1. Epidemiology and Global Health

Future life expectancy in Europe taking into account the impact of smoking, obesity, and alcohol

  1. Fanny Janssen  Is a corresponding author
  2. Anastasios Bardoutsos
  3. Shady El Gewily
  4. Joop De Beer
  1. Netherlands Interdisciplinary Demographic Institute - KNAW/University of Groningen, Netherlands
  2. Population Research Centre, Faculty of Spatial Sciences, University of Groningen, Netherlands
https://doi.org/10.7554/eLife.66590
Share this article
Cite this article
  1. Fanny Janssen
  2. Anastasios Bardoutsos
  3. Shady El Gewily
  4. Joop De Beer
(2021)
Future life expectancy in Europe taking into account the impact of smoking, obesity, and alcohol
eLife 10:e66590.
https://doi.org/10.7554/eLife.66590
  2. Download BibTeX
  3. Download .RIS

Abstract

Introduction: In Europe, women can expect to live on average 82 years and men 75 years. Forecasting how life expectancy will develop in the future is essential for society. Most forecasts rely on a mechanical extrapolation of past mortality trends, which leads to unreliable outcomes because of temporal fluctuations in the past trends due to lifestyle ‘epidemics’. Methods: We project life expectancy for 18 European countries by taking into account the impact of smoking, obesity, and alcohol on mortality, and the mortality experiences of forerunner populations. Results: We project that life expectancy in these 18 countries will increase from, on average, 83.4 years for women and 78.3 years for men in 2014 to 92.8 years for women and 90.5 years for men in 2065. Compared to others (Lee–Carter, Eurostat, United Nations), we project higher future life expectancy values and more realistic differences between countries and sexes. Conclusions: Our results imply longer individual lifespans, and more elderly in society. Funding: Netherlands Organisation for Scientific Research (NWO) (grant no. 452-13-001).

eLife digest

On average, in Europe, men can currently expect to live till the age of 75 and women until they are 82. But what will their lifespans be in the next decades? Reliable answers to this question are essential to help governments plan for future health care and social security costs.

While medical improvements are likely to further extend lifespans, lifestyle factors can result in temporal distortions of this trend. Yet, most estimates of future life expectancy fail to consider changing lifestyles, as they only use past mortality trends in their calculations. This can make these projections unreliable: for example, increases in smoking rates among Northern and Western European men led to stagnating male life expectancies in the 1950s and 1960s, but these picked up again after smoking declined. The same pattern is showing for women, except it is lagging as they took up smoking later than men. Based simply on the extrapolation of past mortality trends, current projection models fail to consider the past and predicted modifications of life expectancy trends prompted by changing rates of health behaviours – such as increases followed by (anticipated) declines in alcohol consumption and obesity rates, similar to what was observed with smoking.

To produce a more reliable forecast, Janssen et al. incorporated trends in smoking, obesity, and alcohol use into life expectancy projections for 18 European countries. The predictions suggest that life expectancy for women in these countries will increase from 83.4 years in 2014 to 92.8 years in 2065. For men, it will also go up, from 78.3 to 90.5 years.

In the future, this integrative approach may help to track the effects of health-behaviour related prevention policies on life expectancy, and allow scientists to account for changes caused by the COVID-19 pandemic. In the meantime, these estimates are higher than those obtained using more traditional methods; they suggest that communities should start to adjust to the possibility of longer individual lifespans, and of larger numbers of elderly people in society.

Introduction

Mortality projections are essential for forecasting how long people will live on average, for predicting the future extent of population ageing, for determining the sustainability of pension schemes and social security systems, for setting life insurance premiums, and for helping governments plan for the demand for services such as healthcare (European Commission, 2009). Recently, the societal and academic relevance of obtaining reliable and robust mortality forecasts has increased due to the revived debate on the existence and proximity of a limit to human life expectancy (Fries, 1980; Manton et al., 1991; Couzin, 2005; de Beer et al., 2017; Olshansky and Carnes, 2019; Vaupel et al., 2021); the linkage between life expectancy developments and retirement age in many European countries (Carone et al., 2016; OECD, 2016); the stagnation of increases in life expectancy in the United States, the United Kingdom, and other European countries since 2011 (Murphy et al., 2019; Raleigh, 2019); and the recent COVID-19 pandemic.

The growing relevance of obtaining reliable mortality forecasts has led to important advances in mortality forecasting by professionals and researchers from – among others – demography, actuarial sciences, and public health (Janssen, 2018). In the past, mortality forecasts were often based on expectations fuelled by the debate on the limit to life expectancy (Oeppen and Vaupel, 2002; Booth and Tickle, 2008). The majority of the currently existing mortality forecasting methods, including those used by statistical offices in Europe, are mainly based on extrapolations of past trends in age-specific mortality (Booth and Tickle, 2008; Stoeldraijer et al., 2013a). These extrapolative methods rely on the regularities observed in mortality trends over time and in the age patterns (Booth and Tickle, 2008). An extrapolative approach to mortality forecasting is considered more objective, easier to apply, and more likely to result in reliable forecasts than the previously employed expectation approaches and explanation approaches (mortality forecasting by cause of death or with an explanatory model) (Booth and Tickle, 2008). The stochastic Lee–Carter mortality projection methodology (Lee and Carter, 1992) has long been the benchmark extrapolative mortality forecasting method (see Janssen, 2018).

It is, however, increasingly being acknowledged that mechanically extrapolating past national mortality trends can result in unreliable outcomes (e.g., Janssen and Kunst, 2007; Janssen et al., 2013; Bongaarts, 2014; Stoeldraijer, 2019; Li and Raftery, 2021; Blakely, 2018). First, when past trends in mortality are non-linear, the projection outcome will depend to a large extent on the period on which the extrapolation is based (Janssen and Kunst, 2007; Janssen et al., 2013; Stoeldraijer, 2019). Second, extrapolative mortality forecasts made for individual populations often project unrealistic crossovers or large future differences in life expectancy between populations (Li and Lee, 2005; Janssen et al., 2013; Hyndman et al., 2013). For example, extrapolating the large mortality declines observed in Eastern European countries – which have historically had relatively low life expectancy, but have experienced improvements since 2005 (Trias-Llimós et al., 2018) – would result in long-term future life expectancy values being higher in Eastern than in Western European countries with historically high life expectancy values and more moderate mortality declines.

We argue that to obtain a reliable and robust mortality forecast, it is essential to distinguish between (1) the general and gradual long-term mortality decline due to socio-economic and medical progress that can be extrapolated into the future, thereby taking into account the mortality experiences of other countries, and (2) remaining factors that cause deviations from this general mortality decline as well as country and sex differences in this trend.

While life expectancy has increased overall as a result of socio-economic and medical progress (Omran, 1998; Oeppen and Vaupel, 2002; Mackenbach, 2013), past trends in mortality and life expectancy were characterised by periods of stagnation or even decline and exhibited large differences between countries and sexes (e.g., Vallin and Meslé, 2004; Leon, 2011). Lifestyle factors, particularly smoking, alcohol abuse, and obesity, contributed greatly to these deviations and differences (e.g., Lhachimi et al., 2016; WHO, 2018a; Janssen et al., 2021). For example, the decelerating life expectancy improvements among men in many north-western European countries in the 1950s and 1960s, and, in later decades, in other European countries and among women, can be largely explained by the smoking epidemic (Rostron and Wilmoth, 2011; Janssen et al., 2015; Lindahl-Jacobsen et al., 2016). The pattern of a rapid increase and a subsequent decline in smoking prevalence occurred first among men in Anglo-Saxon countries, and was followed later in other countries and by women (Lopez et al., 1994). These developments resulted in similar patterns in smoking-attributable mortality about 30–40 years later (Lopez et al., 1994; Janssen, 2020). The large increases in alcohol consumption and alcohol-attributable mortality between 1990 and 2005 in Eastern Europe contributed to the stalling of life expectancy improvements in Eastern Europe in this period (Trias-Llimós et al., 2018). The more favourable alcohol-related trends in Eastern Europe since 2005 have contributed to the convergence in life expectancy levels between Eastern and Western Europe. The stagnation in life expectancy improvements since approximately 2011 in the United States and the United Kingdom, but also in other selected European countries (Murphy et al., 2019), can be partly attributed to the high obesity prevalence and obesity-attributable mortality in these countries (Raleigh, 2019; Janssen et al., 2020a; Janssen et al., 2020b ) and to the increasing impact on life expectancy levels of past large increases in obesity prevalence (Vidra et al., 2019).

Previous mortality forecasts that incorporated lifestyle information (Bongaarts, 2006; Pampel, 2005; Wang and Preston, 2009; King and Soneji, 2011; Preston et al., 2014; French and O'Hare, 2013; Foreman et al., 2018; Janssen and Kunst, 2007; Janssen et al., 2013; Li and Raftery, 2021) generally included information on only one lifestyle factor, usually smoking. Moreover, these forecasts employed very different techniques and ignored the mortality experiences of other countries. The previous so-called coherent or multi-population forecasting methods that take into account the mortality experiences of other countries (Li and Lee, 2005; Hyndman et al., 2013; Enchev et al., 2017; Eurostat, 2020b; Bergeron-Boucher et al., 2018) have, despite their growing popularity, so far been applied almost exclusively to all-cause mortality. The two previous projections that incorporated both elements simultaneously only included the effect of smoking (Janssen et al., 2013; Li and Raftery, 2021). The approach developed for this purpose by Janssen et al., 2013 consists of the combination of a coherent projection of non-smoking-related mortality with an age-period-cohort projection of smoking-attributable mortality. This approach was applied to the Netherlands (Janssen et al., 2013) and was adopted by Statistics Netherlands as part of their official population forecast (Stoeldraijer et al., 2013b). Li and Raftery, 2021 applied an amended version of this approach to 69 countries.

Our objective is to project future life expectancy in 18 European countries, while simultaneously taking into account the time-varying impact of smoking, obesity, and alcohol on mortality, and the mortality experiences of forerunner populations.

Materials and methods

Our approach to mortality forecasting relied on an analysis of past mortality trends and their determinants. In line with the evidence gathered, we distinguished between (1) the general and gradual long-term mortality decline not affected by the three lifestyle factors that could be extrapolated into the future, while taking into account the mortality experiences of other countries, and (2) deviations from and differences in this general mortality decline caused predominantly by the time-varying impact of smoking, obesity, and alcohol on mortality, which required the use of more advanced projection techniques.

The projection involved four steps. First, we determined the long-term decline in mortality and life expectancy without the combined effect of smoking, obesity, and alcohol. To this end, we used existing recent age- and sex-specific estimates of mortality after excluding the combined impact smoking, obesity, and alcohol (Janssen et al., 2021), which we refer to as non-lifestyle-attributable mortality. Second, we projected this long-term underlying mortality decline into the future, while taking into account the mortality experiences of other populations. Specifically, we performed a Li–Lee coherent projection (Li and Lee, 2005) of non-lifestyle-attributable mortality in which we regarded the high life expectancy values and more favourable long-term trends among women in France, Spain, and Italy as the values and trends towards which the other populations will converge, since we regard these three populations as forerunners. Third, we obtained future estimates of mortality attributable to smoking, obesity, and alcohol, which we refer to as lifestyle-attributable mortality. We did so by utilizing recently published projections of smoking-, obesity-, and alcohol-attributable mortality that were both data and theory driven (Janssen et al., 2020c; Janssen et al., 2020b; Janssen et al., 2020d). Fourth, we combined the projections of non-lifestyle-attributable mortality and lifestyle-attributable mortality by extending the approach that Janssen et al. developed for combining the separate projections of non-smoking- and smoking-attributable mortality (Janssen et al., 2013).

We applied our projection approach to the national populations of 18 European countries (see Table 1 for the included countries) using country, sex, and age-specific all-cause mortality and exposure data for the 1990–2014 period from the Human Mortality Database (HMD and University of California, Berkeley (USA), and Max Planck Institute for Demographic Research (Germany), 2019), and similar data for lifestyle-attributable mortality and non-lifestyle-attributable mortality from a recent study (Janssen et al., 2021). We were restricted in both the starting year and the end year because lifestyle-attributable mortality data was available only for this period.

Table 1
Comparison of observed gains in life expectancy at birth (e0) between 1990 and 2014 with the gains when the effects of smoking, obesity, and alcohol are removed ( = for non-lifestyle-attributable mortality), 18 European countries, by country and sex.
CountryGain e0 1990–2014
ObservedNon-lifestyle-attributable mortality
MenWomenMenWomen
Austria6.694.875.055.29
Belgium5.884.184.224.90
Czech Republic8.196.335.506.62
Denmark6.544.944.985.01
Finland7.194.965.195.56
France6.564.474.954.84
Germany6.524.924.995.45
Greece3.804.183.444.23
Hungary7.125.475.616.16
Ireland7.035.525.975.90
Italy6.924.914.674.97
Netherlands6.053.204.204.50
Norway6.584.296.105.21
Poland7.416.136.567.04
Slovenia8.206.005.626.39
Sweden5.543.665.094.35
Switzerland6.984.405.254.76
United Kingdom6.414.494.894.67
Average6.644.835.135.33
Min3.803.203.444.23
Max8.206.336.567.04
Variance0.920.660.510.62
Annual change0.280.200.210.22
Forerunners*4.814.815.005.00

  1. *French, Italian, and Spanish women (unweighted average). For Spanish women, the gain in e0 (1990–2014) was 5.05 years for all-cause mortality and 5.20 years for non-lifestyle-attributable mortality.

Table 1—source data 1

Data behind Table 1.

https://cdn.elifesciences.org/articles/66590/elife-66590-table1-data1-v2.xlsx

We obtained estimates of life expectancy at birth (e0), including 95% projection intervals up to 2065, by applying standard life table techniques to the projected mortality rates for ages 0–130 (Preston et al., 2001). We focussed on life expectancy at birth because it is a very common summary measure of health, and the most common output measure of mortality forecasts. We have chosen a relatively large projection horizon (2015–2065) given the comparatively short historical time series available (1990–2014) to illustrate that our approach is able to generate reliable outcomes for the long-term future.

We compared our projection outcomes with the outcomes of the benchmark Lee–Carter extrapolation (Lee and Carter, 1992) applied to all-cause mortality and examined the separate effects of incorporating lifestyle factors, and of including the mortality experiences of other countries. We report the outcomes of additional projections and comparisons in the Results section ‘The differences explained’. We also compared our outcomes to the official forecasts by Eurostat and United Nations (see the Discussion section ‘Comparison with other projections’).

More detailed information on the data and methods can be found in Appendix 1.

Results

Underlying long-term increase in life expectancy

Over the 1990–2014 period, life expectancy at birth (e0) increased, on average, across the 18 European countries studied by 6.6 years for men (from 71.8 to 78.3 years) and by 4.8 years for women (from 78.6 to 83.4 years) (see Table 1). These values translate into a yearly increase in e0 of 0.28 years for men and 0.20 years for women. This difference between men and women can largely be attributed to lifestyle factors. Without the impact of smoking, obesity, and alcohol, the increase in e0 over the 1990–2014 period was more similar between men (5.1 years; 0.21 annually) and women (5.3 years; 0.22 annually). In addition, the increase in e0 was more similar between countries for non-lifestyle-attributable mortality (variance: 0.5 for men, 0.6 for women) than for all-cause mortality (variance: 0.9 for men, 0.7 for women).

The greater increase in e0 for all-cause mortality than for non-lifestyle-attributable mortality among men can be explained by the significant declines in lifestyle-attributable mortality men experienced (see Figure 1). These declines stemmed predominantly from large declines in smoking-attributable mortality (Figure 1—figure supplement 1), after a period of sharp increases. In contrast, the past increases in lifestyle-attributable mortality among women were driven by increases in all three factors (except alcohol in Eastern Europe) (Figure 1—figure supplement 2) and resulted in smaller increases in e0 for all-cause mortality than for non-lifestyle-attributable mortality.

Figure 1 with 2 supplements see all
Download asset Open asset
Past and projected age-standardised lifestyle-attributable mortality fractions (LAMF) (20–84), 1990–2065, by sex, for 18 European countries.
Figure 1—source data 1

Data behind Figure 1.

https://cdn.elifesciences.org/articles/66590/elife-66590-fig1-data1-v2.xlsx

Figure 2 and Figure 2—figure supplement 1 show that whereas trends in e0 since 1950 have been rather unstable, the trends in e0 without lifestyle factors can be regarded as more stable. This is indicated by (1) the more stable trend in e0 for non-smoking-related mortality for men in the 1950s and 1960s, which seems to be in line with the more recent trends in e0 for non-lifestyle-attributable mortality and (2) the close correspondence among women between the recent trends in e0 for non-lifestyle-attributable mortality and the trends in e0 for all-cause mortality in the 1950s–1960s that were probably affected only minimally by lifestyle factors.

Figure 2 with 1 supplement see all
Download asset Open asset
Comparison of trends in observed life expectancy at birth (e0) ( = all-cause) with trends in e0 when the effects of smoking, obesity and/or alcohol are removed ( = non-smoking, non-obesity, non-alcohol, non-lifestyle combined), 1950–2014*, men.

*Based on the available information. Reproduced from various panels in Supplementary Figure 3a from Janssen et al., 2021, under the terms of a Creative Commons Attribution License (CC-BY 4.0; https://creativecommons.org/licenses/by/4.0/).

Figure 2—source data 1

Data behind Figure 2.

https://cdn.elifesciences.org/articles/66590/elife-66590-fig2-data1-v2.xlsx

Projected future level of life expectancy

We extrapolated the more universal and more stable mortality trends that we observed for non-lifestyle-attributable mortality, while assuming that the trends for men and women in the individual countries will eventually move towards the more favourable long-term trends for women in France, Spain, and Italy. We selected these populations as the forerunner populations in terms of life expectancy in Europe because they exhibit both very high recent e0 values and very favourable long-lasting past trends in e0 (Stoeldraijer, 2019).

We added to these projections the projected levels of lifestyle-attributable mortality (see Figure 1). For men, the past declines in age-standardised lifestyle-attributable mortality fractions (LAMF) over the 1990–2014 period are projected to further decline until 2065, albeit at a different pace than in the past. Among Eastern European men in particular, the decline is projected to decelerate. For women, we project that the past increases in LAMF will (eventually) turn into declines. Figure 1—figure supplements 1 and 2 show the projections of smoking-, obesity-, and alcohol-attributable mortality (Janssen et al., 2020c; Janssen et al., 2020b; Janssen et al., 2020d). The projections follow the observed wave-shaped dynamic of the smoking epidemic (Lopez et al., 1994; Janssen, 2020) and the hypothesised wave shape of the obesity epidemic (Xu and Lam, 2018; Jaacks et al., 2019). For alcohol-attributable mortality, unrealistic future differences between countries were avoided by assuming that the current increases observed for selected countries will eventually turn into declines (Janssen et al., 2020d).

We project that in the 18 European countries studied, life expectancy at birth will increase from, on average, 78.3 years for men and 83.4 years for women in 2014 to 90.5 years for men and 92.8 years for women in 2065 (Table 2; Figure 3). This represents an average increase per year of 0.24 years for men and 0.18 years for women. The projected increase is greater for men than for women because all-cause mortality levels are currently further away from non-lifestyle-attributable mortality levels for men than for women and because lifestyle-attributable mortality is higher among men than among women. There is, therefore, more room for improvement for men than for women.

Figure 3 with 3 supplements see all
Download asset Open asset
Observed and projected life expectancy at birth (including 95% projection intervals) using our projection approach, which takes into account the impact of smoking, obesity, and alcohol on past and future mortality trends, and the mortality experiences of forerunner countries, 18 European countries, by country and sex, 1990–2065.
Figure 3—source data 1

Data behind Figure 3.

https://cdn.elifesciences.org/articles/66590/elife-66590-fig3-data1-v2.xlsx
Table 2
Observed (2014) and projected (2065) life expectancy at birth (e0), by country and sex, according to our projection methodology, which takes into account the impact of smoking, obesity, and alcohol and the mortality experiences of forerunner countries (‘lifestyle and coherent’); the benchmark Lee–Carter extrapolative mortality projection applied to all-cause mortality (‘Lee–Carter’); and when purely accounting for smoking, obesity, and alcohol (‘adding lifestyle’).
CountryMenWomen
e0 2014e0 2065e0 2014e0 2065
Lee–CarterAdding lifestyleLifestyle and coherentLee–CarterAdding lifestyleLifestyle and coherent
Austria78.9188.9189.0290.3883.7391.4292.5792.54
Belgium78.5787.7688.2090.9383.5290.0992.4392.98
Czechia75.7287.7787.2388.8681.7390.8792.2891.37
Denmark78.5687.7288.4890.5082.6789.6693.3393.08
Finland78.1388.2588.3090.2283.8590.9092.6592.69
France79.2889.1190.0391.4085.4492.7794.2494.01
Germany78.4388.3688.6790.1883.3590.7892.6092.62
Greece78.4685.0786.3290.9783.8389.7590.5192.44
Hungary72.2681.7485.8988.6479.2487.0990.2190.96
Ireland79.1589.5990.5790.9583.2391.6794.3893.40
Italy80.5589.2589.1491.4385.1692.3093.3493.53
Netherlands79.8788.6788.7391.1183.2988.8092.5993.18
Norway80.0388.8690.2390.8784.0990.7293.4893.35
Poland73.6685.2188.2889.3781.4190.4492.7491.88
Slovenia77.9688.6888.9589.9183.6992.2093.7992.60
Sweden80.3588.3789.1890.6484.0589.9292.2793.01
Switzerland80.9389.7789.3190.8885.1191.4192.7093.16
United Kingdom79.2588.9989.3591.0482.9989.9392.7093.34
Average78.3487.8988.6690.4683.3590.5992.7192.78
Min72.2681.7485.8988.6479.2487.0990.2190.96
Max80.9389.7790.5791.4385.4492.7794.3894.01
Variance4.983.771.410.622.021.731.070.55
Table 2—source data 1

Data behind Table 2.

https://cdn.elifesciences.org/articles/66590/elife-66590-table2-data1-v2.xlsx

Compared to the average past increases (1990–2014) in e0 (0.28 years for men; 0.20 years for women), these projected increases are lower, particularly for men. This is because – as mentioned – the increase in e0 for men over the 1990–2014 period is actually an acceleration compared to the underlying trend in non-lifestyle-attributable mortality because of the decline in lifestyle-attributable mortality, predominantly in smoking-attributable mortality.

