NBER WORKING PAPER SERIES
IS WELL-BEING U-SHAPED OVER THE LIFE CYCLE?
David G. Blanchflower
Working Paper 12935
NATIONAL BUREAU OF ECONOMIC RESEARCH
1050 Massachusetts Avenue
Cambridge, MA 02138
We thank two referees, Andrew Clark and Richard Easterlin for helpful suggestions. The second author's
work was funded by an ESRC professorial research fellowship The views expressed herein are those
of the author(s) and do not necessarily reflect the views of the National Bureau of Economic Research.
© 2007 by David G. Blanchflower and Andrew Oswald. All rights reserved. Short sections of text,
not to exceed two paragraphs, may be quoted without explicit permission provided that full credit,
including © notice, is given to the source.
Is Well-being U-Shaped over the Life Cycle?
David G. Blanchflower and Andrew Oswald
NBER Working Paper No. 12935
JEL No. I1,J0
Recent research has argued that psychological well-being is U-shaped through the life cycle. The
difficulty with such a claim is that there are likely to be omitted cohort effects (earlier generations
may have been born in, say, particularly good or bad times). Hence the apparent U may be an artifact.
Using data on approximately 500,000 Americans and Europeans, this paper designs a test that makes
it possible to allow for different birth-cohorts. A robust U-shape of happiness in age is found. Ceteris
paribus, well-being reaches a minimum, on both sides of the Atlantic, in people's mid to late 40s.
The paper also shows that in the United States the well-being of successive birth-cohorts has gradually
fallen through time. In Europe, newer birth-cohorts are happier.
David G. Blanchflower
Department of Economics
6106 Rockefeller Hall
Hanover, NH 03755-3514
Department of Economics
Coventry CV4 7AL
Is Well-being U-Shaped over the Life Cycle?
A large social-science literature is emerging on the determinants of happiness and
mental well-being. As would be expected, this topic has attracted the attention of
medical statisticians, psychologists, economists, and other investigators (including
Easterlin 2003, Frey and Stutzer 2002, Lucas et al 2004, Layard 2005, Smith et al
2005, Ubel et al 2005, Gilbert 2006, and Kahneman et al 2006). However, a
fundamental research question remains poorly understood. What is the relationship
between age and well-being?
Traditional surveys of the field, such as Myers (1992), Diener et al (1999) and
Argyle (2001), argue that happiness is either flat or very slightly increasing in age.
Mroczek and Kolarz (1998) provide a discussion of psychologists’ earlier writings.
However, much new work has argued that there is evidence of a U-shape through the
life cycle. In cross-sections, even after correcting for potentially confounding
influences, there is now thought to be a convex link between reported well-being and
age. This modern literature includes Clark and Oswald (1994), Gerlach and Stephan
(1996), Oswald (1997), Theodossiou (1998), Winkelmann and Winkelmann (1998),
Blanchflower (2001), Di Tella et al (2001, 2003), Frey and Stutzer (2002),
Blanchflower and Oswald (2004), Graham (2005), Frijters et al (2004, 2005), Senik
(2004), Van Praag and Ferrer-I-Carbonell (2004), Shields and Wheatley Price (2005),
Oswald and Powdthavee (2005), Propper et al (2005), Powdthavee (2005), Bell and
Blanchflower (2006), Uppal (2006), and Blanchflower and Oswald (2007). Clark et
al (1996) makes a similar argument for job satisfaction equations. Pinquart and
Sorensen (2001) develops an equivalent case for a measure of loneliness, and Hayo
and Seifert (2003) does so for a measure of economic subjective well-being. Jorm
(2000) reviews the psychiatric evidence and concludes that there are conflicting
results on how the probability of depression alters over the life course.
There is an important difficulty with the conclusion that well-being is U-
shaped in age. As Easterlin (2006) points out, the effect of an age variable is likely to
be contaminated by omitted cohort effects (earlier generations may have been born in,
say, particularly good or bad times). Hence the U-shape in age, uncovered now by
various authors, could be an artifact of the data.
