Causal Effects of Breast Cancer Risk Factors across Hormone Receptor Breast Cancer Subtypes: A Two-Sample Mendelian Randomization Study Open Access

Previous case–control and cohort studies provide some evidence for breast cancer subtype–specific risk factors, including reproductive factors (1–3), body mass index (BMI; refs. 2, 4, 5), and alcohol consumption (2, 6, 7), although reported associations are inconsistent across studies (see Supplementary Table S1 for a detailed overview). A likely explanation for this inconsistency is the relatively small number of cases for rarer subtypes, such as HER2-enriched and triple-negative breast cancer. In general, breast cancer studies that collected both risk factor data and detailed pathology data for a large number of women are limited. A second challenge is that results from observational studies are often subject to bias due to (residual) confounding, measurement error, and reverse causation (8). As a result, it remains unclear whether previous results reflect causal associations with breast cancer subtypes.

Mendelian randomization (MR) is a specific type of instrumental variable (IV) analysis that minimizes the risk of these biases through the use of germline genetic variants, provided that certain assumptions are valid (8). Previous MR studies on breast cancer risk supported causal associations for height (9), BMI (10), age at menarche (11), age at menopause (12), breast density (13), and physical activity (14) but not for type 2 diabetes (T2D; ref. 15), alcohol consumption, and ever having smoked regularly (16, 17). Because of the heterogeneity of breast cancer, also within estrogen receptor (ER)–defined subtypes, it is essential to assess whether these associations with hormone receptor subtypes are causal. Thus far, only a handful of MR analyses have included data on these biologically more homogeneous subtypes (18–23). However, the validity of their findings is currently unclear, as most previous MR studies did not perform a full assessment of the different MR assumptions (see Supplementary Table S2 for a detailed overview). Consequently, a comprehensive evaluation of causal evidence for breast cancer subtype–specific risk factors is still lacking.

Therefore, the aim of this study was to estimate and compare the causal effects of nine established risk factors, including anthropometric, reproductive, and behavioral exposures, across five hormone receptor breast cancer subtypes using a two-sample MR design.

We used a two-sample MR (RRID:SCR_019010) study design and summary-level data for both the risk factors and outcomes of interest. Specifically, we extracted summary statistics for genetic variant–risk factor and genetic variant–breast cancer subtype associations from different genome-wide association studies (GWAS), which are described in more detail below. All included GWAS conducted comprehensive quality control of the genetic data. To yield valid causal estimates, selected genetic IVs should meet the following MR assumptions: (i) IVs are robustly associated with the risk factors of interest (relevance assumption), (ii) IVs are not associated with confounders of the studied associations (independence assumption), and (iii) IVs do not affect the risk of breast cancer subtypes through mechanisms that do not involve the risk factors of interest (exclusion restriction assumption; ref. 8). An important additional assumption of two-sample MR is that the study participants included in both samples are from similar underlying populations (homogeneity assumption; ref. 24). We performed extensive secondary analyses to assess whether these assumptions were reasonable.

In 2020, Cancer Research UK published a list of established breast cancer risk factors based on scientific evidence up to that moment (25). From this list, along with a meta-analysis of prospective cohort studies (9), we selected risk factors for which GWAS data were available. As a result, we included the following nine breast cancer risk factors: height, BMI, T2D, age at menarche, age at menopause, percent breast density, alcohol consumption, regular smoking, and overall physical activity. We extracted data for these risk factors from the largest published GWAS, including (mostly) participants of European ancestry (12, 13, 26–33) that were published before September 2021. Details for each included data source, including the percentage of female participants and age, are presented in Table 1, and details about the association models specified by each risk factor GWAS are included in Supplementary Table S3. Due to risk factor transformations that included GWAS performed, estimated MR ORs correspond to a 1 SD increase in risk factors, except for age at menarche and age at menopause, for which ORs correspond to a 1-year increase, and for T2D and smoking, for which ORs correspond to a unit increase in the log odds.

