Parsimonious covariate selection with censored outcomes

Area of Science:

• Statistics
• Biostatistics
• Data Science

Background:

• Selecting relevant covariates is crucial for understanding censored outcomes.
• Existing methods may lack objectivity or struggle with dynamic covariate effects.

Purpose of the Study:

• To propose an objective methodology for selecting parsimonious sets of important covariates associated with censored outcomes.
• To develop a method that simplifies for uncensored outcomes and controls the false selection rate.

Main Methods:

• Iterated forward covariate selection controlled by bounds on the number of covariates and false selection rate.
• Fitting working regression models and estimating conditional prediction error processes.
• Utilizing a novel adequacy measure based on prediction error processes.

Main Results:

• The proposed method asymptotically controls the false selection rate at the nominal level.
• It consistently ranks covariate sets, enabling high-probability recruitment of all important covariates.
• Simulation studies confirm analytical results and compare favorably to competitors.

Conclusions:

• The new objective methodology effectively identifies important covariates for censored outcomes.
• The method offers robust control over false selections and improved covariate set ranking.
• It provides a valuable tool for statistical analysis in various scientific fields.