Smoking Methylation Marks for Prediction of Urothelial Cancer Risk

This article is featured in Highlights of This Issue, p. 2141

Urothelial cell carcinoma (UCC) is a type of malignancy arising from the urothelium. Although UCC accounts for more than 90% of urinary bladder cancers (1), some can also be found in the proximal urethra, the transitional epithelium of the renal pelvis, and the ureter (2). According to Global Cancer Statistics, bladder cancer was in 2020 the 12th most common cancer worldwide, with an estimated 573,000 new cases and 212,000 new deaths (3). Cigarette smoking has been established as a strong risk factor for UCC with approximately half of newly diagnosed patients reporting a history of smoking (4, 5). Many studies (6–9) have investigated the association between smoking and risk of UCC, and a meta-analysis of 89 observational studies reported an increased risk of bladder cancer for current smokers [odds ratio (OR) = 3.1; 95% confidence interval (CI) = 2.5–3.7] and former smokers (OR = 1.8; 95% CI, 1.5–2.1), compared with never smokers (10). However, information on smoking history used in most epidemiologic studies, such as smoking status (never, former, or current smoker) or pack-years, is typically collected via self-report and may be prone to substantial measurement error. The accuracy of self-reported information has also been questioned because of declining response rates and the increasing social stigmatization of smoking (11). Furthermore, such information cannot reflect secondhand smoke exposure during childhood or adulthood. Therefore, such less accurate information would have potential impact on studies of disease association and risk prediction.

Serum or urinary cotinine (12) and blood DNA methylation (13–16) have been established as valid biomarkers of cigarette smoking exposure. Although cotinine and methylation markers showed similar accuracy in distinguishing current from never smokers, only methylation markers can distinguish former from never smokers with high accuracy (17). Therefore, DNA methylation markers measured in blood, which may also reflect different individuals' responses to lifetime exposure, can be used to augment self-reported smoking data to help refine individual risk profiling of smoking-induced diseases (18–20).

Authors of several studies (21–23) have evaluated the association of genome-wide cytosine-guanine (CpG) methylation in blood DNA with risk of UCC. Jordahl and colleagues (23), for example, identified potential methylation-based markers of susceptibility to urothelial carcinoma of the bladder, using the Illumina Infinium HumanMethylation450 Bead Array (∼450,000 probes) on prediagnostic blood collected in the Women's Health Initiative (WHI). They subsequently found that two previously identified smoking-associated CpG sites mediated the effect of smoking on bladder cancer risk (24). With the current study, we aimed to expand on previous research by identifying associations between smoking-associated DNA methylation and bladder cancer risk and by developing a predictor of UCC risk using smoking-associated DNA methylation measures.

The Melbourne Collaborative Cohort Study (MCCS) is an Australian prospective cohort study of 41,513 people recruited between 1990 and 1994 in the Melbourne metropolitan area. All participants were of white European origin. DNA was extracted from prediagnostic peripheral blood taken at recruitment (1990–1994) or at a subsequent follow-up visit (2003–2007) in participants free of UCC. More details about the cohort, blood collection, DNA extraction, and cancer ascertainment can be found elsewhere (22, 25). Information on tobacco use was self-reported by participants using questionnaires (24, 25). In this study, we utilized a case–control data set of urothelial cancer nested within the MCCS. Controls were matched to incident cases on age at blood draw, year of birth, sex, country of birth (Australia/New Zealand/UK/other, Italy, or Greece), sample collection period (baseline at recruitment or the follow-up visit), and sample type (peripheral blood mononuclear cells, dried blood spots, or buffy coats) using incidence density sampling. To minimize batch effects, samples from each matched case–control pair were plated to adjacent wells on the same BeadChip, with plate, chip, and position assigned randomly. We excluded from the analysis sex-discrepant and failed samples for DNA methylation measures. Case–control pairs with any missing values for the confounders measured were also excluded. Overall, 404 case–control pairs were included in the present study.

