Generalized Promotion Time Cure Models

• GPTCMs are cure-rate survival models defined by a latent promotion-time mechanism that yields a non-mixture cure fraction from competing causes.
• The model generalizes through subtype-heterogeneity, transformation families, and flexible regression functions to address a variety of censoring schemes and data structures.
• Advanced estimation paradigms like Bayesian, semiparametric, and deep learning techniques are employed to robustly model time-to-event data within GPTCM frameworks.

Generalized Promotion Time Cure Models (GPTCMs) are cure-rate survival models built around a promotion-time or latent competing-risks mechanism in which the observed event time is generated by the earliest activation among an unobserved number of latent causes. In the canonical non-mixture form, if $N\sim \mathrm{Poisson}(\theta)$, the latent activation times share cdf $F(t)$, and $T=\min_i Z_i$, then the population survival is $S_{\text{pop}}(t)=\exp[-\theta F(t)]$, so that $\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$ is a cure fraction. Contemporary uses of the label GPTCM do not refer to a single unique extension: the literature uses it for biologically structured heterogeneity inside the promotion-time mechanism, transformation families containing promotion-time and mixture cure models, generalized covariate-rate formulations, flexible Bayesian and semiparametric baseline specifications, and implementations tailored to current-status, interval-censored, clustered, zero-inflated, and high-dimensional right-censored data (Zhao et al., 2024, Pal et al., 2023, Portier et al., 2018, Hariharan et al., 2024).

1. Promotion-time foundation and cure-model semantics

The promotion-time cure model is motivated by a latent process in which a subject retains a random number of active carcinogenic or failure-generating agents after treatment or exposure. In the classical formulation, $N$ is Poisson, the latent times $Z_i$ are i.i.d. with cdf $F(t)=1-S(t)$, and the recurrence or failure time is $T=\min\{Z_i:i=1,\dots,N\}$, with the convention that $N=0$ implies no recurrence. This yields

$F(t)$0

and hence

$F(t)$1

The positive limiting survival probability is therefore generated endogenously by the latent-count mechanism rather than by a separate cured-versus-susceptible mixing weight (Zhao et al., 2024, Louzada et al., 2015).

This non-mixture interpretation is central to GPTCMs. In the canonical promotion-time setting, the population hazard is proportional in the latent intensity parameter, for example $F(t)$2 when $F(t)$3, and the cure fraction depends only on the latent count parameter. Several later formulations preserve this mechanism while changing the regression map, the latent promotion-time law, the observation scheme, or the dependence structure, but they retain the basic logic that cure corresponds to the event of having no effective latent causes (Medina-Olivares et al., 2023).

2. What is generalized in a GPTCM

One line of generalization changes the biology of the latent process rather than the basic cure mechanism. In the subtype-heterogeneous construction, the total active cell count is decomposed as $F(t)$4 with $F(t)$5, $F(t)$6, and subtype-specific promotion-time survivals $F(t)$7. The resulting population survival is

$F(t)$8

equivalently $F(t)$9 with $T=\min_i Z_i$0. Standard PTCM is recovered when all $T=\min_i Z_i$1 are identical. Here the generalization is a mixture of subtype-specific promotion-time survivals inside the exponential survival generator, intended to represent intratumor or intra-sample heterogeneity rather than a population mixture of latent classes (Zhao et al., 2024).

A second line of generalization is transformation-based. The Box-Cox transformation cure model defines

$T=\min_i Z_i$2

with $T=\min_i Z_i$3 yielding the promotion-time cure model and $T=\min_i Z_i$4 yielding the mixture cure model. In this sense, promotion-time cure becomes one endpoint of a broader transformation family, and intermediate $T=\min_i Z_i$5 values interpolate between the two regimes (Pal et al., 2023).

A third line generalizes the regression-side rate function. The model

$T=\min_i Z_i$6

replaces the classical one-index form by a user-chosen positive covariate map $T=\min_i Z_i$7 and a separate scalar $T=\min_i Z_i$8. This preserves the promotion-time interpretation while moving the baseline object from a normalized cdf to $T=\min_i Z_i$9, which eliminates the Lagrange-multiplier constraint that complicates classical nonparametric maximum likelihood estimation (Portier et al., 2018).

A fourth line generalizes the family of promotion-time laws under a common Bayesian engine. The R package "bayesCureRateModel" presents a GPTCM family indexed by $S_{\text{pop}}(t)=\exp[-\theta F(t)]$0 and a promotion-time law $S_{\text{pop}}(t)=\exp[-\theta F(t)]$1, with built-in exponential, Gompertz, log-logistic, Weibull, Gamma, Lomax, Dagum, finite mixtures of gamma distributions, and user-defined proper distributions. In that family, $S_{\text{pop}}(t)=\exp[-\theta F(t)]$2 with $S_{\text{pop}}(t)=\exp[-\theta F(t)]$3 yields the promotion-time cure model, $S_{\text{pop}}(t)=\exp[-\theta F(t)]$4 the negative binomial cure model, and $S_{\text{pop}}(t)=\exp[-\theta F(t)]$5 the mixture cure model; the case of no cured subjects is represented at $S_{\text{pop}}(t)=\exp[-\theta F(t)]$6 rather than as a boundary case (Papastamoulis et al., 2024).

