Content uploaded by Lorenz Goette
Author content
All content in this area was uploaded by Lorenz Goette
Content may be subject to copyright.
IZA DP No. 1002
Do Workers Work More When Wages Are High?
Evidence from a Randomized Field Experiment
Ernst Fehr
Lorenz Götte
DISCUSSION PAPER SERIES
Forschungsinstitut
zur Zukunft der Arbeit
Institute for the Study
of Labor
Januar 2004
Do Workers Work More When Wages
Are High? Evidence from a
Randomized Field Experiment
Ernst Fehr
University of Zurich, CEPR
and IZA Bonn
Lorenz Götte
University of Zurich, CEPR
and IZA Bonn
Discussion Paper No. 1002
January 2004
(revised October 2005)
IZA
P.O. Box 7240
53072 Bonn
Germany
Phone: +49-228-3894-0
Fax: +49-228-3894-180
Email: iza@iza.org
This paper can be downloaded without charge at:
http://ssrn.com/abstract=500908
An index to IZA Discussion Papers is located at:
http://www.iza.org/publications/dps/
Any opinions expressed here are those of the author(s) and not those of the institute. Research
disseminated by IZA may include views on policy, but the institute itself takes no institutional policy
positions.
The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center
and a place of communication between science, politics and business. IZA is an independent nonprofit
company supported by Deutsche Post World Net. The center is associated with the University of Bonn
and offers a stimulating research environment through its research networks, research support, and
visitors and doctoral programs. IZA engages in (i) original and internationally competitive research in
all fields of labor economics, (ii) development of policy concepts, and (iii) dissemination of research
results and concepts to the interested public.
IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion.
Citation of such a paper should account for its provisional character. A revised version may be
available on the IZA website (www.iza.org) or directly from the author.
IZA Discussion Paper No. 1002
January 2004
ABSTRACT
Do Workers Work More When Wages Are High?
Evidence from a Randomized Field Experiment∗
Most previous studies on intertemporal labor supply found very small or insignificant
substitution effects. It is not clear, however, whether these results are due to institutional
constraints on workers’ labor supply choices or whether the behavioral assumptions of the
standard life cycle model with time separable preferences are empirically invalid. We
conducted a randomized field experiment in a setting in which workers were free to choose
their working times and their efforts during working time. We document a large positive wage
elasticity of overall labor supply and an even larger wage elasticity of labor hours, which
implies that the wage elasticity of effort per hour is negative. While the standard life cycle
model cannot explain the negative effort elasticity, we show that a modified neoclassical
model with preference spillovers across periods and a model with reference dependent, loss
averse preferences are consistent with the evidence. With the help of a further experiment
we can show that only loss averse individuals exhibit a significantly negative effort response
to the wage increase and that the degree of loss aversion predicts the size of the negative
effort response.
JEL Classification: J22, C93, B49
Keywords: labor supply, intertemporal substitution, loss aversion
Corresponding author:
Lorenz Götte
Institut für Empirische Wirtschaftsforschung
University of Zurich
Blümlisalpstrasse 10
8006 Zürich
Switzerland
Email: lorenz@iew.unizh.ch
∗ The authors acknowledge support from the Swiss National Science Foundation under project number
1214-051000.97. This paper greatly benefited from the comments of David Card and two excellent
referees. In addition, we thank George Akerlof, Henry Farber, David Huffman, Reto Jegen, Rafael
Lalive, George Loewenstein, Jennifer Lerner, Stephan Meier, Matt Rabin, Jason Riis, Alois Stutzer,
Richard Thaler, and George Wu for their helpful comments.
1
The intertemporal substitution of labor supply has far-reaching implications for the interpretation
of important phenomena. If, for example, the intertemporal substitution of labor supply is high,
one may interpret the large variations in employment during business cycles as voluntary choices
by the workers rather than involuntary layoffs. Intertemporal substitution also plays a crucial role
in the propagation of shocks across periods (Romer, 1996; King and Rebelo, 1999). Previous
studies have found little evidence for intertemporal substitution of labor, however; the estimated
elasticities are often small and statistically insignificant, and sometimes even negative (see, e.g.,
Mankiw, Rotemberg and Summers 1985; Pencavel 1986; Altonji 1986; Blundell, 1994; Card
1994 and Blundell and MaCurdy 1999).1
However, the low estimates of intertemporal substitution are difficult to interpret because
of serious limitations in the available data. The life cycle model of labor supply predicts
intertemporal substitution with regard to transitory wage changes or wage changes the workers
anticipate. Yet, the typical wage changes are not transitory; hence they are associated with
significant income effects. In addition, it seems almost impossible to infer reliably from existing
data whether the workers anticipated the wage change. Furthermore, serious endogeneity
problems arise, as both supply and demand conditions determine wages.2 Thus, the typically
available data require many auxiliary assumptions when testing the life cycle model of labor
supply.
Another issue arises if labor markets are characterized by a significant amount of job
rationing or other constraints on workers’ labor supply. In fact, there is strong evidence
suggesting that workers are not free to set their working hours (Ham 1982, Kahn and Lang 1991,
Dickens and Lundberg 1993), rendering the identification of the source of small intertemporal
substitution effects difficult, even if the above mentioned problems could be solved. A small
intertemporal substitution effect could be due to these constraints or it could be that the
behavioral assumptions behind the life cycle model are wrong. Indeed, Camerer, Loewenstein,
Babcock and Thaler (1997) put forward the view that New York City cab drivers’ daily labor
supply is driven by nonstandard, reference dependent, preferences that exhibit loss aversion
1 After reviewing a sizeable part of the literature, Card (1994) concludes, for instance, that the “very small magnitude
of the estimated intertemporal substitution elasticities” can only account for a tiny fraction of the large person-
specific year-to-year changes in labor supply.
2 Oettinger (1999) shows that if one neglects the endogeneity of wage changes, estimates of labor supply elasticities
are severely downward-biased.
2
around a target income level. This view has recently been called into question by Farber (2004,
2005).
In this paper, we use an ideal data set to study workers’ responses to transitory wage
changes. We conducted a randomized field experiment at a bicycle messenger service in Zürich,
Switzerland. The bicycle messengers receive no base wage that is independent of effort, and are
paid solely on commission. We have precise information for all the workers on the number of
shifts they work, and the revenues they generate per shift. A shift always comprises five hours
and workers in our sample worked at most one shift per day. A key feature of our experiment is
the implementation of an exogenous and transitory increase in the commission rate by 25 percent.
Therefore, we can be sure that unobserved supply or demand variations did not induce the change
in the commission rate (i.e., the “wage” change). Each participant in the experiment knew ex-ante
the precise duration and size of the wage increase. Since the wage was only increased during four
weeks, its impact on the workers’ lifetime wealth is negligible.
In the firm under study, the messengers can freely choose how many shifts (hours) they
work, and how much effort they exert (to generate revenues). This means that our setting also
provides an ideal environment for studying the behavioral foundations of labor supply: in our
context, the absence of intertemporal substitution effects cannot be attributed to institutional
constraints on labor supply. The exogenous change in the commission rate raises the returns from
both the number of shifts and effort per shift. Therefore, we have the unique opportunity of
studying how hours and effort respond to the wage increase and how overall labor supply (i.e.,
the number of hours times the effort per hour) is affected.
Our experimental results show that the wage increase caused a large increase in overall
labor supply. Our estimate of the intertemporal elasticity of substitution with regard to overall
labor supply is between 1.12 and 1.25. This large effect is exclusively driven by the increase in
the number of hours worked. In fact, the elasticity of hours worked with regard to the wage is
even higher than the elasticity of overall labor supply; it lies between 1.34 and 1.50, considerably
in excess of that found in previous studies (e.g., Oettinger 1999). The fact that the elasticity of
hours worked is larger than the overall labor supply elasticity suggests that the effort per hour
decreased in response to the wage increase. And indeed, a detailed analysis indicates that effort
3
per shift decreased by roughly 6% in response to the wage increase, which implies a wage
elasticity of effort per shift of – 0.24.
These results confirm the non-experimental evidence in previous studies of intertemporal
labor substitution based on samples where workers were largely unconstrained in choosing hours
and effort. Camerer et al. (1997) and Chou (2002) examined how cab drivers, after having
decided to work on a given day, vary their daily working time (which is a good proxy for daily
effort) in response to wage variations. Both studies report that workers work fewer hours (provide
less effort) on high-wage days, indicating a negative effort elasticity. Oettinger (1999)
investigated how stadium vendors adjust their probability of working in response to transitory
wage variations across different baseball games. He develops a good set of ex-ante predictors of
game attendance, which can be used to instrument for the wage: temperature, day of the week,
the ranking of the home team, the quality of the opposing team, etc. In his IV estimates, Oettinger
finds a positive and significant wage elasticity of participation, ranging from 0.53 to 0.64. The
data in all these studies, however, do not allow making inferences on the overall labor supply
elasticity: the data by Camerer et al. (1997) and Chou (2002) do not reflect the participation
decision and Oettinger’s data do not allow reliable inferences about effort per game.
When viewed through the lens of a standard neoclassical model with separable time
preferences, the reduction in effort seems puzzling. After all, the rise in the commission rate
provides strong economic incentives for working more hours and for working harder during those
hours. Our results are, however, immune against the criticisms that have recently been raised
against the study of Camerer et al. (Farber 2004, 2005). One problem is that the source of the
variation behind cab drivers’ wages is not completely clear. If, for example, there are common
supply side shocks (e.g. most drivers don’t work on the 4th of July), then the supply of cabdriver
hours will be small on these days and the wage will be high. As a result, there will be a negative
correlation between wages and hours that has nothing to do with loss averse preferences. This
criticism does not apply to our study because we vary workers’ wages experimentally. A second
concern is a possible selection effect: higher wages may induce cab drivers to work a few hours
on days when they otherwise would not have worked. Such an effect may generate a negative
correlation between daily wages and daily hours even though all individuals behave exactly as
the standard model with time separable preferences predicts. Our data enable us to solve this
problem as well.
