team incentives and performance: evidence from a … team incentives and performance: evidence from...
TRANSCRIPT
1
Team incentives and performance: Evidence from a retail chain1
November 2014
Guido Friebel, Matthias Heinz, Miriam Krüger, Nick Zubanov
Abstract: There is substantial field evidence that incentive pay increases the performance
of workers when individual performance is measurable. Comparable evidence for teams,
however, is scarce. We fill this gap by a randomized experiment on team incentives in a
retail chain of roughly 200 shops and 1200 employees. It is technologically impossible to
measure individual performance, but the firm measures team (shop) performance along
various dimensions. Using stratified randomization, we introduced a team bonus
conditioned on sales targets fixed well before the team incentive was discussed. Treated
shops increase their sales on average by around three percent, wages increase by around
two percent on average (and up to 13 percent). The team incentive works best for (i) shops
in larger towns and cities where arguably the marginal productivity of effort is highest; (ii)
shops with younger employees, for whom the marginal costs of effort is likely to be
lowest, and (iii) shops that did not reach their targets regularly before the introduction of
the bonus, for whom the effect of effort on the marginal probability of success is likely to
be largest.
Keywords: randomized experiment, controlled trial (RCT), natural field experiment, team
bonus, insider econometrics
JEL codes: D23, J33, M52
1All authors are at Goethe University Frankfurt, except for Heinz who is affiliated with the University
of Cologne. Friebel is also affiliated with CEPR and IZA, and Zubanov with IZA. We are grateful for
the support of Deutsche Forschungsgemeinschaft (DFG). We would like to thank for their comments:
Iwan Barankay, Oriana Bandiera, Nick Bloom, Johan Lagerlöf, John List, and participants in seminars
at Bergen, Cologne, Columbia, Copenhagen, Frankfurt, Maastricht, the European Bank for
Reconstruction and Development, London, a conference organized by the university of Arhus, a
workshop organized by LMU Munich, the Annual GEABA conference and the NBER Organizational
Economics Working Group Meeting 2014. We would also like to praise the team spirit of our partners
in the retail chain, and of Artur Anschukov, Sandra Fakiner, Larissa Fuchs, Andre Groeger, Daniel
Herbold, Malte Heisel, Robin Kraft, Stefan Pasch, Jutta Preussler, Elsa Schmoock, Patrick Schneider,
Sonja Stamness, Sascha Wilhelm, who provided excellent research assistance.
2
1. Introduction
“How can members of a team be rewarded and induced to work efficiently?” This is the
question that Alchian and Demsetz (1972) asked more than 40 years ago in one of the most
influential contributions to the economic literature on organizations. Alchian and Demsetz’
focus was on input monitoring; an alternative would be incentives conditioned on joint output.
The very nature of teamwork, however, blurs the performance of individuals into a common
performance signal, weakening the effect of monetary incentives (Holmström, 1982). While
there is substantial evidence that incentives work quite well provided individual performance
is measurable (Lazear, 2000, Shearer, 2004, Bandiera et al, 2009), a number of questions are
open: under what conditions do team incentives raise efficiency in the field, and by how much
(Bloom and van Reenen, 2011), do such incentives lead to unintended reactions and gaming,
and what mechanisms may affect performance?
We address this research gap through a randomized, controlled experiment in a retail
chain in Germany. Our study is the first one in which the effect of team incentives is analyzed
in a natural field experiment combining both realism and randomization (List and Rasul,
2011). In particular, the employees are working in an ongoing firm and they are carrying out
their normal day-to-day job. Besides the change in the compensation scheme, there is no other
intervention, and we took great care in ensuring that employees would not consider
themselves as part of an experiment. Except for our partners in management and the worker
council, no one was aware of our involvement and communication was taken care of by
management, not by us. The firm used the term “pilot”, which they also use when introducing
new HR or marketing practices for a limit period of time, and we conditioned incentive pay
on the existing performance measurement system used for the compensation of middle and
lower management.
Using the stratified randomization method developed by Barrios (2014), we
introduced a team bonus conditioned on sales targets that were fixed well before the team
incentive was discussed. Treated shops increase their sales on average by around three
percent, wages increase by around two percent on average (and up to 13 percent). The team
incentive works best for (i) shops in larger towns and cities where arguably the marginal
productivity of effort is highest; (ii) shops with younger employees, for whom the marginal
costs of effort is likely to be lowest, and (iii) shops that did not reach their targets regularly
before the introduction of the bonus, for whom the effect of effort on the marginal probability
of success is likely to be largest. A fourth result is owing to an institutional specificity in
Germany: roughly a third of the workers in our shops are so-called mini-jobbers, registered
3
unemployed with an income (on top of unemployment benefits) of 450 or less. For tax
reasons, these employees were not eligible for the bonus. We find that the treatment effect is
lower in shops with a higher proportion of mini-job workers, although the eligible team
members would receive a larger share of the team bonus (holding other things equal). This
result points to the importance of complementarities between team members.
The main effect of the team incentive consists of an increase in the customers served,
so incentives seem to increase operational efficiency, rather than increasing sales by up-
selling activities. We find no effect of team size on treatment effect, which seems
counterintuitive from a free-riding perspective, but is in line with the “group-size paradox”
analyzed by Esteban and Ray (2001). We also find no evidence that the incentive is gamed as
measured by additional orders of bread or higher return rates of unsold bread.
Our study distinguishes itself from the existing literature on the effect of group
incentives and team work. First, in contrast to much of the literature, we look at small work
teams in which people interact on a regular basis, and not on agencies or divisions like
Propper et al (2011), Courty and Marschke (1997), or other papers surveyed by Prendergast
(1997). Second, we are dealing with a technology in which job design necessarily builds on
team rather than individual work. (Why this is the case, we explain in the paragraph below).
Our question is not whether team rather than individual work is preferable for incentive and
efficiency reasons (Itoh, 1991, Che and Yoo, 2001). Rather, we ask whether given
technologically determined work organization in teams, an incentive can raise output (sales),
in what magnitude, and what the impact of incentives depends on. Third, we control for
variables widely believed to interact with team incentives, such as organizational
commitment, job, and context satisfaction and perception of leadership, but find no significant
results.
Our retail chain consists of shops where employees bake and sell pre-fabricated bread
and cakes, prepare and sell sandwiches, snacks, and hot and cold beverages. On average,
seven full- and part-time employees work in each shop, a third of the employees are mini-
jobbers. Wages are low (on average around ten Euros), at a level slightly above the currently
debated, but not yet implemented, nation-wide minimum wage. Individual work organization
and performance measurement is impossible, because there is a broad variety of tasks each
person has to carry out, including handling the goods delivered, preparing food in the oven,
taking care of the customers, and handling the cash register. The time workers spend on each
task varies much, people work in overlapping shifts and are supposed to help each other. The
need to deal with different tasks of high volatility makes it too costly to have highly
4
specialized agents who would be idle most of their time. Furthermore, providing individual
incentives would lead to measurement and gaming problems and productivity losses because
of forgone help efforts among the members of the team (cf. Itoh, 1991, Auriol et al, 2002).
