Violations of CPT in Mixed Gambles

Guido Baltussen Thierry Post Pim van Vliet

Abstract This study conducts a classroom experiment and an online experiment to examine individual decision-making under risk. Like Levy and Levy (2002), the experiment uses pairs of mixed gambles with moderate probabilities to avoid the framing effect and certainty affect that may affect non-mixed gambles with extreme probabilities. Also, we use Stochastic Dominance criteria to avoid parametric specification of decision-maker preferences. Explicitly accounting for the Wakker (2003) comment, we find several serious violations of Cumulative Prospect Theory (CPT). In fact, in a head-to-head competition between Second-order Stochastic Dominance (SSD) and CPT, most individuals choose the SSD alternative.

Keywords: cumulative prospect theory; expected utility; mixed gambles; probability weighting; stochastic dominance.

Pim van Vliet is corresponding author: Erasmus University Rotterdam, P.O. Box 1738, 3000 DR, Rotterdam, The Netherlands, email: wvanvliet@few.eur.nl, tel: +31-104088931. We appreciate the comments by Jaap Spronk and Philippe Versijp. Financial support by Tinbergen Institute, Erasmus Research Institute of Management and Erasmus Center of Financial Research is gratefully acknowledged. Any remaining errors are our own.

Electronic copy of this paper is available at: http://ssrn.com/abstract=506582

1. Introduction Cumulative prospect theory (CPT) is currently the most popular alternat ive paradigm to the traditional expected utility theory (EUT). CPT is consistent with rank and sign dependent models of choice under risk ( Quiggin (1982) , and Tversky and Kahneman (1992)) and captures two well-documented violations of EUT: loss aversion and probability weighting. According to CPT, individuals choose according to two functions that interact, namely the value function and the probability weighting function. The most common pattern for the value function is an S-shape (convex for losses and concave for gains). The most common pattern of probability transformation is an inverse S-shape. 1 Unfortunately, the empirical evidence for CPT may heavily depend on research design. First, several studies (for instance, Conlisk (1989) , and Camerer (1992)) suggest that violations of EUT are concentrated in prospects with very low probability events. With moderate probabilities, and hence less subjective probability distortion, evidence for CPT is weaker than for very small probabilities. In fact, extreme probabilities (in combination with branch splitting) leads to violation of monotonicity, the key assumption underpinning both EUT and CPT (Birnbaum and McIntosh (1996) and Birnbaum and Thompson (1996)). These findings suggest that the distribution of the probabilities in the prospects is important. Second, the certainty equivalent (CE) method, which is usually employed to characterize the shape of the utility function, possibly suffers from the framing effect (Tversky and Kahneman (1986) because subjects are presented with an unnatural situation; the possible outcomes are restricted to either the positive domain only or the negative domain only. Furthermore, the CE method is sensitive to the certainty effect (see Kahneman and Tversky (1979)), because one of the outcomes is certain and involves no risk. Recently, Levy and Levy (2002) use a new experimental design which considers moderate probabilities only. Furthermore, they focus on pairs of risky gambles (no riskless alternative) that involve both gains and losses. At least for investment choice, these aspects are required in order to ensure that the experiments are realistic; investment alternatives are risky and involve losses as well as gains. Levy and Levy claim to have found a series of violations of Prospect Theory and they argue that the choices of the respondents are best summar ized by a reverse S-shape utility function such as proposed by Markowitz (1952). This provides a rationale for using their novel Markowitz Stochastic Dominance (MSD) criterion. However, as demonstrated by Wakker (2003) , the Levy and Levy experiments actually support CPT after accounting for the CPT-type of probability distortion. In other words, for the Levy and Levy experiments MSD and

