Do Maximizers Predict Better than Satisficers? - PDF Document

Presentation Transcript

  1. Do Maximizers Predict Better than Satisficers? _______________ Kriti JAIN J. Neil BEARDEN Allan FILIPOWICZ 2011/18/DS/OB

  2. Do Maximizers Predict Better than Satisficers? Kriti Jain * J. Neil Bearden** Allan Filipowicz*** * PhD Candidate in Decision Sciences at INSEAD 1, Ayer Rajah Avenue, Singapore 138676, Singapore. Email: Kriti.jain@insead.edu Corresponding author. ** Assistant Professor of Decision Sciences at INSEAD 1, Ayer Rajah Avenue, Singapore 138676, Singapore. Email: neil.bearden@insead.edu *** Assistant Professor of Organisational Behaviour at INSEAD 1, Ayer Rajah Avenue, Singapore 138676, Singapore. Email: allan.filipowicz@insead.edu A Working Paper is the author’s intellectual property. It is intended as a means to promote research to interested readers. Its content should not be copied or hosted on any server without written permission from publications.fb@insead.edu Click here to access the INSEAD Working Paper collection

  3. Abstract We examined the relationship between maximizing (i.e. seeking the best) versus satisficing (i.e.seeking the good enough) tendencies and forecasting ability in a real-world prediction task: forecasting the outcomes of the 2010 FIFA World Cup. In Studies 1 and 2, participants gave probabilistic forecasts for the outcomes of the tournament, and also completed a measure of maximizing tendencies. We found that although maximizers expected themselves to outperform others much more than satisficers, they actually forecasted more poorly. Hence, on net, they were more overconfident. The differences in forecasting abilities seem to be driven by the maximizers’ tendency to give more variable probability estimates. In Study 3, participants played a betting task where they could select between safe and uncertain gambles linked to World Cup outcomes. Again, maximizers did more poorly and earned less, because of a higher variance in their responses. This research contributes to the growing literature on maximizing tendencies by expanding the range of objective outcomes over which maximizing has an influence, and further showing that there may be substantial upside to being a satisficer. KEYWORDS: Maximizing; Satisficing; Forecasting; Predictions; Overconfidence

  4. Maximizing vs. Satisficing and Accuracy Do Maximizers Predict Better than Satisficers? Simon (1955, 1993) proposed satisficing as a descriptive alternative to the normative maximizing objective that guides the behavior of neo-classical agents. According to him, actual people are more apt to search for something that is good enough (i.e., that satisfies by sufficing) than they are to try to find the thing that is the absolute best (i.e., that maximizes). More recently, Schwartz and colleagues argued that individuals varied stably in their tendency to maximize vs. satisfice (Schwartz, Ward, Monterosso, Lyubomirsky, White, & Lehman, 2002), with more recent literature examining how this individual difference leads to differences in objective outcomes (Bruine de Bruin, Parker, & Fischhoff, 2007; Iyengar, Wells, & Schwartz, 2006; Polman, 2010). Following Simon’s original emphasis of satisficing involving choice behavior, this empirical work on individual differences in satisficing tendencies has focused on choice and decision behavior. In the current paper, we examine whether satisficing (versus maximizing) tendencies are associated with judgment quality. In particular, we test whether satisficers or maximizers do better in a forecasting task with true exogenous uncertainty: predicting the outcomes of the 2010 FIFA World Cup. Where one stands on the satisficing - maximizing continuum has an impact on decision making outcomes. Maximizers report making more poor decisions based on a self-report decision outcome inventory (Bruine de Bruin et al., 2007; Parker et al., 2007). Polman (2010) found that maximizers simply seek and chose more alternatives, ending up with both better and worse decisions than satisficers. The maximizers in Polman's study reported making worse decisions, as assessed by Bruine de Bruin's (negative) Decision Outcome Inventory, but they also reported making better decisions on a measure containing items relating to positive decisions. And in an impressive demonstration of the objective benefits of maximizing, Iyengar et al. 3

  5. Maximizing vs. Satisficing and Accuracy (2006) found that graduating students who scored higher on Schwartz et al.’s (2002) maximizing measure secured higher paying jobs than did lower scoring students. While data is accumulating that maximizing tendencies influence decision making, several researchers have asked about the breadth of behavior influenced by maximizing. Schwartz noted that "it remains to be determined whether maximizers also consistently act differently from satisficers" (2002, p. 1195), while Iyengar wondered whether maximizing tendencies were global individual difference measures or "simply a set of learned behaviors or search strategies designed specifically for decision-making tasks, and not necessarily even all decision-making tasks" (2006, p. 148). We explore the breadth of impact of this individual difference by looking at an important antecedent of decision making, probabilistic forecasting, as well as a related judgment task, estimating one's relative performance on those probabilistic forecasts. Accurate probabilistic forecasting, the ability to correctly assign high probabilities to events that will occur, and low probabilities to events that will not, is essential to effective decision making. Consider the most proximal example, sports betting, an activity with millions of participants (e.g. in the UK alone, the turnover from betting on the FIFA 2010 World Cup was estimated to have been between 1 and 3 billion pounds). Obviously, successfully assessing odds is crucial to good (long-term) performance. In a health context, a physician with highly inaccurate assessments of outcome likelihoods would have difficulty recommending alternative courses of treatment. And in an organizational context, correctly assessing the probability that interest rates (or the stock market, or currency exchange rates, or even product demand) will move in a certain direction, would allow one to adequately take advantage of and also protect against such movement. 4

