Origin: 70% confidence

Hello Jérôme,

In BOHS-NVvA Guidance (p. 21):
“Work by the Institut National de Recherche et de Sécurité in France (INRS, 2008), and by Jérôme Lavoué of the University of Montreal (Lavoué, to be published), has shown that the test can be applied with a modest number of samples if we require 70% confidence that less than 5% of the exposures are above the OEL. This gives a balance between probability of declaring non-compliant a distribution which would prove compliant if a large number of samples were taken, and declaring compliant a distribution which would prove non-compliant if there were a large number of samples.”

In WebExpo (p. 25-26):
“[…] For σ, we took inspiration from the model described by McNally et al. (McNally et al., 2014), in which they used the population of values observed in a dataset presented by Kromhout et al. (Kromhout et al., 1993) and Rappaport et al. (Rappaport et al., 1993). The authors described variability estimates for close to 200 exposure groups. From Table A1 in Kromhout et al. we tabulated 165 values of σ. Graphical assessment suggested a lognormal shape for the distribution of these values. Fitting the data to a lognormal distribution yields a GM of 0.84 for σ (which corresponds to a GSD of 2.32 for exposure levels), and a GSD of 1.87 (this quantity expresses variability in the sigma values, not in exposure levels). […] Seventy percent of the distribution is comprised between 1.5 and 4.5. The prior for variability in the [SEG.informedvar] model is therefore expressed as a lognormal distribution for the log-transformed GSD of the exposure distribution.”

Could you comment on the work to be published cited in the BOHS-NVvA Guidance?
Is the UTL95,70 based on the GSD distribution from the Kromhout et al.(1993) [70% between 1.5 and 4.5]?

Hello Marcus,

The two discussions are unrelated, despite the repeated use of “70%”.

In the first case, the BOHS_NVVA document discusses the usefulness of requiring 70% confidence instead of 95% for decision on compliance. I hadn’t checked the text closely and it is unfortunate that the document cites a “future” paper from me : such a paper was never planned. What I did for the committee with some help from a student, and was partially reproduced in a paper by Trevor Ogden, president of committee (Testing compliance with occupational exposure limits: development of the British-Dutch guidance - PubMed), was to perform computer simulation to show the impact of choosing various confidence levels on the compliance decision . This is an unpublished report to the committee which I shared online. The first researchers studying this problem were from France at INRS, but their work is only available in French. Their work was the basis for the use of 70% confidence in the French regulation. In a way I reproduced their theoretical results using computer simulation.

The second bit of text you mention is simply a description of the GSD values in the paper by Kromhout et al. : if one fit a lognormal distribution to the log(GSD) values from that paper, the resulting distribution has 70% of it’s values between 1.5 and 4.5. This 70% is unrelated to the one above.

I hope this clarifies a bit




Thank you very much for the clarifications!

Best regards.


1 Like