Confidence in Dyadic Decision Making: The Role of Individual Differences

Groups typically express more confidence than individuals, yet how individual-level confidence combines during collaborative decision tasks is not well understood. We prescreened 686 community members using a novel confidence measure (a true/false trivia test) intentionally designed to be difficult (accuracy rates were not significantly better than chance) and randomly assigned 72 individuals to collaborate on a matched version of the same test in dyads composed of two low-confidence individuals, two high-confidence individuals, or one of each (“mixed”). Consistent with past research, we found that the confidence expressed by dyads was higher than the confidence expressed by individuals; importantly, however, this pattern varied markedly by dyad type, with low-confidence dyads showing the largest increase, mixed dyads showing a moderate increase, and high-confidence dyads showing no increase—despite the fact that all dyads showed similarly low accuracy (about 55%). These results highlight the conditions under which groups express greater confidence than individuals and offer insights for the composition of collaborative decision-making teams. Copyright © 2015 John Wiley & Sons, Ltd. key words confidence; overconfidence; collaborative teams; dyads; group decision making What factors affect a group’s level of confidence in its decisions? This question has long been a topic of inquiry among scholars of group dynamics, from the seminal work on risky shift within the group polarization literature (e.g., Stoner, 1961; Myers & Lamm, 1975) to more recent efforts to illuminate the conditions under which “two heads are better than one” (Bahrami et al., 2010; Koriat, 2012; Turner & Pratkanis, 1998). Accordingly, much of this research has focused on relating a group’s confidence to the accuracy of its decisions, with the familiar finding that groups routinely fail to perform as well as their best individuals because of a variety of pitfalls (or “process losses”) that can plague group settings (e.g., groupthink; Janis, 1972; see Kerr & Tindale, 2004, for a review). Although this focus on the relationship between group confidence and accuracy has yielded important insights, consequential decisions are routinely based on the confidence expressed by groups long before the accuracy of those decisions is known, highlighting a need to better understand the factors that shape group confidence itself. For instance, consider the George W. Bush Administration’s famously high confidence that Iraq possessed weapons of mass destruction in the lead up to the 2003 U.S. invasion. According to journalist Bob Woodward (2004), CIA Director George Tenet expressed his agency’s confidence to the president by declaring, “It’s a slam dunk case! ... Don’t worry, it’s a slam dunk!” In turn, White House Press Secretary Ari Fleisher told the public, ... we have high confidence that they have weapons of mass destruction. This is what this war was about and it is about. And we have high confidence it will be found (Fleischer, 2003). In this vein, the present research focuses primarily on confidence itself, as opposed to the relationship between confidence and accuracy (or confidence realism; Adams & Adams, 1961) in exploring how confidence expressed by decision makers acting individually may shape the confidence expressed by groups they comprise. In doing so, we draw on research to suggest that confidence is equally or perhaps more reliable over time and across domains than is overconfidence, and that confidence is a suitable construct for the kind of individual difference research we pursue here. We first review work on the reliability and trait-like nature of individuals’ confidence judgments before turning to the present work: an experimental study in which participants were prescreened to assess their expressions of confidence on a general-knowledge test when working individually (as relatively low, medium, or high) and were later assigned to complete a similar test as a member of a collaborative dyad. Depending on experimental condition, dyads were composed of two low-confidence individuals (low condition), two highconfidence individuals (high condition), or one of each (mixed condition), allowing us to explore whether and how the confidence expressed by two people making joint decisions varies as a function of their individual-level confidence expressions. Individual-level confidence and overconfidence The bulk of research on confidence comes from studies that seek to relate an individual’s subjective confidence in his or her decisions to a measure of task performance or accuracy (e.g., Lichtenstein, Fischhoff, & Phillips, 1982). The relationship between subjective confidence and some objective outcome measure, or confidence realism (Adams & Adams, 1961), is typically represented by subtracting overall performance from overall confidence, such that positive numbers *Correspondence to: Jonathon P. Schuldt, Department of Communication, Cornell University, Ithaca, NY, USA. E-mail: [email protected] These authors contributed equally to the work. Copyright © 2015 John Wiley & Sons, Ltd. Journal of Behavioral Decision Making, J. Behav. Dec. Making (2015) Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/bdm.1927 signal overconfidence and negative numbers signal underconfidence (Lichtenstein et al., 1982; Yates, 1990). Research in this domain reliably finds that individuals are unjustifiably confident in their decisions, exhibiting marked overconfidence in a wide array of domains including tests of general knowledge (Fischhoff, Slovic, & Lichtenstein, 1977; Koriat, Lichtenstein, & Fischhoff, 1980), cognitive ability (Pallier et al., 2002), and various social predictions (Dunning, Griffin, Milojkovic, & Ross, 1990). Although overconfidence itself is widely observed, its magnitude has been shown to vary widely by task difficulty and domain, with substantial overconfidence reported in low-accuracy