The Relationship Between Subjective Wellbeing and Mortality within Discordant Twin Pairs from Two Independent Samples

Abstract

Prior research has shown robust associations between greater subjective wellbeing (SWB) and reduced mortality. Whether this observed association is causal in nature, or due instead to confounding genetic or environmental factors affecting both SWB and mortality, is not well understood. We used a combined sample of 6,802 twins drawn from two cohorts: the Longitudinal Study of Middle-Aged Danish Twins (MADT; N=2,815, baseline age between 45 and 69 years, M=56.8 years, SD=6.4 years) and the Longitudinal Study of Aging Danish Twins (LSADT; N=3,987, baseline age between 70 and 97 years, M=76.6 years, SD=4.9 years). The relationship between SWB, encompassing measures of life satisfaction and affect, and all-cause mortality was evaluated using survival analyses at both the individual level and within twin pairs. Twin difference analyses were completed within 1,053 monozygotic (MZ) twin pairs and 1,143 dizygotic (DZ) twin pairs to control for genetic and shared environmental confounding. As expected, the individual level results showed that higher levels of SWB were associated with reduced mortality (affect hazard ratio (HR) = 0.90, 95% confidence interval (CI) = [0.87–0.94]; life satisfaction HR = 0.88, 95% CI = [0.84–0.92]). The effect of SWB on reduced mortality remained significant within both MZ and DZ pairs, suggesting that the association is independent of genetic and non-shared environmental confounding factors. These findings, which generalized across both younger (MADT) and older (LSADT) cohorts of adults, remained significant when accounting for demographic factors, physical health, and cognitive functioning.

Subjective wellbeing (SWB) can be broadly conceptualized as a range of subjective life evaluations and positive life feelings. The terms wellbeing, life satisfaction, quality of life, and happiness are often used interchangeably. There is evidence to suggest that this is appropriate, as many aspects of SWB share an underlying commonality. For example, most measures of SWB are moderately to highly intercorrelated, with estimates ranging from r=0.35 to r=0.77 (Bartels & Boomsma, 2009; Krueger & Schkade, 2008). Additionally, aspects of SWB, excluding happiness, appear to load on the same common factor (Kapteyn, Lee, Tassot, Vonkova, & Zamarro, 2015; Linley, Maltby, Wood, Osborne, & Hurling, 2009). Happiness may represent a separate dimension of SWB primarily due to reflecting a distinct, more temporally-limited mood state than other, more stable, aspects of SWB (Moons, Budts, & De Geest, 2006).

Furthermore, SWB is moderately stable and heritable. In early research on SWB, test-retest correlations between r=0.39 and r=0.53 were found over a two-year period (Atkinson, 1982). More recently, the stability coefficient of SWB was estimated at 0.56 over one year, declining to 0.24 over a 16-year interval (Fujita & Diener, 2005). Additive genetic factors largely account for this stability (Nes, Røysamb, Tambs, Harris, & Reichborn-Kjennerud, 2006). For example, a meta-analysis of different measures of SWB reported a moderate weighted heritability estimate of 0.36 (Bartels, 2015). Moreover, phenotypic correlations between different measures of SWB appear largely due to common underlying genes as well (Bartels & Boomsma, 2009). This indicates that studies based on different measures of SWB can be compared, as they measure biologically overlapping constructs.

Changes in SWB may be affected by developmental and cultural factors. As individuals age, a U-shaped pattern of SWB is commonly observed, with the highest ratings of SWB occurring in early adulthood and later life, and lowest ratings occurring in midlife (Blanchflower & Oswald, 2008; Shankar, Rafnsson, & Steptoe, 2015; Steptoe, Deaton, & Stone, 2015; Van Landeghem, 2012). Other studies find relatively little change in average SWB over the lifespan (Frijters & Beatton, 2012), including one using the current sample (Vestergaard et al., 2015). While the U-shaped pattern may hold in high-income countries, other countries show different developmental patterns. For example, in less affluent and developing countries, mean SWB often remains flat, or decreases, across age (Steptoe et al., 2015).

Of practical importance among aging populations there is a robust association between greater SWB and increased longevity (Diener & Chan, 2011; Steptoe et al., 2015). A meta-analysis investigated the protective association of SWB with reduced mortality, including 35 studies each of healthy and disease populations (Chida & Steptoe, 2008). Overall, increased SWB was associated with reduced mortality. This association was significant in both healthy populations (combined hazard ratio [HR] of 0.82; 95% confidence interval [CI] of 0.76–0.89, p<0.001), and in disease populations (combined HR of 0.98; 95% CI of 0.95–1.00, p=0.03).

