Atmospheric methane levels increased over geological formations in Northern Siberia following a heat wave in 2020, suggesting that permafrost thaw releases methane from reservoirs.
Image credit: Wikimedia Commons.
Edited by Shengwu Li, Harvard University, Cambridge, MA, and accepted by Editorial Board Member Paul R. Milgrom June 27, 2019 (received for review December 14, 2018)
Significance
A large literature in behavioral and experimental economics has identified a long list of robust phenomena that are hard to explain within the classic model of economic choice. These are, however, typically analyzed independently. We study the joint distribution of 11 of the most prominent behaviors using an incentivized laboratory experiment involving undergraduate students. Our main aim is to provide empirical guidance in the construction of a unified, parsimonious model of economic behavior, an important step in integrating behavioral insights into broader economic and policy analysis. We find that some of these phenomena are strongly related to each other, while others are independent. This provides evidence in support of some of the proposed attempts at unification, but not others.
Abstract
We study the joint distribution of 11 behavioral phenomena in a group of 190 laboratory subjects and compare it to the predictions of existing models as a step in the development of a parsimonious, general model of economic choice. We find strong correlations between most measures of risk and time preference, between compound lottery and ambiguity aversion, and between loss aversion and the endowment effect. Our results support some, but not all attempts to unify behavioral economic phenomena. Overconfidence and gender are also predictive of some behavioral characteristics.
Over the past 30 y, behavioral and experimental economists have made great strides in identifying robust phenomena that are hard to explain within the classic model of economic choice. These include high short-term discount rates and small stakes risk aversion, present bias, loss aversion, the endowment effect, aversion to ambiguity and compound lotteries, the common ratio and the common consequence effects in choice under objective risk, and sender/receiver behavior in trust games. In response to these findings, there has been an enormous amount of research within experimental economics aimed at understanding the nature of each of these behaviors. However, much less attention has been paid to understanding the links between them.
In this paper we document the empirical relationship found in an incentivized laboratory experiment between the 11 phenomena listed above, as well as gender, cognitive ability, and personality measures.† Our primary aim is to provide data with which to test and develop parsimonious models of economic choice that can capture many features of behavior with a small number of parameters. The construction of such models is of first-order importance for the systematic application of behavioral economic insights to mainstream economic analysis. Previous authors have lamented the twin problems of “model” and “anomaly” proliferation in behavioral economics (1). Tractable workhorse models would also greatly benefit policy design and cost–benefit analysis. At the same time, there is no point in developing theoretical models that tie different behaviors to the same underlying parameter if these behaviors are not, in fact, related. In this paper we aim to document the empirical links between different behaviors as a first step toward informing this important theoretical project. We take 2 approaches to analyzing the data. In Results, we report on the main empirical patterns we observe. In Discussion, we outline the implications for existing models that make predictions about the links between different behavioral biases (see also SI Appendix, section S3) and describe the potential limitations of our study.
Our basic premise is that if 2 behaviors are driven by some common underlying cause, then they should be correlated in the population. So, for example, if present bias and ambiguity aversion are both due to probability weighting (2⇓–4), then variance in this trait across individuals should cause the 2 behaviors to be correlated in our data. Implicit in our strategy is the assumption that not only the same cause, but also the same parameter is driving behavior in both cases. This is not without loss of generality, as in principle it could be that probability weighting is responsible for both behaviors, but different, independent weights are used in each domain, leading the 2 variables to be potentially uncorrelated.
