3
Health Authority Perspectives on the Synthesis of Epidemiologic Evidence

Rebecca Nachman of the U.S. Environmental Protection Agency (EPA) surveyed example applications of triangulation from EPA’s Integrated Risk Information System (IRIS) Program. To anchor her talk, Nachman relied on the definition of triangulation provided by Lawlor (2016). She noted that triangulation is “a pretty well-established approach to evidence integration in EPA risk assessment,” such as when the three streams of evidence (epidemiologic, toxicological, and mechanistic) are integrated to inform causal determinations. Work remains, however, in applying triangulation in the context of multiple analyses within a single study, or within a single stream of evidence. She shared examples to focus on these latter areas.

First, she referenced the 2014 assessment of Libby Amphibole Asbestos, which derived an inhalation unit risk for lung cancer using data from an occupational cohort with incomplete smoking data. The potential for uncontrolled confounding by smoking was evaluated using a method that uses a negative control outcome to examine an association expected to be null (Richardson, 2010). “You can have negative control exposures. In that case, the association between the control exposure and outcome has the same common causes as the association between the target exposure and the outcome. An example of that would be a placebo in a controlled exposure trial,” Nachman explained. Specifically, the negative control was chronic obstructive pulmonary disease (COPD), which is related to smoking but is not believed to be associated with Libby Amphibole Asbestos exposure. EPA evaluated the relationship between Libby Amphibole Asbestos and the risk of COPD, using an extended Cox proportional hazards model with two different exposure metrics, finding that the estimated slope for COPD was small and not statistically significant. The analysis provided greater confidence that the relationship between Libby Amphibole Asbestos and lung cancer was not due to uncontrolled confounding by smoking in the occupational study that had incomplete smoking data, she concluded.

Second, Nachman referenced EPA’s trichloroethylene (TCE) assessment. Associations between TCE and multiple types of cancer were investigated in the 2011 IRIS Toxicological Review,¹ the International Agency for Research on Cancer (IARC) Monographs 2012 TCE assessment,² and in the National Toxicology Program’s (NTP’s) 2015 Report on Carcinogens.³ In its assessment, EPA performed a meta-analysis and evaluated the role of confounders across studies, finding an overall meta-risk ratio (RR) of 1.27 with no significant heterogeneity. A meta-analysis using a priori criteria to select studies with greater overall confidence yielded a summary RR of 1.58 for the highest exposure groups. EPA explored whether the meta-analyses results could be explained by uncontrolled confounding, such as co-exposures to mineral oils, hydrazine, and other solvents to which degreasers are commonly exposed. Mineral oils and other co-exposures were included as covariates in some studies, and the included studies varied in the pattern, level, and specific types of co-exposures. In addition, cutting oil exposure had not been associated with kidney cancer in other cohort or case-control studies. EPA also explored whether the results could be explained by known risk factors for kidney cancer, specifically high body mass index and smoking. However, most case-control studies controlled for these factors, and the cohort study found no pattern of increased lung cancer, indicating that a correlation with smoking did not drive the effect of

___________________

¹ See https://cfpub.epa.gov/ncea/iris_drafts/recordisplay.cfm?deid=237625.

² See https://monographs.iarc.who.int/wp-content/uploads/2018/06/mono106.pdf.

³ See https://ntp.niehs.nih.gov/ntp/roc/monographs/finaltce_508.pdf.

TCE. Overall, EPA concluded that co-exposures to other solvents or uncontrolled confounding are unlikely to provide an alternative explanation for the kidney cancer findings for TCE.

Third, Nachman described EPA’s analyses of formaldehyde. She noted that multiple agencies and the 2022 draft IRIS assessment have concluded that formaldehyde inhalation causes nasopharyngeal cancer in humans, based in part on consistent findings across studies. EPA explored whether the consistency in these findings could be explained by an uncontrolled confounder. Known risk factors for nasopharyngeal cancer include consumption of Chinese salted fish, exposure to wood dust, smoking, alcohol consumption, and Epstein-Barr virus.⁴ “Are any of these risk factors associated with exposure?” she asked, noting that dietary exposures and alcohol consumption are unlikely to be consistent confounders across all studies. On the other hand, multiple studies controlled for wood dust and for smoking, but neither were found to be a confounder of the association with formaldehyde. Consistency was demonstrated across multiple studies by a pattern of increased risk in many different populations, exposure scenarios, and time periods. Unmeasured confounding or chance are unlikely to be alternative explanations to the observed associations.

