Alcohol and Alcoholism, 2020, 1–9 doi: 10.1093/alcalc/agaa107 Article Article Future Alcohol-Attributable Mortality in France Downloaded from https://academic.oup.com/alcalc/advance-article/doi/10.1093/alcalc/agaa107/5934607 by guest on 05 March 2021 Using a Novel Generalizable Age-Period-Cohort Projection Methodology 1 ,2 , Sergi Trias-Llimós *, Anastasios Bardoutsos1 , and Fanny Janssen 1 ,3 1 Population Research Centre, Faculty of Spatial Sciences, University of Groningen, Landleven 1, 9747 Groningen, The Netherlands, 2 Center for Demographic Studes, Centres de Recerca de Catalunya (CERCA), Carrer de Ca n’Altayó, Edifici E2, 08193 Bellaterra, Spain, and 3 Netherlands Interdisciplinary Demographic Institute, KNAW/University of Groningen, Lange Houtstraat 19, 2511 The Hague, The Netherlands *Corresponding author: Center for Demographic Studes, Centres de Recerca de Catalunya (CERCA), Carrer de Ca n’Altayó, Edifici E2, 08193 Bellaterra, Spain. Tel.: 0034 935813060; E-mail: strias@ced.uab.cat Received 11 May 2020; Revised 22 September 2020; Editorial Decision 22 September 2020; Accepted 22 September 2020 Abstract Aim: To forecast age- and sex-specific alcohol-attributable mortality in France for the period 2015– 2050 using a novel generalizable methodology that includes different scenarios regarding period and cohort change. Methods: For the French national population aged 25–90 years (1979–2014), we estimated alcohol- attributable mortality by mortality from the main causes of death wholly attributable to alcohol, plus liver cirrhosis mortality. We modelled sex-specific alcohol-attributable mortality by adjusting for age, period and birth cohort. We forecasted the model parameters to obtain future age- and sex-specific alcohol-attributable mortality up until 2050 using a conventional baseline, scenario I (favourable period change) and scenario II (unfavourable cohort change). Results: Alcohol-attributable mortality is clearly declining in France, with the decline decelerating from 1992 onwards. In 2014, the age-standardized alcohol-attributable mortality rates, in deaths per 100,000, were 34.7 among men and 9.9 among women. In 2050, the estimated rates are between 10.5 (prediction interval: 7.6–14.4; scenario I) and 17.6 (13.1–23.7; scenario II) among men, and between 1.1 (0.7–1.7; scenario I) and 1.8 (1.2–2.9; scenario II) among women; which implies declines of 58% for men and 84% for women (baseline). Conclusion: Alcohol-attributable mortality in France is expected to further decline in the coming decades, accompanied by age pattern changes. However, France’s levels are not expected to reach the current lower levels in Italy and Spain for 15 years or more. Our results point to the value of implementing preventive policy measures that discourage alcohol consumption among people of all ages, but especially among adolescents. INTRODUCTION As alcohol consumption has a marked impact on overall mortality Alcohol consumption is a major public health issue in Europe (Franco levels and trends (Trias-Llimós, Kunst, et al., 2018), understanding 2015; WHO 2019), where the levels of both alcohol consumption how this impact will develop in the future is relevant for public and attributable mortality are the highest worldwide (WHO 2019). health professionals and policy-makers aiming to reduce alcohol- © Oxford University Press and Foundation for the European Review of Agricultural Economics 2020; all rights reserved. For permissions, please e-mail: journals.permissions@oup.com This is an Open Access article distributed under the terms of the Creative Commons Attribution NonCommercial-NoDerivs licence 1 (http://creativecommons.org/licenses/by-nc-nd/4.0/), which permits non-commercial reproduction and distribution of the work, in any medium, provided the original work is not altered or transformed in any way, and that the work is properly cited. For commercial re-use, please contact journals.permissions@oup.com 2 Alcohol and Alcoholism, 2020 attributable mortality, and, consequently, the impact of alcohol on Database. More detailed national cause-of-death data, available for overall mortality. However, there is a lack of knowledge about future France, are therefore required. trends in alcohol-attributable mortality. When estimating future alcohol-attributable mortality, it is essential to understand past trends, which differ widely across MATERIALS AND METHODS European countries depending on their drinking cultures. Alcohol- Setting and study population attributable mortality has been gradually decreasing in Southern For the total population of men and women in France aged 25–90, Europe since around the 1970s, but the long-term trends in other we studied past trends in overall and age-specific alcohol-attributable European countries have been more irregular (Kraus et al., 2015; mortality (1979–2014) and estimated future overall and age-specific Trias-Llimós et al., 2017). In examining past trends in alcohol- alcohol-attributable mortality (2015–2050). Downloaded from https://academic.oup.com/alcalc/advance-article/doi/10.1093/alcalc/agaa107/5934607 by guest on 05 March 2021 attributable mortality, birth cohorts must be taken into account (Kraus et al., 2015; Trias-Llimós et al., 2017). Birth cohort differences in the adoption of unhealthy lifestyles, including heavy drinking, are Data often attributed to certain birth cohorts facing similar contextual We used detailed cause-of-death data by single year of age and sex situations at