We conducted a comprehensive search using keywords across PubMed, Web of Science, and Embase databases. Our search encompassed meta-analyses that examined the association between non-genetic risk factors for BC from database inception to July 27, 2023. Supplementary Table 1 outlines the complete search strategy. We also manually reviewed reference lists of eligible studies.

This study followed the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guideline (16) and was registered at PROSPERO (https://www.crd.york.ac.uk/PROSPERO/) with CRD42023470151.

Two authors independently searched the databases, screened titles and abstracts, and identified meta-analyses meeting the inclusion criteria through full-text reading. A third author resolved any discrepancies. Inclusion criteria involved meta-analyses with: 1. observational cohort study designs; 2. non-genetic risk factors as the exposure of interest; 3. BC as the reported outcome; 4. available risk estimates between risk factors and BC (risk ratio, odds ratio, hazard ratio) with 95% confidence intervals (CIs), number of cases/controls, or total population size; 5. publication in English; and 6. study population comprising women. We excluded meta-analyses of genetic markers, systematic reviews without quantitative analyses, animal or laboratory studies, and reviews lacking study-specific data (risk ratio, odds ratio, hazard ratio) that could not be retrieved from the original studies. We also excluded studies with baseline populations already diagnosed with cancer.

When two or more reviews examine the same exposure and outcome, overlapping associations may result in multiple primary studies being included in the reviews. Additionally, research indicates that half of published reviews become outdated within 5.5 years. We first identified meta-analyses with identical risk factors and outcomes to mitigate bias in interpreting such outcomes. Subsequently, we:

Overlap was categorized as follows: 1. very high: CCA >15%, 2. high: CCA=11-15%, 3. moderate: CCA=6-10%, and 4. slight: CCA=0-5% (17). Both reviews were retained for slight or moderate overlap. For high or very high overlap, the review with the highest quality based on the AMSTAR 2 tool was prioritized. The most recent review was included when quality ratings were identical (18, 19).

Two reviewers independently extracted data, including the first author’s name, journal name, publication year, study design, exposure factors, health outcomes, included studies, cases, total participants, and the estimated summary effects with 95% CIs, from eligible studies. Additionally, if available, information relevant to dose-response relationships was retrieved from meta-analyses. To ensure transparency, a third author reviewed the discrepancies between the two reviewers and resolved them by considering all relevant data and methodologies.

Two reviewers (WZ and XYQ) evaluated the methodological quality of the included studies using the AMSTAR-2 tool (20), assigning an overall score. In case of disagreement, a third author (GS) was consulted. AMSTAR-2 assesses 16 items, with 7 considered critical. Deficiencies in any critical item may affect the overall review quality. The key areas deemed crucial included: protocol registered before the commencement of the review; adequacy of the literature search; justification for excluding individual studies; risk of bias in the included studies; appropriateness of meta-analytical methods; considering bias risk when interpreting results; and assessment of the presence and likely impact of publication bias. The final ratings were classified into four levels ranging from “high” to “very low”: 1. high: no defects or one non-critical area with defects; 2. moderate: more than one non-critical area with defects; 3. low: one critical area with or without non-critical areas with defects and 4. very low: more than one critical area with or without non-critical areas with defects. AMSTAR-2 scoring results assess the quality of the included studies and account for potential biases and methodological limitations, providing a more reliable interpretation of results.

BC risk factors were categorized into four evidence-based classes: class I (strong evidence), class II (highly suggestive evidence), class III (suggestive evidence), and class IV (weak evidence) (21, 22) ( Table 1 ). This classification system offers an objective, standardized approach, consistent with other grading schemes in cancer epidemiology. This classification method enables the study to rank the strength of evidence for each risk factor. class I and II evidence correspond to factors with a high level of confidence, whereas class III and IV represent factors with lower confidence. Specifically, class IV (weak evidence) factors are associated with reduced confidence, which may be influenced by factors such as data heterogeneity, insufficient sample size, or methodological limitations.

We focused on cohort studies, recalculating summary effects and 95% CIs using random- or fixed-effects models (23). Heterogeneity was evaluated using the I² statistic and 95% predicted intervals (PIs) (24–26) to account for variability across studies and reduce its impact on the results. We evaluated the presence of publication bias by assessing small-study effects and excess significance bias. Egger’s regression asymmetry test (P<.10) and whether the summary estimate of random effects exceeded the point estimate of the largest study in the meta-analysis were used to assess small-study effects (27). Additionally, sensitivity analyses were conducted on all eligible cohort and case-control studies using the same criteria as those used in the primary analysis. R software version 4.3.0 was used for all statistical analyses (http://cran.r-project.org/).

We performed a two-sample MR analysis using genetic variants as proxies for exposure to explore the causal impacts of non-genetic risk factors on BC (28) Genome-wide association study (GWAS) catalogs were queried for exposures rated as weak evidence or higher in the umbrella review to find relevant GWAS offering summary-level genetic information (29). We used data from previous GWAS to ascertain the association between single nucleotide polymorphisms (SNPs) and BC risk (30).

