Worked Analysis of Long-Term Survivor Data

From a Cohort to a Cure Model Analysis

A long-term plateau in observed survival can be scientifically plausible when some individuals will never experience the event of interest. A mixture cure model represents this possibility by separating the probability of belonging to a cured group from the event-time distribution among susceptible individuals. The analysis must do more than produce a fitted cure fraction: it should check whether the data support the model, examine calibration over time, and report predictions for clearly defined covariate profiles.

Consider a hypothetical cohort of 240 patients followed for up to 10 years after treatment. There are 120 patients in each of two treatment groups. During follow-up, 83 experience the event, 34 are censored before 10 years, and 123 are known to be event-free through 10 years, including those administratively censored at that time. These summaries make a long-term survivor component plausible enough to investigate, but they do not establish that any particular patient is cured. The fitted model below is assumed to use the individual event or censoring times and group indicators, not merely these aggregate counts.

Definition: In a mixture cure model, $\pi(x)$ is the probability of belonging to the cured component for covariates $x$. Susceptible individuals have event-time survival $S_u(t\mid z)$, where $z$ denotes covariates in the latency component. The marginal event-free survival is $$ S(t\mid x,z)=\pi(x)+\{1-\pi(x)\}S_u(t\mid z). $$

This analysis uses treatment indicator $z=0$ for the reference group and $z=1$ for the other group. The incidence model is logistic, and the susceptible event-time distribution is Weibull with a proportional-hazards treatment effect:

\operatorname{logit}\{\pi(z)\}=\alpha_0+\alpha_1z, \qquad S_u(t\mid z)=\exp\left[-\left(\frac{t}{4}\right)^{1.4}\exp(\beta z)\right].

Time is measured in years. The baseline susceptible cumulative hazard is $(t/4)^{1.4}$, and $\exp(\beta)$ is the susceptible-group hazard ratio for $z=1$ versus $z=0$. The fitted coefficients used for illustration are $\hat\alpha_0=-0.4055$, $\hat\alpha_1=0.4055$, and $\hat\beta=-0.8419$. These are hypothetical fitted estimates, not results from a real study.

Assumptions Before Interpreting the Fit

The model treats cured status as a latent, fixed membership: cured individuals do not experience the modeled event during the time horizon represented by the model, while susceptible individuals follow the specified latency distribution. Here the Weibull susceptible survival tends to zero as $t$ increases, so the model’s limiting marginal survival is $\pi(z)$. That limiting interpretation depends on the chosen model and is not a direct classification of observed survivors.

The right-censoring assumption is that, conditional on the modeled covariates, the censoring time is independent of the event time. If patients with different unmeasured event risks are systematically lost to follow-up at different times, the observed data may not identify the event-time distribution without additional modeling. The logistic incidence form and Weibull latency form must also be adequate for the data; neither is guaranteed by the presence of a plateau. Finally, there must be enough follow-up and enough late observations to distinguish a genuine cure fraction from a susceptible distribution with a long tail. These conditions connect directly to the earlier material on identifiability and follow-up in mixture cure models.

For an observed event at time $t$, the event density comes from the susceptible component. For a person censored at $t$, the likelihood contribution is the probability of remaining event-free to $t$, which includes both cured and still-event-free susceptible individuals:

L_i= \left[\{1-\pi(z_i)\}f_u(t_i\mid z_i)\right]^{\delta_i} \left[\pi(z_i)+\{1-\pi(z_i)\}S_u(t_i\mid z_i)\right]^{1-\delta_i}.

Here $\delta_i=1$ indicates an observed event and $\delta_i=0$ indicates right censoring. The likelihood relies on the stated censoring assumption and the specified incidence and latency models. It is the individual event and censoring times, together with covariates, that are used to estimate the coefficients.

Worked Example: Estimating the Incidence Component

For the reference treatment group, use $\hat\alpha_0=-0.4055$. Its fitted cure probability is obtained by applying the inverse logit:

\hat\pi(0)=\frac{\exp(-0.4055)}{1+\exp(-0.4055)} =\frac{0.6666}{1+0.6666} \approx 0.4000.

Thus the model estimates that 40.00% of patients with $z=0$ belong to the cured component. For $z=1$, the fitted log odds are $-0.4055+0.4055=0$, so

\hat\pi(1)=\frac{\exp(0)}{1+\exp(0)}=0.5000.

The fitted incidence component therefore predicts a 10.00 percentage-point higher cure fraction in group $z=1$. This is a model-based covariate association; it does not by itself show that treatment caused a change in cure probability.

Calculating Covariate-Specific Survival

The marginal prediction combines the incidence and latency predictions. At a finite time, susceptible individuals can still be event-free, so marginal survival is generally greater than the cure probability. At the limit as $t$ tends to infinity, the Weibull susceptible survival tends to zero and the marginal survival tends to the fitted cure probability.

Worked Example: Ten-Year Predictions for Two Profiles

For the reference group at 10 years, the baseline susceptible cumulative hazard is

\left(\frac{10}{4}\right)^{1.4} =2.5^{1.4} \approx 3.607.

The susceptible survival is therefore $\exp(-3.60675)\approx 0.02714$. Combining it with the estimated cure probability gives

\hat S(10\mid z=0) =0.4000+(1-0.4000)(0.02714) =0.4163.

For $z=1$, the fitted susceptible hazard ratio is $\exp(-0.8419)\approx 0.4309$. Its susceptible cumulative hazard at 10 years is $3.60675(0.4309)\approx1.554$, giving susceptible survival $\exp(-1.554)\approx0.2113$. Hence

\hat S(10\mid z=1) =0.5000+(1-0.5000)(0.2113) \approx 0.6057.