Nachman closed by saying that the examples shared “were done in the spirit of triangulation, with the goal of synthesizing results within an evidence stream. Negative control outcomes or exposures can be used to identify potential for residual or uncontrolled confounding or to rule out uncontrolled confounding. And it is possible to look at groups of studies for patterns of results by study attributes such as study design, level of exposure, and study setting.”

During the next presentation, Mary Schubauer-Berigan discussed applications of triangulation by the IARC Monographs. The IARC Monographs identify the causes of human cancer. Over its 50-year history, the Monographs have evaluated more than 1,000 agents, classifying 121 as carcinogenic to humans (Group 1) and >400 as probably (Group 2A) or possibly (Group 2B) carcinogenic to humans. National and international health agencies use the IARC Monographs to support preventive actions and regulation.

The Preamble to the IARC Monographs,⁵ which lays out the guiding principles for the evaluations, has been published in each volume since 1971. The Preamble has been updated periodically by external advisory groups, most recently in 2018. The Preamble provides procedural guidelines for selecting experts who are free from conflicts of interest to perform the evaluations, involving stakeholders, and for conducting the evaluation meetings. It also describes the five-step systematic process for reviewing, synthesizing, and integrating the three individual streams of evidence. The strength of these streams of evidence—on cancer in humans, cancer in experimental animals, and mechanistic evidence—is first categorized using pre-specified criteria. The conclusions on the three streams of evidence are then integrated in a formal decision process to reach uniform classifications.

The 2019 Preamble update concerned procedural advances in systematic review principles, as well as the scientific advances in cancer epidemiology and carcinogen mechanisms. For the mechanistic evidence evaluation, the Preamble describes an approach to address the notable challenges identified as part of the Volume 100 review of all the Group 1 agents and two subsequent workshops.⁶ One outcome of these workshops was identifying the 10 key characteristics (KCs) of carcinogens, detailed later by Smith, which comprise the chemical and biological properties of the established human carcinogens. These KCs provide the basis for an approach to search systematically for evidence on relevant mechanisms, bringing uniformity across assessments, facilitating efficient analysis of the large literature, and importantly, avoiding bias toward any favored mechanisms.

Another new feature of the updated Preamble is that the three evidence streams are integrated in a single step “in a process that is truly triangulation by design,” noted Schubauer-Berigan. This step results in classification of the agent into one of four groups: Group 1, carcinogenic to humans; Group 2A,

___________________

⁴ See https://monographs.iarc.who.int/wp-content/uploads/2019/07/Classifications_by_cancer_site.pdf.

⁵ See https://monographs.iarc.who.int/wp-content/uploads/2019/07/Preamble-2019.pdf and https://pubmed.ncbi.nlm.nih.gov/31498409.

⁶ See https://monographs.iarc.who.int/monographs-available and https://publications.iarc.fr/578.

probably carcinogenic; Group 2B, possibly carcinogenic; and Group 3, not classifiable. As previously noted, sufficient evidence of cancer in humans leads to a Group 1 classification. Agents with sufficient evidence in experimental animals together with strong mechanistic evidence based on KCs are also classified in Group 1. Group 2B is based on a single stream of evidence, such as limited evidence in humans, sufficient evidence of cancer in experimental animals, or strong mechanistic evidence based on KCs. Group 2A is based on two or more streams of evidence, at least one of which is in humans or human cells. Since 2011, more than 140 environmentally relevant agents have been evaluated by the Monographs, Schubauer-Berigan said. Notably, nearly all Group 1 agents have contributions from multiple evidence streams. Human cancer evidence rarely forms the sole basis for an evaluation. Many Group 2A agents had limited human cancer evidence, combined with either strong mechanistic or sufficient bioassay evidence, she explained.

Schubauer-Berigan next addressed triangulation in the context of human cancer evidence synthesis. Observational epidemiology studies of different designs and aims may provide relevant information but also differ in their quality, informativeness, and potential for bias. IARC evaluations consider both study quality and study informativeness—or the ability to detect the presence of a true association or the absence of a true null association. This analysis incorporates study quality and power, and factors such as adequacy of latency between exposure and cancer outcome, and exposure to the target organ. To ensure consistency and transparency, several software tools are used to search and organize the studies, extract study data, and summarize study characteristics in a standardized format to address strengths, limitations, and exposure quality considerations (Shapiro et al., 2018).

