Does Employee Happiness Have an Impact on Productivity?
Abstract
This paper provides evidence from a natural experiment on the relationship between
positive affect and productivity. We link highly detailed administrative data on the behaviors
and performance of all telesales workers at a large telecommunications company with survey
reports of employee happiness that we collected on a weekly basis. We use variation in
worker mood arising from visual exposure to weather—the interaction between call center
architecture and outdoor weather conditions—in order to provide a quasi-experimental test
of the effect of happiness on productivity. We find evidence of a positive impact on sales
performance, which is driven by changes in labor productivity – largely through workers
converting more calls into sales, and to a lesser extent by making more calls per hour and
adhering more closely to their schedule. We find no evidence in our setting of effects on
measures of high-frequency labor supply such as attendance and break-taking.
Introduction
A large number of employers are increasingly claiming to care about how their employees feel at
work, and have begun to invest in management and organizational practices aimed at creating
and maintaining a happier workforce. There may be various reasons for this – such as an
increased ability to attract and retain high quality workers – but at least one motivation is
a belief that happier workers will be more productive. When surveyed, for example, around
79% of U.S. managers reported an expectation that unhappiness in their workforce is likely
to hurt productivity.1 While this recent focus on employee happiness may seem a relatively
new development, the relationship between happiness and productivity has in fact been the
keen focus of both researchers and practitioners for many decades (see, e.g., Fisher and Hanna,
1931; Hersey, 1932, for early examples). One reason for this long-running interest is that the
relationship has potentially key implications for how firms manage their employees and organize
work (Edmans, 2012; Wright and Cropanzano, 2000). However, isolating the causal effect of
happiness on productivity has remained an empirical challenge, particularly in field settings.
Indeed, despite multiple generations of research on the topic, the literature has been bedeviled
by inconsistent findings (Iaffaldano and Muchinsky, 1985; Judge et al., 2001; Tenney, Poole and
Diener, 2016), such that claims surrounding the organizational benefits of worker happiness
have frequently been met with skepticism (Wright and Cropanzano, 2000).
In this paper, we study the effects of employee positive affect on productivity. We use data
on the universe of telesales workers at British Telecom (BT), allowing us to observe objective,
granular information about the behaviors and performance of 1,793 workers at one of the United
Kingdom’s largest private employers. We link this administrative data to a survey instrument
designed to measure the week-to-week affect of employees using a well-established measure of
happiness. We use variation in exposure to visual weather conditions while at work, arising out
of the interaction between weather and architecture, in order to provide a quasi-experimental
test of the effect of happiness on sales performance.
We find that a one unit increase in happiness, on a standard 0-to-10 scale, leads to around 3
additional weekly sales, around a 12% increase over a base of 25. Using detailed data on worker
behaviors, we rule out any effect going through employees working more (e.g. by taking fewer
breaks or working overtime). Instead, we are able to study three potential productivity channels
behind the effect of happiness on sales performance. First, workers could be better at organizing
their time while happier at work and in doing so adhere more closely to their prescribed daily
workflow schedule. Second, they could work faster when happier – that is, answer a higher
number of calls per hour. Third, they could be more efficient at converting calls into sales when
in a better mood. Although we find evidence for all three potential channels, the magnitude
of the third channel is much stronger, such that the estimated effect on sales can be almost
entirely explained by workers converting more of their calls into sales during weeks when they
feel happier. We interpret this as suggesting that, at least in our context, much of the effect
of happiness on productivity can be explained by better moods augmenting the ability to solve
ore complex cognitive tasks, where social and emotional skills may also matter more. In line
with this, when breaking down the analysis by sales type, we find strong effects of happiness
on sales when the worker is selling bundles of products and, in particular, when re-contracting
– but negligible effects when doing more routine order taking.
Our identification strategy relies on four key features of the empirical setting. First, the
11 call centers are dispersed geographically across the whole of the United Kingdom, such
that there is significant variation, within week, in weather in the local vicinity of each of the
call centers. Second, workers take incoming calls that are allocated to them based on the
call-type and handler availability, but not based in any way on the location of the caller –
meaning that, in this two-sided market, workers based in a given location will take calls from
customers all over the country. Third, the call centers vary significantly in terms of their
architecture and, in particular, their window coverage – ranging from fully glass-covered tower
buildings all the way to warehouse-style buildings with almost no windows at all. Finally,
despite these major exterior differences, all are laid out internally in the same open-plan way,
such that any given worker within a call center has roughly equal visual exposure to the external
walls – be they fully windowed, solid, or anywhere in between. We hypothesize that only the
most psychologically-relevant features of weather—namely, the extent to which it is bright or
gloomy—will impact worker happiness. Such aspects are visual in nature, meaning that any
effect should be contingent on a worker’s observable exposure to them. We show that visibly
gloomy weather has a strong negative impact on our measure of worker mood, on average,
but that this effect is dependent on the window coverage of each call center – in other words,
workers’ exposure to it. We exploit this plausibly exogenous variation in happiness to provide
a quasi-experimental test of the effect of happiness on sales performance in a real-world field
setting.
We make three main contributions. First, we provide causal field evidence on the relationship
between happiness and productivity. In doing so, we build on an extensive literature that spans
many decades as well as several disciplines across the social and behavioral sciences. Despite
a long line of research, the literature has been plagued with inconsistencies (Tenney, Poole
and Diener, 2016). Although experimentally-induced positive affect has been shown to improve
performance on stylized productivity tasks in the lab (see, e.g., Erez and Isen, 2002; Oswald,
Proto and Sgroi, 2015), it remains unclear whether such effects translate to real-world, largescale organizational settings – which are typically very different along a number of different
dimensions.
Second, we make a methodological contribution in terms of how we identify mood effects in
the field. While affective states are increasingly seen as a potentially important factor in driving
economic behavior (Loewenstein, 2000), demonstrating this in natural settings has proven difficult. We join a growing literature that typically proceeds by estimating the effects of various
proxies for mood such as weather patterns on outcomes like stock returns, consumption, real
estate transactions, and voting (e.g. Agarwal et al., 2020; Edmans, Garcia and Norli, 2007; Hirshleifer and Shumway, 2003; Hu and Lee, 2020; Li et al., 2017; Meier, Schmid and Stutzer, 2019;
Saunders, 1993). We label this existing method the reduced-form approach to mood effects in
the field, and build on it in three ways. First, by studying participants in a two-sided market
who are located in different places, we are able to identify weather-induced worker mood effects
aside from overall shocks that may, for example, affect national demand or customer mood. Second, implicit to the reduced-form approach is an instrumental variables (IV) set up, whereby
weather i) has an impact on mood and ii) affects behavior solely through that mood mechanism.
However, whereas this is often assumed rather than fully estimated, we follow Guven (2012) in
collecting a measure of happiness and estimating both stages of the IV specification (see also
Coviello et al., 2020; Guven and Hoxha, 2015). This is particularly important since the prior
literature on the effects of weather on happiness suggests the relationship is not necessarily a
given, threatening the implied first stage of the IV equation (Feddersen, Metcalfe and Wooden,
2016; Frijters, Lalji and Pakrashi, 2020). Third, an ever-growing number of economic phenomena have been instrumented for using weather patterns, casting doubt on the validity of the
exclusion restriction (Gallen, 2020; Mellon, 2020). Our novel use of variation in architecture
coupled with a focus on only the visual and mood-relevant aspects of weather allows us to overcome many of the usual objections related to this assumption. We rely on detailed institutional
knowledge of our workplace setting, gleaned from a mixture of image coding, employee surveys,
semi-structured interviews with managers, and site visits. This mixed methods approach means
we can design an identification strategy in which differential visual access to outside weather
patterns allows us to “turn on and off” the treatment – and in doing so better isolate any causal
effect.
