A Reassessment of the Relationship between GDP and Life Satisfaction

The scientific debate on the relation between Gross Domestic Product (GDP) and self reported indices of life satisfaction is still open. In a well-known finding, Easterlin reported no significant relationship between happiness and aggregate income in time-series analysis. However, life satisfaction appears to be strictly monotonically increasing with income when one studies this relation at a point in time across nations. Here, we analyze the relation between per capita GDP and life satisfaction without imposing a functional form and eliminating potentially confounding country-specific factors. We show that this relation clearly increases in country with a per capita GDP below 15,000 USD (2005 in Purchasing Power Parity), then it flattens for richer countries. The probability of reporting the highest level of life satisfaction is more than 12% lower in the poor countries with a per capita GDP below 5,600 USD than in the counties with a per capita GDP of about 15,000 USD. In countries with an income above 17,000 USD the probability of reporting the highest level of life satisfaction changes within a range of 2% maximum. Interestingly enough, life satisfaction seems to peak at around 30,000 USD and then slightly but significantly decline among the richest countries. These results suggest an explanation of the Easterlin paradox: life satisfaction increases with GDP in poor country, but this relation is approximately flat in richer countries. We explain this relation with aspiration levels. We assume that a gap between aspiration and realized income is negatively perceived; and aspirations to higher income increase with income. These facts together have a negative effect on life satisfaction, opposite to the positive direct effect of the income. The net effect is ambiguous. We predict a higher negative effect in individuals with higher sensitivity to losses (measured by their neuroticism score) and provide econometric support of this explanation.

with the coefficient increasing until the 23 rd quantile -corresponding to a GDP interval of 25K-26Kand then decreasing. Figure 1 plots the estimated quantile coefficients presented in the first column of table 2. Furthermore, we repeat the above exercise by partitioning the country-wave observations into 50 quantiles. We present the resulting quantiles' coefficients in Figure 2 and we interpolate a cubic line.
The relationship is clearly monotonic until the 25 th quantile, then it flattens for richer countries. The quadratic and cubic coefficients of the interpolating line are both significant at 1% level, and we can observe a maximum around the 40th quantile, corresponding to a GDP interval 28.3K-2.8.5K. In Figure   3, we display only the coefficients of the 25 richest quantiles, corresponding to the top 50% GDP, and its quadratic interpolation with the 95% confidence interval. We observe that the quadratic interpolation features a peak at the 40 th quantile; from a visual inspection of the figure we note that a monotonic pattern within the 95% confidence interval can be rejected.

Region Based Analysis
In table 4, we partitioned the regional data into 10 quantiles to check the robustness of the results for a different partition of the analysis presented in table 3 of the main text. From column 1, we note that there is an increasing positive effect until the 7th quantile, then the coefficients of the quantile dummies decrease. However, this is true until the 9th quantile since the coefficient of the 9th quantile is negative, reversing the decreasing pattern. Column 2 and 3 show that this reversion at the last quantile disappear when we control for either town size or country effect (in the top panel of Figure 4, we display the value of the coefficient of of the 10 quantile dummy relative to the estimation of column 2). We note a pattern that seem monotonically increasing (apart for the exception of the 2nd quantile) until the 7th quantile, then it is decreasing. This suggest a hump shaped pattern with a maximum in the 7th quantile, corresponding to a regional GDP within the interval 30K-33K. Columns 4 and 5 finally show that the non monotonic pattern is robust to the introduction of a number of individual controls as in the previous table.
In order to check how much of the above results are dependent from the outliers we observed in Figure   1 of the main text , we repeated the analysis above by excluding the 10th quantile (containing both Paris and Brussels) and using the 9th as base level. Results are displayed in table 3 and in the bottom panel of Figure 4, where we note a similar pattern in the analysis with all 10 quantiles; the pattern is generally increasing in the first 7 quantiles, then decreasing.  1981-1984, 1989-93, 1994-99, 1999-04, 2005-08. Dummy of the last quantile (the 15 th ) is omitted. Emplostat represents dummies variables for: Unemployed, Full time, Part time, Self Employed, Retired, House-Keeper. Education is a series of 10 dummies controlling for different years of schooling. GDP is the per capita GDP in PPP, in 10K, 2005 USD. Standard errors are clustered at country and wave level (in brackets).  1981-1984, 1989-93, 1994-99, 1999-04, 2005-08. Dummy of the last quantile (the 15 th ) is omitted. Emplostat represents dummies variables for: Unemployed, Full time, Part time, Self Employed, Retired, House-Keeper. Education is a series of 10 dummies controlling for different years of schooling. GDP is the per capita GDP in PPP, in 10K, 2005 USD. Standard errors are clustered at country and wave level (in brackets).  1996-2006 1996-2006 1996-2006 1996-2006 1996-2006 1996-2006 1996-2006 1996-2006 1996-2006 1996-2006 1994-99, 1999-04, 2005-08. Dummy of the last quantile (the 10 th ) is omitted. Reg.GDP is the per capita regional GDP in PPP, in 10K, 2005 USD. Standard errors are clustered at the quantile level (in brackets).  . Effect of regional GDP quantiles on life satisfaction in the 10-quantile partition of EU14 data. Coefficients on the dummies indicate the different 10 quantiles-with the 95% confidence intervals and errors clustered at quantile level -derived from the basic ordered probit regression. GDP is in 10K, 2005 USD, PPP adjusted.
A: All data: The base level is the 10 th quantile.

S5: Factor Analysis to Determine the Personality Traits
Determination of traits from the score in each question was necessary because no existing imputation to traits on the sample of questions in the data exists.
Trait determination was realized with exploratory factor analysis (statistical software Stata, release 11). We retained factors with eigenvalues larger than a threshold value of 1 as suggested by different sources. We selected all the personality questions in the WVS dataset, such questions were available only for the wave 1989-93. For completeness, we also included the variable e065, the answer 'none of the above' to the residual question. In Figure 5 we present the list of the questions and some descriptive statistics.
In Figure 6 we present the Stata log showing the eigenvalues of all factors and the factor loadings.
We plot the factors' loadings with eigenvalues larger than 1 in Figure 7, where we note that variables are clustered into two main groups. A high score in the group of variable with high loading on factor 1 represents high excitement and assertiveness, high seeking of stimulation and other peoples' company, and a pronounced engagement with the external world. We therefore defined factor 1 as Extraversion.
A high score in the group of variables with high loading on factor 2 represents negative emotions like depression, loneliness, boredom, anxiousness, and anger. We defined factor 2 as neuroticism.
To complete the analysis we also present the rotated matrix in Figure 8 and the Kaiser-Meyer-Olkin measure of sampling adequacy in Figure 9. This test generates values between 0 and 1 for each variable included, with smaller values meaning the variables have too little in common to warrant a factor analysis.
All our variables show adequacy levels well above 0.7, generally considered the acceptable threshold.   Figure 7. Factor Loadings of the Personality Factor Analysis. Factor 1 has been defined as extraversion, Factor 2 has been defined as neuroticism, variable e065 is the answer 'none of the above' to the residual questions