Social Preferences and the Response to Incentives: Evidence from Personnel Data
Author(s): Oriana Bandiera, Iwan Barankay and Imran Rasul
Reviewed work(s):
Source: The Quarterly Journal of Economics, Vol. 120, No. 3 (Aug., 2005), pp. 917-962
Published by: Oxford University Press
Stable URL: http://www.jstor.org/stable/25098759 .
Accessed: 17/05/2012 13:11
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].
Oxford University Press is collaborating with JSTOR to digitize, preserve and extend access to The Quarterly
Journal of Economics.
http://www.jstor.org
SOCIAL PREFERENCES
INCENTIVES: EVIDENCE
Oriana
AND THE RESPONSE TO
FROM PERSONNEL DATA*
Bandiera
iwan barankay
Imran Rasul
on whether workers
We present
have social preferences
evidence
by compar
individual
under relative
effort im
incentives, where
ing workers'
productivity
on others, with their productivity
under piece rates,
poses a negative
externality
of the average worker
is at least
where
it does not. We find that the productivity
incentives. We show that
50 percent higher under piece rates than under relative
the negative
their effort
this is due to workers
internalizing
externality
partially
when working
incentives,
imposes on others under relative
alongside
especially
their friends. Under
does not affect
among workers
piece rates, the relationship
that workers
Further
reveals
analysis
productivity.
others and be monitored.
only when they can monitor
as
the underlying
motive
of workers'
internalize
This
rules
the externality
out pure altruism
behavior.
I. Introduction
This paper uses personnel data to present evidence on social
in the workplace,
inter
preferences
namely on whether workers
on their colleagues. While
nalize
the effects of their behavior
extensive evidence from experimental
economics
indicates that
individuals take account of the effect of their actions on others in
individuals exhibit social preferences
laboratory games, whether
in the workplace
is largely unknown. The issue is of great prac
since workers' productivity under several incen
tical relevance
tive schemes, such as relative performance evaluation
and team
*
We
from discussions
have
with Heski
V. Bhaskar,
benefited
Bar-Isaac,
David
Bertrand,
Card, Christopher
Timothy
Besley, Marianne
Flinn, Edward
Alan Manning,
John Pencavel,
Amil P?trin, Canice
Lazear,
Dilip Mookherjee,
Matthew
at the Uni
Rabin, Debraj
Prendergast,
Ray, and seminar participants
at Berkeley,
Boston University,
Graduate
School of
versity of California
Chicago
Columbia
Essex University,
the University
of Georgetown,
Business,
University,
London
School of Economics,
New York University,
Purdue University,
Stanford
the University
of Toronto, University
the University
London,
University,
College
ofWarwick,
Yale University,
and the 2004 European
Economic
Meet
Association
We also thank the editor, Lawrence
and three anonymous
ings in Madrid.
Katz,
referees
for providing
useful
comments.
The
research
started when
the first
author was visiting the University
of Chicago Graduate
whom
School of Business,
she thanks
for their hospitality.
Financial
from the Royal
Economic
support
and the Suntory
and Toyota
International
Centres
for Economics
and
Society
Related
is gratefully
We
thank all those involved
in
Disciplines
acknowledged.
the data. This paper has been screened
to ensure
that no confidential
providing
data are revealed.
All errors remain our own.
? 2005 by the President and Fellows ofHarvard College
Technology.
The Quarterly Journal ofEconomics, August 2005
917
and the Massachusetts
Institute of
QUARTERLY JOURNAL OF ECONOMICS
918
pay, depends crucially on whether and to what extent they inter
nalize the effects of their actions on coworkers' payoffs.1
We use detailed personnel data from a leading farm in the
United Kingdom
that first paid its workforce according to a rela
to piece rates. The
tive incentive scheme and then switched
workers' task is to pick fruit, and their individual productivity is
recorded daily. To identify whether workers have social prefer
ences, we compare their productivity under relative incentives to
relative incentives,
their productivity under piece rates. Under
workers' daily pay depends on the ratio of individual productivity
to average productivity among all coworkers on the same field
and day. In contrast, under piece rates individual pay only de
pends on individual productivity.
under relative incen
The comparison
is revealing because
tives individual effort imposes a negative externality on cowork
ers' pay, whereas under piece rates individual effort has no effect
on others' pay. The difference in workers' performance under the
two schemes, if any, then provides evidence on whether and to
internalize
the externality they impose on
what extent workers
their colleagues.
can be
It is important to stress that social preferences
of either altruism or
thought of as a reduced-form representation
collusion. Namely, workers might internalize the externality ei
ther because
they truly care about colleagues' payoffs, or because
In this paper we first
and retaliation.
they fear punishment
the
internalize
whether
workers
externality their effort
analyze
on
and
then
shed
others,
light on their underlying
imposes
motives.
in four stages. First, we
analysis
proceeds
empirical
of
the
causal
effect
the
change in incentive schemes on
identify
workers' productivity. Second, we derive the empirical distribu
that is consistent with the observed
tion of social preferences
across
in
incentive schemes. Third, we ana
productivity
change
internalize the exter
to
workers
extent
which
the
whether
lyze
on
To address
between
coworkers.
the
relationship
nality depends
to workers
this issue, we use data from a survey we administered
to collect precise information on their social network of friends on
The
the farm. Fourth, we present
evidence
to distinguish
between
two
evidence
1. See Fehr and G?chter
[2000a] for an overview of the experimental
has
on social preferences.
in the workplace
That human
relations matter
long
literature
been noted in the sociology
[Mayo 1933; Roy 1952], and the organiza
a
and O'Reilly
literature
tional behavior
review).
[1998] for
(see Williams
SOCIAL PREFERENCES,
hypotheses
ences?pure
on why workers
and
altruism
RESPONSE
behave
as
TO INCENTIVES
if they have
919
social prefer
collusion.
The data have four key features that help identify the causal
effect of the change in incentive schemes on worker productivity,
to attribute this change to workers'
reaction to the externality
created by the relative incentives scheme, and to investigate the
motives behind the workers' behavior. First, we observe the daily
productivity of the same workers before and after the introduc
sources of unobservable
tion of piece rates. Time-invariant
indi
vidual heterogeneity
are, therefore, controlled for.
face an identical work environ
the same workers
Second,
ment throughout, except for the change in incentive schemes. In
sorting of new workers into the
particular, there is no endogenous
attrition ofworkers out of the sample.
sample and no endogenous
Moreover,
tices were
tasks,
technology,
management,
and
other
farm
prac
the same under both incentive schemes.
Third, the group of coworkers each individual works with,
changes on a daily basis. This allows us to identify the effect of
group composition on worker productivity from the comparison of
the behavior
of the same worker working
different
alongside
coworkers.
of the same workers using
Fourth, we observe a subsample
an alternative
not
that
does
allow monitoring
of co
technology
workers. The ability to monitor coworkers creates differences in
on whether workers
observed behavior depending
cooperate be
cause
of collusion
or
pure
altruism.
The
comparison
of workers'
to the introduction of piece rates under the two moni
responses
then provides evidence to distinguish
between
toring scenarios
collusion and pure altruism.
We find that the change in incentive scheme had a significant
and permanent
impact on productivity. For the average worker,
increased
productivity
by at least 50 percent moving from relative
incentives to piece rates. Calibration
of the first-order conditions
for worker's efforts reveals that the observed change in produc
that
tivity is too large to be consistent with the assumption
workers ignore the negative externality they impose on others. At
the same time, the observed change in productivity
is also too
small to be consistent with the assumption
that workers maxi
mize the welfare of the group and fully internalize the negative
externality.
We, therefore, posit that workers place some weight on the
payoffs accruing to their coworkers and derive the distribution of
920
QUARTERLY JOURNAL OF ECONOMICS
social preferences among workers that fits the observed change in
productivity. We find that this is consistent with the average
worker placing a weight of .65 on the benefits accruing to all other
that they place a weight of one on their own
coworkers, assuming
benefits. Further analysis reveals that under relative incentives
workers internalize the externality more when the share of their
personal friends in the group is larger and this effect is stronger
in smaller groups.
In line with the interpretation
that social
the
difference in productivity across the two
preferences explain
schemes, we find that the relationship
among workers does not
affect productivity under piece rates.
Finally, we find that productivity under relative incentives
was significantly lower only when workers were able to monitor
is necessary
to enforce collu
each other. Given that monitoring
it does not affect altruism, we take this finding to
sion while
that workers are able to sustain implicit
support the hypothesis
collusive agreements when relative incentives are in place.
our findings imply that worker productivity
is sig
Overall,
different
two
schemes
the
because
in
under
workers
nificantly
ternalize
the externality
they impose on others under relative
and they do so more when
incentives,
they work alongside
friends. Importantly, workers
internalize
the externality
only
when they are able to monitor each other, which rules out pure
altruism as the underlying cause of workers' behavior.
The results demonstrate
the importance of understanding
on cowork
in the presence of externalities
how workers behave
incentive
schemes.
For
when
ers,
instance, the results
designing
on
to
how
Lazear's
observation
[1989]
rarely workers are
speak
to rank-order tournaments,
compensated
according
despite the
incentive ef
fact that theory suggests they may have desirable
fects [Lazear and Rosen
1981]. One reason why such schemes
may not be observed in practice is that they lower productivity
when workers behave as if they have social preferences.
The paper contributes to two strands of the literature. First,
we contribute to the literature on incentive schemes and workers'
1999] .2We
existing re
[Prendergast
complement
productivity
the effects of two different relative
2. Knoeber
and Thurman
[1994] analyze
of rank order tournament
incentive schemes on chicken ranchers. The predictions
in experimental
data
incentive, have been examined
theory, a type of relative
and Bog
and Weigelt
1987] and sports tournaments
[Bull, Schotter,
[Ehrenberg
nanno
and Shearer
and Huselid
1992]. Lazear
1990; Becker
[2000], Paarsch
from fixed pay
[1996], and Shearer
[2004], find sizable productivity
gains moving
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
921
suits by providing the first evidence on the comparison between
relative incentives and piece rates. Our work is also closely re
and Owan
lated to Hamilton,
Nickerson,
[2003], who present
evidence suggesting that productivity is higher under team pay,
where workers'
effort imposes a positive externality on others,
than under piece rates. Consistent with their results, we find that
productivity is lower under relative incentives where the exter
nality is negative.
the large experimental
lit
Second, our results complement
erature on social preferences, showing that these matter outside
our findings are in line with the
the laboratory. In particular,
literature on public good games where
individual
experimental
are generally
contributions
found to be halfway between
the
selfish Nash
equilibrium
(complete free riding) and the group
optimum [Ledyard 1995; Fehr and G?chter 2000b].
as follows. Section
The paper
is organized
II models
the
workers' effort choice under relative incentives and piece rates.
the data. Section IV presents reduced-form
Section III describes
estimates of the effect of the change in incentives on productivity.
Section V brings alternative
structural models ofworkers' behav
ior to the data and derives the distribution of social preferences
that are consistent with the observed
in productivity.
change
Section VI analyzes
the effect of the identity of coworkers on
productivity under the two incentive schemes. Section VII pro
vides evidence to distinguish
between pure altruism and collu
on the design of
sion. Section VIII
concludes with a discussion
incentive schemes when workers have social preferences.
II. Theoretical
Framework
This
section makes
affect
precise how social preferences
workers' effort choice under relative incentives and under piece
rates. Theory suggests that relative schemes have the advantage
of differencing out common shocks to productivity
and could,
a
for
the
when
shocks
such
therefore, yield
higher payoff
principal
are important [Lazear and Rosen 1981; Green and Stokey 1983;
to piece
rate incentives.
Laffont
and Matoussi
[1995] find worker
Similarly,
to be 50 percent higher in farms operated
under fixed rent contracts
productivity
those under
with
contracts.
Foster
and Rosenzweig
compared
sharecropping
of body mass net of calorie intake,
[1994] show that effort, proxied by the depletion
is 22 percent higher for rural laborers paid by piece rates compared with those
paid hourly wages.
QUARTERLY JOURNAL OF ECONOMICS
922
Nalebuff
and Stiglitz 1983]. This literature, however,
typically
assumes
that workers ignore the externality their effort imposes
on others under relative incentives. The model that follows shows
how, other things equal, a worker's response depends on whether
she has social preferences. Therefore, the empirical comparison of
can provide evidence
on
productivity under the two schemes
whether
social preferences are relevant in practice.
