Abstract
We investigate lifetime socioeconomic health disparities through longitudinal biomarkers from the Dutch Lifelines cohort study and biobank. By constructing an allostatic load index from 12 biomarkers, we analyze the dynamics of health and its association with socioeconomic status (SES) over the life cycle. Our findings reveal that health risks linked to lower SES emerge early and precede chronic disease onset. Further analyses investigate the drivers of allostatic load and emphasize health behaviors. Our research highlights the need for early interventions targeting SES-related health disparities and provides new insights into the physiological pathways linking SES to long-term health outcomes.
I. Introduction
The narrowing of socioeconomic health disparities has become a consensus objective for governments and international organizations; see, for instance, the EU4 Health Programme 2021–2027 (European Commission 2021) and the Netherlands Global Strategy 2023–2030 (Ministry of Foreign Affairs 2023). These policy programs argue that reducing the health gap is a major way to improve population health. To design effective preventive interventions, we need to know when and how associations between health and socioeconomic status (SES) arise.
One of the main challenges in this endeavor concerns the lack of consensus on how to measure health. Much of the prior literature on health disparities across socioeconomic groups has employed morbidity, mortality, or self-rated health as outcome measures (see, for example, Van Kippersluis et al. 2010; Hosseini, Kopecky, and Zhao 2022). This has provided valuable insights, but self-rated health (SRH) is inherently subjective and nonspecific, while morbidity measures often capture outcomes realized later in life, after much of the cumulative wear and tear has already occurred. Drawing on the link between chronic disease and mortality in old age, the ground-breaking recent study by Danesh et al. (2024) uses a range of Dutch population-wide register data to show that the income-related health gap in chronic disease, which translates into mortality differences, already emerges earlier in life. We extend this insight by considering biomarkers, which are objective indicators that reveal “under-the-skin” health risks that potentially precede the onset of diagnosable disease.
Specifically, we study the evolution of health and the dynamics of socioeconomic health disparities over the life cycle by using longitudinal data on biomarkers derived from blood, electrocardiograms (ECG), anthropometric measurements, and blood pressure within a large-scale, population-based, prospective cohort and biobank. Biomarkers are normally seen as objective, quantifiable indicators of biological processes (Strimbu and Tavel 2010). The dynamics of biomarkers are often linked to the aging process, the onset of diseases, and mortality (Arbeev, Ukraintseva, and Yashin 2016). Consequently, biomarkers not only reflect an individual’s current health status but also serve as predictive indicators of morbidity or mortality. Complementing clinical health assessments, the longitudinally observed biomarkers from prospective cohort studies offer an opportunity to investigate socioeconomic health gradients before the emergence of diseases. Moreover, by tracking the accumulation of physiological health deficits, these biomarkers provide deeper insights into the interplay between SES, biological processes, and clinical outcomes over the life cycle (Arbeev, Ukraintseva, and Yashin 2016).
An emerging literature in economics and epidemiology has started to employ biomarkers to investigate health disparities. Prior studies have identified significant SES-related disparities in biomarkers associated with diabetes and cardiovascular disease, as well as body mass index (BMI) (Baum and Ruhm 2009). Furthermore, systematic combinations of biomarkers that indicate cumulative health risks reveal considerable disparities in biomarker-related risks1 across SES groups (for example, Seeman et al. 2004; Carrieri, Davillas, and Jones 2020; Davillas and Jones 2020).
Following this, we examine the dynamics of health disparities using “dynamic” (longitudinally observed) biomarkers from the Dutch Lifelines cohort study and biobank, which includes data from more than 167,000 individuals at baseline. By linking the longitudinally observed biomarkers with information on chronic diseases, health-related behaviors, and sociodemographic factors, we are in the unique position to study the evolution of socioeconomic health disparities across the life cycle and the role of biomarkers in the relationship between SES and health outcomes.
Following the approach of Seeman et al. (1997, 2004), we adopt the concept of allostatic load and construct an allostatic load index (ALI) based on 12 biomarkers from cardiovascular, metabolic, and kidney systems, representing cumulative physiological dysregulation due to stress and aging. To assess the relevance of the ALI in predicting health outcomes, we examine its relationship with aging-related chronic disease,2 with a particular focus on how early biomarker-related risks predict chronic disease prevalence. To engage with the literature focusing on mortality outcomes, notably the abovementioned study by Danesh et al. (2024), we also conduct analyses that investigate whether biomarker-related risks can predict mortality.
Our analyses suggest that biomarker-related risks emerge rather early in adulthood, preceding the onset of aging-related chronic diseases, which typically become prominent only in middle age. In addition, we conduct age-group-specific regressions to examine the role of the ALI in chronic disease development and three-year mortality. Those results demonstrate that both ALI and lagged ALI are significantly associated with an increased risk of chronic diseases and three-year mortality.
Next, we examine how educational disparities in biomarkers and allostatic load evolve over the life cycle, using graphical analysis. Our findings show that allostatic load disparities emerge in early adulthood, widen with age, and peak in late middle age, stabilizing thereafter. Gender differences are significant, with males consistently exhibiting higher ALI levels than females throughout the life course. Additionally, we analyze the prevalence gap for individual biomarkers and biomarker-related risks. The results reveal that disparities in biomarker-related risks emerge early, often before age 30, and exhibit pronounced educational and gender differences, with males generally showing higher risk levels for most biomarkers. Our findings highlight the early onset and cumulative nature of socioeconomic health disparities, with virtually no stage of adulthood at which predictors of future mortality do not differ across educational groups.
Finally, we investigate factors driving allostatic load levels and the growth of allostatic load over the life cycle. We employ age-group-specific regression by gender and decompose the total R-squared using the Shapley and Owen decomposition method. The decomposition reveals that alcohol consumption and physical activity are important contributors to the ALI across genders and age groups. Educational attainment and employment also play notable roles, with education having a persistent impact and employment being more influential during working years. These results vary by gender. For females, alcohol consumption and education have stronger effects, while for males, physical activity and smoking are more important contributors, particularly before age 55. When examining the growth of the ALI, behavior towards a healthy lifestyle remains a key driver, but the importance of various types of behavior shifts, where smoking plays a more substantial role for males. While these findings provide valuable insights into the relative importance of these factors, they reflect correlations rather than causation, and they are affected by the choice of biomarkers included in the study.
