Socializing at Work:
Evidence from a Field Experiment with Manufacturing Workers∗
- Job Market Paper Sangyoon Park†
Northwestern University
December 7, 2015
Click Here for Most Recent Version
Abstract
This paper examines how working alongside friends affects employee productivity, how much
employees value socializing at work, and how these results are heterogeneous with respect to an
employee’s non-cognitive skills. I designed and implemented a field experiment at a seafoodprocessing plant in Vietnam in which, each day, workers were randomly assigned to work stations within the processing room. This generated exogenous variation in their proximity to their
friends: sometimes, but not always, a worker’s friend was assigned to work right next to her.
This paper presents three main findings. First, worker productivity, as measured by the weight
of seafood processed per hour, declines when a friend is close enough to socialize with. Second,
combining the experimental results with pre-experiment data, I find that when workers chose
their own work stations, workers were willing to forgo 6% of their wage to be close enough to
socialize with friends; their wage is a function of their output and they earn less when working
alongside friends, yet systematically choose to do so. Using data on worker productivity before
and after the intervention, I find that the random assignment policy led to a 5% increase in the
firm’s productivity. Finally, based on assessments of workers’ Big Five personality factors administered prior to the experiment, I show that workers who are higher on the conscientiousness
scale show smaller declines in productivity when working alongside a friend and also place less
value on socializing at work. Socializing behaviors account for two-thirds of the 7% productivity
gap observed between high- and low-conscientiousness workers.
JEL Codes: C93, J24, J31, M54, O14.
∗
I am extremely grateful to Seema Jayachandran, Lori Beaman, Jonathan Guryan, and Cynthia Kinnan for
encouragement and support throughout this project. I am also thankful to Stephanie Chapman, Erika Deserranno,
Mitchell Hoffman, Aanchal Jain, Lee Lockwood, Samir Mamadehussene, Matthew Notowidigdo, Uta Schoenberg,
Dean Yang, Noam Yuchtman, Ariel Zimran and seminar participants at the Northwestern Applied Micro Seminar,
U-Michigan Development Day, and J-PAL Conference on Field Experiments in Labor Economics and Social Policies
for providing insightful comments and suggestions. I also thank Nguyen Thi Diem for providing excellent research
assistance. This work was made possible in part by a research grant from the Equality Development and Globalization
Studies (EDGS) program, funded by the Rajawali Foundation in Indonesia, the Buffett Center Summer Research
Grant, and Graduate Research Grant at Northwestern University. All errors are my own.
†
Email: [email protected] Web: sites.northwestern.edu/spa136
1
1
Introduction
Human resource allocation policies have substantial impacts on employee productivity (Mas and
Moretti, 2009; Bandiera et al., 2010; Hjort, 2014; Bloom et al., 2015). The question of how working alongside friends affects the productivity of a worker is especially important for developing
economies. Evidence suggests that friends are essential providers of information (Banerjee et al.,
2013), job referrals (Beaman and Magruder, 2012), and informal credit (Fafchamps and Lund,
2003) for individuals in less developed regions. In my sample of seafood-processing workers in Binh
Thuan, Vietnam, 75% reported to have learned about their current job from their friends and at
least 80% reported to have recently engaged in lending and borrowing with their friends at work.
While the relevance of friendships in the workplace is convincing, evidence on the causal effect of
friendships on job performance is limited.
This paper investigates how working alongside friends affects employee productivity, how much
employees value socializing at work, and how these effects differ with respect to an employee’s
personality factors, or non-cognitive skills. I designed and implemented a field experiment that
changed the firm’s policy for assigning work stations to workers in a seafood-processing plant in
Vietnam. The new policy randomly varied where workers were physically standing during the
work day. I exploit this exogenous variation to estimate the causal effect of having socially tied
coworkers nearby on worker productivity. I then estimate how much workers value socializing with
friends by combining these results with pre-intervention data on workers’ choices of whether to
work alongside their friends. Finally, I examine heterogeneity in these results using assessments on
workers’ personality factors.
Studies indicate that peers influence a worker’s productivity largely through emotional or psychological mechanisms. For instance, Kandel and Lazear (1992) suggest that shame or social norms
can serve as motives for workers to exert greater effort in the presence of their peers. Recent empirical studies find evidence of social pressure or social incentives affecting worker productivity
(Falk and Ichino, 2006; Mas and Moretti, 2009; Bandiera et al., 2005, 2010; Amodio and MartinezCarrasco, 2015). Working with friends may create a sense of competition or help cope with boredom,
leading to greater motivation. Alternatively, the presence of a friend could also lead to goofing off
2
during work and become less productive. I complement previous studies by allowing heterogeneity
in these underlying mechanisms that are driven by an individual’s non-cognitive skills.
A substantial body of evidence documents how non-cognitive skills relate to job performance
(Barrick et al., 1998; Schmidt and Hunter, 1998; Callen et al., 2015) and labor market outcomes
(Heckman and Rubinstein, 2001; Mueller and Plug, 2006; Heckman et al., 2006). In particular,
conscientiousness, one of the Big Five personality domains that measures a person’s tendency to
exert self-control, is consistently shown to have strong relevance to job performance (Salgado, 1997;
Schmidt and Hunter, 1998; Barrick et al., 1998, 2003; Hogan and Holland, 2003). Yet, the channels
through which one’s conscientiousness influences one’s job performance are less well understood;
a particular skill may help one thrive in certain work tasks and environments but be a liability
in others. Identifying these channels can be useful in guiding public policies on human capital
development and firms’ policies on human resources management.
This field experiment was conducted in collaboration with the management of a seafoodprocessing plant. The plant only hires female processing workers. Workers fillet fish in rooms
while standing at work tables. Compensation is a combination of a fixed daily wage plus a piece
rate based on each worker’s individual output. Prior to the experiment, workers could choose their
work positions at the start of the workday. I focus on friendship ties between workers in the same
processing room.
During the five-month experiment period (August 2014 to December 2014), processing workers
were randomly assigned to work at different positions each day. This experiment created random
variation in the presence of friends at various levels of spatial proximity to the worker. The outcome
data on daily worker productivity and the data on work position were collected from the firm’s
employee records database. Prior to the experiment, data on each worker’s friendship ties at the
plant, personality characteristics, and other background data were collected via a baseline survey.
As the first main result, I find that a worker is on average 6 percent less productive when a
friend is working alongside her. Yet, I find no effect when friends are working at positions further
away from her (for example, at the same table but not immediately adjacent). These findings
suggest that the presence of friends affects worker productivity mainly through socializing, such
3
as engaging in chit-chats. Since workers are partially paid based on individual performance, the
productivity loss implies an average 4 percent decline in the daily wage when a friend is present
alongside a worker.
In the second main result, guided by a probabilistic choice model, I find that the average worker
positively values working alongside friends. While workers earn less when working alongside friends,
they choose to do so with positive probability. Observations on worker positions during the preexperiment period indicate that a worker chooses to work alongside a friend in half of the workdays
in that period. This implies that a worker is willing to forgo a positive amount of potential earnings
to socialize with friends during work. A back-of-the-envelope calculation shows that workers with
positive valuations are on average willing to forgo 6 percent of their wage to socialize with friends
at work.
In the third main result, I find that both the magnitude of productivity loss associated with
working alongside friends and the value of working alongside friends depend on the worker’s level
of conscientiousness. Specifically, I observe a 9 percent loss in productivity from working alongside
a friend among workers with low conscientiousness scores (one standard deviation lower than the
average) compared to the 3 percent loss among workers with high conscientiousness scores (one
standard deviation above the average). More conscientious workers also tend to place lower valuations on working alongside friends. I also find that workers with high emotional stability are less
likely to choose to work alongside friends the larger the magnitude of the wage loss. Somewhat
surprisingly, extraversion is not related to the productivity loss (or gain) from working alongside
friends or how often an individual chooses to work close to friends.
This study joins a number of field experiments that investigates how social interactions influence worker behavior.1 Most notably, Bandiera et al. (2005, 2007, 2009, 2013) implement field
experiments to study how social connections within a firm affect worker performance across a wide
array of incentive schemes in the context of a U.K. fruit farm. In the same context, Bandiera et al.
(2010) exploit a quasi-random feature of assigning workers to different fields and document pacing
behaviors among socially tied workers; workers slow down when working alongside lower-ability
1
Bandiera et al. (2011) provide an overview of the literature on field experiments in firms.
4
friends and speed up when working alongside higher-ability friends.2 I examine the effect of friendships on productivity in a context where workers’ physical locations are fixed throughout the day.3
This feature allows estimation of friendship effects in a setting (low-skilled manufacturing) where
there is no need for friend pairs to synchronize their work speeds in order to socialize. This paper
also contributes to this literature by showing that heterogeneity in individual personality matters
for social interaction effects.
This study also contributes to the literature on compensating differentials by providing empirical
within-firm evidence that demonstrates that workers are willing to forgo pecuniary benefits for a
desirable non-pecuniary attribute of the job.4 In a model of compensating wage differentials, a
worker is willing to accept a lower wage for a favorable non-pecuniary job attribute (Rosen, 1986).
In this study, a worker’s productivity varies with the attribute of working alongside her friends.
This variation in productivity drives the wage difference through the performance pay component
of the compensation scheme. By choosing to work alongside friends, workers forgo a higher average
wage. Since other attributes of the workplace can be seen as orthogonal to the daily, experimentallyinduced variation in friend’s proximity, this study provides quantitative evidence of the value of
socializing with friends at work.
Finally, I also add to recent interdisciplinary research on the relationship between non-cognitive
skills, or personality traits, and socioeconomic outcomes (Bowles et al., 2001; Heckman et al., 2006;
Borghans et al., 2008; Almlund et al., 2011). Earlier findings in the personality psychology literature point to conscientiousness, among the Big Five factors, as the strongest predictor of job
performance (Schmidt and Hunter, 1998; Barrick et al., 1998, 2003; Hogan and Holland, 2003) and
2
In a recent study, Amodio and Martinez-Carrasco (2015) investigate productivity spillovers in an egg production
plant, at which worker compensation is largely based on a fixed wage, and find that working next to a friend mitigates
free-riding behavior that occurs from working next to high producing coworkers.
3
Cornelissen et al. (2014) show that peer effects, in general, differ across occupation types. For instance, when
considering the full set of jobs available at a local municipal level, the authors find peer effects only among jobs that
involve routine tasks.
4
Hamilton et al. (2003) find similar evidence at a garment manufacturing facility: high ability workers are
willing to participate in team production despite receiving wages lower than what they would earn under individual
production. However, this decision to join team production is confounded by other changes to the production setting,
such as standing instead of sitting in chairs and specializing in certain tasks. More recently, Bloom et al. (2015)
suggest qualitative evidence from a working from home experiment that workers are willing to forgo a nontrivial
amount of commute time by choosing to work in the office rather than at home in order to socialize with their
coworkers.
5
earnings (Nyhus and Pons, 2005; Mueller and Plug, 2006). I test this hypothesis in the context
of a manufacturing plant and find that workers who score above the median on the conscientiousness scale are nearly 7 percent more productive than workers who score below. Moreover, I find
that socializing behaviors explain about two-thirds of this productivity gap; high-conscientiousness
workers have smaller productivity declines when working alongside friends (explains 2.4 percentage
points) and lower likelihoods to choose to work alongside friends (explains 2 percentage points)
relative to low-conscientiousness workers.
This paper proceeds in six parts. Section 2 describes the field context and experimental design.
Section 3 presents the data and descriptive statistics. Section 4 outlines the conceptual framework
that guides the empirical analysis. Section 5 presents the empirical strategy and reduced form
estimates of the effect of working with friends on productivity. Section 6 provides a structural
analysis based on a probabilistic choice model and derives estimates on workers’ valuations of
workplace socializations with friends. Section 7 delivers the conclusion of the paper.
2
2.1
Field Context and Experimental Design
Field Context
For this study, I partnered with a seafood processing plant in Vietnam. The plant manufactures
canned and pouched seafood products which are mainly exported to U.S. markets. One of the main
tasks in producing seafood products is the semi-processing job, where tasks range from gutting to
filleting fish. Because this job requires deftness at using a processing knife, the plant hires processing
workers who specialize in this task. I studied these workers who were, at the time of the study,
regular employees of the firm.
The plant operates three teams of processing workers that specialize in the filleting process.
Each team works in their own processing room. Images of processing rooms are provided in Figure
1. The filleting job takes place on rectangular work tables which are all identical in size and
positioned side-by-side. Exact dimensions of the work table are provided in Figure 2. Each table
provides work space for up to six workers although typically four workers occupy one table.
The filleting process is an individual task and the plant records individual worker output.
6
Processed fillets are placed on individual trays which are weighed at one side of the room using an
electronic scale and recorded by a designated worker. Weighed trays are placed on racks for quality
inspections. Managers make routine inspections and trays that pass inspection are sent to the next
production stage, whereas rejected trays are returned to the worker for supplementary work; in
this sense, I observe quality corrected worker outputs.
Work materials (i.e., steamed fish) arrive at the processing room in large tray carts and managers
distribute trays containing fish to tables based on the number of workers at the table; the general
rule is one large tray per worker in each round of arrival. Workers jointly process the stock of fish
allocated to their table. While workers are not allowed to take fish from other tables, managers
would reallocate fish across tables according to each table’s work speed. In Appendix A, I show
that a one percent increase in average ability of workers at the table was, on average, associated
with an approximate one percent increase in the per-capita amount of fish allocated to that table.5
Worker compensation is based on a two-part wage system: base wage and performance wage.
Base wage is determined by whether the worker was on site for half or full-day. Workers are not
paid for days they are absent. Performance wage is based on a piece rate which pays workers
a fixed amount per kilogram of fish processed. Thus, the compensation scheme only pertains to
individual worker attendance and productivity. There was no change in the hourly-pay or piece
rate throughout the study period.6
Before the intervention, workers could choose their work positions at the start of each workday.
In general, there was no restriction on how workers should be positioned or whom they could work
with, except that the maximum number of workers allowed at one table was six. This enabled
workers to choose whether or not to work closely with their friends.
2.2
Randomization of Worker Positions
The field experiment was designed to randomly assign workers to work positions, including the
table and position at that table, within their processing rooms on a daily basis on days when there
5
I measure individual worker ability using the estimated worker fixed effect from a panel data specification that
takes into account the composition of coworkers, in terms of both physical and social distance to that worker.
6
Wages are paid once a month, typically on the second Saturday of the following month. Workers often request
advance payments which are typically made on the fourth Saturday of the month.
7
was processing work. To this purpose, I developed a random sequence generator tailored to each
processing room. Each worker was given a unique ID number and the positions in the room were
numbered from 1 to N , where N is the total number of processing workers in that team. For each
team and workday, a random sequence with length N was generated and workers were assigned to
work positions according to the order of their number in the sequence. In Appendix C, Figure D1
provides a sample of worker position assignment forms for each processing room. During the study
period, workers were neither allowed to switch positions nor fill in positions that were empty due
to worker absences.
Human resources staff at the management office generated the sequences and recorded the
resulting worker positions typically a week before the actual assignment.7 However, due to unexpected supply or demand shocks, such as shortage of fish among local suppliers or requests for
expedited shipping from buyers, it was difficult to forecast the exact dates of which there will be filleting work. On days with processing work, prior to entering the processing rooms, team managers
came to the office to pick up their team’s daily position assignment form from a human resources
staff and was responsible for arranging worker positions according to the form. Human resources
staff made unexpected visits to the processing rooms twice a day, usually once in the morning and
once in the afternoon, to check if the position assignments were implemented accordingly.
3
Data and Descriptive Statistics
3.1
Employee Records Data
The firm manages an employee records database that contains daily information on the tasks
performed, hours of work, and weight of fish processed by each individual worker. From this
database, I collected worker level data on hours of work and weight of fillet processed. I use this
data to construct each worker’s daily productivity, measured as kilograms of fillet processed per
hour.
Worker positions were recorded as part of the study for six months. Specifically, human re7
This was intended to account for possible new hires and job turnovers during the experiment period. Nonetheless,
there were no new hires and only 4 job turnovers during this period.
8
sources staff visited the processing rooms and recorded the ID number of the worker occupying
each position. The firm started recording worker positions six weeks prior to the randomization
phase and continued until the end of the study period. I combine worker position records with the
employee records data to produce a data set that consists of 104 workers and approximately 7,800
worker-workday observations.
Figure 3 presents plant level statistics on the fraction of workers working alongside friends and
worker productivity across time. Before the experiment about half of the workers were working
alongside friends. With the start of the experiment this statistic reduces to less than a quarter.
Average productivity jumps at the start of the experiment. Note, however, that I am not interpreting this as the causal estimate of the policy change. Overall, an average worker produces between
7-7.5 kilograms per hour during the study period. The average hourly wage is 17,000 Vietnamese
Dong (approximately $0.80).
3.2
Survey
A baseline survey was administered two weeks prior to the start of randomization to collect information on workers’ socioeconomic statuses, social ties, and personality measures. To recruit survey
participants, fliers containing information about survey eligibility, duration, compensation, and location of the survey office were posted on walls outside the canteen and inside changing rooms a
week prior to the survey. The survey office, set up exclusively for this study, was located outside
the plant at a five minute walking distance. Participants made walk-in visits to the office, mostly
during their two-hour lunch time period from 11am - 1pm.8 The baseline survey consisted of three
modules and took workers about an hour to complete all three. Upon completion, workers were
paid 30,000 Vietnamese Dong (approximately $1.50) as a token of appreciation for participating in
the survey.
