Schedule - Parallel Session 4 - Field Experiments 1WMG IMC Room 246 - 11:00 - 12:30
Bayesian Updating: Evidence from the Field
Constantinos Antoniou; Christos P. Mavis
To apply Bayes Rule when forming expectations agents must correctly assess the inherent “process variance” of each relevant cue, and place a larger weight on those cues with a smaller process variance. We test this notion by analysing subjective probabilities inferred from odds on the outcomes of tennis matches, exploiting natural variation in process variance related to the format with which tennis matches are played. Specifically, men’s tennis matches are played in either a best-out-of-three set format (Master-series matches) or a best-out-of-five set format (Grand-Slams). In the longer Grand-Slam matches the more skilful player is more likely to win, therefore, an indicator of skill (like the players official ranking) is a signal with relatively lower process variance for those matches. Our tests examine whether bookmakers sufficiently adjust their subjective probabilities for the high-skill player in Grand-Slam matches to reflect this reduction in process variance. Our results are consistent with “process variance neglect”, i.e., bookmakers are not adjusting their subjective probabilities for the high-skill player sufficiently, which results in a higher probability bias for the best-out-of-five matches. Although we mainly use bookmaker (fixed) odds to infer subjective probabilities, our results continue to hold when we infer probabilities from odds achieved on a person-to-person betting exchange. Because the bias in Grand-Slams implies that bookmakers are offering overly attractive odds for the high-skill player, the bias is costly and bookmakers are on average earning significantly less in best-out-of-five matches. To control for other factors that vary between Master and Grand-Slam matches, other than match length, we conduct a placebo test using data for women matches, where both Master and Grand-Slam matches are played in a best-out-of-three match format. In these tests we find insignificant differences in probability biases and insignificant differences in profits between Master and Grand-Slam matches.
Decision-Making Under the Gambler's Fallacy: Evidence from Asylum Judges, Loan Officers, and Baseba
Daniel L. Chen; Martin Schonger
We find consistent evidence of negative autocorrelation in decision-making that is unrelated to the merits of the cases considered in three separate high-stakes field settings: refugee asylum court decisions, loan application reviews, and major league baseball umpire pitch calls. The evidence is most consistent with the law of small numbers and the gambler’s fallacy “people underestimating the likelihood of sequential streaks occurring by chance “leading to negatively autocorrelated decisions that result in errors. The negative autocorrelation is stronger among more moderate and less experienced decision-makers, following longer streaks of decisions in one direction, when the current and previous cases share similar characteristics or occur close in time, and when decision-makers face weaker incentives for accuracy. Other explanations for negatively autocorrelated decisions such as quotas, learning, or preferences to treat all parties fairly, are less consistent with the evidence, though we cannot completely rule out sequential contrast effects as an alternative explanation.
Risk Attitudes, Sample Selection and Attrition in a Longitudinal Field Experiment
Longitudinal experiments allow one to compare inferences about behavior over time for the same individuals, and evaluate the temporal stability of latent preferences. But longitudinal experiments entail the possibility of sample selection and sample attrition over time, confounding inferences about temporal stability. We evaluate the hypothesis that risk preferences are stable over time using a remarkable data set combining administrative information from the Danish registry with longitudinal experimental data we designed to allow better identification of joint selection and attrition effects with respect to risk attitudes. Our design builds in explicit randomization on the incentives for participation. We show that there are significant sample selection effects on inferences about the extent of risk aversion, but that the effects of subsequent sample attrition are minimal. Ignoring sample selection leads to inferences that subjects in the population are more risk averse than they actually are. Correcting for sample selection and attrition affects utility curvature, but does not affect inferences about probability weighting. Properly accounting for sample selection and attrition effects leads to findings of temporal stability in overall risk aversion. However, that stability is around different levels of risk aversion than one might naively infer without the controls for sample selection and attrition we are able to implement. This evidence of “randomization bias” from sample selection is important given the popularity of field experiments that rely on randomization, and the effects that risk attitudes have for economic behavior in general.
Do Customers Return Excessive Change in a Restaurant? A Field Experiment on Dishonesty
We conducted a field experiment on dishonesty in a restaurant. Customers who paid with cash received excessive change, and we examined whether customers behaved honestly and returned the excessive change to the waiter, and how this depended on various factors. In total, we collected data from 192 tables. The excessive change was either 10 or 40 extra Shekels (about 2 or 8 Euros). The two levels of excessive change allow us to test whether higher stakes (in a zero-sum game) increase dishonesty. To examine gender differences in our experiment, we documented the customers’ gender and limited our sample to tables with one or two diners. We hypothesized that female tables will return the excessive change more often than male tables. Assuming that honesty is a valuable trait, we hypothesized that people would have more incentive to behave honestly and return the extra change given to them when they dine with someone else compared to the case that they dine alone. To examine the issue of closeness, we recorded for every customer whether he is a repeated customer or not, and whether he holds a membership card of the restaurant. A repeated customer gets to know the waiters and therefore is closer to them, and is also likely to return in the future to the restaurant compared to one-time customers. Therefore, we hypothesized that repeated customers and restaurant members will return the excessive change more often than others. The dependent variable we are interested in is whether the customer returned the extra change that he received. Only 64 customers out of 192 (33%) returned the excessive change. We found, in line with our hypothesis, that repeated customers return the extra change much more often than one-time customers and that diners who have the restaurant club membership return much more often than non-members. The hypothesis that two diners will return the change more often than one diner is not supported. Female tables returned the extra change much more often than male tables. Interestingly, the behavior of mixed tables (a man and a woman) is much closer to male tables than to female tables. Only 15.6% of the people who got 10 extra Shekels returned it, whereas 51% of those who received 40 extra Shekels returned it, in opposite direction to our hypothesis. This seems to suggest that the psychological disutility from behaving dishonestly increases when the amount involved increases, and can even increase more than linearly. That is, the disutility from keeping 40 Shekels that should not be yours is so much higher than the disutility of keeping 10 Shekels (and in particular more than four times higher), that despite the economic gain increasing by a factor of four, many people behave dishonestly in the 10-Shekels case but honestly in the 40-Shekels case (in the experiment the comparison is between subjects).