Arsenic Epidemiology and Cancer Risk

RoseyFields evaluates the human epidemiological evidence for Arsenic as follows: this summary is based on the U.S. EPA’s 2025 IRIS Toxicological Review of Inorganic Arsenic (final report), which conducted a systematic review and evidence synthesis of the epidemiological literature, and on the underlying peer-reviewed primary studies identified and evaluated within that review. Citations below are to the original peer-reviewed epidemiological publications rather than to the IRIS document itself.

1. Study Populations and Data Sources

The human evidence base for inorganic arsenic is unusually large and geographically diverse, encompassing several distinct categories of study populations and exposure-assessment approaches. Understanding this diversity is important because convergent findings across populations that differ in genetic background, diet, co-exposures, exposure route, exposure magnitude, and study design substantially strengthens confidence that observed associations reflect a causal relationship rather than an artifact of any single population’s characteristics or a single study design’s biases.

1.1 High-Exposure Community/Contamination-Site Cohorts

A substantial portion of the evidence base derives from populations with naturally elevated arsenic in groundwater, principally in Bangladesh, Taiwan, China, Chile, and Mexico, where drinking-water arsenic concentrations have historically ranged from tens to, in some areas, over 900 μg/L — up to several hundred-fold higher than typical U.S. background levels. The Bangladesh Health Effects of Arsenic Longitudinal Study (HEALS) cohort (e.g., Chen et al., 2011b; Argos et al., 2014) and the Matlab-based Maternal and Infant Nutritional Intervention (MINIMat) cohort (e.g., Hamadani et al., 2010, 2011) are examples of large, prospective cohorts in high-exposure settings. Southwestern Taiwan represents a distinct and heavily studied population, in part because of a well-documented historical “natural experiment” in which arsenic-contaminated artesian well water use was discontinued following a public health intervention, allowing before/after and dose-response analyses with well-defined exposure periods and limited population migration (e.g., Chen et al., 1996; Hsueh et al., 1998; Wu et al., 2013). Northern Chile (Antofagasta region) similarly experienced a sharp rise and subsequent decline in municipal drinking-water arsenic concentrations, providing another natural-experiment population used to study both cancer and noncancer outcomes with long latency (e.g., Ferreccio et al., 2000; Steinmaus et al., 2013).

1.2 General-Population and Lower-Exposure Cohorts (Including U.S. Populations)

A second major category comprises general-population cohorts, often in the United States and Europe, with substantially lower background arsenic exposure (typically <100 μg/L, and frequently <20 μg/L, in drinking water), which are essential for establishing that adverse associations are not confined to the extreme historical exposure levels seen in South and Southeast Asia. The U.S. Strong Heart Study, a prospective cohort of American Indian adults in Arizona, Oklahoma, and North and South Dakota (e.g., Moon et al., 2013; Grau-Pérez et al., 2017), and the New Hampshire Birth Cohort and New Hampshire Skin Cancer Study (e.g., Farzan et al., 2015a, 2015b, 2016; Signes-Pastor et al., 2019a) are prominent U.S. examples using biomarker (urinary, toenail) exposure assessment at low-to-moderate exposure levels. European cohort and registry-linkage studies — for example, the Danish Diet, Cancer and Health cohort (Monrad et al., 2017; Bräuner et al., 2014) and an Italian municipal cohort linking geocoded residential drinking-water histories to mortality outcomes (D’Ippoliti et al., 2015) — similarly evaluate arsenic effects at exposure levels directly relevant to U.S. regulatory contexts.

1.3 Occupational Cohorts

Occupational cohorts, generally involving inhalation exposure in smelting, mining, glass manufacturing, and pesticide-application settings, contributed importantly to the historical evidence base establishing arsenic as a human lung carcinogen and continue to inform hazard identification, although the current IRIS assessment’s quantitative dose-response analysis is centered on oral exposure given the predominance of the oral route for general-population exposure and the greater availability of recent oral-exposure epidemiological data.

1.4 General-Population Biomonitoring and Birth-Cohort Studies

Cross-sectional biomonitoring studies and prospective birth cohorts that measure maternal and/or infant biomarkers (urine, cord blood, toenail) at defined points in pregnancy and childhood constitute a further major category, particularly important for the fetal, newborn/infant, and neurodevelopmental outcome literature. These include cohorts in Bangladesh (Hamadani et al., 2010, 2011; Vahter et al., 2020), Taiwan, China, Mexico, Spain, and the United States, using a range of biomarkers collected at various points across gestation and early childhood.

