Population
EPIC is a large prospective cohort study conducted in 23 centers in 10 European countries [France, Italy (Florence, Varese, Ragusa, Turin, Naples), Spain (Asturias, Granada, Murcia, Navarra, San Sebastian), The Netherlands (Bilthoven, Utrecht), United Kingdom (UK; Cambridge, Oxford), Greece, Germany (Heidelberg, Potsdam), Sweden (Malmö, Umea), Norway, and Denmark (Aarhus, Copenhagen)]. In most centers, the participants were recruited from the general population. However, the French cohort comprises female members of a health insurance program for school and university employees. Spanish and Italian participants were recruited among blood donors, members of several health insurance programs, employees of several enterprises, civil servants, but also the general population. In Utrecht and Florence, participants in mammographic screening programs were recruited for the study. In Oxford, half of the cohort consisted of 'health conscious' subjects from England, Wales, Scotland, and Northern Ireland. The cohorts of France, Norway, Utrecht, and Naples include women only [18]. Participants were recruited between 1992 and 2000 depending on the study center. At recruitment, men were 40 to 70 and women 35 to 70 years old [18]. All participants gave written informed consent to use their questionnaire data and the Internal Review Boards of the International Agency for Research on Cancer and all EPIC recruitment centers approved the analyses based on EPIC participants.
Of 511,781 apparently healthy participants at baseline, we excluded individuals with a ratio for energy intake versus energy expenditure in the top or bottom 1% (n = 10,197) and those with self-reported cancer, stroke or myocardial infarction at baseline (n = 29,300). We further excluded participants with unknown smoking status at baseline (n = 23,716). The analytical cohort included 448,568 participants.
Exposure assessment
Following the results of several methodological studies conducted in the early 1990s, habitual diet over the previous twelve months was measured at recruitment by country-specific instruments designed to capture local dietary habits and to provide high compliance [18]. Seven countries adopted an extensive self-administered dietary questionnaire, which can provide data on up to 300 to 350 food items per country. In Greece, Spain and Ragusa, the dietary questionnaire was very similar in content to the above, but was administered by direct interview. A food frequency questionnaire (FFQ) and a seven-day food record were adopted in the UK. In Malmö, Sweden, a quantitative questionnaire combined with a seven-day menu book and an interview was used. Baseline food consumption, as well as ethanol and energy intake, was calculated from the dietary instruments applied in each center.
For this analysis, meats were grouped into red meat (beef, pork, mutton/lamb, horse, goat), processed meat (all meat products, including ham, bacon, sausages; small part of minced meat that has been bought as a ready-to-eat product) and white meat (poultry, including chicken, hen, turkey, duck, goose, unclassified poultry, and rabbit (domestic)). Processed meat mainly refers to processed red meat but may contain small amounts of processed white meat as well, for example, in sausages.
A set of core questions posed at recruitment that was similar in all participating centers ensured comparability of non-dietary questions and assessed information on education, medical history (including history of stroke, myocardial infarction, and cancer), alcohol consumption, physical activity, lifetime history of consumption of tobacco products including smoking status (current, past, or never smoker), type of tobacco (cigarettes, cigars, or pipe), number of cigarettes currently smoked, and age when participants started and, if applicable, quit smoking [18]. Height and weight were measured in all EPIC centers except for France, Norway, and Oxford, for which self-reported height and weight was recorded. In Oxford, self-reports were improved by using prediction equations [19].
Outcome assessment
Information on vital status and the cause and date of death were ascertained using record linkages with cancer registries, Boards of Health, and death indices (in Denmark, Italy, the Netherlands, Norway, Spain, Sweden, and the UK) or active follow-up (in Germany, Greece, and France). Active follow-up included inquiries by mail or telephone to participants, municipal registries, regional health departments, physicians, and hospitals. Participants were censored as follows: June, 2005 (Cambridge), December 2006 (France, Varese, Turin, Naples, Granada, Murcia, Malmo, and Denmark), December 2007 (Florence, San Sebastian, Umeå and Norway), December 2008 (Ragusa, Asturias, Navarra, and the Netherlands); June 2009 (Oxford). For Germany and Greece, the end of the follow up was considered to be the last known contact or date of death, whichever came first. Cause of death was coded according to the 10th Revision of the International Classification of Diseases (ICD-10). The underlying causes of death were used to estimate the cause-specific mortality: cancer (ICD-10: C00 to D48), cardiovascular diseases (I00 to I99), respiratory diseases (J30 to J98), digestive diseases (K20 to K92), and other diseases. Currently, vital status is known for 98.4% of all EPIC subjects.
