Content uploaded by Anita Tucker
Author content
All content in this area was uploaded by Anita Tucker on May 21, 2019
Content may be subject to copyright.
This article was downloaded by: [129.64.171.135] On: 09 December 2015, At: 07:26
Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA
Management Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org
The Diseconomies of Queue Pooling: An Empirical
Investigation of Emergency Department Length of Stay
Hummy Song, Anita L. Tucker, Karen L. Murrell
To cite this article:
Hummy Song, Anita L. Tucker, Karen L. Murrell (2015) The Diseconomies of Queue Pooling: An Empirical Investigation of
Emergency Department Length of Stay. Management Science 61(12):3032-3053. http://dx.doi.org/10.1287/mnsc.2014.2118
Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.
The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.
Copyright © 2015, INFORMS
Please scroll down for article—it is on subsequent pages
INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
MANAGEMENT SCIENCE
Vol. 61, No. 12, December 2015, pp. 3032–3053
ISSN 0025-1909 (print) ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.2014.2118
© 2015 INFORMS
The Diseconomies of Queue Pooling: An Empirical
Investigation of Emergency Department Length of Stay
Hummy Song
Harvard University, Boston, Massachusetts 02163, hsong@hbs.edu
Anita L. Tucker
Brandeis University International Business School, Waltham, Massachusetts 02453, atucker@brandeis.edu
Karen L. Murrell
Kaiser Permanente South Sacramento Medical Center, Sacramento, California 95823, karen.l.murrell@kp.org
We conduct an empirical investigation of the impact of queue management on patients’ average wait time
and length of stay (LOS). Using an emergency department’s (ED) patient-level data from 2007 to 2010, we
find that patients’ average wait time and LOS are longer when physicians are assigned patients under a pooled
queuing system with a fairness constraint compared to a dedicated queuing system with the same fairness
constraint. Using a difference-in-differences approach, we find the dedicated queuing system is associated with a
17% decrease in average LOS and a 9% decrease in average wait time relative to the control group—a 39-minute
reduction in LOS and a four-minute reduction in wait time for an average patient of medium severity in this ED.
Interviews and observations of physicians suggest that the improved performance stems from the physicians’
increased ownership over patients and resources that is afforded by a dedicated queuing system, which enables
physicians to more actively manage the flow of patients into and out of ED beds. Our findings suggest that
the benefits from improved flow management in a dedicated queuing system can be large enough to overcome
the longer wait time predicted to arise from nonpooled queues. We conduct additional analyses to rule out
alternate explanations for the reduced average wait time and LOS in the dedicated system, such as stinting and
decreased quality of care. Our paper has implications for healthcare organizations and others seeking to reduce
patient wait time and LOS without increasing costs.
Keywords: pooling; fairness; queue management; strategic servers; empirical operations; healthcare
History : Received July 23, 2013; accepted October 23, 2014, by Serguei Netessine, operations management.
Published online in Articles in Advance May 28, 2015.
1. Introduction
Improving efficiency and customer experience is a key
objective for service organizations. Skillful applica-
tion of operations management principles may help
achieve these goals. In particular, queue management
decisions—such as queue structure and job routing
policies—may impact how long customers have to
wait for service and their service times.
Prior work has demonstrated through analytical
models that pooling separate streams of identical cus-
tomers into a single queue served by a bank of iden-
tical servers is more efficient than having a set of
dedicated queues, because pooling results in shorter
wait times for service (Eppen 1979,Kleinrock 1976).
Having a pooled queue structure leads to a reduction
in wait time because it enables customers to be pro-
cessed by any available server from a bank of servers,
rather than having to wait for a specific server to
become available. That said, prior analytical research
also suggests that pooling queues may not always
yield the expected performance improvements (Debo
et al. 2008,van Dijk and van der Sluis 2009,Hopp et al.
2007,Jouini et al. 2008,Loch 1998,Mandelbaum and
Reiman 1998). For example, combining streams of cus-
tomers who have different processing requirements
can introduce inefficiencies that erode the benefits
of pooling (Benjaafar 1995,Green and Nguyen 2001,
Mandelbaum and Reiman 1998,Rothkopf and Rech
1987). In addition, the perceived unfairness of a pooled
queue, in which faster servers are assigned more cus-
tomers than their peers, may negatively impact the
speed at which servers work (Doroudi et al. 2011).
Thus, the overall impact of queue pooling in service
settings is ambiguous.
To our knowledge, there have been few field-based,
empirical studies on the impact of pooled versus ded-
icated queue management systems on the speed of
service. This is an important omission because, in ser-
vice settings, servers can adjust how they manage
their work to increase or decrease their service rate
(Doroudi et al. 2011,Hopp et al. 2009). Operations
management scholars advocate for more studies that
3032
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3033
examine how human behavior can alter the dynam-
ics between operational variables and performance
(Boudreau et al. 2003,Jouini et al. 2008). Thus, empir-
ical research that examines the impact of queue struc-
ture on servers’ behaviors can provide new insights
for operations management theory and increase the
relevance of queuing theory and research to practice.
To address this gap, we leverage the introduction of
a new policy that changed the queuing system in only
one part of a hospital’s emergency department (ED),
but not the other, from a pooled system to a dedicated
system. The parallel trend in performance of the two
parts of the ED before the queuing system change,
and the fact that the change affected only one part
of the ED, allows us to use a difference-in-differences
approach to empirically test the impact of a change
in the structure of the queuing system on the average
wait time to be seen by an ED physician and the aver-
age length of stay (LOS) in the ED. LOS is a measure
of service time and starts with the time the physician
begins delivering care to the patient and ends with
either a bed request for admission to the hospital or
the discharge of a patient to their home or an outside
facility. We use the term LOS rather than service time
to more clearly convey that this measure encompasses
both (a) the value-added time when clinicians are pro-
viding care, as well as (b) the time that the patient is
occupying an ED bed but is not receiving active care
(e.g., when the physician is waiting for test results or
treating other patients).
The ED under study switched from a pooled to a
dedicated queuing system to be able to handle the
larger volume of patients predicted to occur because
of the closing of a nearby ED. For both the pooled
and dedicated queuing systems, a fairness constraint
in the form of a round robin (RR) routing policy
was used to assign patients to physicians, in which
patients were evenly distributed across physicians
independent of physician speed or idle time. The ED
had this policy because physicians were paid a fixed
salary and did not receive additional compensation
for treating more patients or working more hours than
scheduled. As a result, there were few financial incen-
tives available to increase physician productivity, and
instead, work was allocated equally among physi-
cians. Using a difference-in-differences approach, we
find that, on average, the use of a dedicated queu-
ing system with a RR routing policy as a fairness
constraint—after controlling for individual patient,
physician, time, and ED characteristics—is associated
with a 17% decrease in patients’ average LOS and a
9% decrease in their average wait time relative to the
control group. This represents a 39-minute reduction
in LOS and a four-minute reduction in wait time—
a meaningful time savings for the ED.
Operations management theory suggests a possi-
ble reason why the pooled queuing system with a
fairness constraint is associated with a longer aver-
age LOS than the dedicated queuing system with a
fairness constraint. Similar to workers in other ser-
vice settings (Debo et al. 2008,Hasija et al. 2010,
Tan and Netessine 2014), physicians in the dedicated
queuing system are strategic servers who change
their behaviors in response to their assigned respon-
sibilities and ownership over the work routines and
resources needed to accomplish those responsibilities
(Cachon and Zhang 2007;Gilbert and Weng 1998;
Hopp et al. 2007,2009). Interviews with physicians
suggest that, in this context, the increased ownership
that stems from a dedicated queuing system with
a fairness constraint leads to a situation in which
the improvements in service rates due to better flow
management are greater than the variability-buffering
benefits of a pooled queuing system with a fairness
constraint.
This paper makes a contribution to the literature on
queue pooling because prior research has emphasized
customer behaviors that reduce the process losses of
dedicated queues, but fewer papers have empirically
examined the impact of employee behaviors on the
performance of dedicated versus pooled queuing sys-
tems (Boudreau et al. 2003,Hopp et al. 2007,Jouini
et al. 2008). Our work thus informs the debate over
the benefit of a pooled queue, which enables flexibil-
ity in the routing of jobs to servers, and a dedicated
queue, which enables improvement in wait times and
service times through better flow management.
2. Prior Research and Hypotheses
2.1. Prior Research on Queue Management and
Service Times
Operations scholars have investigated at least two dif-
ferent contexts in which pooling may occur: inven-
tory waiting to be processed (production-inventory
systems) and customers waiting for service (queuing
networks). Most closely related to our research con-
text, studies of queuing networks focus on the effect
of pooling queues of customers, servers, and tasks
in service organizations (Mandelbaum and Reiman
1998). Much of this research has been conducted
with call centers, and has shown that the benefits
of flexible servers and pooled queues can outweigh
potential drawbacks (Anupindi et al. 2005,Bassam-
boo et al. 2010,Gans et al. 2003,Jouini et al. 2008).
Researchers have reached similar conclusions in other
settings, such as mail delivery, finding that pool-
ing can improve quality while concurrently reduc-
ing costs (Ata and Van Mieghem 2008). Furthermore,
prior research has found that pooling is beneficial and
wait time reductions are achieved even when work
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3034 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
is allocated fairly among servers using a RR rout-
ing policy (Hyytiä and Aalto 2013,Raz et al. 2006).
In fact, Armony and Ward (2010) find that pooling
with a fairness constraint outperforms classical pool-
ing when the arrival rate of customers is high because
faster servers have an incentive to slow their service
rate under systems in which work is allocated based
on server availability instead of a fair distribution
across servers.
On the other hand, some analytical models have
shown that the behavioral responses of servers and
customers can reduce the expected benefits of queue
pooling (van Dijk and van der Sluis 2008,Hopp
et al. 2007,Loch 1998,Mandelbaum and Reiman
1998,Rothkopf and Rech 1987). Most pertinent to our
study, strategic servers may reduce the effectiveness
of queue pooling (Cachon and Zhang 2007;Debo et al.
2008;Hopp et al. 2007,2009;Jouini et al. 2008). First,
they may manipulate customer service times to be
higher or lower by managing their tasks differently
when it benefits them to do so (Hopp et al. 2007,Link
and Naveh 2006,Tan and Netessine 2014). For exam-
ple, in the restaurant industry, Tan and Netessine
(2014) find that wait staff adjust the services offered
to customers so that customers spend less time in
the restaurant when their workload is high. Similarly,
Oliva and Sterman (2001) find that bank employees
reduce the steps they go through to approve loans
when their workload is high, even though this erodes
bank profitability. Second, strategic servers can also
slow down their work pace. Using analytical mod-
els, Debo et al. (2008) show that when workers are
paid by the quantity of work completed, such as taxi-
cab drivers and lawyers, they add unnecessary tasks
when business is slow, thereby increasing service time
for their customers. Similarly, Hasija et al. (2010) find
that call center agents take more time to answer cus-
tomers’ queries when they have low workloads if
their contract rewards them for keeping utilization
above a minimum threshold. Collectively, these stud-
ies suggest that service time is impacted by strategic
servers’ responses to incentives and responsibilities.
Even when strategic servers do not have direct
financial incentives to adjust their service rates, they
may still manipulate their service times if they
have a high degree of perceived ownership over
their assigned jobs. Employees feel higher levels of
ownership when they are given the resources and
responsibility to manage the complete workflow of
a meaningful task (Hackman and Oldham 1976). By
design, dedicated queuing systems with a fairness
constraint afford higher levels of ownership than
do pooled queuing systems with the same fairness
constraint because in the former, each server has
been explicitly assigned the responsibility for effi-
ciently completing the work waiting in his or her
queue. In contrast, pooled queuing systems provide
lower levels of ownership because the responsibility
for depleting the queue is dispersed over multiple
servers. Thus, strategic servers in dedicated queuing
systems with a fairness constraint may be more moti-
vated to efficiently manage their workload than those
in pooled queuing systems with a fairness constraint
(Doroudi et al. 2011,Gilbert and Weng 1998).
2.2. Queue Management and Strategic Physician
Behavior in the Emergency Department
ED physicians are strategic servers, as defined by
Cachon and Zhang (2007). To illustrate how physi-
cians operate as strategic servers, consider an ED
physician who has a patient with a headache. The
physician can treat the patient using any combi-
nation of the following tasks: obtain a detailed
medical history to generate possible causes of the
headache, order a computed tomography scan, or pre-
scribe aspirin. The physician’s choice can impact the
patient’s LOS because of the variance in time required
for the different options. In addition, the physician
can influence patient LOS by proactively pulling for
information, such as x-ray results, rather than waiting
for that information to be pushed. The physician can
also control his or her own utilization because there
are usually multiple patients under the care of an
ED physician. Thus, physicians can reduce their own
idle times and further increase the flow of patients
through the system.
