Censoring Trace-Level Environmental Data: Statistical Analysis Considerations to Limit Bias

Area of Science:

• Environmental Science
• Analytical Chemistry
• Biostatistics

Background:

• Trace-level environmental data often contain values below detection limits, posing challenges for accurate health risk assessment.
• Low-concentration exposures to single or multiple chemicals can have significant health implications.

Purpose of the Study:

• To assess bias in statistical measures (mean, standard deviation) of trace-level environmental data using various censoring approaches.
• To compare the performance of different statistical methods for handling censored data in environmental exposure assessments.

Main Methods:

• A cook stove case study analyzed dibenzo[a,h]anthracene concentrations using gas chromatography-mass spectrometry.
• Evaluated censoring methods including threshold/2 substitution, maximum likelihood estimation, robust regression, Kaplan-Meier, and data omission.
• Simulated measurement data from log-normal distributions at varying censoring levels (30%, 50%, 80%) to compare statistical approaches.

Main Results:

• Threshold/2 substitution demonstrated the least bias among simple methods but was inferior to modern after-censoring statistical approaches.
• All after-censoring methods were less accurate than analyses that included all measurement data.
• Differences in group means varied significantly based on censoring decisions and chosen statistical methods.

Conclusions:

• Censoring environmental data can introduce significant bias, impacting the accuracy of exposure assessments and health risk evaluations.
• Investigators should prioritize statistical methods that minimize bias, ideally by including all available measurement data.
• Careful consideration of censoring and distributional assumptions is crucial to avoid erroneous conclusions in environmental health research.