We contribute to the literature in several ways. First, we contribute to the existing literature on socioeconomic health gradients by investigating these gradients before the onset of clinical diagnoses. Typically, individuals with higher SES enjoy longer and healthier lives. Socioeconomic-status -related health differences have been found in mortality (Deaton 2003; Cutler and Lleras-Muney 2006; Van Kippersluis et al. 2010; Chetty et al. 2016) and in most diseases and conditions (Kivimäki et al. 2020; Pallesen et al. 2024; Danesh et al. 2024). However, morbidity and mortality differences are often only prominent at middle and higher ages, which opens the question of how the differences in health develop across SES before reaching the clinical endpoints. Evidence suggests that the socioeconomic health gap in chronic disease, which translates into mortality differences, has already emerged early in life (Danesh et al. 2024). Our paper explicitly builds on the latter study. Indeed, the observation of biomarkers offers an objective means to assess health risks and enables an investigation of how socioeconomic health disparities emerge before the onset of disease (Arbeev, Ukraintseva, and Yashin 2016).
Second, our study contributes to the literature on health dynamics more in general. While socioeconomic health gradients have been widely explored in prior studies, data limitations make it challenging to achieve a consensus on how to define and measure these gradients over the life cycle (Hosseini, Kopecky, and Zhao 2022; Danesh et al. 2024). Previous work examined health evolutions across the life cycle using indicators such as SRH, morbidity, and mortality. Among these, SRH is often employed as a health measure in studies on socioeconomic health disparities and is generally regarded as a reliable predictor of other health outcomes. For instance, Van Kippersluis et al. (2010) use SRH to analyze the life-cycle profile of adverse health by income in the Netherlands. However, SRH has inherent limitations—it is subjective, lacks specificity, and does not provide a cardinal metric (Hosseini, Kopecky, and Zhao 2022). Recent studies have attempted to overcome these issues by using more objective health indicators. Danesh et al. (2024) introduce a chronic disease index (CDI), constructed from prescription medication data, to investigate whether the mortality observed in old age has already materialized by early adulthood. Hosseini, Kopecky, and Zhao (2022) develop a frailty index, incorporating factors such as medical diagnoses, mental health conditions, and cognitive impairments, to predict health dynamics over the life cycle. We expand on this by relying on biomarkers to examine the progression of socioeconomic health disparities over the life cycle. Importantly, we observe that there is a nonnegligible share of individuals who do not report clinically elevated biomarkers or chronic disease, respectively, even at high ages. An advantage, therefore, of relying on population-based cohort and biobank data is that we are able to observe individuals, and their biomarkers, even if they are not in touch with the healthcare system as a consequence of a disease.
Third, our study offers insight into the drivers of health disparities in biomarkers. Economic and epidemiological research underscores the important role of health-risk behaviors—such as smoking, alcohol consumption, physical inactivity, and poor dietary habits—particularly among adults. These behaviors serve as critical pathways linking SES to health outcomes (Adler and Stewart 2010). Previous studies have estimated that health behaviors account for approximately 40 percent of premature mortality (McGinnis, Williams-Russo, and Knickman 2002) and significantly influence the prevalence and incidence of chronic diseases (Danesh et al. 2024). Among these behaviors, smoking has been identified as having a particularly detrimental impact on both physical and mental health. Furthermore, health-risk behaviors are closely associated with allostatic load. For instance, Suvarna et al. (2020) review 26 studies examining the relationship between health behaviors and allostatic load and find robust evidence of significant associations. Specifically, 65 percent of studies on obesity and substance abuse, 75 percent of studies on sleep, and 62.5 percent of studies on combined lifestyle factors report significant correlations with allostatic load. In the current study, we contribute to understanding how these factors contribute to the allostatic load and the growth of allostatic load and how they differ across age and gender.
In what follows, Section II describes the data. Section III outlines the methodology for constructing the ALI and explores its role of allostatic load to the aging-related chronic disease and mortality. Section IV provides graphical evidence of the evolution of socioeconomic allostatic load disparities over the life cycle. Section V presents the decomposition results and discusses their interpretation. Section VI concludes. Results of sensitivity analyses are reported in the Online Appendix.
II. Data
A. Lifelines
We utilize data from the Dutch Lifelines cohort study and biobank. Lifelines is a multidisciplinary prospective population-based cohort study examining in a unique three-generation design the health and health-related behaviors of 167,729 persons living in the North of the Netherlands. It employs a broad range of investigative procedures in assessing the biomedical, sociodemographic, behavioral, physical, and psychological factors that contribute to the health and disease of the general population, with a special focus on multi-morbidity and complex genetics. This prospective cohort study is designed to explore the complex relationships among various factors in the development of chronic diseases and healthy aging (Scholtens et al. 2015). The Lifelines study began in 2006, and by 2023, three main waves (including the baseline) and three follow-up questionnaires were completed.3
Every five years, participants are invited to Lifelines facilities for physical examinations, during which biomaterials are collected (Scholtens et al. 2015). These samples are promptly processed for analysis and preserved for long-term biobanking. Additionally, every 1.5 years, participants complete questionnaires that gather information on demographics, health status, lifestyle, environmental exposures, and psychosocial factors. All examinations are conducted by trained nurses following medical standards, and all assessments take place at the University Medical Center Groningen laboratory center, which is certified according to international, European, and Dutch standards.
B. Variables
1. Biomarkers
Biomarker data were obtained during physical examinations and biomaterial collection as part of the Lifelines cohort study. As of the end of 2023, three waves of biomarker data have been made available, encompassing a wide range of measurements, including anthropometric data, blood analyses, blood pressure, and ECG, among others. Specifically, the first wave was collected between 2006 and 2013, the second wave between 2014 and 2018, and the third wave between 2019 and 2023. These longitudinal data enable us to follow individuals’ health over a relatively long period.
For our analysis, we selected 12 biomarkers related to cardiovascular, metabolic, and kidney functions. Table 1 summarizes the selected biomarkers, along with brief descriptions and clinically defined thresholds.4 To capture the cumulative dysregulation of physiological systems, we employ the concept of allostatic load and construct an index to be the indicator of biological health status. Section III provides a detailed description of the allostatic load and the construction of the index.
Definitions and Clinical Cutoff Points of Biomarkers Used to Construct the ALI
2. Chronic disease
In the Lifelines self-reported questionnaires, administered every 1.5 years, participants were asked whether they had been diagnosed with specific diseases.7 These diseases are categorized into groups, such as cancer, cardiovascular diseases, diabetes, kidney and bladder diseases, mental illnesses, and neurological disorders. Within each category, specific conditions are further detailed. For example, cardiovascular diseases include stroke, heart failure, and heart attacks. To evaluate the overall burden of aging-related chronic diseases among Lifelines participants, we use a composite score as a proxy measure. The selection of chronic diseases is guided by the design of the Lifelines questionnaires, their definitions, and the availability of corresponding data.8 We include most of the chronic conditions from Lifelines’ list, while considering the timing of disease onset and the data availability in Lifelines.9 The complete list of 19 aging-related chronic diseases is provided in Table 2.10
Categories of Aging-Related Chronic Diseases Used to Construct the CDI
3. Mortality
Lifelines continuously receives mortality updates for participants from the Personal Records Database (BRP), which contains personal data of individuals residing in the Netherlands as recorded by municipalities. Participants’ mortality information is updated even if they withdraw from Lifelines, and the cutoff date for death records in our current data set is September 2024.
