Publications
An inquiry into the nature and causes of the Description – Experience gap
Journal of Risk and Uncertainty, October 2022, Volume 65, Issue: 2, pp 105–137
with Chris Starmer and Robin Cubitt
Abstract
The Description-Experience gap (DE gap) is widely thought of as a tendency for people to act as if overweighting rare events when information about those events is derived from descriptions but as if underweighting rare events when they experience them through a sampling process. While there is now clear evidence that some form of DE gap exists, its causes, exact nature and implications for decision theory remain unclear. We present a new experiment which examines in a unified design four distinct causal mechanisms that might drive the DE gap, attributing it respectively to information differences (sampling bias), to a feature of preferences (ambiguity sensitivity) or to aspects of cognition (likelihood representation and memory). Using a model-free approach, we elicit a DE gap similar in direction and size to the literature’s average and find that, when each factor is considered in isolation, sampling bias stemming from under-represented rare events is the only significant driver of the gap. Yet, model-mediated analysis reveals the possibility of a smaller DE gap, existing even without information differences. Moreover, this form of analysis of our data indicates that, even when information about them is obtained by sampling, rare events are generally overweighted.
Keywords: Decisions from experience; Decisions from description; Risk preferences; Cumulative prospect theory; Uncertainty; Ambiguity
JEL Codes: D81, C91, D91
Presented at:
- Subjective Probability, Utility and Decision Making (SPUDM), Haifa, 2017, Symposium presentation.
- Economic Science Association (ESA) European Meeting, Bergen, 2016.
- NIBS workshop: The nature of preferences and their relation to choice, Potsdam, 2016
The role of information search and its influence on risk preferences
Theory and Decision, , Volume 84, Issue 3, pp 311–339
Abstract
According to the ‘Description–Experience gap’ (DE gap), when people are provided with the descriptions of risky prospects they make choices as if they overweight the probability of rare events; but when making decisions from experience after exploring the prospects’ properties, they behave as if they underweight such probability. This study revisits this discrepancy while focusing on information-search in decisions from experience. We report findings from a lab-experiment with three treatments: a standard version of decisions from description and two versions of decisions from experience: with and without a ‘history table’ recording previously sampled events. We find that people sample more from lotteries with rarer events. The history table proved influential; in its absence search is more responsive to cues such as a lottery’s variance while in its presence the cue that stands out is the table’s maximum capacity. Our analysis of risky choices captures a significant DE gap which is mitigated by the presence of the history table. We elicit probability weighting functions at the individual level and report that subjects overweight rare events in experience but less so than in description. Finally, we report a measure that allows us to compare the type of DE gap found in studies using choice patterns with that inferred through valuation and find that the phenomenon is similar but not identical across the two methods.
Keywords: Decisions from experience; Decisions from description; Risk preferences; Cumulative prospect theory; Uncertainty; Source method; Information search
Presented at:
- Foundations of Utility, Risk and Decision Theory (FUR), Warwick, 2016.
Cite this article as:
Kopsacheilis, O. Theory Decis (2018) 84: 311. https://doi.org/10.1007/s11238-017-9623-y
Limits of the social-benefit motive among high-risk patients: a field experiment on influenza vaccination behaviour
BMC Public Health (2020) 20:240
with Burcu Isler, Ozan Isler and Eamonn Ferguson
Abstract
Influenza vaccine uptake remains low worldwide, inflicting substantial costs to public health. Messages promoting social welfare have been shown to increase vaccination intentions, and it has been recommended that health professionals communicate the socially beneficial aspects of vaccination. We provide the first test whether this prosocial vaccination hypothesis applies to actual vaccination behaviour of high-risk patients. In contrast to the literature observing intentions of low-risk populations, we found no evidence that social-benefit motivates actual vaccination behaviour among a high-risk patient population. Instead, those who selfcategorize as being in the high risk group are more motivated by the self-benefit message. Our results suggest that a stratified approach can improve coverage: even if an emphasis on social-benefit could be effective among lowrisk groups, an emphasis on self-benefit holds more promise for increasing vaccination in medical organizational settings where high-risk groups are prevalent.