For many years prior to our intervention, shop performance has been measured along a
number of dimensions, such as sales, personnel costs, and qualitative indicators, all of which
have traditionally been used to incentivize top, middle and shop managers. Prior to the
intervention, however, the more than 1,000 sales agents in the shops never received any
performance related pay. In April 2014, in half of the almost 200 shops, we introduced a
monthly team bonus. Shops were assigned to treatment and control groups through an optimal
stratification procedure developed by Barrios (2014). Management informed the teams in the
treatment group through personal communication, letters and posters about the incentive
scheme.
As usual in sales people compensation we used a step bonus function (Figure 1). We
were aware that step functions have many issues, but linear compensation Teams that reached
or surpassed by up to one percent the sales target defined by top management at the beginning
of the year (well before the decision in favor of the team incentive was made) would receive a
100 Euros bonus. For each additional one percent of sales beyond the target, an extra bonus of
50 Euros was offered. The bonus was capped at 300 Euros when the sales exceeded the target
by more than four percent. Teams were initially informed that the bonus would be paid for a
pilot period of three months ending on the 30th
of June 2014. The teams were informed that
the bonus would be shared among the full- and part time employees including the shop
manager, according to the hours they worked in the respective month, compared to the total
work hours of the team.
We find a treatment effect of roughly three percent on sales over the period between
April and June 2014. Many of the teams reach sales levels beyond which the bonus is capped.
Interestingly, and contrary to the free-riding argument but in line with the literature on the
group size paradox (Esteban and Ray, 2001), larger shops in the treatment group fare no
worse in terms of sales than smaller ones. Shops in cities compared to smaller municipalities
feature treatment effects of around six percent, arguably because consumers react more
intensively to increased sales activities. Shops that, in the past, were less likely to reach the
sales targets react more intensively to the bonus than shops that were more likely. Job and
context satisfaction, and organizational commitment, as measured by a firm-wide survey prior
and unrelated to the introduction of the team bonus, plays no role for the effect of the team
bonus. Neither does the treatment affect these measures in a second survey during the
5
treatment period. The firm made extra profits of around 50,000 Euros per month from the
treated shops, and the wage payments to the sales agents increased by around 12,000 Euros
per month. We carry out a comparison with investments in renovations if the point and sales
and argue that the team bonus provides much higher returns to the firm. Following a
successful implementation of the team bonus, the firm decided to roll out the scheme to all
shops for the period from July to December 2014, signing a respective agreement with the
worker council at the end of June.
For many reasons retailing is a natural candidate to test for the effectiveness of team
incentives. Many people in retail chains work in teams, that is, the individual efforts can only
be mapped into some joint output signals, such as sales.2 Demand volatility is high, so people
must be prepared to switch tasks frequently, and even if sales people can specialize to some
extent, they must be willing to help each other, all of which makes it complicated, if not
impossible, to provide them with individual incentives. A study of team incentives study in
retailing, such as ours, will also make a rather general case because retailing is one of the
largest industries in the world in terms of employment. In Germany, the country in which our
experiment took place, more than 4 million people work in retailing, that is, almost 10% of
the country’s active labor force. Most of these people work in larger groups or chains,3 just as
in our study. Many other service industries, in particular restaurants, hotels, airlines, are
similar because, again, technology forbids individual incentives, people carry out many tasks,
individual performance measurement is difficult, if not impossible, and people do not get
individual incentives.4
Our paper contributes to the existing literature on incentives by providing clean causal
evidence on the performance effects of team incentives, thus filling an important gap in
empirical research. In our study, all shops have the same technology, team incentives are
randomly assigned, people cannot self-select into teams, because hiring is centralized, and
there are no moves between shops. This makes our study different from other papers in which
the adoption of teamwork or team incentives may be endogenous. Boning et al (2007),
Hamilton et al (2003), Bandiera et al (2009) and Bandiera et al (2013) all find supportive
evidence that team work and incentives raise efficiency. However, the first study shows that
the decision in favor of teamwork and its effects depend on technology, while the others
2 http://job-descriptions.careerplanner.com/Retail-Salespersons.cfm 3
http://de.statista.com/statistik/daten/studie/261517/umfrage/beschaeftigte-im-deutschen-einzelhandel-
nach-unternehmensgroessen/ 4 Other type of restaurants and cafés a very different type of business. Here, workers are much more
specialized and they receive individual incentive pay in the form of tips.
6
observe sorting of workers of similar productivities into teams. In particular, Hamilton et al
(2003) are instructive as they find that people even forgo individual earnings in exchange for
working in a team, which is a question orthogonal to ours, where team work is
technologically fixed, but compensation schemes vary. Also complementary to our work is
the paper from Delfgaauw et al (2013), which look at competition between teams,
incentivized through tournaments. The authors focus on gender, the effects of prize spread,
how far shops are away from targets and of social cohesiveness in teams. Our paper rather
shows the pure effects of an incentive conditioned on team output (not on relative output), has
teams that almost entirely consist of women, finds substantial effects of the bonus of six
percent in bigger municipalities, but no measurable interactions with organizational
commitment, perceived leadership and other factors. CITE ALSO DELFGAAUW AT AL
2014; THEY FIND IN A TOURNAMENT THAT SHOPS THAT ARE FAR BEHIND DO
NOT REPSOND TO RELATIVE INTENCIVES; WE FIND THAT SHOPS THAT LAG
BEHIND RESPOND THE MOST.
Is a six percent sales increase a large effect? We would argue yes. First of all, as
noticed by a substantial literature by Bloom and co-authors, Germany is one of the countries
with the highest level of managerial efficiency, second only to the US, and that also applied to
retailing (Bailey and Solow, 2001). Second, in stark contrast to France (Bertrand and
Kramarz, 2002), entry barriers are low, and competitive pressure, in particular triggered by
aggressive discounters such as Aldi and Lidl, is high. It is actually precisely the entry of these
firms into the market of “our” chain that triggered the change in incentives that we analyze
here.
2. Background
In 2013, we were contacted by the general manager of a bakery chain who sought
advice on how to cope with the challenges of a rapidly changing market. Since the 1980s,
bakery chains, some of them owning hundreds of shops and with sales numbers of up to a
billion Euro, had successfully built their business model on attractive locations (including
supermarkets and malls), and economies of scale. The chains had crowded out many of the
existing small master bakeries whose number and market shares had steadily declined. In
2011, however, discounters Aldi and Lidl had begun to bake and sell fresh bread, rolls and
related products in their dense network of shops, with large success. Their bread is widely
believed to be of similar quality as the one of the chains, but sold at much lower prices, hence
forcing the incumbent chains to rethink their business model. As a consequence, many of the
7
chains were differentiating their product portfolio moving into the market for snacks and
sandwiches and beverages, traditionally covered by cafés and fast food chains. They also tried
to intensify their sales activities on freshly baked cakes, a market the discounters have not
(yet) entered. These moves were accompanied by substantial investments in point of sales,
making them more modern, better designed and often equipping them with a “café section”
were customers can sit down to eat and drink.
