Title: Transformation methods for smoothed estimation from interval-censored data
Authors: Rebecca Betensky - New York University (United States) [presenting]
Jing Qian - University of Massachusetts Amherst (United States)
Abstract: Interval censoring is common in longitudinal studies of Alzheimer's disease due to observation of monotonic processes at periodic visits, e.g., time to Clinical Dementia Rating scale score of 0.5 and amyloid at a given time. In addition, monotonic AD markers (e.g., amyloid) may not be observed at a time of interest, and thus are interval-censored at that time. Additionally, it may be of interest to estimate the time between events (latency), where the initiating event is interval-censored, such as time from CDR of 0.5 to death. The nonparametric information contained in interval-censored data lies in minimal intersections of the observed intervals. Thus, data that are highly intersecting are less informative than those that are not. We propose a novel solution through a linear transformation of the unobserved event time using a discrete Uniform random variable and a scalar parameter selected to satisfy an independence condition. We then calculate the NPMLE for the distribution of the transformed time, which yields a smoothed estimator of the original time.