Identifiability and Follow-Up in Mixture Cure Models

Why Follow-Up Determines What Can Be Learned

A mixture cure model separates two explanations for remaining event-free: a person may be cured, or may be susceptible but have a long time to the event. These explanations can look alike when observation ends before many susceptible individuals experience the event. The key issue is not merely whether a fitted model returns a cure fraction; it is whether the observed follow-up contains enough information to distinguish that fraction from the latency distribution among susceptible individuals.

The earlier material on mixture cure models defines the cure component and the latency component, and gives the marginal survival function. Here the focus is the information available from finite follow-up. In particular, a curve that has not declined much by the end of observation does not by itself establish that its remaining survivors are cured.

Definition: Let $\pi$ be the cure probability, and let $S_u(t)$ be the survival function of event time among susceptible individuals. The population survival function is $$ S_{\mathrm{pop}}(t)=\pi+(1-\pi)S_u(t). $$ Assume $S_u(0)=1$, $S_u(t)$ is nonincreasing, and susceptible individuals eventually experience the event, so $S_u(t)\to 0$ as $t\to\infty$. Under this definition, the population survival curve has limiting value $\pi$. The event time for a cured individual is infinite, while a susceptible individual has a proper event-time distribution.

The definition makes a strong distinction between cure and delayed event: cure means no event ever under the model, not simply no event during the study. In practice, a study observes each person only up to their event time or censoring time. If follow-up is bounded, the data generally say nothing directly about the shape of the survival curve beyond the time at which observation effectively ends.

Finite Follow-Up and Non-Unique Decompositions

Suppose, first, that the population survival curve is known perfectly only up to a finite time $\tau$. For any proposed cure probability $0\leq\pi<1$, rearranging the mixture equation gives the latency survival curve over the observed time range. If $\pi=1$, the mixture equation gives $S_{\mathrm{pop}}(t)=1$ throughout the observed range, so this case is compatible only when the observed curve is identically 1 there:

S_u(t)=\frac{S_{\mathrm{pop}}(t)-\pi}{1-\pi}, \qquad 0\leq t\leq \tau.

For this to be a valid nonnegative survival curve over that range, $\pi$ cannot exceed the smallest observed population survival value, $S_{\mathrm{pop}}(\tau)$. Many different values of $\pi$ can therefore yield valid latency curves on the observed interval. Each can be extended beyond $\tau$ in a way that eventually reaches zero. Those extensions imply different long-term cure fractions, even though they agree with the same observed curve through $\tau$.

Key result: With an unrestricted latency distribution and follow-up only through $\tau$, the data do not nonparametrically identify the cure probability from the survival curve on $[0,\tau]$. A cure fraction may become distinguishable if sufficiently long follow-up reveals a stable tail while susceptible individuals are still observable, or if additional model assumptions restrict the possible latency curves. The latter can provide model-based identification, but the distinction may then depend heavily on those assumptions.

This is a statement about the information in the observation window, not a claim that all cure models are impossible to estimate. A specified parametric or otherwise constrained latency model may mathematically determine a unique parameter value from the observed portion of the curve. But if the data do not reach the part of the curve where competing assumptions make noticeably different predictions, the estimate can be weakly identified in practice: small changes in the data or model assumptions may produce large changes in the estimated cure fraction.

Worked Example: Two Decompositions of the Same Observed Curve

Suppose a hypothetical population survival curve is observed through year 4, with values $1.00$, $0.80$, $0.55$, and $0.32$ at years 0, 1, 2, and 4. Compare decompositions using cure probabilities $\pi=0.10$ and $\pi=0.25$.

For $\pi=0.10$, the latency survival at year 4 is $$ S_u(4)=\frac{0.32-0.10}{1-0.10} =\frac{0.22}{0.90} =0.2444. $$ The corresponding values at years 1 and 2 are $(0.80-0.10)/0.90=0.7778$ and $(0.55-0.10)/0.90=0.5000$.

