Explaining inequalities in quality of life: a longitudinal study of health disparities in Norway Authors: Jan Abel Olsen, Gang Chen, Ola Løvsletten Published: 2026 ABSTRACT PurposeStudies of health inequalities often examine the association between a single health indicator and a single socioeconomic indicator. Recent research has increasingly explored social disparities in self-reported quality of life. However, most evidences are based on cross-sectional survey and the role of health behaviours is frequently overlooked in these analyses. This study aims to integrate broader sets of socioeconomic and behavioural variables to better explain health disparities that emerge in later adulthood.MethodsUsing longitudinal data from the Norwegian population-based Tromsø Study, we followed 11,313 adults aged 25–54 at baseline over 21 years. A series of self-reported outcomes were studied, particularly EQ-5D-5L values, EQ-VAS scores, and Self-Rated Health. Predictors included three socioeconomic indicators (childhood living standard, own and spouse’s educational attainment) and three health behaviours (smoking, physical activity, obesity). Shapley value decompositions were used to quantify each predictor’s contribution to explained variance.ResultsFor all three health outcomes, health behaviours accounted for the majority of the explained variance compared to socioeconomic factors: 80% for EQ-5D-5L, 75% for EQ VAS, and 67% for Self-Rated Health. When adjusting for differences in baseline health, the relative importance of health behaviours increased slightly.ConclusionThis study underscores the critical role of health behaviours in explaining inequalities in quality of life, providing new insights into the mechanisms driving these disparities. While preventing childhood deprivation would have significant impact on adult health, potential health benefits of promoting healthy behaviours are substantial. Both types of policy interventions could thereby contribute to reducing health inequalities.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s11136-026-04311-y. Introduction Health inequalities persist across all countries, with evidence suggesting they are not only enduring but also widening in rich nations like Norway and the United States [1]. Studies of health inequalities often examine the association between a single socioeconomic indicator, such as education or income, and a single health indicator, such as life expectancy or morbidity. With the increasing availability of population health survey data, research has expanded to explore social disparities in quality of life. A recent review identified 38 studies examining the relationship between ‘social determinants of health’ and health-related quality of life (HRQoL) among chronic disease populations, showing significant associations between socioeconomic indicators, and HRQoL, such as EQ-5D or SF-36 [2]. Notably, the vast number of studies focusing on social inequalities in health often overlook the role of health behaviours, often referred to as ‘lifestyle factors’ (e.g., smoking, physical activity). This omission is surprising given the well-documented causal links between various health behaviours and health outcomes [3]. The literature explaining health inequalities, or disparities, tends to emphasize either of the two key sets of health determinants: the social environment or individual health behaviours. In a recent study of Chinese rural adults, Liao et al. makes an interesting contribution by including five lifestyle factors in a healthy lifestyle score, showing consistent positive associations between a healthy lifestyle and EQ-5D [4]. However, the study did not include variables on socioeconomic position in a way that allowed for a comparison of their relative importance. In this paper, we aim to bridge this gap by integrating broader sets of socioeconomic and behavioural variables to better explain disparities in health and wellbeing. Our study leverages unique data to provide a more comprehensive measure of socioeconomic position (SEP). Specifically, we include an indicator of past SEP, that is, respondents’ childhood living standards. This reflects the extensive evidence linking childhood socioeconomic circumstances to adult health outcomes (e.g., Case, Fertig, and Paxson [5]). We also include respondents’ educational attainment, the most widely used SEP-proxy in this field, as well as the educational attainment of their spouses. This inclusion acknowledges the growing trends in assortative mating and educational homogamy [6, 7]—where individuals with similar educational levels are more likely to pair. Such patterns of social stratification in resources, habits, and behaviours [8] amplifies social disparities in health [9–12] and subjective wellbeing [13]. While systematic differences in health-related lifestyles are evident across socioeconomic strata, substantial individual variation persists within each group. Given the critical role of "lifestyle diseases" on overall public health, we include three key health behaviour indicators: smoking, physical activity and obesity. These are the most widely included variables in a Healthy Lifestyle Index (see e.g. Loef and Walach [3]), and they were also included in the healthy lifestyle score used by Liao et al. [4]. In the current paper, the three variables are collectively referred to as health behaviours (HB). Longitudinal studies examining the impact of health behaviours on health and wellbeing are relatively scarce and often limited to short timeframes. For example, an Australian study spanning eight years found that modifiable behaviours, such