We extracted publicly available summary-level data from the largest Breast Cancer Association Consortium (BCAC) GWAS to date for total breast cancer (n = 247,173; 133,384 cases, 113,789 controls) and the five hormone receptor breast cancer subtypes: luminal A–like (45,253 cases), luminal B–/HER2-negative–like (6,350 cases), luminal B–like (6,427 cases), HER2-enriched (2,884 cases), and triple-negative breast cancer (8,602 cases; ref. 34). These subtypes were classified based on tumor grade, ER, progesterone receptor (PR), and HER2 status as follows: luminal A–like (ER⁺ and/or PR⁺ and HER2⁻ and grade 1 or 2), luminal B–/HER2-negative–like (ER⁺ and/or PR⁺ and HER2⁻ and grade 3), luminal B–like (ER⁺ and/or PR⁺ and HER2⁺), HER2-enriched (ER⁻, PR⁻, HER2⁺), or triple-negative (ER⁻, PR⁻, HER2⁻). The same set of controls (91,477 controls) was used for all subtype-specific GWAS analyses. Participants included in this BCAC GWAS were all female and of European ancestry. All GWAS that generated the used summary-level BCAC data received ethical approval from qualified institutional boards, and all study participants provided informed consent, mostly in accordance with the Declaration of Helsinki and, in a few studies, in accordance with the U.S. Common Rule. Additional details about the study population are included in Table 1. Genetic variant–total breast cancer and genetic variant–subtype associations were estimated using standard logistic regression models and two-stage polytomous logistic models, respectively. Accordingly, the number of cases per hormone receptor subtype represents the effective number of cases per subtype; see the Supplementary Note of Zhang and colleagues (34) for details. We estimated the maximum percentage of overlap with the selected risk factor GWAS based on the number of individuals from studies that were included in both risk factor and BCAC GWAS analyses. For this calculation, we used data provided in the supplemental information of risk factor GWAS and similar details about our BCAC summary statistics (kindly provided by Dr. Haoyu Zhang). We calculated the maximum proportion of overlap by dividing the number of study participants in overlapping studies by the total number of risk factor GWAS participants. For example, between the smoking behavior GWAS and the BCAC GWAS, the only overlapping study was the Women’s Health Initiative. The smoking behavior GWAS included 17,868 Women’s Health Initiative study participants, while the BCAC GWAS included 9,553 participants. Consequently, the maximum percentage overlap of study participants would be (9,553/total smoking behavior GWAS sample size) × 100%.

From each risk factor GWAS, we selected genome-wide significant genetic variants (P < 5 × 10⁻⁸). For height and BMI, an even more stringent P value threshold (P < 1 × 10⁻⁸) was used by the GWAS authors (27). The T2D GWAS (30) was the only data source that included a trans-ethnic population. To avoid confounding due to population stratification (8), we only included genetic variants that reached genome-wide significance in European ancestry–specific T2D analyses. For genetic variants that were not available in the BCAC GWAS summary statistics, we searched for proxy genetic variants [linkage disequilibrium (LD) r² ≥ 0.8] using the NIH LDlink API implemented in the LDlinkR R package (35). To maximize statistical power to detect causal effects, we did not exclude genetic variants based on pairwise correlation (i.e., LD) for our primary analyses but instead accounted for variant correlation in the analysis. Prior to conducting MR analyses, we performed harmonization of alleles and effect estimates between the risk factor and BCAC GWAS using the TwoSampleMR R package (36). At this step, we excluded palindromic genetic variants with intermediate allele frequencies (i.e., adenine/thymine or cytosine/guanine) genetic variants with an effect allele frequency ranging from 0.40 to 0.60) because the harmonization of these specific variants between different data sources can be error-prone. In addition, genetic variants that were not available in the 1000G phase III reference panel were excluded during harmonization. Supplementary Figure S1 presents an overview of this selection process.

All analyses were performed using R version 4.0.5 (https://www.R-project.org/). Prior to performing our primary MR analyses, we set out to calculate LD matrices for all genetic variants for the specific risk factors using the ld_matrix function implemented in the TwoSampleMR R package. However, due to the substantial proportion of highly correlated genetic variants for height and BMI, the correlation matrices that we calculated for these risk factors were near-singular. Therefore, we used a previously published method (37) that performs unscaled principal component analysis on a weighted version of the genetic correlation matrix instead. This method results in transformed values for the genetic variants–risk factor and genetic variants–outcome associations and a transformed correlation matrix. We included these transformed objects in our primary inverse-variance weighted (IVW) analyses.