  6. Maximizing vs. Satisficing and Accuracy In contrast to the positive relationship between maximizing and decision making outcomes (Iyengar et al., 2006; Polman, 2010), two preliminary findings hint that maximizers might be worse at making accurate probabilistic forecasts. Bruine de Bruin and colleagues (2007) argued that maximizers have worse decision making processes overall, based on maximizers' lower scores on a seven-scale behavioral decision making competence inventory. Most relevant to probabilistic forecasting, maximizers did worse on the consistency of risk perceptions sub-scale. In this subscale, maximizers were more likely to assign higher probabilities to events happening in the next year than to events happening in the next five years (i.e. happening in the next year plus four other years), and assigning higher probabilities to events in a subset (e.g. dying in a terrorist attack) than to events in the corresponding superset (e.g. dying from any cause). Second, and related to the lack of consistency of risk perceptions finding, maximizers tend to show higher variance in responding (Polman, 2010). Polman showed that maximizers alternated more between decks in the Iowa Gambling Task (Bechara et al., 1994), and that this inflated response variability drove down their earnings, as their switching increased their sampling from the “bad deck.” In making probabilistic forecasts, higher response variance that is driven by superior discriminating skills, e.g. by a forecaster who can discriminate precisely between teams that win their matches and those that do not, will not affect the accuracy of the forecasts. All other sources of variance, for example variability driven by “needlessly maximizing”, as Polman observed, would lead to worse forecasting accuracy. In the studies described below, we examine whether maximizing tendencies are related to the objective outcome of probabilistic forecasting accuracy, and test the effect of variance of responding as a possible mechanism linking maximizing tendencies to forecasting accuracy. 5

  7. Maximizing vs. Satisficing and Accuracy And since variance of responding driven by superior discriminating skills should not affect the accuracy, we look both at overall response variability, and at the unnecessary variability left after factoring out the variability that can arise from superior discriminating skills. Given that we can measure how well participants did on the forecasting task, we are also able to examine a related and well-documented judgment bias, people’s tendency to overestimate their relative performance. In one classic demonstration, Svenson (1981) asked drivers to estimate their driving abilities (relative to others in the experiment), and found that 93% of a US sample and 69% of a Swedish sample put themselves in the top 50% of drivers. This pattern of self-enhancing judgment has been observed across a range of tasks with a wide variety of samples (e.g., Zuckerman and Jost, 2001), and has been dubbed the better-than-average effect (Alicke et al., 1995). Yet it remains an open empirical question whether maximizers would overestimate their relative performance. Bruine de Bruin and colleagues' (2007) behavioral measure of decision making competence included an underconfidence/overconfidence scale, which assessed participants' abilities to recognize the extent of their knowledge. Maximizers were less aware of the extent of their knowledge, but the scale does not allow one to adjudicate between underconfidence (respondents do better than they thought) and overconfidence (respondents do worse than they thought). Maximizers self-report less life satisfaction, optimism and self-esteem and report more regret and depression (Schwartz et al., 2002), which suggests that they may bias their self evaluations downward. However, since maximizers have highest standards for themselves and prefer not to settle for second-best (see sample items from the maximizations scale by Diab, Gillespie, and Highhouse, 2008, which we present below), it is also reasonable to expect that they might bias their self evaluations upwards. But overconfidence is a combination 6

  8. Maximizing vs. Satisficing and Accuracy of self-evaluation (how one thinks one did) and objective measures (how one actually did), making the net effect indeterminate. With no a priori hypotheses, we carried out a pilot study to test the relationship between maximizing and the better-than-average effect, before studying overconfidence in the context of probabilistic forecasting. We report below on this pilot study and three subsequent experiments. The pilot study looked at the relationship between maximizing tendencies and the better-than-average effect. Study One tested whether maximizing tendencies were associated with actual forecasting performance, and the extent to which this relationship was driven by differences in response variability. Given the results of the pilot study, we also examined the relationship between maximizing tendencies and overconfidence. Study Two replicated Study One, but with a time lag on the forecasting data, by using a second wave of predictions gathered 2 weeks later. In Study Three we then examined whether maximizing tendencies were linked to outcomes in a decision making (betting) task that implicitly required judgmental forecasts, and again tested the mediating effect of response variability. Apart from the pilot study, all studies were conducted just prior to or during the 2010 FIFA World Cup, and all of the forecasts and decisions were linked to World Cup outcomes. Using tasks linked to real-world events with truly uncertain outcomes allows us to better assess the degree to which maximizing might be linked to judgment in uncertain economic settings – compared to general knowledge type tasks that involve only epistemic uncertainty. PILOT STUDY To examine whether a relationship exists between maximizing tendencies and overconfidence, we included a better-than-average type question and a measure of maximizing 7

