Nine BAs were included in the present study based on data availability in SATSA. Leukocyte telomere length was measured by quantitative polymerase chain reaction (qPCR) and presented as a relative telomere length (Berglund et al., 2016). Genome-wide methylation levels in whole blood were assessed by Infinium HumanMethylation 450K BeadChip, and four types of DNAmAge (Horvath, Hannum, PhenoAge, and GrimAge) were then calculated using established algorithms based on the methylation levels of 353, 71, 513, and 1030 CpG sites, respectively (Hannum et al., 2013; Horvath, 2013; Levine et al., 2018; Lu et al., 2019). Physiological age was derived from a list of age-correlated biomarkers (Pearson correlation coefficient with CA > 0.10) measured by blood tests and physical examinations (see Materials and methods and Supplementary file 1B). Biomarkers were transformed into a calibrated BA value using principal component analysis and the Klemera and Doubal (2006) methods. Cognitive function measured performance in verbal (crystallized) ability, spatial (fluid) ability, memory, and perceptual speed using a cognitive testing battery. The four specific domains were combined into an overall score by principal component analysis (Reynolds et al., 2005). FAI incorporated four functional aspects from self-reported information and physical examinations, that is sensory (vision and hearing), pulmonary, strength (grip strength), and movement/balance (gait speed). The four indicators were standardized separately and summed to create a composite score (Finkel et al., 2019). Lastly, a set of 42 self-reported deficits, covering a range of health domains, were taken into account in the development of the FI (Jiang et al., 2017). (Table 2 and Supplementary file 1B–C).

For each BA, we calculated a corresponding BA residual by regressing out CA and sex effects as well as individual- and twin-pair related components from the BA using a linear mixed model (see Materials and methods and Figure 1). The BA residuals in the present study are commonly referred to as age acceleration in the DNAmAge-related literature.

Figure 1 with 1 supplement see all

Download asset Open asset

Of the 845 study participants, 59.5% were women, 42.2% attained above primary education, and 25.0% were current or former smokers; the average body mass index (BMI) was 25.7 kg/m², and mean chronological age was 63.6 years at baseline. On average, the levels of BAs were 0.73 for telomere length (T/S ratio), 60.4 years for Horvath DNAmAge, 65.2 years for Hannum DNAmAge, 63.8 years for DNAmPhenoAge, 69.4 years for DNAmGrimAge, 64.7 years for physiological age, 51.5 for cognitive function, 48.3 for FAI, and 0.10 for FI. The baseline characteristics in individuals with complete measurements were similar to those in all individuals. (Table 3 and Supplementary file 1D).

On average, BAs were assessed between 2.5 times (telomere length) to 4.2 times (FI) in each participant with information on the respective BAs. (Table 3) Longitudinal changes in BAs were modeled as functions of CA (as a natural spline with three degrees of freedom) and sex, with random effects introduced at the individual and twin-pair levels (mixed models, see Materials and methods). Both individual-level BAs and population BA means over CA in men and women are presented in Figure 1. We found that with increasing age, average telomere length and cognitive function declined, while all types of DNAmAges, physiological age, FAI, and FI increased. The profiles for the three functional BAs (cognitive function, FAI, and FI) show clear curvature, indicating an accelerated rate of change around the age of 70, whereas the other BAs exhibit minimal or no curvature, corresponding to a constant change (linear growth) over the age span. In addition, we observed small to moderate sex differences in the mean levels of BAs. Women exhibited longer telomere length (p=0.001) and lower DNAmAge (p=0.013 for Horvath, p=0.001 for Hannum, and p<0.001 for GrimAge) compared to men, but also worse functioning (p<0.001 for FAI and p=0.001 for FI). (Figure 1).

To examine the possible effect modification, we next introduced an interaction term between CA and sex in the mixed models. We found no evidence that the shape of the BA curves differed between men and women, except for in physiological age (p<0.001 for sex interaction). However, the difference was mainly observed at the end of the CA spectrum (i.e., before the age of 50 and after the age of 85). (Figure 1—figure supplement 1).

