Marital transitions during earlier adulthood and subsequent health and well-being in mid- to late-life among female nurses: An outcome-wide analysis

Abstract

Comparing outcomes for individuals remaining married to those for single or divorced individuals might overstate the positive effects of the decision to marry, since marriage carries an inherent risk of divorce and its associated negative outcomes. While a growing literature has examined marital transitions, confounding by past marital history remains a concern and only a limited set of outcomes have been examined. To address these issues, this study examined incident first-time marriage and incident divorce/separation in relation to multiple subsequent physical health, health behavior, psychological distress, and psychosocial well-being outcomes in a large sample of female nurses in the U.S.. Data from the Nurses' Health Study II were studied (1993 to 2015/2017 questionnaire wave, N_{marriage analyses} = 11,830, N_{divorce/separation analyses} = 73,018, interquartile range of baseline age = 35 to 42 years). A set of regression models were used to regress each outcome on marital transition status, adjusting for a wide range of initial health and wellbeing status in addition to other covariates. Bonferroni correction was performed to account for multiple testing. Among the initially never married, those who became married had lower mortality (RR = 0.65, 95%CI = 0.50, 0.84), lower risks of cardiovascular diseases (e.g., RR_stroke = 0.64, 95%CI = 0.50, 0.82), greater psychological wellbeing and less psychological distress (e.g., ß_{depressive symptoms} = −0.10, 95%CI = -0.15, −0.06). Among the initially married, those who became divorced/separated had lower social integration (β = −0.15, 95%CI = -0.19, −0.11), greater psychosocial distress (e.g., RR_depression = 1.23, 95%CI = 1.10, 1.37), and possibly greater risks of mortality, cardiovascular diseases, and smoking. Future research could study similar questions using data from more recent cohorts, examine potential mechanisms and heterogeneity, and also examine alternative social relationship types.

Introduction

“Marriage represents the keystone institution for most…societies,” observes anthropologist Joseph Henrich, “and may be the most primeval of human institutions [1].” Marriage, in its diverse forms, arguably remains our species' most durable solution to the interconnected problems of nurturing children, socializing sexuality, and furnishing social support in many parts of the world [2]. Marriage trends, however, have undergone drastic changes in recent decades. Marriage rates in the U.S. have decreased by more than half as compared to a century ago [3], and around half of the first marriages now end in divorce [4]. These changes have been associated with increased single parenthood and risk of poverty especially among women [5,6], which has prompted a search for policy approaches to support single parents and improve family well-being [7].

In addition to its societal effects, marriage also shapes individual health and well-being. A large body of literature on prevalent marital status has suggested that married individuals often have lower rates of mortality and better physical and mental health outcomes such as lower risks of cardiovascular diseases or depression as compared to those not currently married, with the associations stronger in men than women [[8], [9], [10]]. Among those not currently married, those who are divorced, separated or widowed often have poorer health than the never married [11]. In addition, a growing literature has examined marital transitions, and also found that entering marriage is generally associated with improved health behaviors and wellbeing, whereas marital dissolution is related to unfavorable health outcomes and health behaviors [[12], [13], [14], [15], [16]]. A number of theories have been proposed to help understand the dynamics of marriage and wellbeing. For instance, the Health Protection Model posits that marriage may protect health and wellbeing through improving one's financial stability, increasing social support, as well as enhancing adherence to social norms against risky behaviors via increased social control from the spouse and an enhanced sense of responsibility for the family [17,18]. The model is supported by the empirical evidence that the married versus the unmarried often have higher levels of material wellbeing, social support, psychological wellbeing, and engagement in healthy lifestyles [12,19,20]. Conversely, marital dissolution is often associated with financial hardships, loneliness, depression and increased risky behaviors, leading to unfavorable biological profiles and adverse health [10,13,19]. In comparison, the Health Selection Model hypothesized that healthier individuals are more likely to get married and stay married, which leads to a positive correlation between marriage and health [21]. The two above-mentioned mechanisms are not mutually exclusive, and there is evidence suggesting that the associations between marriage and some health and wellbeing outcomes are likely reciprocal [22].

