Professional Documents
Culture Documents
on an Assembly Line
Kenneth L. Schultz
S. C. Johnson Graduate School of Management
Cornell University
Tobias Schoenherr
College of Business, Eastern Michigan University
David Nembhard
Harold and Inge Marcus Department of Industrial and Manufacturing Engineering,
Penn State University
The authors wish to thank Richard Garrett, Chris Albright and Srinagesh Gavirneni for
help with this research. Dr. Nembhard’s work on this project was supported in part by a
grant from the National Science Foundation SES0217666.
Deleted: 2/12/2006
5/9/2006
Worker Motivation on an Assembly Line
Abstract
Models and understanding of line design depend on accurate assessments of the effects of
design parameters on human actions. Equity theory predicts that workers would react to
the speed of people around them. Experimental work has repeatedly failed to find this
effect with additive tasks. We use archival data from a manufacturing line to show that
workers do react to the speed of their coworkers but that individual reactions vary widely.
This implies that models using randomly drawing times from the set of worker nominal
processing times will overestimate the size of within team variance while
underestimating the range of between team variance. Because workers vary both in speed
and an heuristic rearrangement of workers and develop a modified heuristic that performs
well.
2
Worker Motivation on an Assembly Line
1. Introduction
Researchers in the field of Operations Management have studied questions of line
dynamics and concluded that factors such as line length, variability and buffer size all
affect the productivity of a line. This work has relied upon mathematical modeling and
generally uses the assumption that processing times of successive workers are
independent of each other (Dallery and Gershwin, 1992). Other work has demonstrated
that processing times may, in fact, not be independent of the state of the system (Edie,
1954; Doerr, Mitchell, Klastorin and Brown, 1996; Schultz, Juran, Boudreau, McClain
processes. New designs are based on the models and the understanding of line dynamics
developed from these models. Currently, most models assume the behavior of individuals
is independent of the major design factors. The motivation of the individuals on the line
is considered separately from the design of the process. While few would argue that there
is no connection between design and motivation, we currently do not know enough about
the interactions to inform the design choices. Scientific work at this intersection of
behavioral and OM theory is needed. Wallace Hopp (2004, p. 5) in the 50th Anniversary
Article for Management Science writes “understanding a manufacturing plant does not
just require a theory of human motivation and a theory of material flow; it also requires a
means for describing the interaction between the two.” It is exactly this interaction that
we explore in this paper; the interaction between line design and human motivation. If we
can improve our understanding of how these design parameters affect the motivation of
the workers we can construct better models and make better choices.
3
Worker Motivation on an Assembly Line
Equity Theory predicts that workers will be motivated to more closely match their
work peers. Experimental work with conjunctive tasks has found that the effort of the
slower workers will increase while that of the faster workers will decrease. Conjunctive
tasks occur when the performance of the group depends on the least performing worker.
These are contrasted with additive tasks where group performance is based on the sum of
the performance of all workers. For additive tasks, experiments seem to indicate that
workers do not react to the speed of their coworkers. Almost all of this work has been
conducted in experimental labs and there is little confirmation from industrial settings.
If equity theory is a factor, and workers are motivated to adjust their speed to the
speed of their coworkers, there are implications for the modeling of work lines. Within
team variation will be reduced as the faster workers in each group slow down and the
slower workers speed up. Additionally, if the theory holds, adjustments can be made in
the work organization and workplace setting to improve the motivation of workers.
We use data from 148 workers collected over six months. Tasks were arranged in
parallel on the bottleneck station of a radio assembly line, an additive work organization.
Our analysis supports the predictions of Equity Theory. Workers who are slower than the
group tend to speed up while those who are faster tend to slow down. However, the
differences in human reactions are quite varied. Since there are two parameters to
balance, work speed and reaction, a simple heuristic for arranging workers in is not
We test that model using data collected from the workers on the radio line. The heuristic
performs well compared to the optimal in most cases and presents a measure for an upper
limit of improvement of 4.8%. Analysis of the data indicates a simple modification of the
4
Worker Motivation on an Assembly Line
heuristic to improve performance when the heuristic does not perform well compared to
the optimum.
The paper is organized into 6 sections. Section 2 reviews the literature to derive a
testable hypothesis. The data collection is described in section 3. The analysis of that data
is presented in section 4. Section 5 presents the results while section 6 uses models to
motivational losses or gains from the development of the group itself. Some of this
research has focused on the task organization. Task organization can be classified as
being either conjunctive or additive. A conjunctive task organization exists when the
outcome is the result of the least performing member. This is similar to the team time
trial in the Tour de France where the official time of the first six riders on the team is the
time of the sixth rider. An additive task organization exists when the outcome is the sum
(or average) of all the members of the group. This is similar to a relay race where the
time of the team is the sum of the times of the individual racers.
shown that people will work less hard in group situations than in individual ones.