as physical activity, influence self-rated health over time, though the relationship varies by behaviour [14]. Similarly, a Swiss study conducted over two years during the COVID-19 pandemic highlighted significant associations between HB and self-rated health [15]. Beyond health outcomes, a Finnish study over nine years demonstrated positive effects of changes in health behaviours on subjective wellbeing [16]. Our primary research question is: How do three socioeconomic proxies and three health behaviours contribute to explaining inequalities in quality of life? Specifically, we begin with a simple reference model that examines the association between the most widely used socioeconomic indicator—educational attainment—and health outcomes. We then incorporate two additional, unique socioeconomic indicators—childhood living standard and spouse education—to assess their added contributions. Next, we expand the model to include three health behaviours: smoking, physical activity and BMI, to evaluate the relative importance of socioeconomic factors vs. health behaviours in explaining health disparities. Lastly, we adjust for differences in baseline health reported in early adulthood, allowing us to investigate into how these variables collectively explain health disparities that emerge in later adulthood—given similar level of baseline health. Research on social inequalities in health often relies on register data, which typically lack detailed information on health behaviours (HB). A key contribution of this study is to isolate the influence of socioeconomic variables and examine the role of HB—and changes in HB—over a 21-year period. By adjusting for baseline health (measured at age 25 or older), our analysis uniquely investigates how HB contribute to health disparities in later adulthood. Methods Data We utilized data from a prospective cohort study of the population residing in the largest municipality in Northern Norway. The Tromsø Study, which has completed seven surveys to date, is broadly representative of the Norwegian adult population, with individuals holding a university degree being slightly overrepresented [17, 18]. The current paper focuses on individuals who participated in both the fourth survey (1994/95, referred to here as t0 for year 0, or baseline) and the seventh survey (2015/16, referred to as t21 indicating that follow-up was 21 years after baseline). All inhabitants of Tromsø Municipality aged ≥ 25 years and ≥ 40 years, respectively, were invited to these surveys. Attendance rates were high, with 77% in the fourth survey and 65% in the seventh. The study sample includes individuals aged 25–54 years at t0, who were aged 46–75 years at t21. Of the 19,181 participants in this age range at t0, 14,179 were invited to t21, and 80% participated, resulting in a final study sample size of n = 11,313 individuals. The study was approved by the Regional Committee for Medical and Health Research Ethics (ID, 2016/607). All participants provided written informed consent. Variables Outcome variables Health-related quality of life (HRQoL) was measured by the EQ-5D-5L index value and the EQ VAS (Visual Analogue Scale) [19, 20]. The EQ-5D-5L is a generic preference-based descriptive system that includes five dimensions: mobility, self-care, usual activities, pain/discomfort, and anxiety/depression, each with five severity levels (no problems, slight problems, moderate problems, severe problems, or unable to/extreme problems) [21]. We applied the recently published Norwegian value set [22]. The EQ-5D-5L is anchored on a [0 (death) to 1 (full health)] scale. The EQ-VAS score is based on respondents’ direct valuations of their overall health on a visual analogue scale that ranges from 0 (worst imaginable health) to 100 (best imaginable health). We rescaled the EQ-VAS scores to [0–1]. In addition to these two primary outcomes, we consider two additional outcomes for comparability. Self-rated health is based on a widely used single-item question: How do you in general consider your own health to be?, with five response levels. We dichotomised by collapsing those that reported Good, Very good = 1, for Good health vs. a reference category that reported Very bad, Bad, Neither good nor bad = 0, for Not-good health. Furthermore, we consider a measure of subjective wellbeing, assessed by the Satisfaction With Life Scale [23]. We use the first three items: In most ways my life is close to my ideal; The conditions of my life are excellent; I am satisfied with my life. The response options ranged from 1 (strongly disagree) to 7 (strongly agree). The total sum of scores is linearly transformed onto a [0–1] scale. The omitted two items are sensitive to age, as they implicate the experience of life satisfaction in the past [24, 25], and they have poorer psychometric properties [26, 27]. All four outcome variables were measured at follow-up only. Explanatory variables The first set of predictors pertains to socioeconomic position (SEP): Childhood living standard, was assessed using the question: How was your family’s financial situation during childhood?, with response options: very good, good, difficult, very difficult. Due to few respondents in the extreme categories (less than 2% very difficult; less than 6% very good), they were merged with the respective difficult and good childhood living standards. Similar indicators have been used in epidemiological studies to proxy childhood socioeconomic circumstances [28, 29]). Educational attainments, own and spouse, were reported along four levels: 1) primary (including lower secondary), 2) secondary (including