Addressing immortal time bias in precision medicine: Practical guidance and methods development

Area of Science:

• Biostatistics
• Epidemiology
• Health Services Research

Background:

• Immortal time bias is a significant challenge in observational studies, particularly in precision medicine evaluations.
• Existing adjustment methods have theoretical limitations that can impact the validity of real-world evidence.

Purpose of the Study:

• To compare the strengths and limitations of common immortal time adjustment methods.
• To propose and evaluate a novel approach using multiple imputation (MI) for immortal time bias adjustment.
• To provide practical guidance for implementing MI in precision medicine research.

Main Methods:

• Comparative analysis of landmark analysis, time-distribution matching, and time-dependent analysis against the proposed MI method.
• Development of practical guidance for MI application, including imputation method selection, model specification, and analysis pooling.
• A real-world case study using matched cohort design to evaluate survival benefits of whole-genome and transcriptome analysis in advanced cancers, applying both time-distribution matching and MI.
• Bootstrap simulations to assess imputation sensitivity to missing data and sample size.

Main Results:

• Multiple imputation (MI) theoretically offers advantages over other methods by minimizing information loss and better characterizing statistical uncertainty.
• MI explicitly accounts for patient characteristics influencing immortal time distributions, reducing potential bias.
• In the case study, MI and time-distribution matching yielded similar survival analysis results, though MI produced higher standard errors.
• Imputed immortal time remained stable across simulation scenarios.

Conclusions:

• Robust immortal time adjustment methods are crucial for generating unbiased, decision-grade real-world evidence in precision medicine.
• Multiple imputation (MI) presents a promising and effective solution for addressing immortal time bias in these evaluations.