Comparing Methods for Measurement Error Detection in Serial 24-h Hormonal Data

Area of Science:

• Endocrinology
• Biostatistics
• Physiological Measurement

Background:

• Measurement errors, appearing as outliers, are common in 24-hour hormonal data.
• No universally accepted automatic method exists for detecting these outliers.
• Accurate outlier detection is vital for reliable study outcomes in hormonal research.

Purpose of the Study:

• To compare the performance of four different outlier detection methods in 24-hour serial hormonal data.
• To evaluate the impact of outlier removal on statistical outcomes.
• To identify the most suitable method for automatic outlier detection in this context.

Main Methods:

• Four outlier detection methods were compared: eyeballing, Tukey's fences, stepwise approach, and expectation-maximization (EM) algorithm.
• Hormonal data (glucose, insulin, TSH, cortisol, growth hormone) were collected every 10 minutes for 24 hours from 38 participants.
• Performance was assessed by the number of outliers detected and changes in statistical results post-removal.

Main Results:

• The EM algorithm detected the most outliers (11.0%), while eyeballing detected the fewest (1.0%).
• Outlier removal did not significantly alter mean hormone levels but affected minima.
• Individual-level glucose-insulin correlations were affected, but not the averaged correlations.

Conclusions:

• The EM algorithm is not recommended due to excessive outlier detection.
• Eyeballing is time-consuming, and Tukey's fences have data limitations.
• The stepwise approach is recommended for its balance of physiological knowledge, automation, and effectiveness.