Accommodating detection limits of multiple exposures in environmental mixture analyses: an overview of statistical approaches

Myeonggyun Lee; Abhisek Saha; Rajeshwari Sundaram; Paul S Albert; Shanshan Zhao

doi:10.1186/s12940-024-01088-w

Accommodating detection limits of multiple exposures in environmental mixture analyses: an overview of statistical approaches

Environ Health. 2024 May 16;23(1):48. doi: 10.1186/s12940-024-01088-w.

Authors

Myeonggyun Lee¹, Abhisek Saha², Rajeshwari Sundaram², Paul S Albert³, Shanshan Zhao⁴

Affiliations

¹ Biostatistics and Computational Biology Branch, National Institute of Environmental Health Sciences, National Institutes of Health, Durham, NC, USA.
² Biostatistics and Bioinformatics Branch, Division of Population Health Research, Eunice Kennedy Shriver National Institute of Child Health and Human Development, National Institutes of Health, Bethesda, MD, USA.
³ Biostatistics Branch, Division Cancer Epidemiology and GeneticsBiostatistics Branch, Division of Cancer Epidemiology and Genetics, National Cancer Institute, National Institutes of Health, Bethesda, MD, USA.
⁴ Biostatistics and Computational Biology Branch, National Institute of Environmental Health Sciences, National Institutes of Health, Durham, NC, USA. shanshan.zhao@nih.gov.

Abstract

Background: Identifying the impact of environmental mixtures on human health is an important topic. However, such studies face challenges when exposure measurements lie below limit of detection (LOD). While various approaches for accommodating a single exposure subject to LOD have been used, their impact on mixture analysis has not been thoroughly investigated. Our study aims to understand the impact of five popular LOD accommodation approaches on mixture analysis results with multiple exposures subject to LOD, including omitting subjects with any exposures below LOD (complete case analysis); single imputations by LOD/ $\sqrt{2}$ , and by estimates from a censored accelerated failure time (AFT) model; and multiple imputation (MI) with or without truncation based on LOD.

Methods: In extensive simulation studies with high-dimensional and highly correlated exposures and a continuous health outcome, we examined the performance of each LOD approach on three mixture analysis methods: elastic net regression, weighted quantile sum regression (WQS) and Bayesian kernel machine regression (BKMR). We further analyzed data from the National Health and Nutrition Examination Survey (NHANES) on how persistent organic pollutants (POPs) influenced leukocyte telomere length (LTL).

Results: Complete case analysis was inefficient and could result in severe bias for some mixture methods. Imputation by LOD/ $\sqrt{2}$ showed unstable performance across mixture methods. Conventional MI was associated with consistent mild biases, which can be reduced by using a truncated distribution for imputation. Estimating censored values by AFT models had a minimal impact on the results. In the NHANES analysis, imputation by LOD/ $\sqrt{2}$ , truncated MI and censored AFT models performed similarly, with a positive overall effect of POPs on LTL while PCB126, PCB169 and furan 2,3,4,7,8-pncdf being the most important exposures.

Conclusions: Our study favored using truncated MI and censored AFT models to accommodate values below LOD for the stability of downstream mixture analysis.

Keywords: Detection limits; Environmental mixtures; Imputation; Left-censored exposures.

MeSH terms

Environmental Exposure* / analysis
Environmental Monitoring / methods
Environmental Pollutants* / analysis
Humans
Limit of Detection
Models, Statistical
Nutrition Surveys

Grants and funding

ZIA ES103307/ES/NIEHS NIH HHS/United States