Identifying treatment effects using trimmed means when data are missing not at random.

Area of Science:

• Biostatistics
• Clinical Trials
• Data Analysis

Background:

• Patient discontinuation in clinical trials can bias outcome results.
• Missing data that cannot be explained by observed data are termed missing not at random (MNAR).

Purpose of the Study:

• To derive conditions for the trimmed means approach to identify average population treatment effects.
• To evaluate the trimmed means approach for handling MNAR data in clinical trials.

Main Methods:

• The trimmed means approach sets missing values as the worst observed outcome and trims data before efficacy calculation.
• Simulation studies were conducted to assess the method's performance.
• The approach was compared to existing methods using chronic pain trial data.

Main Results:

• The trimmed means approach effectively estimates treatment efficacy for MNAR data strongly associated with unfavorable outcomes.
• The method fails when data are missing at random.
• Combining multiple imputation with trimmed means can improve estimates when discontinuation reasons are known.

Conclusions:

• The trimmed means approach is a valuable tool for identifying treatment effects in the presence of MNAR data.
• Justifiable assumptions are crucial for the reliable application of this methodology.
• An R package, trim, is available to implement the trimmed means method.