4
In addition, we would like to point out that a reasonable extension of the standard model
can, in principle, explain a negative effort elasticity. In the theory part of our paper, we show that
a neoclassical model, in which last period’s effort raises this period’s marginal disutility of effort,
is consistent with our evidence: workers who work in more periods may rationally decide to
reduce effort per period. However, we also show that a rational choice model, with reference
dependent preferences exhibiting loss aversion around the reference point, is also able to explain
the evidence. The intuition behind this model is that workers with loss averse preferences have a
daily reference income level.3 Daily incomes below the reference level are experienced as a
“loss”, and the marginal utility of income is large in the loss domain. In contrast, the marginal
utility of income at and above the reference level decreases discontinuously to a lower level.
Workers who temporarily earn higher wages are more likely to exceed the reference income
level, hence reducing their marginal utility of income and ultimately inducing them to provide
less effort. At the same time, however, workers with higher wages have a higher overall utility
from working a shift so that they can more easily cover the fixed costs of getting to work. Hence,
they are more likely to work.
There are thus two competing theories which are consistent with the facts. In order to
discriminate between the two theories, we conducted another experiment based on the idea that
loss aversion is a personality trait which affects behavior across several domains (Kahneman and
Tversky 2000; Gaechter, Hermann and Johnson 2005). In this experiment, we measured the
individual worker’s loss aversion in lottery choices. We then used these measures to examine
whether the negative response of effort per shift is due to the existence of loss averse workers.
We indeed find that the degree of a worker’s loss aversion contributes significantly to the
negative effort elasticity. Moreover, it turns out that workers who do not show loss aversion in
the lottery choices also do not have a significantly negative elasticity. Only workers with loss
aversion reduce effort per shift significantly when paid a high wage.
Thus, the result of our second experiment favors the model with reference dependent
preferences over the neoclassical model with “disutility spillovers” across periods. Of course, the
evidence from the second experiment is not the ultimate arbitrator, but it suggests that future
work should not disregard the loss aversion model because it could contribute to a deeper
3 Heath, Larrick and Wu (1999) provide evidence that goals often serve the function of a reference point.
5
understanding of effort choices. At the same time, we should also point out that one third of the
workers in our sample did not exhibit loss aversion and a negative effort elasticity. Thus future
work should take the possibility of heterogeneous preferences more seriously. In addition, the
results of our first experiment unambiguously show that whatever behavioral forces worked
against the intertemporal substitution of labor, they were apparently not capable of generating a
negative elasticity of the overall labor supply. The behavioral forces that worked in favor of
intertemporal substitution far outweighed any opposing forces.
The remainder of this paper is structured as follows. Section I describes the institutional
environment and the details of the field experiment. Section II discusses the implications of
different models of labor supply. Section III reports the results from the field experiment. We
first report the impact of the wage increase on overall labor supply and then discuss how shifts
responded; we finally present the evidence on how the wage increase affected the effort per shift.
This section also describes the follow-up experiment and discusses the link between individual
loss aversion and workers effort responses. Section IV concludes the paper.
I. Experimental Set-up
Our study is based on the delivery records of two relatively large Swiss messenger services –
Veloblitz and Flash Delivery Services (henceforth “Flash”) – which are located in Zurich. Each
firm employs between 50 and 60 bicycle messengers. The available records contain information
about when a messenger worked a shift, all deliveries he conducted during a shift, and the price
of each delivery. Thus, we know which messengers worked a shift and how much revenue they
generated during the shift for each day in the observation period. We first describe the
organization of work at a bicycle messenger service below and then present our experiment in
more detail.
A. Work at a Messenger Service
Unless pointed out below explicitly, the arrangements are the same for the two messenger
services, Veloblitz and Flash. When a potential worker applies for a job with one of the
messenger services, an experienced messenger evaluates him or her with respect to fitness,
knowledge of locations, names of streets, courtesy, and skills regarding handling the CB radio.
6
Once accepted as an employee, messengers can freely choose how many five-hour shifts they
will work during a week. There are about 30 shifts available at Veloblitz, and about 22 at Flash
on each workday from Monday to Friday. In principle, messengers could work more than one
shift per day, but none of them chose to do so during the experiment and the months prior to the
experiment. The shifts are displayed on a shift plan for every calendar week at the messenger
service’s office. There are two types of shifts, called ''fixed'' and “variable”. A “variable” shift
simply means that a shift is vacant at a particular time. Any messenger can sign up to work that
shift, e.g., on Wednesday from 8 am to 1 pm. If a messenger commits to a ''fixed'' shift, he has to
work that shift every week. For example, if a messenger chooses Wednesday, 8 am – 1 pm as a
fixed shift, he will have to fill that shift on every Wednesday for at least six months. Thus, fixed
shifts represent a commitment for several months and can only be cancelled with at least four
weeks notice. Roughly two-thirds of the shifts are fixed. It is also import to note for our
examination that the number and the allocation of fixed shifts across messengers remained the
same during the whole experiment; the company refused to change the fixed shifts just because of
the experiment. All shifts that are not fixed are variable shifts; they are available for any
messenger to sign up for. All workers participating in our study worked both fixed and variable
shifts.
Two further items are worth mentioning. First, there is no minimum number of shifts that
the messengers have to work at either messenger service. Second, both messenger services found
filling the available shifts difficult. There is almost always at least one unfilled shift and, on
average, almost 3 shifts per day remain unfilled. For example, during the period before the
experiment, from September 1999 – August 2000, approximately 60 shifts remained unfilled
every month. This implies that messengers are unlikely to be rationed in the choice of shifts.
Messengers receive no fixed wage. Their earnings are given solely as a fixed percentage w
of their daily revenues. Hence, if a messenger carries out deliveries that generate revenues r
during his shift, his earnings on that day will be wr. An important feature of the work
environment concerns the fact that messengers have substantial discretion on how much effort to
provide during a shift. They only stay in contact with the dispatcher at the messenger service's
office through CB radio. In order to allocate a delivery, say, from location A to location B, the
dispatcher will contact the messenger whom he thinks is closest to A to pick up the delivery. All
messengers can listen in on the radio. If they believe that they are closer to A than the messenger
7
who was originally contacted, they can get back to the dispatcher and say so and will then be
allocated to that delivery. Conversely, if the messenger does not want to carry out the delivery
from A to B, he may just not respond to the call. Messengers have, therefore, several means of
increasing the number of deliveries they complete. They can ride at higher speed, follow the radio
more actively, or find the shortest possible ways to carry out a delivery.
Thus, work at a bicycle messenger service closely approximates a model where individuals
are unconstrained in choosing how many shifts (hours) to work, and how hard to work (i.e., how
many deliveries to complete during a shift).
B. The Experimental Design
In order to evaluate the labor supply effect of a temporary wage increase, we randomly assigned
those Veloblitz messengers who were willing to participate in the experiment to a treatment and a
control group and we implemented a fully anticipated temporary increase in the commission rate
by (roughly) 25% for the treatment group. The commission rate for men in the treatment group
was temporarily increased from w = 0.39 to w = 0.49 and the rate for women was temporarily
increased from w = 0.44 to w = 0.54. The additional earnings for the messengers were financed
by the Swiss National Science Foundation.
In order to participate in the experiment, all messengers had to complete a questionnaire at
the beginning and at the end of each experimental period. The messengers were informed that a
failure to complete all questionnaires meant that they would not receive the additional earnings
from the experiment. All messengers who finished the first questionnaire also filled in the
remaining questionnaires.4 Thus, the group of messengers who participated in the experiment was
constant during the whole experiment, i.e., there was no attrition. Randomization into a treatment
and a control group was achieved by randomly allocating the participating messengers into a
group A and a group B. The randomization was based on the administrative codes that the
messenger service uses to identify a messenger in its accounting system. All messengers at
Veloblitz are assigned a number depending on the date when they started working for the
company. The first messenger who worked at Veloblitz was assigned the number 1, the second 2,
4 The messengers at Veloblitz who did not participate in the experiment were almost exclusively workers who were
already quite detached from the company or who where on probationary shifts. The “detached” workers typically
worked roughly one shift per week during the experiment and the months prior to the experiment.
8
and so forth. The participating messengers with odd numbers were assigned to group A,
participating messengers with even numbers to group B.
The messengers did not know that the purpose of the experiment was the study of labor
supply behavior, nor did they realize that we received the full (anonymous) records of each
messenger about the number of shifts and the number of deliveries completed. If pressed, we told
the participants that we wanted to study the relation between wages and job satisfaction. The
purpose of our study was credible because the questionnaires contained several questions related
to job satisfaction.5
We implemented a 25 percent increase in the commission rate during four weeks in
September 2000 for group A. The messengers in group B were paid their normal commission rate
during this time period so that they can be used as a control group. In contrast, only the
individuals in group B received a 25 percent increase in the commission rate during four weeks in
November 2000, while the members of group A received their normal commission rate and
therefore served as a control group. Thus, a key feature of our experiment is that there are two
experimental periods that lasted for four weeks and that both group A and B served as a treatment
and a control group in one of the two experimental periods. This characteristic of our experiment
enables us to provide a very clean isolation of the impact of the temporary wage increase. If, for
example, the implemented wage change increases labor supply, then we should observe this
increase both in the first and the second experimental period. In the first experimental period, the
members of group A (who receive the higher wage in this period) should exhibit a larger labor
supply than the members of group B while the reverse should be true in the second experimental
period – members of group B (who receive the higher wage in this period) should supply more
labor. Moreover, since members of both groups are in the treatment and the control group we can
identify the treatment effect within subjects by controlling for individual fixed effects.
Our experimental design also enables us to control for the income effect of the wage
increase, i.e. we can identify the pure substitution effect for the participating messengers. We
announced the experiment in the last week of August 2000, and all additional earnings from the
5 These features of the experiment ensure that our results cannot be affected by the Hawthorne effect. The Hawthorne
effect means that subjects behave differently just because they know that the experimenters observe their behavior.
Yet, our subjects did not know that we could observe their behavior during the wage increase. Moreover, since both
the treatment group and the control group are part of the overall experiment, and since our key results rely on the
comparison between these groups we control for a potential Hawthorne effect.
