Noninferiority testing with censoring when the event rate is low

Area of Science:

• Clinical Trials
• Biostatistics
• Infectious Diseases

Background:

• Drug-susceptible tuberculosis treatment trials often have low event rates (3-5% relapse).
• Noninferiority testing in low-event settings requires robust statistical methods.
• The PREDICT TB trial evaluates an abbreviated tuberculosis treatment regimen.

Purpose of the Study:

• To determine the optimal statistical method for testing noninferiority in tuberculosis trials with low event rates.
• To compare the operating characteristics of five different statistical tests.

Main Methods:

• Conducted simulations and theoretical analyses of five noninferiority tests.
• Included normal approximation, exact, and simulation-based tests.
• Utilized Kaplan-Meier estimators to account for variable follow-up and loss to follow-up.

Main Results:

• Normal approximation tests showed good performance with minor type I error inflation.
• Kaplan-Meier methods demonstrated higher statistical power compared to other tests.
• The advantage of Kaplan-Meier methods was most pronounced with 10-30% loss to follow-up.

Conclusions:

• Kaplan-Meier based estimators are recommended for noninferiority testing in tuberculosis trials with low event rates and significant loss to follow-up.
• Statistical methods must carefully account for patient follow-up duration and loss.
• Further evaluation of statistical approaches in low-event clinical trials is warranted.