An analytic method for the placebo-based pattern-mixture model.

Area of Science:

• Statistics
• Biostatistics
• Clinical Trial Methodology

Background:

• Pattern-mixture models are crucial for sensitivity analyses of nonignorable missing data in longitudinal studies.
• The placebo-based pattern-mixture model provides a clinically interpretable approach to handling missing data.
• Standard multiple imputation is commonly used but may yield overly conservative variance estimates.

Purpose of the Study:

• To evaluate the performance of standard multiple imputation for placebo-based pattern-mixture models.
• To develop and propose alternative inference methods for the placebo-based pattern-mixture model.
• To compare the efficiency and power of proposed methods against multiple imputation.

Main Methods:

• Derivation of an analytic expression for treatment effect in the placebo-based pattern-mixture model.
• Proposal of posterior simulation and delta methods for treatment effect inference.
• Conducting simulation studies to assess variance estimation and statistical power.

Main Results:

• Rubin's variance estimate from multiple imputation can be overly conservative.
• The proposed posterior simulation and delta methods yield consistent variance estimates.
• Simulation studies show the proposed methods outperform multiple imputation in power.

Conclusions:

• The developed analytic and simulation-based methods offer improved inference for placebo-based pattern-mixture models.
• These methods provide more accurate variance estimates and increased power compared to multiple imputation.
• The findings are illustrated using data from a major depressive disorder clinical study.