Effect of Mood on Workplace Productivity - HEC...
Transcript of Effect of Mood on Workplace Productivity - HEC...
Effect of Mood on Workplace Productivity*
Decio Coviello, Erika Deserranno, Nicola Persico, Paola Sapienza
December 24, 2017
Abstract
We leverage unique data on call-center workers to explore the causal effect of mood
on their productivity in the field. Mood is measured through an online “mood question-
naire” which the workers are encouraged to fill out daily. We find that better mood actually
decreases our call-center workers’ productivity. This finding holds both at a correlational
level and in two IV settings, where mood is instrumented for by weather or, alternatively,
by whether the local professional sports team has played the day before. We interpret this
finding through the lens of a model where, consistent with experimental evidence, good
mood increases sociability. Thus improving workplace mood may make “work downtime”
more appealing, at the expense of productivity. The effect of mood is more muted for the
subset of call-center workers whose compensation depends on productivity (high-powered
incentives). We rule out a number of threats to the exclusion restrictions. To our knowl-
edge, this is the first evidence of a causal link between worker mood and productivity in the
field. JEL Codes: J24, M52
*Decio Coviello: HEC Montreal, [email protected]. Erika Deserranno: Kellogg School of Management,Northwestern University, [email protected]. Nicola Persico: Kellogg School of Man-agement, Northwestern University, [email protected]. Paola Sapienza: Kellogg School of Man-agement, Northwestern University, [email protected]. We thank Shumiao Ouyang andAthanasse Zafirov for excellent research assistance. This research was conducted in collaboration with WorkforceScience Project of the Searle Center for Law, Regulation and Economic Growth at Northwestern University. Thispaper has been screened to ensure no confidential information is revealed. Data and institutional background willbe provided such that we do not disclose information that may allow the firm to be identified.
1
1 Introduction
It is a popular notion that good mood improves workplace productivity. A Google search of
“mood and productivity,” for example, turns up a lot of managerial literature supporting this
idea.1 However, the managerial literature is based on non systematic evidence. The economics
literature largely ignores mood as a determinant of productivity (with a few pioneering excep-
tions discussed later).
A causal link between mood and productivity, if established, would have profound conse-
quences for the economic analysis of incentives in the workplace. Firms routinely choose how
much to invest on workplace mood (by allowing more time off, longer breaks, office celebra-
tions and events, use of social media at work, etc) and on compensation. If, as we will argue,
improving workplace mood decreases effort but increases worker welfare, knowing the effect
of mood on productivity may help the firm calibrate the rest of its compensation scheme.
What causal evidence is available for the link between mood and productivity comes from
laboratory experiments. In these experiments a subject’s mood is manipulated and then the sub-
ject’s performance in an experimental task (e.g., performing long additions) is measured. But
performance in these experiments may be a poor proxy for real-world workplace performance
because the experimental setting fails to provide the opportunities for “social downtime” (the
water cooler conversations) that are available in real-world workplaces. Since it is known that
good mood increases sociability and vulnerability to distractions, it is possible that improv-
ing workplace mood may render social downtime more appealing, thus reducing productivity.
Therefore, obtaining causal evidence from the field, i.e., effect of mood on actual workplace
productivity, is of critical importance. To our knowledge, such evidence is lacking.
In this paper we leverage unique data on call-center workers to explore the causal effect of
mood on their productivity in the field. What is unique is that we have data on worker mood.
Mood is measured through an online “mood questionnaire” which the workers are encouraged
to fill out daily: see Figure 1.2 Productivity is measured by the number of calls per worker/hour,
1Gallup Inc. has measured workplace well-being for decades, and has long supported the notion of a linkbetween wellbeing and productivity. Jim Harter, Chief Scientist of Gallup’s Workplace Wellbeing Practices, writesthat “Investigation of the happy productive worker clearly links emotional well-being with job performance.”
2The mood questionnaire arises from the company’s desire to measure worker engagement.
2
Figure 1: Screenshot of Mood Questionnaire
and by other measures including downtime, time the customer is put on hold, and customer
satisfaction.
The panel structure of the data allows us to use worker fixed effects. Identification therefore
leverages within-worker variation in mood. Controlling for worker time-invariant characteris-
tics (like worker ability), we find that better mood is negatively correlated with productivity.
Though provocative, this finding is consistent with the experimental psychology literature on
distraction and sociability. But we seek causal effects. The call-center setting is especially suit-
able for our purposes because variation in call-center demand (a likely confounder of produc-
tivity) is national, and thus independent of local shocks to mood which can be used to estimate
the causal effect of mood.
We instrument for mood with local weather on the same day and, separately, with whether a
local professional sports team played the day before. The first-stage estimates are as expected:
rain worsens mood, and the local sports team playing improves mood (though the latter effect
is somewhat weak, perhaps due to some heterogeneity in the response: the male workers’ mood
improves mainly after wins). Using these two instruments we estimate that mood has a very siz-
able, and similarly-sized for both instruments, negative causal effect on our call-center workers’
productivity. Both IV estimates are much larger than the OLS estimates (direct effect of mood
on productivity). We provide direct evidence of a reverse-causation bias in the OLS estimates
that may partly account for this difference.
We find that the effect of mood is more muted for the subset of call-center workers whose
compensation depends on productivity (high-powered incentives). In other words, mood affects
productivity more so when incentives are low-powered. This finding is consistent with a possi-
3
ble distraction effect of mood: workers for whom getting distracted has a higher monetary cost
are less likely to do so when they are in a good mood. This may be interpreted as indicating
that introducing monetary incentives crowds out the impact of mood.
The causal interpretation of our estimates rests on the assumption that the effect of weather
or sports games on productivity is mediated by mood alone. A first concern is that demand
might be related to weather (and maybe also to sports events). However, our call centers face a
national demand: calls from all over the U.S. are first centrally directed then routed to individual
call centers; thus we are able to show that demand is uncorrelated with our instruments. A
second concern is that our instruments might affect the number of hours a worker shows up at
work (e.g., bad weather may increase traffic; sports events may increase the likelihood that a
worker shows up late); and this may affect productivity per hour. However, we show that the
results hold if we control for the “number of hours at work,” or if we replicate the analysis on
the subsample of workers who live close to the office. A third concern, which is specific to
our weather instrument, is that forecasted weather might require workers to waste productive
time rearranging their schedules (if rain is forecasted, cancel the BBQ, and vice versa). The
idea is that if rain is forecasted tomorrow, a worker might have to spend some time today in
order to rearrange her personal schedule. To assess the importance of this concern, we regress
productivity at time t−1 on rain at time t; but we find no effect.
The paper proceeds as follows: Section 2 discusses the related literature; Section 3 presents
statistics and explains our institutional context. Section 4 presents a conceptual framework.
Sections 5 and 6 identify the correlation and the causal effect of mood on productivity: OLS and
IV results, respectively, and discusses potential threat to the IV identification strategy. Section
7 concludes by discussing the external validity of our results.
2 Related Literature
Mood as a precursor of engagement. Worker engagement is defined as “a positive,
fulfilling, work-related state of mind that is characterized by vigor, dedication, and absorption”
(Bakker 2008). This state of mind is seen as desirable in workers. Engagement is important to
4
HR practitioners because they believe that engagement can be measured and, also, influenced
by management.3 Importantly for our purposes, some academic literature supports the idea that
positive mood is a precursor of work engagement (Albrecht and Su, 2012); but to our knowledge
the literature should be read as supporting correlation, not necessarily causation. We interpret
our results as casting doubt on the notion that positive mood necessarily causes engagement
to increase. In other words, we believe that mood and engagement are two separate (though
related) concepts.
Mood and productivity. Our “mood instrument” captures a form of self-reported pos-
itive affect at work. Positive affect is a form of “subjective well-being” (SWB). There is a
large literature on the relationship between SWB and work performance. Tenney et al. (2015)
provide an excellent survey. Almost all observational studies in this literature report a posi-
tive correlation between SWB and a host of outcomes including: subjective and objective work
performance metrics, unemployment, health, relationship outside of work, etc. However, most
of the observational studies are cross-sectional and correlational in nature and thus not conclu-
sive about causality (Tenney et al. 2015, p. 40). Closest to our setting is Rothbard and Wilk
(2011), which studies (among other things) the correlation between call-center worker mood
at the beginning of workday and worker productivity during the workday. Rothbard and Wilk
(2011) do not find a statistically significant relationship between worker mood and productivity
as measured by the number of calls per hour. Like ours, this paper is based on administra-
tive data for two call centers. The main difference in research designs is that while Rothbard
and Wilk’s (2011) research design cannot exclude the possibility of omitted factors affecting
both beginning-of-day mood and productivity, our research design leverages two separate in-
struments for mood. So we can make stronger causality claims.
Many experimental studies exist where mood is manipulated through gift-giving or by show-
ing uplifting videos. Improving mood through priming has been shown to improve impulse con-
trol.4 Because impulse control should help resist the temptation to slack off, individual worker
3For example, the Gallupp Workplace Audit asks employees a battery of 12 questions including “I know whatis expected of me at work,” “At work, I have the opportunity to do what I do best every day,” “I have receivedrecognition or praise for doing good work,” etc. These questions are seeing as measuring worker engagement.Also, such questions pertain to dimensions of the workplace that are under management’s control.
4For example, Fry (1975) showed that children whose mood is positively primed (asked to think about pleasur-
5
productivity is presumably improved by better mood.5 Overall, these experiments afford strong
causality claims, but collectively their findings are somewhat ambiguous with regards to pro-
ductivity, and it unclear how these findings extend to actual workplaces. In particular, these
experiments do not allow experimental subjects to have “social downtime” which they can sub-
stitute for work time, and so the effect of mood on this substitution cannot be observed.6
Mood and Sociability. Priming a better mood has been shown to increase the sub-
jects’ vulnerability to distractions (Pacheco-Unguetti & Parmentier 2016), and to increase socia-
bility (see Cunningham 1988 and the literature cited therein). Neither trait necessarily promotes
productivity “in the wild,” as shown in the next paragraph.
Sociability and Productivity. A number of papers study how socializing affects em-
ployee productivity. In the context of a seafood-processing plant in Vietnam in which worker
productivity is individual, Park (2016) shows that productivity declines by up to 9% when work-
ers have their friends nearby and socialize. In the context of call-center workers in China, Bloom
et al. (2014) show that workers who work from home are 13% more productive. The effect is at-
tributed, partly, to the relative quiet at home and, possibly, to the absence of socializing. Overall,
this literature suggests that in a number of work settings sociability decreases performance.
Summary of our contribution to the literature. In sum, the current research does
not study the causal effect of mood on individual productivity in the field. Our paper is the
first to address this question, and it does so through an instrumental variable strategy. A further
contribution of our paper is to demonstrate that mood affects productivity more strongly when
incentives are soft (fixed salary) than when they are hard (pay for performance).
able events) are better able to resist temptation (play with a forbidden toy) than children whose mood is neutral oris negatively primed. Fry’s early insight has since been validated in a many different settings. Inducing emotionaldistress has been shown to impair the ability to: moderate food consumption; stop smoking; stop gambling; avoidcompulsive shopping; and generally delay gratification. See Tice et al. (2001) for a review of this literature.
5This is the interpretation favored by Oswald et al. (2015), who shows that inducing happiness in experimentalsubjects by showing them humorous videos causes the subjects to perform long additions faster.
6Consistent with our reduced-form findings, Lee et al. (2014) find that bad weather increases individual pro-ductivity in three experimental designs and one field observational study, in a setting where productivity is notdirectly affected by weather. However, Lee et al. (2014) stop short of using weather as an instrument for mood.
6
3 Data and institutional setting
Our call-center data cover 2,749 workers located in 9 call centers across 9 different US states
from January 2015 to February 2016. 73% of call centers workers are females. They are 34
years of age on average and mean tenure is 39 months (see Table 1).
