Predecisional Information Distortion and the Self-Fulfilling Prophecy of Early Preferences in Choice

The Ubiquity of Predecisional Information Distortion

Predecisional information distortion is a pervasive phenomenon, documented not only in medical decisions (Kostopoulou, Russo, Keenan, Delaney, & Douiri, 2012; Nurek, Kostopoulou, & Hagmayer, 2014), legal decisions (Carlson & Russo, 2001; Engel & Glöckner, 2013; Holyoak & Simon, 1999; Simon et al., 2001; Simon, Snow, & Read, 2004), and consumer decisions (Carlson, Meloy, & Russo, 2006; Russo et al., 2008; Russo et al., 1996; Russo et al., 1998) but also in employment, professional, entrepreneurial, and educational decisions (for additional references, see Miller, DeKay, Stone, & Sorenson, 2013; Russo, 2015). It occurs in choices between real goods (Carlson & Pearo, 2004) and for both experimentally manipulated and naturally emerging beliefs and preferences.

Information distortion is also relatively difficult to curtail. For example, it persists in the presence of monetary incentives for accuracy (Engel & Glöckner, 2013; Meloy, Russo, & Miller, 2006). In one successful debiasing effort, distortion was eliminated when participants considered detailed descriptions of the options’ attributes and their levels prior to the choice task (Carlson & Pearo, 2004). In terms of Figure 1, this procedure limits distortion by more firmly establishing the “undistorted” evaluations of attribute information. Russo (2015) has described other countermeasures, both successful and unsuccessful.

This section highlights three additional advances. First, many findings regarding distortion in riskless choices (e.g., between backpacks) also extend to risky choices. For example, DeKay, Patiño-Echeverri, and Fischbeck (2009b) reported that participants considering personal and policy decisions involving risk (e.g., deciding whether to evacuate upon receiving a dam-failure warning) distorted new information about outcomes and probabilities in whichever direction favored their emerging preference. Similar results have been obtained for choices between monetary gambles (DeKay, Stone, & Miller, 2011; DeKay, Stone, & Sorenson, 2012; Glöckner & Herbold, 2011; Miller et al., 2013) and other risky prospects (DeKay, Patiño-Echeverri, & Fischbeck, 2009a; Miller et al., 2013; Russo & Yong, 2011). These studies and others (Simon, Krawczyk, & Holyoak, 2004) have also demonstrated that precise numerical information is routinely distorted.

Second, information distortion can be decomposed into two components: positive distortion of information about the tentatively leading option, called proleader distortion, and negative distortion of information about the trailing option, called antitrailer distortion. In Figure 1, an emerging preference for Option A distorts information about that option positively (darker blue shading) and information about Option B negatively (lighter pink shading), with the antitrailer distortion resulting from the inhibitory link between the two options. Blanchard, Carlson, and Meloy (2014) and DeKay, Miller, Schley, and Erford (2014) found that proleader and antitrailer distortion were roughly symmetric in choices between two dry cleaners, backpacks, or apartments (see Figs. 2a and 2b). In Nurek et al.’s (2014) studies, however, physicians choosing between two diagnoses exhibited substantial antitrailer distortion but little if any proleader distortion (see Fig. 2c). The authors hypothesized that antitrailer distortion was more prevalent because physicians are trained to rule out competing diagnoses or because they are especially cautious in accepting a leading diagnosis.

OPEN IN VIEWER

Fig. 2.

Positive distortion of information about the leading option (proleader distortion) and negative distortion of information about the trailing option(s) (antitrailer distortion) in studies involving two options (panels a–c) or multiple options (panels d and e). In these studies, distortion was assessed relative to mean “undistorted” evaluations from separate control groups. Original evaluations were on 9-point scales (panels a, b, d, and e) or 11-point scales (panel c, rescaled here to match the others). Error bars represent 95% confidence intervals for the means. Data are from Blanchard, Carlson, and Meloy (2014; a and d); DeKay, Miller, Schley, and Erford (2014, Study 2; b and e); and Nurek, Kostopoulou, and Hagmayer (2014, Study 1; c). Panel (d) is adapted from “Biased Predecisional Processing of Leading and Nonleading Alternatives,” by S. J. Blanchard, K. A. Carlson, and M. G. Meloy, 2014, Psychological Science, 25, p. 815. Copyright 2014 by the Association for Psychological Science. Adapted with permission. Panels (b) and (e) are adapted from “Proleader and Antitrailer Information Distortion and Their Effects on Choice and Postchoice Memory,” by M. L. DeKay, S. A. Miller, D. R. Schley, and B. M. Erford, 2014, Organizational Behavior and Human Decision Processes, 125, p. 140. Copyright 2014 by Elsevier. Adapted with permission.

