Contemporary issues in static and dynamic prediction: some applications and evaluation in the clinical context

Prediction models that estimate the probabilities of developing a specific disease (diagnostic model) or a specific endpoint of disease (prognostic model) given a set of subject’s characteristics are closely connected to personalized medicine of which the key idea is to base medical decisions on individual patient characteristics rather than on population averages. Depending on decision point, prediction models can be divided into two categories: static prediction models (making one-off decision) and dynamic prediction models (making dynamically updated decisions). While multivariable logistic and Cox regression are commonly used to develop prediction models, they are not the master key to every situation. Various issues such as clustered data, competing risks and time-dependent variable may occur when a simple logistic or Cox model cannot estimate the risk correctly in static and dynamic prediction. Although adapted or more advanced approaches have been developed to address those issues in medical statistics field, they are not appropriately applied in clinical research. To fill this gap, this thesis illustrated how sophisticated statistical models can be appropriately applied to obtain better predictions using a series of clinical case studies.