1.5 Value of Using Multiple, Diverse Populations

The use of multiple, independent populations that differ in genetic ancestry (and therefore in arsenic-methylation capacity, governed substantially by AS3MT genotype), dietary patterns (including seafood consumption, which affects urinary arsenic speciation), background disease rates, co-exposures (e.g., smoking prevalence, other metals), exposure route and magnitude, and study design (cohort, case-control, cross-sectional, and “natural experiment” ecological designs) strengthens causal inference in several specific ways. First, replication of an association across populations with different sources and patterns of confounding makes it less plausible that a single unmeasured confounder explains the pattern in every study. Second, populations spanning a wide range of exposure levels — from the very high historical exposures in Taiwan and Bangladesh to the low exposures typical of U.S. and European populations — allow evaluation of whether a dose-response gradient is present across the full exposure range, and specifically whether effects persist at exposure levels below 100 μg/L that are most relevant to current U.S. regulatory decision-making. Third, studies using markedly different exposure-assessment methods (drinking water measurements versus urinary, toenail, hair, or blood biomarkers) and different outcome-ascertainment methods (death certificates, ICD-coded hospitalizations, clinical/biomarker diagnosis, standardized neurocognitive testing) that nonetheless converge on similar associations provide evidence that the finding is not an artifact of any one measurement approach. Finally, the arsenic evidence base includes several “natural experiment” ecological studies (e.g., the southwest Taiwan and northern Chile exposure cessation cohorts) that, despite lacking individual level exposure data, provide unusually well-defined exposure periods, limited population migration, and large sample sizes that reduce the influence of individual-level confounding — a distinctive strength for arsenic that is not typical of ecological designs generally, and one that adds meaningfully to the overall weight of evidence when combined with individual-level cohort and case-control data.

2. Overview of Health Outcomes Investigated, Organized by Body System

The epidemiological literature on inorganic arsenic spans several hundred publications addressing a wide range of health outcomes. The table below organizes the principal outcome categories evaluated in EPA’s 2025 evidence synthesis by body system and evidence strength, using EPA’s own evidence-integration terminology: “robust” evidence supports a conclusion that the evidence demonstrates causation; “moderate” evidence supports a conclusion that the evidence indicates (likely) causation; and “slight” evidence, while suggestive, does not support a stronger conclusion, generally because of unexplained inconsistency across studies and/or imprecision (wide confidence intervals spanning the null). Bladder, lung, and skin cancer and skin lesions are treated separately below because they were designated “accepted hazard” outcomes prior to this assessment (i.e., their hazard status was not re-litigated in the current evidence synthesis, which instead focused on updating the quantitative dose-response estimates).

2.1 Cancer (Accepted Hazard Outcomes)

• Bladder cancer — established human carcinogen target organ; extensive case-control and cohort literature with quantitative dose-response modeling (see Section 3.4).
• Lung cancer — established human carcinogen target organ; extensive case-control and cohort literature, including occupational (inhalation) and environmental (oral) exposure studies, with quantitative dose-response modeling (see Section 3.4).
• Skin cancer (basal cell and squamous cell carcinoma) — among the earliest and most consistently reported arsenic-associated cancers in high-exposure populations; classified as an accepted hazard, though not a focus of updated quantitative dose-response modeling in the current assessment.
• Other cancers investigated with more limited or inconsistent evidence: kidney, liver, prostate, and other internal cancers have been examined in several cohort and ecological mortality studies (e.g., D’Ippoliti et al., 2015; García-Esquinas et al., 2013), with some positive associations reported, particularly in high-exposure populations, but the evidence for these sites is less extensive and less consistently modeled quantitatively than for bladder, lung, and skin cancer.

2.2 Cardiovascular / Circulatory System

• Ischemic heart disease (IHD)/coronary heart disease — STRONG, ROBUST evidence (see Section 3.1).
• Hypertension — STRONG, ROBUST evidence (see Section 3.2).
• Cerebrovascular disease/stroke — SUGGESTIVE BUT INCONSISTENT (moderate) evidence; some well conducted studies report positive associations, but findings across the full literature are mixed, including inverse and null results.
• Peripheral vascular disease (including Blackfoot disease in Taiwan) — well documented in high-exposure Taiwanese populations but with limited generalizability and quantitative modeling for lower-exposure settings; considered supportive/coherence evidence rather than a primary quantitatively modeled outcome.
• Subclinical/intermediate cardiovascular endpoints (carotid intima-media thickness, QT prolongation, endothelial dysfunction biomarkers) — generally consistent positive associations, used as coherence and biological-plausibility evidence supporting the IHD and hypertension findings rather than as independently modeled outcomes.

2.3 Endocrine / Metabolic

• Type 2 diabetes — STRONG, ROBUST evidence (see Section 3.3).
• Metabolic syndrome and its individual components (elevated fasting glucose, dyslipidemia, hypertension, waist circumference) — SUGGESTIVE BUT INCONSISTENT; some individual components (e.g., fasting plasma glucose) show positive associations in specific cohorts (Spratlen et al., 2018), while composite metabolic syndrome outcomes are frequently null.
• Gestational diabetes — investigated in a smaller literature with generally positive but less extensively replicated findings (e.g., Farzan et al., 2016).
• Type 1 diabetes — investigated in a small number of studies (e.g., Grau-Pérez et al., 2017 SEARCH study); evidence is limited and not a focus of quantitative dose-response modeling.