vocational), 3) tertiary low (less than 4 years of university study), and 4) tertiary high (4 years or more of university study). The spouse education variable included not having a spouse as the reference [8, 10]. Thus, this reference category includes all participants who, for whatever reason, do not have a current partner, i.e., they might be single, widowed, separated, or divorced. We then dichotomized the spouse education levels into Non-tertiary (levels 1 & 2) and Tertiary (levels 3 & 4). Spouse education was reported at follow-up. However, considering the average marriage age of 32 for women and 35 for men in Norway in 2005, it is reasonable to assume that this variable has been relevant for a significant period, though not necessarily throughout the entire 21-year interval. The second set of predictors encompasses key health behaviours commonly included in a ‘healthy lifestyle index’ [30]: i) smoking status, ii) physical activity, iii) healthy weight status, categorized based on body mass index (BMI, calculated as weight/height2 in kg/m2). Height and weight were objectively measured rather than self-reported. BMI is not a 'health behaviour' but more accurately conceptualized as an outcome of behaviours associated with diet and physical activity. However, due to the challenges inherent in obtaining accurate self-reported dietary intake data in surveys, obesity, as measured by BMI is frequently employed as a proxy for assessing health risk. Smoking status was categorized into three groups: never-smokers (reported “no” to daily smoking at both surveys), quitters (daily smokers at t0, but not at t21), persistent smokers (reported “yes” to daily smoking at both t0 and t21, including a small number of new smokers who had started between t0 and t21). Physical activity (PA) was measured differently at t0, and t21. At t21, respondents were grouped into three categories based on the World Health Organization (WHO) guidelines recommending at least 150 min of moderate-intensity PA per week [31]: Active (> 150 min/week), Moderate (60–150 min/week), Inactive (< 60 min/week). At t0, PA was measured using two separate questions on light and hard PA, each reported as hours per week across four levels: (i) none, (ii) < 1 h, (iii) 1–2 h, and (iv) ≥ 3 h. Participants were categorised as Active if they reported ≥ 3 h/week of hard PA, or 1–2 h/week of hard PA combined with at least 1–2 h/week light PA. The Inactive group included those reporting no hard PA and < 3 h/week of light PA, or < 1 h/week of hard PA with no light PA. Remaining combinations were categorized as Moderate. In the regression analyses, we used categories at t21 for BMI and PA, and further categorized the patterns on HB changes over 21 years into whether it was persistent, i.e. the same category as measured in t0. This approach allowed us to examine whether past (un)healthy behaviour had lasting health impacts beyond what could be explained by current levels. For BMI, a value < 25 at t21 was split into Persistent normal BMI or Newly normal BMI, whilst a BMI > 35 at t21 was split into Increased to 35 + or Persistent obese, defined as 30 + at t0 and 35 + at t21. The same logic was applied to the PA categories, in which we split the Active category into Persistent active and Newly active. Conversely, the Inactive category at t21 was split into Persistent inactive and Newly inactive. This categorization enabled us to explore the long-term effects of sustained or changed health behaviours. To examine the lasting effects of health behaviours and changes over the 21-year period, we adjusted for differences in baseline health. This adjustment ensures that any observed effects of health behaviours are not confounded by pre-existing differences in health status. While the baseline data only included self-rated health (SRH) as a relevant measure, the EQ VAS in essence records the patient's self-rated health on a vertical visual analogue scale; so they are intended to measure the same underlying concept of health. Derkach et al. [32] reported a Spearman correlation coefficient of -0.67 between SRH and EQ-VAS, indicating a strong relationship between the two measures. Additionally, as shown in Figure A1 there is a strong association between SRH and EQ VAS at follow-up, further supporting the use of SRH at t0 as a reliable proxy for baseline health. Statistical analyses Figure 1 illustrates our analytical framework where the solid arrows are estimated in regressions. The dashed arrows have been documented elsewhere (see e.g. [8, 10].), while the dotted arrows follow the same reasoning as the solid ones, except that we have no information on spouse education at baseline. Arrow denoted ⓪ represents a simple reference case model (Model 0), on the link between own educational attainment, and health outcomes. Arrows ① link childhood living standard and spouse educational attainments to health outcomes (Model I). Arrow ② links the additional contribution of health behaviours (Model II). Finally, arrow ③ links variations in baseline health, included in the full Model III. These models were run separately to explain each outcomes. Analytical framework All regressions were adjusted for sex and age. Results were presented as beta-coefficients for continuous dependent variables (except for the binary outcome of self-rated health, where odds ratios were used). To assist the interpretation of key predictors in explaining the variance in health and wellbeing, Shapley decompositions were applied. Shapley values attribute the explained variance in regressions to individual predictors, providing insights into their relative contributions