BAs were broadly categorized into four groups according to the main biological structural levels where the BA measurements were implemented, that is telomere length, DNAmAges, physiological age, and functional BAs (Panel A in Figure 2). We estimated correlation coefficients between BAs and BA residuals and adjusted for repeated measurements using the rmcorr function in R (Figure 2 and Supplementary file 1E) (Bakdash and Marusich, 2017). Using all complete measurements, we estimated repeated measures correlation (rmcorr, 22) between age and all nine BAs, which accounts for non-independence among observations and captures the common intra-individual association between age and each of the BAs in turn. Telomere length showed low correlations with both CA and BAs (absolute correlation coefficients ≤ 0.16); the other BAs were correlated with CA and each other to varying degrees, with absolute correlation coefficients ranging from 0.24 to 0.87. As indicated by the correlation matrices of BAs and BA residuals (Panel B and C in Figure 2), after regressing out CA from BAs, most of the original correlations were attenuated and moderate correlations remained only between Horvath and Hannum DNAmAges (r = 0.35), cognitive function and FAI (r = −0.32), and FAI and FI (r = 0.31). Correlation patterns of BAs and BA residuals in men and women were similar to those in the whole population. (Data not shown).

Figure 2 with 1 supplement see all

Download asset Open asset

Next, we transformed correlation coefficients into Euclidean distances and then performed hierarchical cluster analysis. The same types of BAs, that is methylation BAs and functional BAs, tended to be more closely related. GrimAge and PhenoAge, however, were somewhat separated from the other two DNAmAges, especially using BA residuals. (Figure 2—figure supplement 1).

To study the association between BAs and all-cause mortality, we used Cox regression models with attained age being time scale and accounted for left truncation and right censoring. To make the hazard ratios (HR) comparable, we rescaled all BAs and BA residuals by their standard deviation (SD), so that HRs express changes in mortality risk due to a one-SD increment of the respective BA or BA residual. Fully adjusted models controlled for sex, educational attainment, smoking status, and BMI. All models were stratified by participants’ birth year (categorized in 10 year intervals into five groups) and implicitly adjusted for CA through the choice of attained age as the underlying time scale. When repeated measurements were available, only the first BA assessment was included in the survival analysis and referred to as the baseline BA.

First, we looked at the effect of baseline BAs separately (i.e., one-BA models). During a median follow-up time of more than 15 years, we observed statistically significant BA-mortality associations for DNAmAges (GrimAge, PhenoAge, and Horvath) and functional BAs (FI, FAI, and cognitive function), where a one-SD increase of BAcorresponded to a change in mortality risk by 39, 26, 17, 32, 27, and 15%, respectively. Increased mortality risks were also found with changes in Hannum DNAmAges and physiological age, although the effects were generally weaker and came with a slightly higher level of imprecision, with HRs (95% CI) being 1.17 (0.98, 1.40) and 1.13 (0.97, 1.31). We found no evidence for telomere length as associated with mortality risk. (Table 4) The corresponding HRs for the BA residuals were comparable to BAs (Supplementary file 1F).

Second, we explored the baseline BAs in relation to mortality risk by taking all BAs into consideration simultaneously in 288 individuals with complete measurements (i.e., nine-BA models). To avoid biased estimation due to collinearity (as all nine BAs were correlated to varying extents), BA residuals instead of BAs were used. During a median follow-up time of 16.5 years, we observed that three BA residuals, Horvath DNAmAge, DNAmGrimAge, and FI, out of nine were associated with the risk of all-cause mortality independently of all other BAs. A one-SD increment of Horvath DNAmAge, DNAmGrimAge and FI residuals was associated with a 31%, 43% and, 58% higher risk of death, respectively, independent of the level of CA, all other BA residuals and common risk factors. The HRs for the other DNAmAges (Hannum and PhenoAge), physiological age, and the other functional ages (cognitive function and FAI) were attenuated towards one after controlling for all BAs in the same model (Table 5).

We replicated the first part of the survival analyses (i.e., one-BA models) in subgroups classified by sex, age group, and smoking status, as well as in the sub-cohort of individuals with complete measurements. In the subgroup analyses, we observed evidence for potential effect modification. The associations of BAs with mortality risk were generally stronger in women (except for Horvath DNAmAge, and physiological age), more pronounced in the younger individuals (except for Horvath DNAmAge, physiological age and cognitive function), and a bit stronger in current or ex-smokers (for Horvath DNAmAge and DNAmGrimAge) compared to those in men, older individuals, and non-smokers, respectively. (Figure 3 and Supplementary file 1G–I) Furthermore, as with the results observed in the whole population, we found the same general pattern of BA associations with mortality in the sub-cohort with complete measurements. (Supplementary file 1J) We also adjusted for the presence of the previous diseases, including heart failure, stroke, diabetes, and cancer in the survival model and the observed results were largely unchanged. (Supplementary file 1K).

