Income reliably predicts daily sadness, but not happiness: A replication and extension of Kushlev, Dunn, & Lucas (2015)

Abstract

Kushlev, Dunn, and Lucas (2015) found that income predicts less daily sadness—but not greater happiness—among Americans. The present study used longitudinal data from an approximately representative German sample to replicate and extend these findings. Our results largely replicated Kushlev and colleagues’: income predicted less daily sadness (albeit with a smaller effect size), but was unrelated to happiness. Moreover, the association between income and sadness could not be explained by demographics, stress, or daily time-use. Extending Kushlev and colleagues’ findings, new analyses indicated that only between-persons variance in income (but not within-persons variance) predicted daily sadness—perhaps because there was relatively little within-persons variance in income. Finally, income predicted less daily sadness and worry, but not less anger or frustration—potentially suggesting that income predicts less “internalizing” but not less “externalizing” negative emotions. Together, our study and Kushlev and colleagues’ provide evidence that income robustly predicts select daily negative emotions—but not positive ones.

Does income predict subjective well-being? This is a deceptively complex question—primarily because there is no one single, unified definition of well-being (Lucas & Diener, 2008). Indeed, there are at least two elements of well-being: (1) people’s general cognitive evaluation that their lives are going well, and (2) the extent to which they tend to experience positive and negative emotions throughout their days (Diener, 1984). Moreover, these components of well-being only partially overlap (Kim-Prieto, Diener, Tamir, Scollon, & Diener, 2005; Lucas, Diener, & Suh, 1996). Therefore, it is possible that income might have different associations with people’s global evaluative judgments of well-being (e.g., overall life satisfaction) and their experiential well-being (e.g., experienced positive and negative emotions).

Most previous research has tested the associations between affluence and people’s global, evaluative well-being. These studies have found that income has positive—albeit generally small to moderate—associations with life satisfaction (Diener & Biswas-Diener, 2002; Diener, Ng, Harter, & Arora, 2010; Kahneman & Deaton, 2010; Sacks, Stevenson, & Wolfers, 2012). Thus, wealthier individuals generally perceive their lives somewhat more positively than do poorer persons. Nevertheless, increasingly greater wealth is associated with diminishing returns in life satisfaction; and consequently, especially for richer individuals, substantial increases in income may be necessary to have an appreciable impact on life satisfaction (Lucas & Diener, 2008; Lucas & Schimmack, 2009).

In contrast to the associations between money and evaluative well-being, research suggests that income has even weaker associations with the extent to which people experience positive emotions (Diener et al., 2010; Kahneman & Deaton, 2010; Kahneman, Krueger, Schkade, Schwarz, & Stone, 2006; Kushlev, Dunn, & Lucas, 2015). Thus, despite reporting greater life satisfaction, wealthier individuals may not feel greater amounts of daily happiness, as compared to less affluent people.

But what about negative emotions? Recently, Kushlev and colleagues (2015) argued that income may be associated with lower levels of sadness. Specifically, positive and negative emotions are not necessarily polar opposites, and they can consequently vary independently of one another (Watson, Clark, & Tellegen, 1988). Kushlev and colleagues (2015) proposed that lower-income individuals may feel less empowered and able to control their environment, as compared to wealthier persons (Johnson & Krueger, 2006; Kraus, Piff, & Keltner, 2009). This lack of perceived efficacy to mold one’s circumstances may lead to negative emotions such as helplessness or sadness in the face of challenging life events (Roseman, Antoniou, & Jose, 1996). For example, wealthier individuals have greater capacity to afford unexpected repair or medical bills. In contrast, lower-income persons may need to forestall necessary repairs or treatments, or may be required to sacrifice in other areas of their lives to pay for them—which may spur negative emotions, such as sadness. Stated more succinctly, wealth may shelter people against the vicissitudes of life.

Supporting this line of reasoning, Kushlev and colleagues (2015) replicated prior findings that income is unrelated to experiential reports of happiness (e.g., Kahneman et al., 2006), but found that wealth does, in fact, predict lower levels of sadness (though the effect was relatively small in size). The purpose of the present paper was to test whether this pattern of findings is replicable. To do so, we directly replicated Kushlev and colleagues’ statistical analyses as closely as possible, and then extended them in several ways using a longitudinal dataset of German participants.