For replication and meta-analysis, we included the study sample previously used by Jordahl and colleagues (23, 24), which comprises 440 cases diagnosed with urothelial carcinoma of the bladder and 440 cancer-free controls matched on year of enrollment, age at enrollment (±2 years), follow-up time greater than or equal to their matched case, trial component and DNA extraction method (5-Prime, phenol, Bioserve, or PurGene). This case–control study was nested within the WHI, which includes 161,808 postmenopausal women recruited from 1993 to 1998 across the United States (26).

The study was approved by the Cancer Council Victoria's Human Research Ethics Committee, Melbourne, VIC, Australia, and the Institutional Review Board and Publications and Presentations Committee of WHI–Clinical Coordinating Center in the Fred Hutchinson Cancer Research Center, Seattle, Washington. All participants provided informed consent in accordance with the Declaration of Helsinki.

Quality control (QC) details for measures of genome-wide DNA methylation in the MCCS have been reported previously (22). Briefly, we removed probes with missing rate > 20% and probes on Y-chromosome, and ultimately retained 484,966 CpG sites with their beta values for each sample. Methylation M-values, calculated as log₂[beta/(1-beta)], were used for analysis as these are thought to be more statistically valid for detection of differential methylation (27). In the replication data of WHI, similar data processing on DNA methylation was performed, e.g., QC on CpGs sites using probe missing rate (> 10%) and beadcount (<3) in at least 10% of samples, and M-value transformation, as described previously (23, 24).

An epigenome-wide association study (EWAS) based on the 404 case–control pairs in MCCS was conducted, using conditional logistic regression to estimate OR and 95% CI of UCC risk per SD at each of the 484,966 CpG sites. A first model (model 1) was adjusted for white blood cell composition (percentage of CD4⁺ T cells, CD8⁺ T cells, B cells, NK cells, monocytes, and granulocytes, estimated using the Houseman algorithm; ref. 28), and a second model (model 2) was additionally adjusted for smoking status (current/former/never) and pack-years (log-transformed). As a sensitivity analysis, we evaluated a third model (model 3) with additional adjustment for alcohol consumed in the previous week (in grams/day), body mass index (in kg/m²), height (in meters), educational level (pseudo-continuous score ranging from 1 for “primary school only” to 8 for “tertiary or higher university degree”), physical activity (categorized score based on time spent doing vigorous/less vigorous activities), socioeconomic status (deciles of the relative socioeconomic disadvantage of area of residence index), and diet quality (Alternative Healthy Eating Index 2010). We also stratified analyses by sex and clinical subtype (muscle invasive or non–muscle invasive) and tested heterogeneity of the associations using the likelihood ratio test, by comparing models with and without interaction terms for these variables. The Bonferroni correction was applied to account for multiple comparisons (P < 0.05/484,966 = 1.03 × 10⁻⁷).

Among the 484,966 probes, we focused on 1,061 sites that were found to be strongly associated with a comprehensive smoking index in the MCCS (P < 10⁻⁷) and also reported to be associated with smoking at this threshold P < 10⁻⁷ in any of six large studies, as described in our previous publication (see Supplementary Table S1; ref. 29). For the replication study, we also used conditional logistic regression (models 1 and 2) to estimate associations of the 1,061 smoking-associated DNA methylation measures with risk of UCC in the WHI. For the WHI study, models 1 and 2 were additionally adjusted for race/ethnicity (Asian/Pacific Islander, Black/African American, Hispanic/Latino, non-Hispanic white, or other). The Bonferroni correction was applied to account for multiple comparisons (P < 0.05/1,061 = 4.7 × 10⁻⁵).

A fixed-effects meta-analysis with inverse-variance weights was conducted to combine associations with UCC risk at the 1,061 smoking-associated CpGs from the analyses of MCCS and WHI, using the metagen function in the R package meta (30). The I-square statistic was used to assess heterogeneity across the two studies.

A predictor of UCC risk was developed using the data of 270 case–control pairs from the MCCS cohort for which blood was collected at baseline (1990–1994) as the training set (discovery phase), and 134 case–control pairs for which blood was drawn at follow-up (2003–2007) as an independent testing set in the testing phase. We used penalized logistic regression models with UCC risk as the outcome and the M-values at the 1,061 smoking-associated CpGs as the independent variables, applied to the training set using the R package glmnet (31). Five-fold cross-validation was used, and the mixing parameter (alpha) was set to 1 to apply a Lasso (least absolute shrinkage and selection operator) penalty. The covariates used in model 3 were forced in the penalized logistic models. Coefficients of the logistic Lasso model with the lambda value corresponding to the minimum mean cross-validated error were extracted and used as weights of the selected CpGs to construct a smoking methylation score (MS) for each participant. The smoking MS was then evaluated as a predictor of UCC risk in conditional logistic regression models (adjusted for covariates in model 3 for MCCS data and in model 2 for WHI data, respectively) in the testing sets.

