The Age Profile of Life Satisfaction After Age 65 in the U.S.

Abstract

Although income and wealth are frequently used as indicators of well-being, they are increasingly augmented with subjective measures such as life satisfaction to capture broader dimensions of the well-being of individuals. Based on large surveys of individuals, life satisfaction in cross-section often is found to increase with age beyond retirement into advanced old age. It may seem puzzling that average life satisfaction does not decline at older ages because older individuals are more likely to experience chronic or acute health conditions, or the loss of a spouse. Accordingly, this empirical pattern has been called the “paradox of well-being.” We examine the age profile of life satisfaction of the U.S. population age 65 or older in the Health and Retirement Study (HRS), and find that in cross-section it increases between age 65 and 71 and is flat thereafter; but based on the longitudinal dimension of the HRS, life satisfaction significantly declines with age and the rate of decline accelerates with age. We reconcile the cross-section and longitudinal measurements by showing that both differential mortality and differential non-response bias the cross-sectional age profile upward: individuals with higher life satisfaction and in better health tend to live longer, and, among survivors, individuals with higher life satisfaction are more likely to remain in the survey, masking the decline in life satisfaction experienced by individuals as they age. We conclude that the optimistic view about increasing life satisfaction at older ages based on cross-sectional data is not warranted.

1. Introduction

Although income, wealth and labor market participation are frequently used as indicators of population well-being, the Sarkozy Commission called attention to a much broader list of measures such as health, education, and subjective measures to better capture the life experiences that shape the well-being of individuals (Stiglitz et al. 2009). An important example of a subjective measure is life satisfaction, which gauges “people’s explicit and conscious evaluations of their lives, often based on factors that the individual deems relevant” (Diener et al. 2018). As measured in surveys, life satisfaction varies as expected with measures of socioeconomic status, but it exhibits additional variation reflecting, among others, individuals’ subjective perceptions of their health, family, and social networks. Because of population aging, the well-being of the older population and its determinants will increasingly become a matter of intense policy interest.

Based on data from large surveys of individuals, life satisfaction in cross-section often exhibits a U-shaped pattern with age: average life satisfaction is high at younger ages, reaches a minimum at about age 40, which is sometimes called the “midlife crisis,” after which it monotonically increases. This U-shaped pattern has been confirmed in many datasets and across many countries (Blanchflower 2020a, b, Blanchflower and Oswald 2008, Deaton 2008, Ulloa et al. 2013, Stone et al. 2010).

Some researchers have called the empirical pattern of increasing well-being after the age of 40 the “paradox of well-being” (Hansen and Slagsvold 2012, McAdams et al. 2012, Swift et al. 2014), but the label “paradox” would only seem to apply to the age-pattern of well-being at ages past retirement, say, past age 65. As individuals progress into advanced old age they are more likely to experience challenging life events, such as developing new chronic or acute health conditions, the loss of their spouses, friends, and siblings, or economic distress. Thus, partly the paradox may be due to the age range studied: many studies have few observations on individuals beyond age 65, and some studies restrict their samples to younger individuals. Yet, between ages 40 and 65 many of the negative events associated with old age happen infrequently, while other positive events that might increase well-being happen frequently, retirement being a leading example. This paper focuses on the age-pattern of well-being after the age of 65, when, indeed, an increase with age might be a paradox.

The available evidence about the cross-sectional shape of life satisfaction after age 65 is limited and mixed. While some studies documented increases (Deaton 2018, Graham and Ruiz-Pozuelo 2017, Stone et al. 2010), a few studies reported a decline in life satisfaction among the oldest old based on cross-section data from Germany, Britain, and Norway (Baird et al. 2010, Frijters and Beatton 2012; Gwozdz and Sousa-Poza 2010, Hansen and Slagsvold 2012, Kunzmann et al. 2000, McAdams et al. 2012). Blanchflower and Oswald (2019) compared patterns in life satisfaction between ages 20 and 90 in seven datasets from 51 countries. They found that the cross-sectional patterns after 65 greatly vary across countries and datasets.

Several studies that have used longitudinal data and methods to investigate the age profile of life satisfaction found that life satisfaction increased significantly less in panel compared to the cross-section. However, the available evidence is mixed on whether life satisfaction increases or decreases with age in panel at advanced age. Some studies continued to find in these panel models increases in life satisfaction (or similar measures) with age beyond typical retirement ages (Cheng et al. 2017, Gana et al. 2013, Shankar et al. 2015, Zhang et al. 2017); other studies found that the age patterns were flat (Costa et al. 1987), or declined at older ages (Baird et al. 2010; Frijters and Beatton 2012; Hansen and Slagsvold, 2012; Jivraj et al. 2014; Kunzmann et al. 2000; Sharifian and Gruhn 2019).

