Hexavalent chromium and stomach cancer: a systematic review and meta-analysis

Abstract

Hexavalent chromium [Cr(VI)] is known to cause lung cancer in workers of certain industries, but an association with stomach cancer is uncertain and widely debated. Systematic review and meta-analyses were conducted to assess the risk of stomach cancer mortality/morbidity in humans and experimental animals exposed to Cr(VI). In accordance with the protocol (PROSPERO #CRD4201605162), searches in PubMed and Embase^®, and reviews of secondary literature bibliographies, were used to identify eligible studies. Critical appraisal of internal validity and qualitative integration were carried out using the National Toxicology Program’s Office of Health Assessment and Translation (OHAT) approach; meta-analyses were conducted based on the occupational data (the only data suitable for quantitative assessment). Forty-seven publications (3 animal, 44 occupational, 0 non-occupational) met the eligibility criteria. Stomach cancer was only observed in one high risk of bias animal study, and in the low risk of bias studies no stomach cancer was observed. Thus, confidence in this evidence base is high. Environmental epidemiology studies did not meet eligibility criteria because exposure and outcome were not measured at the individual level. Meta-analyses of human data resulted in overall meta relative risks of 1.08 (95% CI: 0.96–1.21) including all studies and 1.03 (95%CI: 0.84–1.26) excluding studies associated with the highest risk of bias. Because most occupational studies have high risk of bias for confounding and exposure domains, the overall confidence in this evidence base is low to moderate. Combining the streams of evidence per the OHAT approach, Cr(VI) does not pose a stomach cancer hazard in humans.

Keywords:

• Hexavalent chromium
• stomach cancer
• systematic review
• meta-analysis
• bias

Introduction

The risk of cancer associated with occupational exposure to hexavalent chromium [Cr(VI)] has been studied for over 100 years, in hundreds of studies, from a wide spectrum of industries (IARC 1990, 2012; NIOSH 2013). Among workers in certain industries, such as chromate production, pigment production, and chrome plating, a significant increase in lung cancer risk has long been recognized (IARC 1990; OSHA 2006; ATSDR 2012; NIOSH 2013; Proctor et al. 2014). Cr(VI) is classified as a known human carcinogen by the International Agency for Research on Cancer (IARC). IARC’s conclusion is based on sufficient evidence in humans that Cr(VI) compounds cause cancer of the lung, and positive associations observed with cancer of the nose and nasal sinuses (IARC 2012). IARC indicated, “There is little evidence that exposure to chromium(VI) causes stomach or other cancers” (IARC 2012). However, recent reviews and meta-analyses report conflicting findings. Two of those meta-analyses found no evidence of an association between stomach cancer and Cr(VI) exposure (Cole and Rodu 2005; Gatto et al. 2010), but a more recent meta-analysis reported a significantly greater risk of stomach cancer in Cr(VI)-exposed workers (Welling et al. 2015). Each meta-analysis used different inclusion criteria, and only Cole and Rodu (2005) considered socioeconomic status (SES) and study quality. Interestingly, Cole and Rodu (2005) reported a significantly decreased risk of stomach cancer in studies that adjusted for SES (Meta-SMR: 82, CI: 69–96), but a significantly increased risk in studies that did not consider differences in SES (Meta-SMR: 137, CI: 123–153).

Although Cr(VI) is rapidly absorbed into cells, before that absorption occurs, it may be converted through extracellular reduction to the trivalent form [Cr(III)] in biological fluids and tissues; notably, gastric fluid, blood, and liver have significant reducing capacity (De Flora et al. 1997). Because Cr(III) is not well absorbed and has not been shown to be carcinogenic, and perhaps for this reason, the potential for carcinogenicity due to Cr(VI) outside the lung is limited. The question of whether reduction of Cr(VI) to Cr(III) in stomach contents is sufficient to protect against carcinogenicity has been studied for several decades (Donaldson and Barreras 1966; De Flora and Boido 1980; Finley et al. 1997) and has been a controversial issue considered in regulatory guidelines and several reviews (U.S. EPA 1991; De Flora 2000; Proctor et al. 2002; Sedman et al. 2006; OEHHA 2011; De Flora et al. 2016). Still, stomach cancer in association with Cr(VI) exposure has long been suspected among highly exposed workers, because oral exposures may occur through swallowing particles and hand-to-mouth contact. Observations from historical industry exposures, in highly contaminated work environments, have included reports of gastritis, ulcers, and stomach upset among workers of the chromium chemical production industry (Mancuso 1951; PHS 1953). Further, historical accounts of working conditions include evidence of oral exposures based on appearance of yellow-stained teeth and tongues (PHS 1953). Although not observed consistently, some occupational studies have reported a significant increase in stomach cancer risk among subcohorts of Cr(VI)-exposed worker populations (e.g. Davies et al. 1991; Korallus et al. 1993). Thus, occupational data are thought to provide important evidence for evaluating the risk of stomach cancer among populations exposed to Cr(VI).

In addition to evidence from occupational epidemiology studies, data from animal toxicology studies contribute insight into the overall assessment of hazard and risk. Typical of animal toxicology studies, the administered doses of Cr(VI) in these studies far exceed potential human exposures. Nonetheless, animal studies offer an evidence stream traditionally used for both hazard identification and risk assessment. For the purpose of this review, important characteristics of animal data include low risk of bias in exposure characterization, because exposure occurred by oral administration with relatively precise measures of individual dose, as compared to the human data. The animal study data also include conflicting and somewhat controversial findings because tumors of the stomach have only been reported in the forestomach of mice, a structure of the stomach that humans lack.

Finally, several studies of human populations with environmental exposures to Cr(VI) have been published, and conflicting results exist in these studies as well. Most notable among these literature are three studies of a population of villagers in China exposed to Cr(VI) in drinking water. The original publication (Zhang and Li 1987) was followed-up by two subsequent analyses of the original data that reached different conclusions (Beaumont et al. 2008; Kerger et al. 2009). While Beaumont et al. (2008) reported an increase in stomach cancer among villagers exposed to Cr(VI) in drinking water, Kerger et al. (2009) found no increase in stomach cancer among the same villager population when the comparison rates were from villages with uncontaminated drinking water. Kerger et al. (2009) concluded that differences in risk factors and demographics of the industrialized area, used as the source of control rates in the Beaumont et al. (2008) study, influenced the findings.

To date, no review or meta-analysis of Cr(VI) has utilized a structured guidance for systematic review of the primary studies, such as those of the National Toxicology Program (NTP) Office of Health Assessment and Translation (OHAT) and the Cochrane Handbook for Systematic Review of Interventions (The Cochrane Collaboration 2011; Rooney et al. 2014; NTP OHAT 2015). Given these methodological considerations, the inconsistent results, conclusions, and interpretations of the previously published reviews suggest that the literature on the association between Cr(VI) exposure and stomach cancer risk needs to be reexamined using methods consistent with the current state of the science. Thus, the overall objective of this research is to conduct a systematic review and meta-analysis to determine whether there is a significantly increased risk of stomach cancer mortality and morbidity associated with Cr(VI) exposures in humans or experimental animals using current methods. To that end, the NTP OHAT (2015) Handbook for Conducting Systematic Reviews was used to conduct a systematic review of human and animal evidence, and NTP OHAT criteria were used to critically appraise the risk of bias in each study. A meta-analysis of occupational exposure literature was then conducted to quantify the risk of stomach cancer among worker populations (the only data suitable for such quantitative assessment).

Materials and methods

Protocol development

A multidisciplinary research team was assembled with expertise and experience consistent with standards for conducting a systematic review (Eden et al. 2011). The research team included two subject-matter experts (MS and DP), a systematic review expert (DW), and a scientific advisory board (SAB) with expertise in epidemiology and meta-analysis (LL and MG). Please see tables in the acknowledgments and declaration of interest for roles of each collaborator and author, including the SAB members.

This systematic review was conducted in accordance with the registered protocol (PROSPERO CRD42016051625). Elements of the protocol include the research question, literature search syntax and strategy, types of studies to be included, specifications of the population, exposure, comparator, and outcome (PECO), details of the risk-of-bias assessment and strategy for data synthesis, meta-analysis of subgroups or subsets, and conflict-of-interest information for each research team member. As described in the registered protocol, this systematic review focused on studies of the following populations and exposure conditions: (1) workers with occupational inhalation or ingestion exposure to Cr(VI), (2) non-occupational populations with ingestion exposure to Cr(VI), and (3) experimental animals with ingestion exposure to Cr(VI). The comparator groups included: (1) workers with no or low occupational exposure to Cr(VI) per specifications by the authors of the primary studies, (2) non-occupational populations identified by the authors of the primary studies having no ingestion exposure to Cr(VI), and (3) control animals that were not orally exposed to Cr(VI). The outcome of interest for the human data was stomach cancer morbidity and/or mortality as reported by the primary study authors. For animal data, we evaluated the number of incident stomach cancers per dose or experimental group, patterns of statistical significance, and dose-response as reported by the authors of the primary studies.