The manager told us that for this strategic move to be successful, the behavior of the
personnel should be changed, and that they needed to be more actively involved in sales
activities. The English saying “something sells like hotcakes” has a German equivalent
“something sells like warm rolls” or “something sells like sliced bread”, and, indeed, many of
our partners in the firm (“the bakery”) complained that sales agents took a rather passive
stance in their interactions with customers. Turnover in the bakery was relatively high,
making training activities a questionable investment, also because personnel consisted mainly
of very low-skilled workers without comprehensive vocational training (only one fourth of the
staff have finished an apprenticeship). The bakery had experimented before with hiring more
qualified employees, but concluded that there was little if any increase in the quality of
service, at the cost of a substantially increasing wage bill. So, the challenge was to find a lever
through which the existing staff could be motivated to spend more effort on their sales
activities.
We agreed to help the bakery under the condition that we would have access to all
relevant data, and could design a possible intervention through a randomized controlled
experiment. In particular, we explained that our randomization would be more successful, if
we had access to historic sales data. We received sales data over more than two years,
allowing us to carry out a very precise stratified assignment, which is explained in more detail
in the next section. In exchange we offered our advice free of charge, and covered some of the
costs, in particular the one of research assistant needed to carry out an employee survey
before and during the treatment period.
When our partners presented the existing compensation structure to us, we realized
that there was a very detailed system of key performance indicators (KPI) according to which
managers were evaluated, and on the basis of which they were compensated. These KPIs
include sales, personnel costs, specific strategic goals such as enhancing the share of a certain
product in total sales etc. Most importantly, at the end of each year, sales targets are
determined for the next year, and these targets are never adjusted during the year. For the year
2014, management had based targets on 2013 sales, corrected by an expected decrease of 2%
8
on average. District managers, who are responsible for 10 to 15 shops, are incentivized
through a number of targets along these dimensions, and the 193 shop managers receive
bonuses when they reach certain sales and personnel cost targets and certain grades in
regularly carried out mystery shopper visits (the topic of another paper of ours). However,
sales agents, representing the bulk of staff, only receive fixed wages, which for most workers
are regulated by collective agreements. There is also a second group of workers, on so-called
mini jobs who can earn up to 450 Euro on top of their welfare benefits.
At the end of February 2013, we proposed introducing a team bonus conditional on
reaching or passing the existing sales targets. Our partners first reacted surprised by this
suggestion. A member of the management team put it bluntly: “We have never seriously
thought about this.” Other members of the management team were afraid that payments could
turn out to be a burden on the firm. We provided our partners with some simulations showing
that the expected payments would likely to be lower than 20,000 Euro per month in case (i)
half of the shops were treated; (ii) a step function capped at four percent sales above the target
was used; with (iii) a top monthly bonus of 300 Euro. While this convinced top managers to
try a “pilot” study with half of the shops assigned to the team bonus, district managers were
afraid that wage costs would rise, meaning that they could not reach the targets of keeping
personnel costs low. The general manager reacted by suggesting that the extra bonus
payments would be paid from a different budget and would not affect the personnel costs
relevant for district managers’ performance. District managers were quick to realize that in
such a setting they were likely to benefit as well, if the team bonus resulted in an increase of
sales of the shops under their supervision. The worker council also was in favour of the
bonus, in particular, because it was designed as a pure add-on payment. Also, trust between
worker council and management was high, and a new collective agreement had been written
concerning fixed wages, such that ratchet effects were unlikely.
To reduce the risk of information leakage, middle management was informed about
treatment and control shops under their purview only some days before introduction of the
team bonus. Another concern was raised about the possibility of envy between treatment and
control shops. We discussed this issue with the worker council who suggested that we should
explain that the intervention was “just a pilot” and that everybody would have the same 50-50
chance of taking part in it. This, the worker council argued, would be acceptable for the non-
treated shops in case they would learn about the bonus scheme in other shops. In the next
section, we explain in more detail how we communicated the scheme and how we minimized
the risk of contamination between the treated and non-treated shops.
9
3. The experiment
In March 2014, a month prior to the introduction of the team bonus, and before its
announcement to the employees and managers of the treatment group of shops, we carried out
an employee survey in order to measure employee satisfaction with the job context, general
situation, and organizational commitment, following some influential work in industrial
psychology (Allen and Meyer, 1990). In May, we carried out a second wave. This latter
survey entailed the same questions plus additional items in relation to the team bonus, and
some social interactions within and between shops.
Our partners and we thought the survey would be useful for a number of reasons. First,
it would allow an additional check whether treatment and control groups were balanced
samples; second, we hypothesized that the treatment effect may depend on employees’ pre-
treatment satisfaction and commitment levels; third, because of the team bonus, employees’
satisfaction levels could increase. The survey was distributed through the district managers
and collected by our research assistants in sealed envelopes. Arguably, because the
anonymous surveys were collected on site, and because of ours’ and the district managers’
substantial communication efforts, the feedback rates were almost 80 percent in the first and
above 60 percent in the second wave of the survey.
In preparation of team bonus introduction, we designed information leaflets to be
posted in the back offices of the treatment shops, and letters that were distributed by the
district managers to the employees. In contrast to the employee survey, the logo of Goethe
University did not show on these materials (see Appendix 2), so that people would not
perceive themselves as part of an experiment. In fact, there was no mention of our research
team in any communication regarding the bonus.
We trained district managers in how to explain the team bonus to shop managers. We
also instructed them about how to react to questions of employees in control group shops
about the bonus. In that case, district managers would explain that “this is a pilot. Everybody
had the same chance to be drawn with a 50-50 chance. The work council agreed to this
procedure.” We were afraid that there could be contamination between the treatment and
control group. Envy or frustration of control group shops could lead them to reduce their
efforts, which could then be picked up erroneously as a positive treatment effect. To monitor
this risk, we regularly called the district managers and asked them whether employees in the
control group had asked them questions about the team bonus. We actually only heard about
three cases in April, all of which were satisfied with their district manager’s answer. In May
10
and June, nobody inquired. In addition, we also regularly checked the Facebook page of the
bakery on which customers and employees alike tend to discuss issues such as product quality
but also (sometimes to the dissatisfaction of management) internal issues such as stress at the
workplace, quality issues of products, or problems of leadership and organizational culture.
We could not find a single entry on the team bonus. However, we also built in some questions
into the second wave of the employee survey to check for potential channels of
contamination, in particular, we asked people how frequently they interacted with colleagues
in the same and in other shops, and we found that contacts to employees in other shops are
very rare: only 20% of respondents indicated that they ever spoke to a colleague from another
shop. Hence, it seems fair to say that contamination between treatment and control shops is
unlikely to be an issue.
FIGURE 1 ABOUT HERE
Figure 1 depicts the bonus offered to the treatment shops. Shops that reach the sales
target for the month receive a bonus of 100 Euro to be shared between the part-time and full-
time employees in the shop. The bonus increases by 50 Euro for each percent point above the
target and is capped at 300 Euro per month at 4 percent points above the target. Hence, the
team in a shop can make extra earnings of up to 900 Euro in the treatment quarter April to
June. We provided the employees with examples of what sales increases would mean in terms
of additional goods to be sold per day (for instance a one-percent increase above the sales
target for a mid-sized shop would be tantamount to selling per day ten additional rolls, two
loafs of bread, some sandwiches and some cups of coffee). The fact that a number of shops
failed to reach the target by small amounts (for instance, in April, one shop failed to reach the
target by 16 Euros, and another one by 8 Euros) is an indicator that there was no manipulation
and that at least in the beginning of the treatment phase, employees found it hard to estimate
their likelihood to reach the various targets, although district managers regularly communicate
to all shops, treatment and control shops alike, their current sales figures.