Consider a group ofN workers. Each worker i exerts e? > 0
units of effortwhich determines her productivity. Without
loss of
in what
that effort equals productivity
generality, we assume
follows. Each worker's payoff is <)>( )
0?e2/2, where 4>() is the
benefit derived from pay (which depends on effort), and 0?e2/2 is
the cost of effort.We assume that 4>() is a differentiable concave
= ??-The
function, with lmx^o^'W
parameter
0? is interpreted
as the inverse of the workers
innate ability. We assume
that
are heterogeneous
can
workers
this
and
be
dimension,
along
ordered such that 0X < 02 < . . .< 0^, where 0? > 0 for all i. The
effort and
between
relationship
scheme as explained below.
H.A.
Relative
pay
depends
on the incentive
Incentives
Under relative incentives, a worker's pay depends on how she
specifically, in our setting,
performs relative to her peers. More
workers' benefit from pay takes the form <$>(e?/e)for all i, where
e = ( 1/N) 2? e? is the average effort of all N workers.3 The relative
that other things equal, an
scheme has the key characteristics
in worker
increase
?'s effort (i) increases her pay; and (ii) in
creases average effort and, hence, imposes a negative externality
by reducing the pay of everybody else in the group.
The choice of effort under relative incentives then depends on
internalize this externality,
whether and to what extent workers
on
have
social
whether
preferences. Denote by tt? the
they
namely
on
coworker's payoffs. The equi
i
her
that
worker
weight
places
librium effort for worker
i solves
incentive
scheme
3. This relative
and
benefits are based on their cardinal
similar to a "linear relative performance
and Thurman
Knoeber
[1994]. Under
a and jjlare parameters
?), where
[L(e?
Worker
is not a rank order tournament.
It is, however,
not their ordinal ranking.
as
in
scheme
evaluation"
studied
(LRPE)
a LRPE
worker's
is, a +
compensation
as
taken
given by workers.
RESPONSE
SOCIAL PREFERENCES,
TO INCENTIVES
923
Such social preferences can be thought of as a reduced-form
of a number ofmodels. They depict behavior con
representation
sistent with reciprocity or altruism
[Fehr and Schmidt 1999], or
of a repeated Prisoner's Dilemma
the evolutionary
equilibrium
game inwhich workers learn which strategies to play [Levine and
Pesendorfer
1999]. In Section VII
2000; Sethi and Somanathan
we present some evidence to distinguish between models
inwhich
workers' preferences display altruism toward others, and models
in which workers behave as if they are altruistic because,
for
to
enforce
collusive
instance, they play trigger strategies
implicit
agreements.
that worker i chooses her effort taking the effort of
Assuming
effort of worker i solves
others as given, the Nash equilibrium
?
II.B.
*'(?)f(wB^'(?)&
Piece Rates
Under piece rates, individual effort is paid at a fixed rate ?
i chooses her effort
social preferences, worker
per unit. With
under piece rates as follows:
cj>(?e-) + tt?Jj[
(3) max
The
j*i
equilibrium
"
?K?e,-)
J?T
?
V
effort level solves
(4) *'(?Ci)?
=
*&
the first-order condition,
e?e?.
as worker ?'s effort does not affect the pay of her
Naturally,
her
coworkers,
optimal choice of effort is independent of the social
she
weight
places on others.
II. C. Comparing
Relative
Incentives
and Piece Rates
To compare
under piece rates,
lie so that for a
effort is the same
condition under
(5)
The
effort choices under the relative
scheme and
we evaluate the first-order condition (4) at ? =
given average effort level, the pay per unit of
under both incentive schemes. The first-order
piece rates then is
*'(?)?=e*
difference between
the first-order conditions
(2) and
(5)
QUARTERLY JOURNAL OF ECONOMICS
924
can then be ascribed to two sources. The first and most important
i imposes on others under
difference is the externality worker
relative incentives. By increasing her effort, she also increases
the average effort, and this reduces the pay of coworkers other
of this effect depends on the extent
things equal. The magnitude
towhich workers internalize the externality, that is, on the social
weight tt?. This is captured in the second term on the left-hand
is lower under
side of (2). When
ir? > 0, worker ?'s productivity
relative incentives compared with that under piece rates. Second,
by exerting more effort, each worker lowers the pay she receives
for each unit of effort under relative incentives. This effect, cap
also reduces productivity under
tured by the 2y#?
eJiN?) term,
in large groups.4
relative incentives but is negligible
This highlights that the difference between productivity un
der relative incentives and piece rates depends on the extent to
the
and so internalize
which workers have social preferences
on
This
is
minimized
others.
difference
externality they impose
= 0 for all
when
?, namely when workers do not take into
tt?
account the effect of their actions on others. In this case, effort
identical for
levels under the two schemes
are, thus, almost
large N.
In contrast, when tt? ? 1 for all ?, the first-order condition
under relative incentives (2) coincides with the first-order condi
that is, when effort
tion of the social optimum among workers;
the welfare of the group as a whole.
levels are chosen tomaximize
In the remainder of the paper we establish whether productivity
is indeed different under the two schemes and then derive impli
cations
for the workers'
underlying
III. Context
III.A.
and Data
preferences.
Description
Context
We analyze personnel data from a leading United Kingdom
in the sample are
based fruit farm for the 2002 season. Workers
to pick fruit across a number of fields on the
hired seasonally
farm. They are paid according to a relative incentive scheme for
Then
the
workers.
is seen most
clearly in the case of homogeneous
is e* = eR =
incentives
relative
effort level under
((1
equilibrium
=
under
effort
rates.
The
ratio
of
under piece
(<f>'(l))1/2
l/ADcJ>'(l))1/2 and e*=ep
are heterogeneous,
the ratio
the two systems
1/N)1/2. If workers
is, thus, (1
that
it can be shown
on
size
and
workers'
although
ability,
group
depends
= ep.
lim^^e*
4. This
Nash
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
925
the first half of the season and according to piece rates for the
second half. In both cases workers face a compensation
schedule
of the form,
compensation
=
$K?,
where K? is the total kilograms of fruit picked by worker ? on the
field-day. Throughout we define individual productivity y ?as the
of fruit picked per hour.5
number of kilograms
Under the relative scheme, the unit wage ? is endogenously
determined by the average productivity of all workers in the same
field on the same day. In particular,
? is set according to
(6)
?
-
w/y,
where w is the minimum wage plus a positive constant fixed by
at the beginning
the management
of the season, and y is the
on the field-day. At the
of
all
workers
average hourly productivity
an ex ante
announces
start of each field-day, the field manager
unit wage based on her expectation
of worker productivity. This
unit wage is revised at the end of each field-day to ensure that a
worker with productivity y earns
the preestablished
hourly
wage
w.
In line with the relative
in Section
scheme analyzed
II,
worker Vs compensation
depends on her productivity relative to
the average productivity of her coworkers. In particular,
given
=
in a field,
K?
y?h> where h is the number of hours worked
worker ?'s pay is (y?/y)hw. Note that an increase in worker
?'s
on the field-day and,
effort increases
the average productivity
thus, imposes a negative externality on her coworkers by reduc
ing the unit wage ? in (6).
Under piece rates, the unit wage is set ex ante, based on the
assessment
of productivity that field-day, and is not
manager's
revised. The manager
aims to set the unit wage at the level such
that the worker with average productivity receives the minimum
wage plus a fixed positive constant. The data under piece rates
indicate that the rate was set correctly in the sense that the unit
5. To comply with minimum
is supple
wage
laws, workers'
compensation
mented
whenever
In practice,
the
wage.
$K? falls below the pro-rata minimum
farm management
makes
clear that any worker who needs
to have her compen
sation increased
to the minimum
we
level
would
be
fired.
wage
Indeed,
repeatedly
observe
less than 1 percent
of all worker-field-day
observations
pay
involving
to meet the minimum
increases
Of
46
wage requirements.
these,
percent occurred
under relative
54
occurred
under
rates.
incentives,
percent
piece
926
QUARTERLY JOURNAL OF ECONOMICS
wage set ex ante is very close to the unit wage that would obtain
ex post if the relative formula in (6) were used instead.6
The key difference between the two systems is that under the
relative incentives, workers' effort on the day determines the unit
for each kilogram picked on that field-day. Under
wage
piece
rates, workers' effort does not affect the unit wage.7
in the sample are hired on a casual basis, namely
Workers
work is offered daily with no guarantee
of further employment.
The majority
of workers hired are from Eastern
and Central
and live together on the farm for the duration of their
Europe
are issued with a farm-specific work permit for a
stay.8 Workers
of six months,
maximum
implying that they cannot be legally
in the United Kingdom. Their outside option
employed elsewhere
is, therefore, to return to their home countries. The vast majority
in the sample report their main reason to seek tem
of workers
to be financial, which
porary employment in the United Kingdom
are much
in light of the fact that wages
is hardly surprising
in
in
their
home
the
United
than
countries.9
Kingdom
higher
We analyze productivity data on one type of fruit only and
and the end
focus on the season's peak time?between
mid-May
of August. Fruit plants are lined up in rows, and each worker is
one or more adjacent
rows to pick. The productivity of
assigned
each worker depends exclusively on her effort and on the amount
on her rows, namely workers'
efforts are not
of fruit available
in
complements
production.
are recorded electronically.
on workers'
Data
productivity
on the
in part depends
the unit wage
under both incentive schemes
6. Hence,
and other determinants
of expected productivity.
field life cycle, field conditions,
to productivity. We
related
The unit wage
is, therefore, mechanically
negatively
the
in the unit wage
that allow us to estimate
do not observe exogenous
changes
of labor supply for example.
wage
elasticity
over the unit wage under relative
incen
face more uncertainty
7. Workers
ex ante, this can be revised ex post to
a rate is announced
tives because
although
reflect the productivity
of the average worker. Under
piece rates, the ex ante unit
to have a large impact
is unlikely
cannot be revised. However,
wage
uncertainty
on effort choices
workers
because
daily and have many
play the same game
under
relative
to learn the ex post adjustment
of the unit wage
opportunities
can explain
the
In Section
IV we further discuss whether
incentives.
uncertainty
in the data.
in productivity we observe
change
must
and have at
individuals
be full-time students,
8. In order to qualify,
must
Workers
least one year before graduation.
(i) return to the same university
in the United
in the autumn,
(iii) have not worked
(ii) be able to speak English,
19 and 25.
before, and (iv) be aged between
Kingdom
at the United
9. As
of January
gross monthly
Kingdom
2003,
earnings
in
minimum
wage
(Euro 1105) are five times as high as at the minimum
wage
come from, and almost twenty
of workers
the majority
Poland
(Euro 201), where
our
times higher than in Bulgaria
(Euro 56), the poorest country in
sample.
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
927
Each worker is assigned a unique bar code, which is used to track
the quantity of fruit they pick on each field and day in which they
error in recorded
work. This ensures
little or no measurement
to
The
is
restricted
those
workers who
sample
productivity.
worked at least ten field-days under each incentive scheme. Our
level observa
working sample contains 10,215 worker-field-day
142
22
and
108
tions, covering
fields,
workers,
days in total.
The compensation
scheme changed from relative incentives
to piece rates for all workers midway
through the season. Rela
tive incentives are in place for the first 54 days in the sample,
54. The change was
piece rates are in place for the remaining
on the same day it was first implemented. No other
announced
organizational
change took place during the season, as reported
in the next subsection.
and as documented
by farm management
revealed that the rela
Finally, interviews with management
tive incentive scheme was adopted because
it allowed
them to
difference out common productivity shocks, such as those derived
from weather
and field conditions, that are a key determinant
of
productivity in this setting. The management
eventually decided
to move
to piece rates because
productivity had been lower
than they initially expected at the start of the season. Whether
to piece rates had the desired effect is analyzed
the move
in
Section IV.
III.B.
Descriptive
Analysis
I provides
information on unconditional
worker pro
rose significantly
ductivity by incentive scheme. Productivity
from an average
of 5.01 kg/hr in the first half of the picking
season under relative incentives to 7.98 kg/hr in the second half
of the season under piece rates, an unconditional
increase of 59
Table
percent.
Figures I and II show disaggregated
productivity data across
time and across workers under the two schemes. Figure I shows
the mean ofworker productivity over time in the two fields that
were operated for the most days under each incentive scheme.
these fields contribute one-third of the total worker
observations.
