Statistical methods for joint modeling and dynamic predictions for clustered data

点击翻译按钮获取中文摘要

Joint modeling offers great flexibility in capturing the temporal dynamics of longitudinal and recurrent outcomes on right-censored event times and provides dynamically updated risk prediction using all available information. Although statistical methods for joint modeling and dynamic prediction are widely available for independent data, few methods have been developed for clustered data. The main goals of my research program are to develop and evaluate statistical methods for joint modeling and dynamic predictions and address the statistical challenges arising from the analysis of clustered data in family-based genetic studies. The proposed research has the following specific aims: 1.Develop efficient statistical methods for joint modeling for clustered data including longitudinal, recurrent and survival outcomes collected during follow-up that account for study design, multiple terminal events, time-dependent covariates, and within-cluster correlation; 2.Develop statistical methods for measuring dynamic prediction accuracy that account for right-censoring, competing events, and within-cluster correlation and evaluate their performance using simulation studies; and 3.Integrate high-throughput genomic data into joint models to improve individualized dynamic predictions. In Aim 1, I will develop various types of joint models for clustered data: 1) trivariate joint models for survival, recurrent and longitudinal outcomes measured during follow-up that incorporate a generalized linear model for various types of longitudinal data and 2) multistate joint models that account for competing events and successive events along with recurrent and longitudinal outcomes. In addition, I plan to address two important issues in joint modeling: 3) modeling complex dependence structure and 4) adjusting ascertainment bias due to sampling schemes. Finally, I plan to 5) develop composite likelihoods to reduce computational burdens for joint modeling with complex correlation structure and ascertainment correction. In Aim 2, I will derive dynamic predictions for the joint models proposed in Aim 1 by incorporating individual and familial history of events and assess their predictive accuracy based on the time-dependent area under the receiver operating characteristic curve (AUC) and Brier scores (BS). As the AUC and BS were derived for independent data, I plan to develop modified AUC and BS estimators to account for within-cluster correlation, censoring and competing events. In Aim 3, I will integrate next generation sequencing data into joint models to identify genetic polymorphisms associated with the event(s) of interest and longitudinal outcomes to improve dynamic predictions of the event risk. The proposed research will improve risk estimation and individual's dynamic predictions accounting for multiple events, longitudinal outcomes and high-throughput data with the potential for a number of applications in genetics and health sciences.