Given that firms are increasingly at least claiming to focus on the happiness of their employees, the third main contribution of our paper is to provide more fine-grained and potentially
useful evidence for managers on the source of the happiness-productivity relationship. A recent
survey of U.S. executives suggests, for example, that while a large proportion of U.S. firms say
they are considering investing in managerial practices designed to foster a happier workforce,
only a small number currently have any sort of strategy in place to move in this direction (HBR
Analytical Services, 2020).2 Since work is one of the most unhappy activities people do in their
day-to-day lives in countries like the USA and UK (Bryson and MacKerron, 2016; Krueger
et al., 2009), there is ample room for improvement that could unlock potential productivity
gains. In one sense, our empirical approach suggests the significance of a managerially important but often overlooked aspect of work life – space and the physical workplace environment.
But, even more importantly, a long line of research has found that a wide range of management
and organizational practices influence the happiness of workers – suggesting a large number of
potential levers for firms to pull.3 By probing mechanisms through which happiness may impact
productivity, we are able to discuss the types of tasks and jobs where happiness is most likely to
be an important factor in explaining productivity differentials. In particular, with the number
of jobs requiring workers to interact socially with customers increasing rapidly (Deming, 2017),
87% of executives agreed that workplace happiness can provide their firm with a competitive advantage, but
only a third of the organizations in the survey of executives noted above say their organization sees employee
wellbeing as a strategic priority. Not only this, fewer than 20% of these firms actually have any sort of strategy
in place to measure or improve the wellbeing of their workforce.
While evidence on the effectiveness of employee “wellness programs” is mixed (Gubler, Larkin and Pierce,
2018; Jones, Molitor and Reif, 2019), a growing literature demonstrates the more fundamental point that the
ways in which work is managed and organized by firms—as well as the cultures they create—has a significant
impact on employee wellbeing (Bloom et al., 2014; Clark, 2010; Gosnell, List and Metcalfe, 2020; Krekel, Ward
and De Neve, 2019; Moen et al., 2016).
Background & Theory
For over a century, the relationship between happiness and productivity has been the keen
focus of both researchers and practitioners. One reason for this long-running interest is that
the relationship has potentially key implications for how firms manage their workers and for the
place, more broadly, of human resource management in firms’ overall business strategy (Edmans,
2012; Wright and Cropanzano, 2000). Yet, despite a long line of research, the literature is
bedeviled by inconsistent findings – many of which may be traceable to i) the multiplicity of
ways that researchers approach—theoretically and empirically—the concepts of both happiness
and performance, as well as ii) a number of inherent empirical difficulties in estimating any
causal effect, particularly in natural field settings.
2.1
Happiness
Although the study of human happiness has a long history (see Diener et al., 1999), confusion
can arise insofar as the term “happiness” is sometimes used loosely as a catch-all term referring
to subjective wellbeing (SWB) – which has both affective and cognitive dimensions (Krueger
et al., 2009). Cognitive measures of SWB are evaluative in nature, refer to global judgements
people make about how things are going overall, and are typically assessed in the workplace
context using survey questions on job satisfaction. Affective SWB, on the other hand, refers to
people’s emotional or hedonic experience. Kahneman, Wakker and Sarin (1997) refer to this as
‘experienced utility,’ in the Benthamite tradition, and note that it can be measured either in
real time or via people’s recollections – for example, by asking how happy they feel or have felt
during a given day, week, or month.
One potential reason for the inconsistent nature of the literature on happiness and productivity is that researchers have approached the question often with differing notions of what
“happiness” means (Wright and Cropanzano, 2000). Whereas much of the early literature focused on measures of job satisfaction (e.g., Brayfield and Crockett, 1951; Fisher and Hanna,
1931; Lawler and Porter, 1967; Locke, 1969), a more recent body of work has generally turned
toward the study of affect in the workplace (see, e.g., Barsade and Gibson, 2007; Brief and
Weiss, 2002; Knight, Menges and Bruch, 2018). This distinction is particularly important as
the theoretical links between happiness and performance are, as we will discuss below, arguably
much stronger when thinking in terms of affect than satisfaction (Côté, 1999; Lucas and Diener,
2003).
In this paper, we focus on affective wellbeing – specifically, workers’ feelings of happiness as
they experience it week-to-week. We see this as a general measure of positive affect that can also
reasonably be referred to as “mood.” Within the broad category of affect or affective wellbeing,
there is a key distinction between moods and emotions (see, e.g. Frijda, 1986). Emotions
typically refer to a specific feeling that is a (relatively short-lived) reaction to a particular
and usually known) stimulus. Moods, on the other hand, are less specific and are typically
less intense. They are not directed at a particular person, task, or situation, but are rather a
more diffuse general feeling. While it is often easy for people to trace the root of a particular
emotion, they are usually not aware of the source of a good or bad mood (Russell and Barrett,
1999). Given this, there is little reason to expect the effects of a weather-induced good or bad
mood to be any different from the effects of a mood state induced by other factors, ranging
from working arrangements to line manager behavior. This is important, since although we use
weather-induced mood shocks for identification, we are nevertheless able to say something more
broadly about managerial and policy implications.
In addition to being influenced by a range of management and organizational factors (Gosnell, List and Metcalfe, 2020; Krekel, Ward and De Neve, 2019), SWB is also influenced by
the weather, one of the most pervading background variables in human life. However, the literature suggests that the link is not straightforward (Feddersen, Metcalfe and Wooden, 2016;
Frijters, Lalji and Pakrashi, 2020), since the empirical relationship between weather and affect
can be unstable (Denissen et al., 2008). A key explanation for the instability of these findings
is that effects are contingent on the extent to which weather is visible. Keller et al. (2005)
find weather affects mood, for example, but only when people are experimentally assigned to
be outdoors. A related literature shows that the visual pleasantness of weather improves mood
and prosocial behavior in studies with outdoor settings (Cunningham, 1979), though this does
not replicate using time-series data on tipping in an indoor restaurant (Flynn and Greenberg,
2012).4 Overall, one of the key findings of this literature is that the visibility of weather is key
to any relationship with subsequent mood and behavior.5 Given this contingent relationship,
in order to convincingly use weather as a mood proxy or instrument, it is (a) useful to have
variation in weather that is visual in nature and (b) even more useful to have variation in visual
exposure to any given weather, in order to eliminate concerns related to the direct effect of
weather on non-mood related drivers of productivity.
2.2
Performance
In addition to confusion surrounding conceptual definitions and measurement of happiness, a
further potential explanation for the inconsistent state of the literature is that many different
definitions of the performance outcome have been studied. Here we study labor productivity
– that is, the residual variation in output that cannot be fully explained by observable inputs
(Syverson, 2011). Call centers provide a particularly good setting in which to study this, since
we not only observe detailed data on a large number of labor inputs, but workers in our setting
also largely do the same telesales job, which is to take incoming calls from new and existing
customers and sell them various products using the same phones and computer system.
Rind (1996) also studies an indoor setting – a casino hotel in Atlantic City – in which hotel rooms all have
dark, limo-tinted windows that make it look like it is cloudy outside regardless of the brightness of the weather.
Having a server inform customers of the weather outside, experimentally varying how bright it is reported to be,
the authors find that (a belief in) sunny skies increases tipping. That is, though this experiment does not vary
visual exposure to weather, it does vary a related concept, namely its salience.
This is in line with a large medical literature on seasonal affective disorder (SAD), which shows that experimental exposure to sunlight improves mood (e.g. Kripke, 1998), even among the non-depressed (e.g. Leppämäki,
Partonen and Lönnqvist, 2002).