Electronic copy of this paper is available at: http://ssrn.com/abstract=506582

CPT are indistinguishable; reverse S -shaped utility predicts the same choice as CPTs combination of a reverse S-shaped weighting function and a S-shaped value function. While the Wakker comment directly calls into question whether the Levy and Levy experiments violate CPT, it does not imply that CPT gives a good description of choice between pairs of mixed gambles with moderate probabilities. In this paper, choice experiments are conducted to shed new light on the descriptive value of CPT for mixed gambles and on the exchange between Levy and Levy and Wakker. We perform two separate experiments, a classroom experiment among undergraduate students (n=289) and an online Web survey among colleagues and students (n=194). Following Levy and Levy (2002), our experiments employ pairs of mixed gambles with moderate probabilities to avoid the above problems with research design. We also use Stochastic Dominance criteria so as to avoid a full parametric specification of decision-maker preferences. A unique feature of our experiment is the use of a novel choice alternative that allows us to distinguish between MSD/CPT and Second-order Stochastic Dominance (SSD), the decision rule associated with traditional EUT with concave utility. Our conclusions are as follows. First, consistent with Wakkers (2003) comment, we find that MSD and CPT are indistinguishable in our experiments if we account for subjective probability distortion. Second, we confirm the results of Levy and Levy; a majority of our respondents prefers the MSD/CPT prospect to the PSD prospect. Third and most important, we find strong evidence against CPT/MSD for mixed gambles with moderate probabilities. In a head-to-head competition between SSD and CPT/MSD, most individuals (63%) choose the SSD alternative. Therefore, we find evidence against CPT for mixed gambles (conform the claim of Levy and Levy), but also against MSD (unlike Levy and Levy). Our results suggest that CPT/MSD misses some basic aspect of individual decision making in mixed gambles with moderate probabilities. This paper is organized as follows. Section 2 explains the experimental design. Section 3 discusses with the experimental results. Section 4 examines the robustness of the results. Finally, our conclusions are summarized in Section 5.

2. Experimental design Following Levy and Levy (2002), we prefer mixed gambles without certain or extreme outcomes to minimize any biases with regards to framing and the certainty effect. Also, as Levy and Levy (2002), we use rules of Stochastic Dominance (SD). These rules allow one to fix the general shape of decision-maker preferences without using a fully specified parametric functional form. Thus, by employing SD we do not try

to figure out the precise shape of decision maker preferences, but we can test whether it belongs to a given family. Furthermore, stochastic dominance rules can be employed with mixed prospects (in contrast to CE) and with change of wealth rather than total wealth (see Levy and Levy (2002) and Le vy and Levy (2002b)). SD rules are typically formulated in terms of EUT. However, this is mostly for notational convenience; the rules are also economically meaningful in the context of many non-expected utility theories. For example, SD rules are invariant to some patterns of probability weighting; see Levy and Wiener (1998) . Our experiments are designed to test for consistency with five different SD rules: FSD, SSD, RSD, PSD and MSD. First-order Stochastic Dominance (FSD) assumes an increasing utility function and it is the cornerstone of most expected and non-expected utility models. Second-order Stochastic Dominance (SSD) assumes an increasing and concave utility function and it is consistent with the assumption of global risk-aversion. Risk-seeking Stochastic Dominance (RSD) covers all non-decreasing risk seeking utilitymaking functions. Prospect Stochastic Dominance (PSD) includes all S-shaped utility making functions, risk seeking for losses and risk-averse for gains. Markowitz Stochastic Dominance (MSD) includes all inverse S-shaped utility functions, which are risk-averse for losses and risk seeking for gains. For a more detailed description of the different SD rules we refer to Levy (1998) . We deviate from Levy and Levy (2002) in two important ways. First, we provide an additional task in the experiments, which makes it possible to discriminate between the MSD criterion (risk seeking for gains) and the SSD criterion (global risk aversion). Second, we include subjective probability weighting in our analysis and calculate the CPT values for each alternative for different CPT parameters. This allows us to control for the Wakker critique. The participants in our classroom experiment are first-year undergraduate students in economics at the Erasmus University Rotterdam. Each subject answers a questionnaire with six tasks and finishes with a gender question, to control for this variable. 2 The different tasks are given in Table I. A second control group of respondents, which will be discussed in the robustness section, do the first three tasks only. Insert Table I about here Task 1 is identical to Experiment 2 of Levy and Levy (2002). The alternatives are designed such that alternative G dominates alternative F by MSD and F dominates G by PSD. Although this task distinguishes between MSD and PSD preferences, no alternative SSD dominates the other. Hence, depending on the outcomes, we may reject S-shaped