Under relative incentives, there is no dis
field-day
cernible trend in productivity. With
the introduction of piece
rose
a higher level until the
at
and
remained
rates, productivity
end of the season.
Together,
Figure II shows kernel
ductivity by each incentive
of individual pro
density estimates
scheme. The productivity of each of
QUARTERLY JOURNAL OF ECONOMICS
928
TABLE
Mean,
Relative
Worker
(.208)
[4.53, 5.49]
Confidential
41.1
23.2***
-.475
38.1
1.80
-.105***
Confidential
picked
-3.11
(1.29)
Confidential
per kilogram
2.97***
[7.57, 8.39]
Confidential
field
Difference
7.98
(2.38)
Daily pay
Unit wage
Interval
rates
Piece
(.243)
Kilos
picked per day
worked
per day
of workers
in same
Number
incentives
Hours
Variables
Confidence
5.01
(kg/hr)
productivity
and Other
and
in Parentheses,
in Brackets
Errors
Standard
I
in Productivity
Differences
Unconditional
*** denotes
significanceat 1 percent. Sample sizes are the same as those used for the productivity
regressions. Standard errors and confidenceintervals take account of the observations being clustered by
field-day.Productivity ismeasured in kilograms per hour. Daily pay refersto pay frompicking only.Both
daily pay and the unitwage per kilogram picked aremeasured inUK Pounds Sterling. Some informationin
the table cannot be shown due to confidentialityrequirements.
the 142 workers in the sample is averaged within each incentive
scheme in this figure. The mean and variance of productivity both
rise moving from relative incentives to piece rates.
The second and third rows of Table I reveal that the increase
First Field
Second
Field
Introduction
ofPieceRates
June30th
Figure
Productivity
(kilogram/hour)
31st
July
I
over the Season
Aug31st
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
929
Productivity Under
Relative Incentives
.12
.1
.08
Productivity Under
Piece Rate
.06
.04
.02
0
0
4
12
8
16
20
Productivity (kilogram/hour)
Figure
Distribution
of Productivity
II
(kg/hr) by Incentive
Scheme
in productivity was entirely due to workers picking more fruit
over the same time period, rather than working shorter hours. On
average, workers picked 23.2 more kilograms per day under piece
rates?a
level. Hours
significant difference at the 1 percent
worked did not significantly change across incentive schemes.
in Section II makes
The discussion
clear that the size of the
over
relative
is
which
group
pay
computed is key for understand
workers'
behavior
incentives. The fourth row of
under
relative
ing
Table I reports the average number of people each worker worked
with on a given field-day. This remained constant throughout the
season. The fact that under relative incentives, the unit wage
on the average productivity of 40 workers has two im
depends
First, the effect each worker has on her own pay is
plications.
and so if workers
do have social preferences
there
negligible,
should be no difference in their productivity under the two incen
tive schemes. Second, by exerting effort, the worker reduces the
ifwork
pay ofmany coworkers under relative incentives. Hence,
ers behave as if they have social preferences, there may be a large
in their productivity between when
divergence
they are paid
a
to
relative
incentive
and
scheme
according
piece rates.
farm level data reveals that the
of aggregate
analysis
in
not coincide with a wave of new
incentive
scheme
did
change
nor
did
it
hasten
the
arrivals,
departure ofworkers. Indeed, very
The
930
QUARTERLY JOURNAL OF ECONOMICS
few workers
left before
or just
after
the change
in incentive
schemes.10
Second, we find that the total kilograms
picked per day
shows no discernible
trend under either incentive scheme, a con
sequence of the deliberate
timing of planting of fields to ensure a
constant stream of fruit throughout the season. Third, the total
man-hours
spent picking are higher under relative incentives.
This is due entirely to more workers picking, rather than each
worker picking for longer hours. Under
piece rates the total
as
are required to
man-hours
falls
fewer
workers
spent picking
each
pick
day.
total kilograms
Overall, while
picked and the time spent
constant
remained
the sea
picking per field-day
throughout
to picking fell mov
allocated
son, the total number of workers
to piece rates. Under piece rates,
incentives
ing from relative
the management
had some workers
pick less frequently and
to the
instead had them perform other tasks, mostly
related
and
of
fruit.
These
workers
had
the
transportation
packaging
same productivity as workers who continued regularly on pick
such as
ing tasks. They also did not differ on characteristics
gender and nationality.
Figure Ilia shows the wage paid per kilogram over time?the
unit wage
deviation
from its mean.11 Under
?, in percentage
relative incentives the unit wage rises gradually as productivity
declines. This is as expected given that under the relative incen
tive scheme, the unit wage is set endogenously
according to (6).
With the introduction of piece rates there is a one-off fall in
the unit wage. Table I shows that the difference in average unit
wages between the two halves of the season is significant at the 1
percent level. It is, therefore, unlikely that the observed rise in
of higher returns to the marginal
is a consequence
productivity
unit of effort under piece rates. To the contrary, the pay per unit
of effort is lower under piece rates.
Figure Illb then shows the daily pay from picking over the
from its mean.
Given
that
deviation
season, as a percentage
productivity and unit wages are inversely related to each other,
average workers'
pay remained
relatively constant over time.
Table
I shows
that the difference
in average
10. This and other results not reported here are
and Rasul
2004].
paper
[Bandiera,
working
Barankay,
we cannot
to confidentiality
11. Due
requirements
daily pay between
available
show
in a companion
this series
in levels.
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
931
100
75
Ei
E
=
O) i
? E
Introduction
ofPieceRates
50
25
o c
ti
-50
-100
June30th
Figure
Unit Wage
Illa
over the Season
100
75
50 H
25
OH
-25
-50
-100
Figure Illb
Daily Pay over the Season
sizes are the same as those used for the productivity
The
Sample
regressions.
over all fields operated
on each day. This
series for the daily wage
is an average
on each field-day. The series for
is weighted
ofman-hours
average
by the number
over all workers
each day. This average
is weighted
daily pay is averaged
by the
on each day.
hours worked
per worker
relative incentives and piece rate is positive but not significantly
different from zero. Overall, the average worker became worse off
under piece rates: their productivity rose, while total compensa
QUARTERLY JOURNAL OF ECONOMICS
932
tion did not increase significantly. The regression estimates
in
the next section, therefore, most likely provide a lower bound of
in incentives on productivity holding
the effect of the change
the top third of workers
did have
utility constant.12 However,
significant increases in pay moving to piece rates (not reported),
which is as expected ifworkers are of heterogeneous
ability.
in four stages. Section
The empirical analysis proceeds
IV
on
in
whether
incen
the change
presents reduced-form evidence
tive schemes causes the increase in productivity. Section V pre
on workers'
sents structural-form
estimates
social preferences
that are consistent with the observed
change in productivity.
Section VI provides reduced-form evidence on how the relation
ship among coworkers affects individual productivity under the
two incentive schemes. Section VII provides evidence to distin
behind
guish between pure altruism and collusion as motives
workers'
behavior.
IV. Reduced-Form
IVA.
Empirical
Evidence
of the Response
to Incentives
Method
We first identify the effect of the change in incentives on
individual worker productivity. We estimate the productivity of
worker / on field f on day t, ytft, using the following panel data
are in logarithms:
regression, where all continuous variables
(7)
=
yift ai+\f+
yPt + hXm + i\Zft+ Kf + uift.
de
Worker fixed effects a? capture time-invariant, worker-level
such as innate ability and intrinsic
terminants of productivity
field
Field fixed effects Xf capture time-invariant,
motivation.
of productivity such as soil quality and plant
level determinants
spacing. Pt is a dummy equal to one when piece rates are in place
of
and zero when relative incentives are in place. The parameter
from
the effect of the move
is 7, namely
interest throughout
relative incentives to piece rates on individual productivity.
across all fields,
As piece rates are introduced simultaneously
to control for day fixed effects. Instead, we
it is not possible
in the incentives
that
literature
12. We maintain
the standard
assumption
in the piece rate. We discuss
level of effort is increasing
the utility-maximizing
can explain
in
increase
the observed
income effects or income targeting
whether
in Section
IV.
productivity
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
933
and
control for time-varying factors at both the individual
(Xift)
field
level and for a farm level trend t.
(Zft)
term uift captures unobservable
determi
The disturbance
nants of productivity at the worker-field-day
level. Worker obser
vations within the same field-day are unlikely to be independent
since workers face similar field conditions. We account for this by
clustering standard errors at the field-day level in all productivity
regressions.13
TV.B. Baseline
Results
II presents the baseline
estimates of the effect of the
Table
(1)
change in incentive scheme on worker productivity. Column
regresses worker productivity on a dummy for the introduction of
piece rates, clustering standard errors by field-day. Productivity
rises significantly by 70 percent when moving from relative in
centives to piece rates.
Column
(2) controls for worker fixed effects, so that only
variation within a worker over time is exploited, while column (3)
controls for field fixed effects, so only variation
additionally
within a worker picking on the same field over time is exploited.
Controlling forworker heterogeneity
improves the fit of the model
fixed
effects
almost
double the explained
considerably?worker
variation in productivity. In contrast, field heterogeneity
appears
to be much less important. The estimated effect of the change in
incentives on individual productivity remains significant and of
as in column (1).
similar magnitude
Column
of
(4) controls for other time-varying determinants
productivity at the level of the farm, field, and individual. We
include a linear time trend to capture farm level changes over
of each field's life cycle to capture field level
time, a measure
of each worker's picking expe
changes over time and a measure
rience. We measure
the field's life cycle as the number of calendar
in time, divided by
days the field has been picked at any moment
the total number of days the field is picked over the season. Each
field is picked for a predetermined
number of days that depends
on the number and the age of the plants, which were planted
earlier in the season or in the previous years. The field's life cycle
13. We also allowed
to be clustered
at the work level and at the
observations
worker-incentive
scheme
level to account
for idiosyncratic worker
characteristics
over different field-days
that lead to worker productivity
being correlated. Doing
so caused
errors to fall considerably.
the estimated
standard
934
QUARTERLY JOURNAL OF ECONOMICS
TABLE
The
Effect
of the Change
dependent
robust
standard
variable
picked
(kilogram
errors
reported
at
(1)
Unconditional
Piece
.530***
Field
per hour
per field-day)
allowing
in parentheses,
field-day
level
(2)Worker
heterogeneity
for clustering
(3) Field
heterogeneity
(4)
Controls
rate dummy
(.059)
Time
II
in Incentives
on Individual
Productivity
= log of worker's
productivity
.515***
(.056)
.460***
(.070)
trend
.577***
(.098)
.004
(.003)
-1 16***
life cycle
(.362)
Worker
experience
(.031)
Worker
Field
fixed effects
fixed
Adjusted
Number
effects
R2
No
Yes
Yes
Yes
No
No
Yes
Yes
.1607
.2925
.3407
.3640
10215
10215
10215
10215
of
observations
(worker-field-day)
***denotes
**
*
significanceat 1 percent, at 5 percent,and at 10 percent. Standard errorsare clustered
at the field-day level.All continuousvariables are in logs.The piece rate dummy is set equal to zerowhen
relative incentivesare in place, and set equal to onewhen piece rates are in place. The sample is restricted
toworkerswho have worked at least ten days under both incentiveschemes.The field lifecycle is definedas
thenumber ofdays the fieldhas been operated at anymoment in time,divided by the totalnumber ofdays
thefield is operated over the season.Worker experience is definedas thenumber offield-daystheworker has
picked for.There are 142workers, 22 fields,and 108 days in the sample.
the natural
trend in productivity
that
variable,
thus, captures
occurs within each field as it depletes over time.
is defined as the number of field-days the
Picking experience
worker has picked for. To identify the effect of the field's life cycle
and the worker's picking experience from the general farm trend,
we use the variation
arising from the fact that different fields
start being picked at different points in time, that fields are not
picked every day, and that different workers arrive on different
dates.
We find that there is no trend in productivity over time at the
level of the farm. This is consistent with the fact that different
fields are operated at different times to ensure a constant stream
of output throughout
the season. Within
each field, however,
as the field is picked
declines
later in its cycle.
productivity
SOCIAL PREFERENCES,
Moreover,
there are positive
RESPONSE
returns
TO INCENTIVES
to picking
experience
935
as
expected.14
The results
confirm the quantitative
of the
importance
change in incentives on productivity. A one standard deviation
increase in the field life cycle reduces productivity by 20 percent,
in picking experience
while a one standard deviation
increase
increases productivity by 7 percent. In comparison, the move from
relative incentives to piece rates significantly increases produc
tivity by 78 percent.