We employed the IVW method using a multiplicative random-effects model to calculate primary MR estimates for all nine risk factors in relation to each breast cancer subtype. For these analyses, we included genetic variant–risk factor estimates from the largest GWAS sample available. Consequently, we used estimates from sex-combined analyses for the risk factors height, BMI, T2D, smoking behavior, alcohol consumption, and physical activity. If analyses from conditional and joint GWAS analyses (i.e., analyses that identify index and independent secondary genetic variants) were available, we used these estimates to weigh genetic variant–risk factor associations (Supplementary Table S3). We performed post hoc power calculations for subtype-specific associations using a publicly available web application (https://sb452.shinyapps.io/power/; ref. 38). Specifically, we estimated statistical power within each subtype to detect a causal effect estimate equal to the magnitude of the causal effect estimate that we observed for overall breast cancer, as this estimate would be the most accurate under the null hypothesis that there is no heterogeneity across subtypes (see Supplementary Table S4 for used parameters). In addition to estimating causal effects of each risk factor on total breast cancer and breast cancer subtypes, we calculated the I² index to quantify heterogeneity (%) in primary MR estimates across subtypes. We calculated I² estimates through meta-analysis of the five subtype-specific effect estimates per risk factor using random-effects models implemented in the metafor R package (39). MR estimates for the different subtypes were not formally independent, as they were calculated using the same set of controls. However, because the case populations were different, the correlation resulting from this overlap is expected to be minimal. As our results suggested consistently different effect estimates for triple-negative tumors, we also calculated heterogeneity in subtype-specific estimates after exclusion of this subtype as a post hoc analysis.

We additionally performed IVW analyses restricted to uncorrelated genetic variants (LD r² ≤ 0.001). Our rationale for this was twofold: direct comparison with previously published MR estimates for breast cancer and, direct comparison with robust MR analyses, which are not all extended for the inclusion of correlated IVs.

To yield valid causal estimates, genetic IVs have to meet the relevance, independence, and exclusion restriction assumptions. The selection of genetic IVs based on the genome-wide significance threshold is an accepted approach to ensure that genetic variants meet the relevance assumption (40). To quantify the strength of included genetic IVs, we calculated F-statistics based on r² (i.e., the variance explained in the respective risk factor) estimated in independent study populations, if available (Supplementary Table S5). We performed the following robust MR methods to check how consistent our findings were under less stringent assumptions about the pleiotropic effects of the included genetic IVs: MR-Egger regression (41), weighted median (42), mode-based estimator (43), and MR-PRESSO (44). Altogether, the results of these different methods give some insight into the plausibility of the exclusion restriction assumption. A comprehensive overview of each method’s assumptions, strengths, weaknesses, and statistical power was previously published (45). Based on previous MR findings about breast cancer (11), we also performed multivariable MR analyses for BMI and age at menarche. We performed two separate multivariable MR analyses: The first included summary-level data for BMI from the GWAS by Yengo and colleagues (27), and the second included data from the largest BMI GWAS in the UK Biobank by Elsworth (https://gwas.mrcieu.ac.uk/datasets/ukb-b-19953/). The reason for this was that ∼25% of the genetic IVs for age at menarche were missing in the most recent BMI GWAS (27), whereas all variants were available in the UK Biobank GWAS summary statistics (https://gwas.mrcieu.ac.uk/datasets/ukb-b-19953/).

In addition to these three fundamental assumptions, the homogeneity assumption should be valid in two-sample MR studies (24). However, the BCAC GWAS only included women, whereas six of the nine selected risk factor GWAS included both females and males (Table 1). Accordingly, the homogeneity assumption is not met by design. To assess potential bias because of this violation, we conducted secondary analyses in which we replaced the betas of the genetic variants that reached genome-wide significance in sex-combined analyses with female-specific betas. For height and BMI, we extracted female-specific betas from a previous GIANT GWAS (46, 47). Of the 3,146 genetic variants for height, only 3,002 genetic variants were available in this female-specific data source. For the remaining 144 genetic variants, we included weights from sex-combined GWAS analyses. For physical activity, we received female-specific betas from the GWAS authors (33). For T2D, alcohol consumption, and smoking behavior, female-specific estimates were not publicly available.