Past studies on marriage, health and wellbeing have substantially advanced our understanding of the dynamics and have significantly contributed to the literature. In this paper, we sought to address a few unresolved issues and criticisms in the literature, hoping to further strengthen the evidence base. First, one may argue that comparing individuals currently remaining married to those for single or divorced individuals has the potential to overestimate the effect of the decision to marry, since marriage carries an inherent risk of divorce, with its associated negative outcomes. Analyses that separately compare those currently married, never married, and divorced effectively set aside the issue that the decision to marry may itself eventually lead to divorce. One commentator has gone so far as to refer to this as the “cheater technique” in a recent critique [23]. It is not that these standard analyses on prevalent marital status are “wrong”; it is simply that they are considering a different question. The standard analyses are attempting to address the effects on average of current marital status and the decision to remain married, whereas one might alternatively try to assess the effects of the decision to become married (which may also for some subsequently result in divorce). While studies on marital transitions helped address this issue, residual confounding by past marital history remains a concern [24]. For instance, in some prior studies on marital transitions, the group of transition into marriage often included participants who entered marriage from diverse unmarried states (e.g., the previously divorced/separated, the widowed, the cohabitated, or the never married), whereas such variation in prior marital status was not always adequately accounted for in the analyses. Second, the marriage and wellbeing associations may vary across outcomes of interest [16]. Examining a single outcome separately in each study may limit the possibility of comparing the dynamics across broader aspects of one's life within the same sample, and may potentially mask some exceptions to the general pattern [19,25]. An outcome-wide approach [26] that examines multiple categories of outcomes within the same study may be an alternative approach to provide a broad view of the marriage-health dynamics.

To address these issues, this study performed an outcome-wide longitudinal analysis [26,27] to examine the associations of incident first-time marriage (i.e., restricting to the previously never-married to reduce confounding by past marital history) with a wide range of subsequent physical health, psychosocial well-being, psychological distress, and health behavior outcomes simultaneously in a sample of U.S. female nurses. For comparison purposes, we also examined the association of incident divorce/separation during the same time period for those married at baseline in relation to the same sets of outcomes as a secondary analysis. In both sets of analyses, various indicators of initial health and wellbeing status were adjusted for to reduce potential reverse causation and selection.

Methods

Study population

This study used longitudinal data from the Nurses' Health Study II (NHSII) [28]. Established in 1989, NHSII enrolled 116,412 female registered nurses aged between 25 and 42 years from across the U.S to study the dynamics between lifestyle factors, behaviors, personal characteristics, and health. The participants have been followed up since then through mail or web-based questionnaires biennially, and the response rates are approximately 90% at each follow-up cycle [28].

Marital status was first queried in year 1989 (mean age = 34 years) in NHSII, and the status was updated every 4 years thereafter. To get closer to the average age of first marriage among women in the 1990s [29], in the present study the sample for analyses on incident first-marriage was drawn from those who were never married in 1989 but became married or stayed never married in the subsequent wave in which marital status was again queried (i.e. the 1993 wave) (N = 11,830). The incident first-marriage group thus includes those who married between 1989 and 1993 but then might become divorced sometime later possibly experiencing negative effects of divorce that itself in some sense arose from the decision to marry. We did not use data on marital status during the follow-up waves for two reasons. First, we sought to rule out confounding by prior marital history by examining first-time marriage. The participants were moreover closer to the national average age of first-time marriage at the earliest waves of data collection in this sample. Second, in this study we sought to understand the effects of the decision to become married, whereas considering marital status over the follow-up as a time-varying variable would effectively address a different research question (i.e., the effects of remaining married averaged across time). Similarly, the analytic sample for incident divorce/separation was drawn from participants who reported being married in 1989 but became divorced/separated or stayed married in 1993 (N = 73,018) (For the analysis on incident divorce/separation, we were, however, unable to control for prior marital history due to lack of data on marital status prior to their entering in the cohort in 1989; nevertheless, we tried to get closer to the national average age of first-time divorce by using the earliest wave of marital transition data in this sample). For both analyses, the outcomes were taken from the most recent questionnaire waves for which data was available (principally 2015 or 2017). Potential confounders (detailed below) were assessed at baseline primarily in the 1989 questionnaire. Missing data on the outcome variables and the covariates were imputed using multiple imputation (see details about the multiple imputation procedure in the Statistical Analyses section). Further details regarding the sample derivation process were reported in the Supplementary Figure. This study was approved by the Institutional Review Board at the Brigham and Women’s Hospital. Informed consent was obtained from all individual participants included in the study.

Exposure assessment

Incident first-marriage

Current marital status was first queried in the 1989 questionnaire, and then again in 1993. The response options included “never married”, “married”, “divorced/separated”, and “widowed”. Among participants who were initially never married in 1989, those who reported being married in 1993 were considered as “became married”, whereas those who stayed never married in 1993 were considered as “remained never married”. Participants who were married, divorced/separated, or widowed in 1989 and those who were divorced/separated or widowed in 1993 were excluded from analyses on incident first-marriage.

Incident divorce/separation

Among participants who initially reported being married in the 1989 questionnaire, those who reported being divorced/separated in 1993 were considered as “became divorced/separated”, whereas those who stayed married in 1993 were considered as “remained married”. Participants who were never married, divorced/separated, or widowed in 1989 and those who reported being never married or widowed in 1993 were excluded from all analyses on incident divorce/separation.