Overview of Kushlev et al. (2015)

Kushlev and colleagues (2015) analyzed data from a large sample of Americans who provided reports of their income, as well as their experiential happiness and sadness during specific episodes of a single day. In their study, experiential happiness and sadness were measured via the Day Reconstruction Method (DRM; Kahneman, Krueger, Schkade, Schwarz, & Stone, 2004). The DRM is a low-cost, easy-to-implement alternative to experience sampling methods (ESM; Shiffman, Stone, & Hufford, 2008). Unlike ESM, which is intrusive and burdensome for participants and can require researchers to purchase costly technologies, DRM can be administered in a standard survey format and at least some versions of the DRM can be completed in as few as 10-15 minutes (Anusic, Lucas, & Donnellan, 2016a, 2016b). When completing DRM measures, participants systematically reconstruct their prior day, listing all activities in which they engaged. Subsequently, they rate the extent to which they experienced various emotions during either every activity (Kahneman et al., 2004), or a randomly selected subset of the activities (e.g., Anusic et al., 2016b; Kushlev et al., 2015). Despite the DRM entailing retrospective reporting, participants appear to be able to accurately reconstruct their emotions from the prior day (Kahneman et al., 2004)—something they appear unable to do over longer periods of time (e.g., Robinson & Clore, 2002a, 2002b, 2007). Indeed, the patterns of affect reported across the day in DRM studies appear to closely match those found in typical ESM studies (Anusic et al., 2016b; Kahneman et al., 2004). Moreover, DRM measures of affect exhibit both convergent and predictive validity (Anusic et al., 2016a, 2016b).

Using DRM measures of experiential well-being, Kushlev et al. (2015) found that income was negatively related to daily averages of experiential sadness, but was uncorrelated with happiness. Moreover, they found that the association between wealth and sadness could be not explained by controlling demographic variables such as gender, relationship status, and employment status. Similarly, systematic variation in daily stress could not explain the link between income and sadness—suggesting that poorer individuals do not feel greater sadness simply due to greater experienced stress. Finally, although Kushlev and colleagues (2015) found numerous differences in how people of varying affluence spent their time (e.g., wealthier people spent more time working and commuting)—statistically controlling for these differences did not attenuate the relationship between income and sadness. Thus, wealthier individuals were not less sad simply because they spent greater time engaged in physical exercise, for example.

Overview of the Present Study

The purpose of the present study was to replicate Kushlev and colleagues’ (2015) findings from an American sample in a different culture. To do so, we used an approximately nationally representative German sample. Participants reported their income and completed DRM measures of experiential affect once annually for up to three years. These data were used to directly replicate Kushlev and colleagues’ (2015) statistical analyses. Specifically, we first examined whether income correlated with daily happiness and sadness. We subsequently tested whether these correlations withstood controlling demographics and daily stress. Finally, we investigated whether individuals with varying levels of wealth differed with respect to daily time usage—and whether these differences in time allocation could explain the links between income and daily affect.

In addition to replicating Kushlev and colleagues’ (2015) analyses, we also expanded upon their findings in two ways. First, we leveraged the repeated-measures nature of our data to examine whether income predicted happiness and sadness both between-persons and within-persons. Second, because participants rated a wide range of positive and negative emotions in our study (e.g., happiness, enthusiasm, anger, frustration), we examined whether income predicted generalized composites of positive and negative affect.

We expected to replicate Kushlev and colleagues’ (2015) findings that income would predict sadness, but not happiness. Moreover, we expected this association to emerge even with demographics and daily time usage held constant. In contrast, we did not have strong a priori expectations regarding how income might relate to composites of general positive and negative affect—or the extent to which within-person fluctuations in income might predict changes in emotions.

Method

Preregistration

Prior to conducting any analyses, this project, including our sample, planned analyses, and expected results, was preregistered on Open Science Framework (https://osf.io/d7r8p/).