Third, people distort information not just in binary choices but also in choices with multiple options. This extension greatly increases the relevance of information distortion for consumer decision making, as well as for decisions in business, medicine, public policy, and other domains. When choosing among four apartments, DeKay et al.’s (2014) participants exhibited proleader and antitrailer distortion in roughly equal amounts, though the study did not assess distortion separately for each of the trailing apartments (see Fig. 2e). When choosing among six restaurants, Blanchard et al.’s (2014) participants exhibited positive distortion of information about the leading option, essentially no distortion for the second-ranked option, and increasingly negative distortion for lower-ranked options (see Fig. 2d). This orderly decrease may have resulted partly from participants’ ranking the full set of options after learning about each attribute (arguably a somewhat unnatural task) rather than simply focusing on the most competitive options. Nonetheless, these results and those of Nurek et al. (2014) indicate that information distortion need not be symmetric or uniform.

The Effects of Predecisional Information Distortion

Manipulating the order of information items so that participants are initially steered toward one option or the other has a substantial effect on their choices, even though all participants eventually see the same information (Carlson et al., 2006; DeKay et al., 2011; DeKay et al., 2012; Miller et al., 2013; Russo & Chaxel, 2010). This self-fulfilling prophecy of initial preferences—a form of primacy effect in which early information has undue influence on final choices—is mediated by the predecisional distortion of later information (DeKay et al., 2011; Miller et al., 2013). Similarly, Engel and Glöckner (2013) reported that assigning participants to the role of prosecutor or defense counsel in a criminal case affected participants’ (incentivized) verdict predictions and that this effect was mediated by distorted interpretations of the evidence. Primacy effects are not universal, however: Studies involving physicians have found either recency effects (Nurek et al., 2014) or neither primacy nor recency effects (Kostopoulou et al., 2012). A possible methodological explanation for these anomalous results is that the initial three-item “steer” in these studies was followed by a single question regarding diagnostic likelihood, whereas the three opposing information items, which appeared much later, were presented separately and were followed by two or three questions each (six or nine questions in all, depending on the study). This difference in emphasis may have accorded the later information items greater influence on the final diagnosis.

Of course, people’s evaluations of relevant information would be expected to affect their choices even in the absence of information distortion. It is therefore useful to disentangle the effect of information distortion from the effect of people’s “undistorted” evaluations of the information items (undistorted by the consideration of specific choice options, that is). Individual differences in undistorted evaluations cannot account for the effect of information order on choice in the studies cited above because participants were randomly assigned to information orders. But what if participants’ early leanings are not manipulated? DeKay et al. (2014) provided strong evidence for the link between distortion and choice in such cases. Specifically, both proleader and antitrailer distortion predicted choice over and above the effect of the same participants’ undistorted evaluations, which were assessed in a separate no-choice control task. DeKay et al. (2014) also analyzed identical information items that did not distinguish between the options (e.g., when two apartments both had “tasteful, newly renovated interiors”). Although participants’ undistorted evaluations should have been the same for identical items, participants still distorted the information to favor the tentative leader, and this distortion still predicted participants’ final choices.

Distortion also predicts other choice-related variables, such as participants’ final strength of preference (DeKay et al., 2011; DeKay et al., 2012), the difference between participants’ subjective monetary values for two alternatives (DeKay et al., 2012), participants’ final confidence in their verdicts (Simon, Snow, & Read, 2004), and physicians’ final and near-final judgments of diagnostic likelihood (Nurek et al., 2014). In DeKay et al.’s (2014) studies, predecisional distortion also predicted participants’ postdecision memories of the original information (attributes of apartments), with positive and negative distortions predicting corresponding memory errors.

The effects of information distortion can be sizable. For example, manipulating information order typically yields swings of 20 to 30 points (or more) in the percentage of participants choosing one option or the other (DeKay et al., 2011; Miller et al., 2013). Using U.S. dollars as the metric, DeKay et al. (2012) estimated the effects of distortion to be about 30% as large as the average value of the options in choices between monetary gambles. In other risky choices (e.g., responding to a dam-failure warning), evaluations of possible decision outcomes are sometimes distorted enough to imply falsely that one option is always better, regardless of the probabilities involved (DeKay et al., 2009a, 2009b). Distortion can also lead people to choose options that they otherwise judge to be inferior (Russo, Carlson, & Meloy, 2006).

Outstanding Issues

There are several promising avenues for future research. For example, although choices among several options are commonplace outside the lab, only two studies have assessed the distortion of multiple options (Blanchard et al., 2014; DeKay et al., 2014). Blanchard et al.’s results (see Fig. 2d) raise a number of interesting questions. In particular, the increasingly negative distortion of information about the less attractive options appears consistent with coherence-based accounts, but it seems hardly worth the trouble if the goal is to choose the best alternative. Ideally, this issue can be studied without requiring participants to rank the options explicitly, though that may prove difficult. The absence of distortion for the second-ranked option is also intriguing. If the decision maker (or another party) winnows the field from several options to two, does distortion of the second-ranked option remain near zero, or does it become negative, as in decisions that start with only two options?

