On the Use of Buckley and James Least Squares Regression for Survival Data

Janez Stare, Harald Heinzl, and Frank Harrell

The method of Buckley and James (1979) for fitting linear regression models to censored data has been shown to have good statistical properties under usual regularity conditions. Nevertheless, even after 20 years of its existence, it is almost never used in practice. We believe that this is mainly due to lack of software and we state three reasons for using the method. We argue against some findings by Heller and Simonoff (1992) and in remainder of the paper briefly explore the method’s behaviour under model misspecification. We conclude that at present there are no good procedures for checking the model’s assumption of homoscedasticity under censoring and therefore do not recommend using this method with censoring higher than 20%.