  9. Maximizing vs. Satisficing and Accuracy tendency as a filler task in another, unrelated study. Two-hundred subjects completed the Maximizing Tendency Scale (Diab, Gillespie, and Highhouse, 2008), and also estimated the percentage of participants completing the survey who would have driving skills inferior to their own. In line with the results reported by Svenson (1981), 75% of the respondents placed themselves above the median (50%) in driving ability. More importantly, their judgments of relative driving skill were positively related to their score on the maximizing scale (r = 0.21, p < 0.01). In short, the maximizers showed a greater degree of better-than-average effect than did satisficers in the pilot study, suggesting that we should also examine overconfidence in our studies of probabilistic forecasting. STUDIES 1 AND 2: WORLD CUP PREDICTION TASK We will report the results from Studies 1 and 2 together. In both, subjects made probabilistic forecasts for a number of outcomes of the 2010 FIFA World Cup. Study 1 was conducted in the week prior to the start of the tournament, while Study 2 was conducted in the roughly 1 day window between the Round of 16 and Quarter-finals. Both studies employed incentive-compatible payoffs and standard experimental economics protocols (e.g., we did not use any deception). Method Participants and Procedure Eleven-hundred and ten subjects (906 males, 204 females), recruited from the INSEAD and Singapore communities, took part in Study 1. Thirty-seven percent of these participants reported that they would be betting on the World Cup results using real money. Each participant was informed that five out of every 100 participants in the study would be selected at random 8

  10. Maximizing vs. Satisficing and Accuracy and paid up to SGD 100 based on his or her own individual performance. The subjects in Study 2 were recruited from those who participated in Study 1. Four-hundred and thirty-seven subjects took part in Study 2. The incentive scheme used in Study 2 was the same one used in Study 1, except that one out of every 100 participants was selected at random for payment. In total, we paid out nearly SGD 5000 (USD 3800) in performance-based incentives (including Study 3, which we describe later). Measures Maximization Scale We used Diab et al.’s (2008) nine item Maximizing Tendency Scale (hereafter MT) to measure the maximizer construct. Diab et al. showed that the scale has a greater reliability and internal consistency than the original scale developed by Schwartz et al. (2002). The MT items are relatively straightforward and have very high face validity. Some representative items are: −“No matter what I do, I have the highest standards for myself.” −“I never settle for second best.” −“I never settle.” Respondents rated the items on a standard 5-point Likert-type scale (1 = strongly disagree to 5 = strongly agree). Individual item ratings were then summed to create a single, scalar composite score (M = 30.82; SD =5.29; α = 0.84). Higher values correspond to a greater tendency to maximize. World Cup Predictions In the week prior to the tournament, each subject in Study 1 made forecasts for 20 of the 32 World Cup teams. The twenty teams were selected randomly and independently for each subject. We used a subset of the 32 teams to keep the time required to complete the study 9

  11. Maximizing vs. Satisficing and Accuracy manageable (around 40 minutes). For each of the 20 teams, the subject estimated the probability of the team making it to each of the following five stages of the tournament: 1. Round of 16 (where 16 of the 32 teams remain) 2. Quarter-Finals (where 8 of the 32 teams remain) 3. Semi-Finals (where 4 of the 32 teams remain) 4. Finals (where 2 of the 32 teams remain) 5. Overall Winner (where 1 of the 32 teams remain) Study 2 was run just after the Round of 16 and prior to the first Quarter-Finals match. In it, each subject gave four probabilistic forecasts for each of the remaining 16 teams: the chances of the teams making it to the Quarter-Finals, Semi-Finals, Finals, and of winning the tournament. In both studies, the subjects expressed their judgments by selecting one of ten response categories: 0-10%, 11-20%, …, 91-100%. For each of the two studies separately, subjects also estimated the percentage of participants they would perform better than in the forecasting task (we call this Estimated Standing). This question was asked after the subjects had made all of their team forecasts. Performance Measures Using the forecast probabilities, we computed Brier scores as well as its two components, calibration and resolution, for each subject (Brier, 1950; Murphy, 1973). We used the mid-point of the selected probability interval to compute the scores. For instance, for the 0-10% interval, we used 5%; for the 11-20% interval, we used 15%; and so on. Higher Brier scores indicate poorer accuracy, and higher calibration scores indicate poorer calibration. Higher resolution scores, on the other hand, indicate greater (better) resolution. The overall Brier score will serve 10

  12. Maximizing vs. Satisficing and Accuracy as our primary performance measure. The scheme under which the subjects gave their estimates was based on the Brier score and was therefore incentive-compatible (Winkler, 1969). We computed an overconfidence measure based on the difference between estimated relative performance (Estimated Standing) and actual relative performance (Actual Standing). For example, if a subject estimated that he would do better than 85% of the forecasters and he actually did better than only 30%, then his score would be 85 – 30 = 55. For shorthand, we will refer to this measure as OC. We also measured how the subjects used the response scale, that is, how they assigned probabilities. In particular, we wanted to measure the degree of variance in the assigned probabilities. Based on the results from Polman (2010), we conjectured that the maximizers’ probabilities would have greater variance. We therefore computed a composite measure of response variability for each subject. Since the average assigned probabilities tend to decrease with the rounds of the tournament (as there are fewer and fewer slots left at later stages), for each subject we computed the variance in assigned probabilities separately for each round (since the grand mean would not be representative across rounds). We then averaged these variances across rounds to form a single measure of response variance for each subject. We will denote the composite measure by VAR. Results Tables 1 and 2 shows zero-order correlations between maximizing tendency (MT), Brier score, calibration score, resolution score, estimated standing, actual standing, overconfidence (estimated standing – actual standing), and VAR. The results demonstrate that MT had a significant positive correlation with both the Brier score (Study 1: r =0.08, p < 0.01; Study 2: r = 0.12, p < 0.05) and the calibration score (Study 1: r = 0.10, p < 0.001; Study 2: r = 0.14, p < 11

