Daruvala, D. "Gender, Risk, and Stereotypes." Journal of Risk and Uncertainty 35, 2007, 265-83.

Dave, C., C. C. Eckel, C. Johnson, and C. Rojas. "Eliciting Risk Preferences: When Is Simple Better?" Journal of Risk and Uncertainty, 41, 2010, 219-43.

Eckel, C. C., and P. J. Grossman. "Sex Differences and Statistical Stereotyping in Attitudes toward Financial Risk." Evolution and Human Behavior, 23, 2002, 281-96.

--. "Forecasting Risk Attitudes: An Experimental Study of Actual and Forecast Risk Attitudes of Women and Men." Journal of Economic Behavior and Organization, 68, 2008, 1-17.

Eckel, C. C., M.A. El-Gamal, and R. K. Wilson. "Risk Loving after the Storm: A Bayesian-Network Study of Hurricane Katrina Evacuees." Journal of Economic Behavior and Organization, 69, 2009, 110-24.

Gill, M. J. "When Information Does Not Deter Stereotyping: Prescriptive Stereotyping Can Foster Bias under Conditions that Deter Descriptive Stereotyping." Journal of Experimental Social Psychology, 40, 2004, 619-32.

Gneezy, U., and A. Rustichini. "Gender and Competition at a Young Age." American Economic Review Papers and Proceedings, XCIV, 2004, 377-81.

Gneezy, U., M. Niederle, and A. Rustichini. "Performance in Competitive Environments: Gender Differences." Quarterly Journal of Economics 118, 2003, 1049-74.

Grossman, P. J., and O. Lugovskyy. "An Experimental Test of the Persistence of Gender-Based Stereotypes." Economic Inquiry, 49, 2011, 598-611.

Hsee, C. K., and E. U. Weber. "A Fundamental Prediction Error: Self-Others Discrepancies in Risk Preference." Journal of Experimental Psychology, 126, 1997, 45-53.

--. "Cross-Cultural Differences in Risk Preferences by Cross-Cultural Similarities in Attitudes toward Risk." Management Science, 44, 1998, 107-21.

Jianakoplos, N. A., and A. Bernasek. "Are Women More Risk Averse?" Economic Inquiry, 36, 1998, 620-30.

Kunda, Z., L. Sinclair, and D. Griffin. "Equal Ratings but Separate Meanings: Stereotypes and the Construal of Traits." Journal of Personality and Social Psychology, 72, 1997, 720-34.

Kunda, Z., and P. Thagard. "Forming Impressions from Stereotypes, Traits, and Behaviors: A Parallel-Constraint-Satisfaction Theory." Psychological Review, 103, 1996, 284-308.

Lavy, V. "Do Gender Stereotypes Reduce Girls' or Boys' Human Capital Outcomes? Evidence from a Natural Experiment." Journal of Public Economics, 92, 2008, 2083-105.

Levin, I. P., M. A. Snyder, and D. P. Chapman. "The Interaction of Experiential and Situational Factors and Gender in a Simulated Risky Decision-Making Task." Journal of Psychology, 122, 1988, 173-81.

List, J. A., R. P. Berrens, A. K. Bohara, and J. Kerkvliet. "Examining the Role of Social Isolation on Stated Preferences." American Economic Review, 94, 2004, 741-52.

Mussweiler, T. "The Durability of Anchoring Effects." European Journal of Social Psychology, 31, 2001, 431-42.

Niederle, M., and L. Vesterlund. "Do Women Shy Away from Competition? Do Men Compete Too Much?" Quarterly Journal of Economics, 122, 2007, 1067-101.

Nisbett, R. E., H. Zukier, and R. E. Lemley. "The Dilution of Effect: Nondiagnostic Information Weakens the Implications of Diagnostic Information." Cognitive Psychology, 13, 1981, 248-77.

Powell, M., and D. Ansic. "Gender Differences in Risk Behaviour in Financial Decision-Making: An Experimental Analysis." Journal of Economic Psychology, 18, 1997, 605-28.

Sherman, J. W. "Development and Mental Representation of Stereotypes." Journal of Personality and Social Psychology, 70, 1996, 1126-41.

Siegrist, M., G. Cvetkovich, and H. Gutscher. "Risk Preference Predictions and Gender Stereotypes." Organizational Behavior and Human Decision Processes, 87, 2002, 91-102.