Further analysis, not reported for reasons of space, shows
that this result is robust to controlling for other time-varying
factors including contemporaneous
condi
and lagged weather
fixed
field
and
the
ratio
of
tions,
effects,
supervisor
supervisors to
workers.15
TV.C
Robustness
Checks
As the change in incentives occurs at the same time in all
fields, identification of the effect of this change on productivity
over time of the same worker. The
arises from a comparison
estimated
effect 7 is then biased upward
to the extent that it
captures factors that cause productivity to rise through the sea
son regardless
of the change in incentive schemes and that are
not captured by the farm level trend or the field-specific life cycle.
Table III presents a series of robustness
checks to precisely
address
this concern. First, we augment
the sample by adding
observations
from the same farm in 2004 when
worker-field-day
workers were paid piece rates throughout. We restrict the sample
to workers who picked at least ten days before and after July 8.
The 2004 sample, thus, consists of 55 workers, 18 fields, and 3664
observations. We estimate
the productivity
of
worker-field-day
worker i on field f on day t, yift, using the following panel data
regression:
14. To check robustness
to functional
we also ran the
form specifications,
in log-linear
on the levels of picking expe
form (i.e., log productivity
regression
on
form (i.e., log productivity
rience, field cycle, and trend) and in log-quadratic
the levels and the squares
of picking experience,
field cycle, and trend). Finally, we
also allowed
for field-specific
trends. The estimated
effect of the introduction
of
is robust to these alternative
are not
These
results
piece rates
specifications.
reasons
for
are
of
and
on
available
from
the
authors
space
reported
request.
15. Each
a group of 15 to 30 workers.
is assigned
The supervisor
supervisor
is primarily
for
that
fruit
is
taken
from
the
field
for pack
responsible
ensuring
in production.
are paid a fixed wage
bottlenecks
aging and preventing
Supervisors
the season.
throughout
QUARTERLY JOURNAL OF ECONOMICS
936
TABLE
The
Effect
of the Change
III
in Incentives
on
Individual
Productivity-Robustness
Checks
= log of worker's
dependent
variable
productivity
robust
standard
picked
(kilogram
errors
reported
at
per hour
(1) Difference
in difference
with 2004
rate dummy
piece
(2)
(3)
Placebo
(.098)
.096
rate
dummy (2004)
(.086)
.156
(.196)
-.009
(0.91)
on number
Field
of days
on the farm
fixed effects
fixed
Other
Adjusted
Number
days
(.110)
piece rate based
on field life cycle
Placebo
piece rate based
Worker
(4) Twenty
.387***
Placebo
present
piece rate:
workers
.577***
(2002)
Placebo
for clustering
Placebo
piece rate:
fields
season
Piece
per field-day)
allowing
in parentheses,
field-day
level
effects
controls
Yes
Yes
Yes Yes
Yes
Yes Yes
Yes
Yes Yes
R2
of observations
No
Yes
.4149
.4927
13879
.5921
2863
.2922
2969879
(worker-field-day)
*** denotes
**
*
significanceat 1 percent, at 5 percent,and at 10 percent. Standard errorsare clustered
at the field-day level.All continuousvariables are in logs.The piece rate dummy is set equal to zerowhen
relative incentivesare in place, and set equal to one when piece rates are in place. Other controls include
worker picking experience,field lifecycle,and a linear time trend.The sample in column (1) coversworkers
who have worked at least ten days under both incentiveschemes in 2002 and workers who have worked at
least ten days either side ofJuly 8 in 2004. Control variables are year specific.The sample in column (2)
coversworkers who have worked at least ten days under both incentiveschemes and fields thathave been
operated exclusivelyunder one incentivescheme.The sample in column (3) coversworkerswho have arrived
at the farmafter the introductionofpiece rates. The sample in column (4) is restrictedto the last ten days
under relative incentivesand the first ten days under piece rates.
=
(8) yift
a? + kf+ X lsP4st + 2 KXiftdst
s=l
s=l
2
+ 2
where
season
2
+
T\8Zfld8t S
Kstdst+ um,
= 1 for the 2002
{1,2} identifies the season and d\
=
1 for the 2004 season and 0
and 0 otherwise, while df
s E
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
937
otherwise. We define a placebo piece rate dummy Pt for the 2004
season to be equal to one after July 8; that is, when piece rates
were
in 2002. Thus, y1 measures
introduced
the effect of the
introduction of piece rates in 2002 on individual productivity and
the effect of the placebo piece rate in 2004. As in the
72 measures
previous specification we include workers' fixed effects a?, fields
fixed effects Xf, workers' experience
the field life cycle
Xift,
Zft,
and a farm level trend, allowing these variables
to be different in
the
two
seasons.16
in column (1) indicate
The difference-in-difference
estimates
that the previous result was not due to seasonality:
the placebo
dummy for piece rates for the 2004 season has no effect on
coef
productivity. Interestingly, we also find that the estimated
ficients of the farm level variables
and
life
field
(trend
cycle) are
the same in the two seasons.
In contrast, returns to experience
are twice as large in 2004, in line with workers holding back effort
under relative incentives in 2002.
The second and third tests simulate the introduction of piece
rates in fields and forworkers that did not actually experience the
change in incentive schemes in 2002.
For fields, we note that the two main fields operated most
the change
frequently under both incentive schemes, experienced
in incentive scheme at one-quarter of their life cycle. We construct
a placebo piece rate dummy for each field, set equal to one after a
field has passed 25 percent of its life cycle and zero otherwise. We
then examine whether this placebo dummy affects the productiv
ity of the sample of fields that are only operated under either
relative incentives or piece rates. The result in column (2) shows
no evidence of a natural jump in productivity on fields after they
pass 25 percent of their life cycle.
Column
(3) exploits the same idea at the worker level. In the
baseline
of
sample, workers had been picking for an average
nineteen days before the change in incentives. For the placebo
test we exploit information on workers who arrived after the
introduction of piece rates. We create a placebo piece rate dummy
for each such worker set equal to one after that worker has been
picking for nineteen
days. The result in column (3) shows no
some of the fields observed
16. While
in 2004 are the same as those in 2002,
we allow the fixed effect to be different to capture the fact that the plants are at
a different stage of their life in the two years. None
of the workers
is present
in
both seasons.
QUARTERLY JOURNAL OF ECONOMICS
938
evidence of the natural jump in worker productivity at this level
of picking experience.
Finally, in column (4) we restrict the time window to ten days
the
before and after the introduction of piece rates to eliminate
in
effect of natural
This
long-term changes
productivity.
specifi
cation yields our lowest estimate of 7. The effect is nevertheless
large: for the average worker, productivity was 47 percent higher
during the ten days after the introduction of piece rates compared
with that during the last ten days under relative incentives.17
Further
indicates
that results are robust to the
analysis
of
different
into picking over time, and
selection
workers
possible
to controlling for changes in task composition in the two halves of
the season. We also find that the introduction of piece rates was
in the sense that productivity did not significantly
not anticipated
change in the week prior to the introduction of piece rates, that
reacted slowly to the change
workers
(possibly in the hope of
incentive
scheme), and that the quanti
reinstating the relative
to piece rates on productivity
lasted
tative effect of the move
the
season.18
throughout
the increase in productivity
A remaining
issue is whether
came at the expense of the quality of fruit picked. Pickers are
as supermar
expected to classify fruit as either class 1?suitable
as market
or class 2?suitable
is
ket produce,
produce. This
to
in
due
this
context
because,
technological
especially pertinent
of fruit cannot be traced back to
restrictions, misclassifications
the
individual workers. To check for this, we analyze whether
after the introduction of piece
of fruit worsened
misclassification
2 show that this was not the
rates. Results, reported inAppendix
case.
IV.D.
Income Targeting
and Other Hypotheses
Taken together, the results show that moving from a relative
incentive scheme to piece rates significantly increased worker
productivity by at least 50 percent. As workers' pay remained
constant on average under both incentive schemes while produc
is most
increase in productivity
tivity increased, this estimated
on
a
in
effect
of
the
the
lower
bound
pure
incentives,
change
likely
holding worker
utility constant.
the shorter
time
17. Given
controls in this specification.
checks
18. These
robustness
and Rasul
[2004].
frame,
we
are discussed
do not
include
in detail
other
in Bandiera,
time-varying
Barankay,
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
939
for the change in
We now discuss alternative
explanations
workers' behavior from relative incentives to piece rates. While
the negative externality is the most evident source of difference
between the two schemes, workers' behavior might possibly re
flect income targeting, their reaction to uncertainty, and ratchet
concerns. We address these below.
in Section III, the unit wage per kilogram
First, as discussed
decreased by 12 percent moving from relative incentives to piece
rates. If workers
adjust their effort to reach a constant daily
income target, the fall in the unit wage may cause the observed
increase in productivity. To judge the empirical relevance of in
come targeting
in our context, it is important to stress that
workers cannot choose the number of hours they work, implying
that the standard income-leisure
trade-off does not arise. In other
can
for
workers
words,
adjust only on the intensive margin,
instance by choosing towork harder when the unit wage is low, to
achieve their income target.19
Three pieces of evidence cast doubt on the empirical rele
vance of income targeting in this setting. First, we find that
workers who face higher piece rates work harder. To establish
this, we exploit the fact that the real value of piece rates varies
among workers who come from countries with dif
exogenously
save most of
ferent levels of GDP per capita, and that workers
their earnings
workers
come
to bring back to their home
from
Eastern
Europe,
there
countries. Although
are
large
all
cross-country
differences. For instance, gross monthly
earnings at the mini
mum wage are four times higher in Poland
(euro 210) compared
with Bulgaria
that more
(euro 56). In line with the assumption
we
find workers
high powered incentives result in higher effort,
who come from poorer countries have higher productivity, all else
equal. To the extent that the worker pools from different coun
tries are not selected differently, the cross-sectional
evidence does
not lend support to the income targeting hypothesis.
to exoge
Second, we find that workers' daily pay responds
nous variation
in weather
conditions. In particular, workers earn
more when the temperature
is milder. This is in contrast to the
19. Other
in different settings
of income targeting
reach mixed
analyses
conclusions.
have a strong
increases
[1999] finds that exogenous
wage
Oettinger
and positive
effect on the labor supply of stadium vendors, which
is not consistent
with daily income targeting. Camerer
et al. [1997] find that New York cabdrivers
work fewer hours when
the observed
is higher and interpret this as
daily wage
in favor of income targeting.
evidence
In contrast, F?rber
similar
[2004] using
evidence
such income targeting by cabdrivers.
data, presents
against
QUARTERLY JOURNAL OF ECONOMICS
940
that workers adjust their effort levels to achieve the
hypothesis
same absolute daily income target.
Finally, as we show in Section VI, productivity under relative
incentives depends on the social relationships
between workers in
the field and on the number ofworkers in the field, while neither
of these two factors affect productivity under piece rates. These
that income targeting is
findings cast doubt on the hypothesis
in
the
difference
for
fully responsible
productivity under the two
incentive schemes, as income targeting does not predict that
social connections and group size should have a different effect
under the two schemes (if at all).20
Another difference between
the relative scheme and piece
rates is that under the latter, the unit wage is set ex ante at the
of the field-day, whereas
under the former the unit
beginning
at
is
the end of the field-day,
determined
wage
endogenously
based on workers' productivity. Workers may then work less hard
under relative incentives because
of uncertainty over the ex post
unit
wage.
Such uncertainty may play a role in the first days a worker
picks, but is unlikely to be driving the difference in productivity
on the unit wage based on
given that workers form expectations
re
observations
each field-day they pick. Simulation
repeated
can
the
observed
sults further show that uncertainty
only explain
of the variance are
change in productivity ifworkers' expectations
orders ofmagnitude
larger than is observed in the data.
and piece rates also differ because
incentives
relative
Finally,
if they believe that
under piece rates workers may underperform
in
hard
will
result
management
setting lower piece rates
working
in the future. In such a dynamic framework, productivity under
piece rates is lower than implied by (4). This ratchet effect does
not occur under relative incentives because the unit wage is based
on a given field-day.
on the average
productivity
exclusively
in the presence of such ratchet effects, the true effect of
Hence,
the change in incentives on productivity is underestimated.