2.4 Reproductive, Fetal, Newborn, and Infant Outcomes

• Fetal and infant mortality — SUGGESTIVE, MODERATE evidence, strongest in high-exposure South Asian cohorts.
• Reduced birth weight — SUGGESTIVE, MODERATE evidence, generally consistent inverse associations across diverse geographic regions (see Section 3.5).
• Fetal growth parameters (head circumference, femur length, biparietal diameter, chest circumference) — SUGGESTIVE BUT INCONSISTENT (slight) evidence, with conflicting results even among studies with overlapping exposure levels.
• Prematurity/preterm birth — SUGGESTIVE BUT INCONSISTENT (slight); most studies report no association, though a minority report significant positive associations, some with a dose-response pattern.
• Postnatal growth — SUGGESTIVE BUT INCONSISTENT (slight); limited number of studies with mixed findings.

2.5 Neurodevelopmental / Nervous System

• Developmental neurocognitive deficits (IQ, memory, learning, executive function) — SUGGESTIVE, MODERATE evidence, generally consistent across diverse populations and outcome measures (see Section 3.6).
• Social, behavioral, and emotional effects (including autism-spectrum indicators) — SUGGESTIVE BUT INCONSISTENT (slight); roughly half of case-control studies examining autism found an association, with more mixed findings for other behavioral/emotional endpoints.
• Motor effects — SUGGESTIVE BUT INCONSISTENT (slight); most studies from diverse regions identify an association with adverse motor skills/scores, but the evidence base has unexplained inconsistency across studies.
• Developmental delays — SUGGESTIVE BUT INCONSISTENT (slight); based on a small evidence base of nail and urinary biomarker studies concentrated in Taiwan.

2.6 Other Outcomes Investigated

• Immune system effects — investigated in a smaller epidemiological literature (in addition to the animal/mechanistic literature); findings are limited and not incorporated into EPA’s current quantitative dose response outcomes.
• Respiratory disease (nonmalignant) — investigated in several occupational and environmental cohorts, with some positive associations for chronic respiratory symptoms and reduced pulmonary function, but this outcome was not selected as a priority outcome for updated quantitative synthesis in the current assessment.
• Renal/kidney function (noncancer) — investigated in a more limited literature; findings are inconsistent and this outcome was not developed as a quantitatively modeled endpoint in the current assessment.
• Hepatic (liver) noncancer effects — investigated to a lesser degree in the epidemiological literature relative to the animal literature; human data are more limited and largely descriptive.
• Skin lesions (hyperkeratosis, hyperpigmentation) — classified as an accepted hazard based on a long, consistent historical literature, particularly from high-exposure South Asian populations, though these lesions are generally treated as a well-established but non-priority outcome for updated dose-response modeling in the current assessment relative to cancer, DCS, diabetes, fetal/infant, and neurodevelopmental outcomes.

3. Strong-Evidence Outcomes: Key Studies, Exposure Metrics, and Effect Sizes

This section details the outcomes for which EPA’s evidence synthesis reached a “robust” (cardiovascular disease, hypertension, diabetes) or “moderate” (fetal/infant mortality, birth weight, developmental neurocognitive effects) evidence-integration judgment, along with the accepted-hazard cancer outcomes, presenting representative key studies, the exposure metrics used, effect sizes and direction, and the consistency of findings across populations.

3.1 Ischemic Heart Disease (Robust)

Some of the strongest evidence for arsenic’s cardiovascular effects comes from large prospective cohort and case control studies with individual-level exposure data conducted in multiple countries and diverse populations. In the U.S. Strong Heart Study (n = 3,575 American Indian adults), Moon et al. (2013) compared the highest and lowest quartiles of urinary arsenic (>15.7 vs. <5.8 μg/g creatinine) and found a hazard ratio for incident IHD of 1.71 (95% CI: 1.19, 2.44; p for trend <0.001), with a statistically significant dose-response gradient. In the Bangladesh HEALS cohort (n = 11,746), Chen et al. (2011b) found a dose-response relationship between baseline well-water arsenic and IHD/heart-disease mortality, with hazard ratios of 1.00 (reference), 1.22 (0.65–2.32), 1.35 (0.71–2.57), and 1.92 (1.07–3.43) (p = 0.0019 for trend) across increasing water-arsenic quartiles (0.1–12.0, 12.1–62.0, 62.1–148.0, and 148.1–864.0 μg/L).

