World Happiness Report 2023 70 Endnotes 1 A country’s average answer to the Cantril ladder question is exactly equivalent to a notion of average underlying satisfaction with life under an assumption of “cardinality:” the idea that the difference between a 4 and a 3 should count the same as the difference between a 3 and a 2, and be comparable across individuals. Some social scientists argue that too little is known about how people choose their answer to the Cantril ladder question to make this assumption and that if it is wrong enough, then rankings based on average survey responses could differ from rankings based on underlying satisfaction with life (Bond & Lang, 2019). Other researchers have concluded that answers to the Cantril ladder question are indeed approximately cardinal (Bloem & Oswald, 2022; Ferrer-i-Carbonell & Frijters, 2004; Kaiser & Oswald, 2022; Krueger & Schkade, 2008). 2 For any pair of countries, the confidence intervals for the means (depicted in Figure 2.1 as whiskers) can be used to gauge which country’s mean is higher than the other, accounting for statistical uncertainty in the measurement of each. The confidence interval for a country’s rank (given in Figure 2.1 as text) represents a range of possible values for the ranking of their mean among all countries, accounting for uncertainty in the measurement of all of the means (following Mogstad et al., 2020). The ranges are constructed so that the chance that the range does not contain the country’s true rank is no more than 5%. 3 Not every country has a survey every year. The total sample sizes are reported in Statistical Appendix 1, and are reflected in Figure 2.1 by the size of the 95% confidence intervals for the mean, indicated by horizontal lines. The confidence intervals are naturally tighter for countries with larger samples. 4 Countries marked with an * do not have survey information in 2022. Their averages are based on the 2020 and/or 2021 surveys. 5 This can be seen as part of a more general Baltic phenomenon. The increase in Estonia’s rank was even larger, from 66th in 2017 to 31st in 2023. Latvia’s increase was also significant, but smaller, from 54th in 2017 to 41st in 2023. These increases reflect the general increases in life evaluations in Central and Eastern Europe shown in Figure 2.3, with the Baltic countries converging faster than average toward Western European levels. 6 The statistical appendix contains alternative forms without year effects (Appendix Table 9), and a repeat version of the Table 2.1 equation showing the estimated year effects (Appendix Table 8). These results continue to confirm that inclusion of the year effects makes no significant difference to any of the coefficients. In these aggregate equations, adding regional or country fixed effects would lower the coefficients on relatively slow moving variables where most of the variance is across countries rather than over time, such as healthy life expectancy and the log of GDP. With equations based on individual observations, such as in Table 2.2 of World Happiness Report 2022, where income and health are measured by individual-level variables, adding country fixed effects makes little difference to any of the coefficients. 7 The definitions of the variables are shown in Technical Box 2, with additional detail in the online data appendix. 8 T he model’s predictive power is little changed if the year fixed effects in the model are removed, with adjusted R-squared falling only from 0.757 to 0.752. 9 For example, unemployment responses at the individual level are available in most waves of the Gallup World Poll. While they show an effect size similar to that found in other research, the coefficient has never been significant, and its inclusion does not influence the size of the other coefficients. 10 Below, we use the term “effect” when describing the coefficients in these regressions; some caveats to this interpretation are discussed later in this section. 11 In the equation for negative affect, healthy life expectancy takes a significant positive coefficient, despite its positive simple correlation with life evaluations in this aggregate dataset. 12 T his influence may be direct, as many have found, e.g. De Neve et al. (2013). It may also embody the idea, as made explicit in Fredrickson’s broaden-and-build theory (Fredrickson, 2001), that good moods help to induce the sorts of positive connections that eventually provide the basis for better life circumstances. 13 S ee, for example, the well-known study of the longevity of nuns, Danner et al. (2001). 14 See Cohen et al. (2003), and Doyle et al. (2006). 15 The prevalence of these feedbacks was documented in Chapter 4 of World Happiness Report 2013, De Neve et al. (2013). 16 W e expected the coefficients on these variables (but not on the variables based on non-survey sources) to be reduced to the extent that idiosyncratic differences among respondents tend to produce a positive correlation between the four survey-based factors and the life evaluations given by the same respondents. This line of possible influence is cut when the life evaluations are coming from an entirely different set of respondents than are the four social variables. The fact that the coefficients are reduced only very slightly suggests that the common- source link is real but very limited in its impact. 17 The coefficients on GDP per capita and healthy life expectancy were affected even less, and in the opposite direction in the case of the income measure, being increased rather than reduced, once again just as expected. The changes were very small because the data come from other sources, and are unaffected by our experiment. However, the income coefficient does increase slightly, since income is positively correlated with the other four variables being tested, so that income is now able to pick up a fraction of the drop in influence from the other four variables. We also performed an alternative robustness test, using the previous year’s values for the four survey-based variables. Because each year’s respondents are from a different random sampling of the national populations, using the previous year’s average data also avoids using the same respondent’s answers on both sides of the equation. This alternative test produced similarly reassuring results as
RkJQdWJsaXNoZXIy NzQwMjQ=