Cure Models for Long-Term Survivors

Introduction

Standard survival analysis assumes that every patient is ultimately susceptible to the event if follow-up is long enough. That assumption is often reasonable, but it can be inadequate in diseases where a meaningful subset of patients may remain event-free indefinitely.

Examples include some cancers after curative treatment, certain hematologic malignancies, and other settings in which the survival curve appears to flatten at a nonzero level after prolonged follow-up.

A cure model explicitly represents this possibility. Instead of treating every long-term survivor as merely censored, the model separates the population into patients who are susceptible to the event and patients who are considered cured, or no longer at risk for the event of interest.

Key idea: A long-term survival plateau is not automatically proof of cure. A cure model is appropriate only when the clinical definition of the event, the follow-up duration, and the data provide enough information to support a distinct long-term survivor component.

Why a Cure Model?

Suppose a Kaplan–Meier curve declines rapidly during the first several years and then appears to level off around 25%. A conventional survival model can describe the observed survival curve, but it does not necessarily answer the clinical question: what proportion of patients are expected never to experience the event?

A cure model addresses that question directly. It estimates a population-level cure fraction while also modeling the event-time distribution among patients who remain susceptible.

Incidence component: estimate the probability that a patient belongs to the cured versus susceptible group.

Latency component: among susceptible patients, model the time until the event occurs.

Combine the components: obtain the overall population survival curve, which can approach a nonzero asymptote.

What Does "Cured" Mean?

In a cure model, "cured" is a statistical and clinical concept tied to the specified event. It does not necessarily mean that every disease process has been biologically eliminated.

For example, in an oncology study the event might be disease-specific death, recurrence, or progression. A patient classified as cured for that endpoint is modeled as having no future occurrence of that particular event.

This distinction matters because the cure fraction depends on the endpoint. A patient might be cured with respect to recurrence but remain at risk for death from another cause.

The Mixture Cure Model

The most intuitive formulation is the mixture cure model. Let \(\pi\) denote the probability that a patient is cured, and let \(S_u(t)\) denote the survival function among susceptible patients.

The population survival function is then:

S(t)=\pi+(1-\pi)S_u(t)

The first term represents patients who never experience the event. The second term represents susceptible patients who have not yet experienced the event by time \(t\).

If the susceptible survival function eventually approaches zero, then:

\lim_{t\rightarrow\infty}S(t)=\pi

Interpretation: Under the mixture cure model, the asymptotic survival probability is the estimated cure fraction, provided the susceptible survival distribution eventually tends to zero and the model is identifiable from the available follow-up.

Susceptible Versus Cured Patients

Introduce a latent indicator \(C\), where \(C=1\) denotes a cured patient and \(C=0\) denotes a susceptible patient. The observed event time is available only when the susceptible patient experiences the event before censoring.

Latent group	Event risk	Observed contribution
Cured	No future event of interest	Survival contribution \(\pi\)
Susceptible	Positive event risk	Survival contribution \((1-\pi)S_u(t)\)

The latent group is not directly observed. A patient who remains event-free at the end of follow-up could be cured or simply susceptible but censored before the event occurs. This is the central statistical challenge of cure modeling.

Adding Covariates to the Cure Probability

The cure fraction can depend on baseline characteristics. A common approach is a logistic regression for the probability of being cured:

\operatorname{logit}(\pi_i) = \gamma_0+\gamma_1 z_{i1}+\cdots+\gamma_q z_{iq}

For example, suppose \(z_i\) includes treatment, disease stage, and a baseline biomarker. A positive coefficient means that increasing the corresponding covariate increases the odds of belonging to the cured group, under the parameterization above.

Important: The cure-model incidence coefficient is not automatically a hazard ratio. It describes the odds of cure. Its interpretation is fundamentally different from a Cox-model hazard ratio.

Modeling Survival Among Susceptible Patients

The second component describes the event-time distribution among susceptible patients. Several distributions can be used, including Weibull, exponential, log-normal, log-logistic, and more flexible parametric distributions.

For a Weibull susceptible population, one convenient survival function is:

S_u(t)=\exp\left[-\left(\frac{t}{\lambda}\right)^k\right]

where \(k\) is a shape parameter and \(\lambda\) is a scale parameter under this parameterization.

The resulting population survival curve is:

S(t)=\pi+(1-\pi) \exp\left[-\left(\frac{t}{\lambda}\right)^k\right]

This formulation makes the long-term plateau explicit: as \(t\) becomes large, the susceptible component disappears and the survival curve approaches \(\pi\).