Literature search

The literature search was conducted using two databases (Embase^® and PubMed), with the search syntax specified in the protocol (PROSPERO CRD42016051625). The literature search syntax was developed via multiple pilot exercises involving a series of validation checks and several iterations and considered all search terms used in the previous meta-analyses (Cole and Rodu 2005; Gatto et al. 2010; Welling et al. 2015). The final literature search was executed by an information specialist (SF) on March 20, 2018. All literature search results were de-duplicated and uploaded into the DistillerSR software for subsequent screening. Hand searching was the primary method used for identification of animal studies and supplemental for human studies. These included reviews of the citations in previously published meta-analyses and government agency documents on Cr(VI), such as those of the National Institute for Occupational Safety and Health (NIOSH), Agency for Toxic Substances and Disease Registry (ATSDR), IARC, California Office of Environmental Health Hazard Assessment (OEHHA), and USEPA (IARC 1990; USEPA 2010; OEHHA 2011; ATSDR 2012; IARC 2012; NIOSH 2013). A final validation review of the search findings was conducted by the subject-matter experts (MS and DP) prior to the initiation of study screening.

Study screening and selection

Table 1 presents the study’s inclusion and exclusion criteria. Studies in languages other than English were included if they met the inclusion criteria. If multiple studies were published on the same cohort population, the most recent/updated results were included (Langard 1990; Lipworth et al. 2011; Gibb et al. 2015; Proctor et al. 2016). Proportionate mortality ratio (PMR) and proportionate cancer mortality ratio (PCMR) studies were excluded—it is widely acknowledged that these types of studies are difficult to interpret, because they cannot measure risks or rates due to lacking population denominator data (Aschengrau and Seage 2003; Rothman et al. 2008; Guha et al. 2010). Additionally, studies that did not assess risk at the individual level (i.e. ecologic studies) were excluded. For completeness, a brief characterization of the excluded PMR and ecologic studies is provided in Supplemental Tables A and B.

Table 1. Inclusion criteria.

Category	Details
For inclusion	Experimental animals with chronic administration of Cr(VI) by drinking water or oral intake
	Humans with non-occupational (environmental) exposures to Cr(VI) via ingestion
	Humans with occupational exposures to Cr(VI) Workers in chromate production, stainless-steel welding, chrome pigment production, chrome plating/electroplating, ferrochrome production industries Leather tanners included if the authors of primary studies specifically indicate exposure to Cr(VI) or process such as “two bath” process Cement workers included if authors of primary studies indicated that the occupation involved cement production because Cr(VI) is known to potentially produced in the kiln
	Other occupations (not specified above) will be included if the authors of primary studies indicated that workers were exposed to Cr(VI). This distinction was made because of evidence of increased cancer risk from Cr(VI) exposure has not been specifically reported for these other occupations generally, and other risk factors including asbestos, solvent, and arsenic exposures, have been indicated.
For exclusion	Workers with no occupational Cr(VI) exposure. Individuals with no non-occupational (environmental) Cr(VI) exposure. Human occupational and non-occupational studies, where exposure to Cr(VI) is not specifically evaluated nor stated. Experimental animals without chronic ingestion of Cr(VI).
	Studies that do not provide quantitative data of stomach cancer risk morbidity or mortality, or risks cannot be reasonably calculated based on the information provided by the authors of the primary studies
	Proportionate mortality ratio (PMR) and proportionate cancer mortality ratio (PCMR) studies Studies that do not assess risk at the individual level (e.g. registry studies based on occupational titles, ecological studies)

Title and abstract screening was conducted independently by SF. An iterative process was ultimately employed for screening titles and abstracts along with full text, because many epidemiologic studies of Cr(VI) are focused on lung cancer, and titles and abstracts often presented information for this cancer type and not stomach cancer. As a critical additional step, MS and SF hand searched reference citations from published meta-analyses and government agency reports and publications issued after these agency reports (IARC 1990; Cole and Rodu 2005; Gatto et al. 2010; U.S. EPA 2010; OEHHA 2011; ATSDR 2012; IARC 2012; NIOSH 2013; Welling et al. 2015) to identify studies missed by the electronic search and title and abstract screening. Seven studies were included based on hand searching and full text review (Moulin et al. 1990, 1992; Moulin, Wild, Haguenoer, et al. 1993; Moulin et al. 1995; Sorahan and Harrington 2000; Smailyte et al. 2004; Proctor et al. 2016).

Data extraction

Data extractions were performed independently by three reviewers (see tables in the acknowledgments and declaration of intent) using Microsoft Excel for Mac (2018 Version). One Japanese study (Itoh et al. 1996) required the use of a translation service (Honyaku USA, Inc., Torrance, CA). MS and DP did not assess Proctor et al. (2016) because they are authors of that study. Three studies in French (Moulin et al. 1992; Moulin, Wild, Toamain, et al. 1993; Moulin et al. 1995) were translated to English by a team member. For each study, information was extracted on population/animal descriptions and size, study design, exposure assessment and dosing methods, and results, including the number of stomach cancer cases; information was also extracted for relative risk (RR) estimates, including SMR, standardized incidence ratio (SIR), risk ratio, or odds ratio (OR). For experimental animal studies, stomach cancer numbers and incidence rates were included. Information on dose-response and/or statistical significance, as discussed by the primary study authors, was also included along with discussions by the authors of the primary studies regarding confounding factors and limitations.

Some occupational studies reported multiple stomach cancer RR estimates for workers in separate jobs or in different plants; in those instances, estimates were extracted for each category when applicable. In other studies, stomach cancer estimates were provided for the overall worker population, as well as for several sub-cohorts (e.g. Sorahan et al. 1987; Davies et al. 1991; Becker 1999). Per specifications of the registered protocol (PROSPERO CRD42016051625), stomach cancer estimates for the different occupations and plants were extracted. In cases where qualitative exposure categories were provided, the RR estimate for high-exposure categories noted by the primary study authors was extracted, in addition to the RR estimate for the entire cohort population. Therefore, multiple RR estimates for non-overlapping workers from a single study were incorporated when applicable. When warranted, Open Source Statistics for Public Health was used to calculate mortality risk ratios (available at: http://www.openepi.com/SMR/SMR.htm). Quality check was performed to ensure that the details extracted (including the RR estimates) were accurate, and that information for overlapping workers was not included. For one pair of follow-up studies (Korallus et al. 1993; Birk et al. 2006), the cohorts partially overlapped. Of the 901 workers in the cohort of Birk et al. (2006), 678 were the same as those in the Korallus et al. (1993) study, which had a total of 1417 workers. Birk et al. (2006) examined only workers hired after a change in the chromate production processes in the same two plants studied by Korallus et al., but also included new workers with hire dates after the inclusion cutoff in the Korallus et al. (1993) study. The findings of Korallus et al. (1993) were used in the meta-analysis, because the cohort is larger; however, in a sensitivity analysis, the results reported by Birk et al. (2006) were substituted for those of Korallus et al. (1993).

Critical appraisal via risk-of-bias assessment (internal validity)

After data extraction, three reviewers performed a risk-of-bias assessment (see tables in the acknowledgments and declaration of intent). NTP OHAT recommends assessing risk of bias by considering various methodological aspects relevant to specific study design to “address the extent to which results of included studies should be relied on” (NTP OHAT 2015). Risk of bias was assessed according to the NTP OHAT Risk of Bias Rating Tool for Human and Animal Studies (NTP OHAT 2015).

The NTP OHAT approach includes 11 risk-of-bias questions or domains for evaluating internal validity, and each question is applicable to one to six study design types (Supplemental Table C) (NTP OHAT 2015). NTP OHAT recommends that each study be given the following ratings¹ for each question: “− −” (definitely high risk of bias), “−” (probably high risk of bias), “+” (probably low risk of bias), or “+ +” (definitely low risk of bias). The specific instructions for scoring detailed in the NTP OHAT Risk of Bias Rating Tool (NTP OHAT 2015) were used to determine the ratings of the individual question components for all studies that were evaluated. Topic-specific refinements and/or interpretations (also recommended as part of using the tool) included:

Question 3: comparison groups for human studies—As outlined in the PECO, the comparison groups were the general populations or workers in the plant who had limited or no exposures to Cr(VI); these comparator groups were specified by the authors of the primary studies. In occupational studies, the primary study authors used industrial hygiene data with airborne concentrations measured/estimated, and/or occupation and job titles, to identify the control or low-exposure workers. Those in administrative duties (e.g. secretaries, office clerks) were commonly indicated as the control or low-exposure groups. In studies with standardized mortality or incidence ratios, the comparator group was the standard population with age- and sex-specific mortality or incidence rates. In non-occupational studies, the comparator groups were populations indicated by the authors of the primary studies as being minimally exposed to Cr(VI) (i.e. not in proximity to the exposure source).

