Limitations of evaluating risks of all cancers combined in etiological epidemiology research Authors: Julie E Goodman, Satori A Marchitti, Nicole S Olgun, David L Eaton Published: 2026 ABSTRACT AbstractThe International Agency for Research on Cancer (IARC) was the first public health agency to conclude that an agent – in this case, 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD) – could be carcinogenic based on associations with all cancers combined, even though there were no consistent associations with any particular cancer type. Most types of cancer each represent an etiologically distinct disease. Epidemiology associations with all cancers combined may be false-positive findings, which lack biological meaning and do not adequately consider confounders, or could mask true associations with specific and/or rare cancer types. We propose that, if investigators report results of etiological analyses with all cancers or groups of etiologically distinct cancers (or other diseases), they consider the heterogenous nature of cancer and that their observations are hypothesis-generating statistical associations without an immediate mechanistic inference.Keywords: All cancers combined, Epidemiology risk assessment, Key characteristics of carcinogens, Hallmarks of cancer Background Globally, cancer is the second leading cause of death [1]. Although cancer cells share certain defining characteristics, cancer is not a singular disease [2], [3]. As such, investigators should consider the limitations of analyses of cancers grouped together, particularly in etiological epidemiology studies, when interpreting results, as they can lead to erroneous conclusions. For example, in 1997, The International Agency for Research on Cancer (IARC) published Monograph Volume 69, in which the agency concluded that “the strongest evidence for the carcinogenicity of 2,3,7,8-TCDD [2,3,7,8-tetrachlorodibenzo-p-dioxin] is for all cancers combined, rather than for any specific site” [4]. The Monograph went on to say: There are few examples of agents which cause an increase in cancers at many sites; examples are smoking and ionizing radiation in the atomic bombing survivors (for which, however, there are clearly elevated risks for certain specific cancer sites). This lack of precedent for a multi-site carcinogen without particular sites predominating means that the epidemiological findings must be treated with caution; on the other hand, the lack of precedent cannot preclude the possibility that in fact 2,3,7,8-TCDD, at high doses, does act as a multi-site carcinogen [4]. To our knowledge, this is the first time that a public health agency concluded that an agent could be carcinogenic based on associations with all cancers combined, when there were no consistent associations with any cancer type in particular. Over a decade later, IARC [5] published Monograph volume 100, stating: Dioxin (2,3,7,8-tetrachlorodibenzo-para-dioxin, TCDD) was previously classified in Group 1, based on limited evidence of carcinogenicity in humans, sufficient evidence in rodents, and strong evidence in humans and animals for a mechanism via initial binding to the aryl-hydrocarbon receptor (AhR), which leads to changes in gene expression, cell replication, and apoptosis. There is now sufficient evidence from epidemiological studies for all cancers combined, making TCDD the first agent classified initially in Group 1 based on sufficient evidence of carcinogenicity in experimental animals and strong mechanistic data, to be later confirmed by increased cancer incidence in humans. IARC [5] concluded, “The evidence for specific cancers is strongest for lung cancer, soft-tissue sarcoma and non Hodgkin lymphoma [NHL], but confounding cannot be ruled out for lung cancer, while the findings on soft-tissue sarcoma are based on small numbers.” Interestingly, IARC [5] also acknowledged, “The lack of complete consistency among the cohorts and the weak effect detected in most of the positive studies, however, caution against a causal interpretation of the findings” for NHL and that, in addition to small numbers, there may be additional issues with confounders and outcome misclassification for soft-tissue sarcoma.“ These statements by IARC show that its conclusions from the epidemiology evidence rest almost entirely on findings for “all cancers” combined, with insufficient support for any individual type of cancer. In these Monographs, IARC set a precedent for future evaluations of dioxins. For example, US EPA relied on the 1997 evaluation, although a National Research Committee [6] noted “that the argument provided by EPA in the 2003 Reassessment to support its position that the epidemiological data met the criterion of ‘strong evidence of an association’ between dioxin exposure and cancer risk was unconvincing.” In addition, NTP [7] relied solely on the IARC evaluations of dioxins for its conclusions regarding epidemiology studies in humans. These IARC analyses also set a precedent for the interpretation and communication of results from other epidemiology studies that evaluate all cancers combined, or groups of cancers in aggregate. As discussed below, analyses of cancers grouped together do not account for the fact that most types of cancer represent etiologically distinct diseases. Therefore, epidemiology associations with all cancers combined – and sometimes some cancers grouped together – lack biological meaning and may be false positive findings. They also may not adequately consider confounders, or could mask true associations with specific cancer types. As such, they are not useful for causal determinations. Cancer cell characteristics Hanahan and Weinberg [8] proposed “six capabilities [that] are shared in common by most and perhaps all types of human tumors.” These characteristics, termed the “Hallmarks of Cancer” comprised the most commonly observed