Williams, Harkens and Latané (1979) refer to “Social Loafing”. Social Loafing is the
tendency for individuals to give less effort when working in a group then they do when
working alone.
5
Worker Motivation on an Assembly Line
(1926, 1927 as referenced in Hertel, Kerr and Messé 2000) experimentally demonstrates
situations where dyads with conjunctive tasks perform better than the least performing
individual. Hertel, Kerr and Messé (2000) further examine the Köhler effect and
hypothesize that the basic result is a motivational gain by the weaker individual in the
dyad and is due to feeling particularly indispensable to the group’s success. Williams and
Karau (1991) use Social Compensation theory to explain motivation gain of the stronger
members in situations where there is an expectation that coworkers will perform poorly
and the task is highly meaningful. Thomas (1957) discusses the role of interdependence
group relative to their performance as individuals. They do not explain how different
people will react differently to the same group situation. We wish to explore the effects of
coworker speed on individual motivation. We want to explain why people will work at
different speeds in the same task situation with different group members.
processing line. The line can work no faster than the bottleneck. Doerr et al. (1996)
studied such a line in a fish packing plant and it appeared that workers in the low
inventory line (higher conjunctive task organization) would be idle more often. They
would run out of work more often. However, productivity remained the same as in the
high-inventory case (higher additive task organization). It appeared from the data that the
6
Worker Motivation on an Assembly Line
The authors found that in the low-inventory (conjunctive) treatment, slower workers
worked harder than slower workers did in the high-inventory (additive) treatment.
However, they only found significant effects in the case of the slower workers. Faster
workers did not appear to change their work pace, or, alternatively, did not slow down at
the level sufficient for the statistics to discriminate. Schultz, Juran and Boudreau (1999)
showed that the effect was due to the interdependent nature of the task and the feedback
Hertel et al. (2000) also analyzed how different people reacted differently in a
conjunctive task situation. Using a subjective measure of effort the authors found that the
effort of the weaker member increased as the difference between the abilities of the two
members of the team increased. They also found that the stronger member reported less
Doerr et al. (1996), Schultz et al. (1998) and Hertel et al. (2000) all found
increasing effort by the slowest or least capable team members. Hertel at el. also found
decreasing effort by the stronger members. This can be referred to as a regression to the
mean effect. The three studies described were experiments. While Doerr et al. did their
work in an actual fish packing plant, the work of Schultz et al. and Hertel et al. were
laboratory experiments with student subjects. Additionally, the tasks discussed so far
A regression to the mean effect has not been found in situations involving an
additive task organization. Edie (1954) studied the processing rates at tollbooths on the
George Washington Bridge, an additive group task. He found that all workers sped up
7
Worker Motivation on an Assembly Line
with an increase in traffic intensity. The high inventory treatments in Doerr et al. (1996)
and Schultz et al. (1998) were tasks with an additive task organization. No worker ever
ran out of work to do or a place to put their output (no starving, no blocking) making
rewards were based on the combined output of all workers. No regression to the mean
effect was noted. Doerr et al. did not have the opportunity to compare task times by faster
and slower workers. Checks made by Schultz et al. indicated that task times in the high
inventory situation were independent of each other. Hertel et al. (2000) also did an
additive task treatment and found no overall motivational gain in the additive task.
the faster workers was detected. The conclusion would appear to be that workers in
additive tasks do not adjust their speed based on the speed of their coworkers.
Equity Theory suggests that there will be a regression towards the mean, even in
additive tasks (Adams, 1965). According to Equity Theory, people desire the ratio of
their outcomes and investments to be roughly equivalent to the ratio for some other
referent person or group. If ratios are not equal, and the difference is beyond some
individual threshold level, emotional responses occur and the desire to restore equity
manifests itself. In order to restore equity, people have a number of choices. They can
change their outcomes, change their inputs, distort the value of either their outcomes or
inputs, leave the field, change the outcomes of inputs of another, or change the referent
person or group. In a work setting, one method to restore balance would be to change
inputs by changing how hard the person works. For those who adopt this mechanism, it
would lead to correlations in processing times as workers adjust their effort to the
8
Worker Motivation on an Assembly Line
perceived effort of those around them. They would work faster when their coworkers are
faster, and slower when their coworkers are slower. In other words, Equity Theory would
predict a regression toward the mean, similar to that found in conjunctive tasks, but
Our archival data is from a situation of an additive task organization, and allows
us to check for correlations among the processing times of workers as their workgroups
change. The findings of Hertel et al. (2000) predict that, with additive tasks, workers
would not change their work rates in response to the speed of others. This would lead to
be due to a lack of a sufficient sample size to overcome variability. It is also possible that
the experimental situation in the work of Hertel et al. did not allow for the development
of the types of social relationships that could lead to a regression toward the mean effect.