Question 4: confounding and modifying variables—We identified asbestos, smoking, and SES as variables to be assessed to produce low risk of bias. It is well recognized (by authors of the primary studies) that asbestos and smoking can affect the measured associations between Cr(VI) exposure and stomach cancer. Increased risk of stomach cancer from smoking was reported to be significant in one systematic review and meta-analysis of cohort studies (Ladeiras-Lopes et al. 2008). The Institute of Medicine Committee on Asbestos evaluated occupational cohort and case-control studies and observed some evidence of dose-dependence and consistent pattern of fairly moderate increases risk of stomach cancer from asbestos exposure; however, causal inference between asbestos and stomach cancer was indicated as suggestive but not sufficient (IOM 2006). As noted earlier, differences in results among studies with SES control were found in the meta-analysis of Cole and Rodu (2005).

Question 11: other threats to internal validity—This question was developed to represent potential bias associated with experimental conditions of animals that threatened the validity of the study but were not covered in other domains (i.e. Q5—identical housing and husbandry practices and Q7—attrition). In this evidence base, Borneff et al. (1968) and Mackenzie et al. (1958) reported infections (e.g. mousepox, respiratory) and in Borneff et al. (1968), cannibalism that resulted in early mortality for the experimental animals. As such, Q11 herein addressed the health of the animals from the context of potentially biasing response due to biological differences associated with sickness (vs. differences in conditions between control/exposed groups or accounting for attrition from early mortality due to diseases; Q5 and 7, respectively).

Quality assurance was conducted with overlapping reviews for a subset of studies (i.e. a second reviewer conducted data extraction and risk-of-bias assessment). These overlapping evaluations were mostly consistent. For human occupational data, a few differences for Question 8 (exposure characterization) were resolved by MS.

Following critical appraisal of internal validity, we used NTP OHAT’s tier system to characterize the overall risk of bias for each study as a way of comparing the internal validity across the evidence base. Tier 1 studies were determined to have definitely low or probably low risk of bias, and Tier 3 studies had definitely high or probably high risk of bias (NTP OHAT 2015). Tier 2 studies were those that met neither the criteria for Tier 1 nor the criteria for Tier 3. Although NTP OHAT recommendations propose these classifications, they do not provide specific guidance on how the tiers should be assigned. We developed two approaches for tier classification. Both approaches were used for tiering human data; for animal data, only the second approach was used. Approach 1 placed emphasis on the ratings for the key questions (Question 4, 8, and 9) as identified by NTP OHAT. For a study to be classified as Tier 1, there had to be no “− −” or “−” ratings in any key question and no “− −” rating in other questions. If a study had two negative ratings for the key questions (“− − ”or “−”), it was classified as a Tier 3 study. Question 4 was not evaluated for experimental animal studies, and thus this approach was not considered for tiering the animal data.

Approach 2 considered the overall average ratings across all questions, and thus a numerical value must be given to each score to calculate an average. If a particular question was assigned a “− −” (definitely high risk of bias) on the basis of the OHAT tool, it received a −2, “−” (probably high risk of bias) was assigned a −1, and “+” (probably low risk of bias) and “+ +” (definitely low risk of bias) received 1 and 2, respectively. The numbers across all questions were summed and a mean rating per study was used to assign each study to one of the three tiers using the cutoffs of >0.7, from 0.7 to −0.6, and less than −0.6 for Tiers 1–3, respectively. These cutoffs were based on division by three of the range of possible ratings (−2 to 2) for each question. Unlike Approach 1, Approach 2 equally weights each domain in the assignment of studies to a tier. Per NTP OHAT guidance regarding weighting of key domains, Approach 1 was the primary approach to determine study tiers used in the meta-analyses described below. Approach 2 was included as an additional measure of comprehensiveness and transparency. NTP OHAT indicated that Tier 3 studies could be excluded due to high concern about bias on the key elements such as exposure assessment, outcome assessment, and confounding/selection; they may be “too problematic to provide any useful evidence and should not be included in any synthesis” (Sterne et al. 2014; NTP OHAT 2015). As such, Tier 3 studies were not included in several meta-analyses.

Meta-analysis

Meta-analyses were conducted based on human data. All meta-analyses were conducted by LM with oversight of CR and MS. Meta-analysis was conducted using the “metafor” package in R (R Core Team 2016). Meta-RR and 95% confidence intervals (95% CI) were calculated using random-effects models. Weights of the studies contributing to the meta-RRs were also determined in R for each analysis. Forest plots were developed using R.

A funnel plot was developed in R to visualize whether results were symmetrical. Meta-analyses are based on published studies, and it is known that studies with null results are less likely to be published (Thornton and Lee 2000; Dwan et al. 2013). To assess the possibility of publication bias, a Trim and Fill test (Duval and Tweedie 2000) and Egger’s regression test (Egger et al. 1997) for funnel plot asymmetry were performed. The Trim and Fill test was essentially the results of an alternative random-effects meta-analysis model, with studies that made the funnel plot asymmetrical trimmed, and new “studies” added to make the funnel plot symmetrical. The meta-RR from a Trim and Fill test should not be interpreted as a more valid estimate, but if it is markedly different from the original meta-RR, it may suggest that publication bias could be influencing the results. With the Egger's test, we examined the relationship between normalized study outcomes (effect size/standard error) and study precision (inverse of study standard error) where a significant regression coefficient indicates asymmetry in the funnel plot, which in turn implies publication bias.

Meta-analyses were conducted as outlined in the protocol (PROSPERO CRD42016051625). They included evaluation of the overall evidence stream (Analysis 1; n = 44, no studies excluded) as well as a series of analyses (Analysis 2 through 8) based on study type, industry type, exposure category, and study quality (as measured by risk of bias) (Table 2).

Table 2. Meta-analysis descriptions.

Analysis	Description
1	All Studies
2*	From Analysis 1, eliminate Tier 3 studies with -2 or -1 ratings in the key questions (Approach 1 for tier classification)^a^,c
3*	Tier 1 studies only – Based on overall risk of bias ratings (Approach 2 for tier classification) ^b,c
4a4b	Cohort studies only Case-control studies only^d
5	Sub-groups: Identified by the study authors as high exposure groups^e
6	Industries associated with increased lung cancer risk from Cr(VI) (i.e., chromate production, pigment production, plating, ferrochromium production)
7*	From Analysis 4a, eliminate Tier 3 studies (Approach 1 for tier classification)^a^,c
8*	From Analysis 6, eliminate Tier 3 studies (Approach 1 for tier classification)^a^,c

*Integration of the risk of bias assessment results for the meta-analysis.

^a Approach 1 for tier classification: Greater emphasis is placed on the ratings for the key questions as identified by NTP OHAT (2015). There were 31 studies classified as Tier 3.

^b Approach 2 for tier classification: overall average ratings for the risk of bias questions is considered for tiering.

^c No Tier 3 studies were identified with Approach 2. Approach 1 only identified one study for Tier 1 (Iaia et al. 2006).

^d One case-control study was classified as Tier 3 (Moulin et al. 1992), which only left two studies (Krstev et al. 2005; Xu et al. 1996). Thus, additional analyses for the case-control study category was not conducted.

^e All studies for Analysis 5 were classified as Tier 3 studies in the risk of bias assessment (Approach 1 for tier classification). Thus, additional analyses for the high exposure subcohort category were not conducted.

To account for internal validity as suggested by NTP OHAT, Analysis 2 excluded Tier 3 studies (i.e. those with high risk of bias) from the overall meta-analysis, and Analysis 3 was limited to all Tier 1 studies (i.e. those with low risk of bias). Analyses 4a and 4b were limited to cohort or case-control studies, respectively, to characterize the results by study type. One study reported 0 and 12 stomach cancer deaths among female and male chrome plating workers, respectively, in the UK (Sorahan and Harrington 2000). The data for female workers were not included in the meta-analysis, but there were 167 female workers (in comparison to 920 male workers) and no stomach cancers reported so the impact of excluding females from this study was inconsequential.

Analysis 5 was conducted to evaluate the association between Cr(VI) and stomach cancer among workers in high exposure groups. Analysis 6 focused on industries that have been identified by IARC and United States Occupational Safety and Health Administration (OSHA) (IARC 1990; OSHA 2006; IARC 2012) to have evidence of significantly elevated lung cancer risk due to Cr(VI) exposure. These industries included chromate production, pigment production, chrome plating, ferrochromium production. Although welders were observed to have increased lung cancer risk, this increase in risk does not appear to be attributable to Cr(VI) exposure (IARC 1990; OSHA 2006; IARC 2012). For this reason, studies of welders were not included in Analysis 6.

Analysis 7 excluded Tier 3 studies from Analysis 4a; Analysis 8 excluded Tier 3 studies from Analysis 6. These analyses (along with Analyses 2 and 3) were used to integrate the risk-of-bias assessment results into the quantitative examination of the data.