In the first module, each worker was asked about her socioeconomic background, experience
on search and applying for her current job, and experience as a processing worker. The second
module asked each worker to report social ties within her processing team and for each reported
8
Surveyors were on-site to assist the workers with the survey process.
9
social tie workers were asked to report on the details of the relationship, such as whether the person
is a family member or friend, the duration of the relationship, whether the tie was formed before
working at this plant, and the frequency of activities shared inside and outside of the workplace.9
The last module was a Vietnamese-translated version of the Big Five Inventory(BFI), which is
a self-report inventory with 44 short-phrase questions designed to measure factors along the five
factor-analytically-derived dimensions of personality: extraversion, agreeableness, conscientiousness, neuroticism, and openness.10
To collect qualitative information on each worker’s preference with regard to working with
their social ties, an endline survey was conducted during the third week of December, 2014. The
procedures were similar to the baseline survey, except that the endline survey took approximately
five minutes to complete and participants were paid 20,000 Vietnamese Dong (about $1.00) as a
token of appreciation.
3.3
Descriptive Statistics
Summary statistics on survey participation for both surveys are presented in Table 1. Overall, I
was able to survey processing workers who were regularly showing up for work during the survey
periods. Two workers did not participate in the baseline survey because of long absences and, as a
result, are not included in the data set. Approximately, 10 percent of the workers, or 11 out of 112
workers, who participated in the baseline survey left their jobs during the study period, though 7
of these 11 workers had quit prior to the start of the experiment.11 Moreover, workers were only
notified about the position assignments on the first day of implementation. The other four workers
left their jobs over the five month period of the experiment.
Summary statistics on the socioeconomic statuses of workers and job application process are
provided in Table 2. Processing workers are all females because the firm only hires women for this
9
I only find three family ties among workers in the same processing room and, due to the small sample size, count
these as friendship ties.
10
The questionnaire, originally from John et al. (1991) and John et al. (2008), was translated in Vietnamese and
back translated in English by a professional translation company. Both versions were additionally checked by a
native Vietnamese with experience in Vietnamese-English translations. The original version of the BFI is available
for research purposes at http://www.ocf.berkeley.edu/~johnlab/bfi.php. The Vietnamese-translated version of
the BFI is available from the author upon request.
11
These were mostly involuntary job separations due to the firm’s internal decisions.
10
particular job. An average processing worker has about three years of experience in processing fish
and twenty months of tenure at their current job. Three quarters of the surveyed workers learned
about job openings through friendship ties. Moreover, almost 20 percent of the surveyed workers
reported that they had received help from a friend currently working at the plant when applying
to the job.
Table 3 provides a summary of friendships within processing rooms reported by surveyed workers. The median worker reported to have 4 friendships within her processing team. Two workers
reported to have no friends within their team. The median worker was mentioned as a friend four
times. It is also reasonable to expect friendships to form or dissolve over time, especially if there is
a change in the population with whom one regularly interacts. One of the limitations of this study
is that this study analyzes worker behaviors assuming that friendships are static over the study
period.
The details of the characteristics of reported friendships are shown in Table 4. I define a
friendship to exist between two workers if either one of the workers reported the other as a friend.
The top panel indicates that about 15 percent of the friendships formed prior to their current job
and the average duration of a friendship is approximately 16 months. Although not shown in this
table, 98 percent of reported friendships formed at least a month before the collection of worker
position data in the pre-experiment period. The bottom panel reports survey responses with respect
to the frequency of activities shared with friends during the past 3 months. Although the survey
questionnaire used a 5 point scale for all activities, except exchange of money, the actual word
descriptions varied across items. Importantly, 75 percent of the friendships reported to have been
working at the same table everyday. This suggests that, prior to the intervention of randomizing
positions, there was a tendency for friends to work in close proximity.
Summary statistics on workers’ personality scores along the Big Five domains are presented in
Table 6. The correlation matrix provided in the last five columns of the table show high correlations
between four traits; extraversion, agreeableness, conscientiousness, and neuroticism. While the Big
Five factors were initially constructed to be orthogonal to each other, between-factor correlations
are commonly found in empirical studies of the Big Five factors (Anderson et al., 2011).
11
4
Conceptual Framework
In order to guide the empirical analysis, I first present a model of worker behavior embedded with
social ties and skills, in which skills take the form of processing skills and non-cognitive skills.
Next, I introduce the literature on non-cognitive skills and present hypotheses on the relationship
between each non-cognitive skill and worker outcomes.
4.1
Model
4.1.1
Worker’s choice of effort
Denote e as worker’s choice of effort for production. Productivity of worker i, yi , is measured in
kilograms of fish processed per hour. For simplicity, assume that productivity of worker i is given
by yi = ei . As in the empirical setting, I assume that workers are paid a combination of a fixed
base wage plus a piece rate wage. Accordingly, worker’s wage is an affine function of individual
productivity. Workers derive utility from wage, W (), where We > 0 and Wee < 0.
Workers are heterogeneous in fish processing skills, α, and non-cognitive skills, denoted by the
vector θ.12 Denote C(e, α, θ, f ) as the worker’s cost function from exerting effort level e, where f
is an indicator for the presence of friends.13 Assume Ce > 0 and Cee > 0 such that the cost of
effort is increasing in effort and the marginal cost of effort is also increasing in the current level of
effort. Then, the worker’s utility maximizing effort level in each state with regard to the presence
of friends can be characterized as follows,
enf
ef
∈ arg max W (e) − C(e, α, θ, nf )
(1)
∈ arg max W (e) − C(e, α, θ, f )
(2)
e
e
12
In general, processing skills and non-cognitive skills are likely to influence each other and as a result be correlated.
For instance, processing ability may influence personality characteristics if better processing ability leads to being more
sociable. Conversely, high conscientiousness may positively affect training motivation and acquire better processing
skills. For instance, Colquitt et al. (2000) conducts a meta-analysis on the training motivation literature and provides
evidence on positive correlations between training motivation and internal locus of control, conscientiousness, and
self-efficacy, and a negative correlation between training motivation and anxiety.
13
As in Almlund et al. (2011), the optimal effort level also depends on non-cognitive factors. Here, however, I
introduce non-cognitive factors as parameters of cost rather than the production function to facilitate the idea that
non-cognitive skills influence the cost of exerting effort.
12
where enf denotes optimal effort in the absence of friends and ef denotes optimal effort in the
presence of friends.
I assume that exerting effort is less costly for workers with better skills. As a result, optimal
effort levels in equations (1) and (2) are each increasing in the worker’s production skill and noncognitive skills. However, it is a priori unclear how the presence of friends may influence the worker’s
marginal cost of effort. It is possible for working with friends to motivate the worker and, therefore,
make exerting effort less costly (Ce (nf ) > Ce (f )). On the other hand, working with friends may
make exerting effort more costly if workers engage in idle-chats or friends spread negative work
behaviors (Ce (nf ) < Ce (f )).
I also consider worker heterogeneity in behavioral responses to the presence of friends. The
difference in marginal cost of effort between the two states Ce (nf ) − Ce (f ) may be a function of
the worker’s production skill or non-cognitive skills. An example of production skill would be if,
for instance, high productivity workers are less sensitive to the friendship status of workers nearby
where as low productivity workers experience social pressure when a friend is working nearby
relative to when no friend is present.14 In addition, as shown in Bandiera et al. (2010), there exists
a possibility for Ce (nf )−Ce (f ) to depend on the relative productivity between the reference worker
and her friends present at nearby positions.
Non-cognitive skills may play an important role in determining whether or not to exert less
effort in the presence of friends. For example, workers that are more talkative may have a greater
urge to socialize in the presence of friends and, therefore, incur higher marginal costs than taciturn
workers. Also, workers with better self-discipline skills or the ability to focus on personal goals may
experience smaller increases in their marginal costs of effort, compared to less disciplined workers,
when working in the presence of friends relative to when working without friends compared to
workers with less self-discipline. I discuss this in more detail with respect to the Big Five Factors
in the next section.
14
While peer pressure has mostly been discussed in workplace contexts with free-riding incentives (Kandel and
Lazear, 1992; Mas and Moretti, 2009) or fixed-wage schemes (Falk and Ichino, 2006), the possibility of peer pressure
under compensation schemes with partly piece rate wages may still arise if there are status concerns with respect to
wages within social networks (Frank, 1985).
13
4.1.2
Worker’s selection on the presence of friends
Next, I model how the worker selects whether or not to work with friends using a binary choice
framework. The goal is to build a model that allows estimation of each individual workers’ value
of working with friends. Let yif and yinf denote worker i’s utility maximizing productivity levels
derived from equations (1) and (2), respectively. Suppose workers have heterogeneous intrinsic
valuations for working with friends. Worker i derives a constant utility, si , from working with
friends. I specify worker i’s utility from working with, and without, friends using the following
random utility model:
ufi
unf
i
where
= wif − cfi + si + fi
(3)
nf
= winf − cnf
i + i
(4)
nf
wif = W (h + ρ · yif ), winf = W (h + ρ · yinf ), cfi = C(yif , α, θ, f ), cnf
i = C(yi , α, θ, nf ) (5)
and fi and nf
i are mean-zero stochastic error terms. Fixed hourly wage and piece rate wage are
each denoted by h and ρ, respectively. Denote Fi as the indicator variable equal to one if worker i is
working with a friend and, zero otherwise. Then worker i chooses the presence of friends according
to
Fi =


 1 if uf > unf
i
i
(6)

 0 otherwise.
An implication of the theory of compensating wage differentials is that workers are willing
to forgo wage in exchange of desirable work attributes that are nonpecuniary and consumed as
part of the work (Rosen, 1986). In the current context, the attribute of interest is working with
friends. Wage is a function of individual productivity. If working with friends has a negative
impact on productivity then working with friends entails a monetary cost in the amount of the
forgone potential wage. Yet, if workers consider working with friends as a desirable attribute then
we would observe workers choosing to work with friends even in the presence of negative impacts
on productivity and, therefore, lower earnings.
This model implicitly assumes that workers are knowledgeable about their utility differences
14
between the two states and that they are able to make rational decisions. In fact, either of these
assumptions may fail. Unfortunately, I currently do not have sufficient data to independently test
these assumptions.15 In section 6, however, I return to this issue by investigating how the probability
to work with friends varies across subgroups with different levels of education, work experience, and
non-cognitive skills. Differences in selection behavior along these attributes should be indicative
of why some workers choose to work with their friends more often than their counterparts with
similar costs and benefits.
4.2
Non-Cognitive Skills
For a brief introduction to the taxonomy of the five most commonly known non-cognitive skills,
known as the Big Five Factors, Table 5 provides a description of each skill and the associated facets.
I use the term skills rather than traits to reflect findings from recent studies that non-cognitive
skills can be shaped and developed throughout a person’s life (Kautz et al., 2014; Blattman et al.,
2015). Here I focus on describing earlier findings in empirical studies on non-cognitive skills and
its relation to socioeconomic behaviors and outcomes. Based on this literature, I formulate specific
hypotheses on the relationship between each non-cognitive skill and socializing behaviors in the
workplace.
Ample evidence, often corroborated through different studies and researchers, suggests that
non-cognitive skills influence academic achievement (Duckworth and Seligman, 2005; Heckman
et al., 2006; Duncan et al., 2007) as well as labor market outcomes (Bowles et al., 2001; Heckman
and Rubinstein, 2001; Nyhus and Pons, 2005; Heckman et al., 2006; Carneiro et al., 2007; Heckman
et al., 2010; Heineck and Anger, 2010).16 In addition, a substantial body of evidence documents
the predictiveness of the Big Five Factors on job performance and wages (Barrick and Mount, 1991;
Barrick et al., 1998; Salgado, 1997; Schmidt and Hunter, 1998; Hogan and Holland, 2003; Nyhus
and Pons, 2005; Mueller and Plug, 2006; Callen et al., 2015).
15
In an ideal setting, the first assumption can be tested through comparing worker position data before and after
the randomization experiment if I were to assume that friendships have not changed during this period. An alternative
which I have not considered to do is to have asked workers prior to the experiment about their productivity differences
between when working with a friend and without a friend.
16
For a literature review on the evidence of the effect of non-cognitive skills on educational achievement and
post-education outcomes, I refer the reader to Kautz et al. (2014).
15
With respect to skill-specific evidence, individuals with high extraversion, a measure of how
talkative and sociable a person is, might have stronger preferences to socialize at work. Whether
or not it is positively correlated with job performance can be task specific. As Barrick and Mount
(1991) finds in their meta-analytic study, extraversion can be a positive skill in work tasks involving
social interactions, such as managerial and sales related occupations. Yet, overall evidence on its
relationship to earnings is weak (Heineck and Anger, 2010). In the current manufacturing context,
one would expect extraversion to have a negative relation to job performance since socializing
reduces the time allocated to production activities. Moreover, workers with high extraversion
might have higher valuations to work with friends because of preferences to socialize. On the
other hand, extroverts might also talk to friends and coworkers equally. This would imply that an
extrovert would show no difference in productivity between when working with friends and when
working without friends nor more likely to choose to work with friends.
Meta-analytic studies in the personality psychology literature on Big Five personality factors
and job performance typically single out conscientiousness as a strong and positive predictor of job
performance across a wide array of occupation groups (Barrick and Mount, 1991; Barrick et al.,
1998, 2003; Salgado, 1997; Larson et al., 2002). Thus, the empirical evidence fits well with the
definition of conscientiousness as having the ability to exert self-control and self-discipline and a
preference to follow rules and goals.17 Accordingly, in terms of the model, I expect individuals with
high conscientious to experience smaller increases in their marginal costs of effort, compared to low
conscientious individuals, when working with a friend. This, in turn, implies that high conscientious
individuals are expected show smaller negative effects on productivity from the presence of friends
relative to low conscientious individuals. Whiteside and Lynam (2001) also finds from a psychological survey of young adults that low conscientiousness is associated with a lack of premeditation.
This suggests that if working with friends is costly, workers with high conscientiousness should be
less likely to choose to do so relative to low conscientious workers.
Personality studies report weak associations between neuroticism, or the reverse measure of
emotional stability, and job performance (Barrick and Mount, 1991; Bono and Judge, 2003). Based
17
Empirical evidence on how conscientiousness predicts earnings is rather mixed. See, for instance, Heineck and
Anger (2010).
16
on this evidence, it is not obvious whether the change in marginal costs of effort from the presence
of friends depends on neuroticism. On the contrary, a number of studies document evidence that
neuroticism is negatively correlated with earnings. For instance, Nyhus and Pons (2005) finds
among Dutch households that neuroticism is negatively correlated with wages. Heckman et al.
(2006) provides evidence that locus of control, a measure outside of the Big Five but known to be
correlated with neuroticism, predicts earnings at a similar level to the predictiveness of a person’s
cognitive ability.18
Studies suggest that neuroticism, or the inverse of emotional stability, is predictive of decision
making behavior. Costa and McCrae (1992) attribute impulsive decision making as one of the
facets of neuroticism. According to the authors, people with higher neuroticism are more likely to
make decisions in favor of their instant emotions rather than the future payoffs that arise from those
decisions.19 Also, Denburg et al. (2009) find from a gambling experiment that high neuroticism
adults are more likely to make disadvantageous choices, in the sense of lower expected payoffs,
relative to low neuroticism adults. Combining these earlier findings, I expect workers with high
neuroticism to show decision making behaviors consistent with being more impulsive (i.e. choice
outcomes accord more with their consumption value of socializing, whether they like it or not)
relative to low neuroticism workers.
The personality psychology literature does not provide compelling evidence on the predictiveness of agreeableness and openness on job performance. With regard to economic preferences,
Dohmen et al. (2008) finds positive correlations between agreeableness, as well as openness, and social preferences, such as trust and positive reciprocity.20 However, it is not clear in this setting how
a preference for trust, or reciprocity, might affect socializing behaviors in the workplace. In terms
of earnings, Nyhus and Pons (2005) and Heineck (2011) find negative associations between agree18
In a meta-analytic study, Judge et al. (2002) find that locus of control, self-esteem, neuroticism, and self-efficacy
are all strongly related to each other. Cebi (2007) and Semykina and Linz (2007) find that locus of control is positively
related to earnings using the National Longitudinal Survey of Youth dataset and Russian employee data, respectively.
19
Relatedly, Shafer (2000) and Gati et al. (2011) find that individuals with high neuroticism report to have
greater difficulty in making career decisions. On the other hand, these studies do not find associations between
conscientiousness and indecisiveness, a report of having difficulty in making decisions.
20
Experimental studies find that other types of economic preferences, such as time-preference or risk-preference,
are not significantly associated with the Big-Five factors but rather work as complements in determining life time
outcomes (Dohmen et al., 2010; Becker et al., 2012; Kautz et al., 2014).
17
ableness and earnings while Mueller and Plug (2006) find positive correlations between openness
to experience and women’s earnings using the Wisconsin Longitudinal Study (WLS).
5
Reduced Form Analysis on the Effect of Working With Friends
5.1
Empirical Strategy
5.1.1
Spatial Contiguity and Proximity
The workspace, or work positions, surrounding a worker can be broadly partitioned into two,
depending on whether or not there is spatial contiguity between the worker and that position.
Figure 2(a) presents a diagram with worker positions at three tables. The shaded areas represent
work spaces that are contiguous to worker B and L, respectively. Since work tables are positioned
side-by-side, table boundaries are irrelevant in this setting.
