Happiness and Productivity
Andrew J. Oswald, Eugenio Proto and Daniel Sgroi
WARWICK ECONOMIC RESEARCH PAPERS
DEPARTMENT OF ECONOMICS
Happiness and Productivity
Andrew J. Oswald, Eugenio Proto and Daniel Sgroi
Department of Economics
University of Warwick
21 December 2008
Keywords: Labor productivity; emotions; well-being; happiness; positive affect;
Corresponding author: email@example.com.
Address: Department of Economics, University of Warwick, Coventry CV4 7AL, United
Telephone: (+44) 02476 523510
Acknowledgements: For fine research assistance and valuable discussions, we are indebted
to Malena Digiuni, Alex Dobson and Lucy Rippon. For helpful advice, we would like to
record our deep gratitude to Alice Isen. Seminar audiences at PSE Paris and Zurich
provided insightful suggestions. Thanks also go to Eve Caroli, Emanuele Castano, Andrew
Clark, Alain Cohn, Ernst Fehr, Justina Fischer, Bruno Frey, Dan Gilbert, Amanda Goodall,
Greg Jones, Michel Marechal, Aldo Rustichini, Daniel Schunk, Claudia Senik, and Tanya
Singer. The first author thanks the University of Zurich for a visiting professorship. The
ESRC provided research support.
Little is known by economists about how emotions affect productivity. To make
persuasive progress, some way has to be found to assign people exogenously to
different feelings. We design a randomized trial. In it, some subjects have their
happiness levels increased, while others in a control group do not. We show that a
rise in happiness leads to greater productivity in a paid piece-rate task. The effect is
large; it can be replicated; it is not a reciprocity effect; and it is found equally among
males and females. We discuss the implications for economics.
Happiness and Productivity
Andrew J. Oswald, Eugenio Proto, Daniel Sgroi
There is a large economics literature on individual and economy-wide
productivity. There is also a fast-growing one on the measurement of mental well-
being. Yet economists know little about the interplay between emotions and human
productivity. Although people’s happiness and effort decisions seem likely to be
intertwined, we lack evidence on whether, and how, they are causally connected.
This paper makes two contributions. First, it attempts to alert economists to a
psychology literature in which happiness (or more precisely what psychologists
describe as positive affect) has been shown to be associated with higher human
creativity and performance. The work of the psychologist Alice Isen has been
particularly important1. The second, and main, contribution of our paper is to design
an empirical test that has not been performed in the psychology literature. By doing
so, we address a question of special interest to economists (and arguably to economic
policy-makers): Does happiness make people more productive in a paid task?
The paper finds that it does. We demonstrate this experimentally in a piece-
rate setting with otherwise well-understood properties2.
The links between productivity and human well-being are of interest to many
kinds of social scientists.
Argyle (1989, 2001) points out that little is understood about how life
satisfaction affects productivity, but that there is (mixed) evidence that job
satisfaction exhibits modestly positive correlations with measures of worker
productivity. Work by Wright and Staw (1999) examines connections between
worker affect and supervisors’ ratings of workers. Depending on the affect measure,
the authors find interesting but mixed results. Amabile et al (2005) uncovers
evidence that happiness appears to provoke greater creativity. Baker et al (1997),
Boehm and Lyubomirsky (2008), Paterson et al (2004), Steele and Aronson (1995)
and Tsai et al (2007) detect influences of emotion and affect upon performance. In
1 We list a number of them in the paper’s references; they include a series of papers in the 1980s, Ashby et al
(1999), Erez and Isen (2002), and the recent work of Hermalin and Isen (2008). Our study also connects to the
broaden-and-build approach of Fredrickson and Joiner (2002) and to the ideas of Lyubomirsky et al (2005).
2 The psychology experimenters have not examined productivity in paid piece-rate tasks.
contrast to our paper’s later argument, Sanna et al (1996) suggests that those
individuals who are in a negative mood put forth the most effort.
There is an analytical literature by economists that is especially relevant to our
later empirical findings. Although not directly about mood or happiness, it examines
the interconnections between psychological forces (in particular, biased perception)
and human performance. The paper by Benabou and Tirole (2002) focuses on the
interactions between self-deception, malleability of memory, and ability and effort.
