The role of intuition and reasoning in driving aversion to risk and ambiguity

Ellsberg (1961) was the first to show that individuals tend to prefer prospects whose probabilities are known over the same prospects with unknown probabilities.

See Alvarez et al. (2012) for details about the survey.

See Sloman (1996), Evans et al. (1993), Hammond (1996), Stanovich and West (2000), Gilovich et al. (2002), Kahneman (2003), and Slovic et al. (2002). Much of the literature, most notably regards intuitive (System 1) thinking as a source of mistakes and biases, while a handful of researchers disagree (see, e.g., Gigerenzer et al. 1999; Klein 2003; Dijksterhuis 2004).

Recent research comparing how fraternal and paternal twins make decisions suggests that reliance on these decision modes has a genetic component and is a stable, individual, trait, see Bouchard and Hur (1998). Our experiment shows that decision mode is stable across contexts, and that participants who take longer to reach decisions involving uncertain monetary outcomes also took longer to make choices in decisions free of monetary consequences (see Table A2 in Appendix—Supplementary material).

Conducting the experiment on-line rather than in a more traditional laboratory setting has pros and cons, of course. The primary benefit in this case was scalability: to have enough observations to examine primarily intuitive thinkers we anticipated it would be necessary to collect more observations than would be practical using in-lab sessions. An additional benefit is to lessen the salience of the experimenters, ameliorating common concerns about experimenter demand effects. The primary downside of conducting the experiments on-line is the limited interaction opportunities between experimenters and participants that make it difficult to directly address individual’s specific concerns or doubts. For example, individuals who doubt the veracity of the instructions and suspect that the experimenters strategically rig outcomes may have a preference for options involving less ambiguity for reasons having nothing to do with ambiguity preferences (see, e.g., Schneeweiss 1973; Kadane 1992). To allay such concerns, decision situations were described clearly and thoroughly; randomness was (accurately) described as being resolved impartially and not being contingent on participant’s choices–e.g., “...the computer will draw a chip ... each chip has the same probability of being drawn....” Finally, to lessen participants’ more general concerns about the possibility of being deceived, the invitation to participate originated from a trusted institution (EIEF) and was signed by a faculty member (Butler), whose contact information was provided in the e-mail and who promptly replied to individuals’ questions concerning the experiment.

All raw experimental data and statistical code are available on request. Data from the UCS are also available upon request but may require permission from Unicredit.

The question on the decision mode could have been framed in the context of a financial decision, for instance “Think of when you make a decision about financial matters.” Specific focus on a financial decision would have increased our confidence that the decision mode applies to financial decision rather than to other contexts.

Stanovich and West (2000) also provide evidence supporting this assumption. They argue that the systematic differences in performance along a large variety of tasks that they document in a sample of individuals can be traced to differences in the prevalence of one of the two systems of thinking, System 1 (based on intuition) or 2 (based on reasoning). Similarly, Klein (2003) offers many examples consistent with the idea that individuals differ systematically in their willingness to rely on intuition to make decisions.

A recent literature on eliciting preferences from survey data shows that qualitative questions on risk aversion are informative and have predictive power on behavior, see Barsky et al. (1997), Guiso and Paiella (2008), Dohmen et al. (2011).

These individuals could simply believe that the ambiguous urn has a more favorable distribution, for whatever reason. While such unwarranted optimism with respect to the ambiguous urn seems akin to ambiguity loving, it does not fit with any theoretical definition of ambiguity-loving that we know of and hence we do not classify it as such.

There are several alternatives to obtaining an index of ambiguity aversion. One is to ask individuals, as we do, to choose between a risky lottery and an ambiguous lottery; an alternative, followed for example by Guiso et al. (2008) and Halevy (2007), is to ask the willingness to pay for lotteries involving risk and involving ambiguity and then back out a measure of ambiguity aversion from the reported prices. A third, recently developed methodology (Bossaerts et al. 2010; Ahn et al. 2014; Choi et al. 2007) faces individuals in lab experiments with a large number of simple portfolio choices involving risky but non-ambiguous and ambiguous assets with varying prices. Individual preference parameters are then retrieved from these choices. Each of these approaches has pros and cons discussed in Sect. 6.

