Adiposity and Endometrial Cancer Risk in Postmenopausal Women: A Sequential Causal Mediation Analysis

Excess adiposity is an important risk factor for endometrial cancer (1). In 2012, 34% [90% confidence interval (CI), 32%–36%] of diagnosed endometrial cancer cases were attributable to overweight and obesity (2). Disturbed adipocytokine production and inflammation, hyperinsulinemia, and sex-steroid hormones are hypothesized to underlie the adiposity–endometrial cancer link (3). In women with obesity, adipose tissue secretes less adiponectin (an anti-inflammatory adipocytokine) and more inflammatory adipocytokines (4). This inflammatory status may have mitogenic, antiapoptotic, and angiogenic effects (5, 6); reduce insulin sensitivity (7); or dysregulate aromatase expression and increase estrogen levels (8). Insulin sensitivity may also be reduced through increased hepatic glucose production in response to excess free fatty acids (7). The resulting hyperinsulinemia and higher circulating insulin levels are causally linked to endometrial cancer (9). It may have mitogenic effects; increase free insulin-like growth factor-1 (IGF-1) levels (10, 11); increase aromatase activity via IGF-1, thus increase estrogen levels (8); or downregulate sex-hormone–binding globulin (SHBG) production and increase bioavailable estrogens (12). Bioavailable estrogens, especially when unopposed by progesterone (3), may increase endometrial cancer risk through mitogenic effects in endometrial tissue (13).

Causal mediation analysis (14) can quantify the role of biological pathways involved in the effect of adiposity on endometrial cancer risk. As highlighted, it is improbable that the influences of the involved biomarkers are siloed. When measures for multiple biomarkers are available, mediation analysis approaches should account for correlations between biomarkers (14). Failing to do so and assessing the mediating roles of biomarkers individually may result in biased estimation of the effect explained by the biomarkers (14).

We used data from a case–control study nested within the European Prospective Investigation into Cancer and Nutrition (EPIC; refs. 15, 16) to quantify the mediating roles of biomarkers representing inflammatory status, hyperinsulinemia, and estrogens in explaining the effect of adiposity on endometrial cancer risk in postmenopausal women. We estimated path-specific mediated effects using a sequential causal mediation analysis approach that allowed us to take dependences between biomarkers into account (17). The analysis relied on an assumed causal ordering between the pathways (17), which was decided based on the existing evidence as described above. Although adiposity increases pre- and postmenopausal endometrial cancer risk (1), we excluded premenopausal women because the mechanisms may be different before and after menopause (18).

Established in 1992, EPIC is a cohort study including approximately 370,000 women recruited from 10 European countries. Details of baseline data collection, blood sample collection and storage, ascertainment of cancers, and follow-up are published (16). The present study made use of data from a nested case–control study that investigated associations between endometrial cancer risk and sex-steroid hormones, metabolic factors, and inflammatory biomarkers measured in baseline blood samples, for which follow-up ended between December 1999 and November 2003 (15).

The eligibility criteria for the original case–control study included no history of hysterectomy or diagnosis of cancer (except keratinocyte skin cancer) and no use of oral contraceptives or hormone therapy at blood collection. It included 233 incident primary endometrial cancer cases and 446 controls (incidence density sampling), matched on recruitment center, age and fasting status at blood draw, time of blood draw, and menopausal status (15). We used the updated EPIC database for this study, in which one case was found to have prevalent cancer at recruitment, and seven were no longer classified as endometrial cancer cases based on additional information on tumor histology collected through pathology reports. We excluded these women and their matched controls. In addition, women who at recruitment were not postmenopausal (46 cases and 86 controls), had a history of diabetes (9 cases and 16 controls), and had body mass index (BMI) <18.5 kg/m² (1 case and 3 controls) were excluded from the analysis. Finally, rather than performing multiple imputation to handle missing confounder data (the method used for missing biomarker data), we excluded women who had missing values for any of the selected confounders (see Confounder selection for further detail) because they comprised a small proportion of all participants (6 cases and 17 controls; 5% of the participants).