The authors consider the possibility that self-confidence enhances the motivation to
act, so their framework is consistent with the idea that there can be a connection
between mood and productivity. They develop an economic model of why people
value their self-image, and they use this specifically to justify seemingly irrational
practices such as handicapping self-performance or the practising of self-deception
through selective memory loss. Compte and Postlwaite (2004) extends this line of
work, by seeking to identify circumstances in which biased perceptions might
increase welfare. The authors model perceptions as an accumulation of past
experiences given gradual adjustment. Benabou and Tirole (2003) provides a formal
reconciliation of the importance of intrinsic motivations with extrinsic (incentivised)
motivations. Such writings reflect an increasing interest among economists in how to
reconcile external incentives with intrinsic forces such as self-motivation. Our later
results also have implications for standard microeconomics as described in sources
such as Laffont and Tirole (1993). This body of work assumes -- in contrast to later
evidence in the paper -- that choices can be viewed as independent of emotions. 3
We shall not attempt in the paper to distinguish in a sharp way between
happiness and ‘mood’. For simplicity, we shall take the distinction, in a short run
experiment like the one to be described, to be largely semantic.
Nor shall we discuss the possibility that other stimuli such as music, alcohol or
sheer relaxation time -- all mentioned by readers of early drafts -- could have the same
or equivalent effects. Nor shall we measure how long-lasting are the effects of
emotion upon labor productivity. Our instinct, however, is that these are important
topics for future research.
3 A review paper in psychology is Diener et al (1999). A considerable literature in economics has studied
happiness and wellbeing as a dependent variable – including Blanchflower and Oswald (2004), Clark et al (2008),
Clark and Oswald (1994), Di Tella et al (2001, 2004), Easterlin (2003), Frey and Stutzer (2002, 2006), Kahneman
and Sugden (2005), Luttmer (2005), Oswald (1997), Van Praag and Ferrer-I-Carbonell (2004), and Winkelmann
and Winkelmann (1998). For related work on emotions, see Frank (1988), Elster (1998), and Loewenstein (2000).
3. A model of work and distraction
This section describes a theoretical framework. Its aim is partly taxonomic.
The main comparative static result stems from a form of internal resource-allocation
by a worker.
Think of individuals as having a finite amount of energy. Within any period
of time, they must choose how to distribute that across different activities. In one
version of the later model, a happiness shock can be seen as raising the psychological
resources available to a worker. At the margin, the shock frees an overall energy
constraint. That, in turn, allows an individual to devote more effort to solving
problems for pay, and to act as though switching away from other distractions.
Let the worker’s (randomly distributed) ability be z. This has a density
function f(z). Denote p as the piece-rate level of pay. Denote u and v as two different
sources of utility to the individual. Let e be the energy the worker devotes to solving
tasks at work. Let w be the energy the worker devotes to other things -- to
‘distractions’ from work. Let R be the worker’s psychological resources. Hence (e +
w) must be less than or equal to R.
We assume that u, the utility from work, depends on both the worker’s
earnings and effort put into solving work problems. Then v is the utility from
attending broadly to the remaining aspects in life. For concreteness, we shall
sometimes think of this second activity as a form of ‘worrying’. But it can be viewed
as a generalized concern for issues in the worker’s life that need his or her cognitive
attention. In a paid-task setting, it might be realistic to think of a person as
alternating, during the working day, between concentrating on the work task and
being distracted by the rest of his or her life. There is a psychic return from the
energy devoted to distraction and worry -- just as there is a return from concentrating
on the paid task.
Consider an initial happiness shock, h. For the sake of clarity, assume
separability between the two kinds of utility going to the individual. People then
solve the problem: Choose paid-task energy e to
The first-order condition for a maximum in this problem is
The comparative-static result of particular interest here is the response of
productivity, given by work effort e, to a rise in the initial happiness shock, h.
Formally, it is determined in a standard way. The sign of de*/dh takes the sign of the
cross partial of the maximand, so that:
Sign de*/dh takes the sign of
Without more restrictions, this sign could be positive or negative. A happiness shock
could increase or decrease the amount of work done on the maths task.
To get some insight into the likely economic outcome, consider simple forms
of these functions. Assume that workers know their own productivity, so are not
subject to the uncertainty. Let R be normalized to unity. Set z to unity.
Assume that the u and v functions are concave and differentiable. This is not
strictly necessary, but it follows the economist’s modelling tradition, and leads to
natural forms of interior solutions. The analysis is easily generalized.
How then might an exogenous happiness perturbation, h, enter a person’s
objective function? In stylized form, consider three alternative maximands:
u(.) + v(.) + h
u(h, .) + v(h, .)
hu(.) + (1 – h)v(.)
The additive model is -- we conjecture -- what most economists would write
down when asked to think about exogenous emotions and choice. They would view a
happiness shock as a vertical shift upwards in the utility function.
