OPERATIONS RESEARCHVol. 00, No. 0, Xxxxx 0000, pp. 000–000ISSN 0030-364X | EISSN 1526-5463 |00 |0000 |0001

INFORMSDOI 10.1287/xxxx.0000.0000

c© 0000 INFORMS

The impact of size and occupancy of hospital on the extentof ambulance diversion: Theory and evidence

Gad Allon, Sarang Deo, Wuqin LinKellogg School of Management, Northwestern University, Evanston, IL 60208

g-allon, s-deo, wuqin-lin@kellogg.northwestern.edu

In recent years, growth in the demand for emergency medical services along with decline in the number of hospitals withemergency departments (EDs) has raised concerns about the ability of the EDs to provide adequate service. Many EDsfrequently report periods of overcrowding during which they are forced to divert incoming ambulances to neighboringhospitals, a phenomenon known as “ambulance diversion”. The objective of this paper is to study the impact of keyoperational characteristics of the hospitals such as the number of ED beds, the number of inpatient beds, and the utilizationof inpatient beds on the extent to which hospitals go on ambulance diversion. We propose a simple queueing networkmodel to describe the patient flow between the ED and the inpatient department. We analyze this network using twodifferent approximations – diffusion and fluid – to derive two separate sets of measures for inpatient occupancy and EDsize. We use these sets of measures to form hypotheses and test them by estimating a sample selection model using data ona cross-section of hospitals from California. We find that the measures derived from the diffusion approximation providebetter explanation of the data than those derived from the fluid approximation. For this model, we find that the fraction oftime that the ED spends on diversion is decreasing in the spare capacity of the inpatient department and in the size of theED, where both are appropriately normalized for the size of the inpatient department. In addition, controlling for thesehospital-specific factors, we find that the fraction of time on diversion at a hospital increases with the number of hospitalsin its neighborhood. We also find that certain hospitals, owing to their location, ownership and trauma center status, aremore likely to choose ambulance diversion to mitigate overcrowding than others.

Key words: emergency department, empirical research, sample selection model, queueing approximation

1. IntroductionAnnual visits to emergency departments (EDs) in the US increased by 18% from 1994 to 2004 due bothto the growth in population and in per capita consumption of emergency medical services. In the sameperiod, the number of hospitals operating 24 hour EDs declined by 12% (Burt and McCaig 2006). Thishas led to a growing number of hospitals reporting overcrowding situations in their EDs. One of the mostimportant consequences of this overcrowding is the inability of the EDs to accept incoming ambulancesfor extended periods of time, a phenomenon known as ambulance diversion. A number of root causes, bothwithin (shortage of ED beds and staff) as well as outside the ED (shortage of inpatient beds, delays indiagnostic services), have been proposed to explain ED overcrowding and ambulance diversions. However,most of the discussion is limited either to qualitative commentary and surveys (Derlet and Richards 2000,The Lewin Group 2002, GAO 2003, Burt and McCaig 2006) or to single hospital empirical studies (Schullet al. 2003a, Han et al. 2007, McConnell et al. 2005, Forster et al. 2003).

The objective of this paper is to inform this discussion by conducting a cross-sectional analysis about theimpact of various operational characteristics of the hospital on the extent of diversion. Since the hospitalsin the cross-section display considerable heterogeneity with respect to size, occupancy, ownership, locationetc., we find that raw measures of hospital size and occupancy are not sufficient to explain the variation inthe extent of diversion in our data. Hence we make two significant improvements over this basic approach:(i) we employ analytical models to develop appropriate measures of hospital size and occupancy for ourempirical study, and (ii) we endogenize hospitals’ decision to use diversion to mitigate overcrowding. Webriefly describe our approach below.

We develop a two-station queueing model to analyze the flow of patients in the ED and the inpatientdepartment of the hospital. Patients arrive at the ED either by ambulance or by self-transportation and are

1

Allon, Deo, Lin: Ambulance Diversion2 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

either discharged or admitted to the inpatient department after treatment. If inpatient beds are unavailable,these admitted patients continue to occupy beds in the ED thus blocking them for new arrivals. With theobjective of meeting a pre-specified level of delay probability (probability that an arrival has to wait fora bed), the ED decides to go on ambulance diversion if the number of blocked beds reaches a threshold1.Since our objective is to derive appropriate analytical measures of inpatient capacity and ED size for ourempirical study, we cannot use existing queueing approaches as described later.