A large number of existing studies rely on employee self-reports of productivity or subjective
managerial evaluations (e.g. Staw and Barsade, 1993; Zelenski, Murphy and Jenkins, 2008). In
the case of using self-reported performance, there are well-known empirical difficulties associated
with regressing one subjective report on another, and in the case of managerial reports there
is a strong possibility that performance will be subject to a ‘halo’ effect whereby the happier
employee is rated more highly by the managerial rater precisely because they are happier and
more agreeable, and not because of any real performance differences. We are able to use
administrative data on a clear and objectively measurable output that is unambiguously positive
for the firm: sales performance. Moreover, fine-grained objective data on worker behaviors also
allows us to investigate channels through which any effect of happiness may translate into sales.
Given the various ways in which workers’ mood may affect performance, having a clear
understanding of which performance indicator is being used in any specific context is especially
important. For instance, a task involving complex interactions with coworkers or customers
may involve different skills than a more autonomous and repetitive task. The most recent
field-experimental literature shows that management practices can have simultaneously positive
impacts on i) productivity as well as ii) employee happiness and satisfaction (Gosnell, List and
Metcalfe, 2020). From pay inequality (Breza, Kaur and Shamdasani, 2017; Cullen and PerezTruglia, 2019) to gift exchange (DellaVigna et al., 2020) and work autonomy (Bloom et al.,
2014), this line of research suggests employee wellbeing as one possible channel through which
workplace organization may feed through to productivity. However, it is unable to isolate the
happiness channel in a causal manner. As a result, little discussion has been paid to the types
of tasks or psychological channels through which improvements in worker wellbeing itself may
mediate the positive effects of managerial practices on productivity.
2.3
Happiness and Performance
Whereas much of the earlier work on happiness and workplace performance is based on crosssectional comparisons, more recent studies have been able to leverage within-worker variation
in longitudinal research designs that go some way to assuage concerns related to unobserved
heterogeneity between workers, and in doing so have demonstrated that prior affective states
predict subsequent performance (see, e.g. Koys, 2001; Miner and Glomb, 2010; Rothbard and
Wilk, 2011; Staw and Barsade, 1993; Staw, Sutton and Pelled, 1994).6 But although this
temporal ordering is consistent with a causal effect, it still may be the case that time varying
third factors could be driving both. To infer causality, Oswald, Proto and Sgroi (2015) rely on
a series of mood-inducement experiments in the lab,7 using as their outcome an incentivized,
math-based productivity task designed to be similar to something one might find in a real-world
work setting – and, ultimately, find a positive effect of happiness (see also Erez and Isen, 2002,
for similar findings using a non-incentivized task).
Yet, while this evidence using a stylized, piece-rate productivity task moves the literature
forward, the extent to which these affect-induced productivity effects translate into real-world
employment settings remains an open question. Indeed, real-world jobs typically involve bundles
of tasks, decisions on how to focus time and energy between them, as well as a number of other
factors that make it difficult to generalize from a math task in the lab to a real job in the
field. The external validity of existing lab-evidence can therefore be built upon, particularly
for tasks involving social interactions with customers or coworkers (which are becoming ever
more common in modern economies (Deming, 2017)). In such cases, the positive effect of mood
on performance found in a constrained lab-environment may either be reduced (e.g. if happier
workers spend less time at work and more time socializing with co-workers) or magnified (e.g.
if happier workers are better able to deal with customers). It is therefore important to be
more specific about the exact mechanisms through which positive affect may lead to greater
performance (for extensive discussion of the theoretical mechanisms between happiness and
performance, see, Lucas and Diener, 2003; Tenney, Poole and Diener, 2016).
First, good mood may affect performance by augmenting cognitive skills, independently of
the social interactions in which workers might be involved. In particular, positive affect can
influence how we think and process information. The influential broaden-and-build framework
suggests that positive affect signals to people experiencing it that the environment is nonthreatening and that things are generally going well. As a result, positive mood states tend
to broaden people’s thought-action repertoires as well as allow them to build longer-lasting resources (Fredrickson, 2001). In line with this, laboratory evidence suggests that people induced
into positive mood states tend to think in ways that are more flexible (Isen and Daubman,
1984), creative (Isen, Daubman and Nowicki, 1987), integrative (Isen, Rosenzweig and Young,
1991), open to information (Estrada, Isen and Young, 1997), and efficient (Isen and Means,
1983). Relatedly, it has been shown that the thoughts of happier people are less likely to “wander” (Killingsworth and Gilbert, 2010), a mechanism that has been formalized into an economic
model in which happiness reduces the amount of time spent worrying about negative aspects of
people’s lives, and thus drives productivity (Oswald, Proto and Sgroi, 2015).
Second, a further potential channel for the happiness-performance link is that happier workers may be more motivated (Erez and Isen, 2002). People in a positive mood state may have
greater prior expectations about the task, for example, and since people experiencing positive
affect are more likely to expect to enjoy an upcoming task, they are likely to be more motivated
to initiate or engage with it. Moreover, happier people might attribute their positive mood
state to the task at hand, and in doing so make them feel that they are enjoying the task – and,
ultimately, make them want to complete it (Forgas, 1995).
Third, positive affect may also influence behavior and outcomes by improving people’s social
and emotional skills. Outward indicators of happiness such as laughter signal that a person is
friendly and open. People induced into positive mood states (in the lab) are more likely to
engage in social contact with others (Isen, 1970) and people in happier moods tend to be more
cooperative and less aggressive with others (Isen and Baron, 1991). The positive impact of
affect may hence be even stronger when the task involves interacting directly with customers or
coworkers. For instance, Carnevale and Isen (1986) show that positive affect improves bilateral
negotiation skills in bargaining tasks, with participants employing less contentious tactics and also finding more integrative solutions. Related to this, the sociological literature on emotional
labor (see, e.g., Hochschild, 1983) has long argued that in tasks involving interactions with
customers, happier workers may be better at negotiating with angry customers, as it becomes
easier to manage their emotions.
Given the differences in these psychological mechanisms, the extent to which affective happiness is likely to lead to greater productivity in real-world jobs will depend on a number of
factors, including the type of tasks and the nature of the workplace in question. For example,
being more integrative in thought may be helpful to productivity in jobs where novel or more
complex solutions to specific problems are performance-boosting, but less so where performance
is best accomplished by following precise instructions or routines. However, to the extent that
positive affect boosts motivation, task persistence, and energy, we may still expect it to lead
to greater productivity regardless of the job being performed. Equally, since happiness leads to
greater social rewards it may be a particularly strong predictor of productivity in jobs where
employees frequently interact with others, where social contact is beneficial, and where cooperation or negotiation with others is required in order to be successful. But in jobs where
social contact is less integral to the successful carrying out of the job, or where there is little
supervision or structure, happiness may be less important – or may even reduce productivity
by providing a distraction.
Institutional Setting & Data
Administrative Firm Data. We use detailed individual-level administrative data from BT,
a large multinational telecommunications company based in the United Kingdom. We focus our
attention on sales workers, whose job it is to take incoming calls and sell BT products, at 11
call centers across the United Kingdom (see Figure 1 for a map). The vast majority of the work
(91% of time and 82% of tasks) carried out by the employees in the sample are incoming calls
from potential or existing customers, with the remainder consisting of outgoing calls (4% of
time and 12% of tasks) and “other” activities (which includes tasks such as dealing with letters,
online customer chats, and SMS messaging). Workers are paid a fixed hourly wage, with a
potential bonus if they meet their target.8 At the worker-day level we observe the number of
sales, the distribution of which can be seen in panel (a) of Figure 2. As is typically the case
with sales data, the distribution is both right-skewed and also contains some zero values. In
addition, we also observe a host of information about worker behaviors such as break-taking,
attendance, number of calls, average length of calls, and so on.