Assuming additively separable functions, and that the worker gets the h
happiness shock whether or not he or she subsequently works or instead worries about
other things, the worker solves:
)1 ()( (3)
and at an interior maximum
0) 1 (
This establishes a mathematically elementary but economically useful benchmark
case: here the optimal work effort e* is independent of the happiness shock, h. Thus
as the parameter h rises or falls, the marginal return to effort is unaffected. Happiness
therefore does nothing. It can be seen as orthogonal to choice. In passing, a variant
on this is the simple multiplicative form:
)] 1 (
)( )[ 1 (
where shocks to h again have no effect on optimal work-energy e*.
A concavity case
Another, and arguably more plausible, form of utility function has a happiness
shock operating within a concave structure. Imagine the worker solves
Maximize )1 (
which is the assumption that h is a shift variable inside the utility function itself,
rather than an additive part of that function.
Now the first-order condition is
0) 1 (
In this case, the optimal level of energy devoted to solving work problems, e*, does
depend on the level of the happiness shock, h:
The sign of de*/dh takes the sign of
) 1 (
Its first element is thus negative and its second is positive. By the first-order
condition, we can replace the piece rate wage term p by the ratio of the marginal
utilities from working and worrying.
Hence, after substitution, the sign of the comparative static response of work
effort, e, with respect to the size of the happiness shock, h, is greater than or equal to
These terms can be viewed as unconventional versions of the degrees of absolute risk
aversion in two domains -- the utility from work and the utility from worrying. If the
marginal utility of worry declines quickly enough as energy is transferred from
working to worrying, then a positive happiness shock will successfully raise the
worker’s chosen productivity, e*. Put intuitively, as the individual become happier,
that allows him or her to divert attention away from other issues in life.
A convex-combination case
A final approach is to think of happiness as tilting people’s preferences away
from distractions. For instance, assume that the worker solves
Maximize ) 1 (
which is the assumption that h acts as part of a convex combination outside the utility
function itself -- rather than within it or as an additive part of that function.
In such a circumstance, the first-order condition is
0) 1 (
) 1 ()(
It can be seen that the sign of de*/dh under such a setup takes the sign of the
which is automatically positive because it is the sum of expression ,)1 ()(
two marginal utilities. A positive happiness shock therefore lifts work effort e*.
These later approaches, in which effort is not independent of h, also
potentially offer economists a way to think about stress in the workplace. Work-life
strain could be conceived of as the (rational) need to devote energy and attention
away from the job. Happier workers need to do so less, and thus have higher
4. Experimental design
We now explain the structure of the experiment. We start with a motivation
for the choices made within the design, and then provide a description of the tasks and
a time-line for the trial. The experimental instructions, the GMAT MATH-style test
and the questionnaires are all set out in an appendix.
The experimental design was built around the desire to understand the
productivity of workers engaged in a task for pay. Our focus is the consequences, for
their output, of different starting levels of happiness.
We employ the task previously used in a number of existing papers (for
example, Niederle and Vesterlund, 2007), which entails asking subjects to add
sequences of five 2-digit numbers under timed conditions. This task is comparatively
simple but is taxing under pressure. It might be thought of as representing in a highly
stylized way an iconic white-collar job: both intellectual ability and effort are
Since we are trying to evaluate the relationship between happiness and
productivity, we wish ideally to disentangle the effort component and ability
component. To this end, we also included two control variables that we hoped would
capture underlying exogenous but heterogeneous ability as opposed to effort --
although we were also open to the possibility that changes in underlying happiness
might induce shifts in ability or change the nature of the interaction between ability
and effort to alter overall productivity. Our control variables came from (i) requiring
our subjects to do a brief GMAT MATH-style test (5 multiple choice questions)
along similar lines to that of Gneezy and Rustichini (2000) and (ii) obtaining
information in a final questionnaire to allow us to construct a measure of subjects’
prior exposure to mathematics. The aim was to allow us to control for heterogeneous
A key concern was to examine the consequences that happiness has for
productivity (be it through effort or ability). We therefore needed some means of
inducing an exogenous rise in happiness. The psychology literature offers evidence
that movie clips (through their joint operation as a form of audio and visual stimulus)
are a means of doing so. They exogenously alter people’s feelings and mood. For
example, Westermann et al (1996) provides a nice meta-analysis of the methods
We used a 10-minute clip based on composite sketches taken from various
4 We deliberately kept the number of GMAT MATH-style questions low. This was to try to remove any effort
component from the task so as to keep it a cleaner measure of raw ability: 5 questions in 5 minutes is a relatively
generous amount of time for an IQ-based test, and casual observation indicated that subjects did not have any
difficulty completing the GMAT MATH questions, often well within the 5-minute deadline.
comedy routines enacted by a well-known British comedian. In order to ensure that
the clip and subjects were well matched, we restricted our laboratory pool to subjects
of an English background who had likely been exposed to similar humor before. As
explained later, whether subjects enjoyed the clip turned out to be important to the
effects on productivity.
