Elizabeth Sigworth Westerberg dissertation defense – November 2
PhD candidate Elizabeth Anne Sigworth Westerberg will be defending her dissertation on Wednesday, November 2, at 10 a.m. Central Time, on-site and online. Her advisor is Qingxia (Cindy) Chen. All are invited.
The in-person event will be held in Room 1402 (14th floor) at 2525 West End Avenue.
For access to the event stream, please contact the department at biostatistics[at]vumc[dot]org.
Improving Inferential and Computational Efficiency for Real-World Data
Large clinical and epidemiological studies allow researchers to study drug and exposure effects across populations and increase the chance of observing rare events. However, with large studies come increased data volume, follow-up time, and overall complexity, necessitating analytical techniques that scale with increasing data sizes or perform efficient analyses on subsets or summaries. In this dissertation, we explored several facets of inferential and computational efficiency in the context of big data. First, we considered a case where complete data is unavailable due to data sharing limitations and developed a Bayesian meta-analysis model that uses clinical trial summary data to estimate equivalent dose pairs of two drugs based on adverse event rates. We demonstrated its efficiency when compared to models fit on individual subject data via simulation study, including under model misspecification. We then applied our method to clinical trial data on two taxane chemotherapy drugs known to cause peripheral sensory neuropathy. Next, we compared methods for estimating time-varying effects in the Cox model, evaluating via simulation five different options in terms of inferential and computational efficiency when applied to effects of varying complexity. We then applied the best performing methods to tumor survival data pulled from Vanderbilt University Medical Center electronic health records, using available ICD-9 codes and drug exposures to estimate the potential time-varying effect of metformin on survival in several tumor types. Finally, we applied the findings of our time-varying effect exploration to two-phase study designs, where inexpensive covariates are available for all participants, but an expensive covariate that is believed to have a time-varying effect is available only for a subset. This work extends on the previous development of a two-phase approach for the Cox proportional hazards model, and uses B-spline sieves to estimate the conditional density of expensive covariates given the available inexpensive covariates. We demonstrated its performance via simulation and comparison to existing methods that incorporate only Phase II data, and finished with an application to the Shanghai Women’s Health Study on the effects of a biomarker of interest over time on colorectal cancer development.