capabilities acquired by individual cells as they transitioned from normal cells into malignant tumor cells [9], [10], [2]. At the cellular level, these now-established characteristics include resisting cell death, evading growth suppressors, and enabling replicative immortality. These tumor cell features define cancer at its most basic level, thus allowing for “cancer” to be used as an umbrella term for the many types of individual cancer that have been identified. Subsequent to the development of the Hallmarks of Cancer, Smith et al. [11] noted that agents that cause cancer had no similar set of established characteristics or defining physicochemical properties, aside from perhaps genotoxicity. Thus, Smith et al. [11] developed a set of 10 key characteristics (KCs) of carcinogens during two 2012 IARC workshops. This KC paradigm is based on the concept that human carcinogens often demonstrate activity for at least one of ten KCs: is electrophilic; is genotoxic; alters DNA repair or causes genomic instability; induces epigenetic alterations; induces oxidative stress; induces chronic inflammation; causes immunosuppression; modulates receptor-mediated effects; causes immortalization; and alters cell proliferation, cell death, or nutrient supply. While the KCs are intended to provide the basis of a uniform approach “to facilitate systematic consideration of mechanistic evidence” to support cancer hazard identification [12], in practice, some investigators have since proposed that evidence for one or more KCs support carcinogenicity (e.g., DeMarini et al., 2025 [13]). As discussed below, these assertions do not hold. Though it has been established that tumor cells share common characteristics that serve to define them as cancer cells, the KC concept, based on empirical observations of certain chemical and biological properties, is an oversimplification with regard to the vast chemical space and to human health risk assessment in general. While it may be tempting to try to group all carcinogens together in the same way the Hallmarks of Cancer grouped tumor cells together, the KC concept does not take into account the complexity of the chemical world and living organisms. Evidence for any given KC does not automatically imply that an agent could (or will) cause cancer in living organisms, which are complex and dynamic systems with a high capacity for detoxification and biological adaptation [10]. For example, many compounds increase oxidative stress (KC #5), but so do normal physiological processes, such as pregnancy and aging [14], [15]. Becker et al. [15] characterized the use of KCs in isolation as too simplistic of an approach to address the complexity of cancer hazard identification, stating that the KCs, alone or in combination, were no better than chance in distinguishing carcinogens from non-carcinogens. While KC activity may be useful for interpreting the human relevance of observational epidemiological data, or understanding potential cancer modes of action for agents identified as carcinogens through animal testing, the inference that an agent that exhibits one or more KCs is a likely risk for any form of cancer in humans, much less all forms of cancer, would be misguided. In a similar manner, it is generally incorrect to assume that epidemiological or toxicological evidence showing a causal association between a given agent and a specific form of cancer would also represent a risk for multiple additional forms of cancer (or all cancers). Cancers as etiologically distinct diseases Although cancer cells share certain defining characteristics, cancers can arise almost anywhere in the body, and they behave differently depending on the organ of origin [2]. Accordingly, cancers are primarily categorized based on the cell type and the organ in which they first originated, even if the cancer ultimately metastasizes to other sites [16]. For example, lung cancers can be classified as either small cell lung cancer or non-small cell lung cancer [17]. Similarly, skin cancers can be classified as basal cell carcinomas, squamous cell carcinomas, or melanomas [18]. Most agents identified as carcinogens typically cause cancer in specific target tissues that they or their metabolites can access, with few (if any) carcinogens capable of causing all cancers (e.g., as noted above, even radiation, a ubiquitous carcinogen, is not capable of causing all types of cancers) [3]. Just as each cancer is etiologically distinct, and each carcinogen typically has target tissues, another way in which cancers are unique is in how they respond to treatment (e.g., radiation or immunotherapy); what is effective for one cancer or subtype is not necessarily effective for another, and no one treatment is effective for all cancers. While some common risk factors have been identified for certain cancers, such as sex (e.g., breast cancer, lung cancer) and alcohol use (e.g., cancers of the oral cavity, pharynx, larynx, esophagus, liver, colorectum, and female breast) [19], [20], these cancers all have additional independent risk factors that also increase risk (e.g. genetic mutations, diet, radiation exposure). Age is the only risk factor identified that applies to all cancers, as it “can be considered a surrogate measure for the complex biological processes associated with aging” [21]. There are certain instances for which there is some logic to evaluating related cancers as a group (e.g., respiratory cancers), particularly when they have the same mode of action. However, when the risk factors and etiology differ among cancer types, it does not make sense to combine cancers either as a whole or in subgroups (e.g., lymphohematopoietic cancers), as if they were the same disease, in statistical analyses. Just as it would not be meaningful to analyze the risk for the autoimmune disease lupus and colorectal cancer together, it does not make