A second set of questions concerns the embedding of information distortion in the larger choice process. For example, even if the proximal effects of proleader and antitrailer distortion on final choices are roughly equal (DeKay et al., 2014), antitrailer distortion may still matter less if people focus their attention or information search on the leading option, as has been suggested (Carlson & Guha, 2011). Engel and Glöckner (2013) have explored these issues in legal decisions, finding that participants’ assigned roles (prosecution or defense) affected information distortion but not information search. However, differences in search patterns were surely dampened by the task-appropriate requirement that participants read each piece of evidence at least once. Decisions in other contexts or with more options may allow greater latitude for leader-based search processes.

Third, it is clear that predecisional information distortion and evaluative coherence should be added to the list of empirical findings (regarding, e.g., choice patterns, confidence judgments, and response times) that are used to evaluate and compare psychological models of choice. These additional criteria imply immediately that many of the most popular models need to be modified. As already noted, connectionist models that include bidirectional links between information items and options are currently distinctive in their ability to account for information distortion and its effects. Such models can (a) explain how information distortion leads to greater coherence (Glöckner et al., 2010; Holyoak & Simon, 1999; Simon et al., 2001; Simon, Snow, & Read, 2004), (b) describe choice patterns (and some other results) at least as well as other decision models when all information is presented simultaneously (Glöckner & Herbold, 2011; Glöckner, Hilbig, & Jekel, 2014), and (c) account rather naturally for primacy effects when information is presented sequentially (Hagmayer & Kostopoulou, 2013).

More work remains, however. For example, current implementations of connectionist models do not clearly differentiate between changes in information evaluation and changes in information weighting, both of which are known to occur (Simon, Krawczyk, & Holyoak, 2004). In addition, different authors structure and interpret their networks differently. In inference tasks, for example, Glöckner et al. (2010; Glöckner et al., 2014) located the subjective validity of the evidence in the information nodes, whereas Hagmayer and Kostopoulou (2013) located subjective validity in the links that connect the information nodes (the evidence) to the option nodes (the hypotheses). Structures and interpretations can also differ between inference tasks (Glöckner et al., 2010; Glöckner et al., 2014) and preference tasks (Glöckner & Herbold, 2011). While such diversity is unsurprising, and perhaps healthy, restructuring or generalizing connectionist models to provide a more systematic account of the choice process should be an important goal.

Finally, there is the question of rationality. Although numerous authors have argued that bidirectional reasoning is irrational or that it can lead to poor outcomes (DeKay et al., 2009a, 2009b; Engel & Glöckner, 2013; Russo, 2015; Russo et al., 2006; Russo & Yong, 2011; Simon, 2004), it could still be adaptive in some circumstances if it leads to faster, easier, good-enough decisions that are less likely to be reconsidered or regretted. This broader view of rationality acknowledges that choosing the option with the best combination of features is only one of several goals and constraints that decision makers typically consider. Information distortion may be especially beneficial in situations where early information is expected to be valid and diagnostic rather than irrelevant or misleading, as when a potential moviegoer seeks a recommendation from his or her favorite film critic or from a friend with similar tastes. It may also help in situations where evaluatively similar features tend to co-occur, so that an option’s attractiveness on one feature predicts its attractiveness on others. More generally, whether information distortion is helpful, harmless, or harmful may hinge on the fit between the decision process and the decision environment (Hogarth & Karelaia, 2007). Recent evidence from simulations has indicated that choices based on coherence-based processes like those underlying information distortion can mimic choices based on rational processes relatively well—better, in fact, than several competing decision strategies—in a variety of situations (Jekel, Glöckner, Fiedler, & Bröder, 2012). Although distortion could still foster unjustified confidence or a tendency to overpay for the chosen option, these results suggest that the consequences of bidirectional reasoning may not be as bad, or as frequently bad, as is commonly assumed.

Concluding Remarks

Two decades of research on information distortion and its effects tell a consistent and compelling story: Once decision makers start to lean toward one option or another, they evaluate subsequent information in a manner that tends to reinforce their developing preference. This positive feedback loop increases the coherence between their preference and the information on which it is ostensibly based and increases the chance that their final choice will match their initial leaning. This self-fulfilling prophecy is not perfect, of course, but the influence on choice can be considerable. Although interesting questions remain, predecisional information distortion and its effects are widespread, robust, and well established. The phenomenon warrants inclusion as an essential feature in any theory or model that aspires to describe the choice process.