Use of updated mortality study of ethylene oxide manufacturing workers to inform cancer risk assessment

Abstract

The two most recent cancer risk assessments for ethylene oxide (EO) are based on the same epidemiologic study of sterilant workers conducted by the National Institute of Occupational Safety and Health (NIOSH) but result in cancer risk estimates with three orders of magnitude difference, despite relying on the same assumption of a default linear (non‐threshold) extrapolation. A major reason for the difference is the use of different exposure‐response models (i.e., the standard Cox proportional hazards [CPH] versus a two‐piece linear spline model with a steep initial slope) to derive the inhalation unit risk. The purpose of this research is to utilize analysis of a 10‐year update of the Union Carbide Corporation (UCC) EO 2053 chemical worker cohort to examine the epidemiological evidence for the shape of the exposure‐response model for EO. This updated UCC study provides an external dataset that is informative given high average cumulative exposures (67 ppm‐years), extensive average follow‐up of over 40 years, and number of male lymphoid cancer deaths (25) comparable to that observed in the NIOSH cohort. This independent analysis of a different cohort using continuous dose response modeling with cumulative or log cumulative exposure metrics provides no empirical support for a steep curve at low exposures. Furthermore, analyses of the categorical odds ratio estimates across different updates of the UCC cohort and for each sex in the NIOSH cohort provide further epidemiological evidence that the standard CPH model more plausibly describes the relationship between EO exposures and lymphoid mortality for both cohorts.

1. INTRODUCTION

Ethylene oxide (EO) is a chemical widely used in the production of other chemicals, in the sterilization industry, and in the preparation of spices. EO is characterized by the US Environmental Protection Agency's (EPA's) Integrated Risk Information System (IRIS) as “carcinogenic to humans” (EPA IRIS, ²⁰¹⁶, p. 1). Thus, deriving an adequately protective cancer risk value is of importance to protecting workers and the general population.

The putative mode of action for carcinogenicity is mutagenicity (EPA IRIS, ²⁰¹⁶; Gollapudi et al., ²⁰²⁰). EO is a direct‐acting alkylating agent that does not require metabolic activation to form DNA adducts, the hypothesized initial key molecular initiating event. However, EO is a weak genotoxicant based both on the magnitude of the observed response and the doses and exposure durations needed to elicit a significant response, as described in greater detail by Gollapudi et al. (2020). For example, the mutagenic response in the bone marrow of Big Blue (lacl transgenic) mice is significantly increased following chronic (48 week) EO exposures of 100 and 200 ppm but not 25 and 50 ppm (Gollapudi et al., 2020; Recio et al., ²⁰⁰⁴). The toxicokinetics of EO indicates a linear relationship between EO blood concentrations and EO exposure concentrations (50–200 ppm) in mice, rats, and humans (Fennell & Brown, 2001). The EPA's default approach for quantitative risk assessment (QRA) for any presumed mutagenic agent is to assume a linear low‐exposure extrapolation from a point of departure derived from exposure‐response models that are applied to the selected data (EPA, ²⁰⁰⁵; EPA IRIS, ²⁰¹⁶).

Cancer risk assessments for EO conducted by EPA IRIS (2016) and the Texas Commission on Environmental Quality (TCEQ, ²⁰²⁰), while relying on data from the same cohort study of sterilizer workers published by authors from the National Institute of Occupational Safety and Health (NIOSH) (Steenland et al., 2004, 2003), derived different lymphoid cancer risk values based on linear extrapolation from two different models. In both assessments, low‐exposure linear extrapolation was applied from the 95% upper confidence limit on the slope to derive a cancer unit risk factor (URF). The URF derived by EPA (9.1 per ppm) is three orders of magnitude higher than the URF derived by TCEQ (4.1E‐03 per ppm). A major reason for the difference between the TCEQ and EPA risk values is the selection of very different exposure‐response models applied to the NIOSH study. In addition, EPA combined model estimates from both lymphoid mortality and breast cancer incidence, whereas TCEQ included only lymphoid mortality. However, lymphoid mortality contributes 87% to the EPA's cancer risk value. Thus, the important question for EO cancer risk assessment centers on the shape and steepness of the exposure‐response model that should be applied to the NIOSH lymphoid dataset.

This paper provides the first in‐depth consideration of the plausibility of exposure‐response models for the lymphoid data based on weight of evidence from epidemiology studies, including consideration of both continuous and categorical models. In this context, plausibility refers to the consistency of the most informative epidemiological evidence, the NIOSH and updated Union Carbide Corporation (UCC) studies, with exposure‐response model selection. Toward that purpose, an update of the UCC EO workers cohort is reported herein, in comparison with earlier updates of this cohort and with NIOSH findings reported for sterilizer workers. The UCC study and this new update provide an external dataset to inform the plausibility of the different continuous dose‐response models applied to the NIOSH lymphoid data used by EPA and TCEQ for cancer risk assessment.

EPA IRIS used a steep two‐piece linear spline Cox proportional hazards (CPH) model in which two linear functions are joined at a single point of inflection at 1600 ppm‐days, called a “knot,” such that the first slope at lower exposures is much steeper than the second¹ (Figure 1).

FIGURE 1.

Visual representation of Environmental Protection Agency (EPA) Integrated Risk Information System (IRIS) (2016) two‐piece linear spline model.

This knot was selected by evaluating knots in multiple increments of 100 ppm‐days and considering those with the largest overall and local maximum likelihood (EPA IRIS, ²⁰¹⁶, pp. 4–15). This type of two‐piece spline model was selected because the published NIOSH study by Steenland et al. (2004) reported statistically significant results of the standard CPH model with log cumulative exposure. In contrast, TCEQ (2020) applied a single standard (log‐linear) CPH model to the NIOSH mortality study, which is a single slope² that is very nearly linear (EPA IRIS, ²⁰¹⁶, p. D‐2).

The EPA cancer risk assessment guidelines (EPA, ²⁰⁰⁵, pp. 3–14) state that any statistical model based on individual data that is to be considered reasonable for risk assessment should not only have a sound statistical basis but also have an exposure‐response form that is “both biologically plausible and consistent with the observed data” (U.S. Environmental Protection Agency Science Advisory Board [SAB], ²⁰¹⁵, p. 2, 9, 11). Neither the EPA two‐piece linear spline model nor the TCEQ standard CPH model with cumulative exposure as the dose metric are statistically significantly different from the null model of no increase in risk with increasing exposure (p = 0.14³ and 0.22, respectively) (TCEQ, ²⁰²⁰). Given the comparable statistical results a weight of evidence approach based on the epidemiological and toxicological data should play a primary role in the model selection. This manuscript focuses on the plausibility of the dose‐response models based on consistency with epidemiological studies that include individual quantitative exposure estimates.

The UCC cohort is the only EO occupational cohort, other than the NIOSH cohort, with individual quantitative exposure estimates that is suited for quantitative exposure‐response analysis. The cumulative exposures to EO from the 25th to 75th percentiles of the UCC study subjects are greater than those in the corresponding percentile ranges of the NIOSH study (Table S1). While it is not our purpose to derive a new URF based on the UCC cohort, the exposure‐response analyses of the data from this study population, using the standard CPH model and the CPH model with log cumulative exposure, inform the plausibility of the exposure‐response models applied to the NIOSH cohort by EPA IRIS and TCEQ.

