ARTICLE https://doi.org/10.1038/s41467-018-08002-2 OPEN Feasible future global scenarios for human life evaluations Christopher Barrington-Leigh1 & Eric Galbraith 2,3,4 Subjective well-being surveys show large and consistent variation among countries, much of which can be predicted from a small number of social and economic proxy variables. But the 1234567890():,; degree to which these life evaluations might feasibly change over coming decades, at the global scale, has not previously been estimated. Here, we use observed historical trends in the proxy variables to constrain feasible future projections of self-reported life evaluations to the year 2050. We find that projected effects of macroeconomic variables tend to lead to modest improvements of global average life evaluations. In contrast, scenarios based on non- material variables project future global average life evaluations covering a much wider range, lying anywhere from the top 15% to the bottom 25% of present-day countries. These results highlight the critical role of non-material factors such as social supports, freedoms, and fairness in determining the future of human well-being. 1 Institute for Health and Social Policy; and School of Environment, McGill University, Montreal H3A1A3 QC, Canada. 2 ICREA, Pg. Lluís Companys 23, 08010 Barcelona, Spain. 3 Institut de Ciència i Tecnologia Ambientals (ICTA) and Department of Mathematics, Universitat Autonoma de Barcelona, 08193 Barcelona, Spain. 4 Department of Earth and Planetary Sciences, McGill University, Montreal, QC, Canada. Correspondence and requests for materials should be addressed to C.B.-L. (email:Chris.Barrington-Leigh@McGill.ca) NATURE COMMUNICATIONS | (2019)10:161 | https://doi.org/10.1038/s41467-018-08002-2 | www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-018-08002-2 C omplex policy decisions, ranging from international cli- variables using both static (across countries, XS) and dynamic mate change negotiations to investments in education or (within countries comparing two periods, 2P) models (Fig. 1, see infrastructure, often rely on projections of readily quan- Supplementary Note 2 for details). As compared with the static tified material outcomes such as per capita income to assess the XS model of differences across countries (Fig. 1a), the dynamic 2P impacts on human welfare, for example, refs. 1,2. An alternative model of within-country changes shows a relatively greater approach, developed over the past few decades, aims to apply emphasis on freedom to make decisions, availability of social more direct measures of human experience that integrate many support, and perception of corruption. The coefficients in Fig. 1a, dimensions of life according to those living it. These well-being b are normalized to standard deviations, meaning that a 1 s.d. measures have the disadvantage of being subjective, and change in income predicts a 0.19 s.d. change of life evaluations, can therefore be difficult to interpret. Nevertheless, they have holding other factors constant, while a 1 s.d. change in social been shown to be consistent with external evaluations at the support predicts a 0.29 s.d. change in life evaluation. In our individual level, are reproducible over time within populations, projections, below, we use the dynamic model (2P) to predict and are increasingly embraced by decision makers as a leading changes in life evaluation over several decades, though as dis- objective3–7. They also show consistent predictive relationships cussed later, our main findings would be even more pronounced with other societal variables. were the XS model coefficients to be used. We emphasize that the Here, we focus on one measure of subjective well-being (SWB) projections that emerge are not predictions of the future, but with particularly consistent predictive relationships: the cognitive illustrate the range of the most feasible futures that might occur, evaluation of life. Annual, near-global national samples of self- depending on human actions. reported life evaluations, from Gallup’s World Poll, have recently become available through the World Happiness Report. While the Projected range of global mean changes. The range of feasible Poll’s extensive questionnaire is not designed expressly for explaining outcomes spanned by the three sets of scenarios is shown in differences in life quality, it does contain several questions which Fig. 2. Two sets of scenarios relate to changes in material con- address dimensions of life known to be important to life evaluation. ditions. For the first of these sets (Organization for Economic Co- Prior work in a range of contexts has shown that, among these, operation and Development (OECD) projections), both the income differences have significant explanatory power in accounting optimistic (strong-growth) and pessimistic (weak-growth) pro- for variation in life evaluations, while less tangible aspects of human jections show improvements of global average experienced life experience can account for as much or more of the variation8; see evaluation compared with the 2016 value of 5.24 out of 10. The Supplementary Note 1. Following Helliwell et al.8, we account for a optimistic case of the second material-factors pair (Material portion of international differences in life evaluation using four proxy Trends) is characterized by steady growth at the 90th percentile of variables derived from the Gallup World Poll, which we will refer to recently observed trends (4%/year for income and 0.55 years/year as reflecting non-material factors (corruption9, freedom, giving10, for healthy life expectancy), and this scenario’s outcome agrees and social support), and two proxy variables reflecting material very closely with the optimistic OECD estimate. However, not all factors (per capita gross domestic product (GDP) and life expec- countries have actually experienced positive real economic tancy). When these measures are included in a simple linear model growth. As a result, the corresponding pessimistic Material of cross-country differences, they explain roughly three-quarters of Trends scenario, using 10th percentile observed trends, admits the variation among annual national averages of life evaluations, and the possibility of a small decline of global average life evaluations, parameters are estimated with high statistical precision, as shown by to 5.1, by 2050. prior work8; see Supplementary Note 2 for details. The range of feasible outcomes encompassed by these material Static inter-country comparisons such as these have been the scenario sets is dwarfed by the range of outcomes in the non- focus of much attention, but they do not address how the life material scenario set. In the top 10% of recent observed trends, evaluations of humans might change in the future, an important freedom and social support grew at 2%/year and 0.6%/year, consideration for motivating and evaluating policy decisions. respectively, while corruption decreased at 1.3%/year. At these Although there is no mechanistic understanding on which to rates, our Non-material Trends scenario projects a radical global reliably predict these future changes directly, the time-span cov- mean improvement of life evaluation to 6.9, which—for ered by the Gallup World Poll (2005–2016) is now sufficient to comparison—is close to the levels actually reported by Belgium use as an empirical constraint on the feasible rates at which proxy and Costa Rica in 2016, when they had the 18th and 14th highest variables might change, and reveals that strong trends have life evaluations17. Conversely, if the least favorable 10th percentile occurred in some countries (Supplementary Figure 1). The por- of observed non-material trends were to prevail in all countries, tion of life evaluations that can be predicted from the proxy our projections suggest that the drop in life evaluations by 2050 variables would be expected to change accordingly, providing a would take the global average to 3.4, below the level of Egypt and means by which to estimate the range within which the future India in 2016, when they ranked 118th and 120th out of 157 trajectories of life evaluations are most likely to fall. countries. Here, we use the historical survey data to develop a dynamic Figure 2 also shows our projections based on 30th and 70th statistical model and evaluate feasible rates of change, and con- percentile recent trends. These provide very similar qualitative struct simple scenarios for human life evaluations in 2050, within conclusions as our primary analysis, in that the Non-material the time frame commonly considered in forward-looking policies Trend scenarios encompass more extreme positive and negative informed by climate model projections11–16. We find that sig- possibilities than the material trends. Notably, the 30th to 70th nificant changes in global average life evaluations are feasible, and percentile Material Trend projections span a similar range of life could be either positive or negative. In addition, we find these evaluations as that of the OECD-derived projections, although the future outcomes to have a markedly larger dependence on non- trend-based projections are approximately 0.3 points lower than material proxy variables as compared with material ones. the OECD projections. Results Geography of projected changes. We map the distribution of Proxy model of life evaluations. Our statistical models predict a projected changes in life evaluation in our material OECD and portion of the observed changes in life evaluations from the proxy Non-material Trend scenarios, according to population density 2 NATURE COMMUNICATIONS | (2019)10:161 | https://doi.org/10.1038/s41467-018-08002-2 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-018-08002-2 ARTICLE a b XS standardized coefficients 2P standardized