Each call center representative works in a cubicle with a computer and a headset. Whenever
a representative is ready to accept calls, she is asked to clock in to the IT system and calls
are automatically routed into her headset. A call from any location in the US is randomly
allocated to whichever worker in any of the locations happens to be available. To take a break,
a worker temporarily pauses the system. In this case she stops receiving calls and is logged as
not available to receive calls. At the end of the working day, the employee is asked to clock out
of the system.
These IT records provide us with detailed information on worker’s daily productivity (see
Table 1). For each worker, we know the number of hours she shows up at work (mean is 6.23)
and the proportion of these hours that are “unproductive” (i.e., downtime: off the phone and
unavailable to receive a call; mean is 10%). We also have information on the number of calls
per hour handled by each worker (mean is 6.9), average call duration (7.2 minutes per call on
average) and the proportion of time a customer is put “on hold” (14% on average). Finally, the
company provided us with information on average daily customer satisfaction (Likert scale 1-
10, average 8). Customer-reported productivity measures are available for relatively few calls;
this may be because few customers are selected to answer these questions, or because few
customers choose to answer them. In the latter case an issue of selection arises, but we have no
visibility of customers non-response, so we take these numbers at face value.
Workers are divided into two positions: customer service representative, and sales represen-
tatives. Customer service representatives have the role of providing information about products
and services, take orders, respond to customer complaints, and process returns. Sales represen-
tatives evaluate consumer needs, recommend and sell products. The two call-center positions
differ in the compensation scheme. Customer service representatives are paid a fixed hourly
rate (mean is 11.5 dollars) and earn almost no commission on top of that. Sales representatives
7
Table 1: Summary Statistics
VARIABLES Obs. Mean S.D.
A) Demographics and Position (N=Workers)=1 if Female 2,749 0.73 0.45Age 2,749 33.66 13.81Position= Customer Service Representative 2,749 0.54 0.50Position= Sales Representative 2,749 0.37 0.48Position= Other 2,749 0.09 0.28Tenure (in months) 2,736 38.62 59.71Distance from home to the office (in km) 2,723 15.43 11.57
B) Productivity (N=Workers*Days)No. calls per hour 219,279 6.90 8.10No. hours at work 219,279 6.23 1.96Proportion `unproductive' time 219,279 0.10 0.08Average call duration (in minutes) 219,277 7.19 4.98Proportion of call duration "on hold" 219,271 0.14 0.13No. customers answering customer survey 219,277 0.58 1.02
Average daily customer satisfaction (1 to 10)[Conditional on at least 1 answer] 74,513 8.00 2.54
C) Earnings (N=Workers*Months)Daily earnings (gross) Customer Service Representative 11,107 1123.03 513.65
2,571 1208.60 593.70 Sales Representative Daily earnings per hour (fixed + variable pay; gross) Customer Service Representative 11,072 11.82 1.05 Sales Representative 2,568 13.65 4.71Proportion of earnings that are "incentivized" (vs. fixed) Customer Service Representative 11,072 0.02 0.06 Sales Representative 2,568 0.42 0.51
Notes: All variables in Panel A correspond to the most recent observation of each worker (one observation per worker). Panel B displays the mean and standard deviation of daily-level productivity measures (one observation per day and per worker). No. calls per hour = total number of daily calls divided by total hours at work. Proportion `unproductive' time= % time not spent on the phone with customers or not spent being available to receive phone calls. Proportion of call duration "on hold" = % time the worker puts the customer on hold vs. talk to the customer. Customer satisfaction score calculates the average daily customer satisfaction score for each worker (score 1 to 10). This variable is missing if none of the customer were asked to fill the survey and/or none of the customers answered the survey. Panel C presents information on worker earnings for customer representatives and sales representatives separately, at the monthly level (data is available at the monthly level).
8
earn a lower fixed hourly salary (mean is 7.9 dollars) with commissions on top (5.7 dollars per
hour on average). While the incentive scheme differs across the two positions, the aggregate
month pay is similar (1123 vs. 1208 dollars respectively). Workers are also similar in terms of
tenure, age and gender. Neither position is segregated in specific call center locations.
For both positions, our preferred measure of productivity is the “number of calls per hour.”
(Productivity is recorded hourly, rather than “per day” or “per shift,” and workers are compen-
sated hourly in this firm.)7 As a measure of downtime, we also report “the proportion of time a
worker is unproductive (not available to receive calls)” and “the proportion of the call duration
which is on hold.” In Table A.1, we show that the raw correlation between these proxies of
downtime and the “number of calls per hour” is negative, while we find a positive correlation
between the number of calls handled per hour and average customer satisfaction.
Mood is measured through an online “mood questionnaire” which the workers are encour-
aged to fill out: see Figure 1 and Table 2. Individual answers are anonymous; each call center
manager is provided with monthly summary statistics aggregated at the call-center level. The
questionnaire is presented to the worker upon logging into a particular software platform and
is asked only once per day. Logging in is required to access a number of HR functions includ-
ing tracking their pay information, accessing online training, setting one’s quarterly goals, and
giving and receiving performance feedback. Accordingly, we assume that the login choice is
largely determined by considerations other than mood and we restrict our sample to the 77,514
worker-days in which the worker logged in the platform. We provide evidence in support of
the assumption in Table 7 (Column 1) where we show that our weather and sports instruments
(which are related to mood) have either a very small effect or no effect at all on the login choice.
Conditional on logging into the platform, a worker answers the mood question 46% of the
time, while skips the question – by pressing an “exit” button – the rest of the time (see Table 2).
This requires taking a stand on how to code non-responses. It is believed by HR managers in the
organization that non-response to the mood question is an indication of bad mood. In personal
communication with one HR manager, the authors learned that workers may be uneasy reporting
7We do not focus on the “number of hours an employee shows up at work” as a key outcome variable because:(1) workers are compensated hourly and (2) schedules are set by the firm a week in advance and are thus unaffectedby daily mood.
9
Table 2: Summary Statistics of Mood Question
VARIABLES Obs. Mean S.D.
Daily Worker Mood (N=Workers*Days)
=1 if worker logs into platform 219,279 0.35 0.48Conditional on logging into platform…
=1 if worker answers mood question 77,514 0.46 0.50
% who feel Frustrated 35,783 0.08 0.26% who feel Exhausted 35,783 0.07 0.25% who feel SoSo 35,783 0.17 0.38% who feel Good 35,783 0.35 0.48% who feel Unstoppable 35,783 0.33 0.47
35,783 3.80 1.19between-workers S.D= 1.02
within-worker S.D= 0.75
Conditional on answering Mood Question …
Notes: Upon logging into an online platform, workers are asked the mood question; "How do you feel today: Frustrated, Exhausted, So so, Good or Unstoppable?" The question is asked maximum one time per day. The worker has the option of answering the mood question or skipping it. We report here the mood score conditional on answering the mood question (coding the no answer as missing).
Mood score 1 to 5 [1= frustrated ... 5=unstoppable]
10
bad mood despite the organization’s assurance of anonymity of survey results. Consistent with
this view, “bad mood” is underreported (see below). Moreover, the percent of “no answers” is
lower on Fridays, when mood is believed to be higher (start of the weekend).
We report all results following two approaches and we show that the results are consistent
regardless of how the missing responses are coded. In the first approach, we assume that “no
answer” means “bad mood” (frustrated). In the second, we discount the selection concern and
code the non-responses as missing observations, thus effectively dropping non-responses and
halving the sample. We also show that our results are robust to imputing intermediate mood
scores for “no answer” in the appendix.8
Conditional on answering the mood question, 68% of respondents report feeling either
“good” or “unstoppable”, while only 15% report feeling “exhausted” or “frustrated.” Mood
score (which takes value 1 for “frustrated”, 2 for “exhausted, 3 for “so so”, 4 for “good” and 5
for “unstoppable”) takes an average value of 3.8 among the respondents. As a validation check
of our mood data, we correlate reported mood with “days of the week” in Table A.2. As one
would expect, mood is higher on Fridays and lower during weekends (consistent with the notion
that employees do not like to work during weekends). Importantly, the variation in mood score
exists both between workers (s.d. 1.01) and also within workers (s.d. 0.75). The within-worker
portion of the variation is sizable. Because we use worker fixed effects, identification will come
from within-worker variation: we compare the productivity of a given worker in days in which
she is in good mood to days in which she is not.
4 Conceptual framework
This section makes two conceptual points. First, it presents a micro-foundation that rationalizes
why a worker’s productivity might be decreasing in mood. The basic insight is that better mood
makes the worker more sociable, thus increasing the value of work (down)time spent socializing
relative to work time spent on solitary productive activities (refer to the literature on mood and
sociability discussed in Section 2). This effect might be stronger when the productive activity8 Reassuringly, our instruments do not affect the dispersion of the mood answer (available upon request), in
which case coding no answers as “bad mood” may be incorrect.
11
is solitary whereas downtime is social, as is the case in our call-center setting. Second, this
section studies a firm which may decide how much to invest in improving workplace mood
vs compensation. The firm chooses the optimal investment mix on mood and compensation
in order to maximize effort conditional on retaining the worker. Improving workplace mood
decreases effort, but it increases worker’s welfare, which helps the firm retain workers.
A worker allocates a unit of time (workday, e.g.) between effort e and downtime d. Com-
pensation is: a fixed salary W plus a piece-rate w. The worker’s problem is:
maxe,d
u (W +we)+ g (d,m)
s.t. e+ d = 1.
where u (·) is the (concave) utility from wealth, and g (·, ·) is the utility of downtime which is
assumed to be concave in d.
Substituting from the constraint into the utility function we get:
maxe
u (W +we)+ g (1− e,m) .
Let e∗ (W ,w,m) denote the solution to this problem.
Proposition 1. Fix the compensation scheme (W ,w) . Suppose g1,2 > 0, that is, every ad-
ditional unit of downtime is more valuable when mood is better. Then optimal effort
e∗ (W ,w,m) is decreasing in mood m.
Proof. Because of concavity, the first order conditions characterize optimal effort choice, and
so optimal effort e∗ solves:
(1) wu′ (W +we) = g1 (1− e,m) .
Because g1,2 > 0, increasing m to m′ raises the right hand side which, by concavity, is an in-
creasing function of e. Therefore the optimal e gets smaller.
12
This proposition rationalizes why a worker’s productivity might be decreasing in mood.
The proposition takes the compensation scheme as fixed, and it does not ask what optimal
compensation scheme might be put in place by a firm who can also, at a cost, affect workplace
mood. Next we sketch the problem of a firm which sets the compensation scheme and work-
place mood to maximize profits. Workplace mood can be manipulated in a variety of ways,
including: more time off, longer breaks, office celebrations and events, allowing the use of
social media at work, free food in the workplace, etc.
The firm’s problem is :
maxW ,w,m
(1−w)e∗ (W ,w,m)−W −K (m)(2)
s.t. U∗ (W ,w,m) ≥Ω
In this formulation K (m) represents the cost of creating mood m (we assume it is increasing
in m); the function U∗ represents the worker’s indirect utility function, that is, the worker’s
maximum attainable utility given (W ,w,m) ; and Ω represents the value of the worker’s outside
option, which is a measure of labor-market tightness. The reason why the firm might want to
incur a cost to increase mood is that mood improves U , and thus makes it easier to meet the
worker’s participation constraint.
In the appendix, we solve (2) within a specific functional-form example where: W ≡ 0,
that is, workers are paid a piece rate w; u (x) =√
x;g (d,m) = m√
d, and m ≥ 1. Given these
primitives we compute:
e∗ (w,m) =w
w+m2
U∗ (w,m) =√
w+m2.