EPA IRIS (2016) justified use of a steep two‐piece linear spline model based on a significant log cumulative exposure model. We agree with EPA IRIS (2016) that such models are biologically implausible because the initial slope of these models at lower exposures approaches infinity (Valdez‐Flores et al., 2010). However, since EPA IRIS (2016) used this model as the justification for triggering statistical modeling using two‐piece linear spline models, we applied this log cumulative exposure model to the UCC study update.

In terms of total number of workers, the NIOSH study is a much larger cohort compared to the UCC study (18,235 male and female workers vs. 2053 male workers). This is true even when only male workers (7634) are considered. However, two major determining factors for statistical power are the length of follow‐up and the actual number of cancer cases found in the cohort. The UCC cohort is not only the largest cohort of highly exposed EO workers other than the NIOSH cohort (average cumulative EO exposure levels of 67 ppm‐years compared to 27 ppm‐years for the NIOSH cohort) (Steenland et al., 2004; Swaen et al., ²⁰⁰⁹), but also has extensive average follow‐up of over 40 years, compared to 25 years for the NIOSH cohort, resulting in a comparable number of cases in the most susceptible sex (males), for the cancer having greatest impact on the cancer risk assessment (25 lymphoid deaths in the UCC cohort vs. 27 in NIOSH cohort).

A limitation of the UCC study is that it cannot address female cancers such as breast cancer mortality because the cohort is comprised of only male workers, whereas the NIOSH sterilant worker cohort included both males and females (45% and 55%, respectively). The NIOSH study subjects were employed at 14 different sterilizing facilities and followed up for mortality initially to 1987 (Steenland et al., 1991) and later updated to 1998 (Steenland et al., 2004). However, the lymphoid cancers in the NIOSH cohort play a predominate role in the derivation of URFs, contributing approximately 87% and 100%, to the EPA IRIS and TCEQ URFs, respectively (EPA IRIS, ²⁰¹⁶; TCEQ, ²⁰²⁰). Lymphoid cancers, as defined by Steenland et al. (2004), include non‐Hodgkin lymphoma (NHL), multiple myeloma (MM), and lymphocytic leukemia (LLK) under the hypothesis that these tumors might share a common etiology. Importantly, Steenland et al. (2004) concluded that increases in lymphoid cancer mortality in the NIOSH cohort were found in males only and only at the highest exposure category, with female lymphoid cancers having a deficit and an apparent inverse relationship with exposures.

Thus, for the purpose of considering the epidemiological weight of evidence to inform the dose‐response models that were applied to the NIOSH lymphoid cancers, there is no other EO epidemiological study with comparable numbers of cancers of interest and reasonable individual estimates of EO exposures (discussed below) than the UCC study. The results of this UCC study update are compared with those of previous UCC updates and the NIOSH study using the same dose‐response modeling approaches. These comparisons are also discussed within the context of the EPA IRIS (2016) and original Steenland et al. (2004) analysis.

2. MATERIALS AND METHODS

2.1. Use of the UCC cohort as an external validation data set to inform exposure‐response analysis

The UCC study was derived originally from a cohort mortality study of 29,139 UCC male chemical workers in the Kanawha Valley of West Virginia (Table 1). This study was conducted by NIOSH (Rinsky et al., 1988) and included complete work histories and a chemical dictionary, linking chemicals to work areas. Greenberg et al. (1990) published the findings of the subpopulation of workers in this cohort exposed to EO as early as 1925. Workers hired from 1940 on were included in the study and followed up for mortality through the end of 1978. Teta et al. (1993) and Swaen et al. (2009) updated this UCC study to 1988 and 2003, respectively. In the present study, we updated the study by an additional 10 years through the end of 2013, increasing the power of the study by including a longer observation period and therefore a larger number of decedents from the causes of interest. Included are 2053 EO‐exposed male workers active in 1940 or hired between 1940 and 1988. Since 70% of study subjects are deceased by 2013, longer updates are not expected to add new information. In addition, longer postexposure follow‐up could potentially make it more difficult to distinguish background disease from exposure‐related disease.

TABLE 1.

Comparison of events of primary interest for internal quantitative exposure‐response analysis based on National Institute of Occupational Safety and Health (NIOSH) and Union Carbide Corporation (UCC) data sets for male workers.

	NIOSH	UCC
Characteristic	Steenland et al. (2004)	Greenberg et al. (1990)	Teta et al. (1993)	Swaen et al. (2009)	Updated analysis
Final year of follow up	1998	1978	1988	2003	2013
All male workers	7634	2174	1896	2063	2053 a
Percent deceased	19%	14%	23%	51%	70%
Average follow up	25 years	NR2	>27	37 years	41 years
Person‐years	190,850	NR	>51,192	76,331	84,173
PPM‐year exposure	27	NR	NR	67	67
Primary endpoint for quantitative exposure response analysis
Lymphoid cancers b	27	NR	NR	17	25
Subcategories of lymphoid cancers ( EPA IRIS, 2016)
Non‐Hodgkin's lymphoma	18	2	NR	12	15
Related blood cancer endpoints of interest c
All lymphatic and hematopoietic	37	9	7	27	41
Leukemia	10	7	5	11	20
EPA IRIS (2016) animal inhalation carcinogenicity endpoints
Respiratory cancers (including lung)	143	19	44	99	120
Central nervous system	NR	4	6	13	16

Abbreviations: EPA, Environmental Protection Agency; IRIS, Integrated Risk Information System.

^a

The Swaen et al. (2009) update included 10 workers who were later recognized to be less than full time employees, and so the current update included 2053 male ethylene oxide (EO) workers.

^b

NR = Not reported.

^cLymphoid tumors are defined by Steenland et al. (2004) as Non‐Hodgkin's Lymphoma, Multiple Myeloma and Lymphocytic Leukemia (ICD 9^th revision codes 200, 202, 203, 204).

Teta et al. (1993) removed 278 workers from the UCC EO cohort examined by Greenberg et al. (1990) who worked in the chlorohydrin production unit (men assigned to unit making chlorohydrin, not making EO by the chlorohydrin process), due to potential confounding by other chemicals, as described in Benson and Teta (1993). These men with significantly elevated leukemia and pancreatic cancer death rates had low potential for EO exposure but heavy exposure to the by‐products of ethylene chlorohydrin (EC) production. The remaining differences in numbers of study subjects in the various updates are minor due to improved information. The number of deceased workers increases with longer follow‐up, resulting in a total of 1453 decedents.

2.2. Selection of cancers of interest for exposure‐response analysis

As described above, the focus of this exposure‐response analysis of the updated UCC cohort is on lymphoid cancer mortalities utilized by EPA IRIS (2016) and TCEQ (2020) for exposure‐response analysis using the NIOSH sterilant worker study. Subcategories of lymphoid cancers (NHL, MM, and LLK), leukemia, and the larger category of all lymphatic and hematopoietic cancers (lymphohematopoietic or LH), as well as cancers of interest identified by EPA IRIS (2016) in animal studies, were also considered for exposure‐response analysis. EPA IRIS (2016) listed lung, mammary gland, uterus, lymphoid cells, brain, and tunica vaginalis testis as tissues that had increased incidence of cancers in rats and mice following chronic inhalation exposures to EO. Since the UCC cohort did not include females, only lung, central nervous system (CNS), and testis were considered for further analysis. Testis was not included because there were no cases in the UCC cohort. Internal exposure‐response analyses were applied to cancers of interest with at least 15 cancer mortalities needed to adequately characterize the exposure‐response pattern for both continuous and categorical modeling (Table 2).