In southwestern Taiwan, Chen et al. (1996), a case-control study nested in the Blackfoot-disease-endemic region, reported relative risks of 2.5, 4.0, and 6.5 for cumulative arsenic exposure of 0.1–9.9, 10.0–19.9, and ≥20.0 mg/L years, respectively, compared with unexposed referents. Hsueh et al. (1998) reported odds ratios of 2.55 (95% CI: 1.02–6.37) and 2.89 (95% CI: 1.01–8.29) for 13–29 and ≥30 years of high-arsenic artesian well-water consumption, respectively, relative to <13 years. Importantly, comparable associations have also been observed in populations with far lower exposure: in a case-cohort study using a geospatial drinking-water exposure model in the San Luis Valley, Colorado (n = 555; exposure range 10–100 μg/L), James et al. (2015) reported an IHD hazard ratio of 1.38 (95% CI: 1.09, 1.78) per 15 μg/L increase in lifetime time-weighted average exposure. In rural Texas, Gong and O’Bryant (2012) found coronary heart disease associated with low-level groundwater arsenic exposure (mean 6.2 μg/L; OR: 1.10, 95% CI: 1.00–1.21). In Italy, D’Ippoliti et al. (2015), following 165,609 residents of 17 municipalities exposed on average to 19.3 μg/L for 39.5 years, found associations of both first-year and cumulative arsenic exposure with IHD and coronary atherosclerosis in both sexes after adjustment for age, calendar period, socioeconomic status, smoking, and radon. A toenail-biomarker study in the U.S. New Hampshire Skin Cancer Study population (Farzan et al., 2015a) found the strongest evidence of an IHD mortality association among long term smokers specifically, illustrating a potential effect-modification pattern relevant to co-exposure considerations (Section 4).

Overall, EPA’s evidence synthesis rated the IHD/cardiovascular disease evidence “robust,” citing consistency within and across study types (including meta-analyses), dose-response gradients observed in most studies, large effect magnitudes (odds ratios >4–6 in some high-exposure studies), and coherence with related endpoints such as hypertension and carotid atherosclerosis, with no identified factors that decreased certainty in the evidence base.

3.2 Hypertension (Robust)

Cohort studies across geographically diverse populations report generally positive associations between arsenic exposure and hypertension or blood pressure, more pronounced at higher exposure (>50 μg/L), with a dose-response gradient observed in most studies and, in some studies, large odds ratios (2–4-fold). In the Bangladesh HEALS cohort, associations with elevated systolic and diastolic blood pressure were observed (Jiang et al., 2015), and pulse pressure associations were observed in subgroups with low nutrient intake (Chen et al., 2007). In the U.S. New Hampshire Birth Cohort, arsenic exposure (mean well-water concentration 4.3 μg/L) was associated with increased blood pressure among pregnant women (Farzan et al., 2015b), and cross-sectional associations with blood pressure were also observed among children in Mexico exposed to 3–135 μg/L drinking-water arsenic (Osorio-Yáñez et al., 2015). Coherence is provided by findings of QT-interval prolongation (a repolarization abnormality that co-occurs with hypertension and left ventricular hypertrophy) associated with arsenic exposure in multiple studies. As with IHD, EPA’s evidence synthesis rated the hypertension evidence “robust,” with no factors identified that decreased certainty.

3.3 Type 2 Diabetes (Robust)

The diabetes literature is large (112 identified publications; 57 of medium or high confidence) and generally positive and consistent, with hazard/odds/relative-risk ratios in the highest exposure categories (typically compared with a reference group at or below 10 μg/L) usually around 2, and dose-response gradients observed in several though not all studies. In a prospective cohort in an arseniasis-endemic village in Taiwan, Tseng et al. (2000) found a relative risk of 2.1 (95% CI: 1.1, 4.2) for diabetes incidence at cumulative drinking-water exposures ≥17,000 μg/L-years. In Denmark, Bräuner et al. (2014) found an incidence rate ratio of 1.03 (95% CI: 1.01, 1.06) per 1 μg/L increase in average drinking-water arsenic, rising to an IRR of 1.19 (95% CI: 1.09, 1.31) in the highest exposure quartile (>1.82 μg/L) compared with the lowest (<0.57 μg/L) — an association observed at exposure levels far below the 100 μg/L benchmark historically used to define “high” arsenic exposure. In the U.S. Strong Heart Family Study, Grau-Pérez et al. (2017) reported a hazard ratio for incident diabetes of 1.57 (95% CI: 1.18, 2.08) per interquartile-range increase in urinary arsenic among participants without baseline prediabetes, with arsenic metabolism efficiency (influenced by AS3MT genotype and vitamin B intake) modifying the magnitude of the association with insulin resistance. Not all studies were positive: two Norwegian studies (Hansen et al., 2017; Simić et al., 2017) found no association, and results for metabolic syndrome as a composite outcome were largely null (Chen et al., 2012a), even though EPA’s evidence synthesis rated the core type 2 diabetes evidence itself “robust,” citing consistency across diverse populations and study designs, dose-response evidence in several studies at relatively low exposures (<100 μg/L), and coherence with the more limited type 1 and gestational diabetes literature.