In additive interdependent work situations, workers will adjust their speed toward
There are other effects that can lead to correlated processing times, which must be
controlled for in our analysis. One such, learning, induces correlation as all workers
function (e.g., Argote, 1999), indicates that the underlying workforce progressively
contemporaneously, will tend to have correlated processing times (see e.g., Nembhard
and Uzumeri, 2000). Correlation of processing times can also be induced by common
9
Worker Motivation on an Assembly Line
Psychology provides a large body of literature substantiating both of these effects (see
3. Data
identical production lines of a major automobile manufacturer. These data were examined
by Shafer, Nembhard and Uzumeri (2001) from the perspective of measuring individual
unrelated to learning, we use the residuals from this learning curve model. While we
support the findings of Shafer et al., our goal is not to pick a learning model with
universal applicability, but to develop the best model for removing learning effects from
One hundred forty eight workers performed the operation during six months.
Typically, each employee in this unionized plant had been with the company for several
years performing a variety of other tasks. During the data collection period, employees
worked on the three lines for a total of 25.2 days on average and produced 2,389 units.
Task times were recorded for each unit, workstation, and worker on the line. Most
of the work on the assembly line involved machine paced tasks. The primary human
paced work was the functional and cosmetic inspections toward the end of the line. This
was also the bottleneck operation that effectively determined the output of the line. Up to
eight employees worked in parallel on the same task, picking radios from a common
10
Worker Motivation on an Assembly Line
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
Place Figure 1 Approximately Here
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
The task involved 130 distinct evaluation criteria and procedural steps based on a
QS9000 documented process (QS9000 is similar to ISO9001 but adapted for use in the
American automobile industry). The average processing time per item was 3.2 minutes.
A typical process commenced when the operator took a unit from a conveyor system and
began a series of cosmetic inspection and lighting tests (see figure 2). She or he then
plugged the unit into a fixture to initiate a series of electronics tests. Additional manual
and cosmetic examination were performed on another unit while the first unit was in the
fixture. When complete, the worker removed the unit from the fixture and repeated the
process with the second and third units. Since the manual tasks took longer than the
electronics testing, the consecutive time stamps recorded by the electronic tester,
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
Place Figure 2 Approximately Here
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
The original data set consisted of 176,000 items in 11,066 records on 148
workers. We define an episode as one work group coming together for a period of time.
Each record then is the average processing time for a particular worker during a particular
episode. If the composition of the group changes a new episode is begun. The
composition of the group may change if someone goes on break, is moved to a different
task in the plant, to one of the other lines, or as each worker finishes their last piece for
the day. All episodes are less than 4 hours in duration. If, at some later point, the same
11
Worker Motivation on an Assembly Line
group of people recombined, a new episode is recorded, since, due to learning, workers
would be facing a new set of processing times and a different set of motivational
parameters, we treat this as a new episode. Three hundred twelve records, consisting of
only one person working, were discarded. One hundred twelve records consisting of five
or fewer units produced were also deleted because of potential estimation error. We also
excluded any worker who was in the dataset for less than one month. Furthermore, we
discarded all episodes for which there was only one record available, since computation
of episode variables would not have been possible. This led to the removal of 1,561
additional records. We were left with a total of 9,081 records for 78 individuals in a total
of 1,445 episodes.
additive task organization. Work is done in parallel. Since it is the bottleneck workstation
worker production is not physically dependent on each other. Further, processing times
are not confounded with idle time. The workers were starved infrequently. The process
under consideration was the system bottleneck, the only labor dominated process, and by
far the most expensive process on the line, both in terms of labor and equipment. Typical
line operations entailed running the upstream automated equipment sporadically until the
pre-inspection buffer was full. Once the process input buffer started to draw down,
operators would restart the upstream process. It was not allowed to starve the workers.
