Confidence intervals construction for difference of two means with incomplete correlated data

Area of Science:

• Statistics
• Biostatistics
• Clinical Trials

Background:

• Incomplete data are common in clinical trials, particularly for continuous correlated outcomes.
• Existing methods for confidence intervals (CI) primarily address binary data, leaving a gap for continuous data.

Purpose of the Study:

• To develop and evaluate methods for constructing CIs for the difference of two means with incomplete continuous correlated data.
• To address the limitations of existing statistical approaches in handling such data.

Main Methods:

• Proposed methods include large sample, hybrid, and Bootstrap-resampling approaches (B 1-B 4).
• Bootstrap methods were based on maximum likelihood estimates and Ekbohm's unbiased estimator.
• Performance was assessed via simulation studies evaluating coverage probability, interval width, and non-coverage probabilities.

Main Results:

• Bootstrap-resampling-based CIs (B 1, B 2, B 4) demonstrated satisfactory performance for small to moderate sample sizes.
• These methods maintained coverage probabilities close to nominal levels.
• The ratio of mesial to distal non-coverage probabilities was well-controlled within [0.4, 0.6].

Conclusions:

• Bootstrap-resampling methods provide reliable confidence intervals for incomplete continuous correlated data.
• The Bootstrap-resampling method B 1 is recommended for achieving the shortest interval width while maintaining desired coverage properties.