While individuals tend to behave consistently within a given setting as documented in revealed preference analysis, they exhibit considerable inconsistency across different settings as shown in behavioral economics literature. We investigate this narrowly rational behavior—consistency within each setting and inconsistency across settings. In a series of experiments, we compare portfolio allocations between two Arrow securities in one setting and between safe and risky assets in another. We observe that choices are consistent within each setting but largely inconsistent across settings. This inconsistency is partially due to a diversification heuristic —the tendency to allocate evenly across two securities in one setting and across two assets in another. We explore the underlying mechanisms in two additional experiments and show that the inconsistency across settings can be reduced by framing the two settings similarly but not by further decreasing the likelihood of the securities to a low level. Our study links revealed preference analysis and heuristic-based decision-making, and offers new insights into behavioral consistency across different contexts.

This study proposes a nonparametric revealed preference test to measure preference heterogeneity. Our method is assumption-free and applicable to any choice environment with revealed preference features. We develop two measures of preference heterogeneity at both the individual and group levels. We apply the method to test the much-debated greater male variability hypothesis in risk and social preferences using representative samples from the Netherlands and United States. Results indicate significantly greater male heterogeneity in risk preferences, but not in social preferences. Our study highlights the potential of the revealed preference approach as a powerful tool to understand heterogeneity.

In the context of the 2024 U.S. presidential election, experimental participants evaluate lotteries based on electoral outcomes (Trump or Harris winning), economic outcomes (improvement or decline), and the conjunctions of these outcomes. We document a conjunction fallacy in choices: participants value lotteries on conjunctive events more than single events. This pattern is stronger when conjunctive outcomes are congruent with participants' partisan identities—for example, a Harris victory combined with improved economic condition for Democrats—than when they are incongruent. Our results challenge models that satisfy dominance and point to preference-based explanations encompassing the source and the valence of uncertainty.

We extend social identity theory to study human-AI trade-offs in hiring contexts. In a baseline experiment, representative U.S. participants allocate tasks between a human worker and either another human, ChatGPT, or the foreign model DeepSeek, with explicitly varied productivity. On average, people incur costs to under-allocate tasks to AI, particularly DeepSeek. Two additional experiments, inducing identities on workers via minimal-group and political affiliations, reveal a hierarchy: in-group humans receive the most, followed comparably by out-group humans and in-group AI, and out-group AI receives the fewest. Individual perceived social distance to AI and groupy tendency jointly shape the task allocation decisions.