4. Health behavior
Health behavior data are collected from the Lifelines questionnaire. Participants were queried about various health-related behaviors, including alcohol consumption, tobacco use, smoking habits, sleeping disorder, and overall physical activity levels. To measure drinking behavior, we use self-reported data on both the frequency of alcohol consumption over a month and the number of glasses consumed over a day. These variables capture both the frequency and intensity of drinking. Smoking behavior is represented by a dummy variable based on responses to the question, “Do you smoke now, or have you smoked in the past month?” Physical activity is proxied by the average number of days per week participants engaged in activities, such as cycling, doing odd jobs, gardening, sports, or other strenuous tasks for at least 30 minutes. We calculate the average for these physical activities across winter and summer seasons. Finally, sleep disorder is captured by the question: “Do you have trouble sleeping nearly every night?”
5. Socioeconomic status
Lifelines provides socioeconomic data on education, income, and occupation. For this study, we use the highest educational attainment as the primary measure of SES. Educational attainment is categorized into two groups based on the Dutch school system: low (no education, primary education, lower or preparatory secondary vocational education, junior general secondary education, secondary vocational education or work-based learning pathway, senior general secondary education, pre-university secondary education) and high (higher vocational education and university education).11 For participants under the age of 25, we use their parent’s highest educational attainment as a proxy, given that individuals typically complete their education in their mid-20s. Additionally, we consider household net income and parental education in the sensitivity analyses as alternative measures of SES (Online Appendix D).
6. Covariates
Demographic factors, such as age, gender, cohort, and province of residence, are available for all participants in Lifelines. In addition, we obtain the degree of urbanization information at Postal Code-4 level from Statistics Netherlands (Centraal Bureau voor de Statistiek).
C. Sample Selection and Summary Statistics
We build an unbalanced panel based on the data from waves 1a, 2a, and 3a of the Lifelines study, covering 2006–2023. Table 3 presents our sample selection process. The baseline sample includes 150,605 observations of participants aged 18–80.12 Among these, 99,608 participants from the baseline sample participated in wave 2a, and 60,794 participated in wave 3a. Observations with missing values for any of the 12 biomarkers of interest are also excluded, resulting in the removal of 21,215 observations. Further, we exclude observations with missing values for chronic diseases or demographic characteristics or those who did not fast before blood sampling.13
Sample Selection Process for the Analytical Sample
Our final sample consists of 137,110 individuals. Of these, 46,796 participated in only one wave, 50,906 participated in two waves, and 39,408 participated in all three waves. As shown in Table 4, there is attrition across the three waves. Approximately 3.1 percent of participants passed away during the study period, which extended until 2024. Other reasons for withdrawal include time-intensive participation requirements, loss of interest, relocation from the research area, or enrollment in a regular healthcare program (Sijtsma et al. 2022).14
Sample Summary Statistics
Table 4 presents selected summary statistics on demographic and socioeconomic information, biomarkers, and chronic diseases across three waves of Lifelines participants in our sample.
III. Allostatic Load, Chronic Disease, and Mortality
A. Allostatic Load Index
Aging is a complex process involving numerous biological changes and interactions that gradually result in physiological dysregulation, disease, and ultimately death (Arbeev, Ukraintseva, and Yashin 2016). Although individual biomarker changes may seem small, the cumulative effect of multiple dysregulated biomarkers can significantly deteriorate health, impacting various body systems over time. To measure this cumulative biological dysregulation, we construct an index that captures the overall burden of dysregulated biomarkers. To do so, we follow established research to construct an index by summing biomarkers for which individual values deviate from clinical thresholds (Seeman et al. 2004; Howard and Sparks 2016; Davillas and Jones 2020).
Allostatic load, often referred to as “wear and tear,” represents the cumulative dysregulation of physiological systems over time due to stress (Seeman et al. 2004; Beckie 2012).15 Allostatic load has been widely used in health research, especially in studies on health measurements and inequalities, as it potentially provides insight into biological mechanisms underlying health disparities. The ALI is essential for understanding how sociodemographic factors and environmental stressors influence both physical and mental health, shaping individual aging trajectories (Beckie 2012).
Depending on the data availability, the number of biomarkers varies between studies. Seeman et al. (1997) calculate the ALI using ten biomarkers associated with the cardiovascular and metabolic systems and the hypothalamic-pituitary-adrenal (HPA) axis. Subsequent studies expanded this scope as additional biomarker data became available; for example, follow-up research employed 16 biomarkers to assess allostatic load (Seeman et al. 2004). More recent studies, such as those by Howard and Sparks (2016), use ten biomarkers, while Karimi et al. (2019) include 16 biomarkers spanning four body systems and two organs. Among 26 representative studies reviewed by Johnson, Cavallaro, and Leon (2017), all included at least one biomarker from the cardiovascular and metabolic systems. Despite variation in its construction, the ALI remains a valuable tool for understanding the physiological pathways linking SES to morbidity and mortality.
In our study, we use 12 biomarkers (see Table 1) to construct the ALI,16 focusing on three physiological systems: cardiovascular (n = 3), metabolic (n = 8), and kidney function (n = 1).17 The ALI is calculated by applying clinically established threshold cut points to each biomarker and counting the number of biomarker-related risks18 that individual i has at age a:
1where k denotes the biomarkers and Ii,a,k is a binary variable indicating whether the level of biomarker k in individual i at age a is above the threshold.19 The Ii,a,k is equal to one if an individual is identified as “at-risk” based on a certain biomarker’s cut point. The value of the ALI for individual i at age a represents the current number of biomarker-related risks based on 12 selected biomarkers. The allostatic load algorithm based on clinical cutoff points is the most commonly used approach to construct the ALI, and it has been validated as an effective method to capture health risks and to predict future health outcomes (McLoughlin, Kenny, and McCrory 2020).
To gain a preliminary understanding of the ALI without considering any other factors, we visualize the dynamics of these biomarker-related risks across age. We pool the observations from three waves and group them by the number of risks. Figure 1 illustrates the shares of observations at different ages with differing numbers of biomarker-related risks, depicting the evolution of biomarker-related risks throughout the life cycle. As shown, first, the number of biomarker-related risks increases with age, with a different speed by the categories in the number of risks. Notably, there is a significant rise in the development of risks after the age of 40. Additionally, biomarker-related risks are relatively prevalent even among young adults, and more than 40 percent of participants under the age of 25 have at least one risk. Interestingly, there is a nonnegligible share of individuals who display no biomarkers above clinically relevant cutoff points even at high ages.
Evolution of Biomarker-Related Risks over the Life Cycle
Notes: This figure illustrates the share of observations with varying numbers of biomarker-related risks across ages using pooled observations from an unbalanced panel data set. The sample covers ages 18–80. The number of biomarker-related risks refers to the count of biomarkers exceeding the clinical threshold (Table 1) for an individual at a given age. The different colors (gray shades in black and white version) represent different numbers of risks, with the lowest risk at bottom and highest risk at the top. The proportion of the color/shade at each age is the share of observations at this age.
Despite the insights gained from clinical cutoff points, this approach may overlook health risks that fall below the threshold, and it does not capture the relative importance of different biomarkers. Therefore, in sensitivity analyses, we adopt two alternative approaches to construct the ALI. First, following previous studies (for example, Seplaki et al. 2005; Hawkley et al. 2011), we calculate the ALI based on z-scores, which provides an index derived from continuous biological variables rather than categorical ones. Compared to the clinical cutoffs, the z-score-based ALI (z-ALI) captures risks both below and above the threshold and reduces the influence of extreme values and outliers. Second, inspired by McLoughlin, Kenny, and McCrory (2020) and Danesh et al. (2024), we employ a data-driven approach to construct a weighted ALI, in which weights are assigned to biomarkers according to their associations with chronic diseases at older ages. The weighted ALI is straightforward and informative, as it incorporates biomarker-specific weights and could improve the efficiency of predicting health outcomes. Online Appendix C provides a detailed comparison of the three different allostatic load algorithms, as well as the procedures for calculating the z-ALI and weighted ALI.
B. Chronic Disease
Chronic diseases are widely recognized as a substantial burden on healthcare systems, with many conditions becoming prominent in middle adulthood. These diseases significantly contribute to socioeconomic disparities in healthcare expenditures and mortality rates, further exacerbating health inequalities later in life (Danesh et al. 2024). In this study, we adopt a variant of the CDI proposed by Danesh et al. (2024) to measure the overall burden of chronic health conditions using different data and methodology.
We use a count-based approach to construct a CDI based on self-reported information covering a broad range of aging-related chronic diseases, including cardiovascular conditions, diabetes, and neurological disorders.20 The proposed CDI reflects the cumulative number of adverse health events an individual has experienced at the time of participation in the Lifelines.21 The resulting index can be treated as a continuous variable or normalized to a scale ranging from zero to one. The main advantages of the count-based method lie in its simplicity, transparency, and ease of replication, as it does not rely on other health outcome measures.
Specifically, 19 aging-related chronic diseases from the Lifelines data set are used to construct the CDI (Table 2).22 Each chronic disease is represented by a binary variable, taking a value of either zero or one for each individual, indicating whether the individual currently has or has previously had the disease. The CDI is calculated as the total number of chronic diseases an individual has experienced by a given age.
Similar to the analysis of ALI, we employ a stacked area graph to examine the progression of chronic disease prevalence across age groups by pooling all observations. As illustrated in Figure 2, the onset of aging-related chronic diseases typically occurs after early adulthood. Among young adults, the majority do not have any of the included chronic diseases, while only a small proportion have one chronic condition. Furthermore, the prevalence of these chronic diseases becomes substantial after age 35, with a marked increase observed only after around age 45. Again, we observe a nonnegligible share of individuals who do not report any chronic diseases, even at high ages. Therefore, an advantage of Lifelines is that we observe health indicators of individuals even if they are not in healthcare, thus providing an insight into health as well as disease.
Evolution of Aging-Related Chronic Diseases over the Life Cycle
Notes: This figure illustrates the share of observations with varying numbers of aging-related chronic diseases across ages using pooled observations from an unbalanced panel data set. The sample covers ages 18–80. The number of chronic diseases refers to the count of conditions an individual currently has or has previously had at a given age. Details on the list of chronic diseases are provided in Table 2. Different colors (gray shades in black and white version) represent different numbers of chronic diseases, with the lowest at the bottom and the highest at the top. The proportion of the color/shade at each age is the share of observations at this age.
Next, we compare the trajectories of ALI and CDI across the life cycle. Figure 3 illustrates these trajectories by age. To ensure comparability of scale, the indexes are rescaled using the min–max scaling approach.23 The figure reveals that biomarker-related risks emerge early in adulthood and precede the onset of aging-related chronic diseases, prompting the question: to what extent can ALI predict these chronic diseases? From an intervention perspective, understanding whether early intervention before the onset of diseases is necessary is essential.24
Life Cycle Profile of ALI, CDI, and Three-Year Mortality
Notes: This figure presents the trajectories of ALI, CDI, and three-year mortality rate over the life cycle using pooled observations from an unbalanced panel data set. The sample covers ages 18–80. Due to limitations in death record availability, the number of observations for three-year mortality rate is smaller than for ALI and CDI (see Section III.C for details). For CDI, the figure (and subsequent figures that stratify by gender and education) follows the corresponding figures in Danesh et al. (2024). In our setting, to ensure comparability between CDI and ALI, we rescale the two indexes using a min–max scaling approach. We apply the lpoly smoothing method, a kernel-weighted local polynomial regression, to smooth the trajectories of ALI, CDI, and three-year mortality. Specifically, we use the Epanechnikov kernel with a zero-degree polynomial, which minimizes noise while preserving the underlying data pattern. The 95 percent confidence intervals are shown. The larger variation at younger and older ages reflects the smaller sample sizes in these groups.
C. Mortality
Mortality, as the final stage of the life cycle, typically becomes significant only later in life, occurring well after the onset of chronic diseases. To connect with previous research on mortality, we also examine whether ALI can predict mortality. We determine three-year mortality based on the survey date and the recorded date of death from Lifelines and BRP. However, due to data limitations, we are unable to identify the three-year mortality of participants who took the survey after September 2021 and remained alive until the cutoff date for death records. As a result, our sample size for mortality analysis is smaller than that for biomarkers and chronic diseases.25
Figure 3 also illustrates the trajectory of three-year mortality among Lifelines participants. Compared to previous research, the absolute three-year mortality rate is relatively low. This can be attributed to the fact that the Lifelines cohort consists of a relatively young, noninstitutionalized population, with older individuals not being the primary focus. Additionally, mortality rates derived from survey data tend to be lower than those observed in the general population (Keyes et al. 2018). As we can see, the three-year mortality remains relatively low before age 50 but rises sharply after age 60. In contrast, CDI increases steeply before age 50, while biomarker-related risks become significant and escalate already before age 30. This progression demonstrates a transition from biomarker-related risks to the onset of chronic diseases and, ultimately, to mortality.
D. Allostatic Load and Disparity in Chronic Disease and Mortality
As shown above, the number of aging-related chronic diseases increases later in life, while biomarker-related risks have already emerged early in adulthood before chronic diseases happen. Previous research has highlighted a significant association between the ALI and mortality risk and has shown that it explains a significant portion of the SES-related mortality gap, with findings suggesting that the ALI accounts for a substantial portion of the differences in mortality risks across SES groups (for example, Seeman et al. 2004; Howard and Sparks 2016). Here, we build on prior work by investigating whether a cumulative index of biological risk, namely ALI, can predict the prevalence of chronic disease and three-year mortality.
We start with a linear regression of CDIi,a on a set of controls Xi,a, including age, age-squared, age group, gender, cohort, urban, province, and survey year. Then, we add the ALIi,a and the lagged term of ALIi,a into the regression. To do that, we restrict our sample to individuals who participated in at least two consecutive waves and pool all the observations. Afterward, we repeat the regression by age groups with ten-year intervals.
2Table 5 examines the extent to which the ALI and its lagged term predict the CDI and three-year mortality when sequentially added into the model. The analysis reveals two key points. First, ALI and its lagged term demonstrate a significant positive association with the CDI and three-year mortality, indicating that higher contemporary and lagged ALI correspond to increased CDI and mortality values. This finding supports the role of ALI as an early indicator of aging-related chronic disease prevalence and mortality. Additionally, in Table 6 and Table 7, we present the regression results by ten-year age groups. The coefficients for ALI and its lagged term increase with age group, indicating that their association intensifies in later life.
Regression Results of CDI and Three-Year Mortality on ALI and Lagged ALI
Regression Results of CDI on ALI and Lagged ALI by Age Group
Regression Results of Three-Year Mortality on ALI and Lagged ALI by Age Group
As a sensitivity analysis, we use z-ALI and weighted ALI to predict CDI and three-year mortality. The results, reported in Online Appendix Tables C.1 and C.2, are consistent with those obtained using ALI. This finding aligns with previous studies (for example, McLoughlin, Kenny, and McCrory 2020) suggesting that the choice of allostatic load scoring algorithm has a relatively small impact on predicting general health outcome.
IV. Socioeconomic Health Disparities over the Life Cycle
We start by considering educational disparities in health by graphical analysis. We present the trajectory of educational disparities in allostatic load over the life cycle and display disparities in each biomarker and related risks among young adults. This analysis provides insights into the timing through which socioeconomic health disparities emerge and evolve.
The health gap between educational groups is defined as ΔHealtha = Healtha,high − Healtha,low, representing the difference in health outcomes between individuals with high and low levels of education at a given age. Simply tracking how this gap changes with age offers insight into when and how the educational disparities in biomarkers and allostatic load open. This simple comparison is valuable because it provides information on the timing of education-related health disparities before endpoints.
A. Trajectory of the Allostatic Load Disparity over the Life Cycle
At the outset, we examine how allostatic load disparity across education groups evolves over the life cycle. This analysis aims at understanding the onset of health disparities in allostatic load and the pattern over the life cycle.
Figure 4 presents the trajectory of educational allostatic load disparity by age, pooling observations from three waves of data, separately for males and females. The figure highlights that educational disparities in allostatic load become evident in early adulthood and consistently increases with age, peaking in late middle age.
Educational Disparities in the ALI by Age and Gender
Notes: This figure presents the trajectories of educational disparities in ALI by age and gender using pooled observations from an unbalanced panel data set. The sample covers ages 18–80. High education is defined as higher vocational or university education according to the Dutch education system, while the rest of the sample is classified as low education (see Section II.B for details). We apply the lpoly smoothing method, a kernel-weighted local polynomial regression of y on x, using the Epanechnikov kernel function with a zero-degree polynomial. The 95 percent confidence intervals are shown. Larger variation at younger and older ages reflects smaller sample sizes in these groups.
To interpret the magnitude of these differences, consider the ALI for the low education group at age 30, which is comparable to that of the high education group at around age 44 for females and age 37 for males. This indicates a substantial biological health gap across socioeconomic groups. Additionally, the absolute gap in allostatic load reflects the cumulative burden of biological risks. For instance, at age 40, females in the low education group have, on average, 0.31 more biological risks than their high education counterparts. Given that the ALI for the high education group for females at this age is 0.89, this represents a relatively large disparity.
Gender differences are also evident in both the levels and trajectories of allostatic load over the life course. For females, the ALI is consistently lower than that for males throughout the life cycle. The rate of increase in the ALI for females begins to accelerate slightly before age 40 and rises more sharply till age 60. The educational disparity in ALI for females continues to widen until around age 50, after which it stabilizes. For males, the average ALI is higher than that for females across the entire lifespan. The increase in ALI and the corresponding educational disparities occur more rapidly for males compared to females and tend to stabilize around age 55.
For both females and males, the increase in allostatic load tends to slow down around the ages of 50–60. This pattern aligns with the findings of Lleras-Muney and Moreau (2022), who demonstrate that SES gradients in mortality widen with age but eventually decline after a certain point. Several factors may explain this decline. First, the prevalence of chronic diseases at late middle age often leads individuals to begin treatment, which can reduce the levels of certain biomarkers, such as HbA1c.26 Second, as individuals experience health issues, they tend to place higher value on maintaining their health. This shift in priorities often leads to increased investment in healthier behaviors and lifestyles, such as engaging in more physical activity or adopting healthier lifestyle habits. Third, health-based attrition may contribute to this trend, as individuals with higher allostatic load are more at risk of dropping out of the Lifelines study, for instance, because of deteriorating health or death.
These patterns underscore that education-related health disparities emerge early and widen with age, with no stage of adulthood at which future mortality predictors converge across educational groups. Our findings are consistent with those of Danesh et al. (2024), who show that the income-related health gap, which translates into mortality differences at older ages, has already materialized by early life.27 By focusing on biomarker-based health measures, our study complements their study and contributes new evidence on the early emergence and progression of socioeconomic gradients in health.
The static comparisons of health gaps could be confounded by factors such as cohort effects, health-based sorting, and health-based attrition. These factors will shape our graphic analysis. For example, SES measures may be endogenous to individual health, as poor health may lead to lower educational attainment, potentially shaping the pattern of SES-related health disparities over the life cycle.28 In our study, we are mainly interested in the differences in the health evolution by educational attainment. Educational attainment often becomes stable after early adulthood, which reduces the issue of health-based sorting.
Another concern is the cohort effects. While the longitudinal nature of the Lifelines study supports cohort analysis, the currently available biomarker data include only three waves, allowing us to track individuals for an average of 11.2 years and a maximum of 17 years. Ideally, we would construct a cohort specific to each birth year, but this results in too few observations for each cohort in each wave. Instead, to test for the significance of cohort effects, we create eight cohorts using ten-year birth intervals and then compare the average health outcomes of different cohorts at the same age across waves. Due to limited observations in the oldest cohort, we exclude individuals born before 1930 from the analysis. Although we do not aim to capture the cohort effect, this setting allows us to observe the extent of cohort effects by comparing the average ALI at the same age but in different cohorts. We present the graphical analysis in Figure 5 by showing the extent of cohort effect.
Educational Disparities in ALI Across Age and Birth Cohorts
Notes: This figure examines the importance of cohort effects by grouping individuals into ten-year birth cohorts and presenting educational disparities in ALI across ages using pooled observations from an unbalanced panel data set. For example, “80s” refers to participants born in the 1980s. However, the figure still reflects a mix of cohort and age effects because individuals are not followed continuously over time. We apply the lpoly smoothing method, a kernel-weighted local polynomial regression of y on x, using the Epanechnikov kernel function with a zero-degree polynomial. Large variation is observed for the 30s and 90s cohorts due to smaller sample size.
Additionally, we present sensitivity analyses using alternative approaches to construct the ALI, as well as alternative SES measures to examine SES-related ALI disparities over the life cycle. Specifically, we conduct analyses based on the z-ALI and weighted ALI (Online Appendix Figures C.1 and C.2), and we use parental education and household net income (Online Appendix Figures D.1 and D.2) as alternative SES indicators. The sensitivity analyses show that regardless of how these measures are constructed, the conclusions remain consistent with our main findings. Further details are provided in the respective Online Appendixes.
B. The Prevalence of Biomarker-Related Risk
Next, we examine the prevalence gap in biomarker-related risks across education groups. Specifically, an individual i at age a is considered to have a risk for a specific biomarker if their biomarker value exceeds the threshold. The prevalence gap is defined as Bioa,high − Bioa,low, where Bioa,high represents the average value of individuals with high educational attainment at age a, and Bioa,low represents those with low educational attainment.
Figure 6 presents the difference in the share of high-risk biomarker observations between individuals with low and high educational attainment among those under age 30. Several key observations can be drawn from this. First, the figure reveals a noticeable biomarker-related risk for young adults, which contrasts with the more substantial morbidity and mortality typically observed in middle-aged and older individuals. For instance, the prevalence of BMI and waist-to-hip ratio (WHR) risks for males is 9.8 percent and 41.4 percent, respectively, indicating that 9.8 percent of males under age 30 have BMI values above 30, and 41.4 percent have WHR values above 0.85. Both BMI and WHR are commonly associated with diseases such as diabetes and metabolic syndrome, which generally manifest later in life.
Educational Disparities in the Prevalence of High-Risk Biomarkers Among Individuals Aged Under 30
Notes: This figure illustrates differences in the prevalence of each biomarker-related risk across education groups, segmented by gender. Observations from all three waves are pooled for this analysis. For example, a value of 0.027 for females’ heart rate indicates that 2.7 percent of females have a heart rate above 90 beats per minute (see Table 1 for clinical thresholds for other biomarkers). Among females, the differences in total cholesterol, HbA1c, glucose, and diastolic blood pressure between low- and high-education groups are statistically insignificant at the 10 percent level. Among males, the differences in heart rate and diastolic blood pressure are statistically insignificant at the 10 percent level. All other differences are significant at the 10 percent level.
Second, we observe a significant educational gradient in the prevalence of biomarker-related risks before age 30. For example, the prevalence gap in HDL cholesterol risk is approximately 3.8 percent lower among females with high educational attainment compared to those with low educational attainment. Among males, the gap is around 6.7 percent between individuals with high and low educational attainment.29 Additionally, gaps are also observed for LDL cholesterol, systolic blood pressure, creatinine, triglycerides, total cholesterol (for males), WHR, and BMI. One exception is creatinine, where the lower education group exhibits lower values than the higher education group for females.
Third, we find a pronounced gender difference in the prevalence of biomarker-related risks. The risk tends to be higher in males for LDL cholesterol, systolic blood pressure, creatinine, triglycerides, total cholesterol, HDL cholesterol, and WHR. In contrast, females exhibit slightly higher prevalence rates for BMI and heart rate. Furthermore, the prevalence gap is generally larger for males across most biomarkers compared to females. Additionally, we also present the results for age older than 60 and the whole sample (Online Appendix Figures A.2 and A.3).
V. Drivers of Allostatic Load over the Life Cycle
A. Framework
Our prior graphical analysis demonstrates educational disparities in biomarkers and allostatic load that begin to manifest in early adulthood and progressively widen until late middle age. In this section, we focus on factors contributing to allostatic load and its growth, as well as how the contribution of these factors differs by gender and age groups. Although our analysis is primarily descriptive, it offers valuable insights into the relative importance of various determinants and highlights how targeted health interventions can mitigate biomarker-related risks. Specifically, our analysis emphasizes the role of health-related behaviors in driving the gradient of ALI.
To estimate the role of these factors in determining allostatic load, we estimate the following linear regression for each age group g (defined in ten-year age bins),30 using individuals whose age a falls within group g:
3In this equation, ALIi,a denotes the ALI of individual i at age a within age group g. Xi,a,j represents the jth explanatory variable that includes health behaviors (smoking, drinking, physical activity, and sleep disorders), educational attainment, employment status, neighborhood SES, and basic demographic controls (age, age-squared, urban residence, and partnership status).
To assess the relative contribution of these variables, we decompose the total R-squared of these regressions using Shapley and Owen decomposition methods (Huettner and Sunder 2012). This approach enables us to evaluate the average contribution of each predictor to the explained variance across all possible sequences of regressors.31
B. Decomposition Results
Figure 7 illustrates the relative importance of different determinants of ALI using gender- and age-specific regression decompositions. Each column represents the total R-squared contributed by the listed factors in the corresponding regression. Health behaviors are consistently among the most important contributors to ALI across both genders and age groups.
R-Squared Contribution of Modifiable Factors to ALI by Gender and Age Groups
Notes: This figure presents the Shapley–Owen decomposition results of the relative importance of modifiable factors for ALI by age group. For each age group, we run a linear regression of ALI on education, neighborhood SES, employment, a set of modifiable factors, and basic demographics. The stacked bars display the contribution of each factor to the overall R-squared of the models. Details on these modifiable variables are provided in Section II.B.