different ages than other birth cohorts. In addition, we for the 1979–2014 period from the Center for Epidemiology on the know that alcohol use in younger adulthood is associated with overall Medical Causes of Death (CépiDc) at the French National Institute of drinking patterns over the life course (Pitkänen et al., 2005) and with Health and Medical Research (Inserm) (Rey 2016). In line with previ- the development of alcohol-related problems (Hingson et al., 2006; ous studies, alcohol-attributable mortality was defined as the sum of Dawson et al., 2008). Thus, when forecasting alcohol-attributable deaths from the main causes wholly attributable to alcohol (mental mortality, it may be useful to take birth cohorts into account. and behavioural disorders due to alcohol use, alcoholic liver disease Despite the important role birth cohorts play in determining and alcohol poisoning) and liver cirrhosis deaths (ICD-9 codes: 291, alcohol-attributable mortality trends, only one previous study barely 303, 571 and E860-E866; ICD-10 codes: F10, G31.2, K70, K73, K74 touched upon this dimension when discussing future levels of alcohol- and X45), as similarly done in previous studies (Jasilionis et al., 2011; attributable mortality (Rosén and Haglund 2019). By assuming Trias-Llimós, Martikainen, et al., 2018). We used the corresponding age effects and cohort effects to be constant over time, which is population data from the Human Mortality Database to estimate arguable over the short term, it has been projected that alcohol- alcohol-attributable mortality rates (HMD 2017). attributable mortality will decline moderately in Sweden between 2015 and 2025 (Rosén and Haglund 2019). Other efforts to project Approach alcohol-attributable mortality overlooked the cohort dimension. For example, Pruckner et al. (2019) linearly projected declines between We applied a generalized methodology to forecast (age-specific) 1979 and 2015 in age-standardized death rates for selected alcohol- alcohol-attributable mortality that uses APC modelling and includes related causes of death up until 2030 for 29 European countries, but different likely scenarios for the period and cohort dimensions. A they did not account for the cohort dimension. Other projections of crucial part of this methodology is the careful assessment of past alcohol-attributable mortality failed to simultaneously account for trends prior to the formulation of the assumptions for the base- both period and cohort effects. Sheron et al. (2011, 2012) used linear line projection and the likely scenarios based on observed data. extrapolation techniques to project alcohol-attributable liver deaths In the remainder of the methodology section, we will describe the in the UK up to 2030 while also applying past time trends from other methodology used to study past trends (APC modelling), and describe European countries to the UK. Thus, projections of future alcohol- the methodology and assumptions used to estimate future alcohol- attributable mortality are currently available for two countries only attributable mortality levels and trends (projection methodology). and for relatively short time periods. It should also be noted that in addition to being based on arguable assumptions; these results were APC modelling not provided disaggregated by age group. We estimated yearly sex-specific age-standardized alcohol-attributable We aim to forecast overall and age-specific alcohol-attributable mortality, using the total French population of 2014 as the standard. mortality in France by sex over a long period (2014–2050) using We modelled alcohol-attributable mortality by means of APC age-period-cohort (APC) modelling while including different likely modelling scenarios for the period and cohort dimensions. To do so, we have log μalc x,t = αx + κt + γt−x , used a methodology that can be applied to other European countries. France is a country of particular interest as, even though where μalcx,t are alcohol attributable mortality rates by age (x) and alcohol consumption has declined in France in recent decades year (t). The parameters αx , κt and γt−x capture the age pattern, the (Gual and Colom 1997; WHO 2019), its current levels of alcohol overall time trend (period) and the cohort patterns, respectively. We consumption are still high (WHO 2019). Moreover, France has higher assume that alcohol-attributable deaths follow a Poison distribution. levels of alcohol-attributable mortality than most other Western To deal with the linear dependency between period and birth European countries (Trias-Llimós, Kunst, et al., 2018). According cohort (age = period − cohort) in the APC modelling, we used to recent estimates for France, 13% of deaths among men and the approach by Cairns et al. (2009). This method builds on the 5% of deaths among women are attributable to alcohol (Guerin commonly applied Clayton and Schifflers APC approach (Clayton et al., 2013). Finally, in order to carefully study past trends and and Schifflers 1987), which distinguishes between period, cohort and to adequately capture cohort effects, information about long-term drift. One additional step is performed to move the drift (=shared lin- trends in alcohol-attributable mortality is needed. However, because ear trend between period and cohort) completely to either the period of changes in the International Classification of Diseases (ICD) or the cohort trend. This is especially beneficial when the aim is to (Janssen and Kunst 2004), these trends