Traditionally, experimental papers have focused on measuring small subsets of the behaviors we measure.‡A key innovation of our paper is that we study a wide range of behaviors in the same group of subjects. This has a number of advantages. First, it allows us to test a large number of theoretical predictions on the same dataset, including relationships that to our knowledge have not been previously looked at. Second, we can control for one behavior when examining the relationship between others. Third, we can differentiate between explicitly bilateral relationships between 2 phenomena and a general tendency toward “irrational” behavior. Fourth, we can compare the magnitude of different estimated relationships while keeping the sample constant. Finally, there are direct practical benefits in understanding the correlation between multiple behaviors independent of any modeling analysis: Understanding the joint distribution of such phenomena could help with empirical analysis that allows for more than one of our behaviors of interest—for example by informing about their orthogonality. The identification of the main factors of economic decision making could help in the design of efficient tests of a person’s “economic makeup,” along the (incomplete) set of dimensions we measure, much in the same way that personality research has allowed psychologists to come up with tests that characterize a person’s personality.
A few papers attempt a broader analysis closer to our approach. Refs. 21 and 23 make use of a large-scale experiment carried out on a group of newly recruited truck drivers and use parametric methods to measure a subset of our behaviors. A literature in psychology studies the empirical relationship between different kinds of violations of rationality (27), albeit focusing on different behaviors including specific violations of reasoning (e.g., base rate neglect and syllogistic reasoning). Ref. 28 studies the relationship between different methods of risk elicitation, finding that the choices in “behavioral” tasks are only weakly correlated with each other or with self-reported risk attitudes or risky lifestyle choices. In contrast, in this study we keep the elicitation method fixed and vary the lotteries over which we extract preferences.
More recently, a few papers have followed subsequent to our initial research (29⇓⇓⇓⇓–34). Some of these papers are complementary to ours, having different aims, and therefore making use of different questions and analysis. Refs. 29 and 30 focus on how to combine behavioral factors into a single measure of how “behavioral” a person is, while ref. 34 is mainly concerned with studying international variation in preferences. Unlike these studies, in our experiment decisions are actualized, and so subjects face real financial incentives (Materials and Methods). Other studies build explicitly on the approach of this paper; this is especially the case for ref. 32, which elicits a similar (but not identical) set of measures in a representative sample. Unlike that study, we elicit behaviors in a laboratory environment, in a group that is highly selected and homogenous (e.g., in age and cognitive abilities). This sacrifices representativeness, but allows for the much greater degree of control that only the laboratory guarantees. For example, we are better able to time the receipt of payments for time preference questions. Moreover, the homogeneity in the population reduces the concern that the correlations we observe could be due to some unmeasured background circumstance (e.g., familiarity with the interface or attention), but are instead more likely to derive from a deep connection between the behaviors. Overall, we see the 2 approaches as being largely complementary.
Results
The analysis in this paper is based around the measurement of 11 economic behaviors, as well as cognitive and personality measures (details and experimental instructions in SI Appendix). Where possible, our measures are derived directly from choice, rather than from an underlying behavioral model. For example, we measure risk aversion as the percentage difference between a lottery’s certainty equivalence and its expected value, rather than the estimated curvature of a parameterized utility function. In general, we use multiple questions to measure each behavior; this allows us to control for measurement error in regressions (SI Appendix, section S2.1).
Two measures relate to intertemporal choice. “Present discount” measures the discount rate between money sent today and money sent in the future. “Discount” measures the discount rate between 2 future dates. This also allows us to calculate “present bias”—the frequently observed phenomenon in which subjects exhibit higher discount rates for choices involving immediate payment (e.g., ref. 35).§
Seven measures relate to attitudes toward risk and uncertainty. “Risk aversion” measures the degree to which the subject’s valuation of a lottery with 2 equally likely positive monetary prizes is discounted with respect to the expected value. The “common consequence” and “common ratio” effects are versions of the Allais paradox (36) and measure the degree to which subjects violate expected utility by exhibiting a preference for certainty. “Uncertainty attitude” and “compound lottery attitude” measure the degree to which subjects discount gambles based on prospects in which, respectively, probabilities are not explicitly stated or are 2-stage lotteries. In turn, this allows us to measure “ambiguity aversion” and “reduction aversion”: Violations of expected utility theory based on the degree to which uncertain prospects and compound lotteries are discounted more than simple lotteries (17). “Mixed risk” measures risk aversion for binary lotteries with 1 positive and 1 negative prize, which in turn allows the estimation of “loss aversion,” the degree to which losses are weighted more heavily than gains (37, 38). “Buy risk” replicates the risk aversion measure, but frames the decision as one in which the subject must spend money to buy the lottery. From this we can also estimate the “endowment effect,” or the disparity between the willingness to pay for a lottery and the amount at which the subject will sell the same lottery (e.g., ref. 39).