Participants

We analyzed data from participants in the 2012 through 2014 waves of the Innovation Sample of the German Socioeconomic Panel (GSOEP; Wagner, Frick, & Schupp, 2007). This sample is an approximately nationally representative subsample of the larger GSOEP study, in which new and innovative questions are administered. Participants completed DRM measures once annually in 2012, 2013, and 2014. Across these three years, a total of 2,504 unique participants (52% female; age M=51.78, SD=18.00) provided at least one wave of data. The respective individual sample sizes for 2012, 2013, and 2014 were 2,303, 1,920, and 1,763.

On average, participants provided 2.39 waves of data (SD=0.85)—with 1,898 participants (76%) providing at least two waves of data. Attrition analyses revealed that people tended to provide fewer waves of data if, collapsing across waves, they reported greater levels of daily stress (r=−.08, 95% confidence interval [CI] [−.12, −.04]).

Measures

Income

At each wave, participants self-reported their monthly household net income in euros.¹

Daily emotions

At each time point, participants were asked to systematically reconstruct their prior day by reporting all activities that had occurred. Specifically, participants were first asked what time they awoke. Afterward, they were queried, “What did you do next?” Participants selected a general activity from a predetermined list (e.g., personal care, commuting, preparing food, watching TV, socializing) and indicated what time the episode began and ended. This procedure was repeated (i.e., participants were asked, “What did you do next?”) until participants had accounted for their entire day—ending with either their bedtime or midnight.

After providing a basic account of all of their activities during the previous day, three of the provided episodes were randomly selected for each participant. For each of these three episodes, participants rated the extent to which they felt several emotions during the episode: happy, enthusiastic, satisfied, angry, frustrated, sad, worried, and stressed. Each emotion was rated on a scale from 1 (not at all) to 7 (very much). Having participants rate three randomly selected episodes—rather than every episode (e.g., Kahneman et al., 2004)—dramatically reduces the time required to complete the measure, yet nevertheless appears to produce similar patterns of findings (Anusic et al., 2016b).

We formed daily composites for each of the eight emotions by averaging the ratings from the three episodes together with equal weighting. For example, we computed a single daily happiness composite for each participant at each wave—which was an average of their reported happiness during each of the three episodes they had rated.²

To directly replicate Kushlev and colleagues’ (2015) statistical methods as closely as possible, in our primary analyses, we examined zero-order correlations between income and daily happiness and daily sadness. Only when explicitly noted in the Results, daily stress was used as a control variable. For subsequent analyses, we also formed composites at each time point for daily positive affect and daily negative affect. Daily positive affect was an average of daily happiness, enthusiasm, and satisfaction (2012 α=.85). Daily negative affect was an average of daily anger, frustration, sadness, worry, and stress (2012 α=.87).³

Demographic controls

Only when explicitly noted below, participants’ age, gender, relationship status (married/partnered vs. not),⁴ number of children, and unemployment/retirement/student status were used as control variables.⁵

Results

Does Income Predict Daily Happiness and Sadness?

Table 1 contains the descriptive statistics and intercorrelations for all variables in 2012. For our first series of analyses, we examined whether income predicted daily happiness and sadness in each individual wave. We examined the zero-order associations within each time-point separately (rather than the aggregate associations across time) to directly replicate Kushlev and colleagues’ (2015) statistical methods as closely as possible (and thus, any potential differences between our results and theirs cannot be attributed to different analytic methods).

Table 1.

Descriptive statistics and intercorrelations for all study variables at 2012

Variable	M	SD	Correlations
			1	2	3	4	5	6	7	8	9	10	11	12
1. Monthly income (euros)	2737.89	1684.33	–
2. Daily happiness	4.00	1.51	−.02	–
3. Daily sadness	1.30	0.70	−.11	−.02	–
4. Daily stress	1.61	0.98	.05	.00	.41	–
5. Daily positive affect	3.82	1.28	.01	.89	−.03	.03	–
6. Daily negative affect	1.52	0.72	−.02	−.05	.76	.79	−.06	–
7. Age	51.79	17.99	−.16	−.03	.06	−.22	−.02	−.11	–
8. Male	0.48	0.50	.04	.00	.00	−.01	.01	.02	−.03	–
9. Married/partnered	0.61	0.49	.24	.05	−.02	.00	.03	−.03	.26	.03	–
10. Number of children	0.62	0.97	.28	.02	−.05	.16	.02	.08	−.46	−.03	.10	–
11. Unemployed	0.45	0.50	−.28	.04	.03	−.20	.03	−.10	.46	−.08	.00	−.21	–
12. Retired	0.003	0.06	.01	.00	−.02	−.04	.01	−.03	−.03	.01	.03	−.03	−.05	–
13. Student	0.02	0.15	−.01	.02	−.05	.00	.02	−.02	−.26	.04	−.16	−.12	−.14	−.01

