Subjective Well-Being Around the World: Trends and Predictors Across the Life Span

Translation

In each nation, bilingual speakers translated the survey into one or more widespread languages. Depending on the country, one of two methods was used. For some nations, two independent translations were completed, a third party resolved the differences, and a final professional back-translated the survey into the source language. For others, an initial translation was made, a second translator back-translated it, and a third party reviewed both versions and gave revisions when necessary. On the basis of an inductive approach, recent analyses of emotion terms revealed that the underlying dimensions were not substantially different across cultures (Tay, Diener, Drasgow, & Vermunt, 2011).

Subjective well-being

Subjective well-being is often conceived of as three different metrics: life satisfaction, positive affect, and negative affect (Diener, 1984). We had measurements of all three. Our life-satisfaction measure was the Cantril Self-Anchoring Striving Scale (Cantril, 1965), which asks respondents to evaluate their current life on a ladder scale with steps ranging from 0 (worst possible life) to 10 (best possible life). The scale has shown test–retest reliabilities (rs) of .70 over a 2-year period (Palmore & Kivett, 1977). Measures of positive and negative affect were based on items that asked whether respondents had experienced particular emotions the previous day. The emotion items were all dichotomous; the positive items were “happiness,” “enjoyment,” and “smile or laugh,” and the negative items were “worry,” “sadness,” “stress,” “fear,” and “depression.” ¹ Averaging these items for a participant yielded scores between 0 and 1, which were then rescaled from 0 to 10 for ease of interpretation and so they would be on the same scale as the life-satisfaction scores. These measures had internal consistency reliabilities (αs) of .72 and .71, respectively.

Subjective-well-being predictors

All participants were asked for their age. When the age provided was over 98, it was entered as missing. In addition to age, we examined four life variables. Marriage was measured by an item on marital status. The responses were coded into two categories (1 = married or in a domestic partnership, 0 = single/never been married, separated, divorced, or widowed), which are referred to as married and unmarried, respectively, for simplicity. For employment, individuals indicated their employment status, which we also coded dichotomously (0 = unemployed, 1 = employed full-time for an employer, employed full-time for self, employed part-time but want full-time, and employed part-time and do not want full-time).

Our measure of prosociality was constructed from three items from the Gallup survey. These all asked whether the respondent had performed a different prosocial behavior in the past month: (a) donating money to an organization, (b) volunteering time for an organization, and (c) helping a stranger. If respondents had performed at least one of these behaviors, they were coded as 1; they were coded as 0 otherwise. We kept this measure dichotomous because we were interested in subjective-well-being differences between participants who were active prosocially and those who were not, rather than the possible degrees of prosociality. Keeping the variable dichotomous also kept it consistent with our other measures and made interpreting the model parameters much easier, especially because we were examining many interactions. In an additional validation attempt for the current study, we found that this Gallup measure correlated (r = .45, p < .001) with the Self-Report Altruism Scale (Rushton, Chrisjohn, & Fekken, 1981; see the Supplemental Material). Finally, life meaning was assessed with a single dichotomous item that asked, “Do you feel your life has an important purpose or meaning?” (1 = yes, 0 = no). This measure correlated (r = .75, p < .001) with the three-item Meaning and Purpose subscale of the Comprehensive Inventory of Thriving (Su, Tay, & Diener, 2014).

Models

The data were structured such that individuals were nested within countries, and countries were nested within the world regions listed above. We therefore used three-level multilevel models for all of our analyses and fit them using the nlme package in the R programming environment (Pinheiro, Bates, DebRoy, Sarkar, & R Core Team, 2014). Because we wanted to model the possibility of different effects across regions (and expected different effects), we used a random-intercepts and random-slopes model. In every fitted model, the intercepts and slopes of the predictors were allowed to vary by nation and region. We controlled for gender and survey year, ² coded as dummy variables, but did not use any time-varying covariates (for the reasons why, see Glenn, 2009). These controls were not specified as random effects (i.e., their coefficients were not allowed to vary across nations or regions).

When analyzing the data, we initially fitted regular linear models but then realized that more flexibility was needed. First, when examining the relationship between subjective well-being and age, it was clear that curvilinear effects were possible (e.g., a negative slope that reduced in strength across age). Therefore, we added a quadratic term to the multilevel models (also allowed to vary across nations and regions) to potentially account for any curvilinear effects.

When looking at the effects of predictors, we realized that more modeling flexibility was needed here, too. We wanted to test whether the effects of certain predictors (e.g., marriage) changed throughout the life span (i.e., were moderated by age). This can be modeled via an interaction term between the predictor and age. However, a traditional product term allows only for the effect to be either increasing or decreasing with values of the moderator. In our context, however, it was possible for a variable such as marriage to have an effect that grew until middle age and then become weaker at older ages. This is an example of a curvilinear interaction (Li, 2018), in which the interaction does not have a simple “increase-or-decrease” form across values of the moderator. Therefore, when examining the effects of our four predictors, we always used curvilinear interaction models. These models are straightforward extensions of traditional interaction models. In addition to the usual product term, two terms are added: (a) the moderator (here, age) squared and (b) the product of this squared term with the other predictor (e.g., marriage). Again, these analyses were always in the context of a random-slopes and random-intercepts multilevel model so that the nature of this curvilinear interaction was allowed to change across nations and regions. All of our software code used for these analyses is available at the Open Science Framework at https://osf.io/pr68j/.