cannot be directly estimated extrapolate past trends because you end up with two time-dependent from cause-of-death data from the international WHO Mortality factors that are easy interpretable and that can meaningfully be Alcohol and Alcoholism, 2020 3 extrapolated into the future (Janssen et al., 2020). In our application For each scenario, we forecasted the period and cohort parameter of the Cairns et al.’s approach, we moved the drift to the period by using autoregressive integrated moving average (ARIMA) time trend. The period parameter thus captures the entire linear time series models. We selected the best ARIMA model based on the trend (including the drift), while the cohort parameter captures the corrected Akaike’s information criterion (Hurvich and Tsai 1989) cohort variations from this overall trend (see (Cairns et al., 2009) for and the Hyndman-Khandakar algorithm (Hyndman and Khandakar further details). We used, for this purpose, the Stochastic Mortality 2007). For the implementation of the algorithm, we used the Modelling (StMoMo) package (Villegas et al., 2015) in R (R Core auto.arima function from package forecast in R (R Core Team 2018). Team 2018). See Table 1 for further details on the specific time series choices for To ensure the robustness of the cohort estimates, we excluded the different scenarios. from the model the first and the last five birth cohorts, which We projected age-specific alcohol-attributable mortality rates up Downloaded from https://academic.oup.com/alcalc/advance-article/doi/10.1093/alcalc/agaa107/5934607 by guest on 05 March 2021 were highly uncertain (wide confidence intervals) due to the reduced to 2050 for the baseline and the two scenarios and estimated the number of available data points (see Fig. S1). This resulted in the respective 95% prediction intervals by performing 10,000 simula- inclusion of the cohorts born between 1894 and 1984. tions. The projected rates were subsequently age-standardized using the total French population of 2014 as the standard. Projection methodology Based on the past trends analysis (see results section—past trends), RESULTS we formulated a baseline projection and two likely scenarios. Past trends The baseline projection used common settings in APC mortality Alcohol-attributable mortality was clearly declining in France (1979– projections and in general (mortality) forecasting. That is, we 2014), but with two distinct trends: (a) a pronounced decline from assumed that the birth cohort parameter was constant at the mean 1979 to 1992 and (b) a slower decline from 1992 to 2014, especially observed over the past period (Stoeldraijer et al., 2015; Rosén and among men (Fig. S2). Haglund 2019). For the period parameter, we considered the trend When looking at the period trend in our APC model, we see that for the most recent period (1992–2014), which represents the period the decline was larger among women than among men (see Fig. 1). after the clear trend break in 1992 (see Fig. 3) (Janssen and Kunst When examining the general decline in alcohol-attributable mortality 2007; Sheron et al., 2011). over time, we observe that alcohol-attributable mortality increased Scenarios I (period) and II (cohort) account for likely changes among the cohorts born before 1955, then decreased among the in the period and the cohort dimension, respectively, relative to the cohorts born between 1955 and 1975, and increased for the more baseline projection. Scenario I represents a positive period change recent birth cohorts (Fig. 1). (stronger decline), whereas scenario II is a negative cohort scenario. In scenario I (period), we assumed a more pronounced decline in alcohol-attributable mortality than in the more recent period, given Future trends the possibility that favourable contextual changes or more effective The results for the projected period and birth cohort parameter for preventive policies will favourably affect the drinking behaviour the different scenarios are shown in Fig. 2. of people in all age groups. This is an optimistic, yet plausible, The results showed that the age-standardized alcohol-attributable scenario given the growing awareness of the health damage of alcohol mortality rates will continue to decline for both men and women, consumption (Rehm et al., 2017), and the potential of contextual albeit at different speeds depending on the period and cohort assump- and policy changes to reduce alcohol use among adolescents (Noel tions (Fig. 3, Table S1 for the prediction intervals). 2019) and in France (Cogordan et al., 2014). In addition, France The baseline scenario projected a decline in age-standardized has currently higher alcohol consumption levels as compared with alcohol-attributable mortality, by deaths per 100,000, from 34.7 in similar countries (Spain and Italy) (WHO Global Health Observatory 2014 to 14.7 (prediction interval: 11.2–19.6) in 2050 among men and Data Repository, n.d.), and therefore higher possibilities for reducing from 9.9 in 2014 to 1.6 (1.0–2.4) in 2050 among women; or by 58% alcohol consumption. In this scenario, we therefore extrapolate the among men and 84% among women. The gender gap in alcohol- stronger declines over the whole 1979–2014 period compared with attributable mortality is projected to narrow over time in absolute the less strong declines over the more recent period (1992–2014). terms, from 25 deaths per 100,000 in 2014 to 13 in 2050 (baseline In scenario II (cohort), we assumed that the increase in mortality