Outcome assessment

A wide range of outcomes in middle adulthood were assessed (primarily in the 2015 or 2017 questionnaire wave). Such outcomes included multiple indicators of physical health (i.e., all-cause mortality, type 2 diabetes, stroke, coronary heart disease, cancer, overweight/obesity, number of physical health problems [sum of the above 5 physical health conditions]; data on physical health conditions were obtained from medical records, and data on mortality was obtained from the state vital records and the National Death Index), psychological well-being (i.e., positive affect, optimism, purpose in life, hopefulness, social integration, emotional support), psychological distress (i.e., depressive symptoms, depression diagnosis, anxiety symptoms, loneliness), and health behaviors (i.e., preventive healthcare use, heavy drinking, current smoking, physical activity, dietary quality, sleep duration). Details regarding assessment of all outcome variables were reported in the Supplementary Text.

Covariate assessment

Based on the disjunctive cause criteria for identifying confounders in the context of outcome-wide studies [27,30], we tried to control for all variables that are likely causes of the exposure or of any outcome wherever data were available, in addition to conventional sociodemographic covariates. Specifically, the covariates, assessed primarily in the 1989 questionnaire (unless otherwise specified below), included participants' age (in years), race/ethnicity (non-Hispanic white, others), geographic region (Northeast, South, West, Midwest), census-tract median income (<$50,000, $50,000–$74,999, $75,000–$99,999, ≥$100,000), census-tract college education rate (continuous), night shift work schedule over the past two years (none, 1–9 months, 10–19 months, 20+ months), and childhood abuse victimization (retrospectively reported in a 2001 supplementary questionnaire, a summary score ranging from 0 to 5 was created) [31]. To reduce the possibility of reverse causation [27], we also controlled for baseline values of the outcome variables simultaneously wherever data were available. These included the baseline values of: preventive healthcare use, heavy drinking, cigarette smoking, physical activity, dietary quality, and the number of physical health problems.

Statistical analyses

All statistical analyses were performed in SAS 9.4 (tests of statistical significance were two-sided). Descriptive analyses were first used to examine the distribution of participant characteristics across marital status at study baseline.

The primary analyses respectively examined the associations of 1) incident first-time marriage among those initially never-married in 1989 and then comparing the “became married” (N = 3272) with the “remained never married” (N = 8558) and 2) incident divorce/separation among those currently married in 1989 and then comparing those who “became divorced/separated” (N = 3527) with those who “remained married” (N = 69,491) on a wide array of subsequent health and well-being outcomes. In both analyses a set of regression models were used to regress each outcome on the exposure variable (i.e., marital transition status) separately, adjusting for all covariates. Specifically, we ran one of the following models, depending on the nature of the outcome variable: 1) modified Poisson regression models for all binary outcomes to estimate risk ratios [32]; and 2) linear regression models for continuous outcomes to estimate mean difference. We standardized all continuous outcomes at mean = 0 and standard deviation = 1, such that the effect estimates were reported in terms of standard deviations in the outcome variables. We also reported the unstandardized effect estimates in the supplement. To account for multiple testing, we performed Bonferroni correction. However, the practices for correction for multiple testing vary widely and it is an evolving research area in the field [33,34]. To acknowledge the differently types of cutoffs that can be used in interpreting the results, we marked multiple p-value cutoffs both with and without Bonferroni correction.

In the full analytic sample that includes participants with data on marital status at both the 1989 and 1993 questionnaire, the proportion of missing data on covariates was <12.75% (except for the covariate of childhood abuse which was assessed in a later wave), and <34.37% of participants had additional missing data on the outcome variables or were lost to follow-up over the 20+ years of study period. Because multiple imputation for missing data is often more flexible than many other methods of handling missing data [[35], [36], [37]], we performed multiple imputation by chained equations (with 20 imputed datasets created) to impute missing data on the outcome variables and the covariates. As a sensitivity analysis, we performed complete-case analysis to reanalyze the primary sets of models.

We also performed a number of other sensitivity analyses. First, we evaluated the extent to which the associations of marital transitions with various outcomes were robust to potential unmeasured confounders [38,39]. To this end, we calculated E-values [39], which represent the minimum strength of association that an unmeasured confounder(s) would need to have with both the exposure and the outcome variables on the risk ratio scale to fully explain away the exposure-outcome associations, above and beyond the measured covariates. Second, to reduce concerns about potential non-linearity between the continuous covariates and outcomes, we also reanalyzed the primary sets of models using quintiles or indicator variables for continuous covariates. Third, because depression was not measured prior to the assessment of marital transitions, we adjusted for depression assessed contemporaneously with marital transitions (in the 1993 questionnaire wave) as a sensitivity analysis. Without distinctions in the timing of measurements, however, it is possible that marital transition occurred before depression, and depression was a mediator affecting subsequent health. Next, to also examine relations to prior literature, we examined the associations of prevalent marital status (i.e., currently married versus never married, divorced/separated, or widowed) with various outcomes, both with and without adjustment for prior marital status as a covariate. Specifically, the analyses used the full analytic sample that included all participants with data on marital status in both the 1989 and 1993 questionnaire, and the primary sets of models were reanalyzed to regress each outcome on prevalent marital status (as assessed in the 1993 questionnaire) separately, both with and without adjustment for prior marital status (as assessed in the 1989 questionnaire).

