This paper compares causal forests with varying-coefficient penalized spline estimators for heterogeneous treatment effects. We conduct simulations with randomized treatment assignment that vary heterogeneity, predictor dimension, and discrete nuisance structure. Performance is measured by out-of-sample mean squared error. Forests perform best in Wager–Athey designs with sharp treatment-effect transitions. Penalized splines dominate when heterogeneity is smooth, when the baseline mean and treatment effect differ in structure, and in mixed designs with continuous, binary, and group-level covariates. As group cardinality rises, forest performance deteriorates, while mgcv random-effect smooths handle group structure effectively and remain easy to implement.

Large Language Models (LLMs) such as ChatGPT are transforming how scientists conduct and validate research, offering promise as tools to improve scientific reproducibility. However, computational reproducibility and error detection remain expensive and labor-intensive. We experimentally test how collaboration between researchers and LLM assistants influences the reproduction of quantitative social science findings across different levels of AI autonomy. We randomly assigned 288 researchers to 103 teams working under three conditions: human-only, AI-assisted (using ChatGPT as a collaborative tool), or AI-led (ChatGPT operating with minimal human oversight). Teams reproduced published results from leading social science journals, detected coding errors, and proposed robustness checks. Human-only and AI-assisted teams achieved comparable reproduction rates (94% vs. 91%) and performed similarly on most outcomes, except human-only teams identified significantly more major coding errors. Both substantially outperformed AI-led teams, which achieved only a 37% reproduction rate, detected fewer errors across all categories, proposed weaker robustness checks, and required more time. This autonomous approach, however, likely represents only a lower bound of AI capabilities. Despite rapid model advances, expert human judgment currently remains indispensable for reliable empirical verification. While AI assistance did not degrade most outcomes, it provided no measurable advantages and was associated with reduced detection of major errors. However, the 37% autonomous reproduction rate indicates that AI could provide value in settings where scale or cost constraints preclude human review of papers, even though general-purpose LLMs offer no immediate advantages for human-supervised verification.

This project examines how mandatory diversity statements in faculty job applications influence hiring and student outcomes. Using comprehensive text from JOE and APSA postings linked to institution–year–discipline hiring records, we classify DEI-related requirements and estimate their effects using staggered-adoption DiD estimators with entropy-balancing weights.

We study whether Italy’s Superbonus 110% created incentives for anomalous billing by construction firms. The program subsidized home renovation through transferable tax credits, separating payment from verification of the underlying work. Using panel data on 66,920 Italian construction firms from 2015–2024 and the February 2023 closure of credit transfers to new projects, we find that firms with unusually high revenue per employee and elevated profit margins became substantially more common during the program, with the increase concentrated among firms that retained access to credit transfers. Predetermined bank-branch density also strongly predicts credit-driven firm formation, linking the response to local capacity to intermediate the credits. Because accounting data cannot distinguish inflated billing from legitimate price increases under subsidized demand, we interpret the evidence as anomalous rather than fraudulent billing.

We develop a statistical framework for diagnosing parameter identifiability and uncertainty in stochastic agent-based models (ABMs). The approach combines Monte Carlo experiments with simulation-based confidence intervals, providing generalizable tools for calibration, validation, and sensitivity analysis in complex ABMs.

We extend the Schelling segregation model by replacing traditional, rule-based agents with Large Language Model (LLM) agents that make residential decisions using natural language reasoning grounded in social context. To our knowledge, this is the first application that substitutes the mechanical agents of the Schelling model with LLM-driven agents. We compare LLM agent behavior across four social and two other contexts: Income (High vs. Low), Ethnic (Asian vs. Hispanic), Racial (White vs. Black), Political (Liberal vs. Conservative), two pairs of colors to serve as control groups, and a mechanical baseline (the standard Schelling model). We measure the resulting segregation via a Dissimilarity Index, which measures how dissimilar neighborhoods are in an overall geographical area. We find LLM agents partially reproduce empirical segregation patterns collected at the census block level and measured across counties.

This project examines how the expansion of dollar stores and low-nutrition retail environments contribute to obesity and metabolic health risks. Using store rollouts and quasi-experimental variation in food environments, I study how changes in access to calorie-dense, nutrient-poor options affect chronic disease outcomes.