While the association between SWB and mortality appears to be robust, a major limitation of the existing literature is the lack of causally informative research designs. Most studies are based on observational data, in which it is not possible to infer a causal association between increased SWB and reduced mortality. Rather than being causal, it is possible that confounding variables explain the link between SWB and mortality. For example, both SWB and lifespan are moderately heritable, with heritability estimates ranging from 0.32 to 0.36 for SWB (Bartels, 2015) and from 0.20 and 0.30 for lifespan (Herskind et al., 1996; Skytthe et al., 2003). Thus, common genetic influences could account for the association of SWB with length of life.

The etiology underlying this association may be addressed more directly through the use of discordant twin pairs (Sadler, Miller, Christensen, & McGue, 2011). Accordingly, Sadler et al. (2011) used a co-twin control design to test whether the relationship between SWB and mortality was consistent with a causal interpretation. Co-twin control models make use of discordant monozygotic (MZ) twin pairs to determine whether an observed association is consistent with a causal effect of an exposure on an outcome (McGue, Osler, & Christensen, 2010). If MZ twins who differ on an exposure (in this case, in degree of SWB) also differ on an outcome (i.e., longevity), we can be confident that this association cannot be due to the genetic and rearing environmental factors they share (e.g. socioeconomic status). Thus, if an MZ twin with higher SWB lives longer than his or her co-twin with lower SWB, this suggests a possible causal influence of SWB on mortality because we are inherently controlling for all genetic and shared environmental factors within a twin pair (though not for non-shared experiences, unless they are specifically measured). If differences within dizygotic (DZ) but not MZ pairs are found, this suggests genetic factors influence both SWB and mortality, as DZ pairs share only 50% of their segregating genetic material. Applying these models, Sadler et al. (2011) found that twin differences in SWB predicted differential mortality within discordant pairs (i.e., through significant within-twin pair effects), consistent with a causal relationship between SWB and mortality.

The current study aims to replicate Sadler et al.’s findings using a larger number of informative twin pairs and incorporating an additional cohort of twins, thereby increasing the total sample size substantially. Because Sadler et al.’s findings were published in 2011, more twin pairs now exist in which at least one member of the pair has died, resulting in over 600 additional informative pairs for survival analysis. Furthermore, the addition of a second, younger, large cohort of twins completely independent of the sample used by Sadler et al. allows us to see whether previous findings replicate in a younger sample covering a different birth cohort, as well as to combine samples, increasing the power to detect within-pair effects. Lastly, because a non-shared experience (e.g., a health condition affecting one twin only) might partially account for within-pair differences in both SWB and mortality (McGue et al., 2010), a measure of physical health functioning was added to disentangle the effects of SWB on mortality from the those due to twin differences in physical health status (McFadden et al., 2009). Consistent with previous findings (Sadler et al., 2011), we hypothesize that greater SWB will be associated with reduced mortality at both the individual level (i.e., when treating twins as individuals) and within discordant twin pairs. We expect that findings in the combined sample will generalize across both cohorts as well.

Method

Sample

Participants included twins from two longitudinal Danish samples, both of which were ascertained through the population-based Danish Twin Registry (Skytthe et al., 2013). The first sample was from the Longitudinal Study of Aging Danish twins (LSADT), a cohort-sequential study of Danish twins aged 75 years and older (born prior to 1920) that began in 1995. The LSADT cohort was followed at 2-year intervals through 2003. At each assessment, additional twins who had aged into the study’s age range were added. Over the course of the study, the minimum age threshold for participation was decreased until it reached age 70. Twins were invited to participate in LSADT even when their co-twin was deceased. As a result, 1,152 complete pairs and 2,427 individual twins completed an intake assessment. This assessment was conducted in participants’ homes by trained interviewers and included demographics, medical and psychological health, and physical and cognitive functioning. Members of each twin pair were interviewed separately by different interviewers. Additional details on LSADT ascertainment and assessment, including an analysis of non-participants, can be found in Christensen et al. (1999) and McGue and Christensen (2007).

The second sample was drawn from the Middle-Aged Danish Twins (MADT) study (Skytthe et al., 2013). The MADT sample consisted of 120 intact twin pairs (i.e., with both members alive and living in Denmark at intake), randomly sampled from each of 22 consecutive birth years (1931–1952) in the Danish Twin Registry, and equally divided by sex and zygosity. A total of 4,314 participants (83% of the 5,190 surviving twins), including both members from 1,884 pairs, completed the MADT intake assessment in 1998. The protocol and assessment used in MADT closely paralleled that used in LSADT. Research for both cohorts was approved by the Danish Research Ethics Committee system.