This is more than a theoretical possibility. Suicide levels seem to vary across
cohorts (Stockard and O’Brien 2002). Moreover, Blanchflower and Oswald (2000)
find some evidence of rising well-being among young people. There is also evidence
-- for example, in Sacker and Wiggins 2002 -- that levels of depression and
psychiatric distress, measured consistently across cohorts, have risen in countries
such as Great Britain. Nevertheless, these matters are still the subject of debate
(Murphy et al 2000, Paykel 2000).
New work by Clark and Oswald (2007) argues that in British panel data on
well-being it can be shown that the U-shape in age is identified entirely from the
longitudinal element of the data set. The authors’ study can be thought of as literally
following the aging process of particular individuals at different points in the lifespan.
Nevertheless, such research is rare and does not allow cohort effects to be examined,
and it seems important to inquire into the foundations of the U-shape in other nations.
2. Testing for Cohort Effects
This paper offers some of the first cross-country evidence that the curvilinear
relationship is robust to cohort effects. We draw upon randomly sampled data on
approximately 500,000 Americans and Europeans. These data come from the
General Social Surveys of the United States and the Eurobarometer Surveys, and,
necessarily for the design of the test, cover some decades.
One point, however, should perhaps be made clear from the outset. It is that
the paper can examine only simple so-called single-item measures of well-being, so
cannot allow subtle differentiation -- as favoured in psychology journals -- into what
might be thought of as different types of, or sides to, human happiness or mental
health. Nevertheless, the patterns that emerge are perhaps of interest.
After controlling for different birth-cohorts, the paper finds that ceteris-
paribus well-being reaches its minimum in a person’s 40s. This U-shape is similar for
males and females, and on each side of the Atlantic Ocean. Moreover, because of the
size of our data sets, the turning point in well-being -- the age at which happiness
begins to lift back up -- is fairly precisely determined.
The paper’s concern is with the ceteris paribus correlation between well-being
and age, so we later partial out other factors, such as income and marital status, that
both alter over a typical person’s lifetime and have effects upon well-being. This
follows one particular tradition of empirical research. We read the effect of a
variable’s coefficient from a long regression equation in which other influences have
been controlled for as effectively as possible.
Despite the commonness of this convention in modern social-science research,
such a method is not inevitable. A valid and different approach is that of, for
example, Easterlin (2006), who focuses on the raw or reduced-form link between
happiness and age. Interestingly, he finds evidence of an inverted U-shape. As
Easterlin points out, and as explained also in Blanchflower and Oswald (2004), if few
or no control variables other than age are included in an American happiness
regression estimated from the General Social Survey, the effect of age is concave and
not convex. A related result is that of Mroczek and Spiro (2005), who establish in a
data set on American veterans, where the youngest person in the data set is 40 years
old (making it hard to do an exact comparison with research on the GSS), that
happiness rises to the early 60s, and then appears to decline.
As common observation shows, the quality of a person’s health and physical
abilities depends sensitively on the point in the life cycle. Most diseases, and the
probability of getting them, worsen with age. A 90 year old man cannot in general do
the same number of push-ups as a 20 year old man. Hence an important issue is
whether in happiness equations it is desirable to control in some way for health and
physical vitality. There is here no unambiguously correct answer, but the approach
taken in the paper is not to include independent variables that measure physical
health. This is partly pragmatic: our data sets have no objective measures and few
subjective ones. But the decision is partly substantive: it seems interesting to ask
whether older people are happier once only simple demographic and economic
variables are held constant.
3. Conceptual Issues
There is relatively little social-science theory upon which to draw (though
mention should perhaps be made of Carstensen’s theory, which, put informally, is that
age is associated with increasing motivation to derive emotional meaning from life
and decreasing motivation to expand one's horizons: Carstensen et al 1999 and
Charles et al 2001).
Conventional economics is in principle capable of making predictions about
the life cycle structure of happiness if conceptualized as utility in the normal
economist’s framework. However, in practice, economists’ standard life-cycle theory
does not generate a U-shape in a straightforward way. Instead, the natural conclusion
is that well-being might be predicted to be independent of age.