3.4 Bladder and Lung Cancer (Accepted Hazard)

Bladder and lung cancer are the two cancer sites with the most extensive epidemiological literature and the most fully developed quantitative dose-response analyses in the current IRIS assessment; both are longstanding “accepted hazard” outcomes, consistent with prior classifications by IARC, WHO, NTP, and EPA (see Section 4). For bladder cancer, EPA’s dose-response meta-analysis drew on 11 case-control and cohort data sets spanning the U.S., Taiwan, Argentina, Chile, and Finland, including Chen et al. (2010b) (a cohort study from northeastern Taiwan finding elevated urinary tract cancer risk with increasing arsenic exposure), Steinmaus et al. (2003) and Bates et al. (2004) (western U.S. and Argentina case-control studies, respectively, both reporting positive drinking-water arsenic bladder cancer associations), Bates et al. (1995) (an earlier western U.S. case-control study establishing the association), Steinmaus et al. (2013) (documenting persistently elevated bladder cancer risk in northern Chile some 40 years after peak exposure and subsequent exposure cessation), and two urinary-biomarker case-control studies from Taiwan, Wu et al. (2013) and Chang et al. (2016), both linking urinary arsenic to urothelial carcinoma risk. For lung cancer, six studies were selected for meta-analysis, including three cohort studies — Chen et al. (2010a) (histological subtypes of lung cancer in relation to well-water arsenic and consumption characteristics in northeastern Taiwan), Argos et al. (2014) (lung disease mortality in the Bangladesh HEALS cohort), and García Esquinas et al. (2013) (cancer mortality in the U.S. Strong Heart Study) — and three case-control studies, including Ferreccio et al. (2000) (the foundational northern Chile study establishing a dose-response relationship between drinking-water arsenic and lung cancer) and Dauphiné et al. (2013) (a case-control study in California and Nevada).

Across this literature, effect sizes for bladder and lung cancer at the highest arsenic exposure categories are generally substantial (often several-fold elevated risk in the highest-exposure historical cohorts, such as northern Chile and southwestern Taiwan), with evidence of a dose-response gradient extending down to exposure levels well below 100 μg/L in some U.S. and biomarker-based studies. Both cancer sites also benefit from coherence with mechanistic evidence (Section 5, toxicology summary) and from the fact that arsenic’s carcinogenicity at these sites has been independently recognized by multiple national and international health agencies for decades (see Section 4).

3.5 Birth Weight and Fetal/Infant Mortality (Moderate)

Thirty-six medium- or high-confidence studies from diverse geographic regions (including China, the United States, Chile, and Bangladesh), using a variety of exposure-assessment methods, provide generally consistent evidence of statistically significant and non-significant inverse associations between arsenic exposure and birth weight. In an urban Chinese population, Guan et al. (2012) found a statistically significant inverse association between maternal blood arsenic (median 5.30 μg/L) and birth weight (β = −0.19 kg; p = 0.015). Several other Chinese cross-sectional studies similarly reported inverse associations between maternal or cord blood arsenic and birth weight, though a number of studies from Bangladesh, Mexico, and Romania reported associations whose confidence intervals included the null, illustrating the imprecision noted in EPA’s evidence-profile table for this outcome. Thirteen medium- or high-confidence cohort and cross-sectional studies, concentrated in Bangladesh and India (where drinking-water arsenic commonly exceeds 200 μg/L) but supplemented by lower-exposure studies elsewhere, provide the strongest evidence for fetal and infant mortality, with generally consistent findings across geographic regions, study types, and outcome metrics, though a dose-response gradient was observed in only some of these studies. EPA rated both birth weight and fetal/infant mortality “moderate,” reflecting consistency across diverse populations and studies of medium or high confidence, offset by imprecision in some individual studies (wide confidence intervals including the null).

3.6 Developmental Neurocognitive Effects (Moderate)

Forty-five medium- or high-confidence cohort and case-control studies evaluated developmental neurocognitive outcomes (learning, memory, verbal comprehension, executive function, and related domains) in children chronically exposed to a range of arsenic levels across Bangladesh, Taiwan, China, Nepal, the United States, Spain, Tanzania, and Canada, with generally consistent evidence of cognitive deficits across diverse populations, exposure metrics, and outcome-assessment tools. In the Bangladesh MINIMat cohort, Hamadani et al. (2011) followed children to age 5 and found a statistically significant inverse association between verbal IQ and both maternal urinary arsenic at gestational week 8 (β = −0.9, 95% CI: −1.7, −0.13) and child urinary arsenic at 1.5 years (β = −0.9, 95% CI: −1.7, −0.10); when stratified by sex, the association with decreased verbal and full-scale IQ was significant in girls but not boys. In an earlier assessment of the same cohort at 18 months, however, Hamadani et al. (2010) found no association between maternal or infant urinary arsenic and mental development — illustrating that effects in this literature may emerge or strengthen at later ages of assessment, a pattern the study authors themselves noted. In a separate Bangladesh cohort, Vahter et al. (2020) found that maternal urinary arsenic in early (but not late) pregnancy was inversely associated with full developmental scores and verbal comprehension at 10 years of age, with effect estimates for the highest exposure quintile of approximately −6.0 points (95% CI: −10.2, −1.77) relative to the lowest quintile — a pattern of stronger effects for early-gestation exposure that is broadly consistent with the greater vulnerability of the fetus to inorganic arsenic before maternal methylation capacity increases later in pregnancy (Section 1).