We then show that if labor market tightness Ω2 is below√
2 then the participation constraint
in (2) is not binding. In this scenario the firm optimally chooses to make no investment in im-
proving mood, and the optimal piece rate is w∗ =√
2−1≈ 0.41. If the participation constraint
binds, i.e., labor market tightness Ω2 exceeds√
2, then the optimal piece rate w∗ > 1/2 and
13
the firm chooses to invest in improving mood. Finally, we show that as labor market tightness
grows above√
2 then the optimal piece rate w∗ grows. As for the optimal investment in mood,
it will grow or shrink with Ω depending on the functional form assumed for K. Therefore, in
general mood can be a complement or a substitute to the optimal piece rate as labor market
tightness grows.
5 Negative correlation between mood on productivity:
OLS results
Table 3 reports the correlation between mood and productivity in our sample of call-center
workers. As explained above, we have daily-level individual mood and productivity data. The
panel structure of the data allows us to include worker fixed effects, thus controlling for any
endogeneity that may arise across workers and is fixed through time. We also add date fixed
effects (day x month x year), and control for worker tenure.9 Standard errors are clustered
at the worker level and at the call-center*date level (two-way clustering). This accounts for
correlation within a given worker and across workers of a given call-center in a given day. In
the appendix Table A.3, we show that our results remain identical if we allow for autocorrelation
at short horizon by clustering standard errors at the call center*week level.
Regardless of how mood is coded, we find that a higher mood score is negatively correlated
with productivity: a one unit increase in mood decreases the number of calls per hour by 6pp
(Table 3, Column 1).10 The correlation is relatively linear across the different moods: the
highest the mood score, the lowest the number of calls per hour (see Table A.7). Finally, mood
is found to be positively correlated with call duration while the correlation with “the proportion
of unproductive time” and customer satisfaction is very small (at least in these OLS regressions).
While not causal, these results indicate that within-worker variation in mood is negatively
correlated with productivity (as measured as “calls per hour”) for call-center operators. Is this
negative correlation true in other work settings? We can answer this question for more than9We do not include workers’ position and call-center fixed effects as these are fixed within a worker (we do not
observe position or call-center switches).10See Table A.4 for alternative coding assumptions.
14
Table 3: Mood and Productivity, OLS Results
(1) (2) (3) (4) (5)
Dep. Var # calls per hour
% un- productive
time
Average call duration (minutes)
% of call duration "on hold"
Average customer
satisfaction (1 to 10)
Panel A: Conditional on answering Mood Question
Mood Score (1 to 5) -0.061*** -0.002*** 0.081*** -0.001 0.019(0.015) (0.001) (0.020) (0.001) (0.027)
Observations 33,461 33,461 33,461 33,461 14,725Rsquared 0.819 0.312 0.762 0.755 0.269Mean Dep Var 6.956 0.0981 7.866 0.150 8.140
Panel B: Assuming that not answering Mood Question = Bad Mood
Mood Score (1 to 5) -0.056*** -0.000 0.048*** -0.001 0.004(0.011) (0.000) (0.014) (0.001) (0.015)
Observations 73,268 73,268 73,268 73,268 32,305Rsquared 0.815 0.297 0.752 0.743 0.250Mean Dep Var 7 0.0960 7.719 0.151 8.049Worker FE Day x Month x year FE Notes: All regressions control for worker tenure. Standard errors are clustered (twoway) at worker & call center*date level. *** p<0.01, ** p<0.05, * p<0.1.
15
20,000 sales associates in more than 500 retail stores covering the entire US who used the same
online platform from September 2013 until August 2015.11 Unlike call-center workers, we can-
not link individual mood with individual performance, but we can link store-level productivity
(at the monthly level) with average store-level mood in that month. Controlling for store fixed
effects, month x year fixed effects and for the number of workers in the store, Table A.6 shows
that the correlation is negative: higher mood score is associated with lower average store profits
and revenues. This shows that the negative correlation observed among our call center workers
generalizes to a larger and more representative pool of workers.
A limitation of OLS estimates is that they are subject to potentially large endogeneity, re-
verse causality and strong measurement error. Because of these concerns, we turn to an IV
strategy in the next section. We focus on the call-center dataset as the store-level data are not
granular enough to perform any analysis beyond a simple OLS.12
6 Negative effect of mood on productivity: IV estimates
There are two reasons to believe that OLS estimates may underestimate the size of the effect of
mood on productivity. First, reverse causality: a worker who happens to be highly productive
on a given day may feel happier because of that. Table A.7 (Column 2) provides suggestive
evidence of a feedback effect of work environment on our mood variable. Workers who an-
swered the mood question were then asked a follow-up question: “What contributed the most
to your mood?” and could identify the source of their mood as work-related (“boss,” “work
environment,” “co-workers,” etc.); or “non-work related.” We believe that work-related mood
is more likely to be subject to reverse causality. Indeed, work-related mood turns out to be
positively correlated with productivity, whereas non-work-related mood is negatively correlated
(coefficient of -0.74). Therefore, there is reason to believe that OLS estimates are significantly
attenuated by reverse causality.
11The proportion of workers who answer the mood question conditional on logging in and the average moodscore among sales associates is very similar to the one of call-center workers, with a similar distribution of answers(see Table A.5)
12Productivity data are aggregated at the store level and does not vary at the daily level.
16
The second reason to believe that OLS estimates underestimate the impact of mood is clas-
sical measurement error in the mood variable. Mood is intrinsically hard to measure, especially
when captured through surveys.
Due to these concerns about downward bias of the OLS estimates, we now present IV esti-
mates based on two separate instruments for daily mood: daily weather and professional sports
events. Both instruments yields quantitatively similar estimates for the effect of mood.
6.1 First-stage results
Weather instrument. We use weather as an instrument for worker mood, because we expect
bad weather to cause worse mood. The existing literature offers support for this notion. Seasons
are known to affect mood: in some people, the winter months bring bad mood and depression
(seasonal affective disorder). Higher-frequency weather (daily, rather than seasonal) has also
been found to affect mood (Keller et al. 2005, Braga et al. 2014).13
The weather data come from the National Oceanic and Atmospheric Administration (Global
Historical Climatology Network-Daily Dataset). The data contain four weather variables at the
daily and zipcode levels: precipitation, maximum and minimum temperatures, and snowfalls.
As an instrument, we choose the weather variable that is found to be most positively correlated
with mood: whether it rains or not during the day, i.e., whether precipitations are strictly pos-
itive, which is known to correlate with sunshine. As shown in Table A.8, the “rain dummy”
negatively affects mood with an F-statistics of 16.25 when no response is coded as “bad mood,”
and of 9.4 when no response is coded as “missing.” Using all four weather variables as instru-
ments for mood, or using “rain precipitation” (in ml) alone leads to lower F-statistics and hence
we prioritize “rain dummy” as our instrument.
In our sample, 61% of the days were rainy with considerable variation across days (s.d.
0.49). Rain also varies across localities, but not enough to precisely estimate our coefficients
controlling for date fixed effects. In all specifications in which we use rain as our instrument
for mood, we control for: day-of-the-week fixed effects and month x year fixed effects. We
also control for the historic amount of rain in each calendar day (average in the past 5 years) to13Braga et al. (2014) show that on rainy days college students are harsher in their teaching evaluations.
17
further control for seasonality.
Professional sports games instrument. For each call center, we collected information on
whether the local sport team (football, baseball, basketball, or hockey) played on any given
day.14 As shown in Table A.8, employee mood is higher if the local team played the day before.
However, the first stage has limited power: statistical significance is only achieved in Panel B
(where no answer is coded as bad mood) and, even then, the F-statistic is relatively low (=6).
Despite this limitation we feel that this second instrument adds depth to our story, and so in
what follows, we will use this instrument in Panel B.15
6.2 Second-stage results
Our second-stage estimates are presented in Table 4. Regardless of how we code mood “no
answers,” we find that one unit increase in mood score reduces the “number of calls per hour”
by roughly 1.5, equal to 20% of the average. This result holds both with the rain and with the
sports instrument.
A reduction in the “number of calls handled per hour” can be explained by two possible
channels: either calls become longer or workers spend less of their time on the phone. Table
4 shows that both channels are at play. A one unit increase in the mood score increases the
proportion of “unproductive time” (downtime) by 4 to 5 percentage points depending on the
specification. This corresponds to a 50% increase in unproductive time. Call duration also
increases with mood, although not precisely and not in all specifications. This increase in
call duration is due to customers being kept on hold: the “percentage of minutes on hold”
increases by 4 to 6 percentage points, corresponding to a 40% increase relative to the average.
While “longer calls” and “more time on hold” may not necessarily reflect lower productivity
(but could reflect a higher willingness to help the client for instance), we find that customer
satisfaction scores decrease by up to 41% (though the coefficients are not significant due to the
14We use “whether a sports team played” rather than whether the team won or lost, because winning and losingappear to have similar first-stage effects (both positive) on mood. We were puzzled by this finding until we brokedown the first stage by gender. We found that, whereas for women good mood was associated with the teamplaying the day before irrespective of winning or losing, for men, the mood after a win was three times as high asafter a loss.
15Adding a second-stage weak instrument correction (LIML estimator) does not change the second-stage results.
18
small number of observations in which customer satisfaction is recorded). Moreover, the fact
that the proportion of “unproductive time” increases necessarily means that workers spend less
of their time effectively working.
Table A.9 presents the same results with alternative coding strategies. While the exact co-
efficients change from one coding strategy to another, the coefficients are consistently negative
and large in magnitude. Table A.11 shows that the results survive with standard errors clustered
at call-center*week level.
The overall picture, then, is one of fewer number of calls per hour, a reduction in “productive
working time” and an increase in the “proportion of time a customer is placed on hold.” Our
conclusion is that an exogenous increase in mood causes productivity to decline and this decline
seems to be explained by an increase in downtime.
We now introduce the distinction between customer service and sales representatives. The
former are almost entirely compensated by fixed compensation per hour worked. The latter
have a significant part of their compensation based on “sales per hour” and “number of calls
per hour.” In our sample, customer service representatives earn an average of 1,123 dollars
per month (gross), 98% of which comes from a fixed hourly pay. Sales representatives earn
slightly more: 1,209 dollars per month and 41% of these earnings is “variable” and based
on performance (see Table 1). Because sales positively correlate with the number of calls,
the organization evaluates workers in both positions with the same performance metrics: the
“number of calls per hour.”
Table 5 indicates that the effects of mood on productivity are weaker for the subsample of
sales representatives, in that the coefficients on the interaction term “Mood Score x Sales Rep-
resentative” has the opposite sign to the “Mood Score” variable. Using rain as an IV for mood,
it appears that the “number of calls per hour” is 20% less responsive to mood for sales represen-
tatives than customer representatives (Panel B). This result is stronger (although less precise)
when using sports as an instrument for mood: sales representatives as 43% less responsive to
mood. Similarly, the “% unproductive time” and “% call duration on hold” are roughly 30%
less responsive to mood for sales representatives than for customer representatives with the rain
instrument (with the sports instrument, the effect on the “% call duration on hold” becomes
19
Tabl
e4:
Moo
dan
dPr
oduc
tivity
,Sec
ond
Stag
eIV
Res
ults
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Dep
. Var
# ca
lls
per h
our
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
# ca
lls
per h
our
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
Inst
rum
ent =
>
Pane
l A: C
ondi
tiona
l on
answ
erin
g M
ood
Que
stio
nM
ood
Scor
e (1
to 5
)-1
.760
*0.
053*
1.27
10.
060
-4.8
85-
--
--
(0.9
61)
(0.0
32)
(0.9
81)
(0.0
38)
(6.8
68)
--
--
-
Obs
erva
tions
33,4
6133
,461
33,4
6133
,461
14,7
25-
--
--
Mea
n D
ep V
ar7
0.09
607.
719
0.15
18.
