Outline

EPIC is an on-going multicentre prospective cohort study designed to investigate the associations between diet, lifestyle, genetic and environmental factors and various types of cancer. A detailed description of the methods has previously been published [20], [21]. In summary, 521,448 participants (∼70% women) mostly aged 35 years or above were recruited between 1992 and 2000. Participants were recruited from 23 study centres in ten European countries: Denmark, France, Germany, Greece, Italy, the Netherlands, Norway, Spain, Sweden, and United Kingdom (UK). Participants were recruited from the general population, with the following exceptions: the French cohort were teacher health insurance programme members; the Italian and Spanish cohorts included members of blood donor associations and the general population; the Utrecht (the Netherlands) and Florence (Italy) cohorts contained participants from mammographic screening programs; the Oxford (UK) cohort included a large proportion of vegetarians, vegans, and low meat eaters; finally, only women participated in the cohorts of France, Norway, Naples (Italy) and Utrecht (the Netherlands). Written informed consent was provided by all study participants. Ethical approval for the EPIC study was obtained from the review boards of the International Agency for Research on Cancer (IARC) and local participating centres. Exclusions prior to the onset of the analyses included: participants with prevalent cancer at enrolment (n = 28,283); participants with missing dietary, lifestyle, and anthropometric data (n = 6,253); participants in the highest and lowest 1% of the distribution for the ratio between energy intake to estimated energy requirement (n = 9,600); and finally participants with extreme total dairy intakes above 2000 g/day (n = 190). Our study therefore included 477,122 participants (334,981 women and 142,141 men).

Diet and lifestyle questionnaires

Dietary information over the previous 12 months was obtained at study baseline using validated country/centre specific dietary questionnaires. In Malmö (Sweden), a dietary questionnaire was combined with a 7-day food registration and interview. In Greece, two Italian centres, and Spain, interviewers administered the dietary questionnaires. In all other centres/countries, the questionnaires were self-administered. In Spain, France, and Ragusa (Italy) questions were structured by meals, while in other countries the structure was by food groups. Also at baseline, standardized computer-based single 24-hour dietary recalls (24-hdr) were collected from 36,994 study participants. This additional dietary assessment was used to calibrate for differences in questionnaires across countries [22]. Individual dairy products were categorized as milk, cheeses, and yoghurts. Due to relatively low intakes and incomplete measurements across centres, other individual dairy products such as ice cream, cream desserts and milk-based puddings, milk beverages, dairy creams and creamers for milk and coffee were not analysed individually. Total milk was assessed as the sum of all types of milk consumed (whole-fat, skimmed, semi skimmed, and not specified). Semi-skimmed milk was defined as milk containing 0.5–2.5% fat, and skimmed milk was defined as having <0.5% fat content. Milk subtype information was unavailable in Norway, and only partially available in Germany, Greece (both whole-fat milk only), and three Italian centres (Florence, Varese, Turin; whole-fat and semi-skimmed milks only). Cheese included all kinds of fresh, fermented, and matured cheese. Yoghurt included natural and flavoured in all cohorts, and additionally fermented milk in Sweden, Norway, and Denmark. Intakes of calcium were obtained from the EPIC Nutrient Data Base (ENDB); in which the nutritional composition of foods across the different countries has been standardized [23].

Lifestyle questionnaires were used to obtain information on education (used as a proxy for socioeconomic status), smoking status and intensity, alcohol consumption, and physical activity levels. Height and weight were measured at the baseline examination in all centres apart from part of Oxford, and all of the Norway and France sub-cohorts, where measurements were self-reported via the lifestyle questionnaire [20].

Ascertainment of colorectal cancer incidence

Population cancer registries were used in Denmark, Italy, the Netherlands, Norway, Spain, Sweden and the United Kingdom to identify incident cancer diagnoses. In France, Germany and Greece cancer cases during follow-up were identified by a combination of methods including: health insurance records, cancer and pathology registries, and by active follow-up directly through study participants or through next-of-kin. Complete follow-up censoring dates varied amongst centres, ranging between 2005 and 2010.