Results

Participant characteristics

In the full analytic sample that included all participants with data on marital status in both the 1989 and 1993 questionnaire, the participants had a mean age of 38.16 years (interquartile range = 35 to 42 years) in 1993, were predominantly white, had relatively high socioeconomic status (SES), and were generally healthy and were mostly married (Supplementary Table S1).

Among those who were never married in 1989, 28% of them became married in 1993. Compared to those who remained single, participants who became married were younger, had fewer experiences of childhood abuse, worked fewer nightshifts, and were healthier at baseline (Table 1). Among participants who were married in 1989, around 5% of them became divorced/separated in 1993. Compared to those who remained married, participants who became divorced/separated reported more experiences of childhood abuse and had more unfavorable health behaviors at baseline (Table 1). The distribution of all outcome variables at the follow-up wave was reported in Supplementary Table S2.

Table 1.

Participant characteristics by baseline marital transition status (The Nurses' Health Study II 1993 Questionnaire Wave, N = 11,830 for analyses on incident marriage, N = 73,018 for analyses on incident divorce/separation).

	Baseline Marital Transition Status
	Analytic sample for incident marriagea		Analytic sample for incident divorce/separationb
Participant characteristics	Became married (n = 3272)	Remained never married (n = 8558)	Became divorced/separated (n = 3527)	Remained married (n = 69,491)
Age in years (range: 28 to 48)	33.74 (3.97)	37.14 (4.78)	38.54 (4.56)	38.24 (4.56)
Non-Hispanic white, %	92.08	91.51	93.54	95.62
Geographic region, %
Northeast	38.01	38.44	30.10	35.18
Midwest	30.98	29.49	32.03	34.63
South	14.91	14.29	21.11	17.81
West	16.10	17.78	16.75	12.38
Census tract median income, %
<$50‚000	35.31	39.53	37.56	33.92
$50‚000-$74‚999	43.44	41.93	42.65	44.67
$75‚000-$99‚999	15.35	12.89	15.28	15.88
≥$100‚000	5.90	5.64	4.51	5.53
Census tract college education rate (range: 0 to 1)	0.33 (0.19)	0.32 (0.19)	0.28 (0.16)	0.29 (0.17)
Childhood abuse victimization (range: 0 to 5)	1.60 (1.42)	1.78 (1.52)	2.15 (1.57)	1.72 (1.48)
Night shift work schedule, %	66.08	66.65	60.60	60.50
Depression, %c	10.72	17.11	22.54	12.46
Prior preventive healthcare use, %	73.94	66.14	69.49	72.02
Prior heavy episodic drinking, %	15.87	9.72	6.70	4.33
Prior cigarette smoking, %	13.66	17.05	17.20	10.71
Prior physical activity (reach recommended level), %	77.76	72.75	66.17	66.06
Prior dietary quality score (range: 13.82 to 95.68)	50.17 (10.86)	49.96 (11.34)	48.34 (10.77)	47.27 (10.58)
Prior number of physical health problems (range: 0 to 2)d	0.19 (0.40)	0.40 (0.49)	0.30 (0.46)	0.29 (0.45)

Notes: Values are means (SD) for continuous variables; percentages for categorical variables. Values of polytomous variables may not sum to 100% due to rounding.

^aThe analytic sample for incident marriage was restricted to participants who were never married in 1989 questionnaire and became married or remained never married at wave 1993.

^bThe analytic sample for incident divorce/separation was restricted to participants who were married in 1989 questionnaire and became divorced/separated or remained married in 1993.

^cDepression was assessed simultaneously with the assessment of marital transition. Depression was not assessed prior to the 1993 questionnaire wave in this cohort.

^dThe score sums the total number of the following prior physical health conditions: prior overweight/obesity, type 2 diabetes, myocardial infarction, stroke, and cancer.