Data integration, overall evaluation of confidence in the body of evidence, and development of conclusions

Confidence in the evidence and generation of conclusions followed the guidelines in the NTP (2015) Handbook for Conducting a Literature-Based Health Assessment Using OHAT Approach for Systematic Review and Evidence Integration. Per this approach, an initial rating for confidence, also referred to as the quality of evidence, was assigned per key features of study design elements including controlled exposure, exposure prior to outcome, individual outcome data, and comparison group used. The designations for initial confidence rating (high, moderate, low, very low) were based on Table 8 of NTP OHAT (2015). Once the initial confidence rating was established, consistency, directness, precision, publication bias, magnitude, and confounding were evaluated to adjust the initial confidence rating (either up- or down-grading). The final confidence ratings were assigned for animal and human evidence as well as overall. Continuing through the NTP OHAT process, the confidence ratings were integrated with the findings for the body of evidence to develop hazard conclusions by evidence stream and overall.

Results

Overall summary of the evidence base

Consistent with reporting requirements of systematic reviews (Moher et al. 2009), Figure 1 delineates the process of screening the literature to identify relevant studies. After excluding articles that did not meet the inclusion criteria based on review of titles and abstracts, 127 publications were found to be eligible for full-text review. Of those, 47 were included for data extraction. Of those included, 44 were occupational studies, and 3 were experimental animal studies. Figure 1 also presents different categories of exclusions; citations of the 80 excluded studies that underwent full text review, and specific reasons for exclusion, are listed in Supplemental Table D. For non-occupational populations with ingestion exposures to Cr(VI) (n = 4), all studies employed ecological designs and were excluded (Fryzek et al. 2001; Beaumont et al. 2008; Kerger et al. 2009; Linos et al. 2011). As such, they were not included for critical appraisal but were summarized as contextual information (Supplemental Table B).

Figure 1. Literature Search Results (in accordance to PRISMA reporting). *Hand searching was based on the reference citations of previously published meta-analyses and recent government documents, i.e., ATSDR 2012; Cole and Rodu 2005; Gatto et al. 2010; NIOSH 2013; Welling et al. 2015.

Experimental animal data characterization

There were four experimental animal studies (reported in three publications, Mackenzie et al. 1958, Borneff et al. 1968 and NTP 2008) evaluating the carcinogenic effects of ingested Cr(VI) in drinking water. Two studies were 2-year cancer bioassays of Cr(VI) in drinking water described in one publication (NTP 2008). Third one was the 1-year drinking water study (Mackenzie et al. 1958). The fourth was a three-generation drinking-water study with a total duration up to 2.4 years; dosing duration varied by generation and was affected by early termination due to the virus (Borneff et al. 1968). Table 3 presents the experimental details and results for each individual study. Both males and females were evaluated across multiple species/strains, including B6C3F1 mice, F344/N rats, white NMRI mice, and Sprague–Dawley albino rats. In the Borneff et al. (1968) and Mackenzie et al. (1958) studies, other test groups were studied. Animals were given either tap water, detergent dissolved in drinking water, detergent plus Cr(VI) in drinking water, benzo[a]pyrene plus Cr(VI) in drinking water, Cr(III) in drinking water, or Cr(VI) in drinking water. For the Borneff et al. (1968) and Mackenzie et al. (1958) studies, the Cr(VI)-exposed groups were detergent plus Cr(VI) in drinking water and Cr(VI) in drinking water, respectively, and control groups were detergent in drinking water and tap water, respectively.

Table 3. Data extraction of experimental animal studies (n = 4).

Study	Species (strain)	Gender (number)	Test compound and dose groups	Route, Cr(VI) concentration/dose^a	Study duration	Stomach tumor incidence	Overall tier for risk of bias^b	Limitations (as stated by author)
NTP (2008)	Rat, F344/N	Males (50 per dose group), females (50 per dose group)	Cr(VI) as sodium dichromate dihydrate (4 dose groups per gender per species) Negative control group also included per gender per species (tap water)	Ingestion (drinking water) Concentration: 0, 5, 20, 60, 180 mg/L (M/F) Dose: 0, 0.21, 0.77, 2.1, 5.9 mg/kg/day (M); 0, 0.24, 0.94, 2.4, 7.0 mg/kg/day (F)	2 years	Male: forestomach - 0/50, 0/50, 0/49, 0/50, 0/49; glandular stomach - 1/50,0/50, 0/49, 0/50, 1/49 Female: forestomach - 0/50, 0/50, 0/50, 0/50, 0/50; glandular stomach - 0/50, 0/50, 0/50, 0/50, 0/50 No evidence of Cr(VI)-related stomach cancer	1	None stated
NTP (2008)	Mouse, B6C3F1	Males (50 per dose group), females (50 per dose group)		Ingestion (drinking water) Concentration: 0, 5, 20, 60, 180 mg/L (F); 0, 5, 10, 30, 90 mg/L (M) Dose: 0, 0.38, 0.91, 2.4, 5.9 mg/kg/day (M); 0, 0.38, 1.4, 3.1, 8.7 mg/kg/day (F)	2 years	Male: forestomach - 1/50, 0/50, 0/48, 4/50, 0/48; glandular stomach - 1/50, 0/48, 0/48, 2/50, 0/47 Female: forestomach - 1/49, 2/49, 0/50, 0/48, 3/50; glandular stomach - 0/49, 1/48, 0/50, 0/48,0/50 No evidence of Cr(VI)-related stomach cancer	1	None stated
Borneff et al. (1968)	Mouse, White NMRI	Males (40), females (480)	4 groups. Group 1: detergent, Group 2: detergent plus Cr(VI) as potassium chromate, Group 3: detergent plus benzo[a]pyrene, Group 4: benzo[a]pyrene plus Cr(VI) as potassium chromate	Ingestion (drinking water) Household detergent "Pril" added to drinking water: 3% Benzo[a]pyrene (Group 4): 10 μg/ml (range 7-15) Cr(VI): 134 mg/L chosen as max tolerated without developing any damage; remained constant during the first 3 days. Was 80 mg/L after 3 weeks. Average uptake in Group 2: 13.5 mg/kg/day; Group 4: 12 mg/kg/day	3-generation (880 days)	Group 2: 2/66 F developed forestomach carcinoma; 9/66 F, 1/35 M developed forestomach papilloma. Group 4: 23/80 F, 12/26 M developed forestomach carcinoma. 26/80 F, 4/26 M developed forestomach papilloma. Vehicle controls also developed forestomach papilloma (F, 2/79; M, 3/47) but not carcinoma. The incidence of forestomach tumors in treated mice was not significantly higher than controls; study authors concluded evidence of carcinogenicity was equivocal.	3	Authors mentioned that groups that were given Cr(VI) (especially Group 4), cannibalism occurred. It was also mentioned that mice only reluctantly accepted the mixture of detergent, benzo[a]pyrene, and Cr(VI) in their drinking water (taking up a fluid closer to their requirements). Continuous decrease in the average weights from 7 month and onward observed for Group 4. In the 8th month of the experiment, an ectromely epidemy occurred in mice. Within 3 months, 512 animals died.
Mackenzie et al. (1958)	Rat, Sprague-Dawley albino	Experiment 1: 8 males and 8 females except the control group which contained 10 rats of each sex. Experiment 2: 9 females and 12 males	Experiment 1: control group given distilled water, Groups 2 to 6 given Cr(VI) as potassium chromate Experiment 2: Group 7 (control) given distilled water, Group 8 given Cr(VI) as potassium chromate, Group 9 given Cr(III) as chromic chloride	Ingestion (drinking water) Groups 2 to 6 in experiment 1: 0.12, 0.58. 1.2, 2.1, 3.0 mg/L Group 8 in experiment 2: 6.7 mg/L	1 year	No stomach tumors were reported	2	Mortality from respiratory infection occurred in both experiments.

M: Males; F: Females.

^aConcentrations/doses were converted based on molecular weight of Cr(VI) and molecular weight of the test compound. For Borneff et al. (1968), average bodyweight of 0.02 kg assumed for mice in calculation of doses.

^bBased on Approach 1 for tier classification.

Morbidity and carcinogenicity associated with Cr(VI) exposure were examined in all four studies. No pathological changes or carcinogenic effects in tissues including the stomach were reported in Mackenzie et al. (1958). Limiting the analysis to PECO, in the NTP (2008) bioassays, no increase in stomach cancers were observed in mice or rats at any drinking-water concentrations. In Borneff et al. (1968), white NMRI mice exposed to detergent plus Cr(VI) in drinking water developed forestomach carcinoma (2/66 females) and/or papilloma (9/66 females, 1/35 males). Vehicle controls (those exposed to detergent alone in water only) also developed forestomach papilloma, but no carcinomas. The study authors indicated that incidence of forestomach tumors in treated mice was not significantly higher than controls and concluded that evidence of carcinogenicity was equivocal (Borneff et al. 1968). Considering the small number of studies and the lack of relevant events, further quantitative evaluation of the evidence (i.e. a meta-analysis) was not warranted.