In the infrequent cases when there was a problem restarting the upstream process,
management would move workers to a different line (~0.5 - 2 times/per day). The data
were not confounded with starving. The time data we use were generated by the
electronic testing which was done after the cosmetic inspection. When a worker ran out
12
Worker Motivation on an Assembly Line
of radios to process, they processed their last unit and, when the electronic test was
complete, they would fail to place a new unit in the tester. This showed up in the data as
the worker leaving the group. The team would then go from 7 to 6 workers and a new
episode would begin. However, as soon as the next worker finished, they too would clock
out of the group and, since the episode was too short, these data would not be included in
the analysis. Therefore, on the rare occasions where starving did happen, data were not
4. Analysis
speeds. In the model we control for learning, the day of the week and the time of day. In
order to separate the correlation due to coworker speed from the correlation due to
learning, we fit a learning/forgetting curve to each individual using the data and
methodology described by Shafer et al. (2001). The residuals from this process were then
used to represent how workers’ processing rates varied over time. We refer to the
processing times corrected for learning and forgetting as the nominal processing times.
While the same people may recombine in more than one episode, due to differences in
learning, their speeds will be different each time. Each time a group meets they will each
have, and will each be reacting to, a different set of relative speeds.
For the dependent variable we use the processing time p of worker i during an
episode j, pij, and subtract the nominal processing time for that worker (i) during that
episode (j), fij. We use ‘nominal processing time’ in this case to mean the processing time
predicted by the learning model. The resulting residuals contain variation in that worker’s
13
Worker Motivation on an Assembly Line
processing time that is not explained by learning. It is the variance in this set of residuals
We calculated the processing time of the coworkers at any given time by taking
the average of the actual times of every worker in episode j, excluding the individual
worker i under consideration. We denote this value as cij . A slow worker for this episode
would be a worker whose pace is slower than this average. A fast worker would have a
pace faster than this average. Therefore, the independent variable is the individual
worker’s nominal time, fij, subtracted from the average time of his or her coworkers, cij .
In other words, we are predicting that workers react to the difference in speed between
The formulation of the regression model follows. Potential effects of the time of
day or day of week on worker i’s processing time are incorporated in this model in the
last two terms of equation (1). We refer to this as the primary model.
where
fij = nominal processing time for worker i at the time of episode j, adjusted for
learning
shj = hour block variable for episode j, (one block h equals two hours)
tdj = day of week block variable for episode j (one block d equals one day)
14
Worker Motivation on an Assembly Line
variance. We replace cij , the average of the actual processing times of all coworkers of
worker i in episode j, with ĉij , the average of the nominal processing times of all
predictions drawn from the entire data set and are independent of episodic variance. On
average, we assume that the workers have been in the plant long enough to have some
level of knowledge about their coworkers. According to this model workers would react
to how fast they expect their coworkers to work. We refer to this as the secondary model,
since equity theory predicts that workers react to actual and not expected speed, which is
5. Results
The results of the primary model are shown in Table 1. The overall correlation
coefficient is 0.407. Roughly, this means that the average person will adjust their time by
41% of how much faster, or slower, their nominal time is in relation to the coworkers
with whom they are working during that period. The correlation is strongly significant
with the t statistic for the coefficient at 38.61. However, the model has a relatively low R2
of 0.147. While the model is significant, it explains little of the total variation. We
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
Place Table 1 Approximately Here
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
15
Worker Motivation on an Assembly Line
Partial F-tests show that adding hours of the day, (F11,9053 = 4.22, p = 0.000),
contributes to the model but that adding days of the week, (F6,9058 = 0.524, p = 0.790),
does not.
While the 0.407 correlation provides an estimate of how responsive workers are
to the speed of their coworkers, it is only an average. It would also be instructive to look
at the range of worker reactions. How great is the difference among workers? Do some
workers have negative slopes, reacting contrary to their coworkers? Do some overreact
To look at these questions we ran 78 regression models, one for each worker. The
number of data points was equal to the number of episodes in which each worker was
involved. Due to the results of the partial F-test we included coefficients for hours of the
day but not for days of the week. The results for the 78 workers are shown in Table 2,
organized in order of their coefficients from highest to lowest. All but four have positive
coefficients. The four negative coefficients are not significant at p-values less than 0.39.
Of the 74 workers with positive coefficients, 66 (85% of the total) are significant at 0.1.
Of those, 60 (77%) have coefficients with p-values less than 0.05, 52 (67%) are at 0.01 or
less, and 49 (63%) are 0.001 or less. On the high end, workers will react by changing
their times approximately 127% of the difference between their nominal times and the
times of their coworkers. On the low end, workers appear not to react at all. Several
standard errors of the slopes are over 0.1 so care needs to be taken in interpreting the ends
of the range. In particular, the workers at the low end of the range are generally
characterized by few observations. However, even there, workers 24 and 113 have 70
observations or more and still do not have slopes that are statistically significant. On the
16
Worker Motivation on an Assembly Line
high end, the top three workers have few observations and large standard errors. Thus, a
more conservative estimate of the strongest reaction among workers may be the 0.869
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
Place Table 2 Approximately Here
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
the data for four different workers. In Table 2, the results corresponding to these four
workers are highlighted. The first one, worker 127, was chosen to illustrate a large and
significant reaction. Worker 113, on the other hand, shows no discernable reaction at all.