Work positions that are spatially contiguous to a certain position can be separated into finer
categories depending on their proximity and orientation to that position. I use Figure 2(b) to
illustrate this point. Positions that are spatially contiguous to worker B and L are now shaded
in gray with three different levels of darknesses which represent different levels of proximities and
orientations to worker B and L. Positions that are alongside are shaded in dark gray, positions that
are not on the same side but directly oriented towards the worker are shaded in moderate gray,
and positions that are diagonally oriented towards the worker are shaded in light gray.
I combine these spatial features of the workplace and adopt the following terminologies which
are used throughout the remainder of the paper.
Definition 1. A and B are in high proximity if A and B work at positions alongside each other
(i.e. positions shaded in dark gray).
Definition 2. A and B are in medium proximity if A and B work at positions that are directly
oriented towards each other (i.e. positions shaded in moderate gray).
Definition 3. A and B are in low proximity if A and B work at positions that are diagonally
oriented towards each other (i.e. positions shaded in light gray).
18
5.1.2
Econometric Specification
To identify the effect of workplace friendships on worker productivity, I exploit the within-worker
variation in productivity and proximity to friends across workdays caused by randomized position
assignments. First, to check if the presence of friends at spatially contiguous positions has any
affect at all on a worker’s productivity, I estimate the following panel data specification:
yirt = β · Contiguousirt + Xirt + αi + λrt + irt
(7)
where yirt is the log productivity (log of kilograms of fish processed per hour) of worker i in room r on
day t. Contiguousirt is an indicator variable equal to one if worker i has at least one friend working at
a spatially contiguous position in room r on day t, and zero otherwise. Xirt contains information on
the number of workers and their average productivities working at positions spatially contiguous to
worker i in room r on day t, and αi and λrt are the worker and room×day fixed effects, respectively.21
The worker fixed effect accounts for unobserved time-invariant worker characteristics while the
room×day fixed effect accounts for time-varying productivity shocks occurring at the room×day
level. The latter type of shock may be especially relevant to this setting since each processing room
is associated to different production lines with own pre-processing and post-processing facilities
operated by different groups of workers.22 The error term for individual i in room r on day t is
represented by irt . The sole parameter of interest in equation (7) is the coefficient on the variable
Contiguous, β. Under this model specification, β can be interpreted as the effect of a friend’s
presence nearby on worker productivity.
To better understand the behaviors through which friends may influence one another’s productivity, I turn to a specified model that makes use of differing levels of proximity. Accordingly, I
estimate the following specification,
yirt = γL · Low Proxirt + γM · Med Proxirt + γH · High Proxirt + Xirt + αi + λrt + irt
21
(8)
The estimation procedure for individual worker ability is presented in Appendix B. Xirt is the average of that
estimate among workers that are at contiguous positions, excluding the reference worker.
22
Regressions results from including alternative fixed-effect specifications are separately reported as part of a
robustness check.
19
where Low, Med and High Prox are indicator variables equal to one if there is at least one friend
working at low, medium, and high proximity, as defined in the previous section, respectively. All
other variables are defined as above. In equation (8), γL is the effect on a worker’s productivity
from having a friend working at a spatially contiguous position but with low proximity. γM is the
effect on productivity from the presence of a friend working at medium proximity and γH is the
effect on productivity from the presence of a friend working at high proximity.
Although randomized position assignments prevented workers from self selecting into certain
proximities to friends, proximity variables can yet be considered to be endogenous given that they
can only be realized, or appear in the dataset, if the worker actually showed up on the day of
assignment. This would be problematic if work attendance is both correlated with the position
assignment and some unobserved determinant of worker productivity. For instance, if workers were
informed about the position assignments in advance or were able to make accurate predictions on
the days that they will work closely to friends, workers might show up even on days when they
are under-motivated or fatigued if a friend is assigned to work at close proximity but not under
a position assignment with no friend working nearby. In general, any possibility of correlation
between unobserved determinants of worker productivity and realized worker positions would bias
the estimate of interest.
Accordingly, for the main specification, I use an instrumental variables strategy that exploits
variation in the assigned proximities to friends which includes both realized and unrealized proximities. The idea is to instrument three endogenous variables (low, medium, and high proximity
positions that were realized) with three exogenous variables (assignment to low, medium, and
high proximities). Since positions were randomly assigned ahead of time, the assigned proximity
variables can be considered to be exogenous and not subject to worker decisions.
5.2
Results
Throughout the analysis, I consider two workers to be friends if either one of the workers reported
the other worker as a friend. Table 7 reports summary statistics on unconditional and conditional
probabilities of having at least one friend present at each level of proximity during the randomization
20
period. On average, the probability of working with at least one friend at a spatially contiguous
position was around 0.41. Conditional on at least one friend working at a contiguous position, the
probability of at least one friend at low, medium, and high proximity was, on average, 0.48, 0.28,
and 0.40, respectively.23 Conditional probabilities in the right three columns of table 7 show that
on some days a worker was simultaneously exposed to multiple levels of proximities.
Table 8 shows baseline results with OLS estimates. Columns 1 and 2 use data from the preexperiment period and find negative correlations between the presence of at least one friend at
contiguous and high proximity positions, respectively. Columns 3 and 4 each estimate equation (7)
and (8) using the experiment period data. The estimates are relatively smaller than those from the
pre-experiment period. This is possible if workers were selecting to work with their closest friends
before the experiment whereas the experiment only captures how productivity change is associated
with the presence of an average friend.
Next, I present IV estimates based on assigned proximities to friends. The top panel of Table
9 presents first stage results. Each level of realized proximity is strongly related to assignment to
that level of proximity. On average, assignment to low proximity with a friend increased a worker’s
likelihood of working at low proximity with a friend by 0.84. The reason this is not one is because
of worker absences. Likewise, assignment to medium and high proximity increased the likelihood
of working at medium and high proximity by 0.86 and 0.83, respectively. The negative associations
between different assigned proximities and realized proximities are due to the fact that workers can
be assigned to multiple levels of proximities with friends and since the friend that has already been
assigned at some proximity cannot appear again at a different proximity. Concerning issues with
weak instruments, F-statistics from tests of joint significance of the three instruments are reported
at the bottom of the panel. The statistics are sufficiently large.
The bottom panel of Table 9 reports the second stage results. The estimate in Column 1
suggests that workers are on average 5.8 percent less productive when at least one friend is present
at a high proximity position relative to when no friend is present at a contiguous position. The
estimates for other proximities are insignificant and close to zero. This result suggests that friends
23
It is least likely to have a friend at medium proximity because there is only one position that has medium
proximity to a worker.
21
affect productivity only when they are close enough to socialize. Also, the estimates are similar
to the OLS estimates which may indicate that workers’ decisions to show up at work was not
systematically related to their assignments to high proximity positions. Column 2 indicates that
on average workers lose about 4 percent of their hourly wage when working with friends at high
proximity. When converted into monthly terms, this is commensurate to a loss of a full day’s wage.
To estimate the impact of the policy intervention on the firm’s average productivity, I restrict
the data set to workdays within a certain bandwidth of days to and from the date of implementation
and measure the difference in productivity before and after the policy change. Since workers did not
anticipate this change and other factors of production should have been relatively stable between
these two periods, the productivity difference should be capturing the policy impact along with
plausible hawthorne effects from workers responding to the implementation itself.
For a given bandwidth size, I use the following specification:
yirt = µ · Interventiont + αi + λrt + irt
(9)
where Interventiont is an indicator variable equal to one if date t is after intervention. Figure 4
plots coefficient estimates (µ̂) in the order of increasing bandwidth size from one to twelve. The
estimate at bandwidth of size one shows a 27 percent increase in worker productivity and then
drops to approximately 5 percent after I increase the size to more than four days.24 This suggests
that the initial increase observed within the first two or three days can be largely attributed to
effects other than the change in worker positions. Comparing the estimate on the three day interval
to the estimate on the five day interval shows that this additional effect accounts for around 55
percent of the productivity increase. Part of this behavior can be explained, loosely speaking, as
hawthorne effects if the policy change led workers to temporarily think that the firm could take
further actions, such as firing, if they were not highly productive, termed by Levitt and List (2007)
as experimenter demand effects, or if the implementation process simply reminded workers that
24
An alternative calculation that uses the change in the likelihood to work alongside friends and the estimate on
the average effect of working alongside friends also shows that the policy change led to about a 4 percent increase in
the firm’s average productivity. This value is derived by taking the mean of the product of the change in probability
of working alongside friends and the difference in productivity when working alongside friends and when without
working alongside friends across all workers.
22
they are under observation. Yet, the current experimental setting limits further investigation of
the channels through which these effects may arise.
5.3
Robustness Checks
While specification (8) provides a simple formulation for characterizing how the presence of friends
at different proximities may affect worker productivity it may be naive to not consider how the
presence of other friend pairs (i.e. friend pairs not connected to the focal worker) that are working in
close proximity may affect the focal worker’s productivity as well. A friend pair may be disruptive
to others if they engage in loud chats during work. Goofing off with one another may spawn
negative worker behavior among workers nearby. Failure to take into account potential spillovers
from other friend pairs may result in misinterpreting the effect of friends on productivity.
In order to allow for possible externalities from other friend pairs, I include variables that
indicate whether any of the workers at contiguous positions have a friend nearby at each level of
proximity. This can be written as
yirt = γL · Low Proxirt + γM · Med Proxirt + γH · High Proxirt
X
+
(δL · Low Proxjrt + δM · Med Proxjrt + δH · High Proxjrt )
(10)
j∈C (i)
+ Xirt + αi + λrt + irt
where Low Proxjrt , Mex Proxjrt , and High Proxjrt are indicator variables equal to one if worker
j in room r on day t has a friend working at low proximity, medium proximity, and high proximity, respectively. These externality variables are summed over the set of workers at spatially
contiguous positions to worker i, denoted by C (i). Then, δL is the externality on productivity
from an additional contiguous worker working with a friend at low proximity. Similarly, δM and
δH are externalities arising from an additional worker at a contiguous position working at medium
proximity and high proximity with respect to her friend, respectively. If both workers of a friend
pair are working nearby, this is counted as two rather than one. In this way, a friend pair with both
workers being spatially contiguous to the reference worker can be considered as twice as influential
23
than that with only one worker being contiguous.
Column 1 of Table 10 provides IV estimates on worker’s own proximity variables and other workers’ proximity variables. Estimates on own proximity variables are almost unchanged. However, the
estimate on other workers’ high proximity variable suggests that working at positions contiguous
to a high proximity friend pair is associated with a 1.8 percent loss in productivity. This suggests
that social interactions between high proximity friends negatively impact other coworkers as well.
The result is significant with standard errors clustered two-way by worker and room×day level to
allow for arbitrary correlations of error terms across workers in the same room and day. Failure
to account for this is especially problematic when the regressor of interest is also correlated across
workers within the room×day level, also referred to as the clustering problem, which is known to
lead to over-rejections of true null hypotheses (Wooldridge, 2003; Angrist and Pischke, 2009).25
Work materials are initially distributed at the table level and then reallocated across tables
according to each table’s progress. Under this distribution process, workers may behave differently
depending on the friendship status of coworkers at the same table. For instance, a friend’s presence
may influence worker productivity when both are at the same table but not when at different tables.
To examine whether working at the same table or not played a significant role in determining
worker behavior when working in the presence of friends, proximity variables are interacted with
an indicator variable that is equal to one if at least one friend at that proximity is at the same
table and zero otherwise.26 Then, these interaction terms should be capturing the difference in
productivity effects from the presence of a friend at the same table relative to the presence of a
friend at a different table within each level of proximity. Column 2 of Table 10 suggests that the
associated productivity decline from the presence of a friend at high proximity is robust to whether
or not the friend is assigned to the same table.
Not all work positions have two alongside positions. Positions along the front and back side of
the processing rooms only have one alongside position and, therefore, a lower likelihood of working
25
To see why the regressors are correlated, if worker i has a friend pair working at a contiguous position, then that
friend pair also appears in the regressors of all other workers that are contiguous to the friend pair. The room×day
cluster correlation for other workers’ proximity variables at low, medium, and high proximity is 0.12, 0.07, and 0.12,
respectively.
26
A worker at medium proximity can only be at the same table. Therefore, I include only two interaction terms.
24
alongside a friend. If workers are naturally more productive at these positions because of, say,
more workspace then this will lead to a negative association between working alongside a friend
and productivity. To check for this possibility, I include interaction terms between each proximity
and an indicator variable equal to one if the assigned position is a corner position and zero otherwise.
Column 3 indicates that the results are not driven by spurious correlations associated with working
at corner positions. Yet, it also shows that working at a corner position leads to a smaller decline
in productivity relative to non-corner positions when a friend is working alongside.
Next I perform a series of robustness checks on the baseline estimate by including different
sets of dummy variables at the room and day level. Estimates derived from using alternative fixed
effects specifications are reported across columns 4-6. The results suggest that the baseline estimate
is robust to different fixed effects specifications.
An additional concern is that friendship reports were collected once two weeks prior to the
start of the experiment. Changes in friendship status may have occurred during the experiment
period. This is likely to bias the estimates toward zero if existing friendships dissolve while new
friendships form. Column 6 checks if the effect on worker productivity from the presence of reported
friends declines over the five month experiment period. The estimate on the interaction term, High
proximity × Month, suggests that the effect persisted, at least during the duration of the study.27
The attrition rate during the experiment period was low: 101 of the 105 workers were still
working at the end of the experiment. Nonetheless, worker absence causes a missing data problem.
The mean absence rate during the experiment was 14.8%.28 Comparing days assigned to work
alongside friends and to work without friends alongside, the absence rate is 13.0% for days when
assigned to work alongside friends and 16.7% for days when not assigned to work alongside friends.
To examine the robustness of my main result to this differential absence rate, I use the bounding
approach of Lee (2009) to construct lower and upper bounds for the effect of working alongside
friends. The key identifying assumption required for implementing these bounds is a monotonicity
assumption that assignment to work alongside friends affects the decision to show up at work only
27
Though not reported in this paper, interacting proximity variables with month dummies, instead of a linear
month trend, produces similar results.
28
During the six weeks before the experiment, the mean absence rate was 18.7%. Absence rate went down by 3.9%
after starting the experiment (p = 0.022).
25
in one direction. In my context, this requires assuming that for a given worker the assignment to
friends homogeneously affects that worker’s attendance decision. Heterogeneous effect of treatment
on attendance across workers is not an issue since treatment assignment (assignment to work
alongside a friend or not) varies within the worker.
To construct Lee (2009) bounds I trim both ends of the distribution of each worker’s productivity
on days when the worker was assigned to work alongside friends by the difference in the attendance
rate between days when assigned to work alongside friends and days when not assigned to work
alongside friends as a proportion of the worker’s attendance rate on days when assigned to work
alongside friends. Taking the weighted average of this then gives a lower bound for the effect of
working alongside friends of -0.031 and an upper bound of -0.098, compared to the IV estimate of
-0.058 in Table 9. In terms of the effect on hourly wage, which is estimated at -0.040 in Table 9,
the bounds are -0.022 and -0.066, respectively. The lower bound is relevant if workers were able
to anticipate their assignments and decided not to show up on days when none of their friends
were assigned to work alongside and were also expecting to be less productive. However, position
assignments were randomized on a daily basis and were informed only at the start of each workday.
The probability of correctly anticipating days of assignment to work alongside friends should have
been low.
5.4
Heterogeneous Effects on Productivity from Working with Friends
The previous section provides strong evidence that workers socialize with their friends when they
are in close proximity and, as a result, are less productive and also earn less when working alongside
friends. Yet, workers might respond differently to the presence of friends, especially when there is
a monetary cost to socializing. In this section, I examine individual heterogeneity by focusing on
two crucial skills in the workplace: non-cognitive skills and production skills.
5.4.1
Non-Cognitive Skills
Performing analysis using self-reported personality measures may be potentially problematic for
several reasons. First, measurement error in the personality measure and reserve causality with the
26
outcome variable complicate interpreting the results (Heckman et al., 2006). For this reason, all results in this study regarding the relationships between personality traits and outcome variables are
prohibited from having causal interpretations. Second, one may question the robustness of the Big
Five personality structure across cultures and languages. Yet, the 44-item Big Five Inventory used
in this study is shown to be well replicated and generalizable across various cultural regions and
languages (Benet-Martinez and John, 1998; Schmitt et al., 2007).29 Third, participants may make
false reports on the test. This might be the case if test scores are believed to have actual consequences, such as in job interviews or promotion evaluations. Nonetheless, impression management
behaviors should have been minimal since survey participants in this study were job incumbents and
informed about confidentiality of the results prior to the test. Another problem with self-reports
is that participants may report based on how they wish to be perceived rather than how they
actually think and behave, known in personality assessment as self-deception. However, studies
in the personality psychology literature provide evidence that the influence of response distortions
on the predictive validity of the relationship between personality factors and job performance is
minimal (Hough et al., 1990; Ones et al., 1996, 2007). Overall, the survey context and evidence
from previous studies support the reliability of using self-reported personality measures as a tool
to understanding differences across individual behaviors and outcomes.
To examine whether the effect of the presence of friends on worker productivity depends on the
reference worker’s non-cognitive skills, I estimate the following specification,
yirt =
X
γK · K Proximityirt + ρK · K Proximityirt × Non-cognitive Skilli + Xirt + αi + λrt + irt
K=L,M,H
(11)
where Non-cognitive Skilli is worker i’s standardized score of that non-cognitive skill.
