Model diagnostics for censored regression via randomized survival probabilities

TL;DR

This paper introduces normalized randomized survival probability residuals for censored regression, providing a new diagnostic tool that effectively detects non-linearity and subtle model misfits, outperforming traditional methods in certain scenarios.

Contribution

The paper proposes a novel residual-based diagnostic method for censored regression using randomized survival probabilities, with proven uniform distribution under the true model.

Findings

NRSP residuals are uniformly distributed under the true model.

The non-linear test based on NRSP residuals has higher power in detecting non-linearity.

NRSP residual diagnostics detect subtle non-linear relationships in real data.

Abstract

Residuals in normal regression are used to assess a model's goodness-of-fit (GOF) and discover directions for improving the model. However, there is a lack of residuals with a characterized reference distribution for censored regression. In this paper, we propose to diagnose censored regression with normalized randomized survival probabilities (RSP). The key idea of RSP is to replace the survival probability of a censored failure time with a uniform random number between 0 and the survival probability of the censored time. We prove that RSPs always have the uniform distribution on $(0,1)$ under the true model with the true generating parameters. Therefore, we can transform RSPs into normally-distributed residuals with the normal quantile function. We call such residuals by normalized RSP (NRSP residuals). We conduct simulation studies to investigate the sizes and powers of statistical tests based on NRSP residuals in detecting the incorrect choice of distribution family and non-linear effect in covariates. Our simulation studies show that, although the GOF tests with NRSP residuals are not as powerful as a traditional GOF test method, a non-linear test based on NRSP residuals has significantly higher power in detecting non-linearity. We also compared these model diagnostics methods with a breast-cancer recurrent-free time dataset. The results show that the NRSP residual diagnostics successfully captures a subtle non-linear relationship in the dataset, which is not detected by the graphical diagnostics with CS residuals and existing GOF tests.