Using Logistic Regression and Variable Selection to Model Time-To-Event Data:Applications to Tree Phenology and Graduation Time of Engineers

Abstract: The day of bud burst (DBB) and leaf senescence are two examples of time-to-event phenological processes influenced by climate factors. Time to graduation or quitting for engineering students is another example of time-to-event data, with the added complication of having multiple possible outcomes, or absorbing states. This master thesis elaborates upon the models presented in Song (2010) "Stochastic Process Based Regression Modeling of Time-to-event Data". The time-to-event model is extended to use many different covariates, and Lasso regularization techniques are used for variable selection, resulting in compact and statistically relevant models. Models with multiple outcomes are shown to be able to perform classification of students sequentially over time. For the phenological examples, DBB is predicted with an accuracy of a couple of days while leaf senescence proves to be a harder problem, possibly in need of additional climate data not included in this analysis. Overall the model of Song is shown to have great promise and versatility for modeling of time-to-event data.