Column 1 of Table 11 reports ordinary least square estimates from including all five factors
in equation (11). The estimates suggest that on days when a friend is working at high proximity
workers with low conscientiousness scores are more likely to be associated with larger productivity
29
Schmitt et al. (2007) provides evidence that the English BFI can be successfully replicated across cultures
and languages by examining a dataset containing the BFI translated in 29 different languages and surveys 17,408
individuals across 56 nations in 6 continents.
27
declines than workers with high conscientiousness scores. A worker with a conscientiousness score
that is one standard deviation below the sample mean is likely to be associated with a nine percent
decline in productivity in the presence of a friend at high proximity. On the other hand, a worker
with a conscientiousness score one standard deviation above the mean is likely to experience a
three percent decline in productivity. This supports recent findings in Callen et al. (2015) that
conscientiousness is one of the strongest predictors of job performance among doctors, inspectors,
and health officials in Pakistan’s public health sector.
Columns 2-6 present estimates from separate regressions on each personality factor. Column 3
suggests a positive relationship between agreeableness and productivity in the presence of a friend
at high proximity. However, this is plausibly due to a positive correlation between agreeableness
and conscientiousness since the estimate on agreeableness loses significance in Column 1. Column
4 reports a similar estimate on conscientiousness to that reported in Column 1.
5.4.2
Production Skills
Several studies suggest that production skills or ability (henceforth, ability), may have a crucial
influence on job performances of friends (Bandiera et al., 2010), co-workers (Mas and Moretti,
2009; Falk and Ichino, 2006), and managerial decisions (Bandiera et al., 2009). In a fruit picking
setting, Bandiera et al. (2010) finds that the effect of the presence of friends significantly depends
on the relative ability between friends working alongside on the same field and that this is due to
pacing with friends of different abilities in order to socialize. While in the current context work
speed pacing with friends is unnecessary for workers to socialize since work positions are fixed along
the workday, it is still possible for workers to adjust their work speeds according to that of their
friends if workers have social preferences, such as aversion to inequity (Andreoni and Miller, 2002;
Charness and Rabin, 2002) or status concerns (Frank, 1985).
Examining whether the effect of working in the presence of friends is heterogeneous to the
reference worker’s ability or that of her friends’ requires a measure of production ability for all
workers. Building on the approach of Mas and Moretti (2009), for a measure of ability, I use
estimates of worker fixed effects from a specification that includes a set of dummy variables to
28
account for physical proximities and social relationships between the focal worker and coworkers in
contiguous positions.
Figure B1 in Appendix B shows the distribution of kernel density estimates on worker ability.
Note that the estimates are mean zero by construction. The standardized ability estimate of workers at the 25th and 75th percentile is -0.065 and 0.073, respectively, implying an ability differential
of about 15 percent. Table 12 checks if non-cognitive skills have significant relationships with production ability. Column 1 reports estimates without controlling for individual covariates, worker’s
age and experience. Column 2-8 controls for individual worker covariates using a quadratic polynomial function. Column 3-8 reports estimates from separate regressions of estimated ability on
each non-cognitive skill and the Big Five Index.30 Among the Big Five factors, only extraversion
shows a statistically significant correlation with estimated worker ability.
The lack of a significant correlation between conscientiousness and production ability may
indicate that ability depends on physical attributes, such as the worker’s height, since workers have
to adjust to the table height, or hand size. Likewise, the positive association between extraversion
and ability might be simply reflecting an underlying relationship between innate physical ability
and outgoing behaviors or a matter of reverse causality if, for instance, being adept at work leads
to greater social confidence.31
Next, to check for heterogeneous effects, I estimate the following equation:
X
yirt =
\ j + Xirt + αi + λrt + irt (12)
γK · K Proximityirt + ξK · K Proximityirt × Ability
K=L,M,H
where K Proximity is an indicator variable of the presence of at least one friend at each proximity
\ j denotes the estimated production ability of worker j. I follow
(K = Low, Med, High) and Ability
the instrumental variables strategy used in the previous section and, therefore, all results are IV
estimates.
Column 1 of Table 13 reports estimates on the high proximity parameter and the interaction
30
The Big Five Index is constructed as an equally weighted average of the standardized scores of the five personality
traits. In accordance with conventional methods, I reverse score Neuroticism when constructing this index.
31
Heckman et al. (2006) and Urzua (2008) exemplify the importance of reverse causality issues with respect to
interpreting results from personality tests on socioeconomic outcomes.
29
term using the reference worker’s ability estimate.32 The estimate on the high proximity term
interacted with worker’s ability is positive but statistically insignificant. Column 2 checks whether
worker productivity is differentially affected by the ability of friends at high proximity positions.33
Column 2 indicates that a high proximity presence of a friend with ability corresponding to the
75th percentile of the ability distribution is associated with a productivity decline of 5.1 percent,
whereas the presence of a friend at the 25 percentile is associated with a decline of 6.8 percent. This
suggests that productivity effects from the presence of friends at high proximity positions depends
on the ability of friends and that working with low ability friends tends to be associated with
slightly larger declines in worker productivity relative to working alongside high ability friends.
In the current context, a possibly more relevant type of measure would be the relative ability
between friends. I use two different measures of relative ability that would be pertinent to this setting. First, I measure relative ability as the ordinal relationship between friends’ abilities: whether
or not worker i has higher ability than her friends at contiguous positions. The second measure of
relative ability is a cardinal measure that also takes into account of the absolute ability difference
between friends. Column 3 indicates that the associated decline in productivity from the presence
of a friend at high proximity does not depend on the relative ability with friends. Column 4 does
not find differential effects associated with the size of ability difference with friends.
6
Structural Analysis on Working With Friends
The findings presented so far indicate that workers are, on average, six percent less productive
in the presence of friends at high proximity. Converted into wages, this translates into a loss of
four percent in daily wage. This raises several questions: Would workers select to work alongside
friends even at the cost of losing potential income? To what extent are workers willing to forgo
wage in order to work alongside friends? Is there a relationship between how much a worker values
socializing with friends and worker characteristics, such as ability and the Big Five personality
factors, and the value placed on working alongside friends? In order to explore these questions, I
32
For succinctness, all full reports on regression estimates, including parameter estimates for low and medium
proximities, are provided in Appendix D.
33
In case of more than one friend at a certain proximity I use the mean of friends’ abilities.
30
draw on the model introduced in Section 4.
To derive estimates on how much workers value working near friends, I estimate a structural
model using data from both the experiment period and the pre-experiment period. In the first stage,
I use experiment period data to derive estimates of each worker’s cost of effort parameters when
working with friends and when working without friends. In the second stage, I estimate worker’s
valuation of working with friends using data on worker positions during the pre-experiment period.
Then, I test hypotheses related to non-cognitive skills based on earlier findings in the economics
and personality psychology literature.
6.1
Estimating Worker’s Valuation on Working Alongside Friends
To estimate how much workers value working alongside friends, I present a structural model based
on the framework introduced in Section 4. Since effort is not directly observable I substitute effort
with productivity. Assume that the wage benefit function is of the following CRRA type,
W (y) =
(h + ρ · y)1−δ
1−δ
(13)
where h is the fixed wage (per hour), ρ is the piece rate (per kilogram), and y is productivity
(kilograms per hour). In the estimation, I set δ equal to 21 .34 Cost function is a quadratic function
of productivity,
C(y, θ, f ) =



1 f 2
2θ y
if friend is present alongside the worker,


1 nf 2
2θ y
otherwise.
(14)
where θf and θnf are the cost of effort parameters. I substitute θf with the inverse of the estimate
on the worker fixed effect when working alongside friends, α̂iH , and θnf with the inverse of the
estimate on the worker fixed effect when working without friends, α̂i . This way the cost of effort
parameters are interpreted as the reciprocals of the worker’s permanent productivity, or skill, when
working alongside friends and when working without friends, respectively. In Appendix B, I present
34
Estimating with different values of δ provide similar results. The results are available upon request.
31
the estimation procedure for αi and αiH .
I model choices on work positions using a random utility model assuming that error terms,
(fi , nf
i ), are distributed i.i.d according to a Type-1 extreme value distribution. Then, from equations (3)-(6), I obtain the probability of worker i working alongside a friend as follows,
Pr(Fi = 1) = Λ(si − xi )
(15)
f
f
where Λ is the cdf of the logistic distribution, xi = (winf − cnf
i ) − (wi − ci ). The right hand side
indicates the worker’s preference to socialize with friends based on the cost and benefit she incurs.
The maximum likelihood estimator for si in equation (15) is obtained in explicit form as
T
1X
I{Fit =1}
si = xi + logit
T
!
(16)
t=1
where I{Fit =1} is an indicator variable equal to one if on day t worker i has a friend at high proximity,
and zero otherwise.35
Calculating si in equation (16) requires knowledge on xi , and Fi1 , . . . , FiT . I use individual
worker data from the experiment period to derive estimates on xi ’s.36 For data on the probability
of working alongside friends, Fi1 , . . . , FiT , I use worker position records from the pre-experiment
35
To see this, the log-likelihood function for worker i is
!
T
1 X
I{Fit =1} · log Λ(si − xi ) +
log L ∝
T t=1
T
1 X
1−
I{Fit =1}
T t=1
!
· log(1 − (Λ(si − xi ))
(17)
λ(si − xi )
= 0.
1 − Λ(si − xi )
(18)
and the first order condition is
d logL
=
d si
T
1 X
I{Fit =1}
T t=1
!
λ(si − xi )
·
−
Λ(si − xi )
T
1 X
1−
I{Fit =1}
T t=1
!
·
Rearranging equation (18), I obtain
T
1 X
I{Fit =1} = Λ(si − xi ).
T t=1
Under the assumption that the scale parameter of the logistic distribution is one, I can derive si as
!
T
1 X
si = xi + logit
I{Fit =1}
T t=1
(19)
(20)
since logit function is the inverse of the cdf of the logistic distribution.
36
Specifically, since xi is unobservable I approximate xi with x̂i which is the difference in mean net utility, excluding
valuation on working with friends, between when working with friends and when working without friends at high
32
period. However, because the number of positions in the processing room is fixed, this may have
imposed a constraint on the worker’s choice set of available positions. For instance, a worker may
have wanted to work alongside a friend but in case there is no empty position she may not have been
able to. Moreover, for an indifferent worker who chooses her position randomly, the probability of
observing that worker alongside a friend will be based on how many of her friends and non-friends
are present in the room. Ideally, this can be addressed if I have data on all available positions
whenever a worker makes a decision on where to work at the start of the day. Unfortunately, I did
not collect this type of data. Instead, I adjust a worker’s observed probability by the difference
between 0.5, the median of the logistic distribution, and the predicted probability of working
alongside a friend when positions are randomized.37
Table 14 provides summary statistics on the probability of working with at least one friend
at various levels of proximity during the pre-experiment period. Column 1 reports the mean
probability of working with at least one friend in each of the proximities when at least one friend
was present at the room. On average, workers work with at least one friend in high proximity half
of the time, or in 9 out of 18 workdays. Column 2 reports the mean of predicted probabilities
derived from 300 simulations of worker positions assuming random assignment. The differences
between observed and randomized probabilities are large and significant.
Since I do not observe differences in productivity when friends are working at low or medium
proximities relative to when no friend is present, I only focus on workers selecting into working
with friends at high proximity. Figure D2 plots cumulative distribution functions of both observed
and predicted probabilities of the presence of at least one friend at high proximity in the preexperiment period. The cumulative distribution function obtained from observed data lies largely
to the right of the cumulative distribution function generated by simulations of random worker
proximity. That is,
x̂i =
Tnf
Tf
1
1
2 o
2 o
1 Xn
1 Xn
1
1
2 h + ρ · yit 2 − θinf yit
2 h + ρ · yik 2 − θif yik
−
Tnf t=1
2
Tf
2
(21)
k=1
where t = 1, . . . , Tnf denote days without friends at high proximity and k = 1, . . . , Tf denote days with friends at
high proximity.
37
To derive the predicted probability, I run 300 simulations of randomly assigning workers to positions for all
workdays in the pre-experiment period. I exclude days on which there was no friend present in the room.
33
positions. A Kolmogorov-Smirnov test of equality significantly rejects at the 1 percent level that
the two distributions are equal. Overall evidence suggests that on average workers had a tendency
to work alongside friends before the experiment.
6.2
Worker’s Valuation on Socializing with Friends
Figure 5 presents the empirical cumulative distribution function of estimates of worker’s valuation
on socializing with friends. Approximately 10 percent of the workers show a negative value on
working alongside friends. A comparison on the cost and likelihood of working alongside friends
between workers with positive and negative values show that workers with negative values are those
who are more productive, and therefore earn more, when working alongside friends relative to when
no friend is alongside but, yet, choose to work without friends less frequently than the probability
predicted under randomization.
This study defines two workers as friends if either one of them reported the other as a friend
in the baseline survey. While 55 percent of the reported friendships were mutually reported, 45
percent were unilateral reports. Including unilaterally reported friendships may be lead to an
overestimation of worker’s valuation of socializing if only the reporting worker enjoys the presence
of the reported friend. To be conservative, I restrict the friend sample to only mutually reported
friendships and repeat the estimation on workers’ valuation of socializing. Figure D3 of the Online
Appendix presents the empirical CDF. The Kolmogorov-Smirnov equality of distribution test fails
to reject that the unrestricted (all friendships) and the restricted (mutual friendships) distributions
are equal (p-value = 0.994). This suggests that the estimates in Figure 5 are not overestimates.
For a better understanding on which types of workers are more likely to possess higher valuations on socializing with friends, I next check how individual estimates on the value of socializing
with friends depends on worker characteristics, including non-cognitive skills. I run separate ordinary least squares regressions of the estimate of si on processing skills and non-cognitive skills,
controlling for background characteristics in all regressions. Regression results are reported in Table 15. Column 1 indicates that workers who are married and with more experience on the job
are likely to place less value on socializing with friends at work. Column 2 suggests a negative
34
relationship, though not statistically significant, between processing skill and value of socializing
with friends. In terms of the Big Five, Column 3 suggests that workers with high conscientiousness
scores are likely to place lower values on socializing at work. This is consistent with the description
of conscientiousness as possessing skills to exert self-discipline and thus being industrious. However, I do not find a significant association between extraversion and the value of socializing. A
possible explanation is that extroverts are sociable to non-friends as much as they are to friends
which makes them not necessarily more likely to have higher values on socializing with friends.
Next, I use estimates on value of socializing to obtain a back-of-the-envelope calculation of the
percentage of wage that workers are willing to forgo to socialize with friends. The idea is to derive
the amount of wage loss, as a percent of wage when working without friends, that leaves a worker
indifferent between working without friends and working alongside friends, where workers derive
utility ŝi in the latter. For this exercise, I exclude workers with negative values (10 percent of
the sample). Figure 6 presents a histogram of the percentage of wage that workers are willing to
forgo to socialize with friends. Conditional on positive values, an average worker is willing to forgo
approximately 6 percent of her wage.38 Workers in the top quartile are willing to forgo at least 8
percent of their wage.
6.3
Are High-Cost (Benefit) Workers Less (More) Likely to Work Alongside
Friends?
In this section, I explore the relationship between the cost, measured in terms of the potential
wage loss, and the likelihood to work alongside friends. In a usual price-demand framework, we
would expect lower demand under higher prices (a downward slope on the price estimate). In this
context, this is not clear. Price, or cost, is determined by a worker’s choice on how much effort she
would reduce in order to socialize when a friend is alongside. This implies that cost can also be
interpreted as a measure of how much the person enjoys socializing; the more one enjoys, the more
she will use her time socializing. Without an instrument for cost, such as an exogenous variation
38
For comparison, in the garment factory setting of Hamilton et al. (2003), high productivity workers under
individual production switched to team production at the cost of forgoing approximately 8 percent of their wage.
Team production enabled workers to socialize with coworkers in their team but, at the same time, accompanied others
changes as well; for instance, standing instead of sitting in chairs and specialization in production.
35
in the piece rate, the direction of how demand is related to cost in this setting is ambiguous. As
a result, the goal here is not to test the sign of the slope per se but to examine whether there is
heterogeneity in the slope estimates between groups with different levels of skills.
To first examine average cost-demand behavior, I estimate the following equation:
Pr(F = 1) = η · Costi + Xi + i
(22)
where Costi is constructed as the percentage change in expected hourly wage when working alongside friends relative to when working without friends.39 Xi is a vector of worker covariates: worker’s
age, education level, job tenure, marital status, number of reported friends, and weekly time spent
with friends outside the workplace. The dependent variable is the probability that worker i works
alongside friends. For this measure, I use the fraction of days worker i works alongside friends
during the pre-experiment period.40
Regression estimates are reported in Table 16. To account for sampling variability of the Cost
variable, I use bootstrapped standard errors from a Bayesian parametric bootstrap procedure introduced in Mas and Moretti (2009). Column 1 indicates no significant association between the
cost and likelihood to work alongside friends. Column 2 uses a sample restricted to workers with
more than 3 months of experience at their current job. The estimate is similar to the previous
column suggesting that workers with more experience are not likely to have had responded differently compared to less experienced workers. Also, people with more years of education or better
processing skills may be more likely to make sound judgements or have better perceptions on the
productivity loss when working alongside friends. Column 3 shows that more-educated workers
are not significantly different from less-educated workers in terms of choosing to work alongside
friends. Likewise, from Column 4, I do not find a significant difference on the cost coefficient
39
More specifically, Costi is derived using the following formula:
Costi =
(h + ρ · α̂i ) − (h + ρ · α̂iH )
× 100
(h + ρ · α̂i )
where h and ρ are the fixed hourly wage and piece rate wage, respectively. α̂i is worker i’s fixed effect estimate when
working without friends and α̂iH is worker i’s fixed effect estimate when working alongside friends.
