Deterministic variational inference for hierarchical Bayesian Cox regression: the VB–CAVI method based on quadratic upper bounds

The Cox proportional hazards model is a standard tool for analyzing right-censored survival data, but estimation may become unstable when covariates are highly correlated and event information is limited. In addition, the non-conjugate structure induced by the Cox partial likelihood makes hierarchical Bayesian inference computationally expensive. This paper develops VB–CAVI, a deterministic variational Bayes procedure implemented through coordinate ascent variational inference (CAVI), for hierarchical Bayesian Cox regression based on a fixed-curvature global quadratic upper bound for the risk-set log-sum-exp term. The bound yields a computable surrogate objective, closed-form coordinate updates, and fixed risk-set quadratic quantities that can be reused during repeated fitting. Across scenarios S1–S9, VB–CAVI showed discrimination broadly comparable to CoxPH and Ridge–Cox. Paired analyses found statistically detectable but practically modest reductions in the inverse-probability-of-censoring-weighted (IPCW) Brier score, providing no evidence of uniform predictive superiority. In a targeted S5 benchmark, uncorrected mean-field VB–CAVI substantially underestimated posterior dispersion, whereas post-convergence local-curvature covariance calibration brought the Gaussian marginal approximation much closer to the MCMC reference while preserving comparable predictive performance. Across the simulations under the raw representation, calibration increased the overall empirical coverage of nominal 95% Gaussian marginal HPD intervals from 0.145–0.460 to 0.846–0.941, although undercoverage remained in the most challenging settings. In WHAS500, repeated out-of-fold evaluation showed predictive performance comparable to the frequentist baselines and descriptively ordered risk groups. Overall, VB–CAVI provides a deterministic and computationally practical approximation for hierarchical Bayesian Cox regression, while its calibrated coefficient-level uncertainty summaries should be interpreted as approximate rather than uniformly exact in finite samples.

Authors

Institutions

Publication Details

Journal
Communications in Statistics - Simulation and Computation
Published
2026-11-14
DOI
https://doi.org/10.1080/03610918.2026.2757049
Primary Topic
Statistical Methods and Bayesian Inference
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Deterministic variational inference for hierarchical Bayesian Cox regression: the VB–CAVI method based on quadratic upper bounds

Xiaohong Wang, Yanqiu Yang, Jinqiu Li
Communications in Statistics - Simulation and Computation
Statistical Methods and Bayesian Inference
article

Deterministic variational inference for hierarchical Bayesian Cox regression: the VB–CAVI method based on quadratic upper bounds

Xiaohong Wang, Yanqiu Yang, Jinqiu Li
article en

Abstract

The Cox proportional hazards model is a standard tool for analyzing right-censored survival data, but estimation may become unstable when covariates are highly correlated and event information is limited. In addition, the non-conjugate structure induced by the Cox partial likelihood makes hierarchical Bayesian inference computationally expensive. This paper develops VB–CAVI, a deterministic variational Bayes procedure implemented through coordinate ascent variational inference (CAVI), for hierarchical Bayesian Cox regression based on a fixed-curvature global quadratic upper bound for the risk-set log-sum-exp term. The bound yields a computable surrogate objective, closed-form coordinate updates, and fixed risk-set quadratic quantities that can be reused during repeated fitting. Across scenarios S1–S9, VB–CAVI showed discrimination broadly comparable to CoxPH and Ridge–Cox. Paired analyses found statistically detectable but practically modest reductions in the inverse-probability-of-censoring-weighted (IPCW) Brier score, providing no evidence of uniform predictive superiority. In a targeted S5 benchmark, uncorrected mean-field VB–CAVI substantially underestimated posterior dispersion, whereas post-convergence local-curvature covariance calibration brought the Gaussian marginal approximation much closer to the MCMC reference while preserving comparable predictive performance. Across the simulations under the raw representation, calibration increased the overall empirical coverage of nominal 95% Gaussian marginal HPD intervals from 0.145–0.460 to 0.846–0.941, although undercoverage remained in the most challenging settings. In WHAS500, repeated out-of-fold evaluation showed predictive performance comparable to the frequentist baselines and descriptively ordered risk groups. Overall, VB–CAVI provides a deterministic and computationally practical approximation for hierarchical Bayesian Cox regression, while its calibrated coefficient-level uncertainty summaries should be interpreted as approximate rather than uniformly exact in finite samples.

Communications in Statistics - Simulation and Computation
Jilin Normal University (CN)
Statistical Methods and Bayesian Inference
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon