Time trends of the association of body mass index with mortality in 3.5 million young Swedish adults

Purpose: We investigated time trends of the obesity-mortality association, accounting for age, sex, and cause-specific deaths. Methods: We analysed pooled nationwide data in Sweden for 3,472,310 individuals aged 17 – 39 years at baseline in 1963 – 2016. Cox regression and flexible parametric survival models investigated BMI-mortality associations in sub-groups of sex and baseline calendar years (men: < 1975, 1975 – 1985, ≥ 1985 and women: < 1985, 1985 – 1994, ≥ 1995). Results: Comparing men with obesity vs. normal weight, all-cause and “ other-cause ” mortality associations decreased over periods; HR (95% CI) 1.92 (1.83 – 2.01) and 1.70 (1.58 – 1.82) for all-cause and 1.72 (1.58 – 1.87) and 1.40 (1.28 – 1.53) for “ other-cause ” mortality in < 1975 and ≥ 1985, but increased for CVD mortality; HR 2.71 (2.51 – 2.94) and 3.91 (3.37 – 4.53). Higher age at death before 1975 coincided with more obesity-related deaths at higher ages. Furthermore, the all-cause mortality association for different ages in men showed no clear differences between periods (p-interaction = 0.09), suggesting no calendar effect after accounting for attained age. Similar, but less pronounced, results were observed in women. Associations with cancer mortality showed no clear trends in men or in women. Conclusions: Accounting for differences in age and death causes between calendar periods when investigating BMI-mortality time trends may avoid misinterpreting the risks associated


Introduction
Obesity (body mass index, BMI ≥30 kg/m 2 ) is a major cause of death, [1] and was responsible for over five million deaths globally in 2019.[2] However, the association of BMI with mortality is complex.Observational, prospective studies have established the association with all-cause mortality to be J-or U-shaped, with the highest risks observed for BMI in the obesity range.[3][4][5][6][7][8][9][10][11] Obesity has also been suggested to be a weaker risk factor for all-cause mortality during later calendar years, which has been speculated to be due to better treatment of other cardiovascular disease (CVD) risk factors over time, particularly in individuals with obesity.[3,12] However, age, sex, and different causes of death are all differently associated with BMI [6,8] but were not always fully accounted for in previous time-trend analyses.For example, two studies from the US and Denmark reported a decreased risk of all-cause mortality in obesity over time but did not stratify by sex and had a large variation in baseline age.[3,12] The BMI-mortality association is strong in young adulthood and weakens with higher baseline age.[8,12,13] It is possible that less influence by potential confounders in young adulthood, such as by smoking due to a shorter time exposed, together with the absence of age-related comorbidities affecting weight in younger age, contributes to a stronger BMI-mortality association in young than in old adulthood.BMI is a stable long-term marker with small measurement error and intra-individual variation over decades [5,14], further validating the investigation of BMI in young adults on mortality.
Using data from a large, pooled cohort from Sweden, we investigated sex-specific calendar time trends in the association between BMI and allcause and cause-specific mortality.We focused on BMI measured in young adults  years of age at baseline), which allowed for further exploration of possible causes for the observed time trends, while accounting for the potential influence of age and causes of death compositions on the associations.

Study Design and Setting
We used pooled nationwide data from the Obesity and Disease Development Sweden (ODDS) study, which consists of Swedish cohorts and national registers with individual-level data on height and weight [15].Men in the Swedish Military Conscription Register [16] and women in the Medical Birth Register (MBR) [17], made up the majority of the study population (84 % of the overall sample).Height in the MBR was self-reported while weight was measured at the first antenatal care visit (around gestational weeks 8-10) for approximately 90 % of pregnant women from 1983 onwards, before any substantial pregnancy-related weight gain.[18] We observed a larger-than-expected increase in body weight between 1989 and 1992 in the MBR, likely due to a change in recording in 1992.[19,20] Therefore, women's weight in this register between 1982 and 1989 was adjusted by estimating annual weight gain in the full MBR.The conscript's data contains objective measures of height and weight of all Swedish conscripts (aged 17-22 years) who enlisted for mandatory military service between 1969 and 2010, or for voluntary military service between 2011 and 2014.[16] Other included cohorts in the study are listed in Supplementary Table S1.The Swedish Ethical Review Authority approved the ODDS study (no: 2020-03846).

End points
We used the unique identification number of residents in Sweden to link each individual in ODDS to the Cause of Death Register, covering virtually all deaths in Sweden.[22] From this register, we obtained information about cause-specific deaths until December 31, 2020, including the primary underlying cause.We classified causes of death into the three primary categories; CVD, cancer, and all other causes combined, using the International Classification of Diseases (ICD) codes (ICD-8 to ICD-10) as outlined in Supplementary Table S2.Post-hoc analyses were also performed for deaths from specific CVDs (ischemic heart diseases vs. all other CVDs combined), obesity-related [23] vs. other cancers, and external mortality causes vs. all other mortality causes combined (Supplementary Table S2).

Study population
The pooled ODDS cohort consists of 4,295,859 men and women with 7,733,901 assessments of weight at the age of ≥17 years.We excluded assessments with recalled historical weight, missing height, extreme anthropometric values, and a study entry later than the date of death, emigration, or end of follow-up (Supplementary Fig. S1).In the remaining 4,163,959 individuals, we selected the first available assessment at baseline, giving priority to the one with smoking information if that was available.To optimise the final participant selection, we applied Cox regression analysis, as described in Supplementary Figs.S2 and S3, to investigate the potential effect of baseline age and confounding by smoking on the BMI-mortality association.[8] We found that the BMI-mortality association changed from J-to U-shaped across age (Supplementary Fig. S2), and the older (≥40 years), but not the younger (<40 years), population, was affected by confounding by smoking (Supplementary Fig. S3).In younger individuals, the adjustment for smoking and the analysis of never-smokers resulted in similar BMI-mortality associations compared to the smoking-unadjusted analysis, which justified the final study population of younger individuals and the focus on smoking-unadjusted analyses.The final selected study population consisted of 3,472,310 individuals 17-39 years of age with > 5 follow-up years (Supplementary Fig. S1), to reduce the potential effect of reverse causation.[8,24,25] The distribution of individuals in the final study population across cohorts is available in Supplementary Table S1.

Statistical analyses
Individuals were followed up from five years after baseline  to the date of death, emigration, or December 2020, whichever occurred first.We used Cox proportional hazards regression with attained age as the time scale and estimated hazard ratios (HR) and 95 % confidence intervals (CI) for the BMI-mortality association.We used Cox regression to investigate the shape (using restricted cubic splines with five knots at Harrell's recommended percentiles [26,27]) and strength of BMI-mortality association using the WHO categories (<18.5, 18.5-24.9,25.0-29.9,and ≥30 kg/m 2 ).All predicted BMI splines were centred around a reference of 22.5 kg/m 2 .Models were adjusted for the baseline calendar year and age, the highest attained education level during follow-up, marital status, birth country for individuals and parents, modes of weight and height measurement, and for having the weight measurement originating from the MBR (i.e. in early pregnancy).Departures from the proportional hazards assumption for covariates were assessed graphically by plotting the log-log survival vs. analysis time (attained age), and evaluated by looking at the scaled Schoenfeld residuals with no major violations observed.Associations between BMI and mortality were examined using Cox regression (i.e., restricted cubic splines and in categories) and flexible parametric splines (to account for the time-varrying effect of BMI in categories) over the attained age scale.Due to strong sex differences in the association between BMI splines and WHO categories with mortality [6,8], all analyses were performed separately for men and women.
We performed analyses in sub-groups of calendar year at the baseline examination (<1975, 1975-1985, ≥1985 in men and <1985, 1985-1994, ≥1995 in women) for the BMI association with mortality.The 10-year difference in cut-points for calendar periods in men vs. women was used because of younger ages and earlier years of baseline examinations in men than in women.A test for interaction to compare the estimated HR of obesity vs. normal weight across calendar periods was calculated and obtained from the standard normal distribution tables.[28] Age-adjusted mortality rates were estimated using marginal post-estimation predictions from Poisson regression models adjusted for age categories (17-19, 20-29, and 30-39).[29] The HRs for the BMI-mortality association from the Cox regression models are summary estimates based on the event distribution over the full follow-up time.To further investigate the observed time trends of the HR for obesity vs. normal weight with mortality, we analysed BMI categories and their interaction with time (attained age) using flexible parametric survival regression models, adjusted for the above-mentioned covariates.All statistical tests were evaluated using a significance level of 0.05, and analyses were performed using Stata/MP version 18.0 (StataCorp LLC., College Station, Texas).[26].

Characteristics across calendar periods
Age at death from all causes was overall low with a median (IQR) of 56 (46− 65) years in men and women combined.However, age at death was much lower in later calendar years owing to the shorter follow-up time; for example, 62 (53− 69) years in the < 1975 calendar period vs. 38 (31− 46) 1).

Time trends of BMI and mortality
Across all calendar periods combined, associations between BMI and mortality were J-or U-shaped for all-cause and cause-specific deaths, with the strongest associations for BMI in the obesity range observed for CVD mortality (Figure 1).
Time trend analyses of BMI in categories revealed that compared to normal weight, the increased risk of all-cause mortality with obesity decreased over time in men, HR (95 % CI); 1.92 (1.83-2.01)before 1975, and 1.70 (1.58-1.82) in 1985 and later, (p-interaction=0.001)(Figure 2).This trend was driven by the results for mortality due to other-causes (p-interaction=0.001).The opposite trend was observed for CVD mortality, where mortality risk in men with obesity increased over the calendar years (p-interaction<0.001).The association of obesity vs. normal weight with the increased risk of cancer mortality showed no clear trends across the calendar years (p-interaction=0.34).Similar, but less clear, results were observed in women (Figure 3).

BMI and mortality across attained age and death causes composition
We further investigated if the trends in the association of obesity and mortality risk in men could be influenced by differences in age at death and composition of death causes between the calendar periods.For allcause and other-cause mortality, higher age at death in the first calendar period (<1975) coincided with higher HRs for obesity at higher ages (i.e., more obesity-related deaths at higher ages) analysed for all periods combined (Figure 4).
The opposite trend, i.e., lower HRs for obesity at higher ages (which were more prevalent in the first vs. last calendar period) was observed for CVD mortality.Furthermore, death causes more strongly associated with higher BMI were CVD amongst all mortality causes (Figure 1), ischemic heart diseases amongst CVD mortality causes, obesity-related cancers amongst cancer mortality causes, and "other non-external causes" (such as infectious and parasitic diseases, non-malignant neoplasms, and endocrine, nutritional, and metabolic diseases) amongst "other" mortality causes (Supplementary Fig. S4).Comparing the percentage distribution of specific death causes with HRs changing across the attained age scale, these mortality causes were more prevalent in the earliest calendar period and more common at an older age, at which the highest risks for obesity vs. normal weight were observed (Figure 4).Similar, but less clear results were observed for women (Supplementary Fig. S5).Interval.Hazard ratios were derived from Cox regression models with BMI modelled in WHO categories on attained age time scale.Individuals were followed from five years after baseline until death, emigration, or end of follow-up, whichever occurred first.Hazard ratios adjusted for the age and calendar year at baseline examination, education level, marital status, modes of height and weight measurement, and birth country.P-interaction to compare the estimated HR were calculated and obtained from the standard normal distribution [28].For all-cause, cancer, CVD, and other-causes mortality, p-interaction= 0.01, 0.34, < 0.001, and 0.001, respectively, for the HR of obesity vs. normal weight in < 1975 vs. ≥ 1985 calendar periods.Hazard ratios were derived from Cox regression models with BMI modelled in WHO categories on attained age time scale.Individuals were followed from five years after baseline until death, emigration, or end of follow-up, whichever occurred first.Hazard ratios adjusted for the age and calendar year at baseline examination, birth register (Yes' for women in the Medical Birth Register and 'No' if otherwise), education level, marital status, modes of height and weight measurement, and birth country.P-interaction to compare the estimated HR were calculated and obtained from the standard normal distribution [28].For all-cause, cancer, CVD, and other-causes mortality, p-interaction= 0.01, 0.18, 0.25, and 0.10, respectively, for the HR of obesity vs. normal weight in < 1985 vs. ≥ 1995 calendar periods.
A time-trend of the BMI association with all-cause mortality (Figure 2) independent of attained age should reveal an overall stronger association between obesity and mortality across attained age in the first vs. the last calendar period.However, an additional post hoc analysis in men revealed no consistent differences between calendar periods in the HRs of all-cause mortality for obesity vs. normal weight across attained age (p for interaction=0.09, Figure 5), suggesting that the observed time-trends for obesity on mortality were essentially due to variations in compositions of death causes and attained ages.Similar analyses of other-causes mortality in men, and analyses in women, lacked statistical power and were not meaningful for interpretation.

Discussion
In this large, pooled study, we investigated BMI-mortality time trends among young adults in Sweden for baseline periods ranging from < 1975 to ≥ 1985 in men and < 1985 to ≥ 1995 in women.We observed that the obesity-associated risk for all-cause and other-causes mortality decreased over time but increased for CVD mortality, most clearly in men, but also in women.Further analyses in men showed that the investigated calendar periods differed by age at death and the distribution of main death causes, which were differently associated with BMI.Also, obesity-related risk of all-cause mortality for different ages showed no consistent differences between calendar periods.Altogether, these findings indicate that the observed time trends of obesity with mortality are likely explained by methodological artefacts influenced by different ages at death and compositions of death causes between the analysed periods.
Relative to normal weight, the mortality risk associated with obesity is reported to be lower in later calendar years [3,12] while the optimal BMI (i.e., the BMI associated with the lowest mortality risk) [3,11,13,30] has increased across decades ranging from the 1970s to the early 2000s.One interpretation of these findings has been that improved treatment of CVD risk factors over time may have benefitted individuals with obesity relatively more, hence, reducing the mortality risks associated with obesity over time.[3,12] This notion does not hold in our study, as obesity was a stronger risk factor for CVD mortality over time.The differences between our findings and previous investigations concerning CVD mortality are likely explained by differences in important characteristics such as age, sex, and variations in the distribution of baseline calendar periods, follow-up time, age at death, and death causes.[3,8,25] For instance, whilst individuals in the last calendar period in our study were younger at the age of death compared to those in the first calendar period, individuals in the Danish study instead were slightly older at baseline overall and in the last calendar period.[3] This, together with implicitly different proportions of CVD mortality causes Obesity-related Cancers refer to incident cancers with sufficient evidence of an association with obesity [23].External causes refer to death from external morbidity and mortality causes such as accidents and injuries.Hazard ratios were derived from flexible parametric survival regression models with BMI modelled in categories as a function of time (time-changing covariate) on attained age as time scale.Vertical dashed lines show the median age at death for the earliest (before 1975) and later (1985 onwards) calendar years for all-cause and cause-specific mortality.The secondary y-axis (right-hand side of the plots) is the percentage of specific death causes from the respective death cause at attained age categories (i.e., <30, then every 5 years, and ≥85).Individuals were followed from five years after baseline until death, emigration, or end of follow-up, whichever occurred first.Hazard ratios adjusted for the age and calendar year at baseline examination, education level, marital status, modes of height and weight measurement, and birth country.
between our young population and other epidemiological studies, is likely to have contributed to the contrasting time trends.Yet, two studies in Denmark and the US observed similar time trend patterns for all-cause mortality.[3,12] These findings in our study were not driven by CVD mortality, which was less common in our young population, but by the trends of other-causes mortality.
Studies focusing on the change of optimal BMI did identify the effect of follow-up time as important on BMI-mortality time trends.Two studies observed that the optimal BMI increased in later cohorts, but these cohorts also had a shorter length of follow-up, and the time-trend effect was almost diminished when fixing follow-up time across the cohorts.[11,13,25] The authors speculated that prevalent disease at baseline and a more accurate, i.e., not changed over time, measure of BMI, might be the reason for a stronger association of BMI with mortality for shorter follow-up time.[25] The observed effects may indeed be attributed to the time since baseline as investigated in these studies, but they may also depend on the age at death and the specific death causes, which inherently are correlated with follow-up time.It is established that the BMI association with all-cause mortality is affected by sex, age at death, and death cause compositions.[8] When restricting the baseline age to < 40 years, analysing men and women separately, we found a decreased risk of all-cause and other-causes mortality in obesity at higher ages, but a decreased risk for CVD mortality.Further investigations in men showed that the observed time trends were largely explained by large differences in age at death across calendar periods, and the fact that CVD mortality, which was most strongly associated with obesity, was a less common death cause in later calendar periods due to lower age of death.All together, these results show that obesity has remained a strong mortality risk factor over time, including for cardiovascular mortality.
Potential limitations of this study are the lack of adjustment for smoking habits and not accounting for obesity-related comorbidities.[31] However, smoking-adjusted analyses and analyses of never-smokers in a subpopulation with this information available showed only minimal differences in the association with BMI as compared to the smoking-unadjusted analysis.Furthermore, obesity-related comorbidities are rare in young adults.[32] Despite adjustments made, there could still be residual confounding by factors not fully accounted for or measured, such as parity in women, dietary patterns, physical activity, and other lifestyle factors that can influence both obesity and mortality.BMI does not distinguish between lean body mass and fat mass, whose composition changes with age, even within the baseline age of our study population.[33] However, BMI has shown stronger correlations with both body fatness and cardiometabolic risk markers, such as serum insulin levels, in younger than in older adults [33] further supporting the relevance of our study.Nevertheless, the low prevalence of obesity in our young, historical study population (i.e., 2 % in men and 7 % in women) limits statistical power to detect time trends over calendar periods.Furthermore, the young average age at death of 56 years in our study is not representative of life expectancy in Sweden.[34] When younger individuals die, the causes are less likely obesity-related (or less CVD-related, due to its strong association with obesity) than in older individuals.However, due to the large population size and long follow-up, sufficient numbers of deaths in older age were included to allow modelling risks also across higher attained age in the population.Although the autopsy rates have declined in Sweden, potentially leading to higher misclassification of deaths over time, the agreement between the causes of death from death certificates and those on patient case summaries is high, especially up to middle age (e.g.91 % agreement at 45-64 years).[22] A major strength of this study is that it constitutes a large young adult population (<40 years) with long follow-up (median of 28 years), enabling to study men and women and cause-specific mortality separately.Despite this, the calendar periods investigated differed in terms of follow-up time and age of death, and, inherently, by major death causes.While we have provided plausible explanations for our findings of BMImortality time trends, the ideal would have been a direct comparison of three populations differing only in terms of the investigated calendar periods, which has yet to be done in future, and even larger, studies.