9
experiment – regardless of whether subjects were members of group A or group B – were paid
out after the end of the second experimental period in December 2000.6 Thus, the budget
constraint for both groups of participating messengers was affected in the same way. Due to the
randomization of the participating messengers into groups A and B, the income effect cancels out
if we identify the treatment effect by comparing the labor supply of control and treatment group.
As demand for delivery services varies from day to day and from month to month, it is
useful to control for time effects. The available information about Flash enables us to identify
possible time effects across treatment periods because both Veloblitz and Flash operate in the
same market. There is a strong correlation between the total daily revenues at Veloblitz and
Flash. When we compute the raw correlation between total revenues at the two firms over the two
experimental periods plus the four weeks prior to the experiment, we find a correlation of 0.56
(Breusch-Pagan
χ
2(1) =18.93, p < 0.01, N = 60 days). Even after removing daily effects from
both series, the correlation is still 0.46 (Breusch-Pagan
χ
2(1) =13.16 , p < 0.01, N = 60 days).
This shows that the revenues at the two firms are highly correlated, even over quite short a time
horizon.7
We believe that our experiment represents a useful innovation to the existing literature for
several reasons. First, it implements a fully anticipated, temporary and exogenous variation in the
(output based) wage rates of the messengers, which is key for studying the intertemporal
substitution of labor. The experimental wage increase was large and provides a clear incentive for
increasing labor supply. Moreover, the participating messengers are experienced, and daily
fluctuations in their earnings are common. Hence, we experimentally implement a wage change
in an otherwise familiar environment. Second, the data we obtained from Veloblitz allows us to
study two dimensions of labor supply: Hours as measured by the number of shifts and effort as
measured by the revenues generated per shift or the number of deliveries per shift. No other study
6 In the time period between the announcement of the experiment and the beginning of the first treatment period no
new regular workers arrived at Veloblitz. Only workers who worked on probationary shifts arrived during this time
period and they were not allowed to participate in the experiment because they often leave the firm after a short time
and lack the necessary skills. Including them in the experiment would have created the risk of attrition bias.
7 If we add the 8 months prior to the experiment we find a correlation of about 0.75. To check the robustness of our
results we also include – in some of our regressions – the non-participating messengers at Veloblitz in the non-
experimental comparison group that is used to identify time effects.
10
that we are aware of can look at these two dimensions simultaneously. Third, we can combine the
data set with the full records from a second messenger service operating in the same market. This
will prove useful for investigating any effect that the experiment might have had on the non-
participating messengers at Veloblitz, and helps to control for demand variations over time.
II. Predictions
In this subsection, we derive predictions about labor supply behavior in our experiment. We use
two types of models –neoclassical models and a model of reference dependent utility with loss
averse workers. In view of our results, we are particularly interested in the question of which kind
of model is capable of predicting an increase in shifts (hours) worked and a decrease in effort per
shift.
A. Neoclassical Model with Time-Separable Utility
In this subsection, we integrate the institutional setting at our messenger service into a canonical
model of intertemporal utility maximization with time-separable utility. We define the relevant
time period to be one day. Consider an individual who maximizes lifetime utility
Uo = ∑=
T
t0
δ
t u(ct,et,xt) (1)
where
δ
< 1 denotes the discount factor, u( ) represents the one-period utility function, ct denotes
consumption, et is effort in period t and xt denotes a variable that affects the preference for
working on particular days. For example, a student who works a few shifts per week at Veloblitz
may have higher opportunity costs for working on Fridays because he attends important lectures
on Fridays. The utility function obeys uc > 0, ue < 0 and is strictly concave in ct and et. The
lifetime budget constraint for the individual is given by
∑=
T
t0t
p
ˆct(1+r)-t = ∑=
T
t0(t
w
ˆet + yt) (1+r)-t (2)
where t
p
ˆ denotes the price of the consumption good, t
w
ˆ the period t wage per unit of et and yt
non-labor income. For convenience we assume that the interest rate r is constant and that there is
no uncertainty regarding the time path of prices and wages. The sign of the comparative static
predictions is not affected by these simplifying assumptions.
11
In appendix A, we show that along the optimal path, the within-period decisions of a
rational individual maximizing a time-separable concave utility function like (1), subject to
constraint (2), can be equivalently represented in terms of the maximization of a static one-period
utility function that is linear in income.8 This static utility function can be written as
v(et, xt) =
λ
wtet – g(et,xt), (3)
where g(et,xt) is strictly convex in et, and measures the discounted disutility of effort and xt
captures exogenous shifts in the disutility of effort.
λ
measures the marginal utility of life-time
wealth and wt represents the discounted wage in period t. Thus,
λ
wtet can be interpreted as the
discounted utility of income arising from effort in period t.9
Workers who choose effort according to (3) respond to an anticipated temporary increase in
wt with a higher effort et. A rise in wt increases the marginal utility returns of effort,
λ
wt, which
increases the effort level *
t
e that maximizes v(et, xt). The situation is, a bit more complicated in
our experiment, however, because the messengers can choose the number of shifts and the effort
during the shift. Theoretically the existence of shifts can be captured by the existence of a
minimal effort level e
~
that has to be met by the worker or by the existence of fixed costs of
working a shift. Intuitively, if there is a fixed cost of working a shift, an employee will only work
on a given day if the utility of *
t
e, v(*
t
e,xt), is higher than the utility of not going to work at all. As
a wage increase raises v(*
t
e,xt) workers are more likely to work on a given day, i.e., the number
of shifts worked will increase. 10
8 Our characterization is inspired by the results in Browning, Deaton and Irish (1985) who show that the within
period decisions can be characterized in terms of the maximization of a static profit function. However, the present
exposition is more convenient for our purposes.
9
λ
is constant along the optimal path of ct and et. This has the important consequence that an anticipated temporary
wage variation does not affect the marginal utility of life-time wealth. Thus, anticipated temporary variations in
wages (or prices) have no income effects. Yet, if there is a non-anticipated temporary increase in the wage,
λ
changes
immediately after the new information about the wage increase becomes available and remains constant at this
changed level afterwards. For our experiment, this means that the income effect stemming from the temporary wage
increase has to occur immediately after the announcement of the experiment in August 2000. Thereafter, the
marginal utility of life-time wealth again remains constant so that there are no further changes in
λ
during the
experiment. The difference in behavior between the treatment group and the control group during the two treatments
can thus not be due to changes in
λ
. Note also that (3) does not only describe the optimal effort choice in period t but
is also based on the optimal consumption decision in period t. For any change in effort, the consumption decision
also changes in an optimal manner (see appendix).
10 More formally, the wage increase raises the utility of going to work for all x. Hence the participation condition will
be met for more states x.
12
B. Neoclassical Model with Non-Separable Utility
The prediction of the previous subsection is, however, not robust to the introduction of non-
separable utility functions. To illustrate this, consider a simple example where
v(et,et−1)=
λ
etw
−
g(et(1
+
α
et−1)). (4)
This example captures the intuition that if a worker worked yesterday, he has higher marginal
cost of effort today. We assume, for simplicity, that e0
=
0, that there are only two further time
periods, period 1 and period 2 and that the wage is constant across time. If we ignore discounting
the two-period utility is given by U = 1
(,0)ve + 21
(,)ve e . Therefore, if the wage is high enough to
induce the worker to go to work in both periods the worker chooses effort e1
** and e2
** according
to
λ
w=g'(e1)
+
α
e2g'(e2(1
+
α
e1)) (5)
λ
w=g'(e2(1
+
α
e1))(1
+
α
e1). (6)
If work is supplied in both periods, an increase in e1 causes a higher disutility of labor in period 2
which lowers e2. Of course, rational workers take this effect into account when they decide on e1
which means that the overall marginal disutility of e1 is higher if e2 is positive compared to when
it is zero. In particular, if wages are low enough so that it is no longer worthwhile to work in
period 2 (e2 = 0), the first order conditions are given by
λ
w = g’(e1) (5’)
λ
w < g’(0)(1 + αe1). (6’)
A comparison of conditions (5) and (6) with conditions (5’) and (6’) shows that it is possible that
the optimal effort e1 according to (5’) is higher than e1
** and e2
** . In appendix B we provide an
explicit example that proves this point. This possibility arises because the marginal disutility of
working in each of the two periods, which is indicated by the right hand side of (5) and (6), is
higher than the marginal disutility of working only in period 1 which is given by g’(e1). In the
13
context of our experiment, this means that messengers who work more shifts when the wage is
high may rationally decide to reduce the effort per shift.
The simple model above does not predict that workers who work more shifts (days) will
necessarily reduce their effort per shift. It only allows for this possibility. If the wage increase is
large enough, it is also possible that workers who behave according this model raise their effort
per shift. There is, however, one prediction that follows unambiguously from a neoclassical
approach regardless of whether utility is time separable or not. Browning, Deaton, and Irish
(1985) have shown that a general neoclassical model predicts that overall labor supply et
∑
increases in the high wage periods in response to a temporary increase in wages. Applied to our
context, this means that during the four-week period where the wage is higher for the treatment
group, the total revenue (or the total number of deliveries) of the treatment group should exceed
the total revenue (or the total number of deliveries) of the control group.
C. Reference Dependent Utility
Another potential explanation for why effort per shift might decrease in response to a temporary
wage increase is that individuals could have preferences that include a daily income target y
~
that
serves as a reference point. The crucial element in this approach is that if a person falls short of
his or her target, he or she is assumed to experience an additional psychological cost which is not
present if income varies above the reference point. This explanation is suggested by the large
number of studies indicating reference dependent behavior (for a selection of papers on this see
Kahneman and Tversky 2000). Evidence from psychology (Heath, Larrick, and Wu, 1999)
suggests that the marginal utility of a dollar below the target is strictly higher than the marginal
utility of a dollar above the target.11 A daily income target seems plausible for bike messengers in
our sample because their daily incomes are a salient feature of their work environment. The
11 See Goette and Huffman (2003) for survey evidence on this point. They present bike messengers with direct
survey scenarios to elicit whether messengers care more about making money in the afternoon if they had good luck
in the morning than after a bad morning. In their scenarios good luck means that messengers had the opportunity of
making particularly profitable deliveries in the morning. For example, good luck means that a delivery just crosses
an additional district boundary; such deliveries command a substantially higher price without much additional effort.