In this paper, we consider two different types of
queuing systems in the context of an ED. In a pooled
queuing system—which is typical for most EDs in the
United States—a physician is assigned to a patient
only once the patient is placed in an ED bed. This
means patients in the waiting room remain in a
pooled queue while waiting for an open bed. In a
dedicated queuing system, physicians are assigned to
patients at the point of triage. Here, patients in the
waiting room are, in effect, waiting to be seen by a
specific physician. In the dedicated queuing system,
each physician thus has a greater ownership over his
or her workload even before the patient is placed in
an ED bed.
In the ED that we study, each physician in the ded-
icated system also controls his or her own bank of
resources (e.g., beds and nurses) necessary to facil-
itate the flow of his or her own patients. Physi-
cians are assigned patients in a RR fashion that fairly
allocates patients among all physicians independent
of physicians’ service rates. In addition, they can
only go home when all of their assigned patients
are discharged or nearly discharged (e.g., awaiting a
test result), and are not paid extra for working past
the scheduled end of their shift. Therefore, physi-
cians have an incentive and the ability to manage
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3035
their workload as efficiently as possible. For example,
physicians can coordinate the care of their patients
with their nurses to prioritize getting test results
back for a patient so he can be discharged, and then
quickly move a patient from the waiting room into
that vacated bed. In contrast, in the pooled queu-
ing system, physicians do not “own” patients in the
waiting room, nurses and beds are shared among all
physicians, and they rely on a triage nurse, called
the “internal triage” nurse, to manage the flow of
patients into available beds for the entire ED. Thus,
in the pooled queuing system, physicians’ have own-
ership over a much smaller portion of the patient
flow process. Based on our interviews with physicians
and observations of their practice patterns, we suspect
that the higher level of ownership of one’s workload
and the resources necessary to manage that workload
afforded by the dedicated queuing system increases
physicians’ perceived ownership over patient flow.
This results in physicians having a faster rate of dis-
charging patients throughout their entire shift than
when in the pooled system.
Prior theoretical operations management research
suggests that when strategic servers have ownership
and responsibility for managing flow, it can lead to
lower service times. Gilbert and Weng (1998) and
Cachon and Zhang (2007) construct analytical models
of a buyer’s choice of queue structure for allocating
demand among two suppliers. They find that suppli-
ers in a dedicated system produce the goods faster
than those in a pooled system because the dedicated
system’s suppliers have more incentive to invest in
production capacity. The dedicated system provides
certainty that they will benefit from their capacity
investments, which can be thought of as having own-
ership over a demand stream in combination with
the responsibility over production resources needed
to meet that demand. Similarly, in the context of a
hospital’s inpatient department, Best et al. (2015) use
a stylized model to show that a patient flow director
with increased ownership and responsibility for man-
aging flow is able to attain a significant decrease in
patient LOS. The authors suggest that this decrease is
attained from increased motivation to cut non-value-
added time and better coordinate patient care among
doctors, nurses, and case managers.
In the context of an ED, switching from a pooled
to a dedicated queuing system should similarly affect
the behavior of physicians by increasing the degree
of ownership physicians have over their patients’
flow through the ED. Specifically, we hypothesize
that ED physicians may attain a shorter average LOS
for their patients when they work in an ED with a
dedicated queuing system with a fairness constraint.
Prior research suggests that servers work slower at
low workloads because there is no need to work fast
because of the slack capacity (Tan and Netessine 2014).
However, in our ED setting, workloads are typically at
high levels because of the ED’s ability to staff accord-
ing to historical demand and to send clinicians home
early during periods of unexpectedly low demand.
Therefore, we hypothesize a direct, positive effect of a
dedicated queuing system on LOS.
Hypothesis 1. LOS is shorter in the ED when physi-
cians are working in a dedicated queuing system as opposed
to a pooled queuing system.
We further consider how dedicated queuing sys-
tems may affect patients’ average wait times. A priori,
it is unclear whether dedicated queues with strate-
gic servers will result in shorter or longer wait times
for customers. On one hand, when under a dedicated
queuing system, if patients currently being cared for
spend less time in an ED bed and if a physician
proactively places the next patient from his or her
queue into the newly available bed, the next patient’s
wait time may decrease because of an indirect queu-
ing effect. In other words, the benefits of a dedicated
queue—fair assignment of work and ownership over
patients, resources, and patient flow—may overcome
the negative impact on wait time of using a dedi-
cated rather than a pooled queue. Thus, we predict
the following:
Hypothesis 2A. Wait time is shorter in the ED when
physicians are working in a dedicated queuing system as
opposed to a pooled queuing system.
On the other hand, switching from a pooled to a
dedicated queue may result in an increase in wait
time, due to the well-known inefficiency of forcing
customers’ whose server is busy to wait for that
server to be free, even if another server is idle (Eppen
1979,Kleinrock 1976). The inefficiency of dedicated
queues might overpower the possible reduction in
wait times because of faster service times. Therefore,
we test the following competing hypothesis.
Hypothesis 2B. Wait time is longer in the ED when
physicians are working in a dedicated queuing system as
opposed to a pooled queuing system.
To understand the behavioral mechanism through
which different queuing systems may impact LOS,
we explore the rates at which physicians discharge
patients during different time periods throughout
their shifts. We hypothesize that the higher level of
ownership over patient flow afforded by a dedicated
queuing system, as opposed to a pooled queuing sys-
tem, motivates physicians to more efficiently man-
age patient flow throughout the duration of the entire
shift. Physicians in the dedicated system may be able
to efficiently manage patient flow—and thus achieve
higher discharge rates—by proactively “pulling” for
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3036 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
lab, x-ray, and specialty consult results; improving
coordination with nurses to prioritize tasks necessary
for discharge; initiating the discharge process sooner
for patients ready for discharge; and making sure that
nurses place waiting patients into available beds as
soon as possible. This hypothesized increase in dis-
charge rate is in contrast to only speeding up toward
the end of the shift, which would be predicted if physi-
cians were only subject to a deadline effect and were
not better managing patient flow (Deo et al. 2014).
Prior theoretical research suggests that physicians
in the dedicated system will have a greater incentive
to consistently work at a higher rate because they can
reap the benefits that stem from achieving a faster rate
of production (Gilbert and Weng 1998). In our set-
ting, the benefits to physicians of obtaining a higher
discharge rate are (a) more time to spend with cur-
rent patients, which increases both patient and physi-
cian satisfaction (Hopp et al. 2007); (b) idle time if
the physician has no additional patients currently in
queue (Armony and Ward 2010); and (c) less work
remaining for the physician to complete before he or
she can go home. In a pooled queuing system, these
benefits do not necessarily accrue to physicians who
work at a higher rate because the misalignment of
responsibility for patient flow and ownership over
patients and resources prevents physicians from being
able to reap these benefits. Thus, we hypothesize
that physicians working in a dedicated queuing sys-
tem will attain higher rates of discharging patients
throughout the shift. Specifically, we hypothesize that
this increase in discharge rate will emerge a few hours
after the beginning of a shift because the average LOS
is greater than two hours and, therefore, it would not
be possible to discharge many patients in the first two
hours of one’s shift. However, after this initial two-
hour period, the faster discharge rate will be present
throughout the remainder of the shift, rather than only
at the end of the shift.
Hypothesis 3. A physician’s discharge rate of patients
is greater for each noninitial time period of the shift when
physicians are working in a dedicated queuing system as
opposed to a pooled queuing system.
3. Setting, Data, and Empirical
Methods
3.1. Research Setting
Our data comes from the ED of a 162-bed hospital in
Northern California. WeselectthisEDforstudybecause
in August 2008, it experienced an intervention—which
we describe in more detail in §3.2—that transformed a
part of the ED from having a pooled queuing system
to a dedicated queuing system for the patients wait-
ing to be seen in the ED. We use data from a timespan
before and after the intervention (March 2007 to July
2010) to test our hypotheses about the impact of queu-
ing systems on average LOS, wait time, and discharge
rate in the ED.
Depending on the time of day, this ED had an aver-
age of two to five physicians staffing 41 ED beds and
up to nine hallway gurneys. One bed was located in
the resuscitation room and reserved for patients arriv-
ing without a pulse, three beds were in the trauma
bay reserved for trauma intakes, four beds were in the
rapid care area (RCA) for low severity patients, and a
minimum of two beds were reserved for psychiatric
patients. This ED experienced an average 5% increase
in patient volume each year, from approximately
65,000 patients in 2007 to 76,000 patients in 2010.
The average daily patient volume was 178 patients in
2007 and 212 patients in 2010. This was a relatively
large patient volume in comparison to other EDs in
the surrounding areas.
This ED, like many others, had a standardized
patient flow process (Figure 1). On a patient’s arrival,
a registration clerk conducted a brief registration
process. A second triage nurse, called the “external
triage” nurse, obtained vital signs, collected the chief
complaint, and assigned an Emergency Severity Index
(ESI) triage category—a commonly used, standard
ranking of ED patient severity that ranges from lev-
els 1 (highest acuteness) through 5 (lowest acuteness).
This triage process accounted for a patient’s expected
level and type of resource utilization, and was used
to route a patient to either the main area (main ED) or
the RCA. The two areas of the ED each had its own
equipment and staff to deliver care to patients (e.g.,
the RCA had its own computer terminals and vital
sign monitors that were separate from the main ED’s
equipment). Ninety-eight percent of higher acuteness
patients (ESI levels 1, 2, or 3) were treated in the main
ED. Seventy-five percent of lower acuteness patients
(ESI levels 4 or 5) were treated in the RCA. Lower
acuteness patients were treated in the main ED when
main ED beds were available and the waiting room
census was low (15% of lower acuteness patients) or
when they arrived between 11 p.m. and 7 a.m. when
the RCA was closed (9% of lower acuteness patients).
Figure 1 Standard Patient Flow in the Emergency Department
!RRIVAL
REGISTRATION 4RIAGE %$BED )NPATIENTBEDREQUESTAOR
DISCHARGEFROM%$B
%$WAITTIME %$LENGTHOFSTAY
%$SOJOURNTIME
-$START
aFor patients who were admitted to the hospital.
bFor patients who were discharged home or to an outside facility.
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3037
In this ED, a computer system assigned each patient
to a specific attending physician, either on assign-
ment to a bed (pooled queuing system) or at the point
of triage (dedicated queuing system). The assigned
physician assumed responsibility for completing the
set of physician-related tasks for that patient during
the patient’s ED visit, such as taking the patient’s his-
tory, prescribing medications, and ordering tests or
treatments. This physician could consult with other
physicians concerning his or her patient’s care, but
this did not transfer the responsibility for patient care
to the consulting physician. It was common for a
physician to serve multiple patients simultaneously.
In other words, a physician did not need to discharge
one patient before starting work for the next patient.
Physicians arrived at staggered times throughout
the day, such that there was not a certain time
at which all physicians changed shifts (Figure 2).
Physician shift times were determined in advance by
the ED chief, and the ED scheduler assigned indi-
vidual physicians to the predetermined shift times.
Physicians could change shifts on the hour between
5 a.m. and 11 a.m., between 2 p.m. and 5 p.m., and
at 11 p.m. or midnight. Between 7 a.m. and 11 p.m.,
there was usually one physician working in the RCA
and four physicians working in the main ED. During
the overnight shift from 11 p.m. to 7 a.m., there were a
minimum of two physicians and a maximum of four
physicians working in the main ED.
Physicians were assigned to either the RCA or
the main ED for the full duration of a shift by the
ED scheduler. They were paid a flat rate for their
Figure 2 (Color online) Example of Physician Shift Distribution Over a 24-Hour Period
MD 247 50 268 392 96 152 27 319 367 350 270 28
0:00
1:00
2:00
3:00
4:00
5:00
6:00
7:00
8:00
9:00
10:00
11:00
12:00
13:00
14:00
15:00
16:00
17:00
18:00
19:00
20:00
21:00
22:00
23:00
Notes. MD numbers across the xaxis are unique physician identifiers. Shaded bars indicate the duration of a physician’s shift.
shift without any additional compensation for the
services provided or the number of hours worked.
