Bayesian estimation of the probability of asbestos exposure from lung fiber counts

Our plain-English summary of Bayesian estimation of the probability of asbestos exposure from lung fiber counts, by S. Weichenthal, L. Joseph, P. Bélisle, et al (Biometrics, 2010).

The study

• Authors: S. Weichenthal, L. Joseph, P. Bélisle, A. Dufresne
• Journal: Biometrics
• Published: 2010;66(2):603-12
• Cited by: 5 articles, as counted by NIH iCite
• Read it: PubMed · Publisher

Abstract

Asbestos exposure is a well-known risk factor for various lung diseases, and when they occur, workmen's compensation boards need to make decisions concerning the probability the cause is work related. In the absence of a definitive work history, measures of short and long asbestos fibers as well as counts of asbestos bodies in the lung can be used as diagnostic tests for asbestos exposure. Typically, data from one or more lung samples are available to estimate the probability of asbestos exposure, often by comparing the values with those from a reference nonexposed population. As there is no gold standard measure, we explore a variety of latent class models that take into account the mixed discrete/continuous nature of the data, that each subject may provide data from more than one lung sample, and that the within-subject results across different samples may be correlated. Our methods can be useful to compensation boards in providing individual level probabilities of exposure based on available data, to researchers who are studying the test properties for the various measures used in this area, and more generally, to other test situations with similar data structure.

In plain English

This statistics study built Bayesian models estimating each person's probability of asbestos exposure from lung counts without a gold-standard test. The authors applied them to lung analyses from 1996 to 2000 of 78 Quebec workers who had died of lung diseases potentially caused by occupational asbestos exposure; each had three or four tissue blocks. Controls were 41 men who died of accidents or acute myocardial infarction. The tests were counts of short fibers (under 5 micrometers), long fibers (5 micrometers or longer) and asbestos bodies. With one block, the area under the ROC curve (how well a test separates exposed from unexposed people) was 0.92 (95% credible interval 0.76 to 0.98) for long fibers, 0.86 (0.74 to 0.96) for short fibers and 0.88 (0.76 to 0.97) for asbestos bodies. Counts from different blocks in one person were correlated, and models allowing for this gave substantially different results. The authors concluded that analyzing more than one block improves estimates, but for some subjects even 12 data points were not sufficient. They added there is no substitute for a detailed work history.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Statistical methods study: Bayesian latent class models (no test treated as a gold standard) developed and applied to existing lung fiber count data
• Who was studied: 78 Quebec workers (75 men) who had died of lung diseases potentially caused by occupational asbestos exposure, with lung analyses done between 1996 and 2000. 35 had three tissue blocks and 43 had four. Also 41 control men autopsied between 1990 and 1992 after accidental death or acute myocardial infarction, each contributing one block.
• Exposure: Lung counts of short asbestos fibers (less than 5 micrometers), long asbestos fibers (5 micrometers or more) and asbestos bodies, used as diagnostic tests for past asbestos exposure. Results were not reported by fiber type (e.g., chrysotile or amphibole).
• Main results:
  • For long fibers, the model estimated that almost 60% of unexposed subjects had values at the lower limit of detection, compared with under 4% (0.034) of exposed subjects.
  • Using data from block 1 only, the area under the ROC curve for long fibers was 0.92 (95% credible interval 0.76 to 0.98).
  • Using data from block 1 only, the area under the ROC curve for short fibers was 0.86 (95% credible interval 0.74 to 0.96). The authors noted that the credible intervals were very wide.
  • Using data from block 1 only, the area under the ROC curve for asbestos bodies was 0.88 (95% credible interval 0.76 to 0.97).
  • Counts from different tissue blocks within the same subject were correlated (above 0.3) for all three types of counts.
  • The two models that allowed for within-subject correlation gave very similar results to each other, but results differed substantially from models assuming independence.
  • In one example subject with high values on all three tests, a single observation from any test was enough to classify the subject as exposed.
  • In a third example subject with moderate values, results ranged widely depending on the model and the data used. The authors said this showed the danger of relying on one type of fiber count.
  • For short and long fibers, the precision of the estimated difference between exposed and unexposed means improved from one to two blocks but not much after that. For asbestos bodies, it improved little as blocks were added.
• The authors' conclusion: The authors concluded that their models give direct estimates of the probability of exposure and that collecting more than one tissue block per subject improves these estimates. They also concluded that for some subjects even 12 data points are not sufficient for a definitive assessment, and that there is no substitute for a detailed work history.
• Limitations the authors state:
  • The authors did not interpret the estimated prevalence of exposure because the sample was specially selected.
  • ROC curves, and so the area under the curve, could only be estimated from one tissue block.
  • Only a limited number of control subjects was available.
  • The authors stated that lung fiber counts are no substitute for a detailed work history, particularly for chrysotile, which does not accumulate as readily in the lungs.

Topics

Medical terms in this study, explained in our Lexicon: Asbestos Exposure, Work History, Asbestos Bodies.

Related studies

• A Bayesian approach to assess interaction between known risk factors: the risk of lung cancer from exposure to asbestos and smoking — D. Wraith, K. Mengersen (2008)
• Bayesian analysis of the rate of spontaneous malignant mesothelioma among BAP1 mutant mice in the absence of asbestos exposure — D.M. Nielsen, M. Hsu, et al. (2025)
• Examining the association of lung cancer and highly correlated fibre size-specific asbestos exposures with a hierarchical Bayesian model — G.B. Hamra, D. Loomis, et al. (2014)
• A Bayesian approach to strengthen inference for case-control studies with multiple error-prone exposure assessments — J. Zhang, S.R. Cole, et al. (2013)
• A Bayesian Model and Stochastic Exposure (Dose) Estimation for Relative Exposure Risk Comparison Involving Asbestos-Containing Dropped Ceiling Panel Installation and Maintenance Tasks — F.W. Boelter, Y. Xia, et al. (2017)
• The analysis of asbestos count data with "nondetects": the example of asbestos fiber concentrations in the lungs of brake workers — M.M. Finkelstein (2013)

Browse return to medical research.