preferences (if a significant majority selects G) or reverse S-shaped preferences (if a significant majority selects F), but not the standard EUT with concave utility. Therefore, we have added alternative H, which dominates alternative F and G in terms of SSD. With this new alternative we introduce a head-to-head competition between PSD and SSD (Task 2) and a head-to-head competition between MSD and SSD (Task 3). 3 Figure 1 graphically illustrates the SD rules for Task 2 and Task 3 (a similar graph for Task 1 is included in Levy and Levy (2002)). In Figure 1, where we see that any risk averse decision maker would prefer alternative H to F and G; the cumulative area under the cumulative distribution function for H is always smaller than for F and G. The combination of the first three tasks enables us to test for three different types of preferences: S-shaped preferences (PSD), reverse S -shaped preferences (MSD) and risk averse preferences (SSD). The combined tasks also allow for testing transitivity. If the subjects choose FHG or GFH in Task 1 to 3, they violate the transitivity axiom. 4 Insert Figure 1 about here Task 4 and 6 are designed to control for monotonicity, because I dominates J and M dominates N by FSD. These control tasks test whether the subjects comprehend the experimental setup and whether they the standard axiom of monotonicity. Task 6 is similar to, but less transparent than Task 4. In Task 5, L dominates K in terms of SSD and CPT, but K dominates L by RSD. Task 1, 4, 5 and 6 are also conducted by Levy and Levy (2002) and Task 4 and 5 are presented in Levy and Levy (2001). Wakker (2003) points out that the experimental outcomes of Levy and Levy are consistent with Prospect Theory if one controls for probability transformation. To check if the results are consistent with CPT, we compute the CPT value of each choice alternative. For this purpose we use the parameter estimation of Tversky and Kahneman (1992), Camerer and Ho (1994) , Wu and Gonzalez (1996), and Abdellaoui (2000) 5 and Abdellaoui, Vossman and Weber (2003). The results are presented in Table II. Insert Table II about here Our computations indicate that according to CPT, G is preferred to H, H is preferred to F, G is preferred to H, I is preferred to J, L is preferred to K, and M is preferred to N. No differences show up when different estimations of the CPT parameters are used. Unreported robustness tests indicate that the CPT-dominance of the alternatives remains intact for a very wide range of CPT-parameters. All the estimated parameters in the studies fall wit hin the range where G is preferred to H for any CPT deciding subject.

A CPT subject would thus choose GHGILM in tasks 1 to 6. While the CPT calculations are based on the specific parameter assumptions suggested by the literature, the results are also qualitatively consistent with the extreme orientedness predicted by Cumulative Prospect Theory. To have MSD/CPT preferences subjects must not only choose alternative G, but also prefer G to H, this is the case for the joint preference for alternatives GHG in Task 1 to 3. The questions in our experiment are hypothetical and no real incentives are provided. The absence of real incentives is unlikely to affect the results of this experiment, because several studies have argued and presented empirical evidence that hypothetical and real questions give similar results in decision under risk (Tversky and Kahneman 1992, and Beattie and Loomes 1997). In a recent review of the literature on the effect of financial incentives in experiments, Camerer and Hogarth (1999) conc lude that for simple choice problems subjects do not need real incentives to reveal their preferences.