Alternative ways to build methylation-based predictors of UCC risk were explored. We conducted univariate analyses using conditional logistic regression models to the training set to estimate ORs for the individual associations between DNA methylation and UCC risk at each of the 1,061 CpGs. The same covariates as those forced in the Lasso models were included as covariates. We considered three P-value cutoffs (0.05, 0.01, and 0.001) of individual associations at the 1,061 sites, and for each of them we calculated a smoking MS as a weighted average using as weights the logarithm of the OR for each selected CpG.

As a sensitivity analysis, we also used the logistic Lasso method (as described above) to develop a DNA methylation-based smoking predictor of UCC risk using all 404 MCCS case–control pairs. The external 440 case–control pairs from the WHI study were then used as an independent testing set to assess the proposed DNA methylation-based smoking predictors by using conditional logistic regression models (adjusted for covariates in model 2).

The accuracy of the predictive models with the smoking MS as UCC risk predictor was assessed using area under the receiver operating characteristic curve (AUC) estimates with unconditional logistic regression models (models A, B, and C), using the R package pROC (32). Model A used white blood cell composition as independent variables. Model B used white blood cell composition, smoking status, and pack-years (log-transformed) as independent variables. Race/ethnicity was also included in the two models for the WHI sample. Model C used white blood cell composition, smoking status, pack-years, and other covariates (age, sex, country of birth, sample type, alcohol, BMI, height, educational level, physical activity, socioeconomic status, and diet quality) as independent variables. The proposed MSs were then used as additional independent variables in the models to assess the prediction performance by AUC. The DeLong test (33) was used for comparing AUCs.

All methylation scores were rescaled to Z-scores for better comparability of their association with UCC risk. The flowchart of the statistical analysis pipelines and method details are shown in Fig. 1.

The distribution of sociodemographic, lifestyle, anthropometric, and clinical characteristics of the participants in the MCCS is presented in Table 1. Controls were matched to cases on age at blood draw, sex, country of birth (Australia/New Zealand/UK/other, Italy, or Greece) and sample type (peripheral blood mononuclear cells, dried blood spots, and buffy coats). The participants in the MCCS testing set were an average eight years older than in the training set. Compared with controls, cases were more frequently past and current smokers, and had greater smoking pack-years.

For the genome-wide probes tested on the 404 MCCS case–control pairs using models 1–3, there was no significant association between DNA methylation and risk of UCC after Bonferroni correction (P < 1.03 × 10⁻⁷). Nominally significant associations (P < 0.05) were observed for 40,664 (∼8%), 32,137 (∼7%), and 31,319 (∼6%) of the 484,966 CpGs using models 1–3, respectively.

Focusing on the 1,061 smoking-associated CpG sites that we previously identified (29), there was no significant association between DNA methylation and UCC risk in the MCCS after Bonferroni correction (P < 4.7 × 10⁻⁵). Comparing to genome-wide results, there were more methylation markers associated with risk of UCC for the smoking-associated loci, e.g., 19 of the 25 CpGs most strongly with smoking had P < 0.05 in model 1 (Supplementary Table S1). Nominally significant associations (P < 0.05) were observed for 206 (∼19%) and 93 (∼9%) of the 1,061 CpGs in models 1 and 2, respectively (Supplementary Table S1), and the direction of the association was the same as for smoking for 205/206 (100%) and 88/93 (95%) CpG sites. Adjustment for a more comprehensive set of variables (model 3) did not substantially change the associations (Table 2 and Supplementary Fig. S1). Furthermore, the direction of association at 883 (83%, 662 negative and 221 positive, model 1) and 766 (72%, 586 negative and 180 positive, model 2) of the 1,061 CpGs was the same as for their association with smoking (Supplementary Table S1). The results for the 20 most significant associations are presented in Table 2; for all of these associations, the direction of association was the same as with smoking. The stratified results by UCC subtype and sex are shown in Supplementary Tables S2 and S3; we observed no evidence of significant UCC subtype or sex heterogeneity after Bonferroni correction for multiple testing (P < 4.7 × 10⁻⁵).

