In the literature on the relationship between subjective well-being and age, one of the recurrent themes has been the shape of this relationship. While some contributions on this this topic have suggested that the relationship is essentially flat (for example, Kassenböhmer and Haisken-DeNew 2012), many have concluded that it is U-shaped (recent examples using data from a variety of different sources can be found in Cheng et al. 2017; Blanchflower 2020a and 2020b). This literature is now voluminous, and useful summaries appear in three chapters on the subject in the recent book edited by Mariano Rojas (Blanchflower and Oswald 2019; Clark 2019; Helliwell et al. 2019).

While there is probably majority agreement in the literature that the relationship is U-shaped, there is much more debate among scholars about why this is the case, and whether this pattern holds equally for all countries and all groups in society. Here, we contribute to the second of these discussions, and ask whether the U-shaped relationship between life satisfaction and age is flatter for individuals who are partnered. Partnership is obviously a relevant candidate as a variable in explanations of well-being, as being in a partnership requires substantial investments of time and resources and high levels of emotional involvement for the partners (Gove et al. 1990). Moreover, like other characteristics of family life that are important for well-being, being in a partnership is associated with greater financial security and less loneliness. Earlier work has found significant differences in well-being according to partnership status, although the size of these differences has been shown to vary across countries (Verbakel 2012). Moreover, an analysis of World Values Survey/European Values Survey data (Mikucka 2016) has indicated that the size of these differences has declined among men since the 1980s.

We are not the first to examine the question of whether being in a partnership flattens the U-shaped relationship between subjective well-being and age. Helliwell et al. (2019) introduced an interaction between marriage/cohabitation and age in subjective well-being regressions drawn from both the US Gallup Healthways data and the UK Annual Population Survey. They found a substantially flatter U-shaped relationship between well-being and age for partnered individuals based on their analysis of both of these cross-sectional datasets, which was only partly explained by control variables (sex, education, labour force status, log income and the number of children). A flatter U-shape for married people was also found in an analysis of the Gallup World Poll data by Jebb et al. (2020). The authors imposed a quadratic relationship between age and the Cantril ladder well-being question, though they argued that the difference between the two curves was not large.

In the current analysis, we use a very large-scale harmonised dataset covering 32 European countries, the EU-SILC, to first show that the U-shape holds in general in Europe. This is a very standard observation in the empirical literature. Moreover, as in Helliwell et al. (2019), we demonstrate that this U-shape is far flatter for partnered individuals, and that the age U-shape is mostly concentrated among those individuals who are not partnered. This pattern is not found for all of the subjective evaluations in the EU-SILC data. The protective role of being in a partnership is found to be particularly important for people’s overall life satisfaction, their satisfaction with their financial situation and their satisfaction with their personal relationships, but not at all for their job satisfaction or their satisfaction with time use. Thus, being partnered seems to protect people against the typical decline in well-being associated with middle age in at least certain domains of subjective well-being.

This first result is, however, less clear in the panel data. Fixed-effect analyses using the BHPS (UK), the SOEP (Germany) and the HILDA (Australia) reveal a U-shape in age for both partnered and non-partnered individuals, although the U-shape for the partnered individuals was again found to be somewhat flatter. We consider a number of explanations for this gap between the cross-sectional and the panel results. In particular, we do not believe that this discrepancy reflects the different countries in the cross-sectional and panel analyses, or differential mortality (or attrition in general) between the partnered and the non-partnered individuals. However, we do find some evidence of compositional effects that tend to flatten the U-shape in age for the partnered individuals in the cross-sectional data. Specifically, we observe that those individuals who experience a break-up in their forties and fifties are noticeably less satisfied with their lives than those who experience a break-up at younger or older ages. This composition or shift-share effect will mechanically increase the average well-being among those individuals who remain partnered in their forties and fifties, and will tend to produce a less U-shaped age profile.

The remainder of the paper is organised as follows. In Section 2, we describe the European data that we use and the hypotheses that we test, followed by the main cross-sectional results. In Section 3, we present our analysis of panel data in selected countries. Section 4 concludes.

Our main analysis of subjective well-being by age and gender is based on data from the EU Statistics on Income and Living Conditions (EU-SILC) survey conducted by Eurostat, which provides harmonised information on hundreds of thousands of respondents in more than 30 European countries. This annual survey is the reference source for comparative statistics on income distribution and social inclusion in the European Union (EU). The EU-SILC provides (micro-level) information on both households and individuals on topics such as income and tax, material deprivation, housing conditions, employment, childcare, health and education.