In summary, the data collected were on the successful and unsuccessful
numerical additions, a brief GMAT MATH-style test and a questionnaire that
included questions relating to happiness and intellectual ability.
Initially, we do not use a ‘placebo’ film. Hence, the control group start the
task straight away. We vary this later in the paper.
5. Design in detail
We randomly assigned people into two groups:
Treatment 0: the control group who were not exposed to a
Treatment 1: the treated group who were exposed to a comedy
The experiment was carried out on four days, with deliberate alteration of the
morning and afternoon slots, so as to avoid underlying time-of-day effects, as follows:
Day 1: session 1 (treatment 0 only), session 2 (treatment 1
Day 2: session 1 (treatment 0 only), session 2 (treatment 1
Day 3: session 1 (treatment 1 only), session 2 (treatment 0
Day 4: session 1 (treatment 1 only), session 2 (treatment 0
Subjects were only allowed to take part on a single day and in a single session.
On arrival in the lab, individuals were randomly allocated an ID, and made
immediately aware that the tasks at hand would be completed anonymously. They
were told to refrain from communication with each other. Those in treatment 1 (the
Happiness Treatment subjects) were asked to watch a 10 minute comedy clip
designed to raise happiness or ‘positive affect’.5 Those in the control group came
separately from the other group, and were not shown a clip nor asked to wait for 10
minutes. Isen et al (1987) finds that a control clip without positive affect gives the
same general outcomes as no clip.
The subjects in both the movie-clip group (treatment 1) and the not-exposed-
to-the-clip control group (treatment 0) were given identical basic instructions about
the experiment. These included a clear explanation that their final payment would be
a combination of a show-up fee (£5) and a performance-related fee to be determined
by the number of correct answers in the tasks ahead. At the recruitment stage it was
stated that they would make "… a guaranteed £5, and from £0 to a feasible maximum
of around £20 based purely on performance". Technically, subjects received £0.25
per correct answer on the arithmetic task and £0.50 on each correct GMAT MATH
answer, and this was rounded up to avoid the need to give them large numbers of
coins as payment.
An extra reason to pay subjects more for every correct answer was to
emphasize that they would be benefit from higher performance. We wished to avoid
the idea that they might be paying back effort -- as in a kind of reciprocity effect -- to
the investigators for their show-up fee.
The subjects’ first task was thus to answer correctly as many different
additions of five 2-digit numbers as possible. The time allowed for this, which was
explained beforehand, was 10 minutes. Each subject had a randomly designed
sequence of these arithmetical questions. The numerical additions were undertaken
directly through a protected Excel spreadsheet, with a typical example as in Legend 1.
The spreadsheet necessarily contained more such rows that any subject could hope to
add in the ten minutes allowed. The subjects were not allowed to use calculators, and
it was explained that any attempt to use a calculator or any outside assistance was
deemed to be a disqualification offence, resulting in only a show-up fee being paid.
This did not prove to be a problem across the 4 experimental days. The numerical
additions were designed to be reasonably simple, if dull and repetitive, and earlier
literature has deemed this a good measure of intellectual effort (Niederle and
5 The questionnaire clearly indicated that the clip was generally found to be amusing and had a direct impact on
reported happiness levels. More on this is in the results section.
31 56 14 44 87
Legend 1: Adding 2-digit Numbers
The second task for subjects was to complete a simple 5-question GMAT
MATH-style test. These questions were provided on paper, and the answers were
inputted into a prepared protected Excel spreadsheet. The exact questions are given
in an appendix. This test was designed as a brief check on ability, as used before in
the research literature (Gneezy and Rustichini, 2000).
The final task, which was not subject to a performance-related payment (and
subjects were made aware of this), was to complete a questionnaire. A copy of this is
provided in an appendix. The questionnaire inquired into both the happiness level of
subjects (before and after the clip for treatment 1), and their level of mathematical
expertise. The wording was designed to be straightforward to answer; anonymity was
once again stressed before it was undertaken; the scale used was a conventional 7-
point metric, following the well-being literature.
To summarize the timeline:6
Subjects enter and are given basic instructions on experimental
Subjects in treatment 1 are exposed to a comedy clip for 10
minutes, otherwise not.
