Individual dynamic predictions using landmarking and joint modelling: validation of estimators and robustness assessment

TL;DR

This study compares and validates joint modelling and landmarking approaches for individual dynamic predictions of disease progression, emphasizing the importance of proper model specification and estimator validation in clinical follow-up.

Contribution

It provides a formal validation framework, compares prediction methods, and assesses their robustness and accuracy in a clinical context.

Findings

Joint and landmark models have different strengths and limitations.

Proper model specification is crucial for accurate predictions.

Validation techniques improve estimator reliability.

Abstract

After the diagnosis of a disease, one major objective is to predict cumulative probabilities of events such as clinical relapse or death from the individual information collected up to a prediction time, including usually biomarker repeated measurements. Several competing estimators have been proposed to calculate these individual dynamic predictions, mainly from two approaches: joint modelling and landmarking. These approaches differ by the information used, the model assumptions and the complexity of the computational procedures. It is essential to properly validate the estimators derived from joint models and landmark models, quantify their variability and compare them in order to provide key elements for the development and use of individual dynamic predictions in clinical follow-up of patients. Motivated by the prediction of two competing causes of progression of prostate cancer from the history of prostate-specific antigen, we conducted an in-depth simulation study to validate and compare the dynamic predictions derived from these two methods. Specifically, we formally defined the quantity to estimate and its estimators, proposed techniques to assess the uncertainty around predictions and validated them. We also compared the individual dynamic predictions derived from joint models and landmark models in terms of prediction error, discriminatory power, efficiency and robustness to model assumptions. We show that these prediction tools should be handled with care, in particular by properly specifying models and estimators.