There are a number of reasons why in this setting, there are
unlikely to be such large ratchet effects. First, given the stochas
it is difficult for workers to
tic nature of agricultural
production,
disentangle
changes in the piece rate due to changing conditions
and those due tomanagement
learning about workers' true abil
20. We
thank
an anonymous
referee
for suggesting
this point.
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
941
ity.21 Second, the effect of a worker's current performance on the
unit wage
she faces in the future is weak as the unit wage
is
are
on
to
workers
and
reallocated
different
fields
field-day specific
different days.22 Finally, any ratchet effect should become weaker
as the time horizon of the worker becomes shorter. We checked for
this and did not find worker's productivity to change significantly
in their last week of work.
V. Structural-Form
V.A. Empirical
Evidence
on Workers
Preferences
Method
We now use the data on worker productivity to draw impli
in
cations for workers' behavior
in light of the models discussed
Section II. Our first aim is to assess whether the observed change
in productivity
is consistent with the standard assumption
that
on
the
workers
others
under
the
ignore
externality they impose
=
or
relative scheme
whether
internalize
it
0),
they fully
(tt?
=
1).
K
To this purpose, we use the first-order conditions
of the
workers' maximization
problem derived in Section II to compute
an estimate
of each worker's
cost parameter,
0?, under each
incentive scheme and behavioral
Since the workers'
assumption.
are innate, we ought to find the same
cost (ability) parameters
implied distributions of costs across workers under both incentive
schemes if the underlying behavioral
is correct.
assumption
are paid on the basis of their observed productivity
Workers
y which is a function of their effort e. Taking this into account, the
first-order conditions for the choice of effort under relative incen
tives assuming
that workers do not internalize
the externality
=
that they do fully (ir? = 1), and under piece
0), assuming
(tt?
rates are respectively,
concerns
21. Such
have
ratchet
been documented
to exist in workplaces
where productivity
shocks are less common
such as shoe making
and
[Freeman
Kleiner
1998] and bricklaying
[Roy 1952].
22. In particular,
over which
face uncertainty
workers
fields they will be
to in the future?the
a worker works on the same field on two
assigned
probability
are
consecutive
is
.25.
Workers
also uncertain
the identity of their
about
days
future coworkers.
QUARTERLY JOURNAL OF ECONOMICS
942
(10)
dji
BeiGiyd2
*i'??l*-l*p>
dl)
N
^?e"
+'(?yi)?g7=6i?i.
To derive estimates of 0? in each case, we proceed as follows. First,
we assume
that the benefit function is of the following CRRA
type,
> 1.
(12) 4>(y)
pyyp for p
= 2. The results
Throughout we report results obtained with p
choices of p. Second, we derive
are, however, robust to alternative
a Cobb-Douglas
an estimate of worker effort, e?, assuming
rela
tionship between effort and productivity, as explained below.23
Third, we substitute data (on each field-day) for estimated
effort, observed productivity
(?) and group size
(y?), unit wages
then obtain an
(N), into the first-order conditions above. We
estimate of 6? on each field-day the worker picks, 6?^, and take the
estimate of 6?, under each
of these to derive a unique
median
incentive scheme and behavioral
assumption.24
on the
of 0? based
therefore, derive three estimates
We,
calibration of the first-order conditions (9), (10), and (11), respec
that workers do not
tively, (i) under relative incentives assuming
under
relative
incentives
internalize
the externality,
(ii)
?fN;
workers
the
internalize
externality, namely that
assuming
fully
efforts
choose
(iii) under piece
cooperatively,
they
0fc; and
rates,
Of.
Finally, we compare the distribution of of with the distribu
either of these two as
tions of ?fN and Ofc to assess whether
on the underlying behavior
of workers
is consistent
sumptions
with the observed change in productivity.
We assume that workers' effort e translates into productivity
function. To estimate
y through a Cobb-Douglas
production
23. This
implies that the same effort on two different days can lead to two
on other inputs into production,
such as
different levels of productivity
depending
In the first-order conditions
field conditions.
(9) to (11), dy?/de? <*y?/e? with a
so that 6? is identified up to some scalar in each case.
Cobb-Douglas
specification,
the first-order
of the estimated
This does not affect the comparison
0?'s across
conditions.
as sample workers work at least ten field-days
24. The model
is overidentified
as this is less sensitive
to
under each incentive scheme. We use 6? = median
(Qift)
or to estimating
are
of the
robust to taking the mean
outliers. The results
6?^'s
likelihood.
them for each worker using maximum
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
943
effort, we first estimate the productivity regression as in
of productivity as in the
(7) controlling for the same determinants
baseline
II, and interacting
specification of column (4) in Table
each worker fixed effect with the piece rate dummy. The estimate
ofworker ?'s effort in field f on day t under incentive scheme s E
{R,P} is each worker's estimated fixed effect added to the residual
from the regression
(7) when incentive scheme s is in place:
worker
=
+
u\n.
e\n df
(13)
The first term captures the worker's average effort over time
under incentive scheme s. The second term captures how much of
cannot be explained
the worker's
productivity
by observables.
effort (mea
This method provides an estimate of each worker's
on
on
which
sured in kilograms
hour)
every field-day
per
they
pick.
with the actual distribution
of productivity by
Consistent
and variance
of effort
incentive scheme in Figure II, the mean
both rise significantly moving from relative incentives to piece
rates.25 Moreover, we find that nearly all workers put in more
effort under piece rates than under relative incentives and that
there is little churning among workers?those
that exert the most
effort under relative incentives continue to exert the most effort
under
V.B.
piece
rates
Individualistic
and
vice
versa.
versus Fully
Cooperative
Behavior
Figure IVa shows the kernel density estimate of the implied
of workers' cost of effort Of and $fN, namely under
distribution
the assumption
that workers
ignore the externality they impose
on others under the relative scheme (tt? = 0). This shows that the
distribution of cost parameters
under relative incentives lies al
most entirely to the right of the distribution under piece rates,
indicating that the implied cost of effort is higher under relative
incentives than under piece rates.
that cost of effort is an innate parameter,
the fact
Assuming
that the same distribution of costs cannot be fitted to both incen
tive schemes indicates that effort choices are not consistent with
25. Splitting
the estimated
effort (13) into each of its components?the
resid
ual
and the worker fixed effect ?s??we
find that the exponent
of the residuals
usift,
zero under
is centered
around
each scheme,
but the variance
of the residuals
under relative
is significantly
incentives
it is the distribution
of
higher. Hence,
worker
fixed effects, and not the residuals,
that drives
in the
the differences
of effort. See Bandiera,
distributions
and
Rasul
for
details.
[2004]
Barankay,
944
QUARTERLY JOURNAL OF ECONOMICS
Cost of Effort
Parameter (9)
UnderPiece Rate
Cost of EffortParameter (0)
UnderRelative Incentives
1 1.5
Cost of EffortParameter (9)
Figure
Kernel
Density
IVa
Estimates
of Cost of Effort Parameter,
Behavior
Individualistic
Assuming
by Incentive
Scheme
Cost of Effort
Parameter (0)
UnderRelative Incentives
Cost of EffortParameter (9)Under
Piece Rate
Cost of EffortParameter (0)
Figure
Kernel
IVb
of Effort Parameter,
Scheme
by Incentive
Behavior
Assuming
Cooperative
are calculated
Kernel
estimates
kernel. The
density
using an Epanechnikov
to be
benefit function is assumed
underlying
- 2x112.
<p(z)
Density
Estimates
of Cost
to be quadratic
The total cost of effort is assumed
in effort. Under
individualistic
behavior we imply that the worker
chooses her effort to maximize
her own net
benefits. Under
behavior we imply the worker
chooses her effort level
cooperative
to maximize
the sum of all workers
utilities.
SOCIAL PREFERENCES,
workers
ignoring
relative
RESPONSE
the externality
TO INCENTIVES
they impose
945
on others under
incentives.
Next, we estimate the distribution ofworkers' cost of effort Of
that workers fully inter
and ofc, namely under the assumptions
nalize
the externality
their effort imposes on their coworkers
under the relative incentive scheme. Figure
IVb shows the im
cost
of
distributions
the
parameter
plied
6?, by incentive scheme.
The distribution
of of under piece rates is, by definition, un
the distribution of costs
changed to that derived above. However,
under relative incentives ofc now lies almost entirely to the left
of the distribution under piece rates.
If workers
chose their effort levels cooperatively,
then the
cost of effort under relative incentives would have to be signifi
cantly lower under relative incentives to fit the observed produc
tivity data. In other words, productivity is actually too high under
to be explained
relative
incentives
by workers
choosing their
effort levels cooperatively.
IVa and IVb together reveal an interesting pattern.
Figures
The observed change in productivity is too large to be reconciled
with the assumption
of individualistic
behavior but too small to
be reconciled with the assumption
of fully cooperative behavior.
as if they internalize
This suggests
behave
that workers
the
some
next subsection explores
to
extent.
The
negative externality
this idea.
V.C.
Social
Preferences
We now posit that workers have social preferences, namely
that they place some weight on the payoffs of their coworkers
of either pure altruism or collusion, and retrieve the
because
reduced-form "social weights"
(tt?) that fit the observed change in
productivity.
To do so, we assume
the true cost of effort of each worker is
that derived under piece rates of .26Given ?f, we calibrate the
first-order condition of the worker's maximization
problem when
they have social preferences,
(2),
(14)
dy?
deA^y?2
4>lfj23v-^2*'||y,
i
= Tr
6fe,
?
26. Using
this measure
of ability, we find that groups were
equally hetero
in terms of ability, before and after the change
in incentives. Hence,
geneous,
there is no evidence
of management
into fields by
sorting workers
differently
the incentive schemes.
ability across
946
QUARTERLY JOURNAL OF ECONOMICS
-.50.51
SocialWeight(tt)
Figure
Kernel
Density
Estimates
V
of Social
Weight
(tt)
This provides an estimate of each worker's social weight on every
is overidentified, we take
field-day they pick, nift. As the model
the median
of these to derive tt?.
The resulting distribution of social weights that explains the
observed change in productivity is shown in Figure V. The aver
age worker places a social weight of .65 on the benefits of all
others in the same field-day. Less than 3 percent ofworkers have
an implied social weight greater than one, and less than 2 percent
of workers have an implied social weight of less than zero.
The
next
the idea that social preferences
pursues
so
is
much
lower under relative incen
explain why productivity
tives. In particular, we explore whether the identity of coworkers
on the field-day has differential effects on workers' behavior un
der relative incentives?when
worker's effort imposes a negative
on
and
there are no
under
others,
externality
piece rates?when
on
externalities
coworkers.
imposed
VI.
section
Incentives,
A natural
Social
Networks,
and Workers
Productivity
to explain the extent to which workers
candidate
on
the
place weight
payoffs of their coworkers is the relationship
on
the extent to
workers
among
any given field-day. Naturally,
which a given worker is altruistic or able to collude might depend
on the social relations between her and her coworkers on a given
this social relation, we exploit data on each
day. To measure
worker's
self-reported friends on the farm. This information al
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
947
the
that workers
internalize
to explore the hypothesis
are
more
exter
when
less
the
and, hence,
productive
externality
nality hurts their friends rather than other workers.27 To this
in the number of friends each
purpose, we exploit the variation
worker works with on each field-day. We
identify the effect of
on
group composition
productivity by comparing the productivity
of the same worker, on the same field, working alongside different
coworkers on different days.
To obtain information on the workers' social networks on the
a questionnaire
about two weeks after the
farm, we administered
change in incentive schemes. Workers were asked to name up to
five other workers
they were friends with before coming to the
farm and up to five other workers
friends with
they became
during their stay. In the main analysis below we pool these two
lows us
categories.
All but seventeen workers report having at least one friend
on having at least one friend on the
on the farm. Conditional
who are friends ofworker i is on
share
of
coworkers
the
field-day,
4
both
schemes.
The dispersion
is also very
under
average
percent
similar under the two schemes indicating that the results under
in the share of cowork
piece rates are not due to lack of variation
ers who
are
friends.28
To be clear, the composition of the group varies each field-day.