Because of a strong association between life satisfaction and mortality, mortality selection could be a possible explanation for the difference between the cross-sectional variation with age and the longitudinal variation (Blazer and Hybels 2004, Brummett et al. 2006, Gerstorf et al. 2008, Segerstrom et al. 2016; Steptoe et al. 2015). A few articles speculated that mortality selection may bias the cross-sectional age profile of life satisfaction, but they did not quantify its extent (Hansen and Slagsvold, 2012; Steptoe et al. 2015). Non-mortal health conditions could also lead to response bias (Gwozdz and Sousa-Poza 2010, Ried et al. 2006).

The goal of this paper is to resolve the paradox of well-being by relating the age profile of life satisfaction of the U.S. population age 65 or older to mortality and other missing data patterns using longitudinal data from the Health and Retirement Study (HRS). The HRS has several features that provide a unique opportunity for this analysis: It is longitudinal; it has a large sample at those older ages; it has high sample retention rates; it attempts an interview by proxy for respondents who are unwilling or unable to be interviewed; it includes individuals who reside in nursing homes; and it carefully tracks mortality status, even among those who attrit from the sample.

We show that in the HRS the cross-section pattern of life satisfaction is flat in age after age 71 but that the longitudinal pattern is negative. The difference is mainly due to selective mortality, but other types of selective attrition also add to the difference. Our main contribution to the literature is to document these types of attrition and to provide the first quantification of their impact on the cross-section profile of life satisfaction.

2. Data and methodology

2.1. Data

The HRS is a nationally representative, longitudinal survey of the U.S. population over age 50. It is sponsored by the National Institute on Aging (grant number NIA U01AG009740) and is conducted by the University of Michigan.¹ The survey started in 1992 and it has interviewed individuals biennially since then. Refresher cohorts of 51–56-year-olds were added to the survey every six years. It interviews about 20,000 individuals every two years on a wide range of domains such as economic, social, and physical and mental health. Compared with other general-purpose surveys, the HRS has more detailed and more precise information about individuals’ health status.

The HRS is a bilingual (English and Spanish), racially and geographically diverse survey that represents older adults living in the U.S. African Americans and Hispanics are oversampled to increase statistical precision in these minority groups. Survey weights are available to adjust the sample to the American Community Survey. Most statistics reported in this study are weighted, except in a few cases, when weighting is not appropriate. Weighted and unweighted statistics are similar, however.

As any longitudinal dataset, the HRS is subject to panel attrition. Banks et al. (2011) found that panel attrition in the HRS is lower than in similar longitudinal surveys in Europe, and that attrition did not vary much with predictor variables.

Two HRS survey protocols are of particular importance for this paper because they provide information about the correlation between all-cause attrition and life satisfaction. (1) The HRS tracks the mortality status and mortality dates of all sample members, even of those who left the sample in any prior wave. (2) When a sample member cannot participate in the survey in person, either due to an illness or other reasons, the HRS tries to conduct an interview with a proxy informant, such as a spouse or a child. About 8.2% of the HRS sample age 65 or older respond through proxies. While nothing would be known about the current status of these individuals in most other surveys, thanks to the HRS proxy interviews we observe that these individuals tend to be less healthy compared to those who answer in person. In particular at advanced old age they tend to have cognitive limitations, which can help explain the patterns of non-response in other surveys.

In 2008, the HRS introduced the following question about life satisfaction in its core survey:

Please think about your life-as-a-whole. How satisfied are you with it? Are you completely satisfied, very satisfied, somewhat satisfied, not very satisfied, or not at all satisfied?

This is a validated single-item measure of life satisfaction that has been widely used in prior research and correlates strongly with richer, multiple-item life satisfaction measures (Diener et al. 2018). Because life satisfaction is a subjective concept, this question is not asked in proxy interviews. Thus, life satisfaction is not available in a subsample of respondents who tend to be less healthy and less well-off than the general population. To gain insights about the extent of this missing data problem, we use the longitudinal information in the HRS to analyze how missing values in a given wave are related to prior life satisfaction. In the case of a proxy interview, interviewers record whether cognitive limitation of the HRS sample member was the main reason for collecting a proxy interview, which we use to gain insight into the processes leading to nonresponse.

Between 2008 and 2016 (the last wave we use in this study), the HRS collected 93,051 person-wave observations on life satisfaction. In our main analysis we restrict this sample to 48,614 person-wave observations that are reported by those age 65 or older. Overall, 15,183 distinct individuals reported about life satisfaction after reaching age 65 over a maximum of five waves.

We reverse-coded the answers to the life satisfaction question so that higher values indicate greater satisfaction (1 = not at all satisfied … 5 = completely satisfied). In the 65+ year old sample, life-satisfaction has a weighted mean of 3.922 (i.e. slightly below 4 = “very satisfied”) and a standard deviation of 0.848.