Measured biomarkers were adiponectin; IL6, IL1 receptor antagonist (IL1Ra), TNFα, TNF receptor 1 (TNF-R1), TNF-R2, and C-reactive protein (CRP; biomarkers of inflammatory status); C-peptide (hyperinsulinemia biomarker); and calculated free estradiol (proxy for bioavailable estradiol) and estrone (estrogen pathway biomarkers; refs. 19–23). Free estradiol was calculated from SHBG and total estradiol (24). Details of biomarker measurements have been previously published and are summarized in Supplementary Table S1 (19–23). Samples from matched cases and controls were analyzed within the same batch, and technicians performing the assays were blinded to the case status. The German Cancer Research Center (Deutsches Krebsforschungszentrum, Heidelberg, Germany) performed assays for IL1Ra, TNF-R1, and TNF-R2. Other biomarkers were measured at the International Agency for Research on Cancer (IARC, Lyon, France).

We had three measures of adiposity: BMI, waist circumference, and waist–hip ratio. To assist with the interpretations of the mediated effect (see Mediation analysis below), translation to policy, and to relax the parametric assumption of linearity, in primary analyses, these were considered as categorical variables. The cutoff values to categorize women were based on guidelines (BMI ≥18.5–<25, ≥25–<30, and ≥30 kg/m²; waist circumference ≤80, >80–≤88, and >88 cm), or tertile cutoffs (waist–hip ratio ≤0.78, >0.78–≤0.84, >0.84; ref. 25).

To remove the effects of batch, fasting status, and time of blood draw, for each biomarker we: (i) fitted a linear mixed-effects model to the log-transformed biomarker value for the controls with batch as a random effect, fasting status at blood draw and time of blood draw as fixed effects, then (ii) for all women, derived a normalized value by subtracting the difference between the predicted mean batch-specific values and the overall mean from the observed values.

The sequential mediation analysis used (see Mediation analysis) relied on no unmeasured exposure–outcome, mediator–outcome, and exposure–mediator confounding. A causal directed acyclic graph (Fig. 1) was developed with reference to existing evidence (13, 26–28) to identify these confounders (14). These potential confounders included age at recruitment, number of full-term pregnancies, age at menarche and menopause, history of oral contraceptive use and hormone therapy, smoking status, and physical activity. Age at menopause had 7% missing value and was not included as a confounder in the analysis because it was weakly correlated with the exposure and biomarkers (r < |0.17|). Every other variable had <3% missing values. As previously described, participants with missing confounder data were excluded from analysis.

Geometric mean ratios (GMR) of biomarkers (and 95% CIs) in relation to adiposity measures were estimated using linear regression models applied to log-transformed data for the controls, with adjustment for potential confounders.

In models that included the outcome, we broke the matching to avoid losing participants and included the matching variables as covariates. We grouped the 21 recruitment centers into regions to avoid creating combinations with sparse data. Age at recruitment was modeled as restricted cubic splines (2 degrees of freedom corresponding to 3 knots). The ORs and 95% CIs for the association between endometrial cancer and adiposity measures and biomarkers were estimated from models that included confounders. Models for biomarkers included BMI and when applicable, other biomarkers that might have confounded the biomarker–outcome association. The linearity of mediator–outcome associations was checked using models with restricted cubic splines (2 degrees of freedom).

