Abstract:
|
Despite the richness of data collection throughout the disease, most risk prediction studies only included measurement at single timepoint when developing models, let alone make full use of various longitudinal measurements to improve prediction accuracy. Many tools such as joint modeling and landmarking have been developed as remedies to the prior problems, such as Joint Latent Class Model (JLCM) and Shared Random Effects Model (SREM). These approaches present a unified yet flexible framework to incorporate multiple specifications, multivariate events, and multivariate biomarkers. They presumably have great potentials to improve prediction model performance, at the cost of model complexity and computation burdens. In this paper, we focus on the model interpretations and operating characteristics of several model-based dynamic prediction methods, including but not limited to time-dependent model diagnostics, model selection, discrimination, and calibrations. Our work provides useful guidance for choosing the appropriate tool when various data complexity occurs. Both Monte Carlo simulations and data analysis of lung cancer clinical trials are presented for illustration purposes.
|