Comparison of methods for analyzing left-censored occupational exposure data

Area of Science:

• Environmental Health Sciences
• Occupational Health
• Biostatistics

Background:

• The GuLF STUDY investigates health outcomes of Deepwater Horizon oil spill cleanup workers.
• Exposure assessment relies on personal monitoring data, often with values below the limit of detection (LOD).
• Accurate statistical methods are crucial for analyzing censored exposure data.

Purpose of the Study:

• To evaluate three methods (maximum likelihood, β-substitution, Kaplan-Meier) for analyzing censored exposure data.
• To determine the best method for estimating exposure distribution parameters (AM, GM, GSD, X0.95).
• To assess method performance under various conditions of sample size, variability, and censoring levels.

Main Methods:

• A simulation study using computer-generated datasets from lognormal and mixed lognormal distributions.
• Varied sample sizes (N=5-100), geometric standard deviations (GSD=2-5), and censoring levels (10-90%).
• Evaluated methods based on relative bias and relative root mean squared error (rMSE).

Main Results:

• β-substitution generally performed as well as or better than ML and K-M methods.
• ML method suitable for large sample sizes (N≥30) with up to 80% censoring in lognormal distributions.
• K-M method accurate for arithmetic mean (AM) with <50% censoring in lognormal and mixed distributions.
• All methods' accuracy decreased with high variability (GSD=4-5) and small sample sizes (N<20), but β-substitution remained superior.
• ML method requires careful interpretation due to potential for biased results.

Conclusions:

• The β-substitution method is recommended for analyzing censored exposure data, particularly in scenarios with high variability and censoring.
• Maximum likelihood is viable for large datasets with moderate censoring.
• Kaplan-Meier is effective for AM estimation with lower censoring levels.
• Further research is needed to improve estimation accuracy for small sample sizes and high censoring, and to develop uncertainty intervals for the β-substitution method.