to the outcome. All statistical analyses were run in R (version 4.3.2) [33]. Furthermore, we estimated the relative contributions of health behaviours to the education gradient in EQ-5D-5L and EQ-VAS, by comparing the magnitude of the education coefficients in Model-I and Model-II, using a simple formula consistent with previous studies [34, 35]: Contribution of HB (%) = 100 * (βModel-I—βModel-II)/βModel-I, where β refers to the estimated β coefficient of the highest educational attainment (i.e. Tertiary high). Heterogeneities and robustness Heterogeneities were examined using the two primary outcome measures: EQ-5D-5L and EQ VAS. Recognizing the significance of baseline health, we ran Model-II separately for two sub-groups: individuals who reported Good or Very good health at baseline (t0). Additionally, we explored heterogeneities by age groups and sex. As a robustness check, we account for the possibility that individuals in poor health may have limited capacity to engage in regular physical activity. Including such individuals could therefore overestimate the association between physical activity and health, as their poor health may explain their low physical activity levels. To address this, we exclude respondents who reported level 4 or level 5 (i.e. the existence of more severe impairment) on at least one of the first four EQ-5D-5L dimensions. This allows us to focus the analysis on the association between physical activity and health, specifically among individuals who are capable of being physically active. Results Table 1 presents respondent characteristics alongside their unadjusted mean values for EQ-5D-5L, EQ VAS, and SWLS-3, as well as percentage reporting Good/Very good self-rated health. The results reveal a consistent pattern of increasing health outcomes along each of the three socioeconomic variables (childhood living standard, levels of own- and spouse educational attainment), and each of the three healthy behaviours (smoking, physical activity, BMI-categories). Table A1 provides corresponding mean values stratified by the physical activity and BMI categories used in the regression analyses, i.e. when further accounting for past health behaviours.Table 1Respondent characteristics with unadjusted mean health and wellbeing at follow-upCharacteristicN = 11,3131EQ-5D-5L2EQ VAS3SRH:Good/Very good4SWLS-32SexWomen6,047 (53%)0.8820.76368%0.718Men5,266 (47%)0.9060.77069%0.723Age_t025–344,311 (38%)0.8920.76769%0.72235–443,565 (32%)0.8890.77172%0.70345–543,437 (30%)0.8980.75964%0.737Childhood living standard5Good8,349 (74%)0.8980.77471%0.728Difficult2,931 (26%)0.8800.74260%0.699Own educationPrimary3,043 (27%)0.8820.73557%0.720Secondary3,407 (31%)0.8910.76366%0.717Tertiary low2,092 (19%)0.8980.77974%0.712Tertiary high2,618 (23%)0.9050.79680%0.732Spouse educationNo spouse2,130 (19%)0.8790.74963%0.660Non-tertiary5,366 (47%)0.8930.75865%0.728Tertiary3,817 (34%)0.9010.78776%0.740Smoking statusNever smoker7,000 (62%)0.9010.78273%0.737Quitter2,563 (23%)0.8850.74762%0.705New smoker162 (1.4%)0.8890.73766%0.679Smoker1,483 (13%)0.8720.73155%0.674BMI categories_t06 < 256,460 (58%)0.8980.78172%0.721[25,30)3,767 (34%)0.8920.75565%0.725[30,35)701 (6.3%)0.8720.70851%0.700 ≥ 35121 (1.1%)0.8310.68243%0.650BMI categories_t216 < 253,495 (31%)0.9030.79376%0.725[25,30)5,036 (45%)0.8980.77370%0.729[30,35)2,030 (18%)0.8850.73559%0.716 ≥ 35655 (5.8%)0.8480.68242%0.656Physical activity_t07Active3,753 (33%)0.9030.78777%0.735Moderate4,154 (37%)0.8920.76468%0.715Inactive3,302 (29%)0.8840.74559%0.710Physical activity_t217Active4,195 (39%)0.9040.79678%0.735Moderate3,260 (31%)0.9000.77170%0.724Inactive3,205 (30%)0.8820.73658%0.707Self-Rated Health_t0Very good2,222 (20%)0.9300.84189%0.781Good6,978 (62%)0.8980.76970%0.719Not good82,096 (19%)0.8360.67640%0.657The Tromsø Study 1994–95 (t0) and 2015–16 (t21) was used. 1n (%); 2based on Norwegian value set; 3linearly transformed to the interval [0—1], SWLS-3 is the first three items of the Satisfaction with Life Scale; 4Includes the collapsed two top levels on the five-level Self-Rated Health scale; 5This question was asked in both surveys. We used responses at t0, filling in with answers from t21 for those who did not respond at t0; 6BMI = Body Mass Index; the few underweight BMI < 18.5 (< 1% in t0 and < 0.5% in t21) were collapsed into category < 25; 7Different questions at t0 and t21; 8Includes the collapsed two bottom levels (‘not good’ and ‘bad’) on the four-level SRH-scale used at t0 Table 1Respondent characteristics with unadjusted mean health and wellbeing at follow-upCharacteristicN = 11,3131EQ-5D-5L2EQ VAS3SRH:Good/Very good4SWLS-32SexWomen6,047 (53%)0.8820.76368%0.718Men5,266 (47%)0.9060.77069%0.723Age_t025–344,311 (38%)0.8920.76769%0.72235–443,565 (32%)0.8890.77172%0.70345–543,437 (30%)0.8980.75964%0.737Childhood living standard5Good8,349 (74%)0.8980.77471%0.728Difficult2,931 (26%)0.8800.74260%0.699Own educationPrimary3,043 (27%)0.8820.73557%0.720Secondary3,407 (31%)0.8910.76366%0.717Tertiary low2,092 (19%)0.8980.77974%0.712Tertiary high2,618 (23%)0.9050.79680%0.732Spouse educationNo spouse2,130 (19%)0.8790.74963%0.660Non-tertiary5,366 (47%)0.8930.75865%0.728Tertiary3,817 (34%)0.9010.78776%0.740Smoking statusNever smoker7,000 (62%)0.9010.78273%0.737Quitter2,563 (23%)0.8850.74762%0.705New smoker162 (1.4%)0.8890.73766%0.679Smoker1,483 (13%)0.8720.73155%0.674BMI categories_t06 < 256,460 (58%)0.8980.78172%0.721[25,30)3,767 (34%)0.8920.75565%0.725[30,35)701 (6.3%)0.8720.70851%0.700 ≥ 35121 (1.1%)0.8310.68243%0.650BMI categories_t216 < 253,495 (31%)0.9030.79376%0.725[25,30)5,036 (45%)0.8980.77370%0.729[30,35)2,030 (18%)0.8850.73559%0.716 ≥ 35655 (5.8%)0.8480.68242%0.656Physical activity_t07Active3,753 (33%)0.9030.78777%0.735Moderate4,154 (37%)0.8920.76468%0.715Inactive3,302 (29%)0.8840.74559%0.710Physical activity_t217Active4,195 (39%)0.9040.79678%0.735Moderate3,260 (31%)0.9000.77170%0.724Inactive3,205 (30%)0.8820.73658%0.707Self-Rated Health_t0Very good2,222 (20%)0.9300.84189%0.781Good6,978 (62%)0.8980.76970%0.719Not good82,096 (19%)0.8360.67640%0.657The Tromsø Study 1994–95 (t0) and 2015–16 (t21) was used. 1n (%); 2based on Norwegian value set; 3linearly transformed to the interval [0—1], SWLS-3 is the first three items of the Satisfaction with Life Scale; 4Includes the collapsed two top levels on the five-level Self-Rated Health scale; 5This question was asked in both surveys. We used responses at t0, filling in with answers from t21 for those who did not respond at t0; 6BMI = Body Mass Index; the few underweight BMI < 18.5 (< 1% in t0 and < 0.5% in t21) were collapsed into category < 25; 7Different questions at t0 and t21; 8Includes the collapsed two bottom levels (‘not good’ and ‘bad’) on the four-level SRH-scale used at t0 Respondent characteristics with unadjusted mean health and wellbeing at follow-up The Tromsø Study 1994–95 (t0) and 2015–16 (t21) was used. 1n (%); 2based on Norwegian value set; 3linearly transformed to the interval [0—1], SWLS-3 is the first three items of the Satisfaction with Life Scale; 4Includes the collapsed two top levels on the five-level Self-Rated Health scale; 5This question was asked in both surveys. We used responses at t0, filling in with answers from t21 for those who did not respond at t0; 6BMI = Body Mass Index; the few underweight BMI < 18.5 (< 1% in t0 and < 0.5% in t21) were collapsed into category < 25; 7Different questions at t0 and t21; 8Includes the collapsed two bottom levels (‘not good’ and ‘bad’) on the four-level SRH-scale used at t0 The regression models for the three health outcome measures, presented in Tables 2 and 3, and in the appendix tables A2 and A3, reveal some general patterns: Results from the reference case Model-0 confirm a well-established finding in the literature of consistent positive association between educational attainment and health. Model-I includes the two additional SEP-indicators: childhood living standard and spouse education (and implicitly civil status). A consistent and significant gradient can be observed across all three SEP-variables. The coefficients for spouse education suggest that the spillover effect primarily operates through having a spouse rather than their educational attainment—this is particularly evident for EQ-5D-5L. Notably, the coefficients for own education in Model-I show minimal attenuation compared to Model-0, e.g. in Table 2, beta for Tertiary high education is reduced from 0.021 to 0.019, which suggests that the two additional socioeconomic predictors function more as complementary contributors rather than substitutes in explaining health disparities.Table 2Linear regression results for EQ-5D-5L at t21. The Tromsø Study 1994–95 (t0) and 2015–16 (t21)Model 0IIIIIICharacteristicBetaBetaBetaBeta(Intercept)0.924***0.919***0.950***0.950***Own education (ref. Primary)Secondary0.009***0.008**0.0030.001Tertiary low0.014***0.013***0.003− 0.001Tertiary high0.021***0.019***0.008**0.003Childhood living standard (ref. Good)Difficult− 0.016***− 0.013***− 0.009***Spouse education (ref. No Spouse)Non tertiary0.011***0.009***0.008***Tertiary0.013***0.008**0.006*Smoking status (ref. Never smoker)Quitter− 0.009***− 0.007**Smoker− 0.020***− 0.016***BMI (ref. Persistent normal)New normal− 0.010− 0.008Overweight t21− 0.007***− 0.006**30–35 at t21− 0.018***− 0.015***Increase to 35 + − 0.039***− 0.035***Persistent obese, > 35 at t21− 0.057***− 0.053***Physical activity (ref. Persistent active)Newly active− 0.0020.002Moderate t21− 0.0020.001Newly inactive− 0.015***− 0.012***Persistent inactive− 0.017***− 0.010**S-Rated Healthelf_t0 (ref. Good)Not good− 0.043***Very good0.019***R20.0140.0220.0530.101β-coefficients are reported. *p < 0.05; **p < 0.01; ***p < 0.001. All models were adjusted for age and sexModel-0 is a reference case; Model-I includes the three socioeconomic predictors; Model-II further includes three health behaviours; Model-III further adjusts for differences in self-rated health at baselineTable 3Linear regression results for EQ VAS at t21. The Tromsø Study 1994–95 (t0) and 2015–16 (t21)Model 0IIIIIICharacteristicBetaBetaBetaBeta(Intercept)0.735***0.729***0.800***0.793***Own education (ref. Primary)Secondary0.027***0.025***0.019***0.014***Tertiary low0.043***0.039***0.024***0.015***Tertiary high0.061***0.055***0.034***0.024***Childhood living standard (ref. Good)Difficult− 0.028***− 0.024***− 0.016***Spouse education (ref. No Spouse)Non tertiary0.015***0.014**0.012**Tertiary0.025***0.017***0.012**Smoking status (ref. Never smoker)Quitter− 0.020***− 0.014***Smoker− 0.036***− 0.027***BMI (ref. Persistent normal)New normal− 0.019*− 0.014Overweight t21− 0.018***− 0.014***30–35 at t21− 0.048***− 0.042***Increase to 35 + − 0.084***− 0.075***Persistent obese, > 35 at t21− 0.092***− 0.084***Physical activity (ref. Persistent active)Newly active− 0.011*− 0.001Moderate t21− 0.024***− 0.016***Newly inactive− 0.044***− 0.035***Persistent inactive− 0.051***− 0.035***Self-Rated Health_t0 (ref. Good)Not good− 0.081***Very good0.060***R20.0200.0280.0700.138β-coefficients are reported. *p < 0.05; **p < 0.01; ***p < 0.001. All models were adjusted for age and sexModel-0 is a reference case; Model-I includes the three socioeconomic predictors; Model-II further includes three health