The highest e0 in 2065 is projected for France among women (94.0) and for Italy among men (91.4). Hungary is the country with the lowest projected e0 in 2065, at 88.6 years among men and 91.0 years among women. These countries also had in 2014, respectively, the (second) highest and the lowest values. We project the largest gains in life expectancy for Hungary and Poland, particularly among men, in line with past large increases in non-lifestyle-attributable mortality due to a process of catch-up (Table 1), and large projected declines in lifestyle-attributable mortality in these countries (Figure 1). Among women, we project large gains as well for Denmark and the United Kingdom, which can be attributed to their high current levels of lifestyle-attributable mortality, predominantly due to their high levels of smoking-attributable mortality, and the large projected (continued) declines therein (Figure 1—figure supplement 2). The smallest gains are projected for Switzerland, which could be related to the relatively small projected declines in lifestyle-attributable mortality resulting from a long projected continuation of increases in obesity prevalence.

The sex difference in e0 is projected to decline from, on average, 5.0 years in 2014 to 2.3 years in 2065 (Supplementary file 1A; Figure 3). The gap is projected to be smallest in Greece, at 1.5 years; and largest in Slovenia, at 2.7 years. Among the explanations for the projected convergence of the sex gap in e0 is that the sex differences are smaller for non-lifestyle-attributable mortality than for all-cause mortality; the sex differences in lifestyle-attributable mortality are smaller in the long run; and our assumption that the trends in non-lifestyle-attributable mortality will converge.

Comparison with the benchmark Lee–Carter mortality projection

Compared to the benchmark Lee–Carter (LC) extrapolative mortality projection, our projection results in higher future e0 in 2065 for all populations under study (Table 2). On average, the projected e0 values in 2065 are 2.6 years higher for men and 2.1 years higher for women in our projection than in the individual Lee–Carter mortality projection, which projected e0 values of 87.9 years for men and 90.6 years for women in 2065 (Table 2). For men, particularly in Hungary, the differences between our projected values and those of the LC projection are considerable, mainly because of large projected declines in lifestyle-attributable mortality. Also for Dutch women a considerable difference between our projection and the LC projection can be observed. For Dutch women not only large declines in lifestyle-attributable mortality are projected, but they also experienced a more favourable trend in non-lifestyle-attributable mortality than in all-cause mortality.

In addition, our projection results in smaller differences in future e0 between countries and sexes. That is, the LC projection projects that the variance between countries in e0 for 2065 is 3.8 years for men and 1.7 years for women, whereas in our projection, the expected variance is 0.6 years for both men and women (Table 2). In addition, Figure 3—figure supplement 1 illustrates that the LC projection, unlike our projection, can result in unlikely crossovers and divergences between countries (e.g., Greece and Hungary for men; the Netherlands and Slovenia for women).

In addition, the sex difference in e0 in 2065 is projected to be 2.7 years in the LC projection, but 2.3 years in our projection (Table 2). Furthermore, a closer look at the sex differences projected by the LC projection (Supplementary file 1A) reveals large differences in the projected sex differences between countries ranging from 0.1 to 5.4 years. In our projection, by contrast, the range is between 1.5 and 2.7 years. The very small difference in e0 between women and men projected for 2065 in the Netherlands (0.1 years) and in the United Kingdom (0.9 years) indicates that the individual LC projection of all-cause mortality might result in an unlikely crossover between men and women in e0 in the long-term future due to the more favourable past trends in all-cause mortality among men than among women in the 1990–2014 period. Thus, our projection avoids an unlikely crossover in future life expectancy values between men and women.

Moreover, unlike the LC mortality projection, our projection results in non-linear future increases in e0 (Figure 3; Figure 3—figure supplement 2). Particularly for men in Belgium, Greece, the Czech Republic, Hungary, and Poland, but also for women in Hungary, Denmark, and the United Kingdom, we see faster projected increases in life expectancy in the early decades, but slower increases in life expectancy further in the future, in line with the projections of lifestyle-attributable mortality. The LC projection, by contrast, anticipates more constant future increases in life expectancy at birth.

The differences explained

The observed differences between our projection and the Lee–Carter projection can stem from either the lifestyle dimension, the coherent dimension, or both. Within the lifestyle dimension, two mechanisms apply: first, the projection of non-lifestyle mortality instead of all-cause mortality, and, second, the effect of projected lifestyle-attributable mortality itself.

The higher values estimated by our projection than by the LC projection stem from both dimensions (Table 2; Supplementary file 1B). The coherent dimension led to higher values because the French, Italian, and Spanish women combined exhibited more favourable trends in non-lifestyle-attributable mortality than the individual populations combined, after controlling for the underlying age pattern of mortality. The lifestyle dimension resulted in an additional effect, which was particularly large for women. For women, unlike for men, the past trends in non-lifestyle-attributable mortality were more favourable than the past trends in all-cause mortality. For both men and women, an additional positive effect emerges from the projected (eventual) decline in lifestyle-attributable mortality.

The smaller projected differences between countries in our projection than in the LC projection also stem from both elements, although for men, the impact of integrating lifestyle is more pronounced (Table 2; Supplementary file 1B). Particularly for men, the past trends in non-lifestyle-attributable mortality were much more similar between countries than the past trends in all-cause mortality. In addition, for both men and women, country differences in lifestyle-attributable mortality in 2065 are projected to be smaller than they are currently.

The smaller sex differences in projected e0 values in our projection stem purely from the convergence dimension, whereas the nonlinearity of future trends in our projection stems completely from the lifestyle dimension. Figure 3—figure supplement 3 illustrates that the inclusion of the lifestyle dimension in the Lee–Carter projection (thus ignoring the coherent dimension) leads to a future e0 that is moving back upwards towards the future trend in e0 for non-lifestyle-attributable mortality.

Discussion

Appraisal of our projection methodology

An important element of our projection approach is the identification of the underlying long-term mortality trend that can be extrapolated into the future. We illustrated that past trends in non-lifestyle-attributable mortality serve this purpose better than past trends in all-cause mortality. For non-lifestyle-attributable mortality, the differences in past trends between countries and between men and women are smaller, and the trends are more robust. In contrast, the trends in all-cause mortality exhibit a wave-shaped fluctuation due to initial increases in smoking-, obesity-, and alcohol-attributable mortality, followed (eventually) by declines. This pattern leads, first, to a deceleration of increases in e0, and a move away from the underlying increase in e0 in non-lifestyle attributable mortality, followed by an acceleration of increases in e0, and a move back towards the underlying increase in e0 in non-lifestyle-attributable mortality. Thus, extrapolating past declines in all-cause mortality will lead to non-reliable outcomes: i.e., the outcomes will be either too low when decelerating increases are extrapolated or too high when accelerating increases are extrapolated. Similarly, the extrapolation of past trends in all-cause mortality will result in non-robust outcomes, when updating the projection after a few years. Because the past trends in non-lifestyle-attributable mortality are more linear than the past trends in all-cause mortality, especially from a historical point of view, our projection is less dependent on the data period used for the extrapolation.

Our finding of an average annual past increase in e0 for non-lifestyle-attributable mortality of 0.21 years for men and 0.22 years for women across the 18 European countries over the 1990–2014 period is lower than the continuous upward trend in record life expectancy of 0.243 years that Oeppen and Vaupel, 2002 observed over the 1840–2000 period. There are two main reasons. First, whereas record values always increase, the levels in individual countries may not, and since the 1990s, the record life expectancy has been observed for a non-European country (Japan). Second, more recent increases in life expectancy are driven more by declines in mortality from chronic diseases at older ages with a less strong effect on life expectancy than from declines in mortality from infectious diseases at younger ages, which drove increases in life expectancy up to 1950 (Omran, 1998; Vallin and Meslé, 2009). As a result, a deceleration of increases in life expectancy over time is inevitable. Nonetheless, these large past increases in e0 for non-lifestyle-attributable mortality indicate that considerable increases in e0 are likely in the future.

We applied a coherent mortality projection approach to past trends in non-lifestyle-attributable mortality. This resulted in more realistic estimates of the differences in future e0 values between countries and sexes than those of the individual Lee–Carter projection of past trends. Another added value of our inclusion of the mortality experiences of other countries in our projection is that it provided a broader empirical basis for the determination of the most likely future developments (Janssen and Kunst, 2007), which, in turn, improved the robustness of our projection outcomes (Stoeldraijer, 2019). However, next to a potential effect on the outcomes of the exact coherent methodology that is applied (Stoeldraijer, 2019), it is important to take into account the choice of the reference population (Stoeldraijer, 2019; Booth and Tickle, 2008). Using forerunner populations as the reference population, as advocated by among others Oeppen and Vaupel, 2002 and Bergeron-Boucher et al., 2018, instead of an average across many populations (Li and Lee, 2005; Eurostat, 2020a) will generally result in higher future life expectancy values (see the next section as well). However, given that the past trend in e0 for non-lifestyle-attributable mortality is more equal between countries and sexes than the past trend in e0 for all-cause mortality (see Table 1), which reference populations are chosen will likely have a larger impact when coherently forecasting all-cause mortality than when coherently forecasting non-lifestyle-attributable mortality (see also Janssen and Kunst, 2007 and Janssen et al., 2013). In addition, our convergence assumption relies on – at least – a continuation of improvements in socio-economic developments and medical care. Such continued improvements depend on continued investments in these areas, including diminishing socio-economic inequalities, better preparations for additional possible outbreaks of infectious diseases, and for the potential harmful effects of climate change. Although this might be considered optimistic, our assumption is supported by evidence that life expectancy has long followed a continuously increasing trend (Oeppen and Vaupel, 2002), and that this trend is even stronger from a historical perspective if the effects of lifestyle factors are excluded (see the description of Figure 2 and its supplement in the Results section).

To the projected long-term trend in non-lifestyle-attributable mortality, we added estimates of how smoking-, obesity-, and alcohol-attributable mortality are likely to develop in the near future – and, consequently, how long it will take before all-cause mortality trends again follow more closely the underlying trends in non-lifestyle-attributable mortality. We generated these estimates based on previously projected age-specific smoking-, obesity-, and alcohol-attributable mortality (Janssen et al., 2020c; Janssen et al., 2020b; Janssen et al., 2020d). To avoid generating unrealistic outcomes for the (long-term) future (e.g., continuing increases in smoking-attributable mortality for women, in obesity-attributable mortality, and in alcohol-attributable mortality in the few Eastern European countries where it is still increasing), the projections were based on advanced approaches that did not simply extrapolate past trends, but also took into account knowledge about the future progression of smoking-, alcohol-, and obesity-attributable mortality. Nonetheless, the projection of lifestyle-attributable mortality involves more subjectivity than the extrapolation of non-lifestyle-attributable mortality (see Janssen et al., 2020c; Janssen et al., 2020b; Janssen et al., 2020d for a critical evaluation). For example, in the separate projections of smoking-, obesity-, and alcohol-attributable mortality, population-specific lower bounds were implemented to avoid unlikely future estimates of zero smoking, obesity, and alcohol prevalence, in line with the theory (Lopez et al., 1994; Xu and Lam, 2018), and to prevent unlikely crossovers between sexes and between countries in alcohol-attributable mortality. This approach could be considered conservative. On the other hand, however, the use of a wave-shaped pattern for obesity, and the large declines in alcohol-attributable mortality we project for Eastern European countries, could be considered too optimistic.

However, in our view, the added value of integrating lifestyle into mortality projections outweighs the uncertainties that come with it. First, trends in non-lifestyle-attributable mortality provide a better approximation of the underlying mortality trend. Second, our projection is driven not only by data, but also by theory, and is preceded by a careful study of past trends. Third, by distinguishing between the underlying mortality trend and the additional effect of the three lifestyle factors combined, these projections provide much greater transparency and insight than a mere mechanical extrapolation of past trends could. Fourth, our approach resulted in more realistic differences between countries and sexes than those projected by other approaches (see the next section as well). The abovementioned four added values of our mortality projection approach can be clearly linked to important criteria for evaluating the performance of mortality forecasts in addition to accuracy: namely, robustness, plausibility, reasonableness, and transparency (Cairns et al., 2009). Moreover, the outcomes of our projection approach – which rely above all on the coherent projection of non-lifestyle-attributable mortality – are less dependent on the explicit choices that underlie general mortality forecasts (e.g., the choice of the forerunner populations and the historical time period).

Comparison with other projections

Compared to the benchmark Lee–Carter mortality extrapolation, our projection approach results in higher e0 in the long run; smaller differences in future e0 between countries and sexes; and non-linear future increases in e0, in line with the shift towards the higher e0 values for non-lifestyle-attributable mortality.

Moreover, compared to the most recent official mortality forecast by Eurostat, 2020a and the United Nations, 2020, our mortality projection resulted in substantially higher projected e0 values in 2065 (Supplementary file 1C). Our values were, on average, 5.0 years higher for men and 3.2 years (Eurostat) and 4.3 years (UN) higher for women. The UN projects very large and unrealistic differences in future life expectancy between countries and sexes, including higher projected life expectancy for men than for women in a number of countries. These outcomes clearly show the added value of adopting a coherent mortality projection approach. While Eurostat has performed a coherent mortality projection, it used a different methodology. Eurostat made a partial convergence assumption, applied to men and women separately, and considered the aggregate all-cause mortality of 12 countries (Belgium, Denmark, Germany, Spain, France, Italy, the Netherlands, Austria, Portugal, Finland, Sweden, and the United Kingdom) as the target (Lanzieri, 2009; Eurostat, 2020b). In addition, Eurostat did not integrate the role of lifestyle factors.

The added value of our projection approach over pure coherent mortality forecasting is that we projected the more robust past trends in non-lifestyle-attributable mortality rather than in all-cause mortality, and we projected non-linear future trends in line with the projection of lifestyle-attributable mortality. The reason why coherent mortality projections have to be applied to non-lifestyle-attributable mortality rather than to all-cause mortality is that there are important differences in smoking-, obesity-, and alcohol-attributable mortality trends between countries and sexes (Janssen, 2020; Vidra et al., 2019; Janssen et al., 2020d).

Among the advantages of taking the effects of obesity and alcohol in addition to those of smoking into account is that e0 values for non-lifestyle-attributable mortality are higher than e0 values for non-smoking-attributable mortality (Figure 2, Figure 2—figure supplement 1), and thus have the potential for larger future increases. In addition, as was illustrated, the effects of both obesity-attributable mortality and alcohol-attributable mortality (particularly among Eastern European men) on past and future life expectancy trends are large.

Overall conclusion and implications

All in all, the outcomes of our projection approach, which simultaneously takes into account the lifestyle factors smoking, obesity and alcohol, and the mortality experiences of forerunner populations, can be considered more realistic, more robust, and more insightful than previous extrapolative mortality projections, including those by the United Nations and Eurostat.

Our approach, which distinguishes between the underlying long-term mortality decline and the remaining factors that cause deviations from this mortality decline, could also be adopted to account for the impact of the COVID-19 pandemic on mortality and life expectancy. First, it would be illogical to extrapolate the mortality trends up to 2020 to obtain life expectancy estimates for 2021 and beyond, because life expectancy in 2020 has been importantly negatively affected by COVID-19 (Marois et al., 2020), and will likely to continue to do so for at least 2021. This effect cannot be predicted using a simple extrapolation approach. Second, the underlying long-term mortality decline is argued not to be affected by COVID-19 (e.g., Stoeldraijer, 2020), in line with the predominant temporal dip in life expectancy as a result of the Spanish flu in 1918 (see Roser, 2020). Although, as mentioned, the continuation of the underlying long-term mortality decline does depend on important investments to achieve continued improvements in socio-economic developments and medical care.

Our findings of higher projected life expectancy values than those estimated by the benchmark Lee–Carter mortality projection and the official mortality forecasts by Eurostat and the United Nations have important implications for society. First, individuals could consider planning their life course (e.g., education, work) differently given their longer expected lifespan. Second, a higher projected life expectancy suggests that there will be more older people in the population, which will likely bring both opportunities (e.g., an experienced workforce) and challenges in terms of healthcare planning and social security. Our results imply that either the pension age will have to be raised or, if it is not, that the capital requirements of pension providers will need to be increased. For determining the pension age based on projected life expectancy, a projection that is more robust will have significant added value. More generally, having a more stable projection of mortality will allow for more robust population forecasts.

In addition, our findings of (1) higher projected life expectancy values than those currently used in most mortality projection methodologies and (2) of a rather robust long-term increase in e0 after the effects of smoking, obesity, and alcohol are controlled for, indicate that a limit to human life expectancy (average lifespan) is not within reach – an issue that continues to spark scientific debate (see the first paragraph of the introduction).

Data sharing

Some of our original data regarding lifestyle-attributable mortality were based on previous publications, which, in turn, used data that are openly available. The all-cause mortality data and the exposure data can be obtained through the Human Mortality Database. We have provided source data files for all our tables and figures. These comprise the numerical data that are represented in the different figures, and the output on which the different tables are based. In addition, one excel file with all the final numerical/output data that were used as input for the tables and figures will be made available at the Open Science Framework: https://osf.io/ghu45/. In addition, we will upload there the underlying observed age-specific mortality rates (all-cause mortality, non-lifestyle-attributable mortality, lifestyle-attributable mortality) as well as the adjusted and projected age-specific mortality rates (medians and 90% and 95% projection intervals). The different R codes used for the different steps of the data analyses will be shared – as well – through the Open Science Framework link above.

Appendix 1

Supplementary data and methods

Setting

We included data for the national populations, by sex and age, for 18 European countries, over the period 1990 up to 2014: Denmark, Finland, Norway, Sweden, Austria, Belgium, France, Germany, Ireland, the Netherlands, Switzerland, the United Kingdom, Greece, Italy, the Czech Republic, Hungary, Poland, and Slovenia.

We were restricted to the period 1990–2014 because data on alcohol-attributable mortality were only available from 1990 onwards, which meant, in turn, that we only had data on lifestyle-attributable mortality from 1990 onwards (see Janssen et al., 2021). We included the abovementioned European countries because realistic projections for smoking-, obesity-, and alcohol-attributable mortality were generated for these countries (see Janssen et al., 2020c; Janssen et al., 2020b; Janssen et al., 2020d).

Data

For the abovementioned 18 European countries, for the years 1990 up to 2014, we obtained all-cause mortality (both death numbers and rates) and exposure data by country, sex and single year of age (0–100) from the HMD and University of California, Berkeley (USA), and Max Planck Institute for Demographic Research (Germany), 2019 (update May 1 2019). In addition, we used estimated lifestyle-attributable mortality fractions by country, sex, and single year of age (0–100) from a recent study (Janssen et al., 2021). These lifestyle-attributable mortality fractions represent the share of mortality that can be attributed to smoking, obesity, and alcohol. Multiplying these fractions to the age- and sex-specific all-cause mortality rates, we obtained the age- and sex-specific lifestyle-attributable mortality rates. Subtracting these lifestyle-attributable mortality rates from all-cause mortality rates resulted in non-lifestyle-attributable mortality rates, which – consequently – refers to mortality that is not attributable to smoking, obesity, and alcohol.

In addition, we used for our analysis already published projections of age-specific smoking-, obesity-, and alcohol-attributable mortality fractions for ages 0–84 (Janssen et al., 2020c; Janssen et al., 2020b; Janssen et al., 2020d).