About 70% of the messengers respond in a fashion consistent with daily income targeting.
14
messengers keep receipts from each delivery they did on a shift. This makes them acutely aware
of how much money they earn from each completed delivery. The messengers also turn in the
receipts at the end of the shift, making it difficult for them to keep track of how much money they
earned over several shifts. A daily income target may also serve the messengers as a commitment
device for the provision of effort during the shift. Zurich is rather hilly and riding up the hills
several times during a shift requires quite some effort – in particular if the weather is bad or
towards the end of a shift. A daily income target may thus help the messengers overcome a
natural tendency to “shirk” that arises from a high marginal disutility of effort.
Formally, the existence of reference dependent behavior can be captured by the following
one-period utility function.
() ()
(
)
()()
yewxegyew
yewxegyew
ev
tttttt
tttttt
t~
if,
~
~
if,
~
<−−
≥−−
=
γλ
λ
(7)
where γ > 1 measures the degree of loss aversion, i.e., the increase in the marginal utility of
income if the individual is below the income target. Previous evidence (see Kahneman and
Tversky 2000) suggests that γ ≈ 2 for many individuals. Loss aversion at this level creates
powerful incentives to exert more effort below the income target. However, once individuals
attain the target y
~
, the marginal utility of income drops discretely (from
γλ
to
λ
), causing a
substantial reduction in the incentive to supply effort.
The preferences described in (7) imply that workers increase the number of shifts when
they are temporarily paid a higher wage: a rise in wages increases the utility of working on a
given day. Thus, at higher wages it is more likely that the utility of working v(et) exceeds the
fixed costs of working. At the same time, however, the increase in wages makes it more likely
that the income target is already met or exceeded at relatively low levels of effort. Therefore,
compared to the control group, the workers in the treatment group are more likely to face a
situation where the marginal utility of income is
λ
instead of
γλ
, i.e., they face lower incentives to
15
work during the shift.12 As a consequence, members of the treatment group will provide less
effort than members of the control group.
The previous discussion shows that reference dependent preferences and a neoclassical
model with non-separable preferences may make similar predictions. In particular, both models
are consistent with a reduction in effort per shift during the wage increase. However, the
reduction in effort in the income target model should be related to the degree of loss aversion
γ
,
as explained above. Evidence suggests that there is substantial heterogeneity in the degree of loss
aversion between individuals, and that individuals who are loss averse in one type of decisions
are also loss averse in other domains of life (see Gaechter, Herrmann, and Johnson, 2005). Thus,
in principle, the two explanations can be distinguished if one obtains an individual level measure
of
γ
.
III. Results
This section reports the results from our field experiment. Our analysis is based on the four weeks
prior to the first experimental period and the two subsequent experimental periods in which first
group A and then group B received a wage increase. The data contain the day of each delivery,
the messenger’s identification number, and the price for each delivery. Thus, we have, in
principle, two measures of labor supply – the amount of revenue generated and the number of
deliveries completed. Since longer deliveries command a higher price and require more effort, the
revenue is our preferred measure of labor supply. However, our estimates of the treatment effect
are almost identical for either choice of the labor supply measure.
A. The Impact of the Wage Increase on Total Revenue per Messenger
The first important question is whether there is a treatment effect on total revenue per messenger
during the first and the second experimental period. Figure 1a and 1b as well as Tables I and II
12 If
γ
is sufficiently high relative to the wage increase, one may obtain the extreme result that the worker provides
effort to obtain exactly y
~ before and after the increase. In this case, the worker’s effort obviously decreases in
response to the wage increase because at higher wages y
~ is obtained at lower effort levels. In general, the larger
γ
,
the sharper the kink in the objective function and the more likely the worker’s optimal effort choice e* will be at the
kink, i.e., the more likely
γλ
t
w
ˆ> g’(e*) >
λ
t
w
ˆ holds. Note, however, that even if the worker is not a “perfect”
income targeter, i.e., even if before or after the wage increase he does not earn exactly y
~, negative effort responses
may occur.
16
present the relevant data. The figures depict the revenue data for groups A and B at Veloblitz; the
tables show also the data of messengers at Flash and those messengers at Veloblitz who did not
participate in the experiment. Figure 1a and Table I show the “raw” revenue per messenger –
uncontrolled for individual fixed effects. Figure 1b and Table II control for individual fixed
effects by showing how – on average – the messengers’ revenues deviate from their person-
specific mean revenues. Thus, a positive number here indicates a positive deviation from the
person-specific mean, a negative number indicates a negative deviation.
These figures and tables show that group A and B generate very similar revenues per
messenger during the four weeks prior to the experiment. If we control for individual fixed
effects, we find that the revenues per messenger are almost identical across groups and close to
zero. For example, the difference in revenues between group A and B is only CHF 48.9 if we
control for person specific effects with a standard error of CHF 366.6 (see Figure 1b and Table
II). This difference is negligible in comparison to the average revenue of roughly CHF 3400 that
was generated by a messenger during the pre-experimental period. Thus, in the absence of an
experimental treatment the messengers in group A and B behave in the same way.
However during the first experimental period (henceforth, “treatment period 1”), in which
group A received the higher wage, the total revenue generated by group A is much larger than the
revenue of group B, indicating a large treatment effect. On average, messengers in group A
generated roughly CHF 4131 while messengers in group B only generated revenues of CHF 3006
during this period (see Table I and Figure 1a). Then, in the second experimental period, this
pattern is reversed and group B, which receives now the higher wage, generates much more
revenue. In treatment period 2 group B generates revenues per messenger of CHF 3676 while
group A only produces revenues of CHF 2734. If we control for individual fixed effects (see
Figure 1b) we can see that the standard errors are relatively small, suggesting that the differences
across groups are significant.
Insert Figures 1a and 1b about here
To control more tightly for statistical differences across groups, we performed regressions
(1) – (4) in Table III. All regressions are of the form
17
rit =
α
i +
δ
Tit + dt + eit (8)
where rit measures the revenue generated by messenger i during a four-week period t,
α
i is a
fixed effect for messenger i, Tit is a dummy variable that is equal to 1 if the messenger is on the
increased commission rate, dt is a time dummy estimated for treatment period 1 and for
treatment period 2, and eit is the error term.
Regression (1) is based only on the data of groups A and B at Veloblitz. Due to the random
assignment of the participating messengers across groups and due to the fact that both groups
served once as a control and once as a treatment group, this regression allows for a clean isolation
of the treatment effect. The regression indicates that the treatment effect is highly significant and
that the messengers on a high wage generate roughly CHF 1000 additional revenue compared to
the experimental control group.
The other three regressions show that the measured impact of the experimental wage
increase on the treated group remains almost the same if we include in the comparison group
messengers of Flash and non-participants of Veloblitz. Regression (2) compares the treatment
group at Veloblitz with all other messengers at Veloblitz and finds again a large and significant
treatment effect of roughly CHF 1000. Regression (3) uses observations from all messengers at
Veloblitz and the messengers at Flash. The inclusion of the messengers at Flash is suggested by
the strong correlation in revenues between Flash and Veloblitz. Regression (3) also includes a
dummy for the whole non-treated group at Veloblitz, i.e. the messengers in the control group and
those who did not participate in the experiment. Therefore, this dummy measures whether the
non-treated group at Veloblitz behaved differently relative to the messengers at Flash and the
treatment dummy measures whether the treated group at Veloblitz behaved differently relative to
the messengers at Flash. In this regression, the coefficient of the treatment dummy indicates
again a treatment effect of roughly CHF 1000. In addition, the dummy for the whole non-treated
group at Veloblitz is small and insignificant, indicating that the non-treated group was not
affected by the wage increase for the treated group. This result suggests that the wage increase for
the treated group did not constrain the opportunities for working for the non-treated group at
Veloblitz. The result is also consistent with the permanent existence of unfilled shifts and with
18
survey evidence; the overwhelming majority of the messengers stated that they could work the
number of shifts they wanted to work. Finally, regression (4) uses the data from all Veloblitz and
Flash messengers but does not include the dummy for the whole nontreated group at Veloblitz.
Therefore, the treatment dummy in this regression measures whether the treated group at
Veloblitz generates a different revenue from all the other messengers at Veloblitz and Flash.
Again, the treatment effect is of similar size and significance as in the previous regressions.13
In summary, the above results indicate a large and highly significant effect of a temporary
wage increase on the total effort of the treated group. In contrast to many previous studies, our
results imply a large intertemporal elasticity of substitution. The standard way to calculate this
elasticity is to estimate (8) in logs. However, because some participants of the experiment did not
work at all during a four-week period (because they went on vacation) we cannot use this
method. If we did, we would have to drop these observations although the decision to not work
during a whole four week period also represents a labor supply decision. For this reason we
include all participating messengers in our measure and compare the percentage increase in the
revenue per messenger (which is our proxy for overall labor supply per messenger) that is due to
the wage increase with the 25% increase in wages. We have seen that the treatment effect is
roughly CHF 1000. The average revenue across group A and B is CHF 3568 in treatment period
1; in treatment period 2 it is 3205. Thus, the intertemporal elasticity of substitution is between
(1000/3568)/0.25 = 1.12 and (1000/3205)/0.25 = 1.25, which is substantially larger compared to
what previous studies have found (see e.g. Oettinger 1999).14
13 It is also noteworthy that we find a negative effect of time on revenues per messenger in all four regressions.
However, while the time effect is never significant for the first treatment period, it is higher for the second treatment
period and reaches significance at the 5% level in some of the regressions. These time effects suggest that a
comparison of the revenues of the same group over time is problematic because revenue is likely to be “polluted” by
monthly variations in demand. It is thus not possible to identify the treatment effect by comparing how a group
behaved in treatment period 1 relative to treatment period 2.
14 It is even possible that our measure of the elasticity of labor supply with regard to a temporary wage increase
underestimates the true elasticity because we use revenues per messenger as a proxy for labor supply per messenger.
