Computational Methods for Semiparametric Linear Regression with Censored Data

Abstract

Semiparametric linear regression with censored data assumes a linear relationship between failure time and covariates without specifying the distributional form of the error term. This approach has attracted considerable attention recently. Most notably, rank regression methods have been derived for parameter estimation, hypothesis testing, and goodness-of-fit analysis. The implementation of these methods requires minimizing discrete objective functions with multiple local minima. Conventional optimization algorithms cannot be used to solve such minimization problems. We develop computational methods to implement rank regression procedures using simulated annealing. Two real data sets are used for illustration. Applications of the new algorithms to the modified least squares estimator of Buckley and James and several other related problems are also described.