We used variables from the RAND HRS Longitudinal File 2016 (V2) whenever possible. The RAND HRS is a publicly available, cleaned, longitudinal data set based on the most commonly used HRS variables.² Apart from the main variables of interest, age and life satisfaction, our regression models use information about gender, education (less than high school, high school, some college, BA+), race and ethnicity (non-Hispanic white, non-Hispanic black, non-Hispanic other, Hispanic), whether the person currently works, marital status (married, separated/divorce, widowed, never married), self-assessed health (excellent, very good, good, fair, poor), number of limitations in activities of daily living (ADL), number of limitations in instrumental activities of daily living (IADL), a 35-point cognition scale, and reports of experiencing pain (none, mild, moderate, severe).

2.2. Methods

We conduct our analyses in four steps. First, we verify that the cross-section relationship between life satisfaction and age generally found in the literature is also found in the HRS. Then we show that life satisfaction predicts subsequent mortality and that, when we control for the resulting selection, life satisfaction declines as individuals age. We show that other types of missing data such as nonresponse in the succeeding longitudinal wave operate in the same way as the missing data resulting from mortality. To quantify the relationship between life satisfaction and age we estimate first-differences panel regression models of life satisfaction on age, and on other covariates to find how much of the age-life satisfaction relationship can be explained by health and other life events. We use a quadratic function of age in these regression models because we found that the age trajectory of life satisfaction could not be adequately described as linear in age and because we found that a cubic term was not required to fit the data. Thus if the original (i.e. not differenced) relationship is S = k + αa + βa², where S is life satisfaction and a is age, then the first-difference relationship is $\mathrm{\Delta }{S}_{i,t}=S_{i,t+1}-S_{i,t}=\alpha \left(a_{i,t+1}-a_{i,t}\right)+\beta \left(a_{i,t+1}^2-a_{i,t}^2\right)$ where a_i,j is the age of the ith person at wave j measured in years since age 65 to the precision of months. On average a_i,t+1 – a_i,t = 2 years in the HRS, but because of scheduling of the surveys and difficulties in making appointments a_i,t+1 – a_i,t varies at the individual level: it has a standard deviation of 0.39. Based on the relationship between life satisfaction and health that is observed in cross-section, we include specifications that control for transitions in various health measures, marital status, and labor market status, which will likely reduce both α and β.

3. Results

Table 1 shows the distribution of personal and household characteristics and the weighted mean values of life satisfaction among respondents age 65–74. We present descriptive statistics for this narrower age band of relatively younger respondents to control for covariates that vary with age. The sample has more females than males reflecting the greater survival of females. Females have slightly lower levels of life satisfaction. The gradient in life satisfaction across wealth and income quartiles is similar, although somewhat steeper in wealth. The difference in life satisfaction between lowest and highest wealth quartiles is 0.38. The gradient with respect to education is smaller. Self-assessed health shows a very strong gradient: there is more than one full category of life satisfaction between excellent and poor, which is about three times the difference between the bottom and top wealth quartiles. The variation across ADL limitations is also strong. The observed health gradient reinforces the puzzle as to why life satisfaction would increase with age even though health declines with age.

Table 1.

Distribution of characteristics and life satisfaction, ages 65–74

	N	Distribution	Average life satisfaction
Sex
Female	14,439	58.8	3.90
Male	10,113	41.2	3.92
All	24,552	100.0	3.91
Wealth quartile
Lowest	6,190	25.2	3.69
2	6,091	24.8	3.87
3	6,137	25.0	3.95
Highest	6,134	25.0	4.07
All	24,552	100.0	3.91
Income quartile
Lowest	6,142	25.0	3.76
2	6,134	25.0	3.85
3	6,139	25.0	3.93
Highest	6,137	25.0	4.04
All	24,552	100.0	3.91
Education
less than high school	4,598	18.7	3.80
high school	8,755	35.7	3.90
some college	5,751	23.4	3.88
college graduate	5,438	22.2	4.00
All	24,542	100.0	3.91
ADL limitations
0	21,106	86.0	3.97
1	1,780	7.3	3.62
2 or more	1,657	6.8	3.32
All	24,543	100.0	3.91
Self-rated health
Excellent	2,099	8.6	4.34
Very good	7,636	31.1	4.12
Good	8,266	33.7	3.90
Fair	4,883	19.9	3.58
Poor	1,646	6.7	3.13
All	24,530	100.0	3.91

Notes: Wealth and income quartiles are constructed separately among married and single individuals by age bands (65–74; 75–84; 85+).

Figure 1 shows the cross-section variation in life satisfaction from age 51 to 89. The pattern is consistent with the literature in that life satisfaction increases monotonically after age 51. The increase is steepest between age 57 and 65, around the time when most individuals retire. There is a modest increase between age 65 and 71 (from 3.89 to 3.94), and the profile is flat after 71.