To see why, let the individual person be concerned to maximize lifetime
utility V by choosing a consumption path c(a) where a is the individual’s age. Assume
lifespan runs deterministically from time point t to time point T. Assume away
discounting for simplicity (it is straightforward to show here that it makes no
substantive difference, given an efficient capital market where people both discount
utility at rate r and can lend or borrow at interest rate r). Let income y be fixed and
given by the agent’s talent endowment, and for simplicity set this to unity. Then the
agent chooses consumption c at each age a to maximize lifetime happiness
subject to an inter-temporal borrowing constraint
in which the endowment of income to be allocated across all the periods has been
normalized to one. Assume that u, utility or well-being, is an increasing and concave
function of consumption, c. Spending, by assumption, then makes people happier.
This is a so-called isoperimetric problem. The first-order condition for a
maximum is the usual one: it requires the marginal utility of consumption to be the
same at each level of age, a. Therefore, solving a Lagrangean L constructed from (1)
where, from the underlying mathematical structure, the multiplier lambda is
necessarily constant across all the different ages from t to T. Individuals thus allocate
their discretionary spending to the points in time when they enjoy it most.
If the utility function u(c,a) is additively separable in consumption c and age a,
then equation (3) has a simple implication. It is one that is implicit in much of
economic theory. Consumption will be flat through time (because under separability
u = u(c)) + v(a)) and, therefore, utility will also be flat through the lifespan if the non-
consumption part of utility, v(.), is independent of age. In plainer language,
happiness will not alter over a person’s life course.
It is reasonable to suggest that to go from the utility function u = u(c,a) to the
presumption that u(..) is additively separable in its two arguments is a large, and
potentially unwarranted, step. There is no clear reason why the marginal utility of
consumption would be independent of a person’s age. For example, one might
believe that young people wish to signal their status more, and therefore might have a
greater return from units of consumption than the old (so the cross-partial derivative
of u(c,a) would then be negative). Alternatively, one might argue that older people
have more need of health and medical spending, and therefore that the marginal
utility of c is greatest at high levels of a. Then the cross-partial of u(c,a) is positive.
While it would be possible to assume that early in life the first effect
dominates and then in later life the second one dominates, and in this way get
eventually to a model where well-being was U-shaped through the lifespan, to do so
seems too ad hoc (or post-hoc) to be persuasive theoretically. What this means is that
textbook economics -- without making assumptions about v(a) that could
mechanically lead to any desired shape -- is not capable of producing clear
predictions about the nonlinear pattern of well-being through an individual’s life.
4. Empirical Results
To explore this issue empirically, therefore, we draw upon two data sets, which pool
data on approximately half a million randomly selected individuals, and implement a
test that controls for the possible existence of cohort effects. The data do not follow
the same individuals through time. They provide repeated statistically-representative
snapshots, year after year, covering all ages of American and European adults from
age 16 and above.
The key evidence is summarized in four tables.
Table 1 takes all the males in the U.S. General Social Survey (GSS) from
1974-2004. It estimates a happiness regression equation for this sub-sample, and
shows in its early columns that well-being is U-shaped in age. Then cohort variables
are introduced. These take the form of a set of dummy variables – one dummy for
each decade of birth. Although the introduction of the cohort dummies affects the
turning point of the quadratic function in age, it does not do so in a way that changes
the thrust of the idea that well-being follows a U-shaped path. The same statistical
procedure is followed for the analysis of three further sub-samples, namely, the
females in the GSS data set, the males in the Eurobarometer survey, and finally the
females in the same European sample.
The exact wording of the GSS well-being question is: “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?”
In the Eurobarometer survey it is: “On the whole, are you very satisfied, fairly
satisfied, not very satisfied, or not at all satisfied with the life you lead?”
To give a feel for the raw patterns in the data, happiness in the United States
can be expressed in a cardinal way by assigning 1 to 3 to the three answers above,
where ‘very happy’ is a 3. In that case, the mean of US happiness in the data is 2.2
with a standard deviation of 0.6. Similarly, European life satisfaction can be
cardinalized using the integers 1 to 4, where ‘very satisfied’ is a 4. In this case, the
mean of life satisfaction is 3.0 with a standard deviation of 0.8. Well-being answers
are skewed, in both data sets, somewhat towards the upper end of the possible
The paper tests for a U-shape by examining whether the data take a quadratic
form in age. Almost all the coefficients on age-squared variables in the main part of
the paper are statistically significant at the 0.0001 level. We estimate the effects by
using ordered logit equations. The tables report estimated coefficients, which is an
alternative to odds ratios. This option affects only how results are displayed and not
how they are estimated.