We ultimately combined results from our primary and secondary MR analyses to evaluate the strength of evidence for causal effects of each risk factor on each breast cancer subtype. For this evaluation, we used the following recently proposed definition (48): Evidence was considered to be “consistent” if all performed MR methods presented a P value < 0.05; evidence was considered to be “concordant” if at least one method (primary or secondary analysis) had a P value < 0.05 and the direction of the effect estimate was concordant across all methods; evidence was considered to be “inconsistent” if at least one method had a P value < 0.05, but the direction of the effect estimates differed across methods; and evidence was considered to be “inadequate” if all MR methods had a P value ≥ 0.05. We used the same classification as the mentioned reference, but different labels for the categories that we believe are more reasonable.

All R codes, including details on package versions, that were used to generate our results are available at https://github.com/SchmidtGroupNKI/MR_BCsubtypes.

Details about the setting and participants included in each risk factor GWAS and the breast cancer subtype GWAS are presented in Table 1. The total sample size of the included GWAS ranged from 24,192 to 1,232,091 individuals for breast density and smoking behavior, respectively. Except for the T2D GWAS, all data sources were restricted to European study populations (T2D GWAS: 79.2% European). GWAS for height, BMI, T2D, alcohol consumption, regular smoking, and physical activity included both females and males. For the latter three risk factors, ∼50% of the study subjects were female; for the anthropometric-related risk factors, details on biological sex were insufficiently reported. The age distribution of study participants included in the breast cancer subtype GWAS and risk factor GWAS was similar with a reported mean age of more than 55 years. We estimated the maximum overlap in study participants between the risk factor and BCAC GWAS to range from 0% (height, breast density, and physical activity) to 16.3% (age at menarche).

Causal effect estimates (OR per 1 SD increase for all risk factors but per 1 year increase for age at menarche and age at menopause or per unit increase in the log odds for T2D and smoking) for each of the nine breast cancer risk factors across the five hormone receptor breast cancer subtypes are presented in Fig. 1 and Table 2. Statistical power estimates corresponding to our primary MR estimates are presented in Supplementary Table S4. In general, we observed that IVW estimates for luminal A–like and luminal B–/HER2-negative–like subtypes were very similar to IVW estimates for overall breast cancer. Heterogeneity across subtype-specific causal effects was not due to opposite causal effect estimates but due to stronger estimates or the absence of an effect for some subtypes.

For height, primary IVW estimates were similar across breast cancer subtypes (I² = 0%) but suggested only a causal risk–increasing effect of increasing height on luminal A– and luminal B–like tumors. Estimated heterogeneity across hormone receptor subtypes remained 0% after exclusion of the estimate for triple-negative tumors. The combination of our primary and secondary MR analyses provided consistent evidence for a causal risk–increasing effect of increased height on the risk of luminal A–like tumors and concordant evidence for causal associations with luminal B–/HER2-negative–like and luminal B–like tumors (Table 2; Supplementary Fig. S2). Evidence for causal associations with HER2-enriched and triple-negative tumors was inadequate. However, MR analyses including female-specific genetic IVs provided consistent evidence for a causal risk–increasing effect of increased height on luminal B–like tumors and concordant evidence for the other four subtypes (Supplementary Table S6; Fig. 2).

For increasing BMI, primary IVW estimates provided evidence for a causal risk–decreasing effect on luminal B–like and HER2-enriched tumors [OR_lumB = 0.91; 95% confidence interval (CI), 0.84–0.99 and OR_HER2+ = 0.85; 95% CI, 0.76–0.95]. Accordingly, I² estimates indicated moderate heterogeneity across subtype-specific estimates (I² = 31.1% across all subtypes; I² = 40.1% after exclusion of triple-negative tumors). However, based on the combination of primary and secondary MR analyses, evidence for a causal effect of BMI was merely inconsistent for luminal B–like and HER2-enriched tumors and inadequate for the other subtypes (Table 2; Supplementary Fig. S3). In contrast, analyses including female-specific genetic IVs provided concordant evidence for a causal risk–decreasing effect of increasing BMI on all hormone receptor subtypes (Supplemental Table S6; Fig. 2).