EPA rated the developmental neurocognitive evidence “moderate,” citing consistency across multiple geographic regions, exposure-assessment metrics, and outcome measures, and coherence with related evidence for social/behavioral/emotional effects, motor effects, and developmental delays, offset by imprecision in some studies and unexplained inconsistency in others (e.g., the null findings in early-childhood assessments in the Bangladesh MINIMat cohort noted above, which contrast with positive findings at later ages in the same and other cohorts).

4. Study Quality Considerations

EPA’s evidence synthesis incorporated formal risk-of-bias evaluation of each epidemiological study across seven domains (selection, confounding, performance, attrition, detection, selective reporting, and other), rating each study’s overall confidence as high, medium, low, or uninformative, with only medium- and high-confidence studies carried forward into the qualitative hazard synthesis and quantitative dose-response analyses discussed above. Several recurring study-quality issues are particularly important to understanding the strength and limits of specific findings.

4.1 Confounding

Outcome-specific core confounders were defined a priori for each major outcome category: for diseases of the circulatory system, smoking, sex, age, hypertension, cholesterol, obesity/BMI, and socioeconomic status; for diabetes, smoking, sex, age, BMI/waist circumference, antihypertensive medication use, physical activity, and socioeconomic status; for fetal, newborn, and infant outcomes, maternal smoking and alcohol use, maternal age, race/ethnicity, socioeconomic status, infant sex, parity, gestational age, and maternal BMI; and for neurodevelopmental outcomes, maternal smoking, sex, age, socioeconomic status, and birth weight. Smoking in particular was treated as an essential confounder: because tobacco itself is a source of arsenic exposure and independently affects cardiovascular disease, diabetes, and reproductive/developmental outcomes, studies that did not adequately address smoking were generally excluded from dose-response consideration where smoking rates in the source population were expected to be non-negligible. A further complexity specific to arsenic is that certain risk factors — for example, hypertension and diabetes in relation to IHD, or low birth weight in relation to neurodevelopment — may lie on the causal pathway between arsenic exposure and the outcome of interest rather than acting purely as confounders; adjusting for such variables (as mediators) can therefore understate rather than correctly estimate the total effect of arsenic exposure, and EPA’s evidence synthesis specifically favored dose response models that did not over-adjust for likely mediators.

4.2 Reverse Causation and Temporality

Cohort studies are generally less susceptible to reverse-causation concerns because exposure is measured prospectively, before outcome ascertainment. Cross-sectional and some case-control designs are more vulnerable to this concern, but EPA’s evaluation noted that many arsenic cross-sectional studies were conducted in populations with stable, long-duration (5–10+ year) exposure to a relatively constant arsenic source (e.g., a fixed community well), which increases confidence that a concurrent exposure measurement is a reasonable proxy for the relevant etiologic exposure window, particularly for outcomes without long latency. Reverse causation is a more substantial concern where a health outcome could plausibly alter arsenic handling or excretion (e.g., renal impairment altering urinary arsenic elimination), which is one reason EPA’s evidence synthesis and dose-response modeling placed emphasis on drinking-water exposure metrics and on biomarker studies with documented long-term temporal stability (e.g., citing evidence that drinking-water and urinary arsenic concentrations in stably exposed populations remain consistent for upwards of a decade).

4.3 Exposure Misclassification

The arsenic literature uses a range of exposure-assessment approaches, each with distinct limitations. Drinking water concentration alone, without individual water-intake data, is subject to nondifferential misclassification that generally biases effect estimates toward the null. Urinary, blood, hair, and toenail biomarkers better reflect internal dose and integrate multiple exposure routes and sources, but urinary and blood arsenic have short biological half lives (approximately 4 days and a few hours, respectively), such that a single spot measurement may not represent longer-term exposure unless the population has stable, ongoing exposure; hair and nail biomarkers reflect longer exposure windows but are subject to external surface contamination, a concern that is somewhat less pronounced for nails than for hair. Urinary biomarker studies also require correction for urine dilution (via creatinine or specific gravity adjustment); each correction method has its own limitations (e.g., urinary creatinine is influenced by age, sex, muscle mass, BMI, and diabetes status, while specific-gravity correction is affected by renal, metabolic, and cardiovascular disease), and the choice of correction method has been shown in at least one study (Kuo et al., 2022) to affect statistical inferences for cardiovascular and metabolic outcomes specifically. A further exposure misclassification issue specific to arsenic is dietary co-exposure to nontoxic organic arsenicals (arsenobetaine, arsenosugars) from seafood: because both organic and inorganic arsenic species are eliminated primarily via urine, populations with high seafood intake can have elevated total urinary arsenic that substantially reflects nontoxic organic forms rather than the toxicologically relevant inorganic species and its methylated metabolites, and studies that do not speciate urinary arsenic or otherwise account for seafood intake in such populations are at risk of exposure misclassification that generally biases associations toward the null.