049
--
--
-Fs
tat 1
st st
age
9.4
9.4
9.4
9.4
9.4
--
--
-
Pane
l B: A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= Ba
d M
ood
Moo
d Sc
ore
(1 to
5)
-1.5
39**
0.04
3**
0.71
30.
039*
-2.0
20-1
.485
*0.
008
-1.0
060.
069
-0.5
22(0
.675
)(0
.019
)(0
.655
)(0
.023
)(1
.742
)(0
.892
)(0
.020
)(1
.046
)(0
.042
)(1
.556
)
Obs
erva
tions
73,2
6873
,268
73,2
6873
,268
32,3
0573
,268
73,2
6873
,268
73,2
6832
,305
Mea
n D
ep V
ar7
0.09
607.
719
0.15
18.
049
70.
0960
7.71
90.
151
8.04
9Fs
tat 1
st st
age
16.2
16.2
16.2
16.2
16.2
66
66
6W
orke
r FE
Mon
th x
yea
r FE
D
ay o
f the
wee
k FE
D
ay x
mon
th x
yea
r FE
H
isto
ric ra
in
=1 if
Rai
ny D
ay=1
if S
port
Tea
m P
laye
d in
t-1
Not
es: S
econ
d st
age
IV re
gres
sion
s. A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at w
orke
r & ca
ll ce
nter
*dat
e le
vel.
His
toric
rain
= hi
stor
ic a
mou
nt o
f rai
n on
that
day
(ave
rage
in th
e pa
st 5
yea
rs).
*** p
<0.0
1, **
p<0
.05,
* p<
0.1.
20
even stronger, while the effect on “% unproductive time” is less strong).
Overall we find that mood affects productivity more so when incentives are low-powered.
This finding is consistent with the model in Section 4 because in that model changing the
power of incentives will alter the worker’s relative return of allocating a unit of time to ef-
fort vs downwtime. Note, however, that other alternative stories cannot be ruled out here: e.g.,
sales representatives may react more to mood because they are different ‘types of workers’ (e.g.,
less distractable or more money-driven) rather than because their compensation scheme is more
steep.
6.3 Threats to the exclusion restrictions
The size of the IV estimates is consistent across the two instruments, and we have provided
supporting evidence that rationalizes why it is larger than the OLS estimates. Nevertheless,
threats to the exclusion restrictions can never be discounted. Therefore, in this section we
investigate different threats to the exclusion restriction.
Hours worked. A first potential concern is that the number of hours an employee shows
up at work might be affected by weather or by whether the sports team played the day before.
E.g., rain may increase traffic and reduce hours worked, or, alternatively, rain may increase
hours worked by shifting leisure into work (see Connolly 2008). Similarly watching a sports
game the night before, may increase the number of workers late at work the day after. A direct
effect of our instruments on hours worked may violate the exclusion restriction if working more
hours negatively affects productivity per hour.16 To alleviate this concern we first show that
the second-stage results do not change if we control for the number of hours an employee was
at work (see Table 6, Table A.11 for other outcome variables, and Table A.12 for alternative
codings for mood). Second, we show that our rain and sports instruments have no direct effect
on the number of hours at work (intensive margin) and no effect on the number of workers who
are present at work (extensive margin); see Table 7, Columns 3 and 5. Finally, we find that the
results hold if we restrict the sample to workers who live less than 5km from the workplace and
who are therefore less likely to be delayed by traffic in getting to work (Table A.13).16The raw correlation between these two variables is presented in Table A.1 and is negative.
21
Tabl
e5:
Moo
dan
dPr
oduc
tivity
byIn
cent
ive
Stru
ctur
e,Se
cond
Stag
eIV
Res
ults
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Dep
. Var
# ca
lls p
er
hour
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
# ca
lls
per h
our
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
Inst
rum
ents
=>
Pane
l A: C
ondi
tiona
l on
answ
erin
g M
ood
Que
stio
n
Moo
d Sc
ore
(1 to
5)
-2.7
040.
026
2.46
4*0.
073*
-2.8
88(2
.342
)(0
.032
)(1
.306
)(0
.042
)(5
.103
)M
ood
Scor
e* S
ales
Rep
rese
ntat
ive
0.28
1-0
.011
**-0
.097
-0.0
14**
*0.
303
(0.3
48)
(0.0
04)
(0.1
42)
(0.0
05)
(0.2
96)
Obs
erva
tions
35,4
2535
,425
35,4
2435
,424
15,4
10p-
valu
e (M
ood
+ M
ood*
Sales
rep=
0)0.
246
0.60
70.
055
0.14
00.
594
Pane
l B: A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= Ba
d M
ood
Moo
d Sc
ore
(1 to
5)
-3.1
87**
*0.
033*
*1.
301*
*0.
042*
*-1
.569
-1.0
000.
008
-1.5
320.
044
0.85
2(0
.949
)(0
.015
)(0
.623
)(0
.019
)(1
.477
)(1
.225
)(0
.025
)(1
.275
)(0
.037
)(2
.107
)M
ood
Scor
e* S
ales
Rep
rese
ntat
ive
0.65
0**
-0.0
12**
*-0
.175
-0.0
23**
*0.
463
0.43
30.
003
-0.1
44-0
.040
***
0.56
0(0
.289
)(0
.004
)(0
.136
)(0
.004
)(0
.328
)(0
.346
)(0
.009
)(0
.472
)(0
.011
)(0
.501
)
Obs
erva
tions
77,2
6877
,268
77,2
6777
,265
33,7
5177
,268
77,2
6877
,267
77,2
6533
,751
p-va
lue (
Moo
d +
Moo
d*Sa
les re
p=0)
0.00
20.
109
0.03
90.
258
0.35
30.
680
0.70
90.
246
0.90
30.
559
Mea
n D
ep V
ar fo
r Sal
es R
ep8.
606
0.09
86.
065
0.04
08.
951
8.60
60.
098
6.06
50.
040
8.95
1M
ean
Dep
Var
for C
usto
mer
Rep
6.76
40.
088
7.97
80.
180
7.66
06.
764
0.08
87.
978
0.18
07.
660
Wor
ker F
E
M
onth
x y
ear F
E
Day
of t
he w
eek
FE
Day
x m
onth
x y
ear F
E
His
toric
rain
Rain
y D
ay &
Rai
ny D
ay*S
ales
Rep
rese
ntat
ive
Team
Pla
yed
& T
eam
Pla
yed*
Sale
s Rep
rese
ntat
ive
Not
es: S
econ
d st
age
IV re
gres
sion
s. A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at w
orke
r & ca
ll ce
nter
*dat
e le
vel.
His
toric
rain
= hi
stor
ic a
mou
nt o
f rai
n on
that
day
(ave
rage
in th
e pa
st 5
yea
rs).
*** p
<0.0
1, **
p<0
.05,
* p<
0.1.
22
Table 6: Mood and Productivity, Second Stage IV Results with Extra Controls
(1) (2) (3) (4) (5) (6) (7) (8)
Instrument =>
Panel A: Conditional on answering Mood QuestionMood score (1 to 5) -1.785* -1.743* -1.605* -1.546 -1.308
(0.964) (0.937) (0.929) (0.968) (1.087)No. hours at work -0.099*** -0.097*** -0.097*** -0.094***
(0.020) (0.020) (0.020) (0.024)No. incoming calls in call-center 0.043*** 0.043*** 0.046***
(0.014) (0.014) (0.015)Temperature -0.000 0.000
(0.000) (0.001)Nitric oxide (value) [Air pollution] 2.428
(3.063)Nitrogen dioxide (value) [Air pollution] 0.005
(0.003)Ozone (value) [Air pollution] -0.001
(0.005)
Observations 33,461 33,461 33,461 33,461 21,101Fstat 1st stage 0.658 0.667 0.690 0.700 0.727
Panel B: Assuming that not answering Mood Question = Bad Mood
Mood score (1 to 5) -1.542** -1.530** -1.439** -1.430** -1.344** -1.485* -1.480* -1.371(0.666) (0.653) (0.640) (0.633) (0.641) (0.892) (0.888) (0.863)
No. hours at work -0.084*** -0.084*** -0.084*** -0.086*** -0.059*** -0.059***(0.017) (0.016) (0.016) (0.021) (0.012) (0.011)
No. incoming calls in call-center 0.044*** 0.044*** 0.052*** 0.029**(0.012) (0.012) (0.014) (0.013)
Temperature -0.000 0.001(0.000) (0.000)
Nitric oxide (value) [Air pollution] 1.478(2.273)
Nitrogen dioxide (value) [Air pollution] 0.003(0.003)
Ozone (value) [Air pollution] -0.007(0.004)
Observations 73,268 73,268 73,268 73,268 46,761 73,268 73,268 73,268Fstat 1st stage 16.25 16.25 16.25 16.25 16.25 16.25 16.25 16.25Mean Dep Var 7 7 7 7 7 7 7 7Worker FE Month x year FE Day of the week FE Day x month x year FE Historic rain
Dependent variable = # calls per hour
Notes: Second stage IV regressions. All regressions control for worker tenure. Standard errors are clustered (twoway) at worker & call center*date level. Historic rain= historic amount of rain on that day (average in the past 5 years). *** p<0.01, ** p<0.05, * p<0.1.
=1 if Rainy Day =1 if Sport Team Played in t-1
23
Tabl
e7:
The
Red
uced
-For
mE
ffec
tson
Log
ging
-in,
Moo
dA
nsw
er,D
eman
dan
dPr
oduc
tivity
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
=1 if
logs
in
the
plat
form
=1 if
an
swer
s m
ood
ques
tion
# ho
urs a
t w
ork
# da
ily
inco
min
g ca
lls(in
'000
)
# w
orke
rs
pres
ent
at w
ork
=1 if
Rai
ny D
ay0.
006*
-0.0
01-0
.010
0.14
21.
916
0.05
1**
0.06
2**
(0.0
04)
(0.0
02)
(0.0
16)
(0.0
96)
(1.5
91)
(0.0
23)
(0.0
26)
0.03
00.
018
(0.0
20)
(0.0
23)
Wor
ker F
E
Mon
th x
yea
r FE
D
ay o
f the
wee
k FE
H
isto
ric ra
in
Spor
t Tea
m P
laye
d in
t-1
0.00
00.
005
-0.0
26-0
.023
-1.7
70-0
.035
-0.0
44(0
.005
)(0
.003
)(0
.018
)(0
.138
)(2
.596
)(0
.027
)(0
.000
)
Wor
ker F
E
Day
x m
onth
x y
ear F
E
Obs
erva
tions
214,
094
214,
094
214,
094
1,59
81,
598
214,
094
211,
371
75,9
4075
,148
R-sq
uare
d0.
457
0.45
00.
311
0.90
30.
892
0.79
50.
796
0.80
70.
808
Mea
n D
ep V
ar0.
354
0.16
37.
381
8.53
013
7.2
6.78
76.
787
77
Lead
Rai
n D
umm
y (=
1 if
Rain
y D
ay in
t+1)
Not
es: W
orke
r-le
vel r
egre
ssio
ns co
ntro
l for
wor
ker t
enur
e, w
ith st
anda
rd e
rror
s clu
ster
ed (t
wow
ay) a
t wor
ker &
call
cent
er*d
ate
leve
l. C
all-
cent
er le
vel r
egre
ssio
ns a
re co
llaps
ed a
t the
call
cent
er le
vel a
nd p
rese
nt st
anda
rd e
rror
s clu
ster
ed a
t the
call
cent
er*d
ate
leve
l. #
daily
in
com
ing
calls
(in
'000
) = th
e to
tal n
umbe
r of c
alls
rece
ived
in th
e ca
ll ce
nter
in a
giv
en d
ay. "
Lead
rain
dum
my"
=1
if it
rain
ed in
day
t+1.