For increasing risk of T2D, our primary analyses only suggested a causal risk–decreasing effect on luminal A–like tumors although causal effect estimates for the other subtypes were very similar (I² = 0% for analyses including and excluding triple-negative tumors). Altogether, our primary and secondary MR analyses provided concordant evidence for a causal risk–decreasing effect on the risk of HER2-enriched tumors. Evidence for causal associations with luminal A–like and triple-negative tumors was inconsistent but inadequate for luminal B–like and luminal B–/HER2-negative–like tumors (Table 2; Fig. 2; Supplementary Fig. S4).

We observed no evidence for a causal effect of higher age at menarche on any of the hormone receptor subtypes in primary MR analyses (I² = 0%). Secondary univariable MR analyses supported these findings (Table 2; Supplementary Fig. S5), but multivariable MR analyses for higher age at menarche and increasing BMI provided evidence for a direct causal association between higher age at menarche and a decreased risk of luminal A–like, luminal B–/HER2-negative–like, and triple-negative breast tumors (Supplementary Fig. S6). However, corresponding heterogeneity estimates suggested similar effects across subtypes [I² = 15.1% based on data from Yengo and colleagues (27); 0% based on UK Biobank data (https://gwas.mrcieu.ac.uk/datasets/ukb-b-19953/)]. Based on the combination of our primary and secondary analyses, including multivariable MR analyses, evidence for a causal effect of higher age at menarche was only concordant for luminal A–like tumors and inconsistent for luminal B–/HER2-negative–like and triple-negative tumors (Fig. 2).

For higher age at menopause, primary IVW estimates suggested causal risk–increasing effects on luminal A–like, luminal B–/HER2-negative–like, and HER2-enriched tumors but not on luminal B–like and triple-negative tumors (I² = 42.1%). Estimated heterogeneity across hormone receptor subtypes, excluding triple-negative breast cancer, was 0%, which indicates that the absence of a causal effect on this specific subtype [OR = 1.01 (95% CI, 0.98–1.04)] explains the observed heterogeneity across all five subtypes. Collectively, our MR analyses provided consistent or concordant evidence for a causal effect of age at menopause on all subtypes except triple-negative breast cancer (Table 2; Fig. 2; Supplementary Fig. S7).

In primary IVW analyses, genetically predicted higher breast density was only significantly associated with an increased risk of luminal B–/HER2-negative–like breast cancer (I² = 15.6% across all subtypes; I² = 0% after exclusion of triple-negative tumors). Based on the combination of all used MR methods, evidence for a causal effect was inconsistent for luminal A–like and luminal B–/HER2-negative–like tumors and inadequate for the other subtypes (Table 2; Fig. 2; Supplementary Fig. S8).

Primary analyses did not provide evidence for causal effects of alcohol consumption and regular smoking on the risk of any of the hormone receptor subtypes (I²_alcohol = 0%; I²_smoking = 0%). Two out of the five secondary MR analyses suggested a causal risk–decreasing effect of higher alcohol consumption on the risk of luminal B–like breast tumors but not on the other subtypes (Table 2; Supplementary Fig. S9). Consequently, our results only provide concordant evidence for a causal effect of alcohol consumption on this specific subtype (Fig. 2). Secondary analyses for regular smoking supported our primary findings (Table 2; Supplementary Fig. S10), and thus, evidence for a causal association between smoking and the risk of any of the hormone receptor subtypes is inadequate. For overall higher physical activity, our primary results only provided evidence for a causal risk–decreasing effect on the risk of luminal B–/HER2-negative–like breast tumors (I² = 0%; Table 2; Supplementary Fig. S11). However, statistical power was generally low, except for the luminal A–like subtype (Supplementary Table S4). Moreover, CIs around primary causal effect estimates were very wide, indicating a high degree of uncertainty. Based on the combination of our primary and secondary MR analyses, we found concordant evidence for a causal effect of physical activity on luminal A–like breast cancer and inconsistent evidence for luminal B–/HER2-negative–like tumors. For the other subtypes, evidence for a causal effect was inadequate. MR analyses including female-specific IVs provided concordant evidence for a causal risk–decreasing effect on luminal A–like, luminal B–/HER2-negative–like, and HER2-enriched breast tumors (Supplementary Table S6). In these analyses, evidence for a causal association with luminal B–like tumors shifted to inconsistent, whereas the evidence for a causal association with triple-negative tumors remained inadequate (Fig. 2). Yet the results of secondary MR analyses for physical activity should be interpreted with caution due to the very low number of genetic IVs available for this risk factor.