Affective Wellbeing Survey. We link this performance data at the worker-week level with
a survey we administered to capture positive affect. We use a succinct happiness question that
was designed following the OECD’s (2013) guidelines on the measurement of SWB. Employees
were asked “Overall, how happy did you feel this week? ” over a six-month period. This is an
affective question, and, following Kunin (1955) and decades of subsequent work in psychology,
This is neither an explicit piece-rate pay schedule nor is it a commission-based pay system. In each of these
instances, one might expect each individual sale to bring with it a psychological reaction. Rather, the pay system
is a much more slow-moving bonus scheme, in which the majority of pay is paid through a base salary.
we offer five response categories as a Faces Scale that ranges from very sad to very happy
androgynous faces. The use of faces in this way is both intuitive to respondents, and is also
known to strongly pick up the affective component of wellbeing questions (Fisher, 2000). The
survey was sent by email every week on Thursday afternoon, and is shown in Figure 3. The
single-item survey question could be quickly answered within the email. Workers were assured
that their individual happiness responses were being collected externally for the sole purpose of
academic research, and would not be shared with management. Workers were also offered the
opportunity to opt-out of the study at any time, via a simple email click-through. The study
ran for 6 months – the first and final emails were sent on the 20th of July 2017 and the 18th of
January 2018, respectively.
Benjamin et al. (2021) note the importance of being explicit about what happiness researchers are measuring. It is not only important in terms of wellbeing notion – job satisfaction
versus affect in the workplace, for example – but also the time horizon.9 We are purposefully
specific in the question about the time frame, measuring affect during a week-long period and
thus allowing us to match up with contemporaneous weekly data on productivity. One concern
is that respondents may answer solely based on their current mood at the end of the week (or
the peak during it), given that recall of emotions can be biased (Thomas and Diener, 1990).
While affective recall is not always fully accurate, Kemp, Burt and Furneaux (2008) nevertheless show that subjects do respond to the time prompts in such questions: By surveying daily
Figure 3: Happiness Survey Email
Notes: Screenshot of the mood survey, which was sent weekly over a six month period to all workers. Respondents
had to click a face within the email for their response to be registered. See text for more details.
mood over a week-long period and then subsequently asking about total weekly mood, they are
able to show that people are able to provide a good reconstruction of the affective states (for
further discussion of the validity and reliability of SWB measures, see Krueger and Stone, 2014;
Krueger and Schkade, 2008).10
The distribution of responses is shown in panel (b) of Figure 2. Happiness in our setting
is low, with the modal answer being the most unhappy. We use our happiness responses both
ordinally and cardinally, depending on specification. When using happiness as a continuous
measure—as is typically done in the SWB literature—we assign the five categorical happiness
states equally-spaced numerical values between 0 and 10 from least to most happy. This is
done in order to be more aligned in terms of effect size magnitudes with scales typically used
in similar survey measures, such as in household panel surveys widely used in the literature.
When doing so, the mean response is around 4, with a within-person standard deviation of 2.4
(see Table A1).
Sample construction and characteristics. We aggregate all of the administrative data to
the Monday-to-Friday working week. All 1,793 workers were invited to take part in the study
10
In Figure S7 we are able to provide an empirical test of the temporal nature of our happiness measure.
Using daily weather data (described in more detail below), we show that the mood measure taken on a Thursday
afternoon or Friday is meaningfully related to weather exposure not only on that same day, but a few days before
within that same week. But that the relationship strongly declines in the days after the mood report – lending
intuitive support to the weekly nature of the measure.
10
Electronic copy available at: https://ssrn.com/abstract=3470734
Figure 4: Distribution of Visual Weather Index
(a) Raw Data
(b) Within-Week Variation
Note: Visual weather index is a weekly measure that counts the total number of daily instances of fog, rain,
and snow in the vicinity of each call center. Panel (a) shows the raw distribution, while panel (b) shows the
distribution of residuals from a regression of the index on week fixed effects (bin width=0.5).
and were sent weekly mood surveys. Of these employees, 1,438 (around 80%) participated by
answering at least one survey over the subsequent 6 months. Conditional on participating in
the study, workers responded to a mean of 10.3 waves (SD = 7.1). The weekly response rate
of workers who participated was on average around 37%, which rises to 50% if we focus only
on workers who work on Thursday or Friday. We drop any participants who responded to only
one survey wave, since we rely on within-worker variation over time. This leaves us with a final
sample of 1,157 employees. Summary statistics for this final sample are shown in Table A1.
4.1
Reduced Form Effect of Visual Exposure to Weather on Performance
We begin by following the existing literature on mood-in-the-field effects by estimating the
impact of weather on sales. That is, we first assume weather to be a suitable proxy for P Aijt
For an empirical example of a similar approach using panel data, see Acemoglu and Pischke (2003), who
cluster standard errors at the individual and at the region-time level. A week is defined in our case by the date
on which a worker answered our survey and location by the call center.
in the above equation. Given that sales is a right-skewed count variable including some zero
values, we use a Poisson quasi-maximum likelihood estimator. Using local weather station data
that we matched to the locations of the 11 call centers, we coded an index of how visually
gloomy the weather is. This “Visual Weather Index” corresponds to the total number of daily
incidences of fog, rain, and snow during the working week (see Appendix H for more details and
Figure 4 for the distribution).12
In column (1) of Table 1 Panel A, we show that the Visual Weather Index has a negative,
though imprecisely estimated, impact on sales. Going beyond this, we further isolate the mood
effect by using variation not only in weather but also in people’s visual exposure to it, making
use of two additional factors critical to our identification strategy. In our setting, while all
workstations are open-plan across buildings, the type of building varies significantly – all the
way from warehouse-style workplaces with almost no exposure to outdoor conditions to glass
tower buildings or large window buildings with full exposure to the outdoor weather. Coding the
proportion of external walls that is covered by glass windows using image processing software,
we confirm that there is large variation across call centers – with the window share ranging from
.03 to .59 (see Table A1).13 Figure 5 illustrates this variation in architecture with photos taken
from the outside and inside of Doncaster and Swansea call centers, two ends of the external
window coverage spectrum.
In column (2) of Panel A in Table 1, we interact the weather index with visual exposure
to it. We find that the negative effect on sales is much stronger in situations where weather
is visible. The effect of window share itself is not estimated since it is captured by the worker
fixed effects, but recall that it lies between 0 and 1. The coefficient on the main effect of the
weather index thus gives the effect of weather for buildings that have no windows. Here we find
no significant impact. The interaction with window share, on the other hand, suggests that the
effect is much larger in magnitude (i.e. more negative) for buildings that are fully windowed.
To further illustrate what these results mean in terms of magnitude, we re-estimate our
regression equation within equally-sized groups of call centers with below and above median
window share. As can be seen in Figure 6a, the point estimate on sales is close to zero in call
centers with very few windows, but is significantly negative in centers with many windows.
This in theory has a range of 0 to 15, where 0 would mean a likely bright day with no rain, snow, or fog on any
day of the week at all and 15 would mean that all three happened on every single day of the week. Importantly
for our identification, the United Kingdom has sufficiently volatile weather so that there is significant variation
in weather conditions across call centers, even within the same week (Panel (b) of Figure 4).
For an objective measure of external window-coverage, we first collect all wall photos from each of the call
centers in the dataset using Google Street View. For each building, we then code the percentage of wall surface
that is covered by glass windows using the ImageJ software (see Appendix I for more details). We supplement
this data below with a worker-level survey asking about the number of windows and subjective experience of
natural light.
