Social scientists have argued for decades (Campbell et al. 1976; Diener et al. 1999; Easterlin 1974)—but only recently has it become widely recognized—that income is not a good measure of wellbeing. Amartya Sen, Joseph Stiglitz, Jeffrey Sachs and other economists have proposed using happiness in addition to income as a measure of progress (Helliwell et al. 2012; Stiglitz et al. 2009). What matters, however, is not only the level of happiness, but also its dispersion, and that is still largely overlooked.

There is growing interest in inequality; by far the most popular metric of inequality is income inequality. Academics, politicians, and the media pay attention to income inequality (e.g., Attanasio et al. 2012; Cassidy 2013; Delhey and Dragolov 2014; Economist 2011; Frank 2014; Pickett and Wilkinson 2014; Stiglitz 2012). Usually, those addressing it assume or imply that income inequality results in lower wellbeing, but it is striking how little we know about the actual link between income inequality and inequality in subjective wellbeing or happiness,Footnote 1 particularly because we know so much about the relationship between income and happiness.

Equal income does not guarantee equal outcomes because human abilities to use money and other resources differ greatly; the same income results in different functionings/capabilities for different people (see e.g., Nussbaum 2006; Sen 2000). There is also debate about whether inequality of opportunity or inequality of outcome should be the focus of research and policy. In any case, confining the discussion of inequality to inequality in income is superficial, but also potentially misleading. How do we measure inequality? Ruut Veenhoven, one of the pioneers of happiness research, has offered a suggestion: “income difference falls short as an indicator of inequality […] Instead I propose to measure inequality in another way, not by difference in presumed chances for a good life, but by the dispersion of actual outcomes of life, using the standard deviation of life-satisfaction as an indicator” (Veenhoven 2005, p. 457).

This approach and its modifications (Delhey and Kohler 2010; Delhey and Urlich 2006) provide a broader picture than does an income inequality metric such as the gini coefficient. Yet, variance in happiness as a measure of inequality has limitations as well. Such an approach ignores the connection between income inequalities and happiness inequalities for groups. This study examines the happiness gap between the rich and the poor, directly considering the relationship between income inequality and happiness inequality.

1 Income and Income Inequality

America has doubled its per capita income over the past four decades as shown in Fig. 1, Panel (a). Income and economic growth have resulted in a substantial increase of the average living standards (Bok 2010; Fischer 2010; Veenhoven 2005). By one estimate, today’s bottom decile has a better standard of living than everyone other than the top decile did one hundred years ago (Bok 2010). Americans should have become happier. But we know that the increase in income has not been accompanied by an increase in happiness. This is a well known phenomenon known as the Easterlin Paradox (Easterlin 1974; Easterlin et al. 2010, 2012; Oishi et al. 2011, 2012; Veenhoven and Vergunst 2013). Yet, the paradox is less surprising if we consider the income distribution, as shown in Fig. 1, Panel (b). Indeed, as lamented by Fischer (2008), national income is a woefully inadequate measure of the actual income enjoyed by actual people, because it refers to income per capita on average, disregarding the dispersion from that average. The bottom 20 % of the income distribution did not gain anything from doubling the GDP, and a the bottom 40 % (more than 100 million Americans) gained almost nothing: the upper income limit for the bottom 40 % of families has increased by merely 8 %, from $45,000 in 1972 to $49,000 in 2011 in constant dollars.

Fig. 1
figure 1

Income over time. a Gross domestic product (GDP) per capita is gross domestic product divided by midyear population. GDP is the sum of gross value added by all resident producers in the economy plus any product taxes and minus any subsidies not included in the value of the products. Source: World Development Indicators, World Bank http://data.worldbank.org/data-catalog/world-development-indicators. b Upper limits for family income fifths. 2011 CPI-U-RS adjusted dollars. Source: US Census Bureau, Table F1, see the link for data footnotes http://www.census.gov/hhes/www/income/data/historical/families. a National income (GDP), b personal income

Full size image

As noted above, income inequality has been the subject of exceptional attention (e.g., Frank 2012; Piketty 2014; Stiglitz 2012). The key question, which the present study attempts to answer, is how rising income inequality over the past four decades has affected wellbeing. This connection between inequality and wellbeing is of key importance, because inequality is a serious problem if it can be shown that it diminishes wellbeing. Wilkinson and Pickett (2006, 2010) have shown many examples of negative effects of inequality, but has been criticized for oversimplifying and choosing selective evidence (e.g., Snowdon 2010).