TABLE 2.

Observed (Obs) and expected (Exp) numbers of deaths, standardized mortality ratios (SMRs), and 95% confidence intervals (95% CIs) for selected causes of death among Union Carbide Corporation (UCC) ethylene oxide (EO) workers (N = 2053) followed through December 31, 2013.

Cause of death (ICD‐10 codes)	Obs	Exp	SMR	95% CI
All causes (A00‐R99, U01‐Y89)	1435	1616.6	0.89 *	0.84–0.94
All malignant neoplasms (C00‐C97)	392	407.1	0.96	0.87–1.06
Buccal cavity and pharynx (C00‐C14)	5	8.3	0.60	0.19–1.40
Digestive organs and peritoneum (C15‐C26, C48)	111	98.1	1.13	0.93–1.36
Esophagus (C15)	11	11.4	0.97	0.48–1.73
Stomach (C16)	14	11.5	1.21	0.66–2.04
Large intestine (C18)	39	34.0	1.15	0.82–1.57
Rectum (C20‐C21)	6	6.9	0.87	0.32–1.89
Biliary passages and liver (C22, C24)	13	10.2	1.27	0.68–2.18
Pancreas (C25)	24	20.8	1.15	0.74–1.72
All other digestive organs (C17, C19, C23, C26, C48)	4	3.3	1.19	0.32–3.06
Respiratory system (C30‐C39)	120	144.0	0.83 *	0.69–1.00
Larynx (C32)	2	4.5	0.44	0.05–1.59
Bronchus, trachea, lung (C33‐C34)	116	138.3	0.84	0.69–1.01
All other respiratory cancer (C30‐C31, C37‐C39)	2	1.2	1.72	0.19–6.21
Prostate (C61)	27	39.8	0.68 *	0.45–0.99
Testis (C60, C62‐C63)	0	1.0	0.00	0.00–3.67
Kidney (C64‐C65)	15	10.2	1.47	0.82–2.42
Bladder or other urinary organs (C66‐C68)	11	13.1	0.84	0.42–1.51
Malignant melanoma of skin (C43)	3	6.5	0.46	0.09–1.36
Eye (C69)	0	0.2	0.00	0.00–18.3
Central nervous system (C70‐C72)	16	9.6	1.67	0.95–2.71
Thyroid and other endocrine glands (C73‐C75)	1	1.1	0.88	0.01–4.88
Bone (C40‐C41)	1	0.9	1.14	0.02–6.37
All lymphatic and hematopoietic (C81‐C96)	41	40.7	1.01	0.72–1.37
Hodgkin disease (C81)	0	1.8	0.00	0.00–2.04
Non‐Hodgkin lymphoma (C82, C83.0‐C83.9, C84, C85.1‐C85.9)	15	15.3	0.98	0.55–1.62
Leukemia (C91‐C95)	20	16.1	1.24	0.76–1.92
All other lymphopoietic tissue (C88, C90, C96)	6	7.4	0.81	0.29–1.75
Multiple myeloma (C90)	5	7.0	0.71	0.23–1.67
Lymphoid (C82, C83.0‐C83.9, C84, C85.1‐C85.9, C90, C91)	25	26.8	0.93	0.60–1.38
Soft tissue sarcoma (C49)	3	2.0	1.49	0.30–4.35
Nasopharyngeal cancer (C11)	1	0.6	1.65	0.02–9.20
All other malignant neoplasms (C44‐C47, C76‐C79, C80, C97)	38	31.8	1.20	0.85–1.64
Non‐malignant diseases
Diabetes mellitus (E10‐E14)	31	34.4	0.90	0.61–1.28
Cerebrovascular disease (I60‐I69)	82	85.7	0.96	0.76–1.19
Heart disease (I00‐I02, I05‐I09,I11, I13‐I14, I20‐I28, I30‐I52)	486	568.7	0.86 *	0.78–0.93
Non‐malignant respiratory disease (J00‐J99)	117	153.4	0.76 *	0.63–0.91
All external causes (V01‐Y89)	66	92.8	0.72 *	0.55–0.91
Unknown causes	2	–	–	–

Note: Shaded rows are endpoints considered most relevant to EO cancer risk assessment based on animal and human studies (see text).

*Statistically significantly <1 at the 5% significance level.

Standardized mortality ratios (SMR) were calculated for a much larger comprehensive list of potential endpoints (see below). No additional cancers were identified as having statistical significance indicative of further consideration for exposure‐response analyses.

2.3. UCC exposure assessment

The methodology for EO exposure estimation was developed and published in Teta et al. (1993) and described in greater detail in Swaen et al. (2009) (Table S2). UCC workers started producing and using EO in the earliest years of this industry, starting in 1925. The published literature includes documentation of production upsets, working in enclosed buildings, spills, worker acute over‐exposures, lack of use of personal protective equipment or engineering controls, and worker ability to smell EO, implying hundreds of ppm (Greenberg et al., 1990; Joyner, ¹⁹⁶⁴; Sexton & Henson, ¹⁹⁴⁹). Data from the literature related to companies using comparable methods of EO production were relied upon for the 1940–1956 time period. From 1957 to 1988, exposure data were available for >75% of the cohort based on routine monitoring, personal sampling, and a plant‐wide survey in another UCC plant using the same process (see Table S2). Cumulative exposure to EO was expressed as ppm‐days. The distribution of the cumulative exposures did not change noticeably between the current and previous UCC update (Table S1).

2.4. Mortality follow‐up

Vital status and cause of death ascertainment relied on records included in the epidemiology surveillance system (ESS) maintained by The Dow Chemical Company, which acquired UCC and maintained custody of the UCC study data. ESS was a computerized collection of demographics, work history, vital status, and cause of death information for current and former full‐time Dow employees, including the UCC workers. Biennial record linkage was performed with the Social Security Administration; the National Death Index (NDI) maintained by Centers for Disease Control and Prevention National Center for Health Statistics; the vital records offices of most states; and genealogical sources such as Ancestry.com. Cause of death was coded to the version of International Classification of Diseases (ICD) in effect at the time of death (see Table S3 in the Supporting Information for ICDs corresponding to the six endpoints analyzed here). For all workers who died in the United States we relied on the ICD code assigned by NDI or else obtained the death certificate for encoding by a trained nosologist.

2.5. External analyses⁴ of mortality endpoints: SMR

SMRs were calculated using SAS, employing subroutines specifically developed to analyze cohort studies based on a modified life table procedure (SAS, ²⁰¹³). Workers accumulated person‐years at risk beginning on 01‐01‐1940 or the first employment date at an EO facility, whichever came later, and ending on 12‐31‐2013, the date of death, or the date when the worker was lost to follow‐up, whichever came earliest. The number of deaths observed among the UCC EO workers was compared to the number of deaths that would have been expected if the UCC workers had experienced death at the same rates as the United States white males from the general population. SMRs and 95% confidence intervals for 40 causes of death were computed based on the endpoints reported by Swaen et al. (2009), adding lymphoid cancers as a group. US reference rates were used to facilitate comparisons with published SMRs based on the same rates. Greenberg et al. (1990) reported that there was no difference in SMRs when US or regional rates were used.

