A Bayesian View of the Alternative-Outcomes Effect Kuninori Nakamura (knaka@ky.hum.titech.ac.jp) Japan Society for the Promotion of Science/ Graduate School of Decision Science & Technology, Tokyo Institute of Technology Kimihiko Yamagishi (kimihiko@ky.hum.titech.ac.jp) Graduate School of Decision Science & Technology, Tokyo Institute of Technology the judged probability of the focal outcome. They argue that comparing the focal outcome to the strongest alternative can serve as a useful heuristic—that is, a short-hand process that usually yields good results with minimal effort. Like other recently proposed heuristics that suggest that people sometimes use one or two simple but generally valid cues to make seemingly complex decisions (e.g., take-the-best; Gigerenzer & Goldstein, 1996), the comparison heuristic suggests that people’s perceptions of certainty are often heavily or perhaps sometimes exclusively determined by the consideration of only two pieces of information. There are several alternative explanations to the alternative-outcomes effect. Dougherty & Hunter (2003) have suggested that this effect may result from the working memory span. Still others (Teigen, 2001) have pointed out that this phenomenon is based on an assessment of the presence or absence of causal factors, which could produce or inhibit the outcome. This article provides another explanation of the alternative-outcomes effect. The previous explanations presume that people commonly use heuristic short cuts to make quick probability judgments. However, this article, proposes that the alternative-outcomes effect is based on rational Bayesian inference. In the present research, we formalize the cognitive processes of the alternativeoutcomes effect as a result of Bayesian inference. In this formalization, we show that the estimated probability for winning the raffle corresponds to a posterior probability estimate that amended the prior probability estimate, after participants were given additional data on the distribution of multiple outcomes. Abstract The purpose of this study is to provide a new explanation of the alternative-outcomes effect (Windschitl & Wells, 1998). Although the alternative-outcomes effect has been posited as evidence for the use of heuristics, this study suggests that they result from Bayesian inference reasoning. We reinterpret the results of previous studies, and reanalyze existing data. Some theoretical suggestions are also discussed. Introduction How do people commonly judge probabilities? Since many heuristics and biases have been discovered that underlie the process of making informal judgments of probability, the basis for human probability judgments has become one of the most important issues in judgment and decision-making research. Many empirical studies have been conducted and several explanations have been proposed. One of the important findings from recent studies comes from Windschitl and Wells’ (1998) on the alternativeoutcomes effect. They revealed that probability is sensitive to variations in how alternative outcomes are distributed, even when such variations have no bearing on the objective probability of the focal outcome. Windschitl and Wells (1998) clearly demonstrated the alternative-outcomes effect. In one experiment they informed the participants that they held 21 raffle tickets and were asked to estimate their chances of winning. In the first scenario outcome, participants were told that there were five other raffle players, who held 14, 13, 15, 12, and 13 tickets, respectively; in the second outcome, the other five players supposedly held 52, 6, 2, 2, and 5 tickets. Although the objective probability of winning was in fact identical for the two scenarios, participants informed of the first distribution actually estimated a higher probability of winning the raffle than those informed of the second distribution. Windschitl and Wells called this phenomenon the alternative-outcomes effect. Later, Windschitl and others replicated this effect for a variety of experimental conditions (Windschitl & Young, 2001; Windschitl, Young, & Jenson, 2002; Windschitl & Krizan, 2005). Windschitl and Wells (1998) offered an interesting account for the alternative-outcomes effect. According to them, when judging the probability of a focal outcome, people seem to compare the strength of the focal outcome with the strength of the strongest alternative outcome. The more this comparison favors the focal outcome, the greater A Bayesian interpretation In this study, we do not suggest that people use heuristics. Rather, we propose that people estimate subjective certainty for the focal outcome according to a Bayesian outlook. In previous studies, it was assumed that participants’ estimates should be based solely on the ratio of the focal outcomes to the total outcomes. For example, when told that they held 21 raffle tickets, while the five other raffle players held 14, 13, 15, 12, and 13 tickets, respectively, the “rational” answer for their odds of winning should be about 0.24 (=21/ (21+14+13+15+12+13)). The present research, however, indicates that people consider both the distribution of the focal and alternative outcomes and their prior beliefs