In the first column of Table 1 a GSS happiness ordered logit equation is
estimated on the pooled sample of 19,027 American males with age entered as an
independent variable. It has, as further independent regressors, a separate dummy
variable for each year in the data set and for each region of the United States. This is
to mop up year-by-year variation in national well-being and unchanging spatial
characteristics such as regions’ climatic conditions.
The age regressor in the first column of Table 1 has a positive coefficient of
0.0096 and a t-statistic of approximately 11. Hence reported happiness rises as
people get older. In column 2 of Table 1, a set of further regressors are included into
the equation, and the coefficient on age falls somewhat, to 0.0066, with a t-statistic
that indicates it continues to be statistically significantly different from zero at usual
confidence levels. These extra regressors are a variable for the years of education of
the person, two dummies for racial type, 8 dummies for the number of dependent
children of the individual, a collection of different dummy variables to capture the
working status (employed, unemployed, …) of the person, a dummy variable that
takes the value one if the individual reported that his or her parents had divorced by
the time the individual respondent was aged 16, and 4 dummy variables to capture the
person’s marital status. Table 1 goes on to check for a turning point in age. It does so
in the simplest way, by fitting a level and a squared term. Table 1 finds in column 3
that a quadratic form seems to approximate the data well: the equation traces out a
happiness function that reaches a minimum at 36.8 years of age. This is effectively
the U-shaped result in the literature to date.
However, Table 1 then explores the possibility that the U-shape in age is a
product merely of omitted cohort effects. Column 4 of Table 1 extends the
specification by introducing a separate dummy variable for each decade of birth (it
cannot enter a full set of individual birth-year dummies because the result would be
complete collinearity). The outcome is a U-shape in age, but one where the turning
point is now much later in the typical individual’s life. According to the evidence in
column 4 of Table 1, subjective well-being among randomly selected American
males, bottoms out at an estimated 55.9 years. This is to be thought of, of course, as
the minimum-happiness age after controlling for other influences such as education
and marital status.
Finally, column 5 of Table 1 introduces an income measure into the equation
explaining well-being (although the causal interpretation here is open to debate,
Gardner and Oswald 2007 document longitudinal evidence that windfalls raise mental
well-being). For simplicity, and following much of the literature, income is entered
as the natural logarithm of the person’s family income. The coefficient is positive
(with a t-statistic of 6.83), so richer people report higher levels of happiness with their
lives. Here the U-shape in age bottoms out at age 49.5. The sample size is somewhat
reduced, because of missing income observations, to 11,404 people.
The remainder of the paper’s evidence is similar. Table 2 moves to a sub-
sample of females from the US General Social Survey. Compared to Table 1, the
sample size is a little larger (because women live longer than men) at 24,148
individuals. Once again, each reports a well-being answer on a three-point scale from
very happy down to not at all happy, and Table 2 estimates an ordered logit equation
with the same structure as for the males in Table 1.
Perhaps somewhat surprisingly, the analytical structure for American women
is almost the same as for the men.
In Table 2, well-being is at first increasing in age. But once a squared term in
age is introduced, in the third column, it is clear that the data favour a quadratic form,
so once again happiness seems strongly U-shaped in age. When the same set of
cohort dummies are incorporated into the equation, in column 4 of Table 2, the
turning point of the happiness function is at age 44.9 years. This is noticeably less
than the 55.9 years estimated for the male sub-sample. However, allowing for the
separate effect of income upon well-being in column 5 makes women look more like
the men. The minimum in column 5 of Table 2 is reached at age 45.1. Whatever is
going on, in some sense that may not be immediately understandable, these data are
apparently working in roughly but not exactly the same way for American males and
With only minor differences, Tables 3 and 4 tell the same story, but use
Eurobarometer data pooled from 1975 to 1998. Here, of course, the continent is
different and the sample sizes far larger. A slightly different form of well-being
question (on life satisfaction) has to be employed, but as these estimation methods
effectively use only the ordering of well-being answers, the exact wording is unlikely
to matter significantly, and so empirically it seems to prove.