Fig. 1.

Average life satisfaction, cross-section

However, mortality is strongly associated with life satisfaction. Among HRS respondents over age 50, mortality is substantially higher among those who are not or somewhat satisfied compared to those who are very or completely satisfied with their lives (see Figure 2). The average 2-year mortality rate is 5.6% among those with high life satisfaction, while it is 9.3% (or 66% higher) among those who are not or somewhat satisfied with their lives. This large differential suggests that the cross-sectional pattern does not reflect the actual trajectory of life satisfaction of individuals: those who are more satisfied with their lives live longer and make up a larger fraction of the sample at older ages.

Fig. 2.

Two-year mortality by level of life satisfaction

Note: Five response categories collapsed into two. Top line comprises the three least satisfied categories (33% of the sample) and the bottom line represents the two most satisfied categories.

To focus on the paradox of well-being, in the rest of the paper we restrict the sample to those age 65 or older. To find how life satisfaction changes as individuals age, rather than how life satisfaction varies across individuals of different ages, we use the longitudinal dimension of the HRS and select on individuals who report life satisfaction in two consecutive waves. Figure 3 shows for alternating initial ages beginning at age 65, the two-year trajectories of life satisfaction. For example, among those who were observed in adjacent waves at ages 65 and 67, average life satisfaction increased from 3.905 to 3.913. Thus, the slopes of the lines show the average two-year change in life satisfaction by initial age. With the exception of initial ages 65 and 69 the slopes are negative at all ages, showing that on average the two-year change in life satisfaction experienced by individuals who responded in consecutive waves was negative: as individuals aged, life satisfaction declined. The generally higher placement of the lines as age increases reflects the increasing level of life satisfaction shown in Figure 1: across the two-year line segments the sample changes due to differential mortality and other forms of differential non-response.

Fig. 3.

Average life satisfaction in adjacent waves in panel

Note: Wave-to-wave changes (approximately 2-years) in life satisfaction by initial age. Each segment shows the 2-year trajectory of average life satisfaction estimated on balanced samples with valid reports in the two adjacent survey waves.

Table 2 quantifies the relationship between the level of life satisfaction in one HRS wave (t) and the reporting circumstances in the subsequent wave (t+1). It shows that in addition to mortality selection other types of non-response affect the cross-sectional age pattern in life satisfaction, acting in the same direction. We use 39,460 non-missing values on life satisfaction observed in wave t, irrespective of interview status in the next wave. Thus, the row for “All” shows a highly aggregated cross-section (three age bands) of life satisfaction averaged over 39,460 observations from HRS waves 2008, 2010, 2012 and 2014. It shows modestly increasing life satisfaction with age. The first category (row 1) is restricted to those who were also interviewed in the succeeding wave and reported a value of life satisfaction in that following wave, the panel sample. That group provides the data underlying Figure 3, but Table 2 only shows the initial wave values, not the subsequent wave values. The proportion of the sample that is observed in panel decreases with age from almost 89% in the youngest age band to 63% in the oldest age band, pointing to growing importance of mortality and non-response at advanced ages. The average values of life satisfaction reported in the initial wave, which are cross-section averages, increase just slightly with age. The other rows in the table show life satisfaction in the initial wave (t) according to the interview status in the subsequent wave (t+1). Rows 2, 3 and 4 consist of those who were interviewed at wave t but did not report a value at wave t+1. The percentage of non-responders in the total is small in the youngest age band but increases with age, amounting to 10 percent at 85+. In all three age bands, average life satisfaction is statistically significantly lower in this non-response group (All, rows 2–4) compared to the panel sample in row 1, and the differences between them grow with age.

Table 2.

Life satisfaction in wave t stratified by interview status in the following wave

	Percent Distribution				Average Life Satisfaction
Response status in following wave	65–74	75–84	85+	All	65–74	75–84	85+	All
1. Interviewed, value reported	88.7	81.9	63.0	83.1	3.93	3.97	4.02	3.96
					[0.008]	[0.011]	[0.019]	[0.007]
2–4. Interviewed, no value reported
2. Item non-response	0.6	0.9	1.5	0.8	3.61***	3.82*	3.87	3.75***
					[0.089]	[0.086]	[0.113]	[0.055]
3. Proxy interview, no cognitive limitation	0.9	1.1	1.7	1.1	3.98	3.87	3.90	3.92
					[0.069]	[0.081]	[0.090]	[0.046]
4. Proxy interview, cognitive limitations	0.7	2.5	6.8	2.1	3.80	3.88*	3.82***	3.84***
					[0.092]	[0.048]	[0.052]	[0.033]
All, rows 2–4	2.2	4.5	10.0	4.0	3.82**	3.87***	3.84***	3.85***
					[0.048]	[0.038]	[0.042]	[0.024]
5. Not interviewed, alive	4.8	5.0	4.8	4.9	3.86**	3.9**	3.89**	3.88***
					[0.029]	[0.035]	[0.061]	[0.021]
6. Not interviewed, dead	4.3	8.6	22.3	8.1	3.61***	3.7***	3.86***	3.73***
					[0.034]	[0.029]	[0.029]	[0.017]
All, rows 2–6	11.3	18.1	37.1	17.0	3.76***	3.79***	3.86***	3.80***
					[0.020]	[0.019]	[0.022]	[0.012]
All	100.0	100.0	100.0	100.0	3.91	3.94	3.96	3.93
					[0.008]	[0.010]	[0.016]	[0.006]
Number of observations	20,246	14,365	4,849	39,460