Note: Ninety-five percent confidence intervals for correlations in boldface do not include zero.

As seen in the first three rows of Table 2, income predicted lower levels of daily sadness in every wave (correlations ranged from r=−.05, 95% CI[−.09, −.01] to r=−.11, 95% CI[−.16, −.06]), but was unrelated to daily happiness (all |r|s≤.02). Moreover, Steiger’s Z tests (see Meng, Rosenthal, & Rubin, 1992) revealed that the associations between income and sadness were statistically significantly greater than the correlations between income and happiness within every wave, Δrs ranged from .06 (95% CI[.01, .12]) to .08 (95% CI[.02, .14]). Thus, we replicated the basic pattern of results report by Kushlev et al. (2015), although our effect size estimates were somewhat smaller than their estimates (see the bottom row of Table 2).^6,7

Table 2.

Income predicting DRM happiness and sadness

Measurement Occasion	Outcome
	DRM Happiness			DRM Sadness
	β	95% CI		β	95% CI
		LB	UB		LB	UB
2012	−0.02	−0.06	0.03	−0.11	−0.16	−0.06
2013	0.02	−0.04	0.04	−0.05	−0.09	−0.01
2014	−0.01	−0.06	0.04	−0.07	−0.12	−0.02
All, between-persons*	−0.01	−0.04	0.03	−0.08	−0.11	−0.05
All, within-persons*	0.04	−0.03	0.11	−0.02	−0.10	0.06
Kushlev et al. (2015)	0.00	−0.02	0.02	−0.15	−0.16	−0.13

Note: DRM = day reconstruction method; CI = confidence interval; LB = lower-bound; UB = upper-bound; 95% CIs for parameter estimates in boldface do not include zero.

^*These parameters were estimated using multilevel models, and represent the between-persons and within-persons effects across all three waves of data.

Having directly replicated Kushlev and colleagues’ (2015) analyses as closely as possible, we subsequently used multilevel models (MLMs) to examine the associations between income and daily happiness and sadness across all three waves simultaneously. In these MLMs, we estimated the associations between income and affect both between-persons (i.e., individuals’ mean income across all three waves) and within-persons (i.e., fluctuations in people’s income around their individual means), and included a random intercept to control for within-person dependencies in the data.⁸ All variables were standardized across the entire sample before being entered in the model—thus the resultant parameter estimates are standardized regression coefficients. As seen in the lower half of Table 2, between-persons variance in income was associated with less daily sadness (β_between=−0.08, 95% CI[−0.11, −0.05]), but was unrelated to daily happiness (β_within=−0.01, 95% CI[−0.04, 0.03]). In contrast, within-persons variation in income (i.e., year-to-year changes in individual persons’ incomes) predicted neither sadness (β_within=0.04, 95% CI[−0.03, 0.11]) nor happiness (β_between=−0.02, 95% CI[−0.10, 0.06]). As we elaborate in the Discussion, this may reflect the fact that only a tiny portion of variance in income (9%) was within-persons; thus, there may have been too little within-person fluctuations in income to provide a sufficiently powerful test of the within-person associations.⁹

Because our MLMs provide a more powerful test of the between-persons associations and use all available data—yet produce comparable results to the zero-order analyses (as used by Kushlev et al., 2015)—we use MLMs and data from all three waves in all subsequent analyses.

Does the Association Between Income and Sadness Withstand Control Variables?

