To achieve the most accurate prediction of future obesity prevalence is a significant public health concern, especially for health services and health economic reasons in the developed countries. Previously, most researchers based their predictions on the assumption that the trend is linear [7], [8], [9]. The assumption of a linear trend of increase of the prevalence in obesity over a longer time period in the future, however, could result in unrealistic scenarios.

Moreover, most studies on obesity trends are based on cross-sectional survey data [2], [10], [11], [12]. Cross-sectional data, however, has the disadvantage that they only display the status at a certain moment, but do not show the dynamics with respect to age, period and birth cohort. Only data from longitudinal studies can address such issues. Despite this, longitudinal data on obesity trends are rarely found in the literature compared to cross-sectional data.

The data and conclusions about observed and projected obesity prevalences have been derived from a longitudinal study. Conversely, publications dealing with the development of the obesity epidemic over time are mainly based on cross-sectional data [2], [10], [11], [12]. However, cross-sectional data reflect only the situation at a specific time across the age groups. For example, they often suggest that obesity rates peak between the ages of 50 to 60 years in most developed and developing countries. Thereafter, a drop of the rates is observed [2]. However, this drop in the prevalence does not necessarily mean that elderly subjects loose weight. Instead, it could be the result of birth cohort effects and reflect the fact that older birth cohorts had not gained similar weight with age as birth cohorts being born in a later time period. To overcome this problem of birth cohort effects making cross-sectional data based on one-point-in-time estimates less interpretable, longitudinal data are necessary: e.g. multiple sequential cross-sectional surveys with several birth cohorts and several specific ages or preferably prospective cohort studies. Despite this, longitudinal data on obesity trends are rarely found in the literature compared to cross-sectional data. Grinker et al. for example found that obesity was strongly related to age during a follow-up period of 15 years. The largest increase was observed among the 30-44 year old subjects, whereas subjects aged 60 years and older at baseline remained weight stable [19]. In our study population we also observed a relatively strong weight increase in young age groups. However, in contrast to Grinker et al., we could observe an increase in weight also in the age group of 65–75 years between the baseline and follow-up examination.

Moreover, literature about future prediction of obesity is rare and research groups that have investigated future trends in obesity based their projections on repeated survey data [7], [8], [9] [20]. These data is based on a random sample of different individuals at each time period; in contrast to our study that observed individuals of a wide age range over years. However, in contrast to repeated independent surveys, a prospective cohort is ageing during the observation time. To consider the impact of age on future trends in obesity prevalences, age at recruitment was included as a covariate into the regression models. Furthermore, we stratified by baseline age group to evaluate trends in obesity among different age groups.

In the following we compared our study results with attempts of other researchers to predict obesity prevalences in the future based on repeated cross-sectional surveys. Kelly et al. projected the global burden of obesity all over the world in 2030 [7]. Overall, 9.8% of the world's adult population was obese in 2005 (7.7% in men and 11.9% in women). Based on regional secular trends in obesity prevalence estimated from past data, population growth and demographic shifts, they predicted an increase of obesity prevalence of up to 19.7% in 2030. However, obesity prevalences are differing strongly between world regions. For the established market economies which comprise developed and industrialized high income countries the estimated prevalence of obesity in 2030 amounts to 36%. Wang et al. focused on the projection of future trends in the United States based on repeated nationwide survey data (National Health and Nutrition Examination Study, NHANES) collected between 1970 and 2004 [8]. They reported that 32.3% of the American adults were obese in 2004 and that this proportion will increase up to 51.1% in 2030 based on the assumption that trends will continue linearly. The projected prevalence amounts to 37.4% in 2010 and 44.2% in 2020. These prognoses reflect a worse situation than in Europe which is due to considerably higher initial obesity prevalence.