Figure 1 guided the analysis. The association between each adiposity measure and endometrial cancer [total effect (TE; ref. 14)] was decomposed into a natural indirect effect (NIE) through all biomarkers and a natural direct effect (NDE). When comparing women with obesity versus normal weight, the NIE could be interpreted as the average change in the endometrial cancer incidence if all women had obesity and their biomarker level changed from what it would naturally be if they had obesity to what it would naturally be if they had normal weight. The NIE captures the effect of obesity on the endometrial cancer incidence exerted through obesity-induced alterations in biomarker levels. The NDE could be interpreted as the average change in the endometrial cancer incidence, if all women had the biomarker level they would have naturally had when of normal weight, and they changed from having obesity to having normal weight. This effect captures the part of the effect of obesity on the endometrial cancer incidence that operates through pathways other than the biomarkers assessed in the mediation analysis (see also Supplementary Fig. S1; ref. 14). Assuming that reduced adiponectin and increased inflammation biomarkers levels preceded and potentially, but not necessarily, influenced C-peptide and estrogen levels, and C-peptide preceded and potentially influenced estrogens (7, 8, 29), we sequentially (17) decomposed the estimated NIE into NIEs through: (i) pathways originating with reduced adiponectin and increased inflammation biomarkers, (ii) C-peptide beyond the influences of adiponectin and inflammation biomarkers, and (iii) estrogens (free estradiol and estrone), beyond the influences of adiponectin, inflammation biomarkers, and C-peptide. Assuming that adiponectin preceded and influenced inflammation biomarkers (29), we additionally decomposed the first NIE into NIEs through: (i) pathways originating with reduced adiponectin and (ii) inflammation biomarkers beyond the potential influence of adiponectin. Supplementary Fig. S1 shows interpretations of these effects.

The NIEs and NDE were estimated on the log(OR) scale using a regression-standardization approach [where the log(OR)s were standardized to the distribution of confounders for all eligible women in the EPIC; ref. 30]. The estimation was based on nine sequential linear regression models for the biomarkers (mediators), conditional on exposure, baseline confounders, and any preceding biomarker in the sequence and the following four models for the outcome: (i) logistic regression conditional on exposure, confounders, and adiponectin; (ii) logistic regression conditional on exposure, confounders, adiponectin, and inflammation biomarkers; (iii) logistic regression conditional on exposure, confounders, adiponectin, inflammation biomarkers, and C-peptide; and (iv) logistic regression conditional on exposure, confounder, adiponectin, inflammation biomarkers, C-peptide, free estradiol, and estrone. In combination with coefficients estimated from models for the mediators, outcome model i estimates the NIE through pathways originating with adiponectin; model ii estimates the NIE through reduced adiponectin and increased inflammation; models i and ii estimate the NIE through inflammation biomarkers beyond adiponectin; models ii and iii estimate the NIE through C-peptide beyond adiponectin and inflammation biomarkers; models iii and iv estimate the NIE through estrogens beyond adiponectin, inflammation biomarkers, and C-peptide; and model iv estimates the NDE and the NIE through all the biomarkers (see also Supplementary Table S2 for further details; ref. 17). The regression models for which the mediator was the dependent variable were limited to controls to take the case–control design of the study into account (30). The proportion mediated (PM) was calculated on the log(OR) scale as log(OR)^NIE/(log(OR)^NDE + log(OR)^NIE)) (14).

Seventeen percent of women had missing biomarker data, which was multiply imputed based on chained equations with 20 iterations (31). The imputation models included all variables in the mediation analyses, and recruitment center, height, weight, and hip circumference as auxiliary variables. Within each imputed dataset, 1,000 bootstrap samples were used to estimate standard errors for the TE, NDE, and NIEs. These estimates were pooled using Rubin's rules to calculate the final estimates and 95% CI (32).

We also repeated the mediation analyses with continuous adiposity measures, and after excluding cases diagnosed within 2 years after recruitment.

All analyses were performed in Stata version 15.1 (33).

The analytic dataset included 163 endometrial cancer cases and 306 controls. Median age at endometrial cancer diagnosis was 63 years [interquartile range (IQR), 60–68; Table 1]. At baseline, compared with controls, a smaller proportion of cases had used oral contraceptives, were current smokers, and had moderate/high physical activity; a larger proportion had used hormone therapy and had BMI ≥ 30 kg/m² or waist circumference >88 cm. Compared with women with no missing data, a higher proportion of women with missing biomarker values were cases, fasting >6 hours at blood draw, from Northern Europe, current smokers, or with BMI ≥30 kg/m² (Supplementary Table S3).

For BMI (≥30 vs. ≥18.5–<25 kg/m²), positive associations were observed with IL6, IL1Ra, TNF-R1, TNF-R2, CRP, C-peptide, free estradiol, and estrone. Of these, CRP demonstrated the strongest association (GMR, 2.85; 95% CI, 2.13–3.81). An inverse association was observed for adiponectin (GMR, 0.77; 95% CI, 0.67–0.88). There was no evidence for an association with TNFα. Similar patterns were seen for waist circumference and waist–hip ratio (Table 2; Supplementary Table S4 complete-case analysis results).

