Observed Variability and Values Matter: Toward a Better Understanding of Information Search and Decisions from Experience

The search for different options before making a consequential choice is a central aspect of many important decisions, such as mate selection or purchasing a house. Despite its importance, surprisingly little is known about how search and choice are affected by the observed and objective properties of the decision problem. Here, we analyze the effects of two key properties in a binary choice task: the options’ observed and objective values, and the variability of payoffs. First, in a large public data set of a binary choice task, we investigate how the observed value and variability relate to decision-makers’ efforts and preferences during search. Furthermore, we test how these properties influence the chance of correctly identifying the objectively maximizing option, and how they affect choice. Second, we designed a novel experiment to systematically analyze the role of the objective difference between the options. We find that a larger objective difference between options increases the chance for correctly identifying the maximizing option, but it does not affect behavior during search and choice. Copyright © 2013 John Wiley & Sons, Ltd. key words decisions from experience; information search; bounded rationality; maximization; payoff variability; memory In many important real-life decisions, we seek out information about different possibilities before making a choice. For example, most people would not purchase a house without looking at several possibilities or marry a partner without having gotten to know him/her first. While the question of information search was often sidestepped in classical decision literature, significant progress has been made toward a better understanding of search and choice in recent years. For example, we have learned that people seem to generally search for little information before making a choice (for an overview, see Hau, Pleskac, & Hertwig, 2010), and that search can be affected by several characteristics of the decision maker and the choice ecology (for an overview, see Lejarraga, Hertwig, & Gonzalez, 2012). Relatively little is known, however, about how search is related to properties of payoffs that are actually observed during search, and how these properties affect subsequent choice. COSTS OF INFORMATION SEARCH VERSUS ACCURACY OF CHOICE A general finding in the decision-making literature is that people tend to search for “little” information before making a consequential choice between different options (e.g., Hau et al., 2010; Hertwig & Pleskac, 2010). This behavior can be advantageous. For example, relying on smaller amounts of information reduces the explicit (e.g., monetary) and implicit (e.g., cognitive) costs of information search (Hau, Pleskac, Kiefer, & Hertwig, 2008), reduces demands on working-memory capacity (Kareev, 2000; Rakow, Demes, & Newell, 2008), tends to amplify differences between the options and thereby renders choice easier (Hertwig & Pleskac, 2008, 2010), and maximizes the time available for other decisions (Vul, Goodman, Griffiths, & Tenenbaum, 2009). As suggested by the statistical law of large numbers, however, using less information can come at a price: The chance of forming a valid representation of a decision problem’s objective payoff structure strongly decreases with less and less information gathered (Brehmer, 1980; Fiedler, 2000; Hau et al., 2008; Johnson, Budescu, & Wallsten, 2001). Specifically, people are likely to under-experience low-probability payoffs with smaller samples and thereby risk getting a wrong impression about the options’ objective values (Fox & Hadar, 2006; Hertwig, Barron, Weber, & Erev, 2004; Hertwig & Pleskac, 2010). As a result, they might make suboptimal decisions, because the samples do not allow them to correctly identify the objectively maximizing option. This trade-off between the costs of searching and the accuracy of choice has inspired many studies that investigated how much information should optimally be acquired (e.g., Fiedler & Kareev, 2006; Gittins, 1979; Hertwig & Pleskac, 2010; Kareev, 2005; Vul et al., 2009; Wallsten, Budescu, Erev, & Diederich, 1997). Maybe most prominently, the Gittins index (Gittins, 1979) predicts for each point in time which option should be selected to make an “optimal” decision, given the exact history of experiences. Unfortunately, such prescriptive solutions often require vast computational resources and make strong and not always tenable assumptions about the environment and the decision maker’s goals (Cohen, McClure, & Yu, 2007). Here, we evaluate how peoples’ actual experiences during information search are related to search effort and how they, in turn, affect choice. Specifically, we investigate the effects of two key ecological properties of information search and choice: the variability and value of the payoffs. *Correspondence to: Katja Mehlhorn, Dynamic Decision Making Laboratory, Department of Social and Decision Sciences, Carnegie Mellon University, Pittsburgh, PA, USA. E-mail: [email protected] Copyright © 2013 John Wiley & Sons, Ltd. Journal of Behavioral Decision Making, J. Behav. Dec. Making, 27: 328–339 (2014) Published online 4 November 2013 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/bdm.1809 TWO KEY ECOLOGICAL PROPERTIES: VARIABILITY AND VALUE OF PAYOFFS To understand how observed payoff variability and values might affect information search and choice, one first needs to consider the decision-maker’s goals. A considerable amount of research has been devoted to how people make choices based on their experiences (e.g., Bogacz, Brown, Moehlis, Holmes, & Cohen, 2006; Busemeyer & Townsend, 1993; Erev & Barron, 2005; Gonzalez & Dutt, 2011; Hertwig, in press; Tsetsos, Chater, & Usher, 2012). Building on the classic ideas of prospect theory (Kahneman & Tversky, 1979), decision-maker’s goals during choice are generally assumed to be related to the value and the variability of observed payoffs. When deciding between options, people tend to choose the one where they observed higher values (e.g., Busemeyer, 1985; Gonzalez & Dutt, 2012; Tsetsos et al., 2012). Variability in the observed payoffs can affect this value maximization in different ways. Some have concluded that overall it reduces peoples’ reliance on their experiences, by moving choice toward a random choice (the “payoff variability effect”: Busemeyer & Townsend, 1993; Erev & Barron, 2005; Myers & Sadler, 1960). Others have argued that variability moves preferences away from the risky option, resulting in risk aversion (e.g., shown in the “hot stove effect”: Denrell & March, 2001). Recent evidence suggests that payoff variability can cause risk aversion as well as risk seeking behavior, depending on whether high or low outcomes are more salient in the choice environment (Tsetsos et al., 2012). People’s goals during information search are less well understood than during choice (Gonzalez & Dutt, under review). It has been argued that during information search, people explore the different options at random with the single purpose of gathering information (Erev, Ert, & Yechiam, 2008; also see Cohen et al., 2007). This idea of random search is contradicted by evidence from Hills and Hertwig (2010, 2012), who found that people consistently followed one of two distinct exploratory strategies during information search, and by evidence from Gonzalez and Dutt (2011, 2012), who found that the alternation between options during search decreased over time. But could search also be affected by the properties of the observed outcomes? Recent studies suggest this might be the case. Rakow and Newell (2010) proposed that even if people do not receive the observed outcomes’ rewards during search, they might use the outcomes to test their hypotheses about the different options. Such hypotheses might concern the observed variability and value of the outcomes. Lejarraga et al. (2012) found that people tend to search more in options where they observed variable outcomes, relative to options with stable outcomes, suggesting a preference for risky options during information search. Gonzalez and Dutt (2011, 2012) showed that people tend to search more in options with higher observed values than in options where lower values were observed, suggesting a confirmatory search for the maximizing option. While there is thus some evidence for the possible effects of observed variability and value, open questions remain about whether and how these two factors jointly influence search and choice. On the basis of the literature described earlier, we will focus on the following four questions. First, is the total amount of information search related to the variability and value of the observed outcomes? Second, are peoples’ preferences during information search related to the observed variability and value? Third, does the observed variability and value affect the chance of correctly identifying the objectively maximizing option? Fourth, how do variability and value, as observed during search, affect final choice? In a first step toward answering these questions, we reanalyze a large public dataset of a binary choice task. Subsequently, we present a novel experiment designed to test the possible role of the objective difference between the options. REANALYSIS OF THE TECHNION PREDICTION TOURNAMENT DATASET We reanalyzed a dataset from the Technion Prediction Tournament (TPT, Erev et al., 2010), which is especially suited for this purpose because it is a large and representative dataset, using the sampling paradigm (Hertwig et al., 2004). In this paradigm, search and choice are separated into two distinct phases. In an initial sampling phase, participants are free to explore the different options without explicit costs or consequences for as long as they want. Only after they feel satisfied with the obtained information, they proceed to the decision phase where they make a consequential choice between the options (Figure 1). The set of problems used in the TPT was large and representative of a broad diversity of decision situations. Each of 80 participants solved 30 out of 120 choice problems. All problems consisted of a safe option yielding a fixed medium outcome on every draw, and a risky option yielding a high outcome with some probability p, and a low outcome with the complementary probability 1-p, as shown in Figure 1. The problems were generated with a problem selection algorithm, which ensured that for each participant equal Figure 1. The sampling paradigm as used in the TPT dataset (Erev et al., 2010). During the sampling phase, participants obtained information about two options through a self-structured and selfterminated search without explicit costs. In the decision phase, they selected the option fromwhich a single drawwould be made for determining the problems’ payoff. In each problem, one option was “safe” with a medium payoff and a probability of 1 (8 in the example), while the other option was “risky” with a high payoff with probability p and a low payoff with probability 1-p, (17.1 with p= .1 and 6.9 with p= .9 in the example) Information Search and Decisions from Experience 329 K. Mehlhorn et al. Copyright © 2013 John Wiley & Sons, Ltd. J. Behav. Dec. Making, 27, 328–339 (2014)