The body of human cancer evidence is synthesized using an adaptation of the Bradford Hill viewpoints (Hill, 1965; see Box 2-1). The predefined categories for the strength of evidence conclusions are sufficient, limited, inadequate, and evidence suggesting lack of carcinogenicity. For sufficient evidence, a causal relationship has been established, and chance, bias, and confounding can be ruled out with reasonable confidence. For limited evidence, a positive association is credible, but chance, bias, and/or confounding cannot be ruled out with reasonable confidence. For inadequate evidence, studies either permit no conclusion about a causal association, or no data are available. These criteria have been largely the same since the introduction of the phrase “chance, bias, and confounding” in the Preamble revision of 1982 to differentiate between sufficient and limited evidence.

Schubauer-Berigan described how detailed evidence synthesis assesses the role of chance, bias, and confounding on causal conclusions that are specific to each cancer site. Among the many approaches to assess bias are explicit exposure quality evaluations, including on the impact on measures of association. Evidence triangulation also considers whether studies with different biases point to the same conclusion. Use of causal diagrams helps elucidate confounders. Different study designs, such as Mendelian randomization, may help minimize confounding. Negative control outcomes, which are plausibly related to confounders but not the agent, can help to identify possible confounding. Indirect adjustments and worst-case assumptions about confounder exposure distributions can help when information about the confounder is available for none of or only some subjects.

To illustrate the evidence triangulation in the evaluation process, Schubauer-Berigan cited the IARC evaluation of arsenic and lung cancer. For studies on workplace inhalation exposure (consisting of cohort and nested case-control studies of workers with RRs of 2–3 overall), the main bias concerns were non-differential exposure misclassification and residual confounding by smoking and other occupational exposures. For studies of populations ingesting drinking water with high arsenic concentrations, the main concerns were ecological metrics of exposure and lack of information on confounders. Large excess risks were reported for multiple locations globally with clear dose-response patterns in these population studies. Taken together, she noted that the findings supported a causal interpretation of the association between arsenic and lung cancer.

Schubauer-Berigan also cited the IARC evaluation of nightshift work and breast cancer, which assessed studies with different bias concerns and heterogeneous results. Most cohort studies of nightshift workers did not show positive findings. However, they were uninformative to detect associations for specific windows of sensitivity and non-differential exposure misclassification was a concern. On the

other hand, a large and informative pooled case-control study showed positive associations overall but did not observe exposure response for most exposure metrics. Differential exposure misclassification and selection bias were the major concerns. Taken together, the studies provided limited evidence for breast cancer in humans because bias could not be excluded as a reasonable explanation for the findings.

Looking ahead, IARC Monographs are expected to continue to support the clarity and transparency of its evaluations, relying on nuanced approaches to assess direction and magnitude of biases that inform expert judgment. IARC will hold a workshop in October 2022 to develop an IARC scientific publication on a “bias impact toolkit” to aid evaluation of bias in epidemiological studies for cancer hazard identification.

After Schubauer-Berigan concluded her talk, Jonathan Samet of the Colorado School of Public Health provided reflections on causal inference in different contexts and triangulation as proposed by Lawlor and others. He observed that “determining that a factor causes a disease is a big deal,” referencing the 2004 Surgeon General’s report on the Health Consequences of Smoking.⁷ He noted that a causal conclusion has immediate implications for prevention because it conveys inference that changing a given factor will actually reduce a population’s disease burden.

He raised three major questions about causation. First, does a factor increase risk for an outcome? This question falls in the domain of evidence synthesis and integration, and is the focus of his presentation. Questions two and three—how much does a factor increase risk for a disease and how much of the outcome is attributable to the factor?—fall in the domains of causal effect estimation and burden estimation, respectively.

The National Research Council has addressed these questions about causation in various publications on the risk assessment context and framework. The 1983 “Red Book,” which described hazard identification, dose-response, exposure assessment, and risk characterization, was published almost 40 years ago (NRC, 1983). Those concepts have evolved through subsequent reports: Science and Judgment in Risk Assessment (NRC, 1994), focused on uncertainty, and Science and Decisions: Advancing Risk Assessment (NRC, 2009), focused largely on getting the question right for the policy context.

The process behind establishing causal inference consists of developing evidence through research, synthesizing evidence through systematic reviews, meta-analyses, and other approaches, and integrating evidence from different streams to determine the strength of evidence for causation of an outcome by a factor. “Causal criteria” figure prominently in guiding expert judgment to reach causal inference conclusions, he said.

Among key points, Samet noted that triangulation is an emerging area with nascent literature, as shown by the predominance of recently published articles retrieved by a PubMed search on “triangulation and epidemiology.” A review of citation history revealed that the most cited article to date on triangulation (Lawlor et al., 2016) has not reached many other papers. “There’s been a relatively restricted conversation about the article, triangulation, and its implications. This workshop is perhaps going to broaden the reach of the idea,” he said. He noted that the article presented some “sensible” approaches for using triangulation, but that the idea of using different studies and lines of evidence to assess coherence for causation is long-standing. He cited the Preamble to the IARC Monographs⁸ and a 1964 Surgeon General’s Report⁹ to describe the discourse on using different lines of evidence, for which the key tenets are broadly accepted. As stated in the latter, “when coupled with the other data, results from the epidemiological studies can provide the basis upon which judgments of causality may be made.”

Samet highlighted the challenges and opportunities for mechanistic evidence in evidence integration. He cited the 2017 National Academies report Using 21st Century Science to Improve Risk-Related Evaluations (NASEM, 2017), which focused on how to use the evidence emerging from novel approaches to toxicity testing to understand whether an agent poses a hazard. He noted that the revised

___________________

⁷ See https://www.cdc.gov/tobacco/sgr/2004/index.htm.

⁸ See https://monographs.iarc.who.int/iarc-monographs-preamble-preamble-to-the-iarc-monographs.

⁹ See https://profiles.nlm.nih.gov/spotlight/nn/feature/smoking.

IARC Monographs Preamble places the human, animal, and mechanistic evidence as coequals during evidence integration, broadly reflecting triangulation.

Samet next highlighted three EPA systems for evidence synthesis and integration to help identify hazards and estimate any associated risks: the National Ambient Air Quality Standards framework; the IRIS Program and its approaches, based on the 2014 National Academies review; and the recent National Academies Toxic Substances and Control Act review. He also highlighted numerous other National Academies reports providing guidance to EPA since the IRIS Program started in 1984 including, for example, Toxicity Testing in the 21st Century: A Vision and a Strategy (NRC, 2007); Review of the Environmental Protection Agency’s Draft IRIS Assessment of Formaldehyde (NRC, 2011); the 2014 and 2018 reports on the IRIS process (NRC, 2014a; NASEM, 2018); the aforementioned Using 21st Century Science to Improve Risk-Related Evaluations (NASEM, 2017); The Use of Systematic Review in EPA’s Toxic Substances Control Act Risk Evaluations (NASEM, 2021); Review of U.S. EPA’s ORD Staff Handbook for Developing IRIS Assessments: 2020 Version (NASEM, 2022); a pending National Academies report on causal inference related to the national ambient air quality standards¹⁰; and a pending National Academies re-review of formaldehyde.¹¹

Samet noted that multiple EPA frameworks and this new triangulation itself are in evolution. In addition, triangulation in the broad sense is inherent to evidence integration, and whether this new formalism will be helpful for very important and complex decisions remains unclear. He encouraged finishing the work in progress based on the ample guidance noted above.

In closing, Samet quoted George Comstock: “Epidemiologic science can give only general guidance to those who must decide upon acceptable limits of air pollutants. Judgment in this area depends much more on the art of epidemiology, the drawing of reasonable conclusions from imperfect data” (Comstock, 1979).

Next, Ruth Lunn of NTP described the Report on Carcinogens.¹² Noting that the identification of carcinogens is a key step in preventing cancer, Congress mandated a Report on Carcinogens in 1978 to identify substances that pose a cancer hazard to people residing in the United States. Specifically, two listing categories are used: known to be a human carcinogen and reasonably anticipated to be a human carcinogen. Agents listed in the first category have sufficient evidence of carcinogenicity from studies in humans, which can include mechanistic studies in exposed humans (later detailed by Smith). Agents listed in the second category have sufficient evidence of carcinogenicity in experimental animals, or limited evidence of carcinogenicity in humans, or convincing mechanistic information. In this context, triangulation can help strengthen conclusions.

NTP uses a four-part process to prepare its report for the Secretary of the U.S. Department of Health and Human Services. Key elements are external scientific input throughout the process, public comment, and scientific peer review. NTP conducts evaluations and makes decisions using systematic review methods. Within an evidence stream, NTP uses the categories and criteria for sufficient and limited evidence established by the IARC Monographs. It performs scoping to identify topics, creates evidence maps (in HAWC or Tableau),¹³ and develops a protocol for the specific substance and types of cancer under evaluation. Lunn added that NTP assesses potential biases and study sensitivity, using a definition of study sensitivity akin to IARC’s definition of informativeness. Finally, NTP synthesizes evidence across studies from the same evidence stream, and then integrates across evidence streams.

The synthesis of evidence across epidemiological studies is qualitative, rather than quantitative, and relies on multiple considerations. The approach stretches beyond the risks of bias, to address the magnitude, direction, and impact of biases on the study finding. Triangulation-like approaches, and the Hill considerations (see Box 2-1), address consistency, cohesiveness, and sources of heterogeneity. Lunn

___________________

¹⁰ See https://www.nationalacademies.org/our-work/assessing-causality-from-a-multidisciplinary-evidence-base-fornational-ambient-air-quality-standards.

¹¹ See https://www.nationalacademies.org/our-work/review-of-epas-2021-draft-formaldehyde-assessment.

¹² See https://www.niehs.nih.gov/research/atniehs/labs/iha/roc/index.cfm.

¹³ See https://hawcproject.org and https://www.tableau.com.

noted that stratification can be based on informativeness, exposure metrics, effect modifiers, or any key issue identified during the scoping process. Other Hill considerations, such as temporality, the strength of the evidence, and cohesion are also considered.

Lunn explored three triangulation-like examples. The first, on pentachlorophenol (PCP) and byproducts of its synthesis, considered different measures of PCP exposure in Dow production workers primarily exposed to PCP by inhalation (Ramlow et al., 1996; Collins et al., 2009). Using different methods (e.g., PCP, or dioxin by-products as a surrogate for PCP) increased confidence for interpretation of this individual study because diverse exposure assessments may have different types of errors, but biases are most likely in the same direction. Findings are consistent with an informative, large study that assessed dermal exposure to PCP of Canadian sawmill workers (Demers et al., 2006) and other supporting studies.

The second example related to TCE. Lunn referenced the Report on Carcinogens evaluation published in 2015,¹⁴ with additional studies published after the IRIS meta-analysis, including one large negative study. Plotting ever exposure to TCE (different from the high exposure cited by Nachman) revealed a pattern of heterogeneity in the findings. Evaluating informativeness and the different types of biases helped to identify the probable direction of the biases. “We’re mainly concerned with non-differential exposure misclassification,” she explained. Non-differential exposure misclassification biases toward the null, whereas selection bias and confounding bias away from the null. “Basically, both triangulation and other ways of exploring heterogeneity are important for reaching conclusions,” Lunn asserted.

Lunn’s last example assessed the carcinogenicity of nightshift work and light at night. The assessment considered the human cancer studies, and their integration with other evidence comprising mechanistic data on the KCs (as described by Smith), toxicology findings, and breast cancer biology. Similar to IARC, Lunn and her colleagues found differences in case-control studies and cohort studies that have different biases, but systematic evaluation by duration and other factors pointed to more consistent evidence for the carcinogenicity of persistent nightshift work.

In closing, Lunn suggested considering triangulation-based concepts in the evaluation of individual studies, across studies within a stream of evidence, and across evidence streams. Assessing the magnitude of biases and their direction can be important for evaluating heterogeneity, which may also be explained by exposure metrics and substance-specific factors. Consistent findings across evidence streams increases confidence in conclusions. Detailed evidence tables also provide transparency, she added.

Transitioning to a state perspective, Joseph Haney of the Texas Commission on Environmental Quality (TCEQ) described the TCEQ toxicity factor guidelines. Implemented in 2006, the guidelines have undergone external peer review and were updated most recently in 2015.¹⁵ Various TCEQ values are used by other countries and states, and its guidelines are acknowledged by external experts, including at the World Health Organization. TCEQ’s work demonstrates how state health authorities are using triangulation and evidence synthesis.

Haney noted that for chemical health assessments, triangulation may consist of (1) multiple analyses within a single study; (2) synthesis of results within a stream of evidence such as epidemiological studies; and/or (3) integration across evidence streams to help inform causal determinations. He noted that, although the use of the term triangulation is novel, these three applications are intrinsic to chemical dose-response assessments, as illustrated in TCEQ’s hexavalent chromium assessment.¹⁶ The weight of evidence statement commonly used in chemical assessments reflects the synthesis of results and integration of elements to help inform causal determinations. They also reflect Hill considerations (see Box 2-1) such as strength, consistency of association, and coherence.

Occupational exposure to chromium VI (CrVI) is associated with increased risk of respiratory cancer, as acknowledged by IARC and EPA. Chronic inhalation studies in animals also provide evidence

___________________

¹⁴ See https://ntp.niehs.nih.gov/go/711176.

¹⁵ See https://www.tceq.texas.gov/toxicology/esl/guidelines/about.

¹⁶ See https://www.tceq.texas.gov/downloads/toxicology/dsd/final/hexavalent_chromium.pdf.

that CrVI is carcinogenic. Although CrVI is a recognized hazard, EPA had not updated the CrVI inhalation unit risk factor (URF) since 1984 and more recent studies were available to inform dose-response assessment. This motivated the TCEQ update for CrVI in 2014. Two key studies of chromate worker cohorts in Painesville, Ohio (Crump et al., 2003), and Baltimore, Maryland (Gibb et al., 2000), were available. Both studies have extensive lung cancer mortality follow-up as well as documentation of historical CrVI exposure levels and were amenable to dose-response analyses. The effects of smoking and race were analyzed as covariates, and optimal exposure lag was also determined. Separate URFs were calculated for these two key studies, as well as the supporting study, before the preferred URF estimates from each key study were combined through a weighting procedure to derive the final CrVI URF. Among other examples of TCEQ synthesizing results across epidemiological studies are weighted URFs for nickel and inorganic arsenic.

In its assessment of oral exposures to CrVI, TCEQ relied on newly published information on the carcinogenic mode of action and consideration of non-linear, non-threshold and threshold approaches (Haney, 2015a,b,c). TCEQ adopted a reference dose approach based on compensatory crypt enterocyte hyperplasia induced by chronic villous toxicity. Unlike with the URF, epidemiological studies did not play a critical role. It was assumed that findings from NTP’s mouse drinking water study (NTP, 2008) were relevant to humans. Most of the integration across evidence streams pertained to mode of action and weight of evidence and epidemiological studies did not play a critical role. The reference dose (RfD) was internally peer-reviewed and underwent public comment.

In closing, Haney highlighted some hazard identification challenges such as disparate results within an evidence stream, or disparate health endpoint results across evidence streams such as when animal study effects are not seen in epidemiological studies for any reason. The lack of tumor site concordance across species is a notable example. Hill considerations (see Box 2-1) such as biological gradient, consistency of association, and coherence can aid interpretation in these cases.

PANEL DISCUSSION

Bias Impacts and Triangulation Approaches

Schubauer-Berigan highlighted the need to assess the impact of different biases within and across epidemiology studies. She noted that although the direction and magnitude can vary across types of biases, there is imbalance in how they are considered. For instance, non-differential exposure misclassification and measurement error may bias toward the null whereas other factors that bias away from the null may be given greater emphasis. Lunn concurred, noting that even when both the direction and magnitude of bias can be predicted, studies may have multiple biases. The planned IARC workshop is expected to consider this important issue of multiple biases.

Lawlor highlighted being more explicit about different kinds of bias and expectations of their influence on results. A comparison across different approaches may reveal results that are directionally consistent. When results are inconsistent, however, additional considerations may be needed to determine whether different approaches, methods, and data sources address the same questions but have different sources of bias or address different questions. No study is perfect, including randomized controlled trials (RCTs), and studies differ in their underlying assumptions and strengths and weaknesses. “Comparing across different types of studies can potentially improve things,” she said. Nachman and others noted the need to integrate many areas of expertise, because identifying and understanding biases draws on information from different disciplines.

Samet commented on the “distinct challenges” to hazard identification and risk quantification. For quantifying risk, the question remains of whether to draw on one or multiple studies and examples of both approaches exist. Emerging concepts in biology might inform modeling. Building more mechanistically informed models might merit more discussion as mechanistic understanding advances, he said. Schubauer-Berigan agreed, noting that “with mechanistic evidence growing, both because we know more about the mechanisms of carcinogens and because bioassay evidence is getting scarcer, our

programs will evolve in that direction.” She highlighted improved endpoint measurements for some KCs as very helpful in evidence integration. Lunn concurred that mechanistic data offer value for assessing potential cohesiveness and providing great confidence in conclusions. “Putting together animal, human, and mechanistic evidence to help explain epidemiology is true triangulation. But we need to note limitations,” she cautioned, citing the nightshift work study. As noted by Haney, it can be challenging to draw conclusions when there are disparate results within one evidence stream or disparate health endpoint results across evidence streams. Considering the role of laboratory and mechanistic data, Samet called for more multidisciplinary coordination. Many scientists generate evidence that may be relevant in determining whether an exposure is a hazard or in characterizing its risks; sometimes researchers may act in a coordinated way and are part of the same research center. However, in general, such research is not coordinated to assure that uncertainties are explored, he observed. Lunn agreed, noting the need for transdisciplinary teams, coordination, and crosstalk.

Samet highlighted the 1964 Surgeon General’s report as an early example of triangulation that “would fit perfectly well with our conversations about bringing together different lines of evidence. The conclusion was not driven solely or in a large part even by the epidemiology, but what was available from other lines of evidence.” The narrative that provides the rationale for reaching conclusions is important for understanding the basis of the expert judgment. The results of such expert judgments are robust, Schubauer-Berigan noted, citing the reconfirmation in Group 1 of all the human carcinogens that were reevaluated by a completely different working group during the Monographs Volume 100 series. For most agents, the evidence had strengthened in the intervening years. As a different example, she noted that a 1985 workshop in Oxford, England, that assessed 10 Group 2B agents that were carcinogenic in rodents, but not in humans, concluded that these agents were putatively negative for carcinogenicity in humans. Nearly 40 years later, 3 agents had reached Group 1, and 4 agents had reached Group 2A with limited evidence in humans. Over time, studies improved, often bearing out what the animal studies found. These examples validate the important role of expert judgment, she said. Lunn noted that these examples reinforce earlier comments about overemphasis on biases away from the null and the underemphasis on biases toward the null.

Exposure Assessment

Among advances and emerging methodologies, exposure science needs to be taken into account in hazard identification and risk assessments, said Schubauer-Berigan, calling these the “most critical aspects of study quality and informativeness,” in Monographs meetings. Exposure science allows judgments regarding the adequacy of exposure contrast for assessing effects and can reduce uncertainty about other aspects, such as understanding whether a target organ is likely affected. There can be a tradeoff between sample size and the potential precision of an exposure assessment. Nachman concurred, noting that the risk factors are often known but the association among a risk factor, a potential confounder, and the exposure of interest is not. Samet raised the possibility of having more precise biologically relevant indicators of exposure and highlighting some advances in study design, such as recent population-based cohort studies on the health effects of air pollution as an example that brought together data from millions of people.

Lunn concurred, noting that people are exposed to complex mixtures and exposure scenarios, which provides a rationale for moving toward real-world exposures in studies. Lunn highlighted a lack of data on interactions of psychosocial stressors and environmental agents. Samet raised concerns about cumulative risks as foundational to the broader discussion, and the need for understanding of risk drivers and how different populations may experience risk from agents. Considering the influence of nonchemical stressors on chemical health relationships or the combined action of multiple chemical stressors, Samet emphasized both the need to be broader and the potential usefulness of integrative indexes of population health such as life expectancy and all-cause mortality to understand community impacts. He noted a distinction from health impact assessments, that, for example, model emissions from a plant and apply a unit cancer risk coefficient for a particular tumor site. “We are going to have to be

broader,” he said. Haney noted that mechanistic information is critical to developing a more complete understanding about potential interaction effects of exposures to multiple chemicals, and to assessing cumulative risk in a realistic manner.

Publications

Nachman commented on the need to include relevant supplemental material, such as relations between exposure and confounder, in publications to facilitate systematic reviews. Schubauer-Berigan agreed, noting, “Greater consideration in the scientific and epidemiologic communities as to how critical these studies are for the evaluation of health hazards and encouraging the community to make better use of supplemental materials … would be very beneficial, for bias assessment, for many purposes.” Lawlor added that publication of the objectives and plans of a study in a protocol can promote progress through early engagement with a broader range of experts.