Likelihood Contribution

Let \(T_i\) be the observed follow-up time and \(\delta_i\) the event indicator, where \(\delta_i=1\) means that the event was observed.

For an observed event, the patient must be susceptible. The density contribution is therefore proportional to:

L_i = (1-\pi_i)f_u(T_i) \qquad\text{if }\delta_i=1

For a censored observation, the patient may either be cured or still event-free but susceptible:

L_i = \pi_i+(1-\pi_i)S_u(T_i) \qquad\text{if }\delta_i=0

The full likelihood is:

L(\theta) = \prod_{i=1}^{n} \left[(1-\pi_i)f_u(T_i)\right]^{\delta_i} \left[\pi_i+(1-\pi_i)S_u(T_i)\right]^{1-\delta_i}

The parameters are estimated by maximizing this likelihood, or an equivalent log-likelihood.

Why Censored Patients Create the Difficulty

An observed event identifies the patient as susceptible. A censored patient does not. A patient censored at five years could be a cured patient or a susceptible patient whose event simply has not occurred yet.

Consequently, the estimated cure fraction depends on the information contained in the entire follow-up distribution, not simply on the percentage of patients who are event-free at a particular landmark.

Do not equate the observed event-free proportion with the cure fraction. For example, if 30% of patients are event-free at five years, that does not mean the cure fraction is 30%. Some susceptible patients may experience events after five years.

Mixture Cure Model Versus Cox Regression

Feature	Cox model	Mixture cure model
Population assumption	All patients remain in the risk set conceptually	Population contains cured and susceptible subgroups
Long-term survival plateau	Not modeled as a cure fraction	Explicitly represented
Primary effect measure	Hazard ratio	Cure odds plus susceptible-population survival parameters
Latent subgroup	No	Yes
Interpretation of long-term survivors	Still potentially event-free at time \(t\)	Can be modeled as cured with probability \(\pi\)

A cure model is not simply a replacement for Cox regression. If a proportional hazards model adequately describes the scientific question and no meaningful cure fraction is supported, Cox regression may remain the appropriate primary analysis.

Worked Example: Long-Term Survivors After Treatment

Consider an illustrative oncology study with 200 patients followed after definitive treatment. The Kaplan–Meier curve falls substantially during the first several years and then appears to flatten at roughly 25%.

Planning feature	Illustrative value
Total patients	200
Endpoint	Disease recurrence
Follow-up	Up to 8 years
Approximate late survival plateau	25%
Candidate cure model	Mixture cure + Weibull latency

The 25% plateau is a reason to investigate a cure model, not proof that exactly 25% of patients are cured.

A simple illustrative parameterization

Suppose the fitted model estimates a cure fraction of:

\hat{\pi}=0.25

and the susceptible survival distribution is Weibull with \(k=1.5\) and \(\lambda=30\) months. Then the population survival at 60 months is:

S(60) = 0.25+ 0.75 \exp\left[-\left(\frac{60}{30}\right)^{1.5}\right]

Numerically, this is approximately:

S(60)\approx0.294

So the overall predicted 5-year survival is about 29.4%, while the long-run survival approaches 25%. The difference represents susceptible patients who are still event-free at five years but are expected to experience the event later.

This is the central benefit of the model: it distinguishes patients who are predicted to survive indefinitely from patients who are simply surviving longer within the susceptible population.

Interpreting Treatment Effects

Treatment can affect either component of a cure model. It may increase the probability of being cured, alter the event-time distribution among susceptible patients, or do both.

Pattern	Possible interpretation
Higher cure fraction only	Treatment converts a larger proportion of patients into long-term survivors
Different latency only	Treatment delays events among patients who remain susceptible
Both components change	Treatment affects both long-term cure probability and event timing
No clear component difference	Evidence for a cure effect may be weak or the follow-up may be insufficient

This decomposition can be clinically more informative than reporting one overall hazard ratio when the treatment is expected to produce a durable long-term benefit.

Non-Mixture Cure Models

A second important family is the non-mixture, or promotion-time, cure model. Rather than writing the population survival function as a direct mixture of a cured and susceptible survival distribution, the model often uses a form such as:

S(t)=\exp\{-\theta F_0(t)\}

where \(F_0(t)\) is a baseline cumulative distribution function and \(\theta\) controls the expected number or intensity of latent causes.