3. Experimental results Table III includes the main results of the six Tasks of our SD-experiment. We do not observe a significant gender effect, so the gender variable can be considered as irrelevant in our study. 6 In the analysis we perform two-sided Z-tests of population proportions and employ a 5 percent significance level. Insert Table III about here In Task 1, a significant majority of 184 subjects (63.7%) prefer G to F. Recall that alternative G represents the CPT/MSD dominating alternative. This constitutes a rejection of the S -shaped (PSD) decision function because at least 63.7 percent of the subjects do not have this type of preferences. 7 Still, we cannot conclude that 63.7 percent of the subjects behaves according to CPT/MSD. This is because individuals with preferences other than a reverse S-shaped function could still choose alternative G. For example, subjects who employ the mean-variance rule will choose alternative G because with the same expected payoff, alternative G has the lowest variance. Task 2 and Task 3 further scrutinize the preferences of the subjects. In Task 2, 199 subjects (69%) prefer alternative H to F, which is significantly more than 50 percent (p=0.000). Recall that H is the SSD alternative and F is the PSD alternative. Thus, in a head-to-head competition between risk seeking and risk-aversion for the domain of losses

most individuals exhibit risk-aversion. 8 In Task 3, a significant majority of 181 subjects (63%) prefer the SSD alternative H to the CPT/MSD alternative G. This striking result puts the findings of Levy and Levy (2002) and the comment of Wakker (2003) in a different perspective. From Task 1 one might conclude that most individuals obey MSD/CPT. However, in Task 3, the majority prefers the SSD alternative to the MSD/CPT alternative. Task 3 provides direct evidence against both CPT and MSD, because only 108 individuals (37%) prefer G to H. Task 4 and Task 6 are included in order to control for FSD, or consistency with monotonicity. A large majority of 282 (97.6%) and 279 (96.5%) of the respondents prefer I to J and M to N respectively. This basic property underlying both expected and nonexpected utility functions is therefore not rejected. Task 5 yields surprising results. An insignificant minority of 142 (49%) of the subjects chooses the CPT and SSD alternative L. This result provides weak evidence against CPT and SSD. 9 The outcomes are comparable with those of Levy and Levy (2001) and Levy and Levy (2002). Remarkably, Wakker (2003) leaves this violation of CPT (and SSD) unnoticed. A possible explanation would be that alternative K has more positive outcomes than alternative L, which confirms the notion that probability weighting also depends on the number of positive outcomes (Neilson (1992)). Further, alternative K offers the highe st absolute outcome (2,000), which is line with outcome level dependent probability weighting (Rottenstreich and Hsee (2001) ). 10 However, without a clear preference for either alternative, we cannot draw firm conclusions from this task. To summarize, CPT/MSD is severely violated in our experiment. In Task 3, a large majority prefers the SSD alternative to the CPT/MSD alternative, and in Task 5, a small majority violates both CPT and SSD.

4. Robustness of results This section analyzes the robustness of our results. We provide further empirical evidence, test for gender, age and profession effects, and test two basic axioms of choice under risk. Besides the classroom experiment, we have also designed an online questionnaire in order to test the robustness of the results. In this experiment, the subjects are given only the first three tasks of the classroom experiment (Task1, Task 2 and Task 3) . Further, to control for heterogeneity among respondents, a question about the age, gender, and profession of the respondent is added. A total of 194 subjects completed all

