Fairness and Robustness Under Cross-State Deployment: A Study on Tabular ML Models

Algorithmic fairness interventions are often evaluated under static data assumptions, yet deployed models frequently encounter population shifts that can alter their fairness behavior. In this work, we study fairness mitigation and robustness under cross-state deployment for tabular machine learning models. Using logistic regression and gradient boosting, we evaluate baseline performance and apply a simple post-processing mitigation strategy based on group-specific decision thresholds to reduce the Equal Opportunity gap. While threshold adjustment improves fairness on the training population, we observe that fairness properties do not fully transfer when the model trained on California data is deployed on Texas data without retraining. Although predictive accuracy remains relatively stable, disparities measured by Demographic Parity increase under deployment. These results highlight that fairness guarantees observed during model development may not persist across populations, emphasizing the need to evaluate fairness under realistic deployment shifts. This record includes the main paper and accompanying LaTeX source files and high-resolution figure files.