Figure 3

Download asset Open asset

As biological aging does not proceed at a constant pace across time, monitoring longitudinal trajectories of BAs using repeated assessments to depict the characteristics of the aging course is of great importance. We observed an accelerated change in the aging pace in functional BAs, which are in parallel with previous studies focused on trajectories of cognition, physical function, and frailty (Harris et al., 2016; Stow et al., 2018; Zaninotto et al., 2018). This result indicates that an aged population may expect deteriorations of their functional performance to speed up as their age approach 70 years. In particular, we observed that the aging acceleration of cognition (cognitive function) and physical function (FAI) seemed to take place before age 70, while frailty (FI) went up at an increased rate after age 75. This evidence suggests that accelerated cognitive and physical dysfunctions are likely to arise prior to frailty’s, and corresponding interventions and screenings could aim at people of different ages accordingly.

In contrast to functional BAs, DNAmAges and physiological age accrued at a relatively constant rate. Due to their mode of construction, DNAmAges and physiological age were transformed into an age-calibrated scale in the unit of year. Thus, these BAs allow individuals to perceive an intuitive impression on their biological aging status by comparing BA with CA and have the potential to serve as a straightforward-to-use metric in the geriatric practice. However, in order to be clinically meaningful, they need to be validated on the individual level data as well. In addition, a drawback of this mode of construction is that it largely explains the linear shapes of the BA trajectories, making longitudinal BA characteristics less informative as the changes are constant. Another feature of physiological age is the strong correlation with CA; hence, it allows limited biology perception in prediction models on top of CA. This is again due to its mode of construction using a composite score of several biomarkers where each one, by definition, strongly correlates with CA. To improve construction models for BA, composite scores should balance the reliance on biological aspects and weight on CA.

Sex differences in aging are observed at multiple biological levels (Oksuzyan et al., 2008; Ostan et al., 2016). Most notably, women tend to present lower mortality but higher morbidity rates, especially at advanced ages, compared to men of the same age. We found that, compared to men, women presented lower molecular ages (telomere length and methylation age estimators) but higher functional ages (FAI and FI) across the age spectrum, indicating different medical needs are likely to be encountered for men and women. Meanwhile, we observed no evidence for shape differences in the BA trajectories (except for physiological age) by sex. This suggests that aging proceeded at a similar rate in men and women since midlife onwards. For physiological age, the shape difference was mainly observed at the end of the age spectrum and could be influenced by methodological issues as the construction of physiological age was done differently in men and women. Putting our results in the perspective of previous studies, we confirm earlier findings in women, such as better telomere status and worse frailty (Mu et al., 2014; Bartley et al., 2016). One possible explanation for the sex difference in FI is that men and women tend to present different self-perception on aging-related performances due to social and cultural considerations when self-reported data are used (Ostan et al., 2016). However, cognitive function and FAI incorporate objective measurements tested by trained nurses and both of them exhibited sex differences in the same direction. Hence, sex differences in the functional BA spectrum are not, or at least not exclusively, driven by social and cultural factors. Therefore, we suggest that sex differences should always be accounted for in studies of aging and, if possible, presenting stratified analyses should be encouraged.

Aging is a complex process proceeding through interconnected biological mechanisms. BAs, in essence, are underlain by different sets of these mechanisms. Correlations of BAs could provide an idea of how one BA changes in accordance with another. We observed moderate to strong correlations between CA and all BAs, with the exception of telomere length, which was only weakly correlated with CA and the other BAs. These results are in line with previous research observing similar correlation pattern for 11 BAs, as well as low correlations between telomere length and other types of BA (Belsky et al., 2018). As some degree of correlation between BAs is expected simply due to their relation with CA, we also investigated correlations between BA residuals, which by construction minimize correlation with CA. We found clear attenuation of the correlations between different types of BA residuals, that is methylation BAs and functional BAs as well as physiological age and telomere length. In other words, among individuals with the same CA, molecular BAs and functional BAs are only weakly associated. We interpret this as an indication that these categories largely reflect different aspects of the aging process.

