B1618
Title: Positive-unlabeled survival data analysis
Authors: Tomoki Toyabe - Keio University (Japan) [presenting]
Takahiro Hoshino - Keio University and RIKEN (Japan)
Abstract: A novel framework is introduced for the analysis of positive-unlabeled (PU) data where the survival time for subjects with events is observed as positive data. In contrast, the censored time is observed as unlabeled data, with the event occurrence status remaining uncertain. In fields such as medical and marketing, actual event occurrences might not always be accurately observed due to factors like hospital transfers or purchases at different stores. By treating previously misclassified data as negative or unlabeled, analysis results are obtained that reflect the true situation more accurately. Two cases within the PU context of the so-called "case-control scenario" are considered: when the truncation time is observed for positive data, and when it is not. Simulation results indicate that while traditional survival time analyses might yield significantly biased outcomes under such data situations, the proposed estimation method holds the potential to produce valid results.