2.6. Internal analyses⁵: Continuous exposure‐response analyses and model slope estimation

Internal exposure‐response analyses were conducted using SAS (2013). The standard CPH is commonly used in epidemiology to model the relationship between the hazard of cause‐specific mortality and exposure (Richardson, 2010). This model was fit to each of the six cancer endpoints described in Table 1 that are relevant to informing the shape of the exposure‐response model used for cancer risk assessment by EPA IRIS (2016) and TCEQ (2020). In these analyses, the time variable was age (effectively matching on age), and full risk sets were constructed using all workers in the cohort who survived without the cancer of interest to at least the age of the index case. Statistical significance for the continuous modeling was based on the likelihood ratio test for alternative hypothesis that the coefficient of the standard CPH log‐linear model is different than zero.

The standard CPH model assumes that the logarithm of the cause‐specific mortality hazard rate is linearly related to the cumulative exposure to EO (ppm‐days). Exposure‐response models were also fit assuming non‐zero lags and lags of 5–30 years in increments of 5 years with cumulative exposure to EO. Since there were no statistically significant differences in results based on lag, zero lag was used in the final main analyses (see Table S5 for model fits using exposures with non‐zero lags).

The CPH model was not adjusted for any covariates because none of the covariates significantly improved the likelihood of the model. In the UCC dataset, workers’ race was classified as white or Caucasian, black, non‐white, and unknown. Consistent with previous analyses of the UCC data, the primary analyses were not adjusted for race because of the substantial number of subjects of unknown race (Rinsky et al., 1988). The present update of the UCC study was fit to the six cancer endpoints and compared with the results from the previous UCC update reported by Valdez‐Flores et al. (2010), who published statistical analyses of the UCC data updated through 2003. In addition, a post hoc analysis adjusting the standard CPH model for race and year of birth did not change the results, so no further analysis was performed with this adjustment (Table S7).

For comparison purposes, exposure‐response models were fit to lymphoid cancers using the standard cumulative exposure metric and the log‐transformed cumulative exposure metric. The log‐transformed cumulative exposure metric requires imputation of a non‐zero value for zero exposures to avoid taking the logarithm of 0. The model with log‐cumulative exposure was fit, imputing the 0 cumulative exposure using four alternative values.

2.7. Internal analyses: Categorical exposure‐response modeling

Categorical exposure‐response models using internal comparisons were evaluated using the standard CPH analyses for the same six cancer endpoints selected for continuous exposure‐response as described above. Five categories of cumulative exposures to EO were defined so that there were approximately the same number of cause‐specific deaths in each category for each cancer endpoint (see Table S4 footnotes for details of analysis for both UCC and NIOSH studies). Hazard rates at different exposure categories were compared to the hazard rate for the lowest exposure category to estimate rate ratios. Wald's test was used to calculate the p‐value for the alternative hypothesis that the rate ratio is >1.

2.8. Comparisons with NIOSH lymphoid mortality results

Comparisons between the present internal continuous and categorical analyses of the UCC update with those of the NIOSH study were evaluated in the results section. The analysis results of the lymphoid cancers from the NIOSH study were obtained from Table S9 of Valdez‐Flores et al. (2010) and are based on the same methods as the present study (see Table S4 footnotes for details). While the Results section of the present study is primarily focused on lymphoid cancers, additional comparisons for key cancers of interest are included in Table S4. The discussion section expands the comparisons with NIOSH study analyses reported by Steenland et al. (2004) and EPA IRIS (2016) with a 15‐year lag.

Consistent with our analyses of the UCC data, the Valdez‐Flores et al. (2010) analysis of the NIOSH lymphoid cancers is based on the full risk set of the NIOSH study. This differs from Steenland et al. (2004), in which 100 randomly selected controls were chosen for each case from the pool of all those who survived without the lymphoid cancer to at least the age of the index case. Our internal categorical analyses for lymphoid and other LH subcategories of cancers also differ from Steenland et al. (2004) in that five exposure categories were uniquely defined for each cancer based on evenly dividing the number of cases. Steenland et al. (2004) defined four exposure categories based on evenly dividing the number of LH cancers into four exposure categories (0–1199, 1200–3679, 3680–13499, and 13500+) and applying these exposure categories to lymphoid and other subcategories of LH cancers. A comprehensive discussion of different modeling methods is described in Valdez‐Flores et al. (2010).

3. RESULTS

3.1. Standardized mortality ratios

The SMRs in Table 2 show that none of the 40 endpoints analyzed resulted in a statistically significant increase in mortality when compared with the US population. There were statistically significant deficits for three non‐cancer causes of death—heart disease, non‐malignant respiratory disease, and all external causes. The SMR for all malignant neoplasms was 0.96, based on 392 deaths versus 407 expected. The SMRs for lymphoid and LH cancers, the primary endpoints of interest, indicate that the number of observed deaths was comparable with expectation, 0.93 (95% confidence interval [CI]: 0.60, 1.38) and 1.01 (95% CI: 72, 1.37), respectively. There was a statistically non‐significant increase in leukemia deaths (1.24, 95% CI: 0.76, 1.92), based on 20 workers who died from leukemia (including five deaths from LLK). All other subcategories of LH cancers have a non‐statistically significant deficit in the number of observed deaths. There were statistically significant deficits of respiratory and prostate cancers and a non‐statistically significant excess of cancers of the CNS.

Table 3 presents the SMRs for the cancers of regulatory interest by the four UCC cohort study follow‐up periods. No pattern of increasing or decreasing risk with longer follow‐up is observed across the various updates, with the exception of the SMRs for CNS cancers that increased with longer follow‐up, although none of the SMRs were statistically significant. The SMRs for lymphoid cancers and NHL are similar for the two most recent updates of the UCC study.

TABLE 3.

Standardized Mortality Ratios (SMRs) for key cancer endpoints from different updates of the Union Carbide Corporation (UCC) study.

Cause of death	SMR (95% confidence interval)
	1978	1988	2003	2013
Updates of the UCC study	Greenberg et al. (1990, Table 5)	Teta et al. (1993, Table 1)	Swaen et al. (2009, Table 2 b )	Updated analyses
Lymphoid cancers	NR	NR	0.87 b	0.93
			(0.51–1.40)	(0.60–1.38)
Non‐Hodgkin's lymphoma	0.78	NR	1.05	0.98
	(0.04–3.62)		(0.54–1.83)	(0.55–1.62)
All lymphatic and hematopoietic	1.04	0.59	0.89	1.01
	(0.28–2.66)	(0.24–1.22)	(0.59–1.29)	(0.72–1.37)
Leukemia	1.97 a	1.06 a	0.93	1.24
	(0.38–5.76)	(0.35–2.48)	(0.47–1.67)	(0.76–1.92)
Respiratory cancers (including lung)	0.63	0.90	0.86	0.83
	(0.29–1.19)	(0.66–1.21)	(0.70–1.04)	(0.69–1.00)
Central nervous system	0.67	1.50	1.64	1.67
	(0.03–3.19)	(0.55–3.27)	(0.88–2.81)	(0.95–2.71)

Abbreviation: NR, not reported.