After regressing out the effect of CA, we still observed weak to moderate correlations between DNAmAges. These DNAmAges were initially trained to capture methylome differences associated with various aging-related phenotypes, namely CA for Horvath and Hannum, Phenotypic Age for PhenoAge, and mortality risk for GrimAge (Hannum et al., 2013; Horvath, 2013; Levine et al., 2018; Lu et al., 2019). Employing correlated surrogates and measuring the same aging hallmark of epigenetic alteration in the development of these DNAmAges could partly explain the correlations observed among these residuals. Hierarchical clustering suggested GrimAge and PhenoAge were somewhat separated from the other two DNAmAges, which is expected as GrimAge and PhenoAge incorporate additional information on top of CA in the development of their BA algorithms. In addition, Horvath and GrimAge were the least correlated DNAmAge residuals, indicating that these two BAs captured uncorrelated methylome information to a larger degree than other DNAmAges did and might have the potential to inform aging-related outcomes independently (as illustrated in the results of nine-BA survival analyses).

Functional BA indicators included cognitive function, FAI, and FI, which quantify cognitive, physical, and frailty-related functioning. Intriguingly, the present analysis found comparable levels of correlation for these BAs as well as for their residuals, suggesting that most of these correlations were not simply driven by CA. Previous studies have also reported correlations between physical function, cognitive function, and health-related quality of life (Kim, 2016). As functional BAs measure complex phenotypes resulting from intricate biological mechanisms, further investigations are needed to disentangle the mechanisms underlying both the age-dependent and, in particular, the age-independent correlations.

A majority of the included BAs are well-established aging indicators and were found to be associated with mortality in previous studies (Hannum et al., 2013; Levine, 2013; Shamliyan et al., 2013; Wang et al., 2018a). Our results were comparable to the current evidence. We first examined the baseline value of individual BAs in relation to the risk of death and found associations for all BAs except for telomere length, with DNAmGrimAge and FI demonstrating the strongest relationship. We next examined BA-mortality associations with all BAs included in the same model and observed significant evidence for Horvath DNAmAge, DNAmGrimAge, and FI, suggesting they were complementary in the prediction of mortality. In other words, Horvath DNAmAge, GrimAge and FI were likely to capture aging-related information that was both informative in the prediction of mortality risk and independent of other BAs. DNAmAges (Hannum and PhenoAge) and functional BAs (cognitive function and FAI) were likely to reflect some mortality-associated aging aspects that were shared and/or correlated with other BAs, as the corresponding HRs were attenuated to almost one when all BAs were included in the model.

Few studies have focused on the comparison of the predictive aspect of BAs. Kim et al. (2017) found FI to outperform methylation age estimators in survival models by comparing only two types of BAs, frailty, and DNAmAge (Horvath). Zhang et al. (2017) developed a methylation-based mortality score and observed a stronger mortality association than the FI-mortality relationship. Furthermore, Murabito and colleagues compared clinical age, inflammatory age, and DNAmAge (Horvath and Hannum) and concluded that they were complementary in predicting risk for mortality (Murabito et al., 2018). Our study thus supports these findings, both in the way that mortality-oriented DNAmAges and FI were strongly associated with mortality, and in that certain types of BAs could reflect mortality risk independently. In particular, we add information by including other BAs, especially physiological age, a newly trained DNAmAge (GrimAge) and functional BAs.

Physiological age is a composite biomarker of aging, developed from 10 blood and clinical markers. Levine et al. applied a similar method to develop a BA indicator and found it associated with mortality beyond CA in a large US population (Levine, 2013; Klemera and Doubal, 2006). We also observed that physiological age predicted mortality risk independently of age, albeit weaker and at a moderate significance level. Despite its weak mortality association, physiological age presents some benefits with respect to the physiological interpretation and the search for biological mechanism in future explorations, as the component biomarkers contributed to the development of physiological age were explicitly specified.

DNAmGrimAge is derived via a two-step development method, in which methylation data were first used to predict a set of biomarkers (plasma proteins and smoking pack-year), and the methylation-based biomarkers were then used in the prediction of mortality risk. As a result, the DNAmGrimAge is explicitly trained to be a mortality predictor; it also includes the largest number of genomic sites compared with the other three DNAmAges, and allows the methylation level of a single CpG site to contribute to mortality risk via different intermediate biomarkers, as the methylation-based biomarkers used somewhat overlapping CpG sites in the two-step training phase. Lu et al. (2019) reported the advantage of predicting mortality over other existing DNAmAges in a US population. These results point to the potential of utilizing methylation information to reflect complex aging phenotypes.