Two measures relate to behavior in a strategic interaction known as the trust game, in which senders must decide how much of their endowment to send to receivers. Any amount sent is then tripled by the experimenter, and the receiver must decide how much should be returned to the sender. Subjects take part in both roles in the experiment. “Trust” measures the amount sent in the first stage, while “return” measures the percentage of the received amount which is returned.
Additionally we measure “cognitive ability” using Raven’s matrices (40). We use this also to measure overconfidence, by asking subjects to estimate the number of questions they got right and the average number of correct responses in the session. As is standard, “overconfidence” is measured as the difference between their predicted and actual score; “overplacement” is the difference between their predicted score and predicted group score (41). “Anxiety” and “depression” were measured using the Beck Anxiety and Depression indexes.
SI Appendix, Table S2 summarizes our measures for each behavior across subjects. By and large, our estimated behaviors are similar to those reported in the literature. Almost all of the phenomena of interest are exhibited and are statistically significant at the 0.1% level. The only exception is that we do not observe overconfidence in our subjects. This is probably due to the difficult nature of the task, which can lead to underconfidence (41). Subjects do, however, exhibit overplacement; on average, subjects expect to be above the mean in performance.
Table 1 summarizes our main results; it shows the correlations between our measures. For simplicity, Table 1 includes only “raw” measures (e.g., uncertainty attitude), rather than those which are constructed from multiple questions (e.g., ambiguity aversion). This eliminates concerns about spurious correlations caused by the same underlying question being used in the construction of multiple measures.
View this table:
Table 1.
Correlation between behaviors
Several broad patterns are apparent:
i) There are significant pairwise correlations between almost all of our measures of attitudes to risk and uncertainty. However, the bilateral strength of these relationships varies greatly. The highest correlations are between i) risk aversion and attitudes to uncertainty and compound lotteries and ii) mixed risk (which is related to loss aversion) and buy risk (which is related to the endowment effect). Also notable is the relationship between the common consequence and common ratio effects and risk aversion.
ii) There are significant relationships between many of our measures of risk/uncertainty attitude and discount rates, particularly between risk aversion and present discount. Time preferences are, however, not related to common ratio or common consequence.
iii) While sender and receiver behavior in the trust game are strongly related to each other, there is no evidence of a systematic relationship between these variables and other behavioral measures.
iv) Cognitive and personality measures are generally only weakly predictive of behavior. The exception is our measure of overplacement—which is significantly and negatively correlated with many variables—and gender—which shows that females dislike mixed risk and buy risk significantly more than males.
The significance values in Table 1 are reported without correcting for multiple comparisons. These are of relevance because we are at least in part interested in the comparison between these correlations and those predicted by existing theories. We provide Sidak-corrected significance levels in SI Appendix, Table S11. Broadly speaking, the relationships that survive correction are those labeled as significant at the 0.1% level in Table 1.
We next report further analysis looking at various different clusters of variables. In doing so, we use a “multiple indicators” approach, based on an instrumental variable strategy, to control for potential measurement error (SI Appendix, section S2.1).
Attitudes to Risk and Uncertainty.
Table 1 demonstrates that attitudes to different types of risk and uncertainty are related. Of particular interest is whether subjects who violate expected utility in one domain are more likely to do so in another. This is the case in our experiment (SI Appendix, Table S3). Subjects who are ambiguity averse are 86 percentage points (pp) more likely to be reduction averse, 40pp more likely to exhibit the common ratio effect, and 21pp more likely to exhibit the common consequence effect than those who are ambiguity neutral (all significant at the 5% level—Fisher’s exact test).