Risk-of-bias evaluation

Figure 2 presents the individual risk-of-bias assessments and tiering classifications for the animal data. NTP (2008) bioassays were classified as Tier 1 with definitely low or probably low risk of bias for the various domains. NTP (2008) bioassays had comprehensive study designs, including randomization, ascertainment of exposures and doses, blinding of research personnel, and adequate outcome assessments and complete reporting. Additionally, no other potential threats to internal validity were observed. In contrast, the other two studies did not employ randomization and blinding of research personnel and had definitely or probably high risk of bias in several other domains. For Borneff et al. (1968), a definitely high risk of bias was assigned for the exposure domain. The initial concentration of Cr(VI) (134 mg/L) was described as being stable for 3 days; after 3 weeks, the authors indicated that the concentration of Cr(VI) was 80 mg/L, and the rest was reduced to Cr(III). Other threats to internal validity were also observed, with respiratory infection in Mackenzie et al. (1958) and cannibalism and ectromelia epidemy in Borneff et al. (1968). Based on the overall risk of bias, Mackenzie et al. (1958) and Borneff et al. (1968) were classified as Tier 2 and 3, respectively.

Figure 2. Risk of bias assessment and tier classifications for experimental animal studies (n = 4). Colors: red, (−−; definitely high risk of bias), pink (−; probably high risk of bias), light green (+; probably low risk of bias), dark green (++; definitely low risk of bias). Tiers: 1 (Tier 1), 2 (Tier 2), 3 (Tier 3).

Evidence integration and overall evaluation of confidence in the body of evidence

Table 4 summarizes the elements of evidence integration and resulting confidence in the experimental animal data. The initial confidence rating for the experimental animal data was high. Confidence was increased by the consistency in the response (which also related to no dose response in all studies, no/low magnitude of response in all studies, and no unexplained inconsistencies). The evidence contained high-quality studies, including two Tier 1 studies having low risk of bias and thus confidence in the internal validity, and all four studies having a high level of external validity, because the studies’ objectives were to directly evaluate toxicological effects of ingested Cr(VI). Overall, the final level of confidence in the animal evidence stream was high; thus, there is high level of confidence in the evidence base supporting a lack of association between ingestion of Cr(VI) and stomach cancer in experimental animals.

Table 4. Summary of evidence synthesis and confidence in experimental animal studies.

Study type	Number of studies	Findings	Initial confidence rating^a	Risk of bias	Unexplained inconsistency	Indirectness	Imprecision	Magnitude of effect	Dose-response	Consistency across study types	Final confidence rating^b
Oral ingestion studies	4	No increases in stomach cancer incidence, not statistically significant, or no stomach cancers observed	High	– Two studies Tier 1 ↑; single Tier 3 ↓	– No unexplained inconsistencies	– All studies designed to directly evaluate oral exposures to Cr(VI) and cancer	– No imprecision reported in the 4 studies (Findings were negative, increasing confidence that no imprecision observed in any study)	↑ No significant effect observed in the 4 studies (increases confidence in no/low magnitude of effect)	– No dose-response reported in the 4 studies (increases confidence that no dose-response observed in any study)	↑ Results are consistent; no significantly increased stomach cancer observed in the 4 studies	High (++++): high confidence in the lack of effect

– No change in the initial confidence rating.

↑ Upgrade in the initial confidence rating.

↓ Downgrade in the initial confidence rating.

^aBased on NTP OHAT (2015) Table 8 – Study design features for initial confidence rating.

^bAccording to NTP OHAT (2015) Figure 6, factors decreasing confidence include risk of bias, unexplained inconsistency, indirectness, and imprecision. Factors increasing confidence include magnitude of effect, dose response, and consistency across study types.

Human data characterization

Of the 44 worker studies that were included, 3 used case-control design and 41 followed various occupational cohorts. Due to the large number of publications, all included studies are summarized in Supplemental Table E. Worker populations were from various countries in Scandinavia (n = 14,672), other parts of Europe (n = 64,364), East Asia (n = 48,771), and the United States (n = 16,804). The occupational categories of interest included cement packing, cement production, chromate production, pigment production, stainless steel, welding, ferrochromium production, plating, leather tanning, leather goods, foundry, painting, fur dressing, and aircraft manufacturing. No single industry dominated the evidence base, but 19 of the 44 studies were from industries known to have historically high exposures to Cr(VI), such as chromate production, pigment production, plating, and ferrochromium production. Stomach cancer mortality and morbidity estimates reported in studies of these industries were not significantly elevated (e.g. Axelsson et al. 1980; Hara et al. 2010; Huvinen and Pukkala 2013; Gibb et al. 2015).

A limited number of studies (Xu et al. 1996; Lipworth et al. 2011; Gibb et al. 2015; Proctor et al. 2016) performed Cr(VI) exposure reconstruction using job exposure matrices and industrial hygiene monitoring data. For the most part, quantitative measures of Cr(VI) exposure were not available. Several studies (Franchini et al. 1983; Amandus 1986; Sorahan et al. 1994; Becker 1999) placed workers in high-exposure categories based on longer duration of employment, time spent on a job or task, time since first exposure, or years from first employment. Latency was also used in conjunction with work tenure groups or categories of chrome plating (hard or bright) for calculating SMRs in Amandus (1986) and Franchini et al. (1983), respectively.

Lung cancer was the primary focus of most epidemiologic studies of Cr(VI) workers; although stomach cancer was evaluated, exposure-response relationship, when available, was examined only for lung cancer. Additionally, no Cox regression models or hazard ratios of stomach cancer were available. For some cohort studies (n = 33), stomach cancer mortality was examined by estimating SMR. Other cohort studies (n = 11) presented the results by SIR or standardized relative risk (SRR). In the case-control studies (n = 3), ORs were reported.

As noted previously, all studies evaluating non-occupational populations were ecologic (n = 5) (Bednar and Kies 1991; Fryzek et al. 2001; Beaumont et al. 2008; Kerger et al. 2009; Linos et al. 2011) and thus were excluded. For completeness, a description is provided herein for context, and tabular characterizations can be found in Supplemental Table B.

Risk-of-bias evaluation

With respect to the key elements of internal validity (NTP OHAT 2015)—adequacy of control for confounding, quality of exposure assessment, and accuracy of outcome ascertainment—the potential risk of bias in human studies was high for confounding and exposure assessment domains, but low with respect to outcome ascertainment (Supplemental Figure S1). Most studies had a probably high or definitely high risk of bias for the confounding domain; limited information on smoking histories, SES and asbestos exposure were noted. A few studies performed statistical adjustment for confounding variables when calculating the stomach cancer RR estimates (Xu et al. 1996; Krstev et al. 2005; Ahn et al. 2006). Exposure characterization was limited, also leading to probably high or definitely high risk of bias for this domain, because no study quantified relative risk of stomach cancer by exposure level. Only two studies developed individual measures of exposure (Gibb et al. 2015; Proctor et al. 2016), and they were given definitely low risk of bias for the exposure domain.

The ratings were relatively consistent for confidence in the outcome assessment and completeness in reporting of outcomes (Supplemental Figure S1). A “−” rating was given to three occupational studies (Moulin et al. 1990; Korallus et al. 1993; Deschamps et al. 1995) with regard to the confidence in the outcome assessment. For these studies, cause of death could not be ascertained from death certificates and had to be obtained from secondary sources such as doctor offices. For instance, death certificates are confidential and difficult to obtain in Germany, limiting outcome assessment (Korallus et al. 1993).

The potential risk of bias for other domains was generally low. Regarding the appropriateness of the comparison group, “++” ratings were given for the two occupational studies with internal controls (Moulin, Wild, Haguenoer, et al. 1993; Xu et al. 1996). Per instructions by the NTP OHAT handbook (NTP OHAT 2015), most occupational studies were given “+” ratings, because the controls were recruited within the same time frames and had similar participation and response rates. For the studies with “−” ratings (Axelsson et al. 1980; Edling et al. 1986; Iaia et al. 2006; Koh et al. 2013), the authors provided details such as selection of controls from rural areas, presence of healthy-worker effect, and misclassification of workers.

Loss to follow-up or the presence of attrition was apparent in a handful of occupational studies. For example, one study of stainless steel foundry workers indicated that they excluded data for workers of Indian and Arab descent, due to the lack of complete information (Sorahan et al. 1994). Such studies, with attrition or exclusion of workers due to incomplete data, were assigned “− −” ratings. Regarding outcome reporting, nearly all studies reported on multiple causes of death, not just those limited to cancers; thus, completeness in reporting of outcomes was noted, and “++” ratings were given.

Using Approach 1 for Tier classification, only one study (Iaia et al. 2006) was placed in Tier 1; 12 and 31 studies were in Tier 2 and Tier 3, respectively (Supplemental Figure S1). Using Approach 2 for Tier classification, 8 studies were in placed in Tier 1, 36 studies in Tier 2, and no studies in Tier 3.

