Risk Aversion and Incentive Effects: New Datawithout Order EffectsBy CHARLES A. HOLT AND SUSAN K. LAURY*Holt and Laury (2002) used a menu ofordered lottery choices to make inferencesabout risk aversion under various paymentconditions. The main results of that paperwere: (a) subjects are risk averse, even forrelatively small payments of less than $5; (b)risk aversion increases sharply with large in-creases in the scale of cash payoffs; and (c)there is no significant effect from increasingthe scale of hypothetical payment. With a fewexceptions noted in the paper, all treatmentsbegan with a low-payment choice, followedby a choice with hypothetical payments thathad been scaled up (by 20⫻,50⫻,or90⫻),followed by a real-cash decision with thesame high payment scale (20⫻,50⫻,or90⫻), followed by a single, final, low (1⫻)real payment choice. Those in the 90⫻ treat-ment could earn amounts ranging from $9.00to $346.50 in this task. As Glenn W. Harrisonet al. (2004) correctly note, this design con-founds order and treatment effects since thehigh real payment choice was always com-pleted after the low real and high hypotheticalpayment tasks.In a new experiment reported below, we firstseek to replicate Harrison et al.’s finding thatthe order effect (participating in a low-paymentchoice before making a high-payment choice)magnifies the scale effect. In a second treat-ment, each subject completes the menu of lot-tery choices under just one payment condition(1⫻ or 20⫻, real or hypothetical), thereby elim-inating any order effects.I. New DataThe new experiment was conducted in 2004using 216 subjects recruited from undergradu-ate economics classes at the University of Vir-ginia.1As in our previous experiment, eachsession began with a lottery choice “trainer” anda second unrelated experiment. Results are pre-sented in Table 1. For comparison, we includedata from the Holt and Laury (2002) experi-ment, as shown in the top two rows of Table1, and from Harrison et al.’s (2005) experimentin rows three and four.In the first treatment of our new experiment,48 subjects completed a real low-paymentchoice, followed by a real high-payment choice,in which all choices were scaled up by a factorof 20.2Results are presented in Table 1, row 5.The average number of safe choices for the low(1⫻) real treatment is shown in row five as 6.1.When real cash payments are scaled up by afactor of 20, the average number of safe choicesmade by these subjects increased to 7.1. As canbe seen, subjects from the new experiment aresomewhat more risk averse than those used inthe earlier studies; however, the scale effect(from 1⫻ to 20⫻) with cash payments is essen-tially the same as that of our previous experi-ment. In both cases, the average number of safechoices increased by approximately one safe* Holt: Department of Economics, University of Vir-ginia, Charlottesville, VA 22904-4182 (e-mail: [email protected]); Laury: Department of Economics, GeorgiaState University, 35 Broad Street, 1441 Rcb Bldg, Atlanta,GA 30303 (e-mail: [email protected]). This research wassupported in part by the National Science Foundation, SES0094800, and by the University of Virginia Bankard Fund.We thank Ragan Petrie for helpful suggestions and A. J.Bostian, Erin Golub, Kurt Mitman, and Angela Moore forresearch assistance.1Unlike the experiment reported in Holt and Laury(2002), decisions were recorded using a computer interface.The die-throw, however, was still done by the experimenterby hand. Also, the left/right order of the safe and riskyoptions was alternated in successive 12-person sessions.The order of presentation did not matter, and so we pool thedata from both presentation orders.2As in our original experiments, as a rough control forwealth effects, a person had to agree to give up the paymentfrom the first (1⫻) choice in order to participate in thehigh-payment choice. One subject did not agree to partici-pate in the high-payment choice, stating that she felt she hadearned enough in the experiment already. Omitting thissubject from the following analysis has no effect on theseresults.902choice as the scale increased by a factor of 20.We use a Kolmogorov-Smirnov test to identifydifferences between the distributions of thenumber of safe choices made at the low- andhigh-payment levels.3There is a significant dif-ference between the distributions of safechoices between these two payoff-scale condi-tions (p ⬍ 0.01). This test does not, however,explore the extent to which this payoff-scaleeffect is due to the fact that the 20⫻ choice wasmade after the 1⫻ choice.We conducted four additional treatments inwhich each subject completed a single lottery-choice menu that was identical to that describedabove. Unlike our first treatment (with orderedchoices), however, these subjects participated injust one payoff treatment. The four (unordered)treatments tested were: low (1x) real payments,low hypothetical payments, high (20x) real pay-ments, and high hypothetical payments. Therewere 48 subjects in each real-payment treat-ment, and 36 subjects in each hypothetical-payment treatment.4Instructions for all treatmentswere identical, except for the description of theactual choices the subjects faced.5The data from the single-choice treatmentsare summarized in the bottom two rows of Ta-ble 1. Those subjects who completed the low-real-payment decision were slightly less riskaverse than those who completed the orderedtask reported above (those in the low-real-payment treatment made 5.7 safe choices com-pared with 6.1 for those who participated underboth payment conditions). A Kolmogorov-Smirnov test, however, cannot reject the nullhypothesis of equal distributions between thesetwo low-payment treatments (two-sidedp-value ⫽ 0.50). A Wilcoxon rank-sum test alsofails to reject the null hypothesis of equal dis-tributions and central tendency (p ⫽ 0.33).The increase in the number of safe choicesfrom the low-real to high-real payment condi-tions is identical (1 safe choice) between thesetreatments with ordered data in row 5 and thosewith unordered data in row 6, which indicatesthat real payoff-scale effects are important,whether or not decisions are made in an orderedor unordered manner. Again, a Kolmogorov-Smirnov test fails to reject the null hypothesis of3The Kolmogorov-Smirnov test looks for differences inthe two distributions, both in terms of shape and location.It has good power to test for general differences in