On inference of control-based imputation for analysis of repeated binary outcomes with missing data

Area of Science:

• Biostatistics
• Clinical Trials
• Longitudinal Data Analysis

Background:

• Missing data frequently occur in longitudinal clinical trials, impacting treatment effect assessment.
• Control-based imputation is a recent method assuming discontinued patients mirror control group responses.
• Standard variance estimation methods like Rubin's formula may be biased under control-based imputation.

Purpose of the Study:

• To evaluate statistical methods for accurate variance estimation with control-based imputation.
• To address potential bias in statistical inferences from longitudinal binary outcomes with missing data.
• To compare the performance of different variance estimation techniques in simulation studies.

Main Methods:

• Evaluation of statistical methods for variance estimation under control-based imputation.
• Analysis of repeated binary outcomes with monotone missing data.
• Simulation studies to assess method performance under various settings.
• Application of methods to an antidepressant Phase III clinical trial.

Main Results:

• Both the analytical method by Robins & Wang and the nonparametric bootstrap method yielded more appropriate variance estimates.
• These methods demonstrated improved accuracy across various simulation settings compared to standard approaches.
• The study identified preferred methods for variance estimation in this context.

Conclusions:

• The Robins & Wang analytical method and the nonparametric bootstrap method are recommended for variance estimation under control-based imputation.
• Accurate variance estimation is vital for reliable statistical inferences in longitudinal clinical trials with missing data.
• The findings provide practical guidance for sponsors and regulatory agencies in analyzing such trials.