Modifiable causes of premature death in middle-age in Western Europe: results from the EPIC cohort study
Our analysis of the primary causes of premature death among more than 250,000 European adults indicates that the four major risk factors are tobacco smoking, poor diet, obesity, and high blood pressure, which together account for over 50 % of premature deaths. Two other risk factors, physical inactivity and excessive alcohol consumption, have AFs of 7 % and 4 % of premature deaths, respectively. Our study also provides preliminary evidence for an important AF for high cholesterol levels, although the sample size was limited.
The GBD has derived similar estimates for the role of each exposure using an alternative modelling approach, relying largely on published estimates of effect for each exposure and estimates of exposure prevalence in each population [1]. Their estimates of AFs for premature mortality for the European population provide broadly comparable estimates for smoking, poor diet and high blood pressure (Table 4). However, the GBD estimates are approximately twice as high for excessive alcohol consumption (8 % vs. 4 %), and substantially higher for excessive body mass (14 % vs. 3 % based on BMI). Using WHR as a measure of obesity made the estimates more comparable (10 % based on WHR).
Comparison of population attributable fractions (%, 95 % CI) from the Global Burden of Diseases analysis with those from the present EPIC analysis
Estimates from the GBD are taken from the website http://vizhub.healthdata.org/gbd-compare. They are the estimated attributable fractions for death in Western Europe for the age range 50–69 years for each risk factor
There are three possible explanations for these differences. Firstly, the estimates of relative risk used in the calculations might differ – indeed, we estimated modest relative risks for overweight and obesity and physical inactivity. Secondly, the distribution of the risk factors used for the GBD computations might differ from the distribution in EPIC which, for example, includes relatively few very heavy consumers of alcohol or very obese participants. This is a well-known phenomenon in prospective cohort studies, also called “healthy volunteer” effect. Finally, the reference or counterfactual distributions used for the AF calculations might differ. For instance, the GBD used a “theoretical minimum-risk exposure distribution”. On the other hand, we have chosen to not necessarily use a theoretically “optimal” exposure distribution in all cases. For instance, using the lowest risk category for alcohol intake or BMI would involve an increase in alcohol intake or BMI in a proportion of the participants. Instead, we have focused on the AF for high alcohol intake and overweight and obesity per se. Low reported alcohol intake in particular is associated with substantially higher risk of death in EPIC, possibly due to the influence of former drinkers who quit for health reasons or misclassification of heavy drinkers [23, 27, 39]. This misclassification would lead to an underestimation of the AF for alcohol in the EPIC study, and – assuming that the GBD estimates do not suffer from the same problems – may explain the difference between the EPIC and GBD estimates in this case.
On an individual level, the estimated conditional survival curves suggest smoking could have a similar effect on survival to age 70 to that of all other factors combined. Men who were smokers but possessed otherwise healthy characteristics had expected survival of 86 %, similar to the 83 % expected survival for men with unhealthy characteristics but who never smoked. For women, both smokers with otherwise healthy characteristics and unhealthy non-smokers also had similar expected survival (92 % and 90 %, respectively). These estimates reinforce the critical importance of smoking in terms of preventing premature death, and suggest it is as important as the other five major risk factors combined.
A strength of our approach is the ability to estimate AFs for important sub-groups, such as smokers and never smokers, and also the direct estimation of AFs for combinations of risk allowing for their interdependence. Interestingly, our results indicate that, for both current and never smokers, an equivalent proportion of premature deaths can be attributed to poor diet, hypertension, overweight and obesity, and physically inactivity.
The principal limitation of our study is the relatively small number of participants with available cholesterol and HbA1c measurements, leading to imprecise estimates of relative risks, prevalence, and AF for these factors. Further, given that all exposures were assessed only once at recruitment to the study, we could not assess the potential effects of changing exposures over time, such as quitting smoking, gaining or losing weight, or increasing physical activity. As such, our estimated AFs and survival functions cannot be interpreted as the expected effects on mortality if individuals were to change their lifestyle or diet, but rather reflect comparisons of individuals with a given, constant pattern of exposures, or hypothetical scenarios in which no-one in the population is exposed to a given risk factor. Similarly, with only one assessment of exposure we cannot assess the potential effects of measurement error, which are unlikely to be equal across the six factors (e.g. BMI and blood pressure are subject to only modest measurement error, especially compared with self-reported diet and physical activity), and can lead to under- or overestimates of the risk associated with specific risk factors. Finally, the participants in EPIC are not representative of the general European population and may have a different distribution of risk factors than other target populations. In addition to the participants not being a representative sample of any population at baseline, during the follow-up period, the participants have aged, so estimates presented here do not strictly represent the age group 40 to 70.
We choose to include individuals who reported a prevalent chronic disease condition at baseline (e.g. diabetes, heart disease, angina, or a previous diagnosis of cancer), as such a sample would be more representative of the underlying population. Excluding the 7 % of the cohort who did report a prevalent chronic disease at baseline did not have any substantial effect on our results. As an additional sensitivity analysis to evaluate the possibility of sick individuals changing their lifestyle habits relatively recently prior to interview, we also excluded the first 4 years of follow-up. Again, this exclusion did not substantially affect the results.