Bias and variance reduction in nonparametric estimation of time-dependent accuracy measures

Area of Science:

• Biostatistics
• Survival Analysis
• Nonparametric Statistics

Background:

• Censored survival data presents challenges for accuracy measure estimation.
• Existing nearest neighbor estimators can suffer from finite-sample biases and high asymptotic variances.

Purpose of the Study:

• To develop a new nonparametric approach for estimating time-dependent accuracy measures in censored survival data.
• To reduce finite-sample biases and asymptotic variances of accuracy measure estimators.
• To propose practical bandwidth selection procedures.

Main Methods:

• Developed a new nonparametric approach based on estimable survival processes.
• Applied smoothing techniques to marker processes to reduce asymptotic variances.
• Proposed theoretically valid procedures for bandwidth selection.
• Extended methodology for stratified and multi-marker survival data.

Main Results:

• The new estimators demonstrated superior performance compared to nearest neighbor and inverse censoring weighted estimators in simulations.
• The proposed method effectively reduces bias and variance in accuracy measure estimation.
• Illustrative examples using AIDS Clinical Trials Group study 175 and coronary artery disease data confirmed the methodology's utility.

Conclusions:

• The novel nonparametric approach offers improved accuracy for time-dependent measures in censored survival data.
• The methodology is robust, computationally feasible, and extendable to complex survival data structures.
• This work provides a valuable advancement for statistical analysis in survival data.