Thaler, R., and E. Johnson. "Gambling with the House Money and Trying to Break Even: The Effects of Prior Outcomes on Risky Choice." Management Science, 36, 1990, 643-60.

Thorsteinson, T. J., J. Breier, A. Atwell, C. Hamilton, and M. Privette. "Anchoring Effects on Performance Judgments." Organizational Behavior and Human Decision Processes, 107, 2008, 29-40.

Trope, Y., and E. P. Thompson. "Looking for Truth in All the Wrong Places: Asymmetric Search of Individuating Information about Stereotyped Group Members." Journal of Personality and Social Psychology, 73, 1997, 229-41.

Tversky, A., and D. Kahneman. "Judgment under Uncertainty: Heuristics and Biases." Science 185, 1974, 1124-31.

Weber, E. U., A. Blais, and N. E. Betz. "A Domain-Specific Risk-Attitude Scale: Measuring Risk Perceptions and Risk Behaviors." Journal of Behavioral Decision Making, 15, 2002, 1-28.

Wilson, T. D., C. E. Houston, K. M. Etling, and N. Brekke. "A New Look at Anchoring Effects: Basic Anchoring and Its Antecedents." Journal of Experimental Psychology, 125, 1996, 387-402.

Yordanova, D.I., and M.I. Alexandrova-Boshnakova. "Gender Effects on Risk-Taking of Entrepreneurs: Evidence from Bulgaria." International Journal of Entrepreneurial Behaviour & Research, 17, 2011, 272-95.

(1.) See Charness and Gneezy (2007), Eckel, El-Gamal, and Wilson (2009), and Yordanova and Alexandrova-Boshnakova (2011) for evidence of a gender difference and Andersen et al. (2008) and Barasinska (2011) for evidence of no gender difference. Eckel and Grossman (2008) and Croson and Gneezy (2009) review the earlier literature.

(2.) See Bodenhausen. Macrae, and Sherman (1999) and Kunda and Thagrad (1996) for reviews of the evidence regarding dual processes in stereotyping.

(3.) EG's focus, as is the focus of this paper, was on the gender of the target. Predictors may have drawn clues from visual factors other than the gender of the target (i.e., manner of dress, hairstyle, or posture for example). They, and I, did not collect this type of data.

(4.) Gamble 6 has the same payoff as Gamble 5 but with higher associated risk. This gamble is included to distinguish the risk-lovers.

(5.) Instructions and forms can be found at: http://www. buseco.monash.edu.au/eco/staff/grossman-philip.html.

(6.) Subjects are paid for completing the task rather than just being given the money. This is intended to engender a sense of ownership and help minimize any "house money" effect (see Thaler and Johnson 1990).

(7.) List et al. (2004) note the degree of social isolation can bias stated preferences. While subjects were not informed that, later in the session, their responses to two of the survey questions would be revealed, nevertheless, it is possible that responses are biased because subjects surmised that their responses to the survey questions may be revealed at some later time in the session or possibly they did not believe that their responses were sufficiently anonymous (i.e., that the experimenter would be able to connect their names to their responses).

(8.) The instrument, in its original and adapted form, has been used by Dave et al. (2010), Ball, Eckel, and Heracleous (2010), Cardenas and Carpenter (2010), and Carpenter, Garcia, and Lure (2009). Gambles 1-5 are identical to EG's Gambles 1-5.

(9.) As measured by standard deviation from the expected payoff. Having payoff and risk increase linearly eliminates the possibility that a risk-averse subject would be indifferent between two gambles (possible if payoff increased faster than risk). More important is the instrument's simplicity; expected payoff and risk increase linearly and the 50/50 gambles make comparisons easy and require only minimal math skills of the subjects. Dave et al. conclude that the simpler Eckel and Grossman instrument "... may provide better accessibility, and so more accurate measures of risk aversion, for subjects with low levels of analytic proficiency (p. 30)."

(10.) If subjects are sufficiently loss averse, gamble choices may be biased down. Eckel and Grossman (2008) test for this bias by eliminating their $6 payment for completing the survey and by increasing all gamble payoffs by $6 so that all gamble payoffs are non-negative. They compared gamble choices with and without the possibility of negative payoffs and found no significant difference.