Incident marriage and subsequent health and well-being

Among participants who were initially never married, those who became married had a 35% lower risk of all-cause mortality (Risk ratio [RR] = 0.65, 95% confidence interval [CI] = 0.50, 0.84), lower risks of cardiovascular diseases (e.g., RR_stroke = 0.64, 95% CI = 0.50, 0.82), greater psychological wellbeing on multiple indicators including greater positive affect, purpose in life, hope and optimism (e.g., ß_optimism = 0.12, 95% CI = 0.07, 0.17), and less psychological distress including fewer depressive symptoms and less loneliness (e.g., ß_loneliness = −0.10, 95% CI = -0.15, −0.05), as compared to those remaining never married. There was also evidence that entering marriage was possibly associated with greater emotional support, fewer anxiety symptoms, and less smoking; though these latter associations only passed the conventional, but not the Bonferroni-corrected, p < 0.05 thresholds. There was, however, little evidence that marriage was associated with other health behaviors or other incident physical illness outcomes (Table 2, Supplementary Table S3). Results were highly similar when quintiles or indicator variables for continuous covariates were used (Supplementary Table S4). Complete-case analyses yielded similar results (Supplementary Table S5).

Table 2.

Incident marriage and subsequent health and well-being among participants who were initially never married (The Nurses' Health Study II 1993 to 2015/2017 questionnaire, N = 11,830^a).

Subsequent health and well-being outcomes	Marital Transition Status b
	Became married vs. remained never married
	RRc	βd	95% CI	P-value
Physical health
All-cause mortality	0.65		0.50, 0.84	0.001f
Number of physical health problems		−0.03	−0.07, 0.01	0.168
Overweight/obesity	1.01		0.95, 1.07	0.816
Type 2 diabetes	0.94		0.82, 1.09	0.418
Coronary heart disease	0.67		0.52, 0.85	<0.001f
Stroke	0.64		0.50, 0.82	<0.001f
Cancer	0.95		0.85, 1.07	0.433
Psychological Well-being
Positive affect		0.10	0.06, 0.15	<0.001f
Optimism		0.12	0.07, 0.17	<0.001f
Purpose in life		0.10	0.05, 0.14	<0.001f
Hopefulness		0.09	0.04, 0.14	<0.001f
Social integration		0.01	−0.04, 0.05	0.830
Emotional support		0.06	0.01, 0.11	0.022e
Psychological Distress
Depressive symptoms		−0.10	−0.15, −0.06	<0.001f
Depression diagnosis	0.94		0.78, 1.12	0.477
Anxiety symptoms		−0.07	−0.12, −0.02	0.005e
Loneliness		−0.10	−0.15, −0.05	<0.001f
Health Behaviors
Preventive healthcare use	1.01		0.96, 1.05	0.832
Heavy drinking	1.15		0.90, 1.46	0.271
Current smoking	0.71		0.54, 0.92	0.010e
Recommended level of physical activity	1.01		0.95, 1.06	0.807
High dietary quality	0.98		0.89, 1.07	0.617
Short sleep duration	0.93		0.85, 1.00	0.064

Abbreviations: RR, risk ratio; CI, confidence interval.

^aThe analytic sample was restricted to those who were never married at questionnaire wave 1989 and were married or remained never married at wave 1993. The outcome variables were assessed in the most recent questionnaire waves, primarily the 2015 or 2017 wave. Multiple imputation was performed to impute missing data on covariates and outcomes.

^bA sets of regression models were used to regress each outcome on marital transition status separately. All models controlled for age, race/ethnicity, geographic region, night shift schedule, census-tract college education rate, census-tract median income, childhood abuse victimization, and prior values of the outcome variables (including preventive healthcare use, heavy drinking, current smoking, physical activity, dietary quality score [in quartiles], number of physical health problems) wherever data were available.

^cThe effect estimates for all dichotomized outcomes were RR.

^dAll continuous outcomes were standardized (mean = 0, standard deviation = 1), and β was the standardized effect size.

^ep < 0.05 before Bonferroni correction.

^fp < 0.05 after Bonferroni correction (the p value cutoff for Bonferroni correction is p = 0.05/23 outcomes = 0.0022).

The sensitivity analysis additionally adjusting for depression measured contemporaneously with marital transitions also yielded similar results, although the associations with several indicators of psychological wellbeing and with depression became slightly weaker (Supplementary Table S6). Results from the analyses on prevalent marital status comparing the currently married versus the never married were again similar when additional control was made for prior marital status but yielded stronger associations with several outcomes, when prior marital status was not adjusted for (Supplementary Tables S7 and S8).

Incident divorce/separation and subsequent health and well-being

Among participants who were initially married, those who became divorced/separated reported substantially lower levels of social integration (β = -0.15, 95% CI = -0.19, −0.11), greater depression (RR = 1.23, 95% CI = 1.10, 1.37) and loneliness (ß = 0.11, 95% CI = 0.08, 0.15), as compared to those remaining married. Marital dissolution was possibly also related to a 19% higher risk of all-cause mortality and greater risks of cardiovascular diseases and smoking, though these latter associations reached the conventional, but not the Bonferroni corrected, p < 0.05 threshold. Again, there was little evidence of associations with other health behavior outcomes (Table 3, Supplementary Table S9). Results were highly similar when quintiles or indicators for continuous covariates were used (Supplementary Table S10). Sensitivity analysis using complete-case analysis yielded similar results (Supplementary Table S11).

