A1385
Title: A model-free correlation coefficient for censored data
Authors: Linlin Dai - Southwestern University of Finance and Economics (China) [presenting]
Tengfei Li - University of Texas MD Anderson Cancer Center (United States)
Kani Chen - HKUST (Hong Kong)
Abstract: In clinical studies, assessing statistical associations between covariates and survival outcomes is crucial. To date, there has been no formally defined model-free correlation coefficient for right-censored data that can measure the strength of associations. Traditional methods, such as the Cox proportional hazards model, often struggle with the complexities of non-monotonic or nonlinear relationships. A censored rank-based correlation coefficient (CRC) is introduced. It consistently estimates a new dependence measure taking values in [0,1] and equaling 0 or 1 if and only if the variables are independent or one is a measurable function of the other. The CRC is entirely model-free without depending on the distributions of the variables. It facilitates quick computation with a complexity of O(nlogn) and can effectively detect nonlinear and non-monotonic effects, even under heavy censoring. The p-values for testing independence can be obtained using a power-consistent permutation method. The CRC shows strong consistency and asymptotic normality, outperforming the Cox model and other methods in detecting nonlinear associations in both simulations and real data from the Alzheimers Disease Neuroimaging Initiative, successfully identifying proteins that existing methods fail to detect.