This MR study indicates that the causal effects of several established breast cancer risk factors differ across hormone receptor breast cancer subtypes. Specifically, we observed moderate heterogeneity in subtype-specific causal effects for age at menopause and breast density. Although this heterogeneity was explained by null findings for triple-negative tumors, statistical evidence was also inadequate for causal associations of breast density with luminal B–like and HER2-enriched tumors. For height, BMI, risk of T2D, age at menarche, alcohol consumption, regular smoking, and physical activity, causal effect estimates were similar across breast cancer subtypes. However, for height and BMI, the evidence for causal effects was concordant, or stronger, for all five subtypes. In contrast, for regular smoking, statistical evidence for a causal effect was inadequate for all subtypes. For the remaining six risk factors, the strength of causal evidence ranged from concordant to inadequate across subtypes. There was no evidence of opposing effects of any risk factor across hormone receptor subtypes. Altogether, our findings suggest that it is more likely that there is heterogeneity in the presence or absence of causal associations between risk factors and breast cancer subtypes than in the magnitude and direction of causal effects.

Previous MR studies supported height, BMI, age at menarche, age at menopause, breast density, and physical activity as breast cancer risk factors but could not confirm the risk of T2D, alcohol consumption, and smoking behavior as causal risk factors for overall breast cancer. The majority of patients with breast cancer are diagnosed with luminal tumors (49). As a result, established risk factor–breast cancer associations will, in general, represent associations with luminal subtypes and potentially less with HER2-enriched and triple-negative tumors. Our results indeed indicate that for triple-negative breast cancer, there is concordant evidence of causality for only two out of the nine risk factors, whereas for the other subtypes, there is concordant evidence of causality for four to five risk factors. For HER2-enriched tumors, differences compared with luminal tumors were less clear, possibly due to the considerably lower statistical power for the HER2-enriched subtype. For triple-negative tumors, we consider it likely that null findings for age at menopause, breast density, and physical activity are not caused by insufficient statistical power because the sample size for this subtype was similar to that for luminal B–/HER2-negative–like and luminal B–like tumors, and causal ORs were consistently 1.

Considering results from observational studies (i.e., not MR), recent large-scale analyses for height and age at menopause in relation to hormone receptor breast cancer subtypes are in line with our findings. Specifically, increasing height was associated with a higher risk of ER⁺PR⁺, ER⁺PR⁻, and ER⁻PR⁻ postmenopausal breast cancer (50). In a recent BCAC analysis of self-reported data, age at menopause was also not associated with the risk of triple-negative tumors (1). Results from the same two studies for BMI and age at menarche were, however, not in line with our findings. A higher adult BMI was associated with a lower risk of ER⁺PR⁺ premenopausal breast cancer and a higher risk of ER⁺PR⁺ and ER⁻PR⁻ postmenopausal breast cancer (50). Evidence for associations between BMI and other subtypes was less clear although an association between higher BMI and a lower risk of breast cancer was suggested for ER⁺PR⁻ postmenopausal tumors. The results from our analyses, including female-specific weights, suggest that a higher genetically determined BMI is associated with a lower risk of all hormone receptor subtypes. Our finding is in line with an earlier MR study that found that a higher BMI was associated with a lower risk of both pre- and postmenopausal breast cancer (51), which contradicts observational evidence showing that a higher BMI is associated with a higher risk of postmenopausal breast cancer (e.g.; ref. 52). Although summary-level BCAC data stratified by age are currently not available for the five hormone receptor subtypes that we studied, new summary-level data have been calculated by Dr. Kyriaki Michailidou for ER⁺ and ER⁻ breast cancer across three age categories (<40, 40–55, and >55 years). We used these data to further investigate potential differences in the association between BMI and pre- and postmenopausal breast cancer, but these analyses did not support such a difference (Supplementary Table S7). Based on observational BCAC data, a younger age at menarche was associated with all hormone receptor subtypes (1), whereas our MR results only provided concordant causal evidence for an association with luminal A–like tumors. Yet estimated heterogeneity across subtypes in multivariable MR estimates for age at menarche was negligible, and statistical power to detect a causal effect was limited. Two recent observational studies suggested that breast density was also associated with all subtypes (4, 53), but based on our results, evidence for a causal association is weak. Large-scale analyses that studied associations with hormone receptor breast cancer subtypes are currently lacking for the risk of T2D, alcohol consumption, smoking behavior, and physical activity, which hampers a meaningful comparison with our findings for these risk factors. However, systematic reviews of observational studies for several of these lifestyle-related traits indicate that their results for overall breast cancer are likely to be biased by unmeasured confounding (e.g.; ref. 54). This observation highlights the importance of approaches that are more robust to residual confounding and measurement error, such as MR, to understand the etiology of breast cancer subtypes.