While the propensity to violate independence in the domains of risk and uncertainty is related, there is no strong link between the degree to which subjects exhibit ambiguity aversion or reduction aversion and the degree to which they exhibit the common ratio and common consequence effects. (SI Appendix, Table S4). Instead we find an extremely strong relationship between uncertainty attitude and risk aversion, as well as a significant positive relationship with the endowment effect. Similar relationships hold when uncertainty attitude is replaced by ambiguity aversion, compound lottery attitude, or reduction aversion.
A further important question is how the endowment effect relates to other risk attitudes, and in particular to loss aversion, with many theories predicting that the same underlying process drives both behaviors (see Discussion). Indeed, in our data we find that loss aversion in risky choice is strongly related to both buy risk and the endowment effect (SI Appendix, Table S5), even after controlling for risk aversion. This supports the observation in Table 1 that mixed risk (related to loss aversion) and buy risk (related to the endowment effect) are strongly correlated. It should be noted that while ref. 26 also finds a correlation between loss aversion and the endowment effect, ref. 31 does not. We discuss possible causes for this discrepancy in SI Appendix, section S3.2.
Time Preferences and Risk/Uncertainty Attitudes.
SI Appendix, Table S6 reports the results of a multivariate regression of time preferences on risk/uncertainty measures, allowing us to better understand the relationship between these 2 behavioral clusters. We find that risk aversion is weakly related to the discount rate, strongly related to the present discount rate, and therefore significantly related to present bias. The other measures of risk and uncertainty attitudes have little additional explanatory power, although ambiguity aversion is positively and significantly related to discounting at the 10% level. The common ratio and common consequence effects are not significantly related to time preferences, either individually or jointly. Note also that the point estimate of the relationship between these 2 behaviors and present bias is negative.¶
Trust Game Behavior and Risk Preferences.
Sending money in the first stage of the trust game is an uncertain prospect, while keeping the money is not. Thus, it is possible that play in the trust game is related to risk/uncertainty preferences. One of the advantages of the approach taken in this paper is that we can test simultaneously for the relationship between trust game behavior and various different components of risk preferences. However, we find little evidence of such relationships (SI Appendix, Table S7). One possible explanation suggested by previous research is that other factors, such as betrayal aversion, may cause deviations between risk aversion in strategic and nonstrategic settings (42). Only loss aversion is related to sender behavior, with subjects who are more loss averse sending more—a surprising result given that sending money in the trust game represents a gamble with both gains and losses. We find no relationship between risk preferences and return behavior. We do find a strong relationship between sender and receiver behavior. Such a link could occur either because those who return money in the trust game have different beliefs about the behavior of others or because subjects who have stronger other-regarding preferences both send and return more in the trust game.
Gender, Cognitive Ability, Personality Measures, and Overconfidence.
Our data are inconsistent with the hypothesis that nonstandard economic behavior is a proxy for intelligence. Table 1 shows that only discounting and (more weakly) buy risk are related to our intelligence measure. As shown in SI Appendix, Tables S8–S10, however, this does not extend to present bias or the endowment effect. Recent papers (22⇓–24, 43) find stronger relationships with intelligence. However, they use different samples with possibly more variation in intelligence.#
We find overplacement to be strongly related to many of our behaviors. Both Table 1 and SI Appendix, Tables S8–S10 suggest that higher overplacement is related to lower common consequence effect, ambiguity, and reduction aversion; it is also related to lower discounting, mixed risk, and buy risk—although these relationships disappear when controlling for gender (below). Particularly intriguing is the relationship with ambiguity aversion: The more overconfident they are, the more subjects may find it less likely that they have been “outsmarted” by the environment and thus may exhibit lower ambiguity aversion.