Thus, there were no incentives to stretch out treat-
ment times by providing additional services. Prior
to leaving the shift, physicians were expected to dis-
charge or at least complete a care plan for the cohort
of patients assigned to them (e.g., indicate what next
steps should be taken if the lab test comes back pos-
itive versus negative), which incentivized physicians
to get their patients through the system as efficiently
as possible. Physicians were not required to stay if
they had patients who were simply boarding in the
ED, waiting to be transferred to an inpatient unit
or to another facility. To allow physicians enough
time to either complete a care plan or discharge the
patients who had been assigned to them, they were
assigned new patients only up until two hours before
the scheduled end of their shifts. Patients arriving
during the last two hours of a physician’s shift were
assigned to one of the other physicians on shift or, if it
was close enough in time, to the oncoming physician.
Because physician shifts were sufficiently staggered,
there was always a physician available to take newly
arriving patients and this did not induce a greater
variation in system productivity.
3.2. Intervention: Change in the Patient
Assignment System
In August 2008, the main ED implemented an inter-
vention called the patient assignment system (PAS).
PAS restructured the main ED from having a pooled
queuing system to a dedicated queuing system. Prior
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3038 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
to PAS, higher severity patients due to be seen in the
main ED returned to the waiting room after being
triaged, with the exception of ESI level 1 patients who
proceeded directly to the resuscitation room. When
a bed became available in the main ED, the inter-
nal triage nurse placed the next patient of highest
severity in this bed. Our interviews with ED physi-
cians revealed that this process often resulted in a
delay from the bed becoming available to a patient
being placed in the bed because the internal triage
nurse was not responsible for patient flow through
the ED, and the physicians did not feel responsible
for making sure that empty beds were filled quickly.
Once a patient was placed in a bed, the computer
system assigned each patient to a physician using an
RR routing policy, which means that each patient was
assigned to a physician in a set order that evenly dis-
tributed patients among physicians regardless of each
physician’s current workload. Once this assignment
occurred, the physician could see the assigned patient
listed under his or her panel when logged onto the
patient management system on one of the ED com-
puters. Thus, when a patient was waiting in the wait-
ing room, he or she was in a pooled queue waiting to
be assigned to any one of the, on average, four physi-
cians on shift in the main ED.
Prior to the PAS intervention, the patient only
entered a specific physician’s queue after being placed
in an available main ED bed by the internal triage
nurse. It was at this point that the physician had own-
ership of the patient, not before. The only exception
to the RR routing policy was made when a physi-
cian was currently involved in the resuscitation of
an ESI level 1 patient, in which case another physi-
cian could voluntarily take on that physician’s next
patient. In addition, at the beginning of a physician’s
shift, the computer system assigned one, two, or
three consecutive patients to the oncoming physician.
The specific number of consecutive patients to whom
a physician was assigned was automatically deter-
mined by the computer system based on the rate of
patient arrivals. This RR routing policy was instituted
to prevent physicians from unfairly selecting “eas-
ier” patients and to ensure that the faster physicians
would not be unequally assigned more work simply
because of their higher service rates. This simultane-
ously made patient routing to physicians both fair
and nearly random rather than due to a physician’s
seniority or speed of discharging patients. It was fea-
sible to implement because there were two organiza-
tional structures in place to minimize the variation
in workload across the physicians staffing the main
ED: (a) the hospital’s trauma team assumed primary
responsibility for incoming trauma patients and thus
did not disproportionately increase the workload of
an ED physician; (b) the RCA cared for lower severity
patients. Thus, there was limited variation in patient
intensity among the patients being assigned to the
physicians staffing the main ED.
After PAS implementation, the computer system
still used the RR routing policy but assigned each
patient to a physician at the point of triage. This
means that, when a physician logged onto the patient
management system to view his or her panel of
patients, the display showed not only those patients
who were already placed in ED beds but also those
who were still in the waiting room. This increased
physicians’ perceived ownership of their patients
because they were responsible for their patients’ care
and experience from triage onward—which included
their time in the waiting room—rather than just from
placement in an ED bed. In conjunction, it was now
the physicians’ responsibility to make sure their next
patient from the waiting room was placed in an avail-
able main ED bed. To enable physicians to carry out
this additional responsibility, six main ED beds and
two hallway gurneys were allocated to each physician
working in the main ED. In addition, two nurses were
assigned to each physician to help care for patients,
although each physician typically worked with other
nurses outside of these two nurses during the course
of the shift because (a) nurses’ shift change times were
not aligned with that of the physician and (b) nurses
had designated break times during which a relief
nurse substituted in for the duration of the break.
After PAS implementation, the computer system’s
RR routing policy was maintained and adhered to,
even if there was a physician who had waiting
patients while another physician had an available ED
bed and no waiting patients. Hence, patient assign-
ment remained independent of a physician’s speed of
discharging patients. Similarly, the incentive of having
to stay until all patients had been cared for remained
constant, though now physicians also had to care for
the patients who had been assigned to them who were
still in the waiting room.
In the RCA, the process used to assign patients to
the physician working in the RCA did not change
over the course of our study. A lower severity patient
was assigned to a physician when he or she was
called to be seen in the examination room, not while
in the waiting room. Thus, the RCA physician was not
responsible for any patient who was still waiting in
the waiting room at the conclusion of his or her shift;
any patient still waiting became the responsibility of
the next physician coming on to the shift.
3.3. Data
This study uses approximately three and a half years
of de-identified electronic medical record (EMR) data
of all 238,946 patients treated in the ED from March 1,
2007 to July 31, 2010. The data set contains patient-
level information including, but not limited to, the
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3039
following: the patient’s time of arrival and departure,
LOS, ESI level, attending physician, and disposition.
We exclude patients with no attending physician or
ESI level listed on their record, patients who left with-
out being seen by a physician, patients who had a
LOS of zero minutes or less, patients whose records
lacked a time stamp for when the physician began
caring for the patient, and patients who were admit-
ted to the hospital but whose records lacked a time
stamp for when a bed request was made. In addi-
tion, we exclude patients whose LOS was greater
than 48 hours; most of these patients presented with
a psychological condition and were waiting to be
discharged to an appropriate facility. We exclude
these observations from our data set because their
extended LOS was typically driven by placement
logistics rather than by physicians’ levels of produc-
tivity. In addition, we exclude patients of ESI level 1
(i.e., patients needing resuscitation) and patients who
died in the ED because their LOSs were likely to
be driven by factors other than physician productiv-
ity. Finally, we exclude trauma patients because the
hospital’s trauma team, not a particular ED physi-
cian, primarily cared for these patients. Altogether, we
exclude 12,817 patients or 5.4% of the overall sample.
Using this sample of 226,129 patients, we create a
patient-level panel data set that treats the physician
as the panel variable. For our analyses, we exclude
data from August 2008 to account for an acclimation
period because the exact date of PAS implementa-
tion is unknown. In addition, we limit our sample to
the patients seen by physicians who were full-time
employees of this ED. Physicians who worked in this
ED but were not full-time employees tended to be
employees of other hospitals in the hospital’s network
who were brought in to cover small portions of shifts
when the full-time ED physicians were not able to
staff the ED (e.g., during physician staff meetings).
This results in a final sample of 217,213 patients.
In addition to the EMR data, we also gathered qual-
itative data through 86 hours of observations of ED
staff and unstructured interviews about workflow in
the ED with ED physicians, nursing staff, and the ED
unit leadership.
3.4. Dependent Variables
Our key dependent variables are ED wait time, ED
LOS, and patient discharge rate. ED wait time is
defined as the time from a patient’s arrival to the ED
to the time the physician began delivering care. ED
LOS starts with the time the physician began deliv-
ering care to the patient and—for patients admitted
to the hospital—ends with a bed request for admis-
sion to the hospital, thus excluding the time spent
boarding in the ED and any time spent in an inpatient
unit. For patients discharged to home or to an outside
facility, ED LOS ends at the time of discharge. We log-
transform ED wait time and LOS because each of their
distributions are otherwise right skewed. Patient dis-
charge rate is defined as the number of patients dis-
charged per hour by a given physician in a specified
two-hour period of the shift, such as the first two
hours, second two hours, or final two hours.
We employ a set of additional dependent variables
for analyses that extend the main findings and con-
sider possible alternative explanations. These include
binary indicators for whether a lab was ordered, an
x-ray was ordered, a patient was admitted to the hos-
pital, a patient died in the ED, or a patient returned
to the ED within 72 hours, respectively.
3.5. Independent and Control Variables
3.5.1. Patient Assignment Intervention in the
Main ED. The implementation of PAS marks the time
at which the main ED transitioned from having a
pooled queuing system to a dedicated queuing sys-
tem. We capture this transition with a binary inter-
action term, PAS ×main, which is equal to 1 in the
main ED after the implementation of PAS and 0 oth-
erwise (i.e., in the main ED before the implementation
of PAS, or in the RCA at any time). To account for an
acclimation period, we designate the pre-PAS period
to include up to July 31, 2008 and the post-PAS period
to begin with September 1, 2008.
3.5.2. Control Variables. We account for several
factors that may affect our dependent variables and
may be correlated with our independent variables,
PAS and main. These include factors related to the
patient’s condition, the state of the ED, the physician’s
practice experience, and time trends. To account for
the variation in LOS due to the severity of a patient’s
condition, we control for the patient’s acuteness and
age. We account for patient acuteness using a series of
dummy variables that reflect ESI levels 2, 3, 4, and 5,
respectively. The combination of a patient’s ESI level
and age is the best approximation we have for patient
condition and severity because our data set does not
include patients’ specific diagnoses (e.g., diagnosis-
related groups (DRGs)). It is important to control for
patient acuteness because the patient mix in this ED
changed over time, wherein more patients presenting
to the main ED were of higher acuteness and more
patients presenting to the RCA were of lower acute-
ness after PAS implementation.
To capture ED busyness and congestion, we control
for the total number of physicians working during a
given morning, afternoon, or overnight shift; the num-
ber of patients waiting to be seen by this physician
at a given time; the number of patients being seen by
this physician at a given time; whether an ESI level 1
patient was present in the ED; and whether a trauma
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3040 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
patient was present in the ED. Relatedly, to account
for other systematic differences in patients’ LOSs that
would arise from differences in structural elements of
the ED, we control for the general time frame of the
physician’s shift (morning, afternoon, or overnight)
and the location of the shift (main ED or RCA).
To account for systematic differences arising from
differences in physicians’ experience working in this
particular ED, we control for the number of shifts
the physician has worked in this ED since the begin-
ning of the data set up until the point of each patient
encounter. As we explain in more detail in §3.6, we
also include physician fixed effects to account for
other unobserved differences by physician.
Finally, we account for time trends and related
influences by including dummy variables for day of
the week and by using month-year fixed effects.
3.6. Empirical Models
Our main analyses use a difference-in-differences
framework to examine the relative changes in LOS
and wait time for patients seen in the main ED and
the RCA before and after PAS implementation. We
use linear regression models with month-year and
physician fixed effects and clustered standard errors.
We cluster standard errors by physician to account
for within-physician correlations of the error terms,
both within and across shifts, rather than imposing
the usual assumption that all error terms are inde-
pendently and identically distributed. The fixed-effects
models allow us to capture time trends and to control
for unobservable individual physician effects that do
not vary over time, such as level of motivation, innate
ability, and practice routines. These are important to
account for because they may significantly influence a
physician’s productivity level in ways that cannot be
measured (McCarthy et al. 2012).
In addition to the standard assumptions of linear
regression models, fixed-effects models make two key
assumptions, both of which are satisfied in our study.
First, is the assumption of strict exogeneity, which
means the observation-specific error term is uncor-
related with the covariates of the observation and
all other observations belonging to the same cluster
(Wooldridge 2010). This is a plausible assumption in
our context because (a) there is a low likelihood that
patients with multiple visits are treated by the same
physician and (b) the patient error term is unlikely
to be correlated with the covariates for other patients
of the same physician. In addition, the unobservable
random traits of physicians that affect their patients’
average LOS are not likely to be associated with the
key independent variable of interest. Specifically, the
RR routing policy makes it unlikely that the fastest
physicians receive the most complicated cases since
patient assignment to physicians is random and is not
driven by physician speed or physician preference.
We use fixed-effects models rather than random-
effects models because we do not believe that the
random-effects assumption of zero correlation be-
tween the month-year effect or physician effect and
the other covariates (such as the number of shifts
worked by the physician) holds. By using fixed-effects
models, we can account for the unobserved traits of
each month-year and of each physician that are asso-
ciated with a patient’s LOS and also correlated with
the independent variables of interest. Accordingly, we
conduct the Durbin-Wu-Hausman test, which rejects
the random-effects model in favor of the fixed-effects
model (2>169045, p < 00001).