Here in the above median share group, every 1-point increase on the weather scale (which is coded such that
higher numbers mean gloomier weather and has a standard deviation of 1.36) lowers weekly sales by 1.5%. We
also look for potential asymmetries, testing whether sales are more or less sensitive to visual exposure to bright
or gloomy weather. Though both effects go in the expected opposite direction, we find no systematic evidence of
an asymmetrical effect in terms of magnitude (Figure S5).
First Stage Impact of Exposure to Weather on Happiness
The finding of a significant interaction in the reduced form-equation between weather and visual
exposure to it lends credence to the idea that we are picking up a mood effect. However, without
observing mood itself, alternative explanations may still be possible. In column (1) of Panel B in
Table 1, we show there is, on average, a negative effect of gloominess on happiness. Column (2)
suggests that this effect is much stronger in buildings with high glass coverage and non-existent
in buildings without windows. Another way to see the contingent relationship of weather on
happiness is shown in Figure 6b, which shows that the negative effect of gloomy weather on
happiness is concentrated within the buildings with high window coverage. This means at least
two things. First, by using exposure to visual weather, we are effectively relying on variation
in positive affect rather than any physical effect of weather. Second, we provide evidence that
the source of the mood shock is occurring while at work.15
Our identification comes here not from either windows or weather, but rather from the
interaction between the two. Note that the main effect of windows is not estimated, since it
is subsumed into the worker fixed effects. Indeed, we are not here principally interested in the
effect of windows, particularly since we do not have random variation in the architecture of these
buildings – which may, for example, be correlated with local economic conditions. Nevertheless,
we show that the presence of windows is—on average—positively correlated to workers’ mood
(Figure S1). More important for our identification is the within-worker variation in happiness
across the buildings. Consistent with the intuition behind our empirical strategy, the mood of
workers (and their performance) in more heavily glass buildings is more variable than those who
work in call centers that block visual access to outside weather (see Figure S1).16
4.3
Putting it Together: IV Estimates
Control Function Poisson-IV Estimates. Column (3) of Table 1 shows the elements of
our main IV approach. For our main IV specifications, we prefer to use one single instrument –
since there are well-known issues with strength when using multiple instruments. We will thus
rely on the interaction term by itself in those models (such that this measure can be thought
of as a window-weighted or window-adjusted weather index). To aid interpretation when using
the interaction by itself, we z-score it – such that, in Panel A, the coefficient suggests that a one
16
The variance in sales across call centers is also generally higher in locations with windows (see Figure S2). The
overall relationship between windows and sales is negative, though less systematic than the positive relationship
between windows and happiness (see Figure S2).
4.4
Robustness
Window-Based Exposure. Given that our identification relies on workers’ differential windowinduced exposure to weather, one concern with the approach of relying on objectively-measured
external window coverage is that this may not translate to a worker’s visual exposure to outdoor conditions. For example, even if a building is externally fully glass-covered, some workers
inside may nevertheless have work-spaces with no windows. In our context this is unlikely to
be a problem, given the internal layout is similarly open-plan across buildings. This means the
inside is always one large space in which any given worker will have roughly equal visual access,
on average, to the walls – be they glass or solid. Moreover, as a fully open space, natural light
(if it is let in through windows) will spread through the space.
We combine our objective data on window coverage with two additional sources of data:
semi-structured interviews with managers and a short supplementary survey of workers (see
Appendix J for more details on this additional data collection). The survey enabled us to confirm, from the the perspective of workers themselves, what we (a) were told by management as
well as (b) observed on site visits: that despite significant exterior variation, all of the buildings
are open-plan (see Figure 5 for examples of this in both high- and low-window buildings). We
find from the survey that 95.6% (N=318) report working in an open office space as opposed to
a closed office space.
Moreover, the supplemental worker survey allowed us to ask workers directly about their
subjective experience of windows and light. Here we first ask workers to imagine sitting at their
typical workstation, and then ask “Do you see few or many windows? ” – with a 0 to 10 answer
Figure 7: Subjective and Objective Measures of Visual Exposure to External Conditions
Notes: Objective window share is calculated from photos and using image processing software. Call-center-averages
displayed for responses to worker-level survey questions, each on a 0 to 10 scales, with the following wordings.
Subjective windows: “Imagine sitting at your typical workstation. Do you see few or many windows? Use the
slider below. Imagine a 10 being a completely glass office and 0 being a room with no windows at all.” Subjective
light: “While at work, how much natural light do you have access to from your workstation? Use the slider below.
Imagine a 10 being like seating outside while working and 0 being an office with no access to natural light at all.”
scale where 0 is labelled as “no windows at all ” and 10 “fully glass building.” As a further test to
this, we also ask how much natural light they have access to from their workstation, on a scale
where 0 is “no natural light at all ” and 10 is “like sitting outside.” As can be seen in Figure 7, we
find that workers’ subjective impression of windows and natural light is very highly correlated
with our objective measure of external window coverage (r = 0.83 and r = 0.91, respectively).
Using these subjective measures instead of the objective measure in order to construct our main
instrument, we also show that this does not affect our main 2SLS estimates (see columns (6)-(7)
of Table 2).
Product Demand Effects. A further concern in relation to the exclusion restriction is that
weather may have an effect on customer demand (either directly or by, indirectly, affecting
customers’ mood). This would be a threat to the exclusion restriction. However, in this twosided market, we have call center workers in fixed locations, and customers calling in from
all over the country. As we noted in the Introduction, one of the four key factors in our
identification strategy is that our call centers are sufficiently well dispersed across a country
that has variable weather across space, even within a given week. For a map of the spatial
distribution of these call centers, see Figure 1, which shows this dispersion across three nations
of the United Kingdom. We are also able to show in panel (b) of Figure 4 that, conditional on
the time fixed effects, there is variation in the visual nature of the weather across the locales.18
The inclusion of time fixed effects ensures that our key piece of identifying variation is exposure to weather shocks across call centers but within week. We rely on variation in weather
across call centers within any given week, rather than on movements in national weather conditions from week-to-week. Given that call centers do not field calls based on originating location,19 local weather in the vicinity of the focal call center should be independent, on average,
of customer demand.20
Although we cannot observe the origin of calls directly—our data is at the worker-day level
rather than the call level—we provide direct evidence that local demand pressure at a particular
call center measured by the average daily number of incoming calls per worker and the average
duration of a call is not affected by daily weather that is local to that particular call center
(Table S6).21 Moreover, if it were the case that calls were only routed within each of the four
nations of the UK, such that all calls coming from English customers were only routed towards
workstations located within England, then the addition of England-week fixed effects should
significantly affect our point estimate. We find in column (5) of Table 2, however, that this is
not the case.
Additional Controls. Adverse weather could physically be affecting sales performance through
changes in temperature or pollution. Temperature has been shown, for example, to affect student learning and academic test scores (e.g. Park et al., 2020) as well as investment decisions
(Huang, Xu and Yu, 2020).22 We show in column (2) of Table 2 that, when controlling for local
temperatures, all of our main findings are robust. The temperature variable itself does not enter
into the equation in a statistically significant way. Relatedly, air pollution can have a direct
impact upon worker productivity (e.g. Chang et al., 2016, 2019; Graff Zivin and Neidell, 2012).
Although we are not able to measure air pollution directly, pollution correlates with temperature, and there is little reason to suspect that glass-clad buildings would be more susceptible to
air pollution effects than warehouse-style ones.
Although weather also varies within a day across locations, the (normalized) distribution of the weather
index within a day is narrower than within a week (Figure S8). Measuring employee mood on a weekly basis is
hence also consistent with the spatial variability of weather being higher within weeks than within days.
The company have since our study introduced a new system (in Spring 2019) that does allow for geographical
allocation; however, no such technological capability was in place at the time of our study (July 2017 – January
2018).
