Testing for differences in changes in the presence of censoring: parametric and non-parametric methods

Area of Science:

• Statistics
• Biostatistics
• Longitudinal Data Analysis

Background:

• Repeated measures data with incomplete observations are common in various scientific fields.
• Standard statistical methods often assume random censoring, which may not hold true.
• Informative censoring can significantly bias analysis results.

Purpose of the Study:

• To review and compare parametric and non-parametric methods for analyzing repeated measures with incomplete observations.
• To evaluate method performance under random and informative censoring scenarios.
• To identify robust analytical approaches for handling informative censoring.

Main Methods:

• Simulated experiments based on a linear random effects model.
• Comparison of commonly used parametric and non-parametric analysis methods.
• Evaluation of the conditional linear model with bootstrap variance estimation.
• Assessment of a non-parametric procedure using individual summary statistics.

Main Results:

• Standard methods suffer power loss or yield false positives with informative censoring.
• The conditional linear model with bootstrap variance performed well under both random and informative censoring.
• The non-parametric ranking procedure showed relatively good performance, though less efficient.

Conclusions:

• Informative censoring necessitates careful selection of statistical methods.
• The conditional linear model with bootstrap variance is a reliable approach for informative censoring.
• Testing for and accounting for informative censoring is crucial for valid analysis.