tasks of the experiment, of which 85 (44%) are academics from other European universities and research institutions, and 109 (56%) are students. 11 Insert Table IV about here Panel A of Table IV presents the results of the online questionnaire. In Task 1, 66 percent prefer alternative G to F. This number corresponds with the results from our classroom experiment (64%) and with the Levy and Levy results (62%). In Task 2, 69 percent prefers alternative H to F, which is similar to the results of the classroom experiment (69%). Finally, Task 3 confirms our classroom results because 61 percent of the online respondents prefer H to G. Recall that in the classroom experiment this number was slightly higher, 63 percent. Thus, again in a head-to-head competition between MSD/CPT and SSD, the majority prefers the SSD alternative. In the three tasks, the percentages are all significant different from 50 percent. The results of the online experiment closely resemble and confirm the results of the classroom experiment; the majority of the respondents violate CPT/MSD. In the online survey we have also included age, gender, and profession questions to control for possible heterogeneity among respondents. We find no significant gender effect, as in the classroom experiment (p=0.930). We find also no significant age effect (p=0.242). However, students respond differently than academics in the online experime nt (p=0.014). Although the students outcomes are in line with the previous findings, we find slightly different results among academics. In Task 3, a majority (70%) of the students prefers alternative H to G, while this is only a small majority (51%) for the academic respondents. 12 Finally, in Panel B, we compare our classroom results with the online results and the Levy and Levy (2001) and (2002) results. For Task 1, 2, and 3 we find no significant differences with the classroom experiment. For Task 1, 4, and 5 we do not find significant differences with the Levy and Levy results. Only in Task 6 the proportions are significantly different; somehow, our respondents showed significantly less violations of monotonicity.

5. Conclusions Basically, we find three results. First, c onsistent with Wakkers comment , we find, for all plausible value functions and weighting functions, that the MSD and CPT criteria are indistinguishable in our choice experiments. Second, we confirm the results of Levy and Levy; a majority of our respondents prefers the MSD/CPT prospect to the PSD prospect. Third and most important, we find strong evidence against CPT/MSD for mixed gambles. In our Task 3, a large majority prefers our novel SSD alternative to the CPT/MSD alternative, and in Task 5, a small majority violates both CPT/MSD and SSD. In brief, we find evidence against CPT for mixed gambles with moderate probabilities (consistent with the predictions of Levy and Levy), but also against MSD (unlike Levy and Levy). Our results add to the mounting evidence that CPT, and more generally randependent utility theory, misses some key aspect of choice under risk.

References
Abdellaoui, M., 2000, Parameter-free elicitation of utility and probability weighting functions, Management Science 46, 1497-1512. Abdellaoui, M., F. Vossman, and M. Weber, 2003, Choice-based elicitation and decomposition of decision weights for gains and losses under uncertainty, Working paper, University of Mannheim. Birnbaum, M. H., and W. R. McIntosh, 1996, Violations of branch independence in choices between gambles, Organizational Behavior and Human Decision Processes 67, 91-110. Birnbaum, M. H., and L. A. Thompson, 1996, Violations of monotonicity in choices between gambles and certain cash, American Journal of Psychology 109, 501-523. Camerer, C., 1992. Recent tests of generalizations of expected utility theories. (Kluwer Academic Publishers, Dordrecht). Camerer, C. F., and T. H. Ho, 1994, Violations of the Betweenness Axiom and Nonlinearity in Probability, Journal of Risk and Uncertain ty 8, 167-196. Conlisk, John, 1989, Three variants on the Allais example, American Economic Review 79, 392407. Donkers, B., B. Melenberg, and A. Van Soest, 2001, Estimating risk attitudes using lotteries: A large sample approach, Journal of Risk and Uncertainty 22, 165-195. Kahneman, D., and A. Tversky, 1979, Prospect theory: An analysis of decision under risk, Econometrica 47, 263-291. Levy, H., 1998. Stochastic dominance: Investment decision making under uncertainty (Kluwer Academic Publisher, Dordrecht). Levy, H., and M. Levy, 2002, Arrow-Pratt risk aversion, risk premium and decision weights, Journal of Risk and Uncertainty 25, 265-290. Levy, H., and M. Levy, 2002, Prospect theory: Much ado about nothing?, Management Science 48, 1334-1349.