The FI is developed using clinical information and is likely to capture an organismal perspective on aging that is closer to the clinical end-point of death compared to molecular-based BAs. In addition, the FI incorporates information in multiple health domains, in contrast to the other two functional BAs, cognitive function, and FAI, which exclusively measure cognitive or physical performance. The comprehensive nature of the FI could partly explain the strong and robust mortality association observed here and in previous studies (Kojima et al., 2018). Functional BAs, especially FI, measure not only the biological aging process, but also social and psychological wellbeing. Consequently, disentangling the underlying biological mechanisms from the functional measures would be challenging, which impairs the ability of functional BAs to explain the innate biological aging process. Specifically, it is difficult to implement functional BA measurements in animal models, whereas molecular BAs can be easily investigated in animal studies. Additionally, functional BAs in humans tend to show less heterogeneity among the young, making them less ideal instruments to assess the aging process among young adults, compared to using molecular BAs. This is especially true when it comes to FAI as it only takes four functioning indicators into account. However, FI considers a wide range of health deficits (42 in our case) and provides relatively precise scales to differentiate individuals with lower functioning scores. Furthermore, since biological, psychological, and social factors all contribute to the health status of human beings, the comprehensive nature of functional BAs also comes with important advantages for predicting aging-related outcomes, underlining the complementary nature of molecular and functional BAs.

In the subgroup analysis, we found moderate effect modification from age and smoking status. A previous study in a large Swedish cohort from our research group also observed that the effect size of FI with mortality was attenuated with increasing age (Li et al., 2019). DNAmGrimAge includes smoking (pack-year) associated CpG sites in its establishment (Lu et al., 2019), which could partly explain a stronger association observed in smokers than non-smokers as it may capture smoking-based deterioration at the molecular level. In addition, the present results found a majority of the BA-mortality associations to be stronger in women compared to men, suggesting sex differences should be acknowledged in making use of BAs to inform clinically relevant decisions.

It is noteworthy that in comparing the strength of relative risk of mortality related to each BA, we were not aiming to propose any superior BA, as mortality association alone is by no means the gold standard for a qualified BA in aging research. Evidence exists that some BAs have advantages in predicting other aging-related outcomes such as entry into care (e.g., FAI) (Finkel et al., 2019), which may be more relevant for developing policies and interventions.

The present study has several strengths. First, we measured nine BAs in the same population from the molecular to the organismal level, which provided us an opportunity to examine BA characteristics from several molecular and functional aspects. Second, the BAs were assessed in a longitudinal manner to allow for descriptive and correlational analyses considering changes over time. Third, through the linkage to the Swedish national registers, we were able to keep track of study participants’ vital status for nearly 20 years, thus being able to predict the risk of death in relation to BA status prior to the occurrence of most aging-related outcomes.

Several limitations should be taken into consideration when interpreting the present findings. First, only individuals with complete measurements were included in the nine-BA survival models. Complete observations may not be a representative subgroup of the entire study population and lead to selection bias. However, IPT occasions with incomplete BAs were mostly due to administrative reasons, such as methylation data were only assessed in five waves since IPT3, and these factors are likely not related to subject-specific features. In addition, we replicated the one-BA survival models in individuals with complete measurements, and the results did not show major discrepancies. Second, the birth year of SATSA participants spanned a long period of time, from the year 1900 to 1948, thus raising concerns for cohort effects. During this time, improved medical conditions or emergent public health events could affect both BA and mortality risk and cannot be controlled merely through the age scale in the present analysis. All survival models were stratified by participants’ birth year in 10 year intervals to in part alleviate bias due to the cohort effect.

In summary, biological age is a construct built on molecular, cellular, and functional aspects of an individual´s health status and has the potential to change the way risk prediction is performed in the clinic. In the era of personalized medicine, individual health assessments are warranted and should rely on better methods than simply using the CA. In this study, we present the most comprehensive analysis of biological age in humans to date. We explain their correlations and longitudinal patterns. We highlight the sex-specific effects and point to the importance of providing sex-specific estimates in aging investigations and acknowledging sex-specific care in geriatric practices. Furthermore, we found BAs had the potential to provide mortality-relevant information independently of CA and independently of other types of BAs. Further studies are needed to investigate if the present findings could be extrapolated to a younger population, what the specific mechanisms underlying the BA correlations are, and possible modification effects on and between BAs.

The Swedish Adoption/Twin Study of Aging (SATSA) was a population-based study consisting of reared apart and reared together twin pairs (Finkel and Pedersen, 2004). In-person testing (IPT) in SATSA was initiated in 1986 and nine complete waves of in-person testing were conducted through 2014 (IPT1 to IPT10, except for IPT four where only telephone interviews were performed), during which the information of a maximum of nine BAs, including telomere length, four DNAmAges (Horvath’s, Hannum’s, PhenoAge, and GrimAge), physiological age, and three functional ages (cognitive function, FAI, and FI) were available. In total, 859 individuals participated in at least one IPT wave in SATSA, and 846 individuals who had at least one BAs assessed once since IPT1 to IPT 10 (except for IPT4) and the information of vital status through linkage between the Swedish Twin Registry (STR) and the Swedish Population Register were included in the present study. The number of individuals and measurements by IPT and BAs were detailed in Table 1 and Supplementary file 1A.