Overall projection approach

Our approach to mortality forecasting relied on an analysis of past mortality trends and their determinants. In line with the evidence gathered, we distinguished between (1) the general and gradual long-term mortality decline not affected by the three lifestyle factors that could be extrapolated into the future, while taking into account the mortality experiences of other countries, and (2) deviations from and differences in this general mortality decline caused predominantly by the time-varying impact of smoking, obesity, and alcohol on mortality, which required the use of more advanced projection techniques.

More specifically, our mortality approach consists of the combination of two projections:

  1. A projection of non-lifestyle-attributable mortality using the coherent Li–Lee (LL) projection methodology, which takes into account the mortality experiences of other populations; and.

  2. A projection of lifestyle-attributable mortality fractions based on already published projections of smoking-, obesity-, and alcohol-attributable mortality fractions.

Main outcome measure

Our main outcome measure is life expectancy at birth (e0) by country and sex, which we estimated for the period 1990 up until 2014, and which we projected up to 2065. Life expectancy at birth is a very important summary measure of health. In addition, it is the most common summary output of mortality forecasts, although the reporting of other summary output measures, such as lifespan disparity, has recently attracted attention (Bohk-Ewald et al., 2017). We have chosen a relatively long projection horizon (2015–2065) given the comparatively short historical time series available (1990–2014) to illustrate that our approach is able to generate reliable outcomes for the long-term future.

To obtain life expectancy at birth, we applied standard life table techniques (Preston et al., 2001) to either the observed or the projected age-specific mortality rates (0–130) for the different national populations under study. In doing so, we used the exponential method to transfer the age-specific mortality rates (mx) into age-specific probabilities (qx), by using the formula qx=1-exp(-mx).

To obtain age- and sex-specific all-cause and non-lifestyle-attributable mortality rates for ages 0–130 to be used as input in our life table calculations, we obtained the respective mortality rates for ages 101–130 by applying the Kannisto model of old-age mortality (Thatcher et al., 1998) to mortality for ages 80–100 for all-cause mortality and non-lifestyle-attributable mortality, respectively.

We compared our projection outcomes foremost with the outcomes of a Lee–Carter extrapolation (Lee and Carter, 1992) applied to all-cause mortality (see next section). In addition, however, we also performed additional projections, to illustrate the effects of the different steps in our projection approach (see the prefinal section of this Appendix). All our projections are based on 50,000 simulations, generated from the forecasts of the time series model. For future e0 we used the median values, and the 95% projection intervals.

Lee–Carter mortality projection

The Lee–Carter (LC) model decomposes logged age-specific mortality rates over time [ln(mx,t)] into an average age effect across time a(x), an overall time trend across all ages k(t), and an age-specific adjustment b(x) of this overall time trend (Lee and Carter, 1992). Projection into the future involves extrapolating the overall time trend k(t). The formula for the Lee–Carter model reads as:

(1) ln(mx,t)=ax+bxkt+εx,t

We generated the parameter estimates based on Poisson-likelihood (Brouhns et al., 2002; Renshaw and Haberman, 2006). For the projection of the period parameter in the Lee–Carter model, we applied the common assumption of a random walk with drift ( = ARIMA(0,1,0) with drift) (Lee and Carter, 1992; Renshaw and Haberman, 2006).

Steps involved in our projection

Our projection involved four steps.

  1. Determining the long-term decline in mortality and life expectancy without the combined effect of smoking, obesity, and alcohol.

  2. Projecting non-lifestyle-attributable mortality into the future, thereby taking into account the mortality experience of other populations.

  3. Obtaining estimates of future mortality that is attributable to smoking, obesity, and alcohol.

  4. Combining the projected non-lifestyle attributable mortality rates with the projected lifestyle-attributable mortality fractions.

Step (1) Determining the long-term trend in mortality and life expectancy without the effect of smoking, obesity, and alcohol

As a first step in our projection, we assessed the trends over time in mortality and life expectancy without the effect of smoking, obesity, and alcohol. More specifically we assessed the gain in life expectancy at birth (e0) over the period 1990–2014 for non-lifestyle-attributable mortality and compared this with the observed gain for all-cause mortality. To this end, we used existing recent age- and sex-specific estimates of the share of mortality due to smoking, obesity and alcohol (Janssen et al., 2021) for ages 20–100, which we refer to as lifestyle-attributable mortality fractions (LAMF).

These age- and sex-specific lifestyle-attributable mortality fractions were estimated based on age- and sex-specific smoking-, obesity-, and alcohol-attributable mortality fractions from earlier studies (Janssen, 2020; Janssen et al., 2020b; Janssen et al., 2020d). Smoking-attributable mortality fractions (1950–2014; 35–100 M; 40–100 F) were indirectly estimated by applying a simplified version of the commonly applied Peto–Lopez methodology (Peto et al., 1992; Janssen et al., 2013; Janssen, 2020) to lung cancer mortality data from the WHO Mortality Database (update April 11 2018) (WHO, 2018b), and to additional epidemiological information regarding smoking from the American Cancer Society’s Cancer Prevention Study II (ACS-CPSII) (Peto et al., 1992; Thun et al., 1997). Obesity-attributable mortality fractions (1975–2016; 20–100) were estimated by applying the population-attributable fraction (PAF) formula to age- and sex-specific obesity prevalence data from the NCD Risk Factor Collaboration study 2017 (Abarca-Gómez et al., 2017) and age group- and sex-specific relative risks of dying from obesity from a meta-review (DYNAMO-HIA Consortium, 2010; Janssen et al., 2020a). Alcohol-attributable mortality fractions (1990–2016; 20–100) were estimated based on alcohol-attributable mortality rates obtained from the Global Burden of Disease Study 2017 (Stanaway et al., 2018; GBD, 2018). To ages 65+, we applied the age pattern observed for the main group of causes of death wholly attributable to alcohol (Semyonova et al., 2014), using data from the WHO Mortality Database (WHO, 2018b) and the Human Cause-of-Death Database, 2017. In the estimation of smoking-, obesity-, and alcohol-attributable mortality fractions, Loess smoothing was used to convert estimates by five-year age groups into estimates by single years of age. By setting the smoking-attributable mortality fractions for ages 20–34 (men) and ages 20–39 (women) to zero, and by restricting the data to the same time period, we ended up with estimated age- and sex-specific smoking-, obesity-, and alcohol-attributable mortality fractions for the 1990–2014 period for ages 20–100.

We, consequently, estimated the past age-and sex-specific lifestyle-attributable mortality fractions for the three lifestyles combined (LAMF) (1990-2014; 20-100), by multiplicatively aggregating the population-attributable mortality fractions for the individual risk factors (PAFi) using the formula:

(2) LAMF=PAF1.n=1−∏i=1n(1−PAFi)

(Ezzati et al., 2003; Janssen et al., 2021). For ages 0–20 we assumed these fractions to be zero.

By applying these age- and sex-specific lifestyle-attributable mortality fractions to age-and sex-specific all-cause mortality rates, we obtained age- and sex-specific non-lifestyle-attributable mortality rates (1990–2014; 0–100).

To obtain the age- and sex-specific all-cause and non-lifestyle-attributable mortality rates for ages 0–130 to be used as input in our life table calculations, we obtained the respective mortality rates for ages 101–130 by applying the Kannisto model of old-age mortality (Thatcher et al., 1998) to mortality for ages 80–100 for all-cause mortality and non-lifestyle-attributable mortality, respectively.

Step (2) Projecting non-lifestyle-attributable mortality into the future, thereby taking into account the mortality experience of other populations

To project age- and sex-specific non-lifestyle-attributable mortality rates (1990–2014; 0–100) into the future, thereby taking into account the mortality experience of other populations, we applied the often used Li and Lee coherent projection methodology (Li and Lee, 2005; Stoeldraijer, 2019). In doing so, we selected women in France, Spain and Italy as the forerunner populations in terms of life expectancy in Europe because they exhibit both very high recent life expectancy values and very favourable long-lasting past trends in life expectancy (measured by strong increases in e0 over the 1970–2011 period among non-Eastern European countries) (Stoeldraijer, 2019, p. 106). Our decision to use forerunner populations as the reference populations is in line with the seminal work by James Vaupel (e.g., Oeppen and Vaupel, 2002), and our decision to use female populations is in line with the recent work by Bergeron-Boucher et al., 2018. In addition, using three populations instead of one is regarded more robust.

The Li–Lee (LL) coherent projection methodology, in essence, applies the Lee–Carter (LC) methodology twice. First, the LC model is applied to the forerunner populations, resulting in a common time trend (=a common improvement over time across all ages) K(t), a common age effect that adjusts the rate of mortality decline for each age B(x) and a common average age-specific death rate over time A(x). Second, the LC model is applied to the residuals, resulting in population-specific time trends k(t,i), population-specific age effects b(x,i), and population-specific average age-specific death rates a(x,i). These parameters are subsequently combined into separate models for each population, by the formula below:

(3) ln(mx,t,i)=ax,i+BxKt+bx,ireskt,ires+εx,t,i

Similar as to the Lee–Carter model, we generated the parameter estimates based on Poisson-likelihood (Brouhns et al., 2002; Renshaw and Haberman, 2006).

The past trend in the common time trend K(t) is extrapolated using the ARIMA(1,1,0) model with drift, which provided the best fit (=minimum corrected Akaike Information Criterion (AICc)) for both all-cause mortality and non-lifestyle-attributable mortality. For the remaining population-specific time trends k(t,i), we used a random walk with no drift, and thus assumed non-stationarity, which means that we used the last observed difference between the population and the common (instead of mean reverting). This assumption is in line with the results of the statistical test of it in most cases (see Appendix 1—table 1). For comparative purposes, we applied the assumption to all populations. Our assumption implies that the population-specific time trends k(t,i) will rely on the past trend in K(t) among the forerunner populations. However this will not always result in parallel age-specific mortality trends between the respective population and the forerunner populations, because of differences between populations in the average age-specific death rate a(x,i) and the age-specific adjustments of the overall time trend b(x,i).

Appendix 1—table 1
Test for stationarity of k(t,i) within the coherent Li-Lee mortality projections, for all-cause mortality and non-lifestyle-attributable mortality in which French, Italian, and Spanish women are selected as the forerunner populations, for the 18 European countries, by sex.

All-cause mortalityNon-lifestyle-attributable mortality
MenWomenMenWomen
Austria1111
Belgium1111
Czech Republic1111
Denmark1111
Finland1111
France0111
Germany1110
Greece1111
Hungary2111
Ireland1111
Italy1010
Netherlands1110
Norway1111
Poland1111
Slovenia2111
Sweden1111
Switzerland1111
United Kingdom1111

  1. 0 = stationary time series; i.e., the time series does not exhibit a trend.

    1 or 2 = non-stationary time series.

We obtained, for each sex-specific population, 50,000 simulated matrices of age (x) times period (t) of non-lifestyle-attributable mortality rates.

To obtain projected mortality rates for ages 101–130, for each sex-specific population, we extrapolated the age pattern for ages 80–100 up until 130 for each forecast year, by applying the Kannisto model of old-age mortality (Thatcher et al., 1998).

Step (3) Obtaining estimates of future mortality that is attributable to smoking, obesity and alcohol

We obtained projected lifestyle-attributable mortality fractions (0–130) based on previously published data- and theory-driven projections of smoking-, obesity-, and alcohol-attributable mortality (20–84) (Janssen et al., 2020d; Janssen et al., 2020b; Janssen et al., 2020a). See Appendix 1—box 1 for some essential information on these projections.

We obtained the projected fractions by lifestyle factor and sex-specific population, by means of 50,000 simulation matrices of age (x) times period (t), for ages 20 up until 84, and for 2015 (or the first available year in the projection) up to 2065. See Appendix 1—table 2 for the available projection years per country and lifestyle factor. The simulation matrices were generated from the forecasts of the different underlying time series models.

Appendix 1—table 2
Overview of the data used for the projections per lifestyle factor and combined.
AgeSmoking-attributable mortality fractions (SAMF)Alcohol-attributable mortality fractions (AAMF)Obesity-attributable mortality fractions (OAMF)*Lifestyle-attributable mortality fractions (LAMF)†
35–100 M; 40–100 F20–10020–10020–100‡
Years by country
Austria19552014199020141975201619902014
Belgium19542015199020151975201619902015
Czech Republic19532016199020161975201619902016
Denmark19512015199020161975201619902015
Finland19522015199020151975201619902015
France19502015199020151975201619902015
Germany19702015199020151975201619902015
Greece19612013199020131981201619902014§
Hungary19552014199020141975201619902014
Ireland19502014199020141975201619902014
Italy19512014199020141975201619902014
Netherlands19502016199020161975201619902016
Norway19512014199020141975201619902014
Poland19592014199020141975201619902014
Slovenia19852014199020141983201619902014
Sweden19512016199020161975201619902016
Switzerland19512015199020161975201619902015
United Kingdom19502015199020161975201619902015

  1. *We projected obesity prevalence data (1975–2016) into the future. To obtain the OAMF, we only needed the RRs of dying from obesity. However, to obtain age-standardised fractions, we also needed all-cause mortality data.

    †See projection step three for details on how we dealt with projections not being available for all the three lifestyle factors for the years 2015–2016.

  2. ‡By setting the SAMF for ages 20–34 (M) and ages 20–39 (F) to zero.

    §We implemented the median projected AAMF and SAMF for Greece as the observed AAMF and SAMF in 2014 so that we could calculate the lifestyle-attributable mortality up until 2014, and so that we could apply it to the all-cause mortality rates to obtain non-lifestyle-attributable mortality rates. For this purpose, we used all-cause mortality from HMD from 26 February 2020.

Subsequently, we combined the projections (50,000 simulations) of the smoking-, obesity-, and alcohol-attributable mortality fractions to obtain 50,000 simulations of lifestyle-attributable mortality fractions (LAMF) (20–84) by again aggregating them multiplicatively (see Equation 2). That is, for b = 1, …, 50,000 simulations we applied Equation 2 to the matrices SAMF_x,t[b], OAMF_x,t[b], and AAMF_x,t[b].

Because the projections for smoking-, obesity-, and alcohol-attributable mortality are not available for the same years (for obesity from 2017 onwards and for smoking and alcohol in some cases only from 2015 onwards) (see Appendix 1—table 1), we adopted a slightly different procedure for the years 2015 and 2016. That is, for these years, we combined the available observed values for the specific lifestyle factor(s) with the projected values (50,000 simulations) for the specific lifestyle factor(s) for which we already had projections. In practice, we did so by turning the observed values into matrices so that we obtained equally large matrices everywhere.

As a next step, we extrapolated each of the 50,000 simulated age patterns of LAMF (20–84) to age 100 by applying Loess smoothing (for each year-simulation combination separately). More specifically, we applied Loess smoothing with degree = 2 and span = 0.75 on log(LAMF) for ages 20–84 and 130, where the LAMF at age 130 is set to 0.000001 so that a decline is enforced. To obtain LAMF for ages 85–100, we extrapolated using the fitted Loess curve, and subsequently took the exponent. This generated 50,000 simulated matrices of LAMF_x,t for x = 20–100 and t. For ages 0–19 and 101–130, we set the LAMF_x,t to zero.

Appendix 1—box 1.

Projections of smoking-, obesity-, and alcohol-attributable mortality (20–84).

For our estimate of future mortality that is attributable to smoking, obesity and alcohol, we used previously published data- and theory-driven projections of smoking-, obesity-, and alcohol-attributable mortality fractions (20–84) (Janssen et al., 2013; Janssen et al., 2020b; Janssen et al., 2020d). These projections follow the observed wave-shaped dynamic of the smoking epidemic (Lopez et al., 1994; Janssen, 2020) and the hypothesised wave shape of the obesity epidemic (Xu and Lam, 2018; Jaacks et al., 2019). For alcohol-attributable mortality, unrealistic future differences between countries were avoided by assuming that the current increases observed for selected countries will eventually turn into declines (Janssen et al., 2020d). All projections were based on the integration of these wave patterns into either a conventional age-period projection (for obesity prevalence, and consequently obesity-attributable mortality) or age-period-cohort projections (for smoking- and alcohol-attributable mortality). This was done by utilising the fact that a wave pattern occurs when the logistically transformed outcome measure has a quadratic shape.

The projections of obesity-attributable mortality were based on projections of obesity prevalence (Janssen et al., 2020b; Janssen et al., 2020a). For obesity prevalence, a wave shape was obtained by linearly extrapolating the speed of change ( = the first-order difference of the time trend) in the logit of prevalence, from a positive speed (increase) to a zero speed (maximum level of the quadratic shape) to a negative speed (decline). This is implemented by applying the age-period Lee–Carter mortality model to logistically transformed age-specific obesity prevalence data (1975–2016). Subsequently, the first-order difference of the general time trend k(t) is linearly extrapolated by applying a second-order random walk with negative drift. This linear extrapolation is applied from 2000 onwards in the majority of countries, but from 1980 onwards for Central and Eastern European women.

Smoking-attributable mortality fractions (1950–2014/LAY) and alcohol-attributable mortality fractions (1990–2016/LAY) are projected by integrating the wave pattern of epidemics into age-period-cohort mortality modelling and projection (Janssen et al., 2020c; Janssen et al., 2020d). That is, the age-period-cohort model by Cairns et al., 2009 is applied to the respective fractions using a generalised logit function as a link function. Subsequently, the (decelerating) increases in the period parameter (for smoking-attributable mortality among women, and for alcohol-attributable mortality in most Central and Eastern European countries and selected North-Western European countries) are projected by a quadratic curve with correlated errors to obtain future declines. Declines in the period parameter as well as the recent trend in the cohort parameter (after removing unreliable outer cohorts) are extrapolated using time-series forecasting (best ARIMA model under some restrictions).

In performing these projections, upper and lower bounds are implemented to ensure that the projections will not lead to unlikely crossovers between countries or between men and women, and will not lead to zero, in line with the theory (Lopez et al., 1994; Xu and Lam, 2018; Jaacks et al., 2019). More specifically, for obesity prevalence, the population- and age-specific prevalence in 1975 is assumed as the lower bound, except for Eastern European women (other than women in Estonia and Hungary), for whom the age-specific prevalence for men in 1975 is assumed. For alcohol-attributable mortality, different lower bounds by country group and sex were implemented based on their past trends and their past (peak) levels of age-standardised alcohol-attributable mortality fractions (AAMF). For the smoking-attributable mortality fractions (SAMF), a 5% smoking prevalence was assumed as the lower bound for both men and women. In addition, for women, the age-specific SAMF observed for Danish women in the year in which the all-age SAMF was highest is set as the upper bound.

To illustrate our projection of lifestyle-attributable mortality fractions, we calculated past and future all-age estimates of lifestyle-attributable mortality fractions. To avoid dependence on future death or population numbers, we calculated age-standardised mortality fractions by applying the sex- and country-specific age composition of all-cause mortality in 2010. The age-standardised lifestyle-attributable mortality fractions were calculated by directly applying the multiplicative aggregation to the past and future (median) age-standardised smoking-, obesity-, and alcohol-attributable mortality (ages 20–84).

Step (4) Combining the projected non-lifestyle attributable mortality rates with the projected lifestyle-attributable mortality fractions

In the final step of our projection approach, we combined the projected non-lifestyle mortality rates with the projected lifestyle-attributable mortality fraction to obtain our projected all-cause mortality rates. For this purpose, we extended the approach that Janssen et al., 2013 developed for combining the projection of non-smoking-attributable mortality with smoking-attributable mortality fractions.

We ended up with the following formula:

(4) mx,tallcause=mx,tnon-lifestyle∙11-LAMFx,t

Where mx,tallcause stands for age- and year-specific all-cause mortality rates, mx,tnon-lifestyle stands for age- and year-specific non-lifestyle-related mortality rates, and LAMFx,t stands for the age- and year-specific lifestyle-attributable mortality fraction.

We applied this formula to the 50,000 simulation matrices of coherently projected non-lifestyle-attributable mortality rates (0–130) and the 50,000 simulation matrices of projected LAMF (0–130) to obtain the 50,000 simulation matrices for all-cause-mortality based on our projection.

Additional projections

As mentioned, we compared our projection outcomes foremost with the outcomes of a Lee–Carter extrapolation (Lee and Carter, 1992) applied to all-cause mortality. In addition, however, we also performed additional projections, to illustrate the effects of the different steps in our projection approach. That is, we also performed a Lee–Carter projection of non-lifestyle-attributable mortality, a Li–Lee coherent projection of all-cause mortality, and a Li–Lee coherent projection of non-lifestyle-attributable mortality. In addition, we assessed the effect of purely incorporating lifestyle, without employing a coherent mortality projection, which constitutes of applying the Lee–Carter projection to non-lifestyle-attributable mortality and adding the estimated future lifestyle-attributable mortality. For these projections, we adopted similar specifications as for our projection approach, and as for the Lee–Carter extrapolation of all-cause mortality. Supplementary file 1C compares e0 in 2065 for all the different projections we employed.

Software

We used the R software for all elements of our analysis and projection. We implemented the Lee–Carter and Li–Lee models using the StMoMo package in R (Millossovich, 2018).

