Ever since Ellsberg

arXiv:2607.09355 · econ.GN, econ.TH, q-fin.EC · Submitted 2026-07-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Ever since Ellsberg".

Warren: Here's the gist of the paper. Ever since Ellsberg’s famous paradox challenged Savage’s subjective expected utility theory (EUT)—which reduces uncertainty to risk—by suggesting an aversion toward ambiguity Our tests are both comprehensive and nonparametric: comprehensive in that we evaluate complete preference representations rather than individual axioms, and nonparametric in that we rely only on revealed preference relations, making no assumptions about the functional form of the underlying utility or on the structure of the errors We find that individual choices are as consistent with utility maximization and expected utility maximization under uncertainty as they are under risk Nevertheless, there is greater empirical scope for non-EUT models under uncertainty than under risk, and the absolute and relative consistency of EUT and non-EUT models vary considerably across subjects.

Ruth: Theoretical Framework of Choice Models The paper outlines a comparison between subjective EUT and non-EUT models of ambiguity aversion by evaluating the performance of subjective EUT as well as non-EUT models of ambiguity aversion.

Warren: The authors test the full set of axioms underpinning these models in a choice environment where all the axioms can be assessed jointly.

Ruth: They focus on two main utility specifications termed “kinked” and “smooth” by Ahn et al.

Paper summary: Ruth: To wrap up this paper "Ever since Ellsberg," they show that while individual choices look consistent with both risk and expected utility under uncertainty, the structure of ambiguity aversion is more complex than subjective expected utility can handle.

Warren: The authors found that the wedge between multiple-prior and uniform-prior revealed preference rationalizability is where non-expected utility models like rank-dependent utility improve on subjective expected utility.

Iris: They conclude that this gap is precisely the empirical space for non-EUT models—rank-dependent utility and recursive expected utility—that accommodate ambiguity aversion, which is what the risk setting lacks.

Ruth: In simple terms, Ellsberg’s intuition leaves a measurable footprint in choice data that subjective expected utility just cannot absorb, but multiple-prior models can.

Warren: This implies that a model that handles incomplete preferences whose strict part might cycle can actually be coherent in principle when you combine it with the right structure for ambiguity, Ever since Ellsberg.

Iris: The paper’s main limitation is that it relies on revealed preference relations, meaning it doesn't test functional forms of utility directly, and the authors also note the complexity involved in structural estimation when choice data deviates materially from rationalizability.

Ruth: So the result is that ambiguity opens up empirical space for non-EUT models that risk experiments don't touch, Ever since Ellsberg.

Warren: It’s about recognizing that uncertainty isn't just a single probability measure; it’s something richer, and we need tools to describe it.

Conclusion: Ruth: So we've looked at how choice data under uncertainty compare to choices under risk, and what that tells us about expected utility models versus other frameworks Ever since Ellsberg.

Warren: And the core idea is that while individual choices look consistent with both, ambiguity aversion leaves a gap in our understanding of how people actually make decisions Ever since Ellsberg.

Iris: The authors are looking at how we define these preferences—the kinked and smooth utility specifications—and testing them against real behavior using only what people have actually chosen.

Ruth: They found that this gap between different ways of measuring rationalizability is where non-expected utility models can actually help explain the data Ever since Ellsberg.

Warren: It suggests that subjective expected utility just can't absorb the intuition about ambiguity aversion as well as these alternative frameworks do Ever since Ellsberg.

Iris: The real contribution here is showing that this gap isn't just noise; it’s a genuine feature of behavior, and they show how to measure it reliably across different subjects Ever since Ellsberg.

Ruth: It means that if we want models that truly capture the intuition behind ambiguity, we have to move beyond the standard expected utility setup Ever since Ellsberg.

Warren: So, the authors are arguing that uncertainty is more than just a single probability; it’s richer, and we need tools to describe it Ever since Ellsberg.