Sex, educational attainment, and baseline information of BMI, and smoking status were considered as covariates in the survival analyses. BMI was assessed through physical examination and other covariates were acquired through self-reported questionnaire data. Educational attainment was classified as primary education, lower secondary or vocational, upper secondary education, and tertiary education. BMI was derived as the body mass divided by the square of body height, expressed in units of kg/m². Smoking status was categorized as non-smokers, ex-smokers, and current smokers. Among the included individuals and measurements, 28 out of 846 individuals without education information were assigned into an unknown group, and another one and two individuals who missed the BMI and smoking status were imputed with the corresponding information collected from their nearest available IPTs.

Since the participants in the present analysis were born over a long period of time (1900 to 1948), we estimated the mortality associations with stratification on the calendar year of the participants’ birthday to avoid confounding from cohort effects. Individuals’ birth year was treated as a categorical variable with five groups in an interval of 10 years.

All-cause mortality data, including vital status and dates of deaths, were obtained from linkages between the STR and Swedish Population Register through the personal identification number assigned to all residents. All-cause mortality data were updated on August 16, 2018.

Characteristics of baseline measurements were presented as means (standard deviations, SDs) and proportions. Longitudinal changes of BAs were presented in plots with individual level measurements illustrated by dots, lines, and broken lines. Average trajectories of BAs were estimated through mixed linear regression models as described above, each of which included fixed effects of the intercept, CA, and sex, and a random effect of intercept at the individual and twin-pair level. P-values for sex effects were obtained from fixed effects within the mixed linear model. In addition, an interaction term between each BA and CA in natural spline was introduced to the model, and P-values for sex interaction were determined by the likelihood ratio test comparing the models with and without the interaction term.

We estimated the repeated measures correlation coefficients between age and nine BAs using all complete measurements (Bakdash and Marusich, 2017). The correlation analyses were then replicated for CA and nine BA residuals to examine the correlations between the parts of BA residuals. We then transformed correlation coefficients to scaled squared Euclidean distances and performed hierarchical cluster analysis on BAs and BA residuals via Ward’s method.

We used the Cox regression model to estimate the association between baseline BAs and the risk of all-cause mortality. When repeated measurements were available, only the first BA assessment was included and referred to as the baseline BA. In all models, we used attained age as the underlying time scale and the five groups of birth year as strata. Left truncation and right censoring were accounted for in the estimations; that is individuals entering into the cohort is conditional on their survival at baseline ages, and follow-up time may stop before death occurs. Follow-up time started from the first available measurement (baseline measurement) and ended at the date when they died or were censored on August 16, 2018. In addition, robust standard errors were introduced to adjust for relatedness within twin pairs and subjects. To achieve a direct comparison of BA effects in relation to mortality risk, we standardized all nine BAs to the mean of zero and SD of one across all available measurements and replicated the transformation in the BA residuals, such that the estimated hazard ratios (HRs) could be interpreted as the relative risk of death associated with a one-SD increase in the level of BA or BA residual. In the first part of the survival analyses, all models accounted for only one BA (one-BA models). For those who had their BAs assessed for multiple times, the first available measurement was treated as the baseline measurement. For each BA, we estimated the association of BAs and BA residuals with mortality risk using two models, a univariate model and a multivariate model with sex, education, BMI, and smoking status as additional covariates. In the second part of the survival analyses, all nine BAs were accounted for altogether in the same survival model (multi-BA models). Similarly, two models with or without adjustment for common risk factors were estimated. Only BA residuals were taken into consideration in the second part of survival analyses to avoid biased results caused by collinearity within BAs.

P values were two-sided, and statistical significance was defined as p<0.05. All analyses were conducted using Stata 15.1 and R 3.6.0.

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

Human subjects: All participants provided informed consent. This study has received ethical approval from the Regional Ethics Review Board, Stockholm (Dnr 2016/1888-31/1).

1. Received: August 30, 2019
2. Accepted: December 24, 2019
3. Version of Record published: February 11, 2020 (version 1)

© 2020, Li 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.