Missing not at random models for masked clinical trials with dropouts.

Area of Science:

• Biostatistics
• Clinical Trials
• Data Science

Background:

• Missing data is a pervasive challenge in clinical trials.
• Standard methods often rely on the missing at random (MAR) assumption, which is frequently questionable.
• The underlying causes of missing data are often unknown and untestable from observed data.

Purpose of the Study:

• To propose a novel missing data assumption, masked missing not at random (MNAR), for masked clinical trials.
• To develop statistical models for categorical and continuous outcomes under the proposed MNAR assumption.
• To evaluate the performance of the proposed methods through simulations and comparison with existing approaches.

Main Methods:

• Formulation of statistical models for masked clinical trials under a specific masked missing not at random (MNAR) assumption.
• Conducting simulation studies to assess the finite sample performance of the MNAR methods.
• Comparison of MNAR methods against complete case analysis and methods assuming missing at random (MAR).

Main Results:

• Maximum likelihood methods using the MNAR assumption outperformed complete case analysis and MAR methods when MNAR was true.
• MNAR methods showed comparable efficiency to MAR methods when both assumptions were met.
• Analysis of the TRial Of Preventing HYpertension (TOPH) study demonstrated robustness of MAR estimates to MNAR deviations.

Conclusions:

• Methods based on the MNAR assumption are valuable for masked clinical trials, serving as primary analysis or sensitivity analysis.
• MAR analysis may be preferred for efficiency if MNAR and MAR estimates are similar.
• MNAR estimates may be preferred if substantially different, due to the plausibility of the underlying mechanism.