tasks and less overconfidence reported on high-accuracy tasks. Known as the hard–easy effect (Baranski & Petrusic, 1994; Lichtenstein & Fischhoff, 1977, 1980), this apparent dependence of overconfidence on task difficulty has prompted discussion as to whether the phenomenon results from a real and pervasive cognitive bias or whether it is an artifact of the difficult, artificial, and potentially misleading nature of the tasks that are commonly posed to participants in laboratory settings (Gigerenzer, Hoffrage, & Kleinbölting, 1991). Individual differences in confidence versus overconfidence Although a handful of studies report evidence that is consistent with a general overconfidence trait (e.g., Buratti, Allwood, & Johansson, 2014; Jonsson & Allwood, 2003; Stankov & Crawford, 1996; West & Stanovich, 1997), the weight of the evidence appears to suggest that confidence itself, independent of considerations of accuracy, may be the more reliable construct—a notion with deep theoretical roots (Henmon, 1911; Johnson, 1939) and modern empirical support. Support for a domain-general confidence trait has emerged from studies reporting robust intercorrelations for individuals’ confidence, but not necessarily for accuracy and overconfidence, across a range of tasks (Blais, Thompson, & Baranski, 2005; Kleitman & Stankov, 2001, 2007; Pallier et al., 2002; Schraw, 1997; Stankov, 1998). For instance, Bornstein and Zickafoose (1999) examined the relationships among participants’ confidence, accuracy, and overconfidence across eyewitness memory and general-knowledge domains and found that confidence exhibited the strongest correlation reported in the entire study (r= .49). In other work, Blais et al. (2005) examined confidence estimates for forced-choice tasks spanning three domains (vocabulary, general-knowledge test, and a perceptual line length task) and found that, for each pair of tests, the confidence correlation exceeded the overconfidence correlation. From a measurement perspective, observing stronger correlations for confidence than overconfidence is not surprising given that overconfidence is derived from two empirical measurements (i.e., confidence and accuracy) and therefore incorporates two sources of measurement error that combine to decrease the reliability of the resulting measure. Confidence expressions of individuals versus groups Especially relevant to the present research is the question of whether confidence is similarly stable across the individual and group contexts. On this point, numerous studies suggest that groups typically express more confidence in their assessments than do individuals working alone. Allwood and Björhag (1990) recruited participants to complete a generalknowledge test working either individually or as a member of a collaborative dyad. For each response, participants were instructed to assign a confidence estimate on a scale from 50% to 100%. Results showed that dyads expressed significantly greater confidence than individuals but showed no commensurate gain in accuracy, an observation the authors highlight as indicative of risky shift (Stoner, 1961). It should be noted, however, that the between-subjects design of that study did not allow the researchers to compare the confidence levels expressed by the same people across the individual and group settings. In this vein, a subsequent study by Allwood and Granhag (1996) had individuals answer and provide item-level confidence expressions for 30 general-knowledge questions, before doing the same for a different set of 30 questions while providing one argument in support of each chosen answer. Later on, the participants were divided into dyads to collaboratively answer and rate their joint confidence in the second set of questions once more. Results showed that group confidence exceeded individual confidence in both of the individual conditions. In a study explicitly examining the role of individual-level confidence in the confidence expressed by groups, Sniezek and Henry (1989) (see also Sniezek, 1992) used a common confidence measure in which individuals set 99% confidence intervals around frequency estimates for 15 causes of death (e.g., Haran, Moore, & Morewedge, 2010). Immediately thereafter, the same individuals completed the same task, with identical content, as members of randomly assigned triads. Results again showed that groups expressed significantly greater confidence than individuals; interestingly, groups were also more accurate and less overconfident, setting narrower confidence intervals that more frequently contained the correct point value. Although related to the present work, our study departs from these past studies in notable ways. First, by directly assessing both individual-level and dyadic confidence on similar (but not identical) tasks, our design facilitates strong inferences about the influence of trait confidence and decision context on confidence expressions. Also, in contrast to expressions of confidence made by individuals about group Vocabulary and general knowledge, r = .81 for confidence (.62 for overconfidence); vocabulary and line length, r = .30 (.20); general knowledge and line length, r = .38 (.37; only vocabulary and general knowledge significantly correlated in terms of accuracy, r = .57) As Buratti et al. (2014) discuss, inconsistent results regarding the relative stability of confidence and over/underconfidence across studies may be attributable to different methods (e.g., the nature of the decision task) or different analytic techniques. For instance, although they reported stronger correlations for confidence than for overconfidence in a memory task across three timepoints, more advanced multilevel modeling revealed relatively little intra-individual stability for confidence as compared with overconfidence. However, the difference was statistically significant in comparison with the argument condition only. Journal of Behavioral Decision Making Copyright © 2015 John Wiley & Sons, Ltd. J. Behav. Dec. Making (2015)