Multiplicity adjustments for the Dunnett procedure under heterokcedasticity

Area of Science:

• Biostatistics
• Statistical Methods
• Clinical Trial Design

Background:

• The Dunnett procedure is crucial for comparing multiple treatments against a control.
• Standard methods struggle with heteroskedasticity (unequal variances) due to complex test statistic distributions.
• Existing methods for multiplicity adjustment under heteroskedasticity can be inaccurate, leading to conservative or liberal results.

Purpose of the Study:

• To develop a simulation-based method for computing multiplicity-adjusted p-values and critical constants for the Dunnett procedure under heteroskedasticity.
• To accurately control the familywise error rate (FWER) in treatment-vs-control comparisons with unequal variances.
• To address the limitations of existing methods that do not properly account for the correlated denominators of Welch-Satterthwaite test statistics.

Main Methods:

• A novel simulation-based algorithm is proposed to approximate the joint distribution of correlated chi-square variables, representing the denominators of test statistics.
• This approximation is used to derive accurate critical constants and multiplicity-adjusted p-values.
• The proposed method's performance in controlling the FWER is evaluated through simulations under various heteroskedastic scenarios and compared against existing methods.

Main Results:

• The proposed simulation-based method demonstrates superior accuracy in controlling the familywise error rate (FWER) compared to existing approaches.
• Other methods evaluated were found to be either too conservative, too liberal, or less accurate in FWER control.
• The developed method provides a more reliable way to perform multiple treatment comparisons when variances are unequal.

Conclusions:

• The novel simulation-based method offers a robust and accurate solution for Dunnett-type comparisons under heteroskedasticity.
• This approach improves statistical power and reliability in clinical trials and other research settings with unequal group variances.
• The method's effectiveness is validated through simulations and illustrated with a real-world dataset.