scenario). among the most recent cohorts (as seen in Fig. 1) will also have Scenario I (favourable period change) obviously resulted in larger negative effects on mortality in the subsequent birth cohorts, given declines and lower future age-standardized alcohol-attributable mor- the possibility that negative changes in drinking behaviour (and tality in 2050: 10.5 (7.6–14.1) among men and 1.1 (0.7–1.7) among especially binge drinking) will continue across generations of young women. Scenario II (unfavourable cohort change) resulted in smaller people. This scenario is supported by the evidence suggesting that declines and higher future levels in 2050: 17.6 (13.1–23.7) among drinking patterns are becoming riskier among young generations men and 1.8 (1.2–2.9) among women. As scenario II estimated (Richard et al., 2015), which can also explain the increase in mor- higher mortality for the currently young generations only, but cohort tality among the most recent cohorts. Nonetheless, recent evidence mortality that is almost identical to the baseline projection (same suggests that alcohol consumption among young adults has been period assumptions) in the long run (e.g. cohorts born in 2020), the declining over time (Com-Ruelle and Le Guen 2013; WHO 2018). overall estimates from the baseline projection and from scenario II Therefore, we assumed elevated—but converging—mortality levels tended to converge in the long run (Fig. S3). compared with earlier cohorts for—at least—the future 35 cohorts. Whereas in 2014, an inverse u-shaped age pattern of alcohol- We implemented this in our model by allowing the cohort parameter attributable mortality was observed with a peak at around age 65, to gradually converge towards the mean, instead of being fixed at this peak is projected in each projection scenario to occur at higher the mean. ages among men up to about 2040, and among women at least up 4 Alcohol and Alcoholism, 2020 Downloaded from https://academic.oup.com/alcalc/advance-article/doi/10.1093/alcalc/agaa107/5934607 by guest on 05 March 2021 Fig. 1. Period and (non-linear) birth cohort effects of alcohol-attributable mortality in France (ages 25–90, 1979–2014), by sex. Table 1. Summary of scenarios and time-series specifications for the projection of the period and cohort parameters Projection Period parameter Time series Cohort parameter (γt−x ) Time series (κt ) specification period specification cohort Baseline 1992–2014 ARIMA (0,1,0) Men: constant along the mean (0). ARIMA (0,0,0) Women: constant along the mean for the last 10 cohorts (−0.56). Scenario I (period) 1979–2014 ARIMA (0,1,0) Men: constant along the mean (0). ARIMA (0,0,0) Women: constant along the mean for the last 10 cohorts (−0.56). Scenario II (cohort) 1992–2014 ARIMA (0,1,0) Men: gradually converging ARIMA (1,0,2) towards the mean. Women: same pattern as for men, but shifted downwards based on the differences in the mean between men and women (last 10 birth cohorts). ∗ For women, unlike as for men, the average for the most recent birth cohorts is substantially lower than the overall average (zero). In that case, we assumed the mean to be the mean over the last 10 birth cohorts. until 2050, which results in a linear increase with age for women in DISCUSSION 2050 (Fig. 4). Furthermore, a second sharp peak in the age pattern, which occurs at higher ages in later years, was observed. This peak Summary of results was especially visible in the baseline projection and period scenario, Alcohol-attributable mortality was clearly declining among the in which the cohort parameter was set to the mean observed over French population aged 25–90 in the 1979–2014 period, with a the past period, in line with common settings in APC mortality deceleration of the decline from 1992 onwards. In 2014, the age- projections (Stoeldraijer et al., 2015; Rosén and Haglund 2019). In standardized rates, in deaths per 100,000, were 34.7 among men and the cohort scenario, the peak was more extended and more realistic, 9.9 among women. Over the 2014–2050 period, the age-standardized as here it was assumed that cohort mortality would gradually decline alcohol-attributable mortality rates are projected to further decline towards the mean. by 58% for men and 84% for women (baseline). This implies a Alcohol and Alcoholism, 2020 5 Downloaded from https://academic.oup.com/alcalc/advance-article/doi/10.1093/alcalc/agaa107/5934607 by guest on 05 March 2021 Fig. 2. Alcohol-attributable mortality in France (ages 25–90): age, period and cohort parameters and projections of period and cohort parameters, by sex. Fig. 3. Age-standardized alcohol-attributable mortality rates (ages 25–90) in France (1979–2014 data; 2015–2050 projection), by sex. 6 Alcohol and Alcoholism, 2020 Downloaded from https://academic.oup.com/alcalc/advance-article/doi/10.1093/alcalc/agaa107/5934607 by guest on 05 March 2021 Fig. 4. Observed (2014) and projected (2025, 2040, 2050) age-specific alcohol-attributable mortality rates (25–90) in France, by sex. range in 2050, in deaths per 100,000, of between 10.5 (7.6–14.4; mortality trends from the selected causes of death, and especially liver scenario I period) and 17.6 (13.1–23.7; scenario II cohort) among cirrhosis, followed past overall alcohol consumption trends (Ramst- men, and between 1.1 (0.7–1.7; scenario I period) and 1.8 (1.2–2.9; edt 2001). Second, our method for estimating alcohol-attributable scenario II cohort) among women. The peak of the inverse u-shaped mortality