3
Dose-Response

After hazard identification, a subset of studies is selected for dose-response assessment. Thus, the next step in the Texas Commission on Environmental Quality’s (TCEQ’s) risk assessment process was a dose-response analysis for outcomes identified in the preceding hazard identification step of its 2020 Ethylene Oxide Carcinogenic Dose-Response Assessment: Development Support Document (hereafter referred to as the TCEQ DSD). TCEQ identified lymphoid cancer as an outcome of interest for dose-response modeling. However, TCEQ’s approach to hazard evaluation may have resulted in faulty conclusions regarding breast cancer. In Chapter 2, the committee recommended applying systematic approaches to develop hazard determination for these outcomes. If implementing the committee’s Tier 1 recommendations presented in Chapter 2 results in changes to TCEQ’s hazard conclusions, then all relevant endpoints found to be related to ethylene oxide exposure should be brought into the dose-response assessment.

This chapter addresses the committee’s second charge, regarding dose-response assessment for ethylene oxide, specifically including consideration for:

• the appropriateness of the selected cancer endpoint(s), key data set for modeling, dose-response model, shape of the dose-response curve, and model fit criteria, in the context of relevant mode of action information;
• the model accuracy and validation analyses for lymphoid cancer and the conclusions based on those analyses, as well as the additional analyses that assumed a healthy worker effect for lymphoid cancer mortality; and
• any implications of the endogenous production of ethylene oxide, including from the perspective of biological significance, for risk-based air concentrations.

To address this second charge, the following topics are discussed: consideration of mortality and incidence data, reliance on unpublished studies, the use of two cohorts evaluated by the National Institute for Occupational Safety and Health (NIOSH) and the Union Carbide Corporation (UCC), age-dependent adjustment factors (ADAFs), the influence of mode of action (MOA) on model

selection, guidance of model selection, and systematic review quality considerations. Findings and recommendations are then provided.

CONSIDERATION OF MORTALITY AND INCIDENCE DATA

Mortality rates are a poor reflection of incidence rates for most types of lymphoid cancer (Chu et al. 2023; Huang et al. 2022, 2024); thus, modeling dose-response for cancer incidence is generally preferred over mortality (U.S. EPA 2005a). In situations where sufficient epidemiologic data on cancer incidence do not exist, estimation of a unit risk factor (URF) for incidence is possible through modeling mortality data and applying life-table analysis to scale the dose-response function to cancer incidence, as demonstrated in the Integrated Risk Information System (IRIS; U.S. EPA 2016a). Because lymphoid cancer mortality rates are far lower than the incidence rates, a URF for mortality will be lower than the URF for incidence. For example, in the U.S. Environmental Protection Agency’s (EPA’s) assessment of ethylene oxide for the IRIS program, the estimated URF for lymphoid cancer incidence was more than double that for mortality (U.S. EPA 2016a). TCEQ used epidemiologic data on lymphoid cell cancer mortality to model dose-response relationships for URF derivation. Environmental standards based on mortality will not fully protect against cancer incidence, nor does this approach address the core mission of public health to prevent disease and reduce morbidity. Thus, TCEQ should use cancer incidence as the endpoint for its dose-response model.

Total Lymphoid Cancers Versus Subtypes

The grouping of lymphoid cancers used in TCEQ’s dose-response modeling included non-Hodgkin lymphoma (NHL), multiple myeloma, and lymphocytic leukemia (as in Steenland et al. 2004). This type of grouping is not uncommon in epidemiologic analyses because malignancies deriving from the same type of cell may share common etiologies. For example, immune suppression appears to be a common etiology for multiple types of lymphoid cell cancer, as evidenced by increased cancer rates seen in cohorts of HIV/AIDS and organ transplantation patients (Hernández-Ramírez et al. 2017; Krynitz et al. 2013). However, many lifestyle and environmental risk factors for lymphoid cancers have been found to differ among the subtypes (Morton et al. 2014; Thun et al. 2017). Potentially differing associations with ethylene oxide exposure among lymphoid cancer subtypes have implications for the modeled dose-response and derivation of a URF. Heterogeneity across subtypes in the strength (and, possibly, the direction) of the relationship will lead to an overall association that tends to mask associations with more specific lymphoid cancer subtypes in the group. The published data suggest that such heterogeneity exists. For example, estimated standardized mortality ratios (SMRs) among males in the combined NIOSH and UCC cohorts (Valdez-Flores et al. 2010) indicate a stronger relationship with NHL (SMR = 1.18, 95% confidence interval [CI] [0.80, 1.68]) than for all hematopoietic cancer