^a Includes “leukemia and aleukemia.”

^bAll values are from Swaen et al. (2009), except for lymphoid, which was calculated using the US general population as the comparison group, consistent with Swaen et al. (2009).

3.2. Exposure‐response modeling

The slopes associated with the standard CPH model fitted to each of the 6 endpoints selected from the two most recent updates of the UCC study are summarized in Table 4. The slope is the change of hazard rate per unit increase in the cumulative exposure.

TABLE 4.

Slopes of estimated cancer‐specific mortality rates with respect to cumulative exposure to ethylene oxide (EO) from the 2003 and 2013 updates of the Union Carbide Corporation (UCC) study.

Endpoint	Slope: MLE (90% confidence interval)
Last year of follow‐up for the UCC cohort	2003 (Swaen et al., 2009)	2013 (Present study)
Lymphoid cancers	−1.66 × 10−5	−1.32 × 10−5
	(−3.52 × 10−5, 1.99 × 10−6) a	(−2.80 × 10−5, 1.64 × 10−6)
Non‐Hodgkin's lymphoma (NHL)	−9.91 × 10−6	−1.18E−5
	(−2.73 × 10−5, 7.53 × 10−6)	(−3.02 × 10−5, 6.62 × 10−6)
All lymphatic and hematopoietic (LH)	−6.79 × 10−6	−6.09 × 10−6
	(−1.71 × 10−5, 3.51 × 10−6)	(−1.49 × 10−5, 2.73 × 10−6)
Leukemia	4.57 × 10−7	−8.06 × 10−7
	(−1.05 × 10−5, 1.15 × 10−5)	(−1.03 × 10−5, 8.70 × 10−6)
Respiratory cancers (including lung)	1.24 × 10−6	1.26 × 10−6
	(−2.1 × 10−6, 4.58 × 10−6)	(−2.08 × 10−6, 4.60 × 10−6)
Central nervous system	−2.6 × 10−5 *	−2.15 × 10−5
	(−5.51 × 10−5, 3.11 × 10−6)	(−4.60 × 10−5, 3.01 × 10−6)

Abbreviations: MLE, Maximum Likelihood Estimate; SE, Standard Error.

^aThe 95% upper confidence limit on the slope of the model = MLE+ (1.645 × SE).

^*Statistically significantly less than zero at the 5% significance level using the likelihood ratio test, but not statistically different using the Wald's confidence interval test (SAS Institute, ²⁰¹³; Valdez‐Flores et al., ²⁰¹⁰, Table S9).

None of the slopes are statistically significantly increased when compared to the null model (i.e., zero slope). The three components of lymphoid tumors (NHL, MM, and LLK) all have non‐statistically significant negative slopes (results for MM and LLK not shown). The only positive slope for respiratory cancer is also not statistically significant. Similar to what was seen by Swaen et al. (2009), the results of the UCC cohort with 10 more years of observation do not indicate increased exposure‐response trends for the cancers of interest. The upper bound estimate on these non‐significant model slopes for the lymphoid category was positive for both Swaen et al. (2009) and the current updated estimate (1.99 × 10⁻⁶ and 1.64 × 10⁻⁶, respectively) (Table 4). Additional post hoc analysis of the UCC cohort indicates that adjusting the CPH model for race and year of birth has no impact (Table S7).

For the UCC study, the standard CPH model with untransformed (linear) cumulative exposure had a better statistical fit (i.e., less deviance) compared to the log cumulative exposures for lymphoid cancers (Table S6). Thus, there is no evidence that the CPH model with log cumulative exposure as the dose metric fits the cancer mortality in better than the model with cumulative exposure metric. Because the coefficient of the slope of lymphoid cancer death rate with log cumulative exposure was negative (Table S8) and not statistically significantly different from zero, no attempt to fit a two‐piece linear spline model was made.

Applying the same methods to the NIOSH males resulted in a non‐statistically significant positive slope of 3.89 × 10⁻⁶ (p = 0.07; CI = 9.78 × 10⁻⁷, 6.80 × 10⁻⁶) for the standard CPH model of lymphoid mortality.⁶ In contrast, the slope for female lymphoid mortality based on the standard CPH model was non‐statistically significantly negative. The standard CPH model with log cumulative exposure had a better statistical fit (i.e., less deviance) compared to the linear exposures for male lymphoid alone (Table S6). Our analyses of the NIOSH study are consistent with Steenland et al. (2004) for male lymphoid cancers. We agree with the Steenland et al. (2004) approach of analyzing males and females separately. However, EPA IRIS (2016) combined the males and female lymphoid cancers in the final risk assessment. When NIOSH male and female lymphoid cancers are combined, the standard CPH model with linear exposures had less deviance than the log cumulative exposure model (Table S6).

3.3. Categorical analyses based on internal comparisons

The lack of an increasing trend based on exposure‐response modeling was further confirmed by an internal categorical analysis, whereby no specific shape of the exposure‐response relationship is assumed. This analysis compares the hazard rate at each quintile of cumulative exposure to the hazard rate in the quintile with the lowest cumulative exposure to EO. In the second quintile only for LH, there was a statistically significant increase for the most recent update of the UCC study. None of the remaining comparisons (lymphoid, NHL, leukemia, respiratory, and CNS) were statistically significantly increased.

For the NIOSH study, there was a statistically significant excess in the highest quintile for lymphoid cancer for males (p = 0.02) but not for any lower quintile and no statistically significant excesses in any quintiles for female lymphoid cancers (p = 0.39–0.91; Table S4).

In NIOSH males, NHL and LH categories also were statistically significantly increased in the highest category. Since NHL is the major component of the lymphoid group (15 of 25 cases) in males and since lymphoid cases are the major component of the larger all LH category (25 of 37 cases), these larger categories of lymphoid and LH reflect the NHL excess in the highest quintile (Table S4). There was no increase in the lower quintiles for lymphoid, NHL, and LH, indicative of a steep exposure‐response pattern at lower cumulative exposures. None of the quintiles in the categorical analyses for leukemia or CNS had a consistent increase in the relative risks that indicate an increasing trend (i.e., positive slope) of cancer mortality with increasing exposure.

4. DISCUSSION

The purpose of this manuscript is to investigate the plausibility of two exposure‐response models applied to NIOSH lymphoid data based on consideration of the epidemiological evidence from the results of the present update of the UCC cohort and the published results of the NIOSH study. As described above in the Introduction, the two exposure models of interest are the standard log‐linear CPH model and a two‐piece linear spline model with an initial steep slope followed by a shallower slope. Slopes for both models are essentially linear at the lower exposure levels, but the slope of the latter model is very steep. This discussion section first considers the consistency of exposure‐response models with the UCC cohort update and the original findings of the NIOSH study. Next, limitations of the UCC cohort that might weaken its suitability for addressing the plausibility of exposure‐response models are considered. Finally, analyses and rationale that conflict with our conclusions regarding model selection are evaluated.