Data availability

Some of our original data regarding lifestyle-attributable mortality were based on previous publications, which, in turn, used data that are openly available. The all-cause mortality data and the exposure data can be obtained through the Human Mortality Database. We have provided source data files for all our tables and figures. These comprise the numerical data that are represented in the different figures, and the output on which the different tables are based. In addition, at the Open Science Framework (https://osf.io/ghu45/) we uploaded (i) one excel file with all the final numerical / output data that were used as input for the tables and figures, (ii) the underlying observed age-specific mortality rates (all-cause mortality, non-lifestyle-attributable mortality, lifestyle-attributable mortality) as well as the adjusted and projected age-specific mortality rates (medians and 90% and 95% projection intervals), and (iii) the different R codes used for the different steps of the projection.

The following data sets were generated
    1. Janssen F
    (2020) Open Science Framework
    ID ghu45. Future life expectancy in Europe taking into account the impact of smoking, obesity and alcohol.
    https://osf.io/ghu45/

References

    1. Abarca-Gómez L
    2. Abdeen ZA
    3. Hamid ZA
    4. Abu-Rmeileh NM
    5. Acosta-Cazares B
    6. Acuin C
    7. Adams RJ
    8. Aekplakorn W
    9. Afsana K
    10. Aguilar-Salinas CA
    11. Agyemang C
    12. Ahmadvand A
    13. Ahrens W
    14. Ajlouni K
    15. Akhtaeva N
    16. Al-Hazzaa HM
    17. Al-Othman AR
    18. Al-Raddadi R
    19. Al Buhairan F
    20. Al Dhukair S
    21. Ali MM
    22. Ali O
    23. Alkerwi A
    24. Alvarez-Pedrerol M
    25. Aly E
    26. Amarapurkar DN
    27. Amouyel P
    28. Amuzu A
    29. Andersen LB
    30. Anderssen SA
    31. Andrade DS
    32. Ängquist LH
    33. Anjana RM
    34. Aounallah-Skhiri H
    35. Araújo J
    36. Ariansen I
    37. Aris T
    38. Arlappa N
    39. Arveiler D
    40. Aryal KK
    41. Aspelund T
    42. Assah FK
    43. Assunção MCF
    44. Aung MS
    45. Avdicová M
    46. Azevedo A
    47. Azizi F
    48. Babu BV
    49. Bahijri S
    50. Baker JL
    51. Balakrishna N
    52. Bamoshmoosh M
    53. Banach M
    54. Bandosz P
    55. Banegas JR
    56. Barbagallo CM
    57. Barceló A
    58. Barkat A
    59. Barros AJD
    60. Barros MVG
    61. Bata I
    62. Batieha AM
    63. Batista RL
    64. Batyrbek A
    65. Baur LA
    66. Beaglehole R
    67. Romdhane HB
    68. Benedics J
    69. Benet M
    70. Bennett JE
    71. Bernabe-Ortiz A
    72. Bernotiene G
    73. Bettiol H
    74. Bhagyalaxmi A
    75. Bharadwaj S
    76. Bhargava SK
    77. Bhatti Z
    78. Bhutta ZA
    79. Bi H
    80. Bi Y
    81. Biehl A
    82. Bikbov M
    83. Bista B
    84. Bjelica DJ
    85. Bjerregaard P
    86. Bjertness E
    87. Bjertness MB
    88. Björkelund C
    89. Blokstra A
    90. Bo S
    91. Bobak M
    92. Boddy LM
    93. Boehm BO
    94. Boeing H
    95. Boggia JG
    96. Boissonnet CP
    97. Bonaccio M
    98. Bongard V
    99. Bovet P
    100. Braeckevelt L
    101. Braeckman L
    102. Bragt MCE
    103. Brajkovich I
    104. Branca F
    105. Breckenkamp J
    106. Breda J
    107. Brenner H
    108. Brewster LM
    109. Brian GR
    110. Brinduse L
    111. Bruno G
    112. Bueno-de-Mesquita HB
    113. Bugge A
    114. Buoncristiano M
    115. Burazeri G
    116. Burns C
    117. de León AC
    118. Cacciottolo J
    119. Cai H
    120. Cama T
    121. Cameron C
    122. Camolas J
    123. Can G
    124. Cândido APC
    125. Capanzana M
    126. Capuano V
    127. Cardoso VC
    128. Carlsson AC
    129. Carvalho MJ
    130. Casanueva FF
    131. Casas J-P
    132. Caserta CA
    133. Chamukuttan S
    134. Chan AW
    135. Chan Q
    136. Chaturvedi HK
    137. Chaturvedi N
    138. Chen C-J
    139. Chen F
    140. Chen H
    141. Chen S
    142. Chen Z
    143. Cheng C-Y
    144. Chetrit A
    145. Chikova-Iscener E
    146. Chiolero A
    147. Chiou S-T
    148. Chirita-Emandi A
    149. Chirlaque M-D
    150. Cho B
    151. Cho Y
    152. Christensen K
    153. Christofaro DG
    154. Chudek J
    155. Cifkova R
    156. Cinteza E
    157. Claessens F
    158. Clays E
    159. Concin H
    160. Confortin SC
    161. Cooper C
    162. Cooper R
    163. Coppinger TC
    164. Costanzo S
    165. Cottel D
    166. Cowell C
    167. Craig CL
    168. Crujeiras AB
    169. Cucu A
    170. D'Arrigo G
    171. d'Orsi E
    172. Dallongeville J
    173. Damasceno A
    174. Damsgaard CT
    175. Danaei G
    176. Dankner R
    177. Dantoft TM
    178. Dastgiri S
    179. Dauchet L
    180. Davletov K
    181. De Backer G
    182. De Bacquer D
    183. De Curtis A
    184. de Gaetano G
    185. De Henauw S
    186. de Oliveira PD
    187. De Ridder K
    188. De Smedt D
    189. Deepa M
    190. Deev AD
    191. Dehghan A
    192. Delisle H
    193. Delpeuch F
    194. Deschamps V
    195. Dhana K
    196. Di Castelnuovo AF
    197. Dias-da-Costa JS
    198. Diaz A
    199. Dika Z
    200. Djalalinia S
    201. Do HTP
    202. Dobson AJ
    203. Donati MB
    204. Donfrancesco C
    205. Donoso SP
    206. Döring A
    207. Dorobantu M
    208. Dorosty AR
    209. Doua K
    210. Drygas W
    211. Duan JL
    212. Duante C
    213. Duleva V
    214. Dulskiene V
    215. Dzerve V
    216. Dziankowska-Zaborszczyk E
    217. Egbagbe EE
    218. Eggertsen R
    219. Eiben G
    220. Ekelund U
    221. El Ati J
    222. Elliott P
    223. Engle-Stone R
    224. Erasmus RT
    225. Erem C
    226. Eriksen L
    227. Eriksson JG
    228. la Peña JE
    229. Evans A
    230. Faeh D
    231. Fall CH
    232. Sant'Angelo VF
    233. Farzadfar F
    234. Felix-Redondo FJ
    235. Ferguson TS
    236. Fernandes RA
    237. Fernández-Bergés D
    238. Ferrante D
    239. Ferrari M
    240. Ferreccio C
    241. Ferrieres J
    242. Finn JD
    243. Fischer K
    244. Flores EM
    245. Föger B
    246. Foo LH
    247. Forslund A-S
    248. Forsner M
    249. Fouad HM
    250. Francis DK
    251. Franco MdoC
    252. Franco OH
    253. Frontera G
    254. Fuchs FD
    255. Fuchs SC
    256. Fujita Y
    257. Furusawa T
    258. Gaciong Z
    259. Gafencu M
    260. Galeone D
    261. Galvano F
    262. Garcia-de-la-Hera M
    263. Gareta D
    264. Garnett SP
    265. Gaspoz J-M
    266. Gasull M
    267. Gates L
    268. Geiger H
    269. Geleijnse JM
    270. Ghasemian A
    271. Giampaoli S
    272. Gianfagna F
    273. Gill TK
    274. Giovannelli J
    275. Giwercman A
    276. Godos J
    277. Gogen S
    278. Goldsmith RA
    279. Goltzman D
    280. Gonçalves H
    281. González-Leon M
    282. González-Rivas JP
    283. Gonzalez-Gross M
    284. Gottrand F
    285. Graça AP
    286. Graff-Iversen S
    287. Grafnetter D
    288. Grajda A
    289. Grammatikopoulou MG
    290. Gregor RD
    291. Grodzicki T
    292. Grøntved A
    293. Grosso G
    294. Gruden G
    295. Grujic V
    296. Gu D
    297. Gualdi-Russo E
    298. Guallar-Castillón P
    299. Guan OP
    300. Gudmundsson EF
    301. Gudnason V
    302. Guerrero R
    303. Guessous I
    304. Guimaraes AL
    305. Gulliford MC
    306. Gunnlaugsdottir J
    307. Gunter M
    308. Guo X
    309. Guo Y
    310. Gupta PC
    311. Gupta R
    312. Gureje O
    313. Gurzkowska B
    314. Gutierrez L
    315. Gutzwiller F
    316. Hadaegh F
    317. Hadjigeorgiou CA
    318. Si-Ramlee K
    319. Halkjær J
    320. Hambleton IR
    321. Hardy R
    322. Kumar RH
    323. Hassapidou M
    324. Hata J
    325. Hayes AJ
    326. He J
    327. Heidinger-Felso R
    328. Heinen M
    329. Hendriks ME
    330. Henriques A
    331. Cadena LH
    332. Herrala S
    333. Herrera VM
    334. Herter-Aeberli I
    335. Heshmat R
    336. Hihtaniemi IT
    337. Ho SY
    338. Ho SC
    339. Hobbs M
    340. Hofman A
    341. Hopman WM
    342. Horimoto ARVR
    343. Hormiga CM
    344. Horta BL
    345. Houti L
    346. Howitt C
    347. Htay TT
    348. Htet AS
    349. Htike MMT
    350. Hu Y
    351. Huerta JM
    352. Petrescu CH
    353. Huisman M
    354. Husseini A
    355. Huu CN
    356. Huybrechts I
    357. Hwalla N
    358. Hyska J
    359. Iacoviello L
    360. Iannone AG
    361. Ibarluzea JM
    362. Ibrahim MM
    363. Ikeda N
    364. Ikram MA
    365. Irazola VE
    366. Islam M
    367. Ismail A-S
    368. Ivkovic V
    369. Iwasaki M
    370. Jackson RT
    371. Jacobs JM
    372. Jaddou H
    373. Jafar T
    374. Jamil KM
    375. Jamrozik K
    376. Janszky I
    377. Jarani J
    378. Jasienska G
    379. Jelakovic A
    380. Jelakovic B
    381. Jennings G
    382. Jeong S-L
    383. Jiang CQ
    384. Jiménez-Acosta SM
    385. Joffres M
    386. Johansson M
    387. Jonas JB
    388. Jørgensen T
    389. Joshi P
    390. Jovic DP
    391. Józwiak J
    392. Juolevi A
    393. Jurak G
    394. Jureša V
    395. Kaaks R
    396. Kafatos A
    397. Kajantie EO
    398. Kalter-Leibovici O
    399. Kamaruddin NA
    400. Kapantais E
    401. Karki KB
    402. Kasaeian A
    403. Katz J
    404. Kauhanen J
    405. Kaur P
    406. Kavousi M
    407. Kazakbaeva G
    408. Keil U
    409. Boker LK
    410. Keinänen-Kiukaanniemi S
    411. Kelishadi R
    412. Kelleher C
    413. Kemper HCG
    414. Kengne AP
    415. Kerimkulova A
    416. Kersting M
    417. Key T
    418. Khader YS
    419. Khalili D
    420. Khang Y-H
    421. Khateeb M
    422. Khaw K-T
    423. Khouw IMSL
    424. Kiechl-Kohlendorfer U
    425. Kiechl S
    426. Killewo J
    427. Kim J
    428. Kim Y-Y
    429. Klimont J
    430. Klumbiene J
    431. Knoflach M
    432. Koirala B
    433. Kolle E
    434. Kolsteren P
    435. Korrovits P
    436. Kos J
    437. Koskinen S
    438. Kouda K
    439. Kovacs VA
    440. Kowlessur S
    441. Koziel S
    442. Kratzer W
    443. Kriemler S
    444. Kristensen PL
    445. Krokstad S
    446. Kromhout D
    447. Kruger HS
    448. Kubinova R
    449. Kuciene R
    450. Kuh D
    451. Kujala UM
    452. Kulaga Z
    453. Kumar RK
    454. Kunešová M
    455. Kurjata P
    456. Kusuma YS
    457. Kuulasmaa K
    458. Kyobutungi C
    459. La QN
    460. Laamiri FZ
    461. Laatikainen T
    462. Lachat C
    463. Laid Y
    464. Lam TH
    465. Landrove O
    466. Lanska V
    467. Lappas G
    468. Larijani B
    469. Laugsand LE
    470. Lauria L
    471. Laxmaiah A
    472. Bao KLN
    473. Le TD
    474. Lebanan MAO
    475. Leclercq C
    476. Lee J
    477. Lee J
    478. Lehtimäki T
    479. León-Muñoz LM
    480. Levitt NS
    481. Li Y
    482. Lilly CL
    483. Lim W-Y
    484. Lima-Costa MF
    485. Lin H-H
    486. Lin X
    487. Lind L
    488. Linneberg A
    489. Lissner L
    490. Litwin M
    491. Liu J
    492. Loit H-M
    493. Lopes L
    494. Lorbeer R
    495. Lotufo PA
    496. Lozano JE
    497. Luksiene D
    498. Lundqvist A
    499. Lunet N
    500. Lytsy P
    501. Ma G
    502. Ma J
    503. Machado-Coelho GLL
    504. Machado-Rodrigues AM
    505. Machi S
    506. Maggi S
    507. Magliano DJ
    508. Magriplis E
    509. Mahaletchumy A
    510. Maire B
    511. Majer M
    512. Makdisse M
    513. Malekzadeh R
    514. Malhotra R
    515. Rao KM
    516. Malyutina S
    517. Manios Y
    518. Mann JI
    519. Manzato E
    520. Margozzini P
    521. Markaki A
    522. Markey O
    523. Marques LP
    524. Marques-Vidal P
    525. Marrugat J
    526. Martin-Prevel Y
    527. Martin R
    528. Martorell R
    529. Martos E
    530. Marventano S
    531. Masoodi SR
    532. Mathiesen EB
    533. Matijasevich A
    534. Matsha TE
    535. Mazur A
    536. Mbanya JCN
    537. McFarlane SR
    538. McGarvey ST
    539. McKee M
    540. McLachlan S
    541. McLean RM
    542. McLean SB
    543. McNulty BA
    544. Yusof SM
    545. Mediene-Benchekor S
    546. Medzioniene J
    547. Meirhaeghe A
    548. Meisfjord J
    549. Meisinger C
    550. Menezes AMB
    551. Menon GR
    552. Mensink GBM
    553. Meshram II
    554. Metspalu A
    555. Meyer HE
    556. Mi J
    557. Michaelsen KF
    558. Michels N
    559. Mikkel K
    560. Miller JC
    561. Minderico CS
    562. Miquel JF
    563. Miranda JJ
    564. Mirkopoulou D
    565. Mirrakhimov E
    566. Mišigoj-Durakovic M
    567. Mistretta A
    568. Mocanu V
    569. Modesti PA
    570. Mohamed MK
    571. Mohammad K
    572. Mohammadifard N
    573. Mohan V
    574. Mohanna S
    575. Yusoff MFM
    576. Molbo D
    577. Møllehave LT
    578. Møller NC
    579. Molnár D
    580. Momenan A
    581. Mondo CK
    582. Monterrubio EA
    583. Monyeki KDK
    584. Moon JS
    585. Moreira LB
    586. Morejon A
    587. Moreno LA
    588. Morgan K
    589. Mortensen EL
    590. Moschonis G
    591. Mossakowska M
    592. Mostafa A
    593. Mota J
    594. Mota-Pinto A
    595. Motlagh ME
    596. Motta J
    597. Mu TT
    598. Muc M
    599. Muiesan ML
    600. Müller-Nurasyid M
    601. Murphy N
    602. Mursu J
    603. Murtagh EM
    604. Musil V
    605. Nabipour I
    606. Nagel G
    607. Naidu BM
    608. Nakamura H
    609. Námešná J
    610. Nang EEK
    611. Nangia VB
    612. Nankap M
    613. Narake S
    614. Nardone P
    615. Navarrete-Muñoz EM
    616. Neal WA
    617. Nenko I
    618. Neovius M
    619. Nervi F
    620. Nguyen CT
    621. Nguyen ND
    622. Nguyen QN
    623. Nieto-Martínez RE
    624. Ning G
    625. Ninomiya T
    626. Nishtar S
    627. Noale M
    628. Noboa OA
    629. Norat T
    630. Norie S
    631. Noto D
    632. Nsour MA
    633. O'Reilly D
    634. Obreja G
    635. Oda E
    636. Oehlers G
    637. Oh K
    638. Ohara K
    639. Olafsson Örn
    640. Olinto MTA
    641. Oliveira IO
    642. Oltarzewski M
    643. Omar MA
    644. Onat A
    645. Ong SK
    646. Ono LM
    647. Ordunez P
    648. Ornelas R
    649. Ortiz AP
    650. Osler M
    651. Osmond C
    652. Ostojic SM
    653. Ostovar A
    654. Otero JA
    655. Overvad K
    656. Owusu-Dabo E
    657. Paccaud FM
    658. Padez C
    659. Pahomova E
    660. Pajak A
    661. Palli D
    662. Palloni A
    663. Palmieri L
    664. Pan W-H
    665. Panda-Jonas S
    666. Pandey A
    667. Panza F
    668. Papandreou D
    669. Park S-W
    670. Parnell WR
    671. Parsaeian M
    672. Pascanu IM
    673. Patel ND
    674. Pecin I
    675. Pednekar MS
    676. Peer N
    677. Peeters PH
    678. Peixoto SV
    679. Peltonen M
    680. Pereira AC
    681. Perez-Farinos N
    682. Pérez CM
    683. Peters A
    684. Petkeviciene J
    685. Petrauskiene A
    686. Peykari N
    687. Pham ST
    688. Pierannunzio D
    689. Pigeot I
    690. Pikhart H
    691. Pilav A
    692. Pilotto L
    693. Pistelli F
    694. Pitakaka F
    695. Piwonska A
    696. Plans-Rubió P
    697. Poh BK
    698. Pohlabeln H
    699. Pop RM
    700. Popovic SR
    701. Porta M
    702. Portegies MLP
    703. Posch G
    704. Poulimeneas D
    705. Pouraram H
    706. Pourshams A
    707. Poustchi H
    708. Pradeepa R
    709. Prashant M
    710. Price JF
    711. Puder JJ
    712. Pudule I
    713. Puiu M
    714. Punab M
    715. Qasrawi RF
    716. Qorbani M
    717. Bao TQ
    718. Radic I
    719. Radisauskas R
    720. Rahman M
    721. Rahman M
    722. Raitakari O
    723. Raj M
    724. Rao SR
    725. Ramachandran A
    726. Ramke J
    727. Ramos E
    728. Ramos R
    729. Rampal L
    730. Rampal S
    731. Rascon-Pacheco RA
    732. Redon J
    733. Reganit PFM
    734. Ribas-Barba L
    735. Ribeiro R
    736. Riboli E
    737. Rigo F
    738. de Wit TFR
    739. Rito A
    740. Ritti-Dias RM
    741. Rivera JA
    742. Robinson SM
    743. Robitaille C
    744. Rodrigues D
    745. Rodríguez-Artalejo F
    746. del Cristo Rodriguez-Perez M
    747. Rodríguez-Villamizar LA
    748. Rojas-Martinez R
    749. Rojroongwasinkul N
    750. Romaguera D
    751. Ronkainen K
    752. Rosengren A
    753. Rouse I
    754. Roy JGR
    755. Rubinstein A
    756. Rühli FJ
    757. Ruiz-Betancourt BS
    758. Russo P
    759. Rutkowski M
    760. Sabanayagam C
    761. Sachdev HS
    762. Saidi O
    763. Salanave B
    764. Martinez ES
    765. Salmerón D
    766. Salomaa V
    767. Salonen JT
    768. Salvetti M
    769. Sánchez-Abanto J
    770. Sandjaja S
    771. Sans S
    772. Marina LS
    773. Santos DA
    774. Santos IS
    775. Santos O
    776. dos Santos RN
    777. Santos R
    778. Saramies JL
    779. Sardinha LB
    780. Sarrafzadegan N
    781. Saum K-U
    782. Savva S
    783. Savy M
    784. Scazufca M
    785. Rosario AS
    786. Schargrodsky H
    787. Schienkiewitz A
    788. Schipf S
    789. Schmidt CO
    790. Schmidt IM
    791. Schultsz C
    792. Schutte AE
    793. Sein AA
    794. Sen A
    795. Senbanjo IO
    796. Sepanlou SG
    797. Serra-Majem L
    798. Shalnova SA
    799. Sharma SK
    800. Shaw JE
    801. Shibuya K
    802. Shin DW
    803. Shin Y
    804. Shiri R
    805. Siani A
    806. Siantar R
    807. Sibai AM
    808. Silva AM
    809. Silva DAS
    810. Simon M
    811. Simons J
    812. Simons LA
    813. Sjöberg A
    814. Sjöström M
    815. Skovbjerg S
    816. Slowikowska-Hilczer J
    817. Slusarczyk P
    818. Smeeth L
    819. Smith MC
    820. Snijder MB
    821. So H-K
    822. Sobngwi E
    823. Söderberg S
    824. Soekatri MYE
    825. Solfrizzi V
    826. Sonestedt E
    827. Song Y
    828. Sørensen TIA
    829. Soric M
    830. Jérome CS
    831. Soumare A
    832. Spinelli A
    833. Spiroski I
    834. Staessen JA
    835. Stamm H
    836. Starc G
    837. Stathopoulou MG
    838. Staub K
    839. Stavreski B
    840. Steene-Johannessen J
    841. Stehle P
    842. Stein AD
    843. Stergiou GS
    844. Stessman J
    845. Stieber J
    846. Stöckl D
    847. Stocks T
    848. Stokwiszewski J
    849. Stratton G
    850. Stronks K
    851. Strufaldi MW
    852. Suárez-Medina R
    853. Sun C-A
    854. Sundström J
    855. Sung Y-T
    856. Sunyer J
    857. Suriyawongpaisal P
    858. Swinburn BA
    859. Sy RG
    860. Szponar L
    861. Tai ES
    862. Tammesoo M-L
    863. Tamosiunas A
    864. Tan EJ
    865. Tang X
    866. Tanser F
    867. Tao Y
    868. Tarawneh MR
    869. Tarp J
    870. Tarqui-Mamani CB
    871. Tautu O-F
    872. Braunerová RT
    873. Taylor A
    874. Tchibindat F
    875. Theobald H
    876. Theodoridis X
    877. Thijs L
    878. Thuesen BH
    879. Tjonneland A
    880. Tolonen HK
    881. Tolstrup JS
    882. Topbas M
    883. Topór-Madry R
    884. Tormo MJ
    885. Tornaritis MJ
    886. Torrent M
    887. Toselli S
    888. Traissac P
    889. Trichopoulos D
    890. Trichopoulou A
    891. Trinh OTH
    892. Trivedi A
    893. Tshepo L
    894. Tsigga M
    895. Tsugane S
    896. Tulloch-Reid MK
    897. Tullu F
    898. Tuomainen T-P
    899. Tuomilehto J
    900. Turley ML
    901. Tynelius P
    902. Tzotzas T
    903. Tzourio C
    904. Ueda P
    905. Ugel EE
    906. Ukoli FAM
    907. Ulmer H
    908. Unal B
    909. Uusitalo HMT
    910. Valdivia G
    911. Vale S
    912. Valvi D
    913. van der Schouw YT
    914. Van Herck K
    915. Van Minh H
    916. van Rossem L
    917. Van Schoor NM
    918. van Valkengoed IGM
    919. Vanderschueren D
    920. Vanuzzo D
    921. Vatten L
    922. Vega T
    923. Veidebaum T
    924. Velasquez-Melendez G
    925. Velika B
    926. Veronesi G
    927. Verschuren WMM
    928. Victora CG
    929. Viegi G
    930. Viet L
    931. Viikari-Juntura E
    932. Vineis P
    933. Vioque J
    934. Virtanen JK
    935. Visvikis-Siest S
    936. Viswanathan B
    937. Vlasoff T
    938. Vollenweider P
    939. Völzke H
    940. Voutilainen S
    941. Vrijheid M
    942. Wade AN
    943. Wagner A
    944. Waldhör T
    945. Walton J
    946. Bebakar WMW
    947. Mohamud WNW
    948. Wanderley RS
    949. Wang M-D
    950. Wang Q
    951. Wang YX
    952. Wang Y-W
    953. Wannamethee SG
    954. Wareham N
    955. Weber A
    956. Wedderkopp N
    957. Weerasekera D
    958. Whincup PH
    959. Widhalm K
    960. Widyahening IS
    961. Wiecek A
    962. Wijga AH
    963. Wilks RJ
    964. Willeit J
    965. Willeit P
    966. Wilsgaard T
    967. Wojtyniak B
    968. Wong-McClure RA
    969. Wong JYY
    970. Wong JE
    971. Wong TY
    972. Woo J
    973. Woodward M
    974. Wu FC
    975. Wu J
    976. Wu S
    977. Xu H
    978. Xu L
    979. Yamborisut U
    980. Yan W
    981. Yang X
    982. Yardim N
    983. Ye X
    984. Yiallouros PK
    985. Yngve A
    986. Yoshihara A
    987. You QS
    988. Younger-Coleman NO
    989. Yusoff F
    990. Yusoff MFM
    991. Zaccagni L
    992. Zafiropulos V
    993. Zainuddin AA
    994. Zambon S
    995. Zampelas A
    996. Zamrazilová H
    997. Zdrojewski T
    998. Zeng Y
    999. Zhao D
    1000. Zhao W
    1001. Zheng W
    1002. Zheng Y
    1003. Zholdin B
    1004. Zhou M
    1005. Zhu D
    1006. Zhussupov B
    1007. Zimmermann E
    1008. Cisneros JZ
    1009. Bentham J
    1010. Di Cesare M
    1011. Bilano V
    1012. Bixby H
    1013. Zhou B
    1014. Stevens GA
    1015. Riley LM
    1016. Taddei C
    1017. Hajifathalian K
    1018. Lu Y
    1019. Savin S
    1020. Cowan MJ
    1021. Paciorek CJ
    1022. Chirita-Emandi A
    1023. Hayes AJ
    1024. Katz J
    1025. Kelishadi R
    1026. Kengne AP
    1027. Khang Y-H
    1028. Laxmaiah A
    1029. Li Y
    1030. Ma J
    1031. Miranda JJ
    1032. Mostafa A
    1033. Neovius M
    1034. Padez C
    1035. Rampal L
    1036. Zhu A
    1037. Bennett JE
    1038. Danaei G
    1039. Bhutta ZA
    1040. Ezzati M
    (2017) Worldwide trends in body-mass index, underweight, overweight, and obesity from 1975 to 2016: a pooled analysis of 2416 population-based measurement studies in 128·9 million children, adolescents, and adults
    The Lancet 390:2627–2642.
    https://doi.org/10.1016/S0140-6736(17)32129-3
  15. Report
    1. European Commission
    (2009)
    Dealing with the Impact of an Ageing Population in the EU (2009 Ageing Report)
    Brussels: Commission of the European Communities.
  17. Report
    1. Eurostat
    (2020b)
    Methodology of the Eurostat population projections 2019-based (EUROPOP2019)
    European Commission/Eurostat.
  22. Report
    1. GBD
    (2018) Global Burden of Disease Study 2017 - Results
    Seattle, United States: Institute for health metrics and evaluation.
    http://ghdx.healthdata.org/gbd-2017
  24. Website
    1. Human Cause-of-Death Database
    (2017) Human Cause-of death database (HCD)
    French Institute for Demographic Studies (France) and Max Planck Institute for Demographic Research (Germany) . Accessed June 30, 2017.
    https://www.causesofdeath.org/cgi-bin/main.php
  38. Conference
    1. Lanzieri G
    (2009)
    EUROPOP2008: a set of population projections for the European Union
