Optimal Cox regression under federated differential privacy: coefficients and cumulative hazards
We study two canonical problems in distributed survival analysis under federated differential privacy (FDP): estimation of cumulative hazard functions and Cox regression coefficients, allowing for heterogeneous per-server sample sizes and privacy budgets. To quantify the fundamental cost of privacy, we derive minimax lower and upper bounds that match up to poly-logarithmic factors, revealing server-level phase transitions between private and non-private regimes. Our algorithms are computationally-efficient and are based on adapting differentially private binary tree and gradient descent algorithms to handle estimating equations with dependent terms. We also study three related problems: estimation of the cumulative baseline hazard and survival functions in the Cox model under FDP, and estimation of regression coefficients under the federated label differential privacy, a relaxed privacy notion. For these problems, we develop efficient algorithms, establish theoretical guarantees and investigate minimax optimality. Numerical experiments, including a real-data application, support the theoretical findings. The proposed methods are implemented in the accompanying R package FDPCox.