Meta-analysis

Meta-analyses were conducted based on the occupational data. When assessing all included studies, there was minimal to moderate heterogeneity (I² = 25.1%) (Table 5), and the funnel plot appeared symmetrical (Figure 3), as supported by the results of Egger’s Test (p = 0.57), and the Trim-and-Fill test (Supplemental Table F). The meta-RR across all 44 occupational studies was 1.08 (95% CI: 0.96–1.21) (Analysis 1; Table 5, Figure 4). Approximately half of the studies reported RRs < 1 when comparing exposed workers to the reference groups; RR estimates >2, were imprecise, and exhibited wide 95% CIs in comparison to those with RR estimates <2. As a sensitivity analysis, one study of foundry workers that contributed 23% weight (Sorahan et al. 1994) was removed, considering that Cr(VI) exposure may be insignificant compared to exposures to other chemicals in the foundry. The results were not changed (meta-RR = 1.06, 95% CI: 0.94–1.19, Analysis 1 all studies). Additionally, when replacing the data from Korallus et al. (1993) with those of Birk et al. (2006), the results were also not changed (meta-RR = 1.07, 95% CI: 0.95–1.20, Analysis 1 all studies). Table 5 also presents the results of subgroup analyses. All meta-RRs were of low magnitude and not statistically significant. The lowest I² values were achieved in the analyses that accounted for internal study validity. Figure 5 displays a forest plot of Analysis 2, which excluded Tier 3 studies; the meta-RR was 1.03 (95% CI: 0.84–1.26) with I² of 19.8%. The sub-analyses assessing the potential impact of study validity were based on Approach 1, because Approach 2 produced no Tier 3 studies.

Figure 3. Funnel plot of all occupational studies (n = 44) included for meta-analysis.

Figure 4. Meta-analysis 1: All human occupational studies (n = 44). For the individual studies, 95%CIs are calculated intervals based on the effect size and parameters of the random-effects model.

Figure 5. Meta-analysis 2: Tier 3 removed from human occupational studies (remaining studies, n = 13). For the individual studies, 95%CIs are calculated intervals based on the effect size and parameters of the random-effects model. Tier 3: Studies with definitely high or probably high risk of bias for at least two of the key NTP OHAT questions (Q4, Q8, and Q9 related to exposure, confounding, and outcome, respectively).

Table 5. Meta-analysis results (Human occupational data).

Analysis	Description	Studies (N)	Meta-RR	95% CI	I^2
1	All Studies	44	1.08	0.96 to 1.21	25.1
2*	Eliminate Tier 3 studies from Analysis 1 (Approach 1 for tier classification)^a^,c	13	1.03	0.84 to 1.26	19.8
3*	Tier 1 studies only – Based on overall ratings (Approach 2 for tier classification)^b,c	8	0.90	0.74 to 1.10	0.0013
4a	Cohort studies only	41	1.05	0.94 to 1.19	26.2
4b	Case-control studies only^d	3	1.82	0.98 to 3.40	33.8
5	Sub-groups: Identified by the study authors as high exposure groups^e	13	1.30	1.00 to 1.67	29.5
6*	Industries with evidence of increased lung cancer risk^f	19	1.11	0.90 to 1.37	28.8
7*	Eliminate Tier 3 studies from Analysis 4a (Approach 1 for tier classification)^a^,c	11	0.94	0.78 to 1.13	3.71
8*	Eliminate Tier 3 studies from Analysis 6 (Approach 1 for tier classification)^a^,c	7	1.14	0.77 to 1.68	34.9

*Integration of the risk of bias assessment results for the meta-analysis.

^a Approach 1 for tier classification: Greater emphasis is placed on the ratings for the key questions as identified by NTP OHAT. There were 31 studies classified as Tier 3.

^b Approach 2 for tier classification: The overall average ratings for the risk of bias questions is considered for tieri ng.

^c No Tier 3 studies were identified with Approach 2. Approach 1 only identified one study for Tier 1 (Iaia et al. 2006).

^d One case-control study was classified as Tier 3 (Moulin et al. 1992), which only left two studies (Krstev et al. 2005; Xu et al. 1996). Thus, additional analyses for the case-control study category was not conducted. All studies for Analysis 5 were classified as Tier 3 studies.

^e All studies for Analysis 5 were classified as Tier 3 studies in the risk of bias assessment (Approach 1 for tier classification). Thus, additional analyses for the high exposure subcohort category was not conducted.

^f Chromate production, pigment production, plating, and ferrochromium production are industries in which Cr(VI) exposure specifically has been shown to be associated with increased cancer risk.

Forest plots for Analyses 3 through 8 are presented as Supplemental Figures S2–S8. Analysis 3 included Tier 1 studies only (n = 8), and the meta-RR was 0.90 (95% CI: 0.74–1.10) with I² of 0.0013%. Analysis 4a included cohort studies only (n = 41); the meta-RR was 1.05 (95% CI: 0.94–1.19) with I² of 26.2%. Analysis 4 b included case-control studies only (n = 3); the meta-RR was 1.82 (95% CI: 0.98–3.40) with I² of 33.8%.

Analysis 5 (n = 13) evaluated sub-groups identified by study authors of the primary literature as having high Cr(VI) exposures; as discussed previously, there were no quantitative measures of Cr(VI) exposure for most occupational cohorts. The meta-RR was 1.30 (95% CI: 1.00–1.67) with I² of 29.5%. Analysis 6 (n = 19) evaluated the specific industries associated with high exposures to Cr(VI); the meta-RR was 1.11 (95% CI: 0.90–1.37) with I² of 28.8%. When excluding Tier 3 studies from Analysis 4 b (Analysis 7), the meta-RR was 0.94 (95% CI: 0.78–1.13) with I² of 3.71%. When Tier 3 studies were excluded from Analysis 6 (Analysis 8), the meta-RR was 1.14 (95% CI: 0.77–1.68) with I² of 34.9%.

Evidence integration and overall evaluation of confidence in the body of evidence

Table 6 summarizes the elements of evidence integration and resulting confidence in the human occupational data. Initial confidence was determined using the meta-analysis descriptions, specifically study design, exposure, and industry types. Per NTP OHAT (2015) categories, the initial confidence ratings ranged from very low to moderate for the human occupational data. Confidence decreased due to the high risk of bias present for the human occupational data; potential for bias in exposure and confounding domain was apparent. Confidence increased from low magnitude of effect and consistency across study types. Meta-analyses resulted in similar findings, with meta-RRs that were not substantially/significantly elevated. All the individual studies except (Sorahan et al. 1994) showed consistency in the lack of significantly elevated risk of stomach cancer. There was external validity, because the studies were fit-for-purpose with evaluations of disease mortality risk due to Cr(VI) exposure. Overall, the final level of confidence in the human database was low to moderate.

Table 6. Summary of evidence synthesis and confidence in human occupational studies.

Study type^a	Number of studies^a	Findings^b	Initial confidence rating^c	Risk of bias	Unexplained inconsistency	Indirectness	Imprecision	Magnitude of effect	Dose-response	Consistency across study types	Final confidence rating^d
All studies	44	Not statistically elevated meta-RR	Low to moderate	↓ Probably or definitely high risk of bias in exposure and confounding domains	– Meta-analyses resulted in similar findings, including sub-analyses that assessed potential influence of study quality or size One study (Sorahan et al.1994) reported significantly elevated stomach cancer risk, but this Tier 3 study was of foundry workers with high risk of bias in exposure and confounding domains	– Studies were fit-for-purpose, evaluating disease mortality risk (including stomach cancer) from Cr(VI) exposure	– Confidence intervals around meta-RR are narrow	↑ Magnitude of effect was not large and not statistically elevated	NR Data are not available for most studies. For one study with exposure estimates, unable to establish dose-response in Proctor et al. (2016)	↑ Results are consistent across the meta-analyses (results are not statistically elevated). All individual studies (except Sorahan et al.1994) show consistency in the lack of statistically elevated risk of stomach cancer	Low (++) to moderate (+++) confidence in the human database
Cohort	41	Not statistically elevated meta-RR	Low to moderate
Case-control	3	Not statistically elevated meta-RR	Low to very low
Sub groups: Identified by study authors as high exposure groups	13	Not statistically elevated meta-RR	Low to moderate
Industries associated with increased lung cancer risk from Cr(VI)	19	Not statistically elevated meta-RR	Low to moderate

NR: Not reported.

– No change in the initial confidence rating.

↑ Upgrade in the initial confidence rating.

↓ Downgrade in the initial confidence rating.

^aBased on meta-analysis descriptions listed in Table 2.

^bBased on Meta-RRs presented in Table 5.

^cBased on NTP OHAT (2015) Table 8 – Study design features for initial confidence rating.

^dAccording to NTP OHAT (2015) Figure 6, factors decreasing confidence include risk of bias, unexplained inconsistency, indirectness, and imprecision. Factors increasing confidence include large magnitude of effect, dose response, and consistency across study types.

