Why Cure-Model Diagnostics Need Special Care
A mixture cure model describes two related features: whether an individual is susceptible to the event and, among susceptible individuals, when the event occurs. A model can predict the overall survival curve reasonably well while misrepresenting one of these components. Conversely, an apparent mismatch in the late survival tail may reflect limited follow-up rather than a demonstrable error in the cure probabilities. Diagnostics should therefore examine overall survival, latency among susceptible individuals, and cure-probability calibration as distinct questions.
Let \(T\) be the event time, \(C\) the censoring time, \(Y=\min(T,C)\), and \(\delta=I(T\leq C)\). For covariates \(x\), write \(\pi(x)\) for the model-predicted probability of being cured, and \(S_u(t\mid x)\) for the event-free survival function among susceptible individuals. The model's marginal survival is
A person with an observed event is known to be susceptible under this model. A person censored while event-free is not known to be cured: that person may be cured or may be susceptible and yet to experience the event. This distinction is central to interpreting the diagnostics below.
Observed and Predicted Survival
Within a group of individuals, compare a nonparametric estimate of survival—usually the Kaplan–Meier estimator—with model-predicted marginal survival over the same follow-up range. This comparison relies on censoring being independent of the event time within the group (or independent conditional on the covariates used in the analysis), along with appropriate handling of delayed entry if present. The model curve should be averaged over the group's covariate distribution when the goal is a group-level comparison:
This average is generally not the same as evaluating the model at the group's mean covariate values. The plotted curves should use the same time origin, event definition, and target population. Where uncertainty matters, show confidence intervals for the observed curve and uncertainty bands for predictions, ideally accounting for estimation of all model parameters, for example by resampling individuals and refitting the model.
Worked Example: Comparing Group Survival Curves
Hypothetical data. In a group of 100 patients, suppose the fitted model gives an average predicted cure probability of \(0.30\). At two years, the average fitted susceptible survival in the group is \(0.60\); at four years it is \(0.30\). The observed Kaplan–Meier estimates are \(0.68\) at two years and \(0.50\) at four years.
At two years, the average predicted marginal survival is
The observed estimate, \(0.68\), is \(0.04\) lower. At four years, the prediction is
The observed estimate, \(0.50\), is \(0.01\) lower. These point comparisons do not by themselves establish misfit: uncertainty in both curves matters, and the Kaplan–Meier comparison requires suitable censoring assumptions. The differences are small in this example, but systematic separation over time or across covariate groups would warrant further investigation.
A close match in overall survival is reassuring but not decisive. Different combinations of cure probability and susceptible survival can produce similar marginal curves over a limited time range. Inspect the timing and amount of censoring as well as the curve shapes: few individuals remaining under observation in the tail make apparent plateau agreement especially weak evidence.
Event-Time Residuals for the Susceptible Component
Let \(H_u(t\mid x)=-\log S_u(t\mid x)\) be the fitted cumulative hazard among susceptible individuals. If the susceptible latency model is correct and the event-time distribution is continuous, then for a truly susceptible individual with a fully observed event time \(T\), the transformed time
has a unit exponential distribution. To see this, for \(r\geq0\), let \(H_u^{-1}(r\mid x)\) denote the corresponding latency time. Then
Thus a plot of the empirical event-residual distribution against the unit exponential distribution, or a comparison of its observed cumulative hazard with the reference line, can reveal latency-model departures. For example, curvature can signal an incorrect time pattern, while systematic differences across covariate groups may indicate that the latency model does not capture covariate-related variation.
Right censoring changes how this diagnostic must be carried out. For an observed failure, \(\delta=1\), the individual is known to be susceptible and \(H_u(Y\mid x)\) is an observed latency residual. For a censored individual, \(H_u(Y\mid x)\) is only a censoring point; susceptibility is unknown. Treating every censored person as susceptible, or discarding all censored individuals and comparing only observed failures with an uncensored exponential distribution, can give a misleading result. Censoring selectively hides later event times.
A practical approach is to compare the observed residual patterns with patterns from simulated data under the fitted model, reproducing the observed covariates and censoring process and refitting the model in bootstrap replicates when parameter uncertainty is important. This keeps the censoring and latent susceptibility structure in the reference comparison. Alternatively, use a likelihood-based or posterior-weighted residual procedure that explicitly accounts for uncertain susceptibility among censored individuals. In either case, state which procedure was used; a raw plot of failure residuals alone is not a complete censoring-adjusted assessment.
Worked Example: Transforming Susceptible Event Times
Hypothetical data. Suppose the fitted susceptible latency model has an exponential distribution with rate \(0.40\) events per year, so \(H_u(t)=0.40t\). Three individuals have observed events at \(0.5\), \(1.0\), and \(2.0\) years. A fourth individual is censored at \(1.5\) years.
For the observed failures, the residuals are \(0.40(0.5)=0.20\), \(0.40(1.0)=0.40\), and \(0.40(2.0)=0.80\). Under the fitted latency model, the corresponding unit-exponential cumulative probabilities are \(1-e^{-0.20}=0.181\), \(1-e^{-0.40}=0.330\), and \(1-e^{-0.80}=0.551\). For example, the model places about \(55.1\%\) of susceptible event times at or before the third individual's residual value.
The censored individual's transformed time is \(0.40(1.5)=0.60\), but this is not an observed event residual. The individual may be cured or susceptible with an event time beyond 1.5 years. The four transformed values should therefore not be treated as four completed susceptible event times. A censoring-aware assessment would compare the observed pattern with model-based simulated or appropriately weighted patterns, preserving this distinction.
