Mixture Cure (Long-Term Survivor) Models

Description

A survival model in which a fraction of people will never experience the event, estimating that fraction and the timing for everyone else.

Standard survival models assume everyone eventually has the event, but many trial users never upgrade and many sign-ups never churn within any realistic horizon. In the mixture cure model the survival curve is a weighted sum: a proportion who never have the event plus the remainder following a parametric survival curve, so the overall curve levels off above zero. Each observed event contributes the non-cured share times its density; each case still event-free at the end of follow-up contributes the cured share plus the non-cured share times its survival probability. EM fits the model, with covariates entering a logistic model for the cured fraction and a proportional-hazards or parametric model for timing. The two parts answer different questions: who will ever act, and how soon those who will act do so.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Days

Steps

  1. Define the event, the start of the clock and censoring for each person.
  2. Plot Kaplan-Meier curves; a clear plateau with long follow-up suggests a never-event fraction.
  3. Choose a parametric timing distribution (Weibull, Gompertz) for those who will have the event and a logistic model for the fraction.
  4. Fit by EM (or ECM), with likelihood contributions for observed events and for censored cases as described.
  5. Interpret covariates separately for the fraction and for the timing.
  6. Get confidence limits by bootstrap, resampling separately within event and censored cases.

Sources

  • Finite Mixture Models (McLachlan & Peel, 2000)
  • Statistical Methods for Reliability Data (Meeker & Escobar, 1998)

Also known as

  • Cure rate model
  • Long-term survivor model
  • Split-population survival model
  • Never-event fraction model
  • Mixture cure model
  • Limited failure population model
  • Defective-subpopulation model

Categories

  • Analytics & Behaviour
  • Statistics & Modelling