If wages w affect effort e and effort affects revenue r the elasticity of e with respect to w, which we denote by
ε
ew, is
given by
ε
rw/
ε
re where
ε
rw is the elasticity of r with respect to w (which is observable to us) and
ε
re is the elasticity of
r with respect to e (which is not observable to us). Thus, our measure
ε
rw implicitly assumes that the elasticity
ε
re is
equal to one. If
ε
re is less than one our measure even underestimates the true labor supply elasticity.
ε
re is less than
one if the production function r = f(e) is strictly concave and f(0) = 0 holds.
19
B. The Impact of the Wage Increase on Shifts worked
After we documented the strong impact of the wage increase on total labor supply, the natural
question is whether both the number of shifts and the effort per shift increased. In this section we
examine, therefore, the impact of the wage increase on the number of shifts worked while in the
next section we have a closer look at effort per shift. Figure 2a and 2b provide a first indication of
how the wage increase affected shifts. Figure 2a shows the absolute number of shifts per worker
in group A and B during the four-week period prior to the experiment and the two treatment
periods. Figure 2b controls for person specific effects by showing the average deviation of the
number of shifts from the person specific means.
Figure 2a shows that group A worked roughly 12 shifts in the pre-experimental period and
group B worked roughly 11 shifts (see also Table I for the precise numbers). However, the
standard errors are very large due to large differences between the workers, suggesting
insignificant differences across groups. If we control for person specific effects (see Figure 2b)
we find that the average deviation from person specific means is very small in both groups and
close to zero during the pre-experimental period. Table IV, which shows the concrete numbers
relating to Figure 2b, indicates that in group A the deviation from person specific means is 0.22
(with a standard error of 1.29) and in group B it is -0.35 (with a standard error of 0.98). Thus,
there are almost no differences in shifts across groups before the experiment.
During the first treatment period, however, the messengers in group A, who are paid the
high wage, worked almost 4 shifts more than did the messengers in group B. Likewise, in the
second treatment period the messengers in group B, who receive now the high wage, work
substantially more shifts than the messengers in group B. Moreover, if we control for person
specific effects (see Figure 2b and Table IV), the standard errors become very small, suggesting
that the differences across groups are significant.
Insert Figures 2a and 2b about here
To test more rigorously for significant differences, we performed regressions (5) – (8) in
Table III. The independent variable in these regressions is sit, the number of shifts that messenger
20
i worked during the four week period t. The right hand side of these regressions is the same as in
equation (8), i.e., we included a treatment dummy, individual fixed effects and time dummies for
treatment periods 1 and 2. Regression (5) estimates the impact of the treatment by using only data
from group A and group B. It shows a large and highly significant treatment effect; the treated
group works on average four shifts more than the control group. Regression (6) uses data from all
messengers at Veloblitz; the treatment dummy thus compares the treated with the whole group of
untreated messengers at Veloblitz. This regression basically replicates the results of regression
(5). In regression (7), we use data from all messengers at Veloblitz and at Flash. In addition, we
include a dummy variable that takes on a value of 1 if a messenger belongs to the whole non-
treated group at Veloblitz (which comprises the experimental control group and the messengers
who did not participate in the experiment). As in regression (3), this dummy measures whether
the experiment had an effect on the whole non-treated group at Veloblitz by comparing this group
with Flash messengers. The coefficient of this dummy is highly insignificant, suggesting that the
experiment had no effect on the non-treated group at Veloblitz. The treatment dummy in
regression (7) compares the treated group with the Flash messengers and again indicates a
significant treatment effect of similar size as in the previous regressions. Finally, regression (8)
compares the treated group to all untreated messengers at Veloblitz and Flash; again the
treatment effect is roughly 4 shifts per treatment period and significant.
In summary, Figure 2a and 2b, Table IV, and regression (5) – (8) indicate a clear positive
treatment effect of the wage increase on shifts. On average, workers supplied about four shifts
more if they receive a high commission rate. Since the average number of shifts worked during
the two treatment periods is 11.925 and 10.64, respectively, the wage elasticity of shifts is
between (4/11.925)/0.25 = 1.34 and (4/10.64)/0.25 = 1.50. Thus, the shift choices are even more
responsive to the wage increase than total revenue per messenger. By definition, the wage
elasticity of total revenue is equal to the elasticity of shifts plus the elasticity of the revenue per
shift. Therefore, the higher wage elasticity of shifts compared to the elasticity of total revenues is
a first indication that the elasticity of effort per shift is negative.
21
C. The Impact of the Wage Increase on Effort per Shift
When examining the revenue per shift, it is useful to restrict attention to behavior during fixed
shifts. Recall that the management at Veloblitz did not allow workers to change their fixed shifts
after the announcement of the experiment and during the experiment. The increase in the supply
of shifts is fully borne by the variable shifts. Therefore, our experiment could not induce any kind
of selection effect with regard to the fixed shifts and the revenue change during the fixed shifts
identifies the impact of the treatment on effort per shift. 15
In Figure 3a, we show the log of revenue per shift in group A and B during the four weeks
prior to the experiment and in the two treatment periods. We control for person effects in Figure
3b by showing the deviation from person-specific means. If we control for person-specific
effects, we find that both groups generated almost identical revenues per shift during the four
weeks prior to the experiment. However, group B, which receives the lower wage, generates
roughly 5 percent more revenue per shift than group A during the first treatment period.
Likewise, in the second treatment period, group A, which receives now the lower wage, exhibits
roughly a 6 percent higher revenue per shift than group B. Thus, Figures 3 suggests that the
wage increase caused a reduction in revenue per shift.
Insert Figure 3a and 3b about here
The impression conveyed by Figures 3 is further supported by the two regressions presented
in Table V, which are based on observations from group A and B during fixed shifts. The
dependent variable is log revenue of messenger i at day t. We include a treatment dummy in both
regressions that takes on a value of 1 if messenger i at day t is in the treatment group and we
further control for daily fixed effects and i’s tenure. Daily fixed effects are important because of
demand variations across days; tenure is important because experienced messengers usually have
higher productivity. We do not control for individual fixed effects in regression (1), but for a
messenger’s gender. This regression shows that the wage increase leads to a reduction in revenue
per shift by roughly 6 percent. We control for individual fixed effects in regression (2). The
15 We should, however, mention that the results remain the same when we examine revenue per shift over all (fixed
and variable) shifts.
22
treatment effect in this regression is again significantly negative and indicates a reduction in
revenues by roughly 6 percent.
Thus, the temporary wage increase indeed reduced revenue per shift. The implied wage
elasticity of revenue per shift is – 0.06/0.25 = – 0.24, which is consistent with our neoclassical
model with preference spillovers across periods and the target income model based on loss
aversion. It is also worthwhile to point out that this estimate neatly fills the gap between the
elasticity of total revenue and the elasticity of shifts. The intermediate value (between the lower
and the upper bound) of the elasticity of total revenue is 1.18; the intermediate value for the
elasticity of shifts is 1.42. Thus, according to this difference the elasticity of effort per shift
should be – 0.24. Our estimates in Table V precisely match this value.
D. Does Loss Aversion Explain the Negative Impact on Effort per Shift?
We provide additional evidence in this section that helps us understand the forces behind the
negative impact of the wage increase on effort per shift. Our strategy is to measure individual
level loss aversion and to examine whether these measures have predictive value with regard to
individuals’ response of effort per shift. In other words, we ask the question whether the loss
averse messengers drive the negative effect of the wage increase on effort per shift or whether the
messengers who are not loss averse drive this effect? If mainly the loss averse messengers show a
negative effort response, the loss aversion model is supported; if the negative effect on effort is
not related to individual’s loss aversion, the neoclassical model provides the more plausible
explanation.
Loss aversion and reference dependent behavior have implications in a variety of domains.
Loss averse choices have been documented, in particular, in the realm of decision-making under
uncertainty (Kahneman and Tversky 1979). Therefore, we measured the messengers’ loss
aversion by observing choices under uncertainty in an experiment that took place eight months
after the experimental wage increase. As part of this study, we presented the messengers with the
opportunity to participate in the following two lotteries:
Lottery A: Win CHF 8 with probability 1/2, lose CHF 5 with probability 1/2. If subjects reject
lottery A they receive CHF 0.
23
Lottery B: This lottery consists of six independent repetitions of Lottery A. If subjects reject
lottery B they receive CHF 0.
Subjects could participate in both lotteries, or only in one lottery, or they could reject both
lotteries.
The above lotteries enable us to construct individual measures of loss aversion. In
particular, the observed behavior in these lotteries enables us to classify subjects with regard to
their degree of loss aversion
γ
. If subjects’ preferences are given by (7), subjects who reject
lottery A have a higher level of
γ
than subjects who accept lottery A and subjects who reject
lottery A and B have a higher level of
γ
than subjects who reject only lottery A. In addition, if
subjects’ loss aversion is consistent across the two lotteries, then any individual who rejects
lottery B should also reject lottery A because a rejection of lottery B implies a higher level of loss
aversion than a rejection of only lottery A. We derive these implications of (7) explicitly in
appendix C.
Among the 42 messengers who belong either to group A or B, 19 messengers rejected both
lotteries, 8 messengers rejected only lottery A, 1 messenger rejected only lottery B and 14
messengers accepted both lotteries. Thus, with the exception of the one messenger who rejects
only lottery B, all messengers who rejected lottery B also rejected lottery A. These results can be
neatly captured by a simple loss averse utility function that obeys equation (7).16
In principle, one might think that the rejection of A and/or B is also compatible with risk
aversion arising from diminishing marginal utility of lifetime income. Rabin’s calibration
theorem (Rabin 2000) rules out this interpretation, however. Rabin showed that a theory of risk
averse behavior based on the assumption of diminishing marginal utility of life-time income
implies that people essentially must be risk neutral for low stake gambles like our lotteries.
Intuitively, this follows from the fact that risk averse behavior for low stake gambles implies
ridiculously high levels of risk aversion for slightly higher, but still moderate, stake levels. Yet,
such unreasonably high levels of risk aversion can be safely ruled out. For example, we show in
appendix D that if one assumes that the rejection of lottery A is driven by diminishing marginal
utility of life time income, then the subject will also reject a lottery where one can lose $32 with
16 These results are qualitatively similar to the results obtained in a many other studies (e.g., Read, Loewenstein, and
Rabin, 1999; Cubbit, Starmer and Sudgen, 1998; Hogarth and Einhorn, 1992; Keren and Wagenaar, 1987).