(SMR = 1.00, 95% CI [0.78, 1.27]). The SMR for Hodgkin lymphoma was also elevated among men (SMR = 1.13, 95% CI [0.37, 2.63]) (Valdez-Flores et al. 2010), but this less common type of lymphoid cell cancer was not included in TCEQ’s grouping of lymphoid cancers for derivation of a URF. Furthermore, slopes for the relationship between cumulative ethylene oxide exposure and cancer mortality were positive (i.e., increasing mortality with exposure) for NHL and lymphocytic leukemia but were negative for multiple myeloma (Valdez-Flores et al. 2010). Such differences between subtypes were also suggested in the analysis by Steenland and colleagues (2004) of 10-year lagged cumulative ethylene oxide exposure among males in the NIOSH cohort, in which elevated SMRs were estimated in association with the highest levels of exposure for NHL and Hodgkin lymphoma but not for multiple myeloma. When possible, derivation of separate URFs for incident lymphoid cell subtypes, which have differing relationships with ethylene oxide, and/or exclusion of subtypes which show no association, may result in URFs that are more reflective of relevant risks.

TCEQ’S DOSE-RESPONSE ASSESSMENT FOR LYMPHOID CANCERS

TCEQ concluded that there was epidemiological evidence, albeit inconsistent, for associations between ethylene oxide exposure and lymphoid cancer in exposed workers, and those data were used to derive a URF.¹ The following section provides an overview of the epidemiologic studies performed by NIOSH and UCC. TCEQ relied on the findings of these studies for the dose-response assessment for lymphoid cancers.

NIOSH

NIOSH assembled a cohort of 18,235 workers exposed to ethylene oxide. The cohort averaged 26.8 years of follow-up in the extended follow-up study through 1998, and 2,852 members of the cohort (16%) had died (Steenland et al. 2004). A validated exposure model was developed using individual workers’ personal breathing zone concentrations collected from 1976 to 1985 to estimate exposure levels. Life-table analyses were conducted for the entire cohort using the U.S. population as the referent population, and SMRs were calculated. Categorical analyses by quartiles of cumulative exposure were based on distributions of cumulative exposure for the deaths from either lymphohematopoietic cancer or from breast cancer. Internal exposure-response analyses were conducted using Cox proportional hazards models. The SMR for lymphohematopoietic cancer was 1.00 (95% CI 0.79–1.24), based on 79 cases. Although the external comparison did not show increased risk, there was a significant internal exposure-response relationship between exposure to ethylene oxide and cancers of the hematopoietic

___________________

¹ This sentence was changed after release of the report to clarify TCEQ’s conclusion on the relationship between ethylene oxide and lymphoid cancer.

system. As discussed in Chapter 2, internal results provide more reliable estimates of risk because they compare low-exposed workers to higher-exposed workers and so do not suffer potential bias from the Healthy Hire Effect.

TCEQ used the NIOSH study results for lymphohematopoietic cancer to estimate a URF. To verify that the standard Cox proportional hazards model that TCEQ developed based on the NIOSH cohort adequately predicted the original data, the model was used to predict the number of lymphoid cancer deaths based on the individual exposure estimates for the NIOSH cohort. Rather than developing the model and verifying it with the same data, the committee notes that cross-validation of the model (setting part of the data aside) is the current best practice to verify it (Yates et al. 2023).