  13. Maximizing vs. Satisficing and Accuracy 0.001), but no significant relationship with the resolution score. Put differently, people who scored higher on MT made poorer predictions. For ease of visualization, we compare the bottom 20% of scorers on the MT scale (MT ≤ 27) with the top 20% of scorers (MT ≥ 35), and refer to the two groups as satisficers and maximizers, respectively. Figure 1 shows the calibration curves for the two groups in Studies 1(top panel) and 2 (bottom panel). Notice that for both groups, the subjective estimates were less than the objective relative frequency – indicating miscalibration for most of the range – but that the degree of miscalibration was greater for the maximizers in both studies. Overall, people significantly overestimated their relative performance in both Study 1 (OC: M = 12.13, SD = 33.18, t(1105) = 12.16, p < 0.001), and Study 2 (OC: M = 10.26, SD = 34.37, t(436) = 6.24, p < 0.001). And MT was significantly correlated with OC in both studies (Study 1: r = 0.18, p < 0.001; Study 2: r = 0.17, p < 0.001). Put differently, higher MT scorers were more overconfident about their relative performance. The observed relationship between OC and MT can be better appreciated by examining the relationship between MT and the two component parts of the OC score: MT is positively correlated with estimated standing (Study 1: r = 0.17, p < 0.001; Study 2: r = 0.10, p < 0.05), but negatively related to actual standing (Study 1: r = -0.09, p < 0.01; Study 2: r = -0.14, p < 0.01). Higher MT scorers thought they would perform better, but in fact performed worse. To get a better handle on what might be driving the observed differences in forecasting performance, we looked at the relationship between response variability (measured by VAR) and performance. As anticipated, there is a significant positive correlation between MT and VAR in both studies (Study1: r = 0.19, p < 0.001; Study 2: r = 0.16, p < 0.001). To make clear the strength of the relationship between MT and VAR, the line chart in Figure 2 show the average 12

  14. Maximizing vs. Satisficing and Accuracy VAR by MT for each study. The bars in the figures represent the proportion of subjects in the respective bin with an above median VAR (taken over all values of MT). Clearly, participants with a higher maximizing tendency also had more variable probability estimates. Mediation Analysis Next, we examine whether the observed relationship between maximizing tendency and performance (Brier score) is mediated by the tendency to have more or less variable probability estimates (VAR). Figure 3 shows the hypothesized mediation. For Study 1, all three of Baron and Kenny’s (1986) preconditions for mediation were met. The predictor (MT) predicts the outcome (Brier score) (β = .08, p < 0.01, R2 = .01). The predictor (MT) predicts the mediator (VAR) (β = 0.19, p < 0.001, R2 = 0.03). And when both MT and VARare included, the coefficient on MT becomes non-significant (β = 0.01, p = 0.67) but the coefficient on VARremains significant (β = 0.38, p < 0.001). Hence, VARfully mediates the effect of MT on the Brier score (Sobel’s z = 5.63, p < 0.01). Similarly, for Study 2, all three of Baron and Kenny’s (1986) preconditions for mediation were met. The predictor (MT) predicts the outcome (Brier score) (β = 0.12, p < 0.01, R2 = 0.01). The predictor (MT) predicts the mediator (VAR) (β = 0.16, p < 0.01, R2 = 0.02). And when both MT and VARare included, the coefficient on MT becomes non-significant (β = 0.05, p = 0.22), while the coefficient on VAR is still significant (β = 0.43, p < 0.001). Again, there is full mediation (Sobel’s z =3.21, p < 0.01). Robustness Check An alternative measure of response variability. It is easy to show that having high response variability is not necessarily bad for performance. In fact, a forecaster who can discriminate precisely between teams that win their matches from those that do not would have 13

  15. Maximizing vs. Satisficing and Accuracy accurate forecasts and also have a high VAR. However, our results in Tables 1 and 2 show that maximizers had a higher VAR but no better resolution scores (a measure of discrimination). For robustness, we computed another measure of response variability related to discrimination, a measure of excess “scatter” in the probabilities from the Yates’ (1982) decomposition of the Brier score (see also Yates and Curley, 1985). This measure is the weighted average of the conditional variances of the probabilities given wins and given loses. It can be thought of as the unnecessary variability left after factoring out the variability that can arise from superior discriminating skills. Therefore a higher score on this measure necessarily implies worse overall performance (i.e. a high Brier score). In both studies, MT had a significant positive correlation with this scatter score (Study 1: r = 0.15, p < 0.01; Study 2: r = 0.15, p < 0.01). Therefore, also with this alternative measure, maximizers showed higher response variability. Using scatter score as a measure of variability, we tested whether the maximizer- performance relationship was mediated by the higher response variability. Again the three Baron and Kenny (1986) preconditions were satisfied. The predictor (MT) predicts the outcome (Brier Score) (Study 1: β = 0.08, p = 0.01, R2 = 0.01; Study 2: β = 0.12, p = 0.01, R2 = 0.01). The predictor (MT) predicts the mediator (scatter score) (Study 1: β = 0.15, p < 0.001, R2 = .02; Study 2: β = .15, p = 0.01, R2 = 0.02). And when both MT and scatter score are included, coefficient of MT becomes non-significant (Study 1: β = -0.01, p = 0.75; Study 2: β = 0.02, p = 0.53) and the coefficient of scatter score remains significant (Study 1: β =0.59, p < 0.01; Study 2: β = 0.63, p < 0.01). Therefore, we find that the scatter score mediated the MT-Brier score in both Study 1 (Sobel’s z = 5.13, p < 0.001) and Study 2 (Sobel’s z = 3.16, p < 0.01). Hence, again the higher response variability seems to be playing a role in the MT-Brier Score relationship: 14