Using data on customer satisfaction aggregated at the location-week level, we confirm that a small (negative)
residual correlation between our weather index and customer satisfaction across workstations nationally (r=0.105) is entirely nullified after adding week fixed effect (r=0.001).
Local supply shocks could also affect productivity if, for instance, gloomy weather makes it harder for workers
to answer calls. However, this should affect all call centers regardless of the window coverage of the building.
These effects are more likely to be driven by snowfalls than by fog or rain, but we find the effects are not driven
solely by snowfall (Table S7).
Using an empirical strategy somewhat analogous to ours, Park et al. (2020) show null effects in schools
that have air conditioning (i.e. instead of turning visual weather on and off with windows, they turn heat on
and off with AC). Nevertheless, Heyes and Saberian (2019) presents evidence to suggest a mood effect (though
without measuring mood) using high temperature and decisions of immigration judges who are inside of climatecontrolled buildings. While they argue judges may ‘import’ the mood effect of outdoor temperature when they
move indoors, Spamann (2022) finds no evidence of temperature on judge decision-making when using a larger
sample of years and when looking at all criminal sentencing decisions by US federal district judges. In terms of
studying mood effects, we prefer to focus clearly on visual aspects of weather coupled with observable variation
in exposure to be more sure of the mood mechanism taking place.
In addition to controlling for temperature, in column (3) of Table 2 we control for the main
effect of the visual weather index, while using the interaction of weather and windows as our
instrument. This analysis ensures our identification comes solely from the interaction, and when
doing so we find consistent results, though a slightly higher point estimate. As was suggested
by the reduced form and first stage evidence shown above in column (2) of Table 1, the main
effect of weather is not statistically different from zero in this case.
While we favor a parsimonious specification, our main 2SLS estimate also remains largely
unchanged if we include a much more exhaustive set of controls capturing the detailed daily
work schedule of workers and additional labor supply controls (see column (4) of Table 2).23
Sickness. Adverse weather conditions may cause sickness among workers, and impair their
ability to work effectively if they attend work while ill. Looking directly in the data, the local
weather turns out to be unrelated to the local share of workers under sick leave in any given
day or week (Table S8). We also consider the possibility that such a relationship may occur
with some lag. Weather conditions a day (or a week) before remains unrelated to the frequency
of sick leaves a day (or week) after. Most importantly, however, any sickness argument would
apply whether or not the call center had many windows. Given the relationships shown in
Figure 6, it seems clear that the variation our instrument picks up on is related to psychological
rather than physical health.
Sorting Effects. It is assumed throughout the paper that no other factors correlated with the
share of windows, but unrelated to visual exposure itself may explain heterogeneous sensitivity
to weather. The issue could arise if certain types of workers happen to be more negatively
affected than others by adverse weather conditions (e.g. older or sicker people), and if those
same workers tend to be systematically working in call centers with more windows. Table S9
shows balance across call centers with few or many windows (below vs. above median window
share) in terms of either baseline worker characteristics or the weather index. Of course, one
should only be worried about sorting effects if there exists important sources of heterogeneity
in worker sensitivity to weather in the first place, other than through visual salience. We
investigate this possibility directly, looking at whether weather (non-adjusted for exposure)
affects workers’ mood differently across observable worker characteristics. We look at basic
demographics (gender, age and workers’ tenure), the total number of weekly sales, and how
frequently the worker takes sick leave. We find no evidence of heterogeneity across any of these
dimensions (Table S10).24
This includes a full set of 35 indicator variables capturing the detailed daily work schedule of workers. For
each day of the week, we control for whether workers started (ended) their shift in the morning (7:00-12:00),
afternoon (12:00-17:00), or evening (17:00-24:00), or whether they did not work at all that day. We also add to
this list more detailed working supply controls, namely i) the total amount of hours spent on breaks, ii) the total
amount of hours spent working overtime, and iii) whether the worker was reported sick during the week.
It may still be that workers in call-centers with more windows would have a tendency to report a better (or
worse) mood when they can visualize good (or bad) weather. This effect would be consistent with the effect of
visual weather on mood being stronger in call centers with more windows. However, it is not a threat to our
identification strategy as long as (i) this effect is not biased towards reporting only good or bad mood and (ii)
reporting behavior is correlated to an actual shift in mood, which is further confirmed by the fact visual exposure
to weather does have clear behavioral effects on performance.
Functional Form of Happiness. One concern that we noted above in relation to our happiness survey is that answers are given on an ordinal scale. We make the assumption that this can
be cardinalized into a continuous measure of happiness. We provide a test for the reasonableness of this assumption by replacing continuous happiness with indicator variables for different
levels of happiness in a Poisson regression of sales on happiness (including the same set of fixed
effects and controls as in our main IV specification). Figure S3 reports the coefficients from
this exercise. We interpret the pattern of coefficients as suggestive evidence for being able to
use the happiness survey in a continuous manner in our instrumented analyses, meaning that
we “only” require one valid instrument rather than one for each categorical response.
First Stage Functional Form. In Figure S4, we show a graphical representation of our first
stage regression of happiness on visual exposure to weather. The relationship looks roughly
linear. But in order to explore and account for possible non-linearities more fully, we test
alternative functional forms for our instrumental variable. In Table 2 we use the squared value
of visual exposure to weather as well as the inverse hyperbolic sine transformation. We find in
columns (8) and (9) of Table 2 that the resultant second-stage coefficients are consistent with
our main results.25 Related to this, concerns related to weak instruments should be mitigated
by the reduced form evidence provided in this paper. Indeed, reduced form estimates remains
unbiased estimates, even if the instruments are weak (Angrist and Krueger, 2001).
Heterogeneous Responses. Instead of identifying an average treatment effect (ATE), a valid
instrument identifies a local average treatment effect (LATE) in the second stage (Angrist,
Imbens and Rubin, 1996) – that is, the effect driven by those whose mood is most sensitive
to visual exposure to weather.26 Assuming away the possibility of heterogeneous treatment
effects can be problematic when the causal effect of the endogenous variable is directly related
to the individual’s own choice (Angrist and Imbens, 1995). While this issue has been widely
discussed, for instance when estimating the returns to education (Angrist and Krueger, 1991), in
our case, mood movements are largely “external” to an individual: one does not have a direct
control over them. Heterogeneous treatment effects arising from employees’ selection on the
productivity gains of good mood are unlikely to occur in our context. Table S11 shows the first
stage of our IV strategy this time interacting visual exposure to weather with each of the six
main characteristics described earlier. We find no evidence of heterogeneity across any of these
dimensions.27 We also conduct sub-group analysis splitting between generally unhappy and
We also show a 2SLS regression in column (10) of Table 2 where we include three separate instruments,
one each for the incidence of the three weather phenomena that make up the index. Here we run into potential
problems of weak instruments, since the combined F-statistic is much lower in this case – as is typically the case
when using multiple instruments (see Table S7 for the first stage). However, results remain consistent overall.
As pointed by Angrist, Imbens and Rubin (1996), the LATE (or causal effect of weather-induced happiness
on sales) corresponds to the ratio of the impact of visual exposure to weather on sales (the reduced-form, which
may be referred to as an intent-to-treat effect) and the impact of visual exposure to weather on happiness (i.e.
the fraction of workers whose mood is sensitive to weather) – or alternatively it can be calculated using an IV
estimator. The LATE is of specific policy-relevance if the goal is to target those individuals whose affective state
is most likely to be affected by any policy change.
One reason for the higher salience of weather in buildings with more windows may result from workers paying
more attention to weather over time in those locations as they can more easily blame the bad weather for their
low productivity. Such learning effects should build up over time, so that workers with longer tenure should be
more sensitive to visual exposure to weather, which is not what we find (Table S11). We also found no evidence
happy workers as well as generally low and high productivity workers (below/above median in
each case). We find no statistical difference in the sensitivity of productivity to weather-induced
mood effects across those groups (Table S13).