behaviours; Model-III further adjusts for differences in self-rated health at baseline Table 2Linear regression results for EQ-5D-5L at t21. The Tromsø Study 1994–95 (t0) and 2015–16 (t21)Model 0IIIIIICharacteristicBetaBetaBetaBeta(Intercept)0.924***0.919***0.950***0.950***Own education (ref. Primary)Secondary0.009***0.008**0.0030.001Tertiary low0.014***0.013***0.003− 0.001Tertiary high0.021***0.019***0.008**0.003Childhood living standard (ref. Good)Difficult− 0.016***− 0.013***− 0.009***Spouse education (ref. No Spouse)Non tertiary0.011***0.009***0.008***Tertiary0.013***0.008**0.006*Smoking status (ref. Never smoker)Quitter− 0.009***− 0.007**Smoker− 0.020***− 0.016***BMI (ref. Persistent normal)New normal− 0.010− 0.008Overweight t21− 0.007***− 0.006**30–35 at t21− 0.018***− 0.015***Increase to 35 + − 0.039***− 0.035***Persistent obese, > 35 at t21− 0.057***− 0.053***Physical activity (ref. Persistent active)Newly active− 0.0020.002Moderate t21− 0.0020.001Newly inactive− 0.015***− 0.012***Persistent inactive− 0.017***− 0.010**S-Rated Healthelf_t0 (ref. Good)Not good− 0.043***Very good0.019***R20.0140.0220.0530.101β-coefficients are reported. *p < 0.05; **p < 0.01; ***p < 0.001. All models were adjusted for age and sexModel-0 is a reference case; Model-I includes the three socioeconomic predictors; Model-II further includes three health behaviours; Model-III further adjusts for differences in self-rated health at baseline Linear regression results for EQ-5D-5L at t21. The Tromsø Study 1994–95 (t0) and 2015–16 (t21) β-coefficients are reported. *p < 0.05; **p < 0.01; ***p < 0.001. All models were adjusted for age and sex Model-0 is a reference case; Model-I includes the three socioeconomic predictors; Model-II further includes three health behaviours; Model-III further adjusts for differences in self-rated health at baseline Table 3Linear regression results for EQ VAS at t21. The Tromsø Study 1994–95 (t0) and 2015–16 (t21)Model 0IIIIIICharacteristicBetaBetaBetaBeta(Intercept)0.735***0.729***0.800***0.793***Own education (ref. Primary)Secondary0.027***0.025***0.019***0.014***Tertiary low0.043***0.039***0.024***0.015***Tertiary high0.061***0.055***0.034***0.024***Childhood living standard (ref. Good)Difficult− 0.028***− 0.024***− 0.016***Spouse education (ref. No Spouse)Non tertiary0.015***0.014**0.012**Tertiary0.025***0.017***0.012**Smoking status (ref. Never smoker)Quitter− 0.020***− 0.014***Smoker− 0.036***− 0.027***BMI (ref. Persistent normal)New normal− 0.019*− 0.014Overweight t21− 0.018***− 0.014***30–35 at t21− 0.048***− 0.042***Increase to 35 + − 0.084***− 0.075***Persistent obese, > 35 at t21− 0.092***− 0.084***Physical activity (ref. Persistent active)Newly active− 0.011*− 0.001Moderate t21− 0.024***− 0.016***Newly inactive− 0.044***− 0.035***Persistent inactive− 0.051***− 0.035***Self-Rated Health_t0 (ref. Good)Not good− 0.081***Very good0.060***R20.0200.0280.0700.138β-coefficients are reported. *p < 0.05; **p < 0.01; ***p < 0.001. All models were adjusted for age and sexModel-0 is a reference case; Model-I includes the three socioeconomic predictors; Model-II further includes three health behaviours; Model-III further adjusts for differences in self-rated health at baseline Linear regression results for EQ VAS at t21. The Tromsø Study 1994–95 (t0) and 2015–16 (t21) β-coefficients are reported. *p < 0.05; **p < 0.01; ***p < 0.001. All models were adjusted for age and sex Model-0 is a reference case; Model-I includes the three socioeconomic predictors; Model-II further includes three health behaviours; Model-III further adjusts for differences in self-rated health at baseline Model-II extends the analysis by incorporating health behaviours, revealing strong and consistent effects across all three health outcomes. In addition to current health behaviours, reported at follow-up (t21), past behaviours matter as well: previous smokers have worse health than never-smokers (Table 2: beta = -0.009 for quitters), and individuals who are persistent obese or persistent inactive have worse health than those who are newly obese or newly inactive. While the inclusion of health behaviours attenuates the effects of education, the influence of childhood living standard remains substantial and persistent, cf. the slight reductions in coefficients for childhood living standard between Model-I and Model-II (Table 2: beta reduced by 19% from − 0.016 to − 0.013), as compared to the much larger reductions in the coefficients for own education (Table 2: beta reduced by 58% from 0.019 to 0.008). The relative contribution of health behaviours to the education gradient, as measured by the relative differences in coefficients between Model-II and Model-I for the highest education level, is 58% in EQ-5D-5L (calculated as (0.019–0.008)/0.019, Table 2) and 37% in EQ-VAS. Interestingly, when examining the relative contribution of health behaviours to the spouse education gradient, it is 38% in EQ-5D-5L, and 29% in EQ VAS. In Model-III, we adjust for differences in self-reported health at baseline, thereby explaining disparities in health that have mainly emerged during the past 21-year period. Remarkably, the beta coefficients for health behaviours remain largely unchanged compared to Model-II. Furthermore, and somewhat unexpectedly, childhood living standards continue to exert a significant influence on health later in life, despite the expectation that their effects would be captured through baseline health. When analysing the impacts of the same sets of predictors on subjective wellbeing, Table A3 reveals a somewhat different pattern. Here, the largest relative impacts are