1. Introduction
Over the past three decades, the prevalence of obesity has risen tremendously across the globe [
1] to the point that it is now considered a pandemic [
1,
2]. Obesity constitutes a major health burden [
3] since there is evidence of strong links between obesity and life-threatening chronic diseases such as type II diabetes, cardiovascular disease, stroke, and multiple types of cancer [
4,
5,
6]. As a consequence, the rise in obesity has led to recent declines in survival and life expectancy [
7,
8]. Because the health burden associated with obesity is so significant, its estimation bears high relevance and importance.
In quantifying the health burden of obesity at the population level, the population attributable fraction (PAF) is commonly used [
2]. The PAF is defined as the proportion of total events (e.g., deaths) in a population that could be prevented if a particular risk factor (e.g., obesity) could be eliminated [
2]. The PAF combines information on the proportion of the population exposed to obesity (prevalence) with the relative risk (RR) of dying from obesity [
9].
Over the years, many methodologies for estimating obesity-attributable mortality fractions (OAMF) by means of different PAF formulas have been developed and range from approaches that use RRs for all-cause mortality (all-cause approach) to more recent approaches that use RRs for obesity-related causes of death (cause-of-death approach) (See
Supplementary Material file 1). Within the all-cause approach, there are various methods for estimating OAMF that require varying degrees of data availability (see the
Supplementary Material file 1). Thus, implementing some of these methods can be difficult. The partially adjusted method [
10,
11,
12,
13], which multiplies the adjusted RR of dying from obesity with the obesity prevalence in the studied population, is often used [
14,
15]. In the weighted sum method, unadjusted RRs by age and sex (for instance) are commonly weighted by the obesity prevalence within each subgroup [
16]. The Comparative Risk Assessment (CRA) methodology, which was recently developed by the Global Burden of Disease (GBD) Study, uses cause-specific mortality, cause-specific RRs, and the population distribution of BMI to estimate cause-specific shares of mortality due to a high BMI (≥23 kg/m
2 [
17]. Due to their focus on high BMI, the CRA estimates cannot be readily compared with other estimates that focus strictly on obesity (BMI ≥ 30 kg/m
2).
As previously published research has shown, estimates of obesity-attributable mortality vary depending on which methodology is used [
18]. For example, in 1991, the number of obesity-related deaths in the United States was ~196,000 when the weighted sum method was used and was ~230,000 obesity-related deaths when the partially adjusted method was applied [
15].
The use of different methods and the range of outcomes these methods generate not only cause uncertainty about the true population-level effects of obesity on mortality in a single calendar year but also hamper the construction of time series. First, the use of different methods over time makes it difficult to construct time series of PAFs based on existing studies. Second, data limitations can also pose challenges when estimating time series. In particular, more advanced PAF methods require data that simply are not consistently available over a longer time period (see
Supplementary Material file1). To date, only one previous study has examined the long-term trends in obesity-attributable mortality and did so for Canada using an all-cause approach [
19]. In addition, the GBD study estimates mortality due to high BMI every five years from 1990 to 2015 [
20]. Because the GBD study is updated regularly based on the latest research findings, it is unclear whether the same methodology was applied and the same cause-of-death and RR data were used in each update [
20,
21].
Previous research on obesity-attributable mortality has focused on the United States [
11,
13,
15] in part because of the availability of large cohort studies for the US as a whole. For Europe, by contrast, there is little available information on obesity-attributable mortality levels and even less information on trends. To the best of our knowledge, the influence of the chosen method on estimates of obesity-attributable mortality trends has not previously been assessed.
Our objective is to assess the impact of the use of different estimation techniques on both the levels of and the trends in obesity-attributable mortality. More specifically, we compare approaches that can actually be used to estimate the long-term trends given the data that are available for the European context: namely, the partially adjusted method, the weighted sum method, a combined version of these methods, and the CRA approach. To enable this comparison, we adapted the CRA approach so that it calculates the PAF related to obesity only.
3. Results
The different approaches generated different estimates of the OAMFs for The Netherlands in 2013 (see
Table 3). Using worldwide RRs, the weighted sum method provided the lowest estimates for men and women combined (0.92%) and for men (0.86%). However, the partially adjusted method provided a slightly lower estimate for women (0.94%) than the weighted sum method (0.98%). The adjusted CRA approach using the 2013 world RRs generated the highest estimates for men and women combined (1.46%) and for women (1.62%) while the combined all-cause method using the world RRs provided the highest estimate for men (1.43%). The use of European-specific RRs instead of worldwide RRs resulted in higher estimates. The weighted sum method using the European-specific RRs resulted in the highest estimates overall (1.78%).
Overall, the different approaches—with the exception of the results for women generated by the CRA approaches—showed that the OAMF levels increased over the period 1981 to 2013 and especially from 1993 onwards (see
Figure 1,
Table 4). For men, the trends are parallel for the different approaches even though, in terms of the percentage change, the CRA approaches resulted in larger overall increases (>75%) (
Table 4) than the all-cause approaches (around 50%). For women, the trends identified by the adjusted CRA approaches clearly differed from those found by the other approaches. Over the period 1981 to 2013, both the adjusted CRA approach with recent RRs and the adjusted CRA with less recent RRs resulted in a decline (−1.7% and −9.6%, respectively) as well as with a slight increase from 1993 onwards (4.3% and 0.7%, respectively). For women, the other approaches estimated the overall percentage increase at around 85% even though the partially adjusted method using European RRs resulted in a larger increase (133%) and the weighted sum method using worldwide RRs in a smaller increase (63%). When applied to men and women combined (
Figure S1), the partially adjusted method and the weighted sum method produced very similar levels and trends from 1993 onwards. The same was observed for the two CRA approaches.
5. Conclusions
Estimates of both the levels of, and the trends in, obesity-attributable mortality fractions in The Netherlands differed depending on the method applied as well as on the underlying data and the relative risks (RRs) used. Since obesity prevalence is relatively low in The Netherlands, we would expect to find even larger differences for countries with higher obesity prevalence. In quantifying the health burden of obesity at the population level, it is, therefore, essential to compare different methodologies and different RRs.
Comparisons of obesity-attributable mortality between countries and over time can only be performed accurately by using the same methodology as well as comparable data and RRs. Since the data that are currently available for Europe are limited, we recommend using the weighted-sum method and European RRs to compare obesity-attributable mortality across European countries and over time.