Approximate Negative-Binomial Confidence Intervals: Asbestos Fiber Counts

TL;DR

This paper develops a simple approximation method for negative-binomial confidence intervals to better analyze asbestos fiber counts, accounting for sampling errors and human variability, with validation through simulation and comparison to traditional methods.

Contribution

It introduces a novel approximation technique for negative-binomial confidence limits that improves accuracy in asbestos fiber count analysis.

Findings

The approximation accurately estimates confidence intervals validated by simulation.

It provides a method to derive uncertainty in asbestos fiber concentration estimates.

Comparison shows improved performance over traditional normal-based confidence limits.

Abstract

The negative-binomial distribution is adopted for analyzing asbestos-fiber counts so as to account for both the sampling errors in capturing only a finite number of fibers as well as the inevitable human variation in identifying and counting sampled fibers. A simple approximation to this distribution is developed for the derivation of quantiles and approximate confidence limits. The success of the approximation depends critically on the use of the Stirling expansion to sufficient order, on exact normalization of the approximating distribution, on reasonable perturbation of quantities from the normal distribution, and on accurately approximating sums by inverse-trapezoidal integration. Accuracy of the approximation developed is checked through simulation and also by comparison to traditional approximate confidence intervals in the specific case that the negative-binomial distribution approaches the Poisson distribution. The resulting statistics are shown to relate directly to early research into the accuracy of asbestos sampling and analysis. Uncertainty in estimating mean asbestos-fiber concentrations given only a single count is derived. Decision limits (limits of detection (LOD)) and detection limits are considered for controlling false positive and negative detection assertions and are compared to traditional limits computed assuming normal distributions.