Additional specific topics are covered in ad hoc annual modules, which complement the core variables that appear in every EU-SILC wave by providing supplementary information that highlights different aspects of social inclusion. Examples of the topics covered in these modules are the intergenerational transmission of poverty (2005 and 2011), social participation (2006 and 2015), indebtedness (2008) and the intra-household sharing of resources (2010). The EU-SILC 2013 ad hoc module provided information on well-being for 350,000 individuals in 32 different European countries.¹

EU-SILC is a panel dataset with a four-year rotation (so that 25% of the sample are replaced each year). The 2018 EU-SILC wave repeated parts of the 2013 well-being module, although the gap between the two waves was large enough that none of the 2013 respondents remained in the 2018 sample. However, in this analysis, we carry out a cross-sectional analysis of the 2013 data only, as the 2018 module on well-being did not include all of the well-being variables that we will consider below.

Among the aims of the 2013 well-being module was to place greater emphasis on individuals’ feelings of well-being, which are known to be highly correlated with levels of social inclusion and labour market integration. The module covered subjects like satisfaction with different life domains, the emotional state of mind, social deprivation, and the option to ask for help if necessary. All questions were posed at the individual level (to household members aged 16 and older).

Survey design can have an impact on individual replies. Thus, the placement of the well-being module in the survey questionnaire, the response scales used and how respondents perceive the questions in different countries can influence the responses. As the order of the questions within the module and the position of the module questionnaire in the survey were mostly standardised across countries, Eurostat considers the contribution of these parameters to the overall context–related bias to be minimal. For further details on assessments of the implementation of the 2013 EU-SILC module on well-being, see Eurostat (2012a).

The main dependent variable in our analysis is “overall life satisfaction”, with replies on a scale from zero (not at all satisfied) to 10 (completely satisfied).² The life satisfaction variable is intended to capture the respondent’s broad, reflective appraisal of his/her life as a whole. The term “life” is intended to be understood here as all of the areas of a person’s life at a particular point in time (Diener 2006). The variable captures the respondent’s opinion/feeling about how s/he is feeling “these days”, rather than specifying an explicit longer or shorter time period. Thus, the intent is not to measure the individual’s current emotional state (or affect), but for him/her to make a reflective judgement about his/her satisfaction with his/her life overall (Eurostat 2012b).³

The weighted average overall life satisfaction score in the 2013 EU-SILC, covering all 32 European countries and considering respondents up to age 80, is seven. Table A.1 in the Appendix presents the descriptive statistics of the life satisfaction scores and the percentage of the respondents who are partnered by the different age groups we will analyse.

The average overall life satisfaction score appears in the middle of the ranking of the more specific domain satisfaction variables: satisfaction is lower for “trust in the political system”, “trust in the legal system”, “trust in others”, “satisfaction with the financial situation”, “trust in the police” and “satisfaction with time use”; satisfaction is higher for “satisfaction with recreational or green areas”, “satisfaction with the living environment”, “job satisfaction”, “satisfaction with commuting times”, “meaning of life”, “satisfaction with accommodation” and “satisfaction with personal relationships”. It should, however, be emphasised that the “overall life satisfaction” score comes from a question in its own right in the SILC questionnaire, and is not calculated as a summary measure of the various domain satisfaction scores.

The cross-country distribution of life satisfaction in our analysis is similar to that found in other surveys (see, for example, the distribution of the answers to the Cantril ladder question in the Gallup World Poll, which is presented in the annual World Happiness Report). Overall life satisfaction is generally shown to be higher in Northern Europe and lower in Southern Europe, as illustrated in Figure 1’s heatmap.

The lowest life satisfaction score on the 0–10 scale is found in Bulgaria (4.8), followed by in Serbia (4.9) and Portugal (6.1); and the highest score is found in Switzerland (8.0), followed by in Finland and Denmark (both 7.9).

The relationship of life satisfaction to age and sex is depicted in Figure 2. As throughout the empirical analysis, here we only consider respondents up to age 80. Across all 32 countries in the study sample, women are found to be less satisfied at higher ages and more satisfied at younger ages than men.

However, we find no systematic gender gap in overall life satisfaction within couples, independent of the man’s age, the woman’s age and the size and direction of the age difference within the couple. The distribution of the absolute difference in life satisfaction between the two partners is normally distributed around zero, with 40% of couples reporting zero difference.⁴ It may be the case that the gender gap in life satisfaction at younger and older ages shown in Figure 2 reflects the respondents’ partnership status. Indeed, we find that women are more likely to be partnered at younger ages and less likely to be partnered at older ages, as illustrated in Figure 3.

To find out whether the gender difference in life satisfaction disappears once we control for partnership status, we estimate now, in a first step, life satisfaction as a function of age, sex and partner status. We control for country fixed effects in order to focus on within-country variations only. The results presented in the following are based on OLS regressions.