(11.) Alternatives included: "Investing 10% of your annual income in blue-chip stock" and "Investing 10% of your annual income in government bonds (treasury bills)." From past experience, student subjects were unfamiliar with the term "blue-chip" stock (a low-risk investment) and were not aware that treasury bills are relatively risk-free investments. Neither alternative had the desired air of a gamble.

(12.) Jianakoplos and Bernasek (1998) find that women choose less risky investment portfolios than men. Barsky et al. (1997) find women have a lower propensity toward financial risk than men. In an experiment designed to mimic investment behavior, Powell and Ansic (1997) find that women choose less risky alternatives. Levin et al. (1988) report that men are more likely than women to engage in risky behavior, such as gambling. Barber and Odean (2001) show that, consistent with greater male overconfidence, men trade more aggressively than women in financial markets.

(13.) Sessions were conducted in Saint Cloud State University (SCSU) classes not taught by the experimenter. All sessions were conducted in introductory economics classes using the same classroom but for different classes. Students were informed that in lieu of class the experiment would be conducted and that participation was voluntary. If they did not wish to participate they were free to go. Approximately five students per class elected not to participate. SCSU has an enrollment of approximately 17,000 with a large contingent of nontraditional students and many commuters. For these reasons and based on the experimenter's own observations teaching introductory courses, the experimenter believed it was reasonable to assume that, with few exceptions, subjects did not know each other well.

(14.) Predictors may record the same prediction they made in Task 3.

(15.) Nine subjects were excluded for failing to indicate gender on the post-experiment survey or for not completing one or both of the prediction forms.

(16.) One concern is that since only five subjects per session were paid, subjects may have treated their decisions as only hypothetical. Comparing results from this experiment with those of Grossman and Lugovskyy (2011), in which all subjects were paid for their decisions, suggests this is not so. [chi square] contingency table tests comparing all decisions, decisions by men and decisions by females across the two studies cannot reject the null hypothesis that the gamble decisions are similarly distributed (p-values = .36, .64, and .49, respectively).

(17.) All p-values are lbr two-tailed tests.

(18.) This appears to be a result of the extra gamble choice. EG, who reported a significant difference, had one fewer gamble choice than the instrument used in this paper and Grossman and Lugovskyy (2011). The additional gamble, while redistributing the choices among the riskier gambles (4 and 5 in EG vs. 4-6 in this paper), did not increase the percentages of men choosing one of the riskier gambles (from 60.3% to 64.7%). It did increase the percentage of women choosing one of the riskier gambles (from 30.8% to 66.7%).

(19.) The definition of [P1.sup.*] recognizes that predictions by any predictor i are not independent (i.e., predictions by predictor i for target j = 1 are likely to be dependent on predictions made by i for all other targets j [not equal to] 1). In Table 4, V/S treatment, 4.53 is the mean of the mean prediction made by male predictors for male targets. For each male predictor I calculated the mean prediction made by that predictor for all male targets. I then calculated the mean of those mean predictions.

(20.) How well these factors correlate with the targets' actual gamble choices is reported in Table 5. The Spearman correlation coefficient for targets' Task 2 gamble choices and targets' gender (1 = male) is 0.237; targets' responses to the financial risk statement (Investing 10% of your annual income in a very speculative stock, mean response = 2.55; SD = 1.18) is 0.138; and targets' responses to the social risk statement (Arguing with a friend, who has a very different opinion on an issue, mean response = 3.88; SD = 1.02) is 0.217. All three of the target characteristics except, surprisingly, the targets' responses to the financial risk statement are positively and significantly correlated with targets' gamble choice (p-values = .04 in each case, p-value =.18 for the financial risk statement).

(21.) There was little gender difference in the responses to the social risk statement (p-values <.82 for both a means test and a Wilcoxon two-sample test).

(22.) In this case, and all others, the most conservative result is reported.

(23.) Standard errors are clustered on the individual.

(24.) As with [P1.sup.*], the definition of [P2.sup.*] recognizes that predictions by any predictor i are not independent (i.e., predictions by predictor i for target j = 1 are likely to be dependent on predictions made by i for all other targets j [not equal to] 1). It is calculated in the same way as [P1.sup.*].

(25.) The means test p-value for male predictors and female targets is insignificant (p-value [less than or equal to] .17), but the Wilcoxon Matched-Pairs Signed-Rank test p-value is significant (p-value [less than or equal to] .05).