Cancer incidence data were coded using the 10^th Revision of the International Classification of Diseases (ICD-10) and the second revision of the International Classification of Disease for Oncology (ICDO-2). Proximal colon cancer included those within the caecum, appendix, ascending colon, hepatic flexure, transverse colon, and splenic flexure (C18.0–18.5). Distal colon cancer included those within the descending (C18.6) and sigmoid (C18.7) colon. Overlapping (C18.8) and unspecified (C18.9) lesions of the colon were grouped among colon cancers only. Cancer of the rectum included cancer occurring at the recto sigmoid junction (C19) and rectum (C20).

Statistical analysis

Hazard ratios (HRs) and 95% confidence intervals (CIs) were estimated using Cox proportional hazards models. Age was the primary time variable in all models. Time at entry was age at recruitment. Exit time was age at whichever of the following came first: colorectal cancer diagnosis, death, or the last date at which follow-up was considered complete in each centre. To control for differing follow-up procedures, questionnaire design, and other differences across centres, models were stratified by study centre. Models were also stratified by sex and age at recruitment in 1-year categories. Possible non-proportionality was assessed using an analysis of Schoenfeld residuals, [24] with no evidence of non-proportionality being detected.

Dietary intakes were modelled using either quintiles defined across cohort participants (total milk, total dairy and calcium); pre-defined categories (whole-fat, semi-skimmed, and skimmed milks: non consumers, <100, 100–199, 200–299, ≥300 g/day); and a predefined low intake reference category and quartiles defined across the remaining participants (cheese reference category  = <5 g/day; yoghurt reference category  =  non-consumers). Intakes were also modelled as continuous variables, with HR expressed per increments of: 200 g/day for milk; 100 g/day for yoghurt; 50 g/day for cheese; 400 g/day for total dairy intake, and 200 mg/day for calcium. Trend tests across intake categories were calculated by assigning the median value of each intake quintile/category and modelling as continuous terms into Cox regression models.

Analyses for colorectal, colon, proximal colon, distal colon, and rectal cancers were conducted for both sexes combined as no interactions by sex were observed for intakes of total dairy products (P = 0.26), milk (P = 0.28), cheese (P = 0.58), yoghurt (P = 0.51), and dietary calcium (P = 0.11). The results by sex are in Tables S1, S2, S3, and S4 in File S1. All models were adjusted for total energy intake, using the standard model, to obtain isocaloric risk estimates and partly control for measurement error of dairy products and calcium intake estimates. All models were additionally adjusted for: body mass index (BMI; kg/m²; continuous); physical activity (inactive, moderately inactive, moderately active, active, or missing); smoking status and intensity (never; current, 1–15 cigarettes per day; current, 16–25 cigarettes per day; current, 25+ cigarettes per day; former, quit ≤10 years; former, quit 11–20 years; former, quit 20+ years; current, pipe/cigar/occasional; current/former, missing; or unknown); education level (none/primary school completed, technical/professional school, secondary school, longer education – including university, or unknown); menopausal status (premenopausal, postmenopausal, perimenopausal/unknown menopausal status, or surgical postmenopausal); ever use of oral contraceptive (yes, no, or unknown); ever use of menopausal hormone therapy (yes, no, or unknown); and intakes of alcohol (yes or no; continuous, g/day), red and processed meats, and fibre (both continuous, g/day). Finer adjustment for body shape was attempted by also controlling for waist circumference in a subset of the cohort for which measurements were available. When included in the multivariable models, instead of, or with BMI, the risk estimates were virtually unchanged; and accordingly, we adjusted solely for BMI. In the analyses for whole-fat, semi-skimmed, and skimmed milk, the models included the covariates as detailed above, plus additional adjustment for the other milk subtypes. Similarly, the dairy and non-dairy calcium analyses were mutually adjusted for one another.

To determine whether the dietary calcium-colorectal cancer association differed according to anthropometric, lifestyle, and dietary characteristics, we included interaction terms (multiplicative scale) in separate models. The statistical significance of the cross-product terms were evaluated using the likelihood ratio test.