The regression results appear in Table 1, with first a stripped-down model controlling only for age (as 10-year age group dummies) and country dummies; then a second specification with an additional control for sex; and, finally, a third specification controlling for both sex and partner status. An individual is considered as “partnered” when s/he is observed to be cohabiting with a partner at the time of the survey. Note, however, that the results are robust to considering only those who are married as being “partnered”. As the omitted age category in these regressions is 40–49, the estimated coefficients on the dummy variables for the other 10-year age categories should be read as relative to this age group.

Column 1 indicates that life satisfaction is highest for teenagers and lowest for those in their fifties, and again for those in their seventies.⁵ Column 2 then controls for sex, which barely changes the estimated age coefficients. The difference in life satisfaction between men and women appears to be insignificant once we control for age. Finally, column 3 shows that there is a small but significant gender difference, with women being somewhat more satisfied than men once we control for age and partnership.⁶ Partnership itself turns out to be highly correlated with overall life satisfaction, with partnered individuals being significantly more satisfied than non-partnered individuals. The effect size here is large, at two-thirds of a life satisfaction point, or almost one-third of a standard deviation (see the figures in the last column of the first panel of Appendix Table A.1). A comparison of columns 1 and 3 of Table 1 shows that controlling for partnership deepens the U-shape, which again suggests that partnership plays a role in the age-satisfaction relationship.

To see whether the age pattern of life satisfaction differs significantly between the partnered and the non-partnered respondents, we now run separate regressions for the two groups, and we interact age, sex and partnership status. Table 2 shows the results, with the first two columns referring to the partnered individuals and columns 3 and 4 referring to the non-partnered individuals. Columns 1 and 3 control for age and country-fixed effects only, while columns 2 and 4 also include sex. Finally, column 5 shows the estimation results when we analyse the whole sample together and introduce a full set of interactions between age, sex and partnership.

The results are very different by partnership status. Columns 1 and 2 indicate that for the partnered individuals, the life satisfaction differences between most of the age groups are relatively small, with the U-shape of life satisfaction over age being rather flat. By contrast, for the non-partnered individuals, the effect size of the age coefficients is much larger, indicating a profound U-shaped relationship with life satisfaction, with those individuals in their fifties being the least satisfied. A comparison of columns 2 and 4 shows that the decline in life satisfaction from the teens to the fifties is almost four times larger for the non-partnered individuals than for the partnered individuals.⁷ Finally, women are again found to report slightly higher life satisfaction than men in both partnership groups.

Column 5 of Table 2 shows the estimated coefficients based on a specification with interaction terms between the following three categorical variables: “age category”, “partnership status” and “sex”. The interaction coefficients are difficult to interpret on their own. They can, however, be used to calculate the estimated levels of “overall life satisfaction’ for each possible combination of age category, sex and partnership status. For example, the estimated level of life satisfaction for a non-partnered man aged 40 to 49 (the reference categories) is 7.097, which corresponds to the intercept in column 5. The estimated level of life satisfaction for a non-partnered woman aged 40 to 49 is 7.1698, which is obtained by adding the estimated coefficient of “female’ (0.0728) to the intercept. The estimated life satisfaction level of a partnered women aged 15 to 19 is 8.2128, which is calculated as follows: 7.097 (constant) +0.0728 (estimated coefficient on “female”) +1.69 (estimated coefficient on age category “15–19”) +0.953 (estimated coefficient on “partnered”) - 1.6 (estimated coefficient on the interaction “age 15–19, partnered, female”).⁸

Figure 4 illustrates these estimated life satisfaction scores for each possible combination of age category, sex and partner status using the numbers from column 5 of Table 2. While life satisfaction is clearly U-shaped for the non-partnered individuals, the pattern is relatively flat for the partnered individuals. The U-shape is the most profound for the non-partnered men, who are the least satisfied at ages 50 to 59, and are somewhat more satisfied than the non-partnered women at ages 70–80. Note that the confidence intervals are large for the partnered women and men at ages 15 to 19, as few of the respondents cohabit with a partner before the age of 20.

Figure 4 further shows that within each age and partnership category, the gender difference in life satisfaction is small. Independent of sex, the main difference in life satisfaction found among the groups studied is between the partnered and the non-partnered individuals, especially in mid-life. Thus, based on the raw data, the gender difference in life satisfaction at younger and at older ages, as illustrated in Figure 2, is due to differences in partnership status. Until their early forties, women are happier than men because they are more likely than men to be partnered at these ages. From their sixties onwards, men are happier than women because they are more likely than women to be partnered at these older ages.