In both samples, zygosity was determined by a self-report questionnaire validated by genetic testing, with error rates less than 5% (Hauge, 1981). Information on survival status through January 1, 2015 was obtained from the Danish Central Population Register, which is continuously updated (Pedersen, Gøtzsche, Møller, & Mortensen, 2006). Participants in the current study were included if they had complete data on an affect scale from their intake assessment and had survived at least 18 months beyond the intake assessment. Minimum survival beyond intake was incorporated to ensure that the measures of SWB were not influenced by declining health before death. As we were particularly interested in within-twin pair comparisons, we restricted the sample to same-sex twin pairs. This resulted in a final intake sample of 3,987 twins from LSADT and 2,815 twins from MADT, for a combined sample of 6,802 participants. Only the 4,384 individuals from 2,192 complete pairs (1,051 MZ pairs; 1,141 DZ pairs) were used in the within-pair analyses, as these models incorporate the twin-pair mean. For the purposes of survival analysis, the most informative twin pairs are ones in which one or both twins have died. While this is the case, we did not remove pairs in which both twins were currently alive because, while less informative overall, they provide some information regarding the survival probability. There were 737 twin pairs in which both twins had died (314 MZ pairs; 423 DZ pairs), 507 pairs in which on twin had died (226 MZ pairs; 281 DZ pairs), and 948 pairs in which neither twin had died (511 MZ pairs, 437 DZ pairs).

Measures

Subjective wellbeing (SWB) was assessed with items measuring life satisfaction and affect, which are shown in Table 1. They were adapted from the Cambridge Mental Disorders in the Elderly Examination (CAMDEX) depression section (Roth et al., 1986). Life satisfaction was indexed by item 3: ‘Are you happy and satisfied with your life at present?’, then scored on a 5-point scale. Affect was measured using all nine items, which were derived from a factor analysis of the 21 CAMDEX depression section items (McGue & Christensen, 1997). Two items (items 6 and 8) from the Affect scale were scored dichotomously as 1 = No and 2 = Yes. All other items (except item 3, described above) were scored as 1 = No, 2 = Sometimes, and 3 = Most of the time. Internal consistency reliability estimates for the Affect scale were 0.76 and 0.81 in male and female samples, respectively. Items were reverse-coded and summed so that a higher score corresponded to a higher level of SWB. The life satisfaction and affect measures were highly correlated (r=0.72).

Table 1.

Affect Scale Items (Adapted from the CAMDEX Depression Section)

How often do you feel happy? Do you feel sad, depressed, or miserable? Are you happy and satisfied with your life at present? Do you feel lonely lately? How do you feel about your future? Do you sometimes feel that life is not worth living? Do you feel tense and do you worry more than usual? Do you consider yourself a nervous person? Do you feel worthless?

Wording is a paraphrase of the actual interview item. CAMDEX = Cambridge Mental Disorders in the Elderly Examination.

Covariates were included based on previous research demonstrating their association with SWB and mortality (Inglehart, 2002; Steptoe et al., 2015). For example, base (or ‘unadjusted’) models controlled for age at the intake assessment, sex, and cohort (LSADT vs. MADT). Adjusted models additionally controlled for measures of cognitive functioning and health (i.e., number of current medications), included in the earlier analysis of the relationship between SWB and mortality in the LSADT sample only (Sadler et al., 2011). Cognitive functioning was indexed by the Mini-Mental State Exam (MMSE) in LSADT and through a cognitive composite score in MADT (McGue & Christensen, 2001; Tombaugh & McIntyre, 1992). Number of medications each participant was currently taking was assessed, with a higher number of medications indicating poorer physical health. While the previous LSADT study by Sadler et al. (2011) also included number of reported illnesses as a covariate, we were not able to do so in the analysis of the combined samples, because this variable was not assessed in the same way across cohorts.

Fully adjusted models controlled for all covariates described above, as well as for an additional measure of physical health functioning. Due to the large discrepancy in ages between cohorts, measures of physical health functioning were necessarily different in each cohort. In the LSADT cohort, it was assessed using a 9-item scale with sample items including: “Can you run 400 meters without reset?” and “Are you able to climb two flights of stairs?” (Christensen et al., 2000). In the MADT cohort, a measure of physical activity was used instead, as most participants in the considerably younger MADT sample did not yet show the same physical functioning deficits as LSADT participants. Physical health functioning in this sample was assessed using a 5-item scale with sample items including: “How often do you run, work out, do aerobics?” and “How often do you cycle at least 3 km?” (McGue, Skytthe, & Christensen, 2014).

The unadjusted, adjusted, and fully adjusted results are reported separately in order to provide comparison with the results from Sadler et al. (2011). The current study builds upon this previous analysis of SWB and mortality by incorporating a larger number of informative twin pairs from the LSADT sample (an increase of more than 600 twin pairs), incorporating a second large, independent sample to the analysis (MADT), and including an additional covariate thought to confound the relationship between SWB and mortality in physical health functioning.

