Iterative partial least squares with right-censored data analysis: a comparison to other dimension reduction...

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Linear models with right-censored data often yield unstable or inestimable regression parameters when covariates are numerous.
• High-dimensional data poses challenges for traditional regression techniques, particularly with survival data.

Purpose of the Study:

• To propose an iterative partial least squares (PLS) algorithm for estimating covariate effects in linear models with right-censored responses.
• To develop a method for predicting future responses using a given set of covariates.

Main Methods:

• An iterative PLS algorithm based on the Buckley-James estimating equation was developed.
• Leave-two-out cross-validation was employed to select the optimal number of PLS components.
• Simulations were conducted to compare the proposed method with existing dimension reduction techniques.

Main Results:

• The proposed PLS algorithm provides stable and estimable regression parameter estimates even with many covariates and right-censored responses.
• Leave-two-out cross-validation effectively determined the number of components to minimize prediction error.
• Simulation studies demonstrated the superiority of the proposed method over other dimension reduction techniques.

Conclusions:

• The iterative PLS algorithm offers a robust solution for regression analysis with right-censored data and high-dimensional covariates.
• The method enhances the accuracy of covariate effect estimation and future response prediction.
• This approach is applicable to complex datasets, as demonstrated by its use in AIDS Clinical Trials Group protocol 333.