Marginal Survival and Hazard in Mixture Cure Models

From Two Subpopulations to Marginal Survival

In a mixture cure model, individuals belong to one of two latent groups for the event under study: a cured group that will never experience that event, or a susceptible group whose event time follows a latency distribution. The population survival curve therefore combines a group with no event-time failure and a group whose probability of remaining event-free changes with time. This combination also changes the population hazard: it is not generally the same as the hazard among susceptible individuals.

Let \(q\) be the proportion susceptible, so \(1-q\) is the cure fraction, with \(0\le q\le1\). Let \(S_u(t)\) denote the event-free survival function for susceptible individuals, and let \(f_u(t)=-dS_u(t)/dt\) be their event-time density. Assume \(S_u(0)=1\), \(S_u(t)\to0\) as \(t\to\infty\), and that \(S_u\) is differentiable with density \(f_u\). A cured individual remains free of the event at all finite times. These are model assumptions about the event being studied, not claims that a person cannot experience other health events or die from another cause.

The mixture is formed by conditioning on latent group membership. A cured individual survives beyond any finite \(t\) with probability one; a susceptible individual does so with probability \(S_u(t)\). Hence the marginal survival is

This expression has the expected endpoints: \(S_{\mathrm{pop}}(0)=1\), and, because \(S_u(t)\to0\), \(S_{\mathrm{pop}}(t)\to1-q\). The population survival curve has a positive limiting plateau whenever the model assigns a positive cure fraction.

Deriving the Marginal Hazard

Differentiating the marginal survival function gives the population event-time density. Cured individuals contribute no event-time density, so only the susceptible group contributes:

Substitution into the hazard definition yields

The multiplier has a direct conditional-probability interpretation. Among those still event-free at time \(t\), the probability of being susceptible is

Thus \(h_{\mathrm{pop}}(t)=w(t)h_u(t)\): at time \(t\), only the susceptible share of those who have survived can experience the event, and its hazard is the susceptible-group hazard. For \(0<q<1\), \(w(t)\) is at most one and decreases as susceptible individuals experience events and the remaining population becomes increasingly enriched with cured individuals.

Worked Examples

Why the Population Hazard Can Approach Zero

When \(0<q<1\), susceptible survival tends to zero while population survival tends to \(1-q>0\). Hence \(w(t)\to0\): increasingly few of the event-free survivors are susceptible. If the susceptible hazard remains bounded as time increases, the identity \(h_{\mathrm{pop}}=w h_u\) implies that the population hazard tends to zero.

A useful sufficient condition is that the susceptible density tends to zero. In that case, directly from the density form,

This condition holds for many familiar smooth latency distributions, including the exponential example above. It is important not to claim that a positive cure fraction alone guarantees a zero limiting hazard for every mathematically possible latency distribution: if the susceptible hazard grows without bound, the product \(w(t)h_u(t)\) requires separate examination. The density expression makes the relevant tail behavior explicit.

A declining population hazard can therefore reflect changing composition among survivors rather than a declining hazard within the susceptible group. This is a selection effect built into the mixture: individuals susceptible to the event tend to leave the event-free risk set earlier, while cured individuals remain in it. The result is sometimes described as a long-term survivor plateau, but the plateau is a feature implied by the model, not by itself evidence that the model is true.

Checks, Comparisons, and Limitations

The formula provides several useful checks. At \(t=0\), if \(S_u(0)=1\), then \(w(0)=q\) and \(h_{\mathrm{pop}}(0)=q h_u(0)\). If \(q=1\), there is no cured group and \(S_{\mathrm{pop}}=S_u\), so the population and susceptible hazards are identical. If \(q=0\), everyone is cured, \(S_{\mathrm{pop}}(t)=1\), and the population hazard is zero. If \(0<q<1\) and \(h_u(t)=0\), both hazards are zero at that time.

The susceptible hazard answers a conditional question about members of the latent susceptible group. The population hazard describes the instantaneous event rate among everyone who has remained event-free, without observing their latent group. A standard single-population survival model has no explicit cured component; it should not be interpreted as having the same latent-group decomposition unless that assumption is added. The earlier material on the two components of a mixture cure model defines that decomposition; the equations here show its consequences for marginal survival and hazard.

In practice, an apparent late survival plateau can be difficult to distinguish from a latency distribution with a long tail, especially when few individuals remain under observation. The survival and hazard identities describe what follows if the mixture assumptions hold; they do not establish those assumptions from a plotted curve. Similarly, a fitted marginal hazard that declines does not by itself show that the susceptible-group hazard declines. Conversely, a rising susceptible hazard can coexist with a declining population hazard if the susceptible fraction among survivors falls sufficiently quickly.

A useful diagnostic is internal consistency: the proposed \(q\) and \(S_u(t)\) should give \(S_{\mathrm{pop}}(0)=1\), \(S_{\mathrm{pop}}(t)\ge0\), and a nonincreasing survival curve. The implied \(w(t)\) must lie between zero and one, and the computed hazard must be nonnegative. Violations usually indicate an algebraic, coding, or parameterization error. These checks cannot verify that the cure interpretation is scientifically appropriate; substantive knowledge and adequate follow-up are also needed.

Common Mistake

A related mistake is to say that a positive cure fraction necessarily makes the population hazard decrease at every time or forces it to zero under any latency distribution. It need not decrease monotonically, as the Weibull example shows. A zero limiting hazard follows under conditions such as bounded susceptible hazard or \(f_u(t)\to0\); the displayed population-hazard formula should be used to check the tail behavior.