Statistical analysis
Cox proportional hazards regression was used to examine the association of meat consumption with all-cause and cause-specific mortality. To explore the shape of the risk function, we fitted a Cox proportional hazards model with restricted cubic splines for red and processed meat and poultry intake treated as continuous variables [20, 21]. We specified four knot positions at 10, 20, 40, and 80 g per day of red or processed meat intake. Other knot positions were specified but did not appreciably change the curves. After examining the shape of the association between red and processed meat intake with mortality in restricted cubic spline models, we decided to choose the second category as the reference category in the categorical model (see below) for all three types of meat, that is, also for poultry for consistency reasons.
In a second step, we modeled meat intake as categorical variables as follows: red and processed meat 0 to 9.9, 10 to 19.9, 20 to 39.9, 40 t0 79.9, 80 to 159.9, and ≥160 g/day; poultry 0 to 4.9, 5 to 9.9, 10 to 19.9, 20 to 39.9, 40 to 79.9, and ≥80 g/day. Age was used as the primary time variable in the Cox models. Time at entry was age at recruitment, exit time was age when participants died, were lost to follow-up, or were censored at the end of the follow-up period, whichever came first. The analyses were stratified by sex, center, and age at recruitment in one-year categories. To adjust for lifelong tobacco smoking, we included baseline smoking status and intensity of smoking as one variable (never smokers (reference category); current cigarette smokers (three categories: 1 to 14, 15 to 24 and 25+ cigarettes/day); former smokers who stopped less than 10 years ago, 11 to 20 years ago, 20+ years ago; other smokers (one category including pipe or cigar smokers and occasional smokers)). In addition, duration of smoking in 10-year categories (≤10 (reference category), 11 to 20, 21 to 30, 31 to 40, 41 to 50, >50 years) is added as a second variable in the statistical models. We separately adjusted for the amount of smoking and the duration of smoking instead of using pack-years of smoking to differentiate better between, for example, heavy smokers of a short duration and light smokers for a long duration [22]. Additionally, all analyses were adjusted for body weight and height, energy intake, intake of alcohol (all continuous), physical activity index (active, moderately active, moderately inactive, inactive, missing) [23], and education (none or primary school completed; technical/professional school; secondary school; university degree; missing). We additionally examined the effect of mutually adjusting intake of the three types of meat for each other. We also explored meat intake in models without adjusting for total energy intake. Additionally adjusting for fruit and vegetable consumption did not appreciably change the observed associations and was not included in the main models.
In order to improve the comparability of dietary data across the participating centers, dietary intakes from the questionnaires were calibrated using a standardized 24-hour dietary recall [24, 25], thus, partly correcting for over- and underestimation of dietary intakes [26]. A 24-hour dietary recall was collected from an 8% random sample of each center's participants. Dietary intakes were calibrated using a fixed effects linear model in which gender- and center-specific 24-hour dietary recall data were regressed on the questionnaire data controlling for weight, height, age, day of the week, and season of the year. The confidence intervals (CIs) of the risk estimates, obtained using calibrated data, were estimated using bootstrap sampling to take into account the uncertainty related to measurement error correction. Calibrated and uncalibrated data were used to estimate the association of meat consumption with mortality on a continuous scale.
Results of the 24-hour recalls (mean, standard error of the mean) were also used to describe the FFQ-based intake categories of red meat, processed meat, and poultry.
In our analysis, we considered cause-specific mortality in addition to overall mortality. Therefore, we fitted a competing risk model [27] which, however, resulted in similar associations as those observed in non-competing risk models for deaths from cancer, cardiovascular diseases, respiratory diseases, digestive diseases, and other diseases, and are not shown in the tables.
Results might differ between subgroups of the study population due to different health behaviors in, for example, men and women, or interactions between nutrients in different foods. Therefore, sub-analyses were performed by sex and smoking status (never, former, current), alcohol consumption (dichotomized by sex-specific median), and fruit and vegetable consumption (dichotomized by sex-specific median). Including cross-product terms along with the main effect terms in the Cox regression model tested for interaction on the multiplicative scale. The statistical significance of the cross-product terms was evaluated using the likelihood ratio test. Heterogeneity between countries was assessed using likelihood chi-square tests. We also examined whether the associations differed in the first two years and the succeeding years of follow-up.
The population attributable risk (PAR), which describes the proportion of cases that would be prevented if everyone in the study population had the reference level of the exposure, was estimated based on the formula [28]:
where HRi and Pi are the multivariate adjusted relative risks and the prevalence, respectively, in the study population for the ith exposure category (processed meat consumption 20+ g/day); I = 0: reference group (processed meat 0 to 19 g/day).
All analyses were conducted using SAS version 9.1 (SAS Institute, Cary, North Carolina).
Comments
View archived comments (3)