coefficients 0.4 0.4 0.2 0.2 0.0 0.0 −0.2 Cross-section model −0.2 Two-period change model of life evaluation of life evaluation GDP Log( capita ) Life Corruption | Giving Social GDP Log( capita ) Life Corruption | Giving Social expectancy Freedom support expectancy Freedom support c 1.5 1.0 0.5 ∆ Life evaluation 0.0 Southeast Asia East Asia −0.5 South Asia Sub-Saharan Africa Western Europe −1.0 North America and ANZ Commonwealth of Independent States Latin America and Caribbean −1.5 Middle East and North Africa Central and Eastern Europe −0.6 −0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1.0 Predicted ∆ life evaluation Fig. 1 Predictions of life evaluations from proxy variables. a Six predictor variables are used simultaneously to predict life evaluations in static cross-section between countries (XS). All corresponding country-year observations are used, after removing global year-to-year changes. Effect sizes are shown, with 90% confidence intervals, normalized to standard deviations, in order to compare the estimated coefficients across predictor variables. The strongest predictor is the income variable, but all predictors have significant strength, and their confidence intervals largely overlap. b As in a but for the dynamic, two-period model (2P) which explains changes over time within countries. In the 2P model, confidence intervals are looser, and the income coefficient excludes zero with only ~90% confidence intervals. However, models for annual changes (see Supplementary Information) show similar patterns and have tighter confidence intervals. c The relationship between observed and predicted changes in life evaluation using the 2P model. The changes are the differences in national average life evaluations between 2005–2007 and 2014–2016. Symbol size corresponds to the population size of each country projections for 2050, in Fig. 3 (the corresponding Material Trend scenarios, which are uniform across countries, are shown in Supplementary Figures 2–5). For the non-material scenarios, we OECD material projections show only the most optimistic (90th percentile) and pessimistic (10th percentile) projections. The OECD’s material scenarios are driven by national changes in per capita income, since in our 2P Material trends model life expectancy has no significant effect on life evaluations. In general, the OECD projections show some degree of improvement in all countries, even under pessimistic assump- Non-material trends tions. Differences among countries reflect the details of the OECD’s macroeconomic projections and reflect an expectation of economic convergence, that is, higher economic growth rates in countries with lower initial income. 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 Compared to the OECD projections, the non-material projection maps show a much wider range of possible changes, Global average life evaluation in 2050 as did the global averages. In the optimistic non-material future, Fig. 2 Projected feasible life evaluations in 2050 scenarios. The results of improvements of more than 1.5 on the 11-point life evaluation two sets of material scenarios are shown for the year 2050, based on scale are widespread, including in densely populated regions of macroeconomic projections (OECD, orange) and recent trends (purple), as India, China, eastern Europe, and sub-Saharan Africa. The scope well as a set of non-material scenarios based on recent trends (green). The for improvement in non-material factors is smaller in western central vertical lines indicate projections based on the median trends, the Europe and North America, given that these countries are already thick horizontal bar indicates the range based on the 30th to 70th %ile closer to the maximum values. Nonetheless, the improvements in trends, and the thinner horizontal bar indicates the range based on the 10th predicted life evaluation that would appear to be feasible due to to 90th %ile trends. Outcomes are calculated with coefficients from the 2P non-material changes by 2050 exceed those of even the most model and are weighted by projected population sizes to aggregate across optimistic material changes in all countries. countries. The red arrow indicates the population-weighted average life The pessimistic projections show an even starker difference evaluation recorded in 2016 between material and non-material futures. Whereas the NATURE COMMUNICATIONS | (2019)10:161 | https://doi.org/10.1038/s41467-018-08002-2 | www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-018-08002-2 a OECD material (pessimistic) b OECD material (optimistic) c Non-material (pessimistic) d Non-material (optimistic) –2 –1 0 1 2 Change of life evaluation from 2017 to 2050 Fig. 3 Geographic distribution of feasible life evaluation changes. Projections for 2050 are for the OECD material growth scenarios (a, b), and scenarios in which the non-material predictor variables change at the 10th and 90th percentile rates of recent observations among all countries (c, d). Coefficients used in the projection are based on the two-period model of life evaluations. The shading on the map is weighted by projected population density pessimistic OECD projections do not produce negative outcomes noticed by Easterlin18 is related to the smaller predictive role of in any countries, feasible changes in non-material variables could income in explaining long-run variation as compared with annual produce a large decrease in life evaluation in any country by 2050. changes19. Human life evaluations appear to adapt to the ambient This reflects the fact that, in recent years, the social variables have material affluence over time, in the sense that our mental benefit had large negative trends in many countries, whereas material gradually accommodates to the experienced level20,21. This con- variables are generally expected to improve, or at least remain trasts with social aspects of life, for which there is little evidence stable everywhere. of such adaptation22. In a second phenomenon, humans respond to material consumption norms set by those around them. Both Discussion phenomena help to explain the lack of long-run response of life Our analysis suggests that large future changes in global life evaluations at the national level to rising incomes during eco- evaluations are feasible, and could be either positive or negative, nomic growth, for example, refs. 19,23, and may explain the dif- based on the observed changes of proxy variables within countries ference in raw coefficient sizes mentioned above. Furthermore, during 2005–2016. Our scenarios treat material and non-material life evaluations may be sensitive to national (as opposed to local) proxy variables independently, so that the modeled changes dis- rank of affluence. Such positional concerns would imply an even cussed above are additive. For instance, using the OECD and 10/ smaller long-run direct benefit from aggregate economic growth 90 percentile Non-material Trends, the combined feasible chan- across all countries, because even the cross-sectional distribution ges give a range of projected 2050 global average life evaluations of life evaluations would reflect in part the outcome of a long-run from 3.9 to 7.6. Regardless of whether the material trends or zero-sum game. OECD projections are used, the major part of the summed ranges In further support of the importance of non-material factors, is due to the non-material factors, because the range of material we point out that, during the period of observations, the global rates of change is smaller, relative to their modeled impact on life mean of life evaluations did not change significantly, despite an evaluations, than the range of rates of change of the non-material average increase of incomes by 17% and lengthening of life variables. Thus, non-material variables can feasibly make larger expectancy by 2.6 years (Supplementary Tables 1–3). The global impacts than income on a multi-decadal timeframe. improvements in material conditions, which were accompanied Furthermore, we earlier emphasized that variance-standardized by improvements in average freedom and corruption, were effect sizes are larger for non-material factors than material fac- counteracted by a decrease in average social support, and the tors in the 2P dynamic model, while income is the strongest reported change in global lifeevaluations was negligible. predictor in the cross-sectional model. However, the raw coeffi- Although we have focused here on subjective life evaluation, we cients have the opposite relationship, in that the ratio of raw effect undertook an identical model-building and projection procedure size of income to raw effect size of social support (or to any of for affect measures, which capture day-to-day feelings. A pre- our other non-material variables) is larger for the 2P dynamic dictive model based on the six proxy variables does not capture as model than the cross-sectional model. Therefore, carrying out much of the variation in affect as the models for life evaluation8 projections for long-run changes using coefficients from the (see Supplementary Tables 4–7). Nonetheless, with this caveat in cross-sectional model results in even starker dominance by mind, the trend-based projections suggest that the average affect non-material factors. balance of humans in 2050 could decrease by as much as 0.28 or Speculatively, our model may overestimate the role of GDP per increase by as much as 0.31 on a two-point scale within the 90th capita for another reason. A well-studied phenomenon first to 10th percentile trends, spanning 30% of the total range, very 4 NATURE COMMUNICATIONS | (2019)10:161 | https://doi.org/10.1038/s41467-018-08002-2 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-018-08002-2 ARTICLE similar to the relative magnitude of feasible