For $\pi=0.25$, the year-4 latency survival is $$ S_u(4)=\frac{0.32-0.25}{1-0.25} =\frac{0.07}{0.75} =0.09333. $$ At years 1 and 2, the latency survival values are $(0.80-0.25)/0.75=0.7333$ and $(0.55-0.25)/0.75=0.4000$.

Both latency curves start at 1, are nonincreasing at the stated times, and are nonnegative. Thus both are compatible with exactly the same observed population survival values through year 4. Yet one decomposition attributes 10% of the population to cure and the other attributes 25%. The observed curve alone does not choose between them. Follow-up beyond year 4, or a restriction on the latency distribution, is needed to distinguish their longer-term implications.

Why a Tail Assumption Can Decide the Estimate

A parametric latency model restricts the possible shapes of $S_u(t)$. Such a restriction can make the cure fraction estimable even when the observed curve has not visibly reached a plateau. That estimate is then an extrapolation informed by the chosen model as well as by the observations. The following calculation illustrates the issue using an exponential latency assumption; it is not a recommendation to use that assumption.

Worked Example: Different Cure Fractions, Same Survival at One Time

Suppose a hypothetical study estimates population survival at year 3 to be $0.50$. For illustration, assume susceptible latency is exponential with rate $\lambda$ per year, giving $S_u(t)=e^{-\lambda t}$. The mixture curve is then $S_{\mathrm{pop}}(t)=\pi+(1-\pi)e^{-\lambda t}$. Consider three possible cure probabilities.

If $\pi=0.10$, the year-3 equation is $0.50=0.10+0.90e^{-3\lambda}$. Therefore $e^{-3\lambda}=0.40/0.90=0.4444$, and $$ \lambda=-\frac{\log(0.4444)}{3}=0.2703 \text{ per year}. $$ If $\pi=0.25$, then $e^{-3\lambda}=(0.50-0.25)/0.75=1/3$, giving $\lambda=-\log(1/3)/3=0.3662$ per year. If $\pi=0.40$, then $e^{-3\lambda}=0.10/0.60=1/6$, giving $\lambda=-\log(1/6)/3=0.5973$ per year.

All three parameter pairs reproduce survival $0.50$ at year 3. Their predicted survival at year 6 differs: because $e^{-6\lambda}=(e^{-3\lambda})^2$, the predictions are $0.10+0.90(0.4444^2)=0.2778$, $0.25+0.75(1/3)^2=0.3333$, and $0.40+0.60(1/6)^2=0.4167$, respectively. Data around year 6 could help distinguish these predictions, but a single survival value at year 3 cannot. With finite and sparse data, a fitted curve may select one of these combinations because of the imposed model form, not because the observed point itself uniquely separates cure from latency.

Reading Follow-Up Summaries and Survival Curves

The maximum observed follow-up is not enough to assess identification. A small number of people followed for a long time may contribute little information about the tail. Review the estimated survival curve together with the number still at risk, event counts, censoring patterns, and the amount of follow-up available near the apparent plateau. For estimation of the survival curve from right-censored data, independent censoring (possibly conditional on modeled covariates) and adequate follow-up support are also important: the probability of remaining observable must not vanish in the time range used to estimate the curve.

A curve that is nearly flat while many individuals remain under observation and at risk is more informative about a possible tail than a curve that becomes flat after the risk set has dwindled. Even then, a flat segment is evidence about survival over the observed interval, not proof that the limiting survival is positive. A genuinely late event pattern can mimic a cure plateau if follow-up stops too soon.

Worked Example: A Late Flat-Looking Tail With Few at Risk

Consider a hypothetical cohort of 200 people. For a grouped-time illustration, assume events occur at interval endpoints and censoring occurs between endpoints. The risk set at the start of each interval, the endpoint events, and the intervening censorings are:

Interval (years)	At risk at start	Events at endpoint	Censored before next interval	At risk next interval
0–1	200	40	15	145
1–2	145	35	27	83
2–3	83	20	32	31
3–4	31	8	15	8

The product-limit estimate at year 4 is the product of the fractions surviving each endpoint: $$ \widehat S(4) =\left(\frac{200-40}{200}\right) \left(\frac{145-35}{145}\right) \left(\frac{83-20}{83}\right) \left(\frac{31-8}{31}\right) =0.8(0.7586)(0.7590)(0.7419) \approx 0.3418. $$

There are only 8 people still at risk at year 4, after 103 events and 89 censorings. The estimated survival of about 0.342 at that time is not an estimate that 34.2% are cured. The late estimate is based on a small risk set, and the data shown do not establish what happens after year 4. A longer study might show a continuing decline, or a stable tail; this follow-up summary alone cannot tell which. The calculation also relies on the stated grouped-time convention and on censoring being noninformative for the survival estimation.

