«The Prevalence Heuristic: Mistaking What Has Been Chosen for What Will Be Chosen Emily Reit & Clayton R. Critcher University of California, Berkeley ...»
Running head: PREVALENCE HEURISTIC
The Prevalence Heuristic:
Mistaking What Has Been Chosen for What Will Be Chosen
Emily Reit & Clayton R. Critcher
University of California, Berkeley
Correspondence to: Emily Reit, Haas School of Business, 545 Student Services Bldg., #s300i,
Berkeley, CA, 94720; email: firstname.lastname@example.org
Word Count (Introductions and Discussions): 2000
Submitted on: July 6, 2016
Running head: PREVALENCE HEURISTIC
In predicting what others are likely to choose (e.g., vanilla ice cream or tiramisu), people are led astray by a prevalence heuristic—overestimating how often common (but bland) items (e.g., vanilla ice cream) will be chosen over more unusual (but exciting) items (e.g., tiramisu). Given common items are often chosen merely because they are available, not because they are preferred (tiramisu is rarely offered as a dessert), prevalence is not particularly diagnostic of future choice. Studies 1-3 demonstrate the prevalence heuristic and uncover when and why it emerges. Perceived prevalence is spontaneously used as a guide when forecasting others’ choices (suggesting people confuse what has been chosen with what people will choose), but not when forecasting what others would be pleased to receive. Upon conscious reflection, people realize the prevalence heuristic is unwise. A final, two-part marketplace simulation study found reliance on the prevalence heuristic prompts sellers to misprice goods.
KEYWORDS: choice, social judgment, heuristics, perspective taking, theory of mind 3 Running head: PREVALENCE HEURISTIC The Prevalence Heuristic: Mistaking What Has Been Chosen for What Will Be Chosen People are often tasked with predicting others’ choices. A dinner host must decide how many servings of vanilla ice cream vs. tiramisu to have on hand. A store owner must decide how many of the new dress shoes that come in black vs. blue to stock. A florist must decide how many bouquets of daisies vs. dragon snaps should be available in her pop-up flower shop every Sunday.
Knowing others’ preferences is difficult. In making social judgments, a readily accessible guide is the self (Ross, Greene & House, 1977). Although self-knowledge can helpfully inform social knowledge (Krueger, 2003; Dawes & Mulford, 1996), it is an incomplete guide. Social insight requires not only that people know what part of their own preferences are idiosyncratic (“I have to remember not everyone thinks cilantro tasteslike soap”), but also that they actually expend the effort to adjust from their own egocentric perspective (Epley, Keysar, Van Boven & Gilovich, 2004). And other research—especially in the gift giving literature—has shown people display systematic biases when estimating what others prefer to receive. For instance, gift givers overweight thoughtfulness, cost, and uniqueness. In actuality, gift recipients prefer gifts they explicitly ask for, or simply money (Gino & Flynn, 2011).
Although past research documents clear challenges in understanding what others like, we posit there is a special challenge in predicting what others are likely to choose. This differentiation may seem odd. After all, there are not clear normative reasons to differentiate preferences and choice: It is almost axiomatic that people choose what they would prefer to receive. And here, we don’t challenge this tautology. Instead, we suggest that the task of
scrutiny—is likely to be leaned upon when answering the difficult question of “Will people choose A or B?” A hallmark of a heuristic is that it involves attribute substitution, reliance on an imperfectly valid but readily accessible cue when making a difficult, potentially intractable judgment (Kahneman, 2003; Kahneman & Frederick, 2002). We propose that in estimating others’ choices of A or B, people lean on a prevalence heuristic— their intuitive sense of the relative prevalence of A over B. Our proposal rests on the recognition that this attribute is deceptively similar to the question of interest. The greater prevalence of vanilla ice cream over tiramisu does indeed reflect that the plain frozen treat has been chosen to be eaten more often than the Italian delicacy. But this does not imply that people are likely to choose vanilla ice cream over tiramisu when given the choice between the two. After all, vanilla ice cream is not merely a more common dessert choice, it is (for a variety of reasons) more commonly offered as an option to begin with.
Although the prevalence heuristic has not been identified or tested to date, several lines of research support the tenability of our hypotheses. First, various psychological literatures show that when X (e.g., positive events) is known to cause Y (e.g., positive mood), Y leads to an invalid inference of X (Johnson & Tversky, 1983; Mayer, Gaschke, Braverman, & Evans, 1992).
This leap is invalid when X is not the only cause of Y. That is, the choice of an item (X) does lead to the prevalence of that item (Y), but merely observing an item is prevalent (Y) does not reveal the context in which (X) was chosen. If more grocery stores carry vanilla ice cream than tiramisu, it means that vanilla ice cream may be chosen by shoppers more often, but not chosen
Second, another property of prevalence makes it enticing as a cue to social forecasts: its external observability. People prefer—and often are forced—to make forecasts about others using their observable behavior instead of their (often inaccessible) internal states (Pronin, 2008;
Pronin, Berger, & Molouki, 2007). Reliance on observable behavior is what makes social forecasts frequently superior to self forecasts. That is, social forecasts benefit from leaning on others’ (informative) past performance, whereas self forecasts give weight to the self’s (overlylofty ambitions (Helzer & Dunning, 2012). It is notable that in the present research, we depart from this tradition by highlighting how reliance on observable behavior can be a source of error instead of accuracy.