Some may argue that income inequality is not a problem if people are not unhappy with it. The present study adds direct evidence using a happiness approach: the increasing income inequality gap has been accompanied by an increasing happiness gap. Groups that are further apart on the income dimension are also further apart on the wellbeing dimension. Most studies investigate the link between income inequality and happiness level, not between income inequality and happiness inequality, and find a negative relationship (Alesina et al. 2004; Bartal et al. 2011; Blanchflower and Oswald 2003; Bjørnskov et al. 2013; Graham and Felton 2006; O’Connell 2004; Oishi et al. 2011; Oshio and Kobayashi 2010, 2011; Tricomi et al. 2010; Winkelmann and Winkelmann 2010; Wynne 2004). Two studies have examined the relationship between income and happiness inequalities, but they have reached contradictory conclusions: Berg and Veenhoven (2010) found no effect across countries. Delhey and Kohler (2010) found a negative relationship.Footnote 2

2 Happiness Theories

There are three major theories about how happiness is created. The adaptation theory (Brickman et al. 1978) posits that there is an adjustment to external circumstances, suggesting that people get used to their situations. The second theory, multiple discrepancy theory (MDT; Michalos 1985) asserts that happiness is a result of social comparison or a comparison to various standards, such as in the notion that “It is better to be a big frog in a small pond than a small frog in a big pond” (Davis 1966). The third theory, livability theory (Veenhoven and Ehrhardt 1995) argues that happiness results from objective living conditions and from fulfillment of needs.

According to adaptation theory, inequality should not be much of a problem: people simply adjust to income or lack of it. According to multiple discrepancy theory, inequality will make us less happy because people will feel more relatively deprived compared to others. According to livability theory, inequality will only hurt us to the degree that basic human needs are unsatisfied, that is, that only people in absolute (rather than relative) poverty, or suffering from other deprivation or adversity, will be less happy.

3 Data and Variables

We use American General Social Survey (GSS) data, pooled over 1972–2012. GSS is a cross-sectional, nationally representative survey. GSS was administered almost every year until 1994 when it became biennial. The unit of analysis is a person and data are collected in face-to-face, in-person interviews (Davis et al. 2007). The happiness survey item reads, “Taken all together, how would you say things are these days–would you say that you are very happy, pretty happy, or not too happy?” We coded answers as 1 = “not too happy,” 2 = “pretty happy,”3 = “very happy.”

The key explanatory variable is income. There are many income variables in the GSS. The variable “income” is not informative because half of the respondents are in the highest category, due to its low cutoff point of $25,000. Other income variables, “income72,” “income77,” and so forth, are not comparable because the bins are inconsistent. We use the family real income variable (realinc), which is inflation-adjusted for 1986 dollars. One reason for using family income is theoretical. In technical report no. 64 describing the GSS income variables, Ligon ([1989] 1994) states that the “concept for these two different sorts of measures [respondent and household income] differ as well. Measures of household income attempt to measure income from all sources, while measures of respondent’s income attempt to measure only the respondent’s earnings from a single occupation.”

A person’s happiness is likely to be more affected by family income than by personal income. Indeed, family income correlates twice as highly with happiness as respondent’s income does (0.17 vs. 0.09). Note that Ligon ([1989] 1994) uses the term “household income” to refer to the GSS realinc variable, while we follow the GSS codebook nomenclature and use the term “family income.” Another reason to use family income rather than personal income is logistical. Personal income variables have about 20,000 more missing values than family income variables.