Survival Analyses

To ease interpretation, measures of affect, life satisfaction, and cognitive and physical functioning were standardized prior to analysis (i.e., mean of 0; standard deviation of 1). Data were then analyzed using gamma frailty survival models (Sjölander, Lichtenstein, Larsson, & Pawitan, 2013). Survival analysis, which estimates a hazard function, is appropriate when the outcome of interest is time to an event, in this case survival time. The hazard function at time t_ij is the product of a baseline hazard and the multiplicative effect of an exposure variable. Survival analysis can be extended to clustered (twin) data by incorporating a baseline hazard that is shared within a cluster. Frailty survival models further decompose the cluster-specific baseline hazard into a baseline hazard term and a cluster-specific frailty term. Gamma frailty models used in the current analysis assume that this frailty term follows a gamma distribution with a mean of 1 and unknown variance. Survival was measured starting 18 months post-intake, with censoring occurring in January of 2015. As defined above, unadjusted base models included age at intake, sex, and cohort as covariates. Adjusted models included the existing covariates, as well as number of medications and cognitive functioning. Fully adjusted models include all previous covariates with the addition of physical health functioning.

Individual-level analyses included all singletons and same-sex twin pairs who participated at intake. Gamma frailty models were used and standard errors were adjusted to correct for the non-independence of observations (i.e., correlated data from twins in the same family) by incorporating a twin-pair specific frailty term in each model (Carlin, Gurrin, Sterne, Morley, & Dwyer, 2005). This frailty term implies a mixed-effects model to account for the correlated nature of the data in individual-level analyses.

Within-twin pair analyses included only pairs who were complete at intake and used gamma between-within models as an extension of discordant twin, or co-twin control (CTC) models (McGue et al., 2010), to survival data (Sjölander et al., 2013). CTC models decompose the overall (i.e., individual-level) exposure effect into that which is shared between twins in the pair (i.e., the between-pair term) and non-shared, or different, within pairs (i.e., the within-twin pair term). Because MZ twins share 100% of genetic and rearing environmental factors, the within-MZ twin pair term estimates the exposure effect controlling for all genetic and shared environmental factors. As DZ twins share 100% of rearing environmental factors and 50% of genetic factors, the within-DZ twin pair term estimates the exposure effect partially controlling for genetic factors, while fully controlling for shared environment. If the within-pair term in MZ twin pairs is significantly different from zero, this provides evidence consistent with a causal effect of SWB on mortality; failure to find a within-pair association would be inconsistent with causality. If the within-twin pair term in DZ twins is significantly different from zero while the within-twin pair term in MZ twins is not, this implies that the observed relationship is due primarily to genetic confounding.

All analyses were completed in R using the survival package (Therneau, 2015), and were repeated using stratified Cox proportional hazard models. The results were essentially unchanged, and are not reported here. Gamma between-within models were preferred in this case for their ability to distinguish between- and within-pair effects, smaller standard errors, fewer assumptions, and robustness to model misspecification (Sjölander et al., 2013).

Results

Descriptive Results

Table 2 contains descriptive statistics for the sample separated by cohort and zygosity. Within each cohort, we tested for mean differences between the individual-level sample, the MZ sample, and the DZ sample. In the MADT cohort, MZ twin pairs showed significantly less difference in ages at death within pairs than did DZ pairs, meaning age at death for MZ twins was more similar than for DZ twins (p=0.03), as would be expected given genetic influences on longevity (Hjelmborg et al., 2006). Similarly, in the LSADT cohort, MZ twin pairs also had a significantly less difference in ages at death within pairs than did DZ pairs (p<0.001).

Table 2.

Descriptive Statistics of MADT and LSADT Samples.

	MADT Sample			LSADT Sample
	Total MADT sample M (SD) or %	Complete DZ pairs M (SD) or %	Complete MZ pairs M (SD) or %	Total LSADT sample M (SD) or %	Complete DZ pairs M (SD) or %	Complete MZ pairs M (SD) or %
Sample size3	N=2,815	N=579	N=650	N=3,987	N=562	N=401
Affect1	15.7 (2.1)	15.7 (2.1)	15.8 (2.0)	16.7 (2.8)	16.8 (2.8)	16.8 (2.8)
Life satisfaction1	4.5 (0.7)	4.5 (0.7)	4.6 (0.7)	4.5 (0.8)	4.5 (0.8)	4.5 (0.8)
Mortality by January 2015 (%)	19	19	18	84	81	77
Difference in ages at death (years)	-	4.9 (3.5)	4.5 (3.5)	-	5.5 (3.9)	4.6 (3.7)
Age at intake (years)	56.8 (6.4)	56.5 (6.3)	56.8 (6.4)	76.6 (4.9)	75.4 (4.2)	75.5 (4.4)
Female (%)	49	48	48	59	64	62
Number of medications	1.5 (1.9)	1.5 (1.9)	1.5 (1.9)	2.6 (2.4)	2.5 (2.4)	2.6 (2.4)
Cognitive score2	0.0 (1.0)	0.01 (1.0)	0.02 (1.0)	0.0 (1.0)	0.1 (0.9)	0.13 (0.9)
Physical health functioning1	2.9 (0.5)	2.9 (0.5)	3.0 (0.5)	3.2 (0.7)	3.2 (0.8)	3.2 (0.7)

¹Variables were not yet standardized;

²Variable was standardized within each sample because different measures were used;

³Actual sample sizes differ slightly due to missing data (<1%) in covariate measures.