The UCS asked survey participants to choose between €100,000 1 year from the interview and an immediate sum M \(<\) 100,000. The initial value of \(M\) is set at €95,000. If the respondent accepts (turns down) 95,000 now she is asked whether she would accept 90,000 now (respectively 97,000); if she accepts 90,000 (turns down 97,000) she is further asked whether she would accept 80,000 now (respectively 98,000). If she turns down 80,000 her discount rate is above 20 %; if she turns down 98,000 the alternative offered is to wait 1 year and get 100,000. This allows classifying respondents into six categories with increasing subjective discount rates. In an ordered probit regression of this indicator of subjective discount the dummy for intuition has a small positive coefficient and that for reasoning small and negative but none of them is statistically significant (\(t\)-stat of 0.61 and \(-1.02\), respectively).

We collect decision time data for all sections of each experiment. In our analysis, we focus on decisions involving real monetary stakes and which involve risk and uncertainty. We focus on these questions because they are central to our investigation. However, the patterns in decision time and thinking mode are present in other sections of the experiment that do not involve monetary stakes, risk or uncertainty providing some reassurance that decision time is a stable individual trait. In particular, Table A2 in Appendix—Supplementary material shows that individuals who decide quickly (slowly) when monetary stakes are involved also complete the REI, which does not involve any monetary stakes, quickly (slowly).

Comparisons are strictly within-experiment as the questions used to elicit ambiguity aversion, and hence construct our decision mode measure, differ across experiments. That is to say, an Experiment 1 (2) participant’s response time is only compared to the response times of other Experiment 1 (2) participants.

Such “multiple price lists” (MPL) have become commonly used tools in experimental economics owing largely to their simplicity. The potential drawbacks of common MPLs are also well-recognized and include susceptibility to framing operating through, e.g., a propensity to pick middle rows and a more general weak incentives/hypothetical bias problem as typically only one row is chosen by experimenters to determine earnings. Another recurring issue is how to interpret inconsistent response patterns (multiple switch points). These issues are discussed in more detail in Andersen et al. (2006). To address the first two concerns, in Experiment 1 we implement a one-row-at-a-time version of the mechanism used in Holt and Laury (2002). Specifically, on each of a sequence of ten separate screens participants choose between Lottery A or Lottery B as described above for one value of \(x\), as \(x\) ranges from 0.1 to 1.0 in steps of 0.1. i.e., a typical screen would ask only one question: choose between “Option A: \([x]\) probability of receiving 20 euro; \([1-x]\) probability of receiving 16 euro;” or “Option B: \([x]\) probability of receiving 38.50 euro; \([1-x]\) probability of receiving 1 euro.” The positions of Options A and B (i.e., which comes first) are randomized on each screen. This procedure serves to make each separate decision as salient as possible, thereby ameliorating the weak incentives/hypothetical bias problem. Furthermore, presenting each decision separately and randomizing the order options are presented order plausibly lessens the impact of heuristics thought to be particularly relevant when choices are presented in a more standard ordered table format—heuristics such as picking the middle row, picking the middle option, or picking the first option. The main drawback of this particular mechanism is the possibility of multiple switching points. There is no consensus on how to handle such inconsistent observations—we balance a concern for preserving data against a concern about added noise by classifying individuals only with respect to their first switching point. In Experiment 2, we implement a simpler mechanism that has the advantage of not permitting multiple switching points—the “switching MPL” first used by Harrison et al. (2005). Here, all ten choices between Option A and Option B are presented in an ordered table format as in Holt and Laury (2002), but participants are asked only to specify the first row where they switch from preferring Option A to Option B. This downside of this simpler MPL is that it may be susceptible to the types of framing effects mentioned above. Although previous research finds negligible framing effects overall in these MPLs (e.g., Andersen et al. 2006), one might a priori worry about their influence on the intuitive decisionmakers studied herein.

Theoretically, truth-telling is a dominant strategy in the Becker–DeGroot–Marschak (BDM) mechanism. Whether the BDM mechanism reliably elicits truth-telling in practice, however, has been questioned by several researchers (see inter alia, Harrison 1992; Plott and Zeiler 2005; and Harrison and Rutström 2008). The mechanism has been criticized as being difficult for participants to understand and therefore susceptible to participant’s misconceptions or confusion about the incentives the mechanism provides. It also has been criticized for providing weak incentives for exact truth-telling—i.e., losses from misreporting just a little can be quite minimal depending on how the BDM is implemented. To partially address these concerns, our implementation of the BDM mechanism incorporates features suggested by Plott and Zeiler (2005) as effective at reducing participant misconceptions. We provide a limited form of training through a detailed numerical example demonstrating the optimality of truth-telling, followed by a mandatory quiz on how the mechanism works. We also stress to participants at the beginning of the experiment that their responses are anonymous—individuals are identified only by an experiment code. Finally, to provide adequate incentives for truth-telling, we use a lottery/asset with high variance (binary outcomes: €0 or €20) and participants are told that buying prices may range from the lowest prize to the highest prize (i.e., from €0.00 to €20.00) as suggested in Harrison and Rutström (2008).