3.6.1. ED LOS. To test Hypothesis 1, we estimate
the following difference-in-differences model at the
patient level:
ln LOSijt =0+1mainij +2PASt×mainij +Xij t
+t+MDi+ijt 0(1)
Here, ln LOSij t represents the logged number of min-
utes that patient iof physician jstayed in the ED
in month-year t;mainij indicates whether patient iof
physician jwas seen in the main ED; PASt×mainij
is an interaction term equal to 1 when the patient
was seen in the main ED after the implementation of
PAS; Xijt is a vector of patient, physician, and day-
of-week covariates; tis a vector of month-year fixed
effects, MDiis a vector of physician indicators; ’s
and ’s represent vectors of coefficients; represents
a vector of physician fixed effects; and is the time-
varying error term not already captured. Table 1pro-
vides summary definitions for all variables included
in our models.
In estimating Equation (1), we use the difference-
in-differences estimator, PASt×mainij , to compare the
difference in patients’ average LOS in the main ED and
the RCA before PAS implementation to the difference
after PAS implementation. Because the queue struc-
ture did not change in the RCA, whereas the main ED
moved from having a pooled to a dedicated queuing
system, we consider the shifts worked in the RCA as
comprising the untreated comparison group and those
worked in the main ED as comprising the treatment
group. By using a difference-in-differences approach,
we are able to control for any bias caused by variables
common to the main ED and the RCA, even when
those variables are unobserved. Although the acute-
ness of patients seen in the two parts of the ED dif-
fered, thus implying differences in treatment processes
and levels of patient LOS, the RCA serves as a rea-
sonable control because, as our interviews with ED
leadership and staff indicate, there were no changes
besides PAS during the study period that affected only
one part of the ED and not the other. Furthermore, we
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3041
Table 1 Summary Definition of Variables
Variable Description Level of analysis
Main dependent variable
ED wait time Logged number of minutes elapsed between patient arrival to ED and MD start Patient
ED length of stay Logged number of minutes elapsed between MD start and bed request (for patients admitted to hospital)
or discharge from ED (for patients discharged home or to an outside facility)
Patient
Discharge rate Number of patients discharged per hour by a given physician in a given two-hour period of the shift
(e.g., penultimate two hours, final two hours)
Physician-shift
(two-hour period)
Independent and control variables
ESI level Four indicators for patient’s ESI level (from highest to lowest: 2, 3, 4, 5)aPatient
Age Patient age in years Patient
MDs on shift Number of all physicians working at any point during this shift Physician-shift
Current waiting count Number of patients waiting to be seen by this physician at this time Patient
Current patient count Number of patients being seen by this physician at this time Patient
Shift number Indicator for what number shift this is for this physician in this data set Physician-shift
ESI level 1patient present Indicator for presence of ESI level 1 patient (=1 for present, =0 for absent) Patient
Trauma patient present Indicator for presence of trauma patient (=1 for present, =0 for absent) Patient
Arrival shift type Three indicators for type of shift during which patient arrived (morning, afternoon, overnight) Patient
Months since March 2007 Indicator for what number month this is in this data set Patient
Day of week Seven indicators for day of week of shift Patient
Main ED Shift location (=1 for main ED, =0 for rapid care area) Physician-shift
PAS implemented Indicator for whether PAS was implemented (=1 for preimplementation, =0 for postimplementation) Physician-shift
Interaction PAS ×Main ED Physician-shift
Additional dependent variables
Lab ordered Indicator for whether lab was ordered (=1 for ordered, =0 for not ordered) Patient
x-ray ordered Indicator for whether x-ray was ordered (=1 for ordered, =0 for not ordered) Patient
Admitted to hospital Indicator for whether patient was admitted to hospital on discharge from ED (=1 for admitted,
=0 for not admitted)
Patient
Died in ED Indicator for whether patient died in ED (=1 for died in ED, =0 for did not die in ED) Patient
Revisit within 72 hours Indicator for whether patient returned to ED within 72 hours (=1 for returned, =0 for did not return) Patient
Shift duration Number of hours for which physician worked in ED during this shift Physician-shift
ED sojourn time Logged number of minutes elapsed between arrival to ED and bed request (for patients admitted to
hospital) or discharge from ED (for patients discharged home or to an outside facility)
Patient
ED boarding time Logged number of minutes elapsed between bed request and discharge from ED (if admitted to hospital) Patient
aAlthough the ESI uses five categories, we have four indicators for patient ESI level because we exclude patients of ESI level 1 from our analysis.
find that average LOS in the main ED and the RCA,
respectively, exhibit parallel trends in the 17 months
preceding the implementation of PAS.
After establishing the parallel trend assumption
(Abadie 2005,Duflo 2001), we estimate the effect of
transitioning from a pooled to a dedicated queu-
ing system on patients’ average LOS by examining
the coefficient on the interaction term, PASt×mainij .
We predict that this coefficient, 2, is negative and
statistically significant, suggesting that the dedicated
queuing system is associated with a shorter average
LOS than the pooled queuing system.
3.6.2. ED Wait Time. To test Hypotheses 2A
and 2B, we estimate the following difference-in-
differences model at the patient level:
ln waitijt =0+1mainij +2PASt×mainij +Xij t
+t+MDi+ijt 0(2)
In this model, all variables remain the same as in
Equation (1) with the exception of ln waitijt , which
represents the logged number of minutes that patient i
of physician jin month-year tspent in the waiting
room on arrival to the ED. We use the same difference-
in-differences approach as we do in testing Hypothe-
sis 1. Here, we estimate the effect of PAS on patients’
average ED wait time by examining the coefficient on
the difference-in-differences estimator, PASt×mainij.
Hypothesis 2A predicts that this coefficient, 2, is neg-
ative and statistically significant because of an indirect
queuing effect, suggesting that the dedicated queu-
ing system is associated with a shorter average wait
time than the pooled queuing system. Hypothesis 2B
predicts that 2is positive and statistically significant
because of the inefficiency of dedicated queues, sug-
gesting that the dedicated queuing system is associ-
ated with a longer average wait time than the pooled
queuing system.
3.6.3. Discharge Rate. To test Hypothesis 3, we
estimate the following model at the physician-shift
two-hour period level:
ln dischratekj =0+1PAS +Xkj +MDk+kj0(3)
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3042 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
Here, dischratekj represents the number of patients dis-
charged per hour by physician jin a given two-hour
shift period k;PAS indicates whether PAS had been
implemented; ’s and ’s represent vectors of coeffi-
cients; and all other variables remain the same. For
this analysis, we limit the sample to patients seen
in the main ED and conduct a pre-post analysis. We
do not employ a difference-in-differences approach
because the different discharge processes in the main
ED and the RCA make the comparison difficult, and
because we are interested in the change in discharge
rates during each of the two-hour periods over the
course of a physician’s shift rather than the change
in the average discharge rate of a physician’s shift.
Therefore, we estimate Equation (3) separately for the
first, second, penultimate, and final two-hour peri-
ods of a physician’s shift. This allows us to examine
whether and at what point during a physician’s shift
the implementation of PAS in the main ED affects
the discharge rate of patients. Because the discharge
rate is small and discrete and because the data are
not overdispersed, we employ a Poisson model with
physician fixed effects.
If the dedicated queuing system results in a reduc-
tion in patients’ average LOS, we would expect the
discharge rate in each of the two-hour periods of a
physician’s main ED shift to increase after PAS imple-
mentation. This is because, after PAS implementation,
physicians are more likely to engage in strategic behav-
iors throughout the shift to ensure that their patients’
average LOS is as short as possible. However, because
many of the preliminary tasks may be unaffected by
the post-PAS increase in ownership, we expect that
the discharge rate may be unaffected in the first two-
hour period of a physician’s shift. Accordingly, we pre-
dict that the coefficient on PAS will be positive and
statistically significant for each of the second, penulti-
mate, and final two-hour periods of a physician’s shift,
whereas it will not exhibit a statistically significant
change for the first two-hour period of a shift.
3.6.4. Additional Analyses. To better understand
our main findings and consider possible alternate
explanations, we conduct several additional analy-
ses. We begin by considering two competing explana-
tions that could account for the decrease in average
LOS post-PAS. First, patients might have experienced
shorter LOSs in the ED because physicians “cut cor-
ners” by stinting on care (Oliva and Sterman 2001).
We assess this possibility by estimating Equation (1)
with two different dependent variables, both mea-
sured at the patient level: whether labs are ordered for
a patient and whether x-rays are ordered for a patient.
Data on whether labs or x-rays are ordered for a
patient are obtained directly from the hospital’s EMR
system. For each of these variables, we estimate Equa-
tion (1) as a logistic regression because both are binary
indicator variables. Second, we consider whether the
decrease in LOS stems from physicians shifting their
work onto other clinicians. In the context of the ED,
the most plausible scenario is ED physicians admit-
ting more patients to the hospital, so that patients
appear to stay in the ED for a shorter period of time.
We examine this possibility by estimating Equation (1)
with admission to the hospital as the dependent vari-
able. Data for whether the patient is admitted to the
hospital comes from the EMR system and is mea-
sured at the patient level. Again, we estimate a logis-
tic regression because admission to the hospital is a
binary dependent variable. Next, we examine the pos-
sibility of the quality of care in the ED declining as
an unintended consequence of PAS implementation
in the main ED. As proxies for quality, we exam-
ine whether the patient returned to the ED within
72 hours after an initial visit and whether the patient
died in the ED. We estimate Equation (1) as a logistic
regression with each of these binary indicators as the
dependent variable, respectively. For the analysis of
ED revisits, we employ a 72-hour time period, which
is the standard quality metric used to capture return-
ing ED patients (Keith et al. 1989). For the analysis
of patient mortality in the ED, we include a subset
of previously excluded patient-level observations—
specifically patients of ESI level 1, patients who died
in the ED, and trauma patients.
In addition, we consider the potential impact of
PAS on the duration of a physician’s shift, which
is measured as the number of hours for which a
physician worked in the ED during a particular shift.
Though this does not directly address why having
a dedicated queuing system may decrease patients’
average LOS, it is an important consideration if imple-
menting a similar system at other EDs. If having a
dedicated queuing system results in physicians stay-
ing longer to finish caring for their assigned patients,
it may not be feasible to implement elsewhere for rea-
sons of cost and physician burnout. To assess this
possibility, we estimate a regression of a similar form
as Equation (1) but with the shift duration as the
dependent variable and at the physician-shift level.
We use the shift duration and not the log of shift
duration because the variable is normally distributed.
We estimate this regression at the physician-shift level
because the dependent variable (i.e., shift duration)
is calculated at this level. If physicians are working
longer hours as a result of PAS implementation, we
would expect to see a positive and statistically signif-
icant coefficient on the interaction term, PAS ×main.
Finally, we examine the impact of PAS implemen-
tation on sojourn time, which is the sum of ED
wait time and LOS. We also examine the impact of
the queue structure on ED boarding time to assess
whether the change results in an admitted patient
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3043
waiting longer for an inpatient bed. We estimate
Equation (1) with logged ED sojourn time and logged
ED boarding time, respectively, as the dependent
variable.
4. Results
4.1. Descriptive Statistics
Table 2(a) presents means and standard deviations
for all continuous variables included in the empirical
models, stratified by location (main ED or RCA) and
time period (pre-PAS or post-PAS). Table 2(b) presents
the correlations between all continuous variables
included in the empirical models. Table 2(c) presents
percentages for all categorical or binary variables in
the empirical models stratified by location and time
period. As shown in Table 2(a), the average LOS
for a patient seen in the main ED is approximately
three and a half hours, and it is about 50 minutes
for a patient seen in the RCA. There are, on aver-
age, three or four physicians staffing the main ED
during a given eight-hour period (i.e., morning shift,
Table 2(a) Summary Statistics of Continuous Variables Included in Models
Main ED RCA
Pre-PAS Post-PAS Pre-PAS Post-PAS
Variable Mean SD Mean SD Mean SD Mean SD
1. ED length of stay (minutes) 21207 21007 21003 22703 4606 6501 4608 6103
2. ED wait time (minutes) 4309 4209 3306 3004 5408 4306 4600 3304
3. Discharge rate (patients/hour) 108 009 108 008 303 103 304 104
4. Age (years) 4303 2403 4205 2406 2804 2006 2600 2002
5. MDs on shift 304 009 307 100 100 002 100 003
6. Current waiting count 109 101 107 009 309 206 305 203
7. Current patient count 503 207 503 207 602 304 509 301
8. Shift number 11501 7200 33505 15105 13507 7804 37308 12608
9. Shift duration (hours) 907 105 902 103 1002 102 1000 100
10. ED sojourn time (minutes) 25607 21004 24309 22602 10104 7802 9207 6900
11. ED boarding time (minutes) 32900 41807 16503 25200 25603 39000 12205 24904
Notes.N=2171213. Excludes all observations from August 2008 to account for an acclimation period.