Comparing across the MADT and LSADT cohorts, there were significant mean differences for all variables, except life satisfaction and cognitive score (all ps < 0.05). This was expected for age-related variables, as the LSADT cohort was much older (i.e., ages 70–97 at intake, compared to ages 45–69 for MADT), and thus would be expected to need more medications and have higher mortality by 2015. We also find that the LSADT sample showed slightly higher ratings of affect (p<0.001), consistent with the age-related U-shaped patterns noted in the introduction. There were no significant differences in life satisfaction ratings across cohorts.

Intraclass correlations within MZ and DZ twin pairs for affect and life satisfaction are shown in Table 3. The correlations are displayed for the LSADT, MADT, and combined samples but were remarkably similar across the two samples, with no differences within MZ or DZ twin pairs. Combining both cohorts, MZ twins were significantly more similar than DZ twins on affect ratings but not more similar on ratings on life satisfaction. This finding supports the notion of a non-zero heritability of affect. The lack of finding for life satisfaction may be due to limited variability because it is a single-item measure.

Table 3.

Twin Correlations in the MADT, LSADT, and Combined Samples for Measures of Subjective Wellbeing (SWB).

SWB Measure	MADT Sample		LSADT Sample		Combined Sample
	MZ pairs	DZ pairs	MZ pairs	DZ pairs	MZ pairs	DZ pairs
	N=650	N=579	N=401	N=562	N=1,051	N=1,141
Affect	0.24 (0.17, 0.31)	0.14 (0.06, 0.22)	0.35 (0.25, 0.42)	0.12 (0.04, 0.20)	0.28 (0.22, 0.33)	0.13 (0.07, 0.19)
Life Satisfaction	0.10 (0.02, 0.17)	0.10 (0.02, 0.18)	0.14 (0.05, 0.24)	0.08 (−0.01, 0.16)	0.12 (0.06, 0.18)	0.09 (0.03, 0.15)

Intraclass correlation estimates (95% confidence intervals) within MZ and DZ twin pairs.

Survival model results

Results of survival analyses for the combined sample are shown in Table 4. Unadjusted, adjusted, and fully adjusted results are reported separately to facilitate comparison with Sadler et al. (2011). Hazard ratios (HRs), along with their 95% confidence intervals and p-values, are reported. The HRs are interpreted as the multiplicative effect on mortality risk associated with a one standard deviation (SD) increase in exposure. Thus, an HR less than 1 indicates a protective effect of exposure on mortality risk (e.g., higher positive affect = lower mortality risk). Survival models with the MADT and LSADT cohorts were also fit separately. Hazard ratios (HRs) and p-values for the MADT cohort are listed in Supplemental Table 1; those for the LSADT cohort are in Supplemental Table 2.

Table 4.

Hazard Ratios for Mortality Risk Associated with Higher Levels of Subjective Wellbeing (SWB): Individual-Level and Within-Pair Difference Effects of Life Satisfaction and Affect

