An Evaluation of Statistical Methods for Analyzing Follow-Up Gaussian Laboratory Data with a Lower Quantification...

Area of Science:

• Biostatistics
• Clinical Trial Design
• Epidemiology

Background:

• Laboratory data often have values below the quantification limit (censored data).
• Improper handling of censored data can lead to biased estimates and underestimated variance.
• A vaginal microbicide study in HIV-infected women motivated this research.

Purpose of the Study:

• To evaluate statistical methods for analyzing censored Gaussian data in clinical trials.
• To compare the performance of nonparametric, mixed, mixture, and dichotomous methods.
• To assess methods for a two-treatment parallel-arm or crossover study design.

Main Methods:

• Applied four analysis methods: nonparametric, mixed models, mixture models, and dichotomous measures.
• Utilized simulation to evaluate statistical properties under different variance assumptions.
• Considered scenarios with equal or inflated variance for follow-up data.

Main Results:

• Mixed models demonstrated correct Type I error rates and superior statistical power.
• Mixed models yielded the least biased treatment-effect estimates with appropriate confidence interval coverage.
• Crossover study analysis, including period effects, significantly increased required sample size.
• Existing sample-size estimation methods were found inadequate for achieving desired power.

Conclusions:

• Mixed models are recommended for analyzing censored Gaussian data in clinical trials due to their robust performance.
• Standard sample-size estimation methods require revision for studies with censored data and crossover designs.
• Accurate analysis of censored data is crucial for reliable treatment effect estimation in HIV research and beyond.