Dr. Peter Adamic
Department of Mathematics and Computer Science
Title: " Nonparametric Competing-Risks Modeling"
In this seminar, Self-Consistent (SC) Expectation-Maximization (EM) algorithms will be examined that
nonparametrically estimate the associated single risk survival functions in a competing risks context.
The algorithms are a generalization of the univariate algorithms of Efron (1967) and Turnbull (1974, 1976),
and are carried out in the presence of both double and interval-censored data. Unlike any previous nonparametric
models proposed in the literature to date, the algorithms will explicitly allow for the possibility of masked
modes of failure, where failure is known to occur due to a subset from the set of all possible causes. A second
major step in the model building process is subsequently pursued. Many of the modeling difficulties associated
with censoring are remedied by utilizing a modified SC/EM algorithm that incorporates kernel smoothing at each
iteration of the generalized algorithms. This modification is an extension of similar work done by Li, Watkins,
and Yu (1997), and Braun, Duchesne, and Stafford (2005). The talk will conclude with a brief discussion of the
usefulness of the proposed algorithms within the current state of competing risks scholarship - as well as
exploring areas where derivative work would prove to be fruitful.