	Individual-level analyses		Within-pair analyses
	All twins		DZ twins		MZ twins
	HR (95% CI)	p-value	HR (95% CI)	p-value	HR (95% CI)	p-value
Affect	N=6,802		N=2,283		N=2,102
Affect	0.77 (0.74, 0.80)	<0.0001	0.76 (0.69, 0.83)	<0.0001	0.77 (0.68, 0.87)	<0.0001
Age at intake	0.89 (0.89, 0.90)	<0.0001	0.91 (0.89, 0.93)	<0.0001	0.91 (0.90, 0.93)	<0.0001
Sex (female)	0.56 (0.51, 0.61)	<0.0001	0.64 (0.55, 0.74)	<0.0001	0.68 (0.56, 0.82)	<0.0001
Cohort (MADT)	0.68 (0.57, 0.82)	<0.0001	0.72 (0.54, 0.98)	0.034	0.98 (0.70, 1.39)	0.927
Affect adjusted	N=6,674		N=2,253		N=2,066
Affect	0.85 (0.81, 0.88)	<0.0001	0.82 (0.75, 0.90)	<0.0001	0.82 (0.72, 0.92)	0.001
Age at intake	0.88 (0.87, 0.89)	<0.0001	0.90 (0.88, 0.91)	<0.0001	0.90 (0.89, 0.92)	<0.0001
Sex (female)	0.53 (0.49, 0.58)	<0.0001	0.61 (0.52, 0.71)	<0.0001	0.65 (0.54, 0.79)	<0.0001
Cohort (MADT)	0.60 (0.50, 0.72)	<0.0001	0.65 (0.48, 0.89)	0.008	0.88 (0.61, 1.26)	0.484
Number of medications	1.13 (1.11, 1.15)	<0.0001	1.13 (1.10, 1.16)	<0.0001	1.11 (1.07, 1.16)	<0.0001
Cognitive score	0.82 (0.79, 0.85)	<0.0001	0.82 (0.76, 0.88)	<0.0001	0.80 (0.73, 0.89)	<0.0001
Affect fully adjusted	N=6,669		N=2,233		N=2,065
Affect	0.90 (0.87, 0.94)	<0.0001	0.88 (0.79, 0.96)	0.007	0.85 (0.75, 0.96)	0.011
Age at intake	0.86 (0.86, 0.87)	<0.0001	0.88 (0.87, 0.90)	<0.0001	0.89 (0.87, 0.91)	<0.0001
Sex (female)	0.51 (0.47, 0.56)	<0.0001	0.59 (0.51, 0.69)	<0.0001	0.63 (0.53, 0.76)	<0.0001
Cohort (MADT)	0.44 (0.37, 0.54)	<0.0001	0.48 (0.35, 0.66)	<0.0001	0.70 (0.48, 1.00)	0.050
Number of medications	1.09 (1.07, 1.11)	<0.0001	1.09 (1.06, 1.12)	<0.0001	1.07 (1.03, 1.11)	<0.001
Cognitive score	0.88 (0.84, 0.91)	<0.0001	0.87 (0.81, 0.94)	<0.001	0.84 (0.76, 0.93)	0.001
Physical health rating	0.73 (0.70, 0.77)	<0.0001	0.73 (0.68, 0.79)	<0.0001	0.72 (0.65, 0.79)	<0.0001
Life satisfaction	N=6,800		N=2,282		N=2,101
Life satisfaction	0.75 (0.72, 0.78)	<0.0001	0.75 (0.68, 0.82)	<0.0001	0.78 (0.70, 0.87)	<0.0001
Age at intake	0.89 (0.89, 0.90)	<0.0001	0.91 (0.89, 0.93)	<0.0001	0.92 (0.90, 0.94)	<0.0001
Sex (female)	0.57 (0.53, 0.63)	<0.0001	0.65 (0.56, 0.76)	<0.0001	0.68 (0.57, 0.82)	<0.0001
Cohort (MADT)	0.71 (0.60, 0.85)	<0.001	0.73 (0.54, 0.98)	0.035	1.07 (0.76, 1.51)	0.708
Life satisfaction adjusted	N=6,673		N=2,235		N=2,065
Life satisfaction	0.82 (0.78, 0.85)	<0.0001	0.79 (0.72, 0.87)	<0.0001	0.83 (0.74, 0.92)	0.001
Age at intake	0.88 (0.87, 0.89)	<0.0001	0.90 (0.88, 0.91)	<0.0001	0.91 (0.89, 0.93)	<0.0001
Sex (female)	0.54 (0.49, 0.59)	<0.0001	0.61 (0.52, 0.71)	<0.0001	0.65 (0.54, 0.79)	<0.0001
Cohort (MADT)	0.61 (0.51, 0.73)	<0.0001	0.65 (0.47, 0.89)	0.006	0.92 (0.64, 1.32)	0.653
Number of medications	1.13 (1.11, 1.15)	<0.0001	1.13 (1.10, 1.16)	<0.0001	1.12 (1.08, 1.16)	<0.0001
Cognitive score	0.82 (0.79, 0.85)	<0.0001	0.82 (0.76, 0.88)	<0.0001	0.80 (0.72, 0.88)	<0.0001
Life satisfaction fully adjusted	N=6,668		N=2,233		N=2,064
Life satisfaction	0.88 (0.84, 0.92)	<0.0001	0.84 (0.76, 0.93)	<0.001	0.88 (0.79, 0.99)	0.026
Age at intake	0.86 (0.86, 0.87)	<0.0001	0.88 (0.87, 0.90)	<0.0001	0.89 (0.88, 0.91)	<0.0001
Sex (female)	0.51 (0.47, 0.56)	<0.0001	0.59 (0.51, 0.69)	<0.0001	0.64 (0.53, 0.77)	<0.0001
Cohort (MADT)	0.46 (0.38, 0.55)	<0.0001	0.48 (0.35, 0.67)	<0.0001	0.71 (0.49, 1.03)	0.068
Number of medications	1.09 (1.07, 1.11)	<0.0001	1.09 (1.06, 1.12)	<0.0001	1.07 (1.03, 1.11)	<0.001
Cognitive score	0.87 (0.84, 0.91)	<0.0001	0.87 (0.81, 0.94)	<0.001	0.84 (0.76, 0.92)	<0.001
Physical health rating	0.74 (0.71, 0.78)	<0.0001	0.74 (0.69, 0.81)	<0.0001	0.72 (0.65, 0.80)	<0.0001

