"Is the Golden Ticket Tainted? Testing Standardized Tests with AI," sole authored, Job Market Paper, under review at The Economic Journal. [pdf]

Standardized test scores are signals in college admissions, yet the measurement properties of these instruments receive little empirical attention. This paper uses large language models (LLMs) as stable external benchmarks to audit item-level changes in the SAT math section from 2011 to 2023. Estimating a pooled item response theory model on 1291 math items evaluated by 20 LLMs, I find a downward drift in item difficulty beginning in 2017, with discrimination remaining broadly stable. This shift moves test information towards lower ability regions, reducing signal precision at the upper tail where selective institutions screen applicants. These diagnostics suggest that where the SAT is most informative has changed over time, with implications for efficiency of student institution matching and the distributional consequences of admissions policies.

"Artificial Test-Takers as Transformed Controls: Measuring SAT Difficulty Drift and Student Performance" with Saani Rawat, revise and resubmit at Frontiers in AI. [pdf]

Introduction: Standardized test score trends are widely used to track student performance and inform policy, but they are difficult to interpret when exam content changes over time. We introduce an artificial test-taker framework that uses a fixed large language model as a stable benchmark to measure SAT Math difficulty drift and construct difficulty-adjusted measures of student performance. Results: The artificial test-taker framework indicates a statistically significant decline in SAT Math difficulty of 0.21σ relative to 2012. After adjusting for test difficulty using the transformed-control benchmark, student performance declines by 34 points in Average Difference in Scores (ADS) from 2012 to 2023. Heterogeneity analyses show that these declines are not uniform across racial groups.

"Efficient Analysis of Experimental Panel Data: A Comparison of Categorical and Dynamic Models" with Jeffrey A Mills. [pdf]

This paper evaluates the finite-sample performance of categorical-time versus dynamic specifications for estimating treatment effect trajectories in short experimental panels. Using Monte Carlo simulations and STAR data, we show that autoregressive and log-trend models substantially improve precision and statistical power when treatment effects evolve smoothly, reducing path RMSE by 30-40 percent relative to categorical-time specifications. These efficiency gains arise from restricting the conditional mean rather than modeling error covariance. However, categorical-time models exhibit greater robustness to irregular treatment paths. The results highlight a fundamental bias-variance tradeoff, with specification choice depending on the plausibility of dynamic structure and inferential objectives.

"From Production to Delegation: AI Adoption with Endogenous Restriction Risk," sole authored. [pdf]

I examine how AI adoption reshapes human capital through skill substitution from production to AI-dependent management capabilities. Workers face endogenous restriction risk that increases with aggregate AI adoption, creating a coordination failure: individuals over-invest in AI-dependent skills without internalizing systemic fragility. Incorporating image concerns reveals how social stigma initially delays adoption but generates rapid cascades once norms shift, amplifying vulnerability. The decentralized equilibrium features excessive AI reliance and insufficient investment in robust fallback skills relative to the social optimum.