Synthesis of all evidence streams, and integrated conclusions

According to NTP OHAT, the risk-of-bias assessment and level-of-confidence rating were carried forward to the development of conclusions (NTP OHAT 2015). This involved translating the confidence ratings of each evidence stream and developing overall conclusions (Figure 6). Because the meta-analyses of human data consistently demonstrated a lack of evidence for effect, and all of the data in animals demonstrated a lack of evidence for effect, the data are described in the direction of “no effect.” There was low to moderate confidence in the human data demonstrating lack of evidence for effect, and high confidence in the animal data demonstrating lack of evidence for effect, thus leading to the overall conclusion that Cr(VI) is not identified to be a stomach-cancer hazard to humans. Because of the limitations of the human data related to uncertainty in exposure and confounding, the human data could be categorized as “not classifiable.” However, when combined with the animal evidence, for which there is high confidence, the same hazard conclusion would be reached.

Figure 6. Application of the NTP OHAT (2015) framework of systematic review and evidence integration for developing hazard identification conclusions.

Discussion and conclusions

This is the first systematic review that combined evidence from both experimental animal and observational human data to assess the association between Cr(VI) exposure and stomach cancer. Using the NTP OHAT (2015) framework, we included critical appraisal of study validity as part of the qualitative and quantitative integration of evidence. The available animal evidence demonstrated no discernable evidence for an effect of Cr(VI) ingestion on cancer of the glandular stomach. Quantitative assessment of evidence in humans, which was limited to observational studies of occupational exposures, consistently produced relative risk estimates demonstrating a lack of evidence for an effect, especially after accounting for study validity, study design, and/or industry type. Integration of findings further supports a lack of evidence for hazard for Cr(VI) and stomach cancer in humans.

Evaluation and integration of study validity has become an important element of risk assessment. In the National Research Council’s (NRC’s) review of the USEPA’s IRIS program, the role of evidence evaluation and integration as they relate to development of toxicity values is emphasized (National Research Council 2014). NRC further suggests that a risk-of-bias assessment be conducted on studies used as primary data sources for the hazard identification and dose-response assessments developed by the agency. The European Food Safety Authority, or EFSA, has also started to integrate study quality—measured by risk of bias—in their risk assessments (Mortensen et al. 2017). This systematic review confirms the utility of these types of approaches for assessing validity, thus characterizing the aspects that increase or decrease confidence in the findings. In this assessment, key limitations of the human evidence involved uncertainty in characterization of exposure and control for confounding such as SES and smoking. These aspects, however, are directly addressed by the studies in the animal evidence stream that involved controlled exposure (and lack of confounding). Thus, the data from the animal studies directly inform the shortcomings of the observational studies in humans, lending confidence to the overall conclusions.

With respect to meta-analysis, important methodological challenges include heterogeneity of study attributes and quality (Aschengrau and Seage 2003; Rothman et al. 2008). In addition, the general approach of meta-analysis as a way of summarizing the evidence has been questioned due to a lack of transparency and difficulties associated with replicating reported methods and results (Stroup et al. 2000; Simunovic et al. 2009). For these reasons, we included documentation of all steps and decisions in the registered protocol (PROSPERO CRD42016051625). PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analysis) was also used to preserve transparency in reporting the individual studies and integrity of the approach and conduct of the current analyses (Moher et al. 2009; NTP OHAT 2015). We did not produce a single measure of association; instead, we used the meta-analysis to explore sources of heterogeneity across studies.

Our meta-analyses demonstrated no appreciable or significant increase in stomach cancer risk associated with occupational exposures to Cr(VI); low magnitude and high precision of the risk estimates were observed. In one sub-analysis (Analysis 4b), limited to three case-control studies, the risk ratio was elevated though not significantly (meta-RR = 1.82, 95% CI: 0.98–3.40; I²=33.8%). However, this result was not supported by the more robust analysis based on 41 cohort studies (Analysis 4a: meta-RR = 1.05, 95% CI: 0.94–1.19, I²=26.2%). The Meta-RRs for stomach cancer were not significantly elevated among workers in chromate production, pigment production, chrome plating, and ferrochromium production (Analysis 6; Table 5). Significant exposures to Cr(VI) have been observed for these industries, and OSHA and IARC have identified workers in these industries as having the highest risk of lung cancer associated with Cr(VI) exposure (IARC 1990; OSHA 2006; IARC 2012). Additionally, meta-analysis results were close to the null value when study quality was considered, and integration of the risk-of-bias assessment generally resulted in lower heterogeneity.

As discussed previously, dose-response data were limited or unavailable for stomach cancer. For the Painesville chromate production workers assessed in Proctor et al. (2016), exposure information was accessible. The cumulative Cr(VI) exposures were 0.002, 0.008, 0.15, 0.16, and 0.47 mg/m³-year for five stomach cancer cases, which are relatively low exposures for the Painesville cohort, in which an increase in lung cancer risk was not observed at exposures <1 mg/m³-yrs (Proctor et al. 2016). However, due to the low numbers of stomach cancer cases and lack of quantitative exposure data in other studies, a dose-response pattern could not be established or evaluated.

One limitation of this systematic review is that most occupational studies had a probably high or definitely high risk of bias for confounding and exposure. Statistical adjustment for confounding variables was presented in only three studies (Xu et al. 1996; Krstev et al. 2005; Ahn et al. 2006). Exposure characterization methods in most studies were of poor quality, and as discussed above, only a handful of studies conducted exposure assessment based on quantitative data and methods (Xu et al. 1996; Lipworth et al. 2011; Gibb et al. 2015; Proctor et al. 2016). As a consequence, the level of confidence in the evidence from human studies was low to moderate. Additionally, data for non-occupationally exposed populations were sparse and low quality due to reliance on ecologic study design. For this reason, non-occupational studies could not be included in the systematic review and meta-analysis. It is also important to acknowledge that any systematic review involves an element of judgment. More quantitative bias-adjusted approaches could be employed in the future (Doi et al. 2013). Additional research and discussion are ongoing in the field of evidence-based toxicology, because practitioners recognize that existing tools could be refined to better categorize and integrate data in the context of chemical risk assessment (EFSA 2017).

This systematic review and meta-analysis was compared to the previous publications (Cole and Rodu 2005; Gatto et al. 2010; Welling et al. 2015) (Supplemental Table G); Supplemental Table H also identifies which studies were included in each of the different meta-analysis studies. The previous analyses did not consistently evaluate study quality and validity, likely because systematic review methods in environmental epidemiology and toxicology are relatively new. Only Cole and Rodu (2005) evaluated study quality, and performed meta-analyses according to the presence or absence of control for major cancer risk factors (specifically SES for stomach cancer). Cole and Rodu (2005) observed a statistically significant inverse association between Cr(VI) exposure and stomach cancer mortality (meta-SMR = 82) after controlling for SES; however, when SES was not considered, the association was in the opposite direction (meta-SMR = 137). Cole and Rodu (2005) noted that, of the 14 studies with control for SES, 12 had a stomach cancer SMR below 100; for studies (n = 18) without control for SES, all but two had stomach cancer SMRs greater than 100. Cole and Rodu (2005) concluded that studies that lacked control for the low economic status of Cr(VI) workers largely impacted the association observed between Cr(VI) exposure and stomach cancer. This is not surprising given that SES is an important variable to consider when evaluating stomach cancer. Several studies have consistently shown that higher SES is significantly associated with decreased stomach cancer mortality and incidence in the general population (Fontana et al. 1998; van Loon et al. 1998; Nagel et al. 2007; Donnelly et al. 2013). While we did not attempt to adjust for SES in this analysis, we note that the McDowall (1984) study of cement workers, included in this meta-analysis, reported an increasing trend in stomach cancer SMRs with decreasing SES, and the Sorahan et al. (1987) study of chromium platers, also included herein, concluded that “raised mortality from cancer of the stomach among male chrome platers is due, at least in part, to social class differences” [relative to the general population]. Although it is difficult to generalize across studies, it seems reasonable to infer that many of the primary studies included in this and the other stomach cancer meta-analyses for Cr(VI), have potential for confounding by SES.

The results of the meta-analyses by Gatto et al. (2010) and by Cole and Rodu (2005) are similar to those described here. In contrast, Welling et al. (2015) reported significantly increased risk of stomach cancer in association with occupational Cr(VI). The differences between findings across these reviews and meta-analyses may be attributable to the types of studies that were considered eligible for inclusion. Whereas Welling et al. (2015) included 56 studies with variable likelihood of Cr(VI) exposure, Cole and Rodu (2005) and Gatto et al. (2010) evaluated fewer studies (32 and 29, respectively), but those studies specifically focused on workers with known exposure to Cr(VI). For example, the Welling et al. (2015) meta-analysis incorporated a large Scandinavian survey study (Pukkala et al. 2009) that reported stomach cancer risk for brick layers, a group which original study authors characterized as exposed to asbestos and silica dust, but not chromium. On the other hand data for welders and smelting workers, which Pukkala et al. noted as having chromium exposures (speciation of chromium not indicated), were not included in the Welling et al. (2015) meta-analysis. Most importantly, none of the occupational subgroups were specified by Pukkala et al. as having exposures to Cr(VI). Herein, we have included only studies of workers with documented Cr(VI) exposure. It appears this methodological difference serves as the main explanation for the discrepancy between results reported by Welling et al (2015) and those observed in other meta-analyses.