The third plot, worker 162, has a negative slope that is not statistically significant.
Worker 111 shows an intermediate case. These plots demonstrate the large range of
individual reactions.
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
Place Figure 3 Approximately Here
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
It is possible that there is some other, unknown, factor that would explain the
episodic variance in the secondary model. According to this model, workers react not to
how co-workers are actually performing but to how they are expected to perform. This
approach eliminates any correlation due to common timing with the dependent variable.
The results are shown in Table 3. The correlation is still strongly significant. The
beta coefficient is reduced to 0.190 and the R2 is reduced to 0.017. Workers do react but,
possibly because we are using a less direct measure of coworker speed, the secondary
17
Worker Motivation on an Assembly Line
model explains even less of the total variability. The correlation persists despite removing
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
Place Table 3 Approximately Here
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
6. Analytical Extensions
If workers are reacting to the difference between their nominal time and the times
of their coworkers, we can improve the performance of the line by limiting the referents
available to workers, or encouraging them to pick referents who are faster than they are.
As the system stands now, workers are as likely to speed up as they are to slow down and
the effect on the average speeds of the groups is likely to balance out. In this section, we
explore the managerial policy of restricting the sight lines of workers so they can only see
developing an optimal policy for arranging workers, and compare it to a fastest to slowest
(FTS) heuristic.
Since workers are different in both speed and reaction, a simple fastest to slowest
heuristic is not optimal. Assume we have three workers with processing times of 50, 100
and 101 time units per operation respectively. Their reaction coefficients are 50%, 0%
and 50% respectively. We arrange them so they can only see, and react to, the workers in
front of them. We place worker 1 in the lead; she does not react because she doesn’t see
anyone else. Worker 2 never reacts because she has a reaction of 0%. If we place worker
3 in the third slot, he reacts to an average time of 75, the average of workers 1 and 2, and
he will speed up by 13 time units ( 0.5 * [101 – (50 + 100)/2 ] ). If we place worker 3 in
the second slot, he reacts to a speed of 50, the time of the fastest worker, and speeds up
18
Worker Motivation on an Assembly Line
by 25.5 time units. While this is an exaggerated example to illustrate the point, it
demonstrates why faster workers with smaller reaction coefficients might be placed
behind their slower co-workers. We will compare the optimal to the heuristic to
Let I equal the number of workers and S equal the number of workstations, where
for simplicity, I = S, and the reaction coefficient for worker i is β1i. Also, letting
observe and react to upstream workers, we can construct an expression for hs, the average
∑∑ X pq f p ,
1
hs = (4)
s −1 q=1 p=1
where s = 1 represents the workstation at the front of the line, and s = S the workstation at
Further, the average production rate of workers upstream of worker i, denoted gi,
is given by
S
gi = ∑ Xir hr . (5)
r=1
Thus,
S ⎡ 1 r−1 I ⎤
gi = ∑ Xir ⎢ ∑ ∑ X pq f p ⎥ . (6)
r=1 ⎢⎣ r −1 q=1 p=1 ⎥⎦
19
Worker Motivation on an Assembly Line
we write the optimization model for maximizing worker output by reordering the workers
∑X is = 1 , ∀s ∈ {1..S}
i =1
S
∑X is = 1 , ∀i ∈ {1..I}
s =1
Our FTS heuristic is to order the workers starting at the front of the line from
fastest to slowest such that slower workers can only see the faster workers in front of
them, and not the reverse. In this arrangement, no worker would go slower than their
nominal time. The fastest worker, and any worker with reaction coefficient, β1i = 0,
would work at their nominal speed. All other workers would speed up.
We use Monte Carlo simulation to compare the quality of the heuristic ordering to
the optimal orderings, based on random selections of workers from the automobile plant
data. We restrict our simulation to the 53 workers for whom we have more than 100 data
points each. We choose 5 time periods to cover the full 6 month period and use learning
curves to estimate the nominal times for those workers. Individual worker reaction
coefficients are the estimates obtained from the results of the individual regression
analyses. We used all reaction coefficients, even when they are not statistically different
from zero, as the best estimates of the true coefficients. We generated 10,000 independent
20
Worker Motivation on an Assembly Line
sample trials of workgroups consisting of seven individuals for each time period and
compared the FTS heuristic to the optimal ordering for each workgroup.