H
isto
ric ra
in=
hist
oric
am
ount
of r
ain
on th
at d
ay (a
vera
ge in
the
past
5 y
ears
). Th
e nu
mbe
r of o
bser
vatio
ns is
hig
her i
n th
e fir
st 3
cols
than
in
the
prev
ious
regr
essi
ons b
ecau
se w
e do
*not
* res
tric
t the
ana
lysi
s on
wor
kers
who
logg
ed in
the
plat
form
in a
giv
en d
ay b
ut o
n al
l w
orke
rs (w
heth
er th
ey lo
gged
in o
r not
). **
* p<0
.01,
** p
<0.0
5, *
p<0.
1.
Wor
ker-
leve
l reg
ress
ions
C
all c
ente
r-le
vel
regr
essi
ons
Wor
ker-
leve
l reg
ress
ions
# ca
lls p
er h
our
# ca
lls p
er h
our
(Con
ditio
nal o
n Lo
ggin
g in
)
24
Demand. A second potential concern is that demand might be correlated with local weather,
as would be the case for a number of jobs (farmers, taxi drivers, physical sales positions).
Similarly, demand may be higher or lower the day after a local sports team plays. In our setting
(call centers), the demand our workers face is national, as calls from all over North America
are first aggregated and then distributed across call centers. Accordingly, we see that “number
of calls incoming to a call center” is uncorrelated with weather in that call center or with local
sports games the day before (Table 7 Column 3). The absence of confounding variation from
the demand side is a key advantage of a call-center setting. Finally, Table 6 (Column 3) shows
that the results do not change if we control for the “number of calls incoming.”
Pollution. A third potential concern is pollution. Pollution has been shown to reduce worker
productivity in call-center settings (Chang et al. 2016) and may correlate with rain. In Table
6, we show that the results hold if we control for temperature (which is related with daily
pollution) and for the level of three air pollutants (Nitric Oxide, Nitrogen Dioxide, and Ozone).
Unfortunately, the data are missing for one third of the sample and the sample size is therefore
smaller. Note that pollution is unlikely to be a confounder for our sports instrument.
Others. A final set of potential concerns (for the rain instrument only) is that rain might
have a direct effect on call-center working conditions independent of mood. Two possibilities
come to mind. First, that weather might affect productivity through distraction-on-the-job, i.e,
by looking out a window. Second, that forecasted weather might require changes in the work-
ers’ personal schedules, causing workers to waste time on the job rearranging their schedules (if
rain is forecasted, cancel the BBQ, and vice versa). To guard against the first concern, we have
obtained information about the prevalence of windows in different call center locations. Based
on our information, one third of the call centers have no windows at all while in the others all
workers see natural light. We check in Table A.14 whether workers in the call centers without
windows are less sensitive to rain-induced changes in mood (controlling for worker fixed ef-
fects). We find no significant effect. This indicates that the effect of mood on productivity may
not be affected by the presence of a window in the workplace, and suggests that the effect of
weather on mood is fully achieved in the time spent outside prior to reaching the workplace.17
17Table A.15 shows that the mood of workers without a window reacts to rain as much as the mood of workers
25
Note that the interaction terms in Table A.14 have large standard errors and that this evidence
is thus more suggestive than conclusive.
To assess the importance of the second concern (effect of forecasted weather), we regress
productivity at time t− 1 on rain at time t (which we call “lead rain.”) The idea is that if rain
is forecasted tomorrow, a worker might have to spend some time today in order to rearrange
her personal schedule. Table 7 Columns 6-9 show that the coefficient for “lead rain” is smaller
than the one for “contemporary rain” (roughly half of the magnitude) and is not statistically
significant. The effect of rain which we measure is thus likely not mediated by rescheduling. In
contrast, rain at time t significantly increases the number of calls per productive hour at t by 6.2
percentage points (reduced form).
7 Conclusions
A causal link between good mood and productivity, if established, would have profound con-
sequences for economic theory and for business practice. In this paper we make some initial
progress toward exploring this important link.
Conceptually, we provide a micro-foundation arguing that a worker’s productivity might be
decreasing in mood if better mood makes the worker more sociable, thus increasing the value
of work time spent socializing rather than working. Knowing whether, and how much, mood
decreases productivity is important for firms. Firms, we argue, routinely choose how much to
invest on workplace mood (by allowing more time off, longer breaks, office celebrations and
events, use of social media at work, etc) and in compensation, in order to maximize effort con-
ditional on retaining the worker. Improving workplace mood decreases effort, but it increases
worker’s welfare, which helps the firm retain workers. Thus knowing the effect of mood on
productivity may help the firm calibrate its compensation scheme.
At the correlational level, we present some of the first large-scale evidence of a correla-
tion between mood and productivity, with evidence coming from teams of sales representatives
from a large retailer, and also, separately, from individual-level data from call center work-
with a window.
26
ers. Provocatively, but consistent with some of the experimental psychology, this correlation is
negative.
We then leverage the call-center dataset to explore the causal effect of mood on their pro-
ductivity in the field through an IV strategy. The call center dataset is ideal to investigate the
causal effect of mood because variation in demand (a likely confounder of productivity) is na-
tional, and thus independent of our instruments – rain and sports events the day before. We find
that better mood actually decreases our call-center workers’ productivity. The effect of mood
is more muted for the subset of call-center workers whose compensation depends on produc-
tivity (high-powered incentives). This finding may be interpreted as indicating that introducing
monetary incentives crowds out the impact of mood.
We have ruled out a number of threats to the exclusion restriction: that our instruments
might affect productivity through higher demand, lower pollution, more hours at work, or more
time spent rearranging the workers’ personal schedules.
A number of caveats are in order. Our results concern short-term and individual mood
shifters only. In addition, we do not study worker retention empirically. Despite these limi-
tations, our research suggests that there are good theoretical reasons to believe that improving
workplace mood (by allowing more time off, longer breaks, etc.) entails some trade-offs, and
our empirical estimates have quantified the trade-off with productivity. It is quite possible that
firms have actually evaluated these trade-offs in setting compensation policy. So it may be
appropriate for the literatures in economics and management to include workplace mood as a
variable in the design of incentive schemes.
Finally, our findings relate to a specific workplace environment: call centers. The effects
of pro-sociality, which is our presumed operative channel, may impact differently depending
on how work is organized, e.g., whether the production function is individual, and if not, how
much teamwork there is.
27
References
[1] Bakker, Arnold B., et al. “Work engagement: An emerging concept in occupational health
psychology.” Work and Stress 22.3 (2008): 187-200.
[2] Bloom, Nicholas, et al. “Does working from home work? Evidence from a Chinese exper-
iment.” The Quarterly Journal of Economics 130.1 (2014): 165-218.
[3] Braga, Michela, Marco Paccagnella, and Michele Pellizzari. “Evaluating students’ evalua-
tions of professors.” Economics of Education Review 41 (2014): 71-88.
[4] Chang, T., Zivin, J. G., Gross, T., and Neidell, M. “The effect of pollution on worker
productivity: evidence from call-center workers in China”. American Economic Journal:
Applied Economics, forthcoming (2016)
[5] Connolly, Marie. “Here comes the rain again: Weather and the intertemporal substitution
of leisure.” Journal of Labor Economics 26.1 (2008): 73-100.
[6] Fry, Prem S. ”Affect and resistance to temptation.” Developmental psychology 11.4 (1975):
466.
[7] Keller, Matthew C., et al. “A warm heart and a clear head: The contingent effects of weather
on mood and cognition.” Psychological science 16.9 (2005): 724-731.
[8] Lee, Jooa Julia, Francesca Gino, and Bradley R. Staats. “Rainmakers: Why bad weather
means good productivity.” Journal of Applied Psychology 99.3 (2014): 504.
[9] Oswald, Andrew J., Eugenio Proto, and Daniel Sgroi. “Happiness and productivity.” Journal
of Labor Economics 33.4 (2015): 789-822.
[10] Pacheco-Unguetti, Antonia Pilar, and Fabrice BR Parmentier. “Happiness increases dis-
traction by auditory deviant stimuli.” British Journal of Psychology 107.3 (2016): 419-433.
Cunningham, Michael R. “What do you do when you’re happy or blue? Mood, expectan-
cies, and behavioral interest.” Motivation and emotion 12.4 (1988): 309-331.
28
[11] Park, Sangyoon. “Socializing at work: Evidence from a field experiment with manufac-
turing workers.” Mimeo (2016).
[12] Rothbard, Nancy P., and Steffanie L. Wilk. “Waking up on the right or wrong side of the
bed: Start-of-workday mood, work events, employee affect, and performance.” Academy
of Management Journal 54.5 (2011): 959-980.
[13] Tenney, E., J. Poole, and E. Diener. “Subjective well-being and organizational perfor-
mance.” Research in organizational behavior (2015).
[14] Tice, Dianne M., Ellen Bratslavsky, and Roy F. Baumeister. “Emotional distress regulation
takes precedence over impulse control: If you feel bad, do it!.” Journal of personality and
social psychology 80.1 (2001): 53.
29
A Tables and figures
Table A.1: Correlations between Productivity Measures
# calls per hour
# hours at work
% unprod-uctive time
Average call
duration
% of call duration "on hold"
# calls per hour 1# hours at work -0.0676* 1% unproductive time -0.0276* -0.0639* 1Average call duration -0.2182* 0.0410* 0.1275* 1% of call duration "on hold" -0.0607* 0.0027 0.1644* 0.2117* 1Average customer satisfaction 0.0656* -0.0126* -0.0344* -0.1614* -0.1791*Notes: Simple pairwise correlations. *p-value<0.05. N=Workers*Days
Table A.2: Mood and Days of the Week
(1) (2)
=1 if Weekend -0.041***(0.013)
=1 if Monday 0.043**(0.021)
=1 if Tuesday 0.040*(0.022)
=1 if Wednesday 0.034*(0.020)
=1 if Thursday 0.034(0.021)
=1 if Friday 0.068***(0.021)
=1 if Saturday 0.004(0.023)
Observations 35,425 35,425Rsquared 0.597 0.597Mean Dep Var 3.798 3.798Worker FE Month x year FE
Mood Score (1 to 5) [conditional on answering mood
question]
Notes: All regressions control for worker tenure. Robust standard errors clustered (twoway) at worker & call center*date level are presented in parenthesis. *** p<0.01, ** p<0.05, * p<0.1.
1
Tabl
eA
.3:M
ood
and
Prod
uctiv
ity,O
LS
Res
ults
with
Alte
rnat
ive
Clu
ster
ing
Stan
dard
err
ors
are
clus
tere
d (tw
oway
) at w
orke
r & c
all c
ente
r*w
eek
leve
l(1
)(2
)(3
)(4
)(5
)
Dep
. Var
# ca
lls p
er
hour
% un-
productive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
Pane
l A: C
ondi
tiona
l on
answ
erin
g M
ood
Que
stio
n
Moo
d Sc
ore
(1 to
5)
-0.0
61**
*-0
.002
***
0.08
1***
-0.0
010.
019
(0.0
15)
(0.0
01)
(0.0
19)
(0.0
01)
(0.0
28)
Obs
erva
tions
33,4
6133
,461
33,4
6133
,461
14,7
25Rs
quar
ed0.
819
0.31
20.
762
0.75
50.
269
Mea
n D
ep V
ar6.
956
0.09
817.
866
0.15
08.
140
Pane
l B: A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= Ba
d M
ood
Moo
d Sc
ore
(1 to
5)
-0.0
56**
*-0
.000
0.04
8***
-0.0
010.
004
(0.0
11)
(0.0
00)
(0.0
14)
(0.0
01)
(0.0
16)
Obs
erva
tions
73,2
6873
,268
73,2
6873
,268
32,3
05Rs
quar
ed0.
815
0.29
70.
752
0.74
30.
250
Mea
n D
ep V
ar7
0.09
607.
719
0.15
18.