UCC

The UCC cohort consisted of 2,063 male workers in industrial facilities in the Kanawha Valley of West Virginia where ethylene oxide was produced or used. Swaen and colleagues (2009) conducted an analysis using updated data of the UCC cohort that extended the follow-up period by 15 years to December 31, 2003. The average follow-up time was 37 years, and there were 1,048 total deaths. Exposure estimation for the individual workers was based on an exposure matrix that cross-classified three levels of exposure intensity with four time periods. The exposure estimates have higher uncertainty for 1925–1956 than for later years, as earlier exposures were based on estimates from a Swedish facility separate from the UCC facility. SMRs were calculated for cause-specific mortality. Internal exposure-response analyses were conducted using Cox proportional hazards models. In addition, stratified analyses were conducted for duration of employment and cumulative dose. None of these analyses showed evidence of any statistically significant exposure-response relationships, although the SMR for leukemia was elevated above 1 (SMR = 1.51) and the 95% CI ranged from 0.69 to 2.87 in the early-hire group. The analyses were limited by small numbers of deaths.

In Section A3.3.3, TCEQ used a “2013 update of the UCC data” (TCEQ 2020) to validate the Cox proportional hazards model it fit to the NIOSH study data. TCEQ stated that model parameters estimated using the lymphoid cancer mortality data from the NIOSH cohort can be validated by predicting the number of lymphoid cancer deaths in the 2013 update of the UCC study. While the UCC data analyzed in the last peer-reviewed publication (Valdez-Flores et al. 2010) reported 17 deaths due to lymphoid tumors, TCEQ’s validation of its model relies on 25 lymphoid cancer deaths that are not otherwise reported in the peer-reviewed literature. The committee strongly discourages the reliance on the use of unpublished, non-peer-reviewed data to validate any models.

AGE-DEPENDENT ADJUSTMENT FACTORS

Chemicals operating through a mutagenic MOA have an increased potency during early life stages (Barton et al. 2005). As a result, in the absence of chemi-

cal-specific data describing life-stage differences in an agent’s potency, U.S. EPA has recommended the application of ADAFs to account for the increased susceptibility during early life stages to carcinogens exhibiting a mutagenic MOA (U.S. EPA 2005b). The intent of ADAFs is to modify estimates of cancer potency (including inhalation unit risks [IURs]) to account for increased susceptibility during early life stages; U.S. EPA recommends multiplying IURs by factors of 10 and 3, respectively, for exposures occurring from birth to age two and for exposures occurring from the 2nd birthday until the 16th birthday (IURs for exposures occurring after the 16th birthday require no adjustment).

In the TCEQ DSD (TCEQ 2020, Section 4.3.3.1), TCEQ applies both the U.S. EPA method for applications of ADAFs and a different approach proposed by Sielken and Valdez-Flores (2009). Sielken and Valdez-Flores assert that U.S. EPA incorrectly applied ADAFs in its 2006 evaluation of ethylene oxide (U.S. EPA 2006), because the underlying dose-response evaluation relies on a 15-year lagged cumulative dose metric. As a result, Sielken and Valdez-Flores (2009) suggest that application of ADAFs should exclude consideration of increased cancer potency for ages 0 to <15, and that only the 15th year of life would have increased susceptibility requiring a potency adjustment (changing a lifetime potency increase from U.S. EPA’s 66% to 0.008%). This approach inappropriately conflates a method for identifying the slope of a dose-response relationship from an occupational cohort (which involves disregarding exposures closer in time to the outcome event, not earlier exposures) with an adjustment to a slope factor (SFo) based on evidence of increased biological susceptibility during early life stages to mutagenic carcinogens.

While TCEQ mentions and discusses both the U.S. EPA (2006) and Sielken and Valdez-Flores (2009) methods for using ADAFs, TCEQ ultimately relies on the U.S. EPA method for adjustment, with which the committee agrees.

INFLUENCE OF MOA ON MODEL SELECTION

TCEQ Guidelines to Develop Toxicity Factors (hereafter referred to as TCEQ Guidelines) provides the following advice: “If the MOA is mutagenic or the MOA is unknown, then the default is to determine a POD [point of departure] based on the observed data and perform a linear extrapolation from the POD to determine the URF or SFo” (TCEQ 2015, p. 203). TCEQ Guidelines also indicates that “if there is sufficient MOA information indicating a nonlinear dose-response relationship at low doses, then a nonlinear approach can be used to extrapolate risks to doses below the POD” (TCEQ 2015, p. 203). The committee notes that TCEQ Guidelines also discusses Kirman and colleagues’ (2004) use of a quadratic relationship between concentration and cancer risk to determine a POD and low-dose extrapolation, and states that linear extrapolations “should be presented alongside nonlinear extrapolations even if there is sufficient biological information justifying the low-dose nonlinear extrapolation relationship.”² TCEQ evaluated more than one dose-response model, as described below.

