Data
We used two waves of survey data merged with the diagnostic information on medical conditions included in the medical insurance claims data. Specifically, a group of randomly sampled working adults of a large, national service organization based in the United States provided survey data at two occasions. In the first wave, conducted in June 2018, 2370 individuals provided responses. First wave participants were subsequently invited to the second wave of the study and provided responses in July 2019. The number of participants in both waves amounted to 1,209, which yielded the retention rate of 51.2%. Among participants, females accounted for 84.5% vs. 74.5% for the entire population, which reflected the feminization rate in the organization. Mean age of participants was 43.5 years in the sample compared to 45.6 in the population.
The survey was designed to comprehensively assess well-being and work conditions among employees. It was administered online, which allowed participants to choose a secure and anonymous space to participate in the study. Eligibility criteria for participation included age (i.e., at least 18 years of age) and employment status (i.e., all current employees were considered). Participation was voluntary, confidential, and conditional on the informed written consent that was collected from each participant. Harvard Longwood Campus Institutional Review Board reviewed and approved all protocols for the study. More information about the study and sample is presented elsewhere [6, 28, 29].For respondents who participated in wave 1 (T = 1) and wave 2 (T = 2), we merged their survey records with their medical insurance claims data (T = 0, T = 1, and T = 2) that were provided by the employer. Next to a number of financial measures such as allowed amounts for medical services and pharmacy products, medical insurance data included data on medication prescribed and diagnostic information on medical conditions, which were of interest in this study. Diagnostic information followed the International Classification of Diseases (ICD-10) [30]. It has been also demonstrated to be highly consistent with medical records and useful in epidemiological studies [31, 32]. Merged survey and medical insurance data have been already found useful in other research addressing well-being and health [28]. Table 1 (adapted from [28]) presents the descriptive statistics at baseline (T = 1). Data are available on reasonable request.
Table 1 Participant characteristics at study baseline (T = 0, survey data 2018–2019 merged with health insurance data 2017–2019, United States, N = 1209) Measures
Mental health outcomes
We examined one self-reported mental health outcome from the Well-Being Assessment (WBA)Footnote 1 [29, 33] and the Flourishing Index [4, 34] [‘In general, how would you rate your mental health?’ (0 = poor and 10 = excellent)] and two mental health outcomes captured in health insurance claims data, that is: (1) diagnosis of depression (yes vs. no) and/or (2) diagnosis of anxiety (yes vs. no).
Physical health outcomes
We examined one self-reported physical health outcome from the WBA [29, 33] and the Flourishing Index [4, 34] [‘In general, how would you rate your physical health?’ (0 = poor and 10 = excellent)] and one diagnostic information on medical conditions outcomes derived from the participants’ medical insurance claims data, i.e., a diagnosed cardiovascular disease (yes vs. no).
Strengths of moral character
To measure adherence to high standards of moral behavior, a subscale of the WBA, related to strengths of moral character (SMC), was used. The SMC-WBA instrument was developed based on the concept of human flourishing or complete well-being [4, 35]. The SMC domain, which is of interest in this study, was conceptualized according to a long-standing religious and philosophical tradition, partially adopted by positive psychology in recent years, positing that in order to attain complete well-being, an excellent character and acting in accordance with the virtue, are essential [1, 13, 14, 36, 37]. Consequently, this domain was defined as adherence to high standards of moral behavior reflected in an ability to focus, to maintain consistent thoughts, and to act in a way that contributes to the good of oneself and others [33]. High score in SMC-WBA indicates a self-assessed “strength” in moral character.
SMC-WBA is related to the concept of ‘character strengths’ in general [37] and to one popular measure of character strengths specifically—the VIA Survey of Character Strengths [36]. We refer to our assessment as a ‘measure of strengths of moral character’ to highlight its moral component and to distinguish it from the VIA measure of character strengths. In the Supplementary Information, we present details on similarities and dissimilarities between our measure and the VIA Survey of Character Strengths [36].
Five aspects of adherence to high standards of moral behavior were examined with seven statements from the SMC-WBA [29, 33]:
-
1.
moral compass (‘I always know the right thing to do’),
-
2.
orientation to promote good (‘I am willing to face difficulties in order to do what is right’, ‘I give up personal pleasures whenever it is possible to do some good instead’, and ‘I always act to promote good in all circumstances, even in difficult and challenging situations’),
-
3.
use of strengths (‘I get to use my strengths to help others’),
-
4.
kindness (‘I always treat everyone with kindness, fairness and respect’) and
-
5.
delayed gratification (‘I am always able to give up some happiness now for greater happiness later’).
Respondents could choose an answer on a 0 = ‘not true of me’ to 10 = ‘completely true of me’ scale. The seven items of the SMC-WBA were moderately correlated (r = 0.36–0.61; correlations between main study variables are presented in Table A1 in the Supplementary Information). In addition, SMC-WBA (an aggregate of seven items from the SMC-WBA) was also used as an exposure. This scale was validated and showed satisfactory psychometric properties in terms of reliability (alpha = 0.88), test–retest correlation (r = 0.67), and convergent/discriminant validity in relation to stability over time (r = 0.75), as well as a good fit to the data (confirmatory factor analysis: CFI = 0.962, TLI = 0.943, RMSEA = 0.069) that were invariant over time, gender, age, education, and marital status [a complete psychometric evaluation can be found in 33].