049
Wor
ker F
E
Day
x M
onth
x y
ear F
E
Not
es:
All
regr
essi
ons c
ontr
ol fo
r wor
ker t
enur
e. S
tand
ard
erro
rs a
re cl
uste
red
(twow
ay) a
t w
orke
r & ca
ll ce
nter
*wee
k le
vel.
*** p
<0.0
1, **
p<0
.05,
* p<
0.1.
2
Tabl
eA
.4:M
ood
and
Prod
uctiv
ity,O
LS
Res
ults
with
Alte
rnat
ive
Cod
ing
(1)
(2)
(3)
(4)
(5)
Dep
. Var
# ca
lls p
er
hour
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
Pane
l C: A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= Ba
d M
ood
[Exh
aust
ed]
Moo
d Sc
ore
(1 to
5)
-0.0
71**
*-0
.000
0.06
4***
-0.0
010.
007
(0.0
13)
(0.0
00)
(0.0
16)
(0.0
01)
(0.0
19)
Pane
l D: A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= N
eutr
al M
ood
[So-
so]
Moo
d Sc
ore
(1 to
5)
-0.0
79**
*-0
.001
**0.
077*
**-0
.001
*0.
011
(0.0
15)
(0.0
00)
(0.0
18)
(0.0
01)
(0.0
23)
Obs
erva
tions
73,2
6873
,268
73,2
6873
,268
32,3
05M
ean
Dep
Var
70.
0960
7.71
90.
151
8.04
9W
orke
r FE
D
ay x
mon
th x
yea
r FE
N
otes
: A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at
wor
ker &
call
cent
er*d
ate
leve
l. **
* p<0
.01,
** p
<0.0
5, *
p<0.
1.
3
Tabl
eA
.5:S
umm
ary
Stat
istic
sof
Moo
dQ
uest
ion
fort
he“S
tore
sD
atas
et”
VAR
IABL
ES O
bs.
Mea
nS.
D.
Dai
ly W
orke
r Moo
d (N
=Wor
kers
*Day
s)
Cond
ition
al o
n lo
ggin
g in
to p
latfo
rm…
=1 if
wor
ker a
nsw
ers m
ood
ques
tion
20,1
66,0
18
0.52
0.50
Cond
ition
al o
n an
swer
ing
Moo
d Q
uest
ion
…%
who
feel
Fru
stra
ted
10,4
86,3
29
0.06
0.17
% w
ho fe
el E
xhau
sted
10,4
86,3
29
0.06
0.17
% w
ho fe
el S
oSo
10,4
86,3
29
0.12
0.24
% w
ho fe
el G
ood
10,4
86,3
29
0.38
0.40
% w
ho fe
el U
nsto
ppab
le10
,486
,329
0.
390.
40
10,4
86,3
29
3.99
1.11
Moo
d sc
ore
1 to
5
[1=
frus
trat
ed ..
. 5=u
nsto
ppab
le]
Not
es: U
pon
logg
ing
into
an
onlin
e pl
atfo
rm, w
orke
rs a
re a
sked
the
moo
d qu
estio
n;
"How
do
you
feel
toda
y: F
rust
rate
d, E
xhau
sted
, So
so, G
ood
or U
nsto
ppab
le?"
The
qu
estio
n is
ask
ed m
axim
um o
ne ti
me
per d
ay. T
he w
orke
r has
the
optio
n of
an
swer
ing
the
moo
d qu
estio
n or
skip
ping
it. W
e re
port
her
e th
e m
ood
scor
e co
nditi
onal
on
logg
ing
into
the
plat
form
and
ans
wer
ing
the
moo
d qu
estio
n (c
odin
g th
e no
ans
wer
as m
issi
ng).
Stor
es D
atas
et
4
Tabl
eA
.6:M
ood
and
Prod
uctiv
ity,O
LS
Res
ults
fort
he“S
tore
sD
atas
et”
(1)
(2)
(3)
(4)
(5)
(6)
[All
valu
es in
Col
1-6
are
in
00,
000
USD
] =>
Reve
nues
G
ross
m
argi
n Eb
itda
Reve
nues
pe
r Em
ploy
ee-
Hou
r
Gro
ss
Mar
gin
per
Empl
oyee
-H
our
Ebitd
a pe
r Em
ploy
ee-
Hou
r
-0.3
18**
*-0
.119
***
-0.0
58**
-0.0
61**
*-0
.025
***
-0.0
08(0
.075
)(0
.030
)(0
.026
)(0
.018
)(0
.006
)(0
.005
)
Obs
erva
tions
17,4
0717
,407
17,4
0717
,407
17,4
0717
,407
R-sq
uare
d0.
933
0.89
80.
632
0.83
40.
733
0.52
5M
ean
Dep
Var
Stor
e FE
Mon
th x
yea
r FE
Not
es: T
he o
utco
me
varia
bles
var
y at
the
stor
e * m
onth
leve
l. A
ll re
gres
sion
s con
trol
for t
he to
tal
num
ber o
f wor
kers
in th
e st
ore.
Rob
ust s
tand
ard
erro
rs cl
uste
red
at st
ore*
date
leve
l are
pre
sent
ed in
pa
rent
hesi
s. **
* p<0
.01,
** p
<0.0
5, *
p<0.
1. A
vera
ge m
ood
scor
e =a
vera
ge m
ood
in a
stor
e-m
onth
acr
oss
wor
kers
who
ans
wer
ed th
e m
ood
ques
tion.
'Gro
ss m
argi
ns' i
s sto
re re
venu
e m
inus
cost
of g
oods
sold
. 'E
bitd
a' is
stor
e's e
arni
ngs b
efor
e in
tere
st, t
axes
, dep
reci
atio
n an
d am
ortiz
atio
n fo
r the
mon
th. '
Reve
nue
per E
mpl
oyee
*Hou
r' is
Tot
.Rev
. div
ided
by
the
sum
of a
ll em
ploy
ee h
ours
. 'Eb
itda
per p
er
Empl
oyee
*Hou
r' is
Ebi
tda
divi
ded
by th
e su
m o
f all
empl
oyee
hou
rs.
Aver
age
Empl
oyee
M
ood
Scor
e (1
to 5
)
Hid
den
for a
nony
mity
reas
ons
5
Tabl
eA
.7:M
ood
and
Prod
uctiv
ity,O
LS
Res
ults
byM
ood
(1)
(2)
Ass
umpt
ion
on M
ood
Que
stio
n =>
Moo
d by
moo
d w
ith
"no
answ
er"
as
omitt
ed g
roup
Cond
ition
al o
n an
swer
ing
Moo
d Q
uest
ion
&
answ
erin
g Re
ason
0.12
8(0
.082
)-0
.006
(0.0
58)
-0.0
96*
(0.0
54)
-0.1
08*
(0.0
55)
-0.2
39**
*(0
.051
)
0.05
4***
(0.0
16)
-0.7
39**
*(0
.167
)
Obs
erva
tions
34,1
8831
,709
Mea
n D
epen
dent
Var
iabl
e7
7W
orke
r FE
Day
x m
onth
x y
ear F
E
Dep
ende
nt v
aria
ble
= #
calls
per
hou
r
Not
es:
All
regr
essi
ons c
ontr
ol fo
r wor
ker t
enur
e. S
tand
ard
erro
rs a
re cl
uste
red
(twow
ay) a
t wor
ker &
call
cent
er*d
ate
leve
l. **
* p<0
.01,
** p
<0.0
5, *
p<0.
1.
=1 if
Moo
d=Fr
ustr
ated
(sco
re 5
)
=1 if
Moo
d=Ex
haus
ted
(sco
re 4
)
=1 if
Moo
d=So
so (s
core
3)
=1 if
Moo
d=G
ood
(sco
re 2
)
=1 if
Moo
d=U
nsto
ppab
le (s
core
1)
Moo
d sc
ore
if m
ain
reas
on e
xpla
inin
g th
e m
ood
is W
ork-
rela
ted
Moo
d sc
ore
if m
ain
reas
on e
xpla
inin
g th
e m
ood
is n
ot W
ork-
rela
ted
6
Tabl
eA
.8:M
ood
and
Wea
ther
/Spo
rtga
mes
;Fir
stSt
age
Res
ults
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
=1 if
Rai
ny D
ay-0
.031
***
-0.0
38**
*(0
.010
)(0
.009
)Pr
ecip
itatio
n-0
.001
***-
0.00
1***
-0.0
01**
*-0.
001*
**(0
.000
)(0
.000
)(0
.000
)(0
.000
)Sn
owfa
ll-0
.001
0.00
0(0
.001
)(0
.001
)M
inim
um te
mpe
ratu
re0.
001
-0.0
02(0
.002
)(0
.002
)M
axim
um te
mpe
ratu
re0.
001
0.00
1(0
.001
)(0
.001
)0.
008
0.04
2**
(0.0
15)
(0.0
17)
Fsta
t 1st
sta
ge9.
393
8.53
73.
002
0.28
016
.25
8.94
14.
497
5.95
7
Obs
erva
tions
33,4
6133
,461
33,4
6133
,525
73,2
6873
,268
73,2
6873
,341
R-sq
uare
d0.
599
0.59
90.
600
0.60
30.
684
0.68
40.
684
0.68
6M
ean
Dep
Var
3.80
13.
801
3.80
13.
801
2.28
92.
289
2.28
92.
289
Wor
ker F
E
M
onth
x y
ear F
E
D
ay o
f the
wee
k FE
Day
x m
onth
x y
ear F
E
H
isto
ric ra
in
Dep
ende
nt v
aria
ble
=Moo
d sc
ore
(1 to
5)
Pane
l A: C
ondi
tiona
l on
answ
erin
g M
ood
Que
stio
n
Pane
l B: A
ssum
ing
that
not
an
swer
ing
Moo
d Q
uest
ion
= Ba
d M
ood
Not
es: F
irst s
tage
IV re
gres
sion
s. A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at w
orke
r & ca
ll ce
nter
*dat
e le
vel.
His
toric
rain
= hi
stor
ic a
mou
nt o
f rai
n on
that
day
(ave
rage
in
the
past
5 y
ears
). **
* p<0
.01,
** p
<0.0
5, *
p<0.
1.
=1 if
Spo
rt T
eam
Pla
yed
in t-
1
7
Tabl
eA
.9:M
ood
and
Prod
uctiv
ity,I
VR
esul
tsw
ithA
ltern
ativ
eC
odin
g
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Asu
mpt
ion
on m
ood
ques
tion
=>
Dep
. Var
# ca
lls p
er
hour
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
# ca
lls p
er
hour
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
Moo
d Sc
ore
(1 to
5)
-1.9
33**
0.05
3**
0.89
60.
049*
-2.5
56-2
.598
**0.
072*
*1.
204
0.06
5*-3
.479
(0.8
36)
(0.0
24)
(0.8
17)
(0.0
29)
(2.1
95)
(1.1
35)
(0.0
34)
(1.0
93)
(0.0
39)
(3.1
14)
Obs
erva
tions
73,2
6873
,268
73,2
6873
,268
32,3
0573
,268
73,2
6873
,268
73,2
6832
,305
Mea
n D
epen
dent
Var
iabl
e7
0.09
607.
719
0.15
18.
049
70.
0960
7.71
90.
151
8.04
9Fs
tat 1
st st
age
17.1
217
.12
17.1
217
.12
17.1
215
.08
15.0
815
.08
15.0
815
.08
Wor
ker F
E
M
onth
x y
ear F
E
D
ay o
f the
wee
k FE
His
toric
rain
Inst
rum
ent =
Rai
ny D
ay
(Ass
umin
g th
at n
ot a
nsw
erin
g M
ood
Que
stio
n =
Bad
Moo
d [E
xhau
sted
])(A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= N
eutr
al M
ood
[So-
so])
Not
es: S
econ
d st
age
IV re
gres
sion
s. A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at w
orke
r & ca
ll ce
nter
*dat
e le
vel.
