An umbrella review of the evidence associating diet and cancer risk at 11 anatomical sites

Cancer causes substantial global morbidity and mortality. The WCRF’s Third Expert Report identifies diet and nutrition (including adiposity and physical inactivity) as modifiable risk factors for several cancers, but many reported diet–cancer associations from observational studies may be non-causal due to measurement error, residual confounding, and publication/reporting biases. Dietary intake is often self-reported once via FFQs and nutrient estimation relies on potentially inaccurate food composition databases. To inform public health policy, the authors conducted an umbrella review to systematically evaluate the robustness, strength, and validity of meta-analytic evidence for associations between foods/nutrients and the risk of developing or dying from cancers at 11 anatomical sites, and to assess whether additional research is likely to change current inferences.

The grading results align largely with independent evaluations by the World Cancer Research Fund (WCRF), where many associations supported here as strong or highly suggestive also received convincing evidence. Differences include coffee with skin cancer and fruits/vegetables with head and neck cancer, which were graded higher here (driven in part by case–control evidence) but were limited/suggestive by WCRF. Conversely, red/processed meat with colorectal cancer and salty foods with stomach cancer received convincing ratings by WCRF but were graded as suggestive here due to not meeting the stringent P ≤ 1e−6 threshold. The NutriRECS consortium similarly found positive associations between red/processed meat and colorectal cancer but rated certainty as low using GRADE. Prior umbrella reviews have also documented diet/obesity/physical activity links to cancer but underscore challenges in causal inference from observational nutrition research.

Design: Umbrella review of meta-analyses of observational studies from the WCRF Continuous Update Project (CUP) through 2018, focusing on diet-related exposures and cancer incidence/mortality at 11 sites (head and neck, esophagus, stomach, colorectal, liver, gallbladder, lung, skin, breast [female], kidney, urinary bladder). Prostate and endometrial umbrella reviews were published separately; pancreatic and ovarian reviews from 2011–2013 were considered outdated. Data sources and selection: WCRF CUP meta-analyses since 2015, predominantly prospective cohort studies; case–control studies included only for head and neck cancer. RCTs identified but generally irrelevant to specific dietary intakes and not included in WCRF CUP meta-analyses. Data extraction: From WCRF reports and original studies when needed: dietary factor, exposure contrast (mostly continuous; top vs bottom or use vs no-use when applicable), cancer outcome, study characteristics (author, year, sex, cases, cohort size/controls, design), effect sizes (RRs, 95% CIs). Multiple authors extracted and cross-verified data; disagreements resolved by discussion. Statistical analysis: Random-effects meta-analysis to estimate summary RRs and 95% CIs; heterogeneity quantified via I² and 95% prediction intervals. Small-study effects assessed using Egger’s test (P ≤ 0.10 plus largest study estimate smaller in magnitude than summary). Excess significance evaluated by comparing observed versus expected number of significant results (expected computed from summed powers using the largest study effect as the plausible effect; P ≤ 0.10 indicates excess). Evidence grading: Nominally significant associations were graded into strong, highly suggestive, suggestive, or weak using pre-specified criteria: strong required >1000 cases, random-effects P ≤ 1e−6, I² < 50%, prediction interval excluding null, and no small-study or excess significance bias; highly suggestive required >1000 cases, P ≤ 1e−6, and largest study significant; suggestive required >1000 cases and P ≤ 1e−3; remaining nominally significant were weak. Research synthesis metrics: For non-significant meta-analyses, computed the number of additional average-weight studies needed to achieve ≥80% conditional power to detect the observed summary effect (also using largest-study effect as alternative). For significant meta-analyses, calculated Rosenberg’s fail-safe number (FSN) of average-weight, null-effect studies required to render the meta-analysis non-significant. Analyses conducted in Stata 14 and R 4.0.3.