Effect size

We also discuss our results in terms of effect size because a statistically significant effect does not entail that it is practically significant (Fritz et al., 2012). The effect size for subjective well-being over the life span was simply the predicted mean difference between 20-year-olds (the baseline age) and every other age. For the other predictors, the effect size was simply the predicted mean difference between the group coded as 1 (i.e., married, employed, prosocial, having life meaning) and the group coded as 0 (i.e., unmarried, unemployed, nonprosocial, without life meaning). Because all of these predictors were dichotomous, the effect sizes can easily be interpreted; for example, when discussing the effect of marriage for life satisfaction, an effect size of −0.08 at age 20 and of 0.45 at age 50 means that participants who were married at age 20 had a −0.08 lower mean life satisfaction, compared with those who were unmarried, and those at age 50 had a 0.45 higher life-satisfaction mean.

At some point, an effect size becomes so small that it is truly trivial and lacks practical significance. For our Cantril ladder scale, respondents reported (and probably thought) in terms of the nearest whole scale point from 1 to 10. Therefore, it seemed that differences below 1.00 should be considered quite small. It is ambiguous when an effect size becomes trivial, and we did not apply a strict cutoff for this reason. However, our general heuristic was that effect sizes under 1 scale point were considered very small and became basically trivial when less than roughly 0.60 to 0.70. For positive affect and negative affect, we used the same heuristic because these two constructs are conceptually similar to life satisfaction and their degrees of these constructs were also mapped to the same scale from 0 to 10. For readers interested in standardized effect sizes, the three outcomes did have somewhat different dispersions (SDs = 2.28 for life satisfaction, 3.79 for positive affect, and 3.35 for negative affect).

Reporting

Reporting our results was a challenge because we were examining three subjective-well-being outcomes across 10 regions and five predictors. Furthermore, the curvilinear effects allowed the effects of age and its interaction with predictors to change continuously over the life span. This means that the slope and interactions could not be summarized by single coefficients and that we had a large degree of information to report. To manage this, we rely on graphics and high-level summaries. In the main text, we report coefficients only for the global trends. The regional coefficients are reported in the Supplemental Material. To show how the effect of each predictor changed across the life span, we provide interaction plots. These can be seen in Figures 1 through 3 as well as in Figures S1 through S16 in the Supplemental Material. Finally, although the predictor effects changed continuously over the life span, a good method for giving a sense of this change is to report the effect size at a few discrete points. We therefore report the predicted effect sizes at the ages of 20, 50, and 80 years, 30-year intervals that would roughly convey the nature of change across the entire life span.

OPEN IN VIEWER

Fig. 1.

Predicted life satisfaction across the life span, globally and individually for each region. Each graph shows the overall trend and the trend for 20-year-olds only (baseline). Practically significant effect-size differences between the two trend lines are color coded.

OPEN IN VIEWER

Fig. 2.

Predicted positive affect across the life span, globally and individually for each region. Each graph shows the overall trend and the trend for 20-year-olds only (baseline). Practically significant effect-size differences between the two trend lines are color coded.

OPEN IN VIEWER

Fig. 3.

Predicted negative affect across the life span, globally and individually for each region. Each graph shows the overall trend and the trend for 20-year-olds only (baseline). Practically significant effect-size differences between the two trend lines are color coded.

Supplemental issues: survey weighting, measurement validation, measurement invariance, and response styles

Finally, we examined a range of additional methodological issues, which are reported in the Supplemental Material. First, Gallup uses survey weights to correct for nonresponses (also called “sampling” or “probability” weights). A discussion of how these were used is provided. We also conducted a validity study to investigate the applicability of our Gallup measures. We found acceptable convergent validity. We then conducted measurement-invariance analyses and an analysis of the influence of responses styles across regions. We found that our multi-item measures were invariant and that response-style differences were not very significant across regions.

Life satisfaction

Globally, the trend for life satisfaction was a negative curve (linear term: b = −0.01, p < .001; quadratic term: b = 0.0002, p < .001). However, all changes were very slight; they always stayed within 0.48 of the starting life-satisfaction score at age 20. This is displayed in Figure 1. When looking at the individual regions, one can see that most had changes that were similarly small (also in Fig. 1). One exception was Eastern Europe, where the decreases to life satisfaction reached 1.00 at age 60 and were 1.21 by age 80.

