Biostatistics Dissertation Defense

Xingyuan Li - Modeling Exposure-Time-Response Association in the Presence of Competing Risks

Friday 1/18 1:00PM - 3:00PM
7139 Public Health, Peterson Seminar Room

Xingyuan Li of the Department of Biostatistics defends her dissertation on "Modeling Exposure-Time-Response Association in the Presence of Competing Risks". 

Committee Chairperson: Joyce Chang, PhD, Department of Medicine

Committee Members: 
Stewart Anderson, PhD, Department of Biostatistics
Yu Cheng, PhD, Department of Statistics
Julie M. Donohue, PhD, Department of Health Policy Management 
Robert Krafty, PhD, Department of Biostatistics 

Graduate faculty of the University and all other interested parties are invited to attend


In longitudinal pharmacoepidemiology studies, the exposures may be chronic over a period of time and the intensity, duration, and timing of the exposures may vary among subjects. Further challenges may arise when the data involve competing risks, where subjects may fail from one of multiple events and failure from one precludes the risk of experiencing others.

A model that predicts the risk of a health outcome from the longitudinal pattern of exposures can help the researchers and health care professionals identify high-risk individuals more efficiently. However, methodological challenges arise in at least three aspects: 1) how to account for the time-varying nature of exposures?  2) is there is a cumulative and latency effect such that exposures could contribute to the risk of a health outcome incrementally over time? and 3) is there any competing event that precludes the outcome of interest?

The dissertation focuses on how to overcome these challenges in the development of methods for directly modeling the probability of the main event of interest (aka, cumulative incidence function, CIF). In Section 1 of the dissertation, we propose a subdistribution hazards regression model. The model incorporated weighted cumulative effects of the exposure so the intensity, duration, and the timing of the exposure can be taken into consideration simultaneously. We incorporated penalized cubic B-splines into the partial likelihood equation to estimate the weights. Performance of the proposed model was evaluated through a simulation study.

In Section 2 of the dissertation, we extend the model in Section 1 to a more general form and propose a generalized transformation regression model. In Section 1, the subdistribution hazard ratio was modeled as a function of time lag between exposure initiation and risk estimation. In this section, we allowed the subdistribution hazard ratio to be a bivariate function of both the time lag and the level of time-varying exposures. We also introduced various link functions to model the CIF, such that the method in Section 1 can be considered as a special case of this general definition. We used tensor product splines with penalties to flexibly estimate the bi-dimensional surface for the cumulative effects of exposure, and incorporated an additional ridge penalty to constrain the model behavior at the right tail of lag dimension. Extensive simulations were conducted to evaluate the model performance.

To illustrate our proposed method, we applied the models to investigate the association between patterns of prescription opioid use and the risk of overdose (and other adverse outcomes) for Medicare and Medicaid beneficiaries, treating mortality as a competing risk.

PUBLIC HEALTH SIGNIFICANCE: We introduced novel statistical methods that directly estimate the cumulative effect of time-dependent exposures, to quantify the exposure-time-response association in which the intensity, duration, and timing of an exposure are taken into consideration while the event of interest is subject to competing risks. Using opioid use in Medicare and Medicaid as examples, we showed that the proposed model is able to distinguish different prescription patterns even though these patterns have the same overall intensities during the study period, which allows clinicians and health policy practitioners to better understand the long-term effects of opioid use, identify high-risk individuals for overdose (and other adverse outcomes), and guide decisions that may mitigate the prescription opioid epidemic in the United States. The method is also generalizable to other pharmacoepidemiological studies regarding prescription medication use.


Last Updated On Monday, September 23, 2019 by Valenti, Renee Nerozzi
Created On Friday, January 4, 2019