seems to have a rather minor effect on time trends if we age pattern of alcohol-attributable mortality (currently at around age compare our data to the Global Burden of Diseases (GBD) estimates, 65) is expected to shift towards older ages, and an additional hump which are based on population attributable fraction approaches in the age pattern is projected that moves towards higher ages with (Forouzanfar et al., 2016). That is, the overall decline in alcohol- time, and is more extended in the cohort scenario. attributable mortality between 1990 and 2013 among men is 40.5% in the GBD estimates (Trias-Llimós, Kunst, et al., 2018), and 41.7% Reflection on data in our study (36.5 and 47.3% for women, respectively). Therefore, In line with previous research, we used the main causes of death our estimates of the past and future alcohol-attributable mortality wholly attributable to alcohol, supplemented with liver cirrhosis decline can be considered good approximations of the overall decline mortality, to estimate alcohol-attributable mortality (Jasilionis et al., in alcohol-attributable mortality in relative terms. Explanation of 2011; Trias-Llimós, Martikainen, et al., 2018). The selected causes results. The projected future decline in alcohol-attributable mortality of death wholly attributable to alcohol and liver cirrhosis (F10, is based on the extrapolation of past declines in alcohol-attributable G31.2, K70, K73, K74, and X45) represent >90% of the total mortality. These past declines in alcohol-attributable mortality can deaths wholly attributable to alcohol in France (Trias-Llimós, Mar- be mainly attributed to declines in overall alcohol consumption in tikainen, et al., 2018). Data for the remaining very specific causes France, from 19.4 l of pure alcohol consumption per capita and of death wholly attributable to alcohol (Rehm et al., 2017) were year in 1979 to 12.0 l in 2014 (WHO Global Health Observatory not available to us. Also, the inclusion of liver cirrhosis mortality Data Repository, n.d.). These declines were driven by an overall as a whole, which is mostly—but not always—driven by alcohol reduction in daily wine consumption (Gual and Colom 1997; WHO (Trias-Llimós et al., 2017), was necessary as it was not possible to 2019) in a context of societal and economic changes (e.g. urban- build coherent time trends purely for alcoholic liver cirrhosis. Over- ization and increased female employment) (Cogordan et al., 2014). all, our estimate of alcohol-attributable mortality underestimates In that sense, we assume that societal and economic change will total alcohol-attributable mortality (Trias-Llimós, Martikainen, et al., lead to a continued decline in alcohol-attributable mortality in the 2018) because alcohol has a notable impact on other causes as future. well, such as ischemic diseases, cancers and injuries (Rey et al., Our APC analysis however also revealed a potential important 2010; Guerin et al., 2013; Rehm et al., 2017). The extent of this deviation from the overall decline. That is, the most recent birth underestimation in 2010 for France was estimated to be around cohorts showed relatively higher alcohol-attributable mortality than 58% relative to estimates using attributable fraction methods, and their older counterparts. This relates as well to trends in alcohol between 27% (women) and 42% (men) relative to estimates from consumption. Alcohol consumption among adolescents and young approaches using underlying and contributory causes of death (Tri- adults have also been moderately declining over time (Com-Ruelle as-Llimós, Martikainen, et al., 2018). Therefore, the results of this and Le Guen 2013; WHO 2018), and therefore contributed as well paper on alcohol-attributable mortality levels should be interpreted to the observed decline in alcohol-attributable mortality over time. as strictly referring to past and future mortality due to causes wholly However, recent trends for self-reported drunkenness episodes sug- attributable to alcohol and liver cirrhosis. gested increases among 18–25 year old from 2005 to 2015 (Richard The past and future time trends in alcohol-attributable mortality et al., 2015) and a stagnation or a moderate decline among 15-year- are, however, much less affected by our estimation. First, the past olds (WHO 2018). These more harmful drinking patterns are likely Alcohol and Alcoholism, 2020 7 the result of a complex combination of determinants, including an and the period scenario, this represents a very abrupt peak, mainly earlier onset of drinking and unfavourable changes in social norms because mortality for the younger observed cohorts is higher than the (WHO 2018). At young ages, alcohol-related mortality is driven average level, and because we assumed that the younger cohorts will mostly by external causes of death (Trias-Llimós, Martikainen, et al., have the same cohort mortality as the average level of the previous 2018), which are more closely related with drinking patterns (e.g. cohorts, as commonly done (Rosén and Haglund 2019). In our cohort drunkenness) (compared with overall alcohol consumption levels) scenario, we instead adopted a more gradual change of the cohort (Rehm et al., 2017). Explanations for this complex trend include parameter, which is likely to result in an extended hump and reflects age at drinking onset, family life and societal norms and rules the decline in alcohol use observed among the generations born in (WHO 2018). These somewhat contrasting trends