Cox proportional hazard restricted cubic spline models were used to explore possible deviation from a non-linear calcium-colorectal cancer relationship, with five knots specified at the median of each quintile of intake [25]. Heterogeneity of associations across anatomical cancer sub-sites was assessed by calculating χ² statistics. The heterogeneity across countries was explored by taking a meta-analytic approach [26]. To evaluate possible reverse causality, cases diagnosed within the first 2 and 5 years of follow-up were excluded from the analyses.

To improve comparability of data across study centres and to partially correct the relative risk estimates for the measurement error of dietary intakes, a linear regression calibration model was used utilizing the 24-hdr taken at baseline from a subset of the cohort (n = 34,426 in this analysis) [27], [28]. The 24-hdr were regressed on dietary questionnaire values, with adjustment for the same list of covariates detailed above, and further control for the week day and season of recall measurements. Country and sex-specific calibration models were used to obtain individual calibrated values of dietary exposure for all participants. Cox proportional hazards regression models were then applied using the calibrated values for each participant on a continuous scale. The standard error of the de-attenuated coefficients was corrected through bootstrap sampling. The P-value for the trend of the de-attenuated coefficients was calculated by dividing the de-attenuated coefficient by the bootstrap-derived standard error and approximating the standardized normal distribution. (29).

Statistical tests used in the analysis were all two-sided and a P-value of <0.05 was considered statistically significant. Analyses were conducted using SAS version 9.1 and Stata version 11.0.

Total milk and milk subtypes by fat content

Total milk was similarly inversely related to the cancer risk across all locations of the bowel (colon vs. rectal P Heterogeneity  = 0.83; distal colon vs. proximal colon P Heterogeneity  = 0.76) (Table 3). In calibrated models, colorectal cancer risk was 7% lower for each 200 g/day higher intake of total milk. Over 17% of participants reported consuming more than one milk subtype. The linear inverse associations for colorectal, colon, and rectal cancers were of similar strength for whole-fat and skimmed milk, but there were no significant associations for semi-skimmed milk (Table 4). However, in sensitivity analyses, when the models included only sole consumers of each milk subtype, identical inverse colorectal cancer risk estimates were observed for whole-fat (HR per 200 g/day 0.87, 95% CI: 0.79–0.95), semi-skimmed (HR per 200 g/day 0.87, 95% CI: 0.78–0.97) and skimmed milks (HR per 200 g/day 0.87, 95% CI: 0.76–0.99) (data not tabulated).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Multivariable hazard ratios (95% confidence intervals) of colorectal cancer risk by dairy product consumption categories.

https://doi.org/10.1371/journal.pone.0072715.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Multivariable hazard ratios (95% confidence intervals) of colorectal cancer risk by milk subtype by fat content consumption categories.

https://doi.org/10.1371/journal.pone.0072715.t004

Cheese

Cheese consumption was inversely associated with colorectal cancer in the categorical model (Table 3). The association was significant for colon (≥56 g/day vs. <5 g/day HR, 0.83, 95% CI: 0.71–0.97; P-trend  = 0.047) but not rectal cancer, although this difference was not significant (P Heterogeneity  = 0.39). In the linear calibrated models, non-significant inverse associations were observed for colorectal, colon and rectal cancers. For proximal colon cancer, the highest consumers (>56 g/day) had a 27% (95% CI: 0.58–0.93) reduced risk compared to those consuming <5 g/day, but in the calibrated model, this association was not significant. No association was observed for tumours in the distal region of the colon, and the heterogeneity in association by colonic region was not statistically significant (P Heterogeneity  = 0.82).

Yoghurt

Yoghurt intake was significantly inversely related to colorectal cancer risk in categorical models (≥109 g/day vs. non-consumers, HR 0.90, 95% CI: 0.81–0.99; P-trend  = 0.043) (Table 3). The inverse association was restricted to the colon and not observed for tumours in the rectum, although the difference was not statistically significant (P Heterogeneity  = 0.79). Within the colon the difference in association across the distal and proximal regions was non-significant (P Heterogeneity  = 0.29). No associations were observed in the linear calibrated models for cancers across all bowel locations. After adjustment for dietary calcium intake the inverse association for colorectal cancer using the categorical model was no longer significant (≥109 g/day vs. non-consumers, HR 0.94, 95% CI: 0.85–1.04; P-trend  = 0.33; data not tabulated).