For females, alcohol consumption contributes substantially to allostatic load, while the contribution of physical activity increases gradually after age 25. Educational attainment remains a relatively stable contributor across the life cycle. In contrast, employment mainly matters during prime working ages. Smoking has a moderate contribution across age groups, whereas sleep disorders exhibit a more pronounced effect in middle age.
For males, physical activity is a dominant contributor throughout the life course, especially after age 55. Alcohol consumption steadily increases in importance with age, while the contribution of smoking is substantial only before age 55. The explanatory power of education declines with age.
Online Appendix Figure A.4 presents decomposition results of the growth in ALI. Consistent with the findings in Figure 7, health behaviors appear as significant contributors to the rise in allostatic load for both females and males. In Online Appendix B, we further investigate the role of BMI by removing it from the ALI and treating it as a behavioral variable. BMI accounts for a relatively large share of the gradient in ALI, accompanied by a slight reduction in the contributions of other factors. This underscores the complex role of BMI—it not only shapes overall ALI but also contributes to the observed socioeconomic gradient. These findings are consistent with interpreting BMI as a behavioral factor. However, disentangling this from the biomarker interpretation of BMI remains a topic for future research.
It is important to emphasize that the decomposition results reflect correlations rather than causal relationships and represent a lower bound of each factor’s relative contribution.32 The results are also influenced by the selection of biomarkers included in the ALI. For example, biomarkers related to lung function, the nervous system, immune function, and skeletal health are not available in our data. As a consequence, the ALI measure may be more strongly associated with some health behaviors than others. Nonetheless, the decomposition offers useful insights into how these behaviors relate to key physiological systems, such as cardiovascular system, metabolic system, and kidney function, captured in the ALI.
VI. Conclusions
Using representative data from Dutch Lifelines, we investigate the life-cycle profile of biomarker-related health and its underlying determinants using objective biomarkers obtained from longitudinal biomaterial collection and measurements. We develop an allostatic load index (ALI) to reflect physiological dysregulation in response to stress exposure, which also indicates the cumulative risks of chronic conditions. Complementing previous research, our study underscores the significant role of allostatic load as a predictor of chronic disease development and mortality.
In our life-cycle analysis, we observe that biomarker-related health disparities in SES emerge early in adulthood, well before diseases manifest themselves clinically, for both males and females. For instance, the biomarkers exhibit notable gradients at the onset of adulthood. Educational differences in allostatic load continue to widen during adulthood for both males and females, while also showing a difference in the pattern by gender. Furthermore, we decompose the total R-squared of regression to assess the average contribution of various factors to the allostatic load. We find health behaviors play an important role in allostatic load and the growth of allostatic load, with different behaviors demonstrating distinct relative importance across age groups and genders. Educational attainment emerges as a significant determinant for both males and females throughout the life cycle.
The current analysis has several limitations that need to be acknowledged. First, the construction of the ALI lacks a uniform approach across the literature, making comparisons with other studies challenging, and results might be driven by the number and type of biomarkers selected. Second, potential confounders, such as medication use and health-based attrition, have not been accounted for in the graphical analysis, which could affect the observed patterns of health disparities later in life. Third, the decomposition method employed is relatively straightforward, and the findings are partially shaped by the selected biomarkers and health-related behaviors. The analysis does not account for other factors, such as environmental exposure or early life conditions and parents’ SES, both of which are considered to have a significant influence on physiological dysregulation. Moving forward, addressing these limitations is essential. Future analyses could also put effort into accounting for confounders of graphical analysis and capturing the biological aging speed based on dynamic biomarkers.
Acknowledgments
They thank Hermien Dijk, Marcus Ebeling, and participants at the MPIDR Lab Talk, the FEB Research Institute PhD Conference, the EuHEA Conference, the NBER Determinants of Mortality Conference, the Joint Health Workshop of CPHEB Groningen and ZEW Mannheim, the Rockwool Foundation Seminar in Copenhagen, and the Wenlan Forum at Zhongnan University of Economics and Law, and two anonymous referees for their helpful comments and feedback. The authors express their gratitude to the Lifelines Cohort Study and Biobank, the research centers providing data to Lifelines, and all the participants of the study. The Lifelines initiative has been made possible by subsidy from the Dutch Ministry of Health, Welfare and Sport, the Dutch Ministry of Economic Affairs, the University Medical Center Groningen (UMCG), the University of Groningen, and the Provinces in the North of the Netherlands (Drenthe, Friesland, Groningen). Ailun Shui acknowledges the support of the CSC scholarship program (Grant No. 202108500029) and is grateful for the resources provided by the International Max Planck Research School for Population, Health, and Data Science (IMPRS-PHDS). All funding was used for Ailun Shui’s daily expenses and academic visits during his PhD and was not directly allocated to the research of this paper. The other authors have no information to disclose. This paper uses restricted data from the Dutch Lifelines Cohort Study and Biobank. These data may be obtained from a third party and are not publicly available. Researchers can apply to use the Lifelines data used in this study. More information about how to request Lifelines data and the conditions of use can be found on their website (https://www.lifelines-biobank.com/researchers/working-with-us). The authors are willing to assist with providing guidance on accessing the data (Ailun Shui, a.shui{at}rug.nl.
Footnotes
Contributions were originally presented at the Determinants of Mortality conference hosted by the National Bureau of Economic Research with financial support from the National Institute on Aging (P30AG012810).
↵1. A biomarker-related risk can be viewed as an instance of physiological dysregulation, where the measured biological indicator departs from the clinically defined normal range.
↵2. Aging-related chronic diseases refer to long-term health conditions whose prevalence increases with age and reflects the cumulative physiological deterioration.
↵3. Specifically, two follow-up questionnaires (waves 1b and 1c) were conducted after wave 1a, and one follow-up (wave 2b) took place after wave 2a. Since 2024, the fourth wave of assessment is underway, with plans to include new participants, particularly from younger generations, in the Lifelines cohort. We do not include the data from wave 4a because it is still in process. For more information about Lifelines, please visit https://www.lifelines-biobank.com (accessed December 3, 2025).
↵4. Clinically defined thresholds refer to the biomarker values that define the boundaries between clinically normal and abnormal ranges.
↵5. BMI is an anthropometric measure that is widely regarded as a biomarker in public health and epidemiology. It is objectively measured and strongly associated with various health conditions. Consistent with prior literature, we therefore treat BMI as a biomarker in our analysis.