What rests to be explained is the randomization procedure. We follow Barrios (2014)
who shows that randomizing pairwise by using the predicted outcome variable, in our case
sales, minimizes the variance of the differences in the treatment and control outcomes post
treatment. We use historic observations between January 2012 and December 2013 to run a
regression of log sales on labor input with month and shop fixed effects, from which we
obtain predicted sales. We then rank the shops according to the predicted sales and randomize
within the pairs of shops with adjacent ranks (1-2, 3-4, ..., 192-193). Because we had an odd
number of shops in our sample, we excluded the shop with the median rank from the pairs and
11
assigned it randomly. The resulting treatment and control groups comprised 97 and 96 shops,
respectively. Table 1 summarizes their pre-treatment characteristics.
TABLE 1 ABOUT HERE
Thanks to our randomization procedure, the treatment and control samples are
balanced in the average pre-treatment sales, our key outcome variable. They are also similar
in other potentially relevant characteristics, such as the percentage of unsold goods, number
of customer-visits, frequency of achieving the sales target, location, and employee attitudes.
In fact, none of the averages reported in Table 1 differ significantly between the groups. An
average shop sells over 27,000 Euros worth of goods5, employs 7 people most of whom are
female in their late 30s, unskilled, and working part-time. There is a sizeable share of workers
on a mini-job, around 30%, who for tax reasons were excluded from the team bonus scheme.
Sales are quite variable, with location and size differences explaining 90% of the variance.
There is also a considerable variation within shops, much of which is due to seasonal demand,
temporary closures for renovation, and market dynamics, such as the entry and exit of
competitors, all of which factors we control in our statistical analysis.
FIGURE 2 ABOUT HERE
Figure 2 displays spatial distribution of our control and treatment shops. The region in
which our partner firm operates spans roughly 100 km from West to East and 60 km from
North to South, an economy of more than 3 million inhabitants. Shop locations vary in
population size. However, all shops are placed on the premises of supermarkets owned by the
parent company, or in their immediate vicinity, relying thus mostly on the customer traffic to
and from grocery shopping.
4. Baseline results
Table 2 reports the treatment and control shops characteristics in the treatment period (April
to June 2014), giving a first impression of the treatment effect. Sales and customer-visits have
gone down, reflecting the secular downward trend in the bakery business. Yet, the drop in
sales and customer-visits being more pronounced in the control than in the treatment group
suggests a positive treatment effect. In fact, the difference-in-difference estimated effects on
the log sales and customer-visits are 3.3% and 2.8%, respectively, both significant at
5 One shop, located in the busy Frankfurt airport and assigned randomly to the treatment group, sold
on average 118000 euros worth of goods per month in the pre-treatment period and employed 22
people. Excluding this shop, the average pre-treatment sales in the treatment group are 27176 euros
per month with standard deviation of 10885 euros, which is much closer to the same characteristics of
the control group. Removing this shop from our regression sample does not change the estimated
treatment effects.
12
conventional levels. Since there is no significant treatment effect on other outcomes, we
proceed with a more in-depth analysis of sales and customer-visits.
TABLE 2 ABOUT HERE
To visualize the treatment effect on sales, Figure 3 plots the treatment and control
groups' year-on-year sales growth in the treatment month versus the sales levels in the same
months (April to June) of 2013. Additionally, Figure 4 displays the kernel density graphs of
the year-on-year sales growth for the two groups. There is a shift in the treatment group's
sales growth distribution to the right from the control group's which is fairly uniform across
the growth rates and initial sales levels.
FIGURE 3 ABOUT HERE
FIGURE 4 ABOUT HERE
To identify the treatment effect in a more systematic way, we run the difference-in-
difference estimator in several regression specifications where we control for other factors
that may affect sales and address the estimation issues most frequently discussed in the
experimental econometrics literature. The first issue is serial correlation in the residuals,
which leads to underestimated coefficient standard errors and false positives as a result
(Bertrand, Duflo and Mullainathan, 2004). The second is the correlation between the
treatment status and the baseline outcome, which, despite randomization, may occur in finite
samples, causing the “regression towards the mean” problem (Stigler, 1997).
We start with a specification with shop and time fixed effects:
(1)
ln(𝑠𝑎𝑙𝑒𝑠𝑖𝑡) = 𝛽0 + 𝛽1𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡𝑖 + 𝛽2𝑎𝑓𝑡𝑒𝑟𝑡 + 𝛽3𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡𝑖 ∗ 𝑎𝑓𝑡𝑒𝑟𝑡 + 𝑝𝑒𝑟𝑖𝑜𝑑𝑡
+ 𝑠ℎ𝑜𝑝 𝑓𝑖𝑥𝑒𝑑 𝑒𝑓𝑓𝑒𝑐𝑡𝑖 + 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑡 + 𝑒𝑟𝑟𝑜𝑟𝑖𝑡
where ln(salesit) is the log sales in shop i and period t, the treatment dummy takes the values
1 for the treatment and 0 for the control group shops, the after dummy is 0 for the periods
before treatment and 1 thereafter, 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑡 include the log total hours worked and dummies
for renovation within the last two months, and errorit is the idiosyncratic error term which we
allow to correlate within each shop using the Stata cluster option. Coefficient 3 is the
difference-in-difference estimate of the treatment effect. Columns 1 and 2 in Table 3 are
based on equation (1).
TABLE 3 ABOUT HERE
In addition to clustering errors at the shop level, which may still underestimate
coefficient standard errors in small samples, we implement another solution proposed in
13
Bertrand, Duflo and Mullainathan (2004) – to collapse all observations into the pre- and post-
treatment periods and estimate
(2)
ln(𝑠𝑎𝑙𝑒𝑠𝑖𝜏) = 𝛽0 + 𝛽1𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡𝑖 + 𝛽2𝑎𝑓𝑡𝑒𝑟𝜏 + 𝛽3𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡𝑖 ∗ 𝑎𝑓𝑡𝑒𝑟𝜏
+ 𝑠ℎ𝑜𝑝 𝑓𝑖𝑥𝑒𝑑 𝑒𝑓𝑓𝑒𝑐𝑡𝑖 + 𝑐𝑜𝑛𝑡𝑟𝑜𝑙𝑠𝑖𝑡 + 𝑒𝑟𝑟𝑜𝑟𝑖𝑡
where is the average log sales pre ( =0) and post ( =1) treatment. Column 3 in
Table 3 reports the treatment effect estimated with equation (2). Bootstrapping, another
recommended solution, produces standard errors of similar magnitude.
To control for regression to the mean in we augment eq. (2) with past sales:
(3)
where regression errors are still clustered at the shop level. The estimates from equation (3)
are reported in column 4 of Table 3.