The results above come from a very stripped-down specification. However, they do not change in nature when we add more variables to look for potential confounders in the relationship between age, partnership and life satisfaction. We consider country of birth (Native, European, and other), the degree of urbanisation, living in a relatively rich country (as measured by GDP per capita in PPP terms), education, and both individual and household income. Within each education and income group, the U-shaped pattern of life satisfaction remains much sharper for the non-partnered than the partnered individuals.

We also consider different measures of subjective well-being, replacing “life satisfaction” with “meaning of life”, “satisfaction with financial situation”, “satisfaction with personal relationships” and the “ability to make ends meet”. We find a flatter U-shape for the partnered than for the non-partnered individuals for all four of these variables. In this case, the protection derived from being in a partnership may be partly financial. By contrast, we find no significant differences by age group, sex or partnership status for job satisfaction.⁹ Finally, for satisfaction with time use, we find a U-shape in age, but with a minimum level (at age 30 to 39) that is lower than the figure for overall life satisfaction shown in Table 2. The highest level of satisfaction with time use is observed at ages 60 to 69. Few significant differences in satisfaction with time use between men and women or between partnered and non-partnered individuals are found.

We might wonder whether our results for being in a partnership reflect the presence of children in the household. The EU-SILC only provides information on the number and the ages of children who are living in the respondent’s household. Moreover, having children at home becomes increasingly selective as the respondents grow older, when empty-nesting becomes more prevalent. We therefore controlled for children by limiting the sample to individuals aged 16 to 44.¹⁰ We find that overall life satisfaction does not differ significantly by age for men and women with children, whereas it drops significantly from age 16 to age 44 for both childless men and childless women. However, within the group of partnered individuals, we detect no significant differences in life satisfaction by age between childless individuals and those with children. Thus, the presence of children seems to play a less important role in overall life satisfaction than partnership status, at least for adults up to age 44.¹¹

Finally, we carry out a country-by-country analysis. In each of the 32 European countries in our sample, the pattern of overall life satisfaction over age is found to be relatively flat for the partnered individuals and U-shaped for the non-partnered individuals, independent of their sex.¹²

Thus, our cross-sectional analysis based on the EU-SILC data suggests that the U-shaped pattern of life satisfaction by age holds for the non-partnered individuals, but largely disappears for the partnered individuals. It therefore appears that gender differences in overall life satisfaction are mainly caused by age-specific gender differences in the likelihood of being partnered. These results confirm the role of partnership in well-being, as, for example, reported by Verbakel (2012), whose analysis of the European Values Study showed that married people have the highest level of well-being, followed by those who are cohabiting, dating or single; and, finally, by those who are widowed or divorced. Our results presented above suggest that being partnered also matters for the dynamics of well-being across age groups. To help us understand why this is the case, we can turn to the abundant empirical literature showing that being in a partnership provides a person not only with additional resources, but also with social and emotional support, and with health benefits. First, it has been shown that the economic benefits of being partnered, and, in particular, of being married, increase with income if income is taxed in a progressive way. The observation that income tends to be hump-shaped with age may explain why the protective effect on well-being of being in a partnership is particularly strong for middle-aged people. Second, as noted above, numerous studies have suggested that being married is associated with better psychological health. However, there is no strong evidence that this effect varies with age. In an analysis of data from the Danish Twin Registry, Kohler et al. (2005) found a slight increase in this effect with age for men, but no significant change in this effect for women. Finally, marriage has been shown to be associated with better health and longevity (see Wilson and Oswald 2005, for an extensive survey). For instance, Ben-Shlomo et al. (1993) found that married men have a lower risk of dying from cardiovascular disease and lung cancer, and Helmer et al. (1999) reported that never-married people have a significantly higher risk of developing Alzheimer’s disease. As these pathologies tend to occur in the second part of life, these findings may help to explain why the protective role of partnership is attenuated for younger people.

However, the results of a cross-sectional analysis will not necessarily correspond to the findings of an analysis of panel data if there is selection into partnership. It is not only possible that those individuals who are more satisfied select into partnerships (as in Stutzer and Frey 2006); it also is quite likely that people who are not satisfied with their relationships (and therefore have lower overall life satisfaction scores) select out of partnerships.¹³ Once these individuals find a new relationship that makes them happy, they will re-enter our sample of cohabiting partners. Both of these sources of endogeneity make it more difficult to conclude that being in a partnership protects against the U-shaped pattern, as the people who are partnered (or are non-partnered) are not the same people at different ages.