Table 2(b) Correlations of Continuous Variables Included in Models
Variable 1 2 3 4 5 6 7 8 9 10 11
1. ED length of stay (minutes) 1
2. ED wait time (minutes) −0013∗1
3. Discharge rate (patients/hour) −0022∗0022∗1
4. Age (years) 0030∗−0012∗−0016∗1
5. MDs on shift −0005∗0007∗0013∗−0002∗1
6. Current waiting count −0019∗0050∗0052∗−0018∗0013∗1
7. Current patient count −0004∗0033∗0045∗−0006∗0008∗0061∗1
8. Shift number −0004∗−0007∗0008∗−0005∗0021∗0001∗00004∗1
9. Shift duration (hours) −0011∗0011∗0019∗−0005∗0012∗0016∗0010∗−0004∗1
10. ED sojourn time (minutes) 0098∗0005∗−0015∗0028∗−0004∗−0010∗0002∗−0006∗−0009∗1
11. ED boarding time (minutes) 0040∗0007∗−0002∗0007∗−0002∗0006∗0001 −0017∗−00003∗0041∗1
Notes.N=2171213. Excludes all observations from August 2008 to account for an acclimation period.
∗p < 0005.
afternoon shift, overnight shift), and one physician
staffing the RCA. None of the correlations between
variables in the same regression model have levels
close to or higher than 0.80, minimizing concerns
about multicollinearity (see Table 2(b)). We also check
for multicollinearity by calculating variance inflation
factors (VIF). The largest VIF is 5.45 and the mean
VIF is 2.52 (not shown), both of which fall well below
the conventional threshold of 10, providing addi-
tional evidence that multicollinearity is not a con-
cern (Wooldridge 2012). As Table 2(c) shows, nearly
75% of main ED patients are of ESI level 3, with
the remainder being predominantly split between ESI
levels 2 and 4. About 65% of main ED patients had
a lab ordered compared to less than 9% of RCA
patients.
As expected, patients’ average LOS differs signifi-
cantly by their acuteness. Although, for brevity, we do
not display the numbers in a table, we find that for
patients of ESI levels 2 to 5, the relationship between
LOS and ESI level is a generally monotonically
increasing one, with patients of a higher acuteness
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3044 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
Table 2(c) Percent of Sample by Categorical and Binary Variables
Included in Models
Main ED RCA
Variable Pre-PAS Post-PAS Pre-PAS Post-PAS
ESI level 2 7088 14005 — —
ESI level 3 74010 73070 — —
ESI level 4 17068 11085 96023 96053
ESI level 5 0034 0040 3077 3047
ESI level 1patient present 8078 9016 9069 9092
Trauma patient present 7036 27068 7062 29025
Morning shift 34021 35055 40058 37073
Afternoon shift 44059 43063 53087 55099
Overnight shift 21020 20082 5055 6028
2007 57092 — 56023 —
2008a42008 14039 43077 16016
2009a— 52047 — 54080
2010 — 33015 — 29004
January 5078 8073 6035 8059
February 6018 8045 6063 8018
Marcha12038 9063 11081 9017
Aprila11074 9018 11018 8070
May a12007 9085 12024 9036
Junea11063 8093 11025 8045
July a12012 9019 11044 8087
August a5088 4050 5096 4072
September 5066 8010 5072 8068
October 5055 8001 5077 9017
November 5046 7072 5075 8034
December 5055 7069 5089 7076
Sunday 15015 14075 15001 15009
Monday 14089 15019 15000 15022
Tuesday 14008 14011 14050 14007
Wednesday 13093 13040 13072 13058
Thursday 13084 13086 13089 13040
Friday 13084 13095 13037 13033
Saturday 14027 14073 14051 15031
Lab ordered 64012 66091 8083 8005
x-ray ordered 38044 39046 27037 26092
Admitted to hospital 14011 12042 0038 0030
Revisit within 72 hours 4099 5004 2089 2077
Notes.N=2171213. Excludes all observations from August 2008 to account
for an acclimation period.
aBecause the study period begins on March 1, 2007 and ends on July 31,
2010, it is not surprising that a larger percentage of patients in our data
set presented to the ED in the months between March and July (inclusive)
and in the years of 2008 and 2009, respectively. Because all observations
from August 2008 have been excluded, it is also not surprising that this per-
centage is smaller for the month of August. When these summary statistics
are produced with the inclusion of all observations from January 1, 2007,
to December 31, 2010, we obtain an approximately uniform distribution of
patients across all months of the year.
having a longer LOS. We account for the nonlinearity
of this relationship by adjusting for patient acuteness
using a dummy variable for each ESI level.
4.2. Patient Assignment System
Implementation in the Main ED
Both the qualitative and quantitative data suggest that
PAS was implemented as described, though not with-
out challenges. An ED physician remarked on one
of the key challenges during implementation: “[PAS]
was the hardest thing we have ever done. When we
first started with the PAS system, it was a rocky road
because sometimes there were patients in the waiting
room when there was an open bed.” This comment,
in combination with the first author’s observations
of the ED workflow, suggests that physicians largely
abided by PAS and the RR routing policy. In our EMR
data, we find further support for the general adher-
ence to the RR routing policy. In particular, patient
demographics across physicians are well balanced
and there is little variation in the average acuteness
of patients assigned to each physician, suggesting it is
not the case that certain physicians are being assigned
particular types of patients. Furthermore, on average,
there are only one or two ESI level 2 patients seen
by a physician on a given main ED shift (mean =104,
s0d0=005), suggesting the workload across physicians
remains relatively balanced, thus allowing physicians
to feasibly adhere to the RR routing policy.
However, there are rare situations when the RR
routing policy is violated. Although the internal triage
nurse cannot bypass the patient assignment gener-
ated by the computer system, physicians working in
the main ED can bypass the RR assignment deter-
mined by the computer when another physician has
an exceptionally time-consuming workload of ESI
level 1 patients. One physician stated the follow-
ing: “The expectation is that each physician sees the
patients assigned to him or her. Ninety-nine percent
of the time, this happens 000[but] we help each other
if someone gets slammed with a critical [ESI level 1]
patient 000 I remember one case last year where a
physician got three critical patients in a row. That is
extremely rare. He did not ask anyone, but two of his
colleagues came and took two of the three patients
[onto their panels].” This corroborates our under-
standing of the RR routing policy, in which other
physicians can voluntarily take on the next patient
assigned to a physician caring for an ESI level 1
patient.
4.3. Base Results
4.3.1. ED LOS. We estimate Equation (1) to assess
the impact of having a pooled queuing system (ver-
sus a dedicated queuing system) on patients’ average
LOS in the main ED. Column (1) of Table 3presents
a fixed-effects model that captures the effect of mov-
ing from a pooled to a dedicated queuing system.
We find that the difference in patients’ average LOS
between the main ED and the RCA is greater prior
to PAS implementation. Once the main ED adopts a
dedicated queuing system, this difference in patients’
average LOS is reduced. This difference in differences
is captured by the coefficient on the interaction term,
PAS ×main (2= −0017, p < 00001), and indicates that
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3045
Table 3 Fixed-Effects Models at Patient Level
(1) (2)
Variables Logged ED length of stay Logged ED wait time
Main ED 00642∗∗∗ 00377∗∗∗
40003075 40003305
PAS ×Main ED −00174∗∗∗ −000854∗∗
40002115 40002655
ESI level 3−00401∗∗∗ 00415∗∗∗
40001595 40001295
ESI level 4−10211∗∗∗ 00698∗∗∗
40002035 40002415
ESI level 5−10578∗∗∗ 00617∗∗∗
40002525 40002915
Age 0000773∗∗∗ −0000260∗∗∗
4000002335 4000001885
MDs on shift −0000559 000148
400003025 400008555
Current waiting count 0000184 00189∗∗∗
400001715 400005615
Current patient count 00000909 000201∗∗∗
400001675 400004765
Shift number −00000484∗−2051e−05
4000002365 4000003595
ESI level 1patient present 000169∗∗ 000604∗∗∗
400005285 400008065
Trauma patient present 0000844 000650∗∗∗
400005055 400005625
Afternoon shift −000605∗∗∗ 000726∗∗∗
400007115 40001615
Overnight shift −000731∗∗∗ −00157∗∗∗
40001315 40002835
Constant 40546∗∗∗ 20298∗∗∗
40005385 40006115
Observations 217,161 217,213
Number of ED physicians 40 40
Adjusted R200519 00298
Notes. All regressions are estimated at the patient level and include day of
week controls, month-year fixed effects, and physician fixed effects. Stan-
dard errors (in parentheses) are heteroskedasticity robust and clustered by
physician.
∗p < 0005; ∗∗p < 0001; ∗∗∗p < 00001.
the transition from a pooled queuing system to a
dedicated queuing system is associated with a highly
significant reduction in the difference between the
average LOS in the main ED and the RCA. This 17%
decrease in the difference in average LOS in the main
ED and the RCA after the implementation of PAS
corresponds to a 39-minute decrease in LOS in the
main ED relative to the RCA for an average patient of
ESI level 3 seen by an average physician in the main
ED. In other words, the average patient’s LOS in the
main ED when compared with that in the RCA is sig-
nificantly longer in the pooled queuing system than
in the dedicated queuing system. This result offers
strong support for Hypothesis 1, which predicts that,
in our setting, pooled queuing systems are associated
with a longer average LOS compared to dedicated
queuing systems.
This finding is consistent with strategic changes
in physicians’ behaviors to improve the management
of their overall workflow. After PAS implementation,
physicians change their practice behaviors because
(a) they are aware of their full set of assigned patients,
even those still in the waiting room, and (b) they
have ownership over a designated bank of beds and
nurses. In addition, when one of their designated
beds becomes available because of a patient dis-
charge, physicians post-PAS are responsible for ensur-
ing that their next patient from the waiting room is
placed in that bed as quickly as possible. Specifically,
according to interviews with physicians and obser-
vations of their practice patterns, physicians change
their practice behaviors by (a) proactively “pulling”
for lab results, x-ray results, and specialty consult
results rather than waiting for this information to be
“pushed”; (b) jointly managing their own workflow
with that of the nurses with whom they are paired
to better coordinate various tasks; (c) initiating the
discharge process sooner for patients who are ready
for discharge; and (d) making sure that patients are
brought in from the waiting room as soon as one
of their main ED beds becomes available rather than
waiting for the internal triage nurse to place the next
patient in an open bed. Collectively, these proactive
actions lead to a shorter average LOS for patients in
the main ED and result in a decrease in the difference
in average LOS between the main ED and the RCA.
To confirm that the implementation of PAS only
affected the main ED and not the RCA, a neces-
sary condition for using the difference-in-differences
framework, we conduct two analyses. First, using
a pre-post analysis that is limited to the RCA, we
examine whether there is a discontinuous jump in
LOS in the RCA when PAS is implemented. We find
no evidence of a significant increase or decrease in
LOS in the RCA after PAS implementation (2=0001,
p≈0084). Second, we check for a change in the slope
of LOS trends in the RCA before and after PAS imple-
mentation. A Wald test on the equality of coefficients
also suggests no change in the trend of LOS in the
RCA after PAS implementation (p≈0071). Both of
these findings indicate that the effects of PAS imple-
mentation were limited to the main ED and did not
affect the RCA, thereby validating the use of the
difference-in-differences model.
4.3.2. ED Wait Time. We estimate Equation (2) to
examine the impact of having a pooled queuing sys-
tem on patients’ average wait time in the main ED.
The results are summarized in column (2) of Table 3.
We find that the difference in patients’ average wait
time between the main ED and RCA decreases after
PAS implementation (2= −0009, p < 0001). This 9%
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3046 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
decrease corresponds to a four-minute decrease in
wait time in the main ED relative to the RCA for
an average patient of ESI level 3 seen by an average
physician in the main ED. In other words, the aver-
age patient’s wait time in the main ED when com-
pared with that in the RCA is significantly longer
in the pooled system than in the dedicated system.
This offers strong support for Hypothesis 2A, which
predicts that, in our setting, dedicated queuing sys-
tems are associated with a shorter average wait time
compared to pooled queuing systems. We do not find
support for Hypothesis 2B, which relies on traditional
queuing theory to predict that a pooled queue yields
a shorter average wait time than do dedicated queues.
In the dedicated system, the shorter wait times may
be attained because, instead of waiting for the internal
triage nurse to initiate placing the next patient in an
open bed, physicians operating under PAS are able to
initiate placement of the next patient from their queue
into their newly available bed. Our findings are also
consistent with the expectation of an indirect queuing
effect, where patients experience shorter wait times
because the patients who are receiving care have a
shorter average LOS, which in turn makes beds in the
main ED available sooner.