    Conference: XXVI IUSSP International Population Conference.
  50. Conference
    1. Murphy M
    2. Luy M
    3. Torrisi O
    (2019)
    Stalling of mortality in the United Kingdom and Europe: an analytical review of the evidence
    LSE Social Policy Working Paper. pp. 11–19.
  51. Book
    1. OECD
    (2016) Fragmentation of retirement markets due to differences in life expectancy
    In: Ayuso M, editors. OECD Business and Finance Outlook. Paris: OECD Publishing. pp. 177–205.
    https://doi.org/10.1787/9789264257573-en
    1. Omran AR
    (1998)
    The Epidemiologic transition theory revisited thirty years later
    World Health Statistics Quarterly 52:99–119.
  57. Book
    1. Preston SH
    2. Heuveline P
    3. Guillot M
    (2001)
    Demography - Measuring and Modeling Population Processes
    Oxford: Blackwell Publishers.
  59. Book
    1. Raleigh V
    (2019)
    Trends in Life Expectancy in EU and Other OECD Countries: Why Are Improvements Slowing?
    Paris: OECD.
    1. Stanaway JD
    2. Afshin A
    3. Gakidou E
    4. Lim SS
    5. Abate D
    6. Abate KH
    7. Abbafati C
    8. Abbasi N
    9. Abbastabar H
    10. Abd-Allah F
    11. Abdela J
    12. Abdelalim A
    13. Abdollahpour I
    14. Abdulkader RS
    15. Abebe M
    16. Abebe Z
    17. Abera SF
    18. Abil OZ
    19. Abraha HN
    20. Abrham AR
    21. Abu-Raddad LJ
    22. Abu-Rmeileh NME
    23. Accrombessi MMK
    24. Acharya D
    25. Acharya P
    26. Adamu AA
    27. Adane AA
    28. Adebayo OM
    29. Adedoyin RA
    30. Adekanmbi V
    31. Ademi Z
    32. Adetokunboh OO
    33. Adib MG
    34. Admasie A
    35. Adsuar JC
    36. Afanvi KA
    37. Afarideh M
    38. Agarwal G
    39. Aggarwal A
    40. Aghayan SA
    41. Agrawal A
    42. Agrawal S
    43. Ahmadi A
    44. Ahmadi M
    45. Ahmadieh H
    46. Ahmed MB
    47. Aichour AN
    48. Aichour I
    49. Aichour MTE
    50. Akbari ME
    51. Akinyemiju T
    52. Akseer N
    53. Al-Aly Z
    54. Al-Eyadhy A
    55. Al-Mekhlafi HM
    56. Alahdab F
    57. Alam K
    58. Alam S
    59. Alam T
    60. Alashi A
    61. Alavian SM
    62. Alene KA
    63. Ali K
    64. Ali SM
    65. Alijanzadeh M
    66. Alizadeh-Navaei R
    67. Aljunid SM
    68. Alkerwi A
    69. Alla F
    70. Alsharif U
    71. Altirkawi K
    72. Alvis-Guzman N
    73. Amare AT
    74. Ammar W
    75. Anber NH
    76. Anderson JA
    77. Andrei CL
    78. Androudi S
    79. Animut MD
    80. Anjomshoa M
    81. Ansha MG
    82. Antó JM
    83. Antonio CAT
    84. Anwari P
    85. Appiah LT
    86. Appiah SCY
    87. Arabloo J
    88. Aremu O
    89. Ärnlöv J
    90. Artaman A
    91. Aryal KK
    92. Asayesh H
    93. Ataro Z
    94. Ausloos M
    95. Avokpaho EFGA
    96. Awasthi A
    97. Ayala Quintanilla BP
    98. Ayer R
    99. Ayuk TB
    100. Azzopardi PS
    101. Babazadeh A
    102. Badali H
    103. Badawi A
    104. Balakrishnan K
    105. Bali AG
    106. Ball K
    107. Ballew SH
    108. Banach M
    109. Banoub JAM
    110. Barac A
    111. Barker-Collo SL
    112. Bärnighausen TW
    113. Barrero LH
    114. Basu S
    115. Baune BT
    116. Bazargan-Hejazi S
    117. Bedi N
    118. Beghi E
    119. Behzadifar M
    120. Behzadifar M
    121. Béjot Y
    122. Bekele BB
    123. Bekru ET
    124. Belay E
    125. Belay YA
    126. Bell ML
    127. Bello AK
    128. Bennett DA
    129. Bensenor IM
    130. Bergeron G
    131. Berhane A
    132. Bernabe E
    133. Bernstein RS
    134. Beuran M
    135. Beyranvand T
    136. Bhala N
    137. Bhalla A
    138. Bhattarai S
    139. Bhutta ZA
    140. Biadgo B
    141. Bijani A
    142. Bikbov B
    143. Bilano V
    144. Bililign N
    145. Bin Sayeed MS
    146. Bisanzio D
    147. Biswas T
    148. Bjørge T
    149. Blacker BF
    150. Bleyer A
    151. Borschmann R
    152. Bou-Orm IR
    153. Boufous S
    154. Bourne R
    155. Brady OJ
    156. Brauer M
    157. Brazinova A
    158. Breitborde NJK
    159. Brenner H
    160. Briko AN
    161. Britton G
    162. Brugha T
    163. Buchbinder R
    164. Burnett RT
    165. Busse R
    166. Butt ZA
    167. Cahill LE
    168. Cahuana-Hurtado L
    169. Campos-Nonato IR
    170. Cárdenas R
    171. Carreras G
    172. Carrero JJ
    173. Carvalho F
    174. Castañeda-Orjuela CA
    175. Castillo Rivas J
    176. Castro F
    177. Catalá-López F
    178. Causey K
    179. Cercy KM
    180. Cerin E
    181. Chaiah Y
    182. Chang H-Y
    183. Chang J-C
    184. Chang K-L
    185. Charlson FJ
    186. Chattopadhyay A
    187. Chattu VK
    188. Chee ML
    189. Cheng C-Y
    190. Chew A
    191. Chiang PP-C
    192. Chimed-Ochir O
    193. Chin KL
    194. Chitheer A
    195. Choi J-YJ
    196. Chowdhury R
    197. Christensen H
    198. Christopher DJ
    199. Chung S-C
    200. Cicuttini FM
    201. Cirillo M
    202. Cohen AJ
    203. Collado-Mateo D
    204. Cooper C
    205. Cooper OR
    206. Coresh J
    207. Cornaby L
    208. Cortesi PA
    209. Cortinovis M
    210. Costa M
    211. Cousin E
    212. Criqui MH
    213. Cromwell EA
    214. Cundiff DK
    215. Daba AK
    216. Dachew BA
    217. Dadi AF
    218. Damasceno AAM
    219. Dandona L
    220. Dandona R
    221. Darby SC
    222. Dargan PI
    223. Daryani A
    224. Das Gupta R
    225. Das Neves J
    226. Dasa TT
    227. Dash AP
    228. Davitoiu DV
    229. Davletov K
    230. De la Cruz-Góngora V
    231. De La Hoz FP
    232. De Leo D
    233. De Neve J-W
    234. Degenhardt L
    235. Deiparine S
    236. Dellavalle RP
    237. Demoz GT
    238. Denova-Gutiérrez E
    239. Deribe K
    240. Dervenis N
    241. Deshpande A
    242. Des Jarlais DC
    243. Dessie GA
    244. Deveber GA
    245. Dey S
    246. Dharmaratne SD
    247. Dhimal M
    248. Dinberu MT
    249. Ding EL
    250. Diro HD
    251. Djalalinia S
    252. Do HP
    253. Dokova K
    254. Doku DT
    255. Doyle KE
    256. Driscoll TR
    257. Dubey M
    258. Dubljanin E
    259. Duken EE
    260. Duncan BB
    261. Duraes AR
    262. Ebert N
    263. Ebrahimi H
    264. Ebrahimpour S
    265. Edvardsson D
    266. Effiong A
    267. Eggen AE
    268. El Bcheraoui C
    269. El-Khatib Z
    270. Elyazar IR
    271. Enayati A
    272. Endries AY
    273. Er B
    274. Erskine HE
    275. Eskandarieh S
    276. Esteghamati A
    277. Estep K
    278. Fakhim H
    279. Faramarzi M
    280. Fareed M
    281. Farid TA
    282. Farinha CSE
    283. Farioli A
    284. Faro A
    285. Farvid MS
    286. Farzaei MH
    287. Fatima B
    288. Fay KA
    289. Fazaeli AA
    290. Feigin VL
    291. Feigl AB
    292. Fereshtehnejad S-M
    293. Fernandes E
    294. Fernandes JC
    295. Ferrara G
    296. Ferrari AJ
    297. Ferreira ML
    298. Filip I
    299. Finger JD
    300. Fischer F
    301. Foigt NA
    302. Foreman KJ
    303. Fukumoto T
    304. Fullman N
    305. Fürst T
    306. Furtado JM
    307. Futran ND
    308. Gall S
    309. Gallus S
    310. Gamkrelidze A
    311. Ganji M
    312. Garcia-Basteiro AL
    313. Gardner WM
    314. Gebre AK
    315. Gebremedhin AT
    316. Gebremichael TG
    317. Gelano TF
    318. Geleijnse JM
    319. Geramo YCD
    320. Gething PW
    321. Gezae KE
    322. Ghadimi R
    323. Ghadiri K
    324. Ghasemi Falavarjani K
    325. Ghasemi-Kasman M
    326. Ghimire M
    327. Ghosh R
    328. Ghoshal AG
    329. Giampaoli S
    330. Gill PS
    331. Gill TK
    332. Gillum RF
    333. Ginawi IA
    334. Giussani G
    335. Gnedovskaya EV
    336. Godwin WW
    337. Goli S
    338. Gómez-Dantés H
    339. Gona PN
    340. Gopalani SV
    341. Goulart AC
    342. Grada A
    343. Grams ME
    344. Grosso G
    345. Gugnani HC
    346. Guo Y
    347. Gupta R
    348. Gupta R
    349. Gupta T
    350. Gutiérrez RA
    351. Gutiérrez-Torres DS
    352. Haagsma JA
    353. Habtewold TD
    354. Hachinski V
    355. Hafezi-Nejad N
    356. Hagos TB
    357. Hailegiyorgis TT
    358. Hailu GB
    359. Haj-Mirzaian A
    360. Haj-Mirzaian A
    361. Hamadeh RR
    362. Hamidi S
    363. Handal AJ
    364. Hankey GJ
    365. Hao Y
    366. Harb HL
    367. Harikrishnan S
    368. Haro JM
    369. Hassankhani H
    370. Hassen HY
    371. Havmoeller R
    372. Hawley CN
    373. Hay SI
    374. Hedayatizadeh-Omran A
    375. Heibati B
    376. Heidari B
    377. Heidari M
    378. Hendrie D
    379. Henok A
    380. Heredia-Pi I
    381. Herteliu C
    382. Heydarpour F
    383. Heydarpour S
    384. Hibstu DT
    385. Higazi TB
    386. Hilawe EH
    387. Hoek HW
    388. Hoffman HJ
    389. Hole MK
    390. Homaie Rad E
    391. Hoogar P
    392. Hosgood HD
    393. Hosseini SM
    394. Hosseinzadeh M
    395. Hostiuc M
    396. Hostiuc S
    397. Hoy DG
    398. Hsairi M
    399. Hsiao T
    400. Hu G
    401. Hu H
    402. Huang JJ
    403. Hussen MA
    404. Huynh CK
    405. Iburg KM
    406. Ikeda N
    407. Ilesanmi OS
    408. Iqbal U
    409. Irvani SSN
    410. Irvine CMS
    411. Islam SMS
    412. Islami F
    413. Jackson MD
    414. Jacobsen KH
    415. Jahangiry L
    416. Jahanmehr N
    417. Jain SK
    418. Jakovljevic M
    419. James SL
    420. Jassal SK
    421. Jayatilleke AU
    422. Jeemon P
    423. Jha RP
    424. Jha V
    425. Ji JS
    426. Jonas JB
    427. Jonnagaddala J
    428. Jorjoran Shushtari Z
    429. Joshi A
    430. Jozwiak JJ
    431. Jürisson M
    432. Kabir Z
    433. Kahsay A
    434. Kalani R
    435. Kanchan T
    436. Kant S
    437. Kar C
    438. Karami M
    439. Karami Matin B
    440. Karch A
    441. Karema C
    442. Karimi N
    443. Karimi SM
    444. Kasaeian A
    445. Kassa DH
    446. Kassa GM
    447. Kassa TD
    448. Kassebaum NJ
    449. Katikireddi SV
    450. Kaul A
    451. Kawakami N
    452. Kazemi Z
    453. Karyani AK
    454. Kefale AT
    455. Keiyoro PN
    456. Kemp GR
    457. Kengne AP
    458. Keren A
    459. Kesavachandran CN
    460. Khader YS
    461. Khafaei B
    462. Khafaie MA
    463. Khajavi A
    464. Khalid N
    465. Khalil IA
    466. Khan G
    467. Khan MS
    468. Khan MA
    469. Khang Y-H
    470. Khater MM
    471. Khazaei M
    472. Khazaie H
    473. Khoja AT
    474. Khosravi A
    475. Khosravi MH
    476. Kiadaliri AA
    477. Kiirithio DN
    478. Kim C-I
    479. Kim D
    480. Kim Y-E
    481. Kim YJ
    482. Kimokoti RW
    483. Kinfu Y
    484. Kisa A
    485. Kissimova-Skarbek K
    486. Kivimäki M
    487. Knibbs LD
    488. Knudsen AKS
    489. Kochhar S
    490. Kokubo Y
    491. Kolola T
    492. Kopec JA
    493. Kosen S
    494. Koul PA
    495. Koyanagi A
    496. Kravchenko MA
    497. Krishan K
    498. Krohn KJ
    499. Kromhout H
    500. Kuate Defo B
    501. Kucuk Bicer B
    502. Kumar GA
    503. Kumar M
    504. Kuzin I
    505. Kyu HH
    506. Lachat C
    507. Lad DP
    508. Lad SD
    509. Lafranconi A
    510. Lalloo R
    511. Lallukka T
    512. Lami FH
    513. Lang JJ
    514. Lansingh VC
    515. Larson SL
    516. Latifi A
    517. Lazarus JV
    518. Lee PH
    519. Leigh J
    520. Leili M
    521. Leshargie CT
    522. Leung J
    523. Levi M
    524. Lewycka S
    525. Li S
    526. Li Y
    527. Liang J
    528. Liang X
    529. Liao Y
    530. Liben ML
    531. Lim L-L
    532. Linn S
    533. Liu S
    534. Lodha R
    535. Logroscino G
    536. Lopez AD
    537. Lorkowski S
    538. Lotufo PA
    539. Lozano R
    540. Lucas TCD
    541. Lunevicius R
    542. Ma S
    543. Macarayan ERK
    544. Machado ÍE
    545. Madotto F
    546. Mai HT
    547. Majdan M
    548. Majdzadeh R
    549. Majeed A
    550. Malekzadeh R
    551. Malta DC
    552. Mamun AA
    553. Manda A-L
    554. Manguerra H
    555. Mansournia MA
    556. Mantovani LG
    557. Maravilla JC
    558. Marcenes W
    559. Marks A
    560. Martin RV
    561. Martins SCO
    562. Martins-Melo FR
    563. März W
    564. Marzan MB
    565. Massenburg BB
    566. Mathur MR
    567. Mathur P
    568. Matsushita K
    569. Maulik PK
    570. Mazidi M
    571. McAlinden C
    572. McGrath JJ
    573. McKee M
    574. Mehrotra R
    575. Mehta KM
    576. Mehta V
    577. Meier T
    578. Mekonnen FA
    579. Melaku YA
    580. Melese A
    581. Melku M
    582. Memiah PTN
    583. Memish ZA
    584. Mendoza W
    585. Mengistu DT
    586. Mensah GA
    587. Mensink GBM
    588. Mereta ST
    589. Meretoja A
    590. Meretoja TJ
    591. Mestrovic T
    592. Mezgebe HB
    593. Miazgowski B
    594. Miazgowski T
    595. Millear AI
    596. Miller TR
    597. Miller-Petrie MK
    598. Mini GK
    599. Mirarefin M
    600. Mirica A
    601. Mirrakhimov EM
    602. Misganaw AT
    603. Mitiku H
    604. Moazen B
    605. Mohajer B
    606. Mohammad KA
    607. Mohammadi M
    608. Mohammadifard N
    609. Mohammadnia-Afrouzi M
    610. Mohammed S
    611. Mohebi F
    612. Mokdad AH
    613. Molokhia M
    614. Momeniha F
    615. Monasta L
    616. Moodley Y
    617. Moradi G
    618. Moradi-Lakeh M
    619. Moradinazar M
    620. Moraga P
    621. Morawska L
    622. Morgado-Da-Costa J
    623. Morrison SD
    624. Moschos MM
    625. Mouodi S
    626. Mousavi SM
    627. Mozaffarian D
    628. Mruts KB
    629. Muche AA
    630. Muchie KF
    631. Mueller UO
    632. Muhammed OS
    633. Mukhopadhyay S
    634. Muller K
    635. Musa KI
    636. Mustafa G
    637. Nabhan AF
    638. Naghavi M
    639. Naheed A
    640. Nahvijou A
    641. Naik G
    642. Naik N
    643. Najafi F
    644. Nangia V
    645. Nansseu JR
    646. Nascimento BR
    647. Neal B
    648. Neamati N
    649. Negoi I
    650. Negoi RI
    651. Neupane S
    652. Newton CRJ
    653. Ngunjiri JW
    654. Nguyen AQ
    655. Nguyen G
    656. Nguyen HT
    657. Nguyen HLT
    658. Nguyen HT
    659. Nguyen M
    660. Nguyen NB
    661. Nichols E
    662. Nie J
    663. Ningrum DNA
    664. Nirayo YL
    665. Nishi N
    666. Nixon MR
    667. Nojomi M
    668. Nomura S
    669. Norheim OF
    670. Noroozi M
    671. Norrving B
    672. Noubiap JJ
    673. Nouri HR
    674. Nourollahpour Shiadeh M
    675. Nowroozi MR
    676. Nsoesie EO
    677. Nyasulu PS
    678. Obermeyer CM
    679. Odell CM
    680. Ofori-Asenso R
    681. Ogbo FA
    682. Oh I-H
    683. Oladimeji O
    684. Olagunju AT
    685. Olagunju TO
    686. Olivares PR
    687. Olsen HE
    688. Olusanya BO
    689. Olusanya JO
    690. Ong KL
    691. Ong SK
    692. Oren E
    693. Orpana HM
    694. Ortiz A
    695. Ota E
    696. Otstavnov SS
    697. Øverland S
    698. Owolabi MO
    699. P A M
    700. Pacella R
    701. Pakhare AP
    702. Pakpour AH
    703. Pana A
    704. Panda-Jonas S
    705. Park E-K
    706. Parry CDH
    707. Parsian H
    708. Patel S
    709. Pati S
    710. Patil ST
    711. Patle A
    712. Patton GC
    713. Paudel D
    714. Paulson KR
    715. Paz Ballesteros WC
    716. Pearce N
    717. Pereira A
    718. Pereira DM
    719. Perico N
    720. Pesudovs K
    721. Petzold M
    722. Pham HQ
    723. Phillips MR
    724. Pillay JD
    725. Piradov MA
    726. Pirsaheb M
    727. Pischon T
    728. Pishgar F
    729. Plana-Ripoll O
    730. Plass D
    731. Polinder S
    732. Polkinghorne KR
    733. Postma MJ
    734. Poulton R
    735. Pourshams A
    736. Poustchi H
    737. Prabhakaran D
    738. Prakash S
    739. Prasad N
    740. Purcell CA
    741. Purwar MB
    742. Qorbani M
    743. Radfar A
    744. Rafay A
    745. Rafiei A
    746. Rahim F
    747. Rahimi Z
    748. Rahimi-Movaghar A
    749. Rahimi-Movaghar V
    750. Rahman M
    751. Rahman MHur
    752. Rahman MA
    753. Rai RK
    754. Rajati F
    755. Rajsic S
    756. Raju SB
    757. Ram U
    758. Ranabhat CL
    759. Ranjan P
    760. Rath GK
    761. Rawaf DL
    762. Rawaf S
    763. Reddy KS
    764. Rehm CD
    765. Rehm J
    766. Reiner RC
    767. Reitsma MB
    768. Remuzzi G
    769. Renzaho AMN
    770. Resnikoff S
    771. Reynales-Shigematsu LM
    772. Rezaei S
    773. Ribeiro ALP
    774. Rivera JA
    775. Roba KT
    776. Rodríguez-Ramírez S
    777. Roever L
    778. Román Y
    779. Ronfani L
    780. Roshandel G
    781. Rostami A
    782. Roth GA
    783. Rothenbacher D
    784. Roy A
    785. Rubagotti E
    786. Rushton L
    787. Sabanayagam C
    788. Sachdev PS
    789. Saddik B
    790. Sadeghi E
    791. Saeedi Moghaddam S
    792. Safari H
    793. Safari Y
    794. Safari-Faramani R
    795. Safdarian M
    796. Safi S
    797. Safiri S
    798. Sagar R
    799. Sahebkar A
    800. Sahraian MA
    801. Sajadi HS
    802. Salam N
    803. Salamati P
    804. Saleem Z
    805. Salimi Y
    806. Salimzadeh H
    807. Salomon JA
    808. Salvi DD
    809. Salz I
    810. Samy AM
    811. Sanabria J
    812. Sanchez-Niño MD
    813. Sánchez-Pimienta TG
    814. Sanders T
    815. Sang Y
    816. Santomauro DF
    817. Santos IS
    818. Santos JV
    819. Santric Milicevic MM
    820. Sao Jose BP
    821. Sardana M
    822. Sarker AR
    823. Sarmiento-Suárez R
    824. Sarrafzadegan N
    825. Sartorius B
    826. Sarvi S
    827. Sathian B
    828. Satpathy M
    829. Sawant AR
    830. Sawhney M
    831. Saylan M
    832. Sayyah M
    833. Schaeffner E
    834. Schmidt MI
    835. Schneider IJC
    836. Schöttker B
    837. Schutte AE
    838. Schwebel DC
    839. Schwendicke F
    840. Scott JG
    841. Seedat S
    842. Sekerija M
    843. Sepanlou SG
    844. Serre ML
    845. Serván-Mori E
    846. Seyedmousavi S
    847. Shabaninejad H
    848. Shaddick G
    849. Shafieesabet A
    850. Shahbazi M
    851. Shaheen AA
    852. Shaikh MA
    853. Shamah Levy T
    854. Shams-Beyranvand M
    855. Shamsi M
    856. Sharafi H
    857. Sharafi K
    858. Sharif M
    859. Sharif-Alhoseini M
    860. Sharifi H
    861. Sharma J
    862. Sharma M
    863. Sharma R
    864. She J
    865. Sheikh A
    866. Shi P
    867. Shibuya K
    868. Shiferaw MS
    869. Shigematsu M
    870. Shin M-J
    871. Shiri R
    872. Shirkoohi R
    873. Shiue I
    874. Shokraneh F
    875. Shoman H
    876. Shrime MG
    877. Shupler MS
    878. Si S
    879. Siabani S
    880. Sibai AM
    881. Siddiqi TJ
    882. Sigfusdottir ID
    883. Sigurvinsdottir R
    884. Silva DAS
    885. Silva JP
    886. Silveira DGA
    887. Singh JA
    888. Singh NP
    889. Singh V
    890. Sinha DN
    891. Skiadaresi E
    892. Skirbekk V
    893. Smith DL
    894. Smith M
    895. Sobaih BH
    896. Sobhani S
    897. Somayaji R
    898. Soofi M
    899. Sorensen RJD
    900. Soriano JB
    901. Soyiri IN
    902. Spinelli A
    903. Sposato LA
    904. Sreeramareddy CT
    905. Srinivasan V
    906. Starodubov VI
    907. Steckling N
    908. Stein DJ
    909. Stein MB
    910. Stevanovic G
    911. Stockfelt L
    912. Stokes MA
    913. Sturua L
    914. Subart ML
    915. Sudaryanto A
    916. Sufiyan MB
    917. Sulo G
    918. Sunguya BF
    919. Sur PJ
    920. Sykes BL
    921. Szoeke CEI
    922. Tabarés-Seisdedos R
    923. Tabuchi T
    924. Tadakamadla SK
    925. Takahashi K
    926. Tandon N
    927. Tassew SG
    928. Tavakkoli M
    929. Taveira N
    930. Tehrani-Banihashemi A
    931. Tekalign TG
    932. Tekelemedhin SW
    933. Tekle MG
    934. Temesgen H
    935. Temsah M-H
    936. Temsah O
    937. Terkawi AS
    938. Tessema B
    939. Teweldemedhin M
    940. Thankappan KR
    941. Theis A
    942. Thirunavukkarasu S
    943. Thomas HJ
    944. Thomas ML
    945. Thomas N
    946. Thurston GD
    947. Tilahun B
    948. Tillmann T
    949. To QG
    950. Tobollik M
    951. Tonelli M
    952. Topor-Madry R
    953. Torre AE
    954. Tortajada-Girbés M
    955. Touvier M
    956. Tovani-Palone MR
    957. Towbin JA
    958. Tran BX
    959. Tran KB
    960. Truelsen TC
    961. Truong NT
    962. Tsadik AG
    963. Tudor Car L
    964. Tuzcu EM
    965. Tymeson HD
    966. Tyrovolas S
    967. Ukwaja KN
    968. Ullah I
    969. Updike RL
    970. Usman MS
    971. Uthman OA
    972. Vaduganathan M
    973. Vaezi A
    974. Valdez PR
    975. Van Donkelaar A
    976. Varavikova E
    977. Varughese S
    978. Vasankari TJ
    979. Venkateswaran V
    980. Venketasubramanian N
    981. Villafaina S
    982. Violante FS
    983. Vladimirov SK
    984. Vlassov V
    985. Vollset SE
    986. Vos T
    987. Vosoughi K
    988. Vu GT
    989. Vujcic IS
    990. Wagnew FS
    991. Waheed Y
    992. Waller SG
    993. Walson JL
    994. Wang Y
    995. Wang Y
    996. Wang Y-P
    997. Weiderpass E
    998. Weintraub RG
    999. Weldegebreal F
    1000. Werdecker A
    1001. Werkneh AA
    1002. West JJ
    1003. Westerman R
    1004. Whiteford HA
    1005. Widecka J
    1006. Wijeratne T
    1007. Winkler AS
    1008. Wiyeh AB
    1009. Wiysonge CS
    1010. Wolfe CDA
    1011. Wong TY
    1012. Wu S
    1013. Xavier D
    1014. Xu G
    1015. Yadgir S
    1016. Yadollahpour A
    1017. Yahyazadeh Jabbari SH
    1018. Yamada T
    1019. Yan LL
    1020. Yano Y
    1021. Yaseri M
    1022. Yasin YJ
    1023. Yeshaneh A
    1024. Yimer EM
    1025. Yip P
    1026. Yisma E
    1027. Yonemoto N
    1028. Yoon S-J
    1029. Yotebieng M
    1030. Younis MZ
    1031. Yousefifard M
    1032. Yu C
    1033. Zaidi Z
    1034. Zaman SB
    1035. Zamani M
    1036. Zavala-Arciniega L
    1037. Zhang AL
    1038. Zhang H
    1039. Zhang K
    1040. Zhou M
    1041. Zimsen SRM
    1042. Zodpey S
    1043. Murray CJL
    (2018) Global, regional, and national comparative risk assessment of 84 behavioural, environmental and occupational, and metabolic risks or clusters of risks for 195 countries and territories, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017
    The Lancet 392:1923–1994.
    https://doi.org/10.1016/S0140-6736(18)32225-6
    1. Stoeldraijer L
    2. van Duin D
    3. Janssen F
    (2013b)
    Bevolkingsprognose 2012-2060: model en veronderstellingen betreffende de sterfte
    Bevolkingstrends 61:1–27.
  67. Thesis
    1. Stoeldraijer L
    (2019)
    Mortality Forecasting in the Context of Non-Linear Past Mortality Trends: An Evaluation
    PhD Thesis, University of Groningen.
  69. Book
    1. Thatcher AR
    2. Kannisto V
    3. Vaupel JW
    (1998)
    The Force of Mortality at Ages 80 to 120
    Odense: Odense University Press.
  70. Book
    1. Thun MJ
    2. Day-Lally C
    3. Myers DG
    (1997)
    Trends in tobacco smoking and mortality from cigarette use in Cancer Prevention Studies I (1959 through 1965) and II (1982 through 1988)
    In: Burns D. M, Garfinkel L, Samet J. M, editors. Changes in Cigarette-Related Disease Risks and Their Implications for Prevention and Control. Bethesda, Md: National Cancer Institute. Smoking and Tobacco Control Monograph. pp. 305–382.
  72. Website
    1. United Nations
    (2020) World population prospects 2019
    Department of Economic and Social Affairs - Population Dynamics. Accessed December 22, 2020.
    https://population.un.org/wpp/

