Bias correction for score tests arising from misspecified proportional hazards regression models

When censoring depends on both treatment group and covariates, tests for a randomised treatment effect arising from Cox's proportional hazards model are not in general centred at zero under the null and can be seriously distorted. We propose a corrected test that modifies the at-risk indicator in a way that corrects for this bias. The resulting test is asymptotically valid regardless of whether or not the fitted model is correctly specified. Implementation of the corrected test requires that the dependence of censoring on treatment group and covariates be modelled. Simulations indicate that the bias-corrected test performs well, even when the dependence of censoring on treatment group and covariates is not modelled exactly, and maintains high efficiency relative to the uncorrected test when the latter is valid.