Discussion Paper

No. 2017-113 | December 15, 2017
Subjective well-being and income: a compromise between Easterlin paradox and its critiques


Despite rising popularity of subjective well-being (SWB) as a proxy for utility, its relationship with income is still unresolved. Against the background of debates around the ‘Easterlin paradox’, this paper seeks a compromise between two positions: one that insists on individual relative income, and one that finds similarity between individual and aggregate levels. Proposing a model which puts the emphasis on the interaction between individual and aggregate-level factors, it argues that the effect of relative income on SWB varies across countries as a function of average income, in addition to a relatively small direct effect of the latter, in partial agreement with the two major positions. The model is tested cross-sectionally on the data from the latest wave of World Values Survey. The results from hierarchical mixed-effect models confirm the main argument. But further examination reveals that there is still unaccounted variation especially in middle-income economies.

JEL Classification:

D31, C31


  • Downloads: 329


Cite As

Rusen Yasar (2017). Subjective well-being and income: a compromise between Easterlin paradox and its critiques. Economics Discussion Papers, No 2017-113, Kiel Institute for the World Economy. http://www.economics-ejournal.org/economics/discussionpapers/2017-113

Comments and Questions

Anonymous - Referee report 1
December 17, 2017 - 18:39

This paper contributes to a large literature on the relationship between income and happiness and in particular to multi-level analyses in that context. Its contribution is in a subtly different analysis and use of the most recent dataset. Though the conclusions are not really new, it is worth publishing.
I ...[more]

... recommend two revisions:
1. Acknowledge earlier analyses of this kind, such as Schyns 2002, and indicate the difference with this work.
You can use the Bibliography of Happiness for identifying similar analyses, in particular the subject sections Ff01 ‘Happiness and affluence in nations’ http://worlddatabaseofhappiness.eur.nl/hap_bib/src_pubs.php?mode=1&Subject=190 and Ge ‘personal ‘Income’ http://worlddatabaseofhappiness.eur.nl/hap_bib/src_pubs.php?mode=1&Subject=12

2. The use of control variables requires reconsideration. Why remove indirect effects of income, such as through education? The use of subjective health as a control variable involves the risk of over-control, since the absolute effect of good material conditions reflects about as much in health and happiness. If used at all, control variables should be entered step-wise

Anonymous - Referee report 2
January 29, 2018 - 11:07

I think the findings in this in this paper are both interesting and important. The paper is also nicely written and the results are presented in a clear and concise way.
The paper offers a timely contribution to an interesting discussion by using an important and large dataset. Overall, the ...[more]

... analysis is conducted appropriately.
The paper involves a discussion section where the author contrasts the implications of the model with the data and elaborates on the possible shortcomings and potential effects of cross-national differences which are not accounted for by the available data.
The most important weakness in my opinion is the lack of comparison with the recent findings of Proto and Rustichini (2013) who focus on a very similar question and mention the role of personality traits such as neuroticism.
There are minor typos that needs to be corrected.

Rusen Yasar - Response to reviewers
February 08, 2018 - 11:53

I would like to take this opportunity to thank both referees for their constructive comments.

Both reviews draw attention to earlier works comparable to the present analysis. In the revised version, I will try to integrate the suggested citations into the discussion.

Regarding the use of control ...[more]

... variables, my aim is to keep the estimates as conservative as possible; hence I used a considerable number of control variables backed by previous studies, including certain strong covariates such as subjective health. While education is also in the model, its effect is relatively minor.

Anonymous - Referee report 2
March 08, 2018 - 09:35

This paper uses WVS cross-section data to look at the relationship between own income and country GDP per capita. It is shown that life satisfaction rises with own income (rank) and GDP per capita, but that the interaction of the former with the latter attracts a significant negative coefficient: as ...[more]

... such, the effect of own income (rank) is lower in richer countries.

I believe that we still have a lot to learn from the analysis of income and subjective well-being. However, I did not find the analysis here particularly convincing. I list my concerns in turn below.

A first very general point is that the paper is really quite long for what it does. The introduction in particular is pretty laborious.

The paper is sold as presenting a kind of compromise between those who believe that income is fully relative and those who believe that it is only absolute income that matters. I note in passing that most research papers implicitly take this position, and estimate well-being functions that depend separately on own income, y, and a reference income level, y. Finding that the estimated coefficient on y is negative is evidence that comparisons matter, but we do not impose that the coefficients on y and y* be equal and opposite.

I do think that the paper’s analysis is weakened by the assumption that everyone compares to the same value of y* within a country. In addition, the income measure, income rank, is by construction comparative: if everyone’s income grows, there will be no change in income rank. In this sense, the positive estimated coefficient on GDP per capita that the paper takes as somehow being inconsistent with the Easterlin Paradox (although it is not, as the EP finds no effect of GDP per capita in time series rather) could just be showing you that someone in income decile seven has more dollars in a higher GDP country than in a lower GDP country. What we would really want would be a real dollar income figure for absolute income and then a real dollar figure for relative income, entered separately.

I think that this specification could also easily explain the negative coefficient on the interaction term. It is pretty much universally recognised that the relationship between own income and subjective well-being is concave. The negative income rank X GDP coefficient could just be showing that. In other words, you may not have shown that relative income matters less in richer countries, as the abstract suggests, but just the diminishing marginal utility of own income.

Not everyone will be familiar with mixed effects models, and may wonder how it is possible to estimate a country-level intercept and a coefficient on country GDP at the same time using cross-section data. I think that we need to know what assumptions lie behind this model.

I didn’t find the reasons for the use of only one WVS wave totally convincing. In earlier waves, are the countries richer or poorer? If you are estimating a different regression coefficient per country then I wasn’t sure to see how having twice as many observations (say) in one country relative to another should matter.