Notes: The square brackets show the standard errors of the means that are adjusted for clustering at the person-level. The stars indicate statistically significant differences from the mean in row 1.

^***p<0.01

^**p<0.05

^*p<0.1.

In rows 2, 3 and 4, we stratify by the reason for non-response among those who were re-interviewed in t+1. Row 2 shows life satisfaction among those interviewed in t+1 but whose response is missing (item nonresponse). It is substantially lower than that of panel responders (row 1), although the differences are only statistically significant for the youngest age group (65–74) and overall (65+). Rows 3 and 4 consist of those who were interviewed by proxy in the following wave. Life satisfaction (like other subjective indicators) is not asked in proxy interviews. We have classified the responses by reason for proxy interview. If the reason was not due to a cognitive limitation (according to the HRS interviewer), life satisfaction in wave t was little different from the panel respondents, and the differences were not statistically different from panel respondents (row 1) in any age group. The percentage in this group increased with age, but it remained under 2% even in the oldest subsample. When the reason for proxy interview was cognitive limitation, however, life satisfaction in wave t was substantially lower than in the panel sample, and the differences were statistically significantly different from the panel sample in the 85+ group and in the total sample (65+). The percentage of the sample in row 4 increased with age from 0.7% in the youngest to 6.8% in the oldest age group.

Rows 5 and 6 comprise those who were not interviewed in the subsequent wave because of loss to follow-up (5) or mortality (6). The percentage of survivors is constant with age at about 5% and their life satisfaction in wave t is slightly, but statistically significantly lower than in the panel sample. Row 6 shows that the portion of the sample that died between waves increased with age from 4.3% to 22.3%. In the age band 65–74 life satisfaction at wave t was much below that of panel respondents but the mortality rate was just 4.3%, so that the impact of differential mortality on any cross-section age pattern was relatively minor. By age 85 or older, however, mortality was 22.3%. Average life satisfaction among those who died prior to the next wave was substantially below that of the panel sample and the differences were statistically significant at the 1% level. Also of note, life satisfaction among those who died prior to the next wave increases in cross-section with age, reflecting selective survivorship.

Overall, the percentage of the sample not reporting life satisfaction in the following wave due to any of the five types of nonresponse increased with age from about 11% to 18% to 37%. Their average life satisfaction levels in wave t were about 0.16 less than the life satisfaction levels of those who did report in the next wave, thus depleting the sample of people with lower levels of life satisfaction and increasing the average level of life satisfaction of survivors in the sample. The differences are statistically significant in all age groups. The main component of nonresponse is differential mortality, especially at advanced old age, but the other types contribute and should be part of any analysis that compares cross-section with panel. In particular, many household surveys of the older population do not attempt interviews by proxy and so do not know what fraction of nonrespondents have cognitive limitations. Because a proxy interview due to cognitive limitations likely signals future panel attrition its effect on the cross-section age pattern is similar to mortality.

Qualitatively the relationships between life satisfaction in a wave and nonresponse in the following wave that are shown in Table 2 can also be detected in longer lags. Appendix Table A1 shows life satisfaction in wave t according to response type in wave t+1, t+2, t+3 and t+4. Of those interviewed and reporting a value for life satisfaction in wave t, 50.4% were interviewed and reported a value for life satisfaction in wave t+4 (eight years later); 35.5% had died by wave t+4; and 14.1% of wave 4 responses were missing for other reasons. Among those in the first category, average life satisfaction in wave t was 4.02 whereas it was just 3.85 among those who died by wave t+4. Thus the selection into mortality as a function of the level of life satisfaction is observed even at a lag of eight years, although with diminished force. An implication is that the selection is not just due to reduced life satisfaction associated with health events immediately preceding death.