The replication study using WHI data identified nominally significant associations (P < 0.05) for 229 (∼22%) and 47 (∼4%) of the 1,061 smoking-based CpGs in models 1 and 2, respectively (Supplementary Tables S4 and S5). Among these associations, 51 CpGs (model 1) and 3 CpGs (model 2) were also nominally significant and in the same direction as in the MCCS data.

The meta-analysis of MCCS and WHI results identified nominally significant associations for 387 (∼36%) and 86 (∼8%) CpG sites in models 1 and 2, respectively (Supplementary Tables S4 and S5), and the direction of the association was the same as the association with smoking for 387/387 (100%) and 79/86 (92%) of the CpGs. There were 29 significant associations in model 1 after Bonferroni correction (P < 4.7 × 10⁻⁵), and among these associations, 9 CpGs overlapping the AHRR, GPR15, F2RL3, PRSS23, and GFI1 genes were genome-wide significant (P < 1.03 × 10⁻⁷). The associations were nevertheless substantially attenuated (all P > 4.7 × 10⁻⁵) after adjusting for self-reported smoking variables (model 2). For the majority of the 1,061 CpGs, there was little heterogeneity between MCCS and WHI results (81% and 83% of the CpGs had I² < 0.5 in models 1 and 2, respectively; see Supplementary Tables S4 and S5). The 20 strongest associations in the meta-analyses of models 1 and 2 are shown in Table 3.

The logistic Lasso regression of UCC risk on the 1,061 smoking-based CpGs using the 270 MCCS baseline case–control pairs selected ten CpGs (MS10): cg01324550 (LOC404266), cg02743070 (ZMIZ1), cg07058086 (KIF13B), cg10399789 (GFI1), cg16622061 (chr16: 86888736), cg17924476 (AHRR), cg18979623 (ZBTB46), cg19089201 (MYO1G), cg23110422 (ETS2), and cg24139443 (chr17: 74131549; Supplementary Table S6). The associations with risk of UCC for the 1,061 smoking-associated methylation sites on the training data are shown in Supplementary Table S6. The derived methylation scores based on associations at P < 0.05, P < 0.01, and P < 0.001 included 66 (MS66), 11 (MS11), and 2 (MS2) CpGs, respectively. The associations of these four predictors with UCC risk in the MCCS testing data set (N = 134 cases, model 3) are presented in Table 4. MS10 and MS11 had five overlapping CpGs (cg07058086, cg10399789, cg17924476, cg19089201, and cg23110422) and were associated with risk of UCC in the testing data set (OR = 1.37; 95% CI, 1.00–1.90) and (OR = 1.42; 95% CI, 1.01–1.99), respectively. The distribution of MS10 by smoking status is presented in Supplementary Fig. S2, showing it was elevated in current compared with never smokers. The association of MS10 with UCC risk in the WHI data (model 2) was weaker (OR = 1.09; 95% CI, 0.91–1.30).

Using all 404 case–control pairs of MCCS as the training set, as a sensitivity analysis, the logistic Lasso models selected 18 CpGs (MS18) from the 1,061 smoking-associated CpGs (Supplementary Table S7). MS18 and MS10 had eight overlapping CpGs (cg02743070, cg07058086, cg10399789, cg16622061, cg17924476, cg19089201, cg23110422, and cg24139443). We assessed the resulting predictor MS18 by examining its association with UCC risk in the WHI data, and the result was very similar as for MS10 (OR = 1.09; 95% CI, 0.92–1.30; Table 4). The fixed-effects meta-analysis for MS10 of the two replication sets in MCCS (N = 134) and WHI (N = 440) gave an estimated OR of 1.15; 95% CI, 0.98–1.34, P = 0.08.