40
Using only mutually reported friendships does not change the results significantly. Results are available from
author, upon request.
36
between high-ability and low-ability workers.
Decision making may depend on the person’s psychological attributes such as non-cognitive
skills.41 In particular, personality psychology studies suggest that two of the Big Five Factors are
linked to playing a role in the process of decision making and executing. First, conscientiousness is
associated with exerting greater self-discipline over oneself and premeditating one’s actions and consequences (John and Srivastava, 1999; Whiteside and Lynam, 2001). These attributes are expected
to help people make decisions that are less impulsive and forward-looking. Second, neuroticism is
known to have positive associations with making impulsive decisions (Costa and McCrae, 1992)
and negatively associated with internal locus of control (Judge et al., 2002).42
Columns 5-9 each represent a separate regression with the Big Five Factor of interest interacted
with the Cost variable. Significance levels are corrected by using a Holm-Bonferroni method to
account for multiple hypotheses testing. To be conservative, I control for the Family Wise Error
Rate at the level of all five factors, one for each non-cognitive skill.43 Regarding extraversion
and agreeableness, I do not find significant differences between high-skilled and low-skilled groups.
Column 7 suggests that workers with higher scores on the conscientiousness scale are overall less
likely to work alongside friends than workers with below-average scores. Yet, the cost coefficient
do not differ significantly across the workers in low and high conscientious groups.
Column 8 indicates that high neuroticism workers respond to the cost in a manner significantly
different from low neuroticism workers. The estimates imply that among low neuroticism workers
a one percentage point increase in potential wage loss is associated with a drop in the likelihood of
working alongside friends by two basis points; high neuroticism workers show a three basis point
increase in the likelihood when the cost increases by one percent. Based on a worker with a five
percent wage loss when working alongside friends, the cost-demand estimate implies that a low41
The idea that people’s decisions may depend on psychological factors is well noted in the seminal work of
Kahneman and Tversky (1979).
42
According to Costa and McCrae (1992), people who are more impulsive tend to make decisions in favor of their
instant emotions. Locus of control, conceptually developed by Rotter (1966), describes people who show passive
responses to negative outcomes as to having weak internal locus of control.
43
A less conservative approach to controlling the FWER rate would be to use the free step-down resampling
method (Westfall and Young, 1993) or, even less conservative, controlling for the False Discovery Rate (Anderson,
2008). FWER adjusted p-values obtained from using the free step-down resampling method are available from the
author upon request.
37
neurotic worker earns more than a high-neurotic worker by 1.2 percent due to sorting. Finally,
Column 9 shows that workers with high openness are also less likely to work alongside friends but
do not indicate significant differences on the cost coefficients between the low-skill and high-skill
groups.
The relationship between response to cost and neuroticism is in line with the finding from Callen
et al. (2015) that low neuroticism inspectors and senior health officials show a larger increase in
their performance levels relative to high neuroticism counterparts when the cost of shirking, through
an increase in the probability of being detected, is exogenously increased through implementing
a monitoring technology.44 While this relationship may suggest that high neuroticism workers
make choices that are more reflective of their innate desires or emotions (Costa and McCrae,
1992), in this case the consumption value of socializing, there might be other possible reasons. One
plausible explanation is that neuroticism is negatively correlated with perception. For example, low
neuroticism workers may have better perceptions on the productivity loss from working alongside
friends than high neuroticism workers. Although I do not have data to directly test this hypothesis, I
instead show in Table D4 of the Online Appendix that a high neuroticism worker’s ability to perceive
a friend’s processing skill is at least as good as that of a low neuroticism worker. This may arguably
suggest that the ability to perceive productivity differences between days when working alongside
friends and days when not working alongside friends is not what is mainly driving this relationship.
Another possibility which I am not able to rule out is that neuroticism is negatively correlated with
sophisticated thinking—being aware of her future self-control problems (O’Donoghue and Rabin,
2001). For instance, if low neuroticism workers are more sophisticated than high neuroticism
workers then we would find more commitment type behaviors (choosing to work without friends)
among low neuroticism workers compared to high neuroticism workers.
44
It also relates to findings in the personality psychology literature that high levels of neuroticism are positively
correlated with making choices that consistently yield lower payoffs in a laboratory experiment with gambling tasks
(Denburg et al., 2009).
38
6.4
Does Conscientiousness Actually Predict Job Performance?
Numerous meta-analytic studies in the personality psychology literature report strong and positive
correlations between conscientiousness and job performance (Barrick and Mount, 1991; Barrick
et al., 1998, 2003; Salgado, 1997; Schmidt and Hunter, 1998). This relationship has been corroborated in a recent economic study (Callen et al., 2015). I test this relationship by examining whether
high conscientious workers were actually more productive during the pre-experiment period. Next,
I further explore how this relationship can be explained by socializing behaviors through exploiting data from the two periods; pre-experiment period data on worker’s choices to work alongside
friends and regression estimates on productivity declines when working alongside friends obtained
from the experiment period.
First, to estimate the productivity difference between high and low conscientious workers I run
the following regression using worker level data from the pre-experiment period:
yi = ψs · Non-cognitive Skilli + Xi + i
(23)
where yi is the log average productivity of worker i, Non-cognitive Skilli is a vector of indicator
variables with each variable representing one of the Big Five factor and equal to one if the skill of
worker i is above the median. Xi is a vector that includes variables for the worker’s marital status,
education level, and quadratic functions of worker’s age and experience.
Table D5 in the Online Appendix reports regression estimates. Column 1 suggests that workers
with high conscientiousness were 6.7 percent more productive relative to low conscientiousness
workers during the pre-experiment period. Column 2 shows a wage gap of 4.7 percent between
high and low conscientiousness workers. To better understand why conscientious employees are
more productive, Figure 7 presents a decomposition of the channels of the productivity gap. The
probability to work alongside friends in the pre-experiment period was 0.4 and 0.63 for high and
low conscientiousness workers, respectively. Low conscientious workers show a 9 percent decline
in productivity when working alongside friends as opposed to the 3 percent decline among high
conscientiousness workers. Taking these numbers together, a high conscientiousness worker is 2
39
percent more productive than a low conscientiousness worker due to a lower likelihood of working
alongside friends ((0.63-0.4)×9=2.0) and 2.4 percent more productive due to a smaller productivity
decline ((0.4×(9-3)=2.4). Overall, socializing behavior explains approximately 66 percent (4.4/6.7
× 100) of the 6.7 percent productivity gap observed between high and low conscientiousness workers.
7
Conclusion
In collaboration with the management, I designed and implemented a field experiment at a seafoodprocessing plant in Vietnam. Workers were randomly assigned to different work positions on a daily
basis for five months. I find that workers are less productive when working alongside friends but not
affected when friends are working across the table or at different tables in the same room. These
findings suggest that workers are socializing with friends when they are in close proximity and, as
a result, less productive. Because workers are paid the sum of a fixed daily wage and a piece rate,
based on the amount of fish processed, they also earn less when working alongside friends.
Additionally, I find that, conditional on placing a positive value on socializing, an average
worker chooses to forgo approximately six percent of her wage in order to socialize with friends at
work. This result may not be surprising since working alongside friends can make repetitive work
less boring or less onerous. Another reason to socialize with friends, despite the wage loss, would
be the importance of social networks as sources of information and informal credits to individuals
living in a rural developing region.
An event study analysis on the randomized position policy suggests a 5% increase in the firm’s
average productivity. Since employees were partly paid fixed wages, this implies the possibility
that the management was not maximizing on their profit potentials. However, the random position
policy may not be as profitable in the long run. For instance, random assignment rules may lead
to lower job satisfaction and heighten the number of job turnovers. Both recruiting and training a
new worker requires time and additional supervision from managers. Physically separating friends
may also hamper transfers of skills and knowledge. While it is beyond the scope of this study, it is
an area for future research to assess whether reshaping social dynamics in the workplace can have
impacts on other dimensions of a firm’s human resources, such as the applicant pool, transfer of
40
job skills, and worker retentions.
Finally, I find that non-cognitive skills are predictive of job performance and that a large portion
of this can be explained through social interaction behaviors. Having better self-control skills, as
measured by the conscientiousness scale, is associated with a smaller decline in productivity when
working alongside friends, and a decrease in the likelihood to choose to work alongside friends.
Conversely, while workers with high- and low-conscientiousness show a significant productivity
gap of almost 7 percent, two-thirds of this difference can be explained by socializing behaviors
in the workplace. In light of growing evidence that non-cognitive skills can be developed (Kautz
et al., 2014; Blattman et al., 2015; Deckers et al., 2015), the 5 percent wage difference between
high- and low-conscientiousness workers has potential policy-relevant implications on human capital
development. It would be of interest to explore whether socializing or communication skills can be
an advantage in other contexts, such as jobs in advertising or communications.
41
References
Almlund, M., Duckworth, A., Heckman, J. J., and Kautz, T. (2011). Personaltiy Psychology and
Economics. NBER Working Paper No. w16822.
Amodio, F. and Martinez-Carrasco, M. A. (2015). Input Allocation, Workforce Management and
Productivity Spillovers: Evidence from Personnel Data. McGill University, Working Paper.
Anderson, J., Burks, S., DeYoung, C., and Rustichini, A. (2011). Toward the Integration of
Personality Theory and Decision Theory in the Explanation of Economic Behavior. mimeo.
Anderson, M. L. (2008). Multiple Inference and Gender Differences in the Effects of Early Intervention: A Reevaluation of the Abecedarian, Perry Preschool, and Early Training projects. Journal
of the American Statistical Association, 103(484):1481–1495.
Andreoni, J. and Miller, J. (2002). Giving According to Garp: An Experimental Test of the
Consistency of Preferences for Altruism. Econometrica, 70(2):737–753.
Angrist, J. and Pischke, J.-S. (2009). Mostly Harmless Econometrics: An Empiricist’s Companion.
Princeton: Princeton University Press.
Arcidiacono, P., Kinsler, J., and Price, J. (2014). Productivity Spillovers in Team Production:
Evidence from Professional Basketball. Working Paper.
Bandiera, O., Barankay, I., and Rasul, I. (2005). Social Preferences and the Response to Incentives:
Evidence From Personnel Data. Quarterly Journal of Economics, 120:917–962.
Bandiera, O., Barankay, I., and Rasul, I. (2007). Incentives for Managers and Inequality among
Workers: Evidence from a Firm Level Experiment. Quarterly Journal of Economics, 122(2):729–
74.
Bandiera, O., Barankay, I., and Rasul, I. (2009). Social Connections and Incentives in the Workplace: Evidence from Personnel Data. Econometrica, 77:1047–1094.
Bandiera, O., Barankay, I., and Rasul, I. (2010). Social Incentives in the Workplace. Review of
Economic Studies, 77:417–458.
Bandiera, O., Barankay, I., and Rasul, I. (2011). Field Experiments with Firms. Journal of
Economic Perspectives, 25(3):63–82.
Bandiera, O., Barankay, I., and Rasul, I. (2013). Team Incentives: Evidence from a Firm Level
Experiment. Journal of the European Economic Association, 11(5):1079–1114.
Banerjee, A., Chandrasekhar, A., Duflo, E., and Jackson, M. (2013). The Diffusion of Microfinance.
Science, 341(6144).
Barrick, M. and Mount, M. (1991). The Big Five Personality Dimensions and Job Performance: A
Meta-Analysis. Personnel Pyschology, 44:1–26.
Barrick, M., Mount, M., and Gupta, R. (2003). Meta-analysis of the relationship between the five
factor model of personality and Holland’s occupational types. Personnel Psychology, 56:45–74.
Barrick, M., Mount, M., and Stewart, G. (1998). Five-Factor Model of Personality and Performance
in Jobs Involving Interpersonal Interactions. Human Performance, 11(2):145–165.
Beaman, L. and Magruder, J. (2012). Who Gets the Job Referral? Evidence from a Social Networks
42
Experiment. American Economic Review, 102(7):3574–93.
Becker, A., Deckers, T., Dohmen, T., Falk, A., and Kosse, F. (2012). The Relationship Between
Economic Preferences and Psychological Personality Measures. Annual Review of Economics,
4:453–478.
Benet-Martinez, V. and John, O. (1998). Los cinco grandes across cultures and ethnic groups:
Multitrait multimethod analysis of the Big five in Spanish and English. Journal of Personality
and Social Psychology, 75:729–750.
Blattman, C., Jamison, J., and Sheridan, M. (2015). Reducing Crime and Violence: Experimental
Evidence on Adult Noncognitive Investments in Liberia. NBER Working Paper No. 21204.
Bloom, N., Liang, J., Roberts, J., and Ying, Z. J. (2015). Does Working from Home Work?
Evidence from a Chinese Experiment. Quarterly Journal of Economics, 130(1):165–218.
Bono, J. E. and Judge, T. A. (2003). Core Self-Evaluations: A Review of the Trait and its Role in
Job Satisfaction and Job Performance. European Journal of Personality, 17:S5–S18.
Borghans, L., Duckworth, A. L., Heckman, J., and ter Weel, B. (2008). The Economics and
Psychology of Personality Traits. Journal of Human Resources, 43(4):972–1059.
Bowles, S., Ginitis, H., and Osborne, M. (2001). The Determinants of Earnings: A Behavioral
Approach. Journal of Economic Literature, 39(4):1137–1176.
Callen, M., Gulzar, S., Hasanain, A., Khan, Y., and Rezaee, A. (2015). Personalities and Public
Sector Performance: Evidence from a Health Experiment in Pakistan. NBER Working Paper
No. 21180.
Carneiro, P., Crawford, C., and Goodman, A. (2007). The Impact of Early Cognitive and NonCognitive Skills on Later Outcomes. CEE Discussion Paper 92.
Cebi, M. (2007). Locus of Control and Human Capital Investment. Journal of Human Resources,
42(4):919–932.
Charness, G. and Rabin, M. (2002). Understanding Social Preferences with Simple Tests. Quarterly
Journal of Economics, 117(3):817–869.
Colquitt, J. A., LePine, J. A., and Noe, R. A. (2000). Toward and Integrative Theory of Training Motivation: A Meta-Analytic Path Analysis of 20 Years of Research. Journal of Applied
Psychology, 85(5):678–707.
Cornelissen, T., Dustmann, C., and Schoenberg, U. (2014). Peer Effects in the Workplace. Working
Paper.
Costa, P. T. and McCrae, R. R. (1992). Four ways five factors are basic. Personality and Individual
Differences, 13(6):653–665.
Deckers, T., Falk, A., Kosse, F., and Schildberg-Horisch, H. (2015). Formation of human prosociality: Causal evidence on the role of sociocultural environment. Working Paper.
Denburg, N. L., Weller, J. A., and coauthors (2009). Poor Decision Making Among Older Adults
is Related to Elevated Levels of Neuroticism. Annals of Behavioral Medicine, 37(2):164–172.
Dohmen, T., Falk, A., Huffman, D., and Sunde, U. (2008). Representative Trust and Reciprocity:
Prevalence and Determinants. Economic Inquiry, 46(1):84–90.
43
Dohmen, T., Falk, A., Huffman, D., and Sunde, U. (2010). Are Risk Aversion and Impatience
Related to Cognitive Ability? American Economic Review, 100(3):1238–1260.
Duckworth, A. L. and Seligman, M. (2005). Self-Discipline Outdoes IQ in Predicting Academic
Performance of Adolescents. Psychological Science, 16(12):939–944.
Duncan, G., Dowsett, C., Claessens, A., and coauthors (2007). School Readiness and Later Achievement. Developmental Psychology, 43(6):1428–1446.
Fafchamps, M. and Lund, S. (2003). Risk-sharing networks in rural Philippines. Journal of Development Economics, 71:261–287.
Falk, A. and Ichino, A. (2006). Clean Evidence on Peer Effects. Journal of Labor Economics,
24(1):39–57.
Frank, R. (1985). Choosing the right pond: Human behavior and the quest for status. New York,
NY, US: Oxford University Press.
Gati, I., Gadassi, R., Saka, N., Hadadi, Y., Ansenberg, N., Friedmann, R., and Asulin-Peretz,
L. (2011). Emotional and Personality-Related Aspects of Career Decision-Making Difficulties:
Facets of Career Indecisiveness. Journal of Career Assessment, 19(1):3–20.
Granovetter, M. (1973). The Strength of Weak Ties. American Journal of Sociology, 78(6):1360–
1380.
Hamilton, B., Nickerson, J., and Owan, H. (2003). Team Incentives and Worker Heterogeneity:
An Empirical Analysis of the Impact of Teams on Productivity and Participation. Journal of
Political Economy, 111(3):465–497.
Heath, R. (2015). Why do Firms Hire using Referrals? Evidence from Bangladeshi Garment
Factories. Mimeo. University of Washington.
Heckman, J., Moon, S. H., Pinto, R., Savelyev, P., and Yavitz, A. (2010). Analyzing social experiments as implemented: A reexamination of the evidence from the HighScope Perry Preschool
Program. Quantitative Economics, 1:1–46.