Figure 3 shows that after age 65 life satisfaction declines with age over most age segments. To quantify the implication of those declines for the trajectory of life satisfaction we estimated both a nonparametric trajectory and a parametric trajectory. For the former we tied together the segments of the panel transitions we obtained from the two-year age profiles of life satisfaction shown in Figure 3 and anchored the trajectory at the average level observed at age 65. The resulting trajectory is shown in Figure 4 (solid line). According to this non-parametric estimate of the longitudinal age profile, life satisfaction declines with age from 3.91 at age 65 to 3.52 at age 89, a decline of 0.39 (or 0.45 standard deviation). This is in contrast with the cross-sectional profile (short-dashed line) that shows a modest increase from age 65 to 71 and flatness after that.

Fig. 4.

Estimated trajectories of well-being

Note: The cross-sectional line shows average life satisfaction in the full sample. The three panel lines are restricted to observations with valid reports in two adjacent waves. The non-parametric panel model ties together the segments of the panel transitions in Figure 3 and anchors the trajectory at the average level observed at age 65 in Figure 3. The other two lines are based on first-differences regression models with different sets of controls (columns 3 & 4 of Table 3). The “quadratic age” model only includes age and age-squared. The “more controls” model also includes transitions between marital status, self-assessed health states, ADL and IADL limitations, cognition, pain, and labor market states.

Informed by the shape of the nonparametric trajectory, we specified that life satisfaction follows a path that is quadratic in age and estimated the regression of the first-difference in life satisfaction on the first-difference in age and in age squared.³ The fitted trajectory from this estimate is the orange dashed line in Figure 4. It tracks closely the nonparametric trajectory. We also estimated a model in which we added categorical variables for transitions between states: they include transitions between marital status, self-assessed health states, three levels of ADL limitations, three levels of IADL limitations, three levels of cognition, three levels of pain, and labor market states. The dotted line in Figure 4 is the fitted age trajectory when the age coefficients from this augmented regression are used. Because of the strong correlation between these indicator variables and age, their inclusion reduces the age effect so that the age trajectory declines by about a third as much. For example, the model without detailed control variables predicts a 0.35 decline in life satisfaction between age 65 and 89 compared to a 0.24 decline in the model with detailed controls.

The complete regression results are shown in Table 3. The transition from married to widowed is accompanied by a reduction of life satisfaction of 0.19. This is about the same reduction as aging from 65 to 82. Life satisfaction increased by 0.28 on the transition to married from single. Several of the health transitions are predictive of life satisfaction: declining self-assessed health, increases in ADL and IADL functional limitations all strongly predict declines in life satisfaction. For example, the transition from good to fair/poor health predicts a decline in life satisfaction of 0.22, which is quantitatively comparable to the change that accompanies the transition in marital status. The effects of work status, cognitive abilities, and pain, however, are less related to life satisfaction after adjusting for age, marital status, and the other health measures.

Table 3.

First-differences regression models of life satisfaction on age

	Linear models		Quadratic models
	No controls	With controls	No controls	With controls
Difference in age (years after 65)	−0.01249***	−0.00843	−0.00014	0.00092
	0.00193	0.00521	0.00402	0.00643
Difference in age-squared			−0.00060***	−0.00045**
			0.00015	0.00018
Transitions in work status (ref: not work to not work)
Not work to work		0.047		0.039
		0.035		0.035
Work to not work		−0.029		−0.036
		0.021		0.022
Work to work		0.020*		0.013
		0.011		0.012
Transitions in marital status (ref: married to married)
Married to widowed		−0.191***		−0.185***
		0.035		0.035
Married to divorced		−0.122		−0.125
		0.106		0.106
Single to married		0.279***		0.277***
		0.078		0.078
Single to single		0.040***		0.046***
		0.008		0.009
Transitions in self-rated health (ref: Very good/excellent to very good/excellent)
Fair/poor to fair/poor		−0.055***		−0.057***
		0.017		0.017
Fair/poor to good		0.181***		0.181***
		0.027		0.027
Poor to very good/excellent		0.216***		0.218***
		0.052		0.052
Good to fair/poor		−0.222***		−0.221***
		0.024		0.024
Good to good		0.000		0.001
		0.012		0.012
Good to very good/excellent		0.062***		0.064***
		0.020		0.020
Very good/excellent to fair/poor		−0.273***		−0.271***
		0.042		0.042
Very good/excellent to good		−0.095***		−0.093***
		0.018		0.018
Transitions in #ADLs (ref: 0 to 0)
0 to 1 ADLs		−0.077***		−0.074***
		0.026		0.026
0 to 2+ ADLs		−0.134***		−0.129***
		0.040		0.040
1 to 0 ADL		0.045		0.048
		0.030		0.030
1 to 1 ADLs		0.000		0.003
		0.034		0.034
1 to 2+ ADLs		−0.069		−0.065
		0.049		0.049
2+ to 0 ADL		0.128**		0.129**
		0.064		0.064
2+ to 1 ADLs		0.045		0.048
		0.060		0.060
2+ to 2+ ADLs		−0.008		−0.006
		0.033		0.033
Transitions in # IADLs (ref: 0 to 0)
0 to 1 IADLs		−0.027		−0.022
		0.027		0.027
0 to 2+ IADLs		−0.147***		−0.140***
		0.044		0.044
1 to 0 IADL		0.083**		0.086**
		0.036		0.036
1 to 1 IADLs		0.010		0.015
		0.037		0.038
1 to 2+ IADLs		0.125**		0.133***
		0.051		0.051
2+ to 0 IADL		0.231***		0.233***
		0.076		0.076
2+ to 1 IADLs		0.041		0.046
		0.061		0.061
2+ to 2+ IADLs		0.003		0.011
		0.034		0.034
Transitions in cognitive ability (ref: high to high)
Low to low		0.048***		0.045***
		0.015		0.015
Low to medium		0.012		0.010
		0.023		0.023
Low to high		0.007		0.005
		0.037		0.037
Medium to low		−0.014		−0.015
		0.020		0.020
Medium to medium		0.002		0.002
		0.014		0.014
Medium to high		0.008		0.006
		0.018		0.018
High to low		−0.002		−0.004
		0.029		0.029
High to medium		−0.030*		−0.031*
		0.018		0.018
Transitions in pain (ref: no pain to no pain)
No to mild		−0.019		−0.020
		0.024		0.024
No to moderate/severe		0.008		0.007
		0.021		0.021
Mild to no		0.006		0.005
		0.026		0.026
Mild to mild		0.054*		0.051*
		0.029		0.029
Mild to moderate/severe		−0.054*		−0.057*
		0.029		0.029
Moderate/severe to no		0.019		0.018
		0.025		0.025
Moderate/severe to mild		0.005		0.003
		0.031		0.031
Moderate/severe to moderate/severe		−0.004		−0.008
		0.015		0.015
Observations	32245	32245	32245	32245
R-squared	0.001	0.024	0.001	0.024

