Study population, CRC ascertainment, and sample selection

CRC cases in this study were identified among participants from the EPIC cohort, a large prospective study with over 520,000 participants enrolled in 23 centres in Denmark, France, Germany, Greece, Italy, the Netherlands, Norway, Spain, Sweden, and the United Kingdom between 1992 and 1999. Approval for this study was obtained from the ethical review boards of the International Agency for Research on Cancer and from all local institutions where subjects had been recruited for the EPIC study. Written informed consent was obtained from all participants before joining the EPIC study. The EPIC study methods have been described in detail elsewhere [14,15].

Cancer incidence during follow-up was determined through record linkage with regional cancer registries (Denmark, Italian centres except Naples, the Netherlands, Norway, Spain, Sweden, and the United Kingdom; complete up to December 2006) or via a combination of methods, including the use of health insurance records, contacts with cancer and pathology registries, and active follow-up through study subjects and their next-of-kin (France, Germany, the Italian center of Naples, and Greece; complete up to June 2010). CRC cases were selected among participants who developed colon (C18.0-C18.7, according to the Tenth revision of the International Classification of Diseases, Injuries and Causes of Death (ICD-10)), rectum (C19-C20), and overlapping/unspecified origin tumors (C18.8 and C18.9).

A total of 4,701 CRC cases were identified. Case exclusions included 426 cases diagnosed with CRC after vital status censoring date, 172 cases with in situ or non-primary tumors in the colon, 144 non-adenocarcinoma or tumor of unknown morphology, 21 due to missing date of death or diagnosis, 74 within the extreme ranking (top and bottom 1%) of the ratio energy intake/energy requirement, 37 with missing data on anthropometry, 56 with missing data on diet, 338 with missing data on physical activity (including all participants from Norway), and 141 women with missing data on breastfeeding (including all women from Bilthoven). The final sample included 3,292 CRC cases (1,497 men and 1,795 women; 2,071 colon cancer cases and 1,221 rectal cancer cases).

Exposure assessment: data collection and dietary questionnaires

At recruitment (between 1992 and 1999), before cancer diagnosis, volunteers participating in EPIC filled out medical, dietary, and lifestyle questionnaires, including questions on alcohol use, smoking status, physical activity, education, reproductive history, breastfeeding, exogenous hormones use, and previous illnesses. Body weight and height were measured in all centres except Oxford (health conscious population) and France, where anthropometry was self-reported [16]. Usual food intakes were measured using country-specific validated dietary questionnaires [17] and individual nutrient intakes were derived from foods included in the dietary questionnaires through the standardized EPIC Nutrient Data Base [18]. To correct for any systematic under- or over-estimation of dietary intake between the study centres, a dietary calibration study was conducted. A random sample of 36,308 men and women (7.4% of the sample) completed a detailed computerized 24-h dietary recall, and nutrient intake was calculated using the standardized EPIC Nutrient Data Base [19]. Dietary exposures across centres were scaled using an additive calibration model [17]. Briefly, the difference between the sex- and center-specific mean of the values from the dietary questionnaire and the mean of the 24-h recall values was calculated and added to the questionnaire values. All dietary variables used in the present study were calibrated using additive calibration.

WCRF/AICR score construction

A WCRF/AICR score, incorporating six of the WCRF/AICR recommendations for men (regarding body fatness, physical activity, food and drinks that promote weight gain, plant foods, animal foods, and alcoholic drinks) and seven for women (plus breastfeeding) was constructed. Detailed information on the operationalization of the score was previously published [12] and can be found in Additional file 1: Table S1. Briefly, we assigned, for each component, 1 point when the recommendation was met, 0.5 points when it was half met, and 0 points otherwise. When available, the quantitative criteria provided in the recommendations were used as cut-off points and intermediate cut-off points, defined by the authors, were used otherwise. For the recommendations, including several sub-recommendations (foods and drinks that promote weight gain or plant foods), the final score was the average of each sub-recommendation score (meaning that for these recommendations, plausible scores were 0, 0.25, 0.5, 0.75, and 1). Three recommendations were not implemented: i) the recommendation on preservation, processing, and preparation of foods because of insufficient data available, ii) the recommendation on dietary supplements which could not be operationalized in terms of cancer prevention without further assumptions about type or dose of supplementation, and iii) the special recommendation related to cancer survivors, who were advised to follow the same recommendations for cancer prevention. As the WCRF/AICR recommendations were not ranked according to priority, all major recommendations were summed to contribute equally to the total WCRF/AICR score. Therefore, the total WCRF/AICR score ranged from 0 to 6 for men and from 0 to 7 for women, with higher scores indicating greater concordance with the WCRF/AICR recommendations. The score was further categorized into four categories according to pre-defined cut-off points (0–2, 2.25–2.75, 3–3.75, and 4–6 points in men and 0–3, 3.25–3.75, 4–4.75, and 5–7 points in women).

