Bias and inference from misspecified mixed-effect models in stepped wedge trial analysis

Area of Science:

• Biostatistics
• Clinical Trial Design

Background:

• Stepped wedge trials (SWTs) are increasingly used in public health and medical research.
• The standard mixed-effect model for SWT analysis assumes fixed effects for intervention and time periods.
• Robustness of this standard model to misspecification is not well understood.

Purpose of the Study:

• To investigate the robustness of the standard mixed-effect model for SWT analysis to misspecification.
• To evaluate the impact of period and intervention effects varying between clusters on bias and confidence interval coverage.
• To compare the performance of the standard model with models including additional random effects.

Main Methods:

• Simulated stepped wedge trials with three clusters and two time periods.
• Varied period and intervention effects (common-to-all or cluster-specific).
• Analyzed simulated data using the standard model, a model with random period effects, and a model with random intervention effects.

Main Results:

• The standard model exhibited up to 50% bias in intervention effect estimates when period or intervention effects varied between clusters and were treated as fixed.
• All misspecified models, especially the standard model, showed undercoverage of 95% confidence intervals.
• The standard model gave significant weight to within-cluster comparisons, increasing sensitivity to assumption departures.

Conclusions:

• Mixed-effect models for SWTs are highly sensitive to departures from their assumptions, particularly due to reliance on within-cluster comparisons.
• Trialists should consider incorporating a random effect for the time period in their SWT analysis models to improve robustness.
• Accurate statistical analysis is crucial for reliable intervention effect estimation in SWTs.