Conclusions
In this young Swedish population, a higher all-cause and othercauses mortality risk for obesity vs. normal weight decreased slightly across baseline calendar periods but increased for CVD mortality.Our findings indicate that the observed time trends of obesity with mortality could be explained by methodological artefacts arising from different age distributions during follow-up, leading to different compositions of death causes between the analysed periods.This underscores the importance of accounting for differences in age at death and cause of death when investigating BMI-mortality time trends, which will avoid misinterpretation of the obesity association with mortality risk over time.

Fig. 1 .
Fig. 1.Hazard Ratio of All-cause and Cause-specific Mortality According to Body Mass Index in Men and Women.Hazard ratios were derived from Cox regression models with BMI (range, 15-60 kg/m 2 ) modelled using restricted cubic splines on attained age as time scale.Solid lines are hazard ratio point estimates, and shaded areas indicate 95 % confidence intervals.Follow-up started five years after baseline, and estimates were adjusted for age and calendar year at baseline, sex, birth register (yes/no), education level, marital status, mode of height and weight assessment, and birth country.CVD, Cardiovascular diseases.HR, Hazard Ratio.PYRS, Person years.CI.Confidence Interval.a AMR, Age-adjusted Mortality Rate per 1,000 person-years.

Fig. 2 .
Fig. 2. Hazard Ratio of All-cause and Cause-specific Mortality in Men According to Body Mass Index Categories, in Sub-groups of Baseline Calendar Year.Abbreviations: AMR, Age-adjusted Mortality Rate per 1,000 person-years of follow-up.BMI, Body Mass Index.HR, Hazard Ratio.PYRS, Person Years.CI, Confidence Interval.Hazard ratios were derived from Cox regression models with BMI modelled in WHO categories on attained age time scale.Individuals were followed from five years after baseline until death, emigration, or end of follow-up, whichever occurred first.Hazard ratios adjusted for the age and calendar year at baseline examination, education level, marital status, modes of height and weight measurement, and birth country.P-interaction to compare the estimated HR were calculated and obtained from the standard normal distribution[28].For all-cause, cancer, CVD, and other-causes mortality, p-interaction= 0.01, 0.34, < 0.001, and 0.001, respectively, for the HR of obesity vs. normal weight in < 1975 vs. ≥ 1985 calendar periods.

Fig. 3 .
Fig. 3. Hazard Ratio of All-cause and Cause-specific Mortality in Women According to Body Mass Index Categories in Sub-groups of Calendar Year at Baseline Examination.Abbreviations: AMR, Age-adjusted Mortality Rate per 1,000 person-years of follow-up.BMI, Body Mass Index.HR, Hazard Ratio.PYRS, Person Years.CI, Confidence Interval.Hazard ratios were derived from Cox regression models with BMI modelled in WHO categories on attained age time scale.Individuals were followed from five years after baseline until death, emigration, or end of follow-up, whichever occurred first.Hazard ratios adjusted for the age and calendar year at baseline examination, birth register (Yes' for women in the Medical Birth Register and 'No' if otherwise), education level, marital status, modes of height and weight measurement, and birth country.P-interaction to compare the estimated HR were calculated and obtained from the standard normal distribution[28].For all-cause, cancer, CVD, and other-causes mortality, p-interaction= 0.01, 0.18, 0.25, and 0.10, respectively, for the HR of obesity vs. normal weight in < 1985 vs. ≥ 1995 calendar periods.

Fig. 4 .
Fig. 4. Hazard Ratio (left y-axis), and proportion of specific death causes (right y-axis) in Men, for All-cause, Cardiovascular Disease, Cancer, and Other Mortality Causes in Obesity vs. Normal Weight Across Attained Age.Abbreviations: HR, Hazard Ratio.CI, Confidence Interval, CVD, Cardiovascular Disease.Obesity-related Cancers refer to incident cancers with sufficient evidence of an association with obesity[23].External causes refer to death from external morbidity and mortality causes such as accidents and injuries.Hazard ratios were derived from flexible parametric survival regression models with BMI modelled in categories as a function of time (time-changing covariate) on attained age as time scale.Vertical dashed lines show the median age at death for the earliest (before 1975) and later (1985 onwards) calendar years for all-cause and cause-specific mortality.The secondary y-axis (right-hand side of the plots) is the percentage of specific death causes from the respective death cause at attained age categories (i.e., <30, then every 5 years, and ≥85).Individuals were followed from five years after baseline until death, emigration, or end of follow-up, whichever occurred first.Hazard ratios adjusted for the age and calendar year at baseline examination, education level, marital status, modes of height and weight measurement, and birth country.

Table 1 Characteristics at Baseline and End of Follow-up in Sub-groups of Calendar Year at Baseline Examination and in Total, in Men and Women a .
years in the ≥ 1985 period in men, and 61 (54− 69) years in < 1985 vs. 44 (39− 50) years in ≥ 1995 in women.The proportion of underlying causes of deaths due to CVD decreased from 28 % to 13 % in men in the < 1975 vs. ≥ 1985 calendar period and from 17 % to 9 % in in the < 1985 vs. ≥ 1995 calendar period, whereas the proportion of deaths due to other death causes than CVD and cancer increased from 40 % to 68 % in men and from 31 % to 38 % in women.The median baseline age, BMI, and year of death differed very little between calendar periods (Table Abbreviations: BMI, Body Mass Index.CVD, Cardiovascular diseases.a Numbers are median (interquartile range) unless otherwise indicated.b Calculated as the number of individuals who died from cancer, CVD, and other death causes, divided by number of death from all causes multiplied by 100, separately for calendar years at baseline in men and women.women