Keywords: Vaccination; Influenza; Nudge; Framing; Health behaviour change; Altruism; Risk perceptions
Cite this article as:
Isler et al. BMC Public Health (2020) 20:240 https://doi.org/10.1186/s12889-020-8246-3
In Progress
The Description – Experience gap in cooperation
R&R in Games and Economics Behavior
with Ozan Isler and Dennie van Dolder
Abstract
Many people are conditionally cooperative: they cooperate if others do so as well. Conditional cooperation is usually investigated in experiments where the choices of others are known. In many circumstances, however, there is uncertainty about the cooperativeness of others. Using a novel experimental protocol, we manipulate the information subjects receive regarding the likelihood that their partner cooperates in a Prisoner’s Dilemma, and whether this likelihood is described unambiguously or learned through experience and thus ambiguous. In all treatments, subjects’ cooperation rate increases monotonically with the likelihood that their partner cooperates. Comparing decisions made under description to those made under experience, we observe a description-experience gap in which rare events appear to be more influential under experience than under description. This contrasts with earlier results from the individual choice literature, which typically finds the opposite pattern. Additional measures reveal that the gap is driven by conditional cooperators, who seek and respond to social information more than other types. We argue that stronger priors under social than individual uncertainty can account for this reversal and, in a second experiment, confirm that priors are indeed stronger under social uncertainty.
Keywords: Decisions from experience; Decisions from description; Prisoner’s dilemma; Cooperation; Ambiguity
Presented at:
- Foundations of Utility, Risk and Decision Theory (FUR), Ghent, 2022
- Slides available here.
- CeDEx- Creed- CBSS PhD conference. Amsterdam, 2019
Crowdsourcing the assessment of wine quality – evidence from Vivino ratings
R&R in Journal of Wine Economics
with Pantelis P. Analytis, Karthikeya Kaushik, Stefan M. Herzog, Bahador Bahrami, and Ophelia Deroy
Abstract
Crowdsourcing platforms—such as Vivino— that aggregate the opinions of large numbers of amateur wine reviewers represent a new source of information on the wine market. We assess the validity of aggregated Vivino ratings based on two criteria: correlation with professional critics’ ratings and sensitivity to weather conditions affecting the quality of grapes. We construct a novel dataset consisting of approximately 80,000 Vivino ratings for a portfolio of red wines from Bordeaux. We match our dataset with review scores from professional critics and add weather data from a local weather station. Vivino ratings correlate substantially with those of professional critics, but these correlations are smaller than those among professional critics. This difference can be partly attributed to differences in scope: Whereas amateurs focus on immediate pleasure, professionals gauge the wine’s potential once it has matured. Moreover, both crowdsourced and professional ratings are responsive to weather conditions known to affect wine production, and point to detrimental effects of global warming on wine quality. In sum, our results demonstrate that crowdsourced ratings are a valid source of information that generate valuable insights for both consumers and producers.
Keywords: crowd-sourcing; wine quality; wine-aging; global warming; Vivino
Presented at:
- American Association of Wine Economists (AAWE). Tbilisi/Georgia.
- Experts, markets and crowds: Taste under influence. University of Palermo.
Notes: Visualization of correlation matrix with raters ordered in ascending order according to their average inter-correlation. Reported values correspond to Pearson’s correlation coefficients (r). Right: Correlation network with each rater represented as a separate node. The thickness of the edges is proportional to the strength of correlation between the judgments of two raters. Only edges corresponding to correlations of at least r=0.40 are plotted. Color gradient is proportional to strength of correlation in both panels. Vivino: averaged ratings from Vivino users. DE: Decanter; JS: James Suckling; JR: Jancis Robinson; JL: Jeff Leve; NM: Neal Martin; RG: Rene Gabriel; TA: Tim Atkin; WA: the Wine Advocate..
Order Effects in Eliciting Preferences
with Sebastian J. Goerg
Having an accurate account of preferences help governments design better policies for their citizens, organizations develop more efficient incentive schemes for their employees and adjust their product to better suit their clients’ needs. The plethora of elicitation methods most commonly used can be broadly distinguished between methods that rely on people self-assessing and directly stating their preferences (qualitative) and methods that are indirectly inferring such preferences through choices in some task (quantitative). Alarmingly, the two approaches produce systematically different conclusions about preferences and, therefore, survey designers often include both quantitative and qualitative items. An important methodological question that is hitherto unaddressed is whether the order in which quantitative and qualitative items are encountered affects elicited preferences. We conduct three, pre-registered, studies with a total of 3,000 participants, where we elicit preferences about risk, time-discounting and altruism in variations of two conditions: ‘Quantitative First’ and ‘Qualitative First’. We find significant and systematic order effects. Eliciting preferences through qualitative items first boosts inferred patience and altruism while using quantitative items first increases the cross-method correlation
for risk and time preferences. We explore how monetary incentivization and introducing financial context modulates these results and discuss the implications of our findings in the context of nudging interventions as well as our understanding of the nature of preferences.
