Bayesian Survival Model Based on Moment Characterization

Julyan Arbel, Antonio Lijoi, Bernardo NipotiSpringer Proceedings in Mathematics & Statistics 2015 · DOI · arXiv · OpenAlex

Abstract

Bayesian nonparametric models are powerful tools for survival analysis, yet they frequently struggle to provide full posterior distributions for non-linear functionals, such as median survival time. This paper introduces a novel methodology that leverages posterior moments to reconstruct these distributions, offering more accurate credible intervals. The approach is implemented in the momentify R package, facilitating practical application for researchers.

Introduction

In survival analysis, clinicians and researchers require more than a simple "average" survival curve. Key metrics like median survival time and its associated uncertainty (credible intervals) are critical for decision-making. While Bayesian nonparametric "marginal methods" are popular for simplifying complex infinite-dimensional models, they typically only provide point estimates (the posterior mean). These methods often fail to estimate non-linear functionals or provide valid uncertainty measures, frequently underestimating variability because they ignore the full randomness of the underlying survival process. This research aims to bridge this gap by providing a framework for full posterior density reconstruction.

Methodology

The proposed approach centers on Moment Characterization to recover the posterior distribution of survival functionals:

  1. Hazard Mixture Models: Survival data is modeled using hazard rate functions driven by Completely Random Measures (CRM), such as the gamma process.
  2. Posterior Moments: Instead of relying solely on the mean, the authors derive closed-form expressions for posterior moments of any order for the survival function $S(t)$ using a Gibbs sampler.
  3. Density Reconstruction: The unknown posterior density is approximated using a sequence of these moments via an expansion of Jacobi polynomials. This orthogonal basis is particularly well-suited for densities with compact support.
  4. Rejection Sampling: To ensure the final approximation is a valid probability distribution (integrating to 1 and remaining non-negative), a rejection sampler is employed.
  5. Software Implementation: The methodology is accessible via the momentify R package.

Experimental Validation

The method was validated through controlled simulations:

  • Bimodal Mixture: A mixture of two Beta distributions tested the ability to recover complex, two-peaked densities.
  • Weibull Survival Data: Datasets of sizes $n=20, 100, \text{ and } 500$ were generated from a Weibull distribution (shape=2, scale=2) to compare estimates against a known ground truth median survival time ($m_0 \approx 1.665$).

Key Insights and the "Rule of Ten"

A significant finding from the simulations is the "Rule of Ten": using the first 10 moments ($N=10$) is generally sufficient to capture the shape of the posterior distribution accurately, even for complex bimodal shapes.

Density reconstruction using an increasing number of moments (N=2 to N=10). The black line represents the approximation, while the red line is the target density.

Furthermore, the study highlights that standard marginal intervals are often too narrow. By marginalizing out the underlying random measure, these traditional methods ignore a significant source of uncertainty. In contrast, by reconstructing the full density at each time point, the moment-based method allows for the calculation of any functional—medians, modes, or specific quantiles—with appropriate uncertainty measures.

Results

  • Small Sample Accuracy: For $n=20$, the moment-based 95% credible interval (0.88 - 1.80) successfully captured the true median (1.665), whereas the standard marginal interval (1.09 - 1.57) was too narrow and missed the truth.
  • Convergence: As sample size increased to $n=500$, both methods converged toward the true value, but the moment-based approach maintained a more theoretically sound representation of posterior spread.
  • Visual Fit: As shown in the figures, from seven moments onward, the approximated density becomes nearly indistinguishable from the true density.

Conclusion and Limitations

While powerful, the method involves computational overhead due to the Gibbs sampler required for moment estimation. Additionally, the current closed-form derivations are specific to hazard mixture models; applying this to other Bayesian nonparametric priors would require new derivations. Finally, while $N=10$ serves as a robust heuristic, a formal mathematical criterion for selecting the optimal number of moments remains an area for future research.

Summary updated

BibTeX
@inproceedings{arbel2015bayesian,
  author = {Julyan Arbel and Antonio Lijoi and Bernardo Nipoti},
  title = {{Bayesian Survival Model Based on Moment Characterization}},
  booktitle = {Springer Proceedings in Mathematics \& Statistics},
  year = {2015},
  doi = {10.1007/978-3-319-16238-6\_1},
}
Figure