Method of trimmed moments for robust fitting of parametric failure time models

摘要

Parametric models are frequently used in modeling survival and reliability data in the presence of censoring. In such instances, maximum likelihood estimation is a standard tool used to fit parametric models. However, when the underlying model is misspecified or contaminated the maximum likelihood procedure may be severely affected by that and hence lead to very poor results. Therefore, robust methods which discount the effects of contamination and model misspecification, can provide a desirable alternative for parameter estimation. We propose a robust parametric model-fitting procedure for continuous failure time models. The procedure relies on the underlying principle of the classical method of moments and it can achieve various degrees of robustness and efficiency. Its asymptotic distribution is derived and small-sample performance is studied using simulations. The newly developed method is straightforward to implement in practice as it (typically) does not require numerical solution of non-linear equations. The usefulness of the procedure is illustrated using two real data examples.