Research

We study how effectively an AI agent aligns with a human principal's choices under risk when given stated versus revealed preference information. We conduct an online experiment in which subjects state their preferences through written instructions ("prompts") and reveal them through a series of binary lottery choice problems ("data"). We find that on average, an AI agent given data predicts subjects' choices more accurately than an AI agent given prompts. We attribute this gap to subjects' difficulty in translating their own preferences into written instructions. Subjects also misperceive the relative performance of the two agents and delegate to the less accurate agent too often. Moreover, providing AI agents with more information does not necessarily improve their performance, since different AI models resolve conflicts between stated and revealed preferences differently. Overall, our results highlight revealed preferences as a powerful mechanism for communicating human preferences to AI agents, but their success depends on careful implementation.

We empirically investigate the consideration and choice behaviors of individuals under uncertainty. Our experiment allows subjects to reveal their stochastic or deterministic consideration and choice. Most of the subjects have stochastic (limited) consideration and stochastic choice. Furthermore, consideration and choice data mostly violate monotonicity and regularity axioms, respectively, but they satisfy a proposed stochastic transitivity axiom. We show that observing both consideration and choice is crucial to drawing meaningful welfare inferences.

Economic models are founded on parsimony and interpretability, which is achieved through axioms on choice behavior. We empirically evaluate the predictive accuracy of economic models of choice under risk and ambiguity, and the strength of their axiomatic foundations, using complementary methods of completeness (Fudenberg et al., 2022) and restrictiveness (Fudenberg et al., 2023), respectively. To better understand the tradeoff between the two concepts, we additionally relate their performance to machine learning models. We use budgetary choice environments with three dimensions to provide a strong test of axioms. We show that adding a third dimension of choice marginally reduces completeness of economic models, but significantly increases restrictiveness. Economic models are also more complete than machine learning models, and are significantly more restrictive. These results are robust to considering an environment of choice under ambiguity rather than choice under risk. Overall, economic models capture the behavior of individual subjects well.

This study investigates the transferability of economic models for individual decision-making across different risk domains, specifically comparing performance between two- and three-state budgetary environments. Utilizing within-subject laboratory data, we evaluate the ability of Expected Utility Theory (EUT), Disappointment Aversion (DA), and machine learning models to predict choices when estimated in a simpler two-state environment and applied to a more complex three-state environment at the individual level. Our findings reveal two key insights: (i) there is substantial transferability across domains for the vast majority of subjects; and (ii) EUT demonstrates substantial transferability, maintaining approximately 92.9% of its within-domain predictive accuracy when generalized across domains, outperforming both DA and machine learning models in terms of predictive consistency. These results underscore the robustness of parsimonious economic models, particularly EUT, in providing reliable extrapolations across experimental contexts, suggesting their utility in applications where predictions span diverse risk settings.

We study the measure of restrictiveness proposed in Fudenberg et al. (2023) to evaluate complexity of economic models using synthetic data. We show that Rademacher complexity is an affine transformation of a particular case of a consistent finite-sample estimate of restrictiveness. We highlight that this connection comes with both desirable and undesirable traits regarding bounds on generalization error and (in)ability to distinguish between falsifiable models, which produces a tradeoff in measure selection dependent on the relative importance of each trait to the analyst.

We investigate the predictive value-add of auxiliary covariates in a choice under risk setting. We start with a data set representative of the Dutch population, and simulate different levels of data availability by selectively removing demographic covariates, subject identifiers, or both. We use expected utility theory (EUT) as a benchmark model and evaluate its out-of-sample prediction performance against machine learning (ML) models. We show that identifying information is more valuable than demographic data, although both show significant improvement over choice data alone. EUT is competitive with ML models, in particular outperforming them on subjects whose choices are consistent with (monotonic) utility maximization. There is little heterogeneity across demographic groups. Overall, our results demonstrate the predictive power of simple identifying information while emphasizing the continued relevance of EUT amid advances in ML and AI.

We compare the predictive performance of economic models of choice under risk to various machine learning (ML) models by presenting nearly 1,000 subjects with a consumer decision problem: the selection of a bundle of contingent commodities from a budget line. We compare models' predictions at the individual level and relate them to the consistency of decisions with revealed preference axioms. Using dual measures of completeness and restrictiveness, we show that in our setting Expected Utility Theory (EUT) performs as well as non-EUT and all ML models, with a wider margin as choices align more with utility maximization.

Recent evidence suggests that non-isolation behavior could significantly impact laboratory experiments using the random problem selection (RPS) payment mechanism through lottery integration. Theoretical work also highlights social preferences that can violate statewise monotonicity, a necessary and sufficient condition for incentive compatibility with the RPS payment mechanism in case of lottery integration. Additionally, non-isolation can influence decisions through nonconsequential dynamic concerns. In a series of three simple and parsimonious experiments and three tests, we examine the occurrence of the two kinds of non-isolation and reversal behaviors. We find significant evidence for positive reversal behavior, where subjects are more likely to make a fair choice if there is an alternative possible realization of an unfair outcome (which they chose themselves). In addition, the lower bounds for the prevalence of non-isolation in terms of lottery integration and dynamic non-consequential concern are estimated to be approximately 10% and 20%, respectively.

An r-matrix is a matrix with symbols in {0, 1, ..., r-1}. A matrix is simple if it has no repeated columns. For a family of r-matrices F, we define forb(m, r, F) as the maximum number of columns of an m-rowed, simple r-matrix A such that F is not a row-column permutation of a submatrix of A for all F in F. We investigate the asymptotic bounds of special families of forbidden configurations Sym(F), a symmetric family of matrices based on a (0,1)-matrix F. The previously known lower bound constructions required F to have no constant row. We introduce a new lower bound construction that drops this requirement and is a better bound in certain cases. We will also investigate the effects of constant rows from the upper bound perspective, including upper bounds of block matrices, matrices consisting of entirely constant rows.