4.4 Co-Exposure to Related Chemicals

Beyond organic arsenicals from seafood, several studies in the literature evaluate populations with potential co exposure to other metals (e.g., lead, cadmium, fluoride) or to arsenic in combination with other groundwater contaminants; co-exposure is a particular consideration for some developmental neurotoxicity studies (e.g., combined fluoride-arsenic exposure studies) and for some cardiovascular and renal studies. Where co-exposures were measured and adjusted for, findings for arsenic generally persisted as independent associations, but not all studies characterized or controlled for such co-exposures, which introduces some residual uncertainty into individual study findings, particularly in regions where multiple groundwater contaminants co-occur.

4.5 Overall Implications for Confidence

Taken together, these study-quality considerations mean that individual studies vary in their reliability, but EPA’s evidence synthesis addressed this by relying primarily on medium- and high-confidence studies (i.e., those judged to have adequately addressed the domains above), by explicitly favoring studies and analytic approaches that controlled for core confounders and avoided over-adjustment for likely mediators, and by placing the greatest weight on findings that were consistent across studies employing materially different exposure-assessment approaches, populations, and potential sources of bias — a form of triangulation that is particularly valuable given that no single study design or exposure metric is free of limitations. Where individual studies point in different directions (e.g., stroke, fetal growth, prematurity, and several neurodevelopmental sub-domains), EPA’s synthesis reflected that inconsistency directly in its “slight” evidence-integration judgments, rather than overstating certainty.

5. Independent Expert Panel and Agency Weight-of-Evidence Conclusions

Inorganic arsenic has been evaluated and classified by multiple independent national and international health and regulatory bodies over several decades, and these classifications are broadly convergent.

• International Agency for Research on Cancer (IARC): inorganic arsenic and arsenic compounds are classified as Group 1, “carcinogenic to humans,” based on sufficient evidence in humans for cancer of the urinary bladder, lung, and skin.
• U.S. EPA: arsenic has been classified as carcinogenic to humans by the oral route based on epidemiological evidence since EPA’s 1995 IRIS assessment, a classification retained and updated in the 2025 assessment.
• National Toxicology Program (NTP), 14th Report on Carcinogens: arsenic and inorganic arsenic compounds are listed as “known to be human carcinogens.”
• World Health Organization (WHO) and Agency for Toxic Substances and Disease Registry (ATSDR): both have independently concluded that inorganic arsenic causes bladder, lung, and skin cancer and skin lesions in humans, consistent with the accepted-hazard designation carried forward in the current IRIS assessment.
• National Research Council/National Academies of Sciences, Engineering, and Medicine (NRC, 2013; NASEM, 2019): an independent expert committee convened to review EPA’s IRIS arsenic assessment concluded that low-to-moderate levels of inorganic arsenic (generally discussed as <100 μg/L in drinking water) are associated with diseases of the circulatory system, diabetes, adverse pregnancy and birth outcomes, and neurological/neurodevelopmental deficits based on the human evidence, and specifically recommended that EPA prioritize these noncancer outcomes for updated, quantitative evidence synthesis and dose-response analysis — the recommendation that gave rise to the DCS, diabetes, fetal/newborn/infant, and neurodevelopmental evidence syntheses summarized in Sections 2 and 3 above.

Within the current IRIS assessment itself, EPA applied a structured, transparent evidence-integration framework (using symbols ⊕⊕⊕ for “robust,” ⊕⊕☉ for “moderate,” and ⊕☉☉ for “slight” evidence) that maps directly onto formal causal-language conclusions: robust evidence supports a conclusion that “the evidence demonstrates” causation, applied to diseases of the circulatory system and type 2 diabetes; moderate evidence supports a conclusion that “the evidence indicates (likely)” causation, applied to fetal, newborn, and infant health outcomes and to neurodevelopmental effects generally; and slight evidence, while suggestive, does not support a stronger causal conclusion for the individual sub-outcomes so classified (e.g., fetal growth, prematurity, motor effects). This tiered framework — explicit about both the direction of the conclusion and the qualitative degree of confidence behind it — is directly analogous in structure and function to the “probable link,” “sufficient evidence,” and “limited/suggestive evidence” determinations issued by other expert bodies and panels in analogous toxic-tort and mass-exposure contexts.