___________________

² This sentence was changed after release of the report to clarify TCEQ Guidelines.

DOSE-RESPONSE MODEL SELECTION

TCEQ outlined its rationale for model selection and described its model validation approach. TCEQ considered a log-linear model and a two-piece linear spline model and ultimately selected the log-linear model for URF derivation. TCEQ prioritized consideration of the MOA, model fit, and model accuracy in its model selection.

When evaluating and modeling exposure-response curves, current best practice in dose-response modeling of epidemiology data for risk assessment is to fit and evaluate flexible, nonlinear models (e.g., cubic splines, generalized additive models) to visualize potential nonlinearity in the exposure-response association, as described in Steenland and Deddens (2004).³ Despite evidence from modeling of ethylene oxide epidemiologic data categorically that the exposure-response function may not be linear across all levels of exposure, TCEQ considered linear dose-response models as consistent with its guidance (TCEQ 2015). However, its guidance also allows consideration of additional models when sufficient MOA data are available. TCEQ did not adequately consider flexible, nonlinear models (e.g., splines) when evaluating potential dose-response models to estimate the URF. Given the observed shape of the dose-response association for analyses modeling ethylene oxide categorically, flexible, nonlinear models may provide insight into model selection for the epidemiologic data. For lymphoid cancers, categorical dose-response analyses indicate a nonlinear association (i.e., supralinear, steeper at the lowest end of the exposure-response curve). This is an important consideration since the URF ultimately will be used for environmental ethylene oxide exposures, which occur at the lower end of the exposure distribution relative to occupational exposures experienced in the NIOSH cohort. TCEQ did not prioritize selecting a model that best fits the lowest end of the exposure-response function. Exposure-response data for mortality from occupational cohorts do not adequately represent exposure-response associations for the general population, in part because the lowest end of the occupational exposure range may exceed environmentally relevant concentrations; therefore, the committee notes these are not the most appropriate data for use in setting permissible exposure limits for the general population. The committee also notes that the highest occupational exposures in the NIOSH cohort occurred during earlier time periods when exposure assessment was less robust (prior to extensive personal monitoring), potentially introducing differential and dependent measurement error (TCEQ 2020, Chapter 4). It is therefore possible that the effect estimates for NIOSH participants in the highest categories of exposure may be biased toward the null.⁴

The statistical fit criteria evaluated by TCEQ included comparing Akaike Information Criterion (AIC) values and likelihood ratio test p-values from the two primary models of interest (TCEQ 2020, Table 5). These models were not hierar-

___________________

³ This sentence was changed after release of the report to more accurately reflect Steenland and Deddens (2004).

⁴ This paragraph was changed after release of the report to clarify the consideration of models by TCEQ.

chically nested (the covariates in the smaller were not a subset of the covariates in the larger) and did not use the same model function; thus, TCEQ’s interpretation of the likelihood ratio tests was inappropriate. Nevertheless, the committee agrees that the AIC values indicate that neither the two-piece spline nor the linear model fits the data better than the other.

As is standard practice and described above, TCEQ should prioritize model selection to best fit the data at the lowest end of the dose-response curve because (1) this region of the dose-response curve is most relevant for low-dose extrapolation, and (2) the URF ultimately will be relevant for environmental concentrations well below occupational exposure levels reflected in the dose-response data. Model parsimony should not be prioritized over this consideration. Of the two models, the two-piece spline model may better reflect the supralinear dose-response shape at the lowest end of the exposure distribution. TCEQ guidance states that the MOA must be considered in model selection, and the considered two-piece spline model is not at odds with this guidance. TCEQ also can consider, as feasible, evaluating model fit in a subset of the chosen data that is restricted to the lower end of the exposure distribution. Such approaches have been used to inform low-dose extrapolation when associations are attenuated at the highest exposure levels. Attenuation of the association at the highest exposure levels is not unexpected in occupational studies, further underscoring the need to prioritize model selection based on fit at the lowest exposure levels (Stayner et al. 2003).⁵