(26.) Standard errors are clustered on the individual.

(27.) In the V/S treatment, I interacted Financial Risk Statement and Social Risk Statement with gender of predictor/gender of target dummies. While the null hypothesis that the coefficients were equal across these variables was rejected, the individual coefficients were insignificantly different from zero. In the S/V treatment, I replaced Male Target with gender of predictor/gender of target, but this did not significantly improve the explanatory power of the model. Results are available upon request.

(28.) See Tversky and Kahneman (1974).

(29.) The accuracy rates reported are the percentage of predictions that were correct averaged across all predictors in the specific grouping. For example, for the V/S treatment in Table 9, the 24 male predictors predicting for male targets made correct predictions an average of 30.4% of the time.

(30.) Previous work offers no basis on which to hypothesize regarding the expected findings when comparing the accuracy of predictions by men and women and across treatments. For this reason, subsequent findings are labeled observations rather than results.

(31.) The means test p-value is insignificant (p-value [less than or equal to] .11), but the Wilcoxon two-sample test p-value is significant (p-value [less than or equal to] .05).

(32.) For male predictors and female targets, the means test p-value is insignificant (p-value [less than or equal to] .35), but the Wilcoxon Two-Sample test p-value is significant (p-value [less than or equal to] .10).

(33.) Lavy (2008) reports evidence of discrimination against boys in the grading of high school matriculation exams. Results suggest that the bias is the result of teachers" behavior.

(34.) See Gneezy and Rustichini (2004), Gneezy et al. (2003), and Niederle and Vesterlund (2007).

(35.) See Eckel and Grossman (2002) for a discussion of the literature.

PHILIP J. GROSSMAN *

* This work was supported by a grant from the National Science Foundation (SBR-0136684). I would like to thank the instructors who provided me access to their classes and Mana Komai and Lynn MacDonald for their helpful comments. Thanks are also due to the editor of this journal and to two anonymous referees for their helpful comments and suggestions.

Grossman: Professor, Department of Economics, Monash University, Clayton, Victoria, 3800, Australia. Phone 61-3-990-20052, Fax 61-3-990-55476, E-mail philip.grossman@monash.edu
TABLE 1
Gamble Choices, Expected Payoffs, and Risk

 Probability, Expected
Choice Event % Payoff Payoff Risk (a)

1 A 50 $10 $10 0.0
 B 50 $10
2 A 50 $18 $12 6.0
 B 50 $6
3 A 50 $26 $14 12.0
 B 50 $2
4 A 50 $34 $16 18.0
 B 50 -$2
5 A 50 $42 $18 24.0
 B 50 -$6
6 A 50 $44 $18 26.0
 B 50 -$8

(a) Defined as the standard deviation.

TABLE 2
Subject Characteristics

 SURVEY/ VISUAL/ Test Statistic
 VISUAL SURVEY p value

Male 51.9% 63.2% .29 (a)
Age (mean) 21.7 20.8 .88 (b)
Minority 32.7% 18.4% .13 (a)
Econ/Business 46.2% 31.2% .16 (a)
Regular religious 36.5% 26.3% .31 (a)
 attendance
Class
Freshman 12 12 .85 (c)
Sophomore 22 12
Junior 11 8
Senior 6 5
Graduate 1 1
GPA
<2.00 1 1
2.00-2.49 10 5
2.50-2.99 13 10 .94 (c)
3.00-3.49 15 12
3.50+ 10 9

(a) Binomial proportions test.

(b) Means test.

(c) [chi square] contingency table test.

TABLE 3
Gamble Choice by Treatment and by Gender

 Gamble Choice

 1 2 3 4

SURVEYNISUAL 0 2 15 16
 (0.0%) (3.8%) (28.8%) (30.8%)

VISUAL/SURVEY 2 4 8 9
 (5.3%) (10.5%) (21.1%) (23.7%)

Male 2 4 12 10
 (3.9%) (7.9%) (23.5%) (19.6%)

Female 0 2 11 15
 (0.0%n) (5.1%) (28.2%) (38.5%)

All 2 6 23 25
 (2.2%) (6.7%) (25.6%) (27.8%)