attributed to having a partner (rather than their educational attainment) and smoking status. Figure 2 shows the Shapley value decomposition, illustrating the relative importance of the main sets of predictors to the explained variances in health and wellbeing. Consistent patterns emerge across the three health outcomes: in Model-II, the three health behaviours—smoking, physical activity and healthy weight categories–– collectively account for the majority of the explained variance: 80% of EQ-5D-5L; 75% of EQ VAS; 66% of self-rated health. Next, the columns representing results from Model-III reveal that over half of the explained variance in health at follow-up is attributable to differences in baseline health. However, when the Shapley percentage for baseline health is excluded, the relative importance of SEP vs. HB for the explained health variations remains consistent (see Figure A2 in supplementary material).Fig. 2Shapley value decomposition of main sets of predictors to the explained variances in health and wellbeing. The Tromsø Study 1994–2016 Shapley value decomposition of main sets of predictors to the explained variances in health and wellbeing. The Tromsø Study 1994–2016 Heterogeneities and robustness Recognizing the significance of baseline health, we ran Model-II separately for two sub-groups: the 62% of respondents who reported Good health, and the 20% who had reported Very good health at baseline (t0) (see Table A4 and Figure A3). When in Good health at baseline, the share of the explained variance in health 21 years later, attributable to health behaviours, was 88% for EQ-5D-5L and 86% for EQ VAS. When in Very good health at baseline, these shares were somewhat lower. Heterogeneities by age groups and sex (see Tables A5 and A6) indicate that childhood living standard and health behaviours remain stable and significant predictors across all age groups. However, having a spouse is relatively more important in younger age groups. Regarding sex differences, Table A6 suggests that childhood living standard is relatively more influential for women, whereas having a spouse plays a more significant role in men’s health. Own education and health behaviours appear to have similar effects in both sexes. Acknowledging that individuals in poor health status may have limited ability to engage in regular physical activity, we removed those 514 individuals who reported level 4 or level 5 on at least one of the first four EQ-5D-5L dimensions (Table A7). When comparing Table A7 with Model-II in Tables 2 and 3, respectively, the general patterns remain consistent, albeit with attenuated coefficients as expected. Discussion In the extensive body of literature on health inequalities, individual educational attainment is the most used proxy for socioeconomic position (SEP). In this study, we expand the conceptualization of SEP by incorporating spouse’s educational attainment and past SEP as indicated by childhood living standard. Furthermore, we leverage unique longitudinal data on smoking, physical activity and BMI collected over a 21-year interval, enabling us to examine the long-term influence of changes in health behaviours on health outcomes. Moreover, by adjusting for baseline self-rated health, we also assess the relative importance of modifiable lifestyle factors compared to stable socioeconomic indicators in explaining health disparities that emerge in later adulthood. Our results suggest that health behaviours account for the majority of the explained variance in health disparities. The stepwise adjustments in our regression models revealed several notable findings: First, an important proxy for early life circumstances, childhood living standard, showed a consistent and lasting impact, even after adjusting for health behaviours. Surprisingly, this association remained significant, albeit attenuated, after adjusting for baseline health. Next, own educational attainment exhibited a significant and consistent gradient across all three measures of health, while the educational attainment of the spouse was generally less influential. However, the presence of a spouse, regardless of their educational level, consistently showed significant positive associations with own health. Third, when the three health behaviours were included in the regressions, the coefficients for educational attainments were attenuated, indicating that a substantial proportion of the education-health gradient can be attributed to systematic differences in health behaviours. Among the three HB measures, obesity emerged as the strongest predictor, particularly when obesity was persistent over time. Fourth, when baseline self-rated health was included, Shapley value decompositions revealed that most of the explained variance in health at follow-up was explained by baseline health. Our findings align with prior research demonstrating the persistent influence of socioeconomic position (SEP) on health outcomes (e.g., [1]). Consistent with the literature, we observed a significant positive association between educational attainment and better health. However, our study extends this understanding by showing that childhood living standards continue to exert a significant influence on health outcomes in late adulthood even after accounting for adult health behaviours and baseline health. This finding corroborates evidence that early-life conditions have long-lasting effects on health trajectories (e.g., [5]). The results underscore the importance of addressing health behaviours as modifiable factors that contribute significantly to health disparities [36, 37]. Notably, the strong negative