Outcome assessment: vital status ascertainment

Vital status follow-up was conducted by record linkage with regional and/or national mortality registries in all countries except France, Germany, Greece, and the Italian center of Naples, where data are collected through an active follow-up. Censoring dates for complete follow-up were between June 2005 and June 2009 in Denmark, the Netherlands, Spain, the United Kingdom, Sweden, and Italian centres except Naples. In Germany, Greece, France, and Naples, follow-up was based on a combination of methods, including health insurance records, cancer and pathology registries, and active follow-up through study subjects and their next-of-kin. In these centres, the end of follow-up was defined as the last known date of contact or the date of death, whichever came first. The last update of endpoint information occurred between December 2007 and December 2009.

Mortality data were coded according to the ICD-10. Up to six qualifiers of the cause of death were reviewed. The outcome of interest for the present study (death from CRC) was assigned based on the underlying cause of death [20].

Statistical analyses

Cox proportional hazards regression was used to estimate the association between the WCRF/AICR score and death from CRC (primary endpoint) or death from any cause (secondary endpoint). Age was used as the primary time variable, with entry time defined as the subject’s age at CRC diagnosis and exit time as age censoring or death. All analyses were stratified by country to control for country-specific effects such as follow-up procedures and questionnaire design. The WCRF/AICR score was assessed as a continuous variable (1-point increment) and as a categorical variable (using the four pre-defined categories). The WCRF/AICR categorical variable was scored from 1 to 4, and trend tests were calculated on these scores. Multivariable models were adjusted for sex, year of CRC diagnosis, educational level (coded as no education, primary school, technical school, secondary school, university degree, and unknown/missing), smoking status (never, former, smoker, missing), tumor site (colon, rectum), tumor grade (well differentiated, moderately differentiated, poorly/undifferentiated, missing), and tumor stage (I, II, III, IV, missing); based on a previously described harmonization procedure among different EPIC centres [20]. Models were further adjusted for ‘lag time’ (years between recruitment/exposure assessment and CRC diagnosis), with no change in results; therefore, this variable was not included in the final multivariable models.

Sensitivity analyses were performed excluding participants who died within 6 months of CRC diagnosis and excluding participants with incomplete CRC stage data. Potential effect modifications by sex, mean age at diagnosis (<64.6 years vs. ≥64.6 years), year of diagnosis (1992–2001 vs. 2002–2008), median ‘lag time’ (<6.5 years vs. ≥6.5 years), tumor stage (I, II, III, IV), tumor site (colon vs. rectum), colon tumor sub-site (proximal vs. distal), and smoking status (former smokers, current smokers, and never smokers) were explored by modelling interaction terms (cross-products) between these variables (categorically) and the WCRF/AICR score (continuous), and conducting stratified analyses. Potential heterogeneity between countries in the association between the WCRF/AICR score and CRC-related mortality and overall mortality was assessed by calculating country-specific estimates and using random-effect meta-analyses (I²).

We estimated the independent association of each component of the WCRF/AICR score with CRC-related and overall mortality, after adjusting for all other components of the score. Finally, we evaluated the relative importance of each of the components of the WCRF/AICR score on CRC-related and overall mortality by subtracting alternately one component at a time from the original score, and including this component as a covariate in the model. To be able to compare the risk estimates to that of the total WCRF/AICR score, these alternative scores were assessed as continuous variables per 1 SD increments.

All statistical analyses were conducted using STATA 11 (StataCorp).