Decision letter

  1. Belinda Nicolau
    Reviewing Editor; McGill University, Canada
  2. Eduardo Franco
    Senior Editor; McGill University, Canada
  3. Ugofilippo Basellini
    Reviewer
  4. Collin Payne
    Reviewer

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

Combining a few analytical strategies, this manuscript proposes a new approach to forecast mortality that distinguishes future changes in lifestyle- and non-lifestyle attributable mortality on life expectancy projection. The authors use data from 18 European countries to test their model and the results yield more optimistic forecasts than other well-established forecasting methods; future generations will have an increased life expectancy than previously expected. The proposed methodology could bring significant benefits when applied to other contexts and represents an important contribution to the field of life expectancy forecasting.

Decision letter after peer review:

Thank you for submitting your article "Future life expectancy in Europe taking into account the impact of smoking, obesity and alcohol" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and a Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Ugofilippo Basellini (Reviewer #1); Collin Payne (Reviewer #2).

As is customary in eLife, the reviewers have discussed their critiques with one another. What follows below is the Reviewing Editor's edited compilation of the essential and ancillary points provided by reviewers in their critiques and in their interaction post-review. Please submit a revised version that addresses these concerns directly. Although we expect that you will address these comments in your response letter, we also need to see the corresponding revision clearly marked in the text of the manuscript. Some of the reviewers' comments may seem to be simple queries or challenges that do not prompt revisions to the text. Please keep in mind, however, that readers may have the same perspective as the reviewers. Therefore, it is essential that you attempt to amend or expand the text to clarify the narrative accordingly.

Essential Revisions (for the authors):

1. Line 206: "France, Spain, and Italy". I am unsure why you included Spain in your forerunner population, since you do not have Spain in your results. This seems not very consistent, and it should be better justified and clarified. I suppose that you do not have lifestyle-attributable mortality for Spain? How do your results change if you exclude Spain from the forerunners (I imagine very little)?

2. Line 392-393: "However, in our view, the added value of integrating lifestyle into mortality projections outweighs the uncertainties that come with it." This should, and can, be empirically tested with out-of-sample exercises.

3. SM Methodology: "All our projections are based on 50,000 simulations" and "We obtained, for each sex-specific population, 50,000 simulated matrices". I was not able to understand how you performed these simulations. Are these generated from the forecasts of the time series model? Or do you employ a bootstrap on the deviance residuals of the fitted models? Or both? More details on how you compute variability of your forecasts are needed.

4. The following text is an expanded explanation of the Public Evaluation of Reviewer 2: "My main issue with the paper as written concerns the assumptions built into the Li-Lee projection of non-lifestyle attributable mortality. I wonder how sensitive the results are to the choice of reference populations here--that is, I'm not sure that female populations of France, Spain, and Italy are the best populations to use as reference here.

These populations currently have high life expectancies, but also have histories of substantial mortality shocks and had much higher mortality rates in the near past. My issue here comes down to the essential difference between current mortality rates and current mortality conditions. The low age-specific mortality rates of these populations are a function of both current mortality conditions (e.g. period effects/current conditions in terms of development, medical care, etc) and of the history of cohort mortality selection pressures among those currently in the population. The chosen reference populations all have older cohorts who have experienced fairly severe mortality selection in early life, which may lead to smaller groups of more selected, "hardier" individuals surviving to later life. I worry that choosing these populations may provide an over-optimistic reference point for convergence (indeed, these three countries have among the largest gaps between period e0 and currently attained cohort e0--see https://doi.org/10.1080/00324728.2019.1618480). So I'd be interested in seeing whether results differed by using a set of low-mortality countries that have faced less severe life-course mortality pressures--e.g. Sweden, Norway, and Switzerland. My hunch is that your life expectancy projections will decline a little, though not massively. But I think it's worth pushing on this assumption a bit to see how sensitive it is."

Reviewer #1 (Recommendations for the authors):

The manuscript introduces a new approach to forecast mortality that explicitly considers the role of lifestyle epidemics (smoking, obesity and alcohol) on the dynamics of mortality. Specifically, a four-step projection methodology is proposed, which distinguishes between non-lifestyle-attributable and lifestyle-attributable mortality. The former is forecast using an extrapolative model, the latter is forecast using a data- and theory-driven approach, and the two components are combined to obtain all-cause mortality forecasts.

The model is applied to 18 European countries and compared to three other well-established forecasting methods (Lee-Carter, United Nations and Eurostat). The proposed approach results in more optimistic forecasts than the other methods, suggesting that future individuals will live longer lives than previously expected. Moreover, forecasts of the proposed methodology are more realistic as they do not display implausible crossovers between sexes and countries which characterize the other methods.

For these reasons, the proposed methodology appears to be an important contribution to the mortality forecasting literature. However, the manuscript has some weaknesses that the authors should address to further improve their work.

Weaknesses

While the forecasts of the proposed approach appear to be more realistic than those of the benchmark model of Lee and Carter (1992), the forecast accuracy of the proposed model is not empirically evaluated. In the recent mortality forecasting literature, great attention has been devoted to out-of-sample validation exercises, which are employed to measure the accuracy of the forecasts (see, e.g.,Shang et al. 2011,Bergeron-Boucher et al. 2017). A point and interval forecast accuracy evaluation of the proposed model as compared to the Lee-Carter one would provide concrete evidence of the new model's strengths. Clearly, given the limited time series available to the authors, only a short-term forecast evaluation can be performed (e.g. 5 and 10 years forecasts).

