Changed health behavior improves subjective well-being and vice versa in a follow-up of 9 years

A continuous health behavior sum score (HBSS, range 0–4) is a count of the number of beneficial health behaviors including physical activity, dietary habits, alcohol consumption, and smoking status. Physical activity performed in leisure time or in commuting was first converted into a Metabolic Equivalent Task (MET). A MET score ≥ 2 corresponding to 30 min walking per day was the cut-off value for beneficial behavior and provided one point in the HBSS. Dietary habits were assessed by a non-validated index (range 0–100), but formed in compliance with the Nordic nutritional recommendations [14]. Each choice, if being in line with the recommendations, provided one point in the dietary index according to the following cut-off points: dark bread (≥ 2/day); fat free milk (≥ 1/day); pastries/ sweets, sausages, red meat or chicken/ turkey (each ≤ 1–2/week); fish (≥ 1–2/week); fresh fruits and berries (≥ 2/day); vegetables (≥ 2/day); alcohol use (< 10 g women, 20 g men/day). Then, the sum score ranging from 0 to 10 was multiplied by 10 to give a percentage of compliance to recommendations [15]. Lastly, a cut-off value above median in this dietary index (≥ 60) provided one point for HBSS. Alcohol consumption was dichotomized according to Finnish guidelines [16] where risky consumption for women is ≥ 140 g/week and for men ≥ 280 g/week. Then, the values lower than the relevant cut-off provided one point for HBSS. Smoking status was dichotomized into current smokers vs. non-smokers in combination with former-smokers. The latter choice provided one point in HBSS.

Categorical change in health behavior (i.e., positive, neutral, and negative change) was measured by the change in the health behavior sum score (HBSSchange) during follow-up (i.e. the difference between HBSS₂₀₁₂ and HBSS₂₀₀₃). For evaluating change, individuals reporting zero or one beneficial health behaviors were combined into one group due to the small number of participants. Participants were categorized into three groups according to the direction of change in the HBSS during follow-up: positive, neutral, and negative change.

SWB was measured with the continuous four-item life satisfaction scale (range 4–20) [17, 18] where a lower score indicates better subjective well-being. It has three life assessments (i.e., interest, happiness, and ease in life) representing the cognitive component of SWB [1]. The fourth item is perceived loneliness, which is not typical for a life satisfaction scale, but reflects social well-being. The wordings of the life assessments were “Do you feel that your life at the moment is …”, and for the perceived loneliness “Do you feel that at the moment you are …? The responses were scored as follows: very interesting/ happy/ easy/ not at all lonely = 1; fairly interesting/ happy/ easy = 2; cannot say = 3; fairly boring/ unhappy/ hard/ lonely = 4; very boring/ unhappy/ hard/ lonely = 5 [18, 19]. Previously, the life satisfaction scale has often had three categories: satisfied (score = 4–6); intermediate group (score = 7–11, within ± 1 SD from the mean); dissatisfied (score = 12–20). In the present study, the intermediate group was divided at the mean to create four groups of SWB: high (score = 4–6), high intermediate (score = 7–8), low intermediate (score = 9–11), and low (score = 12–20).

The change in subjective well-being (SWBchange) during follow-up was assessed based on the direction of change between SWB groups with three categories (as in HBSSchange): positive, neutral, or negative change (each: yes/no). Thus, the SWBchange indicated the direction of change from one SWB group to another during the follow-up according to the difference in the group levels of SWB₂₀₁₂ and SWB₂₀₀₃, not in their continuous scores.