Survey Non-Response. Using our final sample of workers, we do not observe a fully balanced
worker-week panel since we are restricted by non-response to the happiness survey instrument.
One concern here is that non-response to the survey is unlikely to occur randomly, and may
indeed relate to our main variables of interest in ways likely to bias our estimates. For example,
it could be that a worker does not respond in a given week because they are either too happy
or miserable to spend time reading the email. To explore this potential issue further, we
regress a dummy for having responded to the survey in a given week on a number of timevarying observables like sales, selling time, local gloomy weather (multiplied or not by window
share) and team average happiness (as well as a set of individual and week fixed effects).
Reassuringly, neither weekly sales performance nor team average happiness (minus the focal
worker) is significantly related to non-response within-individuals over time (see Table S1).
Non-response is, however, positively related to the number of hours worked during the week
and whether or not they work on Thursdays or Fridays, suggesting that workers are less likely
to respond during weeks in which they are scheduled to work less. Importantly, response is also
unaffected by local weather patterns as they vary week to week.28
Timing of Happiness Response. Respondents are explicitly asked to report their average
mood state over the course of the week. Here, we test whether the response to the happiness
question indeed captures respondents’ mood during the week in which the response is reported
(and not the week after or before). As a placebo test, we check whether weekly happiness
significantly relates to visual exposure to weather a week before or a week after the happiness
response week. We find that it is not the case (Table S14). We also find the same result
looking at reported happiness and weekly sales. Moreover, looking at those relationships within
the same week (with reported happiness applied equally to each day up to 3-4 days before vs.
after the response day) using daily weather and sales data around the response day, reported
happiness about “this week” is significantly and strongly related to visual exposure to weather
and sales on the response day and a few days before, but much less so (or not significantly) a
few days after (Figure S7).
Analysis using Daily Data. Although our happiness data is measured at the weekly level,
our productivity data is reported largely at the daily level. In table S15 we regress daily sales
on daily weather exposure at work, together with a full set of individual and date fixed effects,
plus our standard set of daily work schedule controls (equivalent to above). We find that daily
visual exposure to weather has a negative and significant effect on daily sales performance.
for seasonality in the effect of visual exposure to weather for both happiness and sales (see Table S12).
One potential approach to dealing with non-response to the survey is to impute any missing values as the
lowest or the highest category. However, given the suggestive evidence that response behavior is not systematically
related to individuals’ happiness, imputing low or high value is likely to simply add measurement error to the
reported happiness data. When imputing in this way, we find that it does not affect the reduced form effect of
visual exposure to weather on sales, but it does lower the first stage impact of visual exposure to weather on
happiness (Appendix Table S4).
We provide two further analyses, this time using the weekly happiness data combined
with the daily sales and performance data. First, we assume that responses to the happiness
question—which is asked on Thursday and refers specifically to “this week”—apply equally to
each of the weekdays (Figure S7 demonstrates this assumption is supported by the data itself).
We then estimate the 2SLS regression using a daily dataset, and include date (instead of week)
fixed effects. When doing so, in column (11) of Table 2 we find that our results are unchanged.
In column (12) we make the more restrictive assumption that the weekly happiness question
applies only to the day of response. Restricting only to the day of response, we again find
similar results. These effects are less precisely estimated, consistent with the lower variance in
weather within a given day across locations, relative to within a given week, as documented in
Figure S8.
Evidence on Channels
In Section 2.3 we discussed three broad ways in which happiness might be expected to affect work
performance: the impacts of positive affect on cognitive processing (in particular being more
efficient or integrative in thought), work motivation, and social or emotional skills. Although the
main focus of the paper is to provide evidence on the main effect of happiness on productivity
in a field setting, we move in this section to provide some suggestive evidence on the relative
importance of those various mechanisms, at least in our context.
5.1
Cognitive Mechanisms
Though we cannot measure cognition directly, an impact of positive affect on the ways in which
workers think could be reflected in three major labor productivity measures: adherence to daily
workflow schedule, speed, and call-to-sale conversion. First, adherence to workflow – where,
for example, positive affect may lead to greater flexibility in thought and a better ability to
multi-task as well as effectively plan and switch between tasks. In our setting, workers attend
and have their day’s workflow scheduled for them and displayed on their terminal screen (for
example, they may have the first hour scheduled as selling TV bundles, the second selling
internet connections, a 15 minute break, and then an hour selling something else). The firm
routinely records the extent to which employees adhere to this scheduled workflow. We code
our adherence outcome variable here as 1 if the firm’s target is met, zero otherwise.29 How
does workers’ mood affect this outcome? In Table 3 Panel A, we present reduced form evidence
using our measure of visual exposure to weather. In Panel B, we show the results from the
second stage of a 2SLS regression, in which happiness is instrumented for using visual exposure
to weather. We find that happier workers adhere more closely to the workflow that has been
set out for them (see column (1) in Table 3). In our setting, however, conditional on the
total number of hours spent at work (selling or doing other internally scheduled non-productive
activities), adherence happens to have little influence on sales (see Table S16). In other settings
with different types of work and where adherence may be more critical to performance, however,
this mechanism could well be more consequential.
Second, improved efficiency in information processing, brought about by higher positive
affect, should translate into an ability to work faster. We observe, on a daily basis, the total
number of minutes spent on incoming calls as well as the number of calls taken, and show in
column (2) of Table 3 that in happier weeks workers work faster. This “speed” measure is what
would typically be used as a labor productivity metric in the manufacturing industry. However,
in our setting, and in the service industry more generally, it is not clear that taking more,
shorter calls will be beneficial when the goal is selling.30 Table S16 suggests that calls-per-hour
is not a good predictor of productivity – and, if anything, it is associated with a reduction in the
number of weekly sales per worker. In settings where speed may be more crucial, this channel
could have more of an effect.
Relative to adherence or speed, a third productivity metric is arguably more clearly linked
to problem-solving: call-to-sale conversion. Column (3) shows that in happier weeks, workers
convert more of their calls to sales. Out of the three cognitive mechanisms, workers’ improved
ability to solve customers’ problems seems most likely to explain our main productivity effect.
Indeed, the point estimate for adherence and calls per hour are relatively small: Though a
30
This speed-quality trade-off is particularly salient in call center settings (Singh, 2000). Indeed, faster calls
may displease customers and make them less likely to buy if the operator is too blunt or quick with them.
Furthermore, sales calls are likely to be mechanically longer, due to the time it takes to complete an order, take
payment details, and so on.
1-point increase in happiness leads to a rise in the average number of calls per hour from 5
to 5.3 calls, speed is unable to explain much of an increase in sales. The dominance of the
conversion channel is more apparent in Column (4), where we control for adherence and the
number of calls per hour. It confirms that the average number of calls per hour is negatively,
not positively, correlated with the conversion rate.
We saw that an important higher-order mechanism explaining how positive affect may lead
to higher performance is through its impact on workers’ ability to find integrative solutions
to problems. In routine jobs or tasks where there is little space for integrative thinking, this
channel may be much smaller than in solving more complex tasks (or even go in the other
direction). Although we do not observe the amount of time spent (or number of calls) selling
different products, we are able to examine the effect of happiness on different types of sales.
When doing so, we find that, although all of our estimates are less precise, the magnitude is close
to zero for regular order-taking (see Figure 8a). This is consistent with the main mechanism
being call-to-sales conversion rather than working faster (or more efficiently), since line sales are
largely mechanical order-taking. More strongly positive effects are found for TV and cell-phone
contracts, which are also more technical and involve selling bundles with multiple different
options, as well as for re-contracting sales.
