Landmarking 2.0: Bridging the gap between joint models and landmarking

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Longitudinal Data Analysis

Background:

• Dynamic prediction with time-dependent covariates (biomarkers) is crucial in medical research.
• Existing methods include joint modeling and landmarking, each with limitations in efficiency and robustness.
• Joint models can be inefficient if misspecified, while landmarking may be less efficient than correctly specified joint models.

Purpose of the Study:

• To develop methods improving the predictive accuracy of landmarking for time-dependent covariates.
• To retain the relative simplicity and robustness of the landmarking approach.
• To provide reliable dynamic predictions for individual patient outcomes.

Main Methods:

• Fitting a working longitudinal model for the biomarker, incorporating temporal correlation.
• Deriving a predictable time-dependent process for the biomarker's expected value post-landmark.
• Fitting a time-dependent Cox model using the derived predictable time-dependent covariate.

Main Results:

• The proposed method enhances the predictive accuracy of landmarking.
• Dynamic predictions are obtained by estimating biomarker trajectories and using a time-dependent Cox model.
• The approach was illustrated for predicting overall survival in liver cirrhosis patients using prothrombin index.

Conclusions:

• The developed method offers improved predictive accuracy for landmarking with time-dependent covariates.
• This approach balances predictive performance with the inherent simplicity and robustness of landmarking.
• It provides a valuable tool for dynamic prediction in clinical settings, exemplified by liver cirrhosis survival prediction.