Heckman, J. J. and Rubinstein, Y. (2001). The Importance of Noncognitive Skills: Lessons from
the GED Testing Program. American Economic Review: Papers and Proceedings, 91(2):145–149.
Heckman, J. J., Stixrud, J., and Urzua, S. (2006). The Effects of Cognitive and Noncognitive Abilities on Labor Market Outcomes and Social Behavior. Journal of Labor Economics, 24(3):411–482.
Heineck, G. (2011). Does it Pay to be Nice? Personality and Earnings in the United Kingdom.
Industrial and Labor Relations Review, 64(5):1020–1038.
Heineck, G. and Anger, S. (2010). The Returns to Cognitive Abilities and Personality Traits in
Germany. Labour Economics, 17:535–546.
Hjort, J. (2014). Ethnic Divisions and Production in Firms. Quarterly Journal of Economics,
129(4):1899–1946.
Hogan, J. and Holland, B. (2003). Using Theory to Evaluate Personality and Job-Performance
Relations: A Socioanalytic Perspective. Journal of Applied Psychology, 88(1):100–112.
Hough, L., Eaton, N., Dunnette, M., Kamp, J., and McCloy, R. (1990). Criterion-Related Validities
of Personality Constructs and the Effect of Response Distortion on those Validities. Journal of
44
Applied Psychology, 75(5):581–595.
John, O., Donahue, E., and Kentle, R. (1991). The Big Five Inventory–Versions 4a and 54. Berkeley,
CA: University of California, Berkeley, Institute of Personality and Social Research.
John, O., Naumann, L., and Soto, C. (2008). Paradigm shift to the integrative Big Five trait
taxonomy: History, measurement, and conceptual issues. In John, O., Robins, R., and Pervin,
L., editors, Handbook of personality: Theory and research. Guilford Press, New York, NY.
John, O. P. and Srivastava, S. (1999). The Big Five Trait Taxonomy: History, Measurement, and
Theoretical Perspectives. In Pervin, L. A. and John, O. P., editors, Handbook of personality:
Theory and research, pages 102 – 138. Guilford Press, New York, NY.
Judge, T. A., Erez, A., and Bono, J. E. (2002). Are Measures of Self-esteem, Neuroticism, Locus
of Control, and Generalized Self-Efficacy Indicators of a Common Core Construct? Journal of
Personality and Social Psychology, 83(3):693–710.
Kahneman, D. and Tversky, A. (1979). Prospect Theory: An Analysis of Decision Under Risk.
Econometrica, 47(2):263–292.
Kandel, E. and Lazear, E. (1992). Peer Pressure and Partnerships. Journal of Political Economy,
100(4):801–817.
Kautz, T., Heckman, J., Diris, R., Weel, B. t., and Borghans, L. (2014). Fostering and Measuring Skills: Improving Cognitive and Non-Cognitive Skills to Promote Lifetime Success. NBER
Working Paper Series 20749.
Larson, L., Rottinghaus, P., and Borgen, F. (2002). Meta-analyses of Big Six interests and Big
Five personality factors. Journal of Vocational Behavior, 61:217–239.
Lee, D. S. (2009). Training, Wages, and Sample Selection: Estimating Sharp Bounds on Treatment
Effects. Review of Economic Studies, 76(3):1071–1102.
Levitt, S. and List, J. (2007). What do Laboratory Experiments Measuring Social Preferences
Reveal About the Real World. Journal of Economic Perspectives, 21(2):153–174.
Mas, A. and Moretti, E. (2009). Peers at Work. American Economic Review, 99(1):112–145.
Montgomery, J. (1991). Social Networks and Labor-Market Outcomes: Toward an Economic
Analysis. American Economic Review, 81(5):1408–1418.
Mueller, G. and Plug, E. (2006). Estimating the Effect of Personality on Male and Female Earnings.
Industrial and Labor Relations Review, 60(1):3–22.
Nyhus, E. K. and Pons, E. (2005). The effects of personality on earnings. Journal of Economic
Psychology, 26:363–384.
O’Donoghue, T. and Rabin, M. (2001). Choice and procrastination. Quarterly Journal of Economics, 116(1):121–160.
Ones, D., Dilchert, S., Viswesvaran, C., and Judge, T. (2007). In Support of Personality Assessment
in Organizational Settings. Personnel Psychology, 60:995–1027.
Ones, D., Viswesvaran, C., and Reiss, A. (1996). Role of Social Desirability in Personality Testing
for Personnel Selection: The Red Herring. Journal of Applied Psychology, 81(6):660–679.
Rosen, S. (1986). The Theory of Equalizing Differences. In Ashenfelter, O. and Layard, R., editors,
45
Handbook of Labor Economics. North Holland, Amsterdam.
Rotter, J. B. (1966). Generalized Expectancies for Internal Versus External Control of Reinforcement. Psychological Monographs: General and Applied, 80(1):1–28.
Salgado, J. (1997). The Five Factor Model of Personality and Job Performance in the European
Community. Journal of Applied Psychology, 82(1):30–43.
Schmidt, F. L. and Hunter, J. E. (1998). The Validity and Utility of Selection Methods in Personnel Psychology: Practical and Theoretical Implications on 85 Years of Research Findings.
Psychological Bulletin, 124(2):262–274.
Schmitt, D., Allik, J., McCrae, R., and Benet-Martinez, V. (2007). The Geographic Distribution
of Big Five Personality Traits. Journal of Cross-Cultural Psychology, 38(2):173–212.
Semykina, A. and Linz, S. (2007). Gender differences in personality and earnings: Evidence from
Russia. Journal of Economic Psychology, 28(3):387–410.
Shafer, A. (2000). Mediation of the Big Five’s effect on career decision making by life task dimensions and on money attitudes by materialism. Personality and Individual Differences, 28:93–109.
Udry, C. (1994). Risk and insurance in a rural credit market: an empirical investigation in Northern
Nigeria. Review of Economic Studies, 61(3):495–526.
Urzua, S. (2008). Racial Labor Market Gaps: The Role of Abilities and Schooling Choices. Journal
of Human Resources, 43(4):919–971.
Westfall, P. S. and Young, S. S. (1993). Resampling-Based Multiple Testing. New York: Wiley.
Whiteside, S. P. and Lynam, D. R. (2001). The Five Factor Model and impulsivity: using a
structural model of personality to understand impulsivity. Personality and Individual Differences,
30(4):669–689.
Wooldridge, J. (2003). Cluster-Sample Methods in Applied Econometrics. American Economic
Review, 93:133–138.
46
Figure 1: Exhibit of Processing Rooms
(a) Processing Room 1
(b) Processing Room 2
Figure 2: Spatial Contiguity and Proximity
G
A
G
B
H
B
H
C
I
C
I
D
J
D
J
Width = 111 cm
Length = 233 cm
A
<Table Dimensions>
Proximity to
focal workers B, L
E
K
E
K
F
L
F
L
Low
Medium
High
(a) Spatial Contiguity
(b) Spatial Proximity
Notes: Panel (a) - Areas enclosed by dotted lines represent work spaces that are spatially contiguous with worker B and L,
respectively. Note that worker B is at an interior position and, therefore, contiguous with more work spaces than worker L, who
is at a corner position. Panel (b) - Work positions with different levels of proximity to a focal worker (B or L) are shaded in
different brightnesses. Among work spaces that are contiguous with the focal workers, the closest positions are shaded in dark
blue and positions with the least proximity are shaded in light blue. Positions that are not spatially contiguous with either
worker B or L are white.
47
9
Fraction of Workers Alongside Friend
.2
.4
.6
.8
Figure 3: Worker Positions and Productivity across Time
5
6
7
8
Average Productivity (kgs/hr)
− Start Experiment
Jul 1, 2014
Aug 1, 2014 Sep 1, 2014 Oct 1, 2014
Nov 1, 2014 Dec 1, 2014
Working Alongside Friend
Linear Fit of Working Alongside Friend
Locally Weighted Smoothing of Daily Average Productivity
0
Productivity Response to Policy
.1
.2
.3
.4
Figure 4: Impact of Policy Intervention
0
5
10
Bandwidth Size (Day to and from Implementation)
Coefficient Estimate
95% Confidence Interval
48
15
0
.2
Cumulative Density
.4
.6
.8
1
Figure 5: Cumulative Distribution Function of Estimates on Value of Socializing with Friends
−10
0
10
Value of Socializing at Work
20
30
0
5
Frequency
10
15
Figure 6: Utility Value of Socializing Converted Into Percent of Wage
0
5
10
Percentage of Wage (%)
15
20
Notes: A worker’s utility value of socializing (ŝi ) is converted into wage as the amount of wage loss (zi ), as a proportion of
wage when working without friends, that leaves the worker indifferent between working without friends and alongside friends.
Formally, I solve for zi in
q
q
2
Wagenf
i −
Wagenf
i (1 − zi )
= ŝi
for each worker, where utility from wage is assumed to be a CRRA type with the relative risk parameter equal to
with positive productivity gains when working alongside friends are not included in this graph.
49
1
.
2
Workers
Productivity Gap Between
High and Low Conscientiousness Workers (%)
0
2
4
6
8
Figure 7: Socializing Behavior Explains Significant Portion of Productivity Difference
2.4%
6.7%
Total
productivity gap
2%
2.3%
Due to lower
productivity
loss when
near friend
50
Due to
choosing to
work near friends
less often
Due to factors
other than
socializing
Table 1: Summary on Survey Participation (Number of Workers)
Target Pop. Size
Total Surveyed
Participation from
Target Pop.
Participation Rate
114
105
119
101
112
101
98%
96%
Baseline
Endline
Notes: For the baseline survey, Target Population Size is the number of processing workers employed at the company at the time
of survey. For the endline survey, target population size is the number of processing workers, including those who quit during the
study period, who participated in the baseline survey and worked under randomized position assignments. Total Surveyed is the
number of workers surveyed during each survey period. In the baseline survey, there were participation from workers who were
not processing workers. Our survey team misunderstood the employee list from the firm. As a result, we mistakenly surveyed
these workers. Participation from Target Population is the number of processing workers that participated in each of the surveys.
Participation Rate is calculated from (Participation from Target Pop.)/(Target Pop. Size).
Table 2: Summary Statistics - Socioeconomic Variables (N = 112)
Mean
Variable
Panel A. Socioeconomics
Female
Married
Completed secondary school
Age (years)
Tenure at current job (months)
Experience in fish processing (months)
Panel B. Job Search Experience
Learned about job opening through
Friend
Family member
Ex-coworker
Job advertisements
Received help when applying to job
Conditional on help received
Helper is currently working at plant
Relationship with helper
Friend
Family member
Ex-coworker
Type of Help
Information on job opening
Information on job details
Recommendation to manager
1
0.67
0.49
31.43
(9.37)
20.30
(17.40)
33.25
(32.40)
Lives in house with
Water pipe connection
Tiled floors
Cable TV
Refrigerator
Internet connection
Owns motorcycle
Mean
Variable
0.72
0.69
0.56
0.54
0.11
0.59
0.75
0.21
0.13
0.06
0.40
0.90
0.52
0.33
0.13
0.64
0.27
0.05
Notes: Data from survey module 1.
Table 3: Frequency counts on the number of friends reported and the number of times mentioned as a
friend by another surveyed worker (N = 112)
Panel A
Number of friends
reported by worker
Panel B
Number of times
mentioned as a friend
Frequency
Frequency
0
1
2
3
4
5
6
7
8
2
2
8
29
31
28
7
4
1
0
1
2
3
4
5
6
7
8
1
10
14
29
19
17
14
7
1
Mean
Median
S.D.
3.99
4
1.40
Mean
Median
S.D.
3.81
4
1.75
Notes: Data from survey module 2.
51
Table 4: Characteristics of Reported Friendships (N = 112)
Panel A. Basic characteristics of reported friendships
Full sample
Mean
(S.D.)
Total number of reported friendships in sample∗
Friendship formed before employment at current job
Respondent learned processing skills from reported friend
Respondent lives with reported friend
Weekly time spent with reported friend outside workplace (hours)
287
0.15
0.54
0.05
2.57
(3.57)
16.60
(18.33)
3.79
(0.96)
Duration of friendship (months)
Respondent’s subjective rating of closeness with reported friend
(on a five-point scale with 5 being very close)
Panel B. Summary statistics on shared activities in reported friendships†
Frequency in the past three months‡
Shared activities
Very
often
(1)
Go shopping together
Talk during break or lunch times
Work together at same table
Help each other at work
Give advice on personal matters to this person
Receive advice on personal matters from this person
Lent or borrowed money from one other
17.4
33.8
39.4
38.3
24.0
23.3
(2)
Sometimes
(3)
10.5
32.8
35.5
29.6
18.1
14.0
7.3
3.1
10.8
8.7
8.4
11.2
9.8
12.5
Often
(4)
Almost
never
(5)
10.1
8.4
3.5
6.6
9.4
9.4
18.8
58.9
14.3
12.9
17.1
37.3
43.6
61.0
Rarely
∗ Mutually reported friendships are counted as one friendship.
† This part of the table uses the full sample of reported friendships.
‡ All items, except lent or borrowed money from one other, which uses a four point scale, use a five point response
scale. In the survey sheet, response semantics differed across items to account for natural differences in the frequency
of activities. For the questions, how often did you go shopping together, how often did you give advice on personal
matters, and how often did you receive advice on personal matters workers could choose from (1) more than once a
week (2) once a week (3) once every two weeks (4) once a month (5) less than once a month. For questions how often
did you talk to each other during break or lunch, how often did you work together at same table, and how often did you
help each other’s work, workers could choose from (1) more than once a day (2) once a day (3) once every two days
(4) once a week (5) less than once a week. For the question how many times did you lend or borrow money with each
other, workers could choose from (1) more than 5 times (2) 2-5 times (3) once (4) never.
52
Table 5: Definitions and Facets of the Big Five Domains of Personality
Trait
Conceptual Definition
Facets
Extraversion
Energetic approach toward the social and material world
Assertiveness/Leadership
Gregariousness
Social Confidence
Agreeableness
Prosocial and communal orientation toward others
Altruism
Empathy/Sympathy
Modesty
Trust
Conscientiousness
Socially prescribed impulse control that facilitates task and goal-directed behaviors
Industriousness
Orderliness
Self-Discipline
Neuroticism
Negative emotionality
Anxiety
Depression
Irritability
Rumination-Compulsiveness
Openness
Breadth, depth, originality, and complexity of an individual’s mental and experimental life
Adventurousness
Idealism
Intellectualism
Source: John and Soto, 2008
Table 6: Summary Statistics - Big Five Personality Factors (N = 112)
Correlation Matrix
Variable
Extraversion
Agreeableness
Conscientiousness
Neuroticism
Openness
Mean
S.D.
Extra.
Agree.
Consc.
Neuro.
Open.
3.58
4.07
3.84
2.61
2.75
0.56
0.70
0.60
0.71
0.60
1.00
0.36
0.36
-0.40
0.27
1.00
0.63
-0.51
-0.03
1.00
-0.49
0.01
1.00
0.07
1.00
Notes: Data from survey module 3.
Table 7: Summary Statistics - Proximity to Friends (Experiment Period)
Probability (conditional on)
Proximity
Contiguous
Low Prox
Med Prox
High Prox
unconditional
0.413
(0.492)
0.198
(0.398)
0.116
(0.321)
0.166
(0.372)
(Cont. = 1)
(Low = 1)
(Med = 1)
(High = 1)
0.478
(0.5)
0.282
(0.45)
0.401
(0.49)
1
0.211
(0.409)
1
0.177
(0.382)
0.09
(0.286)
1
0.125
(0.33)
0.148
(0.356)
0.127
(0.334)
Notes: This table unconditional and conditional probabilities of having at least one friend at different
levels of proximities. Probabilities are derived from taking the mean of the indicator variables over all
observations (n = 5731). Standard deviations are presented in parentheses.
53
Table 8: Baseline Results - OLS Estimates
Dependent Var.: log(productivity)
Pre-Experiment
(1)
Experiment
(2)
(3)
−0.049∗∗∗
(0.011)
Contiguous
(4)
−0.019∗∗∗
(0.005)
−0.006
(0.008)
−0.010
(0.008)
−0.075∗∗∗
(0.007)
Low Prox
Med Prox
High Prox
Adjusted R2
s.e clustered by
observations
0.65
worker
1,839
0.67
worker
1,839
0.002
(0.005)
0.001
(0.008)
−0.056∗∗∗
(0.006)
0.43
worker
5,731
0.44
worker
5,731
Notes: All regressions include worker fixed effects, room×day fixed effects, and
control variables at the individual spatial level, including the number of workers
working at contiguous positions and the mean of their average productivities. *
p < .10, ** p < .05, *** p < .01.
Table 9: IV Estimates
First stage results
Endogenous variables
Instruments
Low Prox
Med Prox
High Prox
Assigned Low Prox
0.840∗∗∗
(0.013)
−0.001
(0.006)
−0.014 ∗ ∗
(0.006)
0.003
(0.004)
0.862∗∗∗
(0.016)
−0.008
(0.005)
−0.002
(0.004)
−0.022∗∗∗
(0.007)
0.838∗∗∗
(0.012)
4549.7
2901.2
4962.6
Assigned Med Prox
Assigned High Prox
F-statistic
IV/2SLS
Variable
log(prod.)
log(wage)
Low Prox
-0.002
(0.006)
0.004
(0.013)
-0.058***
(0.009)
−0.001
(0.004)
0.002
(0.009)
-0.040***
(0.006)
5,731
5,731
Med Prox
High Prox
observations
Notes: Dependent variables in the bottom panel are the worker’s
log productivity (kilograms/hour) and log hourly wage (daily
wage/hour).