Given the necessity of MR analysis to validate the three core assumptions (31), we established stringent criteria for selecting the instrumental variables (IVs) to ensure the robustness and reliability of the results (32). First, we set a significance threshold of P<5e−08 to choose SNPs as IVs for each exposure. Second, we addressed linkage disequilibrium (LD) between SNPs by excluding those with strong LD (r2 = 0.001, κb=10,000 kb). The PhenoScanner database was used to mitigate the effects of confounding factors. Finally, we computed the F-statistic for all selected SNPs to evaluate weak instrument bias, excluding SNPs with an F-statistic <10, to ensure that all remaining SNPs were strongly associated with exposure (31).

Causal impacts were estimated using the inverse variation-weighted (IVW) MR as the primary analysis. In instances where exposure exhibited significant effects in the main IVW MR analysis, we applied various robust MR methods (such as the weighted median and MR-Egger) in the sensitivity analyses to address potential violations of the IV assumption (33). We systematically employed leave-one-out modeling to evaluate the potential influence of outliers and pleiotropic SNPs (34). We excluded one SNP at a time to determine whether individual SNPs affected the primary causal relationship. Considering the potential for genetic pleiotropy, interactions, and confounding effects among the different exposure factors, we conducted multivariable MR (MVMR) analyses (35) to assess the direct effects of various exposure factors on BC. To further investigate reverse causation, we performed a reverse MR analysis using BC-related SNPs as IVs (treating BC as the exposure and various risk factors as outcomes).

Figure 1 depicts the literature retrieval and selection process. Up to July 2023, we retrieved 58,242 articles. After screening and following the exclusion of duplicate meta-analyses based on assessed exposure and outcome (36–51) ( Supplementary Table 7 ), 218 articles, including 427 meta-analyses remained (52–269).

Among the 427 included meta-analyses, the estimated values from studies ranged from 2 to 69 (median, 10). The median number of cases and individuals for each meta-analysis were 14,055 and 3,328,403, respectively. The minimum and maximum numbers of cases in meta-analyses were 138 and 591,297, respectively. The smallest total population was 1,083, whereas the largest was 18,281,388. Furthermore, 403 of the 427 meta-analyses included >1,000 patients with BC.

From the pool of 427 meta-analyses, 281 consisted exclusively of cohort studies, including at least two that evaluated 202 risk factors and categorized into 10 major groups. ( Figure 2 and Supplementary Table 2 ).

We evaluated the methodological quality of 218 studies, which included meta-analyses of 427 observational studies, using the AMSTAR-2 tool ( Supplementary Table 4 ). Overall, most meta-analyses exhibited low to very low quality owing to various factors such as study design limitations and reduced methodological rigor (such as potential publication bias, indirectness, and inconsistency). Specifically, only a small proportion was rated as having “high” or “moderate” quality.

When cohort and case-control studies were included ( Supplementary Table 6 ), additional four exposure factors associated with increased BC incidence provided strong evidence: autoimmune thyroiditis (2.71; 95% CI: 2.13-3.43), weight gain (1.55; 95% CI: 1.40-1.71), oral progestogen (1.28; 95% CI: 1.19-1.39), and light exposure at night (1.13; 95% CI: 1.09-1.16). Strong evidence for two exposures reducing BC incidence was also found: number of births (0.79; 95% CI: 0.75-0.83) and sex hormone-binding globulin (0.65; 95% CI: 0.58-0.73); both associations were only included in case-control studies and were not evaluated in the main analysis. Conversely, when case-control studies were included, the strong correlations between breast density, BMI ≥25 (high vs. low, postmenopausal), and BC risk were downgraded to highly suggestive evidence (2.89; 95% CI: 2.57-3.25) and not significant (0.86; 95% CI: 0.81-0.91), whereas the remaining three strong associations remained strong.

In the MR analysis, 22 risk factors had available genetic instruments ( Supplementary Table 8 ). The genetically predicted IGF-1 concentrations demonstrated a correlation with BC (1.08; 95% CI: 1.02-1.14) ( Figure 4 , Supplementary Table 9 ). Additionally, higher fruit intake (0.64; 95% CI: 0.46-0.90), cheese intake (0.82; 95% CI: 0.69-0.98), higher serum TG levels (0.91; 95% CI: 0.86-0.97) and higher BMI category (0.82; 95% CI: 0.71-0.95) were found to have protective effects against BC. The sensitivity analysis yielded directions consistent with the main analysis, supporting potential causal effects ( Supplementary Figures 1 , 2 , Supplementary Table 10 ). Reverse MR-IVW analysis ( Supplementary Figure 3 ) indicated a potential bidirectional relationship between calcium channel blockers (CCBs) and BC. No other exposures demonstrated a similar reverse causal association with BC incidence.

In the MVMR analysis, genetic predictions of CCBs and IGF-1 concentrations showed independent associations with BC after adjusting for BMI and serum TG levels. Similarly, the genetic predictions of fruit and cheese intake were independently linked to a decreased risk of BC ( Supplementary Table 11 ).