biological sense to combine all cancers in a statistical analysis. Statistical considerations One might think that aggregating cancer types results in increased statistical power, but that is not the case. If a statistically significant association with all cancers is reported, one may hypothesize that an agent is carcinogenic, but it cannot be determined with any degree of certainty with which (if any) cancer(s) the agent is associated, or whether the result is a false-positive finding. It is also difficult, if not impossible, to adjust for all potential confounders for different cancer types when they are combined in an analysis. In addition, if there is no association observed, it is possible that a true association with one or a few cancers is being masked by null or opposing associations with other cancer types. This is particularly true if associations with common cancers are null and a true association exists with a rare cancer, as the numbers of individuals with the rare cancer would likely be much lower than those of common cancers. For example, the identification of an association between relatively high occupational exposure to vinyl chloride (VC) and the development of angiosarcoma of the liver was possible only because the background incidence of angiosarcoma of the liver is incredibly rare in the absence of VC exposure [22]. Occupational monitoring of VC workers in the early 1970s identified 3 cases of angiosarcoma in a worker population of several thousand. Several additional studies of workers exposed occupationally to VC identified 13 cases of liver angiosarcoma in the United States, with an estimated risk of VC workers for this rare cancer that was 400 times higher than that of the general population [23], [24], [25]. It was the rarity of the disease that allowed detection in a relatively small cohort of workers; it would not have been detected in an all-cancer analysis. This concept is also demonstrated in studies of workers involved in the “Pliofilm” process, during which natural rubber was dissolved in refined benzene and processed by being spread out on a conveyor belt system. Rinsky et al. [26] reported there were no associations between benzene exposure and all cancers combined (SMR = 1.03, 95% CI: 0.88–1.20), but associations with all lymphohematopoietic cancers (SMR = 1.64, 95% CI: 1.06–2.42) and leukemia (SMR = 2.47, 95% CI: 1.38–4.07) were observed. The analysis of all cancers “hid” the associations with lymphohematopoietic cancers and leukemia. According to IARC [27], “Treating all leukaemias as a single entity is a limitation of several…studies, given the evidence of etiological heterogeneity.” In this case, associations with lymphohematopoietic cancers did not hide associations with leukemia, although the risk estimate for the former was smaller. Other studies with this and other cohorts indicate the estimated risks were likely attributable to acute myelogenous leukemia (AML) or possibly its precursor, myelodysplastic syndrome (e.g., Hayes et al., 1997 [28]; Rhomberg et al., 2016 [29]; Linet et al., 2019 [30]). The opposite occurred with studies of dioxin, as noted above. In this case, associations were observed with all cancers combined, but not any individual cancers. It is hard to know what led to this association with all cancers, but it appears to have led to conclusions about the carcinogenicity of dioxin, when these conclusions should have been based on associations (or more accurately, the lack thereof) with specific cancer types. There are also several substances that likely cause certain cancers, while preventing others. As noted above, alcohol consumption has been associated with cancers of the oral cavity, pharynx, larynx, esophagus, liver, colorectum, and female breast (19; 20), but inversely associated with kidney cancer [31]. Similarly, tamoxifen has been shown to prevent breast cancer, but can increase endometrial cancer risk [32]. Studies of all cancers combined could lead to positive and negative associations essentially cancelling each other out. Conclusions While our focus is on analyses of all-cancers combined, we note that this argument holds for any group of etiologically distinct diseases that are grouped together (e.g., cardiovascular mortality, respiratory mortality, kidney disease). To our knowledge, dioxin is the only chemical for which IARC has a carcinogenicity classification based on epidemiological associations with all cancers. However, we continue to see associations with all cancers, or groups of unrelated cancers, reported in peer-reviewed epidemiology studies. Given the issues with interpreting either positive or null results, we propose that if investigators report results of etiological analyses with all cancers or groups of etiologically distinct cancers (or other diseases), they consider the heterogenous nature of cancer and that their observations are hypothesis-generating statistical associations without an immediate mechanistic inference. CRediT authorship contribution statement Julie E. Goodman: Writing – review & editing, Writing – original draft, Funding acquisition, Conceptualization. Satori A. Marchitti: Writing – review & editing, Writing – original draft. Nicole S. Olgun: Writing – review & editing, Writing – original draft. David L. Eaton: Writing – review & editing, Writing – original draft. Declaration of competing interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Gradient received funding from the American Beverage Association (ABA). DLE received no funding for his work on this paper. This paper was written during the authors' normal course of business and represents the professional opinions of the authors alone. Acknowledgments The authors thank Dr. David Savitz for his thoughtful comments on this paper. We also thank Eric Fischbach and Sarah Gulick for their administrative and editorial assistance. References