Identify why federation underperforms for particular clients

Identify the client-specific distributional or consortium-relative factors that cause FedNova federated income estimation to underperform local-only training for certain state-level clients, including low-sample clients such as Hawaii.

Background

The local-only counterfactual shows that FedNova improves out-of-time R2 for most clients, but four of the 50 state-level clients—Hawaii, California, New York, and Texas—perform better with local-only models. Hawaii is especially notable because it belongs to the low-sample group, where federation generally produces the largest gains.

The paper reports that client-specific distributional characteristics and their relationship to the rest of the consortium may explain these exceptions, but the available analysis does not determine the cause. Understanding these cases is important for deciding when personalization or locally calibrated components should supplement the global federated model.