In the next series of specifications, we estimate the treatment effect by comparing
post-treatment sales growth relative to a chosen baseline b in the treatment and control
groups,
(4)
In principle, specification (4) is similar to (3), but some extra flexibility in regression
specification is achieved by varying the baseline, , specifying it as the average
sales across all months before the start of the treatment (column 5 in Table 3), the same
months in 2013 (column 6), and the average sales in January-March 2014 (column 7).
Whatever specification we use, we obtain the average treatment effect estimates of
similar magnitude – around 3% – and significance. This uniformity suggests that neither of
the estimation issues we mentioned above and addressed in our analysis is important on our
sample. Indeed, simply clustering the errors by shop is sufficient on the relatively large
sample such as ours. Regression to the mean is not a concern either since our sample is well
balanced. The estimated average treatment effect on sales of 3% implies an extra 820 Euros
(=[exp(0.03)-1]*27000) worth of sales per month in the average shop. Calculating the
treatment effect in each month with our preferred specification (1) (results reported in Table
4), we find it to be 2.9% in April 2014, 3.7% in May, and 2.9% in June, a steady effect
without noticeable abatement.
TABLE 4 ABOUT HERE
Turning to the treatment effect on salespeople's income, more than 50% of the workers in
the treatment group received a bonus in the treatment period, which averaged at 114.7 Euros
ln(salesit )
ln(salesi1) = b0 +g × ln(salesi0 )+b1 × treatmenti +b2 ×aftert +b3 × treatmenti ×aftert +errorit
ln(salesit ) - ln(salesib) = b0 + g× ln(salesib ) + b3× treatmenti +errorit
ln(salesib )
14
or 3.9% of the average recipient's quarterly earnings. The total bonus payments made by the
company in April to June 2014 amounted to 35,150 Euros.
To gauge the profitability of our team bonus scheme, we compare the implied gains from
it with its implied total costs. With the treatment effect of 820 Euros per average shop per
month, the extra sales amount to 474780 Euros per quarter (=820 Euros times 3 months times
193 shops). Given the historic share of 0.56 of value added in sales, the implied operational
profit gain is 265,880 Euros. On the costs side, there are additional bonus payments to shop
managers and higher ranks (which we estimate at 20,000 Euros per quarter), and our research
expenses, which, though not billed to the company, need to be included in the costs. We
estimate around 60 person days of senior researchers at a rate of 1,000 Euro = 60,000 Euros.
Also we estimate 50 research assistant days at the going rate of 110 Euro = 5,500 Euros.
Material and travel costs were around 12,000 Euros. The total implementation costs of the
bonus scheme are thus 132,650 Euros, or about half its implied gains. Put differently, the
bonus scheme as an “investment in people” project would break even within one quarter from
its start.
In a meeting in June 2014, the management team decided to roll out the team bonus to
all shops as of July 1st 2014. A collective agreement was written with the work council,
according to which the team bonus would be granted until the end of 2014.
5. Treatment effect heterogeneity
Although our treatment and control groups are balanced across a number of characteristics
that might affect sales, the treatment effect may still vary in magnitude between different
shops in the treatment group. We expect the treatment effect to vary along several
dimensions, among which we analyze the following: shop location, workforce size and
composition, success in reaching the sales target in the past, and employee attitudes.
5.1. Shop location
Shop location affects the magnitude of effort’s response to a given incentive by changing the
marginal product of effort. Thus, extra effort pays more in populous, urban locations that have
office workers who might come in for lunch, and visitors who might buy a snack on the go;
incentivized sales agent may “up-sell” to both these groups. On the other hand, smaller
locations have mostly regular shoppers whose demand for bread is harder to affect - hence the
lower marginal product of sales effort in those locations. Besides, shops in urban locations
have more competitors nearby, whose customers may be won over.
15
TABLE 5 ABOUT HERE
Table 5 reports the treatment effect by shop location. As expected, the treatment effect is
largest, at 6%, in shops located in big towns (>60000 inhabitants), going down to 3.8% in
midsize towns, and zero in villages. As before, the treatment effect is fairly stable in time.
5.2. Workforce size and composition
Shop workforce size will influence the magnitude of the treatment effect by increasing
the total effort as the sum of individual efforts, as well as by decreasing the individual effort
through free-riding. Which of these two opposite tendencies will prevail depends on the team
production technology and the individual costs of effort function. Thus, Esteban and Ray
(2001) show, in a collective action setting, that when the costs of effort are quadratic or
steeper larger teams will outperform smaller ones in total effort even if there are no individual
effort complementarities. The presence of complementarities, that is, rewards to individual
team members being nonexcludable, or, equivalently, the total effort being more than the sum
of individual efforts, will reduce the ``steepness'' of the costs of effort function required to
deliver the total effort growing with team size.
TABLE 6 ABOUT HERE
To measure the variation in the treatment effect with shop size, we interact the
treatment dummy with the dummies for the quartiles of the shop-average number of workers
not on a mini-job (the mini job workers did not receive a bonus). Table 6 shows that the
treatment effect is larger in bigger shops (column 1), and that the observed differences in the
treatment effect do not owe to bigger shops being located in bigger towns (column 2). In fact,
the treatment effect increases with shop size faster in big towns than elsewhere.
Turning to the shop workforce composition, we explore treatment effect heterogeneity
with shop workforce tenure, age, and the share of mini-job workers. We expect the treatment
effect to be larger for younger workforce, since younger workers are on average poorer and
thus more susceptible to material incentives. Besides, holding wealth fixed, there may be an
element of resistance to change, which is weaker for younger workers, in the individual
responses to our innovative treatment.
TABLE 7 ABOUT HERE
Table 7 reports treatment effects in the shops below and above the median workforce age and
tenure, on the whole sample as well as separately in big towns and elsewhere. Consistently
with our expectations, ``younger'' shops respond to treatment more strongly. A further
analysis suggests that the differential response to treatment by age and tenure is driven mainly
16
by age: running our preferred difference-in-difference specification (1) with the treatment
effect being interacted with age and tenure separately as well as jointly produces a significant
interaction with age but not with tenure.
The treatment effect should decrease linearly with the share of mini-job workers in
shop team, reflecting a decrease in the size of the team that is incentivized. There will also be
an additional negative influence if there are effort complementarities between mini-job and
ordinary workers, since stronger complementarities increase the weight of each worker's
contribution to the total output6. To accommodate the later, nonlinear, effect, we rerun our
regression specification with the treatment dummy interacted with the quartiles of the shop-
average share of mini-job workers. The results are reported in Table 8.
TABLE 8 ABOUT HERE
We find that the treatment effect goes down with the share of mini-job employees, especially
in the shops located outside big towns. The abrupt drop in the treatment effect to zero past the
second (whole sample) or first (shops outside big towns) quartile of the average mini-job
worker share implies a steeper than linear decrease, which suggests effort complementarities
between mini-job and ordinary workers in shop team.
