Daniel Parra

with Yuliet Verbel

This paper studies the impact of corruption on tax evasion through an experimental study. Participants decide how much of their earned income to report after learning whether there is a high or low probability of being matched with a corrupt public official who misappropriates tax revenues. We find that an environment with a high probability of being matched with a corrupt public official does not significantly increase tax evasion compared to a low-probability environment. To test whether intentions matter, we introduce two additional conditions where the probability of being matched with a corrupt public official remains high or low, but the misappropriation is determined by random assignment rather than an official's deliberate choice. Under these conditions, we find no evidence that the intentionality of corruption affects tax evasion behavior. These results identify the conditions under which neither the probability nor the intentionality of corruption is sufficient to affect compliance: when harm is indirect, beliefs are objective, and there is no strategic interaction between taxpayers and officials. We also find that tax evaders and compliers report differing social norms across all treatments, though whether these differences reflect pre-existing beliefs or post-hoc rationalization remains an open question.

with Sophia Aristizabal

Can large language models (LLMs) reliably code open-ended responses and free-form communication in experimental economics? We apply three LLMs (GPT, Claude, and Gemini) under alternative prompt designs and temperature settings to data from three published experiments that differ in coding structure, textual ambiguity, and the role of communication. In a trust game targeting explicit commitment language, all three models achieve strong agreement with human benchmarks (Cohen's κ > 0.80) and replicate the finding that promise-making predicts cooperation. In thematic coding of explanations for self- versus peer-reported AI use, LLMs recover the dominant category but over-identify rare inductively derived categories by a factor of three to five. In multi-label coding of coordination-game chat, agreement is weaker and downstream regressions do not replicate consistently, with some coefficients reversing sign. Model choice and temperature are not first-order determinants of performance. The primary constraint is the structural fit between the coding scheme and LLM capabilities: LLMs succeed when categories are deductively defined and explicitly marked in the text, and fail when coding relies on pragmatic inference or inductively derived themes. These findings have direct implications for measurement reliability and analytical transparency in experimental research that uses coded text.

This paper examines how prosocial motivations shape lying behavior in a strategic setting where two players privately observe and report a random draw. I develop a theoretical model predicting that individuals will lie less when others can lie on their behalf (strategic substitution), but more when their dishonesty generates positive externalities (prosocial lying). To test these predictions, I design a laboratory experiment using a two-player mind game in which both players benefit if at least one reports a match. Four treatments isolate the roles of strategic avoidance and prosocial incentives. The results show that prosocial motives outweigh lying aversion: participants lie significantly more when their dishonesty benefits others, even when they could avoid lying costs by allowing their partner to lie. These findings suggest that prosocial preferences can substantially offset the psychological costs of lying, but only when individuals’ actions are instrumental in producing gains for others.

with Neha Khandpur, Laura Guerrero Sánchez, Juan Carlos Londoño Roldan & Jeremy C. Young

Academic stress can change eating behavior and often leads to higher consumption of unhealthy foods. Using point-of-sale transaction data from a university supermarket in Colombia — objective purchases rather than self-reports — we estimate a difference-in-differences design with students as the treatment group and university staff as the control. During exam weeks, students increase their purchases of ultra-processed foods by 12.9 percent relative to non-exam periods (p < 0.05), while no significant changes appear among non-students. The results provide causal evidence that exam-related stress or time constraints raise demand for ready-to-heat, ultra-processed foods, and reveal clear temporal patterns with relevant implications for campus nutrition policies and stress-management interventions.

I compare two ways of eliciting cheating behavior in online experiments: an Observed-Cheating Game with on-screen random draws, and a Mind-Cheating Game where participants choose a color in their minds and then randomly draw from ten boxes with question marks. Observed online games are more likely to yield non-significant treatment differences because the effect of observability is particularly strong. Mind games solve the most prominent problems for eliciting lying by making lies unidentifiable at the individual level — and are easily implemented online or in the field.

with Tilman Fries, Uri Gneezy & Agne Kajackaite

Participants in a cheating game draw a random number and report any number they wish, receiving a monetary payoff based only on the report. We study how reports depend on the observability of both the random draw and the report. Increasing the observability of the random draw decreases cheating; increasing the anonymity of reports does not affect average reports.

with Tilman Fries

We study the effect of entitlement on willingness to lie. A model predicts that individuals who feel more entitled are encouraged to lie while others are discouraged. In a lab experiment, when performance determines income, those who earn less money lie less than those who earn more. We find no differences when endowments are windfall.

with Manuel Muñoz-Herrera & Luis Palacio

Does transparency — information precision on budget size — deter embezzlement and bribery when they co-occur? Transparency decreases embezzlement but has no significant effect on bribery. Resource managers strategically use uncertainty on the available budget in their favor, acting as if low public investment rates were a consequence of bad luck rather than misappropriation.

with Uri Gneezy & Agne Kajackaite

This paper examines the effect of noise on cognitive performance using two controlled laboratory experiments (N = 929). Participants complete memory, counting, and cognitive reflection tasks under three conditions: silence (baseline), human noise (e.g., typing, snacking, phone notifications), and intermittent machine-generated noise (air conditioning). We find that noise reduces performance on memory tasks and negatively affects mood and willingness to participate in future experiments. While both human and machine noise have similar detrimental effects on mood and willingness to participate, human noise has significant negative effects on performance, whereas machine-generated noise has rather marginal effects. These findings suggest that workplace noise may generate previously unanticipated productivity losses, reduce worker satisfaction, and decrease willingness to work in such environments.