Estimation and Model Selection for the Competing Risks Model with Masked Causes of Failure
Department of Statistics
University of Toronto
Monday, 6 December 2004
The competing risks model is useful in settings in which subjects may fail/die due to a number of different causes. Complications arise when some of the observed failures have causes known only to belong to a certain group, i.e. they are group-masked. In such instances the cause-specific hazard estimates and the masking probabilities are of interest.
We use the Expectation-Maximization (EM) algorithm to find the maximum likelihood estimates for a semiparametric model in which each cause-specific hazard rate is piecewise constant. The model is flexible enough to handle the case of grouped data and/or time-varying masking probabilities. We show how it can be used to construct likelihood ratio tests for the proportional hazards and symmetry assumptions. We consider the impact of hazard misspecification on the statistical analysis and propose a model-selection strategy based on the minimum description length principle. The method is applied to real and simulated data.
Refreshments will be served at 3:45 p.m. in Room 008 of the Statistics Building