JEL Classification: C83, C91, D01, D91
Keywords: preferences, qualitative vs. quantitative measures, risk, altruism,
patience
Presented at:
- Subjective Probability, Utility and Decision Making (SPUDM), Vienna, 2023
- ARC seminar, Max Planck Institute for Human Development, Berlin, 2023
A horse race between elicitation methods of Prospect Theory
with Dennie van Dolder and Jörg Weber
Abstract
Eliciting risk preferences is crucial for testing and applying economic models. Traditionally, measures of risk preferences focused on Expected Utility Theory (EUT). Much empirical evidence, however, suggests that people often violate EUT’s axioms, and Tversky and Kahneman’s Cumulative Prospect Theory (CPT, 1992) has emerged as the dominant descriptive model of risky choice. To facilitate its application, several methods to elicit CPT’s parameters have been proposed. Unfortunately, different methods lead to different estimates and it remains an open question which of the methods is to be preferred. In this paper, we design a systematic framework to ‘horse-race’ elicitation methods using a host of objective benchmarks for performance. Using our framework, we examine elicitation methods for CPT that we consider representative for the state-of-the-art. The main difference between the considered approaches is whether they elicit preferences through the method of Certainty Equivalents (CE) or through choices between series of Paired Gambles (PG). Using a standard Maximum Likelihood Estimation technique, we find that the two elicitation methods produce markedly different sets of parameters, even under the same error model assumption. In terms of out-of-sample prediction, we find that estimates at the individual level outperform significantly benchmarks such as chance, the expected value model and aggregate estimations of CPT. However, we also find evidence in favour of a “home advantage” effect: the CE methods outperform the PG ones in the CE prediction set but the reverse holds true in the PG prediction set. Using Bayesian Hierarchical Modelling, an estimation technique that moderates individual-level parameters with group-level information, improves the performance of elicited estimates.
Keywords: Decision making; Cumulative prospect theory; Risk preference elicitation; Hierarchical Bayesian modelling
Presented at:
- Economic Science Association (ESA) North American Meeting, Los Angeles, 2019
- Subjective Probability, Utility and Decision Making (SPUDM), Amsterdam, 2019
- Foundations of Utility, Risk and Decision Theory (FUR), York, 2018
Leveraging Machine Learning to harness the wisdom of the crowds
Abstract
The notion that the crowd is – in expectation – wiser than any individual is centuries old. However, social scientists have only recently directed their focus towards examining the efficiency of different aggregation algorithms in harnessing this wisdom. The simplest and most prevalent algorithm relies on simply averaging individuals’ votes (majority rule). More advanced approaches involve weighting these votes either with some function of elicited confidence (e.g. linear confidence weighting) or with a type of social meta-knowledge that relies on forecasts of the average forecasts of others (e.g. Surprisingly Popular Algorithm). In this study, I compile a comprehensive dataset of ‘single-questions’—forecasting problems where the utilization of individual responses to prior problems is infeasible. Subsequently, I employ machine learning techniques to evaluate the performance of these aggregation algorithms. Notably, Machine Learning models based on Random Forests significantly outperform the most effective aggregation algorithms identified, implying the potential for developing novel, more predictive algorithms. Moreover, the results highlight the variability in the optimal method of combining features related to voting, confidence, and meta-knowledge across different question types, such as trivia, sports predictions, election outcomes, or medical diagnoses. This underscores the promise of integrating context-dependent features in the development of new aggregation algorithms.
Keywords: Wisdom Of Crowds; Forecasting; Aggregation algorithms; Machine Learning
Presented at:
- Machine+Behavior, Max Planck Institute for Human Development, Berlin, 2024
- 2-page abstract available here.