Kushlev and colleagues (2015) found that controlling basic demographics and daily stress did not eliminate the association between income and daily sadness. As seen in Table 3, we replicated this finding. Specifically, even holding constant age, gender, relationship status, number of children, and unemployment/retired/student status, between-persons variation in income continued to predict daily sadness (β_between=−0.07, 95% CI[−0.11, −0.03]).^10,11 Similarly, controlling participants’ daily stress did not significantly reduce the association between income and daily sadness (β_between=−0.10, 95% CI[−0.12, −0.07]). Thus, we replicated Kushlev and colleagues’ (2015) findings that the links between income and daily sadness cannot be explained by covariation between income and demographics or stress.

Table 3.

Income predicting DRM happiness and sadness, with control variables

Predictor	Outcome
	DRM Happiness			DRM Sadness
	β	95% CI		β	95% CI
		LB	UB		LB	UB
Demographics Model
Between-persons income	−0.02	−0.07	0.02	−0.07	−0.11	−0.03
Within-persons income	0.05	−0.02	0.12	−0.01	−0.09	0.07
Age	−0.07	−0.12	−0.02	0.04	−0.01	0.08
Male	−0.01	−0.04	0.03	0.00	−0.03	0.04
Married/partnered	0.06	0.03	0.10	−0.02	−0.06	0.01
Number of children	0.00	−0.04	0.04	0.02	−0.02	0.05
Unemployed	0.06	0.03	0.10	0.02	−0.02	0.05
Retired	−0.01	−0.03	0.02	−0.01	−0.04	0.02
Student	0.01	−0.02	0.04	0.00	−0.04	0.04
DRM Stress Model
Between-persons income	−0.01	−0.03	0.04	−0.10	−0.12	−0.07
Within-persons income	0.03	−0.04	0.10	−0.06	−0.13	0.01
DRM stress	−0.01	−0.04	0.01	0.41	0.38	0.44

Note: DRM = day reconstruction method; CI = confidence interval; LB = lower-bound; UB = upper-bound; 95% CIs for parameter estimates in boldface do not include zero.

Does Income Predict Daily Time Usage?

Kushlev and colleagues (2015) found that income predicted differences in the amount of time that people allocated to various activities. For example, in their study, income was positively related to time spent commuting (correlation from Kushlev et al. [r_K]=.17) and negatively correlated with TV viewing (r_K=−.10). Nevertheless, they found that controlling the total time people allotted to each of the various activities could not explain the links between income and sadness.

As seen in Table 4, we largely replicated Kushlev and colleagues’ (2015) findings. Specifically, in both our study and theirs, as compared to their less affluent peers, wealthier individuals spent more time working (β_between=0.23, 95% CI[0.20, 0.26]; r_K=.15), caring for children (β_between=0.06, 95% CI[0.03, 0.10]; r_K=.12), commuting (β_between=0.04, 95% CI[0.01, 0.07]; r_K=.17), and playing sports (β_between=0.04, 95% CI[0.01, 0.08]; r_K=.09), and less time engaging in spiritual activities (β_between=−0.04, 95% CI[−0.08, −0.01]; r_K=v.07), preparing food (β_between=−0.05, 95% CI[−0.09, −0.03]; r_K=v.04), resting/relaxing (β_between=−0.10, 95% CI[−0.13, −0.07]; r_K=−.02), and watching TV (β_between=−0.20, 95% CI[−0.24, −0.17]; r_K=−.10). There were, however, a few differences between our findings and Kushlev and colleagues’ (2015). Specifically, Kushlev and colleagues (2015) found that income predicted time spent socializing (r_K=.03), talking via phone (r_K=−.04), shopping (r_K=.09), and eating (r_K=.08), but not doing housework (r_K=.00)—whereas we found that income was unrelated to time spent socializing (β_between=−0.02, 95% CI[−0.05, 0.01]) and talking via phone (β_between=−0.01, 95% CI[−0.04, 0.03]), and it negatively predicted time spent shopping (β_between=−0.05, 95% CI[−0.08, −0.02]), eating (β_between=−0.07, 95% CI[−0.09, −0.03]), and performing housework (β_between=−0.04, 95% CI[−0.08, −0.01]). Thus, we replicated a total of eight of the thirteen income/time-usage associations found by Kushlev and colleagues (2015).

Table 4.

