Chen-Ibrahim-Sinha (1999) A New Bayesian Model for Survival Data with a Surviving Fraction

Summary

The paper proposes a new Bayesian cure-rate model for right-censored survival data from populations with a surviving ("cured") fraction, as an alternative to the standard Berkson-Gage (1952) mixture model. The new model is derived from a biological argument — a latent number of "carcinogenic" cells $N \sim \text{Poisson}(\theta)$ , each with an i.i.d. promotion time, so the event time is the minimum of the promotion times and a subject is cured when $N=0$ . This yields the population survivor function $S_{\text{pop}}(t) = \exp(-\theta F(t))$ with cure fraction $\exp(-\theta)$ . Unlike the mixture model, it has a proportional-hazards structure through the cure parameter $\theta = \exp(x'\beta)$ , is computationally convenient via latent-variable Gibbs sampling, and — crucially for Bayesian practice — yields proper posteriors under improper (e.g. uniform) priors on the regression coefficients. The authors also develop informative power priors built from historical data, prove propriety theorems, and analyze an E1684 melanoma clinical-trial dataset.

Key Claims

Promotion-time construction. For a subject with $N \sim \text{Poisson}(\theta)$ latent cells and i.i.d. promotion times $Z_j$ with CDF $F(t) = 1 - S(t)$ , the time to relapse is $T = \min(Z_1, \ldots, Z_N)$ (with $T=\infty$ , i.e. cured, if $N=0$ ). The population survivor function is $S_{\text{pop}}(t) = P(N=0) + \sum_{k\ge1} S(t)^k \frac{\theta^k e^{-\theta}}{k!} = \exp(-\theta F(t))$ , and the cure fraction is $S_{\text{pop}}(\infty) = e^{-\theta}$ .
Proportional hazards through the cure rate. The population density and hazard are $f_{\text{pop}}(t) = \theta f(t) e^{-\theta F(t)}$ and $h_{\text{pop}}(t) = \theta f(t)$ ; letting covariates enter via $\theta_i = \exp(x_i'\beta)$ gives a proportional-hazards model in $\theta$ — a property the standard mixture model cannot have in the presence of covariates.
Equivalence with the mixture model. Any standard mixture cure-rate model can be written in the proposed form and vice versa (with $\pi = e^{-\theta}$ and a corresponding non-cured survivor function), so the two are mathematically related but the promotion-time version has the better structural and computational properties.
Posterior propriety under improper priors. With covariates entering the mixture model's cure probability through a binomial regression, standard noninformative (uniform) priors give improper posteriors — so Bayesian inference there essentially requires proper priors. The proposed model instead yields proper posteriors under a wide class of improper noninformative priors, including the uniform prior on $\beta$ ; the paper gives propriety theorems.
Latent-variable Gibbs sampling. Introducing the latent cell counts $N_i$ as auxiliary data makes the complete-data likelihood tractable and allows efficient sampling of $(\beta, \text{baseline parameters}, \{N_i\})$ — a data-augmentation scheme.
Historical-data (power) priors. The construction admits an informative prior built from a previous ("historical") study by raising its likelihood to a power $a_0 \in [0,1]$ that controls how much the historical data are weighted (the Ibrahim-Ryan-Chen power prior). This is especially natural for cure models, since the belief in a surviving fraction usually rests on prior studies.

Concepts Introduced or Extended

Cure-Rate Model — the promotion-time / bounded-cumulative-hazard construction; cure fraction $e^{-\theta}$ ; proportional hazards through $\theta$
Data Augmentation — latent Poisson cell counts $N_i$ as auxiliary variables for Gibbs sampling
Gibbs Sampler — posterior simulation of the cure-rate model
Power Prior — informative prior from historical data via a raised likelihood

Entities Mentioned

Quotes

"We propose a model that is quite different from the standard mixture model for cure rates. … the model has a proportional hazards structure, with the covariates depending naturally on the cure rate."

"The model yields proper posterior distributions under a wide class of noninformative improper priors for the regression coefficients, including an improper uniform prior."

My Take

This is a foundational biostatistics paper — the "promotion-time" (or bounded-cumulative-hazard) cure model that became the standard Bayesian alternative to the Berkson-Gage mixture. Its appeal is that a clean biological story (Poisson latent cells + i.i.d. promotion times) produces a model that is simultaneously proportional-hazards, computationally friendly via latent-variable augmentation, and — the property Bayesians care about most — well-behaved under improper priors, where the mixture model famously is not. The power-prior elicitation from historical data ties it to Ibrahim-Chen's broader programme on informative priors. It sits somewhat apart from this wiki's time-series core, but connects cleanly to the data-augmentation and Gibbs machinery used throughout, and to the wiki's other survival/latent-variable material.