5.3. Past sales target achievement
We expect the treatment effect to vary with the past performance around the sales
target. Historic record of achieving sales targets is informative for shop teams to gauge their
probability of success in the future, since the targets are largely based on past sales (with a
correction for the overall trend, hence the higher frequency of reaching the target in both
groups, recall Table 2) and set in the beginning of the year. However, the pattern of the
treatment effect's variation with past performance is hard to predict, since the marginal utility
of effort, and hence the effort's response to treatment, is influenced by at least two opposing
considerations. First, although for the more successful shops the expected bonus is higher, the
marginal utility of putting more effort than before the treatment is lower because the bonus is
capped. Hence, shops historically performing closer to the target would respond to treatment
less strongly. On the other hand, shops that have been too unsuccessful in reaching their
targets in the past may not react because reaching the target is not realistic enough.
TABLE 9 ABOUT HERE
6 As an example of the empirical framework required here, Iranzo, Schivardi and Tosetti (2008)
estimate a constant elasticity of substitution production function of different workers' skills within
their firms. They find skill complementarity between, and substitutability within, occupational groups.
17
Table 9 reports treatment estimates by quartile of historic distance to the sales target,
which is measured in two ways: i) as the difference between actual sales and sales target
averaged for each shop over the pre-treatment period; and ii) as the frequency of each shop
achieving its target in the pre-treatment period. Shops in the bottom three quartiles of the
distance to the target reacted to the treatment more strongly than did those in the top quartile,
suggesting that rewarding the attainment of too easily achievable targets is not an effective
motivator.
5.4. Employee attitudes
We expect the effort response to treatment to vary with the workers' attitudes towards
their employer. To investigate this possibility, as well as to gather important background data
for our study, we ran questionnaires among the employees before (March) and after (May) the
treatment. In the cover letter sent to every shop, we emphasized that this survey was for our
research purposes and had nothing to do with their employer, and guaranteed anonymity of
employees' response; we also distributed and picked up the questionnaires ourselves rather
than let the company do so for us, as an extra guarantee of anonymity. We had an 80%
response rate for the first survey, and 65% for the second.
There are three aggregate attitudes scores we measured in both waves of the
questionnaire: commitment to the firm, satisfaction with the work context, and overall job
satisfaction. None of these scores were affected by our treatment (recall Table 2). Moreover,
none of them moderates the treatment effect, implying that workers' response to treatment
does not depend on their attitudes towards their firm or their job.
6. Mechanisms
The extra 3% of sales in the treatment shops compared to control may have been achieved by
serving more customers, selling more per customer, having a broader range of goods to better
serve diverse customer demands, selling more side goods (drinks, snacks, etc.), or through a
combination of the above mechanisms. In this section, we go through available empirical
evidence to ascertain the likely role of each of these mechanisms.
TABLE 10 ABOUT HERE
We find a treatment effect on the number of customer-visits that is commensurate with
that on sales (Table 10), so that the effect on sales per customer visit is virtually nil. This
finding implies that the role of mechanisms leading to more sales per customer visit, such as
broader assortment and more side sales, is marginal. With most of the extra sales driven by
18
the increase in the number of transactions, is it greater operational efficiency or better
customer service that drove this increase? Greater efficiency would be part of the explanation
if there were longer queues before the treatment than thereafter. However, comparing the
mystery shopping reports in April to June 2014 with those in January to March 2014
(immediately before treatment), we see no change in the reports on salesperson availability,
which was generally rated as high. Hence, there is no evidence to suggest a significant
decrease in the customer queues.
Turning to better customer service as the remaining explanation, we ran our own
mystery shopping tour of 140 randomly selected shops in our sample (capacity constraints
prevented us from touring every shop). Our research assistants were instructed to act like
ordinary customers and to buy the ``bread of the month'' or the closest substitute to it. After
leaving the shop, they were asked to take note of how friendly the sales staff were, and
whether the question ``Would you like anything else?'' or similar was asked. We found that
while the frequency of asking the ``anything else?" question was comparable in the control
and treatment groups overall (0.72 control vs. 0.79 treatment), in big towns, where the
treatment effect was largest, the treatment shops asked this question significantly more
frequently: 0.82 of the time vs. 0.61 in the control shops. We have also found a positive
correlation, though insignificant, between asking this question and monthly sales in the
treatment period. However, rerunning our regressions with the ``anything else?" question
included as control, we see no reduction in the treatment effect. In sum, although there are
signs pointing to the importance of enhanced customer experience in shaping the treatment
effect, we do not have strong enough evidence to precisely identify the mechanism that drove
our results.
7. Conclusion
Teams are a ubiquitous feature of modern production, and so are monetary incentives.
While the knowledge about the effectiveness of individual incentives is both broad and deep,
much less is known about team incentives. Problems of endogeneity, complementarities and
self-selection into teams make causally interpretable evidence about the effectiveness of team
incentives hard to obtain. We contribute to the incentives literature by providing evidence on
the effectiveness of team incentives. We have designed a fairly large randomized controlled
experiment with 193 shops of a bakery chain in Germany. Power calculations on the basis of
27 months of observations pre treatment and 3 months post treatment informed us that we
would need 70 shops in each group to detect a 3 percent treatment effect at a significance
19
level of 5 percent with the probability 0.9. Our estimated treatment effect is indeed around
3%, and is highly significant. There is also substantial heterogeneity, with the treatment effect
being largest in big towns, shops with younger workforce and few mini-job employees. The
single most important immediate cause of the treatment effect on sales we find is increased
customer traffic; there is no effect on sales per customer visit. We are unable to precisely
distinguish between greater operational efficiency (that is, smaller queues) and better
customer service as the two mechanisms that led to higher sales through increased customer
traffic.
20
References
Alchian, A. A., & Demsetz, H. (1972). Production, information costs, and economic
organization. The American Economic Review, 62(5), 777-795.
Allen, N. J. & Meyer, J. P. (1990). The measurement and antecedents of affective,
continuance and normative commitment to the organization. Journal of Occupational and
Organizational Psychology. 63(1), 1-18.
Auriol, E, Friebel, G. & Pechlivanos, L. (2002). Career concerns in teams. Journal of Labor
Economics, 20(2), 289-307.
Baily, Martin Neil, and Robert M. Solow. "International productivity comparisons built from
the firm level." Journal of Economic Perspectives (2001): 151-172.
Bandiera, O., Barankay, I & Rasul, I. (2013). Team Incentives: Evidence from a Field
Experiment. Journal of the European Economic Association, 11(5), 1079-1114.
Bandiera, O., Barankay, I & Rasul, I. (2009). Social Connections and Incentives in the
Workplace: Evidence from Personnel Data. Econometrica, 77(4), 1047-1094.
Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How much should we trust difference-in-
difference estimates. Quarterly Journal of Economics, 119(1), 249-275
Bertrand, M., & Kramarz, F. (2002). Does entry regulation hinder job creation? Evidence
from the French retail industry. Quarterly Journal of Economics, 117(4), 1369-1413.
Barrios, T. (2014). Optimal stratification in randomized experiments. Mimeo, Harvard
University.
Bloom, N. & Van Reenen, J. (2011). Human resource management and productivity.
Handbook of Labor Economics, 4B(19), 1697-1767.
Boning, B., Ichniowski, C. & Shaw, K. (2007). Opportunity counts: Teams and the
effectiveness of production incentives. Journal of Labor Economics, 25(4), 613-650.