changes in life eva- for example, ref. 26. By considering these measures at the aggregate (country) level, luations. Furthermore, the changes in affect are dependent we best capture the sum of individual effects and those which come from public goods (at national and subsidiary scales) and from social and contextual effects at exclusively on non-material factors, with no discernible influence all spatial scales up to the country level. Our use of country-level data is motivated of material changes (see Supplementary Note 2). Thus, large by the assumption that externalities (e.g., positional effects, or public goods) are changes in affect are also feasible, could also be positive or likely to be large within countries, but are likely to be relatively small between negative, and appear to be exclusively dependent on non-material countries. factors. Summary information about the dataset is provided in Supplementary Note 1, including descriptive statistics of levels (Supplementary Table 3), annual changes We would emphasize that the predictor variables we use are (Supplementary Table 2), and 2P changes (Supplementary Table 1) of our key proxies, rather than direct mechanistic drivers of life evaluation. variables. We also provide estimates of pairwise correlations among levels There are many dimensions of material supports, such as nutri- (Supplementary Table 9) and 2P changes (Supplementary Table 10) in our key tion, shelter, and infrastructure, which may not vary directly with variables. per capita GDP. Provision of these material needs may also help to build trust, support systems among family and communities, Regression models for well-being (overview). We estimate four different models to characterize the relationship between country-level SWB and our six predictor and freedom of choice. Likewise, our non-material proxies do not variables. Each model is suitable for estimation using a weighted least-squares capture all of the dimensions of social experience known to be approach. One model predicts the relationship among countries, another predicts important for life evaluations, and a breakdown of these non- within-country changes between our two time periods, and two approaches predict material conditions likely imperils the material ones24. In order to year-to-year changes within countries. We use the 2P model in our primary ana- lysis in order to remain conservative with our primary conclusions, and in order to address the possibility that these interactions are statistically avoid the short-run effects of the global financial crisis (GFC) in our estimates of important, we repeated our analysis using models augmented long-run trends, but we show results from the other models in Supplementary with a series of interaction (also know as moderation) terms Notes 1 and 3. For completeness, the four models and the standardization are between pairs of predictors with statistically significant effects in explained below. For each of these models, the corresponding estimates are tabulated in Supplementary Note 2 both as raw coefficients and secondly as unitless our 2P model. These tests, which are detailed in Supplementary standardized coefficients. Note 3 and Supplementary Table 8, suggest a statistically sig- As emphasized in the main text, we cannot identify and isolate independent nificant interaction between GDP and freedom of choice, but this causal relationships between supports (predictor variables) and life evaluations. interaction does not alter our general conclusions nor the greater Rather, our set of proxy variables, with their shared variance, captures one projection of the true underlying processes. In the case of our Giving variable, we importance of non-material factors (see Supplementary Figure 1). err on the side of under-estimating the importance of giving by using the residual Nevertheless, we recognize that our results remain limited by the from a regression of the underlying donations variable on an income measure. This use of a global model with six predictors that are, unavoidably, may in turn result in over-estimating the importance of income. inter-related. In the future, the accumulation of larger databases could allow the development of a formal model selection process Regression models for well-being (static model). First, we consider the cross- to choose among broader possibilities of predictor variables, and sectional (XS) relationship may ultimately allow the development of mechanistic models 6 X with greater predictive power. Sit ¼ a þ bj xjit þ ct þ ν i þ εit ; ð1Þ Despite the simplifications inherent in this first attempt to j¼1 project future human well-being, our results show that the greatest benefits to be potentially made over the next decades, as where Sit is the SWB in country i in year t; a is a global constant; xjit are the six well as the most dangerous pitfalls to be avoided, lie in the predictor variables; ct is a constant in a given year t for all countries; and νi + εit is domain of social fabric. Focusing on income among the effects of an error term clustered at the country level. Here, bj are the regression coefficients long-run policy is therefore too narrow, and misses the majority describing the effects of interest. In estimating (1) we use country weighting because we consider the political–cultural–economic dynamics to be distinct across of the human well-being effects that could feasibly occur, based countries and therefore consider each country to be one sample unit. We cluster on past experience. Given that the policies known to support a errors at the country level to take into account the fact that we have multiple strong social fabric can differ from those focused on economic observations (i.e., by year) for each country, but do not capture all country features growth, our results suggest that scarce resources may be better in the model. Our inclusion of year indicators (dummy variables) ct is to remove variation related to global secular trends (e.g., average global economic growth) and prioritized towards explicitly social aims if human well-being is short-run global cycles (e.g., the GFC), so that our estimated bj capture the cross- the goal. sectional variation. In order to quantitatively compare these relationships bj across predictors j, we σ Methods calculate the standardized regression coefficients βj % σ j bj ; that is, we estimate the S Survey data. Following recent World Happiness Reports, we include a measure of standardized variable equation income along with variables representing healthy life expectancy (life expectancy), perceived level of corruption (corruption), freedom to choose (freedom), pre- Sit X6 xjit valence of donating to others (giving), and availability of informal social support aþ ¼~ βj þ ~ct þ ν~i þ ~εit ; ð2Þ σs j¼1 σj (social support). The English wording of the life evaluation question is “Please imagine a ladder, with steps numbered from 0 at the bottom to 10 at the top. The top of the ladder represents the best possible life for you and the bottom of the ladder represents the worst possible life for you. On which step of the ladder would where σs is the standard deviation of Sit across all country-year observations, but again calculated with country weights to take into account the fact that not all you say you personally feel you stand at this time?” The social support variable is countries have the same number of observations (years). The standard deviation σj the national average of dichotomous responses to “If you were in trouble, do you of xjit is calculated the same way. Thus, βj should be interpreted as the number of have relatives or friends you can count on to help you whenever you need them, or standard deviations of change in S associated with one standard deviation change not?” The freedom variable is the national average of dichotomous responses to in xj. “Are you satisfied or dissatisfied with your freedom to choose what you do with your life?” The corruption variable is the national average of dichotomous answers to two questions: “Is corruption widespread throughout the government or not?” Regression models for well-being (dynamic models). While (1) models the and “Is corruption widespread within businesses or not?” In order to isolate the variation across countries, we also generate models of changes over time in order to prevalence for giving from the variation in financial capacities, national mean project future changes based on past rates of change. We do this in three ways. Two responses to “Have you donated money to a charity in the past month?” are of them, fixed effects (FE) and first differences (FD), use the year-to-year changes regressed on GDP per capita, and the residual becomes the Giving variable25. in the predictor variables for each country to explain year-to-year changes in SWB. Further descriptions and important notes for these variables are given in the sta- The third method (2P) considers only the longer-run difference for each country tistical appendix for Chapter 2 of the 2017 report8. Our measure of income is the from early in the Gallup World Poll data (2005–2007) to the most recent years natural logarithm of internationally comparable (i.e., purchasing power parity) (2014–2016). Taking the means during each of these two periods excludes the acute GDP per person; this measure has a relatively good linear fit with life evaluations, effects of the GFC and its recovery. NATURE COMMUNICATIONS | (2019)10:161 | https://doi.org/10.1038/s41467-018-08002-2 | www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-018-08002-2 Regression models for well-being (annual changes). First, we consider the year- rate of change in a proxy variable was observed in 10% of all countries over the 11- by-year changes, captured in an FE model by including an indicator (constant year observation period, it is feasible that the variable could change at this rate in offset) di for each country (this is equivalent to subtracting the mean value for each any country over the next three decades. This differs from the full range of possible country): changes, for which we have no first principles understanding, but which is likely to 6 be greater than our feasible range given the possibility of unforeseeable, radical X transformation. To better characterize the distribution of likely futures, we also Sit ¼ a þ di þ bFE j xjit þ ν i þ εit : ð3Þ j¼1 make projections using the 30th, 50th, and 70th percentiles of recent trends, but we focus on the 10th and 90th percentiles for our primary analysis. For the purpose of illustrating large changes in our predictor variables, Supplementary Figure 6 shows recent trends (with 3-year smoothing) in the nine Alternatively, we may model the 1-year changes in an FD equation, as follows: countries which exhibited the most variation in each variable. 