Article Text

PDF

0300 The NIEHS GuLF STUDY: A comparison of the β-substitution method and a Bayesian approach for handling highly censored measurement data
  1. Tran Huynh1,
  2. Harrison Quick2,
  3. Gurumurthy Ramachandran1,
  4. Sudipto Banerjee3,
  5. Joao Monteiro4,
  6. Caroline Groth3,
  7. Mark Stenzel5,
  8. Aaron Blair6,
  9. Dale Sandler7,8,
  10. Lawrence Engle7,8,
  11. Richard Kwok8,
  12. Patricia Stewart9
  1. 1Department of Environmental Health Sciences, University of Minnesota, Minneapolis, MN, USA
  2. 2Department of Statistics, University of Missouri, Columbia, MO, USA
  3. 3Division of Biostatistics, University of Minnesota, Minneapolis, MN, USA
  4. 4NAMSA, Minneapolis, MN, USA
  5. 5Exposure Assessment Applications, LLC, Arlington, VA, USA
  6. 6National Cancer Institute, Gaithersburg, MD, USA
  7. 7Department of Epidemiology, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA
  8. 8Epidemiology Branch, National Institute of Environmental Health Sciences, Research Triangle Park, NC, USA
  9. 9Stewart Exposure Assessments, LLC, Arlington, VA, USA

Abstract

Objectives Over 150 000 measurements taken on workers responding to the 2010 Deepwater Horizon oil spill are being used to develop exposure estimates for the participants in the GuLF STUDY. A large portion of the measurements, however, has values below the limit of detection (left-censored). The β-substitution method has been shown to provide accurate estimates for handling censored data, but a comparison to a Bayesian method, which permits the estimation of uncertainty and accounts for prior information, is currently lacking. The goal of this research was to compare the two methods.

Method Each method was challenged with computer-generated datasets drawn from lognormal distributions with the geometric mean (GM) = 1, sample sizes = 5–100, geometric standard deviation (GSD) = 2–5, and percent censoring = 10–90%. Percent bias and coverage (the percentage of 95% uncertainty intervals containing the truth) were used as evaluation metrics.

Results For most of our simulation scenarios, estimates of bias from the β-substitution and Bayesian methods were generally comparable for the AM and GM. The β-substitution was generally less biassed in estimating the GSD and the 95th percentile than the Bayesian method. The Bayesian method provided consistently better coverage for the AM than β-substitution. It also provided uncertainty estimates the GM, GSD, and the 95th percentile while β-substitution does not.

Conclusions The β-substitution method generally was observed to have little bias but it only allows the calculation of uncertainty estimates around the AM. The Bayesian approach provided reasonably accurate point and interval estimates (i.e., coverage), but this comes with the cost of additional computation.

Statistics from Altmetric.com

Request permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.