In terms of gender, subjects who self-identify as female report much higher buy risk and mixed risk. Both are extremely robust findings (see SI Appendix, Tables S8–S10 for further analysis). Gender does not seem to impact any other variable.
Finally, we find little evidence of a link between economic behaviors and measured depression or anxiety.
Discussion
Our results so far provide a rich set of observations for any proposed unified model of behavior to match. First, there is evidence consistent with a common factor being at least in part responsible for variation in our underlying measures of risk and time preference (discount rate, risk aversion, uncertainty aversion, etc.), almost all of which exhibit strong bilateral correlations. Second, and in contrast, the biases derived from these measures (present bias, ambiguity aversion, common ratio effect, etc.) seem to have multiple drivers, as the correlation structure between them is much sparser: We see strong relationships between ambiguity aversion and compound aversion, loss aversion and the endowment effect, and sender and receiver behavior in the trust game, but little else. Third, we do not find that any of our behavioral traits appear to be proxies for intelligence: Indeed, of our cognitive or personality measures, only gender and overplacement are meaningfully related to any of our economic behaviors.
Our findings have implications for existing models that make predictions about the relationships between behavioral economic phenomena. Recent theoretical work offers the possibility of a parsimonious, general model that can capture many features of behavior built around the cumulative prospect theory (CPT) of ref. 38. CPT extends the standard model of economic decision making in 2 ways: Loss aversion (which leads people to weight losses more than gains in their utility functions) and probability weighting (which leads people to overweight some probabilities and underweight others). Originally, CPT was used to explain violations of expected utility over objective lotteries: Loss aversion predicts increased risk aversion for lotteries involving gains and losses, while probability weighting can explain the common ratio and common consequence effects as well as small stakes risk aversion. Subsequent papers have suggested that these 2 behavioral traits can explain other phenomena: Refs. 44 and 45 show that loss aversion can explain the endowment effect, refs. 2 and 3 demonstrate that probability weighting can explain ambiguity and compound lottery aversion, and ref. 4 shows that probability weighting can lead to present bias in intertemporal choice if the future is perceived as risky by the decision maker. Thus, an “extended prospect theory” (EPT) model currently appears to be the leading theoretical attempt to parsimoniously link various nonstandard behaviors, adding only 2 parameters to the standard model (46, 47). It is a natural starting point to test in our data. In SI Appendix, section S3 we discuss this model in detail and expand on the implications discussed below.
Our results provide support for some of the elements of EPT, but not others. Our data are consistent with a unified notion of loss aversion linking the endowment effect to risk preferences for lotteries with gains and losses. They are also consistent with a probability weighting function that links risk aversion to the common ratio and common consequence effects. However, our data are not consistent with the hypothesis that the same probability weighting function explains present bias or aversion to ambiguity or compound lotteries.
Other correlations in our data have implications for theoretical links between behaviors that do not depend on probability weighting or loss aversion. The curvature of the utility function can affect attitudes to compound lotteries, uncertain prospects, intertemporal choice, and risk aversion. The strong link we find between these behaviors suggests that the shape of the utility function could be an important determinant of all 4.∥Ref. 48 proposes the concept of a “fear of change” which could link the endowment effect and ambiguity aversion, as is apparent in our data. The correlation we find between ambiguity and compound lottery aversion suggests either that ambiguous prospects are seen as compound lotteries (as in ref. 3) or that compound lotteries are seen as ambiguous prospects. The fact that subjects who are ambiguity averse are also more likely to violate independence in risky choice is problematic for the many theories of ambiguity aversion that rely on risk independence (e.g., ref. 49), suggesting the need for models that simultaneously allow for both types of violation (e.g., ref. 50).
