Abstract
Aim: To propose an updated algorithm with an extra step added to the Newton-type algorithm used in robust rank based non-parametric regression for minimizing the dispersion function associated with Wilcoxon scores in order to account for the effect of covariates.
Methodology: The proposed accelerated failure time approach is aimed at incorporating right random censoring in survival data sets for low to moderate levels of censoring. The existing Newton algorithm is modified to account for the effect of one or more covariates. This is done by first applying Mantel scores to residuals obtained from a regression model, and second by minimizing the dispersion function of these scored residuals. Diagnostic check of the model fit is performed by observing the distribution of the residuals and suitable Bent scores are considered in the case of skewed residuals. To demonstrate the efficacy of this method, a simulation study is conducted to compare the power of this method under three different scenarios: non-proportional hazard, proportional and constant hazard, and proportional but non-constant hazard.
Results: In most situations, this method yielded reasonable estimates of power for detecting an association of the covariate with the response as compared to popular parametric and semi-parametric approaches. The estimates of the regression coefficient obtained from this method were evaluated and were found to have low bias, low mean square error, and adequate coverage. In a real-life example pertaining to pancreatic cancer study, the proposed method performed admirably well and provided a more realistic interpretation about the effect of covariates (age and Karnofsky score) compared to a standard parametric (lognormal) model.
Conclusion: In situations where there is no clear best parametric fit for time-to-event data with moderate level of censoring, the proposed method provides a robust alternative to obtain regression coefficients (both adjusted and unadjusted) with a performance comparable to that of a proportional hazards model.