Positive affect

Globally, positive affect had a negative slope similar to the slope for life satisfaction over the life span (linear term: b = −0.02, p < .001; quadratic term: b = 0.0002, p = .008). However, unlike the effects for life satisfaction, the effects for positive affect were more practically relevant (see Fig. 2). The reduction in positive affect was 0.80 at age 50 and grew to 1.32 at age 80. Most of the regions were similar in form and effect-size magnitude. One exception was Eastern Europe, where the decreases were again much larger (1.76 and 2.90 at ages 50 and 80, respectively). Four regions also seemed to have smaller decreases over the life span (never reaching 1.00): Anglo, Germanic, Confucian Asia, and Latin America. To test this, we fitted a second model with dummy variables that coded these two groups of regions (and their product terms). The significant interaction meant that the slopes of these four regions were significantly less negative than the rest (Groups × Age: b = 0.02, p = .002; Groups × Age ² : b = 0.00002, p = .89).

Negative affect

Globally, the curve across age for negative affect was statistically significant (linear term: b = −0.005, p = .004; quadratic term: b = 0.0001, p = .03), but the changes were all trivial (the maximum difference was a 0.28 increase at age 50 years). The results were slightly different for the regions individually (see Fig. 3). One group of regions was wealthier and had small decreases in negative affect: Anglo, Germanic, Nordic, and Confucian (the effect sizes ranged from 0.16 to 0.22 at age 50 years to 0.74 to 1.30 at age 80 years). The remaining six regions had small increases in negative affect: Latin Europe, Eastern Europe, Latin America, Southern Asia, Arab, and sub-Saharan Africa (effect sizes ranged from 0.42 to 0.74 at age 50 years and from 0.17 to 0.96 at age 80 years). Comparing these two groups of regions yielded a significant interaction in the expected direction (Groups × Age: b = −0.03, p < .001; Groups × Age ² : b = 0.00007, p = .61). However, the changes within regions were almost all very small.

Marriage

The first predictor we examined was marital status. Globally, for life satisfaction, there was a significant interaction between age and marriage; specifically, married people had slightly higher life satisfaction than unmarried people over the whole life span. However, the effect sizes across the life span were almost trivial; the largest difference was about 0.45 at age 50 years. This pattern was replicated in all of the individual regions; people who were married had higher life satisfaction, but these effects were all very small (never exceeding 0.76 in any region).

For positive affect, the same exact patterns were found. Globally and within each region, married individuals had higher positive affect across the whole life span, but these effects were also very small (the largest differences were roughly 0.61 at age 50).

Finally, we examined how marriage might be associated with reductions in negative affect. The results were the same as the other two measures of subjective well-being. Globally and regionally, unmarried people had slightly lower negative affect across the whole life span, but the effect sizes were all small (never greater than −0.67). For all three subjective-well-being outcomes, model coefficients for marriage and estimated effect sizes can be found in Tables S5 through S7, and plots of the Marriage × Age interactions are in Figures S4 through S6.

Employment

Next, we analyzed the effects of employment. There was a remarkable degree of consistency across the measures and regions. For all measures and regions, employed people had higher subjective well-being than unemployed people, with differences that usually peaked at around age 50 years and were lower at younger and older ages. The effect sizes were larger than for marriage and usually exceeded 1.00 by age 50 years. Globally, the effect sizes for life satisfaction at ages 20, 50, and 80 years were 0.43, 1.04, and 0.44, respectively. For the same ages, the effect sizes for positive affect were 0.32, 1.08, and 0.23, and for negative affect, they were −0.48, −1.36, and −0.41. There was some regional variation in these effect sizes, but all regions had qualitatively similar patterns. For all three subjective-well-being outcomes, model coefficients for employment and estimated effect sizes can be found in Tables S8 through S10, and plots of the Employment × Age interactions are in Figures S7 through S9.

Prosocial behavior

The next predictor we looked at was prosociality. Globally, people who engaged in prosocial behaviors had higher life satisfaction across the life span, but these effects were trivial in size, peaking at 0.56 by age 80. This functional form and effect-size magnitude was replicated for every region. For negative affect, there was no main effect of prosociality, and even though the interaction terms were both statistically significant, all effect sizes were trivial here as well (never exceeding −0.40). The only outcome with practical significance was positive affect. Globally, people who had been prosocial had higher positive affect over the entire life span; effect sizes grew continuously across age (0.52, 0.97, and 1.29 at ages 20, 50, and 80 years, respectively). The same functional form was present within each region with similar effect-size magnitudes. For all three subjective-well-being outcomes, model coefficients for prosociality and estimated effect sizes are in Tables S11 through S13, and plots of the Prosociality × Age interactions are in Figures S10 through S12.

Life meaning

Finally, the results of life meaning were very consistent. Across the entire life span and in every region of the world, individuals with life meaning had higher subjective well-being than those without life meaning. This was true for all three subjective-well-being measures. The effect sizes were practically significant and tended to grow across the life span. For ages 20, 50, and 80 years, the predicted mean differences were 0.64, 0.98, and 0.90 for global life satisfaction; 1.38, 1.74, and 1.94 for global positive affect; and −0.97, −1.20, −1.19 for global negative affect, respectively. The individual regions had similar effect sizes, with some variation. For all three subjective-well-being outcomes, model coefficients for life meaning and estimated effect sizes are in Tables S14 through S16, and plots of the Life Meaning × Age interactions are in Figures S13 through S15.