Estimating exposure distributions: A caution for Monte Carlo risk assessment

Abstract

Monte Carlo studies of risk assessment commonly require estimates of exposure distributions. The exposure distribution may be estimated by assuming the distribution follows a specified functional form (lognormal), and estimating parameters of the assumed distribution from observed sample exposure data. Alternatively, to avoid the distributional assumption, the exposure distribution may be estimated directly from the observed exposures measured on a sample of subjects. We discuss problems with this second approach for estimating exposure distributions when exposures are measured with error.

Specifically, we show that when the true exposure varies from day to day, or the observed exposure differs from the true exposure due to measurement error, then the tails of the observed exposure distributions will be biased, with the magnitude of the bias increasing toward the tails of the distribution. The bias may be severe, and lead to overestimation of upper percentile exposure. The size of the bias is directly related to the magnitude of response error. Alternative estimators are discussed that frequently provide closer estimates of a subject's true exposure. Issues regarding choice of estimator, and consequence for exposure distribution estimation are discussed in the context of estimating soil ingestion in children. The biases are illustrated via simulations.

Key words:

• Monte Carlo analysis
• exposure assessment
• soil ingestion
• estimation
• risk assessment

Notes

404 Arnold House, Department of Biostatistics and Epidemiology, School of Public Health, University of Massachusetts at Amherst, Amherst, MA 01003; Tel: [phone]; Fax: [phone]; E‐mail: [email protected]