Nathan Hudson

Doubly-Synthetic Control: Using AI Personality Agents for Policy Counterfactuals

This project develops a doubly-synthetic control methodology that uses AI-based personality agents as synthetic populations instead of real control units. The method creates artificial agents characterized by fundamental personality traits and calibrates them using large language models to match observed survey response distributions. These agents can then be weighted to represent any target population and queried under counterfactual scenarios to estimate treatment effects. The key advantage of this method is its ability to generate counterfactuals without requiring actual control units. Agents are weighted to match the treated population’s pre-treatment characteristics, and their outcomes are then queried under alternative scenarios. Validation tests whether agents can reproduce historical response patterns, establishing that they capture stable behavioral mechanisms suitable for counterfactual inference. This methodology would enable causal inference in settings where traditional synthetic control cannot operate, including evaluating one-time national policies with no comparable units, estimating the effects of hypothetical interventions before implementation, or analyzing historical events where control groups are unavailable. By generating synthetic populations from personality models rather than relying on existing donor pools, his doubly synthetic control approach could substantially expand the scope of evidence-based policy analysis.