There are multiple limitations of our approach. First, our sample is selective as it is composed of undergraduate students. While this is typical in experimental economics, it limits the external validity of our findings; moreover, by possibly reducing variation along some dimensions, it may also reduce our ability to identify some correlations. Second, there is the effect of measurement error; while the regression results reported in SI Appendix use a multiple indicators approach that provides consistent estimates in the face of idiosyncratic measurement error (SI Appendix, section S2.1), noise in our estimates will still affect statistical power, which could lead to false negatives in our correlation table. This may be of particular concern in a long experiment such as ours and one in which we can ask relatively few questions to measure each behavior. Third, false positives could derive from other underlying variation that is not of interest, but might affect behavior across multiple dimensions. For example, if experimental choices are affected by the structure of the Multiple Price List (MPL) (51), and the strength of this effect varies across subjects, this could induce spurious correlations. While in our data the pattern of correlation is not explained by obvious features of the design, such as the distance between mean response and the center of the MPL, the possibility of more subtle effects remains. Fourth, we focus on a very selected subset of behavioral measures, which is far from comprehensive; while feasibility makes selection inevitable, different patterns may emerge once other behaviors are measured. Fifth, we measure time preferences using dated monetary rewards and the certainty effect using choice lists. More recently, both approaches have been subjects of some concern, the former because of the possibility of consumption smoothing and arbitrage (52) and the latter because subjects who do not isolate their decisions may not see such choices as providing certainty (53). This could lead to noisy or inaccurate measurement of these variables and an underestimation of any resulting correlations (see SI Appendix, section S1 for a discussion). Sixth, we focus only on correlations and abstract from nonmonotonic relationships or interactions; a comprehensive analysis of more complex relationships would be an interesting avenue of future research. Finally, we have no evidence linking the constructs measured in our experiment to behaviors of interest outside the laboratory, such as financial, employment, or even political decisions. Subsequent papers, explicitly building on ours, attempt to collect data that address some of these issues, although at times with a different focus (29⇓⇓⇓–33).
Materials and Methods
The experiment was performed on 190 subjects over 8 sessions. All subjects were recruited from Brown University. Approval was obtained from the Institutional Review Boards at Brown University (1004000171) and California Institute of Technology (PO-285). The experiment was administered via a specially written computer program. Written consent was obtained from subjects before the start of the experiment. In total, subjects answered ∼50 questions to estimate their preferences, followed by the cognitive test, overconfidence measures, and gender and personality measures. On average subjects took about 40 min to complete the experiment. A copy of the instructions to subjects is included in SI Appendix.†† Data and replication files are available as part of Datasets S1–S3.
Economic questions were asked in 8 blocks (discounting and present bias, risk aversion, violations of expected utility, uncertainty and compound lottery attitude, risk aversion for losses, risk aversion for gains and losses, trust game, and willingness to pay for lotteries), with specific instructions appearing before each block. For slightly less than half the subjects, the order of these blocks was reversed, allowing us to control for order effects.
For all economic behaviors, other than those involving the trust game, our measures are based on indifferences elicited using a multiple price list method. Subjects were presented with a sequence of between 12 and 18 binary choices on the screen. In each choice, the option on the left remained the same, while the option on the right improved as the subject moved farther down the screen. Subjects had to make a choice on each line. Indifference is estimated as the point at which the subject switched from choosing the option on the left to the option on the right. If the subject did not make exactly 1 switch, then data from that question were discarded. Subjects who make multiple interior switches have behavior which is not compatible with monotonicity in money, while we cannot estimate an indifference point for subjects who have 0 interior switches. Such subjects would also have to have very extreme preferences or, more likely, not be paying attention to the question.
We exclude from our analysis 10 subjects who failed to make a unique switch on more than 15 questions, on the assumption that any remaining information from these subjects is likely to be of little use. For the remaining subjects, the loss of data varies from 0% of observations for the trust game to 22% for the loss aversion measure (constructed using 9 separate choices). On average, we lose about 8% of the subjects for each measure, which is broadly in line with other studies (e.g., ref. 54). For the correlation analysis of Table 1 we report pairwise correlations only for subjects who answered all relevant questions. For the regression analysis reported in SI Appendix we interpolate using the data we have to estimate the missing data—so, for example, if we observed data for 2 of the 3 risk aversion questions for a subject, we estimate the response to the third question using the relationship between the answers to the questions estimated on subjects that answer all 3. Doing so reduces the number of missing observations—for example it allows us to construct a loss aversion measure for all but 10% of our subjects; importantly, this approach does not significantly alter our findings: Similar results are obtained if we just disregard these data, albeit with a smaller sample size.