If people do lean on the prevalence heuristic when forecasting choice, it should lead to systematic biases when two forces are in opposition: the (perceived) prevalence and inherent likeability of options. The prevalence heuristic should lead people to overestimate the extent to which others will choose commonplace but bland options (e.g., vanilla ice-cream, black dress shoes, daisies) over rare but enticing options (e.g., tiramisu, blue dress shoes, dragon snaps). We test whether the prevalence heuristic correctly anticipates such forecasting errors, directly assess the attribute substitution account of our effects, identify what it is about the forecasting task that encourages reliance on the prevalence heuristic, and then ultimately probe an applied implication of the bias.
Method Participants and design. One hundred ten undergraduates at the University of California, Berkeley completed this and other unrelated studies as part of an hour-long session
checks—one that asked for the sum of 2 and 2, one that asked what the study was about (see Supplemental materials for details). The remaining 103 participants are included in the results below.
For Study 1, we aimed to collect at least 100 participants. Because we had key manipulations in Studies 2 and 3, we knew we needed a larger sample size. Instead of prespecifying a specific sample size, we prespecified an amount of time to run the study.
Research assistants recruited as many participants as they could for their scheduled hours and continued to run the experiment until the end of the academic semester. For Study 4, our sample size was based on how many other MTurk studies the lab planned to run that month as well as the lab's monthly MTurk budget at the time. In this way, Study 4 used the largest sample size that the budget permitted.
Procedure and materials. We constructed 11 pairs of items. Each pair was comprised of two options—one relatively prevalent (or common), one relatively rare—from the same category. These materials are presented in the left half of Table 1. To make certain that the items did indeed differ on perceived prevalence, we conducted a pretest on Amazon’s Mechanical Turk with 120 subjects. Nine participants failed one of two attention checks (see Supplemental materials for details). The remaining 111 were used for analyses. For all 11 pairs, the common item was indeed identified as more common for people to use or partake in than the rare item, all
ts 10.97, ps.001. Participants completed two sets of measures in a randomized order:
Choices. For each pair, participants were asked to consider having a choice between two options. For example, one item read, “If you had the choice of the following two options for
common option was “a sandwich.” The order of the two options was randomized, as was the order of the 11 choice pairs.
Forecasts. Participants were asked to forecast the choices of their fellow participants. For exploratory purposes, we elicited such forecasts in one of two normatively-equivalent formats (see Critcher & Dunning, 2013). Some participants were asked to answer, “How many of the 100 other people taking this study would choose one option vs. the other?” Others were instead asked to estimate how likely it was that a specific randomly-selected participant would choose one option or the other (e.g., “what is the percentage chance that Participant 71 would choose one option vs. the other?”) Each question included two sliding scales, one for each item in the pair, that had to add up to 100%. The order of the two items within each pair was randomized, as was the order of the 11 choice pairs.
Results On average, participants chose the common (but bland) over the rare (but exciting) option 50.22% of the time. Did participants realize that others choices would be evenly split between the two options? As predicted by the prevalence heuristic, participants believed that others would gravitate toward the common item (M = 59.12%, SD = 8.57%)—a significant overestimation, t(102) = 10.53, p.001, d = 1.04. The results by item are listed in Table 1. Participants’ forecasts depended neither on the nature of the target (random-selected individual vs. population) nor on the order with which participants indicated their own choice vs. forecasted others choices,
Note. Evidence consistent with the prevalence heuristic is seen when the predicted choice of the common item is greater than the actual choice. The significance level of each prediction bias is based on a one-sample t test on the predicted value compared against the actual value.
Study 2 replicated Study 1, but extended on it in two ways. First, we conducted a more direct test of the prevalence heuristic. More specifically, we estimated that choice forecasts would lean on the perceived prevalence of the options, even when controlling for how much participants thought others would be pleased to receive each option. Second, we had some participants reflect on both the prevalence and others’ liking for each options before they made
to be a valid cue to choice (or to liking, which predicts choice), this manipulation should have no effect. But if the prevalence heuristic is merely an attribute substitution driven by the cue’s accessibility and not its perceived helpfulness, then calling special attention to these two cues should encourage people to lean on the more normatively-defensible liking instead of prevalence (see Critcher & Rosenzweig, 2014, for similar logic).
Method Participants and design. Two hundred twenty-five undergraduates at the University of California, Berkeley completed this and other unrelated studies as part of an hour-long session for which they received course credit. Fourteen participants failed at least one of two attention checks—one that asked for the sum of 2 and 2, one that asked what the study was about (see Supplemental materials for details). Data from the 211 remaining participants are included in analyses reported below. Participants were randomly assigned to a salience or a (non-salience) control condition.
Procedure and materials. Like before, participants made judgments about 11 unique choice pairs. For each, participants indicated which option they would choose (choice), what percentage of other participants would choose one option or another (forecast), how prevalent or common each item was in people’s lives (prevalence), and how much other participants in the study would like to receive each option (liking). Participants in the salience condition completed the perceived prevalence and perceived liking measures (in a counterbalanced order) before completing the choice and forecasting measures (also in a counterbalanced order). Those in the (non-salience) control condition completed the prevalence and liking measures after the choice
Choices. These measures were equivalent to those used in Study 1, except we substituted out two choice pairs and added in two new choice pairs. In place of the questions about wall paint color and musical concerts, we asked participants to choose either an apple or a guava for their next snack and either a traditional landscapes or abstract/modern exhibit for their next art show attendance.
Forecasts. Given in Study 1 it did not matter whether forecasts were elicited for all other participants or a randomly-chosen other participant, all participants in Study 2 estimated what percentage of other participants would select one option or another.
Prevalence. Participants were asked to rate the commonness of having each item when partaking of or consuming an option in the relevant category. For example, participants indicated how common it was to eat curry when having lunch. Ratings were made on a 1 to 10 scale, from 1 (relatively uncommon) to 10 (relatively common). The order of the 22 items was randomized.