Our findings are consistent with the known biological properties of ingested Cr(VI). Several studies have reported that, on ingestion, Cr(VI) is reduced to the less toxic and less bioavailable trivalent form [Cr(III)] (De Flora et al. 1987, 2016; Kirman et al. 2016). These results are also consistent with the conclusions of agencies such as IARC, ATSDR, and NIOSH that have conducted hazard identification to evaluate all disease risks associated with Cr(VI) using human and animal data but have not considered stomach cancer to be of significant or particular concern (NIOSH 1975; IARC 1990; ATSDR 2012; IARC 2012; NIOSH 2013).

In conclusion, a systematic assessment of the body of literature, with updated inclusion/exclusion criteria relative to previous assessments and considerations of internal validity, allows a more robust assessment of the association between Cr(VI) exposure and stomach cancer risk. Combining the streams of evidence, Cr(VI) was not identified to be a stomach cancer hazard in humans.

Table

SAB Member Name	Expertise	Roles and Responsibilities
Michael Goodman Loren Lipworth Tasha Beretvas#	Epidemiology Epidemiology Statistics	Overall: Provide guidance on methods for systematic review and meta-analysis of Cr(VI) exposure and stomach cancer risk. Specific tasks are: Discuss and agree upon the scope and focus of topic.  Approve systematic review framework (i.e. PRISMA) to be used. The framework to be used will be decided by ToxStrategies researchers. Review and approve draft protocol to be written by ToxStrategies researchers. The protocol will include research questions, key search terms, inclusion/exclusion criteria, literature search strategy, risk of bias assessment methods, and strategies for meta-analyses and analyses of subgroups. Review evidence tables (or summary of information) generated by ToxStrategies. Review meta-analyses results. Reviewed evidence synthesis and integration qualitative analysis results. Review and comment on manuscript drafts; serve as coauthor if desired^#

Declaration of interest

The activities of the authors in preparation of the paper are provided in the following table:

Table

Author Name	Activities
Mina Suh	Developed the protocol Reviewed literature search and screening results Conducted data extraction and risk of bias assessment and performed quality assurance/quality control Conducted evidence synthesis and integration qualitative analyses Designed and reviewed meta-analysis results Performed quality assurance/quality control of meta-analysis results Developed the initial manuscript draft, revised the manuscript, and finalized the manuscript
Daniele Wikoff	Developed the protocol Reviewed data extraction, risk of bias assessment, and meta-analysis results Reviewed evidence synthesis and integration qualitative analysis results Developed the initial manuscript draft, revised the manuscript, reviewed the final draft
Loren Lipworth	SAB Member Provided input to the protocol Reviewed data extraction, risk of bias assessment, and meta-analysis results Reviewed evidence synthesis and integration qualitative analysis results Reviewed and provided input on the manuscript (initial draft): development of introduction, study methods, results, and discussion sections Reviewed the revised draft and final manuscript
Michael Goodman	SAB Member Provided input to the protocol Reviewed data extraction, risk of bias assessment, and meta-analysis results Reviewed evidence synthesis and integration qualitative analysis results Reviewed and provided input on the manuscript (initial draft): development of introduction, study methods, results, and discussion sections Reviewed the revised draft and final manuscript
Seneca Fitch	Conducted literature search and screening Conducted data extraction and risk of bias assessment and performed quality assurance/quality control Reviewed the initial draft and final manuscript and provided edits
Liz Mittal	Conducted meta-analyses and subgroup analyses Developed funnel and forest plots in R Reviewed the final manuscript and provided edits
Caroline Ring	Provided oversight of meta-analyses Performed quality assurance/quality control of meta-analysis results Conducted Trim and Fill test Conducted Egger’s regression test Reviewed the final manuscript and provided edits
Deborah Proctor	Developed the protocol Reviewed literature search and screening results Reviewed data extraction, risk of bias assessment, and meta-analysis results Reviewed evidence synthesis and integration qualitative analyses Developed the manuscript draft, revised the manuscript, and finalized the manuscript

The authors include salaried scientists employed by ToxStrategies, which is a private consulting firm, and the paper was prepared during the normal course of employment. ToxStrategies was funded under contract to EPRI to conduct this systematic review and meta-analysis and prepare the manuscript for publication. ToxStrategies continues to provide scientific consulting and research oversight on projects funded by EPRI. With support from EPRI, ToxStrategies’ scientists have published seven studies related to the carcinogenicity of Cr(VI) (Proctor et al. 2012; Proctor et al. 2014; Thompson et al. 2015; Young et al. 2015; Proctor et al. 2016; Thompson et al. 2017; Rager et al. 2019).

ToxStrategies (DP and MS) originally proposed the project to EPRI. EPRI is an independent nonprofit 501(c)3 organization, which is supported primarily by the electric industry in the United States and abroad (www.epri.com). EPRI and its contributors’ interest in Cr(VI) is primarily related to the presence of Cr(VI) in some electrical power generation byproducts (e.g. fly ash), which have the potential to impact environmental media. Dr. Rohr reviewed the draft and final manuscript for clarity, and was the only EPRI reviewer.

DP and MS have given presentations on this and related Cr(VI) topics at scientific conferences and meetings with regulatory agencies with funding provided by EPRI, the American Chemistry Council and ToxStrategies. With funding by ToxStrategies, elements of this systematic review and meta-analysis were presented at the Society of Toxicology 2017 Annual Conference and ToxExpo in Baltimore (Abstract 1308) and at the Joint Annual Meeting of the International Society of Exposure Science and the International Society of Environmental Epidemiology in Ottawa in 2018 (Abstract 2996325). DP and MS have coauthored a previous meta-analysis of occupational exposure to Cr(VI) and cancers of the gastrointestinal tract (Gatto et al. 2010). DP has also been an expert in litigation involving the potential for lung cancer and respiratory effects related to occupational and environmental Cr(VI) exposure.

This research was conducted to inform the scientific and regulatory risk assessment of Cr(VI). The authors have sole responsibility for the writing and content of this paper.

Acknowledgements

We acknowledge Dr. Annette Rohr at the Electric Power Research Institute (EPRI). Dr. Rohr is ToxStrategies’ principal contact for EPRI. We thank Dr. Rohr for organizing the scientific advisory board (SAB), which provided guidance for this systematic review and meta-analysis. Dr. Rohr identified SAB members based on suggestions by ToxStrategies (MS and DP) and sent formal invitations in August 2017. The SAB consisted of three members (Dr. Tasha Beretvas of University of Texas Austin, Quantitative Methods Program in the Department of Educational Psychology; Dr. Loren Lipworth of Vanderbilt University Medical Center; and Dr. Michael Goodman of Emory University Rollins School of Public Health). The following table describes the roles and responsibilities of the SAB:

# Dr. Beretvas indicated that she had limited availability to review the manuscript drafts and contribute to the content. In an email correspondence with Dr. Rohr at EPRI (sent November 2017), she wanted to be acknowledged but not be a coauthor.

Note: All SAB members received compensation from EPRI for participating.

We thank Dr. Rohr for reviewing the initial and final manuscript drafts and commenting on the clarity of presentation. We thank Dr. Tasha Beretvas for providing valuable statistical input for this systematic review. We thank Mr. Rick Nelson of ToxStrategies, for editorial assistance. We thank Ms. Alea Goodmanson, formerly of ToxStrategies, for participating in the data extraction and translating three studies in French to English to facilitate data extraction and risk of bias assessment. The activities of all collaborators are described in the following table:

Table

Collaborator Name	Activities
Tasha Beretvas	SAB member Provided input to the protocol Directed statistical analyses to be performed Reviewed data extraction, risk of bias assessment, and meta-analysis results
Alea Goodmanson	Participated in data extraction for epidemiology studies Translated three studies in French to English to facilitate data extraction and risk of bias (Moulin et al. 1992; Moulin, Wild, Toamain, et al. 1993; Moulin et al. 1995)
Rick Nelson	Reviewed the manuscript draft to provide editorial comments
Annette Rohr	Organized the SAB Reviewed the manuscript draft and suggested wording changes for clarity of presentation

Finally, the authors gratefully acknowledge the comments received from two journal reviewers, selected by the editor, who were anonymous to the authors. We appreciate and acknowledge the guidance offered by the Editor-in-Chief Dr. Roger McClellan for recommendations on the acknowledgments and declaration of interest.

Notes

1 The NTP OHAT (2015) Handbook and the Risk of Bias tool describe the ++, +, −, or − − output as answer formats; for the purposes of this assessment, we have chosen to use the term “ratings” to represent the domain-based output.