The ability of the methylation scores to predict risk of UCC with different models on the testing data sets is presented in Table 5. For the MCCS testing set, the predictions by model C + MS10 and model C + MS11 achieved the highest AUC estimate of 0.66, which was only slightly greater than the same model without methylation information (AUC = 0.64, P = 0.43 for MS10 and 0.39 for MS11). For the WHI testing set, the prediction by model B + MS10 or MS18 achieved an AUC estimate of 0.68, which was of the same as model B alone (P = 0.11 or 0.22).

Most previous studies that investigated the association of smoking with development of urothelial cancer used self-reported smoking history. We included two self-reported variables, smoking status and pack-years, in our analyses. There are other aspects of smoking history, such as age at starting or passive smoking that are typically not or inaccurately captured by questionnaires. As DNA methylation in blood can capture lifetime exposure or different individual responses to smoking, we evaluated the association between smoking-associated methylation and risk of UCC. Although potential associations with UCC were identified at 206 (∼19%) and 93 (∼9%) smoking-based CpG sites in the MCCS in models without and with adjustment for self-reported smoking, respectively, and most associations were in the expected direction, these associations were overall quite weak. In the meta-analysis, DNA methylation at genes, including AHRR, GPR15, F2RL3, PRSS23, and GFI1 (major smoking-related genes), was strongly (P < 10⁻⁷) associated with UCC risk; however, the associations were substantially attenuated after adjusting for self-reported smoking history, likely because these self-reported variables might have captured almost full information of smoking exposure. Thus, these methylation markers added relatively little to the prediction of urothelial cancer risk beyond their association with self-reported smoking. A methylation score combining measures at ten smoking-associated CpG sites developed in the MCCS cohort showed some evidence of association with risk of UCC (OR per SD ∼ 1.4) independently of self-reported smoking in an independent data set of MCCS participants (Table 4). Although these results suggest that the combination of smoking methylation markers may improve the prediction of urothelial cancer risk, limited evidence of replication was found in the WHI cohort (OR per SD ∼ 1.1).

The previous study by Jordahl and colleagues (24) using WHI data investigated three specific smoking-related probes (cg05575921 in the gene AHRR, cg03636183 in F2RL3, and cg19859270 in GPR15) in relation to risk of UCC and showed that methylation alterations at cg05575921 and cg19859270 might mediate the effects of smoking on UCC. Our MCCS data also detected nominally significant associations with UCC risk at these CpGs (cg05575921: OR = 0.78; 95% CI, 0.63–0.97; P = 0.02 and cg19859270: OR = 0.81; 95% CI, 0.68–0.97; P = 0.02) in the adjusted model, which indicate they may add information about risk, in addition to the potential mediation of effect.

DNA methylation at AHRR cg05575921 was previously reported to be strongly associated with lung cancer risk (19, 34–36), e.g., OR = 0.50 (95% CI, 0.43–0.59), P = 4.3 × 10⁻¹⁷ in a pooled analysis of five case–control studies (19). Six CpGs in the AHRR gene also showed nominally significant association (P < 0.05) with risk of UCC in our meta-analysis (model 2): cg05575921 (OR = 0.76, P = 0.003), cg17924476 (OR = 1.19, P = 0.003), cg26529655 (OR = 0.77, P = 0.01), cg12806681 (OR = 0.86, P = 0.02), cg01899089 (OR = 0.88, P = 0.03), and cg03991871 (OR = 0.88, P = 0.04; see Supplementary Table S5). Moreover, cg03636183 in the F2RL3 gene, cg21566642 and cg05951221 in 2q37.1, and cg06126421 in 6p21.33 were also reported to be strongly associated (P = 2 × 10⁻¹⁵) with lung cancer risk (19). Among them, three CpGs also showed nominally significant association with UCC risk in our meta-analyses (model 2): cg21566642 (OR = 0.82, P = 0.009), cg05951221 (OR = 0.86, P = 0.04), and cg06126421 (OR = 0.85, P = 0.03; see Supplementary Table S5). These associations appeared to be weaker than in the lung cancer studies, likely because smoking is not as strong a risk factor for urothelial cancer as it is for lung cancer. In a recent study (37), we showed that GrimAge, a composite biomarker based on several DNA methylation surrogates for plasma proteins and a methylation-based estimator of smoking pack-years (38), is substantially more strongly associated with lung cancer risk (OR per SD = 2.03; 95% CI, 1.56–2.64) than with risk of UCC (OR = 1.22; 95% CI, 0.98–1.52).