4.1. Plausibility of steep versus shallow exposure‐response models based on epidemiology data

The UCC EO chemical worker cohort has been examined multiple times over the past 40+ years to the point where the vast majority of the population is deceased. In this update, we have conducted both traditional external comparisons and detailed worker‐to‐worker internal comparisons, using standard CPH regression modeling, incorporating continuous and categorical cumulative exposure metrics, as well as continuous CPH modeling with log cumulative exposure metric. The results of the updated UCC cohort with 10 more years of observation do not indicate any statistically significant excess deaths nor exposure‐response trends for any of the male cancers of interest. Exposure lag periods (5–30 years) had no meaningful impact on our analyses and conclusions (Table S5). In addition, race and birth year did not have an impact on the standard CPH model fit (Table S7).

Specifically, the results of our detailed analyses of UCC data in no way support an increased risk, particularly no indication of a steep slope at low exposures for the cancers of interest.

• The UCC non‐positive findings persist across multiple updates of the cohort.

• There are no increasing exposure‐response trends, as the non‐statistically significant slopes were negative (Table 4).

• Neither the cumulative exposure nor the log cumulative exposure metrics resulted in statistically significant slopes (Tables S7 and S8).

• Although the upper‐bound slope on the non‐statistically significant CPH model is positive, it is very small in magnitude and does not indicate a steep exposure‐response pattern at lower exposures.

• In the internal comparisons, none of the quintiles of exposure indicate increasing cancer with increasing exposure (Table S4).

• Incorporating exposure lags in the analysis did not alter these patterns (Table S5).

Our analysis of the NIOSH study data are consistent with the Steenland et al. (2004) conclusion that there was little evidence of any excess cancer mortality for the sterilant workers overall, positive exposure‐response trends for lymphoid tumors were found for males only, and there was sufficient variation in the exposure of women to have observed an exposure‐response if one existed (Steenland et al., 2004). More importantly, the excess in males was seen only at the highest exposure category, with no indication of an excess in the lowest exposure group. This singular positive finding was observed despite numerous external and internal analyses, numerous exposure metrics, and lag periods.

The EPA IRIS exposure‐response model below the cumulative exposure knot of 1600 ppm days is the basis for EPA's slope calculation and ultimate unit risk estimate. It is notable that of the 13 lymphoid deaths below the knot, only 3 are males. The remaining male cases above the knot, from which an excess at highest exposure group was determined, are essentially excluded from EPA's low‐exposure slope analysis.

The EPA IRIS (2016) conclusion that the NIOSH study supports a steep risk for lymphoid cancers at low exposures is based mainly on the pattern of four odds ratios for males and females combined, in and of itself a precarious attempt at understanding the shape of continuous exposure response models (Valdez‐Flores & Sielken, 2013). Despite the obvious differences in exposure‐response pattern of the males and females reported by Steenland et al. (2004), EPA IRIS (2016) combined the data from the two genders. Based on EPA's internal categorical results with a 15‐year lag, males show a dip in the low exposure region with a significant increase at the highest exposure category (Figure 2A). Females, who did not show any statistically significant increased risk for lymphoid cancers, appear to increase in the low region and then decline overall (Figure 2B). When combined with males, the female data in the low exposure region erroneously alters the perception of the slope as a steep exposure response at lower exposure (Figure 2C).

FIGURE 2.

Categorical rate ratios and 90% confidence intervals for National Institute of Occupational Safety and Health (NIOSH) lymphoid cancer with 15‐year lag as reported by EPA IRIS (2016; tab. D‐26 and D‐28). (A) NIOSH male workers. (B) NIOSH female workers. (C) Male and female workers combined. Note: Data are from EPA IRIS (2016) tab. D‐28, which reports categories of exposure as 0 (lagged out), > 0–1200, 1201–3680, 3681–13,500, and > 13,500. Midpoints used for the graph are average ppm‐days for the full cohort (both sexes combined), which were 0, 446, 2143, 7335, and 39,927 ppm‐days, respectively (EPA IRIS, ²⁰¹⁶ tab. D‐26). The actual mean and ranges of exposures for males only and females only were not reported by EPA IRIS (2016) for each category of exposure. EO, ethylene oxide.

An exposure‐response model with a steep positive slope in the low‐exposure region indicates a potent carcinogen. Use of the two‐piece linear spline in the EPA IRIS (2016) assessment ultimately results in an EO URF that is among the highest of all inhalation URFs derived by IRIS. If real, a high risk at low concentrations should be easily detected in the most informative occupational epidemiological studies of exposed individuals.

Based on these two most informative epidemiology studies, there is no evidence that EO is a highly potent carcinogen consistent with a steep increase in risk at low exposures. These conclusions are supported by the following integrated analyses of both studies.

Using external analyses, the ratio of overall observed to expected numbers of LH and lymphoid cancer deaths is not in excess in the UCC cohort. Both studies report no increased overall male mortality for all LH cancers and very similar 95% CIs. UCC's SMR was 1.01 (0.72, 1.37) (Table 2), compared to NIOSH's 1.09 (0.79, 1.47) (Steenland et al., 2004). When the NIOSH SMRs were examined by exposure category, males in the highest exposure group had a statistically significant increase in NHL, the major contributor to the lymphoid tumor group (SMR = 2.37, 10‐year lag) based on eight cases (while females had a deficit SMR = 0.37, 10‐year lag, based on one case) (Steenland et al., 2004) (Table 4). Notably, both males and females exhibit deficits of NHL in the low exposure categories. Steenland et al. (2004) did not present SMRs for lymphoid cancers but reported a non‐statistically significant LH deficit in females based on the SMR of 0.31 in the 10‐year lagged‐out group. A deficit in a lagged‐out group can contribute to an apparent steep increase in internal categorical odds ratio at lower exposures.

In internal analyses of the NIOSH data using categorical cumulative exposure metric, Steenland et al. (2004) reported increased risk for lymphoid and LH cancers for males only, and only in the highest exposure category when including a 15‐year lag period. As described in detail above, there was no indication of increased risk in the UCC data internal analyses in any of the update periods.

The overall assessment of EO epidemiology from the two most informative studies clearly fails to identify a potent carcinogen that would be consistent with selection of a model with a steep exposure‐response pattern at lower exposures.

4.2. Suitability of UCC cohort data

The suitability of the UCC data as an external verification for selection of the exposure‐response model should be discussed to address questions raised in the past about the use of this study to support quantitative cancer risk assessment purposes. Human data are a very well accepted methodology for cancer (QRA) provided the epidemiology study(ies) has adequate quality, size, and exposure data at the individual level (EPA, ²⁰⁰⁵). EPA concluded that only the EO data from the NIOSH studies (Steenland et al., 2003; 2004) met these criteria. This calls into question our use of the UCC data as an external study to inform the selection of different exposure‐response models for the NIOSH study. We have examined the updated UCC study and conclude it is adequate for incorporation into the QRA process to support (or not) the selected IRIS exposure‐response model, particularly the assumed shape of the exposure‐response model based on the NIOSH study. Second, it serves in a weight of the evidence assessment, addressing the level of confidence in the result derived from the IRIS (EPA IRIS, ²⁰¹⁶) selected exposure‐response assessment. These key points are discussed in detail below.

4.2.1. Comparability of sample size and historical exposure estimates

With the extended follow‐up of the UCC cohort, the sample sizes for the cancers of interest are comparable to the male numbers in the NIOSH mortality study, despite the larger overall cohort size of the NIOSH study (Table 1). In fact, the total sample size for UCC males with LH (41) exceeds the number of males available for analysis from the NIOSH study (37). However, the distribution of cases by age differs given the much longer follow‐up of the UCC cohort and higher background rates of lymphoid cancers at older ages. The numbers of cancers of interest are key determinants of study power. Both studies suffer, however, from small sample sizes when the data are categorized by levels of exposure, resulting in risk estimates with wide confidence intervals (Table S4; Table 8).