Until now, only one other MR analysis has set out to investigate multiple known risk factors in relation to hormone receptor breast cancer subtypes (18). This previous study reported similar results for age at menopause, which was associated with all subtypes but triple-negative tumors. Furthermore, a subtype-specific causal effect for alcohol consumption was reported, which was suggested to be causally associated only with the risk of HER2-enriched breast cancer. However, their MR-PRESSO and MR-Egger estimates for alcohol consumption were in line with our evidence for a causal risk–decreasing effect on the risk of luminal B–like tumors. This contradiction illustrates the added value of our approach to evaluate causal evidence based on the combination of six different MR methods (seven for age at menarche). Consequently, our conclusions are less likely to reflect invalid causal inferences due to unbalanced horizontal pleiotropy or due to chance findings because of multiple testing. We also assessed the homogeneity assumption through the inclusion of female-specific genetic IVs for height, BMI, and physical activity and showed that these instruments were consistently associated with stronger causal evidence across all breast cancer subtypes compared with combined-sex genetic IVs. In line with this observation, our findings for the risk of T2D, alcohol consumption, and regular smoking should be considered preliminary until female-specific IVs for these risk factors are used in future MR analyses. Altogether, our results underline the importance of an extensive investigation of the MR assumptions.

Another frequently unassessed assumption for two-sample MR analyses is that the risk factor and outcome GWAS samples should be independent, that is, there should be no overlap in study participants (24). As the BCAC includes several studies that also participate in other consortia, this assumption was not completely met in the current analysis. Based on reported details by the included GWAS, we estimated a relatively small overlap in participants, ranging from 0% to 16.3%. Bias due to sample overlap in two-sample MR studies arises in analyses that include weak genetic instruments. In the case of minimal sample overlap, this bias will be toward the null and thus rather increase type 2 error rates than type 1 error rates (55). For each risk factor, we estimated maximum and minimum F-statistics corresponding to primary IVW analyses for luminal A–like and HER2-enriched tumors, respectively. We based these estimations on the variance explained by the included genetic variants in an independent study population, if such independent r² estimates were available. This approach minimized the overestimation of F-statistics due to winner’s curse bias in the discovery GWAS. Minimum F-statistics for height, BMI, alcohol, and smoking were below the arbitrary threshold of 10, which indicates that weak instrument bias may have biased causal effect estimates for these risk factors toward the null. A last important assumption for two-sample MR studies is that regression models employed for the risk factor GWAS and the outcome GWAS should be adjusted for the same covariates (24). Specifically, adjustment for potential confounders in the risk factor GWAS can induce collider bias in two-sample MR analyses. In the current analysis, such bias may have affected our results for breast density because its GWAS was adjusted for BMI. A higher BMI is associated with decreased breast density (56), but without data on the causal association and confounding structure among the genetic IVs, breast density, BMI, and breast cancer subtypes, it is difficult to evaluate the potential impact on our results (24, 57).