5.2
Motivational Mechanisms
Besides purely cognitive mechanisms, positive affect may also increase performance through
higher work motivation – in particular, as noted in Section 2.3, by making work more enjoyable,
happiness may incite workers to put more effort on the least enjoyable tasks and spend more
time at work generally (or take fewer breaks). Although we do not measure work motivation
directly, higher motivation is unlikely to be the main driver of the happiness-productivity effect
in our setting. First, happiness has no impact on simple order-taking, which is the most routine
and likely least fulfilling selling task. Since baseline motivation on those tasks is low, if higher
motivation were a major mechanism then those particular sales should be most reactive to
happiness, which is not what we find. Second, work motivation is arguably linked to the
amount of time spent at work. In addition to labor productivity, we can therefore investigate
the impact of positive mood on short-run labor supply decisions. If positive mood makes work
more enjoyable, we may also expect to find that workers would spend more time at work.
We first look at total number of selling hours, which we used as our main control of labor
inputs throughout the paper. In column (1) of Table 4, we find no robust evidence of any
happiness effects on the amount of time spent selling. We observe additional high-frequency
data on the allocation of time between work and leisure. First, we observe a percentage measure
of weekly attendance, which has a mean of around 93%. Here we code whether the employee
recorded perfect attendance to their scheduled hours during the week. Here too, we find no
robust evidence of any significant mood effects on attendance. This rules out the possibility that
happier workers would find work more enjoyable, and hence be motivated to attend work more
often. We also find negative (but non-significant) coefficients for over-time working. Finally, we
observe whether workers took any paid vacation during the week, and the number and length
of breaks taken by workers. Both coefficients are negative but non-significant. Taken together,
5.3
Social and Emotional Mechanisms
Finally, positive mood could also augment social and emotional skills, especially in a real-life
workplace where workers interact with customers and co-workers. Figure 8a shows that the
strongest positive effect of happiness is on re-contracting sales. In these situations, workers are
negotiating and, consistent with the experimental literature discussed in Section 2.3, may be
better able to find integrative solutions through persuasion – likely because of relying less on
contentious tactics.
In fact, the effect of positive affect on re-contracting likely involves both a cognitive (better
problem-solving) and a social (friendliness and negotiation) channel. Both factors could be
playing a role and we cannot separately identify each factor in this paper. However, we present
some suggestive evidence consistent with a more social channel. In a customer-facing setting,
this is an expected channel – indeed, the sociological literature on emotional labor (see, e.g.
Hochschild, 1983) has long argued that, in tasks involving interactions with customers, it becomes particularly costly for unhappy employees to leverage their social skills and manage their
emotions as they need to “fake” happiness.31 Consistent with those hypotheses, we find that
the happiness effect is stronger during weeks where workers are, on average, most likely to face
unsatisfied customers. Figure 8b replicates our main analysis but split by weeks with above versus below median levels of national customer satisfaction.32 In weeks where national customer
satisfaction is low, being in a good mood has a stronger positive effect on sales than during
weeks where customer satisfaction is high. Going beyond this, we also find that this effect only
manifests in the most complex tasks of TV/cell phone and re-contracting sales: in particular,
in weeks where the average customer is satisfied, the effect of happiness on those sales is null,
which suggests the type of skills driving the effects of positive affect in our setting are not simply
cognitive but also social (Figure S9). This effect may be due to being nicer, better at negotiation (e.g. via persuasion), or a mix of both. The dominance of persuasion in negotiation skills
would suggest an overall negative effect on customer satisfaction as customers would accept
solutions that may not be in their best interest. Using the weekly average customer satisfaction
response for each worker, we show in Table S17 that, within-workers over time, happiness has
a negative (although very imprecisely estimated) relationship with customer satisfaction. We
are hesitant to over-interpret this result, however, given that the customer satisfaction data is
not only endogenous but also very noisy when used at the worker level because each employee
receives very few feedback ratings, on average, per week.33
Finally, as we noted in Section 2.3, happiness may here have an ambiguous theoretical
relationship with performance given that the more social mechanisms may in some contexts
lead to distraction or loafing (cf. Coviello et al., 2020). Indeed, higher sociability may incite
workers to spend less time working and more time socializing with their co-workers while at
work, which would predict negative happiness effects on productivity. In our particular setting,
however, where team work is absent and workers are monitored, we find no evidence suggesting
sociability may be detrimental to productivity. On the contrary, happier workers take calls
slightly faster and the effect of happiness on break taking is negative, effectively ruling out any
large-scale effect whereby happier workers take longer breaks to chat with their co-workers.
Discussion
6.1
Managerial Implications
We employ variation in visual exposure to weather level as an econometric device in order to
isolate the causal effect of happiness on productivity. Our study is thus best thought of as a
form of basic research, albeit in an applied setting. As such, we are not able to say definitively
whether windows, in and of themselves, are good or bad for firms. We show that workers are,
on average, happier in buildings with more windows; however, without plausibly exogenous
variation in window coverage or a setting where workers routinely switch between buildings, it
is difficult to make any strong claims. Nevertheless, our findings do demonstrate the importance
of an often-overlooked aspect of work, namely the physical environment – and the ways that it
is an important factor in mediating the effects of environmental factors on workers.
A natural question is whether or not the effect of weather-induced happiness on productivity
is a useful (i.e. policy- or managerially-relevant) parameter to estimate. One initial thing to
note here is that although the root of an emotion is typically easily traceable for people, the
source of moods is not. There is little reason to expect the effects of a weather-induced mood
to be any different from the effects of a mood state induced by other factors, many of which
are managerially relevant. Indeed, a great deal of research has shown workplace mood and
happiness can be influenced by a range of management practices and other organizationallyrelevant factors (Krekel, Ward and De Neve, 2019). Interestingly, this is a point that is already
well understood by firms themselves: in a recent survey of a large sample of U.S. executives, 95%
believed that they have “some” or “a high degree” of control when it comes to influencing the
happiness of their employees (see HBR Analytical Services, 2020). A growing body of evidence
suggests that employee happiness is at least partially determined by structural factors related to
how firms organize and manage work as well as the workplace cultures they create. Recent fieldexperimental evidence shows, for example, significant effects on wellbeing of various management
practices – including monitoring, performance information feedback, personal targets, and prosocial incentives (Gosnell, List and Metcalfe, 2020). Further work on aspects of work life—
such as manager support and flexibility (Moen et al., 2016), pay inequality (Breza, Kaur and
Shamdasani, 2017; Cullen and Perez-Truglia, 2019), gift exchange (DellaVigna et al., 2020), and
worker autonomy (Bloom et al., 2014)—has also shown that management practices can have
impacts on employee wellbeing.
Although there has been over a century of empirical work on the issue of employee wellbeing
and performance, there remains little causal evidence in the field. Thus our confirmation of
earlier laboratory findings (e.g. Erez and Isen, 2002; Oswald, Proto and Sgroi, 2015) is an
important step forwards in the literature. In particular, managers may worry that findings
inside the lab may not replicate or scale easily in real-life settings, and may not be able to
relate the experimental evidence to the types of tasks, skills and workplace environments that
their workers face. Our paper not only provides confirmation that the type of experimental
evidence generated in the lab replicates in the field, but suggests those effects may be even more
relevant when looking at more complex tasks involving real social interactions with customers.
Conversely, our results should lower managerial concerns that the benefits of happiness on
productivity may necessarily be smaller in real-life settings as more sociable workers would
simply tend to work less.
6.2
Magnitudes & Comparisons with Existing Literature
We estimate an average marginal effect of 3.36 additional sales for each one unit increase in
happiness (from a base of around 25 sales per week), suggesting around a 12% effect. This is for
a one unit increase in the 0-10 happiness scale, which has a within-worker standard deviation
of 2.37