Joint Modeling of Longitudinal and Time-to-Event Data: An Application to Cardiovascular Disease Open Access
Downloadable ContentDownload PDF
Joint modeling of longitudinal and survival data refers to a class of models to link the repeated measurements of a variable to the risk of experiencing an event. The classical survival analysis framework imposes assumptions that may not hold true in practice and considers only the current value of the variable while predicting survival probability. Alternatively, joint models summarize the behaviour of the variable's trajectory over time by a few key parameters and enable the researcher to assess the latent effect of the variable's long-term behaviour, rather than a single reading. This thesis applies the shared parameter model and its modifications to link the history of a subject's total cholesterol and systolic blood pressure to predict the risk of experiencing cardiovascular disease or death over the subject's lifetime. The clinical findings are that the baseline reading of total cholesterol alone performs as well as the shared parameter model in predicting time to event. In the case of systolic blood pressure, the shared parameter model performs best, indicating that there is a significant effect of the variable's long-term trajectory on risk of death. The thesis also compares the performances of these models as estimated from several, frequently spaced measurements and from a few measurements that are spaced far in time. In the case of systolic blood pressure, we find that its trajectory as a linear function of time is adequately estimated by only three randomly sampled observations in lieu of the entire history.