More Unequal in Income, More Unequal in Wellbeing

Abstract

This study focuses on the long-term trend in happiness by income level in the United States. General Social Survey data suggest that in the past, rich and poor Americans were not only more equal in terms of income, but also in terms of their subjective wellbeing: the happiness gap between the poor and the rich has been increasing. Today’s poor suffer greater relative unhappiness than the poor of past decades. The gap between the poor and the rich is substantial, approximately 0.4 on a 1–3 happiness scale. The increase in the happiness gap is striking: comparing the 1970s to the 2000s, the gap has widened by about 40 % between the poor and the rich, and by about 50 % between the middle class and the rich.

Appendices

Appendix 1: Additional Descriptive Statistics

See Figs. 3, 4, 5, 6 and Tables 3, 4, 5.

Fig. 3

Distribution of real income by year. These values are estimated based on a categorical variable and imputed, so caution is necessary in interpretation

Fig. 4

Distribution of happiness by year. This distribution, in contrast to income, does not change dramatically over time. It is consistently true that most people are pretty happy, some are very happy and few are not too happy

Fig. 5

Happiness by quintiles of income: happiness is stable for the rich and declining for the poor. Series are smoothed with a 9-year moving average (4 years back, current year, and 4 years ahead)

Fig. 6

Happiness by deciles of income. Results are similar to those in Fig. 5: happiness is stable for the rich and declining for the poor

Table 3 Variable definitions

Table 4 Summary statistics

Table 5 Correlation matrix

Appendix 2: Other Variables by Income Quintiles Over Time

See Figs. 7, 8 and 9.

Fig. 7

Generalized trust: “Generally speaking, would you say that most people can be trusted or that you can’t be too careful in dealing with people?” The higher the value, the more trust. There is little difference by income, but a very clear pattern in the overall trend exists as documented by many others (e.g., Putnam 2001): Americans report lower trust than in the past

Fig. 8

Support for government action to reduce income differences. The higher the value, the greater the support. Numbers 1–5 shown at the right side of the graph denote income quintiles. As expected, the poorer the group, the more support there is for reducing income differences. Yet, the biggest increase in favor of redistribution is among the richest

Fig. 9

Opinions about fairness. The higher the value, the more agreement that “people try to take advantage of you as opposed to be fair.” Along with the increasing income gap, there was an increase amongst the poorest in the opinion that people are taken advantage of. Numbers 1–5 shown at the right side of the graph denote income quintiles

Appendix 3: Happiness by Original Income Measures

See Fig. 10.

Fig. 10

Happiness by original income measures. 95 % CI shown. The top row shows the two earliest income measures (gss names: income72 and income77). The bottom row shows the latest income measures (gss names: income98 and income06). Vertical lines are drawn to aid interpretation. In all years happiness for the richest is approximately 2.4. But there is a change for the poorest: in the 1970s it was approximately 2.0, but in the last decade it was below 2.0 (about 1.9). There are also some interesting irregularities: for instance, the least happy are not the poorest

Appendix 4: Comparison of Inequality Measures from GSS and from Census

Gini data from the Census Bureau come from Table F-4 “Gini Indexes for Families, by Race and Hispanic Origin of Householder: 1947 to 2014,” available at https://www.census.gov/hhes/www/income/data/historical/inequality. Note that this gini index is for families (in this paper we use family income from the GSS). Furthermore, it is worth noting that from 1947 until 1972, when the GSS series began, the gini decreased only slightly and reached a low of 0.348 in 1968 before increasing, through to the present day. Data are plotted in graph 1. The correlation between the two series is 0.85. There is a spike in the GSS gini, another reason to rerun the analyses without the post-recession years. Yet, both series show a virtually identical linear trend as shown by fitted values (Fig. 11).

Fig. 11

Comparison of the gini calculated from the GSS family income variable used in this paper with the official gini from the Census Bureau

Data on income quintiles from the GSS and the Census are shown in Figs. 12 and 13. The Census data use 2011 constant dollars. The GSS uses 1986 constant dollars. The quintiles are more uneven for the GSS series. Incomes are estimates in the GSS and, for many years, not very continuous but rather, are presented more discretely. This limitation is unfortunate, particularly because these are the only data available that are consistent series and go back as far in time. The series are especially inconsistent in earlier years when there were fewer income categories, and so the cutoff points for quantiles are coarser. The fit between the two series is not close; to check whether the results might have been due to quantization, we reran the analyses using 3 and 7 classes (see “Appendix 5”). Results are similar.

Fig. 12

Quintiles of income from GSS. 1986 constant dollars

Fig. 13

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

Appendix 5: Robustness Checks: Using 3 and 7 Classes

Stata (the statistical software used in this study) performs quantization (assigns variable values into quantiles) in a peculiar way by checking whether it passes the appropriate percentile for given quantile and if it does not, it uses the lower quantile for categorization. This yields non-optimal cutoff points when a variable has few levels. For instance, in 1972 real income has the following distribution:

Income	Frequency	Percent	Cumulative percent
2707	123	8.34	8.34
8122	165	11.19	19.54
13,537	160	10.85	30.39
18,951	160	10.85	41.25
24,366	192	13.03	54.27
30,458	217	14.72	69
37,226	159	10.79	79.78
43,994	103	6.99	86.77
50,763	63	4.27	91.04
60,915	67	4.55	95.59
74,452	33	2.24	97.83
109,355	32	2.17	100
Total	1474	100

Hence, an optimal solution is to classify only the first two categories (2707, 8122) into the first quintile (they constitute 19.54 percent of the distribution), but Stata takes the third category as well (13,573), because it does not reach 20 % after taking the first two. This peculiar approach has been documented on the Stata listserv by Stata’s foremost expert Nick Cox http://www.stata.com/statalist/archive/2012-06/msg01193.html, and at http://www.stata.com/statalist/archive/2012-06/msg01187.html. An obvious question is what if the study’s result—that the poor and rich are more unequal in happiness—is due to this strange clustering? We use 3 and 7 bins as a robustness check. Results hold up as shown in Tables 6 and 7.

Table 6 OLS regressions of happiness: using tertiles

Table 7 OLS regressions of happiness: using tertiles

Appendix 6: Family Income Versus Personal Income

There are theoretical and technical reasons for using family income as opposed to individual income, but for robustness, we discuss here the results using personal income. With respect to family income reported here there is a rather smooth gradient: the richer the group, the smaller the decrease in happiness over time, and as a result the bigger the gap in happiness. With respect to respondents’ income, this is still true when comparing the first quintile to the others: the poorest people became less happy over time. Richer groups remained more flat in their happiness, so the happiness gap between the poorest and the rest has increased. The key difference in results when using personal income is that all richer income groups remained at about the same happiness levels over time, so whatever happiness differences there were in the 1970s, they remained after the 2000s. In addition, the very richest group became slightly less happy, though not nearly by as much as the poorest group. The three middle quintiles remained most stable over time. For the reasons mentioned above, family income is a better measure, and we base our overall conclusions on the family income variable. These results do not generalize entirely when using personal income. In particular, the happiness gap between the richest and middle class has not increased if individual income is used instead of family income.

Finally, we reran models using family income divided by the natural log of the number of people in the household +1 (to avoid zero from ln(1)). The effect of family income on happiness depends in part on the number of people in a household. The rationale for dividing family income by the natural log of the number of people in a household is to scale the effect to reflect the diminishing marginal effect of each additional member of a household on family income (the addition of each subsequent member of a household “costs” less).