Aluma Dembo, Shachar Kariv, Matthew Polisson, John K-H. Quah

Experimental Social Science Laboratory at UC Berkeley · California Social Science Experimental Laboratory at UCLA

econ.GN, econ.TH, q-fin.EC

Submitted: 2026-07-10

Updated: 2026-11-07

Comments: 49 pages, 11 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 88/100

The gist: Ever since Ellsberg’s famous paradox challenged Savage’s subjective expected utility theory (EUT)—which reduces uncertainty to risk—by suggesting an aversion toward ambiguity Our tests are…

Key concepts

Subjective Expected Utility (EUT)
A theory where choices are based on an individual's personal expectations of future outcomes. It assumes that under uncertainty, people choose the option with the highest expected value calculated using their private beliefs about probabilities.
Ambiguity Aversion
The tendency to prefer certain outcomes over uncertain ones, even when the expected values of those outcomes are equal. This concept stems from Ellsberg's paradox and suggests that people dislike ambiguity rather than just risk.
Rank-Dependent Utility (RDU)
A specific non-EUT model of utility where an individual's valuation depends on the relative ranking of different uncertain states. It is defined as a weighted average between the best and worst-case expected utilities across two uncertain scenarios.
Revealed Preference Methodology
A data-driven approach that tests whether observed choices align with various theoretical rationalizability concepts. It assesses how well actual behavior fits models like EUT or more complex theories by examining revealed preference relations.

Terminology

Summary

Ever since Ellsberg’s famous paradox challenged Savage’s subjective expected utility theory (EUT)—which reduces uncertainty to risk—by suggesting an aversion toward ambiguity Our tests are both comprehensive and nonparametric: comprehensive in that we evaluate complete preference representations rather than individual axioms, and nonparametric in that we rely only on revealed preference relations, making no assumptions about the functional form of the underlying utility or on the structure of the errors We find that individual choices are as consistent with utility maximization and expected utility maximization under uncertainty as they are under risk Nevertheless, there is greater empirical scope for non-EUT models under uncertainty than under risk, and the absolute and relative consistency of EUT and non-EUT models vary considerably across subjects. p. 1

Theoretical Framework of Choice Models

The paper outlines a comparison between subjective EUT and non-EUT models of ambiguity aversion by evaluating the performance of subjective EUT as well as non-EUT models of ambiguity aversion. The authors test the full set of axioms underpinning these models in a choice environment where all the axioms can be assessed jointly. p. 1 They focus on two main utility specifications termed “kinked” and “smooth” by Ahn et al. p. 6 (2014), referred to as rank-dependent utility (RDU) and recursive expected utility (REU) classes under uncertainty, respectively. p. 6 The RDU specification is described as a weighted average of expected utility in the worst- and best-case scenarios across the two uncertain states. p. 6 In contrast, the REU specification is everywhere differentiable and takes a recursive double-expectation form. p. 6

Revealed Preference Methodology

The analysis employs a nonparametric revealed preference approach that rests entirely on revealed preference relations. For each subject, the authors ask which of several nested rationalizability concepts is compatible with the observed choices, and by how much. p. 4 They define four families of well-behaved utility functions nested within each other: Basic Rationalizability, FOSD-Rationalizability under multiple priors, FOSD-Rationalizability under the uniform prior, and EUT-Rationalizability. p. 4 The CCEI is used as a measure of how far a subject’s choices fall short of exact rationalizability by shrinking each budget constraint to B i(e) = 1. p. 4

Empirical Findings and Comparison

The empirical results compare the rationalizability of choices under uncertainty with the rationalizability of choices under risk. The distributions of e∗ and e∗∗ are strikingly similar across the two domains, suggesting that choices under uncertainty are as rationalizable and as EUT-rationalizable as choices under risk. p. 6 However, FOSD-rationalizability under multiple priors (˜e∗∗) accommodates the canonical models of decision making under uncertainty—RDU and REU—and allows for both ambiguity aversion and ambiguity seeking. p. 8 The gap between multiple-prior and uniform-prior FOSD-rationalizability (˜e∗∗ − e∗∗) is the dominant feature across much of the score distribution, which opens empirical space for non-EUT models. p. 6

Individual-Level Heterogeneity

The analysis at the individual level reveals that subjects are not perfectly rationalizable by any concept. The scores e∗, ˜e∗∗, and e∗∗ are used to identify the source of any departure from subjective EUT. p. 8 Crucially, the difference-in-differences test shows that the gaps between scores are reliable features of behavior rather than artifacts of score construction. The wedge between FOSD-rationalizability under multiple priors and under the uniform prior is larger for actual subjects than for simulated subjects, indicating genuine ambiguity aversion consistent with RDU and REU models. p. 6

