Imputation of Missing Covariates in Randomized Controlled Trials with Continuous Outcomes: Simple, Unbiased and...

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Statistical Analysis

Background:

• Sophisticated methods like multiple imputation (MI) and maximum likelihood (ML) are typically recommended for missing covariates in nonrandomized studies.
• However, their optimality in randomized controlled trials (RCTs), where treatment assignment is independent of baseline covariates, is less clear.
• Previous research demonstrated this for single missing covariates, but not for multiple missing baseline covariates.

Purpose of the Study:

• To evaluate the performance of MI and ML compared to simple methods for handling multiple missing baseline covariates in RCTs.
• To assess bias and efficiency of treatment effect estimation under various missingness scenarios, including missing completely at random (MCAR) and non-MCAR.
• To provide practical guidance using a chronic low back pain trial example.

Main Methods:

• Derivation of asymptotic relative efficiencies for simple methods under the MCAR scenario.
• Conducting a simulation study to evaluate performance under non-MCAR scenarios.
• Illustrating method implementation with a real-world clinical trial dataset.

Main Results:

• All simple methods yielded unbiased treatment effect estimates, though with increased mean squared residual.
• Mean imputation and the missing-indicator method demonstrated the highest efficiency across all covariate missingness scenarios.
• These simple methods performed at least as well as MI and ML in all evaluated scenarios.

Conclusions:

• Simple methods for handling missing covariates in RCTs are efficient and provide unbiased treatment effect estimates.
• Mean imputation and the missing-indicator method are particularly effective and recommended alternatives to complex MI/ML approaches.
• These findings challenge the conventional recommendation of complex methods for missing data in the specific context of RCTs.