In Table 3, an ordered logit is estimated for 200,848 males from France,
Belgium, Netherlands, West Germany, Italy, Luxembourg, Ireland, Great Britain,
Greece, Spain, and Portugal. To allow comparisons, the aim is to achieve an
econometric specification as close as possible, despite some differences in the data
sets on topics such as the level of detail in the measure of income, to that for the
United States in Tables 1 and 2.
Before the cohort dummies are introduced, the turning point in the male well-
being equation is at a minimum point where age is equal to 43.4 years (see column 3
of Table 3). It is not easy to say why this number might be higher than in the USA
(see column 3 of Table 1), but one possibility is that the Second World War may have
exacted a toll in various ways on this generation of European males. Whatever the
reason, the difference with the United States continues by the time column 4 is
estimated. Now the age at which well-being reaches a minimum is 47.1 years, which
is below the American number.
After the role of income is entered into the specification, the minimum is 44.1
years. Table 4 produces similar figures, and equations, for the female sub-sample of
214,857 randomly sampled European women.
At the suggestion of a referee, the Appendix sets out a number of robustness
checks and inquiries. In the interests of brevity, only the results for males are given.
Table A1 reveals that it is the addition of dummies for marital status that first
makes the U-shape evident in the data of the United States, and this quadratic is
strengthened by a control for years of schooling (see, for example, columns 3 and 4).
It is allowing for an income variable that makes the minimum point of the U-shape in
happiness move considerably further to the right (in the last column of Table A1).
These changes across specifications are less noticeable in European data (as in Table
A3). Table A2 divides the data into sub-samples. It is evident that there is a strong
U-shape in age among the sub-sample of American males who never married. This
suggests that, in the full sample, the quadratic is not merely somehow proxying the
fact that happy people tend to go on to get married more. The same general result is
found for Europe in the final two columns of Table A3, where the minimum point of
well-being is estimated at age 49.1 for single Europeans and 37.6 for ever-married
Europeans. Although they are omitted, equivalent results were found for females in
A full set of interaction terms -- interacting the quadratic in age with the other
independent variables -- was also tried, as a robustness check, but these were found to
have coefficients that were almost always insignificantly different from zero at the
95% confidence level.
5. Measuring the Size of the Age and Cohort Effects on Well-being
Even if statistically significant, is such a U-shape in age large enough to be important
empirically? The data suggest that the answer is yes.
One way to explore this is to compare the levels of well-being between, say,
age 20 and age 45. This difference -- in the equations that control for other factors --
is approximately 0.1 to 0.2 cardinal well-being points, and this is around one fifth of a
standard deviation in well-being scores. At first sight that does not appear
particularly large. But, because the standard deviation is dominated by cross-section
variation in reported well-being, there is a more useful and evocative way to think
about the size of the age and age-squared effect. Going from age 20 to age 45 is
approximately equal to one third of the size of the effect of the unemployment
14 Download full-text
coefficient in a well-being equation. That is suggestive of a large effect on well-
Although the birth-cohort coefficients (on Born<1900, Born 1900-1910, etc)
are not always individually well-defined, there are signs from the Tables that the
United States and Europe differ quite strongly in the time structure of the cohort
effects upon happiness. In Tables 1 and 2, there is evidence that successive American
generations became progressively less happy from 1900 to today. This conclusion is
reminiscent of one of Easterlin’s (2006), although he uses a different statistical
In Europe, by contrast, Tables 3 and 4 suggest that cohort well-being falls
initially from the beginning of the century but, after bottoming out in the 1950s
(which is the omitted base category) has actually been rising throughout the most
recent generations. This is particularly clear for males. The coefficient of 0.3206 (t =
2.36) for the final cohort, in the fifth column of Table 3, implies that, by this criterion
the most recent generation of European men is ceteris paribus the happiest of the 20th
As with the effect of moving along the quadratic function in age, cohort
dummy variables are here large in magnitude; they are not merely different from zero
on a formal significance test. Put loosely, cohort effects are two or three times as
large as the effect from the U-shape in age. The single greatest effect is visible in the
equations for US males in Table 1. Here, comparing the happiest cohort of
Americans to the least happy, the cardinalized well-being difference through the
generations exceeds half of one standard-deviation of the happiness measure. In all
the tables, whilst the details differ, estimated cohort effects are quantitatively
significant and not merely statistically significant.