 Gamble Choice Wilcoxon
 Mean (SD) Two-Sample
 5 6 N Test p value

SURVEYNISUAL 9 10 4.19 .96
 (17.3%) (19.2%) (1.17)
 52
VISUAL/SURVEY 3 12 4.13
 (7.9%) (31.6%) (1.58)
 38
Male 7 16 4.25 .45
 (13.7%) (31.4%) (1.51)
 51
Female 5 6 4.05
 (12.8%) (15.4%) (1.12)
 39
All 12 22 4.16
 (13.3%) (24.4%) (1.35)
 90

TABLE 4
Prediction 1--Average Predicted Gamble Choices for Subjects by
Treatment, Gender of Target, and Gender of Predictor

 Test Statistics for
 Differences in
 Subjects' Mean
 Predictions for Men
 Predicted Gamble for and Women

 Wilcoxon
 Matched-Pairs
 Signed-Ranks
Predictors N Males Females All t-test Test
VISUAUSURVEY
 Male 24 4.53 3.56 4.18 5.40 p < .001
 (1.37) (1.60) (1.40) p < .001
 Female 14 3.83 2.94 3.55 5.34 p < .001
 (0.65) (0.77) (0.63) p < .001
 All 38 4.27 3.33 3.94 7.36 p < .001
 (1.20) (1.38) (1.21) p < .001
SURVEY/VISUAL
 Male 27 3.52 3.08 3.36 2.20 p = .004
 (0.67) (0.81) (0.65) p = .03
 Female 25 3.73 3.36 3.59 1.82 p = .01
 (0.62) (0.81) (0.63) p = .08
 All 52 3.62 3.21 3.47 2.83 p < .001
 (0.65) (0.82) (0.65) p = .006

TABLE 5
Correlation Coefficients for Gamble Choice
and Gamble Predictors

 Correlation Coefficient
 Z -statistic
 (p value)

Target gender 0.237
 2.33
 (0.020)
Social risk statement 0.217
 2.14
 (0.033)
Financial risk statement 0.138
 1.36
 (0.175)

TABLE 6
Determinants of P1--Ordered Probit Results
with Clustered Standard Errors (Dependent
Variable = P1)

 Coefficient (t-stat)

 VISUAL/ SURVEY/
Variable SURVEY VISUAL
Predictor Gamble 0.351 (a) (3.49) 0.186 (a) (2.77)
 Choice
Male Predictor 0.282 (1.34) -0.139 (1.06)
Male Target 0.623 (a) (6.80)
Social Risk Statement 0.222 (a) (5.61)
Financial Risk 0.514 (a) (9.42)
Statement
Constant -0.131 (0.45) -1.222 (a) (4.30)

Log LF -1,336.0 -2,028.1
N 859 1,302
Individuals 38 52

Notes: Standard errors are clustered on the individual
predictor. LF, likelihood function.

(a) Significant at the 5% level or better, two-tailed test.

TABLE 7
Prediction 2--Average Predicted Gamble Choices for Subjects by
Treatment, Gender of Target, and Gender of Predictor

 Test Statistics for
 Differences in
 Subjects' Mean
 Predictions for Men and
 Predicted Gamble for Women

 Wilcoxon
 Matched-Pairs
 Signed-Ranks
Predictors N Males Females All t-test Test
VISUAL/SURVEY
 Male 24 4.55 3.72 4.25 4.85 p < .001
 (1.42) (1.59) (1.43) p < .001
 Female 14 3.55 2.97 3.36 3.03 p = .007
 (0.70) (0.44) (0.56) p =.O1
 All 38 4.19 3.44 3.92 5.84 p < .001
 (1.28) (1.33) (1.25) p < .001
SURVEY/VISUAL
 Male 27 3.67 2.97 3.39 4.35 p < .001
 (0.77) (0.86) (0.70) 17 < .001
 Female 25 4.09 3.21 3.68 7.00 p < .001
 (0.81) (0.78) (0.72) 1) < .001
 All 52 3.87 3.09 3.53 7.62 p < .001
 (0.81) (0.83) (0.72) p < .001

TABLE 8
Tests of Differences Between P2 and P1

 Sign
Treatment Predictor Target N (P2-P1)

VISUAUSURVEY Male Male 24 +
 Male Female 24 +
 Female Male 14 -
 Female Female 14 +
SURVEYNISUAL Male Male 27 +
 Male Female 27 -
 Female Male 25 +
 Female Female 25 -

 p value [less than or equal to]