Table 3.

Incident divorce/separation and subsequent health and well-being among participants who were initially married (The Nurses' Health Study II 1993 to 2015/2017 questionnaire, N = 73,018^a).

Subsequent health and well-being outcomes	Marital Transition Statusb
	Became divorced/separated vs. remained married
	RRc	βd	95% CI	P-value
Physical health
All-cause mortality	1.19		1.02, 1.40	0.031e
Number of physical health problems		0.04	0.01, 0.08	0.020e
Overweight/obesity	1.01		0.96, 1.06	0.776
Type 2 diabetes	1.09		0.99, 1.19	0.072
Coronary heart disease	1.23		1.06, 1.42	0.005e
Stroke	1.21		1.05, 1.40	0.010e
Cancer	1.01		0.92, 1.10	0.825
Psychological Well-being
Positive affect		−0.03	−0.07, 0.01	0.109
Optimism		−0.01	−0.05, 0.03	0.532
Purpose in life		−0.03	−0.07, 0.01	0.108
Hopefulness		−0.01	−0.05, 0.03	0.698
Social integration		−0.15	−0.19, −0.11	<0.001f
Emotional support		−0.04	−0.07, 0.00	0.068
Psychological Distress
Depressive symptoms		0.07	0.03, 0.11	<0.001f
Depression diagnosis	1.23		1.10, 1.37	<0.001f
Anxiety symptoms		0.02	−0.02, 0.06	0.320
Loneliness		0.11	0.08, 0.15	<0.001f
Health Behaviors
Preventive healthcare use	0.98		0.94, 1.02	0.295
Heavy drinking	1.06		0.89, 1.25	0.527
Current smoking	1.21		1.04, 1.40	0.012e
Recommended level of physical activity	0.97		0.93, 1.02	0.213
High dietary quality	0.99		0.92, 1.07	0.803
Short sleep duration	1.03		0.96, 1.10	0.455

Abbreviations: RR, risk ratio; CI, confidence interval.

^aThe analytic sample was restricted to those who were married at questionnaire wave 1989 and were divorced/separated or remained married at wave 1993. The outcome variables were assessed in the most recent questionnaire waves, primarily the 2015 or 2017 wave. Multiple imputation was performed to impute missing data on covariates and outcomes.

^bA sets of regression models were used to regress each outcome on marital transition status separately. All models controlled for age, race/ethnicity, geographic region, night shift schedule, census-tract college education rate, census-tract median income, childhood abuse victimization, and prior values of the outcome variables (including preventive healthcare use, heavy drinking, current smoking, physical activity, dietary quality score [in quartiles], number of physical health problems) wherever data were available.

^cThe effect estimates for all dichotomized outcomes were RR.

^dAll continuous outcomes were standardized (mean = 0, standard deviation = 1), and β was the standardized effect size.

^ep < 0.05 before Bonferroni correction.

^fp < 0.05 after Bonferroni correction (the p value cutoff for Bonferroni correction is p = 0.05/23 outcomes = 0.0022).

The sensitivity analysis additionally adjusting for depression assessed contemporaneously with marital transitions yielded similar results, though some associations became slightly weaker (Supplementary Table S12). The sensitivity analyses on prevalent marital status (comparing the divorced/separated versus the currently married) yielded stronger associations with some outcomes, especially when prior marital status was not controlled for (Supplementary Tables S7 and S8).

Sensitivity analyses for potential unmeasured confounding

To evaluate robustness of the observed associations to potential unmeasured confounding, we calculated E-values (Table 4) [39]. There was evidence suggesting that the associations of incident marriage with several indicators of psychological wellbeing and psychological distress, mortality and cardiovascular diseases, as well as the associations of incident divorce/separation with social integration, depression and loneliness, were at least moderately robust to unmeasured confounding. For instance, to fully explain away the observed association between incident marriage and mortality, an unmeasured confounder associated with both marriage and lower likelihood of mortality by risk ratios of 2.45-fold each, above and beyond the measured confounders, could suffice, but weaker joint confounder associations could not; unmeasured confounding association by risk ratios of 1.67-fold each for marriage and mortality could suffice to shift the confidence interval to include the null value, but weaker joint confounder associations could not. In contrast, the E-values for the confidence intervals for the associations with most health behavior outcomes were 1.00; these associations were not robust to unmeasured confounding.

Table 4.

Robustness to unmeasured confounding (E-values^a) for the associations between marital transitions and subsequent health and well-being (The Nurses' Health Study II 1993 to 2015/2017 questionnaire).