The regression analysis in Table 1 includes a lot of subjective variables on the right-hand side (autonomy, trust, confidence) and not everyone likes subjective on subjective regressions due to the presence of common mood effects. What happens if these variables are dropped?

The discussion of the various countries’ positions around the regression line in Figures 2 and 3 was a bit ad hoc. Any valid explanatory variable here should affect not subjective well-being, but the relationship between subjective well-being and income.

Other points
* I really didn’t see what rationality had to do with the EP at the beginning of Section 2.
* Page 4, para 2. Finding that subjective well-being rises with income in the cross-section is not contrary to the EP: it is rather an essential part of it (if there was no relation in either cross-section or time-series then there would be no paradox to explain).
* Dick Easterlin has a very recent piece out that might be helpful (Easterlin, R. (2017). "Paradox Lost?". Review of Behavioral Economics, 4, 311-339).
* Three lines after equation (1): Why “little role”? This surely depends on the values of beta(1) and A.
* Bottom of page 7. It would be more helpful to state these hypotheses in terms of the coefficients in equation (2).
* Bottom of page 9. Why should beta-hat be larger than gamma?
* The discussion at the bottom of page 10 confuses happiness as a cognitive/evaluative measure with the affect (or emotion) of happy.
* Page 11, para 2. Was that really how the thresholds were determined, so that 10% of respondents would appear in each decile?
* Page 12 talks about individual fixed effects when I think it means controls.
* Page 15, line 5. It is difficult to interpret these results as we do not know what the standard deviations of satisfaction or income rank are.
* Page 15, start of para 2. Why do you care if they are measured on the same scale? I thought that everything was standardised here.

Rusen Yasar - Response to review
March 28, 2018 - 19:41

I would like to thank the last reviewer for the detailed comments. Overall, I will take many of these comments as a sign that I will have to highlight certain aspects of the paper and clarify some others, although I do not agree with the reviewer on all points. Below ...[more]

... I outline my views on the comments.

The introduction and the background sections are written by keeping in mind the readers who are not very familiar with the subject of the paper and why it matters for contemporary debates.

Research that takes into account own income and reference income shows the relevance of relative income. What this paper offers is to take average income (the most widely used type of reference income) not simply as a point of comparison, but also as a contextual effect in addition to relative income. Taking average income as an aggregate-level variable is not innovative in itself; the emphasis is put here on the distinction between being a point of reference and having a contextual effect.

While it is true that individual comparisons might vary, it is nearly impossible to take this into account accurately. Points of reference are likely to vary in terms of locality, gender, age/generation and many other socioeconomic characteristics, which would lead to an extremely high number of cross-cutting groupings. Any choice among these is bound to be arbitrary by assuming the importance of one over the other. In this sense, it is not surprising to see that, unless a research is specifically concerned with the variability of reference income, the majority of economic analyses take the national average as the main, if not the only, reference point.

The income measure is indeed meant to be comparative, the positive coefficient on GDP is meant to show that the same relative income rank with different GDP levels would signify different real incomes, and the negative coefficient on the interaction term is meant to embody diminishing marginal utility of income. The nuance is in separating relative and real income measures in a way that corresponds to individual and aggregate levels respectively, as different from a picture of individual real income values aligning neatly along a universal concave curve or a picture of identical context-specific relative income curves. The empirical support for this view is clarified by showing the relationship between GDP per capita and the coefficients of national relative income curves.

Taking into consideration the length of the paper and its sections, I would still opt for rich discussions on the background and the relevance of the subject, and for minimising textbook-like explanations of methods employed in the analysis.

By including earlier waves of WVS, the aggregate unit could not remain country-only, but would become country*year, entailing further GDP per capita values. This would, in turn, lead to the problem of a skewed sample of countries, since not all countries which took part in the last wave had also taken part in the earlier waves.

The reason why so many subjective variables are included is isolating as much as possible a common subjective component which can be understood as common mood effects. This is why these variables are emphasised as controls, and their coefficients are not attributed a special meaning. Dropping them only leads to a larger variation explained by income variables. As noted in my response to a previous review, their function is to make the model as conservative as possible.

The vertical axis in Figures 2 and 3 represent the intercepts and coefficients of relative income. In this respect, the figures are by definition about the relationship between relative income and SWB. Moreover, the discussion section is essentially about possibly unaccounted country-level variables, hence the relationship between GDP per capita and the relationship between relative income and SWB.

On other points:
- Some of the later works by Easterlin mention Kahneman's approach to rationality as an explanation.
- The EP is defined early on in terms of the asymmetry between individual- and aggregate-level (longitudinal) measures. This is in line with its original formulation. If the main focus has shifted to longitudinal v. cross-sectional difference, this should indeed be clarified in the paper.
- For eq. 1, it is Clark et al's main purpose to make the first term converge to Beta(1), hence almost like a constant except for smaller values.
- I agree that referring to coefficients would clarify the link between hypotheses and the abstract discussions
- If relative income is not a mere reflection of a universal log-linear relationship, then its slope should be steeper than that of such a log-linear curve, hence Beta(hat)>Gamma.
- The last paragraph of page 10 is meant to explain why an evaluative measure is preferable over a measure of affect in this context.
- The choice of the word 'decile' is wrong and the sentences where it is used should be rewritten.
- Individual-level control variables, including controls, are on the 'fixed effects' side, as opposed to 'random effects', in the mixed-effect model.
- SD-interpretation relies on the fact that all measurements are standardised. The next paragraph, when talking about 'same scale', distinguishes variables which are necessarily categorical and those which are treated as numeric.