TCEQ additionally evaluated model accuracy by comparing observed versus predicted cases for the two considered models. Several flaws in TCEQ’s approach limit the utility of and confidence in its findings. First, TCEQ did not prioritize the comparison of observed versus predicted cases at the lowest ends of the exposure distribution, as described above. Second, TCEQ inappropriately applied U.S. general population background hazard rates when calculating the model-predicted number of lymphoid cancer deaths. TCEQ justified this decision by incorrectly concluding that there was no evidence of Healthy Worker Effect (see Chapter 2). Because the expected number of cases was modeled considering the general population cancer rate (which is expected to overestimate the number of cases relative to observed cases in either cohort because of bias due to the Healthy Hire Effect), estimates of expected cases from both models are expected to overestimate the number of observed cases. Third, TCEQ evaluated the predictive ability of each model (developed with NIOSH study data) to predict cases in the UCC cohort. TCEQ’s approach to “ground-truthing” the selected model with UCC data (comparing the expected number of events under each model to the observed number of events) is not best practice, nor is it considered a standard approach in the field. Internal cross-validation is the current standard in the field (Yates et al. 2023). TCEQ’s approach is also subject to bias from measurement error (given the substantial differences in exposure assessment for the NIOSH and UCC cohorts) and the Healthy Hire Effect (given that TCEQ modeled the expected number of cases

___________________

⁵ This paragraph was changed after release of the report to remove content that was beyond the scope of the committee’s review.

under both primary models using cancer rates from the general population, which are subject to bias due to the Health Worker Effect).

TCEQ’S SYSTEMATIC REVIEW FOR DOSE-RESPONSE

The field of risk assessment is evolving rapidly, and the overall trend occurring within many governmental agencies is to increase transparency and rigor, often through the use of systematic review methods. Systematic review methods for the purposes of hazard identification in environmental risk assessments have been developed (NTP 2019; Woodruff and Sutton 2014) and could be applied by TCEQ to the development of toxicity values. Chapter 2 discussed the application of systematic review to hazard assessments. Rather than use systematic review methods in the hazard assessment, TCEQ conducted a systematic review of published data on the dose-response relationship between ethylene oxide and hematopoietic cancer outcomes. The committee found several critical limitations of TCEQ’s systematic review.

TCEQ’s problem formulation and protocol lacked specificity, as noted in Chapter 2. TCEQ provided information regarding its literature search and development of inclusion and exclusion criteria for multiple types of studies. TCEQ also evaluated study quality and risk of bias using a list of study quality domains for epidemiology studies taken from the U.S. EPA Health Assessment Workspace Collaborative (HAWC) tool. The HAWC is a widely used modular web-based tool intended to store and display information to support systematic reviews. The committee agrees with TCEQ’s decision to use the HAWC study quality domains. The study quality domains for epidemiology studies along with a descriptive statement defining each level of the four-point rating criteria are included in the TCEQ DSD (Appendix 1). A four-point-scale rating for each domain for each study is presented in Table 24 of the TCEQ DSD; however, there is no text explaining why a given study received the rating selected, which is a fundamental component of conducting a systematic review to increase transparency (i.e., so that one can determine why the key study was selected). The transparent and unbiased selection of key studies that inform the final assessment conclusions is essential; without it, the credibility and reliability of the overall assessment are fundamentally compromised. The National Academies of Sciences, Engineering, and Medicine (National Academies) have also recommended that “funding sources should be considered in the risk-of-bias assessment conducted for systematic reviews” (see NRC 2014, p. 79), a consideration that is not currently included in the recommended HAWC risk-of-bias assessment tool.

