On the inappropriateness of an EM algorithm based procedure for blinded sample size re-estimation

Area of Science:

• Clinical Trial Design
• Biostatistics
• Statistical Methodology

Background:

• Clinical trial sample size calculations rely on a priori variance estimates.
• Inaccurate variance estimates can compromise trial power.
• Updating variance estimates during a trial is desirable for accurate sample size re-estimation.

Purpose of the Study:

• To critically evaluate an EM algorithm-based procedure for blinded variance estimation in clinical trials.
• To assess the reliability and limitations of blinded variance estimation for sample size re-estimation.

Main Methods:

• Analysis of an existing EM algorithm for blinded variance estimation.
• Simulation studies to investigate estimator properties.
• Evaluation of maximum likelihood estimation for blinded sample size re-estimation.
• Illustration using a clinical trial in asthma.

Main Results:

• The EM algorithm's estimates are sensitive to initialization.
• The stopping rule does not guarantee convergence to the maximum likelihood estimator.
• The procedure is limited to simple randomization, rarely used in clinical trials.
• Maximum likelihood estimation shows bias and high variability for blinded sample size re-estimation.

Conclusions:

• The proposed EM algorithm procedure has significant limitations for blinded variance estimation.
• Maximum likelihood estimation is unsuitable for blinded sample size re-estimation due to poor performance.
• These findings question the utility of current blinded variance estimation methods in clinical trial planning.