 Wilcoxon
 Paired Means Matched-Pairs
Treatment Predictor Target Test Signed Ranks Test

VISUAUSURVEY Male Male .78 .57
 Male Female .14 .20
 Female Male .03 .04
 Female Female .81 .70
SURVEYNISUAL Male Male .06 .01
 Male Female .17 .05
 Female Male .01 .01
 Female Female .02 .03

TABLE 9
Determinants of P2--Ordered Probit Results
with Clustered Standard Errors (Dependent
Variable = P2)

 Coefficient (t-stat)

 VISUAL/ SURVEY/
Variable SURVEY VISUAL

Prediction 1 0.606 (a) (5.54) 0.591 (a) (11.28)
Male Target 0.488 (a) (6.71)
Social Risk Statement 0.017 (0.58)
Financial Risk -0.0003 (0.01)
 Statement
Constant 0.018 (0.06) 0.017 (0.12)
Log LF -1,171.5 -1,895.0
N 859 1,302
Individuals 38 52

Notes: Standard errors are clustered on the individual
predictor. LF, likelihood function.

(a) Significant at the 5% level or better, two-tailed test.

TABLE 10
Accuracy of Predictions: P1 versus P2

 Accuracy Rate (SD)

Treatment Predictor Target N P1 P2

VISUAL/SURVEY Male Male 24 0.30 (0.16 0.30 (0.16
 Male Female 24 0.15 (0.18) 0.19 (0.18)
 Female Male 14 0.16 (0.11) 0.16 (0.09)
 Female Female 14 0.23 (0.13) 0.25 (0.15)
SURVEYNISUAL Male Male 27 0.21 (0.13) 0.19 (0.10)
 Male Female 27 0.19 (0.13) 0.19 (0.11)
 Female Male 25 0.26 (0.14) 0.21 (0.10)
 Female Female 25 0.24 (0.11) 0.28 (0.15)

 p Value [less than or equal to]

 Wilcoxon
 Paired Means Matched-Pairs
Treatment Predictor Target Test Signed-Ranks Test

VISUAL/SURVEY Male Male 0.46 0.64
 Male Female 0.08 0.10
 Female Male 0.75 1.00
 Female Female 0.61 0.81
SURVEYNISUAL Male Male 0.21 0.30
 Male Female 0.92 0.98
 Female Male 0.14 0.16
 Female Female 0.16 0.15

TABLE 11
Accuracy of Predictions: Men Versus Women

 Predictors' P Value [less
 Accuracy Rate than or equal to]

 Wilcoxon
 Means Two-Sample
Treatment Target Prediction Males Females Test Test

VISUAL/SURVEY Males 1 0.30 0.16 0.01 0.01
 2 0.30 0.16 0.01 0.01
 Females 1 0.15 0.23 0.11 0.07
 2 0.19 0.25 0.23 0.17

SURVEY/VISUAL Male 1 0.21 0.26 0.17 0.25
 2 0.19 0.21 0.36 0.39
 Female 1 0.19 0.24 0.12 0.06
 2 0.19 0.28 0.01 0.01

TABLE 12
Accuracy of Predictions across Treatments

 Predictors' p Value [less
 Accuracy Rate than or equal to]

 Wilcoxon
 VISUAL/ SURVEY/ Means Two-Sample
Prediction Predictor Target SURVEY VISUAL Test Test

1 Male Male 0.30 0.21 0.03 0.05
 Male Female 0.15 0.19 0.35 0.10
 Female Male 0.16 0.26 0.02 0.03
 Female Female 0.23 0.24 0.79 0.94
2 Male Male 0.30 0.19 0.01 0.03
 Male Female 0.19 0.19 0.96 0.53
 Female Male 0.16 0.21 0.08 0.06
 Female Female 0.25 0.28 0.54 0.90
COPYRIGHT 2013 Western Economic Association International
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2013 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Grossman, Philip J.
Publication:Economic Inquiry
Article Type:Report
Geographic Code:1USA
Date:Jan 1, 2013
Words:11438
Previous Article:Dynastic political privilege and electoral accountability: the case of U.S. governors, 1950-2005.
Next Article:Parental substance use and foster care: evidence from two methamphetamine supply shocks.
Topics:

Terms of use | Copyright © 2017 Farlex, Inc. | Feedback | For webmasters