  16. Maximizing vs. Satisficing and Accuracy Individuals with higher MT scores tend to show larger response variability, and this effect drives their poorer forecasting performance. Discussion Both Studies 1 and 2 unambiguously show that maximizing tendencies are linked to poorer prediction performance in the World Cup forecasting task. The performance differences associated with maximizing tendencies are fully mediated by response variability. Once we control for variability in responding (which is greater among the maximizers), the performance difference vanishes. Further, we also found strong evidence for increased overconfidence among maximizers: The maximizers predicted they would do relatively better, but in fact they did relatively worse. Next, we test whether the observed differences in forecasting performance affect outcomes in a betting (decision) task. STUDY 3: WORLD CUP BETTING TASK Studies 1 and 2 show that maximizers performed more poorly than satisficers on a probabilistic forecasting task. Here, we examine whether there are concomitant differences in performance on a decision (betting) task. In the betting task, subjects must choose between uncertain gambles whose payoffs are linked to World Cup outcomes. Method Participants and Procedure Subjects who took part in Study 1 were invited again to take part in a betting task two days prior to the start of the first World Cup match. Five-hundred and eleven subjects agreed to take part. As in Studies 1 and 2, we used incentive-compatible payoffs: we selected one out of 15

  17. Maximizing vs. Satisficing and Accuracy every 100 subjects at random and paid them according to their actual betting decisions and outcomes. We paid a total of SGD 150 to a total of five participants. Measures World Cup Betting Task We used a modification of the Holt-Laury (Holt and Laury, 2002) risky choice task to measure decision making under uncertainty. For each of the 32 World Cup teams, the subjects were shown a series of lotteries consisting of a safe and an uncertain option. The safe option offered a sure-thing payoff, and the uncertain option always paid off $100 if and only if the team made it to the Round of 16. Ten gambles were shown for each team, with sure-thing payoffs starting at $10 and increasing in units of $10 (up to a $100 sure-thing payoff). An example using France is displayed in Appendix A. Note that a subject should be more likely to take the uncertain option as her subjective probability of the team advancing to the Round of 16 increases. Risk Preferences One can also suppose that risk attitudes should play a role in the subjects’ choices between the risky and uncertain options. Hence, we also gave the subjects the standard Holt and Laury (2002) risky choice task in order to assess their general risk preferences. The task consists of a series of pairs of risky options in which the subject must choose one option from each pair. The pairs we used are displayed in Appendix B. Notice that the spread of the payoffs for the A options ($40 versus $32) is smaller than the spread of payoffs for the B options ($77 versus $2). Except for the extreme cases where the outcomes are no longer uncertain, the variance of the payoffs for B is greater than the variance for A. A more risk averse person will tend to prefer more of the A options relative to a less risk averse person. Hence, our measure of risk aversion – 16

  18. Maximizing vs. Satisficing and Accuracy following Holt and Laury (2002) – is simply the number of A options selected by the subject. (The Holt-Laury risk measure will be useful below when we examine the relationship between maximizing tendency and average earnings.) Performance Measures We computed the average earnings (henceforth referred to as earnings) for each subject based on his bet choices over all 32 teams. In addition, we computed an analogue of the VAR measure used in Studies 1 and 2. For each of the 32 teams, we calculated the number of risky choices. The variance of the number of risky choices is our measure of response variability, and we refer to it as VAR. Note that, computed this way, VAR in Study 3 is entirely analogous to the VAR measure used for probability estimates in Studies 1 and 2. Results and Discussion Table 3 shows the correlations of earnings and VAR from Study 3 with maximizing tendency (MT), risk aversion (from the Holt-Laury task); and it also has the correlations between the Study 3 measures and the Brier scores and VARfrom Studies 1 and 2. (Note the strong correlation between Brier scores across rounds.) Forecasting performance from Studies 1 and 2 is indeed linked to earnings in Study 3: earnings were negatively correlated with Brier scores (Study1: r = -0.14, p = 0.001; Study 2: r = -0.22, p < 0.001). Further, earnings were negatively correlated with MT (r = -0.13, p < 0.01). Maximizers earned less in the betting task. For visualization, in Figure 4, we show the cumulative distributions of earnings for the top 20% of the scorers on the MT scale (maximizers) and the bottom 20% of the scorers (satisficers). The difference between the groups is stark: in fact, the distribution of earnings for the satisficers first- order stochastically dominates the distribution of earnings for the maximizers. 17