To address the limitations of linear predictions based on cross-sectional data, Wang et al. conducted a further analysis using fitted regression models stratified by sex and race based on the assumption that the cohort is aging 10 years to project shifts in BMI distribution of the 1999-2000 NHANES population to 2010 [20]. These projections were also validated with data from two longitudinal studies (Nurses' Health Study and Health Professional Follow-Up Study) making this approach comparable to our analyses. The expected prevalence of obesity for 2010 was 35% and 36% for white men and women, respectively. Zaninotto et al. investigated different scenarios to project the prevalence of adult obesity in England from 1993-2004 to 2012 based on repeated cross-sectional surveys [9]. Based on the linear trend, the prevalence of obesity among 35–74 year old individuals increased from 15.4–17.8% (men) or 17.9-22.8% (women) in 1993 to 35.1–36.4% for males and 30.6–35.9% for females in 2012. In contrast, the projection based on the best fitting curve allowing to consider a slowing down in the obesity trend lead to prevalences of 28.0–29.8% in men and 25.8–32.0% in women in this age group. As in our analysis, the non-linear modeling resulted in considerably lower estimates than the linear prediction which is not surprising. Similar to us Zaninotto et al. investigated alternatives to the linear model. We also decided to investigate a linear and non-linear scenario to project the situation in the near future, for the year 2015. We could show that differences in linear and non-linear projections could be huge if the observation period was short and the projection period long. Compared with Zaninotto et al., our projections based on the leveling off model appeared to be moderate. However, our prevalence data do not represent a specific region or country and are not representative for the total population.

Overall, the non-linear (leveling off) approach of prediction seems to be more realistic, whereas the linear scenario is more likely to result in an over-estimation of future obesity prevalences. It is likely that age-associated diseases will lead to a leveling off in future weight gain and thus, obesity trends might slow down. Moreover, as BMI reaches a high level, the risk for obesity-related diseases such as diabetes mellitus type 2 or metabolic syndrome increases considerably. Thus, severe obese subjects are forced to control their weight because of medical reasons. Therefore, one cannot assume that people gain weight in a linear way during their whole life. There is also some evidence in the literature based on repeated survey data that the trend in obesity prevalence, which was increasing nearly linearly in the past decades, is now showing a leveling off [21], [22], [23]. This fact in turn adds further support to the leveling off approach of prediction.

Besides investigating different scenarios for future projection a further strength of our study is the large sample size covering cohorts from five European countries and a total of 97,942 participants. Some differences in methodologies used to collect data on body weight might have affected the results. Therefore, we corrected body weight for clothing differences and self-reporting using the equations proposed by Haftenberger et al [18].

We took BMI as a measure of obesity, which can be discussed because it does not take body composition, i.e. fat and lean body mass, into account. However, several studies have shown that even without body weight changes, the amount of body fat significantly increases with age. That is why a BMI ≥ 30 kg/m² is associated with an excess of body fat in any case, thus being an acceptable measure of obesity also in the elderly. Another measure of total body fat is waist circumference, which has been shown to be strongly related to visceral fat with a cut-off for adults of 102 cm in men and 88 cm in women to indicate abdominal obesity. Using these cut-offs led to a great proportion of elderly persons (60+) having high risk values of waist circumference [24]. However, these waist circumference cut-off points still need to be validated as predictors of morbidity and mortality in older ages [25].

One weakness of our study is that only two measurements of anthropometric values were available for each subject in the Diogenes cohort besides the cohort from Potsdam. Therefore, we also calculated trends until 2015 in the EPIC-Potsdam dataset where five measurements per participant were available to investigate which of the two prediction models (linear vs. non-linear) gives the better model fit. Furthermore, we compared models using five data points with models using two measurements to validate our projections. The predicted prevalence estimates for 2015 of these two different models showed a relatively good agreement (data not shown). Thus, we can draw the conclusion that also two observation points might be sufficient for a valid future projection.

However, the results of this study could also be biased by several factors such as selection bias or differential loss-to-follow-up bias. This implies that in general those who are obese or have a poor state of health are less likely to participate in a study for the whole period of time [26]. Hence, our data might underestimate real obesity prevalence trends even though already sizeable increases in obesity prevalence during life course across all study centers were observed. Furthermore, self-reporting of BMI can lead to under-estimations. As already mentioned, we applied correction equations to address this issue. Moreover, the same trends in obesity prevalence were observed in the cohorts that have measured data indicating that it might be valid to rely on corrected self-reported data. A further limitation of the study is that projections are based on a number of assumptions; some of them were simplified scenarios e.g. that the current trend will continue. Potential future policy-, environmental- and behavioral changes may disprove our predictions.