Figure 4 illustrates the summarized results of this comparison, where FTS was
exactly optimal in 21% of the trials. The worst-case sample for the 50,000 trials was
1.75% from optimal, and the average was a 0.1% deviation from optimal. We conclude
that the fastest to slowest heuristic is sufficient in most cases. We note that the fastest
worker will always be assigned to the first workstation, since one would never want to
lower this person’s speed. However, upon examination of the trials for which FTS is sub-
optimal, we observe that a slower worker may be optimally assigned ahead of a faster
worker when their reaction coefficient is significantly higher than that of the faster
worker. This appears when the reaction coefficient is roughly 25% larger for the slower
worker.
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
Place Figure 4 Approximately Here
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
Turning our attention to the differences in the productive output of the FTS
speeds, f(W), and order them from fastest to slowest such that W1 is the fastest worker in
the sample and WN is the slowest. The decrease in the average times of a team is then
N
β n ⎡ n −1 ⎤
∑ (n − 1) ⎢⎣∑ (W − Wn )⎥
1
i (9)
N n=2 i =1 ⎦
independently and identically distributed (iid), as would be the case if f(W) were a
uniform distribution, then we can show that the increase in the speed of a team is a
function of the number of workers on a team and the range, R, of the sample. Let R be the
21
Worker Motivation on an Assembly Line
range of the sample and β be the average β1i, assuming β1i (reaction) and f(W)
(processing time) are independent. Then the average decrease in processing time of the
Rβ (N (n +1) − 2)
. (10)
4 N(N −1)
The results of the simulation show an average increased productivity of 4.8% over
that of a random ordering with respect to worker speed. The paired-sample t-test is
significant at the 0.000 level, and the standard deviation of the improvement is 3.03%.
Additionally, there is a decrease in the standard deviation of within team processing times
of 0.22 units. When comparing the FTS policy to the worst-case orderings, the worst-case
turned out to be an average of 10.5% below the productivity of the FTS policy. This is
the same as the improvement over an arrangement where every worker can see every
other worker.
on a parallel task. On the assembly line from which the data was drawn, every worker
could see every other worker. We also considered a fastest to slowest heuristic, where
workers could only see co-workers who are faster than they are. This heuristic uses
considered an optimal model which considers both speed and reactions. The heuristic
performs 4.8% better than the original configuration. The optimal model performs on
average only 0.1% better than the heuristic. An analysis of the cases where the optimal
model performs significantly better than the heuristic suggests a modification to the strict
fastest to slowest heuristic. When a slower worker has a reaction coefficient that is 25
22
Worker Motivation on an Assembly Line
percentage points higher than the faster worker in front, move the slower worker to a
the speed of his or her coworkers. We have done so using data from an automobile plant.
We have demonstrated that all correlation is not due to a common day-of-week effect nor
is it due to a common time-of-day effect. We have also shown that the correlation persists
when any common source of variance related to time is removed, although this decreases
the discrimination of the test. We also believe that there is no overall effect on average
processing times since workers are as likely to speed up as they are to slow down.
decrease the gap between their work pace and the work pace of others around them. This
is consistent with Equity Theory and similar findings related to conjunctive tasks. Our
results are contrary to Social Compensation Theory. The findings are also in contrast to
the conclusions of Hertel et al. (2000) for other, additive, tasks. This difference might be
among the workers in the plant when compared to students in an experiment. However, it
could also be due to other differences between experimental and archival data, or to some
unexplored difference in the nature of the tasks. Since the effects seem more difficult to
establish in additive than in conjunctive tasks, we suggest that the nature of the task
This effect is important to the modeling of work lines. A model which assumes
23
Worker Motivation on an Assembly Line
estimate of the average processing time of the group. However, estimates of the
difference between members of that group will be underestimated. Slower workers will
be working faster and faster workers will be working slower than previously considered.
We have proposed reorganizing workstations so that workers can only see those
who are faster. Using the data available, we demonstrated how this reorganization could
0.1%. In general, we suggest that varying from the heuristic should be considered when a
slower worker has a reaction coefficient that is more than 25% greater than the faster
These findings assume that modifying the work order and restricting the view of
workers would not change behavior. Workers may, in fact, continue to react to the slower
workers behind them. Even in that case, however, their reactions will be reduced due to a
lack of immediacy. Therefore, while further testing would be interesting, we feel that we
have established a first cut at the lower (no change) and upper bounds (+4.8%) of this
The regression is significant despite an R2 of 0.147. Normally, with such a low R2,
it would be difficult to find statistical significance. In this case that wasn’t a problem as
we have over 9,000 data points. So the question becomes not whether the result is
statistically significant but if it is large enough that we care. Is it significant in the real
sense? In our data set the response by the average employee was to change their
24
Worker Motivation on an Assembly Line
processing time by 41% of the difference between the actual processing times of their
coworkers and their individual nominal times. We believe this is a significant reaction.