The cure fraction is then:

\pi=\lim_{t\rightarrow\infty}S(t)=e^{-\theta}

Non-mixture models can be useful when the scientific mechanism is naturally described in terms of multiple latent causes of failure. They should not be treated as interchangeable with mixture cure models: their parameters and interpretations differ.

When Is a Cure Fraction Identifiable?

The hardest practical issue is often not estimation but identifiability. If follow-up ends while a substantial fraction of susceptible patients could still experience the event, the data may not distinguish a true cure plateau from very long but finite survival.

For a cure model to be credible, investigators should consider:

Length and completeness of follow-up
Number and timing of observed events
Whether the Kaplan–Meier curve has genuinely stabilized
Clinical knowledge about plausible late events
Whether alternative latency distributions give similar cure estimates
Whether the estimated cure fraction is stable under sensitivity analyses

Practical rule: A visually flat Kaplan–Meier curve is evidence to investigate a cure model, not a guarantee that the cure fraction is estimable.

Choosing the Latency Distribution

The susceptible survival distribution can materially affect the estimated cure fraction, especially when follow-up is not extremely long.

Distribution	Useful when
Exponential	A constant susceptible hazard is scientifically plausible
Weibull	A simple monotone increasing or decreasing hazard is reasonable
Log-normal	The susceptible hazard may rise and then fall
Log-logistic	A non-monotone hazard and heavier tail are plausible
Flexible parametric	The event-time shape is not well represented by a standard distribution

Model selection should be based on clinical plausibility, graphical diagnostics, fit statistics, and sensitivity of the estimated cure fraction—not simply on which model gives the lowest information criterion.

Cure Fraction and the Kaplan–Meier Plateau

A useful diagnostic is to compare the estimated cure fraction with the observed tail of the Kaplan–Meier curve.

If the fitted model implies a cure fraction of 20%, but the Kaplan–Meier estimate continues falling substantially late in follow-up, the model deserves scrutiny. Conversely, a fitted cure fraction far above the observed late survival may also indicate model misspecification or insufficient follow-up.

Plot the Kaplan–Meier curve and inspect whether a credible plateau is present.

Fit one or more plausible cure-model specifications.

Compare predicted survival with the nonparametric Kaplan–Meier estimate.

Examine sensitivity of the cure fraction to the latency distribution and covariate specification.

Model Diagnostics

Cure-model diagnostics should be broader than a single goodness-of-fit number. Useful checks include:

Observed versus model-predicted survival curves
Tail behavior and agreement with the apparent plateau
Residual or quantile diagnostics for the latency component
Influence of individual observations
Stability under alternative latency distributions
Calibration of predicted cure probabilities
Sensitivity to covariate functional form

A model can have an acceptable global fit while producing an unstable cure fraction. The tail of the survival curve is especially important because the cure parameter is learned from long-term behavior.

R Implementation

The flexsurvcure package provides a practical framework for fitting parametric mixture and non-mixture cure models in R. The exact model specification should be checked against the package version used for the analysis.

library(survival)
library(flexsurvcure)

fit <- flexsurvcure(
  Surv(time, status) ~ treatment + stage,
  data = dat,
  dist = "weibullPH",
  mixture = TRUE
)

summary(fit)

Here status=1 indicates the event and status=0 indicates censoring. The model estimates the cure component together with the susceptible survival distribution.

Estimating Cure Probability for Individual Patients

When covariates are included in the incidence component, the fitted model can produce patient-specific estimates of the probability of belonging to the cured group.

For a logistic cure model:

\hat{\pi}_i = \frac{\exp(\hat{\gamma}_0+\mathbf{z}_i^\top\hat{\boldsymbol{\gamma}})} {1+\exp(\hat{\gamma}_0+\mathbf{z}_i^\top\hat{\boldsymbol{\gamma}})}

These predictions should be described as model-based probabilities, not as definitive classifications of individual patients. A patient with an estimated 70% cure probability is not literally observed to be 70% cured.

Comparing Two Treatment Groups

Suppose a randomized trial compares treatment A with treatment B. A cure model can estimate separate cure fractions:

\pi_A \qquad\text{and}\qquad \pi_B

A clinically useful presentation may include:

Estimated cure fraction in each treatment group
Difference or ratio of cure probabilities, with confidence intervals
Survival among susceptible patients
Overall model-based survival curves
Restricted mean survival time when a cure interpretation is uncertain

This allows investigators to distinguish a treatment that creates more long-term survivors from one that mainly delays events among patients who remain susceptible.

