12 12 data from the us is not included in the

12
12
Data from the US is not included in the Eurobarometer, but it is available from the early 1970s in
the General Social Survey. However, the measure of subjective wellbeing is “Happiness” rather than “Life
Satisfaction” (used in the Eurobarometer), hence it is not directly comparable. We nevertheless run our
7

2.3 Explanatory variables
Following convention, we use per capita GDP for the first analysis containing only observa-
tions after 1972 from the Penn dataset (version PWT 8.0) where data are in 2005 interna-
tional dollars and are adjusted for purchasing power parity. For the historical analysis we use
data from the Maddison Project (
home.htm , 2013 version.) where data are in 1990 international dollars. 13
The other explanatory variables are historical data from the OECD, available from 1820
onwards (Zanden et al., 2014). As a general rule, we selected all the existing variables with
yearly coverage that are traditionally used in subjective wellbeing literature; they are life
expectancy at birth (as an index of health), internal conflict, external conflict, education
inequality (measured as a GINI index), the index of democracy (originally, from the Polity
IV project; as an index of freedom), total gross central government debt as a percentage of
GDP (as an index of public expenditures), and inflation rates.
14
3 Econometric Analysis
We ran two different analyses, the first aims to show that the valence for each language and
year is a significant predictor of average life satisfaction for the country of origin for the
language. In the second we analyse the historical determinants of valence.
3.1 Valence and Aggregate Life Satisfaction
In figure 4 we present the relationship between the valence of each language and year and
the aggregate life satisfaction of the country of origin of the language in the corresponding
year for the period over which survey data is available on life satisfaction. Both series of
data are presented in the form of residuals after controlling for country fixed effects. The
relationship is clearly positive and highly significant.
In table 2 we show that this positive relationship holds even after controlling for year
fixed effects and words covered (column 1), per capita GDP (column 2), after restricting
the analysis to data before 2000 (columns 1,2) or considering the entire set of data available
(column 3). We also present results excluding Spanish and French data (column 4). In this
analysis by including this measure from the GSS as well, obtaining very similar results which are available
upon request.
13
The results of the next analysis would quantitatively change very little if we used Maddison Dataset
instead of Penn.
14
Ideally, it would have been worthwhile including measures of average education and wealth concentration
as well, but they are not available on a yearly basis in the OECD dataset.