Since a worker is as likely to speed up as to slow down, we measure the effect as the
absolute value of the change in processing rate. The average of the absolute value of the
difference between actual and nominal rate is 0.62 units per hour. This is on a level with
That we have not explained a large percentage of the total variation should not be
surprising. Our data were collected from an actual ongoing operation. Controlled
experiments are often done to avoid this very problem, to control for extraneous sources
of variability. We had to live with the excess variability and rely on the sheer size of our
data set, to find the pattern inside the mass of variation that one is presented with in a
working factory.
Another explanation of the low R2 is that fitting one slope to all workers leaves a
lot of variability unexplained. The lack of explanatory power of the regression is due, in
part, to the large variation among workers. One slope does not fit all workers. Workers do
not react the same as each other, and any model which merely looks for an average
response leaves a lot of unexplained variability. This phenomenon has been noted in
The data suggest that human reactions go as high as 87% of the difference
between a worker’s nominal speed and the speed of their coworkers. On the other end,
some people appear not to react at all. This provides us with preliminary estimates of the
range of human reactions. It also shows that workers are highly likely to be different from
each other in this aspect. This result is in line with an equity theory motivation. As
25
Worker Motivation on an Assembly Line
Adams (1965) stated, reactions to an inequity situation vary. People respond by changing
values of outcomes or inputs, by leaving the field or by changing their referent person.
We have only measured changes in input as translated into effort, and seen changes in
processing times. People who did not adjust, or who did not adjust at some points, may
have chosen a different response mechanism. It is also possible for people to mix their
response mechanisms and use more than one in each inequity situation. Adams points out
that different people have different thresholds at which they begin to respond. Anything
below this threshold will not generate a response. It is possible that some thresholds are
so high that no response would be found in our work situation. Likewise, the individual
does not have to match the speed of their referent person or group, they just match it
within their individual threshold level. Finally, we have used the average speed of
coworkers. People do not necessarily choose the average person as their referent. They
may be reacting to a subgroup, or a single coworker. Therefore, using equity theory, the
range of human responses could be due to physical limits on an individual’s speed, a mix
modes in different inequity situations, or due to the fact that they may have chosen a
referent other than the average of coworkers. In this way, Equity Theory provides a
We have not proven that the correlation demonstrated here is due to workers
reacting to the processing times of their coworkers. We have shown that the correlation
exists in the factory. We have described equity theory to support the hypothesis and
demonstrated that the correlation is not due to learning, common day of the week or time
26
Worker Motivation on an Assembly Line
of day effects and that the correlation persists when common episodic variance is
differentiates the workers who responded to the speed of their coworkers from those who
did not respond? If we can predict that, what do we do with the information? If we wish
slower workers to speed up, do we place them with faster workers or would this cause the
faster workers to slow down for a zero net difference? Powell and Schultz (2004) show
that, on a serial line with limited buffers, more is gained by workers speeding up than is
lost by other workers slowing down by the same amount. However, in the current
situation, we had parallel workstations and the reaction may be different. Can we arrange
the line to encourage or discourage the reaction? If we put the faster workers in the
middle will this help to speed up more workers than if we put them at the ends? Will the
size of the buffer have an impact on processing times? We hope future research will
times of workers and the processing times of their coworkers. The range of reactions
among human workers is quite large. This is consistent with Equity Theory and
inconsistent with other experiments using additive tasks. We have explored one
27
Worker Motivation on an Assembly Line
References
Dallery, Y., S. B. Gershwin. 1992. Manufacturing flow line systems: A review of models
and analytical results. Queuing Systems 12 3-94.
Dreher, G. F., R. A. Ash, P. Hancock. 1988. The role of the traditional research design in
underestimating the validity of the employment interview. Personnel Psychology
41 315-327.
Edie, L. C. 1954. Traffic delays at toll booths. Journal of the Operations Research
Society of America 2(2) 107-138.
Hopp, W. J. 2004. Fifty Years of Management Science. Management Science, 50(1) 1-7.
Köhler, O., 1927. Über den Gruppenwirkungsgrad der menschlichen Körperarbeit und
die Bedingung optimaler Kollektivkraftreaktion. Industrielle Psychotechnik, 4
209-226.
Nembhard, D. A., M. V. Uzumeri. 2000. Experiential learning and forgetting for manual
and cognitive tasks. Int. J. of Industrial Ergonomics 25 315-326.
28
Worker Motivation on an Assembly Line
Schultz, K. L., D. Juran, J. Boudreau. 1999. The effects of low inventory on the
development of productivity norms. Management Science 45(12) 1664-1678.