6 X The estimation of recent trends in our predictor variables and in SWB is ΔSit ¼ bFD j Δxjit þ ν i þ εit : ð4Þ complicated by two issues. First, the panel of available data is incomplete; that is, j¼1 there are data gaps for some country-year variable entries. Table 1 shows the number of countries with data by year or period for our various approaches. In order to estimate trends that are not biased by a changing composition of With finite samples, (3) and (4) do not give identical estimates. Moreover, countries, a subset of countries with complete data can be used, or data can be because our panel is not perfectly balanced, some observations are dropped when appropriately weighted in order to account for features of the incomplete panel. estimating (4). Because of this, and the inherent higher efficiency of the FE Secondly, the GFC had a significant impact on some countries in 2008, yet these estimator, we favor (3) but show estimates of both. effects were somewhat temporary. The short-term changes following the GFC may As described in the Methods section, we wish to be able to compare the not contribute constructively towards estimating the longer-run changes that might estimated effects of one predictor variable to another—that is, to ascribe relative inform scenarios to 2050. Fortunately, the Gallup World Poll time series is now importance to changes in different predictor variables. For the FD estimator, this is long enough that estimates can be made even after excluding several years straightforward, as we can construct normalized versions of observed 1-year following the start of the GFC. changes, in analogy to (3): ΔSit X 6 Δxjit Recent trends (country estimates). We therefore consider the following three ¼ βFD j þ ν~i þ ~εit : ð5Þ estimation models of recent time trends in our key variables: σ ΔS σ Δxj j¼1 First, a differences model: Take averages for each country for each period (2005–2007, 2014–2016), and calculate the difference. This avoids the GFC, but entails a two-step process to calculate confidence intervals for the differences. The term βj in (5) explains the distribution of 1-year changes in S, such that βj is Second, a 2P regression, excluding GFC model: A one-step method which the number of standard deviations of change in ΔS associated with 1 s.d. change in provides estimates of confidence intervals is to carry out a regression Δxj, where these variances are calculated across observed 1-year changes. In order to express the relative importance of estimates from the FE estimator, xit ¼ xi0 þ Bi π t þ εit ; ð9Þ (3), we use the same standard deviations of FD to transform the bFE j in (3) to σ Δxj where πt is an indicator variable for periods after/before the GFC: FE FE βj % b : ð6Þ σ ΔS j ! 1; t & 2014; πt ¼ ð10Þ 0; otherwise Estimating βFE FD j has the advantages over βj mentioned above (efficiency, more inclusion) for FE. and where observations are estimated with weights 0 1 ; t ' 2007; Regression models for well-being (2P changes). The third method constructs an Ni B t' 2007 equation with only two observations per country, the early period and the later wit ¼ B @ 0; 2008 ' t ' 2013; ð11Þ period (2005–2007 and 2014–2016), in order to model longer-run changes δxj and 1 ; t & 2014: i N δS. In this context, the FE and FD estimators are identical; thus, we write simply t& 2014 6 X δSit ¼ b2P j δxjit þ ν i þ εit : ð7Þ j¼1 Here Nt'i 2007 is the number of observations of country i in the early period. The point estimates of Bi in (9) are identical to the difference in period means for country i, but the standard error is calculated correctly. Analogously to the methods above, we define the standardized coefficients In order to convert the estimates Bi of change between the two periods to an estimate of the annual rate of change, we divide by 9, the number of years between σ δxj the middles of the two periods. β2P j % b2P ð8Þ σ δS j Third, and lastly, an all-year (annual) regression model: The annual rate of change can be estimated for country i and variable x using all years t for which xit using the standard deviation of longer-run changes across countries. This estimate has fewer observations than those modeling annual changes, and the confidence intervals are slightly looser. Table 1 Completeness of panel data As mentioned in the Discussion, we also consider moderation effects between variables in the 2P model by testing pairwise interaction terms, detailed in Year Any data All data Balanced ≥2007 Any <2008 Any >2013 Any (2P) Supplementary Note 3. We find a significant moderation effect only between GDP 2005 27 21 118 109 and freedom of choice. 2006 85 77 118 109 2007 99 92 64 118 109 2008 104 100 64 Regression models for well-being (results). Estimates of the models for life 2009 109 105 64 evaluations are described in Supplementary Note 2 and summarized in Supple- 2010 118 110 64 mentary Table 4. In addition, we report analogous estimates for our other measures 2011 138 130 64 of well-being, namely positive affect (Supplementary Table 5), negative affect 2012 134 122 64 2013 130 119 64 (Supplementary Table 6), and affect balance (Supplementary Table 7). We also 2014 137 122 64 129 109 carry out and explain an estimate at the regional level (reported in Supplementary 2015 136 120 64 129 109 Table 11). 2016 133 116 64 129 109 The first column lists the number of countries with coverage within our core set of variables for Recent trends (overview). We define future scenarios by projecting different each year of the Gallup World Poll, after some imputation done in accordance with the data quantiles of recent country trends in predictor variables into the future. Rather appendix of Helliwell et al.8. The All data column lists the number of countries with all of our than defining complex and arbitrary story lines, we represent the range of feasible core variables. The Balanced column shows the number of countries in the largest balanced outcomes by defining idealized scenarios. We denote the total feasible range of (complete, rectangular) panel covering all years after 2006. For our two-period model, we use the countries with at least one observation in the early period and at least one observation in the possible rates of change as that encompassed by the 10th to 90th percentile later period observed range among all countries. Thus, we use the assumption that, if a given 6 NATURE COMMUNICATIONS | (2019)10:161 | https://doi.org/10.1038/s41467-018-08002-2 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-018-08002-2 ARTICLE Table 2 Estimated recent global trends and distribution (1) Life (2) (3) (4) Log(GDP/ (5) Life (6) (7) Freedom of (8) (9) Social today −Affect +Affect capita) expectancy Giving choice Corruption support Global trend (country .003 .002† .0002 .018† .29† −.0004 .004† −.002 −.002* weights): 2P (.005) (.0006) (.0005) (.002) (.020) (.001) (.0010) (.0009) (.0006) Global trend (country −.002 .002† −8e−05 .015† .27† −.0006 .003* −.002 −.002† weights): annual (.006) (.0006) (.0006) (.002) (.022) (.001) (.001) (.0009) (.0007) Global trend (pop’n weights): −.006 .002 .0002 .037† .25† −.0006 .004* −.003* −.0007 2P (.023) (.002) (.0008) (.009) (.023) (.002) (.001) (.001) (.0009) Global trend (pop’n weights): −.012 .003 −.0003 .033* .22† −.002 .003 −.004 −.001 annual (.025) (.002) (.0009) (.011) (.026) (.003) (.002) (.001) (.0009) Fraction p ≤.05: 2P .32 .23 .18 .76 1.00 .27 .27 .34 .21 Fraction p ≤.05: annual .29 .31 .25 .81 .98 .30 .33 .31 .26 10th %ile: 2P −.073 −.005 −.008 −.003 .12 −.016 −.008 −.013 −.008 10th %ile: annual −.099 −.005 −.009 −.004 .12 −.018 −.010 −.018 −.012 30th %ile: 2P −.017 −.002 −.002 .008 .16 −.006 −.002 −.006 −.003 30th %ile: annual −.037 −.0006 −.003 .009 .16 −.007 −.002 −.009 −.004 50th %ile: 2P .006 .001 .0006 .017 .22 −.001 .002 −.003 −.001 50th %ile: annual −.0002 .003 .0009 .020 .23 −.003 .004 −.003 −.001 70th %ile: 2P .034 .005 .003 .028 .32 .004 .011 .001 .0006 70th %ile: annual .031 .009 .005 .029 .33 .003 .011 .002 .003 90th %ile: 2P .077 .010 .009 .043 .55 .016 .017 .010 .006 90th %ile: annual .086 .018 .012 .047 .51 .017 .021 .011 .010 The first four rows show estimated global trends for the period 2005–2016 and between the periods 2005–2007 and 2014–2016. The 2P versions of the five quantiles of the across-countries distributions are used to build our future scenarios Significance: †.1%, *1% 5%, 10% was observed, according to the following equation: our non-material predictor variables. Note that future changes in predictor vari- xit ¼ ai þ bi t þ εit : ð12Þ ables could potentially occur more rapidly than has occurred in any country over the observed timeframe, but our approach focuses on the most likely range based on past experience. The OECD projections use two global economic scenarios, devised to explore Recent trends (global and regional estimates). For directly estimating global (or possible futures for major environmental challenges, including climate change. regional, multi-country) trends, we follow a similar procedure as the 2P and annual From these Shared Socioeconomic Pathways27, we bracket the range of possibilities approaches described above, but using two alternative approaches to weighting for income and population growth by selecting the two most contrasting scenarios, observations. In one, we use weights which count each country equally, and in the rapid population growth, slow economic growth, weak convergence (SSP3) and another we multiply the weights described previously by country populations. the slow population growth, rapid economic growth, strong convergence (SSP5) as Thus, we have four variants of estimated global trends: for each variable there is a our OECD pessimistic and OECD optimistic Material trends scenarios, 2P version and an annual version, and for each of those there are country weights respectively. and population weights. We project each of the sets of trends forward to 2050, starting from 2016 values By contrast, we consider countries as individual observations when thinking in each country. Country level means derived from dichotomous variables have about the global distribution of trends, and thus calculate, for example, the 10th natural saturation points at 0 and 1; otherwise, there are no restrictions on percentile trend from the unweighted set of countries. projected values. We use OECD projections of population change by country in order to calculate globally averaged life evaluations for all humans. Recent trends (results). We estimate the 2P and annual models for each country, To test the robustness of our findings, we undertake each of our projections and for the world as a whole. Table 2 present the results. In general, we find using alternative models. Tabular summary data for the three sets of scenarios excellent agreement between the 2P and annual estimates. Although the annual using 2P, FE, and XS coefficients are given in Supplementary Tables 12 and 13, and estimates use more observations and should therefore be more efficient, they have Supplementary Table 14, respectively. These same values are also presented in slightly higher standard errors, likely reflecting the short-term variation during the summary plots in Supplementary Figure 7, for both weighted and unweighted GFC. means. Overall, the trend in life evaluations, measured as per-year change on an 11- We also provide maps to represent the population-weighted projections in our point scale, is very small (insignificant). The material variables income and life scenarios using our three different sets of coefficients. Supplementary Figure 2 expectancy are globally increasing significantly, according to all our specifications. shows the same cases as in our main text, which use the 2P coefficients, but There are also significant global trends in the non-material variables, whose includes in the middle row the material trend scenarios. They are not included in changes are measured as annual change on a 0–1 scale. Freedom and corruption the main text due to their geographic uniformity. Supplementary Figure 3 and 4 are both improving in all specifications, while social support is getting worse show our projections under the alternative specifications of XS and FE coefficients, (diminishing), and significantly so averaged across countries. Thus, the predictor respectively. Finally, Supplementary Figure 5 shows projections for affect balance, variables are significantly trending but not all in an optimistic or pessimistic using the 2P coefficients. direction. Table 2 also reports, for each variable, the fraction of countries with statistically Code availability. No special code was used in our analysis, which proceeded as significant (p ≤ .05) rates of change. These fractions are in all cases considerably described in our methods. Requests for scripts needed to reproduce our results will >5%, suggesting that trends, both positive and negative, are sustained (and real). be reviewed and made available on a case-by-case basis by the corresponding This fact is central to our argument in that it supports our assumption that we can author. gauge reasonable rates of change based on the distribution of changes in the recent past. The remaining rows of Table 2 show the 10th and 90th, and intermediate, percentiles from these distributions. We consider these to correspond to plausible Data availability optimistic and pessimistic trends in the future scenarios. In general, these values Our country-mean annual time series data come from the 2017 World Happiness are more extreme for the annual model than the 2P model, and we use the latter in Report. Its Statistical Appendix8 contains annual country averages of several order to remain conservative in our scenarios. variables for the entire span of the Gallup World Poll, which began with partial coverage in 2005. Helliwell et al.8 provide further information on the data and on possible imputation methods to fill gaps in some of the coverage. Future scenarios. We present three sets of scenarios, each with more and less optimistic and pessimistic cases. Two sets focus on changes in material circum- stances—the OECD macroeconomic projections, and our material variable trends, Received: 27 April 2018 Accepted: 10 December 2018 which project 10th, 30th, 50th, 70th, and 90th percentile rates of change for all countries. The third scenario pair projects the same quantiles of rates of change for NATURE COMMUNICATIONS | (2019)10:161 | https://doi.org/10.1038/s41467-018-08002-2 | www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-018-08002-2 References 23. Clark, A. E., Frijters, P. & Shields, M. A. Relative income, happiness, and 1. Samir, K. C. & Wolfgang, L. The human core of the shared socioeconomic utility: an explanation for the Easterlin paradox and other puzzles. J. Econ. Lit. pathways: population scenarios by age, sex and level of education for all 46, 95–144 (2008). countries to 2100. Glob. Environ. Change 42, 181–192 (2017). 24. Knack, S. & Keefer, P. Does social capital have an economic payoff? a cross- 2. Hsiang, S. et al. Estimating economic damage from climate change in the country investigation. Q. J. Econ. 112, 1251–1288 (1997). United States. Science 356, 1362–1369 (2017). 25. Helliwell, J. F., Barrington-Leigh, C. P., Harris, A. & Huang, H. in 3. Diener, E., Oishi, S. & Tay, L. Advances in subjective well-being research. Nat. International Differences in Well-Being (eds Diener, E., Helliwell, J. F. & Hum. Behav. 2, 253-260 (2018). Kahneman, D.) 213–229 (Oxford University Press, Oxford, 2010). 4. Global Happiness Council. Global Happiness Policy Report (Global Happiness 26. Layard, R., Clark, A. & Senik, C. in World Happiness Report (eds John, H., Council, 2018). Richard, L. & Jeffrey, F.) Ch. 3 (Earth Institute, 2012). 5. Zhang, X., Zhang, X. & Chen, X. Happiness in the air: how does a dirty sky 27. Dellink, R., Chateau, J., Lanzi, E. & Magné, B. Long-term economic growth affect mental health and subjective well-being? J. Environ. Econ. Manag. 85, projections in the shared socioeconomic pathways. Glob. Environ. Change 42, 81–94 (2017). 200–214 (2017). 6. Tella, R. D., MacCulloch, R. J. & Oswald, A. J. The macroeconomics of happiness. Rev. Econ. Stat. 85, 809–827 (2003). 7. Helliwell, J. F., Huang, H., Grover, S. & Wang, S. Empirical linkages Acknowledgements between good governance and national well-being. J. Comp. Econ. 46, C.B.-L. acknowledges financial support from Canada’s Social Sciences and Humanities 1332–1346 (2018). Research Council grant 435-2016-0531. E.D.G. acknowledges financial support from the 8. Helliwell, J., Huang, H. & Wang, S. in World Happiness Report 2017 (eds European Research Council (ERC) under the European Union’s Horizon 2020 research Helliwell, J., Layard, R. & Sachs, J.) Ch. 2 (Sustainable Development Solutions and innovation programme (grant agreement no. 682602), and from the Spanish Min- Network, 2017). istry of Science, Innovation and Universities, through the María de Maeztu program for 9. Tay, L., Herian, M. N. & Diener, E. Detrimental effects of corruption and Units of Excellence (MDM-2015-0552). subjective well-being: whether, how, and when. Soc. Psychol. Pers. Sci. 5, 751–759 (2014). Author contributions 10. Aknin, L. B. et al. Prosocial spending and well-being: cross-cultural evidence C.B.-L. and E.D.G. designed the research. C.B.-L. performed the data analysis and for a psychological universal. J. Pers. Soc. Psychol. 104, 635–652 (2013). modeling. C.B.-L. and E.D.G. wrote the paper. 11. Pereira, H. M. et al. Scenarios for global biodiversity in the 21st century. Science 330, 1496–1501 (2010). 12. Rahmstorf., S. A semi-empirical approach to projecting future sea-level rise. Additional information Science 315, 368–370 (2007). Supplementary Information accompanies this paper at https://doi.org/10.1038/s41467- 13. Thomas, C. D. et al. Extinction risk from climate change. Nature 427, 145 018-08002-2. (2004). 14. Mastrandrea, M. D. & Schneider, S. H. Probabilistic integrated assessment of Competing interests: The authors declare no competing interests. “dangerous” climate change. Science 304, 571–575 (2004). 15. Warszawski, L. et al. The inter-sectoral impact model intercomparison project Reprints and permission information is available online at http://npg.nature.com/ (ISI–MIP): project framework. Proc. Natl Acad. Sci. USA 111, 3228–3232 reprintsandpermissions/ (2014). 16. Pachauri, R. K. et al. Climate Change 2014: Synthesis Report. Contribution of Journal peer review information: Nature Communications thanks the anonymous Working Groups I, II and III to the Fifth Assessment Report of the reviewers for their contribution to the peer review of this work. Peer reviewer reports are Intergovernmental Panel on Climate Change (IPCC, Geneva, 2014). available. 