Generating single-unit residue concentration distributions based on maximum likelihood estimation from composite data

• Published in: EnvironMetrics (London, Ont.) Vol. 13; no. 5-6; pp. 711 - 724
• Main Authors: Valdez-Flores, Ciriaco, Holden, Larry R., Sielken Jr, Robert L.
• Format: Journal Article Conference Proceeding
• Language: English
• Published: Chichester, UK John Wiley & Sons, Ltd 01-08-2002; Wiley
• Subjects: acute dietary risk assessment; Biological and medical sciences; composite data; Food industries; Food toxicology; Fundamental and applied biological sciences. Psychology; maximum likelihood estimation; Monte Carlo simulation; residue concentrations; Monte Carlo method; Statistical analysis; Residue; Distribution; Pesticides; Frequency; Maximum likelihood; Risk analysis; Unit; Food; Chemical contamination
• ISSN: 1180-4009; 1099-095X
• DOI: 10.1002/env.541

The USDA and FDA databases on pesticide residues in foods contain the residue concentrations in composites (e.g. 5 pounds of apples) as opposed to the value in a single unit (e.g. an apple). However, for assessments of acute dietary intake of pesticides it is more helpful to know the residue concentrations in units that more closely correspond to what is eaten on a single occasion. Rather than assuming that all of the pesticide detected in a composite is concentrated in one single unit or making some other worst‐case assumption, a new statistical procedure based on maximum likelihood estimation and Monte Carlo simulation has been developed to impute the frequency distribution of the single‐unit residue concentration of a chemical from composite residue concentration data. This maximum likelihood imputation procedure (or MaxLIP) incorporates censored observations (e.g. below different limits of detection), mixtures of single‐unit residue concentration distributions, correlations of single‐unit values within a composite, different numbers of units in different composites, mixtures of treated and untreated single units in the same composite, and mixtures of treated and untreated composites in the database.
In several examples with simulated data and in a validation field study, MaxLIP predicts the single‐unit residue concentration distribution well and is quite robust.
MaxLIP provides a practical example of a potentially widely applicable approach in which Monte Carlo simulation makes maximum likelihood estimation possible when the probability density and cumulative distribution functions in the likelihood are not analytically tractable. Copyright © 2002 John Wiley & Sons, Ltd.