Group lasso priors for Bayesian accelerated failure time models with left-truncated and interval-censored data

Area of Science:

• Biostatistics
• Epidemiology
• Genetics

Background:

• Characterizing time-to-event outcomes in health research involves analyzing high-dimensional risk factors.
• Alzheimer's disease (AD) research often uses prospective cohort studies with complex data, including varying ages at enrollment and interval-censored outcomes.
• Existing statistical models face challenges with high-dimensional, grouped covariates common in genetic and clinical risk factor analysis.

Purpose of the Study:

• To propose a novel Bayesian accelerated failure time model tailored for left-truncated and interval-censored time-to-event data.
• To incorporate a group-structured lasso penalty to effectively handle high-dimensional, grouped covariates.
• To develop an R package for estimation using Markov chain Monte Carlo (MCMC) sampling.

Main Methods:

• Development of a Bayesian accelerated failure time model incorporating a group-structured lasso penalty.
• Implementation of an MCMC sampler within an R package for model estimation.
• Simulation studies to compare the proposed method against an ordinary lasso penalty.

Main Results:

• The proposed group-structured lasso penalty method demonstrates effective variable selection and estimation for grouped, high-dimensional covariates.
• The method was successfully applied to identify predictive groups of genetic and clinical risk factors for AD.
• Performance evaluation through simulations confirmed the method's advantages over standard lasso.

Conclusions:

• The developed Bayesian model with group-structured lasso penalty is a powerful tool for analyzing complex time-to-event data in health research, particularly for AD.
• This approach facilitates the identification of important risk factor groups, advancing our understanding of disease etiology.
• The R package provides a practical implementation for researchers studying time-to-event outcomes with high-dimensional, structured covariate data.