The present study was based on an ongoing cohort study in Heidelberg, Germany – a part of the European Prospective Investigation into Cancer and Nutrition (EPIC) . A total of 11,928 men and 13,612 women, mostly 40-year-old or older, were recruited into the Heidelberg cohort from 1994 to 1998. A detailed description of the recruitment procedures has been published elsewhere . The study protocol was approved by the local ethics committee and all cohort participants provided informed consent. In this study, we excluded participants who had pre-existing diabetes, cardiovascular disease or cancer at baseline (n = 3,064). We further excluded 7 participants without complete data, and eventually we had 22,469 participants (10,235 men and 12,234 women) for statistical analyses.
Assessment of lifestyle risk factors
Detailed information on past and current smoking was collected at baseline with a computer-guided interview . According to this information, we categorized smoking status at baseline into five groups: never smokers (reference), long-term quitters (duration of cessation >10 years), short-term quitters (duration of cessation ≤10 years), current light smokers (≤10 cigarettes/day), and current heavy smokers (>10 cigarettes/day).
Data on daily consumption of alcoholic beverages at age 20, 30, 40, 50 and over the 12 months before recruitment were collected by means of a self-administered questionnaire. Alcohol intake was derived using the German Food Code and Nutrient Data Base BLS II.3 (BgVV, Berlin, Germany). Average lifetime alcohol intake was estimated as a weighted average of intakes at different ages, with weights equal to the times of exposure to alcohol at different ages. We converted the average lifetime alcohol intake into standard drinks by assuming that one standard drink contains 12 grams of alcohol. For men, we categorized their daily alcohol intake into three groups: ≤2 (reference), 2.1 to 4, and >4 drinks. For women, the three groups were ≤0.5 (reference), 0.6 to 1, and >1 drink.
Anthropometric data were collected during a baseline physical examination. We used the body mass index (BMI, body weight in kilograms divided by the square of height in meters) to categorize baseline body weight status into four groups: optimal BMI (22.5 to 24.9 kg/m2, reference), low BMI (<22.5), overweight (BMI 25 to 29.9), and obese (BMI ≥30). The choice of BMI 22.5 to 24.9 as the optimal group was made based on the finding from large cohort data showing the lowest mortality rate for this group .
Occupational, household and recreational activity was assessed at baseline using a short questionnaire . The intensity of physical activity was measured with metabolic equivalent (MET) values. A MET is defined as the ratio of work metabolic rate to a standard metabolic rate of 1.0 (4.184 kJ) kg−1.h−1. For each reported activity, a MET value was assigned according to the Compendium of Physical Activity . A large meta-analysis of cohort studies has shown a positive association between leisure time physical activity and life expectancy . Therefore, we were particularly interested in leisure time physical activity, which in the present study was a combination of walking, cycling and other sports. We chose the median of leisure time activity in MET-hours/week to categorize participants into the low group (<36 MET-hours/week) and the high group (≥36 MET-hours/week, reference).
Data on usual consumption levels of 148 food items over the last 12 months before recruitment were collected using a food frequency questionnaire (FFQ) validated by 24-hr dietary recalls. The validity and reproducibility of the FFQ were acceptable, with the correlation coefficients varying from 0.21 for fish to 0.90 for alcoholic beverages . In the present study, we focused on potentially important food groups, namely, processed/red meat, vegetables/fruits, cereals, fish, and dairy products, which, according to previous studies [30–34], might have an association with mortality rates. We dichotomized the consumption of these food groups, except processed/red meat, as high and low by using their medians as the cut-off points. For processed/red meat, we chose 120 g/day as the cut-off point, as it has been suggested that a higher consumption than this amount may significantly increase the mortality rate . Besides these selected food items, a more exploratory analysis covering a total of 16 finer food groups and other dietary factors showed no additional associations with the risk of premature death [see Additional file 1: Table S1].
Ascertainment of deaths
Information on vital status was collected through the official death registry system and reports from next of kin. All reported deaths were verified by obtaining official death certificates. In the present study, we analyzed the mortality data until December 31, 2009, by which time the vital status of all cohort participants had been completely ascertained.
Mortality data were analyzed with two, sex-specific multivariable Gompertz PH models that included smoking status, body weight status, alcohol drinking, consumption of the selected food groups (processed/red meat, vegetables/fruits, cereals, fish and dairy products) and leisure time physical activity. Education and self-reported pre-existing hypertension and hyperlipidemia were additionally included as confounders.
For each of the lifestyle risk factors, we detected no departure from the proportionality assumption by checking the parallelity of the log-log survival plots. Age was modeled as the time scale. As our aim was to estimate the RLE at age 40, we left-truncated the data at this age. Unlike the widely used Cox PH model that leaves the baseline hazard function unspecified, Gompertz PH model is parametric and assumes that the baseline hazard is an exponential function of age.
We predicted lifetime survival probabilities with the multivariable Gompertz PH models given specific baseline status of the lifestyle risk factors. The maximum age attained by cohort participants at the end of the follow-up was 82 years; however, we extrapolated the prediction up to the age of 110 years, which was assumed to be the maximum age that any of our cohort participants could theoretically attain. As the multivariable Gompertz PH models showed no statistically significant association with mortality rate for cereals, fish and dairy products in both sexes, we fixed these factors at the assumedly healthy level (high consumption of cereals and fish but low consumption of dairy products) when we calculated the lifetime survival probabilities. In addition, we fixed the educational degree at the intermediate level (‘secondary/professional’) and pre-existing hypertension and hyperlipidemia at “no”. Assigning a fixed value to these factors, however, did not change the estimated losses of RLE in relation to the other lifestyle risk factors.
We applied the life table method to convert the predicted lifetime survival probabilities into RLEs. In brief, we used the predicted lifetime survival probability on a hypothetical cohort of 100,000 40-year-old subjects to calculate the expected number of deaths (d
) that would occur within each age interval [t, t+1). The number of person-years of survival within [t, t+1), denoted as L
, given the number of subjects who remained alive at age t, denoted as l
, was estimated as follows:
The total person-years of survival remaining at age 40 (T
40) was calculated by summing up L
from the last age interval, [100, 110), backward to the age interval [40, 41). The RLE at age 40 (RLE
40) was then calculated as:
The loss of RLE associated with a lifestyle risk factor was calculated as the difference between the RLEs given the absence and the presence of this factor, respectively, while keeping other factors identical. We estimated 95% confidence intervals for the losses of RLE using the bootstrap method.
As low body weight is likely to be a result of undiagnosed diseases, we performed a sensitivity analysis to examine the possible reverse causality by excluding deaths that occurred within the first two years of the follow-up period. Since previous research shows a stronger obesity-mortality association in never smokers than in current smokers [35, 36], we also examined the effect of body weight on RLE across the smoking categories. To ensure sufficient cases in each category, we combined all the quitters as former smokers and light and heavy smokers as current smokers. We did not consider possible effect modification for other pairwise risk factor combinations because of lack of both biological and statistical evidence (the minimum P-value for interaction was 0.02, which was not regarded as significant evidence from multiple comparisons).
For the male and female sub-cohorts, we predicted their lifetime survival probabilities using a Gompertz PH model that did not include any covariates, and then compared the predicted survival probabilities with the observed ones, which were estimated with the life table method. We also compared the trajectories of the extrapolated survival curves of our cohort with those of the general German population. The latter were derived from the German life table .
All statistical analyses were performed with SAS (version 9.2, SAS Institute, Cary, NC, USA) and the package ‘eha’ in R (version 3.0.1, R Foundation for Statistical Computing, Vienna, Austria). Two-sided P-values <0.05 were considered statistically significant.