Integrative Survival Analysis with Application to Suicide Risk

The concept of integrating data from disparate sources to accelerate scientific discovery has generated tremendous excitement in many fields. The potential benefits from data integration, however, may be compromised by the uncertainty due to incomplete/imperfect record linkage. Survival outcomes with extrinsic or intrinsic uncertainty due to such as inaccurate measurement or diagnosis test with low sensitivity and/or specificity also bring challenges in statistical analysis. In this thesis, we propose integrative statistical approaches for analyzing survival data that consist of uncertain event times or status arising from data integration or inaccurate diagnosis with applications to suicide risk studies. Firstly, we develop an integrative Cox proportional hazards regression, in which the uncertainty in the integrated event times is modeled probabilistically. The estimation procedure combines the ideas of prole likelihood and the expectation conditional maximization algorithm (ECM). Simulation studies demonstrate that under realistic settings of imperfect data linkage, the proposed method outperforms several competing approaches including multiple imputation. Secondly, we propose an approach based on the Cox proportional hazard cure model for survival data with a cure fraction and uncertain event status. We develop the model estimation via an EM algorithm. Variable selection procedure through regularization by elastic net penalty is derived based on cyclic coordinate descent and majorizationminimization (MM) algorithm. Simulation studies demonstrate that the under various settings, the proposed method outperforms several competing approaches in terms of estimation, variable selection, and out-of-sample predictions on cure status and survival outcomes. A marginal screening analysis using the proposed integrative Cox model and variable selection using the proposed Cox cure model is performed, respectively, to identify risk factors associated with death and determined suicide attempts following suicide-related hospitalization in Connecticut. The identied diagnostic codes are consistent with existing literature and provide new insights on suicide risk prediction and prevention. All the aforementioned methods are implemented in an R package named intsurv accessible at https://cran.r-project.org/package=intsurv.