Martingale-residual-based greedy model averaging for high-dimensional current status data.

Area of Science:

• Statistics
• Biostatistics
• Survival Analysis

Background:

• Current status data present unique challenges in failure time analysis.
• Existing variable selection methods for failure time data often overlook model selection uncertainty.
• This uncertainty can impact the accuracy of risk predictions.

Purpose of the Study:

• To propose two optimal model averaging methods for semiparametric additive hazards models with current status data.
• To enhance the prediction accuracy of risk quantities like survival probability.
• To address the statistical inference issues arising from model selection.

Main Methods:

• Development of two optimal model averaging methods.
• Utilizing martingale residuals processes to define a delete-one cross-validation (CV) process.
• Derivation of new CV functional criteria for selecting model weights.
• Implementation using a greedy algorithm.

Main Results:

• The proposed model averaging methods enhance prediction accuracy for risk quantities.
• Asymptotic optimality of the model averaging approaches is established.
• Convergence of the greedy averaging algorithms is demonstrated.
• Simulation experiments confirm the effectiveness and superiority of the proposed methods.

Conclusions:

• The proposed optimal model averaging methods effectively address model selection uncertainty in current status data.
• These methods provide more accurate risk predictions compared to traditional approaches.
• The study offers a robust statistical framework for analyzing current status data.