Hazard ratios (95% confidence intervals) and p-values for both measures of SWB in the combined sample. Unadjusted models include age at intake and sex as covariates, adjusted models include age at intake, sex, number of medications, and cognitive scores as covariates, fully adjusted models additionally include physical health functioning. Sample sizes for within-pair analyses are not necessarily even because, while only complete pairs are used in estimating the effect of SWB, incomplete pairs add information to the effects of the covariates. The within-pair analyses using gamma between-within models are able to estimate the effects of the covariates like sex and age even though these are shared within a twin pair. These estimates are the effect of the covariate holding the twin pair mean and individual score on SWB constant, similar to the individual-level effect of these covariates.

Individual level results

For both affect and life satisfaction in the combined sample, overall (i.e., individual-level) effects in the unadjusted and adjusted models closely replicate those reported by Sadler et al. For all models, greater positive affect and life satisfaction were significantly associated with reduced mortality risk, as indicated by HRs significantly less than 1 (ps < 0.001). In the fully adjusted models (i.e., those adjusted for effects of the newly added rating of physical health functioning), a one SD increase in positive affect was still associated with a 10% reduction in mortality risk (HR = 0.90, 95% CI = [0.87–0.94]), while a one SD increase in life satisfaction was associated with a 12% reduction in mortality risk (HR = 0.88, 95% CI = [0.84–0.92]). Table 4 additionally includes HRs (and p-values) for all covariates in each model. For all models of the relationships of affect and life satisfaction to mortality, a younger age at intake, female gender, being in the younger (MADT) cohort, fewer medications, higher cognitive functioning, and better physical health were protective (HRs less than 1.0). Age at intake was reverse scaled to be consistent with logical protective effects of the covariates.

Within-pair results

The results within MZ and DZ twin pairs (Table 4) were consistent with individual-level results. For all models, the effects of affect and life satisfaction on mortality risk were significant within both MZ and DZ twin pairs, as indicated by HRs significantly less than 1 (ps < 0.03). That is, twins who differed in SWB differed in mortality risk as well. Additionally, the within MZ twin pair and within DZ twin pair effects were generally as strong as, or stronger than, the individual-level effect (Figure 1). For example, in the fully adjusted model within MZ pairs, a one SD increase in positive affect was associated with a 15% reduction in mortality risk (HR = 0.85, 95% CI = [0.75–0.96]), while a one SD increase in life satisfaction was associated with a 12% reduction in risk (HR = 0.88, 95% CI = [0.79–0.99]). The within-pair effects were not significantly different between MZ and DZ twins for affect or life satisfaction. Table 4 additionally includes the HRs (and p-values) for all covariates in the within-pair models. Similar to the individual-level results, the effects of affect and life satisfaction are attenuated slightly from the unadjusted to fully adjusted models, which include measures of physical and cognitive health. All effects, however, remain significant.

Figure 1.

Hazard ratios (HRs) for the individual-level, within dizygotic (DZ) twin pair, and within monozygotic (MZ) twin pair effects from the fully adjusted models for both Affect and Life Satisfaction on mortality risk in the combined sample. Fully adjusted models include age at intake, sex, number of medications, cognitive scores, and physical health functioning as covariates. Error bars denote the 95% confidence intervals.

Discussion

We found that subjective wellbeing (SWB) is associated with mortality risk independent of genetic and shared environmental factors. The current results are consistent with a causal relationship between increased SWB and decreased mortality risk. The strength of the co-twin control (or discordant twin) design used here is that it controls for all confounding effects that members of a twin pair share, even if specific shared confounders were not directly assessed (McGue et al., 2010). Co-twin control analyses cannot unequivocally establish causality, as unmeasured confounders the twins do not share (e.g., death of a spouse) may account partially for these associations, by affecting both SWB and mortality risk. Nevertheless, with shared genes and environment completely controlled, a potentially quasi-causal effect was still evident. These findings, in two large Danish twin samples, replicate and extend previous research by Sadler et al. (2011).

As a first step, a survival analysis at the individual-level showed a significant association between increased SWB and reduced mortality, consistent with previous research (Chida & Steptoe, 2008; Diener & Chan, 2011; Steptoe et al., 2015). Additional covariates of sex, age at intake assessment, number of medications, and both cognitive and physical functioning were included to further disentangle possible causal effects from confounding factors. Furthermore, within MZ twin pairs, the effect of twin differences in SWB on differential mortality risk remained significant (Figure 1). As MZ twin pairs share all genetic and shared environmental confounders, this association must be independent of these factors. In addition to the significant within MZ pair effects, the lack of differences in the within-pair effects between MZ and DZ twins is consistent with causal influence of SWB on mortality risk. We also know that the observed relationship is independent of the measured covariates that were included. Moreover, the pattern of results was very similar for both measures of SWB that were used, i.e., affect and life satisfaction. Results were also consistent across both MADT and LSADT samples. While the confidence intervals within each sample are much wider than those in the combined sample (due to reduced sample size), the point estimates are very similar. This suggests the effect of SWB on mortality risk is not age dependent during the latter half of the lifespan; that is, the same relationship is found in middle-aged as well as older adults.