Income predicting daily time usage

Outcome activity	Between-Persons			Within-Persons			Kushlev et al. (2015)
	β	95% CI		β	95% CI		β	95% CI
		LB	UB		LB	UB		LB	UB
Working	0.23	0.20	0.26	0.12	0.06	0.19	0.15	0.13	0.17
Childcare	0.06	0.03	0.10	−0.10	−0.16	−0.03	0.12	0.10	0.13
Commuting	0.04	0.01	0.07	−0.04	−0.12	0.04	0.17	0.15	0.18
Playing sports*	0.04	0.01	0.08	−0.05	−0.12	0.03	0.09	0.07	0.10
Sexual activity	0.01	−0.02	0.04	−0.03	−0.11	0.05	–	–	–
Phone conversations	−0.01	−0.04	0.03	0.03	−0.05	0.11	−0.04	−0.06	−0.02
Personal care	−0.02	−0.05	0.01	0.08	−0.00	0.16	–	–	–
Socializing	−0.02	−0.05	0.01	0.02	−0.06	0.10	0.03	0.01	0.04
Computer activities	−0.02	−0.05	0.02	−0.02	−0.09	0.05	–	–	–
Reading	−0.03	−0.06	0.01	0.00	−0.08	0.07	–	–	–
Healthcare	−0.03	−0.05	+0.00	0.01	−0.08	0.10	–	–	–
Housework	−0.04	−0.08	−0.01	−0.05	−0.12	0.02	0.00	−0.02	0.02
Spirituality	−0.04	−0.07	−0.00	0.03	−0.05	0.11	−0.07	−0.09	−0.06
Pet care	−0.04	−0.07	−0.01	−0.01	−0.07	0.06	–	–	–
Shopping	−0.05	−0.08	−0.02	0.05	−0.04	0.13	0.09	0.08	0.11
Preparing food	−0.05	−0.09	−0.02	−0.05	−0.11	0.02	−0.04	−0.05	−0.02
Gardening	−0.05	−0.08	−0.02	−0.01	−0.10	0.07	–	–	–
Eating	−0.07	−0.09	−0.03	−0.01	−0.09	0.06	0.08	0.06	0.10
Rest/relaxation	−0.10	−0.13	−0.07	0.04	−0.04	0.12	−0.02	−0.04	−0.01
Watching TV	−0.20	−0.24	−0.17	−0.05	−0.12	0.02	−0.10	−0.12	−0.08

Note: CI = confidence interval; LB = lower-bound; UB = upper-bound; 95% CIs for parameter estimates in boldface do not include zero.

^*This category was “playing sports” in our study, and “exercising/recreation” in Kushlev and colleagues’ (2015) study.

Our participants also provided data on several activities that were not included in Kushlev et al.’s (2015) study. In our study, income predicted less time engaged in gardening (β_between=−0.05, 95% CI[−0.08, −0.02]) and pet care (β_between=−0.04, 95% CI[−0.07, −0.01]), but was unrelated to time allocated to sexual activity, personal care, computer usage, reading, and healthcare (all |β_between|s≤.03).

Despite the slight differences between our study and Kushlev and colleagues’ (2015) with respect to correlations between income and time usage, we replicated their core finding that income continued to predict daily sadness while including all 20 time-usage variables in the model (β_between=−0.06, 95% CI[−0.09, −0.03]). Thus, the links between income and sadness cannot be explained by systematic differences in how people of varying affluence spend their time.

Does Income Predict Positive and Negative Affect More Generally?

For our final series of analyses, we extended beyond the scope of Kushlev and colleagues’ (2015) findings by examining whether income might predict composites of daily positive affect (an average of happiness, enthusiasm, and satisfaction) and daily negative affect (an average of anger, frustration, sadness, worry, and stress). As seen in Table 5, income predicted neither daily positive affect (β_between=0.02, 95% CI[−0.02, 0.05]) nor negative affect (β_between=−0.02, 95% CI[−0.04, 0.03]).

Table 5.

Income predicting DRM positive affect and negative affect composites.