An increased OR for endometrial cancer was observed for BMI ≥30 versus ≥18.5–<25 kg/m² (OR 2.94; 95% CI, 1.71–5.06) and waist circumference >88 versus ≤80 cm (OR 2.10; 95% CI, 1.31–3.36). The evidence for an association between BMI ≥25 versus ≥18.5–<25 kg/m², waist circumference >80–≤88 vs. ≤80 cm, and both categories of waist–hip ratio and odds of endometrial cancer was weak (Table 3). Therefore, we did not attempt to decompose these in the mediation analysis.

An inverse association was observed between adiponectin and endometrial cancer (OR per doubling concentration 0.65; 95% CI, 0.47–0.90), and a positive association for IL1Ra (OR 1.14; 95% CI, 1.00–1.29) and estrone (OR 2.03; 95% CI, 1.33–3.09). There was no strong evidence for departure from linearity for any of the biomarker–outcome associations (Table 3; Supplementary Table S5 complete-case analysis results).

Because we neither observed an association between adiposity measures and TNFα (Table 2) nor between TNFα and endometrial cancer risk (Table 3), we did not include this biomarker in our mediation analysis.

Approximately 72% of the association between BMI (≥30 vs. ≥18.5–<25 kg/m²) and endometrial cancer was mediated through all the biomarkers (OR^NIE 1.95; 95% CI, 1.01–3.74). Following a further decomposition of this NIE, there was suggestion for a 46% PM through pathways originating with reduced adiponectin and increased inflammation biomarkers, 5% PM through C-peptide beyond (the potential influences) of adiponectin and inflammation biomarkers, and 21% PM through estrogens beyond adiponectin, inflammation biomarkers, and C-peptide. A decomposition of the NIE through adiponectin and inflammation biomarkers indicated a 33% PM through pathways originating with adiponectin and 13% PM through inflammation biomarkers beyond adiponectin. The estimated OR^NDE not through any of the biomarkers was 1.29 (95% CI, 0.54–3.09). The OR^NIE point estimates for inflammation biomarkers, C-peptide, and estrogens were suggestive of moderate to weak increase in endometrial cancer OR, but the 95% CIs were wide and also included a decreased OR (Table 4; Supplementary Table S6 complete-case analysis results).

Similarly, for waist circumference (>88 vs.≤80 cm), approximately 76% of the association was mediated through all biomarkers (the OR^NIE 1.73; 95% CI, 1.04–2.90). There was evidence for an NIE through reduced adiponectin and increased inflammation biomarkers (61% PM), as well as through reduced adiponectin (38% PM). The point estimates were also indicative of 24% PM through inflammation biomarkers beyond adiponectin, 4% PM through C-peptide beyond adiponectin and inflammation biomarkers, and 10% PM through estrogens beyond adiponectin, inflammation biomarkers, and C-peptide. However, the 95% CIs around the OR^NIE for these estimates were wide and included a decreased OR. The OR^NDE not through any of the biomarkers was 1.19 (95% CI, 0.59–2.41; Table 4; Supplementary Table S6 complete-case analysis).

Mediation patterns were similar for continuous exposures (Supplementary Table S7) and after excluding cases diagnosed within 2 years after recruitment (Supplementary Table S8).

In this study of postmenopausal women, reduced adiponectin and increased inflammation biomarkers, C-peptide, and estrogens mediated most (>70%) of the increased odds of endometrial cancer in women with obesity compared with normal weight. In the sequential mediation analysis, the largest mediating effect was observed for pathways originating with adiponectin. Based on the point estimates, depending on the measure of adiposity used, the second most important pathway was either inflammation (waist circumference) or estrogens (BMI). Our study had a relatively small size, and there was high uncertainty around the estimate of the NDE, not allowing us to make definitive conclusions about the direction and magnitude of the part of the effect of obesity on endometrial cancer not explained by biomarkers included in our analysis.