The model predicts 10-year event-free survival of 41.63% for the reference profile and 60.57% for the $z=1$ profile. These predictions combine different estimated cure probabilities and different susceptible latency distributions; neither contrast should be reported as a cure-fraction contrast alone.

A useful long-term interpretation is the probability of belonging to the cured component among people who have remained event-free to a specified time. This is not the same quantity as the population cure fraction. By Bayes’ rule, conditional on covariates,

P(\text{cured}\mid T>t,z) =\frac{\pi(z)}{\pi(z)+\{1-\pi(z)\}S_u(t\mid z)}.

The numerator is the probability of being cured and event-free at $t$; the denominator is the probability of being event-free at $t$ from either component. As time increases, susceptible individuals increasingly experience the event, so the cured component can make up a larger share of the remaining event-free group.

Worked Example: Interpreting a Ten-Year Survivor

For a patient in the reference group who is event-free at 10 years, use the values from the previous calculation:

\widehat P(\text{cured}\mid T>10,z=0) =\frac{0.4000}{0.4163} \approx 0.9609.

Under this fitted model, the estimated probability of cured-component membership among reference-group patients who remain event-free at 10 years is about 96.09%. This is not a statement that 96.09% of the original reference group is cured; its fitted cure fraction is 40.00%. Nor is it an observed individual diagnosis: membership is latent, and the probability depends on the model and its assumptions.

Checking Calibration and Model Adequacy

Calibration asks whether predicted event-free probabilities agree with censoring-adjusted observations. One practical check compares the fitted survival curve with a nonparametric survival estimate at a set of times, preferably with uncertainty intervals and attention to the number still at risk. The comparison should be made within relevant covariate groups or after suitable standardization; an overall comparison can conceal opposite errors in different groups.

Worked Example: Comparing Predictions With Observed Survival

Suppose the censoring-adjusted observed survival estimates for the reference group are 0.550 at 5 years, 0.430 at 10 years, and 0.410 at 12 years. The fitted model predicts, using $\hat\pi(0)=0.4000$:

\hat S(5\mid0)=0.4000+0.6000\exp[-1.25^{1.4}] \approx 0.553,

because $1.25^{1.4}\approx1.367$ and the susceptible survival is about 0.2549. At 10 years the prediction is 0.4163, as calculated above. At 12 years,

\hat S(12\mid0) =0.4000+0.6000\exp[-3^{1.4}] \approx0.4057.

The observed-minus-predicted differences are approximately $-0.003$, $0.014$, and $0.004$ at 5, 10, and 12 years, respectively. These small illustrative differences do not establish good calibration: sampling uncertainty, especially in the tail, and the numbers at risk must also be considered. A systematic gap or a trend in the differences would be a warning that the model misses important features.

Calibration is not the only diagnostic. Review the fitted latency shape, the range and distribution of covariates, convergence and uncertainty in the incidence estimates, and the observed survival pattern late in follow-up. A fitted cure fraction that changes sharply under modest changes in the latency specification, or is supported by very few late observations, should be reported cautiously. Residual and calibration diagnostics for cure models provide additional tools, but no diagnostic can prove that a latent cured group exists.

A mixture cure model is useful when a lasting event-free fraction is scientifically defensible and follow-up can distinguish it from a long susceptible tail. A standard survival model without a cure component may be more appropriate when eventual event occurrence is the intended assumption or follow-up does not support estimating a plateau. Other cure-model specifications can encode the long-term component differently. The earlier material on choosing between mixture cure model specifications is relevant when deciding whether the model form is defensible; the fitted predictions here should not be treated as a substitute for that choice.

Reporting the Analysis and Its Limits

A clear report identifies the event, time scale, censoring rule, covariates in each component, and the fitted latency distribution. It presents cure-probability estimates separately from marginal survival predictions, gives uncertainty intervals where available, and reports the time horizon and covariate profile for each prediction. State whether predictions are conditional on a particular profile or standardized over a target population. Include calibration evidence and explain how much late follow-up informs the cure estimate.

The principal limitation is that cured status is not observed directly in a finite follow-up study. An apparently stable tail can also arise from a very slowly declining susceptible survival curve, inadequate follow-up, or misspecified covariate effects. Predictions beyond observed follow-up therefore rely strongly on the parametric latency and cure assumptions. In observational data, covariate-specific differences are associations unless a suitable causal design and assumptions justify causal interpretation.

Key takeaway: A mixture cure analysis reports two linked components: the probability of cured-component membership and the event-time survival among susceptible individuals. Their combination gives covariate-specific marginal survival, while calibration, follow-up, and model-form checks determine how cautiously long-term predictions should be interpreted.

Check Your Understanding

Answer each question, then compare with the solutions.

For a profile with cure probability 0.30 and susceptible survival 0.20 at 8 years, calculate marginal survival at 8 years.
If the susceptible survival tends to zero, what does marginal survival tend to, and what assumption makes that interpretation possible?
For the profile in Question 1, calculate the model-implied probability of cured-component membership among those event-free at 8 years.
Name one reason a close match between predicted and observed survival at a few late time points may still be insufficient evidence of good calibration.

Solutions. 1. $S(8)=0.30+(1-0.30)(0.20)=0.30+0.14=0.44$, or 44%. 2. It tends to 0.30, the cure fraction, provided the susceptible survival tends to zero and the mixture model is an adequate representation. 3. The conditional probability is $0.30/0.44\approx0.6818$, or about 68.18%; it is conditional on being event-free at 8 years, not the population cure fraction. 4. For example, few individuals may remain at risk, making late observed estimates imprecise; alternatively, an overall or sparse-time comparison may conceal miscalibration at other times or within covariate groups.