Factors potentially affecting both SWB and health behavior were included as covariates according to groupings based on the initial random sampling (age) and previous publications of the data. Participants represented four age groups in 2003: 25–29 years (group 1), 35–39 years (group 2), 45–49 years (group 3), and 55–59 years (group 4). Education was also categorized into four groups: (1) no professional education; (2) vocational course/school/apprenticeship contract; (3) college; (4) university degree/university of applied sciences. Health status was categorized according to count of the self-reported diseases into three groups: 0, 1, ≥ 2 diseases. Disease count is commonly used due to its simplicity and the ease of data ascertainment. Even if it does not consider the severity of disease, it is linked with mortality and various health adversities [20]. The complete list of 35 diseases can be found in the Additional file 1. Negative life events before the follow-up survey could affect both SWB and health behavior. From a list of 21 life events in 2007–2012, participants reported burdensome and extremely burdensome life events [21] which were then transformed into a trichotomized covariate: 0, 1, ≥ 2 major negative life events. Details of the life events and the item in the survey can be found in the Additional file 1.

Linear regression models were used to analyze:

As already described, both HBSSchange and SWBchange were measured by the difference between their baseline and end of follow-up grouping divided into three categories (i.e. positive, neutral, and negative change). The first digit of the adjustment models identified the direction (1: from HBSS₂₀₀₃ to SWB_2012; 2: from SWB₂₀₀₃ to HBSS₂₀₁₂) and the second digit was for the adjustment factors. Thus, the Model 1.1. from HBSS₂₀₀₃ to SWB₂₀₁₂ was adjusted for age, gender, education, and disease count (Table 3). The additional adjustments were for either HBSSchange (Model 1.2), or SWB₂₀₀₃ (Model 1.3) or both (Final model 1.4.). The Model 1.4 was then further adjusted with negative life events to create Model 1.5 (For clarification see Table 3). To study the effect of change without baseline level, HBSSchange was made the predictor instead of baseline HBSS₂₀₀₃ to create Model 1.6, which is otherwise parallel to Model 1.1. The Model 1.7 was created by adding also SWB₂₀₀₃ in Model 1.6. (Table 3).

In the opposite direction (from SWB₂₀₀₃ to HBSS₂₀₁₂) (Table 5), the Model 2.1 was adjusted for age, gender, education, and disease count as in Model 1.1., but it had also the SWB₂₀₀₃*education interaction term (Table 5). The additional adjustments were for either SWBchange (Model 2.2) or HBSS₂₀₀₃ (Model 2.3) or both (Final model 2.4). The Model 2.4 was then further adjusted with negative life events to create Model 2.5. To study the effect of change without baseline level, SWBchange was made the predictor instead of baseline SWB₂₀₀₃ and the interaction term SWB₂₀₀₃*education was converted to SWBchange*education to create Model 2.6 from the Model 2.1. The model 2.7 was created by adding also HBSS₂₀₀₃ in the Model 2.6. (Table 5). Data were analyzed with SAS software (version 9.4; SAS Institute Inc. Cary, NC, USA 2016).

To be able to compare the effects of the SWBchange and HBSSchange on their outcomes i.e. HBSS₂₀₁₂ (range 0–4 with 4 steps) and SWB₂₀₁₂ (range 4–20 with 16 steps), the HBSS₂₀₁₂ should be rescaled. Thus, the estimate for SWBchange is multiplied by 4 (4 × 4 = 16) to be comparable for the estimate of HBSSchange.