about the chances of their winning the raffle. Thus, our approach 165 does not consider that the given distributions represent values for the probability of winning. Rather, we assume that the distributions serve as additional “data”, which the participants use to support or revise their prior beliefs. The revision of prior beliefs according to new data can be formalized easily via a Bayesian approach, wherein the probability estimate for the focal outcome may be viewed as a posterior probability. More specifically, the posterior probability p(H|Y) of winning the raffle, Y, after seeing the actual distribution, H, can be obtained from Bayes’ theorem: p (H | Y ) = p (H )pp(Y(Y) | H ) . Application to other phenomena A Bayesian approach to probability judgment can explain not only the alternative-outcomes effect but also other related studies, such as the ratio bias (Denes-Raaji & Epstein, 1994). The ratio bias refers to people's tendency to select betting options that display a greater absolute number but a smaller probability of winning (e.g., 9/100) than an option with a smaller absolute number but better odds (e. g., 1/10). To explain this bias, we consider the ratios (“9/100” or “1/10”) as “data”, which confirm or refute the prior belief. We also assume that the prior distribution is a beta distribution: (1) In this equation, p(H) is the prior belief about winning the raffle; p(Y) is the prior belief about the shape of the distribution; and p(Y|H) is the conditional probability of Y occurring, given H. According to Bayes' theorem, the posterior probability is proportional to the product of the prior probability and the likelihood of H. That is, (2) p H | Y ∝ p (H ) p (Y | H ) . ( f ( x) = b −1 L x )= ∏ n! i x ∏ i p xi i . (4) where xi is frequency of the outcome, pi is the probability of xi occurring ,and n is the total number of outcomes, evident when all the frequencies are summed, (6) Denes-Raj, V., & Epstein, S. (1994). Conflict between intuitive and rational processing: When people behave against their better judgment. Journal of Personality and Social Psychology, 66, 819–829. Dougherty, M. R. P., & Hunter, J. E. (2003). Probability judgment and subadditivity: The role of working memory capacity and constraining retrieval. Memory and Cognition, 31, 968–982. Gigerenzer, G., & Goldstein, D. G. (1996). Reasoning the fast and frugal way: Models of bounded rationality. Psychological Review, 103, 650-669. Teigen, K. H. (2001). When equal chances=good chances: Verbal probabilities and the equiprobability effect. Organizational Behavior and Human Decision Processes 85, 77-108. Windschitl, P., & Krizan, Z. (2005). Contingent approaches to making likelihood judgments about polychotomous cases: The influence of task factors. Journal of Behavioral Decision Making, 18, 281–303. Windschitl, P. D., & Wells, G. L. (1998). The alternativeoutcomes effect. Journal of Personality and Social Psychology, 75, 1411–1423. Windschitl, P. D., & Young, M. E. (2001). The influence of alternative outcomes pn gut-level perceptions of certainty. Organizational Behavior and Human Decision Processes, 85, 109–134. Windschitl, P. D., Young, M. E., & Jenson, M. E. (2002). Likelihood judgment based on previously observed outcomes: The alternative-outcomes effect in a learning paradigm. Memory and Cognition, 30, 469–477. We also assume that p(Y|H) is a multinomial function, 2 (5) References ) x . Thus, the certainty of winning the bet, after seeing the ratio, can be calculated as the posterior mean: a + nhit . (7) p( H | Y ) = a + b + ntotal In this formalization, a qualifying condition that satisfies p{H | (9,100 )} ≥ p{H | (1,10 )} surely exists. We can assume that p(H) remains the same across the set of experimental conditions in the studies of the alternativeoutcomes effect. After all, the posterior probability of winning the raffle is proportional to the conditional probability of seeing the distribution of the focal and alternative outcomes: (3) p H | Y ∝ p (Y | H ) . p(Y | H ) = f (x1 B ( a , b) B(a, b ) = ∫10 π a−1 (1 − π ) dπ . ) ( π a −1 (1 − π ) b−1 ∑x . i Experimental results and discussion The above formalization indicates that the probability estimate for the focal outcome is proportional to p(Y|H). To test this prediction, we performed an experiment in which participants were required to estimate their subjective certainty about winning at a fictitious slot machine. This slot machine showed 6 characters of the alphabet (A, B, C, D, E, F), but only A was a winning letter. Participants were shown 155 multinomial outcomes, which represented the results of turns at the slot machine. Then they were asked to estimate the subjective probability of a focal outcome. We calculated estimated means for each stimulus. We classified the stimuli according to frequency of focal outcome, and calculated the likelihood of the distribution of the alternative outcome for each condition. In these calculations, we set the value of each of the xi at 1/6. The results indicate that the estimated probabilities for the focal outcomes were proportional to the likelihood of Y (rs = 0.54-0.92, all ps < .05). Thus, the results support predictions based on a Bayesian approach. 166