Abstract
In clinical practice, individuals are followed up to predict the outcome event of interest, and their longitudinal measurements are collected on a regular or irregular basis. We aimed to examine the classical approach, joint model (JM), and alternative parameterization structures using data on the effect of time-varying longitudinal measurements on survival. The motivating cohort dataset included 158 consecutive kidney transplant recipients who had baseline and follow-up data. Although the longitudinal log-transformed estimated glomerular filtration rate (log[eGFR]) measurements and graft failure have an association clinically, the 2 processes are analyzed separately in the classical approach. In addition to the extended Cox model, the current value JM, the weighted cumulative effect JM, and dynamic predictions were performed in the study, by taking advantage of R codes. Of the 158 patients, 34.8% were males. The mean age was 29.8 ± 10.9 years, and the median age was 26 years at the time of transplantation. The hazard ratio for graft failure was 8.80 for a 1-unit decrease in log(eGFR) in the extended Cox model, 10.58 in the current value JM, and 3.65 in the weighted cumulative effect JM. The presence of coronary heart disease was also found to be associated with log(eGFR): 0.199 (P = .03) for the current value JM and 0.197 (P = .03) for the weighted cumulative effect JM. The current value JM was identified as a better model than the extended Cox model and the weighted cumulative effect JM based on parameter and standard error comparison and goodness of fit criteria. JMs should be preferred, as they facilitate better clinical decisions by accounting for the varying slopes and longitudinal variation of estimated glomerular filtration rate among patients. Suitable types of models should be practiced depending on baseline biomarker levels, their trends over time, the distribution of the biomarkers, and the number of longitudinal biomarkers.