^***p<0.01

^**p<0.05

^*p<0.1.

Standard errors adjusted for clustering at person-level. Sample: HRS, 2008–2016, age 65+, observations with non-missing life satisfaction reports in two consecutive survey waves. The left-hand side variable is wave-to-wave differences in life satisfaction. The main explanatory variable is difference in age, a_i,t+1 − a_i,t.

To explore heterogeneity in the trajectory of life satisfaction we estimated the regression of the change in life satisfaction on the change in age separately for males, females, ages 65–74, ages 75 or older and four education bands. To be able to make a simple comparison across groups we used a linear specification in age. The coefficient on age can be interpreted as the one-year change in life satisfaction averaged over ages 65–89. With just one exception (some college) the coefficient on age is negative and significant in all the regressions (see Table 4). The coefficients on age are about the same across these groups except that the age coefficient for those 75 or older implies a greater rate of decline than the age coefficient for those 65–74, reflecting the concavity shown in Figure 4. The implication is that variation in the rate of decline is associated with age, not with sex or education.

Table 4.

Age coefficients from first-differences regression models of life satisfaction on age in population subgroups, with and without controls

	No controls		With controls
	Coefficient	Std error	Coefficient	Std error
All	−0.01249***	0.00193	−0.00843	0.00521
Male	−0.01039***	0.00298	−0.00629	0.00794
Female	−0.01404***	0.00253	−0.01015	0.00695
Age 65–74	−0.00588**	0.00272	−0.00081	0.00707
age75 plus	−0.02232***	0.00281	−0.02034***	0.00782
< high school	−0.01484***	0.0053	−0.02999	0.02521
High school	−0.01330***	0.00318	0.00689	0.00946
Some college	−0.00768*	0.00395	−0.00681	0.01118
College	−0.01415***	0.00367	−0.01565**	0.00785

^***p<0.01

^**p<0.05

^*p<0.1.

Standard errors adjusted for clustering at person-level. Sample: HRS, 2008–2016, age 65+, observations with non-missing life satisfaction reports in two consecutive survey waves. Each row of the table corresponds to a different regression model on a different population subgroup. The left-hand variable contains the wave-to-wave difference in life satisfaction. The main explanatory variable is the difference in age, a_i,t+1 − a_i,t.

Age enters the models linearly. The main objective of these estimations is a comparison of the slopes in age by subgroups. Although a quadratic specification in age would fit the data better, the linear specification allows immediate interpretation of the slope. The complete output of these regressions is available from the authors upon request.

4. Discussion and conclusion

Historically, public policy and government programs have relied on objective indicators of well-being, such as income or wealth to judge their success. However, the objective measures capture only a narrow component of overall individual well-being. Understanding the subjective life satisfaction of older individuals is particularly important because older adults are more likely to experience challenging life events such as health problems, and the deaths of their friends and relatives, the effects of which are not captured by measures of material well-being. Moreover, it may be more difficult at older ages to compensate for a health shock and other life events because of health constraints on activity.