Best practices for systematic review for risk assessment include a formal and transparent consideration of the body of evidence (see Box 2-1, Figure 2-1-1). TCEQ did not follow a formalized and transparent process for considering the body of evidence. In its section titled “Evidence Integration,” TCEQ provided a summary of each epidemiologic study and designated each study as “key,” “supporting,” or “informative” for the ethylene oxide carcinogenic dose-response assessment. Unfortunately, no information was provided on the meaning of these

classifications or how they were assigned, thus contributing to a lack of transparency. It is unclear how “supporting” or “informative” studies contribute to TCEQ’s evidence synthesis. TCEQ provided a brief rationale for a single “key” study (Valdez-Flores et al. 2010) but did not provide a comparable rationale for any other study. In addition, TCEQ did not identify and evaluate animal and mechanistic data as part of the body of evidence. This type of data may be useful for dose-response modeling.

OTHER CONSIDERATIONS: ENDOGENOUS ETHYLENE OXIDE

Part (c) of the committee’s second charge is as follows: “[consider] any implications of the endogenous production of ethylene oxide, including from the perspective of biological significance, for risk-based air concentrations.” Production of endogenous ethylene oxide occurs from the metabolism of ethylene. Metabolic pathways include the production of ethylene by gastrointestinal tract bacteria, with subsequent absorption and delivery of ethylene to the liver where metabolism to ethylene oxide can occur (Kirman et al. 2021). Inhalation or ingestion of exogenous ethylene can undergo a similar metabolic fate, resulting in ethylene oxide production. Hepatocyte metabolism of methionine and 2-keto-4-methylthiobutyr-ic acid also contributes to the production of ethylene (Csanády et al. 2000; Fu et al. 1979; Lawrence and Cohen 1985). Isotope labeling (e.g., ¹⁴C-ethylene oxide) can be used to differentiate absorbed ethylene oxide from exogenous sources and endogenously produced ethylene oxide.

Ethylene oxide is an alkylating agent that forms DNA, RNA, and protein adducts. The hemoglobin adduct of ethylene oxide, 2-hydroxyethylvaline (HEV), has been used as an ethylene oxide biomarker (Kirman et al. 2021; Walker et al. 2000). Formation of HEV demonstrates an observed linear relationship between air exposure levels and HEV levels in workers (Angerer et al. 1998). At occupational exposure levels, a relationship exists between exposure to ethylene oxide and increased levels of HEV. Human blood HEV concentrations can be used to estimate endogenous equivalent concentrations and total equivalent concentrations (Kirman and Hays 2017).

In order to address its statement of task, the committee considered TCEQ Guidelines. This document provides minimal advice concerning the risk assessment of an endogenous compound and merely states, “The corresponding toxicity factors may represent an unrealistic characterization of environmental risk (at least from a regulatory compliance perspective). The possibilities of such unreasonableness should be evaluated based on all relevant information (e.g., MOA, background exposure and rates of response, reliable typical human breath concentrations due to endogenous production, and data related to possible thresholds)” (TCEQ 2015, p. 212).

When considering the biological significance of endogenous ethylene oxide production, TCEQ assumes that an equivalent internal dose of ethylene oxide resulting from either endogenous production or exogenous exposure would be equi-

potent in producing a carcinogenic effect. TCEQ relies on analyses performed by Kirman and Hays (2017), who estimated endogenous equivalent levels of ethylene oxide from measured HEV adduct levels in humans. Ultimately, this “reality check” performed by TCEQ did not play a direct role in the agency’s selection of a dose-response model.⁶