Courty, P. and Marschke, G (1997). Measuring government performance: Lessons from a
federal job-training program. American Economic Review, 383-388.
Delfgaauw, J., Dur, R., Non, A. & Verbeck, W. (2014). Dynamic incentive effects of relative
performance pay: A field experiment. Labour Economics, 28, 1-13.
Delfgaauw, J., Dur, R., Sol, J. & Verbeke, W. (2013). Tournament Incentives in the Field:
Gender Differences in the Workplace. Journal of Labor Economics, 31(2), 305-326.
Esteban, J., & Ray, D. (2001). Collective action and the group size paradox. American
Political Science Review, 95(3), 663-672.
European Foundation for the Improvement of Working and Living Conditions (2007),
Teamwork and high performance work organisation.
21
Hamilton, B., Nickerson, J. & Owan, H. (2003). Team Incentives and Worker Heterogeneity:
An Empirical Analysis of the Impact of Teams on Productivity and Participation. Journal of
Political Economy, 111(3), 465-497.
Holmström, B. (1982). Moral hazard in teams. The Bell Journal of Economics, 13(2), 324-
340.
Ichniowski, C., Shaw, K. & Prennushi, G. (1997). The Effects of Human Resource Management Practices on Productivity. The American Economic Review, 87(3), 291-313. Iranzo, S., Schivardi, F., & Tosetti, E. (2008). Skill Dispersion and Firm Productivity: An Analysis with Employer-Employee Matched Data. Journal of Labor Economics, 26(2), 247-285.
Itoh, H. (1991). Incentives to help in multi-agent situations. Econometrica, 59(3), 611-636.
Lazear, E. P. (2000). Performance Pay and Productivity. The American Economic Review,
90(5), 1346-1361.
List, J. & Rasul, I. (2011). Field Experiments in Labor Economics. Handbook of Labor
Economics, 4A, 103-228.
Shearer, B. (2004). Piece rates, fixed wages and incentives: Evidence from a field experiment.
Review of Economic Studies, 71(2), 513-534.
Stigler, S. (1997). Regression towards the mean, historically considered. Statistical Methods
in Medical Research, 6(2), 103-114.
22
Appendix I: Tables and Figures
Figure 1: The Team Bonus
Figure 2: A map of shops by treatment (red) and control group (blue)
23
Figure 3: Scatter plot, year on year sales growth on log sales, April, May and June 2013
Figure 4: Kernel distribution sales growth treatment versus control group
24
Table 1: Characteristics of the control and treatment shops before the treatment
Control Treatment t-test
(n = 96) (n = 97) p-value
Mean monthly sales (SD) 27453 28194
(11481) (14542)
Mean monthly sales (in logs, SD) 10.14 10.15
(0.39) (0.41)
Unsold goods as % of sales (SD) 16.16 (7.0) 15.54 (6.9) 0.331
Mean number of customer-visits (SD) 10028 (3921) 10131 (4018) 0.856
Mean monthly quit rate (SD) 1.9% (4.1%) 1.8% (4.1%)
Frequency of achieving the sales target 35.8% 35.2% 0.860
Mean mystery shopping score 2013 (SD) 96.1% 95.5%
Mean mystery shopping score 2014 (SD) 32.2 32.2
Big town 37.6% 33.6%
Medium/small town 26.0% 29.6%
Village 36.4% 36.7%
Mean age, years 39.8 (6.4) 40.9 (6.3)
Share of females 94.9% 93.0%
Share of full-time employees 71.8% 64.8%
Total number of sales agents 552 580
Mean number of agents per shop (SD) 7.4 (3.2) 7.4 (3.2)
Mean age, years 39.5 (6.1) 39.9 (6.0)
Share of females 93.1 92.4
Share of employees with a permanent contract 66.6% 67.9%
Share of full-time employees 9.7% 10.4%
Share of part-time employees 56.7% 59.7%
Share of employees with a "mini-employment" contract33.6% 29.9%
Share of unskilled workers 77.5% 72.3%
Mean commitment score (SD) 4.50 (1.55) 4.42 (1.69) 0.523
Mean work satisfaction score (SD) 4.45 (1.51) 4.33 (1.57) 0.422
Mean overall satisfaction score (SD) 4.98 (1.63) 4.90 (1.70) 0.548
Standard deviations are in parentheses. Column 3 reports the p-values of the two-sided t-test of
equality of the means for a selection of variables. "Big town", "medium/small town" and "village"
refer to municipalities with more than 90,000; 5,000 to 60,000; and fewer than 5,000 inhabitants,
respectively. Panels D and E are based on the personnel records from the firm as of July 1 2014,
excluding apprentices and interns (18 in the control and 11 in the treatment group). Panel F reports
the means of the commitment, work satisfaction and overall satisfaction scores constructed
according to Allen and Meyer (1990) from the employee survey administered in March 2014. In
total, 563 employees in the control, and 580 employees in the treatment group participated in the
survey (response rate 79.5%).
Panel E: Characteristics of sales agents
Panel F: Employee attitudes
Panel A: Quantitative performance indicators
Panel D: Characteristics of shop managers
Panel B: Qualitative performance indicators
0.695
0.846
Panel C: Shop location
25
Table 2: Characteristics of the control and treatment shops in the treatment period
(April - June 2014)
Control Treatment Diff-in-Diff t-test
(n = 96) (n = 97) p-value
Mean monthly sales (SD) 25376 26995(10708) (15036)
Mean monthly sales (in logs, SD) 10.06 10.10(0.40) (0.42)
Unsold goods as % of sales (SD) 22.88 (9.8) 22.35 (13.3) 0.940
Mean number of customer-visits (SD) 9115 (3582) 9465 (3790) 0.062
Mean monthly quit rate (SD) 1.42% (4.89) 1.69% (5.64) 0.336
Sales targets achieved 44.8% 49.1% 0.442
Mean mystery shopping score 2014 (SD) 32.4% (1.0%) 32.2% (1.2%) 0.295
Mean commitment score (SD) 4.20 (1.28) 4.24 (1.35) 0.468
Mean work satisfaction score (SD) 4.39 (1.34) 4.48 (1.20) 0.245
Mean overall satisfaction score (SD) 3.59 (1.12) 3.72 (1.02) 0.162
Panel F: Employee attitudes
Standard deviations are in parentheses. Column 3 reports the p-values of the two-sided significance test
for the difference-in-difference estimate of the treatment effect. The second employee survey was
administered in May 2014 with a response rate of 76%.
Panel A: Quantitative performance indicators
0.061
0.034
Panel B: Qualitative performance indicators
26
Table 3: Treatment effect estimates
Specification (1) (2) (3) (4) (5) (6) (7)
Treatment effect 0.032 0.033 0.030 0.030 0.032 0.026 0.027
(.013) (.011) (.014) (.014) (.014) (.014) (.014)
Shop fixed effects yes yes yes no no no no
Month dummy vars yes yes no no no no no
Other controls yes yes no yes yes yes yes
Observations 4916 4904 386 193 577 561 577
The table shows the difference-in-difference treatment effect estimates based on several regression specifications with the log sales as the
dependent variable. In all specifications the unit of observation is individual shop. In specification 1, we regress monthly sales from January
2012 until June 2014 on the "treatment group" and "after treatment" dummies and their cross-product. Specification 2 is the same but omits the
outliers, defined as year-on-year sales change exceeding 30% (roughly the top and bottom 1% of the sales growth distribution). The reasons for
such substantial increases or decreases in sales are construction sites close to the bakeries, competitors who enter or leave the market,
temporary closures of shops because of renovations or sunny weather, which affects sales in bakeries located in shopping centers. Specification
3 is the same as 1, except that we use log average sales over the periods before and after the treatment (hence two observations per shop).