We believe this is the first study to investigate the mediating role of multiple biomarkers in the adiposity–endometrial cancer association using formal mediation analysis. The sequential mediation analysis circumvented the assumption of no exposure-induced mediator–outcome confounding and permitted quantifying the indirect effect through all the biomarkers and decomposing this into path-specific indirect effects without assuming the biomarkers were independent (17). The sequential mediation analysis relied on a presupposed causal ordering of the pathways, which was based on existing evidence (7, 8, 29).

We had measures for a range of biomarkers representing the pathways of interest. However, because this was secondary analysis of an existing study, we were limited to biomarkers that had been measured, and these may not have fully captured the entire physiologic impact of these pathways. For the insulin pathway, we had measures for IGF-binding protein (IGFBP)-1 and IGFBP-2 but did not include them because they do not have a clear relationship with adiposity or endometrial cancer (13). We also excluded women who had a history of diabetes at recruitment because long-term hyperglycemia might influence insulin production (34). This might have affected the generalizability of our study findings to women with diabetes. The observed mediating effect through biomarkers might have also been influenced by the quality of the measurements [the intrabatch coefficient of variation (CV) ranged from 2.6% (C-peptide) to 15% (IL1Ra, TNFα), and the interbatch CV from <8% (adiponectin, TNF-R1, TNF-R2) to 27.7% (IL1Ra; refs. 19–23)], and temporal stability of biomarkers. Finally, an observed mediating role through a biomarker does not necessarily indicate that the biomarker has a causal effect on endometrial cancer risk. For example, even assuming that our presupposed causal ordering of the biomarkers held (i.e., adiponectin potentially influenced other biomarker levels but not vice versa), we cannot conclude that a hypothetical intervention to increase adiponectin level in women with obesity would reduce their endometrial cancer risk. The observed NIE through adiponectin might have, at least partly, been because this biomarker performed well at reflecting the inflammatory status in women with obesity, which in turn could have influenced cancer risk through various mechanisms.

We controlled for the known confounders of the associations between exposure, mediators, and outcome, but residual confounding, for example through other unmeasured or mismeasured biomarkers, cannot be ruled out. We were also limited in exploring the potential influence of unmeasured confounding on our results, because sensitivity analysis and bias correction approaches for settings with multiple mediators are not developed yet. Although our analysis assumed that it was adiposity that influenced biomarker levels, we cannot be certain about the temporal exposure–mediator ordering, because they were measured cross-sectionally. Preclinical cancers might have also influenced measures of adiposity and biomarkers. However, results from a sensitivity analysis that excluded cases diagnosed within 2 years after recruitments were comparable with our primary analysis. We used a regression-based mediation analysis that could be adapted to the matched design (30). A limitation was that it did not allow including possible exposure–mediator and mediator–mediator interactions (17). Another limitation of our study was that biomarker measurements were missing for 17% of the participants, and a comparison of women with and without missing data indicated that missingness was not completely at random. We multiply imputed these missing values, making an unverifiable assumption that missingness was at random, with missing data only depending on measured variables. Violation of this assumption would have introduced bias into our estimates of the NIE and NDE, but we attempted to reduce this possibility by defining multiple imputation models that included key variables (including BMI, waist circumference, and waist–hip ratio).

The predominant hypothesis in explaining the adiposity–endometrial cancer association in postmenopausal women is increased estrogen production in adipose tissue (35, 36). To explore the mediating role of estrogens, two studies assessed the degree to which adjusting for estradiol attenuated the BMI–endometrial cancer association (19, 37). The adjustment weakened but did not entirely explain the association [OR for BMI ≥30 vs. <25 kg/m² from 3.97 (95% CI, 2.54–6.21) to 2.25 (95% CI, 1.33–3.81; ref. 37), and from 2.67 (95% CI, 1.63–4.37) to 2.09 (95% CI, 1.22–3.57; ref. 19)], suggesting that other pathways are also likely at play. Similarly, we observed weak evidence of a small to moderate mediating effect through estrogens beyond the potential influence of preceding pathways.