4.3.3. Discharge Rate. To better understand how
dedicated queuing systems impact patients’ average
LOS, we estimate Equation (3). We examine whether,
and at what point during a physician’s shift, the
implementation of PAS affects the discharge rate of
patients in the main ED.
Columns (1)–(4) of Table 4present fixed-effects
models estimated at the physician-shift two-hour
period level for each of the following four time peri-
ods: the first, second, penultimate, and final two-hour
periods of a physician’s shift. We find that in the
second, penultimate, and final two-hour periods of
a shift, the discharge rate in the main ED exhibits a
significant increase after PAS implementation. Specif-
ically, after PAS implementation, the discharge rate is
1.05 times greater (1=1005, p < 0005) in the second
two hours, 1.07 times greater (1=1007, p < 00001)
in the penultimate two hours, and 1.05 times greater
(1=1005, p < 0001) in the final two hours of a physi-
cian’s main ED shift. We also find that this increase
in discharge rate does not manifest in the first two
hours of a physician’s main ED shift (1=1004, p≈
0012). Based on observations in the ED and the fact
that the average LOS of a patient seen in the main
ED is 211 minutes (i.e., approximately three and a
half hours), the lack of significant difference in dis-
charge rates in the first two hours of a shift may be
due to the fact that the baseline amount of time nec-
essary for patient care in the main ED is greater than
two hours and, therefore, it is difficult for physicians
to have a faster discharge rate during the first two
hours of a shift.
Our findings are thus consistent with Hypothe-
sis 3, which predicts that physicians in a dedicated
queuing system exhibit a higher discharge rate that
is sustained throughout the entire shift, which indi-
cates that physicians are engaging in strategic behav-
iors over the entire course of the shift. This may
be attributable to their greater ownership for patient
flow and the resources needed to manage patient flow
that comes with working in the ED’s dedicated queu-
ing system.
4.4. Consideration of Alternate Explanations and
Unintended Consequences
Though our finding of a reduction in the difference
between main ED and RCA patients’ average LOS in
a dedicated queuing system versus a pooled queuing
system is consistent with an increase in physicians’
strategic behavior to more efficiently manage patient
flow, we consider alternate explanations that could
also be consistent with our finding. We also explore
the possibility of unintended consequences arising
when implementing a dedicated queuing system.
4.4.1. Testing for Changes in the Provision of
Care. First, one possibility is that physicians stint
on care after PAS implementation because of the
increased pressure to care for all patients in their
dedicated queues. If fewer services are provided to
patients, they may stay in the ED for a shorter amount
of time. For example, if a patient who would have
otherwise received an x-ray does not, she would
likely stay in the ED for a shorter duration because
she would not need to wait for the x-ray machine
to become available, have the x-ray taken, and wait
for the radiologist to read the films. If physicians are
stinting on care post-PAS, we would be mistaken to
assume that the reduced LOS stems from an increase
in physicians’ strategic behaviors to more efficiently
manage patient flow.
We do not find strong evidence of stinting on care
after the transition to a dedicated queuing system in
the main ED. In column (1) of Table 5, we examine the
change in a patient’s likelihood of having a lab test
ordered. We find that the coefficient for PAS ×main
is not statistically significant (2= −0008, p≈0007),
suggesting that the difference in the likelihood of hav-
ing a lab test ordered for a patient in the main ED
and the RCA does not change significantly after PAS.
Similarly, in column (2) of Table 5, we do not find a
statistically significant change in a patient’s likelihood
of having an x-ray ordered (2= −0003, p≈0050).
This suggests that there is no meaningful change in
the difference in a patient’s likelihood of receiving
an x-ray between the main ED and the RCA before
and after the implementation of PAS. In combination,
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3047
Table 4 Fixed-Effects Models at Physician-Shift Levels
(1) (2) (3) (4) (5)
Discharge rate in first Discharge rate in second Discharge rate in penultimate Discharge rate in last Shift
Variables two hours of shift two hours of shift two hours of shift two hours of shift duration
Main ED — — — — 10060∗∗∗
4001665
PAS 10042 10053∗10069∗∗∗ 10051∗∗ —
40002755 40002455 40001905 40002045
PAS ×Main ED — — — — −000904
40008555
Percent of ESI level 3patients 10001 10002∗∗∗ 10002∗∗∗ 10001∗000137∗∗∗
4000006415 4000004305 4000004405 4000004045 400001625
Percent of ESI level 4patients 10007∗∗∗ 10008∗∗∗ 10005∗∗∗ 10004∗∗∗ 000325∗∗∗
4000007615 4000005355 4000006685 4000005535 400002335
Percent of ESI level 5patients 10012∗∗∗ 10007∗10001 10006∗000311∗∗∗
400003685 400003235 400002995 400002405 400005375
Average age of patients 10000 10000 10001∗∗∗ 10000 −000208∗∗∗
4000002695 4000002675 4000002955 4000003865 400003095
MDs on shift 00982∗00981∗00976∗∗ 00968∗∗∗ −00331∗∗∗
400007365 400008845 400007305 400006625 40002395
Average waiting count 10015 10014 00999 10001 −00530∗∗∗
40001845 400009615 400009255 400006945 40003645
Average patient count 10094∗∗∗ 10123∗∗∗ 10111∗∗∗ 10104∗∗∗ 00661∗∗∗
40001725 400008615 400005105 400004325 40002605
Shift number 10000 10000 10000 10000 7006e−05
4000002315 4000002265 4000001845 4000001425 4000008055
Percent of time ESI level 1patient present 00996 00995 00980 10022 000495
40002795 40001865 40001725 40001615 40005415
Percent of time trauma patient present 00998 00965∗10000 10016 −000234
40001855 40001575 400009945 40001325 40004795
Afternoon shift 00936∗∗∗ 10064∗∗ 10154∗∗∗ 10156∗∗∗ −00134∗
40001335 40002085 40001995 40001525 40005415
Overnight shift 00806∗∗∗ 10031 10094∗∗∗ 00980 −10834∗∗∗
40002035 40003825 40002255 40002355 4001305
Constant — — — — 60917∗∗∗
4003015
Observations 3,922 8,594 10,675 10,905 14,153
Number of ED physicians 38 39 38 40 40
Adjusted R2— — — — 00329
Notes. Columns (1)–(4) are conditional fixed-effects Poisson models estimated at the physician-shift two-hour period level with linear time trends by month,
day of week controls, physician fixed effects, and heteroskedasticity robust standard errors. Discharge rate reflects the number of patients discharged per hour
by a given physician in a given two-hour period of the shift, and coefficients have been exponentiated to show incident rate ratios. Column (5) is a fixed-effects
linear regression model estimated at the physician-shift level with day of week controls, month-year fixed effects, physician fixed effects, and heteroskedasticity
robust standard errors clustered by physician. Shift duration is expressed in hours.
∗p < 0005; ∗∗p < 0001; ∗∗∗p < 00001.
these results suggest that physicians are not system-
atically stinting on care in the main ED as compared
to the RCA as a result of PAS implementation.
4.4.2. Testing for Changes in the Likelihood of a
Patient’s Admission to Hospital. A second possibil-
ity is that ED physicians may be reducing patients’
average LOS in the ED by passing them off to
other hospital departments earlier. If an ED physician
decides to have a patient admitted to the inpatient
unit for further evaluation, rather than taking the time
to conduct further evaluation while the patient is still
in the ED, the patient’s LOS in the ED may appear to
be shorter than it would be otherwise.
We do not find evidence of main ED patients
exhibiting a higher likelihood of admission to the hos-
pital, relative to RCA patients, after PAS. As shown in
column (3) of Table 5, we find that the difference in a
patient’s likelihood of being admitted to the hospital
when in the main ED versus the RCA does not change
significantly after PAS implementation (2= −0019,
p≈0016).
4.4.3. Testing for Changes in the Quality of Care.
Next, we consider two potential unintended conse-
quences of this transition from a pooled to a dedicated
queuing system in the main ED. We assess whether
patients are more likely to return to the ED within
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3048 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
Table 5 Logistic Regression Models at Patient Level for Alternate Explanations and Unintended Consequences
(1) (2) (3) (4) (5)
Variables Lab ordered x-ray ordered Admitted to hospital Revisit within 72 hours Died in ED
Main ED 10451∗∗∗ −00103∗10673∗∗∗ 00167∗∗∗ —
4001205 40005035 4001025 40004465
PAS — — — — −00669∗
4003015
PAS ×Main ED −000847 −000260 −00188 0000770 —
40004685 40003825 4001325 40005065
ESI level 2 — — — — −50270∗∗∗
4004205
ESI level 3−00693∗∗∗ −00508∗∗∗ −10007∗∗∗ 000381 −70457∗∗∗
40003195 40003345 40003225 40003755 4005385
ESI level 4−20550∗∗∗ −00799∗∗∗ −20929∗∗∗ −00374∗∗∗ −80820∗∗∗
40004305 40004765 40008045 40006335 4100085
ESI level 5−30275∗∗∗ −20348∗∗∗ −50300∗∗∗ −00577∗∗∗ —
40007325 4001175 4009945 4001555
Age 000176∗∗∗ 000221∗∗∗ 000389∗∗∗ 0000125∗000284∗∗∗
4000006525 4000009375 4000006925 4000005155 400004355
MDs on shift −000146 −000205 −0000650 −000106 −000915
40001805 40001135 40001275 40001365 40009745
Current waiting count 000131 0000726 −000187 −000232∗∗ −000480
400007235 400005725 40001115 400008035 40007505
Current patient count −000165∗∗∗ 0000217 −9074e−05 −0000847 −0000174
400004625 400003815 400004355 400004365 40003075
Shift number −00000511 −00000515∗−00000314 8050e−05 −7019e−05
4000003035 4000002605 4000002655 4000001225 4000006885
ESI level 1patient present 000316 −000101 −000432 0000331 −000577
40002765 40001635 40002885 40004675 4004395
Trauma patient present 000117 000241 −0000760 000513 −00204
40001835 40001485 40002865 40003695 4001745
Afternoon shift −000999 000286 000460 −0000747 00226
40005525 40002745 40003385 40003275 4001625
Overnight shift −00175∗∗∗ −000177 00000590 000713 00459
40005325 40003305 40004725 40005595 4002825
Constant −00558∗∗∗ −00767∗∗∗ −40317∗∗∗ −30035∗∗∗ −10687∗
4001605 40007715 4001225 4001315 4007295
Observations 193,807 193,807 193,807 193,807 132,952
Pseudo R200331 000679 00257 000110 00564
Notes. All regressions are logistic regression models estimated at the patient level. Columns (1)–(4) include day of week controls, month-year fixed effects,
and physician fixed effects. Column (5) includes linear time trends by month, day of week controls, and physician fixed effects. Column (5) also includes
previously excluded observations—specifically patients of ESI level 1, patients who died in the ED, and trauma patients. Standard errors (in parentheses) are
heteroskedasticity robust and clustered by physician.
∗p < 0005; ∗∗p < 0001; ∗∗∗p < 00001.
72 hours of being seen, which could be an unintended
consequence of physicians providing lower quality
or insufficient care in order to decrease patient LOS.
Similarly, if physicians are providing lower quality
care such that more patients are dying in the ED, this
truncating effect on LOS may result in a decrease in
the average patient’s LOS in the ED.
We do not find evidence of a lower quality of care
as measured by revisits to the ED within 72 hours. As
is summarized in column (4) of Table 5, we find no
statistically significant changes after PAS implemen-
tation in the difference in the likelihood of returning
to the ED within 72 hours of an initial visit (2=0001,
p≈0088). Even when using a more inclusive cut-
off of seven days (results not shown), we find no
statistically significant changes in the difference in the
likelihood of revisit (2=0007, p≈0011).
In addition, we do not find evidence of a lower
quality of care as measured by mortality in the ED.
These results are presented in column (5) of Table 5.
Because of the lack of variation in the dependent vari-
able among patients of ESI level 5 and patients seen in
the RCA, these two categories of patients are omitted
from the analysis. In the resulting analysis, comparing
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3049
patient mortality in the main ED before and after the
implementation of PAS, we find that the likelihood
of dying in the ED decreased after the transition to
a dedicated queuing system (2= −0067, p < 0005).
This suggests that the quality of care, as measured
by patient mortality in the ED, improved after PAS
was implemented, thereby reducing concerns that the
assignment of patients in the waiting room to a spe-
cific physician might adversely affect patients.
