Collaborative Inference for Accelerated Failure Time Model Using Clinical Center‐Level Summary Statistics

Mengtong Hu, Xu Shi, Ziyang Gong, Peter X.‐k. SongStatistics 2025 · DOI

TL;DR

This paper introduces CAFTA, a novel privacy-preserving framework for analyzing time-to-event data from multiple clinical centers using Accelerated Failure Time (AFT) models. CAFTA allows for robust statistical inference by only sharing summary statistics, not raw patient data, overcoming critical privacy concerns in multi-center studies.

What Problem Does This Paper Address?

Multi-center clinical studies are invaluable for increasing sample size and improving the generalizability of findings. However, pooling raw patient data across these centers raises significant privacy concerns, often making centralized analysis infeasible. Existing distributed methods for survival analysis, particularly for the Cox proportional hazards model, have limitations, including potential information leakage even with aggregated data. The Accelerated Failure Time (AFT) model offers a more interpretable alternative to the Cox model, directly modeling the effect of covariates on the time to an event. However, developing privacy-aware, distributed methods for AFT models, especially those that avoid complex partial likelihood constructions and can handle data from centers with varying sample sizes, remains a challenge.

Methods

The core of this paper is the Collaborative Accelerated Failure Time Analysis (CAFTA) framework. CAFTA is an incremental estimation method that allows for distributed AFT model fitting without sharing individual-level data. Instead, each participating clinical center computes and shares summary statistics (specifically, score functions and sensitivity matrices) with the next center in a sequential manner. This approach is inspired by "renewable estimation" techniques.

Key methodological aspects include:

  • Parametric AFT Models: CAFTA focuses on a class of parametric AFT models, specifically those with Weibull, log-normal, and log-logistic error distributions.
  • Generalized Gamma (GG) Distribution: The paper leverages the generalized gamma distribution to unify Weibull and log-normal AFT models, facilitating goodness-of-fit assessments.
  • Distributed Likelihood Ratio Test: A novel distributed likelihood ratio test is proposed to compare candidate parametric AFT models, allowing for model selection without centralized data.
  • Large-Sample Properties: The authors establish theoretical guarantees for CAFTA, demonstrating its consistency, asymptotic normality, and efficiency, comparable to an oracle estimator that has access to all data.
  • No Iterations or Specific Sample Size Requirements: Unlike some prior distributed methods, CAFTA does not require iterative cross-center communication or specific sample size ratios between centers, making it more practical.

Datasets

The paper uses the Scientific Registry of Transplant Recipients (SRTR) data for its real-world application. This dataset contains information on kidney transplant recipients across the United States, organized by 11 geographic regions. The analysis focuses on predicting 5-year death-censored graft failure. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Privacy-Preserving AFT Analysis is Achievable: The most significant insight is that robust and efficient AFT model inference can be performed in a distributed manner, solely relying on summary statistics, thus preserving patient privacy.
  • AFT Models Offer Advantages in Distributed Settings: The AFT model's direct interpretation and avoidance of partial likelihood construction make it more amenable to distributed, summary-statistic-based inference compared to the Cox model.
  • CAFTA's Robustness to Heterogeneity: The simulation studies suggest CAFTA is more robust than traditional meta-analysis methods (Meta-FE, Meta-RE) when dealing with potential heterogeneity across study sites.
  • Practicality of Incremental Updates: The sequential, incremental update mechanism of CAFTA is computationally efficient and avoids the complexities of iterative, multi-round communication often seen in other distributed learning approaches.
  • Goodness-of-Fit is Crucial: The development of a distributed likelihood ratio test for model selection highlights the importance of assessing model fit in a collaborative setting.

Key Results

  • Simulation Studies: CAFTA demonstrated performance comparable to the oracle estimator (centralized analysis) across various event rates and error distributions (Weibull, log-normal, log-logistic). It showed lower bias and better coverage probabilities than traditional meta-analysis methods, particularly in challenging scenarios (e.g., low event rates).
  • Real-World Application (SRTR Data): CAFTA was successfully applied to analyze kidney transplant data, identifying significant risk factors for graft failure. The results from CAFTA were largely consistent with the oracle analysis, while meta-analysis methods showed more deviations, especially for certain covariates. For instance, the Meta-FE and Meta-RE methods failed to detect the significant effect of a Hispanic donor, which CAFTA and the oracle method did.

Limitations and Future Work

  • Parametric Assumption: CAFTA, as presented, relies on parametric AFT models. The authors acknowledge that extending it to semi-parametric AFT models is technically challenging due to the need to compare residuals across sites.
  • Homogeneity Assumption: The current CAFTA framework assumes homogeneity of model parameters across sites. While simulations explore some heterogeneity, future work could explicitly develop methods to account for or model such cross-region heterogeneity.
  • Sequential Order: While simulations suggest CAFTA is robust to the order of sites, the paper recommends starting with the largest site for numerical stability. Further research could explore optimal ordering strategies.
  • Extension to Other Models: The authors suggest future work could incorporate more flexible AFT models, such as those using polynomial splines, into the CAFTA framework.
  • Graphical Diagnostics: The paper mentions the utility of graphical diagnosis for goodness-of-fit and suggests investigating its application within collaborative settings.

Summary updated

Abstract

Multi-center clinical research offers numerous advantages, including the ability to obtain a larger combined sample size and to reduce center-specific insufficiency and imbalance in event rates, ultimately leading to more robust and generalizable findings. This paper develops a collaborative analytic framework for survival data analysis using summary statistics and the Accelerated Failure Time (AFT) model, a popular alternative to the Cox proportional hazards model for the analysis of time-to-event data. The AFT model directly accounts for the effects of the covariates on times to event, rather than through hazard functions, with no proportionality assumption required compared to the Cox model. Given that it bypasses the construction of partial likelihood, it gives rise to more flexibility in integrative analyses of survival data collected from multiple clinical sites. Our proposed distributed inference method focuses on a class of parametric AFT models with Weibull, log-normal, and log-logistic distributions for time-to-event outcomes, with a distributed likelihood ratio test established under the generalized gamma distribution to assess the goodness-of-fit across different candidate parametric AFT models. We present large-sample properties for the proposed distributed method and illustrate its performance through simulation experiments and a real-world data example on kidney transplantation.
BibTeX
@inproceedings{hu2025collaborative,
  author = {Mengtong Hu and Xu Shi and Ziyang Gong and Peter X.‐K. Song},
  title = {{Collaborative Inference for Accelerated Failure Time Model Using Clinical Center‐Level Summary Statistics}},
  booktitle = {Statistics},
  year = {2025},
  doi = {10.1002/sim.70279},
}
Figure