However, workers themselves do not choose which field they work in
and with whom they work. Rather, group composition is decided by
management
taking account of the demand forworkers to perform
nonpicking tasks, and how close workers live relative to the fields
that need to be picked. The way in which workers are allocated to
fields then leads to no systematic relation between individual shocks
to productivity and group composition.29
27. Levine
and Pesendorfer
[2002] show that in an evolutionary
equilibrium
Dilemma
in which workers
of a repeated
Prisoner's
learn which
game
strategies
as if they have social preferences.
to play, players
the weight
behave
Moreover,
on the benefits
on
each player places
of another
relation
the
player
depends
between
will behave more
that, "individuals
They argue
players.
altruistically
when
of their altruism."
they can identify with the beneficiary
28. The mean
who are friends of / is .043 under relative
share of workers
are .042 and .034. The
incentives and .037 under piece rates. Standard
deviations
mean
new workers
share is slightly lower under piece rates because
arrive at the
at least one "old" friend on the farm.
farm. All but 36 workers
report having
on having at least one old friend, the mean
Conditional
share
(standard
deviation)
ofworkers who are old friends of i is .032 (.028) under relative
incentives and .023
(.024) under piece rates.
are more likely to work alongside
29. Unsurprisingly,
workers
those they live
with than other randomly
chosen individuals.
The probability
that a worker works
QUARTERLY JOURNAL OF ECONOMICS
948
TABLE
The
of Group
Effect
dependent
robust
standard
Composition
=
variable
picked
(kilogram
errors
reported
at
Share
of workers
are
who
of workers
Share
are
who
friends
log
fixed
fixed
Other
Number
for clustering
(lb)
Relative
Piece
Piece
incentives
incentives
rates
rates
-1.68***
-5.52**
(2.36)
1.60**
(2a)
(2b)
.072
1.17
(.493)
(1.60)
-.285
(.684)
(.501)
field
in same
effects
effects
controls
Adjusted R2
per field-day)
allowing
(la)
effect of group size
Marginal
friends' share)
(at mean
Field
Scheme
Relative
field
Worker
by Incentfve
productivity
in parentheses,
field-day
level
in the field
X number
of workers
of worker's
per hour
(.647)
in same
of workers
Number
in the field
friends
IV
on Productivity
of observations
.182
.085
(.117)
(.069)
.076
.236**
(.110)
(.065)
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
.3470
.3620
.3065
.3081
2860
2860
4400
4400
(worker-field-day)
*
**
*** denotes
significanceat 1 percent, at 5 percent, and at 10 percent.Robust standard errorsare
calculated throughout,allowing forclusteringat thefield-day level.All continuousvariables are in logs.The
sample is restrictedtoworkerswho have worked at least tenfield-daysunder both incentiveschemes. "Share
ofworkerswho are friends"is equal to thenumber of self-reportedfriendspresent on thefielddivided by the
total number ofworkers on the field.Other controls includeworker experience,field life cycle,and a linear
time trend.
Table IV reports estimates of the productivity regression
(7)
control for
under relative incentives, where we now additionally
group composition at the field-day level as well as the baseline
of worker productivity in column (4) of Table II.
determinants
Column
(la) controls for the share of coworkers in the same
i. Having more friends present
field who are friends of worker
reduces
productivity under relative incentives. The
significantly
estimated coefficient implies that ifworker imoved from a group
with someone
she lives with is .605, with a self-reported
friend, .549, and with
do not,
someone
.344. These
she lives with nor a friend,
neither
probabilities
over the two incentive
schemes.
however,
change significantly
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
949
with none of her friends to a group where five of her friends are
present, her productivity would fall by 21 percent.
in the same
In column (lb) we interact the share ofworkers
field who are friends ofworker ?,with the total number ofworkers
on the field. We find that (i) having more friends present signifi
cantly reduces productivity under relative incentives, and (ii) this
in the same
effect is smaller the greater the number of workers
as
field. The latter effect is
expected given that the externality
imposed by ? on her friends is smaller when the overall group size
is larger.
effect of group size
Column (lb) also shows that the marginal
is positive and significant, indicating that workers internalize the
externality less when they work in larger groups, all else equal.
The result is consistent with the intuition that larger groups may
find it harder to coordinate on the low effort equilibrium.30
in columns
The results
(la) and (lb) have some obvious
their
interpretations: when workers work alongside
exert
less
effort
and
become
less
friends, they might
productive
because
they talk and socialize with their friends. Or, alterna
tively, they might choose to work with their friends when they
feel less prone to work hard.
To shed light on these hypotheses, we analyze whether hav
ing friends around affects productivity under piece rates. Intui
the identity of coworkers and
between
tively, any relationship
to
the
that
is
unrelated
incentive scheme in place,
productivity
such as socializing with friends, will be present under both in
if the relationship
centive schemes. However,
between the iden
tity of coworkers and productivity is related to the externality, it
should affect productivity
incentives. The
only under relative
results in columns (2a) and (2b) lend support to the latter inter
pretation since the share of coworkers who are friends of ?has no
effect on productivity under piece rates.
alternative
In short, the evidence
indicates that under relative
incen
tives workers
internalize
the externality more when they work
their friends. The fact that workers' productivity is not
alongside
affected by the presence
of friends under piece rates, indicates
30. Section
II does not model
costs and, therefore, does not
coordination
this aspect.
In the model,
on workers'
the effect of an increase
in N
capture
behavior
is negligible.
effect on the mean
Indeed, on the one hand, each worker's
is smaller when
the group is larger. On the other hand, each worker affects more
the group is larger.
people when
950
QUARTERLY JOURNAL OF ECONOMICS
that group composition affects productivity only when workers'
effort imposes a negative externality on coworkers.31
One remaining concern is that friendships may be endoge
nous outcomes
of workers
with each
cooperatively
behaving
other. To check for this, we use an alternative measure
of social
networks based only on the friendships that were formed before
came to the farm and that are, therefore, uncorrelated
workers
with any events that took place on the farm itself. Reassuringly,
ifwe use this alternative
the results in Table IV are unchanged
definition of friends.32
VII.
Explaining
Worker
Behavior
The evidence presented so far is consistent with both a model
of altruism where workers care directly about the utility of others
or with a model of collusion where workers agree to cooperate for
benefit. Discriminating
between these hypotheses
is im
mutual
portant. The implications for designing a work environment that
facilitates
between
coworkers will be different de
cooperation
pending on the underlying motive for cooperative behavior. We
now present evidence to distinguish
between the hypotheses
of
pure
altruism
and
collusion.
Both hypotheses
imply that worker i behaves as ifm reduced
as specified in (1). However,
form she has social preferences
are altruistic
if the payoff of one or more coworkers
workers
enters their utility directly so that the structural- and reduced
In contrast, self
of preferences
coincide.
form representation
interested workers would reduce effort under relative incentives
if, as a group, they are able to enforce implicit collusive agree
In our context, such
ments through transfers and punishments.
the finite time horizon
collusive behavior might occur despite
because
workers
are
uncertain
over
when
they,
and
their
cowork
to treat friends
to incentives but causing
individuals
31. Any factor unrelated
over time will be spuriously
to the change
in incentive
attributed
differently
we
under piece rates, the effect of having
examine whether
scheme. To check this,
so
on
more friends
the field is different for those who arrived later and
only worked
under piece rates, compared with those who were also present under the relative
of space,
show that
not reported
for reasons
incentive
scheme.
The
results,
do not react differently to friends when
workers
they first arrive.
check we also exploit the fact that workers who arrive on
32. As an additional
friends. The
the same date from the same country are very likely to become
to any event that took
is orthogonal
is that date of arrival
identifying assumption
links.
on the farm and could have
of friendship
the formation
affected
place
are unchanged
ifwe define the group of friends to include only the people
Results
who arrived within
the same three-day window.
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
951
are also uncertain whether they
ers, will leave the farm. Workers
to different farm tasks in
will keep on picking or will be allocated
the future.
The economic environment we study has a number of fea
tures that facilitate both collusion and altruism. For example,
live and work together, interacting repeatedly both in
workers
it relatively
side and outside the work environment. This makes
a variety of
easy for them to build social ties, and provides
to provide transfers and enforce punishment.
mechanisms
It is important to stress that we focus on the distinction
between collusion and "pure" altruism, namely that workers care
about others regardless of the others' behavior. This is in contrast
to "reciprocal altruism"
[Axelrod 1984; Fehr and Fischbacher
Rabin
1993], whereby individuals are only altruistic toward
2002;
toward them, and, thus, cooperate as
those who act altruistically
as
The
collusion
others
cooperate.
key difference between
long
is that under the latter, cooperation
is
and reciprocal altruism
even in the last period of play. However, we cannot
sustained
in distinguishing
between col
prediction
exploit this particular
lusion and reciprocal altruism because we do not observe any
relative
worker
incentives are still in
leaving the farm while
place.
The difficulty in separately purely altruistic from collusive
to be more
factors that lead individuals
motives arises because
altruistic
toward each other, typically also facilitate collusion,
and vice versa. Here we exploit the fact that the ability tomonitor
coworkers creates differences in observed behavior depending on
of collusion or altruism. We
whether workers cooperate because
on
evidence
workers'
present
productivity for another fruit type
are such that, unlike for the fruit
whose physical characteristics
to monitor
the perfor
type studied so far, workers are unable
mance
of coworkers.
To sustain a collusive agreement, workers must necessarily
on the field-day.
be able to monitor
In
each other's behavior
contrast, the ability tomonitor coworkers is irrelevant ifworkers'
is driven by pure altruism. Under
behavior
altruism, workers
take into account the effect their effort has on others because
it
affects their own utility directly. Hence, they cooperate regardless
of whether
of
they are monitored
by others, and regardless
whether they can monitor coworkers' performance.
so far, monitoring
For the fruit type analyzed
others is cost
one another. Workers
are not
less since workers work alongside
QUARTERLY JOURNAL OF ECONOMICS
952
and so can easily form accurate beliefs on
physically
separated
the performance of coworkers. To establish whether the ability to
monitor
the effect on
coworkers affects behavior, we analyze
individual productivity of the change in incentive schemes for
another fruit type, which we label type 2, as opposed to type 1
fruit that is the focus of the previous analysis.
Type 2 fruit grows on dense shrubs that are 6 to 7 feet high
on average.
In contrast to type 1 fruit, when picking type 2 fruit
are unable
to observe the quantity of fruit picked by
workers
workers in neighboring rows on the field-day. Hence, the physical
of type 2 fruit ensure that workers cannot monitor
characteristics
each other on the same field-day.
Over 80 percent of our sample workers pick type 2 fruit at
some point in the season. Of these workers, 54 pick under both
relative incentives and piece rates. Only these workers are used
in the analysis below. The two samples of workers?those
who
1
ten
at
under
fruit
least
both
incentive
field-days
pick type
schemes and those who pick type 2 fruit under both incentive
nor on their productivity
not differ on observables,
schemes?do
when picking type 1 fruit. Pickers of type 2 fruit are, however,
more likely to be female.
The number of workers who pick on the same field-day is
much smaller for type 2 fruit compared with type 1 fruit?the
average group size is nine under relative incentives and thirteen
under piece rates. Type 2 fruit is picked across twelve fields over
269 field-days. Of these, 112 occur under relative incentives, and
is 2.10 kg/hr under
157 under piece rates. Worker
productivity
relative incentives and 1.62 kg/hr under piece rates.33
To estimate the effect on individual productivity for type 2
the
fruit, zift) of the change in incentive schemes, we estimate
following specification:
=
+ uift,
+
+
+ yPt +
(15)
zift a? + <pf
oXift T)Zft Kf
Pt is a dummy equal to one when piece rates are in place
and zero otherwise. As in previous specifications we cluster the
terms by field-day, and we controls for time trend,
disturbance
worker's picking experience and the field life cycle. The latter two
are of course defined for type 2 fruit.
in Table V. Column (1) reports that
The results are presented
where
piece
33. The
standard
rates is 1.07 and
of productivity
.95, respectively.
deviation
under
relative
incentives
and
RESPONSE
SOCIAL PREFERENCES,
TO INCENTIVES
953
TABLE V
variable
dependent
robust
Monitoring
= log of worker's
picked
(kilogram
errors
reported
standard
at
per field-day)
allowing
in parentheses,
field-day
(1) Fruit
type 2
Piece
rate dummy
(Pt)
-.063
(.129)
Piece
rate X fruit type 2
Piece
rate X fruit type 1
Worker
Field
fixed effects
fixed effects
Other
Adjusted
Number
controls
Yes
Yes
Yes
R2
for
clustering
level
(2) Fruit
type 1
(3) Fruit types
1 and
2
combined
.483***
(.094)
-.100
(.095)
490***
(.092)
Yes
Yes
Yes
Yes
Yes
Yes
.3015
.3777
.6098
934
4224
5150
of observations
(worker-field-day)
productivity
per hour
*** denotes
**
*
significanceat 1 percent, at 5 percent, and at 10 percent.Robust standard errorsare
calculated throughout,allowing forclusteringat the field-day level in all columns.All continuousvariables
are in logs.When picking fruittype 1workers can monitor one another.When picking fruittype2 workers
cannotmonitor one another.The piece rate dummy is defined tobe equal to zerowhen relative incentivesare
in place, and one when piece rates are in place. Other controls includeworker picking experience,field life
cycle,and a linear time trend.The sample in all columns is restrictedtoworkerswho have picked fruittype
2 at least once under both incentiveschemes.
in the baseline
(15), there is no significant effect on
specification
worker productivity moving
from relative
incentives
to piece
rates. The pattern of the other coefficients is, however, similar to
the type 1 fruit.