17. Helliwell, J. F., Huang, H. & Wang, S. in Happiness Report 2016 (eds John, H., Richard, L. & Jeffrey, F.) Ch. 2 (Sustainable Development Solutions Network, Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in 2016). published maps and institutional affiliations. 18. Easterlin, R. A. in Nations and Households in Economic Growth: Essays in Honour of Moses Abramovitz (eds David, P. A. & Reder, M. W.) 98–125 Open Access This article is licensed under a Creative Commons (Academic Press, New York, 1974). Attribution 4.0 International License, which permits use, sharing, 19. Easterlin, R. A., McVey, L. A., Switek, M., Sawangfa, O. & Zweig, J. S. The adaptation, distribution and reproduction in any medium or format, as long as you give happiness–income paradox revisited. Proc. Natl Acad. Sci. USA 107, 22463- appropriate credit to the original author(s) and the source, provide a link to the Creative 22468 (2010). Commons license, and indicate if changes were made. The images or other third party 20. Oswald, A. J. & Powdthavee, N. Does happiness adapt? A longitudinal study material in this article are included in the article’s Creative Commons license, unless of disability with implications for economists and judges. J. Public Econ. 92, indicated otherwise in a credit line to the material. If material is not included in the 1061–1077, 6 (2008). article’s Creative Commons license and your intended use is not permitted by statutory 21. Tella, R. D., New, J. H.-D. & MacCulloch, R. Happiness adaptation to income regulation or exceeds the permitted use, you will need to obtain permission directly from and to status in an individual panel. J. Econ. Behav. Organ. 76, 834–852 the copyright holder. To view a copy of this license, visit http://creativecommons.org/ (2010). licenses/by/4.0/. 22. Helliwell, J. F.& Barrington-Leigh, C. in The Social Cure: Identity, Health, and Well-being (eds Jetten, J., Haslam, C. & Haslam, S. A.) 55–71 (Taylor & Francis 2011). © The Author(s) 2019 8 NATURE COMMUNICATIONS | (2019)10:161 | https://doi.org/10.1038/s41467-018-08002-2 | www.nature.com/naturecommunications Supplementary Information Feasible future global scenarios for human life evaluations Christopher Barrington-Leigh, Eric Galbraith Published in Nature Communications, 2019 1 Supplementary Figures 2 OECD material projections No interactions Material trends Non-material trends 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 OECD material projections GDP⇥Freedom Material trends Non-material trends 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 Conflating interactions OECD material projections Material trends Non-material trends 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 Global average life evaluation in 2050 Supplementary Figure 1: Projections using alternate model specifications incorporating interaction effects. The top panel is based on our preferred model, while the second and third show projections made using alternative specifications (4) and (8) in Supplementary Table 8. 3 A: OECD optimistic material trends (changes) B: OECD pessimistic material trends (changes) C: Material trends (optimistic) D: Material trends (pessimistic) E: Non-material trends (optimistic) F: Non-material trends (pessimistic) Supplementary Figure 2: Geographic distribution of changes in life evalua- tion (2050 scenarios, 2P coefficients). Projections based on two-period (2P) coefficients. 4 A: OECD optimistic material trends (changes) B: OECD pessimistic material trends (changes) C: Material trends (optimistic) D: Material trends (pessimistic) E: Non-material trends (optimistic) F: Non-material trends (pessimistic) Supplementary Figure 3: Geographic distribution of changes in life evaluation (2050 scenarios, XS coefficients). Projections based on cross-sectional (XS) coefficients. 5 A: OECD optimistic material trends (changes) B: OECD pessimistic material trends (changes) C: Material trends (optimistic) D: Material trends (pessimistic) E: Non-material trends (optimistic) F: Non-material trends (pessimistic) Supplementary Figure 4: Geographic distribution of changes in life evaluation (2050 scenarios, FE coefficients). Projections based fixed-effect (FE) coefficients. 6 A: OECD optimistic material trends (changes) B: OECD pessimistic material trends (changes) C: Material trends (optimistic) D: Material trends (pessimistic) E: Non-material trends (optimistic) F: Non-material trends (pessimistic) Supplementary Figure 5: Geographic distribution of changes in affect bal- ance (2050 scenarios, 2P coefficients). Projections based on two-period (2P) coefficients. 7 Supplementary Figure 6: Strong trends in life evaluation, affect balance, and predictor variables. Values are smoothed with a 3-year rolling average. Countries listed are those with the largest changes, in each case, during the observation period. These developments are shown for illustration only; the trends used for our projec- tions are taken from the full distribution across countries of average rates of change over the entire period. 8 A B OECD material projections OECD material projections Material trends Material trends Non-material trends Non-material trends 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 Global average life evaluation in 2050 Global average life evaluation in 2050 C D OECD material projections OECD material projections Material trends Material trends Non-material trends Non-material trends 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 Global average life evaluation in 2050 Global average life evaluation in 2050 Supplementary Figure 7: Range of scenario outcomes for global mean life evaluations in 2050. (A) and (C) show mean life evaluation across countries, on a 0–10 scale , while (B) and (D) show population-weighted global means. (A) and (B) are calculated using our estimated coefficients in the two-period model, while (C) and (D) are from our fixed effects model. In each panel, the coloured bars show the ranges of feasible life evaluations attributable solely to changes spanning the pessimistic to optimistic scenarios for the each of the nonmaterial trends (green), material trends (orange), and OECD factors (blue). 9 Supplementary Tables 10 Country-weighted Pop’n-weighted d d µ std µ std min max 10%ile dy 90%ile dy Life today .027±.049 .53 .21±.19 .56 1.60 1.37 .099 .086 a↵ect .020±.005 .057 .034±.017 .056 .10 .18 .005 .018 + a↵ect .002±.005 .053 .0004±.009 .041 .14 .13 .009 .012 log(GDP/capita) .16±.016 .17 .25±.063 .18 .35 .68 .004 .047 Life expectancy 2.6±.18 1.96 2.4±.19 1.43 .53 13.4 .12 .51 Giving .003±.010 .11 .007±.017 .092 .24 .32 .018 .017 Freedom of choice .034±.009 .095 .038±.013 .081 .30 .24 .010 .021 Corruption .021±.008 .086 .030±.015 .076 .34 .17 .018 .011 Social support .017±.006 .061 .006±.010 .048 .24 .16 .012 .010 Supplementary Table 1: Summary statistics: two-period changes. As in Sup- plementary Table 2, but showing changes between the mean level from 2005–2007 in each country, and its mean between 2014–2016. Country-weighted values show a 17% increase in GDP/capita (i.e., log increase of 0.16), a 2.6-year increase in life expectancy, a 1.7% decrease in affirmative responses to the social support question, and no significant change in life evaluations. 11 Country-weighted Pop’n-weighted µ std µ std min max Life today .002±.015 .39 .014±.031 .34 2.00 1.65 a↵ect .003±.002 .047 .004±.003 .039 .17 .27 + a↵ect .001±.002 .051 .0002±.004 .046 .22 .20 log(GDP/capita) .019±.001 .039 .036±.004 .037 .44 .21 Life expectancy .29±.010 .27 .26±.010 .19 .75 1.73 Giving .002±.002 .070 .0004±.007 .071 .30 .44 Freedom of choice .005±.003 .071 .004±.009 .076 .30 .32 Corruption .004±.003 .059 .003±.003 .044 .27 .25 Social support .0007±.002 .054 .001±.005 .050 .36 .21 Supplementary Table 2: Summary statistics: annual changes. As in Supplemen- tary Table 3, but showing statistics describing the annual changes across countries, generated by taking first differences in every case when a variable is observed in a country over two successive years. All values are given as the change per year. Country-weighted Pop’n-weighted µ std µ std min max N Life today 5.4±.037 1.15 5.3±.069 .95 2.7 8.0 1190 a↵ect .26±.003 .078 .25±.007 .071 .10 .70 1190 + a↵ect .71±.003 .11 .73±.007 .091 .36 .94 1190 log(GDP/capita) 9.1±.044 1.23 9.1±.058 .96 6.4 11.7 1190 Life expectancy 61.0±.31 8.7 61.8±.51 7.2 36.2 76.4 1190 Giving .0003±.005 .16 .004±.013 .16 .32 .54 1190 Freedom of choice .72±.005 .14 .74±.008 .12 .26 .98 1190 Corruption .75±.007 .19 .82±.009 .12 .035 .98 1190 Social support .81±.004 .12 .76±.013 .13 .29 .99 1190 Supplementary Table 3: Summary statistics: annual time series. For each sub- jective well-being metric and predictor variable, the mean (µ) and standard deviation (std) among all countries are given (Country-weighted). The mean and standard de- viation are also given after weighting the value in each country by its population, in order to estimate the average value for all humans (Pop’n-weighted). Also shown are the minimum (min) and maximum (max) among all countries in the dataset, as well as the total number of country-year observations (N). 12 Life today (1) (2) (3) (4) (5) (6) (7) (8) (9) log(GDP/capita) .35† .93† 1.17† .66+ .36† .96† .19+ .14† .095† (.062) (.23) (.30) (.37) (.065) (.24) (.11) (.034) (.024) † Social support 2.2 1.38† 1.05† 2.6† .23 † .15† .29† .14 † .19 † (.40) (.32) (.29) (.65) (.042) (.034) (.072) (.040) (.044) † † Life expectancy .030 .046 .018 .015 .22 .34 .060 .030 .031 (.008) (.019) (.024) (.015) (.060) (.14) (.059) (.025) (.013) ? † † ? ? † ? ? † Freedom of choice 1.03 .83 .65 1.70 .13 .11 .33 .12 .15 (.33) (.24) (.20) (.54) (.042) (.031) (.10) (.036) (.045) † + † + Giving .92 .26 .31 .016 .13 .036 .003 .058 .048 (.27) (.20) (.17) (.39) (.038) (.028) (.080) (.031) (.037) + + + Corruption .67 .43 .18 1.01 .11 .070 .17 .028 .065 (.29) (.25) (.28) (.48) (.047) (.040) (.081) (.041) (.037) model XS FE FD 2P XS FE 2P FD FE normalization XS XS 2P FD FD Year f.e. 12 12 12 12 12 12 12 Nclusters 149 149 139 149 149 139 149 obs. 1212 1212 971 108 1212 1212 108 971 1212 R2 (adj) .755 .130 .060 .347 .755 .130 .347 .064 .130 R2 (overall) .64 .64 .64 Significance: 0.1%† 1%? 