While it is tempting to label as “ambiguity-seeking” those who value the bet related to the ambiguous urn strictly more, this is not possible. For example, an individual could simply believe that there is a larger proportion of red balls in the ambiguous urn, and therefore value a bet of “red from the ambiguous urn” more highly than a bet on either color ball from the non-ambiguous urn. This is a perfectly valid subjective belief that would be consistent with strictly valuing a bet on the ambiguous urn more. We thank David K. Levine for disabusing us of this common misperception.

We implement a procedure to ensure stated preferences are strict (see Appendix—Supplementary material for details).

As in the survey, one may be concerned that our intuitive thinkers are simply more prone to make mistakes and this is why they appear more risk tolerant and less ambiguity averse in our data. As detailed in Appendix—Supplementary material, we implement our risk aversion elicitation procedure in two slightly different ways, so that mistakes should manifest themselves differently across experiments. For example, making a mistake in the risk preferences elicitation procedure used in Experiment 2 is equally likely to misclassify an individual as risk loving as it is to mis-classify an individual as risk averse. Also note that the idea of intuitive thinkers being simply more prone to mistakes is not consistent with our findings in the Iowa Gambling Task experiment (Sect. 8), where we show that intuitive thinkers perform better than deliberative and partially deliberative thinkers.

Here our cognitive ability measure is a participant’s score on a standardized mathematics exam given in the final year of high school in Italy. The correlation between cognitive ability and our main thinking mode variable is non-significant in both experiments: 0.017 \((p\,>\,0.7)\) in Experiment 1 and \(-0.034\) (\(p\,>\,\)0.35) in Experiment 2. This suggests that our behavioral decision mode measure is not simply a proxy for cognitive ability.

The Iowa Gambling Task experiment was programmed and conducted with the software z-Tree (Fischbacher 2007).

The decks were pre-programmed to be identical to the decks used in the original Iowa Gambling Task (Bechara et al. 1994).

It is worth noting that there is evidence of persistence in decision mode. Regressing the total time spent on all 100 card draws on dummies for our behavioral decision mode measure from Experiments 1 and 2 reveals that intuitive thinkers took about 45 fewer seconds to complete all 100 draws than partially deliberative thinkers \((p = 0.009)\) and about 80 fewer seconds than deliberative thinkers \((p = 0.000)\). This pattern is robust to controlling for demographics and/or clustering standard errors by session. Not conditioning on decision mode, the average time spent to complete all 100 draws was about 395 s.

We lose a few observations by controlling for demographics as is evident from the number of observations in the table. However, the results and significance patterns are the same with or without demographic controls, so we report them with these controls.

It is worth noting that, while the authors of this last study claim to find evidence that more rationality is associated with less ambiguity aversion, and more intuitive processes with more ambiguity aversion—i.e., the opposite of what we find—their results are actually not clear on this point. This is primarily because what they describe as an ambiguous urn in one of their treatments is actually non-ambiguous. The urn in question contained “‘...at least two’ red marbles out of the 100, adding that ‘any number of red marbles from exactly two all the way up to 100 is equally likely.”’ This urn, while certainly relatively complex and representing a two-stage lottery, is non-ambiguous because the number of red marbles in it is known to participants to be uniformly distributed from 2 to 100.

In Butler et al. (2013) we further explore this link between preferences for risk and uncertainty and decision mode by experimentally manipulating participants’ willingness to rely on intuition, finding results largely consistent with those presented herein.

As pointed out by Bossaerts et al. (2010) a positive correlation between risk and ambiguity aversion can help explain the “value effect.”

Gollier (2006) shows conditions under which aversion to ambiguity reinforces risk aversion in the sense that it induces investors to invest less in stocks—the risky and ambiguous asset. See also Chen and Epstein (2002), Klibanoff et al. (2005), Mukerji et al. (2005), Gollier and Salanié (2006), and Epstein and Schneider (2010).