24
probability ½ and win any positive prize with probability ½. Thus, there is no finite prize that
induces this subject to accept a 50 percent chance of loosing $32. Similar results are implied by a
rejection of lottery B.
We illustrate the behavior of messengers with and without loss averse preferences in Figure
4. The figure controls for person specific effects by comparing individual log revenues to the
mean of the individual’s log revenues. We show that the messengers who did not display loss
averse preferences do not change their effort per shift across the treatment and the control period.
However, the messengers who displayed loss aversion in the lottery choices exhibit a lower effort
per shift in the treatment period compared to the control period. This pattern suggests that the
negative effect of wages on effort per shift may only be driven by the loss averse messengers.
Insert Figure 4 here
To examine this possibility in more depth, we conducted the regressions in Table VI. In
these regressions, log daily revenue of messenger i at day t is again the dependent variable and
we control for the messengers’ tenure and for daily fixed effects in all four regressions. In the
first two regressions, we generate a loss aversion dummy L that is based on the rejection of
lottery A. If a messenger rejects this lottery L = 1, if lottery A is accepted L = 0. In regression (1)
and (2) we estimate the treatment effect for the loss averse messengers (by interacting the
treatment dummy with L) and for the messengers who did not exhibit loss aversion (by
interacting the treatment dummy with (1-L)). Regression (1), which does not control for
individual fixed effects, shows that loss averse messengers generated a roughly 10 percent lower
revenue per shift when they received the high wage. In contrast, the treatment effect is much
lower and insignificant for the messengers without loss aversion. Regression (2), which controls
for individual fixed effects, shows the same pattern. There is no significant decrease in revenue
per shift for messengers without loss aversion, whereas the messengers with loss aversion exhibit
a significant 10 percent reduction in revenue per shift during the treatment period.
Regressions (3) and (4) provide a further robustness check for these results. We use a finer
scale to indicate a messenger’s loss aversion in these regressions. Here, we capture the absence of
loss aversion, which is indicated by the acceptance of both lotteries, by L’ = 0. If a messenger
25
rejects only one of the lotteries we assign L’ = 1 and if both lotteries are rejected we assign L’ = 2
to this messenger. The variable “treatment dummy × loss averse” is now defined as the
interaction between the treatment dummy and L’. Thus, the interaction term measures how the
degree of loss aversion affects the messengers’ effort responses, whereas the treatment dummy
alone measures the effort response of those who did not display loss aversion.17 Both regression
(3) and (4) indicate that the messengers without loss aversion did not show a significant effort
reduction in response to the wage increase. In contrast, the interaction term is relatively large and
significant; an increase in L’ by one integer unit decreases revenue per shift by roughly 8
percentage points. Thus, messengers who rejected both lotteries generated a 16 percentage point
lower revenue per shift when they received the high wage. This result suggests that the negative
impact of the wage increase on revenue per shift is associated with the messengers’ degree of loss
aversion which lends support to the target income model discussed in Section II.C.
V. Summary
This paper reports the results of a randomized field experiment examining how workers, who can
freely choose their working time and their effort during working time, respond to a fully
anticipated temporary wage increase. We find a strong positive impact of the wage increase on
total labor supply during the two four-week periods in which the experiment took place. The
associated intertemporal elasticity of substitution is between 1.12 and 1.25. The large increase in
total labor supply is exclusively driven by the increase in the number of shifts worked. On
average, messengers increase their working time during the four weeks in which they receive a
higher wage by four shifts (20 hours), which implies a wage elasticity of shifts between 1.34 and
1.50. This is a considerably larger elasticity than what has previously been found on the basis of
daily labor supply data (Camerer et al. 1997, Chou 2000, Oettinger 1999). We also find that the
wage increase causes a decrease in revenue (effort) per shift by roughly 6 percent. However, the
increase in the number of shifts dominates the negative impact on effort per shift by a large
margin such that overall labor supply strongly increases.
17 Thus, in regressions (1) and (2), the variable “treatment dummy × not loss averse” is constructed as “treatment
dummy × (1 – L) whereas the variable “treatment dummy × not loss averse” is given by the treatment dummy alone
in regressions (3) and (4). But the variable “treatment dummy × not loss averse” measures the effort response of the
messengers who did not display loss aversion in the lotteries in all four regressions.
26
The standard neoclassical model with separable intertemporal utility is not consistent with
the evidence because this model predicts that both the number of shifts and the effort per shift
increase in response to the wage increase. However, we show that a neoclassical model with
preference spillovers across periods as well as a target income model with loss averse preferences
is consistent with the observed decrease in effort per shift. In order to discriminate between these
two models, we measured the messengers’ loss aversion at the individual level in the domain of
choices under uncertainty. We use these measures to examine whether the negative impact of the
wage increase on effort per shift is mediated by the degree to which messengers’ are loss averse.
We find that the degree of loss aversion is indeed related to the response of effort per shift: higher
degrees of loss aversion are associated with a stronger negative impact of the wage increase on
effort per shift and workers who do not display loss aversion in choices under uncertainty also do
not show a significant effort reduction. Thus, it seems that loss aversion drives the negative effect
of wages on effort.
We believe that these results contribute to a deeper understanding of the behavioral
foundations of labor supply. Our results certainly do not rule out a role for “neoclassical”
preferences in labor supply decisions. One third of the workers in our sample did not exhibit loss
aversion and the large intertemporal substitution effects on overall labor supply and the supply of
shifts document the power of behavioral forces that have always been emphasized in the standard
life cycle model. Our results also contrast sharply with the small and insignificant substitution
effects that have been found in many previous studies. Therefore, the small effects in these
studies may reflect the constraints workers face in their labor supply decisions and – in view of
our results – may be less likely due to workers’ unwillingness to substitute labor hours over time.
However, our results on the behavioral sources of the negative wage elasticity of effort per shift
also suggest that disregarding reference dependent preferences in effort decisions is not wise.
References
Altonji, Joseph G. (1986), Intertemporal Substitution in Labor Supply: Evidence from Micro
Data, Journal of Political Economy, 94(3), Part 2, S176-S215.
Blundell, Richard (1994), Evaluating Structural Microeconometric Models of Labor Supply. In:
Christopher Sims (Ed.), Advances in Econometrics: Sixth World Congress of the
Econometric Society, Vol. II, 3 – 48.
Blundell, Richard; MaCurdy, Thomas E. (1999), Labor Supply: A Review of Alternative
Approaches, in Ashenfelter, Orley and David Card (eds.): Handbook of Labor Economics
vol. 5.
Browning, Martin, Angus Deaton and Margaret Irish (1985), A Profitable Approach to Labor
Supply and Commodity Demands over the Life-Cycle, Econometrica 53(3), 503-543.
Camerer, Colin, Linda Babcock, George Loewenstein and Richard Thaler (1997), Labor Supply
of New York City Cabdrivers: One Day at a Time, Quarterly Journal of Economics; 112(2),
407-41.
Card, David (1994), Intertemporal Labor Supply: An Assessment. In: Christopher Sims (Ed.),
Advances in Econometrics: Sixth World Congress of the Econometric Society, Vol. II, 49 –
78.
Chou, Yuan K. (2000), Testing Alternative Models of Labor Supply: Evidence from Taxi Drivers
in Singapore, Research Paper No. 768, Dept. of Economics, University of Melbourne.
Cubbit, Robin, Christ Starmer and Robert Sugden (1998), On the Validity of the Random Lottery
Incentive System, Experimental Economics 1(2), 115 – 131.
Dickens, William T. and Shelley J. Lundberg (1993), Hours Restriction and Labor Supply,
International Economic Review 34, 169-192.
Farber, Henry S. (2004), Reference-Dependent Preferences and Labor Supply: The Case of New
York City Taxi Drivers, Working Paper No. 497, Princeton University, Industrial Relations
Section.
Farber, Henry S. (2005), Is Tomorrow Another Day? The Labor Supply of New York City
Cabdrivers, Journal of Political Economy 113, 46-82.
Gaechter, Simon, Herrmann, Andreas and Eric J. Johnson (2005), Individual-level Loss Aversion
in Risky and Riskless Choice. Working Paper, University of Nottingham.
Goette, Lorenz and David Huffman (2003), Reference Dependent Preferences and the Allocation
of Effort over Time: Evidence from Natural Experiments, Working Paper, Institute for
Empirical Research in Economics, University of Zurich.
Heath, Chip, Larrick, Richard and George Wu (1999). Goals as Reference Points, Cognitive
Psychology 38, 79-109.
Hogarth, Robin and Hillel Einhorn (1990), Venture Theory: A Model of Decision Weights,
Management Science 36, 780 – 803.
Kahn, Shulamit and Kevin Lang (1991), The Effect of Hours Constraints on Labor Supply
Estimates, Review of Economics and Statistics 73, 605-611.
Kahneman, Daniel and Amos Tversky (1979), Prospect Theory, Econometrica 47, 263-291.
Kahneman, Daniel and Amos Tversky (2000) Eds., Choices, Values, and Frames. Cambridge
University Press.
Keren, Gideon and Willem Wagenaar (1987), Violation of Utility Theory in unique and repeated
Gambles, Journal of Experimental Psychology: Learning, Memory, & Cognition. Vol 13(3),
387-391.
King, Robert G., Rebelo, Sergio (1999), Resuscitating Real Business Cycles. in: Handbook of
Macroeconomics.
Mankiw, Gregory, Julio Rotemberg and Lawrence Summers (1985), Intertemporal Substitution in
Macroeconomics, Quarterly Journal of Economics, 100 February, 225 – 251.
Oettinger, Gerald S. (1999), An Empirical Analysis of the Daily Labor Supply of Stadium
Vendors, Journal of Political Economy; 107(2), 360-92.
Pencavel, John (1986), Labor Supply of Men: A Survey. in: Ashenfelter, Orley and Richard
Layard (Eds.), Handbook of Labor Economics, Amsterdam: North Holland.
