Kalbfleisch (1978) Non-parametric Bayesian Analysis of Survival Time Data

Summary

This paper gives a Bayesian treatment of Cox's (1972) semi-parametric proportional-hazards regression model by placing a gamma-process prior on the cumulative hazard function — a random-probability-measure (Bayesian nonparametric) prior over the unspecified baseline. With the baseline hazard integrated out, Kalbfleisch derives a marginal posterior for the regression coefficients $\beta$ and shows how to estimate the underlying survival distribution as well. A key result is that Cox's partial likelihood emerges as the limiting case of this Bayesian analysis as the gamma-process prior becomes diffuse, giving a Bayesian justification for the partial likelihood and a coherent way to also estimate the baseline survivor function.

Key Claims

• The Cox proportional-hazards model. Hazard $\lambda(t\mid z)=\lambda_0(t)\exp(z\beta)$, equivalently survivor $P(T>t\mid z)=\exp\{-\Lambda_0(t)\exp(z\beta)\}$, with the baseline cumulative hazard $\Lambda_0$ left unspecified; the survivor form accommodates discrete, continuous, or mixed baselines.
• Gamma-process prior on the cumulative hazard. $\Lambda_0(\cdot)$ is modeled as a gamma process — an independent-increments (Lévy) random measure — so the baseline is a genuinely nonparametric object with a prior, in the spirit of Ferguson's (1973) random probability measures.
• Marginalize the nuisance baseline. Treating $\Lambda_0$ as a nuisance parameter with the gamma-process prior, one obtains the marginal distribution of the data as a function of $\beta$ alone, from which $\beta$ is estimated; the posterior also yields an estimate of the baseline survival curve.
• Partial likelihood as a diffuse-prior limit. As the gamma-process prior is made vague, the marginal likelihood for $\beta$ converges to Cox's (1975) partial likelihood — the same object Kalbfleisch-Prentice (1973) derived as a marginal rank likelihood — so partial likelihood is recovered as a special/limiting case of the Bayesian analysis.
• Coherent baseline estimation. Unlike the partial-likelihood approach, which discards information about the baseline, the Bayesian scheme estimates the regression parameters and the survival distribution within one coherent framework.

Concepts Introduced or Extended

• Parametric Survival Model
• Counting Process

Entities Mentioned

• John D. Kalbfleisch

Quotes

"A Bayesian analysis of the semi-parametric regression and life model of Cox (1972) is given. The cumulative hazard function is modelled as a gamma process."

My Take

The paper is an early and clean instance of Bayesian nonparametrics doing real work: instead of leaving the Cox baseline hazard as an unspecified nuisance (the partial-likelihood move) or forcing it into an exponential/Weibull shape, it puts a prior on the whole cumulative hazard via a gamma process and integrates. The payoff is twofold — a Bayesian rationale for why the partial likelihood is a sensible thing to maximize (it is the diffuse-prior limit), and the ability to report the baseline survival curve with honest uncertainty rather than treating it as an afterthought. It complements the wiki's survival cluster from the Bayesian side: Aalen's Nelson-Aalen estimator is the frequentist nonparametric estimate of exactly the cumulative hazard that Kalbfleisch here endows with a gamma-process prior, and the gamma process is a survival-analytic cousin of the Dirichlet-process priors used elsewhere for nonparametric Bayes. The practical caveat of its era is computational — the marginalization is analytically delicate — which is why the approach became routine only once MCMC made posterior sampling for such Lévy-process priors feasible.