Strategies for imputing missing covariates in accelerated failure time models.

Area of Science:

• Biostatistics
• Epidemiology
• Health Research Methodology

Background:

• Missing data is common in biomedical studies, particularly those with survival outcomes.
• Multiple Imputation via Chained Equations (MICE) is a flexible method for handling multivariate missing data.
• Current MICE application often relies on simplified conditional models, potentially compromising results due to a lack of practical guidelines.

Purpose of the Study:

• To propose and evaluate rationales for appropriate MICE specifications in survival analysis.
• To address the lack of practical guidelines for specifying conditional models in MICE.
• To improve the accuracy and robustness of imputation for survival outcomes.

Main Methods:

• Developed a strategy based on specifying a joint model for involved variables.
• Derived and approximated conditional distributions from the joint model for imputation.
• Proposed using separate models to impute variables, incorporating failure status for survival outcomes.
• Validated the approach through simulations using accelerated failure time models.

Main Results:

• Commonly used simple MICE specifications can yield suboptimal results in survival analyses.
• The proposed strategy, involving careful modeling of conditional distributions, demonstrated good performance.
• Imputations based on the proposed strategy were robust to model misspecifications.
• The strategy is compatible with existing imputation software using fully conditional specifications.

Conclusions:

• Mechanical application of MICE without careful consideration of conditional models can lead to inaccurate results.
• A thoughtful approach to modeling conditional distributions, especially incorporating failure status, is essential for reliable MICE in survival studies.
• The proposed strategy offers a robust and implementable method for improving MICE performance in biomedical research with survival data.