  19. Maximizing vs. Satisficing and Accuracy Again, maximizers had a higher response variability, i.e. VAR was positively correlated with MT (r = 0.23, p < 0.01). Also note the strong correlations between the VAR scores across all three studies. As in Studies 1 and 2, we examined whether VAR mediated the maximizing- performance relationship. Importantly, since earnings are correlated with risk aversion (see Table 3), we must control for risk aversion in the analyses of the betting results. We first regressed earnings onto MT while controlling for risk aversion, and found a significant negative relationship between MT and earnings (β = -0.13, p < 0.01, see Model 1 in Table 4). That is, even after accounting for differences in risk preferences, we find that maximizing is negatively related to earnings. Next, we ran another model with MT, the risk aversion measure, and also VAR as independent variables, and earnings as the dependent variable (Model 2, Table 4). With VAR in the model, the coefficient on MT becomes non-significant (β = -0.03, p = 0.44), while the coefficient on VAR is significant (β = -0.42, p < 0.01). Again, we find that controlling for the variability in responding – which is greater among the maximizers – causes the relationship between maximizing tendency and earnings to go away. Robustness Check As in Studies 1 and 2, we also calculated the scatter score, by taking the weighted average of the conditional variances of the number of risky choices given wins and given loses. As predicted, this scatter score was significantly positively correlated with MT (r = 0.24, p < 0.01) and negatively correlated with earnings (r = -0.54, p < 0.01). Next, we examined whether scatter score mediated the maximizing-performance relationship. This time, we ran a regression with MT, the risk aversion measure, and scatter score as independent variables, and earnings as the dependent variable (Model 3, Table 4). Again, the coefficient on MT becomes non- significant (β = -0.00, p = 0.93), while the coefficient on scatter score is significant (β = -0.54, p 18

  20. Maximizing vs. Satisficing and Accuracy < 0.01). Therefore, we find that this alternative measure of response variability again mediates the relationship between maximizing tendency and earnings. GENERAL DISCUSSION We have shown that individuals who score higher on a self-report measure of maximizing tended to perform more poorly in forecasting the outcomes of the 2010 FIFA World Cup (Studies 1 and 2) and also to earn less in a betting task related to the World Cup outcomes (Study 3). Further, maximizers tended to overestimate their relative forecasting performance to a greater extent than did satisficers (Studies 1 and 2). Importantly, these tasks have many features in common with problems faced by people in genuine economic settings, the type of settings that motivated Simon’s critique of the maximizing objective of neo-classical economics. Most notably, the World Cup forecasting problems involve tremendous genuine aleatory uncertainty – the kind that one finds in asset markets, currency markets, and so on. Glaser and Weber (2007) found that investors who believe themselves above average in terms of investment acumen tend to trade more frequently, and higher volume trading often tends to reduce returns due to the increased transaction costs (see, e.g., Barber and Odean, 2002; Odean, 1998). Hence, with some irony, there is reason to anticipate that maximizers might do less well in some economic settings. Across three studies, we also showed that the negative relationship between maximizing tendencies and both forecasting performance and decision making outcomes on a betting task was driven by differences in response variability. Maximizers had a greater response variability – their probability judgments were more dispersed over the [0, 1] interval – and that led to poorer performance. This is analogous to Polman's (2010) finding, discussed earlier, that maximizers alternated more frequently in the Iowa Gambling Task and in doing so negatively affected their average earnings. It seems reasonable to conjecture that in settings like these (including ours) 19

  21. Maximizing vs. Satisficing and Accuracy that involve judgments about exogenous uncertainty, maximizers’ pursuit for the elusive “best” causes them tremendous anxiety and worriment and this then gets manifested in their higher response variability. Future research could test this emotional proximal cause of maximizers’ response variability. Our findings expand the range of objective outcomes over which maximizing tendencies have an influence. While prior research has focused on decisions making and maximizers' subjective reactions to those decisions, we have shown that maximizing tendencies also influence an important antecedent to decision making, probabilistic forecasting, and a related subjective appraisal, overconfidence. We have also started to examine a potential mechanism through which maximizing tendencies hamper forecasting accuracy, increased response variability. In answering Schwartz's query of whether maximizing tendencies make a difference, we can more confidently say yes, in more ways than one could imagine. 20

  22. Maximizing vs. Satisficing and Accuracy REFERENCES Alicke, M. D., Klotz, M. L., Breitenbecher, D. L., Yurak, T. J., & Vredenburg, D. S. (1995). Personal contact, individuation, and the better-than-average effect. Journal of Personality and Social Psychology, 68, 804-825. Barber, B. M., & Odean. T. (2002). Trading is hazardous to your wealth: The common stock investment performance of individual investors. Journal of Finance, 55, 773– 806 . Baron, R. M., & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51, 1173-1182. Bechara, A., Damasio, A. R., Damasio, H., & Anderson, S. W. (1994). Insensitivity to future consequences following damage to human prefrontal cortex. Cognition, 50, 7–15. Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78, 1–3. Bruine de Bruin, W., Parker, A. M., & Fischhoff, B. (2007). Individual differences in adult decision-making competence. Journal of Personality and Social Psychology, 92, 938– 956. Diab, D. L., Gillespie, M. A., & Highhouse, S. (2008). Are maximizers really unhappy? The measurement of maximizing tendency. Judgment and Decision Making, 3, 364-370. Glaser, M., & Weber, M. (2007). Overconfidence and trading volume. Geneva Risk and Insurance Review,32, 1-36. Holt, C. A., & Laury, S. K. (2002). Risk aversion and incentive effects. American Economic Review, 92, 1644–1655. 21