His
toric
rain
= hi
stor
ic a
mou
nt o
f rai
n on
that
day
(ave
rage
in th
e pa
st 5
yea
rs).
*** p
<0.0
1, **
p<0
.05,
* p<
0.1.
8
Tabl
eA
.10:
Moo
dan
dPr
oduc
tivity
,IV
Res
ults
with
Alte
rnat
ive
Clu
ster
ing
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Dep
. Var
# ca
lls
per h
our
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
# ca
lls
per h
our
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
Inst
rum
ent =
>
Pane
l A: C
ondi
tiona
l on
answ
erin
g M
ood
Que
stio
nM
ood
Scor
e (1
to 5
)-1
.760
*0.
053*
1.27
10.
060
-4.8
85-
--
--
(0.9
05)
(0.0
32)
(1.0
17)
(0.0
42)
(7.0
95)
--
--
-
Obs
erva
tions
33,4
6133
,461
33,4
6133
,461
14,7
25-
--
--
Mea
n D
ep V
ar7
0.09
607.
719
0.15
18.
049
--
--
-Fs
tat 1
st st
age
8.5
8.5
8.5
8.5
8.5
--
--
-
Pane
l B: A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= Ba
d M
ood
Moo
d Sc
ore
(1 to
5)
-1.5
39**
0.04
3**
0.71
30.
039
-2.0
20-1
.485
*0.
008
-1.0
060.
069
-0.5
22(0
.639
)(0
.020
)(0
.699
)(0
.025
)(1
.769
)(0
.883
)(0
.019
)(1
.164
)(0
.042
)(1
.604
)
Obs
erva
tions
73,2
6873
,268
73,2
6873
,268
32,3
0573
,268
73,2
6873
,268
73,2
6832
,305
Mea
n D
ep V
ar7
0.09
607.
719
0.15
18.
049
70.
0960
7.71
90.
151
8.04
9Fs
tat 1
st st
age
1515
1515
155.
75.
75.
75.
75.
7W
orke
r FE
Mon
th x
yea
r FE
D
ay o
f the
wee
k FE
D
ay x
mon
th x
yea
r FE
H
isto
ric ra
in
=1 if
Rai
ny D
ay=1
if S
port
Tea
m P
laye
d in
t-1
Not
es: S
econ
d st
age
IV re
gres
sion
s. A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at w
orke
r & ca
ll ce
nter
*dat
e le
vel.
His
toric
rain
= hi
stor
ic a
mou
nt o
f rai
n on
that
day
(ave
rage
in th
e pa
st 5
yea
rs).
*** p
<0.0
1, **
p<0
.05,
* p<
0.1.
9
Tabl
eA
.11:
Moo
dan
dPr
oduc
tivity
,IV
Res
ults
with
Ext
raC
ontr
ols
(Ass
umin
g th
at n
ot a
nsw
erin
g M
ood
Que
stio
n =
Bad
Moo
d)(1
)(2
)(3
)(4
)(5
)(6
)(7
)(8
)(9
)(1
0)(1
1)(1
2)(1
3)(1
4)(1
5)(1
6)(1
7)(1
8)(1
9)(2
0)
Moo
d sc
ore
(1 to
5)
0.04
1**
0.04
1**
0.04
0**
0.03
9**
0.01
60.
729
0.73
10.
806
0.76
10.
431
0.04
0*0.
040*
0.04
3*0.
049*
*0.
042*
-2.0
13-2
.013
-1.9
97-2
.047
-4.2
55(0
.019
)(0
.019
)(0
.019
)(0
.019
)(0
.019
)(0
.648
)(0
.648
)(0
.644
)(0
.622
)(0
.792
)(0
.023
)(0
.023
)(0
.023
)(0
.023
)(0
.024
)(1
.722
)(1
.706
)(1
.717
)(1
.758
)(2
.684
)N
o. h
ours
at w
ork
0.00
4***
0.00
4***
0.00
4***
0.00
4***
-0.0
20**
-0.0
20**
-0.0
20**
-0.0
24*
-0.0
00-0
.000
-0.0
00-0
.000
0.00
00.
000
-0.0
00-0
.040
(0.0
00)
(0.0
00)
(0.0
00)
(0.0
00)
(0.0
10)
(0.0
10)
(0.0
10)
(0.0
12)
(0.0
00)
(0.0
00)
(0.0
00)
(0.0
00)
(0.0
19)
(0.0
19)
(0.0
19)
(0.0
45)
No.
inco
min
g ca
lls
-0.0
00-0
.000
-0.0
000.
036*
**0.
036*
**0.
028*
0.00
1**
0.00
1**
0.00
10.
008
0.00
80.
019
(0.0
00)
(0.0
00)
(0.0
00)
(0.0
13)
(0.0
13)
(0.0
16)
(0.0
00)
(0.0
00)
(0.0
01)
(0.0
18)
(0.0
18)
(0.0
32)
Tem
pera
ture
0.00
0-0
.000
0.00
00.
000
-0.0
00-0
.000
0.00
00.
002
(0.0
00)
(0.0
00)
(0.0
00)
(0.0
01)
(0.0
00)
(0.0
00)
(0.0
01)
(0.0
02)
Nitr
ic o
xide
0.13
8**
2.14
1-0
.145
3.99
7(0
.062
)(2
.856
)(0
.100
)(7
.218
)N
itrog
en d
ioxi
de0.
000
-0.0
020.
000
-0.0
01(0
.000
)(0
.004
)(0
.000
)(0
.010
)O
zone
-0.0
000.
003
0.00
0-0
.016
(0.0
00)
(0.0
05)
(0.0
00)
(0.0
15)
Obs
erva
tions
73,2
6873
,268
73,2
6873
,268
46,7
6173
,268
73,2
6873
,268
73,2
6846
,761
73,2
6873
,268
73,2
6873
,268
46,7
6132
,305
32,3
0532
,305
32,3
0517
,922
Mea
n D
ep V
ar0.
0960
0.09
600.
0960
0.09
600.
0960
7.71
97.
719
7.71
97.
719
7.71
90.
151
0.15
10.
151
0.15
10.
151
8.04
98.
049
8.04
98.
049
8.04
9W
orke
r FE
Mon
th x
yea
r FE
Day
of t
he w
eek
FE
H
isto
ric ra
in
Not
es: S
econ
d st
age
IV re
gres
sion
s. A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at w
orke
r & ca
ll ce
nter
*dat
e le
vel.
His
toric
rain
= hi
stor
ic
amou
nt o
f rai
n on
that
day
(ave
rage
in th
e pa
st 5
yea
rs).
*** p
<0.0
1, **
p<0
.05,
* p<
0.1.
% o
f cal
l dur
atio
n "o
n ho
ld"
Aver
age
call
dura
tion
(min
utes
)%
unp
rodu
ctiv
e tim
eAv
erag
e cu
stom
er sa
tisfa
ctio
n
Inst
rum
ent =
Rai
ny D
ay
10
Tabl
eA
.12:
Moo
dan
dPr
oduc
tivity
,IV
Res
ults
with
Con
trol
san
dA
ltern
ativ
eC
odin
g
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Moo
d sc
ore
(1 to
5)
-1.9
39**
-1.9
25**
-1.8
09**
-1.8
31**
-1.7
46**
-2.6
13**
-2.5
93**
-2.4
37**
-2.5
44**
-2.4
92**
(0.8
27)
(0.8
10)
(0.7
94)
(0.8
03)
(0.8
41)
(1.1
27)
(1.1
04)
(1.0
81)
(1.1
40)
(1.2
63)
No.
hou
rs a
t wor
k-0
.085
***
-0.0
84**
*-0
.085
***
-0.0
87**
*-0
.086
***
-0.0
85**
*-0
.086
***
-0.0
90**
*(0
.017
)(0
.016
)(0
.016
)(0
.021
)(0
.017
)(0
.016
)(0
.017
)(0
.022
)N
o. in
com
ing
calls
0.
044*
**0.
044*
**0.
052*
**0.
044*
**0.
044*
**0.
052*
**(0
.012
)-0
.012
(0.0
14)
(0.0
11)
(0.0
12)
(0.0
14)
Tem
pera
ture
0.00
00.
001
0.00
00.
001*
(0.0
00)
(0.0
00)
(0.0
00)
(0.0
00)
Nitr
ic o
xide
(val
ue)
1.43
31.
351
(2.2
65)
(2.3
38)
Nitr
ogen
dio
xide
(val
ue)
0.00
30.
004
(0.0
03)
(0.0
03)
Ozo
ne (v
alue
) -0
.006
-0.0
04(0
.004
)(0
.004
)
Obs
erva
tions
73,2
6873
,268
73,2
6873
,268
46,7
6173
,268
73,2
6873
,268
73,2
6846
,761
Wor
ker F
E7
77
77
77
77
7M
onth
x y
ear F
E
D
ay o
f the
wee
k FE
His
toric
rain
Dep
ende
nt v
aria
ble
= #
calls
per
hou
r
Not
es: S
econ
d st
age
IV re
gres
sion
s. A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at w
orke
r & ca
ll ce
nter
*dat
e le
vel.
His
toric
rain
= hi
stor
ic a
mou
nt o
f rai
n on
that
day
(ave
rage
in th
e pa
st 5
yea
rs).
*** p
<0.0
1, **
p<0
.05,
* p<
0.1.
Inst
rum
ent =
Rai
ny D
ay
(Ass
umin
g th
at n
ot a
nsw
erin
g M
ood
Que
stio
n =
Bad
Moo
d [E
xhau
sted
])(A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= N
eutr
al M
ood
[So-
so])
11
Tabl
eA
.13:
Moo
dan
dPr
oduc
tivity
forW
orke
rsL
ivin
gN
eart
heO
ffice
,IV
Res
ults
(1)
(2)
(3)
(4)
(5)
Dep
. Var
# ca
lls
per h
our
% u
n-pr
oduc
tive
time
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (1
to 1
0)
Inst
rum
ent =
>
Pane
l A: C
ondi
tiona
l on
answ
erin
g M
ood
Que
stio
nM
ood
Scor
e (1
to 5
)-1
.140
*0.
028
2.52
6*-0
.002
1.76
3(0
.637
)(0
.030
)(1
.296
)(0
.032
)(2
.604
)
Obs
erva
tions
2,90
62,
906
2,90
62,
906
914
Fsta
t 1st
stag
e6.
481
6.48
16.
481
6.48
16.
481
Pane
l B: A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= Ba
d M
ood
Moo
d Sc
ore
(1 to
5)
-1.9
380.
008
0.89
40.
008
-0.6
49(1
.354
)(0
.036
)(1
.620
)(0
.039
)(1
.681
)
Obs
erva
tions
7,14
07,
140
7,14
07,
140
2,42
8Fs
tat 1
st st
age
3.11
63.
116
3.11
63.
116
3.11
6W
orke
r FE
M
onth
x y
ear F
E
Day
of t
he w
eek
FE
His
toric
rain
Sam
ple =
wor
kers
livi
ng <
5km
from
wor
k
=1 if
Rai
ny D
ay
Not
es: S
econ
d st
age
IV re
gres
sion
s. A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at w
orke
r & ca
ll ce
nter
*dat
e le
vel.
His
toric
rain
= hi
stor
ic
amou
nt o
f rai
n on
that
day
(ave
rage
in th
e pa
st 5
yea
rs).
*** p
<0.0
1, **
p<0
.05,
* p<
0.1.
12
Tabl
eA
.14:
Moo
dan
dPr
oduc
tivity
byA
cces
sto
aW
indo
w,I
VR
esul
ts
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Dep
. Var
# ca
lls p
er
hour
%
unpr
oduc
tiv
e tim
e
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (s
core
1 to
10
)
# ca
lls
per h
our
%
unpr
odu
ctiv
e tim
e
Aver
age
call
dura
tion
(min
utes
)
% o
f cal
l du
ratio
n "o
n ho
ld"
Aver
age
cust
omer
sa
tisfa
ctio
n (s
core
1 to
10
)
Pane
l A: C
ondi
tiona
l on
answ
erin
g M
ood
Que
stio
n
Moo
d Sc
ore
(1 to
5)
-1.9
120.