Levy, H., and Z. Wiener, 1998, Stochastic dominance and prospect dominance with subjective weighting functions, Journal of Risk and Uncertainty 16, 147-163. Levy, M., and H. Levy, 2001, Testing for risk aversion: a stochastic dominance approach, Economics Letters 71, 233-240. Markowitz, Harry, 1952, The utility of wealth, Journal of Political Economy 60, 151-158. Neilson, W., 1992, A mixed fan hypothesis and its implications for behaviour toward risk, Journal of Economic Behavior & Organization 19, 197-211. Quiggin, J., 1982, A theory of anticipated utility, Journal of Economic Behavior & Organization 3, 323-343. Rottenstreich, Y., and C. K. Hsee, 2001, Money, kisses, and electric shocks: On the affective psychology of risk, Psychological Science 12, 185-190. Starmer, C., 2000, Developments in non-expected utility theory: The hunt for a descriptive theory of choice under risk, Journal of Economic Literature 38, 332-382. Tversky, A., and D. Kahneman, 1986, Rational choice and the framing of decisions, Journal of Business 59, 251-78. Tversky, A., and D. Kahneman, 1992, Advances in Prospect-Theory - Cumulative Representation of Uncertainty, Journal of Risk and Uncertainty 5, 297-323. Wakker, P. P., 2003, The data of Levy and Levy (2002) "Prospect theory: Much ado, about nothing?" - Actually support prospect theory, Management Science 49, 979-981. Wu, G., and R. Gonzalez, 1996, Curvature of the probability weighting function, Management Science 42, 1676-1690.

10

Table I Stochastic Dominance Experiment: The choices presented to the subjects

This table shows the six tasks we presented to our student respondents. In each task, the subject was asked to invest a hypothetical 10,000 for one month in one of two stocks. For each stock, the possible 1-month gains/losses and corresponding chances are shown.

Stock Task 1:
Gain/loss -1600 -200 1200 1600

F
Probability 0.25 0.25 0.25 0.25 Gain/loss -1000 -800 800 2000

G
Probability 0.25 0.25 0.25 0.25

Task 2:
Gain/loss -1600 -200 1200 1600

F
Probability 0.25 0.25 0.25 0.25 Gain/loss -1000 -800 1200 1600

H
Probability 0.25 0.25 0.25 0.25

Task 3:
Gain/loss -1000 -800 800 2000

G
Probability 0.25 0.25 0.25 0.25 Gain/loss -1000 -800 1200 1600

H
Probability 0.25 0.25 0.25 0.25

Task 4:
Gain/loss -500 2500

I
Probability 0.33 0.67 Gain/loss -500 2500

J
Probability 0.5 0.5

Task 5:
Gain/loss -500 500 1000 2000

K
Probability 0.25 0.25 0.25 0.25 Gain/loss 0 1500

L
Probability 0.50 0.50

Task 6:
Gain/loss -500 2000 5000

M
Probability 0.30 0.30 0.40 Gain/loss -500 0 500 1000 2000 5000

N
Probability 0.10 0.10 0.10 0.20 0.10 0.40

11

Table II CPT values of each Task

Panel A of this table presents the CPT parameter values for four different empirical studies. Panel B computes the CPT values for each of our tasks for these different parameter values. The CPT values are calculated as follows:

w
i=1

( pi )v ( xi ) +

j = k +1

( p j ) v( x j ) where > 0 is the loss aversion

parameter, w + ( p i ) and w ( p j ) are decision weights that are calculated based on the cumulative rankdependent probabilities. One considers a prospect with outcomes x1 K xk 0 xk +1 K xn having probabilities p1 ,K, p n . The value function is calculated by the following formula: x if x > 0 v ( x ) = 0 if x = 0 ( x ) if x < 0 The decision weights are calculated by the following formula for the positive and negative domain separately with the one parametric specification: p p and with the two parametric specification as: w ( p) = w ( p) = ( p + (1 p ) ) 1/ p + (1 p )

Panel A: CPT parameter values in literature Value function Study:

Tversky and Kahneman (1992) Camerer and Ho (1994) Wu and Gonzalez (1996) Abdellaoui (2000) Abdellaoui (2000)* Abdellaoui, Vossman and Weber (2003) (+) 0.88 0.37 0.52 0.89 0.89 0.91 () 0.88 0.37 0.52 0.92 0.92 0.96 (+) 0.61 0.56 0.71 0.60 0.60 0.78

Weighting function
0.65 0.91 () 0.69 0.56 0.71 0.70 0.65 0.78 0.84 1.20

Panel B: CPT values for different alternatives CPT alternative Study:

Tversky and Kahneman (1992) Camerer and Ho (1994) Wu and Gonzalez (1996) Abdellaoui (2000) Abdellaoui (2000)* Abdellaoui, Vossman and Weber (2003)
*

H -159 -5.29 -15.9 -277 -294 -508

K 168 1.08 9.2 146 126 150

L 262 5.89 20.7 279 264 370

M 908 15.6 53.9 976 940 1271

N 682 6.52 34.2 716 674 961

-215 -137 -5.68 -5.13 -16.2 -15.5 -358 -253 -373 -625 -278 -495

318 169 1.43 -1.71 13.9 298 290 399 0.7 126 104 110

See footnote 5.

12

Table III Results of the class room SD-experiment

This table shows the results of the classroom SD-experiment ( with 289 students ). The table shows the absolute frequency (Number), the relative frequency (Percent), and the p-values for the null that the percentage is 50% (Z-test). In addition, the table combines the results of the first three tasks and gives the absolute and relative fre quency of participants who behave according to CPT/MSD, SSD, PSD, RSD, or who violated transitivity. Finally, the table shows the number of male/female respondents and tests for gender differences (chi-square test).

Individual Tasks
Task no. Task 1 Choice F G Total: Task 2 F H Total: Task 3 G H Total: Task 4 I J Total: Task 5 K L Total: Task 6 M N Total: Gender Male Female Total: 108 181 289 282 7 289 142 147 289 279 10 289 209 80 289 Number 105 184 289 90 199 Percent 36.3 63.7 100.0 31.1 68.9 100.0 37.4 62.6 100.0 97.6 2.4 100.0 49.1 50.9 100.0 96.5 3.5 100.0 72.3 27.7 100.0 0.501** 0.000* 0.384* 0.000
*

Tasks 1-3 combined

p-value 0.000* Choice GHG GHH FHH 0.000
*

Number 89 73 36 56 16 12 6 1 289 89 109 56 18 17 289 16 17

Percent 30.8 25.3 12.5 19.4 5.5 4.1 2.1 0.3 100.0 30.8 37.7 19.4 6.2 5.9 100 5.5 5.9

FFH GFH FFG GFG FHG

0.000*

Total: CPT/MSD SSD PSD RSD Inconsistent Total Monotonicity Transitivity

* Two sided Z-test of population proportions with null hypothesis p=0.5. **Chi-square test (equal proportions)

13

Table IV Results of the online SD -experiment and robustness tests

Panel A of this table shows the results of the online SD-experiment. A total number of 194 respondents have completed the six different Tasks. We provide numbers, percentages, and test for significance (Z -test). Additionally we combine the first three Task results and present the number of participants who behave according to CPT/MSD, SSD, PSD, RSD, or who have inconsistent (transitivity) preferences. Finally we present the average age, gender, and profession and test for differences with the chi-square statistic. Panel B compares the results of the clas sroom experiment with the results of Levy and Levy 2002, 2001 and the Internet experiment. For each study, the sample size, the relative frequencies, the Z-statistics, and the p-values are shown.