Calibrating Predicted Cure Probabilities
Calibration of \(\widehat\pi(x)\) asks whether groups assigned a particular cure probability have a corresponding cured proportion. However, with ordinary right-censored follow-up, cure status is not observed for people who remain event-free at their last contact. An event-free proportion at a finite time is not a cure proportion: it also includes susceptible people whose events have not yet occurred.
A direct calibration assessment is most credible when cure status can be established independently after sufficiently long follow-up. Group individuals by predicted cure probability, compare the mean prediction in each group with the observed cured proportion, and show uncertainty for both quantities. If cure status is instead inferred from a survival plateau or an estimated long-term fraction, the assessment depends on adequate follow-up and on the assumption that susceptible event times do not have a substantial unobserved tail. A plateau estimated from the same fitted model is not independent validation of that model.
If no defensible cure label or long-term estimate is available, assess calibration of a finite-horizon prediction instead. For a specified time \(\tau\), compare observed survival with the predicted marginal survival \(S_m(\tau\mid x)\), using a censoring-adjusted method where needed. This checks survival at \(\tau\), not the probability of being cured.
Worked Example: Grouped Cure-Probability Calibration
Hypothetical validation data. Suppose 200 individuals have independently adjudicated long-term cure status, with 100 in each of two prediction groups. The first group has mean predicted cure probability \(0.20\), and 18 individuals are observed cured. The second has mean prediction \(0.40\), and 48 are observed cured.
The observed proportions are \(18/100=0.18\) and \(48/100=0.48\). The observed-minus-predicted differences are therefore \(0.18-0.20=-0.02\) and \(0.48-0.40=0.08\). The first group is close to its mean prediction; the second has an observed proportion eight percentage points higher.
For scale, approximate binomial standard errors for the observed proportions are \(\sqrt{0.18(0.82)/100}=0.0384\) and \(\sqrt{0.48(0.52)/100}=0.0500\). These quantify sampling variability in the observed proportions, not uncertainty in the model predictions or in the adjudication. The differences alone are not enough to declare miscalibration; uncertainty in predictions and the number and formation of groups should also be considered. If the “cure” labels had instead been assigned merely because no event was seen by a finite follow-up time, these comparisons would not assess cure-probability calibration.
Reading Diagnostic Patterns and Their Limits
No single plot proves that a cure model is correct. Use the diagnostics together and investigate whether discrepancies are systematic, large relative to uncertainty, and relevant over the time range for the intended predictions.
| Diagnostic pattern | Possible concern | Interpret cautiously |
|---|---|---|
| Observed and predicted survival separate over time | Misfit in the cure or latency component, or an unmodeled covariate pattern | Check censoring, group composition, and uncertainty before attributing the gap to model structure. |
| Latency residuals depart systematically from the reference | Incorrect susceptible event-time distribution or covariate effects | Use a procedure that handles censoring and uncertain susceptibility. |
| Observed cure proportions differ from predictions across groups | Potential poor cure-probability calibration | Confirm that long-term cure status is credible and assess uncertainty. |
| Late survival estimates are unstable | Too few individuals remain under observation to assess the tail reliably | Do not treat an apparent plateau as proof of a cure fraction. |
Graphical observed-versus-predicted comparisons are intuitive but may have limited power to reveal modest departures. Residual checks focus more directly on the latency distribution, but are sensitive to how censoring and latent susceptibility are handled. Calibration plots address the agreement of predictions with outcomes but require an appropriate observed target. These approaches complement one another; none replaces checking the data, censoring pattern, and follow-up support for the model's long-term claims.
Common Mistakes
- Calling every event-free person cured. A person censored without an event may still be susceptible. Use cure labels only when supported by long-term information or an explicitly justified estimation procedure.
- Comparing a cure probability with finite-time survival. \(S_m(\tau\mid x)\) includes both cured individuals and susceptible individuals event-free at \(\tau\). It is not generally equal to \(\pi(x)\).
- Using only observed failure residuals as if there were no censoring. Observed failures are selected by the censoring process. Compare residual patterns with a method that accounts for censoring rather than applying an uncensored reference naively.
- Reading a late plateau as conclusive evidence of cure. Sparse risk sets and limited follow-up can produce an unstable apparent plateau. Report how much follow-up and how many individuals support the tail.
- Treating a close fitted curve as proof of calibration. A model can match marginal survival while its cure and latency components are poorly determined. Examine component-specific evidence and uncertainty.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A model predicts cure probability \(0.25\) and susceptible survival \(0.40\) at time \(\tau\). What is its predicted marginal survival at \(\tau\)?
- Why is a person censored while event-free not automatically classified as cured?
- Under a correct continuous susceptible latency model, what is the distribution of \(H_u(T\mid x)\) for a susceptible person with a fully observed event time?
- A group has predicted cure probability \(0.35\), but \(45\%\) are event-free at the end of two years. Does this establish underprediction of cure? Explain.
Solutions. (1) \(0.25+(1-0.25)(0.40)=0.25+0.30=0.55\), so predicted marginal survival is \(0.55\). (2) The person may be cured, or may be susceptible with an event time beyond the censoring time; the event-free observation does not distinguish these possibilities. (3) It has a unit exponential distribution, because \(P\{H_u(T\mid x)>r\}=e^{-r}\) for \(r\geq0\). (4) No. The \(45\%\) is finite-time survival, which includes cured individuals and susceptible individuals who have not yet had an event; it is not an observed cure proportion.