All regressions include worker fixed effects and
room×day fixed effects and control variables at the individual spatial
level, including the number of workers working at contiguous positions and the mean of their average productivities. Standard errors
are two-way clustered by worker and room×day. * p < .10, ** p < .05,
*** p < .01.
54
Table 10: Robustness Checks
IV Estimates - Dependent Variable: log(productivity)
Low Prox
Med Prox
High Prox
P
Low Proxj
P
j∈C (i)
Med Proxj
P
j∈C (i)
High Proxj
j∈C (i)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
-0.002
(0.006)
0.004
(0.014)
-0.060***
(0.010)
0.001
(0.004)
0.000
(0.004)
-0.009*
(0.005)
-0.007
(0.013)
0.004
(0.014)
-0.057***
(0.012)
-0.005
(0.006)
-0.002
(0.015)
-0.063***
(0.010)
-0.002
(0.007)
0.003
(0.013)
-0.060***
(0.011)
0.002
(0.007)
0.003
(0.008)
-0.059***
(0.008)
0.006
(0.010)
0.004
(0.010)
-0.055***
(0.011)
0.011
(0.009)
-0.008
(0.021)
-0.068***
(0.014)
Low Prox × Same Table
0.013
(0.014)
-0.004
(0.011)
High Prox × Same Table
Low Prox × Corner Position
0.019
(0.014)
0.033
(0.023)
0.028
(0.016)
Med Prox × Corner Position
High Prox × Corner Position
Low Prox × Month
-0.007
(0.004)
0.007
(0.006)
0.005
(0.006)
Med Prox × Month
High Prox × Month
Include Xirt ?
Worker FE?
Room FE?
Day FE?
Room × Day FE?
Yes
Yes
No
No
Yes
Yes
Yes
No
No
Yes
Yes
Yes
No
No
Yes
No
Yes
No
No
Yes
Yes
Yes
No
Yes
No
Yes
No
Yes
Yes
No
Yes
Yes
No
No
Yes
Notes: Xirt is a vector of control variables at the individual spatial level, which includes the number of workers working at contiguous
positions and the mean of their average productivities. Same Table is an indicator variable equal to one if the friend at the respective
proximity is assigned to the same table with the reference worker, and zero otherwise. Corner position indicates whether the worker
was assigned to a corner position. Month represents a linear monthly time trend. Standard errors are two-way clustered by worker and
room×day. * p < .10, ** p < .05, *** p < .01.
55
Table 11: Heterogeneous Effects - Worker’s Non-cognitive Skills (Standardized Score)
IV Estimates - Dependent variable: log(productivity)
Variable
High Prox
× Extraversion
× Agreeableness
× Conscientiousness
× Neuroticism
× Openness
Worker FE?
Observations
(1)
(2)
(3)
(4)
(5)
(6)
-0.059***
(0.009)
0.003
(0.007)
0.005
(0.007)
0.031***
(0.005)
0.008
(0.005)
-0.008
(0.004)
-0.060***
(0.007)
0.011
(0.008)
-0.061***
(0.006)
-0.060***
(0.006)
-0.060***
(0.007)
-0.059***
(0.007)
Yes
5731
Yes
5731
0.022***
(0.006)
0.031***
(0.005)
-0.012*
(0.007)
-0.007
(0.006)
Yes
5731
Yes
5731
Yes
5731
Yes
5731
Notes: Dependent variable is the worker’s log productivity (kilograms/hour). The measure of non-cognitive
skills is the worker’s standardized score for each personality factor. Scores are standardized using the sample
mean and standard deviation. All regressions include room×day fixed effects, proximity variables (including
low and medium), proximity variables interacted with Big Five factors, externality variables, and controls
at the individual spatial level. Standard errors clustered at the worker level are provided in parentheses.
Regression results from dropping worker fixed effects and using different degree of polynomials of worker
covariates show similar results to that reported here. Significance levels are corrected to account for multiple
hypothesis testing by using the Holm-Bonferroni method. * p < .10, ** p < .05, *** p < .01.
Table 12: Worker’s Ability and Non-cognitive Skills (N = 104)
OLS Estimates - Dependent variable : Worker’s estimated ability
Variable
Extraversion
Agreeableness
Conscientiousness
Neuroticism
Openness
(1)
(2)
(3)
(4)
0.034***
(0.011)
0.006
(0.013)
-0.007
(0.012)
-0.005
(0.013)
0.004
(0.012)
0.031***
(0.010)
0.001
(0.012)
0.001
(0.012)
-0.006
(0.012)
0.002
(0.012)
0.034***
(0.010)
(5)
(6)
(7)
0.014
(0.011)
0.015
(0.011)
-0.018
(0.011)
0.010
(0.012)
0.043**
(0.018)
Big Five Index
Worker covariates
Adjusted R2
(8)
No
0.07
Quad
0.26
Quad
0.29
Quad
0.21
Quad
0.21
Quad
0.22
Quad
0.20
Quad
0.25
Notes: All regressions, except column 1, include quadratic polynomials as a function of worker’s age and mean days of
working experience as a processing worker. Each non-cognitive skill is a standardized score using the sample mean and
standard deviation. The Big Five Index is obtained as a equally weighted average of the five standardized scores, reverse
scoring for neuroticism. Robust standard errors are presented in parentheses. * p < .10, ** p < .05, *** p < .01.
56
Table 13: Heterogeneous Effects - Ability of Worker and Friends
IV - Dependent variable: log(productivity)
(1)
(2)
(3)
(4)
-0.060***
(0.012)
0.097
(0.063)
-0.060***
(0.012)
-0.054***
(0.012)
-0.061***
(0.013)
Variable
High Prox
× Own Ability
× Friend’s Ability
0.129*
(0.071)
× Moreable
-0.009
(0.013)
× Moreable × |Own Ability - Friend’s Ability|
-0.023
(0.140)
0.066
(0.127)
× Lessable × |Own Ability - Friend’s Ability|
5731
Observations
5731
5731
5731
Notes: Dependent variable is the worker’s log productivity (kilograms/hour). All regressions include worker and
room×day fixed effects along with all externality variables at the individual spatial level. Bootstrapped standard
errors are in parentheses. Standard errors are two-way clustered by worker and room×day level and corrected for
sampling variability of the estimated ability term using a Bayesian parametric bootstrap procedure. In column
1 and 2, proximity variables are interacted with estimates on worker’s own ability and friend’s ability at each
proximity, respectively. Column 3 uses an indicator variable equal to one if the reference worker has higher ability
than her friends at each proximity, and zero otherwise. Column 4 uses a measure of absolute difference in ability
with respect to that of her friends at each proximity. All regressions include a full set of interactions of proximity
variables and each ability measure, although not presented in this table. Full table available in online appendix.
* p < .10, ** p < .05, *** p < .01.
Table 14: Proximity to Friends (Pre-Experiment Period)
Mean (S.D.)
Proximity
Contiguous - Indicator (=1 if yes)
Low Prox
Med Prox
High Prox
observed
(1)
randomized
(2)
Difference
(1) - (2)
0.87
(0.16)
0.51
(0.24)
0.52
(0.24)
0.51
(0.28)
0.40
(0.08)
0.18
(0.08)
0.11
(0.05)
0.18
(0.08)
0.47
(0.02)
0.33
(0.02)
0.41
(0.02)
0.33
(0.03)
Mean number of workdays
P-value
0.00
0.00
0.00
0.00
18
Notes: Columns 1 and 2 report standard deviations in parentheses. Column 3 reports standard errors.
One sided p-values are reported in the last column. Statistics reported in column 2 are obtained from
300 simulations based on a random assignment rule using the number of friends and coworkers present
in the processing room for each worker and day.
57
Table 15: Socializing with Friends (N = 100)
Dep. variable
Structural Estimate of Value of
Working with Friends (ŝi )
Variable
Age
Married
Secondary education
Months on job
Number of friends reported
Weekly hours spent with friend
(1)
0.107
(0.081)
−0.664∗∗∗
(0.255)
0.132
(0.193)
−0.056∗∗∗
(0.018)
−0.042
(0.082)
0.021
(0.052)
(2)
0.101
(0.075)
−0.659∗∗∗
(0.253)
0.103
(0.188)
−0.060∗∗∗
(0.018)
(3)
0.032
(0.063)
−0.441∗∗
(0.217)
0.062
(0.159)
−0.057∗∗∗
(0.018)
−1.285
(1.237)
Production Skill (α̂i )
−0.095
(0.107)
−0.103
(0.117)
−0.425∗∗∗
(0.098)
−0.008
(0.104)
0.021
(0.084)
Extraversion
Agreeableness
Conscientiousness
Neuroticism
Openness
Adjusted R2
0.10
0.12
0.37
Notes: The dependent variable is the standardized estimate of workers’
values on socializing with friends (estimate obtained from equation (16)
standardized by the sample mean and standard deviation). All regressions
include (age)2 and (months on job)2 but not reported in this table. Bootstrapped standard errors with 300 replications are presented in parentheses.
In Column 3, significance levels for the Big Five are corrected to account
for multiple hypothesis testing by using the Holm-Bonferroni correction
method. * p < .10, ** p < .05, *** p < .01.
58
59
0.18
0.013
(0.011)
0.013
(0.009)
0.18
(2)
(1)
0.18
0.016
(0.011)
-0.009
(0.017)
0.102
(0.069)
(3)
Education
0.18
0.016
(0.012)
-0.008
(0.015)
0.092
(0.282)
(4)
Production Skill
(α̂i )
0.18
0.024
(0.011)
-0.017
(0.015)
-0.009
(0.062)
(5)
Extraversion
0.25
0.020
(0.012)
-0.023
(0.015)
-0.024
(0.078)
(6)
Agreeableness
0.39
0.007
(0.010)
-0.022
(0.014)
-0.243**
(0.078)
(7)
Conscientiousness
0.28
-0.017
(0.011)
0.047**
(0.014)
-0.025
(0.072)
(8)
Neuroticism
Separate OLS regressions with Trait variable =
0.23
-0.004
(0.007)
0.027
(0.014)
-0.255**
(0.093)
(9)
Openness
Notes: Costi is constructed as the percentage change in expected hourly wage when working alongside friends relative to when working without friends. All columns
include the following worker covariates: worker’s education level, age, (age)2 , months on job, (months on job)2 , marital status, and number of reported friends from
the baseline survey. All columns, except Column 5 include workers’ standardized scores of extraversion. Column 2 excludes 10 workers who had less than 3 months of
tenure on job at the time of the baseline survey. Column 3 does not control for worker’s education level. Bootstrapped standard errors are in parentheses. Standard
errors are corrected for sampling variability of the relative cost term using a Bayesian parametric bootstrap procedure. Significance levels are corrected to account
for multiple hypothesis testing by using the Holm-Bonferroni correction method. * p < .10, ** p < .05, *** p < .01.
Adjusted R2
Above average in Trait
Cost × Above average in Trait
Cost
Independent variable
Subsample:
Tenure >
3 months
Full Sample
Dependent var.: Mean (High Prox (=1 if friend working alongside))
Table 16: Heterogeneity in the Cost-Demand Relationship (N = 100)
A
Reallocation of Fish
As mentioned in the field context, work material is distributed at the table level, where typically four to six workers are assigned
to, and worker compensation depends on individual performance, a fixed rate per kilogram×kilograms of fish processed by the
worker. Moreover, for each round of arrival of fish, the rule for fish distribution is based on the number of workers per table,
not their performances. This could create negative externalities at the table level since increasing one’s effort would lead
to less amount of fish to work on for other workers at the same table. Yet, an effective reallocation rule based on table
performances could potentially negate this effect. For instance, after the initial distribution, work materials on faster tables
could be replenished by reallocating fish from slower tables to faster tables.
One of the main tasks of managers at the processing rooms is to ensure that every table has fish to work on, which
involves coordinating batch arrivals in a timely manner and redistributing fish across tables according to their work progess.
Nonetheless, if managers are shirking with their job or possess favoritism toward members of a certain ability group (Bandiera
et al, 2009) then the redistribution process would not be effectively in place. For example, if a manager has ties to workers
that are of low ability then she might be more reluctant to take away fish from their tables to replenish fish on faster tables.
Likewise, if a manager favors high ability workers then there would be more fish reallocated to faster tables than the necessary
amount.
In an effort to examine whether the reallocation rule was effectively executed based on table performances, I check whether
the per-capita amount of fish processed at the table, which should be equal to the per-capita amount of fish ultimately allocated
to the table, is directly proportional to the mean of the ability of workers at the table, which is a proxy measure for the table’s
performance. This can be written out as,
fgrt = φ · ᾱgrt + λrt
(24)
where fgrt is obtained by dividing the total amount of fish processed on table g in room r on day t, Fgrt , calculated by adding
the output of workers at table g P
in room r on day t, by the number of workers at the table. The mean ability of workers at the
table is derived from ᾱgrt = n1
i∈g α̂i , which is the mean of the estimated abilities of workers at that table. The estimation
g
procedure of worker ability is presented in Appendix B. Time varying factors of fish supply and processing at the room level is
accounted by including room×day dummy variables, λrt .
Table A1 presents ordinary least square estimates of equation (24) using data from all tables, presented in the first two
columns, and from tables with no friendships, which is presented in the last two columns. Both Columns 1 and 3 suggest that
per-capita allocation is likely to be roughly 1 percent higher when the average ability of table members increases by 1 percent.
Ideally, under an efficient reallocation rule with observations on worker’s contemporaneous productivity, we would expect the
estimate on the proportionality constant in equation (24), φ̂, to be around one. Two sided p-values from testing efficiency
are reported in Columns 2 and 4 which show that we fail to reject this hypothesis. Although this does not provide pin point
evidence on the efficiency of the reallocation rule, it does show that daily allocation of fish across tables was proportionate to
the ability of workers at the tables which implies that a reallocation rule based on table performances was indeed being rolled
out by managers during the study period.
Table A1: OLS Estimates - Effectiveness of Reallocation
All Tables
Tables without Friends
log(per-capita
allocation of
fish)
H0 : φ̂ = 1
(p-value)
log(per-capita
allocation of
fish)
H0 : φ̂ = 1
(p-value)
(1)
(2)
(3)
(4)
Table mean ability (φ̂)
0.98
(0.11)
(0.83)
0.96
(0.19)
(0.85)
Number of tables
R2
1642
0.69
571
0.67
Notes: Dependent variable is the log of per-capita fish allocated to the table. Table mean ability is the table mean of the
estimated worker abilities. The estimation procedure of worker ability is provided in detail in Appendix B. Room×day
fixed effects are included in both regressions. Robust Standard errors are presented in parentheses.
60
B
Estimation of Worker Ability
To estimate workers’ abilities, or permanent productivities, I build on the approach of Mas and Moretti (2009) by exploiting
information on workers’ positions and social ties to take into account of the fact that a worker’s productivity may be affected
by the composition of coworkers at contiguous positions and her social relationship with these coworkers. Thus, I estimate the
αi terms using the following specification:
yirt = αi + Γi + Dirt + πN−i∈C (i) + λrt + irt
(25)
i
i
i
Γi = {γL
, γM
, γH
}.
(26)
where
Γi is a set of worker-proximity specific dummy variables. For example, on day t, if worker i has a friend at high proximity on
i is equal to one on day t. The idea is to account for productivity effects from the presence of friends that
any given day, then γH
may differ across workers as well as the physical proximity of the worker to her friends. Then the sum of a worker’ ability and
worker-proximity specific dummy can be interpreted as that worker’s ability when a friend is present of that proximity. Figure
B1 plots estimates on worker’s ability with respect to the presence of friends at each proximity.
The term Dird is a set of dummy variables, one for each element in the Cartesian product I × T , where, I = {1, . . . , N }
is the set of workers that share the same processing room with worker i, and T = {Same table, Different table}. This term
accounts for worker-table specific influences on the productivities of coworkers at contiguous positions. A worker may have
positive influence on coworkers at the same table if, for instance, she is a high ability worker and exerts pressure to workers at
the same table but at the same time may have no influence on workers at a different table. Regarding notations, suppose that
on day t worker j is working at a contiguous position with workers i and k. If worker j is at the same table with worker i but
at a different table with worker k then the dummy variable di,s = 1 and dk,d = 1 for worker j in room r on day t. The idea is
analogous to that in Arcidiacono et al. (2014) which estimates a player specific effect on other players’ scoring chances in the
professional basketball context. Finally, N−i∈C (i) denotes the number of workers who are spatially contiguous to worker i.
This estimation strategy is a parametric version of the estimation method used in Mas and Moretti (2009) which includes
dummies for all possible combinations of the set of coworkers in the same shift with the reference worker. Instead of including
dummies for coworker combinations, I include separate dummies for each contiguous worker. There are two reasons for doing
this. First, it allows separate estimation of each worker’s effect on other workers’ productivities, regardless of their friendship
status. This is advantageous when examining how workers may individually affect the productivities of their non-socially tied
coworkers which I address in a companion paper. Second, in the setting of this study, coworker compositions only vary across
days. As a result, a change in the composition of coworkers is confounded by unobserved factors of productivity that also
vary across days. While Mas and Moretti (2009) avoids this problem by using combinations of hour of the day and day of the
week dummies, this is not applicable in the current setting since productivity shocks, whether induced by supply side shocks
or demand side shocks, are not as periodic as in the context of supermarkets.
