Separating Covariate Effects on Incidence and Latency

Why Separate Incidence From Latency?

In a population with long-term survivors, a covariate can relate to two different features of outcome. It may be associated with whether an individual is susceptible to the event at all, or with how quickly the event occurs among susceptible individuals. These are distinct questions. A covariate may affect one component, both components, or neither.

A mixture cure model represents this distinction explicitly. Its incidence component models the probability of cure; its latency component models event timing conditional on susceptibility. A coefficient in one component must not be interpreted as though it came from the other. In particular, an association with cure odds is not a hazard ratio, and a latency coefficient does not describe the probability of cure.

The cure status is a latent, or unobserved, class for individuals who have not yet experienced the event. A person who remains event-free at the end of follow-up is not thereby known to be cured: that person could instead be susceptible with an event time beyond the observed period. The interpretation of “cure” should match the scientific setting; if cure means no event over a specified horizon rather than never experiencing the event, that horizon must be stated.

Two Regression Components

A common incidence specification is logistic regression:

For a one-unit increase in a covariate \(x_j\), holding the other incidence covariates fixed, \(\exp(\gamma_j)\) is the multiplicative change in the odds of cure. It is an odds ratio for cure, not a risk ratio and not an event hazard ratio. The corresponding change in cure probability depends on the starting probability. The earlier material on logistic regression for cure probability covers this component on its own; the key here is to keep its meaning distinct from the latency component.

For latency, consider a Weibull accelerated-failure-time model among susceptible individuals:

Here \(\lambda(z)\) is a time scale and \(k\) is a shape parameter. With a common \(k\), increasing \(z_j\) by one unit multiplies the susceptible time scale by \(\exp(\beta_j)\), holding the other latency covariates fixed. Thus, \(\exp(\beta_j)>1\) indicates longer event times among susceptible individuals under this model; it does not mean a larger cure probability. The susceptible median is \(\lambda(z)(\log 2)^{1/k}\), so the same time ratio applies to that median.

The incidence covariate vector \(x\) and latency vector \(z\) need not be identical. A covariate can appear in both, in only one, or in neither. When a covariate appears in both, it has two distinct coefficients. That is a modeling choice about two conditional relationships, not an assertion that the covariate itself has two values or that the component estimates are statistically independent.

Why the Marginal Effect Is Not Either Component Alone

The marginal survival formula follows by conditioning on the latent cure status. A cured individual survives beyond every finite time under the model, while a susceptible individual survives past \(t\) with probability \(S_u(t\mid z)\). The law of total probability therefore gives

This identity also explains why a covariate's marginal survival association can depend on both components. For a continuous covariate \(w\) appearing in both components, the derivative is

The first term represents the incidence contribution, and the second represents the latency contribution. Their relative sizes vary with time through \(S_u(t)\). In particular, as susceptible survival approaches zero at long times, marginal survival approaches the cure probability \(\pi(x)\); at earlier times, latency can contribute substantially to the marginal contrast.

A related result makes clear why a single hazard interpretation can be misleading. If \(f_u(t\mid z)\) is the susceptible event-time density, the marginal event density at finite \(t\) is \(\{1-\pi(x)\}f_u(t\mid z)\). Dividing by marginal survival gives the marginal hazard:

Both components enter this expression. Consequently, a coefficient for cure odds cannot be read as a hazard ratio, and even a latency model with a simple conditional time-ratio interpretation does not automatically imply the same simple interpretation for the marginal hazard.

Assumptions, Diagnostics, and Alternatives

Interpretation requires assumptions about both components and the observation process. The incidence model must adequately describe cure probability as a function of its covariates; the latency model must adequately describe event timing among susceptible individuals. The Weibull example assumes the specified Weibull form and a common shape \(k\), with covariate effects acting through the time scale. In right-censored data, censoring must be non-informative conditional on modeled covariates and relevant model structure. In addition, the data need enough follow-up and information to distinguish a long susceptible tail from a cured fraction; without that, the two components may be weakly identified.

Useful checks include examining observed and fitted survival patterns across covariate groups, especially at late follow-up where a cure component implies a nonzero long-term survival limit. Look for systematic departures in early or middle follow-up that may indicate a poor latency distribution, and departures in the apparent tail that may indicate a poor incidence model or insufficient follow-up. Sparse data, near-complete separation in the cure regression, and few late events can make component estimates unstable. A good fit to overall survival alone is not proof that both components are individually well determined.

The main alternatives answer different questions or make different assumptions. A standard survival model without a cure fraction treats event-free individuals as potentially susceptible, so it does not explicitly represent a nonzero cured proportion. A logistic regression of “event observed” by the study end confounds cure with follow-up duration and censoring; it is not a substitute for modeling latent cure status. A mixture cure model is useful when a lasting event-free fraction is scientifically plausible, but it depends on separating latent susceptibility from delayed events and can be sensitive to follow-up and distributional assumptions.

The model also has limits. “Cured” is a structural assumption about no future event, not a directly observed label in finite follow-up. The selected latency family may be too restrictive, and a common shape may not capture differing timing patterns across covariate groups. Component effects are conditional model associations: their interpretation depends on the covariates included and the functional forms specified. Estimation and uncertainty for the two components require appropriate joint methods; separate coefficient interpretations do not make their estimates independent.

Common Interpretation Pitfalls

• Calling a cure odds ratio a hazard ratio. The incidence coefficient compares odds of cure. The latency component and mixture formula determine event-time behavior.
• Calling a latency time ratio a cure effect. A larger susceptible time scale means later event times among susceptible individuals, not necessarily more cured individuals.
• Interpreting a susceptible median as the population median. The susceptible median is conditional on \(C=0\). Marginal survival includes the cured fraction and may not reach 0.5, in which case a finite marginal median does not exist.
• Equating no observed event with cure. A censored, event-free individual may still be susceptible and experience the event later.
• Reading a marginal contrast from one coefficient. When covariates enter both components, calculate or otherwise examine the marginal survival using both fitted components.