6. Use of Epidemiological Evidence, Alongside Toxicology, in General and Specific Causation

In toxic-tort and products-liability litigation, epidemiological evidence of the kind summarized above typically forms the core of the general causation analysis — i.e., whether the substance is capable of causing the type of injury at issue in the general population, at exposure levels of the kind experienced by the plaintiff population — while individual (specific) causation additionally requires a case-specific evaluation of whether the substance caused a particular plaintiff’s injury.

6.1 General Causation

For outcomes such as bladder cancer, lung cancer, ischemic heart disease, hypertension, and type 2 diabetes, the epidemiological evidence summarized above — replicated across multiple, methodologically diverse, and geographically independent populations; showing dose-response gradients; supported by biological plausibility and mechanistic coherence from the toxicology literature (see RoseyFields’ companion toxicology summary); and independently recognized by IARC, EPA, NTP, WHO, ATSDR, and the National Academies — satisfies the elements courts and experts commonly evaluate under a Bradford Hill-type framework for causal inference (strength of association, consistency, temporality, biological gradient, plausibility, coherence, and, where available, analogy and experimental/mechanistic evidence). For these outcomes, general causation is generally well supported by the combined weight of the epidemiological and toxicological evidence bases, and courts and regulatory bodies have long recognized arsenic as capable of causing bladder, lung, and skin cancer in particular. For outcomes rated “moderate” (e.g., birth weight, fetal/infant mortality, developmental neurocognitive effects), the evidence is generally still considered supportive of general causation but with somewhat less certainty, reflecting greater cross study variability; experts relying on these outcomes should expect closer scrutiny of study selection, exposure response modeling choices, and consistency arguments. For outcomes rated “slight” or characterized by predominantly null or inconsistent findings (e.g., stroke as an isolated endpoint, fetal growth, prematurity, several neurodevelopmental sub-domains, metabolic syndrome as a composite outcome), the current epidemiological evidence base alone is less likely to support a robust general causation opinion without substantial additional case specific support, and such outcomes are more appropriately framed, where relevant, as coherence or corroborating evidence for a related, more strongly supported outcome rather than as freestanding causation claims.

6.2 Specific Causation

General causation evidence of the kind above establishes that arsenic is capable of causing a given outcome at relevant exposure levels; specific causation requires connecting that general capability to a particular individual, and typically draws on several additional lines of case-specific evidence and analysis: (1) reconstruction or documentation of the individual’s actual exposure (e.g., drinking-water source and concentration history, duration of exposure, biomarker data where available, and comparison to the exposure levels and durations reflected in the epidemiological studies relied upon); (2) evaluation of exposure-response consistency, i.e., whether the individual’s estimated dose and exposure duration fall within (or how they compare to) the range of exposures shown to produce the effect sizes described in Section 3; (3) a differential-etiology analysis considering and, where appropriate, ruling in or out other known causes or risk factors for the specific outcome at issue in that individual (e.g., smoking history for IHD or bladder/lung cancer, family history and other metabolic risk factors for diabetes, other pregnancy risk factors for birth-weight or neurodevelopmental outcomes); (4) consideration of individual susceptibility factors identified in the epidemiological and mechanistic literature, such as arsenic-methylation efficiency (influenced by AS3MT genotype and nutritional status), life-stage at exposure (in utero and early childhood exposure windows carry particular weight for developmental and, per the animal carcinogenicity literature, cancer outcomes), and co exposures; and (5) clinical and temporal correlation between the exposure period and the onset, latency, and clinical course of the diagnosed condition. Because specific causation is inherently individualized, the strength of the population-level (general causation) evidence summarized in this report is a necessary but not sufficient component of a specific causation opinion; it establishes the upper bound of what the epidemiological literature can support, while the remaining, case-specific analysis must be developed on the individual record.

6.3 Integration with the Toxicological Evidence Base

Epidemiological and toxicological evidence serve complementary functions in a causation analysis: epidemiology directly measures associations, and where well-designed, exposure-response relationships, in the human population of interest, while toxicology (animal and mechanistic data, summarized in RoseyFields’ companion toxicology page) supplies biological plausibility, mode-of-action support, and evidence bearing on susceptible subpopulations, dose thresholds, and target-organ specificity that individual epidemiological studies alone often cannot resolve. For arsenic, the convergence of these two evidence streams is unusually strong for bladder cancer, lung cancer, ischemic heart disease, hypertension, and type 2 diabetes, where human dose-response data are extensive and are reinforced by a coherent, well-characterized mode of action (oxidative stress, enzyme inhibition, and related pathways) and, for cancer specifically, by animal carcinogenicity bioassay data demonstrating tumor induction at concordant target organs. This convergence is precisely the kind of “coherence” and “biological plausibility” that both scientific weight-of-evidence frameworks and legal admissibility standards (e.g., Daubert-type reliability factors) look for in evaluating the reliability of a causation opinion, and it is the basis on which RoseyFields’ epidemiology and toxicology pages for this chemical are intended to be read together rather than in isolation.

References