To shed light on the causal impact of being in a partnership on life satisfaction from a life-cycle perspective, we need to add panel data to our cross-sectional analysis. If being in a partnership indeed protects against the U-shaped pattern, we should find that the age profile of the partnered individuals in the panel data (i.e., within the same person) is flatter than that of the non-partnered individuals. The EU-SILC data does not help us here, given its four-year revolving design (and with information on life satisfaction only being available in the 2013 and 2018 survey waves). Thus, to examine this question, we turn to panel data from three of the longest-running panel datasets: the HILDA (Australia), the BHPS (the United Kingdom) and the SOEP (Germany).¹⁴

Our analysis of panel data uses the same kinds of specifications as the cross-sectional analysis above. The dependent variable, life satisfaction, is on a 0–10 scale in the HILDA and the SOEP, and a 1–7 scale in the BHPS.¹⁵ We standardise the life satisfaction scores, so that our correlations are in terms of standard deviations. We consider those individuals aged 16–80, and split up age into dummies for the categories 15–19, 20–29, 30–39, 40–49, 50–59, 60–69 and 70–80. Partnered is a dummy for marriage or cohabitation. There is a heated discussion in the literature regarding the controls that should be introduced when analysing the relationship between subjective well-being and age.¹⁶ Here, as in our analysis of the EU-SILC data above, we take a minimalist approach, and introduce no other controls. All of our regressions include an individual fixed effect (which subsumes the effect of sex, plus any other variable that is time-invariant).

The regression to be estimated is

where LSit,m is the life satisfaction of individual i at time t in marital status m, Ageit,m is a set of age dummies, and μi is the individual fixed effect. The presence of the latter implies that the age coefficients, β, are identified by the same individual as s/he moves from one age category to the next in the panel data.

The estimated coefficients from these regressions on our three panel datasets are depicted in Figure 5. The three columns refer to the BHPS, the SOEP and the HILDA, respectively. Within each column, the first graph shows the estimated age coefficients from a regression that includes the whole sample, while the second and the third graphs depict those for the partnered and the non-partnered individuals separately. The Y-axes are on the same scale within each row. It is immediately obvious that the largest age coefficients are found in the SOEP.

Looking at the first row, we can see that the estimated age coefficients suggest that there is something of a U-shaped relationship in both the BHPS and the HILDA. By contrast, the panel results for the SOEP produce a monotonically negative relationship between life satisfaction and age.¹⁷

The most important element of Figure 5 in the context of our analysis here is the comparison of the shapes of the estimated coefficients in the partnered and non-partnered graphs in the second and third rows. A broad characterisation of the results is that the age profile is more U-shaped for the non-partnered than for the partnered individuals, as in the EU-SILC data. However, this distinction is somewhat less apparent in the SOEP, and, in general, there is still a noticeable age profile in life satisfaction for the partnered individuals (whereas the age profile in life satisfaction for the partnered individuals in the cross-sectional EU-SILC data was largely flat).¹⁸

Thus, there is only partial agreement between our pooled EU-SILC results presented in Section 2 and the results from the three panel datasets in Figure 5: i.e., in the panel data, the partnered individuals continue to drop down the first arm of the U-shape at least as much as the non-partnered individuals. Below, we consider a number of potential explanations for the differences between the pooled and panel results.

While the finding in the literature that there is a U-shaped relationship between life satisfaction and age is robust, it has not been fully explored. This relationship has been found for multiple countries and datasets (Blanchflower 2020a and 2020b), and often continues to hold even when large numbers of controls are considered, ranging from observable mid-life burdens to aspirations. Thus, there is still much to be learned about the reasons why the U-shaped pattern exists. In this article, we have analysed the possible role of a rather intuitive factor–-being in a partnership–-and examined the question of whether being partnered could help protect individuals against the mid-life decline in satisfaction. In general, we expected to find that the U-shape would be deeper for those individuals who lacked certain physical or emotional resources.

Our findings were mixed, and depended strongly on the type of data analysed. The life satisfaction of the partnered individuals was shown to be rather flat with age in the cross-sectional data we analysed, but displayed a more U-shaped pattern in the three panel datasets we considered. We believe this difference partly reflects a composition effect: i.e., that those who leave partnerships in their forties and fifties are noticeably less satisfied with their lives than those who break up at younger or older ages.

To study how key demographic, economic and psychological variables change over the life-cycle, reliable longitudinal datasets are needed. As we have seen above, imputing life-cycle changes from cross-sectional data can be hazardous. While developing and supporting panel datasets will require considerable investments by public funding agencies, longitudinal data are essential to research on the evolution of societies, which are composed of individuals who change over time. Having access to panel datasets will also facilitate inter-disciplinary work, with well-being arguably being a good example of a topic that could be analysed jointly by researchers in a number of different fields.