	Incident marriage		Incident divorce/separation
Health and well-being outcomes	Effect estimateb	CI limitc	Effect estimateb	CI limitc
All-cause mortality	2.45	1.67	1.67	1.16
Number of physical health problems	1.20	1.00	1.23	1.07
Overweight/obesity	1.11	1.00	1.11	1.00
Type 2 diabetes	1.32	1.00	1.40	1.00
Coronary heart disease	2.35	1.63	1.76	1.31
Stroke	2.50	1.74	1.71	1.28
Cancer	1.29	1.00	1.11	1.00
Positive affect	1.42	1.28	1.20	1.00
Optimism	1.47	1.32	1.11	1.00
Purpose in life	1.42	1.27	1.20	1.00
Hopefulness	1.39	1.24	1.11	1.00
Social integration	1.11	1.00	1.56	1.45
Emotional support	1.30	1.10	1.23	1.05
Depressive symptoms	1.42	1.28	1.33	1.20
Depression diagnosis	1.32	1.00	1.76	1.43
Anxiety symptoms	1.33	1.16	1.16	1.00
Loneliness	1.42	1.27	1.45	1.33
Preventive healthcare use	1.11	1.00	1.16	1.00
Heavy drinking	1.57	1.00	1.31	1.00
Current smoking	2.17	1.39	1.71	1.24
Recommended level of physical activity	1.11	1.00	1.21	1.00
High dietary quality	1.16	1.00	1.11	1.00
Short sleep duration	1.36	1.00	1.21	1.00

^aSee VanderWeele and Ding (ref no.39) for the formula for calculating E-values.

^bThe E-values for effect estimates are the minimum strength of association on the risk ratio scale that an unmeasured confounder would need to have with both the exposure and the outcome to fully explain away the observed exposure-outcome association, conditional on the measured covariates. For example, an unmeasured confounder would need to be associated with both incident marriage and mortality by risk ratios of 2.45 each, above and beyond the measured covariates, to fully explain away the observed association between marriage and mortality.

^cThe E-values for the limit of the 95% confidence interval (CI) closest to the null denote the minimum strength of association on the risk ratio scale that an unmeasured confounder would need to have with both the exposure and the outcome to shift the confidence interval to include the null value, conditional on the measured covariates. For example, an unmeasured confounder would need to be associated with both marriage and mortality by 1.67-fold each, above and beyond the measured covariates, to shift the lower limit of the confidence interval for the observed association between marriage and mortality.

Discussion

Despite the recent decline in marriage rates and the weakening in traditional marriage, marriage arguably remains one of the strongest social ties that shape individuals' lives [40]. With increasing diversity in marital relationships and changing social norms around family formation, especially among younger cohorts, it is important to continue understanding the long-term impacts of marital transitions and approaches for enhancing individuals' wellbeing across all marital status [41].

Congruent with prior evidence, this study, using data from a large sample of female nurses, suggests that entering marriage is associated with better subsequent health and wellbeing outcomes, but extends the literature by adding evidence on incident first-time marriage (for which confounding by prior marital history is less likely a concern) in relation to a wide range of outcomes. Likewise, this study also adds to the evidence that incident marital dissolution is related to multiple adverse health outcomes. Specifically, in this study incident marriage was associated with a 35% lower risk of all-cause mortality whereas incident divorce/separate was possibly related to a 19% higher mortality. These effect sizes were, however, smaller than some prior studies on prevalent marital status [9,42], which may be in part due to the differences in study design or the relatively low rate of mortality and illness in this middle-aged sample. Similar to other prior findings among women [8,12,43,44], marital dissolution was related to lower social integration, greater depression and loneliness, whereas entering marriage is associated with lower risks of cardiovascular diseases, and greater psychological wellbeing and less psychological distress on multiple indicators. Somewhat contrary to prior evidence [12,13], however, there was limited evidence that marital transition was associated with health behaviors (perhaps except for smoking) in this sample. There are a few possible reasons for the difference in estimates. For example, our analytic approach of focusing on incident marital transition and accounting for the initial levels of various health indicators may partly explain the weaker associations than prior findings on prevalent marital status. The sensitivity analyses on prevalent marital status in this study, especially when prior marital status was not controlled for, indeed yielded stronger associations than analyses on incident marital transitions. Considering prior evidence on marital transitions, many prior studies examined the effects of remarriage or remaining married [13,24,[45], [46], [47]], whereas we examined incident first marriage in this study which addressed a different research question. Furthermore, participants in this study were all healthcare workers who might lead healthier lifestyles than the general population. Therefore, the associations of marital transitions with health behaviors and some physical illness outcomes, if any, may be weaker in this sample. Nevertheless, even within this sample, and even examining incident marriage (that includes the effects of possibly subsequently becoming divorced), we found associations with numerous health and wellbeing outcomes 20 years later. Taken together, this study expands the literature in several ways including the focus on incident first-time marriage, control for initial health status, considering a wide range of outcomes (including some outcomes that have been less often explored in the marriage literature such as optimism and purpose in life), use of a large sample with a long-term follow-up, and use of sensitivity analyses to evaluate robustness of the results to unmeasured confounding.