Despite expectations induced by the increasing frequency of health shocks with age, the literature has typically found a positive association between age and life satisfaction at older ages in cross-sectional studies. This “paradox of well-being” may suggest that older individuals’ subjective well-being is quite resilient, and perhaps the well-being of the older population should not be overly concerning for policy makers despite their declining health and increasing incidence of widowing.

However, most of the literature used cross-sectional methods that embed various forms of selection, particularly mortality selection, or used longitudinal methods based on survey data that cover the entire adult age range, not just old age. Our main contribution is to use better data, the longitudinal data from the HRS, which carefully tracks panel members, and which consistently records vital status. Its large samples at older ages allowed us to estimate models of life satisfaction with flexible functions of age, and to investigate how health conditions, widowing, and other aging-related life events influence this relationship.

We found that the cross-sectional and the longitudinal age profile of life satisfaction are quite different after age 65. In the cross-section, life satisfaction increases slightly from age 65 to 71, but then is flat thereafter. This cross-sectional profile, however, embeds several types of selection. Individuals with higher life satisfaction live longer, and they make up a larger fraction of the population at older ages, which biases the age profile of life satisfaction upward. Differential non-response among survivors has a similar, though less strong effect on the age profile of life satisfaction. When we accounted for these biases using longitudinal models that tracked the life satisfaction of the same individuals over time, we found that life satisfaction significantly falls with age, and the rate of decline accelerates with age. Widowing and health shocks played important roles in this decline.

Studies that used methods closest to ours (Gana et al. 2013, Hansen and Slagsvold 2012, Jivraj et al. 2014, Kunzmann et al. 2000, Shankar et al. 2015) relied on smaller samples from different European countries, and they did not directly analyze the effect of mortality and differential non-response on life satisfaction. Zhang et al. (2017) used HRS data and a mixed effects methodology to study predictors of life satisfaction at old age. They found a positive association between age and life satisfaction. Our results likely differ from theirs because we directly investigate the effects of mortality selection and other types of survey nonresponse. When we corrected for both types of selection, we found that life satisfaction significantly declines with age in longitudinal measurements.

The bias of cross-sectional age profiles induced by the sources of selection that we identified especially at advanced ages affects not just life satisfaction, but likely extends to other dimensions of subjective well-being, and – more broadly – to other outcomes that correlate with mortality and non-response due to cognitive limitations. For example, Grol-Prokopczyk (2017) documented the effect of differential mortality on cross-sectional age profiles of pain (see also Case et al. 2020). Similar selection issues may affect the age patterns of depression (Luppa et al. 2012), and other physical and mental health outcomes (Hantke et al. 2020).

Our results are significant for at least two reasons: First, it is important to resolve the controversy about the trajectory of life satisfaction at advanced ages so that future research does not expend effort to explain a misleading finding. Second, public policy increasingly considers subjective measures of well-being, such as life satisfaction, in decision making. As it aims to improve the well-being of the population it will judge the needs of subpopulations, balancing the needs of the older population with those of the younger population. Our results suggest that the optimistic picture of increasing life satisfaction among older persons based on cross-sectional data is misleading, possibly inducing policy makers to incorrectly conclude that the needs of the older population are of lesser concern.

Highlights.

1. Older people express higher life satisfaction, the “Paradox of well-being.”

2. But in longitudinal data life satisfaction declines with age after 65.

3. Resolution of paradox: individuals with high life satisfaction stay in study longer.

4. Population life satisfaction higher at older ages due to mortality & other selection.

5. Life satisfaction of individuals declines due to health and widowing.

Appendix Table

Table A1.

Life satisfaction in wave t stratified by interview status in the subsequent four waves

	Percent Distribution				Average Life Satisfaction at t
Response status in the later wave	at t+1	at t+2	at t+3	at t+4	at t+1	at t+2	at t+3	at t+4
1. Interviewed, value reported	83.1	72.3	61.7	50.4	3.96	3.97	4.00	4.02
2. Interviewed, value not reported
3. Item non-response	0.8	0.8	0.8	0.7	3.75	3.92	3.90	3.89
4. Proxy interview, no cognitive limitation	1.1	1.0	1.0	0.9	3.92	3.97	3.96	4.18
5. Proxy interview, cognitive limitations	2.1	2.8	3.2	3.2	3.84	3.87	3.87	3.86
6. Not interviewed, alive	4.9	6.5	7.9	9.3	3.88	3.91	3.93	3.99
7. Not interviewed, dead	8.1	16.5	25.3	35.5	3.73	3.77	3.82	3.85
8. All	100.0	100.0	100.0	100.0	3.93	3.93	3.94	3.95
Number of observations	39,460	30,002	20,237	10,422