among young late 1990s and 2000s (WHO 2018). individuals (a decrease in overall consumption, and an increase of Downloaded from https://academic.oup.com/alcalc/advance-article/doi/10.1093/alcalc/agaa107/5934607 by guest on 05 March 2021 riskier drinking patterns) suggest that adolescents drink less often, but more harmfully. We incorporated the potential increase in riskier Appraisal of the methodology drinking patterns in our cohort scenario, leading to—still—a general To forecast (age-specific) alcohol-attributable mortality, we applied decline in alcohol-attributable mortality but to elevated risks among a methodology that uses APC modelling and includes different likely younger birth cohorts and—as a result—a changing age pattern of scenarios for the period and cohort dimensions. alcohol-attributable mortality. Our main rationale for applying an APC projection methodol- Despite the overall expected further decline in alcohol-attributable ogy was that previous research demonstrated that the birth cohort mortality in France, it might take some time before France reaches dimension has clear added value in describing and understanding the currently observed lower levels of alcohol-attributable mortality past alcohol-attributable mortality trends (Kraus et al., 2015; Tri- in Spain and Italy. In fact, we project that France will not reach as-Llimós et al., 2017). Moreover, in previous forecasts of alcohol- Spain’s current levels of alcohol-attributable mortality (22.0 deaths attributable mortality (Rosén and Haglund 2019) and of smoking- per 100,000) until 2032 and will not reach Italy’s current levels attributable mortality (Stoeldraijer et al., 2015), APC projection (17.5 deaths per 100,000) before 2040 (Fig. S4). This is because was applied using the same rationale. Our ‘explanation of results’ France historically has had and currently still has higher levels of section shows the added value of including the cohort dimension both alcohol consumption per capita and attributable mortality when forecasting alcohol-attributable mortality, as cohort effects (Trias-Llimós, Kunst, et al., 2018). This can be related, in particular, to were found to have a substantial impact on the projected age pattern the higher wine and spirits consumption in France as compared with in alcohol-attributable mortality at the population level. Spain and Italy (WHO Global Health Observatory Data Repository, To deal with the linear identification problem (period = cohort n.d.). According to our baseline projection for men, and using the + age) in APC modelling, we applied the approach developed by same definition for alcohol-attributable mortality, France will not Cairns et al. (2009). Their approach allows to allocate the com- reach Spain’s current levels (22.0 deaths per 100,000) until 2032 plete linear trend (including the drift) to the period parameter in and will not reach Italy’s current levels (17.5 deaths per 100,000) a straightforward manner, leaving the cohort parameter as purely before 2040 (see Fig. S4). Thus, even though alcohol-attributable non-linear, which is beneficial when projecting (Janssen et al., 2020). mortality is expected to further decline in France, it will not decrease Other approaches presented in the literature deal with the linear to the current levels in these neighbouring countries in the next identification problem in other, mostly more complicated, ways (Yang decade. et al., 2004; Carstensen 2007). Generally, different APC approaches In addition to a further decline in alcohol-attributable mortality, generate similar outcomes when applied to past trends (Yang et al., we predict a changing age pattern. In particular, we project that 2004). One could therefore also use a different APC model as part of the peak of the inverse u-shaped age pattern of alcohol-attributable our general projection methodology. Applying models with multiple mortality (currently at around age 65) will shift towards older ages. period or cohort parameters, however, will require more assumptions, For men, this seems to be a temporal phenomenon only (up to about could make projections less transparent and could result in different 2040), whereas for women, it lasts at least up until 2050, with the projection outcomes. peak expected to occur at age 85. For example, the cohorts born In contrast to the only previous projection of alcohol-attributable around 1955 (aged 70 in 2025 and aged 85 in 2040), and especially mortality that included the cohort dimension (Rosén and Haglund the women in these cohorts, are at higher risk of alcohol-attributable 2019), we considered different assumptions for both period and mortality than other observed cohorts (Fig. 2), as the observed age birth cohort based on a careful study of past trends. Whereas the patterns show (Fig. 4). In both cases, this is entirely explained by the assumption of the period had an effect on the overall levels (in cohort effects. Thus, these relative changes in age patterns are not this study, lower mortality levels), the assumption of the cohort had observed in the very long term, when cohort differences tend to be an effect on both the overall levels (in this study, higher mortality zero (see Fig. S5). levels) and the age pattern of alcohol-attributable mortality. The Furthermore, the additional humps in the projected age pattern application of our approach allowed us to provide new insights can be entirely explained by cohort effects. That is, for each subse- into the importance