4.4.4. Testing for Potential Impact on the Dura-
tion of a Physician’s Shift. Finally, we consider the
potential impact of PAS on the duration of a physi-
cian’s shift. As summarized in column (5) of Table 4,
we find no statistically significant change in the dif-
ference between the duration of a shift in the main
ED and the RCA before and after PAS implementation
(2= −0009, p≈0030). This suggests that physicians
are not working longer hours in the main ED as a
result of the intervention.
4.5. Specification Tests
To examine the robustness of our main findings about
LOS, we test a variety of other specifications in addi-
tion to the reported models. Because of space con-
straints, these results are not reported in the tables.
First, we use a limited model specification that
includes only patient ESI levels as control vari-
ables. We retain patient ESI levels because the aver-
age acuteness of patients arriving in the main ED
increased over time, whereas that of patients arriving
in the RCA decreased over time. We find that the base
result remains very robust to this limited model spec-
ification (2= −0016, p < 00001), with the magnitude
of the effect decreasing only slightly from 17% to 16%.
We then repeat our estimation of Equation (1) using
nonlogged LOS and bootstrapped standard errors.
With this alternate model specification, we find that
PAS implementation is associated with a 23-minute
reduction in the difference in LOS between the main
ED and the RCA (2= −22073, p < 00001). Even when
not using a log-level specification to account for the
heavily skewed nature of the dependent variable, we
obtain results that are robust to our base findings.
Although our interviews with ED staff suggest that
there were no other interventions besides PAS that
were applied to only the main ED or only the RCA
during the study period (March 1, 2007 to July 31,
2010), we apply our analyses to shorter time frames
around PAS implementation to nullify the possibility
of other effects. When we limit the time frame to three
months, seven months, 12 months, 15 months, and
18 months before and after the intervention, we find
that our base results remain robust to these shorter
time frames (2<−0010, p < 00001).
Next, we repeat our analyses using logged ED
sojourn time to test for the impact of PAS on a more
holistic measure of patient experience. We find that
PAS is associated with a 10% decrease in the differ-
ence between main ED and RCA sojourn times before
and after PAS implementation (2= −0010, p < 00001).
This suggests that when taking both wait time and
LOS into account, PAS is associated with a reduction
in the average time that patients spend in the ED.
In addition, we examine the impact of PAS on logged
ED boarding time, which is the amount of time that
patients being admitted to the hospital spend waiting
for an inpatient bed. We find no statistically signifi-
cant change in the difference in ED boarding times for
patients in the main ED and the RCA before and after
PAS implementation (2= −0025, p≈0009). This is
consistent with our expectation because ED boarding
time is primarily determined by the inpatient unit’s
capacity to admit a new patient rather than ED physi-
cians’ productivity levels.
Next, we limit our sample to those patients seen in
the main ED and conduct a pre-post analysis, com-
paring the average LOS of patients before and after
PAS. We find that our main findings are robust to this
alternate specification that does not use a difference-
in-differences approach, where PAS is associated with
a 5% decrease in LOS in the main ED (2= −0005,
p < 0001).
In addition, our results do not appear to be driven
by differences in patient care delivered in the two
areas of the ED. To examine this, we assess whether
the transition from a pooled system to a dedicated
system differentially affects LOS depending on the
location of a patient’s ED care. To conduct this analy-
sis, we use the same empirical model as Equation (1),
but limit the sample to patients of ESI levels 4 and 5,
and with each independent variable of interest inter-
acted with ESI level 5. We limit the sample to these
patients because they constitute the group of patients
who are potentially seen in both areas of the ED
(because all ESI levels 4 and 5 patients are seen in the
main ED after 11 p.m.). This analysis suggests that
there are no differential effects by the location of a
patient’s ED care (p≈0032).
Furthermore, we examine whether the base results
are sensitive to heterogeneity in patient acuteness. In
other words, we examine whether the transition from
having a pooled queuing system to a dedicated queu-
ing system has a greater impact on patients with a
higher ESI level as opposed to those with a lower ESI
level. Using a similar approach as above, we explore
this possibility by limiting the sample to patients of
ESI levels 2 and 3, and interacting each independent
variable of interest with ESI level 3. For this anal-
ysis, we limit the sample to patients of these two
ESI levels because they exhibit two different groups
with relatively longer LOS (for ESI level 2, mean =
332 minutes, s0d0=330 minutes) and shorter LOS (for
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3050 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
ESI level 3, mean =212 minutes, s0d0=202 minutes).
This analysis suggests that patients of higher acute-
ness (ESI level 2) are likely to experience a greater
decrease in LOS after the implementation of PAS com-
pared with patients of a relatively lower acuteness
(ESI level 3) (=0042, p < 0001). Although it is beyond
the scope of this paper to examine why this hetero-
geneity arises, we speculate that it may be due to the
prioritization of higher acuteness patients (ESI level 2)
within each physician’s dedicated queue.
We also repeat our analyses using several different
exclusion criteria in constructing our sample and find
that our results are robust in all of the following anal-
yses. First, we include all observations that had pre-
viously been excluded as outliers (i.e., patients with a
LOS greater than 48 hours). Then, to test our hypothe-
ses on an even more homogeneous set of patients
and ensure that our findings are not driven by out-
liers, we exclude observations with a LOS greater
than one day (24 hours) and the average duration
of one shift (9.4 hours), respectively. Next, we test
our hypotheses on a sample that includes ESI level 1
patients, which were previously excluded. Finally, we
test our hypotheses on a sample that excludes patients
arriving by ambulance and patients presenting with
a psychological condition, respectively, both of whom
were previously included. All coefficients of inter-
est and their corresponding significance levels remain
robust to these alternate specifications (2<−0013,
p < 00001).
Finally, we use hierarchical linear models, which
specify random effects rather than fixed effects at the
physician level. We conduct this analysis to test each
of our hypotheses with greater efficiency gains. We
use three levels for our multilevel analyses: patient,
physician-shift, and physician. The effect of transi-
tioning from having a pooled queuing system to
a dedicated queuing system remains robust to this
model specification (2= −0018, p < 00001).
5. Discussion and Conclusions
Using three and a half years of data from a hospi-
tal’s ED, we find that patients experience shorter LOS
when physicians work in a dedicated queuing sys-
tem with a fairness constraint as opposed to a pooled
queuing system with the same fairness constraint.
Although we are unable to precisely test the mech-
anism for the shorter LOS in the dedicated system,
we believe that the improved performance stems from
strategic physician behaviors triggered by physicians’
greater ownership over patient flow and the resources
needed to smooth flow through the ED. This sug-
gests that the flow management benefits associated
with a dedicated queuing system with a fairness con-
straint may outweigh the variability-buffering bene-
fits of a pooled queuing system. We consider, but find
no empirical support for, alternate explanations for
this reduction in LOS, such as changes in the provi-
sion of care or lower quality care.
We find evidence that physicians’ strategic behav-
iors persist throughout the entire shift. In particular,
examination of physicians’ discharge rates in two-
hour periods over the course of the shift shows that
physicians exhibit a higher discharge rate when work-
ing in a dedicated queuing system as opposed to a
pooled queuing system soon after beginning the shift.
This increase in discharge rates is sustained through-
out the remainder of the shift. In describing how the
implementation of PAS increases physicians’ ability to
manage patient flow, one physician said,
Before PAS, the physician had no control or responsi-
bility over getting the next patient into an empty bed.
I often had idle time and had more than enough time
to see more patients; I just couldn’t get them to me
from the waiting room. I wasn’t in control so I didn’t
do much to get patient turnover to happen faster. Now,
with PAS, I am responsible for getting my patients
from the waiting room into my beds. I do this by mak-
ing sure that tasks are being done so that I can dis-
charge my current patients0 0 0 0 It changed the whole
responsibility for patient flow from [the] one [internal
triage] nurse onto me to manage my patients.
To quantify the impact of our findings, we calcu-
late effect sizes. We find that moving from a pooled
queuing system to a dedicated queuing system is
associated with a 17% decrease in the difference in
LOS between the main ED and RCA. For an aver-
age patient of ESI level 3 seen in the main ED by
an average physician, this corresponds to a 39-minute
decrease in LOS in the main ED relative to the RCA.
This is a particularly meaningful difference in the con-
text of a hospital’s emergency room. With approx-
imately 200 patients in the ED every day, this is
roughly equivalent to an additional 130 patient-hours
per day that are saved with the dedicated queuing
system. Once we take into account the large costs
associated with emergency room care, it becomes
clear that the time and cost implications are substan-
tial. If these findings are generalizable to other EDs,
this would have significant practical implications for
EDs across the country faced with large increases in
patient volume accompanied by constrained budgets.
Nevertheless, it is important to consider the poten-
tial limitations of dedicated queuing systems. In sys-
tems with less homogenous patient populations, a
dedicated queuing system with fairness constraints
might result in imbalanced workloads among differ-
ent care providers.
5.1. Theoretical Contributions
This paper contributes to the operations management
literature on queue pooling in several ways. Our paper
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3051
is one of a few to use empirical data to examine the
effect of queue management systems on wait times and
service times. We find that when servers have own-
ership over patient flow and key resources, dedicated
queuing systems with a fairness constraint are asso-
ciated with shorter wait times and service times than
pooled systems with a fairness constraint. Our find-
ings illustrate the importance of accounting for the
interaction between human behavior and queuing sys-
tem design when predicting performance (Boudreau
et al. 2003,Jouini et al. 2008). When queuing the-
ory does not account for strategic server behavior, it
suggests that pooling queues should result in shorter
wait times even when fair routing policies are used
(Armony and Ward 2010). In our study, we find that
wait times are longer for the pooled system. Thus, our
paper provides empirical support for prior analytical
models that predict that human behaviors can reduce
the benefits of using a common pool (Best et al. 2015,
Cachon and Zhang 2007,Gilbert and Weng 1998,Hopp
et al. 2007,Jouini et al. 2008,Wang et al. 2010). We
are also able to add quantitative, empirical evidence to
the debate that the benefits that arise from lean man-
ufacturing’s practice of assigning a specific person to
service a specific stream of work outperforms the flexi-
bility benefits from a pooled system (Spear and Bowen
1999). Our paper demonstrates how employees’ will-
ingness and ability to manage flow create an advan-
tage for dedicated systems over pooled systems.
We speculate that queue pooling results in longer
LOS because, in the pooled system, physicians do not
feel completely responsible for patient flow because
the internal triage nurse is responsible for moving
patients from the waiting room to available beds. This
result is similar to, but distinct from, Chan’s (2015)
finding that ED physicians work slower when they
are assigned patients by a triage nurse than when
physicians—collectively as a group—assign patients
to physicians. Chan asserts that this “foot-dragging”
behavior occurs in the nurse-managed system because
physicians delay discharges to overstate their true
workload to the nurse in hopes of avoiding being
assigned another patient. The findings in the Debo
et al. (2008) study are also driven by servers’ mislead-
ing behaviors. Another mechanism in the literature
that explains why dedicated queuing systems have
faster service times than pooled systems is that man-
agers can better supervise the smaller teams of work-
ers that result from splitting up a large pooled system
into a set of dedicated systems and a healthy compe-
tition emerges among the different dedicated systems
(Jouini et al. 2008).
In contrast, we propose a different underlying
mechanism for the improvement in throughput times:
better flow management arising from strategic physi-
cian behaviors. In our study, a computer-automated
RR routing policy fairly assigns patients to physi-
cians both before and after the intervention. Thus,
unlike physicians in Chan’s (2015) study, physicians
in our study are not deliberately working slower to
overstate their workloads. Furthermore, the fact that
only a handful of physicians are working in this
ED at any one time suggests that the Jouini et al.
(2008) emphasis on the challenge of managing a large
pool of employees is not what is driving our results.
Also, physicians were not given any information
about other physicians’ average LOS, so competition
is not the explanatory mechanism (Jouini et al. 2008).
Instead, we propose that making a single physician—
as opposed to a group of physicians—accountable for
efficiently managing patient flow leads to a reduction
in wait time and LOS through better flow manage-
ment practices.
Our findings build on the Schultz et al. (1998) study
of the motivational impact of low inventory levels
on production line workers’ speeds. Schultz et al.
(1998) find that low inventory motivates slower work-
ers to speed up, enough to cancel the productivity
loss due to the blocking and starving that occurs in
low inventory production lines. We examine a differ-
ent lever to increase workers’ motivation: the queue
structure of incoming jobs. We find that, when physi-
cians work in a dedicated queuing system, they are
able to attain shorter average LOS and wait times
for their patients by managing their workloads more
efficiently. We suggest that this may be because the
dedicated system affords physicians a higher level of
ownership over patient flow. We find that the motiva-
tional benefits of the dedicated queuing system out-
weigh the inefficiencies introduced by unpooling the
queue. Thus, our study furthers the Schultz et al.