Williams, K. D., S. J. Karau, 1991. Social loafing and social compensation: The effects of
expectations of co-worker performance. Journal of Personality and Social
Psychology, 61 570-581.
Williams, K., Harkins, S., and Latané, B., 1981. Identifiability as a deterrent to social
loafing: Two cheering experiments. Journal of Personality and Social Psychology,
40 303-311.
29
Worker Motivation on an Assembly Line
Workers
30
Worker Motivation on an Assembly Line
Note: Since the electronic tests are shorter than the cosmetic and lighting tests the
time stamps on successive electronic tests measure the processing time of the
cosmetic and lighting tests.
31
Worker Motivation on an Assembly Line
10
pij − f ij 0
-10
-10
cij − 0f ij 10
32
Worker Motivation on an Assembly Line
10
pij − f ij 0
-10
-10 0 10
cij − f ij
33
Worker Motivation on an Assembly Line
10
pij − f ij 0
-10
-10 0 10
cij − f ij
34
Worker Motivation on an Assembly Line
10
pij − f ij 0
-10
-10 0 10
cij − f ij
35
Worker Motivation on an Assembly Line
0.25
0.2
0.15
Frequency
0.1
0.05
0
100 99.975 99.95 99.925 99.9 99.875 99.85 99.825 99.8
FTS solution relative to optimal
36
Worker Motivation on an Assembly Line
ANOVA
Sum of
Squares df Mean Square F Sig.
Regression 17886.848 18 993.714 88.074 .000
Residual 102074.549 9047 11.283
Total 119961.396 9065
Coefficients
Standardized
Unstandardized Coefficients Coefficients t Sig.
37
Worker Motivation on an Assembly Line
38
Worker Motivation on an Assembly Line
Unstandardized Standardized
R Coefficients Coefficients
Operator N Square B Std. Error Beta t Sig.
36 39 0.315 0.465 0.267 0.307 1.741 0.092
51 181 0.203 0.455 0.091 0.372 4.998 0.000
27 184 0.219 0.448 0.090 0.374 5.005 0.000
46 105 0.368 0.444 0.070 0.560 6.324 0.000
7 191 0.285 0.443 0.060 0.500 7.335 0.000
111 264 0.221 0.442 0.072 0.363 6.125 0.000
100 165 0.154 0.441 0.094 0.358 4.702 0.000
158 185 0.272 0.437 0.062 0.496 7.003 0.000
67 140 0.236 0.428 0.084 0.417 5.111 0.000
117 62 0.195 0.428 0.176 0.321 2.433 0.018
34 121 0.184 0.393 0.106 0.353 3.708 0.000
48 63 0.254 0.377 0.166 0.322 2.271 0.027
107 136 0.188 0.373 0.093 0.333 4.000 0.000
19 288 0.123 0.337 0.075 0.280 4.517 0.000
172 44 0.323 0.334 0.181 0.302 1.843 0.074
141 146 0.228 0.324 0.089 0.288 3.654 0.000
57 118 0.115 0.294 0.139 0.223 2.110 0.037
131 145 0.137 0.287 0.135 0.187 2.131 0.035
64 236 0.150 0.287 0.077 0.236 3.725 0.000
171 36 0.170 0.285 0.231 0.246 1.235 0.227
66 169 0.137 0.277 0.108 0.201 2.565 0.011
97 83 0.126 0.276 0.150 0.230 1.836 0.070
13 93 0.217 0.206 0.115 0.181 1.784 0.078
160 31 0.215 0.182 0.138 0.275 1.316 0.201
113 128 0.116 0.136 0.087 0.142 1.558 0.122
135 11 0.260 0.118 0.226 0.254 0.524 0.623
146 23 0.446 0.104 0.251 0.082 0.416 0.683
78 42 0.249 0.102 0.076 0.208 1.344 0.188
5 39 0.192 0.037 0.228 0.032 0.162 0.872
24 71 0.034 0.029 0.102 0.039 0.287 0.775
179 14 0.436 -0.155 0.393 -0.120 -0.394 0.705
144 16 0.479 -0.213 0.259 -0.244 -0.824 0.434
109 13 0.586 -0.259 0.367 -0.337 -0.706 0.503
162 21 0.395 -0.431 0.480 -0.231 -0.898 0.385
39
Worker Motivation on an Assembly Line
ANOVA
Sum of
Squares df Mean Square F Sig.
Regression 2010.783 1 2010.783 154.520 .000
Residual 117950.613 9064 13.013
Total 119961.396 9065
Coefficients
Unstandardized Standardized
Coefficients Coefficients t Sig.
40