Subjects are given additional instructions, including a statement
that their final payment relates to the number of correct answers, and
instructed against the use of calculators or similar.
Subjects move to their networked consoles and undertake the
numerical additions for 10 minutes.
Results are saved and a new task is initiated, with subjects
undertaking the GMAT MATH-style test for 5 minutes.
Results are again saved, and subjects then complete the final
After the questionnaire has been completed, subjects receive
6 The full instructions provided in the appendix provide a description of the timing.
payment as calculated by the central computer.
6. Principal results
A group of 182 subjects drawn from the University of Warwick participated in
the experiment. Each took part in only one session. A breakdown of the numbers per
day and session is given in Table 1. The subject pool was made up of 100 males and
73 females. Table 2 summarizes the means and standard deviations of the main
variables. The first variable, the key one in our analysis, is the number of correct
additions in the allotted ten minutes. ‘Happiness before’ is the self reported level of
happiness (for the treated group before the clip) on a seven point scale. The variable
‘happiness after’ is the level of happiness after the clip for the treated group; GMAT
MATH is the number of correct problem solved on that; high-school-grades is an
index calculated from the questionnaire. Enjoyment-of-clip is a measure in a range
between 1 and 7 of level of how much they said they liked the movie clip.
According to the data, the clip is successful in increasing the happiness levels
of subjects. As shown in Figure 1, they report an average rise of almost one point
(0.98) on the scale of 1 to 7. Moreover, comparing the ex-post happiness of the
treated subjects with that of the non-treated subjects, we observe that the average of
the former is higher by 0.85 points. Using a two-sided t-test, this difference is
statistically significant (p <0.01). Finally, it is useful to notice that the level of
happiness before the clip for the treated group is not statistically significantly
different (the difference is just 0.13) from the happiness of the untreated group (p =
0.20 on the difference).
In Figure 2 we display the average productivity in the test. The treated group’s
mean performance is higher by 1.71 additions than the average performance of the
untreated group. This productivity difference is approximately ten percent. It is
statistically significantly different from zero (p=0.04).
Interestingly, and encouragingly, the performance of those 16 subjects in the
treated group who did not report an increase in happiness is statistically non-different
from the performance of the untreated group (p=0.67). Therefore, the increase in the
performance may be linked to the increase in happiness rather than merely to the fact
of watching a clip. We return to this.
The clip did not hamper the performance of subjects who did not declare
13 Download full-text
themselves happier.7 For them, the effect is zero.
In Figure 3 we show the performances of male and female subjects. Both
groups feature a similar increase in their arithmetical productivity (1.9 additions for
male, 1.78 for female).
From the cumulative distributions on the number of correct answers for the
treated and untreated groups, shown in Figure 4, we see that the treatment increases
the performances of low and medium performers, while the high performers are
apparently less affected.
We also performed OLS-based regressions to analyze the determinants of the
performances. Table 3 presents the determinants of the number of correct additions;
variable Change-in-Happiness is the difference in happiness before and after the clip;
GMAT MATH is a test score. High school grades measure school performance. Day
2, Day 3 and Day 4 are day-of-the-week dummies.
Consistent with the result seen in the previous session, the subjects’
performances are higher in the session with treatment. As we can see in regression
(1), in the first column, this result holds when we control for subjects’ characteristics
and periods. In regression (2) of Table 3, the performances are increasing in the rise
in elicited happiness (for the case of untreated subjects, by definition, Change-in-
Happiness=0). This result is still true when we restrict the analysis to the treated
subjects as in in regression (3). The size of the effect is only slightly smaller at
approximately eight and a half percentage points.
Because of the well-known skewness in human performance data, it is natural
to use a logged dependent variable. Nevertheless, as a check, Table 4 re-runs the first
two regressions of Table 3 with a dependent variable defined on absolute values
rather than log values. The variable ‘Treatment’ remains large and positive. It is
statistically significant when, as in regression 2 of Table 4, we exclude the outliers
(here we drop the two extreme laboratory subjects, with respectively 2 and 43 correct
additions). The coefficient on the variable Change-in-Happiness is statistically
significantly different from zero irrespective of whether or not we keep in the two
outliers: see regressions 3 and 4.
Could the pattern in the data be a kind of reciprocity effect? Are these
laboratory subjects ‘repaying’, or somehow trying to please, the investigators? Such
7 Also, the 17 subjects who did not declare an increase in happiness enjoyed the clip. In a range of values between 1 and 7, the
average is 5.41, with a minimum of 5 and a maximum of 7.