The result suggests that when the production
technology is
such that coworkers' performance
cannot be monitored, workers
do not internalize
the externality they impose on others under
relative
under
incentives; namely
they are equally productive
relative incentives and under piece rates.34
One possibility is that, in contrast to the average worker, this
34. Further
not reported for reasons
of space, shows that the result
analysis,
to restricting
is robust
the sample
to ten days
either side of the change
in
incentives. Moreover,
results are robust to restricting
to field-days
the sample
where workers
only picked type 2 fruit and performed no other tasks. We also find
that workers
do not anticipate
the change
in incentives, and that piece rates have
no effect on productivity
at any point of the season.
954
QUARTERLY JOURNAL OF ECONOMICS
ofworkers simply do not cooperate, regard
particular subsample
less of the monitoring
technology. To check this, in column (2) we
reestimate
the baseline
for this subsample
when
specification
of coworkers can be
they pick type 1 fruit, where the behavior
monitored.
as
The results show that these workers
cooperate
much as workers in the larger sample when monitoring
is feasi
ble. The effect of the introduction of piece rates is significant and
as in the larger sample.35
of similar magnitude
In column (3) we combine the observations across fruit types for
workers who pick both fruit types under both incentive schemes.
The result shows that there is a significant difference-in-difference
in the response of individual worker productivity to the introduction
of piece rates between finit type 1 and fruit type 2.
the results indicate the effect of the change in in
Overall,
on the
centives on worker productivity depends
fundamentally
ability ofworkers to monitor their coworkers. When workers are
able to monitor each other (type 1 fruit), productivity
is signifi
cantly lower under relative incentives. In contrast, productivity is
identical under both schemes when workers cannot monitor each
other (type 2 fruit). Given that monitoring
is necessary
to enforce
it does not affect altruism,
collusion while
the comparison
of
out
fruit
incentive
and
scheme
rules
the
type
productivity by
that internalize
the externality
because
hypothesis
they are
purely altruistic.
VIII.
Conclusion
This paper provides evidence on social preferences by com
paring workers' productivity under a relative incentive scheme
with the productivity of the same workers under piece rates. Our
estimates
indicate that moving from relative incentives to piece
rates causes productivity to rise by at least 50 percent for the
average worker. We show that the observed change in productiv
the externality their
ity is consistent with workers
internalizing
effort imposes on coworkers to some extent. We also find that
internalize
the externality to a greater extent when a
workers
larger share of their coworkers are their close friends. Finally, we
find that workers internalize the externality only when they can
of workers
this sample
35. When
picks fruit type 1, average
productivity
is 4.80 kg/hr, and 8.01 kg/hr under piece rates. These are
under relative incentives
not significantly
different from the sample used for Table
I.
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
955
These results are among the
monitor others and be monitored.
first to precisely identify an economic environment outside of the
laboratory, where behavior can be explained by individuals' hav
in reduced form.36
ing social preferences
Throughout, we have taken the incentive schemes as given.
to an exogenous
Our focus has been the response
of workers
for their underlying
change in incentives and the implications
issue is whether
the observed incentive
preferences. A separate
schemes are optimally designed by the principal. Two questions
to
arise. First, if the relative incentive scheme was so detrimental
it ever adopted? Second, are piece rates
productivity, why was
optimal in this context?
the use of relative schemes, the farm management
Regarding
that
the relative scheme was mainly adopted to differ
suggested
ence out common shocks that are a key determinant
of workers
in
this
productivity
setting.37
While
this is in line with the predictions of incentive theory,
the superiority of relative incentives relies on the assumption
that workers ignore the externality their effort imposes on others.
Under these conditions the equilibrium
effort choices under rela
tive incentives and piece rates are approximately
equal for large
on worker behavior is not supported
group sizes. This assumption
in our data. Relative
incentives led to lower productivity because,
perhaps surprisingly, workers internalized the negative external
ity to some extent.
The finding that workers place a positive weight on their
coworkers' payoffs also indicates that piece rates might not be
internalize
optimal in this context. To the extent that workers
a
in
and
externalities
similar
way, group incen
negative
positive
schemes
where
the
worker's
and
her coworkers'
tives, namely
pay
are
more
elicit
effort at the
performance
positively related, might
same cost to the principal. To explore this issue further, we use
our estimates
of worker ability and social weights
to simulate
effort levels under
group
incentives.
36. Relatedly,
List
of the sellers of baseball
the behavior
[2004] compares
cards in the laboratory and in the marketplace.
He presents
evidence
that "local"
sellers
in the market)
social pref
(namely sellers who regularly
operate
display
erences both in the laboratory
and in the market,
when
especially
they interact
with buyers with whom
they have a long-term relationship.
37. See Lazear
and Rosen
and
[19811, Green and Stokey
[1983], and Nalebuff
evaluation
to piece
[1983]. Relative
may also be preferred
Stiglitz
performance
as
rates
it lowers informational
rents to high types [Bhaskar
2002] and reduces
incentives
to exert effort in influence activity
of workers
[Milgrom 1988].
956
QUARTERLY JOURNAL OF ECONOMICS
Under
the two incentive schemes we observe in the data,
worker ?'s compensation
is webe?, where w is some constant, e is
the average effort of the group, and e? is ?'s own effort. Under
?
= 0. We
relative incentives 6 =
1, and under piece rates b
illustrate the effect of group incentives within this class of com
schemes by setting b > 0. While
this need not be the
pensation
it
makes
the
comparison with
optimal group incentive scheme,
more
the observed schemes
transparent.
Figure VI shows average effort under these three classes of
incentive schemes where individual pay and group performance
are negatively
uncorrelated
correlated
(relative
incentives),
correlated
(group incentives). We
(piece rates), and positively
that
derive effort levels under the three alternative assumptions
workers are self-interested (tt = 0), fully internalize the external
=
1), or have the average social weight derived in Section V
ity (it
=
w to hold the
(it
.65). Throughout, we adjust the parameter
that
bill constant, and, for simplicity, we assume
total wage
workers are of homogeneous
ability.38
The figure shows that the three types of incentive scheme
yield the same level of effort only in the case of pure self-interest
=
0). In line with the previous findings, when workers have
(it
social preferences
(it > 0), effort is higher under piece rates than
the figure also
under relative
incentives. More
interestingly,
shows that group incentives would, in this context, lead to higher
The estimates
effort at the same total cost to the principal.
increase by 30
effort
would
indicate that if u =
.65, average
a
to
scheme
where indi
rates
from
group
piece
percent moving
in
the
vidual pay increases
average
productivity of the
linearly
=
1).
group (b
The intuition for this is that since workers
place positive
on
of effort is
the
benefit
other
workers'
pay,
marginal
weight
higher when effort benefits
To the extent that workers
their coworkers, other things equal.
the positive externality
internalize
benefit from pay x is 4>(x) =
that workers'
the assumptions
38. We maintain
to its
6
is set equal
The
effort
is
of
the
and
parameter
disutility
8ef/2.
px1/p,
is set to 40. Worker's
value under piece rates, and N
estimated
compen
average
=
sation is kept constant at c
Sterling. The
4.5, the average hourly pay in Pounds
effort level as a function of b then is
Nash
equilibrium
+ b(l/N) + Tib((N- 1)/N))\y2
.
e = fcVp(l
{---j
the effect of b?the
Note
mance?on
effort depends
individual
between
relationship
on the sign and the magnitude
pay and group perfor
tt.
of the social weight
SOCIAL PREFERENCES,
RESPONSE
957
TO INCENTIVES
SocialWeight(it)
-1
1
0
-0.5
0.5
2 1.5
b=-1 Piece rates:b=0
Relativeincentives:
Grouppay:b>0
Figure
Incentive
Estimated
Effort under
the Three
Relative
b =
Effort,
estimated
VI
and Effort, Holding
Schemes
Incentive
incentives:
-1
Total Wage
when
Schemes
Piece
6 =
Bill
rates:
0
Constant
it =
Group
.65
incentives:
b = 1
from
data
productivity
1.67
Effort, calibrated
1.66
2.82
2.86
in Figure V are calculated
estimates
Kernel
using
density
to be
benefit function is assumed
kernel. The underlying
= 2x1/2.
<p(?)
3.68
an Epanechnikov
to be quadratic
in effort. Figure VI shows the
The total cost of effort is assumed
b in
Nash
efforts for it = 0, .65, and 1, as a function of the parameter
equilibrium
the pay schedule:
=
pay
webei.
w is adjusted
to hold constant the total wage bill, and
the parameter
Throughout,
N = 40. We assume
that workers have the same social weight
(either 0, .65, or 1),
0 parameter
is set equal
to its average
and are of homogeneous
ability. The
value under piece rates.
estimated
they impose on others, there is then a rationale for group incen
tives even in settings where the production technology does not
exhibit complementarities.39
are
39. Rotemberg
conditions
under which
[1994] derives
group incentives
are altruistic.
ifworkers
Sen [1966] analyzes
the allocation
rule that leads
optimal
on each
to Pareto
efficiency in a cooperative whose workers
place a positive weight
shows that the optimal
other's material
benefits. He
rule is a combination
of
and Dickson's
individual
and group rewards. Roethlisberger
[1939] results from
QUARTERLY JOURNAL OF ECONOMICS
958
In conclusion, our analysis
that understanding
emphasizes
worker preferences
is key for the optimal choice between alter
native incentive schemes. Clearly,
the magnitude
of the effects
live
might be particularly
large in our context because workers
and work together and, thus, have both solid social ties and
access
to a variety of punishment
mechanisms.
The findings
nevertheless
show that to the extent that workers place some
weight, either positive or negative, on the effect of their actions on
the other workers' pay, group or relative incentive schemes can
outperform piece rates in terms of productivity. The findings,
of incen
thus, provide specific insights for further developments
tive theory and shed new light on an old idea?the
interplay
between
social effects and the provision
of incentives within
firms.40
Appendix
1: Quality
and Quantity
We present evidence to see whether the change in incentives
affected the quality of picking. To do so, we exploit the fact that
as
pickers are expected to classify fruit as either class 1?suitable
as market
produce, or class 2?suitable
supermarket
produce.
While picking, each worker is expected to put class 1 and class 2
1 fruit is the most
fruit into two separate
containers.
Class
for
of
the
85
total
common, accounting
percent
weight of fruit
on
an
relative
and 87 per
under
incentives
average day
picked
cent
under
piece
rates.
to a cooled
it is transported
After fruit has been picked,
In the packhouse
for packing.
each container passes
warehouse
a class 2 fruit is detected in a
through a quality check. Whenever
transferred
class 1 container, it is removed?downgraded?and
to a class 2 container. By the time the fruit picked from a given
for inspection, however,
field-day arrives in the farm packhouse
misclassification
of fruit cannot be traced back to individual
workers. Moreover,
since the electronic system used to record
for
showed that productivity
increased
the Hawthorne
significantly
experiments
to socialize.
and were allowed
workers who were given group incentives
relations
affect workplace
40. The idea that human
goes back to
performance
and Dickson
[1939], and Roy [1952].
[1938], Roethlisberger
[1933], Barnard
Mayo
and Lazear
More
[1989], Rotemberg
[1994], and
[1992], Lazear
recently, Kandel
social
models
and Weiss
[2003] have developed
Fershtman,
Hvide,
incorporating
concerns
of behavior within firms. Fehr and Fischbacher
into the analysis
[2002]
on social preferences
in firms.
evidence
discuss
the experimental
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
959
individual productivity data is not the same used to record mis
at the field-day level in the packhouse,
classification
it is not
to
match
from
the
and
every field-day
possible
productivity
pack
house databases. We are, however, able to retrieve quality infor
mation for 67 field-day observations
of which 29 are under rela
tive incentives and 38 under piece rates.
