Benjamin Langworhy, Postdoctoral Associate at Harvard T.H. Chan School of Public Health, will present:
“Principal Components Analysis for Right-Censored Data”
Abstract: This talk will discuss the estimation of covariance and correlation for multivariate time-to-event data, where a single subject can have multiple non-competing events. The methods are inspired by data from the comparator arm of a clinical trial for pancreatic cancer patients in which there are multiple adverse events. Based on medical knowledge we believe that there are certain groupings of more similar event types (hematologic, gastrointestinal, constitutional), and wish to confirm these groupings using principal components analysis (PCA). However, PCA requires estimation of the covariance or correlation between event types, and this is not possible when data are right-censored. We developed a method to estimate the covariance and correlation between multiple event types by considering counting processes and their associated martingales, which allows for PCA to be estimated for right censored data. We applied the proposed method to the pancreatic cancer clinical trial data and identified two groupings of adverse events that are consistent with prior medical knowledge.
All are welcome.