Measurement Occasion	Outcome
	DRM Positive Affect			DRM Negative Affect
	β	95% CI		β	95% CI
		LB	UB		LB	UB
2012	0.02	−0.03	0.06	−0.02	−0.08	0.02
2013	0.02	−0.02	0.06	0.01	−0.03	0.05
2014	0.04	−0.01	0.09	0.04	−0.01	0.09
All, between-persons*	0.02	−0.02	0.05	−0.02	−0.04	0.03
All, within-persons*	0.05	−0.02	0.11	0.05	−0.02	0.12

Note: DRM = day reconstruction method; CI = confidence interval; LB = lower-bound; UB = upper-bound; 95% CIs for parameter estimates in boldface do not include zero.

^*These parameters were estimated using multilevel models, and represent the between-persons and within-persons effects across all three waves of data.

Post-hoc analyses revealed that the lack of association between income and daily negative affect was driven by stress, anger, and frustration. Income was, in fact, negatively related to daily sadness (β_between=−0.08, 95% CI[−0.11, −0.05]) and worry (β_between=−0.04, 95% CI[−0.07, −0.01]). It was, however, unrelated to daily anger (β_between=0.03, 95% CI[−0.01, 0.06]) or frustration (β_between=0.01, 95% CI[−0.02, 0.05])—and it was positively related to daily stress (β_between=0.04, 95% CI[0.01, 0.07]). This latter association, however, was likely explained by other factors. Specifically, in exploratory analyses, holding constant employment status and total time working/commuting eliminated—and bordered upon reversing—the link between income and stress (β_between=−0.02, 95% CI[−0.05, 0.02]). Thus it does not appear that income per se increases stress. Rather, it appears that any positive link between income and stress is spurious, resulting from their shared covariance with time spent working/commuting.¹²

Discussion

In a national sample of Americans, Kushlev and colleagues (2015) found that income predicted lower levels of daily sadness, but was unrelated to happiness—and the links between income and sadness were robust to controlling demographics, daily stress, and time use. The primary purpose of the present study was to replicate these findings as closely as possible in a national sample drawn from a different country—Germany. We were largely successful. In our study, affluence predicted reduced daily sadness, but was unrelated to happiness. Moreover, the links between wealth and sadness were not attenuated by holding constant basic demographics, daily stress, or daily time usage. Thus, taken together, our study and that of Kushlev and colleagues provide accumulating evidence that income is reliably associated with sadness—but not happiness—and this link cannot be explained by income-based variation in demographics, stress, or time usage. That being said, it is important to note that household income (used in both studies) is only one indicator of wealth and may not fully capture the associations between affluence and affect. Future research should explore whether other indicators of wealth also predict experiential well-being.

Discrepancies Between the Results of the Present Study and Those of Kushlev et al. (2015)

The major divergences between the present findings and those of Kushlev and colleagues pertained to income-based variation in daily time usage. Specifically, Kushlev and colleagues examined the extent to which income predicted the amount of time people allotted to thirteen broad genres of activities. The results of our study and theirs aligned for eight of these activities: in both studies, wealthier people spent more time working, caring for children, commuting, and playing sports, and less time engaging in religious activities, food preparation, resting, and watching TV. In contrast, the two studies produced discrepant findings with respect to the remaining five activities. Kushlev and colleagues found that more affluent Americans spent more time shopping, eating, and socializing, and less time on the phone—whereas we found that income was negatively associated with time spent shopping, eating, and performing housework, and it was unrelated to time socializing or talking via phone in Germany.

The source of these differences should be evaluated in future studies. One possibility is that they represent cultural differences in how wealth is related to daily activities. Kushlev and colleagues’ (2015) sample was from the United States, whereas the present sample was from Germany. Thus, to the extent that the divergence in the studies’ results represents real, cultural variation (as opposed to sampling error), it may be the case, for example, that as compared to their poorer peers, more affluent Americans spent greater amounts of time shopping, whereas wealthier Germans spend less time shopping.

However, it is also important to note that the correlations in both studies were relatively small to begin with; and irrespective of the few differences in the income/time-use associations found across our study and Kushlev and colleagues’ (2015), holding people’s daily time usage constant did not mitigate the link between income and daily sadness in either study. Thus, the primary point—that variation in daily time allotment cannot explain the link between income and sadness— robustly replicated across both studies.