5% 10%+ Supplementary Table 4: Estimates of life evaluations. Estimates of models ac- counting for cross-sectional (XS) and temporal (FE, FD, 2P) variation in life evalua- tions are shown (columns 1–4), along with standardized coefficient versions (columns 5-9) of the same estimates. The estimates in column (4) are used for our main pro- jections to relate changes in our predictor variables to changes in life evaluations. The relative importance of those predictor variables for explaining variation in life evaluations can be gauged by looking at the standardized coefficients. We prefer column (7) as a measure of explanatory power for changes over time. 13 A↵ect balance (1) (2) (3) (4) (5) (6) (7) (8) (9) log(GDP/capita) .018 .031 .13 .13 .14 .24 .23 .100† .016 (.013) (.043) (.050) (.059) (.10) (.34) (.11) (.028) (.023) † † † Social support .50 .27 .20 .049 .40† .22† .033 .15 † .19 † (.071) (.070) (.057) (.15) (.056) (.055) (.098) (.042) (.049) Life expectancy .0004 .008 .005 .008 .020 .43 .19 .038 .027 (.002) (.004) (.005) (.006) (.11) (.21) (.13) (.026) (.013) † ? ? † ? ? ? Freedom of choice .40 .16 .10 .035 .38 .15 .040 .11 .15 (.067) (.049) (.036) (.10) (.064) (.047) (.12) (.036) (.046) † ? † ? Giving .16 .052 .027 .18 .17 .054 .23 .028 .050 (.044) (.049) (.040) (.071) (.046) (.051) (.087) (.042) (.047) + + Corruption .074 .048 .026 .15 .090 .059 .15 .024 .037 (.038) (.045) (.048) (.093) (.047) (.055) (.094) (.039) (.035) model XS FE FD 2P XS FE 2P FD FE normalization XS XS 2P FD FD Year f.e. 12 12 12 12 12 12 12 Nclusters 149 149 139 149 149 139 149 obs. 1208 1208 965 108 1208 1208 108 965 1208 R2 (adj) .468 .083 .043 .151 .468 .083 .151 .049 .083 R2 (overall) .28 .28 .28 Significance: 0.1%† 1%? 5% 10%+ Supplementary Table 5: Estimates of positive affect. See Supplementary Table 4. 14 + a↵ect (1) (2) (3) (4) (5) (6) (7) (8) (9) log(GDP/capita) .003 .070? .076+ .10? .037 .77? .31? .074 .056? (.010) (.022) (.040) (.033) (.11) (.24) (.100) (.037) (.017) † Social support .24 .089 .033 .024 .27† .099 .028 .035 .094 (.052) (.041) (.036) (.095) (.058) (.046) (.11) (.040) (.044) Life expectancy .0006 .002 .002 .004+ .046 .14 .16+ .009 .009 (.001) (.001) (.003) (.002) (.11) (.11) (.093) (.018) (.008) † † † + † † + † † Freedom of choice .33 .11 .090 .092 .44 .15 .18 .13 .16 (.042) (.029) (.027) (.050) (.056) (.039) (.099) (.040) (.041) † ? ? † ? ? ? Giving .16 .085 .077 .11 .24 .13 .22 .12 .12 (.031) (.032) (.028) (.043) (.047) (.048) (.090) (.042) (.046) Corruption .027 .002 .009 .008 .046 .003 .013 .012 .002 (.029) (.030) (.031) (.057) (.049) (.052) (.097) (.037) (.035) model XS FE FD 2P XS FE 2P FD FE normalization XS XS 2P FD FD Year f.e. 12 12 12 12 12 12 12 Nclusters 149 149 139 149 149 139 149 obs. 1209 1209 966 108 1209 1209 108 966 1209 R2 (adj) .482 .089 .036 .219 .482 .089 .219 .037 .089 R2 (overall) .23 .23 .23 Significance: 0.1%† 1%? 5% 10%+ Supplementary Table 6: Estimates of negative affect. See Supplementary Table 4. 15 a↵ect (1) (2) (3) (4) (5) (6) (7) (8) (9) log(GDP/capita) .014+ .040 .047+ .026 .22+ .61 .070 .075† .033 (.008) (.031) (.025) (.040) (.12) (.48) (.11) (.023) (.026) † † † Social support .26 .18 .17 .064 .40† .29† .066 .20 † .20 † (.046) (.052) (.047) (.092) (.072) (.080) (.095) (.055) (.057) + + + + Life expectancy .0009 .006 .003 .004 .10 .65 .13 .050 .032 (.001) (.003) (.003) (.004) (.11) (.34) (.14) (.029) (.017) Freedom of choice .072 .043 .015 .058 .14 .080 .10 .022 .062 (.046) (.035) (.022) (.073) (.085) (.065) (.13) (.035) (.051) + + Giving .004 .034 .048 .078 .008 .070 .15 .080 .050 (.031) (.029) (.025) (.049) (.063) (.061) (.094) (.041) (.044) † + ? † + ? + Corruption .10 .049 .017 .15 .24 .12 .23 .026 .060 (.025) (.027) (.030) (.056) (.061) (.065) (.087) (.038) (.033) model XS FE FD 2P XS FE 2P FD FE normalization XS XS 2P FD FD Year f.e. 12 12 12 12 12 12 12 Nclusters 149 149 139 149 149 139 149 obs. 1211 1211 970 108 1211 1211 108 970 1211 R2 (adj) .221 .088 .046 .045 .221 .088 .045 .052 .088 R2 (overall) .002 .002 .002 Significance: 0.1%† 1%? 5% 10%+ Supplementary Table 7: Estimates of affect balance. See Supplementary Table 4. 16 Life today (1) (2) (3) (4) (5) (6) (7) (8) log(GDP/capita) .19+ .21+ .20+ .24 .19+ .19+ .21+ .24 (.11) (.11) (.11) (.11) (.11) (.11) (.11) (.11) Social support .29† .21 .29† .27† .29† .30† .29† .18+ (.072) (.098) (.072) (.073) (.072) (.079) (.072) (.10) + Life expectancy .060 .055 .062 .10 .061 .064 .064 .091 (.059) (.058) (.061) (.063) (.059) (.059) (.060) (.062) ? ? ? † ? ? ? † Freedom of choice .33 .33 .33 .48 .33 .32 .30 .57 (.10) (.11) (.10) (.12) (.10) (.11) (.11) (.13) Giving .003 .005 .005 .020 .006 .004 .030 .0003 (.080) (.082) (.083) (.076) (.081) (.081) (.10) (.072) + + Corruption .17 .18 .18 .17 .15 .17 .17 .032 (.081) (.081) (.11) (.083) (.087) (.082) (.082) (.096) log(GDP/capita) ⇥ Social support .11 .14 (.10) (.097) + log(GDP/capita) ⇥ Corruption .015 .22 (.11) (.12) log(GDP/capita) ⇥ Freedom of choice .26? .39† (.099) (.12) Social support ⇥ Corruption .037 (.064) Social support ⇥ Freedom of choice .045 (.072) Freedom of choice ⇥ Giving .077 (.097) model 2P 2P 2P 2P 2P 2P 2P 2P AIC 267 268 269 263 269 269 268 263 obs. 108 108 108 108 108 108 108 108 R2 (adj) .347 .346 .340 .376 .341 .342 .345 .387 Significance: 0.1%† 1%? 5% 10%+ Supplementary Table 8: Estimates of life evaluations including interaction terms. Column (1) reproduces our preferred 2P model (column 7 of Supplementary Table 4). Columns (2) through (7) are the same but with the addition of one inter- action term each. Column (8) includes all terms between material and non-material predictor variables. All coefficients are presented normalized to 2P variation. 17 Freedom of choice log(GDP/capita) Life expectancy Social support A↵ect balance Corruption + a↵ect a↵ect Giving Life today .51† .59† .21? .83† .77† .47† .57† .23? .75† 5 5 5 5 5 5 5 (<10 ) (<10 ) (.007) (<10 ) (<10 ) (<10 ) (<10 ) (.004) (<10 ) † † † ? † † † † A↵ect balance .90 .78 .28 .23 .37 .64 .35 .49 5 5 5 5 5 (<10 ) (<10 ) (.0004) (.004) (<10 ) (<10 ) (1e-05) (<10 ) † † † † † † † $+$ a↵ect .44 .35 .30 .33 .69 .42 .49 5 5 5 5 (<10 ) (1e-05) (.0002) (3e-05) (<10 ) (<10 ) (<10 ) † † $-$ a↵ect .075 .057 .28 .33 .12 .31† (.36) (.48) (.0004) (2e-05) (.15) (7e-05) log(GDP/capita) .83† .39 † .41 † .014 .71 † 5 5 5 5 (<10 ) (<10 ) (<10 ) (.87) (<10 ) † † † Life expectancy .29 .36 .059 .63 5 (.0003) (1e-05) (.47) (<10 ) † † Corruption .54 .32 .22? 5 (<10 ) (8e-05) (.006) † † Freedom of choice .41 .42 5 5 (<10 ) (<10 ) Giving .087 (.29) Significance: 0.1%† 1%? 5% 10%+ Supplementary Table 9: Pairwise correlations of levels. Pearson correlation coefficients for our country-mean values of our main variables 18 Freedom of choice log(GDP/capita) Life expectancy Social support A↵ect balance Corruption + a↵ect a↵ect Giving Life today .41† .33† .32† .39† .043 .30? .46† .11 .43† 5 5 5 (<10 ) (.0003) (.0003) (1e-05) (.64) (.001) (<10 ) (.22) (<10 ) † † ? + + + A↵ect balance .78 .80 .30 .11 .21 .17 .17 .15 5 5 (<10 ) (<10 ) (.001) (.24) (.026) (.060) (.072) (.091) † + a↵ect .25? .39 .023 .10 .29 ? .22 .064 (.005) (1e-05) (.80) (.29) (.002) (.018) (.48) + a↵ect .091 .14 .23 .002 .056 .18 (.32) (.12) (.016) (.98) (.55) (.052) † ? log(GDP/capita) .14 .15 .39 .058 .27 (.13) (.13) (2e-05) (.54) (.003) Life expectancy .011 .093 .082 .025 (.90) (.32) (.38) (.79) Corruption .39† .061 .008 (2e-05) (.53) (.94) Freedom of choice .14 .061 (.14) (.51) Giving .063 (.50) Significance: 0.1%† 1%? 5% 10%+ Supplementary Table 10: Pairwise correlations of changes. Pearson correlation coefficients for two-period changes of our main variables 19 Life today (1) (2) (3) log(GDP/capita) .35† .23† .19 (.062) (.069) (.083) † Social support 2.2 1.59† 3.1 † (.40) (.35) (.53) † Life expectancy .030 .012 .009 (.008) (.011) (.035) Freedom of choice 1.03? .77 1.60 (.33) (.32) (1.04) Corruption .67 .25 .18 (.29) (.33) (.53) † Giving .92 1.09 .12 (.27) (.73) (.50) model XS XS XS normalization region All countries Sub-Saharan Africa C.I.S. Year f.e. 12 12 12 obs. 1212 284 116 R2 (adj) .755 .335 .629 Nclusters 149 39 11 Significance: 0.1%† 1%? 