An additional strength of our study compared with previous MR studies that set out to investigate associations between breast cancer risk factors and the hormone receptor subtypes is that we maximized the statistical power of our primary analyses through the inclusion of as many genetic IVs as possible in combination with LD matrices. Although these correlation matrices were estimated based on genetic data of only ∼400 participants, causal estimates from our primary analyses were very similar to estimates from our secondary, more conservative analyses. Despite our efforts, our post hoc power analyses indicated that the statistical power to detect causal risk factors across all subtypes was still suboptimal. Future MR studies including stronger genetic IVs, that is, genetic IVs explaining more variance in the risk factor of interest, could further increase statistical power. However, the identification of additional loci requires even larger GWAS study populations, and this has proved to be challenging for lifestyle-related risk factors such as alcohol consumption, smoking, and physical activity. Another possibility to increase statistical power is the inclusion of larger numbers of breast cancer cases for hormone receptor subtypes, which will be possible through initiatives like the Confluence project (58). Further improvement in the quality of MR studies can be achieved if future risk factor GWAS stratify analyses by biological sex and report full details about the included GWAS settings, participants, and methods.

In conclusion, our results suggest that of the established breast cancer risk factors, height and BMI are likely to exert similar causal effects across all breast cancer subtypes. Our MR analyses also suggest that the majority of established breast cancer risk factors are not causally associated with the risk of triple-negative tumors. These insights are valuable for the development of primary prevention strategies and the improvement of breast cancer risk stratification in the general population. Our findings also emphasize the importance of taking breast cancer subtype into account for the identification of novel breast cancer risk factors.

M. Shokouhi reports grants from the Antoni van Leeuwenhoek Foundation during the conduct of the study. M.K. Schmidt reports grants from the European Union and the Antoni van Leeuwenhoek Foundation during the conduct of the study. No disclosures were reported by the other authors.

R.M.G. Verdiesen: Formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. M. Shokouhi: Formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. S. Burgess: Conceptualization, supervision, validation, writing–review and editing. S. Canisius: Conceptualization, supervision, validation, writing–review and editing. J. Chang-Claude: Conceptualization, supervision, validation, writing–review and editing. S.E. Bojesen: Conceptualization, supervision, validation, writing–review and editing. M.K. Schmidt: Conceptualization, resources, supervision, funding acquisition, validation, project administration, writing–review and editing.

We thank Dr. Haoyu Zhang for providing additional details about the BCAC GWAS summary statistics and Professor Aiden Doherty for providing female-specific summary statistics for overall physical activity. We thank Renee Menezes for help with MR-PRESSO. We acknowledge the BCAC for providing summary data. We gratefully acknowledge the Genetic Epidemiology Group at Cambridge University for their help in generating summary estimates, especially Dr. Kyriaki Michailidou. R.M.G. Verdiesen was funded by the European Union’s Horizon 2020 Research and Innovation Programme (grant number 633784 for the B-CAST project) granted to M.K. Schmidt. M. Shokouhi is funded by the Antoni van Leeuwenhoek Foundation. All research at the Netherlands Cancer Institute is supported by institutional grants from the Dutch Cancer Society and the Dutch Ministry of Health, Welfare and Sport. S. Burgess is supported by the Wellcome Trust (225790/Z/22/Z) and the United Kingdom Research and Innovation Medical Research Council (MC_UU_00002/7, MC_UU_00040/01). The breast cancer genome-wide association analyses for BCAC were supported by Cancer Research UK (PPRPGM-Nov20\100002, C1287/A10118, C1287/A16563, C1287/A10710, C12292/A20861, C12292/A11174, C1281/A12014, C5047/A8384, C5047/A15007, C5047/A10692, C8197/A16565); the Gray Foundation; the NIH (CA128978, X01HG007492: the DRIVE Consortium); the PERSPECTIVE project supported by the Government of Canada through Genome Canada and the Canadian Institutes of Health Research (grant GPH-129344); the Ministère de l’Économie, Science et Innovation du Québec through Genome Québec and the PSRSIIRI-701 grant; the Quebec Breast Cancer Foundation; the European Community’s Seventh Framework Programme under grant agreement number 223175 (HEALTH-F2-2009-223175; COGS); the European Union’s Horizon 2020 Research and Innovation Programme (634935 and 633784); the Post-Cancer GWAS initiative (U19 CA148537, CA148065, and CA148112: the GAME-ON initiative); the Department of Defence (W81XWH-10-1-0341); the Canadian Institutes of Health Research (CIHR) for the CIHR Team in Familial Risks of Breast Cancer (CRN-87521); the Komen Foundation for the Cure; the Breast Cancer Research Foundation; and the Ovarian Cancer Research Fund. All studies and funders are listed in Zhang and colleagues (34).