076
1.29
70.
032
-14.
797
(1.2
68)
(0.0
47)
(1.3
31)
(0.0
45)
(80.
827)
Moo
d Sc
ore
(1 to
5)*
Win
dow
0.48
6-0
.072
-0.0
820.
089
12.2
58(1
.570
)(0
.052
)(1
.476
)(0
.073
)(7
9.01
3)
Obs
erva
tions
33,4
6133
,461
33,4
6133
,461
14,7
25p-
valu
e (M
ood
+ M
ood*
Win
dow
=0)
0.18
90.
899
0.14
40.
054
0.37
5
Pane
l B: A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= Ba
d M
ood
Moo
d Sc
ore
(1 to
5)
-1.4
52*
0.05
4**
0.43
10.
027
-2.4
81-1
.970
0.00
6-1
.130
0.08
80.
305
(0.7
59)
(0.0
24)
(0.7
72)
(0.0
26)
(3.6
95)
(1.4
00)
(0.0
27)
(1.2
82)
(0.0
59)
(5.4
45)
Moo
d Sc
ore
* Win
dow
-0.4
86-0
.062
*1.
568
0.06
70.
937
4.67
60.
020
1.19
9-0
.186
-3.3
59(1
.434
)(0
.034
)(1
.294
)(0
.058
)(3
.861
)(4
.262
)(0
.077
)(3
.010
)(0
.157
)(1
6.85
0)
Obs
erva
tions
73,2
6873
,268
73,2
6873
,268
32,3
0573
,268
73,2
6873
,268
73,2
6832
,305
p-va
lue (
Moo
d +
Moo
d*W
indo
w=0
)0.
126
0.74
60.
056
0.07
90.
079
0.40
00.
640
0.97
40.
402
0.79
3W
orke
r FE
Mon
th x
yea
r FE
D
ay o
f the
wee
k FE
D
ay x
mon
th x
yea
r FE
His
toric
rain
on
that
day
Inst
rum
ents
= R
ainy
Day
& R
ainy
Day
*Win
dow
Inst
rum
ents
= T
eam
Pla
yed
& T
eam
Pla
yed*
Win
dow
Not
es: S
econ
d st
age
IV re
gres
sion
s. W
indo
w=1
if th
e ca
ll-ce
nter
has
a w
indo
w (n
atur
al li
ght).
All
regr
essi
ons c
ontr
ol fo
r wor
ker t
enur
e. S
tand
ard
erro
rs a
re cl
uste
red
(twow
ay) a
t wor
ker &
call
cent
er*d
ate
leve
l. H
isto
ric ra
in=
hist
oric
am
ount
of r
ain
on th
at d
ay (a
vera
ge in
the
past
5 y
ears
). **
* p<0
.01,
** p
<0.0
5, *
p<0.
1.
13
Tabl
eA
.15:
Wea
ther
and
Moo
d,Fi
rsts
tage
byA
cces
sto
Win
dow
s
(1)
(2)
=1 if
Rai
ny D
ay-0
.025
**-0
.033
***
(0.0
11)
(0.0
10)
=1 if
Rai
ny D
ay *
Win
dow
-0.0
39-0
.028
(0.0
30)
(0.0
25)
Fsta
t 1st
sta
ge5.
572
8.90
5
Obs
erva
tions
73,2
6833
,461
Test
Rai
n +
Rai
n*W
indo
w =
0 (p
-val
ue)
0.00
70.
017
Wor
ker F
E
M
onth
x y
ear F
E
D
ay o
f the
wee
k FE
His
toric
rain
Dep
ende
nt v
aria
ble
=Moo
d sc
ore
(1 to
5)
Pane
l A:
Con
ditio
nal o
n an
swer
ing
Moo
d Q
uest
ion
Pane
l B: A
ssum
ing
that
not
ans
wer
ing
Moo
d Q
uest
ion
= Ba
d M
ood
Not
es: F
irst s
tage
IV re
gres
sion
s. W
indo
w=1
if th
e ca
ll-ce
nter
has
a w
indo
w (n
atur
al
light
). A
ll re
gres
sion
s con
trol
for w
orke
r ten
ure.
Sta
ndar
d er
rors
are
clus
tere
d (tw
oway
) at w
orke
r & ca
ll ce
nter
*dat
e le
vel.
His
toric
rain
= hi
stor
ic a
mou
nt o
f rai
n on
that
day
(ave
rage
in th
e pa
st 5
yea
rs).
*** p
<0.0
1, **
p<0
.05,
* p<
0.1.
14
B Calculations for Section 4
The first order conditions (1) specialize to:
√w√e= m
1√1− e
,
so that
1− ee
= m2 1w
1e
= 1+m2 1w
e∗ (w,m) =w
w+m2 .
We see that effort is between zero and one, it is increasing in wages and decreasing in mood.
Substituting into the agent’s utility we get the indirect utility function:
U (w,m) = u (we∗ (w,m))+ g (1− e∗ (w,m) ,m)
=
√w
ww+m2 +m
√1− w
w+m2
=
√w2
w+m2 +m
√m2
w+m2
=w√
w+m2+
m2√
w+m2
=√
w+m2
Denote m2 = µ and K (√
µ) = k (µ) . The firm’s problems rewrites as:
maxw,m
(1−w)w
w+ µ− k (µ)
s.t. w+ µ ≥Ω2.
First case: constraint Ω is not binding
15
Suppose first the constraint is not binding. Then the problem is:
maxw,m
(1−w)w
w+ µ− k (µ) .
Derivatives with respect to the optimization variables:
w.r.t. µ : − (1−w)w1
(w+ µ)2 − k′ (µ)
w.r.t. w : − ww+ µ
+(1−w)
((w+ µ)−w
(w+ µ)2
)
= − ww+ µ
+(1−w)
(µ
(w+ µ)2
)
=−w (w+ µ)+ (1−w)µ
(w+ µ)2
=−w2−2wµ + µ
(w+ µ)2
=−w2−2wµ−µ2 + µ2 + µ
(w+ µ)2
=− (w+ µ)2 + µ2 + µ
(w+ µ)2
= −1+µ2 + µ
(w+ µ)2
In this case the derivative with respect to µ is always negative, so it is optimal to set µ as
small as possible, which we now assume is 1. The derivative with respect to w is decreasing in
w, so we have concavity, and for µ = 1 the derivative equals
−1+2
(w+ 1)2 .
Setting this equal to zero yields:
2
(w+ 1)2 = 1
2 = (w+ 1)2
w∗ =√
2−1 = 0.41421.
16
All of this makes sense. No sense in investing in mood since that only decreases effort and costs
money. Then you get an optimal wage.
To check whether the constraint is binding, substitute this value in the constraint:
Ω2 ≤ w∗+ µ∗ =√
2.
When Ω2 >√
2 the constraint is binding.
Second case: constraint Ω is binding
In this case the firm’s problem is:
maxw
(1−w)w
Ω2 − k(Ω2−w
)=
1Ω2 max
w(1−w)w−Ω2k
(Ω2−w
).
Since the term (1−w)w is increasing in w over (0.41421,0.5) and −Ω2k(Ω2−w
)is increas-
ing in w for all w’s, w∗ will never be below 1/2. This means that µ∗ = Ω2−w∗ is never above
Ω2−1/2.
The derivative with respect to w reads:
1−2w+Ω2k′(Ω2−w
).
If k (·) is a convex function this expression is decreasing in w so the first order conditions
identify an interior maximum w∗ (Ω) . Rewrite the first order conditions as follows:
(3) 1−2w = −Ω2k′(Ω2−w
).
Both sides of this equation are negative. Since −k′ (·) is a decreasing function, the right hand
side is an increasing function of w. As Ω2 increases to (Ω′)2 the right hand side becomes larger
in absolute value, and thus more negative, both because it gets scaled up by Ω2 and because k
is convex. Therefore w∗ (Ω′) > w∗ (Ω) .
17
Suppose K (m) = mα so that k (m) = mα/2 and k′ (m) = α
2 m(α/2)−1. Then (3) reads:
1−2w = −Ω2 (Ω2−w)α
2−1.
Now use the implicit function theorem:
∂
∂ Ω(1−2w∗ (Ω)) = − ∂
∂ ΩΩ2 (Ω2−w∗ (Ω)
)α
2−1
−2w∗′ (Ω) = −[2Ω(Ω2−w∗ (Ω)
)α
2−1+(2Ω−w∗′ (Ω))
(α
2−1)
Ω2 (Ω2−w∗ (Ω))α
2−2]
2w∗′ (Ω) = 2Ω(Ω2−w∗ (Ω)
)α
2−1+(2Ω−w∗′ (Ω))
(α
2−1)
Ω2 (Ω2−w∗ (Ω))α
2−2
2w∗′ (Ω) = Ω(Ω2−w∗ (Ω)
)α
2−1[2+(2Ω−w∗′ (Ω))
(α
2−1)
Ω(Ω2−w∗ (Ω)
)−1]
2w∗′ (Ω) = Ω (µ∗)α
2−1[
2+(2Ω−w∗′ (Ω))(
α
2−1) Ω
µ∗
]2
w∗′ (Ω)
Ω(µ∗)1−α
2 = 2+(2Ω−w∗′ (Ω))(
α
2−1) Ω
µ∗
2w∗′ (Ω)
Ω(µ∗)1−α
2 = 2+ 2(
α
2−1)Ω2
µ∗−w∗′ (Ω)
(α
2−1) Ω
µ∗
2w∗′ (Ω)
Ω(µ∗)1−α
2 +w∗′ (Ω)(
α
2−1) Ω
µ∗= 2+ 2
(α
2−1)Ω2
µ∗
w∗′ (Ω)
[2Ω
(µ∗)1−α
2 +(
α
2−1) Ω
µ∗
]= 2
(1+
(α
2−1)Ω2
µ∗
)
So finally
w∗′ (Ω) = 2(
1+(
α
2−1)Ω2
µ∗
).
1[2Ω (µ∗)1−α
2 +(
α
2 −1) Ω
µ∗
] .
18
Now, we are interested in:
∂ µ∗ (Ω)
∂ Ω=
∂
∂ Ω(Ω2−w∗ (Ω)
)= 2Ω−
2(
1+(
α
2 −1) Ω2
µ∗
)2Ω (µ∗)1−α
2 +(
α
2 −1) Ω
µ∗
= 2
Ω−
(1+
(α
2 −1) Ω2
µ∗
)2Ω (µ∗)1−α
2 +(
α
2 −1) Ω
µ∗
= 2
Ω[
2Ω (µ∗)1−α
2 +(
α
2 −1) Ω
µ∗
]−(
1+(
α
2 −1) Ω2
µ∗
)2Ω (µ∗)1−α
2 +(
α
2 −1) Ω
µ∗
= 22 (µ∗)1−α
2 +(
α
2 −1) Ω2
µ∗ −1−(
α
2 −1) Ω2
µ∗
2Ω (µ∗)1−α
2 +(
α
2 −1) Ω
µ∗
= 22 (µ∗)1−α
2 −12Ω (µ∗)1−α
2 +(
α
2 −1) Ω
µ∗
So ∂ µ∗(Ω)∂ Ω > 0 if and only if:
2 (µ∗)1−α
2 > 1
2 (µ∗)2−α
2 > 1
2(
1µ∗
)α−22
> 1(1
µ∗
)α−22
>12
1µ∗
>1
22
α−2
µ∗ < 2
2α−2
This inequality is never true for α = ∞ and is always true for all values of µ∗ if α = 2.
19