Abstracts

Some reflections on a few of Prof. David Oakes' contributions in survival analysis

Rebecca Betensky
NYU School of Global Public Health 

Prof. David Oakes has made fundamental contributions to the field of Survival Analysis, in addition to clinical trials and neurological applications.  His papers are unique in their focus and exposition.  In celebration of his illustrious career, I will focus on four of his contributions that exemplify these qualities and are all related to the theme of bivariate times: an independence test in the presence of censoring, bivariate survival models, multiple time scales, and the win ratio. I will explain how each of these has influenced my work.

Modeling and Analysis of Disease incidence and Death Rates with Longitudinal Covariates

Ross L. Prentice
Fred Hutchinson Cancer Center and University of Washington 

The joint survivor function for the occurrence of a disease and death, given covariates, is modeled comprehensively using disease-specific hazard rates, marginal death rates, and dual outcome rates for disease incidence and death jointly, with each hazard rate allowing for possible dependence on covariates including longitudinally measured covariates. To do so the failure time data are recast in terms of the minimum of disease and death times and the three hazard rates are shown to specify the joint distribution of these variables using the uniqueness of the solution to the Volterra integral equation given covariates. Longitudinal measures can be analyzed by studying the dependence of these hazard rates on time-varying covariates, for example using Cox models for each of the three hazard rate functions. Extensions to higher dimensional failure time data and death are outlined. An illustration is provided using data from the Women’s Health Initiative hormone therapy trials.

Analyzing wearable device data with threshold-specific temporal trends

Mei-Cheng Wang
Johns Hopkins Bloomberg School of Public Health

Wearable devices are designed to enable objective and continuous monitoring of physical activity in free-living conditions. With the rapid advancement of wearable computing technology, accelerometer-based devices have become widely adopted in physical activity research. Identifying factors associated with physical activity levels holds significant implications for disease prevention and public health.  We propose a unified analytical framework for wearable device data to examine daily activity patterns over a specified time period. By applying varying thresholds to classify activity intensity, we can identify risk factors while accounting for threshold-specific temporal trends conditionally or unconditionally on whether a lower-level threshold is met. Furthermore, we present and leverage an intriguing structural relationship between conditional and unconditional threshold-specific models to derive inferential insights.  The proposed methodology is demonstrated through an analysis of wearable device data from the National Health and Nutrition Examination Survey (NHANES), highlighting its key features and applicability.  

Evaluating the Effectiveness of COVID-19 Vaccines Over Time

Danyu Lin
The University of North Carolina at Chapel Hill

Approximately 800 million COVID-19 cases and 7 million COVID-19 deaths have been reported to the World Health Organization thus far.  Vaccination is a major tool to combat the COVID-19 pandemic, but its effectiveness wanes over time and tends to be lower against new SARS-CoV-2 variants. The knowledge about the waning effects of vaccination can guide boosting strategies. In a series of papers published in New England Journal of Medicine and JAMA, we reported several large cohort studies using COVID-19 vaccination and case surveillance data from the states of North Carolina and Nebraska. We developed a novel statistical framework to evaluate the time-varying effects of the four generations of COVID-19 vaccines produced in the United States on infections with different SARS-CoV-2 variants and on severe outcomes (hospitalization and death). Our findings have been used by the World Health Organization and the U.S. Centers for Disease Control and Prevention and Food and Drug Administration and reported by The New York Times, The Washington Post, ABC News, and NBC News.

Adaptive RMST Methods for Non-Proportional Hazards in Clinical Trials

Douglas Schaubel
University of Pennsylvania Perelman School of Medicine

The Restricted Mean Survival Time (RMST) is an attractive and interpretable alternative to the hazard ratio for analyzing time-to-event data, particularly when the proportional hazards assumption is violated. A key challenge in RMST analysis is the choice of the restriction time L, which directly impacts the estimation of treatment effects. Selecting a small L risks missing late-emerging effects, while a large L increases variance and reduces statistical power.  To address this, we propose an adaptive, data-driven methodology that selects the optimal restriction time, L*, from a continuous range by maximizing a criterion that balances treatment effect magnitude and statistical precision. Our method ensures coherence between hypothesis testing and estimation, maintaining strong performance regardless of whether proportional hazards hold. We provide rigorous theoretical justification for our method, accounting for the uncertainty introduced by adaptive selection. Recognizing the challenges posed by nonregular estimation under the null hypothesis, we introduce two complementary approaches: a convex hull-based estimator for inference and a penalized approach to improve power. Extensive simulations across realistic survival scenarios demonstrate that our method outperforms traditional RMST analyses and the log-rank test, achieving superior power while maintaining nominal Type I error rates. As a practical illustration, we applied our method to a reconstructed clinical trial with transient treatment effects, identifying clinically meaningful benefits that were overlooked by conventional analyses. Our method, implemented in the R package AdaRMST, provides a robust and adaptive tool for survival analysis in a wide range of settings.  This is joint work with Jinghao Sun and Eric Tchetgen Tchetgen. 

Assessing interventions effects on life history processes with observational data

Richard Cook
University of Waterloo

Real-world observational studies and administrative data are viewed as a rich source of information on disease processes and associated treatment effects. The validity of findings from the analysis of such data critically depends on addressing sources of observation bias as well as confounding arising from treatment by indication. Furthermore disease and related features are dynamic processes and changes in disease states, treatment, confounders and risk factors occur at random points in time. We consider challenges arising from the analysis of data on such processes, motivated by a cohort of individuals with psoriatic arthritis in which patients are seen at irregularly spaced clinic visits, the times of which are related to the underlying disease and treatment process. Through the formulation of joint models for disease, marker and treatment processes we address confounding by indication, while an expanded model addresses possible biases arising from a dependent visit process. We specify and estimate average treatment effects, and use models fitted to the psoriatic arthritis registry data estimate average treatment effect based on the probability of developing additional damaged joints over time through g-computation. This is joint work with Jerry Lawless and Lily Zou.