One challenge with TCEQ’s analysis is that it largely relies on HEV adducts as a surrogate for both internal ethylene oxide dose and cancer potency. It is unlikely that HEV adducts are directly involved in the mutagenicity of ethylene oxide. A more likely contributor to the mutagenic effects associated with ethylene oxide is the production of DNA adducts. The type of DNA adduct and the cell type in which it is formed may influence mutation rates. To conduct a risk assessment more accurately, the relative contribution of endogenously versus ex-ogenously derived mutagenic DNA adducts and subsequent mutation rates would therefore be valuable. To date, most studies have focused on N7-(2-hydroxyethyl) guanine (N7-HEG) since it has been the most readily measured DNA adduct. However, N7-HEG is not thought to be directly pro-mutagenic (Tompkins et al. 2009). Studies using dual-isotope approaches and highly sensitive liquid chromatography–mass spectrometry assays have been used to distinguish endogenous and exogenous (¹⁴C-labeled) N7-HEG adducts in rat spleen, liver, and stomach following intraperitoneal administration of ¹⁴C-labeled ethylene oxide (Marsden et al. 2007, 2009). In these studies, levels of ¹⁴C-N7-HEG induced in DNA extracted from treated rat tissues demonstrated a linear dose-response relationship; adducts arising from exogenous administration were negligible compared to the much higher background levels of endogenous N7-HEG. In addition, there is concern that N7-HEG formation may arise from metabolic pathways independent of ethylene oxide formation. Limited information is available regarding O⁶-(2-hydroxyethyl)guanine and N³-(2-hydroxyethyl)adenine and other mutagenic DNA adducts (Tompkins et al. 2008). Thus, a data gap remains for comparative risk assessments based on DNA alkylation rather than hemoglobin adduct formation.

The endogenous production of ethylene oxide has biological relevance for the incidence of cancer. The committee agrees with TCEQ’s decision to use human epidemiologic cohort data with ethylene oxide concentrations assessed in ambient air as the basis for assessing the dose-response relationship for the derivation of the URF for risk-based air concentrations. The committee supports TCEQ’s conclusion that endogenous production of ethylene oxide may not inform the dose-response curve and derivation of the URF for ethylene oxide present in ambient air.

FINDINGS AND RECOMMENDATIONS

Finding 3.1: TCEQ used epidemiologic data on lymphoid cell cancer mortality rather than incidence to model dose-response relationships for URF derivation.

___________________

⁶ This paragraph was changed after release of the report to remove content that was beyond the scope of the committee’s review.

Recommendation 3.1 (Tier 1): The Texas Commission on Environmental Quality should develop unit risk factors for cancer incidence instead of cancer mortality, through dose-response modeling of relationships between ethylene oxide and cancer incidence, where such data are available, or through scaling of dose-response relationships for cancer mortality based on cancer incidence rates and survival.

Finding 3.2: The grouping of lymphoid cell cancers included in TCEQ’s dose-response modeling does not reflect the individual subtype risks and may mask associations with more specific lymphoid cancer subtypes. Furthermore, a positive association between ethylene oxide exposure and Hodgkin lymphoma has been reported, yet Hodgkin lymphoma was not included in TCEQ’s cancer grouping.

Recommendation 3.2 (Tier 1): When data are available, the Texas Commission on Environmental Quality should derive separate unit risk factors for lymphoid cell subtypes, including Hodgkin lymphoma.

Finding 3.3: TCEQ did not appropriately consider alternative nonlinear models (which are suggested by TCEQ guidance). The model selection criteria and the approach to validate the selected model were inadequate and inconsistent with best practice.

Recommendation 3.3 (Tier 1): The Texas Commission on Environmental Quality (TCEQ) should evaluate flexible, nonlinear models (e.g., cubic splines). TCEQ should prioritize selecting a model that best fits the dose-response curve at the lowest end of the exposure distribution rather than a model that best fits occupationally relevant levels of the exposure-response curve or the entire exposure range. TCEQ should use internal cross-validation and should not attempt to validate models built using National Institute for Occupational Safety and Health data with Union Carbide Corporation data. TCEQ should not use general population rates to model expected cancer rates given bias due to the Healthy Hire Effect.

Finding 3.4: The TCEQ DSD appropriately included discussion of different approaches. The committee agrees with the use of U.S. EPA’s ADAFs. The Sielken and Valdez-Flores ADAF method has a fundamental flaw, and its inclusion in the discussion weakens the TCEQ DSD.

Recommendation 3.4 (Tier 2): Discussion of the Sielken and Valdez-Flores method should be removed from the 2020 Ethylene Oxide Carcinogenic Dose-Response Assessment: Development Support Document because of methodological concerns inherent in the study.

As noted in Chapter 2, TCEQ did not complete a robust systematic review. TCEQ should conduct a more thorough systematic review (Recommendation 2.1, Tier 1), following all standard steps recommended in prior National Academies’ reports in a complete and transparent way. This systematic review will also inform the dose-response assessment.