This proposal addresses the development and application of new analytic strategies for three dilemmas in mixed recurrent-event and panel-count data (mixed recurrent-event data). While recurrent-event data and panel-count data are both generated from recurrent-event processes, they have different observation systems. In the former, subjects are observed continuously and in the latter, subjects are observed only at discrete time points. In many single cohort studies, subjects are observed continuously during some data collection periods and only periodically during other data collection points, resulting in mixed recurrent-event data. Meanwhile, the recurrent-event process could be stopped by a terminal event (e.g., menopause prevents the possibility of further pregnancies), the recurrent event could have a nonignorable duration (e.g., each pregnancy lasts for up to 40 weeks), and/or the subjects are clustered (e.g., occurring pregnancies for subjects from the same family will most likely not be independent so the family becomes a cluster). The current common practice for mixed recurrent-event data is to approximate or simplify these complex data, resulting in potentially misleading conclusions. To date, no existing statistical methods have been developed that will allow for comprehensively assessing the three problems in this complex data structure. There is an urgent need to develop intuitive, efficient, and computationally feasible methods for analyzing complex data in event history studies. Our preliminary studies have efficiently and flexibly addressed basic regression analysis of mixed recurrent-event data alone. Based on our previous work, this project proposes to use data from the renowned longitudinal Childhood Cancer Survivor Study (CCSS) and Pediatric Brain Tumor Consortium (PBTC) to: 1) Develop semiparametric methods for regression analysis of mixed recurrent-event data with terminal events; 2) Develop semiparametric methods for regression analysis of mixed recurrent-event data with nonignorable duration; and, 3) Develop semiparametric methods for regression analysis of mixed recurrent-event data with clustering subjects. In all three aims, the proposed methods use marginal models and derive their estimating equations similarly, so we can combine the methods proposed in Aim 1, 2 and 3 to simultaneously deal with these problems should they occur within the same data set. These approaches potentially have strong statistical and clinical relevance for the study of complex event history data.
/Relevance Statement This proposal addresses three dilemmas in mixed recurrent-event and panel-count data: terminal event, nonignorable duration, and/or clustering subjects, which we are confronted with in both motivating examples from the Childhood Cancer Survivor Study and the Pediatric Brain Tumor Consortium. Since there are no existing statistical methods that will comprehensively assess the three problems in this complex data structure, we propose the development and application of new analytic strategies in order to provide useful tools for statisticians and clinicians in cancer and other fields which rely on long-term cohort follow-up studies.
| Xu, Da; Zhao, Hui; Sun, Jianguo (2018) Joint analysis of interval-censored failure time data and panel count data. Lifetime Data Anal 24:94-109 |
| Yu, Guanglei; Zhu, Liang; Sun, Jianguo et al. (2018) Regression analysis of incomplete data from event history studies with the proportional rates model. Stat Interface 11:91-97 |
| Yu, Guanglei; Zhu, Liang; Li, Yang et al. (2017) Regression analysis of mixed panel count data with dependent terminal events. Stat Med 36:1669-1680 |