A typical (there are very small changes in wording over time) question asked of respondents is, “In which of these groups did your total family income, from all sources, fall last year before taxes, that is? Just tell me the letter.” In the earliest surveys there were 12 choices, though the number has increased over time to 25. Ligon ([1989] 1994) describes the income variable in detail. “Getting the most out of the GSS income measures” by Michael Hout, Co-Principal investigator of GSS (2004) and the Ligon Report ([1989] 1994) describe the process of generating the income variable used here, which involved three steps. First, GSS researchers turned categories into dollars by using category midpoints as a measure of central tendency within each income category (including the bottom category). Second, they calculated the mean income in the top category using the Pareto distribution. Third, they adjusted dollars for inflation.

We turned this income variable into quintiles and generated a dummy variable for each category. One reason to generate quintiles is technical, in that it enables us to explore nonlinearities. We chose five bins to have sufficient categories to be able to proxy for five classes: poor, working class, middle class, upper-middle class, and rich, and also have a decent sample size per bin. We include robustness tests using alternative categorizations in appendices. While it may seem that aggregating data to bins results in loss of information, income data for this variable were originally derived from categories, making the variable only artificially (and approximately) continuous. Indeed, Ligon ([1989] 1994) advises: “Because of the crudity of the underlying data, both income measures are expressed in hundreds of dollars. Expressing these income measures in hundreds of dollars still implies a false precision for higher levels of income—for many purposes, the user would do well to round to the nearest thousand dollars.”

Survey measures of income suffer from measurment error. While we demonstrate the robustness of our results, future research should investigate with better measures of income when they become available. Currently, GSS is the longest running US survey containing both happiness and income measures. All variable definitions, summary statistics, correlations, and distributions of income and happiness variables for each year of the survey are in “Appendix 1”. We show that rising income inequality is associated with a rising wellbeing gap even after controlling for predictors of happiness. Two key predictors beyond income are health and unemployment (e.g., Blanchflower and Oswald 2011). We also include a typical set of covariates found to play a role in happiness (that is, happiness is in part a function of these): age, age squared, household size, and education. Finally, we control for a number of happiness predictors suggested in the literature to be of particular importance when dealing with various inequalities. Stevenson and Wolfers (2009) note that female happiness is declining, and we include gender because there are well-known inequalities associated with it. Conservatives are happier than liberals (Okulicz-Kozaryn et al. 2014). Politics matter when exploring inequalities. Indeed, liberals almost universally see inequalities as a problem, while conservatives are more ambiguous about it, and often actually supportive of inequality (Okulicz-Kozaryn et al. 2014). Finally, views about inequality are likely to vary across the US.

4 Results

This study aims to find whether the happiness gap between the rich and the poor has changed over the past four decades. Instead of calculating a standard deviation, gini coefficient, or a similar measure of dispersion for society as a whole, we calculate happiness averages for each income quintile for each year. Contrary to a decline in the variance in overall happiness, the happiness gap between the rich and the poor has increased, as shown in Table 1.Footnote 3

Table 1 Standard deviation by decade (SD)
Full size table

The first column (SD) shows the standard deviation in happiness, which was highest in the 1970s, then declined, then increased again after 2000. Subsequent columns show means of happiness by income quintiles. Happiness for the top quintile is quite stable over time at about 2.365. Happiness for the lowest quintile dropped by 0.1 from 2.06 in the 1970s to 1.96 in the 2000s. Likewise, for the second quintile it dropped by 0.1 from 2.18 in the 1970s to 2.08 in the 2000s. The middle quintile registered a substantial drop, too: about 0.06, from 2.23 to 2.17. Happiness ranges from 1 to 3, and hence, a change of 0.1 for 20 % of the American population (roughly 60 million people) is a substantial change in aggregate happiness. Happiness for the two upper fifths stayed roughly flat, and happiness for the bottom 60 % declined. The upper-middle class and rich stayed happy, and everyone else became less happy. Changes are large in relative terms. In the 1970s the gap between the rich and the poor was 0.3 (2.36–2.06); in the 2000s it was 0.41 (2.37–1.96), which is a 37 % increase (0.41–0.3)/0.3. Even more strikingly, the gap between the middle class (third quintile) and the rich (fifth quintile) increased from 0.13 (2.36–2.23) to 0.2 (2.37–2.17), a 54 % increase (0.2–0.13)/0.13.