In Appendix
2 we assess whether
the trade-off between the
of
and
quality
quantity
picking changed significantly with the
measure
in
incentives.
We
the quality of picking by the
change
2
of
class
fruit
that
is
quantity
wrongly classified as class 1, as a
of the total quantity
of class 2 picked on a given
scale
to
We
this
be
in percentage
measured
field-day.
points
(0-100).
On average, 15 percent of class 2 fruit ismisclassified
as class
1 under relative incentives, and 12 percent under piece rates. The
difference between the two schemes is not significant.
Since class 1 fruit on average
constitutes 85 percent of all
are a negligible
fruit picked, misclassifications
fraction of the
total kilograms of class 1 picked on a given day. On average, 2.32
percentage
percent
of the total class 1 fruit received by the packhouse
is
to class 2 under the relative scheme and 2.28 percent
downgraded
under
piece
rates.
In column
of the quality
of
(1) we regress this measure
picking on a dummy for the introduction of piece rates. In line
with the unconditional
results we find that the share ofmisclas
sified fruit falls by 3 percentage
points, but the effect is not
estimated.
precisely
Including field fixed effects, column
(2),
changes the sign of the piece rate coefficient, but the effect re
mains
of the effect is
statistically
insignificant. The magnitude
also quite small considering that on average there is five times as
much fruit of class 1 as there is of class 2. The estimates
then
fruit as a share of the total of class 1
imply that the misclassified
picked in a field-day increases by 0.4 percentage points.
In column (3) we additionally
control for the quantity of class
1 fruit picked on the field-day and then also for a time trend and
its square. Finally, column (4) adds controls for the field life cycle,
and meteorological
factors. We find that the level of misclassifi
cation of fruit picked increases over time, but at a decreasing rate.
None
of the other controls is significant. Our basic conclusion
remain unchanged?the
coefficient of the piece rate dummy is
always
small and not significantly
different from zero.
QUARTERLY JOURNAL OF ECONOMICS
960
The productivity gains achieved under piece rates were not at
the expense of a lower quality of picking. Combined with the fact
that worker pay remained constant over the season, the change in
made
the farm owners better off.
incentives unambiguously
2: The
Appendix
on the Quality
in Incentives
of the Change
Effect
of Picking
= kilograms
variable
dependent
of total
as a share
robust
of class
points
(percentage
errors
standard
rate dummy
Tons
of class
10~3
Time trend
-3.39
(Pt)
(3)
(2)
X
(5)
(4)
2.19
2.02
2.71
(4.20)
(3.86)
-
.016
(.010)
-
-
.010
(.010)
1.02**
(.438)
-.005**
(.002)
-5.61
life cycle
temperature
.190
Maximum
temperature
.831
.009
(.011)
.846*
(.463)
-.005**
(.002)
-7.49
(10.5)
Minimum
1
(3.00)
2.08
Time trend squared X 10"3
Field
class
field-day
0-100)
in parentheses
(3.78) (2.90)
2 fruit picked
2 on the
reported
(1)
Piece
as
2 fruit misclassified
of class
kilograms
(11.6)
(.584)
(.588)
Hours
?.813
of sunshine
(.834)
Field
R2
Number
Yes
Yes
Yes
Yes
.0142
.0929
.1125
.1702
67
67
No
fixed effects
of observations
67 67
.2156
67
(field-day)
*
**
*** denotes
significanceat 1 percent, at 5 percent, and at 10 percent.Robust standard errors are
calculated throughout.The piece rate dummy is set equal to zerowhen relative incentivesare in place, and
set equal to one when piece rates are in place. Data are based on the packhouse software system. It is
assumed that all fruitarrives in the packhouse twodays after it is picked. The sample is restrictedto those
fields that operated under both incentiveschemes.All right-hand-sidevariables are lagged by two days to
allow fora time lag between picking and packing. Temperature variables correspond to a 0900-0900 time
frame.Hours of sunshine are measured daily.
London
University
University
Economic
School
of Economics
and
Centre
for Economic
Policy
Research
of Essex
of Chicago
Policy
Graduate
Research
School
of Business
and
Centre
for
SOCIAL PREFERENCES,
RESPONSE
TO INCENTIVES
961
References
(New York: Basic Books,
1984).
Axelrod, Robert, The Evolution
of Cooperation
and Imran Rasul,
and Absolute
Iwan Barankay,
"Relative
Bandiera,
Oriana,
on Worker
Incentives:
Evidence
CEPR
Discussion
Productivity,"
Paper No.
4431, 2004.
Chester
MA: Harvard
Barnard,
I., The Functions
of the Executive
(Cambridge,
1938).
Press,
University
Brian E., and Mark A. Huselid,
Effects of Tournament
"The Incentive
Becker,
XXXVII
Science
(1992),
Compensation
Systems," Administrative
Quarterly,
336-350.
Performance
and Limited
Evaluation
Bhaskar,
V., "Relative
mimeo,
Liability,"
of Essex,
1992.
University
"Tournaments
and Piece Rates:
Bull, Clive, Andrew
Schotter, and Keith Weigelt,
An Experimental
XCV
Study," Journal
of Political
(1987), 1-33.
Economy,
and Richard
"Labor
Colin, Linda Babcock,
Camerer,
Thaler,
George Loewenstein,
One Day at a Time," Quarterly
Journal
Supply of New York City Cabdrivers:
CXII
(1997), 407-441.
of Economics,
Ronald
L. Bognanno,
"The Incentive Effects of Tour
G., and Michael
Ehrenberg,
naments Revisited:
Evidence
from the European
PGA Tour," Industrial
Labor
Relations
XLIII
Review,
(1990), 74-89.
"Is
Tomorrow
Another
The
of
Labor
New
York
F?rber, Henry S.,
Day?
Supply
City
Cab Drivers,"
Journal
CXIII
of Political
(2005), 46-82.
Economy,
and
Urs
Social
Matter?The
Preferences
Fehr, Ernst,
Fischbacher,
Impact
"Why
on Competition,
of Non-Selfish
Motives
and Incentives,"
Eco
Cooperation
nomic Journal,
CXII
C1-C33.
(2002),
and Simon G?chter,
"Fairness
and Retaliation:
of
The Economics
Fehr, Ernst,
Journal
XIV
159-181.
Reciprocity,"
of Economic
(2000a),
Perspectives,
and Punishment
in Public Goods
Fehr, Ernst, and Simon G?chter,
"Cooperation
Economic
American
Review, XC (2000b), 980-994.
Experiments,"
and Klaus
M. Schmidt,
"A Theory
of Fairness,
and
Fehr, Ernst,
Competition,
Journal
CXIV
Cooperation,"
Quarterly
of Economics,
(1999), 817-868.
and Yoram Weiss,
Cultural
Status
Fershtman,
Chaim, Hans
Hvide,
Diversity,
Concerns
and the Organization
ofWork,
CEPR
Discussion
Paper No. 3982,
2003.
R. Rosenzweig,
"A Test for Moral
Hazard
in the
Foster, Andrew
D., and Mark
Labor Market:
and Calorie
Review
Effort, Health
of Econom
Consumption,"
ics and Statistics,
LXXVI
(1994), 213-227.
Richard
"The Last American
Shoe Manu
Freeman,
B., and Morris M. Kleiner,
facturers:
the Method
to Survive
of Pay
Changing
Foreign
Competition,"
NBER Working
Paper No. 6750, 1998.
L. Stokey,
"A Comparison
of Tournaments
and
Green,
Jerry R., and Nancy
Journal
XCI
Contracts,"
of Political
(1983), 349-364.
Economy,
Barton H., Jack A. Nickerson,
and Hideo Owan,
"Team Incentives
and
Hamilton,
Worker
An Empirical
on
of the Impact
of Teams
Heterogeneity:
Analysis
and Participation,"
Journal
CXI
Productivity
of Political
(2003),
Economy,
465-497.
P. Lazear,
and Edward
"Peer Pressure
and Partnerships,"
Jour
Kandel,
Eugene,
nal of Political
C (1992), 801-813.
Economy,
Charles
Walter
N.
and
the
of Tourna
Knoeber,
R.,
Thurman,
"Testing
Theory
ments: An Empirical
of Broiler Production,"
Journal
Analysis
of Labor Eco
XII
155-179.
nomics,
(1994),
and Mohamed
Salah Matoussi,
"Moral Hazard,
Financial
Laffont, Jean-Jacques,
Constraints
and Sharecropping
in El Oulja,"
Review
of Economic
Studies,
LXII
381-399.
(1995),
Edward
and
Industrial
Journal
Political
Lazear,
P., "Pay Equality
Politics,"
of
LXXXVII
Economy,
(1989), 1261-1284.
"Performance
and
American
-,
Economic
Pay
Review, XC (2000),
Productivity,"
1346-1361.
Edward
as Optimum
"Rank Order Tournaments
Lazear,
P., and Sherwin Rosen,
Labor Contracts,"
Journal
LXXXIX
of Political
Economy,
(1981), 841-864.
962
QUARTERLY JOURNAL OF ECONOMICS
John O., "Public Goods: A Survey of Experimental
in John H.
Research,"
Ledyard,
and Alvin E. Roth,
Economics
eds., The Handbook
Kagel
of Experimental
(Princeton, NJ: Princeton University
Press,
1995), pp. 111-194.
David
"The Evolution
of Cooperation
Levine,
K., and Wolfgang
Pesendorfer,
at Los Angeles,
2002.
of California
Imitation," mimeo, University
Through
the Market:
Meets
Social Preferences
List, John, "The Behavioralist
Measuring
and Reputation
Effects in Actual
ofMary
mimeo, University
Transactions,"
land, 2004.
The Human
Problems
Civilization
(New York:
Elton,
Mayo,
of an Industrial
1933).
Macmillan,
Paul
Influence Activities,
and Efficient
R., "Employment
Contracts,
Milgrom,
Journal
XCVT
(1988), 42-60.
Organization
Design,"
of Political
Economy,
a
E. Stiglitz,
"Prizes
and Incentives:
Toward
Nalebuff,
Barry J., and Joseph
and Competition,"
General
Bell Journal
Theory of Compensation
of Econom
ics, XIV
(1983), 21-43.
Gerald
of the Daily
Labor
of Sta
S., "An Empirical
Oettinger,
Analysis
Supply
dium Vendors,"
Journal
CVII
(1999), 360-392.
of Political
Economy,
"Fixed Wages,
Piece Rates,
and Intertem
Paarsch,
Harry J., and Bruce Shearer,
in British
A Study of Tree Planters
mimeo
Columbia,"
poral Productivity:
Universit?
1996.
Laval,
"The Provision
of Incentives
in Firms," Journal
Canice,
of Economic
Prendergast,
XXXVII
Literature,
(1999), 7-63.
into Game Theory," American
Fairness
Economic
Rabin, Mathew,
"Incorporating
LXXXIII
Review,
(1993), 1281-1320.
J. Dickson, Management
and the Worker
Fritz, J., and William
Roethlisberger,
MA: Harvard
1939).
Press,
University
(Cambridge,
in the Workplace,"
Journal
Julio J., "Human
Relations
Rotemberg,
of Political
CII (1994), 684-717.
Economy,
a
in
and Goldbricking
Machine
"Quota Restriction
Roy, Donald,
Shop," American
LVII
Journal
(1952), 427-442.
of Sociology,
a
in
Allocation
Review
"Labour
K.,
Sen, Amartya
Cooperative
Enterprise,"
of
XXXIII
Economic
Studies,
(1966), 361-371.
Evolution
and
E.
"Preference
and
Sethi, Rajiv,
Somanathan,
mimeo,
Reciprocity,"
1999.
ofMichigan,
University
Fixed
from a
and
Incentives:
Evidence
Bruce
"Piece
Shearer,
Rates,
S.,
Wages
VXXI
Review
Field Experiment,"
Studies,
(2004), 513-534.
of Economic
in
A. O'Reilly,
and Diversity
Katherine
Y., and Charles
Williams,
"Demography
A Review
in Organiza
of 40 Years
of Research,"
Research
Organizations:
XX (1998), 77-140.
tional Behavior,