5% 10%+ Supplementary Table 11: XS model estimated for regions. Column 1 is repeated from Supplementary Table 4. Column 2 is the same model estimated for the subset of countries in sub-Saharan Africa, and Column 3 is the same but estimated for the Commonwealth of Independent States, ie former nations of the Soviet Union. scenario min max mean (nat’l) mean (pop’n) 2016value (nat’l) 2016value (pop’n) Material trends (OECD optimistic) 4.24 7.69 6.05 6.04 5.35 5.24 Material trends (OECD pessimistic) 3.52 7.73 5.72 5.66 5.35 5.24 Material trends (optimistic) 3.67 8.31 6.12 6.01 5.35 5.24 Material trends (pessimistic) 2.76 7.40 5.21 5.10 5.35 5.24 Non-material trends (optimistic) 4.50 8.61 6.86 6.84 5.35 5.24 Non-material trends (pessimistic) 1.09 5.57 3.52 3.44 5.35 5.24 Supplementary Table 12: Scenario projections using two-period (2P) coefficients scenario min max mean (nat’l) mean (pop’n) 2016value (nat’l) 2016value (pop’n) Material trends (OECD optimistic) 3.97 7.40 5.91 5.95 5.35 5.24 Material trends (OECD pessimistic) 3.47 7.72 5.77 5.74 5.35 5.24 Material trends (optimistic) 3.56 8.21 6.01 5.90 5.35 5.24 Material trends (pessimistic) 2.60 7.24 5.05 4.94 5.35 5.24 Non-material trends (optimistic) 4.15 8.03 6.22 6.16 5.35 5.24 Non-material trends (pessimistic) 1.83 6.39 4.27 4.18 5.35 5.24 Supplementary Table 13: Scenario projections using fixed-effect (FE) coefficients 20 scenario min max mean (nat’l) mean (pop’n) 2016value (nat’l) 2016value (pop’n) Material trends (OECD optimistic) 4.36 8.27 6.39 6.34 5.35 5.24 Material trends (OECD pessimistic) 3.39 7.78 5.71 5.62 5.35 5.24 Material trends (optimistic) 3.98 8.63 6.43 6.32 5.35 5.24 Material trends (pessimistic) 2.98 7.63 5.43 5.32 5.35 5.24 Non-material trends (optimistic) 4.83 8.72 6.97 6.93 5.35 5.24 Non-material trends (pessimistic) 0.99 5.51 3.42 3.33 5.35 5.24 Supplementary Table 14: Scenario projections using cross-section (XS) coefficients 21 Supplementary Note 1 (Descriptive statistics) Our main text focuses on one primary measure of subjective well-being (SWB), namely the cognitive evaluation of life question called “Cantril’s Ladder.” However, this supplement includes analysis of two other dimensions of SWB, positive and negative affect, and the difference between them, “affect balance.” For more details, again see [8]. Supplementary Table 3 shows summary statistics for these data, while Supple- mentary Table 2 shows summary statistics for country changes between successive years (first differences, or FD), where they are available. Supplementary Table 1 shows summary statistics for differences between country means over the period 2014–2016 and corresponding means over the period 2005-2007; we call these our two-period (2P) estimates. Pairwise correlations The simplest measure of the predictive relationships we explore and exploit is a bivariate correlation. Supplementary Table 9 shows pair-wise correlations (Pearson coefficients) among country means for our key variables. Supplementary Table 10 shows pair-wise correlations (Pearson coefficients) among changes in our key variables between the two periods in our two-period (2P) data. While country averages of life evaluations are, among our predictor variables, most closely correlated (Supplementary Table 9) with levels of per capita income, changes in life evaluations are more closely correlated (Supplementary Table 10) with changes in social support (R ⇡ .43) or freedom (R ⇡ .46) than they are with changes in income (R ⇡ .39). 22 Supplementary Note 2 (Model estimates) The four models, described in Methods, are identified as “XS” (static model), “FE” (annual changes, fixed effects), “FD” (annual changes, first differences), and “2P” (two-period) and estimated for cognitive evaluations of life in Supplementary Table 4. In our favored dynamic model (2P), changes in “giving” and life expectancy do not predict significant changes in average life evaluation. Tables 5–7 report analogous estimates for the other measures of SWB. In each case, the first four columns present estimates of raw coefficients (b, in the discussion above), while the remaining columns show coefficients standardized in one of three ways. We consider column (7) to give the best indication of the relative importance of different predictor variables for explaining changes in SWB, and we use the corresponding raw coefficients in column (4) for our primary scenario projections. While the importance of income, as represented by standardized coefficients in our model, is lower in the dynamic model than in the cross-sectional one, the underlying effect size is actually larger. That is, a 10% increase in income per capita accounts for an increase in life evaluations of 0.066 (on an 11-point scale) with the dynamic 2P model (i.e., within a country), while a 10% difference in income per capita accounts for a difference of only 0.035 in the static model (i.e., across countries). In the case of our models of year-to-year changes (FD and FE), the dynamic effect is larger still. The latter finding may reflect the impacts of economic fluctuations in the business cycle and the 2008 financial crisis, which our two-period model expressly avoids. More generally, the quantitative difference we observe is consistent with other evidence that income tends to affect well-being more strongly in the short term than over the long run [e.g., 19, 23]. Despite this possible exaggeration of the effect of GDP, which may overestimate the SWB gains to be made from a longer-term rise of GDP, we use the time-series model coefficients in the scenarios that follow. Estimates for country regions In light of the hypothesis that life evaluations should depend more on income in places with lower incomes, Supplementary Table 11 shows estimates of the XS model for sub-Saharan Africa and for the Commonwealth of Independent States. Both regions have a sufficient number of countries to carry out the estimate; however, fewer coeffi- cients are found to be statistically significant with the smaller samples. Contrary to the hypothesis, the ratio of coefficients on GDP/capita and social support are sim- ilar or smaller in the sub-regions as in the global estimate. For this reason, and to achieve reasonable precision in our estimates, we do not split up the planet according 23 to levels of development in the primary estimates of our models. Supplementary Note 3 (Robustness tests of interac- tion effects) In order to test for interaction (sometimes called moderation) effects between our predictor variables, we focus on the four statistically significant variables in our dynamic 2P model used for our primary analysis. Three of these variables are “non- material”, while one, log(GDP/capita), is “material”. Therefore, of the pair-wise combinations among them, three interactions could indicate a problem with our separation of non-material scenarios from material scenarios, while the remaining three interactions concern only the specification within the non-material model. To best assess the size of any interaction effects in our baseline model, we intro- duce six new normalized variables (i.e., all pair-wise combinations of four predictor variables) one-by-one into the normalized-coefficient 2P model. As shown in columns (2) to (7) of Supplementary Table 8, we find non-significant effects for all but one of them. In the case of log(GDP/capita)⇥Freedom of choice (column 4) a signif- icant interaction effect is estimated. Apart from the interaction itself, there is no qualitative (nor significant quantitative) change in the other estimated parameters, as compared with our baseline model (shown in column 1). Because we recognize that our predictor variables are only proxies for material and non-material drivers of life quality, and that they are likely to exhibit multicollinearity, we interpret these findings as relatively reassuring in that our projections, though necessarily crude, represent a reasonable approach. In our 2P analysis, when only the material vari- ables or the non-material variables, but not both, are changing in any scenario, the significant interaction term will remain zero by construction. We also estimate a version of the model including all the interactions between our (statistically significant) material and non-material RHSvs. This estimate is shown in column (8) of Supplementary Table 8. In this case, there is a particularly significant interaction between GDP/capita and Freedom of choice. We emphasize that our data set (119 countries) is too small to conduct model selection procedures, and that comparing models for quality of fit is not a well- constrained undertaking. A formal test of interaction effects should likely stop at a test of significance of interaction coefficients in columns (2) to (7) in order to limit the likelihood of finding spurious relationships. Were the data extensive enough to expand the model beyond the original 6 predictors, numerous other specifications accommodating nonlinear terms and other available national-level variables would be 24 equally reasonable candidates to specifications that are expanded using interaction terms. For this reason, we constrained our base model to a specification already established in the literature. Nevertheless, in order to compare the alternate specifications, Supplementary Table 8 shows the Akaike information criterion (AIC) for each model. Lower values of the AIC indicate a better tradeoff between goodness-of-fit and model simplicity. Only two models, in columns (4) and (8), have a lower AIC than our baseline model (column 1). To assess the impact on our projections of including interaction terms, we repeated our 2P projection calculations using these two models. The results are shown in Supplementary Figure 1. The top panel shows our preferred projection from the main text. The second panel shows projections based on column (4) of Supplementary Table 8, which includes the interaction term between GDP/capita and Freedom of choice. The primary difference from our preferred model is that the projected range of non-material projection outcomes is even larger than in our pre- ferred version. In contrast, the lower panel shows the projections based on column (8), in which all the “conflating” interaction terms between material and non-material statistically significant predictor variables are included. In this case, there is a reduc- tion in the 90/10 range of projections in the non-material scenario, particularly in the outcome of the most optimistic non-material trends. However, the 30/70 range remains quite similar to our preferred model, and our conclusions remain robust in that the range of our projected distributions in the non-material scenario is far greater than those of the two material scenarios. Given that the AIC is the same as for model (4) which shows a very large range of non-material projections, we main- tain our preference for the intermediate results of model (1), particularly given that it is the simplest model and therefore least prone to over-fitting. 25