Hyperinsulinemia might mediate the adiposity–endometrial cancer association (3). However, we found almost no mediating effect through C-peptide beyond adiponectin and inflammation. Triglyceride glucose product (TyG) was used as a proxy for insulin resistance in a pooled study of six European cohorts. A negligible proportion (3.6%) of the total effect of BMI (≥30 vs. ≥18.5–<25 kg/m²; HR 2.61; 95% CI, 2.29–2.99) was mediated by TyG (HR^NIE 1.04; 95% CI, 0.99–1.09; ref. 38). In both studies, there was little association between TyG or C-peptide and endometrial cancer after adjusting for BMI. However, other studies have found associations between fasting insulin (9, 39) and C-peptide (40) and endometrial cancer, after adjustment for BMI. Put together with the likely limitations of fasting and nonfasting C-peptide in capturing insulin resistance (39), such discrepancies warrant additional research to elucidate the mediating effect of the insulin resistance pathway.

In our analysis, the largest mediating effect was observed through reduced adiponectin and increased inflammation biomarkers. A further decomposition of this indirect effect suggested that a larger mediating effect was through adiponectin, with indication of a smaller mediating effect through inflammation beyond the potential influence of adiponectin. Existing evidence for inverse adiposity–adiponectin (41) and adiponectin–endometrial cancer associations (42) supports our observation. Based on a principal component factor analysis, data from the original case–control study previously demonstrated a reduction in the OR for the association between BMI (≥30 vs. <25 kg/m²) and postmenopausal endometrial cancer from 2.73 (95% CI, 1.66–4.50) to 1.65 [95% CI, 0.92–2.98; ∼50% reduction on log(OR) scale] after adjusting for a factor with >|40%| loading for adiponectin, together with CRP, C-peptide, IGFBP-1, IGFBP-2, SHBG, and HDL cholesterol (15).

In summary, about 70% of the increased odds of endometrial cancer risk in women with obesity compared with normal weight were mediated together through adiponectin, inflammation, C-peptide, and estrogens. We applied a novel mediation analysis approach to quantify the mediating effects through these pathways jointly and the path-specific indirect effects. The applied method was able to handle multiple correlated biomarkers without assuming independence. Pathways originating with reduced adiponectin had the most important role in explaining the link. Future studies with larger sample sizes and a range of biomarkers reflecting the pathways, and preferably repeated measures for adiposity and biomarkers are needed to replicate findings from this study. Larger studies are needed to estimate the mediated effects with more certainty and to allow exploring the possible influence of adiposity–biomarkers and biomarker–biomarker interactions on those effects. Ideally, as has been done in this study, future research in this area would take advantage of the advances in mediation analysis to properly account for the dependences between biomarkers.

S.G. Dashti reports grants from World Cancer Research Fund [this work was supported by the World Cancer Research Fund (grants 2003/18 and 2007/13)] and Australian National Health and Medical Research Council [grant 1150591 (postgraduate PhD scholarship)], personal fees from International Agency for Research on Cancer [IARC; the work reported in this article was undertaken by S.G. Dashti while hosted by the IARC and funded partially by a PhD scholarship by IARC], and other from several (the coordination of EPIC is financially supported by the European Commission and the International Agency for Research on Cancer) during the conduct of the study, as well as grants from Australian National Health and Medical Research Council [grant 1150591 (postgraduate PhD scholarship)] and personal fees from IARC (PhD scholarship by IARC) and Melbourne Research Scholarship (PhD scholarship from the University of Melbourne) outside the submitted work. S. Rinaldi reports grants from World Cancer Research Fund during the conduct of the study. M.B. Schulze reports grants from German Federal Ministry of Science, European Union, German Cancer Aid, and European Community during the conduct of the study. L. Dossus reports grants from World Cancer Research Fund during the conduct of the study. No disclosures were reported by the other authors.

Where authors are identified as personnel of the IARC/World Health Organization, the authors alone are responsible for the views expressed in this article and they do not necessarily represent the decisions, policies, or views of the IARC/World Health Organization.

