Dynamic Prediction in Mixture Cure Models: A Model-Based Landmarking Approach
Mixture cure models are widely used in survival analysis when a portion of patients is considered cured and is no longer at risk for the event of interest. In clinical settings, dynamic survival prediction is particularly important to refine prognosis by incorporating updated patient information over time. Landmarking methods have emerged as a flexible approach for this purpose, as they allow to summarize longitudinal covariates up to a given landmark time and to use these summaries in subsequent prediction. For mixture cure models, the only landmarking strategy available in the literature relies on the last observation carried forward (LOCF) method to summarize longitudinal dynamics up to the landmark time. However, LOCF discards most of the longitudinal information, does not correct for measurement error, and may rely on outdated values if observation times are far apart. To overcome these limitations, we propose a sequential approach that integrates model-based landmarking within a mixture cure model. Initially, longitudinal covariates are modeled using (generalized) linear mixed models, from which individual-specific random effects are predicted. The predicted random effects are then incorporated as covariates into a Cox proportional hazards cure model. We investigated the performance of the proposed approach under different cure fractions, sample sizes, and longitudinal data structures through an extensive simulation study. The results show that the model-based strategy provides more refined predictions compared to LOCF, even when the model is misspecified in favour of the LOCF approach. Finally, we illustrate our method using a real-world dataset on renal transplant patients.