Novel Findings

In addition to replicating Kushlev and colleagues’ (2015) findings, we leveraged several features unique to our dataset to extend their analyses in two ways.

Does income predict affect within-persons?

Given the repeated-measures available in our dataset, we explored the extent to which income might predict daily affect within-persons. In contrast to the between-persons findings that we have summarized thus far—which tap the extent to which wealthier individuals feel different emotions than do poorer people—the within-persons findings capture the extent to which fluctuations in individual persons’ incomes predict changes in their happiness or sadness. For example, if a person receives a raise, do they subsequently experience less sadness?

Contrasting with the between-persons findings, we found no statistically significant within-person links between income and sadness (or any other emotion). This seems to indicate that within-person increases in individuals’ incomes are not generally associated with accompanying reductions in sadness. One limitation of these analyses, however, is that they capture variation over only three consecutive years—a small period of time to expect any large changes in income. Indeed, the vast majority of variance in income (91%) was between-persons in our dataset. Thus, there may have been too few within-person fluctuations in income to obtain an adequately powerful test of the within-person associations. Relatedly, within-person changes in income may have been too small in magnitude to facilitate substantive changes to affect—larger gains in individuals’ income may have been necessary to garner an appreciable reduction in sadness (e.g., Lucas & Diener, 2008; Lucas & Schimmack, 2009).

Future research should therefore more thoroughly explore whether within-person gains in income are associated with lessened sadness. Indeed, associations found on one level of analysis (e.g., income predicting less sadness between-persons) do not necessarily generalize to other levels of analysis (e.g., income predicting less sadness within-persons)(Clancy, Berger, & Magliozzi, 2003). For example, it may be the case that income influences well-being via different processes at different levels of analysis (e.g., between-persons vs. within-persons). Alternatively, it may be the case that income operates upon well-being via similar mechanisms between- and within-persons. Future research should disentangle these possibilities using longer longitudinal designs with greater within-person variation in income.

Does income predict general positive and negative affect?

Finally, given that participants rated multiple positive and negative emotions (e.g., happiness, enthusiasm, anger, frustration) at each time-point, we examined whether income predicted variation in composites of positive and negative affect. Our findings indicated that income was unrelated to composites of both positive affect and negative affect.

To explore why wealth predicted sadness, but not negative affect (an average of anger, frustration, sadness, worry, and stress), we examined the separate associations between income and each individual emotion included in the negative affect composite. We found that more affluent people felt less sadness and worry. In contrast, income was unrelated to anger, frustration, and—holding constant time spent working/commuting—stress.

Collectively, this pattern of results may indicate that income is primarily related to lower levels of what researchers have sometimes referred to as “internalizing” negative emotions—feelings, such as sadness, fear, or anxiety, which are directed inward in lieu of impelling external action (e.g., Chaplin & Aldao, 2013). In contrast, low income may not facilitate “externalizing” or “approach-related” negative emotions, such as anger, contempt, or disgust. Stated differently, emotions such as anger and frustration may motivate action, whereas feelings similar to sadness and anxiety do not involve approach tendencies (e.g., Carver & Harmon-Jones, 2009). Feelings of disenfranchisement and powerlessness associated with lower income (Johnson & Krueger, 2006; Kraus et al., 2009) may contrapose approach-related emotions. Ultimately, however, this explanation is purely speculative. And moreover, it remains possible that the differential associations between income and various negative emotions found in our study are purely attributable to sampling error. To the extent that these associations are robust, future research should explicitly test whether income predicts different types of negative emotions (e.g., those that compel action vs. not).

Conclusion

In sum, the present study replicated the basic pattern of results found by Kushlev and colleagues (2015). Collectively, this study and theirs provides accumulating evidence that income reliably predicts less sadness—despite being unrelated to happiness. Moreover, our study may suggest that income predicts only certain types of negative emotions—potentially internalizing ones, such as sadness and worry, but not externalizing/action-oriented ones, such as anger and frustration. These data reaffirm the idea that subjective well-being is not a single, unitary construct, and instead comprises multiple components. Studying these separable components has the potential to further the development of more sophisticated theories about the processes that underlie well-being.