A Practical Identification Check

The following checks help separate information supplied by the observations from information supplied by assumptions. They diagnose weaknesses; none can prove cure from a finite study.

Locate the effective follow-up limit.
Determine how late event and censoring times extend, and whether people remain observable in the time range where a tail is being interpreted.

Inspect the tail and its support.
Compare the apparent flattening with numbers at risk and events over time. A flat estimate with very few people at risk provides weak evidence about long-term survival.

Ask what identifies the decomposition.
Check whether the cure fraction is supported by a well-observed late tail or mainly by restrictions on the latency distribution. If modestly different plausible latency assumptions change the cure estimate substantially, identification is fragile.

Report the boundary of the evidence.
State the follow-up range and late risk-set size, and distinguish estimates within that range from extrapolations about the eventual cure fraction.

Alternatives, Limitations, and Common Mistakes

The principal alternative to a cure model is a model that assumes everyone remains susceptible but allows event times to be long. Within a cure-model analysis, one can also impose a parametric latency distribution or use a less restrictive latency specification. A more restrictive form can stabilize estimation, but may achieve that stability by stronger extrapolation assumptions. Longer follow-up and larger late risk sets provide more direct information about the tail; neither is replaced by a sophisticated fitting algorithm.

There are limits even with extended follow-up. A susceptible event distribution may have a very long tail, so a finite study may still not reveal a clear plateau. Loss to follow-up can reduce late information, and informative censoring can distort the estimated curve if the censoring mechanism is not adequately addressed. Likewise, a population-level plateau can conceal different patterns across covariate groups. Cure estimates should therefore be interpreted in the population and under the assumptions for which they were defined.

Mistake: treating everyone event-free at study end as cured. The correction is that these individuals are right-censored; they may be cured or susceptible with a later event time.
Mistake: calling a flat curve a cure plateau without checking the risk set. A flat tail based on very few people is weak evidence about the limiting survival probability.
Mistake: assuming a model-based estimate is determined by the data alone. If follow-up is short, tail shape restrictions can drive the estimated cure fraction. Report this dependence rather than presenting extrapolation as direct observation.

Key takeaway: Over a finite follow-up window, the same observed survival curve can often be explained by different combinations of cure probability and susceptible latency. Assess the tail, censoring support, and numbers at risk, and make clear whether the cure estimate is data-supported or assumption-dependent.

Check Your Understanding

Answer each question, then compare with the brief solutions.

If the observed population survival at the end of follow-up is $0.28$, what upper bound on $\pi$ is compatible with a nonnegative latency survival curve over the observed interval?
Why does a person with no event by the administrative study end not count as known cured?
A survival curve is flat late in follow-up, but only 6 people remain at risk. What does this imply for evidence about a cure plateau?
How can a parametric latency assumption make a cure fraction estimable, and what is the corresponding caution?
In the grouped example, how many people are at risk at year 4, and why is the estimated survival at that time not itself a cure estimate?

Solutions. (1) The compatibility bound is $\pi\leq S_{\mathrm{pop}}(\tau)=0.28$; this is a bound, not an estimate. (2) The person may have an event after follow-up ends, so the observation is right-censored rather than evidence of no event ever. (3) The flat tail is weak evidence about the limiting survival because it is based on a very small risk set. (4) The assumption restricts latency-curve shapes and can select a unique decomposition; the resulting cure estimate may depend strongly on that restriction when follow-up is short. (5) Eight people are at risk at year 4. The value $0.3418$ estimates survival through year 4 under the stated censoring assumptions; it does not reveal how many will never experience the event.