S.G. Dashti: Formal analysis, methodology, writing–original draft, writing–review and editing. D.R. English: Supervision, methodology, writing–review and editing. J.A. Simpson: Supervision, methodology, writing–review and editing. A. Karahalios: Supervision, writing–review and editing. M. Moreno-Betancur: Supervision, methodology, writing–review and editing. C. Biessy: Resources, data curation, writing–review and editing. S. Rinaldi: Conceptualization, resources, methodology, writing–review and editing. P. Ferrari: Resources, writing–review and editing. A. Tjønneland: Writing–review and editing. J. Halkjær: Writing–review and editing. C.C. Dahm: Writing–review and editing. H.T. Vistisen: Writing–review and editing. F. Menegaux: Writing–review and editing. V. Perduca: Writing–review and editing. G. Severi: Writing–review and editing. K. Aleksandrova: Writing–review and editing. M.B. Schulze: Writing–review and editing. G. Masala: Writing–review and editing. S. Sieri: Writing–review and editing. R. Tumino: Writing–review and editing. A. Macciotta: Writing–review and editing. S. Panico: Writing–review and editing. A.E. Hiensch: Writing–review and editing. A.M. May: Writing–review and editing. J.R. Quirós: Writing–review and editing. A. Agudo: Writing–review and editing. M.-J. Sánchez: Writing–review and editing. P. Amiano: Writing–review and editing. S. Colorado-Yohar: Writing–review and editing. E. Ardanaz: Writing–review and editing. N.E. Allen: Writing–review and editing. E. Weiderpass: Writing–review and editing. R.T. Fortner: Writing–review and editing. S. Christakoudi: Writing–review and editing. K.K. Tsilidis: Writing–review and editing. E. Riboli: Conceptualization, resources, data curation, writing–review and editing. R. Kaaks: Conceptualization, resources, writing–review and editing. M.J. Gunter: Resources, supervision, methodology, writing–review and editing. V. Viallon: Supervision, methodology, writing–review and editing. L. Dossus: Resources, data curation, supervision, writing–review and editing.

The authors thank thank the National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands, for their contribution and ongoing support to the EPIC Study. This study also made use of data from the EPIC-Norfolk cohort.

This work was supported by the World Cancer Research Fund (grants 2003/18 and 2007/13), Australian National Health and Medical Research Council grants (1150591 to S.G. Dashti and 1104975 to J.A. Simpson), and Melbourne Research Scholarship (to S.G. Dashti). M. Moreno-Betancur is the recipient of an Australian Research Council Discovery Early Career Award (project number DE190101326) funded by the Australian Government. The work reported in this article was undertaken by S.G. Dashti while hosted by the IARC and funded partially by a PhD scholarship by IARC. The coordination of EPIC is financially supported by the European Commission and the IARC. The national cohorts are supported by the following: Danish Cancer Society (Denmark); Ligue Contre le Cancer, Institut Gustave Roussy, Mutuelle Générale de l'Education Nationale, and Institut National de la Santé et de la Recherche Médicale (France); Deutsche Krebshilfe, Deutsches Krebsforschungszentrum, and Federal Ministry of Education and Research (Germany); Stavros Niarchos Foundation and Hellenic Health Foundation (Greece); Italian Association for Research on Cancer and National Research Council (Italy); Dutch Ministry of Public Health, Welfare and Sport, Netherlands Cancer Registry, LK Research Funds, Dutch Prevention Funds, Dutch ZON (Zorg Onderzoek Nederland), World Cancer Research Fund, and Statistics Netherlands (the Netherlands); European Research Council and Nordforsk, Nordic Centre of Excellence Programme on Food, Nutrition and Health (Norway); Health Research Fund, Regional Governments of Andalucía, Asturias, Basque Country, Murcia, and Navarra, and Instituto de Salud Carlos III Red Temática de Investigación Cooperativa en Salud (Spain); Swedish Cancer Society, Swedish Scientific Council, and Regional Governments of Skåne and Västerbotten (Sweden); and Cancer Research UK, Medical Research Council, Stroke Association, British Heart Foundation, Department of Health, Food Standards Agency, and Wellcome Trust (United Kingdom).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.