• Scope: 860 meta-analytic comparisons included; 91% used continuous exposure contrasts; largest counts for colorectal cancer (221) and breast cancer (163). Median number of studies per meta-analysis was 5 (range 2–33); median cases per meta-analysis ≈ 2152 (range varies by cancer).
• Statistical significance and heterogeneity: 247/860 (28.7%) were nominally significant (P < 0.05); 75 (8.7%) at P < 1e−3; 25 (2.9%) at P < 1e−6. High heterogeneity (I² > 50%) in 227 (26.4%); little heterogeneity (I² ≤ 25%) in 450 (52.3%). Only 46 (5.3%) had prediction intervals excluding the null. Small-study effects in 69 (8.0%); excess significance in 121 (14.1%).
• Evidence grading overall: 10 (1.2%) strong, 13 (1.5%) highly suggestive, 42 (4.9%) suggestive, 182 (21.2%) weak, and 613 (71.3%) not significant.
• Strong evidence (10 meta-analyses; all CRC or breast): • Alcohol and colorectal cancer (CRC): RR per 10 g/day ≈1 drink, 1.07 (95% CI 1.05–1.08); similar for beer and for colon cancer. • Postmenopausal breast cancer (including current MHT users): alcohol RR per 10 g/day, 1.12 (1.09–1.15); wine similarly associated. • Inverse with CRC: dairy products RR per 400 g/day, 0.87 (0.83–0.90); milk RR per 200 g/day, 0.94 (0.92–0.96); calcium high vs low RR, 0.83 (0.79–0.87); whole grains RR per 90 g/day (~3 servings), 0.84 (0.78–0.90).
• Highly suggestive evidence (13 meta-analyses): • Alcohol positively associated with: CRC subtypes; esophageal cancer in men RR per 10 g/day, 1.33 (1.22–1.46); head and neck (oral cancer RR 1.15 [1.09–1.22]; upper aerodigestive tract RR 1.18 [1.11–1.26]); liver cancer mortality RR 1.02 (1.01–1.03); breast cancer subtypes. • Coffee inversely associated with liver cancer RR per 1 cup/day, 0.85 (0.81–0.90) and skin basal cell carcinoma RR 0.95 (0.94–0.97). • Fruits/vegetables inversely with head and neck cancers: pharyngeal (high vs low) RR 0.60 (0.52–0.70); oral (high vs low) RR 0.68 (0.60–0.77). After excluding a pooled case–control project, these attenuated (fruit–pharyngeal no longer significant; vegetables–oral downgraded to weak).
• Suggestive evidence (42 meta-analyses; highlights): • Processed meat per 50 g/day and CRC RR 1.16 (1.08–1.26); red/processed meat per 100 g/day and colon cancer RR 1.19 (1.10–1.29). • Breast cancer inversely with total dietary fiber RR per 10 g/day, 0.95 (0.93–0.98) and soluble fiber RR 0.75 (0.63–0.88). • Lung cancer inversely with serum retinol, α-carotene, fruit, folate, vitamin C; positively with alcohol and red/processed meat. • Coffee inversely with melanoma risk in women; alcohol and serum vitamin D positively with skin basal cell carcinoma; pickled vegetables and salty foods positively with stomach cancer.
• FSN and CP metrics: For 65 meta-analyses with strong/highly suggestive/suggestive evidence, median FSN: suggestive 23 (4–159), highly suggestive 111 (38–856), strong 67 (32–369), always exceeding the number of included studies, indicating robustness to additional null studies. For 182 weak associations, median FSN was 4 (1–42), and in 105 comparisons FSN was smaller than current study count, supporting weak evidence. Among 613 non-significant comparisons, ~78% required >10 additional average-weight studies to reach ≥80% conditional power based on random-effects or largest-study effects, suggesting limited value of further similar studies; only 28 (5%) associations might change with fewer studies than currently included (mostly understudied cancers/exposures).

The umbrella review demonstrates that only a small subset of specific foods/nutrients show robust associations with cancer risk when applying stringent criteria and bias assessments. Alcohol exhibits the most consistent positive associations across multiple cancers (colorectal, breast, esophageal, head and neck, liver), though alcohol-related meta-analyses also show higher heterogeneity and indications of bias, necessitating cautious interpretation. Conversely, calcium, dairy, and whole grains are consistently protective for colorectal cancer, while coffee is inversely associated with liver cancer and basal cell carcinoma of the skin. The overall landscape is characterized by modest effect sizes, limited statistical robustness at stringent thresholds, and evidence of heterogeneity and some reporting/small-study biases. Research synthesis metrics (FSN, CP) indicate that for most associations additional similar observational studies are unlikely to change current inferences, except for certain understudied cancer–diet pairs. These findings address the research question by highlighting which diet–cancer associations are most credible and where further efforts should prioritize methodological innovation (better exposure assessment, biomarkers, repeated measures), mechanistic studies, and exploration of dietary patterns and interactions with other exposome components.

This umbrella review of 860 meta-analyses identifies a limited number of diet–cancer associations with strong or highly suggestive evidence: alcohol increases risk of cancers of the colon, rectum, breast, esophagus, head and neck, and liver; calcium, dairy, and whole grains reduce colorectal cancer risk; and coffee reduces risk of liver cancer and skin basal cell carcinoma. Many other associations remain uncertain and are unlikely to be clarified by more of the same observational studies. Future research should prioritize improved dietary assessment (e.g., repeated web-based records, validated intake biomarkers), investigation of early-life diet and overall dietary patterns, mechanistic pathways, molecular subtypes, post-diagnosis outcomes, and integration with the broader exposome. Public health efforts should focus on established diet-related risks, notably reducing obesity and alcohol consumption.

• Evidence base restricted to observational studies, limiting causal inference; few relevant, adequately powered RCTs exist, and those conducted have generally not supported protective effects for many nutrients.
• Dietary assessment predominantly via FFQs (often single time point, mostly European-descent cohorts), subject to measurement error, especially for episodic foods and nutrient estimates reliant on composition databases; measurement error may attenuate or distort associations, particularly when models adjust for multiple imprecisely measured covariates.
• Possible missed studies or unreported sub-analyses, though WCRF CUP searches are comprehensive; the umbrella review relied on meta-analyses available through 2018.
• Bias diagnostics (Egger’s test, excess significance) have low power with few studies; negative tests do not exclude bias; heterogeneity may reflect true differences or bias.
• Many associations in head and neck cancer relied on case–control studies; results attenuated when pooled case–control data were excluded.