A0444
Title: Extending landmarking to mixture cure models with longitudinal covariates
Authors: Marta Cipriani - Sapienza University of Rome (Italy) [presenting]
Marco Alfo - University La Sapienza, Rome (Italy)
Mirko Signorelli - Leiden University, Mathematical Institute (Netherlands)
Abstract: Dynamic prediction models represent an essential class of models for personalized medicine, providing real-time updates on prognosis based on evolving patient information. Among these, the landmarking approach has gained popularity due to its flexibility and conceptual simplicity. However, its integration into cure models remains underexplored. In the context of mixture cure models (MCM), current applications of landmarking rely exclusively on traditional summarization techniques for time-varying covariates, notably based on the last observation carried forward approach. A novel dynamic prediction framework is proposed that extends model-based landmarking to MCMs. The framework separates prediction into two components: (1) the incidence component, which estimates the probability of being uncured using logistic regression and baseline covariates, and (2) the latency component, which estimates post-landmark survival among uncured individuals through a Cox proportional hazards model that incorporates summaries of longitudinal data trajectories. Specifically, the longitudinal trajectories of patient covariates are modeled up to the landmark time using linear mixed-effects models or multivariate generalized linear mixed-effects models. These models allow for the estimation of individual-specific random effects, which provide a compact and informative summary of the patient's covariate trajectory and are then used as (fixed) predictors in the cure model.