Given that the overall standard deviation of reported happiness has decreased, why did the happiness gap increase? The total variance depends on within-group variance and between-group variance. Results imply that there was an increase in between-group variance and/or a decrease in within-group variance. The last five columns in Table 1 show standard deviations by income groups. From the 1970s through the 1990s within-group happiness variance decreased. After 2000, the variability increased. Variance in happiness is the largest among the poorest quintile in all decades, demonstrating the the poor are not a monolithic group.

The rich did not become happier even though they became richer. The poor did not retain their happiness despite the fact that their income stayed flat (and if anything increased slightly). If happiness was derived only from satisfaction of needs according to livability theory, the poor’s happiness should not have decreased. Based on multiple discrepancy theory, however, we expect growing income disparities to result in growing happiness disparities. In other words, relative deprivation matters in addition to absolute deprivation.

The differences in Table 1 are significant and they hold when controlling for key predictors of happiness. Table 2 shows the regression results. Column a0a shows that happiness declined over time. In column a0b, each happiness quintile is progressively happier as compared to the lowest quintile. In column a1, the top two quintiles are significantly happier over time compared to the lowest quintile. Column a2 adds controls for two key predictors of happiness, unemployment and health, though due to many missing observations on the health variable, the sample size drops substantially. Finally, column a3 is a saturated model including a number of additional covariates that have been shown to predict happiness as explained above. Model a3 controls for size of household, a key variable when focusing on a family income variable. In model a3 the sample size drops further due to more missing observations, and the results become less significant; the fourth quintile-year interaction is no longer significant, although the positive sign remains.

Table 2 OLS regressions of happiness
Full size table

To present these results in a more intuitive way, Fig. 2 plots predicted values, showing clearly that the happiness gap between the upper-middle class and rich and everyone else is widening. This is similar to the widening income gap shown earlier in Fig. 1b, and confirms the overall happiness-income-time patterns from Table 1. These figures demonstrate that the widening of the gap persists when adding controls for socio-demographic, political, and regional characteristics in model a3. Based on model a1, the gap between the richest and the poorest in 1972 was 0.31 (2.37–2.06) and increased in 2014 to 0.43 (2.38–1.95), a 39 % increase (0.43–0.31)/0.31; and based on the full model (a3), the gap has increased from 0.22 (2.33–2.11) to 0.31 (2.32–2.01), a 40 % increase (0.31–0.22)/0.22. Based on model a1, the gap between middle class (third quintile) and the rich (top quintile) was 0.11 (2.37–2.26), and increased to 0.2 (2.37–2.17), an 82 % increase (0.2–0.11)/0.11. Given the full model (a3), the gap was 0.08 (2.33–2.25), and doubled, increasing to 0.16 (2.32–2.16). These are the end point estimates (1972 vs. 2012), and conservative estimates comparing decades as in Table 1 are more appropriate. Nevertheless, the magnitude is striking.

Fig. 2
figure 2

Predicted happiness and 95 % confidence intervals for quintiles of income (referenced at the very right of the graph). Predicted values are based on Table 2. a specification a1, b specification a2, c specification a3

Full size image

Using a full set of regressors in model a3, the three bottom quintiles move together, and hence the poor, working, and middle class (lower three quintiles) became equally unhappy as compared to the upper-middle class (fourth quintile), who became only a little less unhappy, and the rich (the top quintile) who largely retained their earlier happiness. Over time, the rich became not only happier than the poor, but also happier than the middle class.

