Schedule - Parallel Session 3 - Risk and Time Preferences 2IDL First Floor Syndicate Room - 15:40 - 17:10
How Real is 'Hypothetical Bias' in the Context of Risk and Time Preference Elicitation?
This paper presents experimental evidence to which extent specific types of hypothetical risk and time preference questions are suitable to induce individuals to unveil their effective decision patterns. My work contributes to the literature in several ways: First, I calculate size and sign of hypothetical bias in the context of stated risk and time preferences. Second, I identify subject-specific factors which allow predicting magnitude and direction of this bias. Third, the previous step enables me to develop bias correction measures and to evaluate their effectiveness. Within the underlying experimental study, all groups have been asked to fill in simple and identical risk (gain and loss dimension) and time preference elicitation items at two points during the session. Whilst for the control groups all associated payoffs remained purely hypothetical, the decisions of participants in the treatment groups at the second point became payoff-relevant. Around the group-level findings I develop a bias model, consisting of four components: a random component, one resulting from task unfamiliarity, another one due to social desirability considerations, and the fourth as hypothetical bias in the narrower sense. I argue that social desirability does not play a prominent role since the study’s designs ensures a sufficiently high level of anonymity, additionally the incentive-compatible payoff for the treatment groups should disperse this component. Those considerations yield the central equation for the estimation of within-subject bias, based on explanatory variables capturing the ability to cope with unfamiliar tasks (thus expected to reduce bias), and a second subset of explanatory variables, likely to exert a certain influence on decisions in an incentivised setup. One can observe the following tendencies: more conscientious individuals with an academic background overstate their degree of risk-affinity in a one-shot hypothetical question. Individuals with a larger available budget tend to understate their risk attitude. The size of an individual’s budget exerts a significant effect on bias in the context of time preference elicitation, too. Financially less restricted subjects tend to underreport their switching point, which implies they actually discount more heavily or have a higher degree of present bias than stated in a one-time, hypothetical framework. The same holds true for respondents who are less self-reliant. This within-subject analysis allows the identification of subject-specific bias predictors, which can be used to test the effectiveness of potential bias correction measures as strategy to improve item reliability. Evaluated bias correction procedures do not yield a gain in precision for stated time preferences and the loss domain in the case of risk attitude. However, there results a substantial increase in precision in the gain domain, which underscores respondents’ sensitivity to the framing of presented items.
I Want to Know it Now: Measuring Preferences Over the Temporal Resolution of Consumption Uncertainty
We design an experiment to elicit preferences over the temporal resolution of consumption uncertainty as axiomatized in Kreps and Porteus (1978) and Epstein and Zin (1989). Subjects consume in the lab by surfing YouTube which is contrasted by a real effort task. Lotteries over consumption at different points in time introduce actual consumption uncertainty – as opposed to income uncertainty. Assessing a series of choices, we find that on average, subjects are willing to forgo about 4% of their total consumption in order to expedite the resolution of consumption uncertainty. A structural estimation suggests that subjects on average indeed prefer an early resolution consumption uncertainty. This, however, is mainly driven by a minority of subjects with a strong preference for early resolution.
Aspiration Levels and Preference for Skewness in Choice under Risk
Francesco Zaffuto; Giorgio Coricelli; Enrico Diecidue
This paper describes an experiment designed to study the effect of aspiration levels on individual choices under risk. Aspiration levels are values greater than a threshold that divides the outcomes in gains or losses. Our experimental conditions can be divided in two groups. In the first group there are conditions with pair of prospects having the same expected value but different probabilities of winning and losing. In one condition both prospects are mixed. In another condition one prospect is mixed and the other one has a probability of winning equal to zero. In a third condition, one prospect is still mixed while the other one has a probability of winning equal to one. Comparing these conditions we observe a preference reversal effect explained by the overall probability of winning. In the second group there are conditions with pair of prospects having the same expected value but different possibilities of achieving aspiration levels. There are conditions where both prospects provide the possibility to achieve aspiration levels, conditions where only one prospect can lead to aspiration levels and other conditions where in both prospects it is not possible to achieve aspiration levels. In conditions where only one prospect provides the possibility to achieve aspiration levels, we observe preferences for that prospect. The resulting choice patterns characterize a heuristic for reducing the complexity of risky decisions. In cases where aspiration levels are not predictive, choices can be explained by preferences for positive skewness. Our results confirm the efficacy of a two-pronged approach that includes both compensatory (e.g. preferences for positive skewness), and simplifying strategies (e.g. aspiration levels) for choosing among risky prospects. Our model fitting suggests that among the standard models of choice under risk, cumulative prospect theory best fits our experimental data.
Bayesian Rapid Optimal Adaptive Design: Method and Applications
Taisuke Imai; Roman Weber; Debajyoti Ray; Daniel Golovin; Andreas Krause; Colin Camerer
Questions used in experiments and surveys in social science are typically developed by intuitive hunches and cumulative search for informative questions that can best separate competing theories. The conventional experimental designs that have emerged are therefore typically a fixed set of test questions. We propose an approach, termed Bayesian Rapid Optimal Adaptive Design (BROAD), in which the sequence of questions is customized for each subject rather than fixed. The subjects themselves tell us, through their answers, the “best” (i.e., most informative) question to ask them next. BROAD method has several advantages. First, the posterior distributions of all theories and parameter values are quickly recomputed for each subject after each trial since it is a necessary step in finding the optimal next question. As a result, when the experiment is over, much of the data analysis is already done. Second, since BROAD method economizes on information gained per minute, they are especially useful for subjects who have scarce time or become bored or habituated quickly. In theory, subjects might prefer to strategically manipulate their early responses in order to get “better” (more economically valuable) future test questions. We pay special attention to the problem of manipulation, and discuss tests to detect it and methods to minimize it. In the first application, we tried to separate six theories of decision making under risk: expected value, expected utility with constant relative risk aversion, prospect theory, cumulative prospect theory, mean-variance-skewness, and normalized mean-variance-skewness. We found that most subjects, after making 30 choices, could be reliably classified as choosing according to expected value maximization or two variants of prospect theory. In the second application, we tried to separate four theories of time preferences: exponential discounting, hyperbolic discounting, quasi-hyperbolic discounting, and generalized hyperbolic discounting. We found that the posterior distributions of the generalized hyperbolic discounting and exponential discounting were significantly higher than the other models. Somewhat surprisingly, the evidence for quasi-hyperbolic discounting is not strong. Finally, we discuss applications of the method to Convex Budget design for risk and uncertainty (Ahn et al., 2014; Choi et al, 2007) or time (Andreoni and Sprenger, 2012).