Cure Models and Competing Risks

Competing risks require particular care. A patient who dies from another cause before recurrence cannot simply be treated as an ordinary non-event if the scientific endpoint is recurrence.

The appropriate analysis depends on the estimand. Cause-specific hazards, cumulative incidence, relative survival, and disease-specific cure models answer different questions.

Do not use "cure" as a shortcut for "event-free." If competing events prevent observation of the event of interest, the cure interpretation must be aligned with the endpoint and estimand.

Cure Models Versus Restricted Mean Survival Time

Restricted mean survival time (RMST) summarizes the average event-free time up to a clinically specified horizon. A cure model instead attempts to separate an indefinitely surviving subgroup from susceptible patients.

Question	Useful method
How much event-free time is gained through 5 years?	RMST
Is there evidence of a long-term survival plateau?	KM curve + cure model
What proportion may never experience the event?	Cure model
Is a proportional hazards assumption plausible?	Cox diagnostics
What happens after a landmark time?	Landmark analysis

These methods are complementary rather than mutually exclusive. In a clinical trial with uncertain cure behavior, reporting both RMST and a carefully justified cure analysis can provide a more complete picture.

Common Mistakes

Calling every long-term survivor cured. A patient who has not experienced an event during observed follow-up is not automatically cured.
Equating the late Kaplan–Meier value with the cure fraction. The Kaplan–Meier curve may still decline after the apparent plateau.
Ignoring insufficient follow-up. A cure fraction is difficult to identify when susceptible patients can still have late events beyond the observation window.
Reporting cure odds as hazard ratios. The incidence component and latency component have different interpretations.
Using only one latency distribution. Cure estimates can be sensitive to the assumed susceptible survival model.
Overinterpreting individual cure probabilities. They are model-based probabilities, not directly observed patient labels.
Ignoring competing risks. The definition of cure must match the event and estimand.
Assuming a cure model is automatically superior to Cox regression. The model should be motivated by the clinical setting and observed survival behavior.

Practical Analysis Workflow

A defensible clinical analysis can follow this sequence:

Define the endpoint. Specify exactly what event is being modeled and whether competing events matter.

Inspect follow-up. Examine the Kaplan–Meier curve, number of events, censoring pattern, and duration of observation.

Assess plausibility. Use clinical knowledge to determine whether a durable cure interpretation is scientifically meaningful.

Fit candidate models. Compare plausible mixture and latency specifications.

Check sensitivity. Evaluate whether cure estimates change materially under alternative distributions and covariate models.

Report uncertainty. Provide confidence intervals and explain the assumptions behind the cure interpretation.

How to Report a Cure-Model Analysis

A clinical report should make the model structure transparent. At minimum, describe:

The endpoint and censoring definition
The rationale for considering a cure model
The cure-model family and incidence link
The latency distribution
Covariates included in each component
The estimated cure fraction with confidence interval
Estimated survival among susceptible patients
Overall model-predicted survival
Model diagnostics and sensitivity analyses
Important limitations caused by follow-up or sparse late events

Clinical interpretation: The strongest cure-model result is not simply "the cure rate was 25%." It is a coherent statement that combines the estimated long-term survivor fraction, uncertainty, observed follow-up, model assumptions, and clinical context.

Key Takeaways

A cure model is useful when a nonzero long-term survival plateau is clinically plausible.
Mixture cure models separate the population into cured and susceptible components.
The cure fraction is the long-term survival limit under the usual mixture-model assumptions.
The susceptible component still requires an appropriate event-time distribution.
Cure and latency effects can provide different clinical information about treatment benefit.
Long follow-up is critical for identifying a cure fraction.
Alternative model specifications and sensitivity analyses are essential.
RMST, Kaplan–Meier analysis, Cox models, and cure models can be complementary.

References and Further Reading

Berkson, J. & Gage, R.P. (1952). Survival curve for cancer patients following treatment. Journal of the American Statistical Association, 47, 501–515.

Boag, J.W. (1949). Maximum likelihood estimates of the proportion of patients cured by cancer therapy. Journal of the Royal Statistical Society: Series B, 11, 15–53.

Farewell, V.T. (1982). The use of mixture models for the analysis of survival data with long-term survivors. Biometrics, 38, 1041–1046.

Maller, R.A. & Zhou, S. (1996). Survival Analysis with Long-Term Survivors. Wiley.

Kuk, A.Y.C. & Chen, C.H. (1992). A mixture model combining logistic regression with proportional hazards regression. Biometrika, 79, 531–541.