Model averaging quantiles from data censored by a limit of detection

Area of Science:

• Environmental Chemistry
• Statistical Modeling
• Risk Assessment

Background:

• Accurate chemical risk assessment requires determining concentration data quantiles.
• Left-censored data (observations below the limit of detection) hinder distribution selection and quantile estimation.
• While log-normal distribution is common, alternative models are necessary.

Purpose of the Study:

• To estimate quantiles, particularly in the left tail, from left-censored concentration data.
• To compare model-averaged quantile estimators for various log-normal related and seminonparametric distributions.
• To extend methods for covariate inclusion and limit of detection uncertainty.

Main Methods:

• Development and comparison of two model-averaged quantile estimators.
• Simulation studies to evaluate model selection and averaging techniques.
• Application to cadmium concentration data and cesium measurements with covariates.

Main Results:

• Model-averaged estimators demonstrate good performance when the true distribution is unknown.
• No significant performance difference observed between direct and indirect estimation methods.
• Accurate estimation of quantiles with high censoring percentages requires the true or a close approximating model.

Conclusions:

• Model-averaged approaches are robust for quantile estimation in chemical risk assessment with censored data.
• The methods are adaptable to include covariates and address uncertainty in detection limits.
• Effective handling of highly censored data relies on accurate model specification.