Not every use of “generalized” introduces a new abstract family. For current-status data, one Bayesian promotion-time paper explicitly states that its generalized aspect is not a new GPTCM class in the abstract sense, but rather the use of a more general, flexible baseline $S_{\text{pop}}(t)=\exp[-\theta F(t)]$7 under Bayesian estimation. In GPTCM terms, that contribution is best viewed as a Bayesian current-status implementation of the promotion-time member of the broader biological cure-model class (Hariharan et al., 2024).

3. Regression structure, cure probability, and hazard interpretation

In the canonical promotion-time regression form, covariates enter through a log-link on the promotion intensity: $S_{\text{pop}}(t)=\exp[-\theta F(t)]$8 This structure couples incidence and latency: the same regression term controls both the long-run cure fraction and the event-time behavior among susceptibles. In that respect, the cure rate is not a separate mixing weight, unlike a standard mixture cure model (Hariharan et al., 2024).

Some GPTCMs allow covariates to affect both the latent promotion count and the latent promotion-time distribution through different linear predictors. In the Bayesian P-spline formulation,

$S_{\text{pop}}(t)=\exp[-\theta F(t)]$9

so that

$\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$0

Because $\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$1 enters the latent distribution, the population hazard ratio is generally time-varying even though the latent susceptible model is Cox proportional hazards (Bremhorst et al., 2013).

Other formulations generalize the predictor itself. In the Deep Promotion Time Cure Model, the classical linear index $\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$2 is replaced by a learned function $\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$3, giving

$\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$4

The model therefore preserves the cure interpretation while allowing nonlinear and interaction effects in high-dimensional covariates (Medina-Olivares et al., 2023).

In clustered marginal promotion-time cure modeling, the same exponential-rate structure is retained subjectwise: $\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$5 This supports two equivalent interpretations of $\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$6: as log hazard-ratio effects on the overall survival distribution and as effects on cure probability through $\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$7 (Xiao et al., 14 May 2025).

4. Observation schemes and likelihood construction

For ordinary right-censored data $\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$8, GPTCM likelihoods usually retain the survival-density factorization

$\lim_{t\to\infty}S_{\text{pop}}(t)=\exp(-\theta)>0$9

often with latent cure indicators $N$0 introduced to form a complete-data likelihood and simplify posterior or EM-based computation. In the generalized Bayesian family implemented in "bayesCureRateModel", this latent-variable augmentation supports Gibbs updates for $N$1 and model-based computation of subject-specific cured probabilities $N$2 (Papastamoulis et al., 2024).

Current-status data change the problem fundamentally. Each subject is observed only once at a monitoring time $N$3, with

$N$4

Under independent non-informative current-status censoring, the likelihood becomes

$N$5

so the data inform the model only through whether the event has occurred by the examination time, not through the actual event time. The paper explicitly treats this as type-I interval censoring and discretizes the baseline on the observed monitoring times by a step-function representation $N$6 parameterized through step increments $N$7 (Hariharan et al., 2024).

Interval-censored GPTCMs use yet another likelihood geometry. For observed data $N$8, the semi-parametric transformation-cure likelihood is written

$N$9

with latent susceptibility indicators introduced only for right-censored subjects in the EM algorithm (Pal et al., 2023).

Some applications require exact mass at time zero. The zero-inflated promotion cure rate model adds a third class beyond susceptibles and cured subjects: $Z_i$0 where $Z_i$1 is the zero-inflated mass, $Z_i$2 the cure fraction, and $Z_i$3 the promotion-model susceptible survival. In that setting, the likelihood distinguishes exact zeros, uncensored positive times, and right-censored times, producing a three-way mixture tailored to immediate-failure phenomena such as bank-loan fraud (Louzada et al., 2015).

5. Estimation paradigms and computational strategies

Bayesian computation has become a major route for GPTCM fitting. In the current-status promotion-time model, priors are placed independently on the regression vector and the step-function baseline parameters, and posterior computation uses an adaptive Metropolis-Hastings algorithm because the full conditionals are not of closed form. The algorithm starts by maximizing the posterior to obtain initial values, uses a Gaussian random-walk proposal with covariance based on the inverse observed Fisher information at the current mode, later updates that covariance adaptively, and summarizes posterior inference through posterior means under squared-error loss. The same paper uses LPML via CPO, DIC, and scaled CPO plots for model assessment (Hariharan et al., 2024).

The generalized Bayesian family in "bayesCureRateModel" employs a Metropolis-coupled Markov chain Monte Carlo sampler embedded in a parallel tempering / MC$Z_i$4 scheme. Its inference engine combines Gibbs sampling for latent cure indicators with either componentwise Metropolis-within-Gibbs log-normal proposals or MALA updates whose gradient is computed numerically. Because the posterior can be multimodal, the authors prefer MAP estimation over posterior means and use heated posteriors to improve exploration of separated modes; the package also supports prediction of survival, cumulative hazard, hazard, conditional cure probability, Cox–Snell residuals, and FDR-based identification of cured subjects (Papastamoulis et al., 2024).