At the end of the study, 2 questions were selected at random for payment. For each of these, 1 line was then selected at random, and subjects received the option that they chose on that line.
In general we use 3 questions to measure each behavior. To aggregate these into a single measure for the correlations in Table 1 we take the first principal component. In regression analysis in SI Appendix we use a multiple indicators approach to control for measurement error (SI Appendix, section S2.1).
Acknowledgments
We gratefully acknowledge support from the National Science Foundation (Grant SES-1156091).
Footnotes
↵1M.D. and P.O. contributed equally to this work.
- ↵2To whom correspondence may be addressed. Email: pietro.ortoleva{at}princeton.edu.
Author contributions: M.D. and P.O. designed research, performed research, analyzed data, and wrote the paper.
The authors declare no conflict of interest.
This article is a PNAS Direct Submission. S.L. is a guest editor invited by the Editorial Board.
↵†These measures do not represent a comprehensive list of behavioral phenomena, which might include other measures of social preferences, statistical biases (such as base rate neglect), and strategic sophistication.
↵‡For example, between trust game behavior and risk (5, 6), discounting and risk (7⇓⇓⇓⇓–12), ambiguity and risk (9, 13⇓⇓–16), ambiguity and compound lottery aversion (17, 18), ambiguity aversion and overconfidence (19), economic behaviors and intelligence (7, 20⇓⇓⇓⇓–25), and loss aversion in risky choice and the endowment effect (26).
↵§For time preference questions extensive steps were taken to equalize transaction costs and risk of nonpayment between amounts received at different dates. See SI Appendix for more details.
↵¶One possible explanation for this is the limitations in our approach to measuring certainty effects. See discussion in SI Appendix, section S2.
↵#Note that there is little evidence that our data are censored by floor or ceiling effects: Less than 10% of subjects got 16 or more questions correct (out of 17), and less than 10% got 7 or more wrong.
↵∥Another possible explanation for the correlation between risk and time preferences is that our subjects see future payments as inherently risky.
↵††Some sessions of the experiment included further questions which were not used in this study. Some subjects answered additional common ratio-type questions aimed at measuring the magnitude of this effect at different probability levels. Some subjects completed the NEO 5-Factor Inventory, aimed at measuring the “big 5” personality traits. As with the Beck anxiety and depression inventories, these measures were not significantly correlated with our behavioral measures. However, because this 5-factor inventory was taken only for 72 subjects, we lack the power to make concrete claims in this regard. We also asked subjects to self-report their SAT scores. These measures were not used, as they appeared not to add much information to the results of the Raven’s matrix questions while reducing the sample size, as several subjects did not self-report.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1821353116/-/DCSupplemental.
Published under the PNAS license.
References
The empirical relationship between nonstandard economic behaviors
Article Classifications
- Social Sciences
- Economic Sciences
Sign up for Article Alerts
Jump to section
You May Also be Interested in
Pottery, burnt clay, and burnt flint from Jordan dated from 7752 BCE to 5069 BCE yielded a record of variations in Earth’s magnetic field strength during the Neolithic period.
Image credit: Thomas E. Levy.
Mechanics of ant tunneling could inspire the development of improved robotic mining techniques.
Image credit: Jose E. Andrade and David R. Miller (California Institute of Technology, Pasadena, CA).
When it comes to potential geoengineering initiatives, researchers and policymakers need to know what to study—and when to stop.
Image credit: Shutterstock/Venera Salman.
The gasdermin protein family has unexpectedly ancient origins.
Image credit: Alex Johnson.