A review and critique of U.S. EPA's risk assessments for asbestos

Abstract

U.S. Environmental Protection Agency (EPA) recently conducted a risk assessment for exposure to Libby amphibole asbestos that is precedent-setting for two reasons. First, the Agency has not previously conducted a risk assessment for a specific type of asbestos fiber. Second, the risk assessment includes not only an inhalation unit risk (IUR) for the cancer endpoints, but also a reference concentration (RfC) for nonmalignant disease. In this paper, we review the procedures used by the Agency for both cancer and nonmalignant disease and discuss the strengths and limitations of these procedures. The estimate of the RfC uses the benchmark dose method applied to pleural plaques in a small subcohort of vermiculite workers in Marysville, Ohio. We show that these data are too sparse to inform the exposure–response relationship in the low-exposure region critical for estimation of an RfC, and that different models with very different exposure–response shapes fit the data equally well. Furthermore, pleural plaques do not represent a disease condition and do not appear to meet the EPA's definition of an adverse condition. The estimation of the IUR for cancer is based on a subcohort of Libby miners, discarding the vast majority of lung cancers and mesotheliomas in the entire cohort and ignoring important time-related factors in exposure and risk, including effect modification by age. We propose that an IUR based on an endpoint that combines lung cancer, mesothelioma, and nonmalignant respiratory disease (NMRD) in this cohort would protect against both malignant and nonmalignant disease. However, the IUR should be based on the entire cohort of Libby miners, and the analysis should properly account for temporal factors. We illustrate our discussion with our own independent analyses of the data used by the Agency.

Keywords:

• amphibole asbestos
• benchmark dose
• inhalation unit risk
• Libby
• localized pleural thickening
• pleural plaques
• reference concentration
• risk assessment
• United States Environmental Protection Agency

Declaration of interest

This work was financially supported by W. R. Grace & Company. Some of the authors (SHM, ELA, and DGH) are consultants to W. R. Grace & Company. W. R. Grace & Company was not involved in the preparation and approval of the manuscript. The findings and conclusions in this manuscript are those of the authors and do not necessarily represent the views of W. R. Grace & Company. The authors are employees of Exponent, Inc., an international consulting company. SHM, ELA and DGH have served as experts in asbestos litigation.

Notes

¹The IUR for asbestos-associated cancer is based on a relative risk model for lung cancer and an absolute risk model for mesothelioma. The model for mesothelioma was originally developed by Julian Peto et al. (1982) based on the multistage model of Armitage and Doll (1954). In the Peto model, the incidence of mesothelioma is a linear function of the intensity (concentration) of exposure and a power function of duration of exposure and of time since exposure ceased. As a consequence, the incidence of mesothelioma cannot be expressed in terms of cumulative exposure. The Peto model explicitly acknowledges the distinct contributions made by intensity and time-related factors.

²EPA updated exposure estimates for the draft risk assessment. Our analyses of both the original Rohs cohort and the restricted cohort (see below) used these updated exposure estimates.

³Here we depart slightly from the notation used in the draft risk assessment. Because the measure of exposure used by the Agency is cumulative exposure, we use BMCE to denote the benchmark cumulative exposure, and BMCEL to denote the lower 95% confidence bound on the benchmark cumulative exposure.

⁴The models allow background rates to be estimated from the data.

⁵The Agency describes these models in the documentation for the BMDS package, which can be downloaded from the Agency website. The package allows fitting of the various models used in the draft risk assessment, with the exception of the Michaelis–Menten model. However, the package does not allow the simultaneous consideration of covariates (i.e., it cannot address confounders, an important limitation for analyses of human data). The BMDL is calculated using the profile-likelihood method for confidence intervals (Crump 1984, Venzon and Moolgavkar 1988).

⁶Lags are often used in analyses of human data on chronic disease to account for latency between exposure and the occurrence of disease. The idea is that any exposure that occurred close to the diagnosis of a chronic disease could not have contributed to its occurrence. With incidence data (i.e., data on the time of occurrence of a condition), the interpretation of a lag is clear. However, with prevalence data (the kind of data available in the Rohs subcohort; i.e., information on the presence or absence of pleural plaques at a single point in time), the interpretation of a lag is problematic, because one does not know when the pleural plaques actually occurred.

⁷The AIC is commonly used to compare the fits of different models to a data set. The model with the lowest AIC is preferred.

⁸The Michaelis–Menten model has three free parameters (i.e., three parameters can be estimated from the data). The Agency chose to fix the background prevalence of pleural plaques at 1% and estimated only two of the three parameters. The Science Advisory Board panel that reviewed the draft risk assessment recommended that the Agency use the dichotomous Hill model, which is a four-parameter model (Science Advisory Board [SAB], August 30, 2012). The Scientific Advisory Board panel suggested that two of the parameters be fixed: the background prevalence at 1%, and the maximum prevalence (plateau) at 85%.

⁹Christensen and Kopylev (2012) analyzed this data set and estimated a BMCE of 0.36 f/cc-y when cigarette smoking was included as a covariate. They did not report a BMCEL in their paper.

¹⁰The Agency repeats the old canard (page 5–78 of the report) that non-differential covariate measurement errors cause risk estimates to be biased toward the null. This statement, though widely repeated by epidemiologists, is incorrect. First, not only must the misclassification be non-differential, it must satisfy other conditions (e.g., Jurek et al. 2005) for the result to hold. Second, the statement applies to the expectation of the risk estimate, not to the value of the estimate from any single study. Thus, it is possible to have non-differential misclassification that satisfies all the required conditions, but the result of a single study may actually overestimate the risk. As Jurek et al. (2005) state, “[E]xposure misclassification can spuriously increase the observed strength of an association even when the misclassification process is non-differential and the bias it produced is towards the null.” Similar discussion is provided by Thomas (1995) and Weinberg et al. (1995).

¹¹Hein et al. (2007) report an RR of about 3 associated with 100 f/cc-y cumulative exposure as compared to an RR of about 1.11 in Libby for the same cumulative exposure.

¹²Moolgavkar et al. (2009) could not estimate the exponent, because they had information only on the number (15) of mesotheliomas in the Sullivan cohort, but not on which individuals had the disease. With this information, only KM can be estimated.