  23. Maximizing vs. Satisficing and Accuracy Iyengar, S., R. Wells & B. Schwartz. (2006). Doing better but feeling worse. Psychological Science, 17, 143-150. Murphy, A.H. (1973). A new vector partition of the probability score. Journal of Applied Meteorology, 12, 595–600. Odean, T. (1998). Volume, volatility, price and profit when all traders are above average. Journal of Finance, 53, 1887–1934. Parker, A. M., de Bruin, W. B., & Fischhoff, B. (2007). Maximizers versus satisficers: Decision- making styles, competence, and outcomes. Judgment and Decision Making, 2, 342–350. Polman, E. (2010). Why are maximizers less happy than satisficers? Because they maximize positive and negative outcomes. Journal of Behavioral Decision Making, 23, 179-190. Schwartz, B., Ward, A., Monterosso, J., Lyubomirsky, S., White, K., & Lehman, D. R. (2002). Maximizing versus satisficing: Happiness is a matter of choice. Journal of Personality and Social Psychology, 83, 1178–1197. Simon, H. A. (1955). A behavioral model of rational choice. Quarterly journal of Economics, 59, 99-118. Simon, H. A. (1993). Satisficing. In D. Greenwald (Ed.), The McGraw-Hill Encyclopedia of Economics (2nd ed., pp. 881-886). New York: McGraw-Hill, Inc. Svenson, O. (1981). Are we all less risky and more skillful than our fellow drivers? Acta Psychologica, 47, 143–148. Winkler, R. L. (1969). Scoring rules and the evaluation of probability assessors. Journal of the American Statistical Association, 64, 1073–1078. Yates, J. F. (1982). External correspondence: Decompositions of the mean probability score. Organizational Behavior and Human Decision Processes, 30, 132-156. 22

  24. Maximizing vs. Satisficing and Accuracy Yates, J. F. & Curley, S. P. (1985). Conditional distribution analyses of probability forecasts. Journal of Forecasting, 4, 61-73. Zuckerman, E. W., & Jost, J. T. (2001). What makes you think you’re so popular? Self evaluation maintenance and the subjective side of the “friendship paradox.” Social Psychology Quarterly,64, 207-223. 23

  25. Maximizing vs. Satisficing and Accuracy TABLE 1 Correlation from World Cup Predictions in Study 1 Maximizing Brier Calibration Resolution Estimated Actual Overcon Response Mean Tendency Score Standing Standing fidence Variance (Standard (MT) (OC) (VAR) Deviation) Maximizing Tendency (MT) 1 30.82 (5.29) Brier Score 0.08** 1 0.13 (0.03) Calibration 0.10** 0.80** 1 0.03 (0.03) Resolution -0.01 -0.46** -0.28** 1 0.06 (0.02) Estimated Standing 0.17** -0.07* 0.03 0.20** 1 62.11 (19.72) Actual Standing -0.09** -0.89** -0.61** 0.43** 0.11** 1 49.98 (28.86) Overconfidence (OC) 0.18** 0.74** 0.55** -0.26** 0.50** -0.81** 1 12.13 (33.18) Response Variance (VAR) 0.19** 0.38** 0.45** 0.07* 0.22** -0.37** 0.45** 1 0.05 (0.03) **p<.01, *p<.05 24

  26. Maximizing vs. Satisficing and Accuracy TABLE 2 Correlation from World Cup Predictions in Study 2 Maximizing Brier Calibration Resolution Estimated Actual Overcon Response Mean Tendency Score Standing Standing fidence Variance (Standard (MT) (OC) (VAR) Deviation) Maximizing Tendency (MT) 1 30.82 (5.29) Brier Score 0.12* 1 0.16 (0.04) Calibration 0.14** 0.88** 1 0.04 (0.03) Resolution -0.03 -0.67** -0.25** 1 0.06 (0.02) Estimated Standing 0.10* -0.02 0.07 0.15** 1 60.26 (19.46) Actual Standing -0.14** -0.89** -0.72** 0.70** 0.03 1 50.00 (28.87) Overconfidence (OC) 0.17** 0.74** 0.65** -0.50** 0.54** -0.82** 1 10.26 (34.37) Response Variance (VAR) 0.16** 0.44** 0.54** -0.05 0.28** -0.46** 0.54** 1 0.05 (0.03) **p<.01, *p<.05 25