of the assumptions for both overall and age- quent period, this hump occurs at a higher age (at age 42 for 2025, specific future alcohol-attributable mortality estimates. In particular, at age 57 for 2040, at age 67 for 2050), at least for the baseline our results showed that the use of a constant cohort pattern, which projection and the period scenario. The birth cohorts causing this is often recommended in the literature, may result in an abrupt— pattern are the youngest birth cohorts we observed; i.e. those born and unlikely—hump in the projected age pattern. Thus, for future in the early 1980s. The finding that alcohol-attributable mortality APC projections, we recommend that more realistic assumptions are is higher among the younger generations than among the preceding used for the continuation of the cohort parameter. To make such generations could be explained by the increase of riskier drinking assumptions, a careful study of past trends is essential. patterns (e.g. drunkenness) among the generations born in the 1980s Our APC approach, which used different assumptions for the and early 1990s (Richard et al., 2015). For the baseline projection period and cohort dimensions, can be generalized and applied to 8 Alcohol and Alcoholism, 2020 other populations. However, to make logical assumptions for the CONFLICT OF INTEREST STATEMENT projection and achieve realistic results for the future, examining past The authors declare that they have no competing interests. trends for the population under study is crucial. ETHICS APPROVAL AND CONSENT TO CONCLUSION PARTICIPATE In conclusion, alcohol-attributable mortality in France is expected Not applicable. to further decline in the coming decades, although France is not expected to reach the current lower levels of neighbouring countries Downloaded from https://academic.oup.com/alcalc/advance-article/doi/10.1093/alcalc/agaa107/5934607 by guest on 05 March 2021 with similar drinking cultures within the next 10 years. Furthermore, the peak of the inverse u-shaped age pattern of alcohol-attributable CONSENT FOR PUBLICATION mortality (currently at age 65) is expected to shift towards older ages Not applicable. because alcohol-attributable mortality is higher among cohorts born around 1955 than among the other observed cohorts (except the male cohorts born in the early 1980s). Given its profound effects on future mortality trends and the REFERENCES unfavourable age patterns for current generations, health policy- Anderson P, Chisholm D, Fuhr DC. (2009) Effectiveness and cost-effectiveness makers should seek to prevent elevated alcohol-attributable mor- of policies and programmes to reduce the harm caused by alcohol. Lancet 373:2234–46. tality in the medium term by tackling alcohol consumption among Cairns AJ, Blake D, Dowd K, et al. (2009) A quantitative comparison of adolescents (cohort measures). In addition, the implementation of stochastic mortality models using data from England and Wales and the preventive policy measures that affect people in all age groups United States. North Am Actuar J 13:1–35. (period measures; i.e. taxation) (Anderson et al., 2009) could pos- Carstensen B. (2007) Age–period–cohort models for the lexis diagram. Stat itively contribute to alcohol-attributable mortality declines in the Med 26:3018–45. short term. Clayton D, Schifflers E. (1987) Models for temporal variation in cancer rates. II: age–period–cohort models. Stat Med 6:469–81. Cogordan C, Kreft-Jaïs C, Guillemont J. (2014) Effects of alcoholic bev- erage control policies and contextual factors on alcohol consumption SUPPLEMENTARY MATERIAL and its related harms in France from 1960 to 2000. Subst Use Misuse Supplementary material is available at Alcohol and Alcoholism 49:1633–45. online. Com-Ruelle L, Le Guen N. (2013) Les jeunes et l’alcool: évolution des comportements, facteurs de risque et éléments protecteurs. In Quest Déconomie Santé. Dawson DA, Goldstein RB, Patricia Chou S, et al. (2008) Age at first drink and AVAILABILITY OF DATA AND MATERIAL the first incidence of adult-onset DSM-IV alcohol use disorders. Alcohol The data were obtained from the following sources: Inserm CépiDc Clin Exp Res 32:2149–60. (contact persons: Mireille EB, Grégoire Rey) and Human Mortality Forouzanfar MH, Afshin A, Alexander LT et al. (2016) Global, regional, Database (publicly available at www.mortality.org). and national comparative risk assessment of 79 behavioural, envi- ronmental and occupational, and metabolic risks or clusters of risks, 1990–2015: a systematic analysis for the global burden of disease study 2015. Lancet 388: 1659–1724. AUTHORS’ CONTRIBUTIONS Franco S. (2015) Tackling Harmful Alcohol Use: Economics and Public Health All authors designed the study. S.T.L. and A.B. carried out the Policy. OECD Publishing. Gual A, Colom J. (1997) Why has alcohol consumption declined in countries empirical analyses. S.T.L. drafted the manuscript. F.J. aided in drafting of southern Europe? Addiction 92:S21–31 Wiley Online Library. the manuscript and critically reviewed the manuscript. All authors Guerin S, Laplanche A, Dunant A, et al. (2013) Alcohol-attributable mortality aided interpreting the results, and approved the final manuscript. in France. Eur J Public Health 23:588–93. Hingson RW, Heeren T, Winter MR. (2006) Age at drinking onset and alcohol dependence. Arch Pediatr Adolesc Med 160:8. ACKNOWLEDGEMENTS Human Mortality Database. (2017) University of California, Berkeley (USA), and Max Planck Institute