Do countries that are already rich become even happier when they become yet richer? This was the essential question on which I entered a gentleman’s bet in 2004 with Andrew Leigh and which just recently got settled.
The reason for the bet was a famous hypothesis in happiness research called the Easterlin hypothesis which held that happiness did not increase when rich countries became even richer. In my ‘Fred Gruen’ presentation on this matter in 2004 I used the following graph to illustrate the happiness income relation across countries:
This graph shows you the relation between average income (GDP in purchasing power terms) and average happiness on a 0-10 scales for many countries. As one can see, the relation between income and happiness is upward sloping for low levels of income, but becomes somewhat flat after 15,000 dollars per person. I championed the idea that this was not just true if you looked across countries, but that this would also hold true over time.
Andrew Leigh’s thinking was influenced by other data, particularly a paper by Stevenson and Wolfers which – he thinks debunks the Easterlin hypothesis. Here’s one of their graphs:
What’s striking about this graph is that the dotted line slopes up in the top right corner. In other words, the relationship between happiness and income becomes stronger, not weaker, for countries with average incomes over $15,000. Andrew thinks that this is because they specify income in log terms (in other words, we’re looking at the effect on happiness of a percentage increase in income rather than a dollar increase in income). I think it’s because the Gallup poll isn’t measuring happiness, but is instead asking people to rank themselves on the Cantrill ladder of life scale.
So our gentleman’s bet was in effect a bet on whether happiness in the world value surveys behaved different to the ladder question of the Gallup polls, and on whether the short-run relation between income and happiness was strong enough to show up in periods of 5 to 10 years as well. Andrew thought it would, I thought 5-10 years would be long enough for the typical long-run no-effects findings to show up and that happiness has a different relation with income than the Cantril-question. So we bet on whether one would get a significantly positive relation between GDP growth and happiness changes for the rich countries when one looked at the World Value data for 2005. We agreed to look at the relation between income and happiness using country-average variation. The winner would get 100 bucks.
Now, both of us forgot about the bet for a few years as the data was supposed to become available. Only recently did Andrew remind me of our bet and asked to check what had happened.
When I (with research assistance from Debayan Pakrashi) started to look into this data again, it quickly became apparent that Andrew and I had been pretty sloppy in formulating the precise conditions of the bet. In many ways, our bet had been far too vague.
For one, the World Value survey is not in fact held in particular years. Rather, some survey is run almost every year in some country that adds to the collection of surveys known as the World Value Survey. Hence there was really no such thing as a ‘2005 wave’. Taken literally, only Australia, Finland, and Japan had a survey in 2005 and were countries that in the previous wave already had a GDP of 15,000 dollars. In all those countries, income had gone up a lot since their previous survey, with Australian happiness down and Japanese and Finnish happiness up. That is a bit meagre as ‘waves’ go.
So the first ‘addition’ was to have a bandwidth of years for the ‘2005’ waves that included 2004, 2005, 2006, 2007, and 2008. That gave 12 countries that were rich enough in the previous wave to qualify. The raw data was:
The next ‘snag’ was of course that there are many ways to define the dependence on income: linear or logarithmic. With logarithmic income one normally gets stronger statistical significance on income, so we went for logarithms.
Then, of course, there are still many other things one can put into the regression. Does one account for effects of particular years (in bands) and for the level of happiness that a country starts? We decided to try it all. Hence the final ‘deciding’ set of regressions were as follows:
Which tells you that the relation between income changes and happiness changes (the last two columns) was either quite insignificantly positive or even negative if one entered year-bands.
When one reflects on the list of countries used in the analysis though, it is clear that the outcome of the bet will have had little to do with the true relation between income and happiness. It will have hinged on hidden aspects of the data. For instance, the Australian world value survey in 1995 was run differently from the 2005 version. Hence the big drop in Australian happiness you see in this period for this data does in fact not show up for other Australian data (like the HILDA). So one suspects some change in the data-gathering to be responsible for it. Indeed, the level of Australian happiness in this data is markedly below the level found for the HILDA (where it is almost 8.0).
Similarly, the big increase in Japanese happiness in this period doesn’t show up either in other Japanese data and so probably has something to do with changes in how the survey was run there. The changes can relate to the months in which the surveys were held, the precise words used for the happiness question, the questions preceding the happiness questions, the cities in which the survey was run, how the survey was run (face-to-face or via telephone), etc.
So I may have gotten lucky and won the bet, but one cannot see the outcome as decisive evidence that income and happiness have no long-run relation within rich countries. The data for the 2010 post-GFC wave might well show the opposite!
I’m sure there are long term happiness surveys in psychology (although unfortunately I can’t remember them well). I think the basic conclusion is that at the individual level, it is often quite strongly biologically driven, although I’m not in a position to evaluate that claim well.
You should also stick in inequality as a covariate (median income would be better than mean income too), since, apart from psychological effects of inequality, presumably if inequality increases with GDP, a lot of people are going to be no better off in terms of what they can have, and indeed some may be worse off since things like housing will simply increase with GDP making the poorer worse off in some respects even if they have a greater amount of money in absolute terms. Without this, I think you will be swimming in a sea of variance. Places like the US, for example, have more or less gone nowhere for most people for ages, but average GDP looks like it has increased a lot. I would think that if happiness was money driven, this should mean they should be becoming less happy.
The other thing you probably want to do (if the data exists) is to break things up into bands. There are reasonably well accepted dips and troughs in happiness across a lifespan, and since the age distribution of countries is changing in a pretty non-random fashion, this could affect the results too.
I’m surprised you didn’t refer to the analysis in the 2013 World Happiness Report.
Log GDP per capita was significant in their regressions (0.283*** (0.073)) with regard to the Cantrill ladder. Other variables were important too such as Social support ,Healthy life expectancy at birth, Freedom to make life choices and Generosity. Interestingly in the special analysis on the Eurozone countries who have had big economic hits in recent years, it was the ‘Freedom to make life choices’ variable that was more important than the ‘Log GDP per capita’ variable.
Yes, John, but in a simple OLS regression the covariance betwen some of the regressors would be very large. In particular, welfare losses from a fall in GDP in a rich welfare state would mostly manifest as lost “freedom to make life choices” rather than as physical deprivation (which is not to deny the welfare loss exists – it’s just that the whole purpose of a welfare state is to change the nature, as well as extent, of deprivation).
Of course derrida. The causality chains in this area are complex. But at least the World Happiness Survey was using multiple variables rather than the very simple regression that Paul used. I’m not sure that such a simple regression as you used Paul is actually enough to test the bet you made with Andrew.
I agree derida. The webs of causality in this area are complex and not easily captured by multiple regression. But at least the World Happiness Survey included multiple variables in its analysis. Paul’s very simple regression is not really a fair test of the bet he made with Andrew.