Rabin, Matthew (2000), Risk Aversion and Expected Utility Theory: A Calibration Theorem,
Econometrica; 68(5), 1281-92.
Read, Daniel; Loewenstein, George; Rabin, Matthew (1999), Choice Bracketing, Journal of Risk
and Uncertainty; 19(1 3), 171-97.
Romer, David (1996), Advanced Macroeconomics, Upper Saddle River: McMillan.
Appendix A
In this appendix, we derive the quasi linear objective function in equation (3) of the paper from
the underlying intertemporal maximization problem. The intertemporal optimization problem is
1
ˆˆ
max ( , ; ) subject to ( )(1 ) 0
tttt tt t tt
Uucex weypcr
δ
−
=+−+=
∑∑
where u is strictly concave and twice differentiable in e and c, e is labor supply in period t, c is
consumption in period t, x is a taste shift variable to allow for periods without work, ˆt
w is the
wage, ˆt
p
is the price of consumption goods,
δ
is the discount rate, and r is the interest rate. We
assume that there are no liquidity constraints, and that the path of wages, prices, and the taste
shifter are known, and that the interest rate is constant.
The first order conditions to this problem are
ˆ
(,, ) (1 )
tt
cttt t
ucex r p
δ
λ
−
=+
ˆ
(,, ) (1 )
tt
ettt t
ucex r w
δ
λ
−
−=+
where
λ
is the Langrange multiplier on the life-time budget constraint. Thus, it can be
interpreted as the marginal utility of lifetime wealth. Define the discounted price as
ˆ
(1 ) tt
tt
p
rp
δ
−−
=+ and the discounted wage wt analogously. The first order conditions then have
the easily interpretable form
(,, )
cttt t
ucex p
λ
= (A1)
(,; )
ettt t
ucex w
λ
−=
(A2)
Equation (A1) implies that, at every date t, the individual equates the marginal utility of
consumption to the marginal utility of lifetime income
λ
times the discounted price of the
consumption good. Similarly, when choosing how hard to work, the individual chooses effort
such that the marginal disutility of effort is equal to the marginal utility of lifetime income times
the discounted wage per unit of effort t
w. The model also allows for non-participation. If
(,0, )
et t t
uc x w
λ
−<
it is optimal to choose e = 0.
It is possible to represent within-period preferences in terms of a static objective function.
This is essentially a reformulation of the results in Browing, Deaton, and Irish (1985). Consider
equation (A1) again. Since u(.) is strictly concave, uc is strictly decreasing in c. Thus, (A1) can be
solved for ct
1(,,)
tc ttt
cu pex
λ
−
= (A3)
Substitute this into (A2) to obtain
1
(( ,,),,)
ec ttt tt t
uu pex ex w
λ
λ
−
−=
(A4)
Now consider the static one-period objective function
() (, , )
tttttt
ve we ge p x
λ
λ
=− (A5)
where λ is the lifetime marginal utility of income along the optimal path. Next we show that
maximizing this static objective function is equivalent to solving the intertemporal maximization
problem, that g( ) is convex and can be interpreted as the monetary equivalent of the disutility of
effort. To see that this, define
1
0
(, , ) ( ( , , ),, )
t
e
t tt ec ttt tt
g
e px uu pex exdx
λλ
−
=−
∫ (A6)
From the construction of g( ) in (A6), it is obvious that the first order condition (FOC) that results
from the static one-period objective function is equivalent to the FOC (A4). To show that g( ) is
convex in e, we need to show that the second derivative w.r.t. e is positive. We proceed in two
steps: First, consider how the individual adjusts consumption to a small perturbation in labor
supply along the optimal path, i.e., λ remains constant. Differentiation of (A1) yields:
dct
det
=−uce
ucc .
Now, take the second derivative of g( ) to obtain
()
22
1
(, , ) tce
eet tt eece ee cceece
tcccc
dc u
g
epx u u u uu u
de u u
λ
−
=− − =− + = −
To determine the signs of the terms, observe that the conditions for concavity of u( ) are ucc < 0,
uee < 0 and uccuee – 2
ce
u > 0. But this establishes the convexity of g( ), as claimed. Thus, in the
canonical life-cycle model, a rational, forward looking individual behaves as if she maximized
the one-period objective function (A5).
Appendix B
In this appendix, we provide a specific example that shows how non-separable time preferences
can induce workers to increase the number of shifts but decrease the effort per shift in response to
a wage increase. We consider a two-period model in which the workers one-period objective
function is given by
11
(, ) (, )
tt t tt
ve e we
g
ee
λ
−−
=− .
We assume that if a worker does not work during a period she has a utility from leisure time of L0
and that the effort cost function g( ) is given by
g(et,et−1)=et(1 +
α
et−1)+0.5get
2.
If we ignore discounting and set e0 = 0, total utility is given by
121
(,0) (,)Uve vee=+ .
(a) If an individual works only one period, the first order condition for effort in this period is
∂U
∂et
=
λ
w−1−get=0⇔et
*(w)=
λ
w−1
g.
Substituting this into the utility function, we get the overall utility of working one shift
U(one shift)=
λ
w−1
(
)2
2g+L0.
(b) If an individual works two shifts, the first order conditions for effort in the two periods are
given by
∂U
∂e1
=
λ
w−1−ge1−
α
e2=0
∂U
∂e2
=
λ
w−1−ge2−
α
e1=0
The two first order conditions imply e1=e2. Therefore,
e1
**(w)=e2
**(w)=
λ
w−1
g+
α
.
Substituting this into the objective function, we get
U(two shifts) =
λ
w−1
(
)2
g+
α
We can now examine the implications of this model for the number of shifts worked and effort
exerted on a shift as a function of the wage w.
(i) Shifts: A rational individual works two shifts if U(two shifts) >U(one shift). This implies
λ
w−1
()
2
g+
α
>
λ
w−1
()
2
2g+L0 (B1)
Notice that, in this model, if
α
>
g
, it is never optimal to work two shifts. The condition
α
>
g
has a straightforward interpretation: In this case, yesterday's effort raises today's marginal costs of
effort by more than today's effort raises today's marginal costs of effort. Simplifying this
inequality, we get
λ
w−1>L0
2g(g+
α
)
g−
α
. (B2)
Denote the wage that satisfies (B2) with equality by w'. As intuition suggests, the individual's
participation is increasing in the wage: If w is large enough such that (B2) is satisfied, she will
work two shifts.
(ii) Effort: To examine how effort responds to a change in wages, we choose two wage levels
wL<wHand set wL=w' , i.e. the low wage is equal to the highest wage at which it is still
optimal to work only one shift. If the wage is low, the individual works one shift, and effort is
equal to
0
*
10
2( )
12()
() ()
L
gg
Lg
wg
ew L
gg gg
α
α
λ
α
α
+
−
−+
== = −
Effort on the high wage is equal to
e1
**(wH)=e2
**(wH)=
λ
wH−1
g+
α
.
(iii) The response to a change from wL to wH: In this example, *
1()
L
ew exceeds
** **
12
() ()
HH
ew ew= if ,
HL L
g
ww w
g
g
α
α
λ
⎛⎤
+
∈−
⎜
⎥
⎝⎦
. Thus, changing the wage from wLto wH may
decrease effort per shift if the wage increase is not too high. Notice also that the effect crucially
depends on L0, the value of leisure. If L0=0, the effect cannot occur, because the wage cancels
from the participation condition (A7). Then the individual always works the same number of
shifts, irrespective of the wage, and effort responds positively to the wage, irrespective of the
strength of the intertemporal spillover
α
. By continuity, this also holds for some L0>0. Thus,
in our example intertemporal spillovers alone can produce the described response of shifts and
effort to the wage only if the value of leisure is large enough.
Appendix C
In this appendix, we derive the conditions under which a loss averse individual whose
preferences obey (7) in the text will reject lotteries A and B. For the purpose of lottery choices
the disutility of effort does not matter so that we can simplify preferences to
v(x−r)=
λ
x−r
()
if x≥r
γλ
x−r
()
if x<r
⎧
⎨
⎩
where x is the lottery outcome, and r is the reference point. We take the reference point to be the
status quo. The individual will reject gamble A if
0.5v(−5)+0.5v(8) ≤v(0)
which simplifies to
0.5( 5 ) 0.5(8) 0
γ
λλ
−+ ≤
This condition is satisfied if
8
5
γ
≥.
The individual will reject gamble B if
1
64 v(−30) +6
64 v(−17) +15
64 v(−4)+20
64 v(9) +15
64 v(22)+6
64 v(35)+1
64 v(48) ≤v(0).
Plugging in our functional form and simplifying, we find that the individual will reject the
gamble if
793
192
γ
≥
As claimed in the text, the degree of loss aversion required to reject gamble B is greater than the
degree of loss aversion needed to reject A.
Appendix D
In this appendix we prove the following: If an individual has a concave utility function u( ) and
rejects a coin flip where she can either win CHF 8 or lose CHF 5 for all wealth levels [m,
∞
),
then she will reject any coin flip in which she could lose CHF 32 no matter how large the positive
prize that is associated with the coin flip.
Proof: We proceed in four steps
(i) We adopt the convention that, if indifferent, the individual rejects the coin flip. Rejecting the
coin flip implies
0.5u(m+8)+0.5u(m−5) ≤u(m)
⇔u(m+8)−u(m)≤u(m)−u(m−5).
It follows from concavity that 8 u(m+8)
−
u(m
+
7)
[]
≤
u(m
+
8)
−
u(m) and.
Define
M
U
(
x
)=u(
x
)
−
u(
x
−1) as the marginal utility of the xth dollar. Putting the last three
inequalities together, we obtain
MU(m+8) ≤5
8MU(m−5) (D1)
and, because of the premise, it is true that MU(x+12) ≤5
8MU(x)for all x > m - 4.
(ii) We now derive an upper bound on u(
∞
). The concavity of u( ) implies
u(m+12) ≤u(m)+12
M
U
(m).