Panel A: The results of the online SD-experiment Individual Tasks

Task no. Task 1 Choice F G Total: Task 2 F H Total: Task 3 G H Total: Age Average Number 66 128 194 60 134 194 75 119 194 29.2 109 61 24 194 124 70 194 56.2 31.4 12.4 100.0 63.9 36.1 100.0 0.930** Transitivity 19 9.8 Percent 34.0 66.0 100.0 30.9 69.1 100.0 38.7 61.3 100.0 0.242
**

Tasks 1-3 combined

p-value 0.000 Choice GHG GHH FHH FFH 0.000 GFH FFG GFG FHG 0.002 Total: CPT/MSD SSD PSD RSD Inconsistent Total Number 49 57 25 27 9 10 13 4 194 49 82 27 17 19 194 Percent 25.3 29.4 12.9 13.9 4.6 5.2 6.7 2.1 100.0 25.3 42.3 13.9 8.8 9.8 100

Profession Student Faculty Other Total: Gender Male Female Total:

0.014**

* Two sided Z-test of population proportions with null hypothesis p=0.5. **Chi-square test (equal proportions)

14

Panel B: Comparison of classroom experiment with the results of Levy and Levy (2001 and 2002) and Internet experiment
Classroom N Task 1 Task 4 Task 5 Task 6 Task 5 Task 4 Task 1 Task 2 Task 3 289 p=0.64 p=0.02 p=0.49 p=0.04 p=0.49 p=0.02 p=0.64 p=0.69 p=0.63 LL02 84, 260 p=0.62 p=0.04 p=0.51 p=0.18 p=0.44 p=0.05 p=0.66 p=0.69 p=0.61 LL01 194 Internet 194 0.279 -1.054 -0.436 -5.585 1.108 -1.520 -0.525 -0.056 0.295 0.780 0.292 0.663 0.000 0.268 0.129 0.600 0.955 0.768 Z-stat p-value*

* Two sided Z-test of difference (D) between two population proportions with null hypothesis D = 0, against alternative hypothesis D 0.

15

Figure 1
Task 2, and Task 3 of the SD -experiment
This Figure graphically shows that alternative H dominates F and G by SSD. Thus we would expect any subject with risk averse preferences to select H in Task 2 and Task 3.

Cumulative probability

0.75

H -/-150 F

0.5

0.25 + 150 0 -1000 -200 600 1400 2200

-1800

payoff (euro)

Cumulative probability

-/-100 0.75 G + 100 0.5 H

0.25

0 -1800 -1000 -200 600 1400 2200

payoff (euro)

16

See Edwards (1996) for a literature review of Cumulative Prospect Theory and Starmer (2000) for an historical introduction of utility theory and an overview of competing theories of choice under risk. 2 Donkers, Melenberg and Van Soest (2001) show that an individuals risk attitude and probability weighting function is influenced by gender, education level, age and income. Since all subjects in the classroom experiment were first year students we may assume no significant differences in age, inco me and education. 3 Since in Task 2 both alternatives have the same outcomes in the gain domain, PSD coincides with RSD, and MSD coincides with SSD. In Task 3 both alternatives have the same outcomes in the loss domain, therefore MSD coincides with RSD, and PSD coincides with SSD. 4 For example, FGH corresponds to F f G , H f F , G f H . 5 In this study Abdellaoui discusses two decision weighting functions. Different parameter values are used for the one parameter and two -parameter weighting functions. 6 This was tested with a chi-square test for equality of proportions. 7 The actual percentage can be even higher as choosing F does not imply that preferences are necessarily Sshaped. 8 We stress that the notion of risk aversion does not apply in case of subjective probability weighting (CPT). 9 In this task there is a CPT dominant alternative, but no MSD dominated alternative. 10 Levy and Levy (2002)b report that with lower outcome levels the preference shifts to alternative L. This finding further supports the outcome dependent weighting function hypothesis. 11 This questionnaire can be found on: http://www.few.eur.nl/few/people/wvanvliet/questionnaire/ invest1/index.htm. We grouped the non-students groups into one group, which we refer to as the academic group. 12 More details are available upon request from the corresponding author.

17