(1998) finding by proposing that queue structure is
another job design factor that interacts with human
behavior in ways that can reverse predicted relation-
ships between work system design and performance.
5.2. Implications for Practice
Our study has important implications for workplace
managers and healthcare policy makers. Our findings
suggest that managers of work settings with strate-
gic servers should design work systems to mitigate
behaviors that benefit the employee to the detriment
of customers or the organization. One possible mech-
anism is to give strategic servers greater ownership
and responsibility for managing their workflow and
to route work evenly across all servers regardless of
differences in work pace, which removes the benefit
of working slower than one’s peers. EDs may benefit
from implementing dedicated, fair queuing systems
in which patients are assigned to physicians immedi-
ately following triage. To our knowledge, this is not
currently in place at most EDs; most EDs employ a
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
3052 Management Science 61(12), pp. 3032–3053, © 2015 INFORMS
pooled queuing system that assigns patients to physi-
cians once placed in a bed. Thus, the potential for
improvement is significant.
5.3. Limitations and Future Research
This study has limitations, and its results should
be interpreted accordingly. First, we note the threat
of omitted variable bias, common to many empir-
ical models. Although it would have been helpful
to include more patient characteristics in our model,
such as patient diagnoses or medical comorbidities,
these data were protected information and not avail-
able for use. However, this is not an important threat
to validity because patients are randomly assigned to
physicians rather than by physician choice. This is
supported by the fact that the average ESI level of
patients seen by each physician is less than one stan-
dard deviation away from the average ESI level of all
patients seen in the ED (mean =3033, s0d0=0064).
Second, our study is of a single hospital’s ED and
its response to a single intervention. The fact that our
data come from a single ED makes it impossible for us
to use another ED as the control in our difference-in-
differences analysis. Although we are confident that
the RCA is a good control for our study, there would
be advantages to using data from another ED with
a similar patient population that did not implement
the PAS system. We were unable to do this in our
study because the PAS system was implemented in all
EDs in the hospital’s network. Though the generaliz-
ability of our findings is limited because we studied
only one ED, we believe our findings have strong the-
oretical underpinnings. Nevertheless, future research
could examine a larger sample of EDs to study a
wider variety of routing policies and queue struc-
tures. Given that prior literature has found a variety
of different mechanisms that may explain the shorter
service times in dedicated systems, such research
might enable greater clarity in which mechanisms are
most powerful and under what conditions. In addi-
tion, these effects and suggested mechanisms could
be studied in different empirical contexts for further
theory development.
Third, our study raises the possibility that better
flow management—arising from ownership over key
resources—enables physicians in dedicated queuing
systems to reduce their patients’ average wait times
and LOS. However, we are unable to precisely identify
and test the mechanisms conclusively. Instead, we sug-
gest these potential mechanisms based on interviews
with physicians and observations of their practice pat-
terns and leave it to future research to disentangle the
mechanisms responsible for the reduced times.
Fourth, future research could consider how dedi-
cated queuing systems affect patient and physician
satisfaction, since changes in wait times and LOS may
be associated with perceptions of fairness and the
general satisfaction of both parties. These data are not
available from the time period of our study, but have
recently become more widely available.
Finally, implementing a dedicated queuing system
is merely one way to try to attain the goal of shorter
wait times and LOS in EDs. Future research should
consider other mechanisms, such as financial incen-
tives or interventions that leverage social pressure
(Chan 2015). For example, do physicians increase their
work rates when provided information about each
other’s average LOS? It may be possible to use a com-
bination of interventions so that EDs can capture the
benefits of pooling while simultaneously avoiding the
slower service rates that seem to arise from queuing
systems where responsibility for customers is shared
across multiple servers.
5.4. Conclusions
Effectively using queue design to create both fairness
and efficiency is an important opportunity for ser-
vice organizations. Although results may differ across
different settings, the mechanisms through which
changes in LOS occur may help shed light on improve-
ment opportunities in other contexts. Our findings are
especially timely and could have significant implica-
tions for healthcare delivery as EDs across the country
contemplate ways to handle the anticipated increases
in ED patient volume as a result of the recent health
reform legislation (Patient Protection and Affordable
Care Act 2010, Pub. L. 111–148).
Acknowledgments
This research would not have been possible without the
collaboration of Kaiser Permanente Northern California. In
particular, the authors thank Mark B. Kauffman for his sup-
port and Brent E. Soon for his assistance in preparing the
ED data. The authors thank Gérard P. Cachon, Laurens G.
Debo, Wallace J. Hopp, Robert S. Huckman, Alexandra A.
Killewald, Rajiv Kohli, Avishai Mandelbaum, Nirup Menon,
Charles Noon, Tom Tan, Jan A. Van Mieghem; participants
in the Longitudinal Data Analysis course at Harvard Uni-
versity; seminar participants at the 2013 INFORMS Health-
care Conference, the 2013 INFORMS MSOM Conference, the
73rd Annual Meeting of the Academy of Management, the
2013 INFORMS Annual Meeting, the Harvard Health Pol-
icy Research Seminar, and the 2014 AcademyHealth Annual
Research Meeting; and the editor, associate editor, and
three anonymous reviewers for their insightful comments.
The authors also thank Simo Goshev, Tomoko Harigaya,
Andrew Marder, and William B. Simpson for their advice
regarding data analysis methods and Lydia Ypsse Kim for
her expert research assistance. The authors acknowledge
support for this research from the Division of Research and
Faculty Development at Harvard Business School.
References
Abadie A (2005) Semiparametric difference-in-differences estima-
tors. Rev. Econom. Stud. 72(1):1–19.
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.
Song, Tucker, and Murrell: The Diseconomies of Queue Pooling
Management Science 61(12), pp. 3032–3053, © 2015 INFORMS 3053
Anupindi R, Chopra S, Deshmukh SD, Van Mieghem JA, Zemel E
(2005) Managing Business Process Flows: Principles of Operations
Management, 2nd ed. (Prentice-Hall, Upper Saddle River, NJ).
Armony M, Ward AR (2010) Fair dynamic routing in large-scale
heterogeneous-server systems. Oper. Res. 58(3):624–637.
Ata B, Van Mieghem JA (2008) The value of partial resource
pooling: Should a service network be integrated or product-
focused? Management Sci. 55(1):115–131.
Bassamboo A, Randhawa RS, Van Mieghem JA (2010) Optimal flex-
ibility configurations in newsvendor networks: Going beyond
chaining and pairing. Management Sci. 56(8):1285–1303.
Benjaafar S (1995) Performance bounds for the effectiveness of pool-
ing in multi-processing systems. Eur. J. Oper. Res. 87(2):375–388.
Best TJ, Sandıkçı B, Eisenstein DD, Meltzer DO (2015) Manag-
ing hospital inpatient bed capacity through partitioning care
into focused wings. Manufacturing Service Oper. Management
17(2):157–176.
Boudreau J, Hopp WJ, McClain JO, Thomas LJ (2003) On the inter-
face between operations and human resources management.
Manufacturing Service Oper. Management 5(3):179–202.
Cachon GP, Zhang F (2007) Obtaining fast service in a queueing
system via performance-based allocation of demand. Manage-
ment Sci. 53(3):408–420.
Chan DC (2015) Teamwork and moral hazard: Evidence from the
emergency department. J. Political Econom. Forthcoming.
Debo LG, Toktay LB, Van Wassenhove LN (2008) Queuing for
expert services. Management Sci. 54(8):1497–1512.
Deo S, Jain A, Pendem P (2014) Pacing work in the presence of
goals and deadlines: Econometric analysis of an outpatient
department. Working paper, Indian School of Business, Hyder-
abad, India.
Doroudi S, Gopalakrishnan R, Wierman A (2011) Dispatching to
incentivize fast service in multi-server queues. ACM SIGMET-
RICS Perform. Eval. Rev. 39(3):43–45.
Duflo E (2001) Schooling and labor market consequences of school
construction in Indonesia: Evidence from an unusual policy
experiment. Amer. Econom. Rev. 91(4):795–813.
Eppen GD (1979) Note—Effects of centralization on expected costs
in a multi-location newsboy problem. Management Sci. 25(5):
498–501.
Gans N, Koole G, Mandelbaum A (2003) Telephone call centers:
Tutorial, review, and research prospects. Manufacturing Service
Oper. Management 5(2):79–141.
Gilbert SM, Weng ZK (1998) Incentive effects favor nonconsolidat-
ing queues in a service system: The principal-agent perspec-
tive. Management Sci. 44(12):1662–1669.
Green LV, Nguyen V (2001) Strategies for cutting hospital beds: The
impact on patient service. Health Services Res. 36(2):421–442.
Hackman JR, Oldham GR (1976) Motivation through the design of
work: Test of a theory. Organ. Behav. Human Performance 16(2):
250–279.
Hasija S, Pinker E, Shumsky RA (2010) Work expands to fill the
time available: Capacity estimation and staffing under Parkin-
son’s law. Manufacturing Service Oper. Management 12(1):1–18.
Hopp WJ, Iravani SMR, Liu F (2009) Managing white-collar work:
An operations-oriented survey. Production Oper. Management
18(1):1–32.
Hopp WJ, Iravani SMR, Yuen GY (2007) Operations systems
with discretionary task completion. Management Sci. 53(1):
61–77.
Hyytiä E, Aalto S (2013) Round-robin routing policy: Value
functions and mean performance with job- and server-
specific costs. Proc. 7th Internat. Conf. Performance Evaluation
Methodologies and Tools (ValueTools ’13) (Institute for Com-
puter Sciences, Social-Informatics and Telecommunications
Engineering, Brussels), 69–78.
Jouini O, Dallery Y, Nait-Abdallah R (2008) Analysis of the impact
of team-based organizations in call center management. Man-
agement Sci. 54(2):400–414.
Keith KD, Bocka JJ, Kobernick MS, Krome RL, Ross MA (1989)
Emergency department revisits. Ann. Emergency Medicine 18(9):
964–968.
Kleinrock L (1976) Queueing Systems, Volume 2: Computer Applica-
tions (John Wiley & Sons, New York).
Link S, Naveh E (2006) Standardization and discretion: Does the
environmental standard ISO 14001 lead to performance bene-
fits? IEEE Trans. Engrg. Management 53(4):508–519.
Loch C (1998) Operations management and reengineering. Eur.
Management J. 16(3):306–317.
Mandelbaum A, Reiman MI (1998) On pooling in queueing net-
works. Management Sci. 44(7):971–981.
McCarthy ML, Ding R, Pines JM, Terwiesch C, Sattarian M, Hilton
JA, Lee J, Zeger SL (2012) Provider variation in fast track treat-
ment time. Medical Care 50(1):43–49.
Oliva R, Sterman JD (2001) Cutting corners and working over-
time: Quality erosion in the service industry. Management Sci.
47(7):894–914.
Raz D, Avi-Itzhak B, Levy H (2006) Fairness considerations of
scheduling in multi-server and multi-queue systems. Proc.
1st Internat. Conf. Performance Evaluation Methodolgies and
Tools (ValueTools ’06) (Association for Computing Machinery,
New York), Article 39.
Rothkopf MH, Rech P (1987) Perspectives on queues: Combining
queues is not always beneficial. Oper. Res. 35(6):906–909.
Schultz KL, Juran DC, Boudreau JW, McClain JO, Thomas LJ (1998)
Modeling and worker motivation in JIT production systems.
Management Sci. 44(12):1595–1607.
Spear S, Bowen HK (1999) Decoding the DNA of the Toyota pro-
duction system. Harvard Bus. Rev. 77(5):96–106.
Tan T, Netessine S (2014) When does the devil make work? An
empirical study of the impact of workload on worker produc-
tivity. Management Sci. 60(6):1574–1593.
van Dijk NM, van der Sluis E (2008) To pool or not to pool in call
centers. Production Oper. Management 17(3):296–305.
van Dijk NM, van der Sluis E (2009) Pooling is not the answer. Eur.
J. Oper. Res. 197(1):415–421.
Wang X, Debo LG, Scheller-Wolf A, Smith SF (2010) Design and
analysis of diagnostic service centers. Management Sci. 56(11):
1873–1890.
Wooldridge JM (2010) Econometric Analysis of Cross Section and Panel
Data, 2nd ed. (MIT Press, Cambridge, MA).
Wooldridge JM (2012) Introductory Econometrics: A Modern Approach,
5th ed. (South-Western Cengage Learning, Mason, OH).
Downloaded from informs.org by [129.64.171.135] on 09 December 2015, at 07:26 . For personal use only, all rights reserved.