Statistical analysis of repeated outcomes of different types

This thesis focused on analyzing data with multiple outcome variables. The motivating data sets comprised longitudinal markers of patients’ disease state (e.g. B cells and CD4+ T cell) as well as information on the time to an event (e.g. death) or (multiple) recurrent event times (e.g. repeated bacterial and viral infections). It was interesting to study how these markers relate with the event times and how their updated values may change prognosis. This could help to guide decision making for patient care.
In part I we applied joint modeling to study the association between longitudinal and survival data. We also performed dynamic predictions of survival probabilities for a new subject, using marker values that were accrued overtime. We present an extension of the application of joint modelling to a setting with multiple markers and multi-type recurring events.
In part II we applied landmarking as an alternative to joint modelling for performing dynamic predictions of survival probabilities. Landmarking circumvents possible computational complications of fitting time-dependent covariates, making it easier to compute survival probabilities compared to using joint models. We present an extension of the application of landmarking to a setting with recurring events of the same type.
In part III we focused on validating prediction models in the presence of multiply imputed data. It was unclear how resampling should be performed over the imputed data sets when internal model validation was performed via bootstrap resampling. We investigated four ways of handling the multiply imputed data sets in the validation procedure.