Using the same logic,
u(m+24) ≤u(m)+12
M
U
(m)
+
12
M
U
(m
+
12)
≤u(m)+12MU(m)1+5
8
⎡
⎣
⎢⎤
⎦
⎥
u(m+36) ≤u(m)+12MU (m)1+5
8+5
8
⎛
⎝
⎜⎞
⎠
⎟
2
⎡
⎣
⎢
⎢
⎤
⎦
⎥
⎥
and so on. Thus, we can develop a geometric series starting from m. Taking the limit, we obtain
u(∞)≤u(m)+12MU 8
3=u(m)+32MU(m)
(iii) Concavity implies u(m−32) ≤u(m)
−
32
M
U
(m).
(iv) Using the results from step (ii) and (iii), we get an upper bound on the value of a coin flip
where the individual would either lose CHF 32 or win an infinite amount:
0.5u(m−32) +0.5u(∞)≤u(m)
This implies that the individual would reject the gamble. This concludes the proof.
Messengers,
Flash
Group A Group B
3500.67 3269.94 1461.70 1637.49
(2703.25) (2330.41) (1231.95) (1838.61)
Mean Shifts 12.14 10.95 5.19 6.76
(8.06) (7.58) (4.45) (6.11)
N21 19 21 59
4131.33 3005.75 844.21 1408.23
(
2669.21
)
(
2054.20
)
(
1189.53
)
(
1664.39
)
Mean Shifts 14.00 9.85 3.14 6.32
(
7.25
)
(
6.76
)
(
4.63
)
(
6.21
)
N22 20 21 65
2734.03 3675.57 851.23 921.58
(2571.58) (2109.19) (1150.31) (1076.47)
Mean Shifts 8.73 12.55 3.29 4.46
(7.61) (7.49) (4.15) (4.74)
N22 20 24 72
Notes a) Standard deviations in parenthesis
b) Group A received the high commission rate in Experimental
Period 1, group B in Experimental Period 2.
TABLE I: DESCRIPTIVE STATISTICS
Mean
Revenues
Four-Week
Period Prior
to Experiment
Treatment
Period 1
Treatment
Period 2
Participating Messengers Non-
Participating
Messengers,
Veloblitz
Mean
Revenues
Mean
Revenues
Messengers,
Flash
Group A Group B
-48.88 -119.91 456.72 305.08
(366.61) (302.61) (179.92) (131.42)
71.03
(475.37)
721.98 -277.95 -160.77 102.85
(192.90) (240.62) (173.89) (105.76)
999.93
(308.40)
-675.32 391.87 -258.95 -342.84
(288.62) (250.55) (137.61) (129.50)
1067.19
(382.20)
Notes a) Standard error of the means in parenthesis
b) Same number of observations as in Table I
c) Group A received the high commission rate in Experimental Period
1, group B in Experimental Period 2.
Treatment
Period 2
Difference:
Group A -
Group B
Difference:
Group B -
Group A
Mean
Revenues
Treatment
Period 1
TABLE II: REVENUES PER FOUR-WEEK PERIOD
AVERAGE DEVIATIONS FROM INDIVIDUAL MEANS
Difference:
Group A -
Group B
Participating Messengers Non-
Participating
Messengers,
Veloblitz
Mean
Revenues
Mean
Revenues
Four-Week
Period Prior
to Experiment
(1) (2) (3) (4) (5) (6) (7) (8)
Observations are
restricted to
Messengers
Participating
in Experiment
All
Messengers
at Veloblitz
All
Messengers
at Flash and
Veloblitz
All
Messengers
at Flash and
Veloblitz
Messengers
Participating
in Experiment
All
Messengers
at Veloblitz
All
Messengers
at Flash and
Veloblitz
All
Messengers
at Flash and
Veloblitz
Treatment Dummy 1033.6*** 1094.5*** 1035.8** 1076.2*** 3.99*** 4.08*** 3.44** 3.9***
(326.9) (297.8) (444.7) (290.6) (1.030) (0.942) (1.610) (0.930)
Dummy for Non- -54.4 -0.772
treated at Veloblitz (407.4) (1.520)
Treatment Period 1 -211 -370.6 -264.8 -290 -1.28 -1.57 -0.74 -1.01
(497.3) (334.1) (239.9) (200.6) (1.720) (1.210) (0.996) (0.781)
Treatment Period 2 -574.7 -656.2 -650.5** -675.9** -2.56 -2.63** -2.19** -2.51**
(545.7) (357.9) (284.9) (238.0) (1.860) (1.260) (1.090) (0.859)
Individual Fixed
Effects Yes Yes Yes Yes Yes Yes Yes Yes
R squared 0.74 0.786 0.753 0.753 0.694 0.74 0.695 0.695
N124 190 386 386 124 190 386 386
Notes:
b) ***, **, * indicate significance at the 1, 5, and 10 percent level, respectively.
Dependent Variable: Revenues per four-week period Dependent Variable: Shifts per four-week period
TABLE III: MAIN EXPERIMENTAL RESULTS
OLS REGRESSIONS
a) Robust standard errors, adjusted for clustering on messenger, are in parentheses
Messengers,
Flash
Group A Group B
0.22 -0.35 1.57 0.98
(1.29) (0.98) (0.75) (0.53)
0.57
(1.62)
2.53 -1.18 -0.48 0.52
(0.65) (0.79) (0.75) (0.42)
3.71
(1.02)
-2.74 1.52 -0.96 -1.27
(0.98) (0.77) (0.57) (0.45)
4.26
(1.24)
Notes a) Standard error of the means in parenthesis
b) Same number of observations as in Table I
c) Group A received the high commission rate in Experimental
Period 1, group B in Experimental Period 2.
Treatment
Period 1
TABLE IV: SHIFTS PER FOUR-WEEK PERIOD
AVERAGE DEVIATIONS FROM INDIVIDUAL MEANS
Difference:
Group A -
Group B
Participating Messengers Non-
Participating
Messengers,
Veloblitz
Mean
Revenues
Mean
Revenues
Four-Week
Period Prior to
Experiment
Treatment
Period 2
Difference:
Group A -
Group B
Difference:
Group B -
Group A
Mean
Revenues
(1) (2)
Treatment Dummy -0.0642** -0.0601**
(0.030) (0.030)
Gender (female = 1) -0.0545
(0.052)
Log(tenure) 0.105*** 0.015
(0.016) (0.062)
Day Fixed Effects Yes Yes
Individual Fixed Effects No Yes
R Squared 0.149 0.258
N1137 1137
TABLE V: THE IMPACT OF THE EXPERIMENT
ON LOG REVENUES PER SHIFT
DEPENDENT VARIABLE: LOG(REVENUES PER
SHIFT) DURING FIXED SHIFTS , OLS REGRESSIONS
Notes: a) Observations are taken from group A and group B
while working on fixed shifts.
b) Robust standard errors, adjusted for clustering on
messenger, are in parentheses
c) ***, **, * indicate significance at the 1, 5, and 10
percent level, respectively.
(1)(2)(3)(4)
-0.0374 -0.0273 -0.0353 -0.0369
(0.034) (0.033) (0.035) (0.033)
-0.0983** -0.105** -0.0827*** -0.0755**
(0.040) (0.046) (0.032) (0.034)
Gender (female = 1) -0.0485 -0.0457
(0.052) (0.052)
Log(tenure) 0.104*** 0.00152 0.102*** -0.00131
(0.016) (0.061) (0.017) (0.061)
Day Fixed Effects Yes Yes Yes Yes
Individual Fixed Effects No Yes No Yes
R Squared 0.14 0.243 0.14 0.243
N1137 1137 1137 1137
Loss Aversion Measure 1 Loss Aversion Measure 2
TABLE VI: DOES LOSS AVERSION EXPLAIN THE REDUCTION IN
EFFORT PER SHIFT?
DEPENDENT VARIABLE: LOG(REVENUES PER SHIFT) DURING FIXED
SHIFTS , OLS REGRESSIONS
Treatment Effect ×
no
t
loss a
v
erse
Treatment Effect ×
loss a
v
erse
Notes: a) Robust standard errors, adjusted for clustering on messenger, are in parentheses.
b) ***, **, * indicate significance at the 1, 5, and 10 percent level, respectively.
c) Loss aversion measure 1 (denoted by L): L=1 if subject rejects lottery A, L=0 if
subject accepts lottery A.
d) Loss aversion measure 2 (denoted by L'): L'=2 if subject rejects both lotteries,
L'=1 if subject rejects one of the lotteries, L'=0 if subject accepts both lotteries.
Figure 1: Revenue per four week period
(a) Amount of Revenue
2000
2500
3000
3500
4000
4500
Period before
Experiment Treatment
Period 1 Treatment
Period 2
CHF per four-week period
Group B
Group A
(b) Deviations of Revenues from Individual
Means
-1500
-1000
-500
0
500
1000
Period before
Experiment Treatment
Period 1 Treatment
Period 2
CHF per four-week period
Group B
Group A
Figure 2: Shifts per four week period
(a) Number of Shifts
7
9
11
13
15
Period before
Experiment Treatment
Period 1 Treatment
Period 2
Shifts per four-week period
Group B
Group A
(b) Deviations of Shifts from Individual
Means
-4
-2
0
2
4
Period before
Experiment Treatment
Period 1 Treatment
Period 2
Shifts per four-week period
Deviation from individual means
Group B
Group A
Figure 3: Log of Daily Revenues on Fixed Shifts
(a) Log of Daily Revenues
5.6
5.65
5.7
5.75
5.8
Period before
Experiment Treatment
Period 1 Treatment
Period 2
log(daily revenues)
Group B
Group A
(b) Deviation of Log (Daily Revenues) from
Individual Means of Log (Daily Revenues)
-0.1
-0.05
0
0.05
0.1
Period before
Experiment Treatment
Period 1 Treatment
Period 2
log(daily revenues) - person mean o
f
log(daily revenues)
Group B
Group A
Figure 4: The Behavior of Loss Averse and Not Loss Averse Subjects during Control
and Treatment Period in Fixed Shifts
Log of Daily Revenues
Deviations from Individual Means
-0.1
-0.05
0
0.05
0.1
Control Period Treatment Period
log(daily revenues) - person mean of
log(daily revenues)
Not Loss Averse
Loss Averse