While the current study closely replicates the findings by Sadler et al. (2011) with an additional cohort of participants (i.e., the MADT cohort), as well as including an additional physical functioning covariate, it is important to note that this is not necessarily a fully independent replication. Both studies contain an overlapping subsample of individuals. However, for within-twin pair survival analysis, twin pairs in which one, or both, of the twins have died provide much more information in understanding the relationship between SWB exposure and mortality. In this way, while the current study contains the same LSADT sample as the previous study, there are a larger number of informative twin pairs, along with a completely independent cohort (MADT) incorporated in the current analysis. Overall, the analysis using the LSADT sample is a partial independent replication, while the analysis using the MADT sample is a fully independent replication.

As there are currently no gold standard measures of SWB, it is possible that these results may not replicate if other measures are used. Most measures of SWB are moderately to highly correlated (Bartels & Boomsma, 2009) and load on the same common factor (Kapteyn et al., 2015), implying that they are all getting at the same underlying trait. Each measure of SWB, however, may be tapping into slight differences in what the underlying construct represents and thus may not show the same relationship with mortality risk. Additionally, one-item measures, like life satisfaction used here, may be inherently problematic (Diener, 1984), though the psychometric literature has found fairly high reliability and validity of single-item measures of wellbeing, which correlate highly with longer versions of wellbeing scales (McDowell, 2010). Single-item measures of life satisfaction, in particular, perform similarly to longer scales (Schimmack & Oishi, 2005).

The observed relationship between SWB and reduced mortality risk is independent of genetic and shared environmental confounders, as well as the effects of included covariates (e.g., non-shared differences in physical health), but there may be other non-shared confounders that explain the relationship between SWB and mortality that were not included here. For example, it may be the case that those with higher SWB are less likely to engage in behaviors such as problematic drinking and tobacco use or more likely to engage in positive social interactions and have stronger social support networks. These behaviors are associated with longevity and may also be correlated with SWB (Dear, Henderson, & Korten, 2002; Holt-Lunstad, Smith, & Layton, 2010; Kujala, Kaprio, & Koskenvuo, 2002; McGue & Christensen, 2007). There may be other variables that act in a similar fashion. For example, we were not able to include number of illnesses experienced in the combined analysis, as the existence of illnesses was not measured in the same way across cohorts. Results using only the LSADT sample indicate that number of illnesses is significantly related to mortality risk, holding other covariates and SWB constant. We re-ran the analysis in the LSADT sample with number of illnesses and found that the effect of SWB on mortality risk was still highly significant. This indicates that, although the combined analysis does not include this variable, we are confident that the results would remain significant if it were able to be included.

The possible mechanisms by which SWB could influence mortality are not yet well understood. It could be the case that increased SWB enhances immune, cardiovascular, and endocrine functioning, which in turn reduces mortality. There is evidence that SWB is linked with both cytokines (most robustly associated with interleukin-6) and C-reactive protein, which are both inflammatory markers (Blevins, Sagui, & Bennett, 2017; Steptoe, O’Donnell, Badrick, Kumari, & Marmot, 2008; Undén et al., 2007). Positive affect has also been linked to overall cortisol levels independent of psychological and lifestyle factors, indicating that SWB may be directly related to physiological functioning (Steptoe, Wardle, & Marmot, 2005). Whether these mechanisms account for the relationship between SWB and mortality requires more research.

There are some additional limitations to note. First, the current sample was largely ethnically homogenous, consisting of participants only from Denmark. These results may not necessarily generalize to other populations. Second, the outcome measure used was all-cause mortality, as information on specific cause of death was not currently available for these analyses. While the occurrence of accidental deaths is likely to be low, these cases are still included in the analysis. Third, co-twin control designs cannot rule out reverse causality. The longitudinal design of the current study, coupled with the removal of participants who died within 18 months of the SWB assessment, help to rule out reverse causality here.

Future research is needed to see if these findings replicate in more diverse populations and when using other measures of SWB. Given that the findings presented here suggest that increasing SWB may increase longevity, whether SWB can be altered through intervention should be explored further. There is some evidence that interventions may work to increase SWB, the vast majority of this research is based on short-term change (Armitage, 2016; Jung et al., 2010; Tomyn, Weinberg, & Cummins, 2015). An ideal intervention would target long-term change in SWB. Whether this is possible is not yet known. However, current evidence suggests that a focus on the mechanisms by which SWB impacts mortality, and what factors affect SWB over the lifespan should prove productive.