Total dairy intake

Total dairy intake was significantly inversely associated to colorectal cancer risk (≥490 g/day vs. <134 g/day, HR 0.77, 95% CI: 0.70–0.86; P-trend <0.001) (Table 3). In calibrated models, each 400 g/day higher intake of total dairy products was associated with a 14% lower risk. The inverse association was of similar magnitude for colon and rectal cancer (P Heterogeneity  = .72); and within the colon, there no evidence of heterogeneity across distal and proximal regions (P Heterogeneity  = 0.66).

Dietary calcium

For dietary calcium, similar strength inverse associations were observed across all locations of the colorectum (colon vs. rectal P Heterogeneity  = 0.56; distal colon vs. proximal colon P Heterogeneity  = 1.00) (Table 5). There was no deviation from linearity for the relationship between dietary calcium and colorectal cancer in the restricted cubic spline model (P = 0.43) (data not shown). Calcium intake from dairy foods was inversely associated to cancer risk across all locations of the bowel. When calcium and milk were included in the same models, the inverse associations for milk weakened and became non-significant, but the significant inverse associations for calcium remained (data not shown). Dietary calcium from non-dairy sources was not inversely associated with colorectal cancer risk. The association between dietary calcium intake and risk of colorectal cancer did not differ by BMI (P = 0.56), waist circumference (men P = 0.74; women P = 0.64), physical activity (P = 0.26), smoking status (P = 0.37 alcohol consumption (P = 0.75), and intakes of red and processed meat (P = 0.50), and fibre (P = 0.65) (data not tabulated).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Multivariable hazard ratios (95% confidence intervals) of colorectal cancer risk by dietary calcium intake categories.

https://doi.org/10.1371/journal.pone.0072715.t005

Between country heterogeneity and inclusion of preclinical disease

There was evidence of significant heterogeneity by country for total dairy products (P = 0.034) (Figure S1 in File S1); although risk estimates ≤1 were observed in most countries. No associations were observed in the Sweden and Denmark cohorts. Non-significant between country heterogeneity was observed for intakes of dietary calcium (P = 0.60; Figure S2 in File S1), total milk (P = 0.13), cheese (P = 0.64), and yoghurt (P = 0.12).

Excluding the participants with less than 2 and 5 years of follow-up (including 502 and 1,483 colorectal cancer cases respectively) from the total dairy, total milk, cheese, yoghurt, and calcium intake analyses resulted in negligible differences in the colorectal cancer associations (data not shown).

Ethical approval review board information for local EPIC centres

The National Committee on Health Research Ethics (Denmark); Comité de Protection des Personnes (France); Ethics Committee of the Heidelberg University Medical School (Heidelberg, Germany); Ethikkommission der Landesärztekammer Brandenburg Cottbus (Potsdam, Germany); University of Athens Medical School (Greece) Comitato Etico Indipendente, Fondazione IRCCS Istituto Nazionale dei Tumori, Milano (Italy); Human Genetics Foundation Torino: Ethics Committee (Turn, Italy); The Medical Ethical Committee (METC =  Medisch Ethische Toetsingscommissie) of the University Medical Center Utrecht (UMCU), Utrecht, the Netherlands (The Netherlands); Regional ethical committee for Northern Norway and the Norwegian Data Inspectorate (Norway); CEIC Comité de Ética de Investigación Clínica (Spain); Ethics Committee of Lundst University (Malmö, Sweden); Umea Regional Ethical Review Board (Umea, Sweden); Norwich District Ethics Committee (Cambridge, UK); Scotland A Research Ethics Committee (Oxford); and the Imperial College Research Ethics Committee [ICREC] (UK).