HBSS₂₀₀₃ predicted the SWB₂₀₁₂ in each model (Table 3), the β being – 0.35 (p < 0.001, Akaike information criterion (AIC) = 48,300) in the final model 1.5 (adjusted for age, gender, education, disease count, HBSSchange, SWB₂₀₀₃ and negative life events). A positive HBSSchange predicted 0.31 (p < 0.001) points lower SWBscore₂₀₁₂ (i.e., better SWB) and a negative change predicted 0.37 (p < 0.001) points higher SWBscore₂₀₁₂ after follow-up compared to those with no change in the HBSS₂₀₀₃ during follow-up (Table 4). In addition, every one point increase in the HBSS₂₀₀₃ (β = 0.35; p < 0.001) resulted in a 0.35 points lower SWBscore₂₀₁₂ (indicating better SWB) and every higher point in the SWBscore₂₀₀₃ (β = 0.42; p < 0.001) resulted in a 0.42 points higher SWBscore₂₀₁₂. With the reference group being at least two major negative life events in 2007–2012, having one or none major negative life event resulted separetely analyzed in overall 0.75 (p < 0.001) and 1.08 (p < 0.001) points lower SWBscore₂₀₁₂ (indicating better SWB), respectively. Further, lower disease count in 2003 (p < 0.001) resulted in better SWB₂₀₁₂ compared to those with at least two diseases. Age as a covariate (p < 0.001) resulted in a U-shaped pattern in SWB₂₀₁₂ (age group 2 having worst SWB₂₀₁₂). In addition, education was a significant covariate (p = 0.038) but not gender. Using HBSSchange as a predictor instead of HBSS₂₀₀₃ in Model 1.1. (Model 1.6) or adding baseline SWBscore₂₀₀₃ in it (Model 1.7), did not improve the AIC or statistical significance of covariates. (Table 3).

In the final model 2.4 (adjusted for age, gender, education, disease count, HBSSchange and HBSS₂₀₀₃) β was –0.025 (p < 0.001, AIC = 23,000). A positive SWBchange resulted in 0.071 (p < 0.001) points higher HBSS₂₀₁₂ and a negative change 0.072 points lower HBSS₂₀₁₂ compared to the ones who experienced a neutral SWBchange during follow-up. Every point towards better SWB₂₀₀₃ (decrease in score) improved HBSS₂₀₁₂ by 0.025 points, and every additional point in HBSS₂₀₀₃ improved the HBSS₂₀₁₂ by 0.48 points. Better HBSS₂₀₁₂ was observed in women (p < 0.001) (vs. men) and those with no diseases (p < 0.03) (vs. ≥ 2 diseases). Further, having a university level education had a statistically significant effect on the interaction SWB₂₀₀₃*education (p = 0.005). Age appeared to have a U-shaped effect on subsequent health behavior, but it lost its significance when covariates were added. Education by itself was not a significant covariate. For details of different models, see Table 5, and for final model 2.4, see Table 6. The model did not improve when adjusting for negative life events (p = 0.52) or when SWBscore₂₀₀₃ was replaced by SWBchange and the interaction SWBchange*education included in the Model 2.6 (Table 5).

The range of the score depicting SWB (4–20 i.e., 16 different points) is four times the magnitude of range of HBSS (0–4). When comparing the magnitudes of effect, the negative HBSSchange had a greater effect (0.37 points) on SWB₂₀₁₂ than a negative SWBchange on HBSS₂₀₁₂ (4*0.072 points = 0.288 points). A positive HBSSchange resulted in slightly greater effect (0.31points) compared to a positive SWBchange (4*0.071 points = 0.284 points). Thus, it appears that changes in HBSS have stronger effect on SWB than vice versa. The maximum effect of SWB₂₀₀₃ (range 4–20) on HBSS₂₀₁₂ (range 0–4) could be 16*0.02548 (i.e., -0.25 in Table 6) ≈ 0.41 and the maximum effect of HBSS₂₀₀₃ on SWB₂₀₁₂ was 4*0.3515 (i.e., -0.35 in Table 4) ≈ 1.4.

The study has also limitations. Attrition might have caused some overrepresentation of positive changes as well as better baseline levels of HBSS and SWB. The magnitude of change was not quantified. Only the direction of change was used in the analysis. Therefore, firm conclusions of the effect of a specific change cannot be drawn. The use of only two follow-up points might be subject to bias of momentary external factors. However, the large data could eradicate most of the effects of individual factors in people’s lives. The study is observational in nature. Thus, definite conclusions about causality cannot be made. The measures are obtained by self-report which especially for health behavior could be subject to reporting-bias [33]. Even if self-reported diseases lack the severity assessments, disease count is the most common instrument to measure the level of multimorbidity being also linked with various health outcomes according to a recent systematic review [20].