A structured framework for assessing sensitivity to missing data assumptions in longitudinal clinical trials.

Area of Science:

• Clinical Trials
• Biostatistics
• Psychiatry

Background:

• Longitudinal clinical trial data often contain missing values, which can bias results.
• Assessing the impact of missing data is crucial for reliable inference in clinical research.

Purpose of the Study:

• To demonstrate a framework for sensitivity analyses of incomplete longitudinal clinical trial data.
• To re-analyze data from a depression clinical trial to assess robustness to missing data assumptions.

Main Methods:

• Primary analysis used a likelihood-based approach assuming data were missing at random (MAR).
• Sensitivity analyses employed various missing not at random (MNAR) models (selection, pattern mixture, shared parameter) and inclusive MAR methods.
• A key sensitivity analysis involved multiple imputation assuming post-dropout trajectories mirrored placebo patients (placebo multiple imputation) to define a worst reasonable case.

Main Results:

• The primary analysis showed a treatment contrast of -2.79 (p = .013).
• The placebo multiple imputation (worst reasonable case) yielded a contrast of -2.17.
• Sensitivity analyses results ranged from -2.21 to -3.87, symmetrically distributed around the primary result, indicating no clear evidence of MNAR bias.

Conclusions:

• A robust treatment effect was confirmed, even under a worst reasonable case scenario where the effect was 80% of the primary result.
• The developed structured sensitivity analysis framework, using a controlled imputation approach and transparent assumptions, proved useful and shows promise for general application.