for Demographic Research (Germany). Avail- The authors thank Mireille EB and Grégoire Rey (Inserm—CépiDc, able at www.mortality.org or www.humanmortality.de. France) for sharing data with us. The authors are grateful to Leo van Hurvich CM, Tsai C-L. (1989) Regression and time series model selection in Wissen (University of Groningen and Netherlands Interdisciplinary small samples. Biometrika 76:297–307. Demographic Institute) and to Joop de Beer (Netherlands Interdis- Hyndman RJ, Khandakar Y. (2007) Automatic time series for forecasting: ciplinary Demographic Institute) for their useful comments on an the forecast package for R. Department of Econometrics and Business earlier version of this work. Statistics, Monash University. Janssen F, Gewily SE, Bardoutsos A. (2020) Smoking epidemic in Europe in the 21st century. Tob Control (Online First: 07 August 2020) . doi: 10.1136/tobaccocontrol-2020-055658. FUNDING Janssen F, Kunst A. (2007) The choice among past trends as a basis for the This work is financed by the Netherlands Organisation for Scientific Research prediction of future trends in old-age mortality. Popul Stud 61:315–26. (NWO) in relation to the research programme ‘Smoking, alcohol, and obesity, Janssen F, Kunst AE. (2004) ICD coding changes and discontinuities in trends ingredients for improved and robust mortality projections’, grant no. 452–13- in cause- specific mortality in six European countries, 1950–99. Bull 001. See www.futuremortality.com. World Health Organ 12:904–13. Alcohol and Alcoholism, 2020 9 Jasilionis D, Meslé F, Shkolnikov VM, et al. (2011) Recent life expectancy Rosén M, Haglund B. (2019) Follow-up of an age-period-cohort analysis on divergence in Baltic countries: récentes divergences d’espérance de vie dans alcohol-related mortality trends in Sweden 1970–2015 with predictions les pays Baltes. Eur J Popul Rev Eur Démographie 27:403–31. to 2025. Scand J Public Health 47:446–51. Kraus L, Østhus S, Amundsen EJ, et al. (2015) Changes in mortality Sheron N, Gilmore I, Parsons C, et al. (2012) Projections of alcohol-related due to major alcohol-related diseases in four Nordic countries, France deaths in England and Wales—tragic toll or potential prize? Lancet and Germany between 1980 and 2009: a comparative age-period- 379:687–8. cohort analysis: changes in alcohol-related mortality. Addiction 110: Sheron N, Hawkey C, Gilmore I. (2011) Projections of alcohol deaths—a wake- 1443–52. up call. Lancet 377:1297–9. Noel JK. (2019) Associations between alcohol policies and adolescent alco- Stoeldraijer L, Bonneux L, van Duin C, et al. (2015) The future of smoking- hol use: a pooled analysis of GSHS and ESPAD data. Alcohol Alcohol attributable mortality: the case of England & Wales, Denmark and 54:639–46. the Netherlands: the future of smoking-attributable mortality. Addiction Downloaded from https://academic.oup.com/alcalc/advance-article/doi/10.1093/alcalc/agaa107/5934607 by guest on 05 March 2021 Pitkänen T, Lyyra A-L, Pulkkinen L. (2005) Age of onset of drinking and the 110:336–45. use of alcohol in adulthood: a follow-up study from age 8–42 for females Trias-Llimós S, Bijlsma MJ, Janssen F. (2017) The role of birth cohorts in long- and males. Addiction 100:652–61. term trends in liver cirrhosis mortality across eight European countries. Pruckner N, Hinterbuchinger B, Fellinger M, et al. (2019) Alcohol-related mor- Addiction 112:250–8. tality in the WHO European region: sex-specific trends and predictions. Trias-Llimós S, Kunst AE, Jasilionis D, et al. (2018) The contribution of alcohol Alcohol Alcohol 54:593–8. to the east-west life expectancy gap in Europe from 1990 onward. Int J R Core Team. (2018). R: A language and environment for statistical computing. Epidemiol 47:731–9. R Foundation for Statistical Computing, Vienna, Austria. https://www.R- Trias-Llimós S, Martikainen P, Mäkelä P, et al. (2018) Comparison of different project.org/. approaches for estimating age-specific alcohol-attributable mortality: the Ramstedt M. (2001) Per capita alcohol consumption and liver cirrhosis mortal- cases of France and Finland. PLoS One 13:e0194478. ity in 14 European countries. Addiction 96:19–33 Wiley Online Library. Villegas A, Kaishev VK, Millossovich P. (2015) StMoMo: an R package for Rehm J, Gmel GE, Gmel G, et al. (2017) The relationship between different stochastic mortality modelling. In 7th Australasian Actuarial Education dimensions of alcohol use and the burden of disease-an update: alcohol and Research Symposium. and disease. Addiction 112:968–1001. WHO. (2018) Adolescent alcohol-related behaviours: trends and inequalities Rey G. (2016) Les données des certificats de décès en France: processus in the WHO European Region, 2002–2014. de production et principaux types d’analyse. Rev Médecine Interne 37: WHO. (2019) Global status report on alcohol and health 2018. 685–93. WHO Global Health Observatory Data Repository. n.d. http://apps.who.i Rey G, Boniol M, Jougla E. (2010) Estimating the number of alcohol- nt/gho/data/node.main.A1022?lang=en (9 November 2018, date last attributable deaths: methodological issues and illustration with French accessed). data for 2006. Addiction 105:1018–29. Yang Y, Fu WJ, Land KC. (2004) A methodological comparison of age-period- Richard J-B, Palle C, Guignard R, et al. (2015) La consommation d’alcool en cohort models: the intrinsic estimator and conventional generalized linear France en 2014. Evolutions 32:6. models. Sociol Methodol 34:75–110 Wiley Online Library.