↵6. Compared to other cohort studies, the triglycerides seem to be lower in Lifelines in all percentiles. This is mainly because current criteria are largely based on the studies that were carried out in the 1970s (Balder et al. 2017).
↵7. In wave 1a, participants were asked, “Have you ever had a certain disease?” For all subsequent waves, participants were asked, “Did any of the health problems listed below begin since the last time you completed the Lifelines questionnaire?”
↵8. For participants aged 18 years and older, the questionnaires assess whether they have experienced any of the specified chronic diseases since their last participation in the Lifelines survey and associated assessments. Our identification of chronic diseases primarily relies on the list provided in the questionnaires.
↵9. Since our focus is on lifetime trajectories, we limit our analysis to aging-related chronic conditions. Specifically, we include only diseases with prevalence that increase with age, excluding chronic conditions primarily observed in childhood and predominantly caused by genetic factors. Additionally, we do not consider chronic diseases that are only available in limited waves of Lifelines.
↵10. While self-reported data on chronic diseases offer valuable health insights, they are subject to limitations, including nonclassical measurement errors and underdiagnosis (Danesh et al. 2024). These issues are particularly prevalent among lower-income or less-educated groups, potentially introducing bias into health assessments.
↵11. In the Netherlands, higher vocational education and university education correspond to levels 6, 7, and 8 of the International Standard Classification of Education (ISCED). Consequently, the low education group in our data set includes individuals with ISCED levels ranging from zero to five. The mandatory nature of most types of secondary education is the primary reason for dividing it into two categories.
↵12. We exclude individuals aged below 18, as most biomarkers in Lifelines are only available for participants aged 18 and above. Additionally, we exclude individuals older than 80 years because they are underrepresented in Lifelines.
↵13. Many biomarkers are influenced by short-term dietary intake. To ensure reliable results, fasting is required before blood sample collection or measurement. It is therefore imperative to exclude individuals who did not fast before blood collection from our sample.
↵14. We currently do not observe data dropout due to enrollment into a regular healthcare facility. However, Sijtsma et al. (2022) show that only a small proportion of participants withdraw in follow-ups.
↵15. The concept of allostatic load was introduced by McEwen and Stellar (1993) and does not directly measure stress itself.
↵16. The use of BMI for assessing clinical obesity has been a topic of ongoing discussion in the literature. Research suggests that BMI might misclassify or overestimate adiposity, potentially leading to inappropriate conclusions. In particular, as suggested by Rubino et al. (2025), BMI should be treated as a surrogate measure of health risk at the population level rather than a direct measure of individual health outcomes. Therefore, rather than treating BMI as a marker of chronic conditions, we use it as an indicator of health risks, aligning with previous research on the construction of allostatic load. To assess the role of BMI, we also constructed an alternative ALI that excludes BMI. Online Appendix B provides additional details on the ALI without BMI. Additionally, we use this alternative ALI in place of the original measure to conduct sensitivity analyses, the results of which are reported in Online Appendix Table B.1 and Figure B.1.
↵17. Some prior studies also employ biomarkers from other systems, including the immune system, the hypothalamic–pituitary–adrenal axis, the respiratory system, and the parasympathetic nervous system. However, biomarkers from the cardiovascular and metabolic systems are the most commonly used to construct the ALI (Johnson, Cavallaro, and Leon 2017).
↵18. The number of biomarker-related risks captures how many biomarkers exhibit values beyond the clinically defined normal range, that is, above or below the respective clinical thresholds.
↵19. One exception is HDL cholesterol, often called “good” cholesterol. A low HDL cholesterol level is considered high risk because HDL cholesterol helps remove excess cholesterol from the blood.
↵20. The count-based approach has also been applied in constructing similar health indicators, such as the frailty index (Hosseini, Kopecky, and Zhao 2022).
↵21. In contrast, Danesh et al. (2024) construct the CDI using a weighted approach based on prescription medication data from the Netherlands. The weights are derived from the observed relationship between chronic diseases and mortality risk in old age, allowing their CDI to reflect how the current chronic disease burden may translate into long-term health risks. Therefore, some caution is warranted when relating our findings to the setting of Danesh et al. (2024).
↵22. There is a clear distinction between the biomarkers we include and the aging-related chronic diseases we consider. While chronic disease indicates a diseased state in the body, biomarkers are biological measures that frequently but not perfectly correlate with an illness. For example, total cholesterol and triglyceride levels exceeding clinical thresholds do not immediately indicate a diagnosis of cardiovascular disease. However, high levels of these biomarkers increase the risk of stroke and heart attack, which offer insights into potential health risks.
↵23. The min–max scaling approach is a data normalization technique used to scale the values of a data set to a specific range, often from zero to one.
↵24. These trajectories are estimated using a pooled analysis of unbalanced panel data and may be confounded by factors such as cohort effects, health-related attrition, and medication interventions.
↵25. Specifically, the three-year mortality for 35,240 observations in wave 3a are missing because these participants took part in the third wave of Lifelines after September 2021.
↵26. In the Lifelines study, we have access to limited self-reported medication data. Unfortunately, the quality of this data is insufficient to test this assumption due to the categorization of medication information and the available sample size.
↵27. By contrast, we also present the educational CDI disparities by age and gender in Online Appendix Figure A.1. Among females, the disparity emerges as early as age 25 and gradually widens with age. For males, the gap becomes apparent only around age 40 and continues to increase steadily until approximately age 70. Our findings are related to those of Danesh et al. (2024), who report that the income-related CDI gap increases around age 25 for both females and males. It is important to note, however, that we employ different SES measures and use a different method and data to construct the CDI.
↵28. For young adults under the age of 25, we use the highest parental education in place of their educational attainment, which may partly mitigate the endogeneity of SES in this group.
↵29. HDL cholesterol is considered “good” cholesterol, so here we report the percentage of individuals with HDL cholesterol below the clinical threshold.
↵30. The ten-year age bins do not apply to all age ranges. Specifically, the 18–24 and 75–80 groups are defined separately due to the age distribution of our sample.
↵31. A similar Shapley–Owen decomposition approach was applied to CDI by Danesh et al. (2024), and our approach is in the spirit of that innovative approach. Apart from the difference between ALI and CDI, the comparison is somewhat complicated due to implementation details, such as differences in the availability of decomposers in the various data sets and differences in age intervals.
↵32. Given the limited set of explanatory variables, the reported contributions should be regarded as conditional lower-bound estimates of the variation attributable to these factors. Additional unobserved determinants, including income, environmental exposures, and dietary behaviors, are likely to explain further variation.
- Received February 2025.
- Accepted November 2025.
This open access article is distributed under the terms of the CC-BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) and is freely available online at: https://jhr.uwpress.org.