Moreover, the results and discussions of the authors are exclusively based on life expectancy, which is certainly a very important summary measure of mortality, but it does not capture everything. No results are shown or discussed in terms of forecast mortality rates and lifespan inequality. The former are very relevant to understand the plausibility of the forecast mortality age-pattern. The latter is another important summary measure of mortality that complements life expectancy (van Raalte et al., 2018), and it should be considered in mortality forecasting (Bohk-Ewald et al., 2017).

The data employed to fit the model is limited to 25 years (1990--2014) because lifestyle-attributable mortality data is available only for this period of time. Based on this (rather short) fitting period, the authors forecast mortality for a period of time twice as long (up to 2065, that is a 51 year forecast). This seems rather unbalanced. Moreover, forecasts of lifestyle-attributable mortality seem to be driven more by theory/expert opinion rather that by observed time series, especially for females. I think these potential limitations should be more clearly specified in the manuscript.

Finally, the principal added value of the manuscript – the new approach to forecast mortality – is currently not available to the general public: "Because of the large size of the many different underlying simulation matrices, we are restricted in sharing these data. … The R codes can be requested from the author." One simple way to address this problem is to share a much smaller set of simulations (say 100 instead of 50,000) and warn the users to increase the number of simulations when running the codes on their machines. This would directly provide the research community with the routines to employ the proposed methodology.

References

Bergeron-Boucher, M.-P., V. Canudas-Romo, J. Oeppen, and J. W. Vaupel (2017). Coherent forecasts of mortality with compositional data analysis. Demographic Research 37, 527-566.

Bohk-Ewald, C., M. Ebeling, and R. Rau (2017). Lifespan Disparity as an Additional Indicator for Evaluating Mortality Forecasts. Demography 54 (4), 1559-1577.

Lee, R. D. and L. R. Carter (1992). Modeling and forecasting US mortality. Journal of the American Statistical Association 87 (419), 659-671.

Shang, H. L., H. Booth, and R. Hyndman (2011). Point and interval forecasts of mortality rates and life expectancy: A comparison of ten principal component methods. Demographic Research 25, 173-214.

van Raalte, A. A., I. Sasson, and P. Martikainen (2018). The case for monitoring life-span inequality. Science 362 (6418), 1002-1004.

Reviewer #2 (Recommendations for the authors):

The authors have conducted an ambitious projection exercise, drawing together a number of methods and approaches to seek to understand the impacts of future changes in lifestyle- and non-lifestyle attributable mortality on prospects for life expectancy increase in a set of European countries. This focus on the underlying dynamics of mortality change--that is, disaggregating the long-term changes in non-lifestyle attributable mortality patterns from the much more variable lifestyle-attributable mortality patterns--represents a real step forward in life expectancy projection modeling. These methods could have considerable benefits for projections in other contexts.

My main issue with the paper as written concerns the assumptions built into the Li-Lee projection of non-lifestyle attributable mortality, where I would encourage the authors to test the sensitivity of their projections to alternative choices of reference countries (looking beyond Italy, Spain, and France).

https://doi.org/10.7554/eLife.66590.sa1

Author response

Essential Revisions (for the authors):

1. Line 206: "France, Spain, and Italy". I am unsure why you included Spain in your forerunner population, since you do not have Spain in your results. This seems not very consistent, and it should be better justified and clarified. I suppose that you do not have lifestyle-attributable mortality for Spain? How do your results change if you exclude Spain from the forerunners (I imagine very little)?

We selected women in France, Spain, and Italy as the forerunner populations in terms of life expectancy in Europe because they exhibit both very high recent life expectancy values and very favourable long-lasting past trends in e0 (measured by strong increases in e0 over the 1970-2011 period among non-Eastern European countries) (Stoeldraijer 2019) (see page 4 of the Supplementary Data and Methods (now Appendix 1)). Our decision to use forerunner populations as the reference populations is in line with the seminal work by James Vaupel (e.g., Oeppen and Vaupel 2002), and our decision to use female populations is in line with the recent work by Bergeron-Boucher et al. 2018. In addition, using three populations instead of one is regarded more robust.

While we had lifestyle-attributable mortality data for Spain, we ended up excluding the projection outcomes for Spain because we were not able to obtain realistic long-term projections of smoking-attributable mortality for Spanish women. Specifically, we found that the projections of smoking-attributable mortality for Spanish women were not robust to the selection of different historical periods (see Janssen et al. 2020 Tobacco Control). Nonetheless, our main argument for why we selected women in France, Spain, and Italy as the forerunner populations in Europe in terms of life expectancy holds.

Given that the gain in e0 over the 1990-2014 period was slightly higher for Spanish women than it was for women in Italy and France (see Table 1), but that the population size was smaller in Spain than in Italy and France, the exclusion of Spanish women from the forerunner populations will most likely result in only slightly lower future life expectancy values. This expectation proves indeed in line with the results of our sensitivity analysis regarding the selection of the forerunner populations that we present in Author response table 1. That is, the coherent projection of non-lifestyle-attributable mortality in which we did not include Spanish women as forerunner population resulted in life expectancy values in 2065 that were – on average – 0.21 years lower for men, and 0.14 years lower for women. For all-cause mortality we see a similar effect (slightly larger difference).

Author response table 1
Effect of the inclusion of alternative reference populations within coherent mortality projections* on the coherently projected e0 values in 2065 for all-cause mortality and for non-lifestyle-attributable mortality.
All-cause mortalityNon-lifestyle attributable mortality
CountryCoh. Li-Lee: sensivity ICoh. Li-Lee: sensivity IICoh. Li-Lee: sensivity IIICoh. Li-Lee: sensivity ICoh. Li-Lee: sensivity IICoh. Li-Lee: sensivity III
Men
Austria-0.20-2.19-0.76-0.13-0.77-0.94
Belgium-0.17-2.08-0.61-0.16-0.81-0.90
Czech Republic-0.24-2.33-0.71-0.23-0.76-1.09
Denmark-0.17-2.11-0.60-0.14-0.71-0.91
Finland-0.26-2.13-0.83-0.22-0.73-1.00
France-0.21-2.12-0.75-0.09-0.78-0.78
Germany-0.23-2.09-0.85-0.22-0.70-0.95
Greece-0.27-2.05-0.80-0.31-0.82-1.01
Hungary-0.19-2.11-0.65-0.24-0.83-1.14
Ireland-0.17-2.05-0.80-0.13-0.82-1.06
Italy-0.28-1.86-0.74-0.18-0.73-0.92
Netherlands-0.16-1.94-0.67-0.17-0.76-0.84
Norway-0.35-1.83-0.77-0.31-0.82-1.03
Poland-0.22-2.14-0.67-0.28-0.79-1.01
Slovenia-0.22-2.03-0.68-0.15-0.67-1.00
Sweden-0.27-1.79-0.71-0.29-0.64-0.89
Switzerland-0.23-1.82-0.67-0.19-0.69-0.89
United Kingdom-0.26-2.02-0.70-0.23-0.80-0.91
Average-0.23-2.04-0.72-0.21-0.76-0.96
Women
Austria-0.19-1.61-0.56-0.16-0.73-0.81
Belgium-0.21-1.67-0.72-0.09-0.76-0.83
Czech Republic-0.24-1.70-0.64-0.10-0.68-0.87
Denmark-0.19-1.77-0.66-0.08-0.53-0.58
Finland-0.22-1.69-0.69-0.14-0.73-0.86
France-0.23-1.37-0.54-0.12-0.71-0.77
Germany-0.21-1.55-0.66-0.16-0.66-0.82
Greece-0.27-1.52-0.72-0.13-0.69-0.85
Hungary-0.18-1.79-0.65-0.16-0.79-1.03
Ireland-0.28-1.83-0.77-0.18-0.81-0.80
Italy-0.25-1.59-0.65-0.16-0.70-0.85
Netherlands-0.25-1.57-0.68-0.15-0.66-0.75
Norway-0.24-1.67-0.66-0.09-0.66-0.85
Poland-0.18-1.85-0.69-0.19-0.71-0.90
Slovenia-0.20-1.62-0.68-0.14-0.60-0.85
Sweden-0.23-1.56-0.65-0.14-0.59-0.74
Switzerland-0.19-1.45-0.59-0.16-0.66-0.73
United Kingdom-0.23-1.65-0.68-0.18-0.56-0.70
Average-0.22-1.64-0.66-0.14-0.68-0.81

  1. *original = using women in France, Spain and Italy as the reference population.

    sensitivity 1 = using only women in France and Italy as the reference population.

  2. sensitivity 2 = using women in Sweden, Norway and Switzerland as the reference population.

    sensitivity 3 = using the average across the 18 countries included in the analysis as the reference population.

  3. Note: the sensitivity analyses were all based on 1000 simulations, and we assumed non-stationarity for the remaining population-specific time trends k(t,i).

In our revised manuscript, we now provide some additional details regarding why we selected women in France, Spain, and Italy as the forerunner populations in Europe (Results section; lines 219-222; Appendix 1 – Supplementary Data and Methods – page 4). We have also included a discussion regarding the selection of forerunner populations in our discussion (see as well our reply to essential revision nr. 4).

2. Line 392-393: "However, in our view, the added value of integrating lifestyle into mortality projections outweighs the uncertainties that come with it." This should, and can, be empirically tested with out-of-sample exercises.

Although out-of-sample exercises can indeed be used in assessing the accuracy of forecasts, we do not think that such an exercise would be able to test whether the added value of integrating lifestyle into mortality projections outweighs the uncertainties that come with doing so. That is, as well as the accuracy of forecasts, which out-of-sample exercises can indeed help to determine, there are many additional evaluation criteria for assessing the performance of forecasts (see Cairns et al. 2009). These include, for example, biological reasonableness, the plausibility of outcomes, and the robustness of the forecast outcomes (Cairns et al. 2009). Because these latter issues are more difficult to quantify, they are often ignored. Instead, accuracy alone is tested by means of out-of-sample (=backtesting) exercises. We believe that our mortality projection approach is particularly promising because its outcomes can be considered more realistic, more robust, and more insightful than previous mortality projections that were purely based on extrapolation.

These points are highlighted in the remainder of the respective paragraph:

“However, in our view, the added value of integrating lifestyle into mortality projections outweighs the uncertainties that come with it. First, it provides a better approximation of the underlying mortality trend. Second, our projection is driven not only by data, but by theory, and is preceded by a careful study of past trends. Third, by distinguishing between the underlying mortality trend and the additional effect of the three lifestyle factors combined, these projections provide much greater transparency and insight than a mere mechanical extrapolation of past trends could.”

Moreover, in response to the comment by reviewer 1 on this issue, we have now added at the end of this paragraph an explicit reference to the different evaluation criteria used for mortality projections, and have also added a reference to our more realistic outcomes (lines 438-446):

“Fourth, our approach resulted in more realistic differences between countries and sexes than those projected by other approaches (see the next section as well). The above mentioned four added values of our mortality projection approach can be clearly linked to important criteria for evaluating the performance of mortality forecasts in addition to accuracy: namely, robustness, plausibility, reasonableness, and transparency (Cairns et al. 2009). Moreover, the outcomes of our projection approach – which rely above all on the coherent projection of non-lifestyle-attributable mortality – are less dependent on the explicit choices that underlie general mortality forecasts (e.g., the choice of the forerunner populations and the historical time period)”.

In addition, in our text, we avoided using the words “accurate” and “inaccurate” as much as possible and replaced these terms with “reliable” and “unreliable”.

3. SM Methodology: "All our projections are based on 50,000 simulations" and "We obtained, for each sex-specific population, 50,000 simulated matrices". I was not able to understand how you performed these simulations. Are these generated from the forecasts of the time series model? Or do you employ a bootstrap on the deviance residuals of the fitted models? Or both? More details on how you compute variability of your forecasts are needed.

Yes, the simulations are generated from the forecasts of the time series model. We did not employ a bootstrap on the deviance residuals of the fitted models. Thus, we did not take into account parameter uncertainty. We have added these details to the Supplementary Data and Methods (page 2 and page 5)(=now Appendix 1).

4. The following text is an expanded explanation of the Public Evaluation of Reviewer 2: "My main issue with the paper as written concerns the assumptions built into the Li-Lee projection of non-lifestyle attributable mortality. I wonder how sensitive the results are to the choice of reference populations here--that is, I'm not sure that female populations of France, Spain, and Italy are the best populations to use as reference here.These populations currently have high life expectancies, but also have histories of substantial mortality shocks and had much higher mortality rates in the near past. My issue here comes down to the essential difference between current mortality rates and current mortality conditions. The low age-specific mortality rates of these populations are a function of both current mortality conditions (e.g. period effects/current conditions in terms of development, medical care, etc) and of the history of cohort mortality selection pressures among those currently in the population. The chosen reference populations all have older cohorts who have experienced fairly severe mortality selection in early life, which may lead to smaller groups of more selected, "hardier" individuals surviving to later life. I worry that choosing these populations may provide an over-optimistic reference point for convergence (indeed, these three countries have among the largest gaps between period e0 and currently attained cohort e0--see https://doi.org/10.1080/00324728.2019.1618480). So I'd be interested in seeing whether results differed by using a set of low-mortality countries that have faced less severe life-course mortality pressures--e.g. Sweden, Norway, and Switzerland. My hunch is that your life expectancy projections will decline a little, though not massively. But I think it's worth pushing on this assumption a bit to see how sensitive it is."

Yes, the choice of the reference or forerunner populations is indeed an important aspect of multi-population or coherent mortality forecasting (Stoeldraijer 2019; Kjærgaard et al. 2016). However, given that the trend in non-lifestyle-attributable mortality is more equal between countries and sexes than the trend in all-cause mortality (see Table 1), this choice will likely have a larger impact when coherently forecasting all-cause mortality than when coherently forecasting non-lifestyle-attributable mortality (see also Janssen and Kunst 2007 and Janssen et al. 2013). A comparison of the average (unweighted) gains in e0 over the 1990-2014 period for women in France, Italy, and Spain (4.8 for all-cause mortality; 5.0 for non-lifestyle-attributable mortality) with the average (unweighted) gains in e0 for women in Sweden, Norway, and Switzerland (4.1 for all-cause mortality; 4.8 for non-lifestyle-attributable mortality) shows that the selection of these latter populations as the forerunner populations would likely result in lower projection outcomes, particularly for all-cause mortality, and to a much lesser extent for non-lifestyle-attributable mortality. This expectation proves indeed in line with the results of our sensitivity analysis regarding the selection of the forerunner populations that we present in Author response table 1.

In response to the remark by the reviewer, we have now added a discussion on the choice of the reference populations in the “Appraisal of our projection methodology” section (lines 384 – 396): “However, as well as the potential effects on the outcomes of the exact coherent methodology that is applied (Stoeldraijer, 2019), it is important to take into account the choice of the reference population (Stoeldraijer 2019; Booth 2020). Using forerunner populations as the reference population, as advocated by among others Oeppen and Vaupel 2002 and Bergeron-Boucher et al. 2018, instead of an average across many populations (Li and Lee 2005; Eurostat 2020) will generally result in higher future life expectancy values (see the next section as well). However, given that the past trend in e0 for non-lifestyle-attributable mortality is more equal between countries and sexes than the past trend in e0 for all-cause mortality (see Table 1), which reference populations are chosen will likely have a larger impact when coherently forecasting all-cause mortality than when coherently forecasting non-lifestyle-attributable mortality (see also Janssen and Kunst 2007 and Janssen et al. 2013)”.

Reviewer #1 (Recommendations for the authors):

The manuscript introduces a new approach to forecast mortality that explicitly considers the role of lifestyle epidemics (smoking, obesity and alcohol) on the dynamics of mortality. Specifically, a four-step projection methodology is proposed, which distinguishes between non-lifestyle-attributable and lifestyle-attributable mortality. The former is forecast using an extrapolative model, the latter is forecast using a data- and theory-driven approach, and the two components are combined to obtain all-cause mortality forecasts.The model is applied to 18 European countries and compared to three other well-established forecasting methods (Lee-Carter, United Nations and Eurostat). The proposed approach results in more optimistic forecasts than the other methods, suggesting that future individuals will live longer lives than previously expected. Moreover, forecasts of the proposed methodology are more realistic as they do not display implausible crossovers between sexes and countries which characterize the other methods. For these reasons, the proposed methodology appears to be an important contribution to the mortality forecasting literature. However, the manuscript has some weaknesses that the authors should address to further improve their work.

Weaknesses

While the forecasts of the proposed approach appear to be more realistic than those of the benchmark model of Lee and Carter (1992), the forecast accuracy of the proposed model is not empirically evaluated. In the recent mortality forecasting literature, great attention has been devoted to out-of-sample validation exercises, which are employed to measure the accuracy of the forecasts (see, e.g.,Shang et al., 2011,Bergeron-Boucher et al. 2017). A point and interval forecast accuracy evaluation of the proposed model as compared to the Lee-Carter one would provide concrete evidence of the new model's strengths. Clearly, given the limited time series available to the authors, only a short-term forecast evaluation can be performed (e.g. 5 and 10 years forecasts).

It is indeed the case that we did not empirically evaluate the forecast accuracy of the proposed model. There are three reasons why we made this choice. First, accuracy is only one of the many criteria for assessing the performance of forecasts (see Cairns et al. 2009). Among the other evaluation criteria are biological reasonableness, the plausibility of outcomes, the robustness of the forecast outcomes, and transparency (Cairns et al. 2009). Because these latter issues are more difficult to quantify, they are often ignored. Instead, only accuracy is tested by means of out-of-sample exercises (=backtesting). However, as we indicated in our discussion, we consider our mortality projection approach to be particularly promising because its outcomes can be considered more realistic, more robust, more theoretically informed, and more transparent than the outcomes of other approaches. Second, given the limited time series for lifestyle-attributable morality, we could only perform a short-term forecast evaluation, whereas our main aim is to produce reliable long-term forecasts. Third, given the profile of the readers of eLife, we believe that applying additional technical exercises would not necessarily be an improvement.

Related to our choice to avoid empirically evaluating the forecast accuracy of our model, we have avoided making any statements in our discussion that would suggest that our model is more accurate than the Lee-Carter model. Moreover, we now avoid the use of “accurate” and “inaccurate” as much as possible in the remainder of our text. In addition, we now explicitly refer to the different criteria for evaluating mortality projections in the last paragraph of the “Appraisal of our projection methodology” section.

Moreover, the results and discussions of the authors are exclusively based on life expectancy, which is certainly a very important summary measure of mortality, but it does not capture everything. No results are shown or discussed in terms of forecast mortality rates and lifespan inequality. The former are very relevant to understand the plausibility of the forecast mortality age-pattern. The latter is another important summary measure of mortality that complements life expectancy (van Raalte et al. 2018), and it should be considered in mortality forecasting (Bohk-Ewald et al., 2017).

We completely agree that in addition to life expectancy, other outcome measures of mortality forecasts are relevant and interesting. However, we decided to focus on life expectancy at birth in our manuscript because it is a very common summary measure of health; it is the most common output measure of mortality forecasts; it is easier to grasp than lifespan variation; and it summarizes the age pattern of mortality using a single measure. Given the more general audience of eLife, we do not consider it particularly helpful to include additional outcome measures as part of our submission. However, in response to the comment of the reviewer, we have clarified why we have chosen to focus on life expectancy, thereby referring to other outcome measures that are important as well (see lines 168 – 170; see Supplementary Data and Methods (now Appendix 1) page 2 – first paragraph).

In addition, we have added figures with the projected age-specific mortality rates (Author response image 1 and Author response image 2) and a table (Author response table 2) with the observed (2014) and projected (2065) standard deviation (SD) of the age-at-death distribution from age zero onwards ( = the dx column of the lifetable). This measure is often used to denote lifespan variation (see Edwards and Tuljapurkar, 2005; Garcia and Aburto, 2019).

Furthermore, we will make available the projected age-specific mortality rates (median and 95% and 90% projection intervals) on Open Science Framework (OSF).