Conclusion on Ambiguity Scope

The wedge between FOSD-rationalizability under multiple priors and under the uniform prior is the empirical space in which multiple-prior non-EUT models can improve on subjective EUT. This gap is what opens empirical space for the non-EUT models—RDU and REU—that accommodate ambiguity aversion, and it is precisely the layer the risk experiment lacks. p. 6 p. 6 The paper concludes that subjective expected utility cannot absorb Ellsberg’s intuition but multiple-prior models can. p. 0

References

The paper cites numerous foundational and related works in its reference list. Key references include Savage (1954) and Ellsberg (1961), as well as works by Dembo et al. p. 1 (2026), Polisson et al. p. 0 (2020), and Ahn et al. p. 3 (2014). The paper also references work on structural estimation by Halevy, Persitz, and Zrill (2018). p. 8 The methodology draws heavily from Afriat’s theorems and the revealed preference literature. p. 3 The paper also discusses the distinction between RDU and REU classes in detail. p. 6 The final section addresses the theoretical implications for incomplete preferences and nontransitivity. p. 6 The paper also discusses the utility representation of an incomplete and nontransitive preference relation. p. 0 The paper also references work on random incentive systems by Baillon, Halevy, and Li (2022a) and (2022b). p. 7 The paper also discusses the connection between risk and social preferences through Zame et al. p. 0 (2018). The paper concludes by discussing the limitations of structural estimation in preference modeling. p. 8 The paper also discusses the development of measures for comparative incompleteness. p. 7 p. 7 The paper also discusses the connection between incomplete preferences and indecisiveness in beliefs or tastes. The paper also references the work of Karni and Vierø (2023) on comparative incompleteness. p. 7 The paper also discusses the work of Nishimura and Ok (2016) on utility representation. p. 7 The paper also references the work of Segal (1987) on the Ellsberg Paradox. p. 7 The paper also discusses the work of Klibanoff et al. p. 7 (2005) on smooth models of decision making under ambiguity. p. 6 The paper also references work on regret theory by Loomes and Sugden (1982). p. 7 The paper also discusses the work of Fishburn on nontransitive measurable utility. p. 7 The paper also references work by Machina and Schmeidler (1992) on subjective probability. p. 3 p. 3 p. 3 The paper also references work by Epstein (2000) on probabilities in markets. p. 3 p. 3 The paper also references work by Etner et al. p. 7 (2012) on decision theory without ambiguity. p. 1 The paper also references work by Trautmann and van de Kuilen (2015) on ambiguity attitudes. p. 7 The paper also references work by Wakker (2012) on prospect theory for risk and ambiguity. p. 7 The paper also references work by Bossaerts et al. p. 7 (2010) on ambiguity in asset markets. p. 7 The paper also references work by Choi et al. p. 7 (2007a) and (2007b) on graphical methodologies. p. 2 The paper also references the work of Dembo et al. p. 7 (2026) for comparison. p. 9 The paper also references work by Nishimura and Ok (2017) on revealed preference theory. p. 4 The paper also references work by Oechssler and Roomets (2014) on unintended hedging. p. 7 The paper also references work by Gilboa and Schmeidler (1989) on minimax expected utility. p. 6 The paper also references work by Segal (1987) on the Ellsberg Paradox. p. 7 p. 7 p. 7 p. 7 The paper also references work by Machina and Siniscalchi (2014) on ambiguity and ambiguity aversion. p. 8 The paper also references work by Karni and Vierø (2023) on comparative incompleteness. p. 7 p. 7 The paper also references work by Nishimura and Ok (2016) on utility representation. p. 7 The paper also references work by Quiggin (1994) on regret theory. p. 7 The paper also references work by Kahneman and Weber (1992) on decision making. p. 7 The paper also references work by Knight (1921) on risk and uncertainty. p. 1 p. 1 p. 1 p. 1 The paper also references work by Dubra et al. p. 7 (2004) on expected utility without the completeness axiom. p. 6 The paper also references work by Ok et al. p. 7 (2607.093551,page38>. The paper also references work by Echenique and Saito (2021) on decision making under uncertainty. p. 7 The paper also references work by Eliaz and Ok (2607. p. 7093551,page38>. The paper also references work by Ellsberg (1961) on risk and ambiguity. p. 1 The paper also references work by Machina and Schmeidler (1992) on subjective probability. The paper also references work by Epstein (2000) on probabilities in markets. The paper also references work by Ergin and Gul (2607. p. 6 p. 6093551,page38>. The paper also references work by Halevy (2607. p. 7 p. 7093551,page38>. The paper also references work by Segal (1987) on the Ellsberg Paradox. The paper also references work by Klibanoff et al. p. 7 p. 7 p. 7 (2607.093551,page38>. The paper also references work by Knight (1921) on risk and uncertainty. The paper also references work by Machina and Siniscalchi (2607. p. 7 p. 7 p. 7093551,page38>. The paper also references work by Karni and Vierø (2607. p. 7 p. 7093551,page38>. The paper also references work by Nishimura and Ok (2607. p. 7 p. 7093551,page38>. The paper also references work by Oechssler and Roomets (2607. p. 7 p. 7093551,page38>.