We performed several robustness checks. “Appendix 3” shows happiness by the original income measures, to show that the transformation of the income variable did not distort the results. The results are similar: the happiness gradient by income was less steep in the past, and specifically, the poor of today are less happy than were the poor in the 1970s. “Appendix 4” compares the inequality measures from the GSS to the “official” measures from the Census. “Appendix 5” uses three and seven bins on the income variable as opposed to the five used above. Finally, “Appendix 6” discusses results using personal income as opposed to family income. We also tried multinomial models and the results were similar; it is well known that when modeling ordinal happiness, OLS performs well (Ferrer-i-Carbonell and Frijters 2004). We also tried dropping years that included oversamples and years when the ordering of the questions was changed: 1972, 1980, 1982, 1985, 1986, 1987. Finally, we reran analyses on a sample excluding the post-2008 years to exclude the Great Recession. Results using all these subsamples were substantively similar.

5 Discussion: Relative Deprivation

An important limitation, and at the same time a direction for future research is to account for secular trends,Footnote 4 which would require a panel dataset across countries. Cross-country investigation is needed, because using multiple countries provides the most variability in contextual factors. Such an approach is beyond the scope of this study, but a brief discussion is in order.

Perhaps, the increasing gap in happiness between income quintiles is not due to increasing income inequality, but due to increasing costs. Over the past four decades, in addition to stagnant wages, other troubling trends have developed. Healthcare costs have skyrocketed, and education costs have increased even more (Chokshi 2009). Household debt increased, and college debt increased even more, surpassing credit card debt (Stiglitz 2013). One could therefore argue that it is rising costs, not inequality, that is responsible for the decline in happiness among the poor. Yet, costs are a problem for happiness only if there is not enough income to cover those costs, and therefore, insufficient income is the obvious explanation, but it is in part likely due to rising inequality as well. While inequality does not invariably mean worse objective conditions for the poor, it does mean worse relative conditions. And in fact, rising inequality in the US since the 1970s has meant not only that the rich are pulling ahead, but that the poor are falling behind; gains are not widely shared (Gilbert 2011). In the US, income inequality has increased as poverty has deepened (Desmond 2015; Edelman 2012; Edin and Shaefer 2015; Gould and Davis 2015; Massey 2007; Putnam 2015; Shaefer et al. 2015; Stiglitz 2012; Yellen 2006).

Future research could use another dataset to investigate alternative explanations and estimate how much of the happiness gap is due to different factors. The goal of this study is exploratory: to document the widening gap in happiness that attends the widening gap in income. The widening gap in happiness is likely due to the widening gap in income.

These results should not be generalized beyond the US. In addition, confirmation of the causal relationship and empirical exclusion of alternative explanations both remain for future research.

The happiness gap in the US between the rich and everyone else has increased; those at the top stayed happy and others became less happy. We can explain these results in terms of relative deprivation as theorized by Michalos in his multiple discrepancy theory (1985), and we conclude that adaptation theory and livability theory do not provide compelling explanations for our results. One possibility is that people tend to compare themselves to the rich and such comparisons makes them unhappy, but the rich do not become happier by comparing themselves to those with less money. Relative loss is felt more than is relative gain (Kahneman and Tversky 1979). Lower income among the poor and middle classes produces more unhappiness than higher income produces increased happiness among those already rich. We know that we compare ourselves to others (Campbell et al. 1976; Michalos 1985; Ng 1997), and therefore, lower income compared to others’ higher income can be felt as a loss. Moreover, once we reach a moderate income, we tend to compare ourselves to those at our level or above (Myers 2004). By one estimate, comparisons to others explain or predict happiness better than do actual resources (Schulz 1995).

We are more unequal in income, and more unequal in wellbeing. Perhaps, this makes sense of the Easterlin paradox: income growth did not cause an increase in happiness because that growth only went to very few people at the top.Footnote 5

Income inequality is only one way of measuring inequality. Happiness inequality is important as well. This study has examined the relation between the two.