Uncertainty in exposure estimation is characteristic of retrospective epidemiology studies that have long observation periods. Both studies have exposure estimates at the individual level that vary over time and are applicable to exposure‐response analysis. The UCC exposure assessment is of comparable reliability to that of the NIOSH study, which had no exposure data prior to 1978. Between 1978 and 1986, the NIOSH exposure assessment was based on a rigorously validated statistical regression model based on personal exposure monitoring measures from 18 plants (Hornung et al., 1994). However, the predictors of exposure from 1978 to 1986 were then used to estimate exposures prior to 1978, when no data were available. Surprisingly, this NIOSH model predicted exposures by job increasing from the 1930s to higher exposures in the 1970s (Bogen et al., 2019). This trend of increasing exposures from earlier to later periods is not consistent with the poorer engineering controls, decreased ventilation, poorer door seals, fewer air washes, combined sterilization operation and sterilized product storage, and very high residue levels of EO on sterilized products in earlier years compared to later years that indicate exposures were likely higher in the 1940s–1960s than in the 1970s (Bogen et al., 2019; Steenland et al., ¹⁹⁹¹). In contrast, the UCC study incorporated actual data from 1957 to 1988 from UCC production plants, followed by reality checks with other contemporaneous exposure levels.

Thus, the UCC study has comparable, if not superior, historical exposure data before 1978, strengthening the use of this study to inform the plausibility of exposure‐response models that have been applied to the NIOSH study.

4.2.2. Quality issues (bias and exposure misclassification)

UCC chemical workers had the potential for multiple chemical exposures in addition to EO, raising the issue of confounding bias. In 1990, Greenberg et al. (1990) reported an excess of leukemia and pancreatic cancer isolated to the EC production unit, a chemical that was needed as the raw material for EO production by the chlorohydrin process in another unit. The EC unit was included in the UCC EO cohort because on occasion EO was used to produce needed by‐products associated with EC production. In a 10‐year update of this cohort with follow‐up through 1988, the 278 EC workers were followed up by Benson and Teta (1993), while the 1896 EO workers in the remaining units were examined by Teta et al. (1993), without the confounding exposures from the EC unit. EC workers were reported to have the same excesses seen by Greenberg et al. (1990) (pancreatic cancer and leukemia), while the EO workers exclusive of these 278 workers showed no meaningful excesses. Positive confounding is no longer an issue for the UCC workers, given the exclusion of the EC workers and the absence of any other positive findings associated with EO exposures in this cohort. Other chemicals cannot explain the absence of non‐positive SMRs, that is, mask a positive finding unless they are protective of disease, which is highly unlikely. Confounding can possibly mask a positive trend in exposure‐response analyses, although there is no known chemical cause of lymphoid tumors in the occupational setting.

The UCC study update failed to observe increased numbers of LH cancers based on external comparisons. The absence of significant SMRs in the UCC study has been questioned based on concern that a healthy worker effect (HWE) bias may be masking an increased risk. However, the epidemiologic literature has shown that the HWE is primarily a concern in occupational cohort studies with short follow‐up and is more serious for diseases that have considerable symptoms at hire than for diseases without pre‐diagnosis symptomatology, such as cancers that develop over a longer time period (Fox & Collier, 1976; Monson, ¹⁹⁸⁶; Pearce et al., ²⁰⁰⁷;  International Agency for Research on Cancer [IARC], 2012). The average follow‐up for the NIOSH study is over 20 years and over 40 years for the UCC study. Steenland et al. (2004, p. 6) concluded in the mortality study that “the HWE would seem an unlikely explanation for the lack of cancer excesses in the exposed versus non‐exposed comparisons.” Thus, plausibility for selection of exposure‐response models for the NIOSH study should consider both internal and external (such as SIRs and SMRs) analyses of epidemiology studies.

We did observe deficits in major non‐cancer causes (e.g., heart and respiratory diseases) that may reflect lower prevalence of smoking in the UCC workers. This is consistent with the absence of excesses in smoking‐related cancers of the respiratory system in this study population.

Criticisms of relying on previous UCC study updates as a basis for informing cancer risk stemmed from concerns that follow‐up was beyond the cancer latency period and possible effects would be attenuated by higher background rates of cancer in this elderly population (EPA IRIS, ²⁰¹⁶). However, the workers have been examined at various lengths of follow‐up, and EO‐related effects did not appear at either shorter or longer observation periods (Table 3). In addition, no decreased risks with additional extensive follow‐up of the UCC cohort were seen, as might be expected with a Healthy Worker Survivor Effect bias.

Study result differences between the UCC and NIOSH studies’ results are evident in the internal exposure‐response analyses. Steenland et al. (2004) suggested an increase in the highest exposure category for male lymphoid tumors compared to the lowest and a statistically significant slope using log cumulative exposure in the continuous model, while the UCC study observed no comparative differences. The number of lymphoid cancer deaths in the highest NIOSH quintile was five among 247 workers at risk, versus five among 534 UCC workers at risk in their highest exposure category (Table S4). Clearly, the difference in results cannot be attributed to sample size. The NIOSH lymphoid tumor analyses were based on 27 male cases compared to 25 cases in the UCC updated cohort.

There is a striking difference in the findings of these two studies that cannot be attributed to sample size, inadequate worker exposures, or confounding. In addition, post hoc analysis controlling for race and year of birth similar to methods used by NIOSH had no impact on the UCC outcome. This calls into question the plausibility of the very high IRIS URF for lymphoid cancers using a two‐piece linear spline model that places EO in the category of a potent carcinogen.

4.3. Consideration of EPA's rationale supporting a steep continuous exposure response model for lymphoid cancers

In the NIOSH cohort, the only exposure‐response model reported to have statistically significant positive slopes for LH and lymphoid cancer mortality in males is the log cumulative exposure model with a 15‐year lag (EPA IRIS, ²⁰¹⁶; Steenland et al., ²⁰⁰⁴). The statistical significance of this supralinear log cumulative exposure model is a key reason EPA IRIS (2016) gives for selecting a two‐piece linear spline model with an initial steep slope. EPA IRIS (2016, pp. 4–10) states that the standard log‐linear CPH model with cumulative exposure “does not reflect the apparent supralinearity of the data demonstrated by the categorical results and the superior fit of the log cumulative exposure model.” It is noteworthy that the EPA IRIS (2016, pp.4–10) ultimately rejected the log‐cumulative model to estimate the risks for lymphoid from the NIOSH study because this model “is inherently supralinear (i.e., risk increases steeply with increasing exposures in the low exposure range and then plateaus), with the slope approaching infinity as exposures decrease toward zero.” In other words, this statistically significant model is implausible because it is too steep, yet no a priori definition of “too steep” was provided.