This study provides evidence for both the health protection (e.g., marriage leads to health) and the health selection theories of marriage (e.g., the healthier are more likely to marry) [12]. Specifically, our findings suggest that marital transitions were associated with subsequent changes in multiple hypothesized pathways (e.g., social support, mental distress) leading to health and longevity, supporting the health protection theory [17]. According to the theory, the presence of a spouse/partner provides a source of social support, helping one to cope with stress and offering a buffer against negative emotions in the face of challenging circumstances. Commitment to a marital relationship also provides a form of social control that encourages individuals' engagement in healthier lifestyles [12]. Moreover, the nature of the social support offered by marriage is arguably different in kind than that offered in other relationships: spouses are bound to one another both by lifelong, often sacred, vows, and by their mutual commitment to their children [48]. In consequence, marriage is often shaped by the deepest intimacy and trust, an expectation formalized in spouses' traditional protection from testifying against one another in court [49]. Marital dissolution, by contrast, entails the loss of this major and distinctive source of social support, leading to increased loneliness, mental distress and engagement in unfavorable coping strategies [12]. On the other hand, the descriptive analysis in this study also suggests that healthier individuals were more likely to get married and were less likely to become divorce/separated, providing some support for the health selection perspective.

This study is subject to several limitations. First, participants in this sample were on average older than the national median age of first marriage and first divorce among women at the time of study [29], although we sought to get closer by using marital transition data from the baseline years only. However, the present sample may not be ideal for studying more generalizable first-time marital transitions concerning the general populations. Second, some new forms of intimate and marital relationships and living arrangements such as cohabitation and same-sex marriage have become increasingly common since the time when marital status was assessed in this study. The social norms and expectations around marital union and dissolution have also been evolving [41]. Future outcome-wide studies on such new forms of marital relationships using a life course approach would help further understand the processes and variation in the ways in which marriage and relationship type shapes health in general and how the dynamics evolve over time. Next, while the main analysis on incident first-time marriage was restricted to participants who were initially never married to reduce concerns about confounding by prior marital history, there may still be residual confounding by factors such as attitudes towards family formation and prior cohabitation history. In addition, while this study adjusted for a wide range of sociodemographic covariates, we were unable to control for potential confounders such as religious service attendance due to lack of data at the time marriage entry was assessed. However, we adjusted for numerous social, demographic, and lifestyle covariates including some that are fairly strongly correlated with religious beliefs such as smoking and heavy drinking [50]. Sensitivity analyses using “E-values” also suggested that the associations with multiple outcomes were at least moderately robust to unmeasured confounding. Finally, participants in this study were all female nurses and were mostly white; as such, the results may not be applicable to other populations, since some prior work suggests that the marriage-health associations may vary by gender, race and SES [51]. Moreover, nurses generally lead healthier lifestyles as compared to the general population. Therefore, it would be helpful to replicate this study using a more diverse and representative sample.

To conclude, this study provides evidence that entering first-time marriage in earlier adulthood is associated with subsequently lower risks of mortality and cardiovascular diseases, and greater psychosocial and mental wellbeing, whereas marital dissolution is related to lower psychosocial wellbeing and greater psychological distress among women, with these associations persisting well into mid- to late- life. Our study also addresses critiques of prior research since it addresses incident marriage (including the effects of the decision to marry on subsequent divorce) rather than only “prevalent” marriage, and reduces concern about residual confounding by prior marital history by focusing on first marriage. Some policy makers and community organizations have sought to promote marriage through approaches such as launching public campaigns about the values of marriage and providing marriage counseling services [52], although a number of these practices were not particularly evidence-based and effectiveness of these programs were not always as desired and require refinement [7]. Online marital counseling programs may also provide cost-effective evidence-based approaches to help address marital problems before they become so severe that divorce seems the only option. Some of these online programs have now been successfully evaluated in randomized trials [53]. Whereas marriage remains one of the strongest social ties for many individuals, there are certainly many other meaningful social relationships that contribute to wellbeing and the dynamics around social relationships continue to evolve over time [54], and these topics likewise merit further attention. While marriage is clearly a powerful social bond, all people need social relationships and community support.

Author contribution

All authors contributed significantly to this study. Drs. Chen, Mathur, Case, and VanderWeele conceptualized and designed the study. Dr. Chen had full access to the data in the study, and takes responsibility for the integrity of the data and accuracy of the data analysis. Dr. Chen drafted the manuscript. Drs. Mathur, Case, and VanderWeele provided critical revisions, and approved the submitted manuscript. Dr. VanderWeele obtained funding for this study.

Declaration of Competing Interest

None.

Appendix A. Supplementary data

Supplementary Text, supplementary figure, and supplementary tables