0
0
Kernel Density
1
2
3
Kernel Density
5 10 15
4
20
Figure B1: Kernel density estimates of worker ability depending on proximity to friends
20
Kernel Density
5 10 15
0
.9
.95
1
1.05
1.1
B. Relative Ability when Friends at Low Proximity
20
.4
Kernel Density
5 10 15
−.2
0
.2
A. Worker Ability without Friends
0
−.4
.9
.95
1
1.05
1.1
C. Relative Ability when Friends at Medium Proximity
.9
.95
1
1.05
1.1
D. Relative Ability when Friends at High Proximity
Notes: All density estimates use an Epanechnikov kernel with ‘optimal’ bandwidths. Panel A uses ability estimates (α̂i )
standardized with respect to the room average. Panel B, C, and D each plots kernel density estimates of the ratio of a worker’s
ability estimate when working with a friend at the corresponding proximity to that worker’s standardized ability estimate when
working without friends.
61
C
Heterogeneous Effects - Strength of Friendships
There is ample evidence that suggests that the type or strength of tie matters in economic decision makings. The importance
of social ties in risk sharing decisions is emphasized through several studies (Udry, 1994; Fafchamps and Lund, 2003). Social
ties have also been proven to impact labor market outcomes (Granovetter, 1973; Montgomery, 1991; Beaman and Magruder,
2012; Heath, 2015). Similarly, friendships of different strength may induce differential worker responses from working in close
proximity. In this section, I examine whether the effect on worker productivity from working with friends depends on reported
characteristics of friendships.
I use survey data on social ties to convey the type and strength of each friendship. First I use information on whether
a friendship was reported mutually and, if not, which of the two made the report. The survey also asked workers whether
the friendship formed before working at their current job and the duration of the friendship. In addition, workers were asked
to report on the intensity of activities shared in the friendship across 8 items using a 5-point scale. Combining this with the
worker’s subjective report on closeness with a reported friend, I construct a measure of strength of friendship which is the
estimate on the first component score obtained from a principal component analysis on the 9 response variables. Specifically,
the variables are responses on the following questions: how often did you go shopping together? how often did you talk to each
other during break or lunch times? how often did you help each other’s work? how often did you give advice/receive advice on
personal matters to/from this person? how many times did you lend or borrow money with each other? what is the average
hour spent together outside the company? how would you rate the closeness of the friendship on a five point scale?
Regression results using friendship characteristics are reported in Table C1. Column 1 suggests that there is no difference
in productivity effects whether or not the friendship was mutually reported. Column 2 and 3 indicate that productivity declines
from the presence of a friend at high proximity tend to be larger, albeit without statistical significance, when the reference
worker reported the friendship and if the friendship formed prior to employment at the current job. However, Column 4 suggests
that duration of the social tie is not correlated with the observed productivity declines. Column 5 suggests that the strength
of friendship, in terms of intensity of shared activities and subjective evaluation of closeness, with the friend at high proximity
does not have differential impacts on worker productivity.
Table C1: Does the effect depend on the strength of friendship?
IV Estimates - Dependent variable: log(productivity)
Variable
High Prox
×Mutual
(1)
(2)
(3)
(4)
(5)
-0.058***
(0.016)
-0.004
(0.016)
-0.052**
(0.021)
-0.054***
(0.009)
-0.061***
(0.011)
-0.058***
(0.012)
×Reporter
-0.009
(0.020)
×Knew Before
-0.027
(0.015)
×Long Duration
0.004
(0.016)
×Strength Index
Worker FE?
Observations
-0.003
(0.014)
Yes
5,731
Yes
5,731
Yes
5,731
Yes
5,731
Yes
5,731
Notes: Dependent variable is the worker’s log productivity (kilograms/hour). Mutual, Reporter, and Knew Before are indicator variables equal to one if the friendship was mutually
reported, reported by the worker, and formed prior to current job, respectively. Long Duration and Strength Index are indicator variables equal to one if the reported duration and
strength index scored above the median, respectively. The strength index is an estimate on the
first component score obtained from a principal component analysis on the following 9 items
on shared activities from the friendship survey; shopping, break talks, work at same table,
help each other, give advice, receive advice, exchange money, hours spent together outside
workplace, and subjective closeness report. All regressions include room×day fixed effects,
externality variables, and controls at the individual spatial level. Standard errors are two-way
clustered by worker and room×day level. * p < .10, ** p < .05, *** p < .01.
62
D
Prepared for Online Appendix
Table D1: [FULL TABLE] Heterogeneous effects - based on ability of worker and friends
IV - Dependent variable: log(productivity)
Variable
Low Prox
×Own Ability
(1)
(2)
(3)
(4)
-0.002
(0.006)
0.073
(0.068)
-0.002
(0.006)
-0.004
(0.010)
-0.005
(0.009)
×Friend’s Ability
0.208**
(0.088)
×Moreable
0.004
(0.011)
×Moreable×|Own ability - Friend’s ability|
×Lessable×|Own ability - Friend’s ability|
Med Prox
×Own Ability
0.004
(0.014)
-0.045
(0.085)
×Friend’s Ability
0.003
(0.014)
0.007
(0.015)
-0.003
(0.067)
×Moreable
-0.007
(0.013)
×Moreable×|Own ability - Friend’s ability|
×Lessable×|Own ability - Friend’s ability|
High Prox
×Own Ability
-0.060***
(0.010)
0.063
(0.068)
×Friend’s Ability
-0.061***
(0.010)
-0.054***
(0.011)
-0.114
(0.148)
-0.049
(0.142)
-0.058***
(0.011)
0.136**
(0.068)
×Moreable
-0.012
(0.012)
×Moreable×|Own ability - Friend’s ability|
-0.103
(0.149)
-0.049
(0.142)
×Lessable×|Own ability - Friend’s ability|
observations
0.037
(0.101)
0.027
(0.157)
0.011
(0.014)
5,731
5,731
5,731
5,731
Notes: Dependent variable is the worker’s log productivity (kilograms/hour). All regressions include worker
and room×day fixed effects along with all externality variables at the individual spatial level. Standard errors
are two-way clustered by worker and room×day level and corrected for the sampling variability of the estimated ability term using a Bayesian parametric bootstrap procedure. In column 1 and 2, proximity variables
are interacted with estimates on worker’s own ability and friend’s ability at each proximity, respectively.
Column 3 uses an indicator variable equal to one if the reference worker has higher ability than her friends
at each proximity, and zero otherwise. Column 4 uses a measure of absolute difference in ability with respect
to that of her friends at each proximity. * p < .10, ** p < .05, *** p < .01.
63
Table D2: [FULL TABLE] Heterogeneous Effects - Worker’s Non-cognitive Skills (Standardized
Score)
IV Estimates - Dependent variable: log(productivity)
Variable
Low Prox
× Extraversion
× Agreeableness
× Conscientiousness
× Neuroticism
× Openness
Med Prox
× Extraversion
× Agreeableness
× Conscientiousness
× Neuroticism
× Openness
High Prox
× Extraversion
× Agreeableness
× Conscientiousness
× Neuroticism
× Openness
Worker FE?
Observations
(1)
(2)
(3)
(4)
(5)
(6)
-0.002
(0.006)
0.014**
(0.007)
-0.006
(0.008)
0.002
(0.006)
0.008
(0.007)
-0.004
(0.007)
0.005
(0.013)
-0.018**
(0.007)
-0.011
(0.008)
0.012
(0.011)
-0.011
(0.011)
0.011
(0.007)
-0.059***
(0.009)
0.003
(0.007)
0.005
(0.007)
0.031***
(0.005)
0.008
(0.005)
-0.008
(0.004)
-0.002
(0.006)
0.009
(0.006)
-0.002
(0.006)
-0.002
(0.006)
-0.002
(0.006)
-0.002
(0.006)
Yes
5,731
-0.004
(0.007)
-0.001
(0.006)
0.004
(0.006)
0.005
(0.013)
-0.012
(0.008)
0.004
(0.014)
0.004
(0.014)
0.003
(0.014)
0.001
(0.007)
0.003
(0.014)
-0.006
(0.009)
0.005
(0.007)
-0.004
(0.008)
-0.060***
(0.010)
0.011
(0.008)
-0.061***
(0.009)
-0.060***
(0.009)
-0.060***
(0.010)
0.008
(0.009)
-0.059***
(0.010)
0.022***
(0.007)
0.031***
(0.005)
-0.012*
(0.007)
-0.007
(0.006)
Yes
5,731
Yes
5,731
Yes
5,731
Yes
5,731
Yes
5,731
Notes: Dependent variable is the worker’s log productivity (kilograms/hour). The measure of non-cognitive
skills is the worker’s standardized score for each personality factor. Scores are standardized using the sample
mean and standard deviation. The Big Five Index is obtained as a equally weighted average of the five
standardized scores, reverse scoring for Neuroticism. All regressions include room×day fixed effects and
externality variables, and controls at the individual spatial level. Standard errors are two-way clustered by
worker and room×day level. Regression results from dropping worker fixed effects and using different degree
of polynomials of worker covariates show similar results to that reported here. * p < .10, ** p < .05, ***
p < .01.
64
Table D3: [FULL TABLE] Heterogeneous Effects - Characteristics of Friendships
IV Estimates - Dependent variable: log(productivity)
Variable
Low Prox
×Mutual
(1)
(2)
(3)
(4)
(5)
-0.020
(0.012)
0.026
(0.014)
-0.030
(0.018)
-0.003
(0.007)
-0.004
(0.008)
-0.017
(0.009)
×Reporter
0.033
(0.019)
×Knew Before
0.004
(0.014)
×Long Duration
0.009
(0.010)
×Strength Index
Med Prox
×Mutual
0.019
(0.017)
-0.023
(0.018)
×Reporter
0.024
(0.025)
0.002
(0.015)
0.001
(0.017)
-0.025
(0.025)
×Knew Before
0.008
(0.022)
×Long Duration
0.011
(0.019)
×Strength Index
High Prox
×Mutual
-0.058***
(0.016)
-0.004
(0.016)
×Reporter
-0.052**
(0.021)
-0.054***
(0.009)
-0.061***
(0.011)
-0.020
(0.018)
-0.058***
(0.012)
-0.009
(0.020)
×Knew Before
-0.027
(0.015)
×Long Duration
0.004
(0.016)
×Strength Index
Worker FE?
Observations
0.035**
(0.013)
0.014
(0.015)
-0.003
(0.014)
Yes
5,731
Yes
5,731
Yes
5,731
Yes
5,731
Yes
5,731
Notes: Dependent variable is the worker’s log productivity (kilograms/hour). Mutual, Reporter, and Knew Before are indicator variables equal to one if the friendship was mutually
reported, reported by the worker, and formed prior to current job, respectively. Long Duration and Strength Index are indicator variables equal to one if the reported duration and
strength index scored above the median, respectively. The strength index is an estimate on the
first component score obtained from a principal component analysis on the following 9 items
on shared activities from the friendship survey; shopping, break talks, work at same table,
help each other, give advice, receive advice, exchange money, hours spent together outside
workplace, and subjective closeness report. All regressions include room×day fixed effects,
externality variables, and controls at the individual spatial level. Standard errors are two-way
clustered by worker and room×day level. * p < .10, ** p < .05, *** p < .01.
65
Table D4: Perception on Processing Skills
Reported Friend’s Estimated Processing Skill (α̂)
Variable
Rate on friend’s skill level
(1)
(2)
(3)
0.019***
(0.008)
0.004
(0.011)
0.028*
(0.015)
0.018*
(0.011)
×Respondent’s ability above mean
×Respondent’s neuroticism above mean
0.002
(0.015)
385
Observations
385
385
Notes: Estimated processing skills is the estimated worker fixed effect (See Appendix B). Respondents
were asked to evaluate their friend’s skill level on a scale of one to seven. Standard errors clustered at
the worker level are provided in parentheses. * p < .10, ** p < .05, *** p < .01.
Table D5: Productivity and Wage Gap in Non-cognitive Skills
Dependent variable:
Variable
Extraversion is above median
Agreeableness is above median
Conscientiousness is above median
Neuroticism is above median
Openness is above median
Observations
Adjusted R2
log(productivity)
log(daily wage)
(1)
(2)
0.044
(0.037)
0.003
(0.041)
0.067∗
(0.040)
−0.037
(0.040)
−0.015
(0.037)
0.010
(0.015)
−0.023
(0.017)
0.047∗∗
(0.016)
−0.026
(0.016)
0.008
(0.015)
100
0.12
100
0.19
Notes: Dependent variables are mean values from the pre-experiment period. Both
columns include the worker’s marital status, level of education and quadratic functions
of age and experience in processing. Robust standard errors are provided in parentheses.
* p < .10, ** p < .05, *** p < .01.
66
Figure D1: Worker Position Assignment Forms
Sơ chế nhóm xưởng 1
Ngày sản xuất:
1
44
2
29
3
61
19
37
20
21
1
4
4
65
5
56
6
5
7
48
8
75
9
63
21
74
22
4
23
23
24
7
2
5
10
49
11
69
12
28
13
59
14
8
15
36
25
35
26
20
27
13
28
41
16
3
17
52
18
15
3
2
3
4
5
6
7
8
9
10
13
14
15
17
18
20
21
23
7
33
10
34
6
Huệ
Hằng
Lớn
Hương
Sáu
Thu Thủy
Hồng
Sang
Năng
Phương
Chín
Bảy
Phụng
Nương
Ngọc Thùy
Thu Hương
Bảy
35
17
36
57
24
28
29
30
31
34
35
36
37
41
44
48
49
51
52
56
57
8
37
51
38
34
Đen
Minh Dũng
Mỹ
Giàu
Trang
Hà
Lắm
Bích Trâm
Kim Thanh
Kim Liên
Đại Nguyên
Thúy
Ngọc Mỹ
Thùy Trang
Hiền
Lào
39
14
40
58
2
58
59
61
63
65
69
74
75
4
1
3
13
41
18
42
9
9
TÊN
2
Loan
Hai Ngọc Liên Thu
Ngọc Nữ Mỹ Duyên Lan Hiền Thu Liên Rớt Hoa Liêm Bich Loan
Phương Thảo Thu Thảo
Bích Hồng
Kim Anh
Bích Hoa Kim Nga
Thu Huệ Huỳnh Kiều Cẩm Hiền Thu Vân Lý Ngọc Trinh Thanh Hoa Thanh Tuyết Hoa
Mùi Ngọc Trinh Hòa Tiểu
Kim Chi
Lâu Hiền Diệu Ngọc Hiệp Kim Thanh
Kim Hoa Thu Hà
Trang
4
1
29
31
30
2
6
MÃ SỐ
3
14
31
24
32
30
Sơ chế nhóm xưởng 2
Ngày sản xuất:
Niệm
Chung
Thuyết
Hời
Hoa
Thanh
Kim Loan
Sáng
9
6
15
16
50
17
14
18
10
5
24
51
25
56
26
46
27
28
28
18
29
23
35
47
36
61
37
33
38
22
33
34
6
3
21
22
69
23
62
5
7
17
8
7
2
10
2
11
16
12
9
19
4
6
20
30
30
31
27
32
5
6
8
9
10
14
16
17
19
20
22
23
24
27
28
29
30
34
36
42
43
8
39
24
40
46
47
50
51
53
56
60
61
62
65
68
69
70
116
140
142
(a) Processing Room 1
(b) Processing Room 2
Sơ chế nhóm xưởng 3
Ngày sản xuất( date):
1
107
2
111
3
79
11
113
12
102
19
117
20
73
4
80
5
75
6
76
7
116
8
77
2
9
120
10
82
3
13
93
14
119
15
109
16
89
4
17
95
18
108
5
21
88
22
112
23
81
24
74
6
25
118
26
105
1
MÃ SỐ
TÊN
73
74
75
76
77
79
80
81
82
88
89
93
95
102
105
107
108
109
111
112
113
116
117
118
119
120
Bảo Trân
Sen
Nga
Mỹ Hảo
Thúy Hồng
Kim Trang
Đài
Hồng Loan
Kiều Phương
Hiền
Minh Sang
Hiền
Phượng Không
Như Thuận
Lành
Bích Thủy
Kim Hằng
Kim Xuân
Phương Thảo
Hồng
Thanh Thuận Như Ngọc
Thu Sương
Yến Nhi
Thanh Vân
(c) Processing Room 3
Notes: The number of tables and worker positions are different across rooms due to differences in room sizes. Pseudonyms are
used to generate these samples and not that of the actual workers.
67
0
.2
Cumulative Density
.4
.6
.8
1
Figure D2: Cumulative Distribution Functions of the Probability of a Friend at High Proximity Positions
0
.2
.4
.6
Pr(Working Alongside Friend)
Observed Data
.8
1
Simulated Under Randomization Rule
Notes: Counterfactual data is from 300 simulations of worker positions with random assignment during the pre-experiment
period. A comparison of the two distributions using the Kolmogorov-Smirnov equality of distribution test rejects that the two
distributions are equal (p-value = 0.00).
0
.2
Cumulative Density
.4
.6
.8
1
Figure D3: Empirical CDF of Estimates on Value of Socializing (Mutually Reported Friends)
−10
0
10
Value of Socializing at Work
20
Notes: Estimates are derived using only mutually reported friendships. A Kolmogorov-Smirnov equality of distribution test
with the distribution obtained using all friendship ties (mutual or not) fails to reject that the two distributions are equal (p-value
= 0.994).
68