  27. Maximizing vs. Satisficing and Accuracy TABLE 3 Correlation from Betting Task in Study 3 Study 3 Maximizing Risk Study 1 Study 2 Study 3 Study 1 Study 2 Mean Average Tendency Aversion Brier score Brier score VAR VAR VAR (Standard Earnings (MT) deviation) Study 3 Average Earnings 1 650.99 (4.99) Maximizing Tendency (MT) -0.13** 1 30.82 (5.29) Risk Aversion -0.12** -0.01 1 5.26 (2.00) Study 1 Brier score -0.14** 0..08** 0.08 1 0.13 (0.03) Study 2 Brier score -0.22** 0.12* 0.05 0.52** 1 0.16 (0.04) Study 3 VAR -0.43** 0.23** 0.02 0.18** 0.24** 1 8.88 (5.19) Study 1 VAR -0.12** 0.19** -0.03 0.38** 0.23** 0.40** 1 0.05 (0.03) Study 2 VAR -0.16** 0.16** 0.02 0.23** 0.44** 0.43** 0.52** 1 0.05 (0.03) **p<.01 26

  28. Maximizing vs. Satisficing and Accuracy TABLE 4 Regression of Average Earnings on maximizing tendency Dependent Variable Average Earnings Average Earnings Average Earnings (Model 1) (Model 2) (Model 3) Maximizing Tendency (MT) -0.13** -0.03 -0.00 Risk Aversion -0.13** -0.12** -0.11** Variance (VAR) -0.42** Scatter score -0.54** F-statistic 8.53** 42.70** 73.67** (df) (2,508) (3,507) (3,507) Adjusted R2 0.03 0.20 0.30 **p<.01 27

  29. Maximizing vs. Satisficing and Accuracy FIGURE 1 Calibration Curves for Maximizers and Satisficers in Study 1 (top panel) and Study 2 (bottom panel). Note: 45 degree line shows perfect calibration 100% 80% Percentage true 60% 40% 20% 0% 0 20 40 60 80 100 Subjective probability Satisficers Maximizers 100% 90% 80% 70% Percentage true 60% 50% 40% 30% 20% 10% 0% 0 20 40 60 80 100 Subjective Probability (%) Satisficers Maximizers 28

  30. Maximizing vs. Satisficing and Accuracy FIGURE 2 Average VAR (line chart) and Percentage of respondents with VAR> median VAR (bar chart) across various maximizing score categories in Study 1 (top panel) and Study 2 (bottom panel) 0.08 100% Percentage of participants with VAR > 90% 0.07 80% 0.06 Average VAR (line) median VAR (bars) 70% 0.05 60% 0.04 50% 40% 0.03 30% 0.02 20% 0.01 10% 0.00 0% <=25 26-30 31-35 36-40 >40 Maximizing tendency 0.08 100% ith VAR > 90% 0.07 80% 0.06 Average VAR (line) edian VAR (bars) Percentage of participants w 70% 0.05 60% 0.04 50% 40% 0.03 30% m 0.02 20% 0.01 10% 0.00 0% <=25 26-30 31-35 36-40 >40 Maximizing tendency 29

  31. Maximizing vs. Satisficing and Accuracy FIGURE 3 VAR mediated the effect of maximizing tendency on Brier score in Study 1 (top panel) and Study 2 (bottom panel) 0.08** (0.01) Brier score (BS) Maximizing tendency (MT) 0.19** 0.38** Response Variance (VAR) 0.12** (.05) Brier score (BS) Maximizing tendency (MT) 0.44** 0.16** Response Variance (VAR) 30

  32. Maximizing vs. Satisficing and Accuracy FIGURE 4 Cumulative Distribution of Average Earnings for Maximizers and Satisficers in Study 3 1 Probability (Average Earnings < X) 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 350 450 550 650 750 X Satisficers Maximizers 31

  33. Maximizing vs. Satisficing and Accuracy APPENDIX A: Betting Task in Study 3 Option A Option B $10 for sure $100 if France makes it to the Round of 16. $20 for sure $100 if France makes it to the Round of 16. $30 for sure $100 if France makes it to the Round of 16. $40 for sure $100 if France makes it to the Round of 16. $50 for sure $100 if France makes it to the Round of 16. $60 for sure $100 if France makes it to the Round of 16. $70 for sure $100 if France makes it to the Round of 16. $80 for sure $100 if France makes it to the Round of 16. $90 for sure $100 if France makes it to the Round of 16. $100 for sure $100 if France makes it to the Round of 16. 32

  34. Maximizing vs. Satisficing and Accuracy APPENDIX B: Holt-Laury Task for Risk Preferences Option A 10% for $40 or a 90% chance for $32 Option B 10% chance for $77 or a 90% chance for $2 20% for $40 or a 80% chance for $32 20% chance for $77 or a 80% chance for $2 30% for $40 or a 70% chance for $32 30% chance for $77 or a 70% chance for $2 40% for $40 or a 60% chance for $32 40% chance for $77 or a 60% chance for $2 50% for $40 or a 50% chance for $32 50% chance for $77 or a 50% chance for $2 60% for $40 or a 40% chance for $32 60% chance for $77 or a 40% chance for $2 70% for $40 or a 30% chance for $32 70% chance for $77 or a 30% chance for $2 80% for $40 or a 20% chance for $32 80% chance for $77 or a 20% chance for $2 90% for $40 or a 10% chance for $32 90% chance for $77 or a 10% chance for $2 100% for $40 or a 0% chance for $32 100% chance for $77 or a 0% chance for $2 33