The samples used in the WHI cohort were all postmenopausal women, and smoking accounts for approximately half of bladder cancer incidence among postmenopausal women (4, 23). Sex is associated with distinct DNA methylation patterns (39). However, we did not find that associations of DNA methylation smoking markers with UCC varied by sex in the MCCS data, nor did we find heterogeneity between MCCS and WHI results. In this study, we used two common methods to develop risk predictors: (i) Lasso and (ii) univariate analysis with weighted average based on individual CpG associations with UCC risk. For the latter, it is difficult to decide on an appropriate P-value cutoff, and our results showed that the Lasso performed well in this setting. Although there was a reasonably large association of the Lasso predictor in the testing set (OR per 1 SD ∼ 1.4), this translated into only moderately improved risk prediction (Table 5).

DNA methylation changes strongly with age (40, 41). In a recent study using methylation case–control studies nested in the MCCS, we have identified and replicated 32,659 age-associated CpGs (42). Among the 1,061 smoking-associated CpGs considered in the current study, methylation at 475 (45%), 328 (31%), and 118 (11%) CpGs was found to be associated with age in never, former, and current smokers, respectively (P < 0.05/1,061 = 4.7 × 10⁻⁵, based on the data set used in (42), results not shown). Specifically, cg01324550, cg16622061, and cg24139443, which were included in MS10, showed significant associations with age in the overall sample and in never smokers, but not other CpGs (42). This implies that aging (or other cancer risk factors; refs. 43 and 44) may affect DNA methylation at the same loci, which may contribute to explain why these methylation marks add information about cancer risk, in addition to unmeasured smoking exposure.

There are several limitations in this study. First, even with prediagnostic blood samples, we cannot rule out the possibility that DNA methylation measures in blood reflected early cancer or development of other smoking-associated diseases. Second, the participants included in the MCCS testing set were an average eight years older than in the training set. We noted that model 1, which included only white blood cell composition variables, achieved an AUC of 0.53 for the training set but an AUC of 0.61 for the replication set (older MCCS participants). It may be that age, a strong cancer risk factor, is associated with changes in white blood cell composition over time (45) that are also associated with cancer risk (46, 47). Third, we considered the two MCCS data sets as independent because there was no participant overlap, and participants with follow-up blood samples were substantially older; however, the samples were drawn from the same cohort and might have a shared environment; thus, the two data sets might not be completely independent, which may have an influence on results of validation and risk prediction. Fourth, the modest improvement of AUC may suggest that other factors, such as germline genetic variation, and incorporation of more environmental exposures, should be considered in the predictive models. Fifth, the biological mechanisms underlying our findings were not assessed because the aim of our study was to improve UCC risk prediction using smoking-associated methylation marks. For example, TET proteins may stimulate and regulate DNA methylation at genes that were included (48), but this requires further investigation using functional studies. Finally, compared with the MCCS cohort, the methylation measures in WHI were produced using different methods of sample collection and storage, DNA extraction, and DNA methylation processing, which may have influenced some findings, e.g., high heterogeneity for some CpGs across the two studies when performing meta-analysis.

In conclusion, our findings suggest that blood-based DNA methylation markers for smoking may be associated, albeit weakly, with risk of UCC independent of self-reported smoking history, and could provide some improvement to the prediction of urothelial cancer risk. The overall utility of our findings needs to be further assessed using additional external data sets.

K.M. Jordahl reports grants from American Cancer Society and NCI during the conduct of the study. J.K. Bassett reports grants from Australian National Health and Medical Research Council (NHMRC) during the conduct of the study. D.R. English reports grants from NHMRC during the conduct of the study. R.L. Milne reports grants from NHMRC during the conduct of the study. G.G. Giles reports grants from NHMRC (Australia) during the conduct of the study. P.-A. Dugué reports grants from NHMRC during the conduct of the study. No disclosures were reported by the other authors.