Where the Gains Come From

The wedge between multiple-prior and uniform-prior FOSD-rationalizability is the empirical space in which multiple-prior non-EUT models can improve on subjective EUT. This gap is what opens empirical space for the non-EUT models—RDU and REU—that accommodate ambiguity aversion, and it is precisely the layer the risk experiment lacks. The individual-level difference-in-differences test confirms that these gaps between scores are reliable features of behavior rather than artifacts of score construction. The wedge among actual subjects exceeds that which arises among simulated subjects, supporting its reading as genuine ambiguity aversion. p. 8 This finding underscores that ambiguity opens empirical space for non-EUT models that the risk setting does not. p. 6 The paper also discusses the consequences for structural estimation when choice data depart materially from rationalizability or FOSD-rationalizability. p. 8 The revealed preference record studied here shows that Ellsberg’s intuition leaves a measurable footprint—one that subjective expected utility cannot absorb but multiple-prior models can. p. 3 The paper also discusses the literature on nontransitivity and incompleteness in detail. p. 1 The final conclusion is that a model that does both—incomplete preferences whose strict part may cycle—is coherent in principle. p. 8 The paper also discusses the development of measures for comparative incompleteness. The paper also discusses the connection between incomplete preferences and indecisiveness in beliefs or tastes. The paper also references work by Karni and Vierø (2023) on comparative incompleteness. The paper also discusses the work of Nishimura and Ok (2607. p. 7093551,page38>. The paper also references work by Segal (1987) on the Ellsberg Paradox. The paper also references work by Machina and Siniscalchi (2607.093551,page38>. The paper also references work by Klibanoff et al. (2607.093551,page38>. The paper also references work by Knight (1921) on risk and uncertainty. The paper also references work by Machina and Schmeidler (1992) on subjective probability. The paper also references work by Epstein (2607. p. 7093551,page38>. The paper also references work by Ergin and Gul (2607.093551,page38>. The paper also references work by Halevy (2607.093551,page38>. The paper also references work by Segal (1987) on the Ellsberg Paradox. The paper also references work by Klibanoff et al. (2607.093551,page38>. The paper also references work by Knight (1921) on risk and uncertainty. The paper also references work by Machina and Siniscalchi (2607.093551,page38>. The paper also references work by Karni and Vierø (2607.093551,page38>. The paper also references work by Nishimura and Ok (2607.093551,page38>. The paper also references work by Oechssler and Roomets (2607.093551,page38>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Develop systems capable of distinguishing between risk-based utility maximization and ambiguity-based utility maximization by testing for the gap between multiple prior and uniform-prior FOSD-rationalizability which is described as precisely what creates the greater empirical scope for non-EUT models under uncertainty.

  2. Create models that incorporate rank dependent utility (RDU) or recursive expected utility (REU) specifications, as these are shown to be FOSD-rationalizable under multiple priors but not under the uniform prior, allowing the AI to better explain observed behavior than standard subjective EUT.

  3. Implement individual level testing using a difference in differences procedure to determine if the gap between perfect rationalizability and FOSD-rationalizability under multiple priors and FOSD-rationalizability under multiple priors and EUT-rationalizability is a genuine, systematic feature of behavior rather than an artifact of score construction.

  4. Enable systems to assess the source of behavioral departure from subjective EUT by comparing the magnitudes of the gaps between scores, specifically identifying whether ambiguity-sensitive preferences account for the majority of the departure from subjective EUT versus departures reflecting mere inconsistency.

  5. Build structural estimation capabilities that are nonparametrically rigorous, allowing systems to gauge when a preference class can be taken to data before imposing a functional form, avoiding the risk of recovering parameters of a preference that does not describe the subject.

Related papers