Semiparametric likelihood theory provides a different computational solution. The extension $Z_i$5 rewrites the model in terms of $Z_i$6, yielding a nuisance maximizer with explicit closed form for each fixed $Z_i$7. The profile objective then resembles Cox-model estimation, and the paper derives consistency, asymptotic normality, weak convergence of $Z_i$8, and closed-form asymptotic variances for $Z_i$9, $F(t)=1-S(t)$0, and $F(t)=1-S(t)$1. This is the main sense in which the model is “closer in spirit to the Cox model” than the classical promotion-time cure model (Portier et al., 2018).

Several semiparametric and Bayesian extensions focus on flexible baseline or covariate effects. Bayesian P-spline estimation models the latent baseline hazard as

$F(t)=1-S(t)$2

places a difference-penalized Gaussian prior on spline coefficients, and uses MCMC with adaptive Metropolis updates for regression and spline parameters and Gibbs updates for smoothing hyperparameters. Local-polynomial estimation in the Yakovlev model instead treats $F(t)=1-S(t)$3 with $F(t)=1-S(t)$4 unknown, derives a strictly concave local likelihood for fixed baseline parameter $F(t)=1-S(t)$5, and proves that the nonparametric component has the usual local-polynomial bias $F(t)=1-S(t)$6 and variance $F(t)=1-S(t)$7, while $F(t)=1-S(t)$8 remains root-$F(t)=1-S(t)$9 consistent under undersmoothing conditions (Bremhorst et al., 2013, Lin et al., 2019).

Deep-learning and correlated-data methods extend GPTCM computation into large-scale and dependent settings. Deep-PTCM trains the cure model end-to-end by backpropagation, jointly optimizing the neural predictor and a piecewise exponential $T=\min\{Z_i:i=1,\dots,N\}$0, and optionally decomposes the predictor into orthogonalized linear and nonlinear parts for interpretability. For clustered right-censored data, the semiparametric marginal promotion-time model estimates regression effects by generalized estimating equations or quadratic inference functions, with QIF shown to be optimal within the stated class of estimating equations and to have smaller asymptotic variance than GEE (Medina-Olivares et al., 2023, Xiao et al., 14 May 2025).

6. Applications, interpretation, and recurring methodological issues

GPTCMs have been used for epidemiological studies, destructive testing, oncology, smoking cessation, marriage duration, mortgage default, kidney transplant survival, periodontal disease, and bank-loan fraud. Recent examples include lung-tumor and breast-cancer current-status analyses, where cure-model interpretation was motivated by a long survival plateau and cure rates varied with environment, tumor stage, grade, nodal status, and cathepsin-D status; a first-marriage duration analysis implemented in "bayesCureRateModel"; a large US mortgage-loan study in which Deep-PTCM improved discrimination and calibration; a smoking-cessation interval-censoring study in which the fitted Box-Cox parameter selected the mixture-cure endpoint; a breast-cancer semiparametric analysis with nonlinear link functions; a melanoma trial in which treatment acted more strongly on the cure component than on latent incubation time; and a periodontal application in which QIF produced the smallest standard errors and revealed effects of mobility, attachment loss, filled teeth, and bleeding on probing (Hariharan et al., 2024, Papastamoulis et al., 2024, Medina-Olivares et al., 2023, Pal et al., 2023, Portier et al., 2018, Bremhorst et al., 2013, Xiao et al., 14 May 2025).

A persistent misconception is to equate GPTCMs with ordinary finite-mixture survival models. The subtype-heterogeneous formulation explicitly contrasts its structure with the naive model $T=\min\{Z_i:i=1,\dots,N\}$1 and argues that this is not biologically meaningful for recurrence, because recurrence is caused by the combined latent propagation of multiple cells rather than by a sample being “from one subtype only.” Likewise, in the standard promotion-time regression form the cure probability is typically $T=\min\{Z_i:i=1,\dots,N\}$2, not a separate logistic or multinomial mixing weight, although exact-zero extensions deliberately add such a component when the application requires a point mass at time zero (Zhao et al., 2024, Louzada et al., 2015).

Methodological difficulties recur across the literature. Identifiability may require at least one continuous covariate with non-negligible effect on the promotion-rate parameter in the generalized Bayesian family, insufficient follow-up can make simultaneous cure-side and latency-side covariate effects hard to disentangle in the Bayesian P-spline model, threshold-based cure classification introduces sensitivity in local-polynomial Yakovlev estimation, and the Box-Cox parameter $T=\min\{Z_i:i=1,\dots,N\}$3 can be difficult to estimate precisely because the likelihood is flat in the $T=\min\{Z_i:i=1,\dots,N\}$4-direction for intermediate values. Posterior multimodality is another practical issue, motivating parallel tempering and MAP-based summaries in some Bayesian GPTCMs. A plausible implication is that “GPTCM” is best understood not as a single fixed model but as a family of non-mixture cure formulations that preserve the promotion-time latent-cause interpretation while varying the latent distribution, regression architecture, censoring regime, and inferential machinery (Papastamoulis et al., 2024, Bremhorst et al., 2013, Lin et al., 2019, Pal et al., 2023).