In considering the statistical significance of the log cumulative model for male lymphoid cancers, it is important to keep in perspective that a large number of curve‐fitting models were tested. Steenland et al. (2004) initially conducted CPH modeling using a variety of exposure metrics (duration, average, maximum, cumulative, and log cumulative exposure) and both continuous and categorical exposure variables. For each of these combinations, a variety of lags were also examined (no lag, 5, 10, 15, and 20). Only the log cumulative exposure model with 15‐year lag in males was reported to be statistically significant for lymphoid cancer (p = 0.02) (Steenland et al., 2004). This analysis is not based on any a priori hypothesis regarding biological mode of action, epidemiological evidence, or toxicokinetics. In addition, this model is not statistically significant (p = 0.07) if the search for the lag is accounted for in the statistical analysis.

Another rationale for applying a supralinear model is based on an apparent “visual fit” of a few non‐parametric categorical rate ratios (i.e., odds ratios) to assume the shape of continuous models based on individual data. Visual fit comparisons are subjective and depend on how graphs are presented. Visual fit comparisons can be especially misleading if the data are normalized along the y‐axis so that the true relative position of the y‐intercept (baseline hazard rate) is obscured. This issue was identified by EPA IRIS (2016, pp. 4–21; Figures 4 and 3) as a footnote in the figure legend that states, “Note that, with the exception of the categorical results…the different models have different implicitly estimated baseline risks; thus, they are not strictly comparable to each other in terms of RR values, that is, along the y‐axis.” This issue was only corrected for by TCEQ (2020; Appendix 6) by approximating the individual data (53 cases) that are modeled and adjusting the CPH model along the y‐axis to account for the differences in the baseline hazard rate implied by different models (Figure 3). The true range of the cumulative exposures of the cohort was also more accurately represented (Figure 3). Inspection of Figure 3 indicate neither model has a superior visual fit to the other, and visual fit should not be a primary basis for determining model fit.

FIGURE 3.

Models adjusted for standard log‐linear Cox proportional hazard (CPH) model background hazard rate. Note: Visual representation of the NIOSH lymphoid individual rate ratios compared to three different models from EPA IRIS (2016), all adjusted for the standard CPH model baseline hazard for improved comparison along the Y‐axis (adapted from TCEQ, 2020, fig. 15). This adjustment is based on an approximation of the difference in y‐intercept between the standard CPH model and the rate ratios, for purposes of addressing EPA IRIS (2016, pp. 4–21) Figures 4 and 3 footnote that states, “the different models have different implicitly estimated baseline risks; thus, they are not strictly comparable to each other in terms of RR values, that is, along the y‐axis.” The individual rate ratios are categorical ratios in which each category was defined to have one lymphoid mortality. See Supplemental Materials Figure 3 Calculations for details.

An objective method to ground truth the model is to compare model predictions with the observed number of deaths overall and within exposure subcategories, as was described in detail by TCEQ (2020) (Appendix 3). Using this approach, the standard CPH model and the EPA IRIS (2016) two‐piece linear spline model for lymphoid mortality were used to predict the number of lymphoid cancer deaths in the NIOSH study. TCEQ (2020) found that while their standard CPH model predicted 52.42 deaths (95% CI: 40.1–70.0), EPA's model statistically significantly overestimated with 91.69 deaths (95% CI: 70.1–122.4) compared to what was observed in the NIOSH study (53 deaths)⁷. Importantly, this was also consistently the case at the lower exposure levels, indicating that the standard CPH model is accurately characterizing the overall and lower cumulative exposure levels. Although HWE is unlikely to impact these analyses, as discussed in detail above, these results did not change even if a theoretical HWE of 15%–16% was introduced in the calculation of the expected number of lymphoid deaths (TCEQ, ²⁰²⁰).

5. SUMMARY AND CONCLUSIONS

In summary, we have compared our results for LH and lymphoid cancers from the UCC cohort with past updates and with the findings of the NIOSH sterilant worker cohort and considered how these data inform the plausibility of exposure‐response models. None of the UCC cohort findings, including a log cumulative exposure analysis, indicate a steep positive slope at low exposures (i.e., a highly potent inhalation carcinogen). In fact, there were no positive slopes for these endpoints, contributing further to the weight of evidence based on epidemiologic evidence that the selection of a steep positive exposure‐response model is implausible.

This updated UCC study is informative for the purpose of evaluating the plausibility of the exposure‐response model for the most critical endpoint for cancer risk based on consistency with the epidemiological evidence. The absence of increased risks in the UCC study (in both external and internal analyses across several updated analyses) and a statistically significant increase limited to males in the highest exposure group in the published NIOSH study call into question the plausibility of the two‐piece linear spline model with a steep supralinear exposure‐response pattern applied in the EPA IRIS (2016) assessment.

The standard CPH model is well accepted by epidemiologists in cancer exposure‐response analysis, essentially linear at the observed exposures levels of the studies, and consistent with an assumption of no‐threshold that reflects the epidemiological weight of evidence. The classical epidemiological approach to identifying a carcinogen is to look for an exposure‐response trend that gradually increases from low to high exposures. Even though EO is a genotoxic agent and is thus assumed by default to operate by a linear mode of action (MOA), human toxicological defense mechanisms nonetheless potentially attenuate responses at very low exposures (Gollapudi et al., 2020; Kirman et al., ²⁰²¹). The standard CPH model follows this MOA concept. Its use in exposure‐response assessment is further supported in the case of EO on the strength of its predictability. The standard CPH model with untransformed cumulative exposure, therefore, is preferred for purposes of EO exposure‐response using epidemiological data based on consideration of the findings from both the NIOSH and the UCC studies.

CONFLICTS OF INTEREST STATEMENT

At the time the study was conducted, Dr. Bender was an employee of The Dow Chemical Company and received a salary and shares of stock in the company. UCC is a wholly owned subsidiary of The Dow Chemical Company. After exiting Dow, Dr. Bender served as a consultant for the American Chemistry Council. Dr. Bender's analysis was limited to the standard mortality rates of the UCC update in Table 2. Dr. Valdez‐Flores and scientists at Exponent Inc. are scientific consultants for the American Chemistry Council, which funded the study. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in any decisions regarding publication of this manuscript.

Supporting information

Table S1: Distribution of the cumulative EO exposures (ppm‐days) in the current UCC update (bold), previous UCC update and NIOSH study

Table S2: Exposure Assessment Matrix for the UCC Cohort, Kanawha Valley (KV), West Virginia

Table S3: The ICD codes corresponding to the different revisions used in the analyses of the endpoints herein

Table S4: Rate ratio analyses and Cox proportional hazards model for cumulative exposure for each combination of endpoint, sex, and study

Table S5: Statistical impact of exposure lag periods (5‐30) on the zero‐lag model fit for UCC male lymphoid mortality

Table S6: Comparison of standard CPH models using linear vs. log‐transformed cumulative exposure in the UCC cohort update through 2013

Table S7: Slopes of estimated cancer‐specific mortality rates with respect to cumulative exposure to EO with models adjusted for race and year of birth based on UCC cohort

Table S8: Slopes of estimated cancer‐specific mortality rates with respect to logarithm cumulative exposure to EO for lymphoid mortality in the UCC cohort update through 2013.

Figure 3 Calculations

Valdez‐Flores, Ciriaco. , Li, Abby A. , Bender, Thomas J. , & Teta, M Jane. (2025). Use of updated mortality study of ethylene oxide manufacturing workers to inform cancer risk assessment. Risk Analysis, 45, 2822–2837. 10.1111/risa.70057

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.