Power Estimation in Planning Randomized Two-Arm Pre-Post Intervention Trials with Repeated Longitudinal Outcomes

Area of Science:

• Biostatistics
• Clinical Trial Design
• Longitudinal Data Analysis

Background:

• Intervention effects are typically estimated using Generalized Least Squares (GLS) models in clinical trials with repeated longitudinal measurements.
• Power estimation for these studies is challenging due to a lack of GLS-based approaches and normative data.

Purpose of the Study:

• To derive Generalized Least Squares (GLS) variance for intervention effects.
• To develop simple power formulas based on the compound symmetry correlation structure.
• To investigate optimal pre-post allocation for maximizing statistical power.

Main Methods:

• Derived GLS variance of the intervention effect.
• Developed power formulas assuming compound symmetry correlation.
• Investigated optimal pre-post allocation for variance minimization.

Main Results:

• Empirical analysis of nursing home and HIV patient data revealed Toeplitz correlation structures, differing from compound symmetry.
• Power calculations using Toeplitz structures showed that one pre-intervention measure often maximized power, contrary to compound symmetry assumptions.
• Approximating Toeplitz structures with compound symmetry led to overestimation of power.

Conclusions:

• The compound symmetry correlation structure, while simplifying power estimation, frequently does not reflect the actual correlation in repeated measures.
• Using compound symmetry to approximate unknown Toeplitz correlations can lead to significant underestimation of the variance of the intervention effect estimate.
• Accurate power estimation requires accounting for the specific correlation structure of longitudinal data, such as the decline in correlation over time.