Control variables
A rich set of control variables was used. Specifically, we controlled for demographic characteristics [gender (male, female), age group (≤ 30, 31–40, 41–50, > 50), race (White, Black/African American, Hispanic/Latino, Asian, other), educational attainment (high school, some college, associate degree, bachelor’s degree, graduate degree), marital status (married vs. not married), having children at home (yes vs. no), taking care of an elderly (yes vs. no)], wealth [owning a house (yes vs. no)], and income [salary (the mid-point salary bands were provided by the employer)]. These variables are classified as social determinants of health and, as shown by previous research [38, 39], have a substantial impact on people’s health, well-being and quality of life. In addition, we controlled for social participation and civic engagement. The variables comprised: (1) voting in the last elections (yes vs. no/not registered voter), (2) religious service attendance (at least once a week, less than once a week, never), (3) spiritual practices (at least once a week, less than once a week, never), (4) volunteering (at least once a week, less than once a week, never), and (5) community work (at least once a week, less than once a week, never). In prior studies, these factors were found to play a predictive role for health and well-being [40–44, 77].
Next, since the impact of work on health has long been recognized in theory [45] and empirical research [46–50, 78], we controlled for work characteristics. We included selected indicators of work resources, work demands and work autonomy: number of work hours, supervisor support [‘My supervisor supports me’ (0–10)], job control [‘I have a lot of freedom to decide how to do my job’ (0–10)], job demand [‘I have too much to do at work to do a good job’ (0–10)], job fit [‘At work, I am able to do what I am good at’ (0–10)] and job meaning [‘I find my work meaningful’ (0–10)] [51, 52]. These variables were controlled for in the first wave (T = 1). In addition, in each regression, the control was made for an outcome and additionally for the number of diagnosed health conditions prior to exposure to further reduce possibility of reverse causality.
Statistical analysis
This study applied an outcome-wide analytic approach [53] and used longitudinal observational data merged with medical insurance claims data. The logistic (for dichotomous outcomes) and linear (for continuous outcomes) regression analysis was applied. All continuous outcomes were standardized (i.e., mean = 0, standard deviation = 1), to report the effect estimates in terms of standard deviations of the outcome variables (i.e., standardized effect sizes). For dichotomous outcomes, we presented odds ratios.
A set of 40 regression models was used to regress each of the five outcomes on each of the eight exposures (i.e., SMC-WBA and its seven items) separately. In particular, the association between a character strength exposure j and a health outcome k for continuous outcomes was modeled as follows:
$${HO}_{i,k}\left(T=2\right)={{\alpha }_{0}+{{\alpha }_{1}{SMC}_{j,i}\left(T=1\right)+\alpha }_{2}X}_{i}\left(T=1\right)+{\alpha }_{3}{HO}_{i,k}\left(T=1\right)+{\eta }_{k,j,i},$$
(1)
and for dichotomous outcomes as follows:
$${prob[HO}_{i,k}(T=2)=1]=\frac{1}{1+{\mathrm{e}}^{-\left({{\alpha }_{0}+{{\alpha }_{1}{\mathit{SMC}}_{j,i}\left(T=1\right)+\alpha }_{2}X}_{i}\left(T=1\right)+{{\alpha }_{3}{\mathit{HO}}_{i,k}\left(T=0\right)\eta }_{k,j,i}\right)}},$$
(2)
where i = 1,…,N; k = 1,…,5; j = 1,…,8.
Subscript i represents an individual, the variable HO indicates one out of five (k = 1,…,5) health outcomes, SMC is one out of eight exposures (j = 1,…,8). X is a vector of control variables. α1 reflects an association between SMC exposure and a subsequent health outcome. α2 shows the association between control variables and the health outcome. α3 shows the association between the health outcome k at T = 2 and T = 1 for self-reported health outcomes and at T = 2 and T = 0 for medical condition outcomes. ηk,j,i is a disturbance term.
All missing exposure, covariate, and outcome variables were imputed using chained Eqs. (20 datasets were generated) [54, 55]. Data were arranged in a wide format as suggested by Allison [56] and all outcome, exposure and control variables were used in the procedure. Consequently, the multiple imputation estimates pooled using the Rubin’s formula [57] are presented. Bonferroni correction was used to correct for multiple testing.
Robustness of the results was examined through a series of robustness analyses. First, for three regressions of diagnosed conditions (i.e., depression, anxiety, and a cardiovascular disease, derived from the medical insurance diagnostic information), supplementary analyses were conducted on a limited sample of those who did not suffer from the health outcome under examination prior to exposure (as opposed to the primary analysis of the entire sample controlling for the outcomes prior to exposure; Table A2 in the Supplementary Information). Second, two additional sets of controls were added to the primary set of analyses: (1) an alternative specification of the overall 2018 well-being index; it was calculated excluding the character strength specific domain in 2018 (Table A3 in the Supplementary Information), (2) all five well-being domain-specific scores in 2018 (the character strength specific domain score was excluded; Table A4 in the Supplementary Information). Third, we reanalyzed the primary sets of models using the complete-case analysis (Table A5 in the Supplementary Information) to examine robustness of the results to the missing data patterns. Fourth, because our choice of using a broad set of controls might have contributed to overfitting the models, we rerun them excluding particular sets of confounders (Table A6 in the Supplementary Information). In model 1, we controlled only for social determinants of health (i.e., demographic characteristics, wealth, and income). In model 2, compared to model 1, we added social participation and civic engagement (i.e., we controlled for demographic characteristics, wealth, income, social participation, and civic engagement). To decrease the risk of reverse causation, both model 1 and model 2 also controlled for the prior outcomes and the history of disease. Fifth, we rerun the primary models using all items of SMC-WBA simultaneously to examine the overall effect of the co-occurrence of different aspects of SMC (Table A7 in the Supplementary Information). Finally, the sensitivity measures—E values—were calculated to assess the robustness of the observed associations to unmeasured confounding [58, 59].
Analyses were performed using Stata/SE 17.0 for Mac.