Author response image 1
Download asset Open asset
Age-specific mortality pattern in 1990*, 2014*, and for different projections in 2065, by country, ages 0-130, men. * based on the adjusted rates up to 130 (because of zeros in the observed rates).
Author response image 2
Download asset Open asset
Age-specific mortality pattern in 1990*, 2014*, and for different projections in 2065, by country, ages 0-130, women. * based on the adjusted rates up to 130 (because of zeros in the observed rates).
Author response table 2
Observed (2014) and projected (2065) lifespan variation from age zero onwards, by country and sex, according to our projection methodology, which takes into account the impact of smoking, obesity, and alcohol and the mortality experiences of forerunner countries (“lifestyle and coherent”); the benchmark Lee-Carter extrapolative mortality projection applied to all-cause mortality (“Lee-Carter”); and when purely accounting for smoking, obesity, and alcohol (“adding lifestyle”).
MENWOMEN
Projected SD* 2065Projected SD* 2065
CountrySD* 2014Lee-CarterAdding lifestyleLifestyle and coherentSD* 2014Lee-CarterAdding lifestyleLifestyle and coherent
Austria14.3310.8811.1110.3113.059.698.738.91
Belgium14.7011.5711.279.7813.3611.099.539.18
Czechia14.3610.9612.0210.6912.619.238.539.22
Denmark14.2510.5110.3710.0913.2610.1510.109.27
Finland14.6911.6111.1810.3512.6610.909.659.03
France15.3211.7511.3510.0413.5911.119.929.27
Germany14.2510.6611.3110.3912.939.618.859.23
Greece15.1014.3114.2610.5612.6710.209.989.25
Hungary15.1112.7711.2311.2113.9011.389.3310.01
Ireland14.5211.4310.849.7813.1810.599.989.25
Italy13.539.1110.959.7612.419.069.018.94
Netherlands13.629.8210.169.6013.1911.129.688.93
Norway13.7410.179.939.6812.5010.229.398.79
Poland15.9313.4912.5211.5013.9211.089.659.86
Slovenia13.8910.0510.9310.3312.289.008.398.96
Sweden13.6010.3110.139.4712.499.889.158.69
Switzerland13.849.4510.299.8112.599.378.798.66
United Kingdom14.6412.2411.8710.1113.6211.0610.589.37
Average14.4111.1711.2110.1913.0110.269.409.16
Min13.539.119.939.4712.289.008.398.66
Max15.9314.3114.2611.5013.9211.3810.5810.01
Variance0.421.810.990.280.250.610.340.12

  1. * SD = Standard deviation of the age at death distribution (age 0-100).

The data employed to fit the model is limited to 25 years (1990-2014) because lifestyle-attributable mortality data is available only for this period of time. Based on this (rather short) fitting period, the authors forecast mortality for a period of time twice as long (up to 2065, that is a 51 year forecast). This seems rather unbalanced. Moreover, forecasts of lifestyle-attributable mortality seem to be driven more by theory/expert opinion rather that by observed time series, especially for females. I think these potential limitations should be more clearly specified in the manuscript.

In general, it is indeed considered optimal to base projections – and, in particular, extrapolations – on a historical period of about 40-50 years when projecting 40-50 years into the future (Technical Panel on Assumptions and Methods, 2003). Thus, it is a pity that because data on alcohol-attributable mortality were only available from 1990 onwards, we could base our forecast only on data from 1990 to 2014/LAY. However, we decided against restricting our projection horizon to 25 years in the future because we considered it essential to show that our approach is able to generate reliable predictions for the long-term future. It should also be noted that the use of a relatively long historical period stems from the desire to select the most reliable long-term past trend on which the extrapolations could be based. However, because of the substantial non-linearity in the past trends in life expectancy in recent decades, many of the approaches currently used rely on a shorter past period (Stoeldraijer et al. 2013a), or rely on the trends for women that were more linear (e.g. Bergeron-Boucher et al. 2018). Moreover, it is in cases in which the past trends were non-linear that the projection outcomes rely most heavily on the historical time period chosen (Stoeldraijer et al. 2013a). Because our approach relies on the extrapolation of non-lifestyle-attributable mortality, which exhibits more linear past trends than all-cause mortality (especially from a historical point of view)(see Figure 2 and Figure 2 —figure supplement 1)(=previous Figure S2), the effect of the choice of the historical time period is smaller than it is in extrapolations of all-cause mortality. However, including only the trends from 1990 onwards has likely resulted in an underestimation of the effects of the separate extrapolation of non-lifestyle-attributable mortality and of lifestyle-attributable mortality, as for men in particular, the differences between the trends in e0 for all-cause and for non-lifestyle-related mortality would have been much larger if a longer past trend had been included. This is primarily because of the large increases in smoking-attributable mortality among men in many Western European countries, which resulted in stagnating increases in life expectancy in the 1960s and 1970s (see Figure 2); and because of the large increases in alcohol-attributable mortality among Eastern European men since 1975 (e.g. Trias-Llimós et al. 2017), which contributed as well to the past stagnating trends in life expectancy in these regions (Meslé et al. 2002).

We have now added a justification for our decision to use a relatively long projection horizon given the comparatively short historical period (see lines 170-173; appendix 1 – supplementary data and methods – page 2 paragraph 1). In addition, in the discussion, we explained that our approach – because it is focused on non-lifestyle-attributable mortality – is less dependent on the choice of the historical period (lines 360 – 363 and lines 443- 446).

The forecasting of lifestyle-attributable mortality indeed involves more subjectivity than the extrapolation of non-lifestyle-attributable mortality, as we acknowledge in our paper (see lines 419-421). To avoid generating unrealistic outcomes for the (long-term) future (e.g., continuing increases in smoking-attributable mortality for women, in obesity-attributable mortality, and in alcohol-attributable mortality in some Eastern European countries), we had to adopt approaches that did not simply extrapolate past trends, but that also took into account knowledge about the likely future progression of smoking-, alcohol-, and obesity-attributable mortality. In doing so, however, we still made as much use as possible of the past trends (for example, by making use of past trends for men in the case of smoking and of past trends in other European countries in the case of alcohol; and by adopting an extrapolation approach, but then applying it to the speed of change in the logit of prevalence for obesity to ensure a wave-shaped pattern, and by applying a quadratic curve to the past trends for populations that were still experiencing increases in smoking and obesity), and implemented our approach in frequently used models (e.g., the Lee-Carter model and age-period-cohort models). See Appendix 1 – Box 1 for a more extended summary of the different approaches used to project obesity-, smoking-, and alcohol-attributable mortality; and see Janssen, van der Broek et al., 2020; Janssen, El Gewily, Bardoutsos, 2020; and Janssen, El Gewily et al., 2020 for the actual projections, including a critical appraisal of each of them.

We have now extended the discussion of the projection of lifestyle-attributable mortality (see lines 413 – 422):

“To avoid generating unrealistic outcomes for the (long-term) future (e.g. continuing increases in smoking-attributable mortality for women, in obesity-attributable mortality, and in alcohol-attributable mortality in the few Eastern European countries where it is still increasing), the projections were based on advanced approaches that did not simply extrapolate past trends, but also took into account knowledge about the future progression of smoking-, alcohol-, and obesity-attributable mortality. Nonetheless, the projection of lifestyle-attributable mortality involves more subjectivity than the extrapolation of non-lifestyle-attributable mortality (see Janssen, van der Broek et al., 2020; Janssen et al., 2020; Janssen, El Gewily and Bardoutsos, 2020 for a critical evaluation)”.

Finally, the principal added value of the manuscript – the new approach to forecast mortality – is currently not available to the general public: "Because of the large size of the many different underlying simulation matrices, we are restricted in sharing these data. … The R codes can be requested from the author." One simple way to address this problem is to share a much smaller set of simulations (say 100 instead of 50,000) and warn the users to increase the number of simulations when running the codes on their machines. This would directly provide the research community with the routines to employ the proposed methodology.

We carefully considered what data we could share in order to make the new approach to forecasting mortality available to the general public. Although much of the information is already provided on Open Science Framework (https://osf.io/ghu45/) in the Excel file with the final data used as input for the tables and figures, we will now also provide information on the underlying observed age-specific mortality rates (all-cause mortality, non-lifestyle-attributable mortality, lifestyle-attributable mortality), as well as the fitted and projected age-specific mortality rates (by means of medians and 90% and 95% projection intervals). In addition, we will make the different R codes used for the different steps of the data analyses public.

In line with this change, we rewrote our data sharing statement: “Some of our original data regarding lifestyle-attributable mortality were based on previous publications, which, in turn, used data that are openly available. The all-cause mortality data and the exposure data can be obtained through the Human Mortality Database. We have provided source data files for all our tables and figures. These comprise the numerical data that are represented in the different figures, and the output on which the different tables are based. In addition, one excel file with all the final numerical / output data that were used as input for the tables and figures will be made available at the Open Science Framework: https://osf.io/ghu45/. In addition, we will upload there the underlying observed age-specific mortality rates (all-cause mortality, non-lifestyle-attributable mortality, lifestyle-attributable mortality) as well as the adjusted and projected age-specific mortality rates (medians and 90% and 95% projection intervals). The different R codes used for the different steps of the data analyses will be shared – as well – through the Open Science Framework link.

References

Bergeron-Boucher, M.-P., V. Canudas-Romo, J. Oeppen, and J. W. Vaupel (2017). Coherent forecasts of mortality with compositional data analysis. Demographic Research 37, 527-566.

Bohk-Ewald, C., M. Ebeling, and R. Rau (2017). Lifespan Disparity as an Additional Indicator for Evaluating Mortality Forecasts. Demography 54 (4), 1559-1577.

Lee, R. D. and L. R. Carter (1992). Modeling and forecasting US mortality. Journal of the American Statistical Association 87 (419), 659-671.

Shang, H. L., H. Booth, and R. Hyndman (2011). Point and interval forecasts of mortality rates and life expectancy: A comparison of ten principal component methods. Demographic Research 25, 173-214.

van Raalte, A. A., I. Sasson, and P. Martikainen (2018). The case for monitoring life-span inequality. Science 362 (6418), 1002-1004.

Reviewer #2 (Recommendations for the authors):

The authors have conducted an ambitious projection exercise, drawing together a number of methods and approaches to seek to understand the impacts of future changes in lifestyle- and non-lifestyle attributable mortality on prospects for life expectancy increase in a set of European countries. This focus on the underlying dynamics of mortality change--that is, disaggregating the long-term changes in non-lifestyle attributable mortality patterns from the much more variable lifestyle-attributable mortality patterns--represents a real step forward in life expectancy projection modeling. These methods could have considerable benefits for projections in other contexts.

My main issue with the paper as written concerns the assumptions built into the Li-Lee projection of non-lifestyle attributable mortality, where I would encourage the authors to test the sensitivity of their projections to alternative choices of reference countries (looking beyond Italy, Spain, and France).

Yes, the choice of the reference or forerunner populations is indeed an important aspect of multi-population or coherent mortality forecasting (Stoeldraijer 2019; Booth 2020). However, given that the trend in non-lifestyle-attributable mortality is more equal between countries and sexes than the trend in all-cause mortality (see Table 1), this choice will likely have a larger impact when coherently forecasting all-cause mortality than when coherently forecasting non-lifestyle-attributable mortality (see also Janssen and Kunst 2007 and Janssen et al. 2013). We selected women in France, Spain, and Italy as the forerunner populations in terms of life expectancy in Europe because they exhibit both very high recent life expectancy values and very favourable long-lasting past trends in e0 (measured by large increases in e0 over the 1970-2011 period among non-Eastern European countries) (Stoeldraijer 2019). Our decision to use forerunner populations as the reference populations is in line with the seminal work by James Vaupel (e.g., Oeppen and Vaupel 2002), and our decision to use female populations is in line with the recent work by Bergeron-Boucher et al. 2018. In addition, using three populations instead of one is regarded more robust.

An alternative approach would be to use the average mortality experience across many countries (Eurostat 2020; Stoeldraijer 2019), or – as indicated by reviewer 2 in the extended (non-public) explanation of this remark – to use the mortality experience of a set of low-mortality countries that have faced less severe life-course mortality pressures--e.g. Sweden, Norway, and Switzerland.

Author response table 1 reports on the outcomes of a sensitivity analysis in which we employed these two alternative selections of the reference population in coherent mortality projection. Our sensitivity analysis revealed that the use of the average mortality experience across all the included countries resulted – in line with our expectations – in lower projected e0 values in 2065, for both non-lifestyle attributable mortality and all-cause mortality. Furthermore our sensitivity analysis revealed that the use of women in Sweden, Norway and Switzerland as the reference population led to lower projected e0 values in 2065, but particularly so for all-cause mortality and much less so for non-lifestyle-attributable mortality. This latter outcome can be largely explained from the difference in the average (unweighted) gains in e0 over the 1990-2014 period for women in France, Italy, and Spain (4.8 for all-cause mortality; 5.0 for non-lifestyle-attributable mortality) compared to women in Sweden, Norway, and Switzerland (4.1 for all-cause mortality; 4.8 for non-lifestyle-attributable mortality).

The results of our sensitivity analysis, confirm that the use of forerunner populations instead of an average across many populations results in lower projected e0 values in 2065, and that the choice of the reference population will likely have a larger impact when coherently forecasting all-cause mortality than when coherently forecasting non-lifestyle-attributable mortality.

We have added a short discussion on the choice of the reference populations in the “Appraisal of our projection methodology” section of our revised manuscript (lines 384 – 396):

“However, as well as the potential effects on the outcomes of the exact coherent methodology that is applied (Stoeldraijer, 2019), it is important to take into account the choice of the reference population (Stoeldraijer 2019; Booth 2020). Using forerunner populations as the reference population, as advocated by among others Oeppen and Vaupel 2002 and Bergeron-Boucher et al., 2018, instead of an average across many populations (Li and Lee 2005; Eurostat 2020) will generally result in higher future life expectancy values (see the next section as well). However, given that the past trend in e0 for non-lifestyle-attributable mortality is more equal between countries and sexes than the past trend in e0 for all-cause mortality (see Table 1), which reference populations are chosen will likely have a larger impact when coherently forecasting all-cause mortality than when coherently forecasting non-lifestyle-attributable mortality (see also Janssen and Kunst 2007 and Janssen et al., 2013)”.

References

Edwards, R. D. and Tuljapurkar, S. (2005). Inequality in life spans and a new perspective on mortality convergence across industrialized countries. Population and Development Review, 31(4):645–674

Garcia, J., and Aburto, J.M. (2019). The impact of violence on Venezuelan life expectancy and lifespan inequality. International Journal of Epidemiology, 48(5), 1593-1601.

Meslé, F. and J. Vallin (2002). Mortality in Europe: The divergence between East and West. Population, 57: 157-197.

Technical Panel on Assumptions and Methods (2003). Report to the Social Security Advisory Board. Washington, D.C.

Trias-Llimós, S., Bijlsma, M. And Janssen, F. (2017). The role of birth cohorts in long-term trends in liver cirrhosis mortality across eight European countries. Addiction ,112(2), 250-258.

https://doi.org/10.7554/eLife.66590.sa2

Article and author information

Author details

  1. Fanny Janssen

    1. Netherlands Interdisciplinary Demographic Institute - KNAW/University of Groningen, The Hague, Netherlands
    2. Population Research Centre, Faculty of Spatial Sciences, University of Groningen, Groningen, Netherlands
    Contribution
    Conceptualization, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing
    For correspondence
    janssen@nidi.nl
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-3110-238X

  2. Anastasios Bardoutsos

    Population Research Centre, Faculty of Spatial Sciences, University of Groningen, Groningen, Netherlands
    Contribution
    Formal analysis, Validation, Investigation, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared

  3. Shady El Gewily

    Population Research Centre, Faculty of Spatial Sciences, University of Groningen, Groningen, Netherlands
    Contribution
    Formal analysis, Writing - review and editing
    Competing interests
    No competing interests declared

  4. Joop De Beer

    Netherlands Interdisciplinary Demographic Institute - KNAW/University of Groningen, The Hague, Netherlands
    Contribution
    Validation, Investigation, Writing - review and editing
    Competing interests
    No competing interests declared

Funding

Netherlands Organisation for Scientific Research (Innovational Research Incentives Scheme Vidi/452-13-001)

  • Fanny Janssen
  • Anastasios Bardoutsos
  • Shady El Gewily

Netherlands Organisation for Scientific Research (Innovational Research Incentives Scheme Vici/VI.C.191.019)

  • Fanny Janssen

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank Mark van der Broek (Research Master Economics, University of Groningen) for creating the final figures in R.

Senior Editor

  1. Eduardo Franco, McGill University, Canada

Reviewing Editor

  1. Belinda Nicolau, McGill University, Canada

Reviewers

  1. Ugofilippo Basellini
  2. Collin Payne

Version history

  1. Received: January 15, 2021
  2. Accepted: June 28, 2021
  3. Accepted Manuscript published: July 6, 2021 (version 1)
  4. Version of Record published: August 4, 2021 (version 2)

Copyright

© 2021, Janssen et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 3,460
    Page views
  • 209
    Downloads
  • 13
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Fanny Janssen
  2. Anastasios Bardoutsos
  3. Shady El Gewily
  4. Joop De Beer
(2021)
Future life expectancy in Europe taking into account the impact of smoking, obesity, and alcohol
eLife 10:e66590.
https://doi.org/10.7554/eLife.66590

Categories and tags

Research organism

Further reading

    1. Epidemiology and Global Health

    The importance of patient-partnered research in addressing long COVID: Takeaways for biomedical research study design from the RECOVER Initiative’s Mechanistic Pathways taskforce

    C Kim, Benjamin Chen ... RECOVER Mechanistic Pathways Task Force
    Review Article

    The NIH-funded RECOVER study is collecting clinical data on patients who experience a SARS-CoV-2 infection. As patient representatives of the RECOVER Initiative’s Mechanistic Pathways task force, we offer our perspectives on patient motivations for partnering with researchers to obtain results from mechanistic studies. We emphasize the challenges of balancing urgency with scientific rigor. We recognize the importance of such partnerships in addressing post-acute sequelae of SARS-CoV-2 infection (PASC), which includes ‘long COVID,’ through contrasting objective and subjective narratives. Long COVID’s prevalence served as a call to action for patients like us to become actively involved in efforts to understand our condition. Patient-centered and patient-partnered research informs the balance between urgency and robust mechanistic research. Results from collaborating on protocol design, diverse patient inclusion, and awareness of community concerns establish a new precedent in biomedical research study design. With a public health matter as pressing as the long-term complications that can emerge after SARS-CoV-2 infection, considerate and equitable stakeholder involvement is essential to guiding seminal research. Discussions in the RECOVER Mechanistic Pathways task force gave rise to this commentary as well as other review articles on the current scientific understanding of PASC mechanisms.

    1. Epidemiology and Global Health

    Nationwide mammography screening participation in Denmark during the COVID-19 pandemic: An observational study

    Tina Bech Olesen, Henry Jensen ... Sisse H Njor
    Research Article Updated

    Background:

    In most of the world, the mammography screening programmes were paused at the start of the pandemic, whilst mammography screening continued in Denmark. We examined the mammography screening participation during the COVID-19 pandemic in Denmark.

    Methods:

    The study population comprised all women aged 50–69 years old invited to participate in mammography screening from 2016 to 2021 in Denmark based on data from the Danish Quality Database for Mammography Screening in combination with population-based registries. Using a generalised linear model, we estimated prevalence ratios (PRs) and 95% confidence intervals (CIs) of mammography screening participation within 90, 180, and 365 d since invitation during the pandemic in comparison with the previous years adjusting for age, year and month of invitation.

    Results:

    The study comprised 1,828,791 invitations among 847,766 women. Before the pandemic, 80.2% of invitations resulted in participation in mammography screening within 90 d, 82.7% within 180 d, and 83.1% within 365 d. At the start of the pandemic, the participation in screening within 90 d was reduced to 69.9% for those invited in pre-lockdown and to 76.5% for those invited in first lockdown. Extending the length of follow-up time to 365 d only a minor overall reduction was observed (PR = 0.94; 95% CI: 0.93–0.95 in pre-lockdown and PR = 0.97; 95% CI: 0.96–0.97 in first lockdown). A lower participation was, however, seen among immigrants and among women with a low income.

    Conclusions:

    The short-term participation in mammography screening was reduced at the start of the pandemic, whilst only a minor reduction in the overall participation was observed with longer follow-up time, indicating that women postponed screening. Some groups of women, nonetheless, had a lower participation, indicating that the social inequity in screening participation was exacerbated during the pandemic.

    Funding:

    The study was funded by the Danish Cancer Society Scientific Committee (grant number R321-A17417) and the Danish regions.

    1. Epidemiology and Global Health
    2. Genetics and Genomics

    Within-host diversity improves phylogenetic and transmission reconstruction of SARS-CoV-2 outbreaks

    Arturo Torres Ortiz, Michelle Kendall ... Louis Grandjean
    Research Article

    Accurate inference of who infected whom in an infectious disease outbreak is critical for the delivery of effective infection prevention and control. The increased resolution of pathogen whole-genome sequencing has significantly improved our ability to infer transmission events. Despite this, transmission inference often remains limited by the lack of genomic variation between the source case and infected contacts. Although within-host genetic diversity is common among a wide variety of pathogens, conventional whole-genome sequencing phylogenetic approaches exclusively use consensus sequences, which consider only the most prevalent nucleotide at each position and therefore fail to capture low frequency variation within samples. We hypothesized that including within-sample variation in a phylogenetic model would help to identify who infected whom in instances in which this was previously impossible. Using whole-genome sequences from SARS-CoV-2 multi-institutional outbreaks as an example, we show how within-sample diversity is partially maintained among repeated serial samples from the same host, it can transmitted between those cases with known epidemiological links, and how this improves phylogenetic inference and our understanding of who infected whom. Our technique is applicable to other infectious diseases and has immediate clinical utility in infection prevention and control.