Specification 4 includes past sales as an additional control, hence one observation per shop. In specification 5, we regress the log monthly sales
in April, May and June 2014 (the treatment period) on the treatment dummy and the baseline sales in the respective shop, defined as the log
average sales over the pre-treatment period. In specification 6, we regress the log monthly sales in the treatment period on the treatment dummy
and the log sales in the respective months in 2013. Specification 7 is the same as 5 except that we use the log average sales in January-Mach
2014 as the baseline. Standard errors are clustered by shop. Cluster-bootstrapped standard errors (available on request) are similar in
magnitude.
27
Table 4: Treatment effect by month
Specifications April 2014 May 2014 June 2014
Treatment effect 0.029 0.037 0.029
(.011) (.022) (.014)
Observations 4532 4532 4532
The regression specification is the same as spec. 1 in Table 3: The log
monthly sales regressed on the "treatment group" and "after treatment"
dummies, their cross-product, and controls.
28
Table 5: Treatment effect by shop location
April 2014 May 2014 June 2014 Overall
Shops located in big towns 0.059 0.055 0.049 0.055
(.019) (.051) (.024) (.025)
Shops in midsize towns 0.023 0.049 0.045 0.038
(.018) (.023) (.026) (.02)
Shops in villages 0.004 0.011 -0.001 0.005
(.019) (.02) (.022) (.019)
The regression specification is the same as spec. 1 in Table 3. The cells in the table give estimated
treatment effect in a given month and location. For example, 0.065 is the treatment effect in April 2014
in shops located in big towns. Standard errors are clustered by shop.
29
Table 6: Treatment effect by quartile of shop size (number of workers)
Whole sample Big towns Elsewhere
Quartile 1 0.001 0.016 -0.006
(.024) (.041) (.029)
Quartile 2 0.022 0.005 0.027
(.022) (.038) (.027)
Quartile 3 0.041 0.046 0.056
(.027) (.054) (.023)
Quartile 4 0.059 0.125 -0.029
(.025) (.043) (.024)
Observations 4916 1760 3156
Shop size is defined as the number of workers employed in a shop
excluding those on a mini job. Quartile of shop size is defined separately
for each subsample (shops located in big towns tend to employ more
workers). Standard errors are clustered by shop.
30
Table 7: Treatment effect by shop-average employee age and tenure in January-March
2014 (the quarter before the treatment)
Below median Above median Below median Above median Below median Above median
Treatment effect 0.043 0.021 0.068 0.033 0.022 0.013
(.019) (.017) (.043) (.032) (.018) (.022)
Observations 2446 2470 873 887 1599 1557
Below median Above median Below median Above median Below median Above median
Treatment effect 0.061 0.001 0.063 0.043 0.034 0.003
(.019) (.017) (.036) (.031) (.024) (.016)
Observations 2453 2463 894 866 1601 1555
Whole Sample Big towns Elsewhere
The samples are split into below and above the median age/tenure of the workforce excluding workers employed in a mini
job. Standard errors are clustered by shop.
Tenure
Whole Sample Big towns Elsewhere
Age
31
Table 8: Treatment effect by the average share of mini-
job employees
Whole sample Big towns Elsewhere
Quartile 1 (<0.06) 0.071 0.052 0.079
(.033) (.076) (.036)
Quartile 2 (0.06 - 0.11) 0.050 0.098 0.005
(.026) (.042) (.031)
Quartile 3 (0.11-0.16) 0.003 0.053 -0.011
(.019) (.039) (.02)
Quartile 4 (>0.16) -0.003 -0.019 0.002
(.021) (.035) (.027)
Observations 4916 1760 3156
The share of mini-job workers is defined as the ratio of the hours worked by
these workers to the total hours worked. Quartiles of the share of mini-job
workers are very similar for every location, and so are defined on the whole
sample. Standard errors are clustered by shop.
32
Table 9: Treatment effect by pre-treatment deviation of sales from the target
Distance measure: pre-treatment average sales/target difference
Quartile 1 (<-8%) Quartile 2 (-8% to -4.5%) Quartile 3 (-4.5% to 0%) Quartile 4 (>0%)
Treatment
effect 0.046 0.036 0.047 0.003
(.026) (.028) (.027) (.017)
Observations 1202 1242 1246 1226
Distance measure: pre-treatment frequency of achieving the target
Quartile 1
(<16%) Quartile 2 (16% to 30%) Quartile 3 (30% to 50%)
Quartile 4
(>50%)
Treatment
effect 0.052 0.048 0.026 -0.009
(.022) (.025) (.03) (.016)
Observations 1256 1255 1228 1117
The regression specification is the same as spec. 1 in Table 3. Standard errors are clustered by shop.
33
Table 10: Treatment effect on the number of customer visits and sales per
customer visit
All shops Big towns Other locations Village
Treatment effect on customer visits 0.027 0.046 0.032 0.006
(.011) (.02) (.019) (.017)
Treatment effect on sales per visit 0.004 0.008 0.004 0.000
(.007) (.018) (.005) (.007)
34
Appendix II
Information leaflet
<LOGO OF THE BAKERY>
AN ALLE VOLL- UND TEILZEITKRÄFTE: VERDIENEN SIE SICH IHREN TEAM-BONUS
In den Monaten April, Mai und Juni 2014 erhält das Team Ihrer Filiale einen Team-Bonus bei Erreichung oder Übererfüllung der Umsatzziele. So sieht das Bonus-Programm für Voll- und Teilzeitkräfte aus:
Bei Erreichung oder Übererfüllung von bis zu 1%, erhält das Filial-
Team einen Bonus von 100€ für den entsprechenden Monat.
Bei 1% bis 2% über dem Umsatzziel erhält das Filial-Team einen
Bonus von 150€.
Bei 2% bis 3% beträgt der Team-Bonus 200€.
Bei 3% bis 4% beträgt der Team-Bonus 250€.
Bei 4% oder mehr gibt es einen Team-Bonus von 300€.
Jedes Filial-Team kann also im Quartal einen Bonus von bis zu 900€ erreichen! Bitte beachten Sie:
Details zur Aufteilung unter den Team-Mitgliedern und Fehlzeiten
finden Sie im Infobrief.
Leider können wir diese Regelung aus steuerrechtlichen Gründen
nicht für geringfügig Beschäftigte anwenden.
Bei Fragen wenden Sie sich bitte an Ihre Bezirksleiter/innen, die Ihnen gerne weiterhelfen und ihnen regelmäßig mitteilen werden, ob sie Ihre Umsatzziele erreicht haben.