association between persistent obesity and physical inactivity highlights the need for public health interventions targeting these behaviours. However, the enduring impact of childhood living standards suggests that interventions aimed at improving early-life conditions may yield long-term benefits for reducing health inequalities [38]. This finding is particularly relevant in the context of policy discussions on the social determinants of health, emphasizing the need for a dual focus on early-life interventions and adult health behaviour modifications. The observed attenuation of the education-health gradient after adjusting for health behaviours supports theoretical frameworks that posit health behaviours as mediators of the relationship between SEP and health [35]. However, the continued significance of childhood living standards even after accounting for baseline health challenges the assumption that early-life SEP effects are fully mediated by adult health behaviours or baseline health. This finding suggests that early-life SEP may influence health through additional pathways, such as biological embedding or cumulative stress, which warrants further investigation. There are several notable strengths of this study. First, it is based on unique longitudinal data spanning 21 years, cf. what Kangas et al. (2025) concluded after their review of the 38 studies on social determinant of HRQoL: “future research should prioritize longitudinal studies”. Second, our study includes three socioeconomic measures and three measures of health behaviours, alongside three alternative measures of health outcomes and one measure of subjective wellbeing. Third, by adjusting for health differences at baseline, reported in early adulthood (aged 25–54), we were able to trace how the relatively stable socioeconomic variables, together with modifiable health behaviours, contribute to explaining the emerging health inequalities in later adulthood. We acknowledge some limitations. First, we cannot ignore the potential influence of reciprocal relationship between the three socioeconomic indicators and health. Own educational achievement might be adversely affected by poor health in early life. Regarding the variable childhood living standard, in the Norwegian institutional context of a generous welfare state, excess costs associated with children’s ill health are paid for by the state and will therefore have little—if any—impact on family’s financial situation during childhood. However, the variable spouse educational attainment might involve some degree of reciprocity, as being in good health can serve as a signal of high human capital, making an individual more attractive to potential partners who themselves possess high human capital often demonstrated through tertiary education. Still, to the extent that such reciprocal effects of health on socioeconomic predictors have occurred, they should have been accounted for in the measure of baseline health. However, given that spouse's educational attainment was reported only at follow-up, and we lack information on how long respondents’ have been together with their current partner, we cannot determine how long they may have been influenced by their current partner’s socioeconomic status or health behaviours associated with their level of education, nor do we know if respondents had any previous partner(s) with different levels of education than the current one. Second, the measures of physical activities and self-reported health were not identical at baseline and 21 years later. This discrepancy may introduce measurement errors over time. However, the variations in the questions were designed to capture the same phenomenon, and we do not anticipate that these differences would significantly affect the conclusions of this study. Moreover, a closer look at Model-III results in Tables 2 and 3 suggest no differences whether being Physical (in)active were persistent or not. Given that health behaviours account for the largest proportion of the explained variance in health outcomes, targeted interventions to improve health behaviours have the potential to not only enhance overall population health but also reduce health inequalities. This is particularly important, as individuals in lower socioeconomic positions are more likely to engage in unhealthy behaviours. For an insightful theoretical discussion on the relationships between socioeconomic circumstances and health behaviours, see [39]. Further empirical research across diverse institutional contexts is needed to assess the generalisability of the findings reported in this paper. Notably, Norway is a rich country with tax-funded national health service and free tertiary education, something that will reduce the influence of two key determinants of health inequalities. As such, Norway represents a ‘best-case’ benchmark against which other countries can be compared. Conclusions This study underscores the critical role of health behaviours in explaining inequalities in quality of life, and the significant—and persistent—associations between having experienced difficult financial circumstances in early life, and experiencing ill health in later life. Our findings carry important policy implications. While preventing childhood deprivation have positive impact on adult health, potential health benefits of promoting healthy behaviours are substantial. Both types of policy interventions thereby contribute to reducing health inequalities. Supplementary Information Below is the link to the electronic supplementary material.Supplementary file1 (DOCX 271 KB) Supplementary file1 (DOCX 271 KB)