C. Yu: Conceptualization, formal analysis, methodology, writing–original draft. K.M. Jordahl: Data curation, formal analysis, writing–review and editing. J.K. Bassett: Data curation, writing–review and editing. J.E. Joo: Data curation, writing–review and editing. E.M. Wong: Data curation, writing–review and editing. M.T. Brinkman: Writing–review and editing. D.F. Schmidt: writing–review and editing. D.M. Bolton: Writing–review and editing. E. Makalic: Writing–review and editing. T.M. Brasky: Writing–review and editing. A.H. Shadyab: Writing–review and editing. L.F. Tinker: Funding acquisition, writing–review and editing. A. Longano: Writing–review and editing. J.L. Hopper: Methodology, writing–review and editing. D.R. English: Methodology, writing–review and editing. R.L. Milne: Resources, methodology, writing–review and editing. P. Bhatti: Resources, funding acquisition, project administration, writing–review and editing. M.C. Southey: Resources, funding acquisition, project administration, writing–review and editing. G.G. Giles: Conceptualization, funding acquisition, project administration, writing–review and editing. P.-A. Dugué: Conceptualization, formal analysis, supervision, funding acquisition, methodology, writing–original draft, and project administration.

This work was supported by the Australian NHMRC project grants 209057, 251553, 504711, 1043616, and 1164455. Cohort recruitment was funded by Cancer Council Victoria (http://www.cancervic.org.au/) and VicHealth (https://www.vichealth.vic.gov.au/). M.C. Southey is a recipient of a Senior Research Fellowship from the NHMRC (GTN1155163). The WHI program was funded by the National Heart, Lung, and Blood Institute (NIH, U.S. Department of Health and Human Services) through grants 75N92021D00001, 75N92021D00002, 75N92021D00003, 75N92021D00004, and 75N92021D00005.

The Melbourne Collaborative Cohort Study (MCCS) cohort recruitment was funded by VicHealth and Cancer Council Victoria. Cases and their vital status were ascertained through the Victorian Cancer Registry and the Australian Institute of Health and Welfare, including the National Death Index and the Australian Cancer Database. The MCCS component of the work was funded by the Australian NHMRC, including grants 209057, 251553, 504711, 1043616 (to G.G. Giles) and 1164455 (to P.-A. Dugué). The WHI program was funded by the National Heart, Lung, and Blood Institute (NIH, U.S. Department of Health and Human Services) through grants 75N92021D00001, 75N92021D00002, 75N92021D00003, 75N92021D00004, and 75N92021D00005 (to the Fred Hutchinson Cancer Research Center).

We would like to acknowledge the WHI investigators: Program Office: (National Heart, Lung, and Blood Institute, Bethesda, MD) Jacques Rossouw, Shari Ludlam, Joan McGowan, Leslie Ford, and Nancy Geller; Clinical Coordinating Center: (Fred Hutchinson Cancer Research Center, Seattle, WA) Garnet Anderson, Ross Prentice, Andrea LaCroix, and Charles Kooperberg; Investigators and Academic Centers: (Brigham and Women's Hospital, Harvard Medical School, Boston, MA) JoAnn E. Manson; (MedStar Health Research Institute/Howard University, Washington, DC) Barbara V. Howard; (Stanford Prevention Research Center, Stanford, CA) Marcia L. Stefanick; (The Ohio State University, Columbus, OH) Rebecca Jackson; (University of Arizona, Tucson/Phoenix, AZ) Cynthia A. Thomson; (University at Buffalo, Buffalo, NY) Jean Wactawski-Wende; (University of Florida, Gainesville/Jacksonville, FL) Marian Limacher; (University of Iowa, Iowa City/Davenport, IA) Jennifer Robinson; (University of Pittsburgh, Pittsburgh, PA) Lewis Kuller; (Wake Forest University School of Medicine, Winston-Salem, NC) Sally Shumaker; (University of Nevada, Reno, NV) Robert Brunner; Women's Health Initiative Memory Study: (Wake Forest University School of Medicine, Winston-Salem, NC) Mark Espeland. For a list of all the investigators who have contributed to WHI science, please visit: https://s3-us-west-2.amazonaws.com/www-whi-org/wp-content/uploads/WHI-Investigator-Long-List.pdf

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.