Joint modeling time-to-event and longitudinal data using Markov chain Monte Carlo methods with application to the Proscar(TM) long-term efficacy and safety study

Many clinical and epidemiologic studies often collect longitudinal repeated measurements of a time-dependent disease marker and time-to-event of a disease. An example of this is total symptom score and urologic surgical interventions for BPH and urinary retention with catheterization in the PLESS study. There have been many approaches suggested for analysis of longitudinal data and time-to-event data separately, but very few for modeling them jointly. This dissertation described the methodology required to approach the complex problem of simultaneously modeling longitudinal repeated measurements of a covariate and relating the covariate to disease risk. Furthermore, this dissertation also described the methodology of assessing the overall treatment effect when the longitudinal repeated measurement was a potential surrogate endpoint for the clinical endpoint of predicting disease risk. Markov chain Monte Carlo (MCMC) techniques were used to estimate parameters of interest. Extensive simulation studies and a data application to PLESS study demonstrated that joint modeling longitudinal and time-to-event data provided a superior alternative to modeling these two types of data separately.; The joint modeling approach was evaluated and compared with the approach of analyzing these two types of data separately through simulation studies. Extensive simulation studies showed that simultaneously modeling time-to-event data and repeated measurement using MCMC produced promising results. The parameter estimates from the disease risk model were improved considerably by accounting for the inherent measurement error of the repeated measurements. Conversely, the parameter estimates from the longitudinal model were also improved by incorporating the informative drop-out information.; A re-parameterized joint model was proposed when the longitudinal repeated measurement might be a potential surrogate endpoint for the time-to-event clinical endpoint. The proposed method was implemented through re-parameterization of the joint model. The implementation of this re-parameterized joint model enabled the estimate of "an overall treatment effect" which combined the contributions from both endpoints. Extensive simulation studies indicated that even when the longitudinal repeated measurement was only a partial surrogate endpoint, combining the contributions from both sources always resulted in an enhanced estimate of the overall treatment effect.