We consider two approximations – diffusion and fluid – that are appropriate for large and busy systems.Each of these approximations allows us to focus on a different source of variability in the arrival stream. Thediffusion approximation focuses on unpredictable variability (inter-arrival times are not deterministic) andthe fluid approximation focuses on predictable variability (some hours are busier than others) (Mandelbaumand Zeltyn 2009). While in reality both effects are present, limitations in current theory preclude us fromdeveloping a unified model with both types of variability. Analysis of these two approximations results inmeasures of inpatient capacity and ED size and hypotheses about their relationship to the extent of diversionthat are different from each other and are empirically testable. Distinguishing between these two sourcesof variability is useful because often different managerial tools are required to address them in practice:unpredictable variability requires EDs to hold enough excess capacity over average demand as a hedge,while predicable variation requires the flexibility to shift resources across different parts of the day.

We formulate a sample selection model in which the EDs’ decision to use ambulance diversion or not,and the hours on ambulance diversion conditional on this decision, are jointly estimated using aggregatedannual data on a cross-section of California hospitals. We use non-operational factors such as ownership,location and trauma center status in addition to the operational variables described above, to explain thedecision to use ED and the extent of diversion conditional on the decision to use it. In order to make a faircomparison, we also empirically test a basic empirical model with raw measures of the number of ED andinpatient beds, and inpatient occupancy within the framework of the sample selection model.

Our paper makes the following contributions to the literature on emergency department operations:1. Our use of queueing approximations to derive empirically testable hypotheses is novel. It provides a

theoretical basis for deriving measures of inpatient capacity and ED size in a cross-sectional sample ofhospitals with different hospital sizes. Using a statistical test for non-nested models, we find that themeasures of inpatient capacity and ED size derived using the diffusion approximation better describethe differences in use of diversion across hospitals than those derived using the fluid approximationand the basic model.

2. To our knowledge, this is the first cross-sectional study to investigate the drivers of ambulance diver-sion in hospitals. While previous single hospital studies have used longitudinal data to estimate theimpact of inpatient occupancy and ED size on extent of diversion, such rich data are not availablefor a cross-section of hospitals. Consequently, we resort to using aggregate annual data and employanalytical models to derive meaningful relationships at that aggregate level.

3. We theoretically highlight and then empirically document the interaction between the capacity of theinpatient department and the size of the ED as a key factor in contributing to ambulance diversion, i.e.,we find that the magnitude of the marginal impact of the spare capacity of inpatient beds on the extentof diversion is decreasing in the size of the ED. It is not possible to observe this interaction in singlehospital studies due to lack of meaningful variation in ED size.

As a consequence of this interaction, we find that the fraction of time spent on ambulance diversion(counter to empirical findings from single hospital studies) does not necessarily increase in the utilizationof the inpatient department. In particular, if the size of the ED relative to the inpatient department is suf-ficiently small, the fraction of time spent on diversion is not significantly associated with the utilization ofthe inpatient department. These findings suggest that one needs to qualify the findings from single hospitalstudies to a larger population and that a “one size fits all” policy aimed at improving inpatient capacitymight not reduce the extent of ambulance diversion across the board.

1 Such policies are routinely followed in practice (Green 2002, Adams 2008) and are discussed in greater detail in Section 4.1.2.

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 3

The remainder of this paper is organized as follows. We review the relevant literature in Section 2. Weprovide a preview of our data and the plan of analysis in Section 3. The queueing model for the underlyingpatient flow is described in Section 4. The models for unpredictable and predictable variability are developedin Section 5 and Section 6 respectively. The empirical model and the related analysis are described inSection 7. Section 8 contains concluding remarks. Appendix A contains background on heavy traffic theory;Appendix B and C include technical results; Appendix D includes the simulation results.

2. Literature ReviewThis paper contributes to the study of ED operations in operations management and emergency medicineliterature from both theoretical and empirical perspectives.

2.1. Theoretical models of ED operationsConsiderable work has been done on various aspects of managing ambulance service operations from theperspective of an emergency medical services (EMS) agency (Savas 1969, Swoveland et al. 1973, Fitzsim-mons 1973, Green and Kolesar 2004) and on capacity management in EDs and inpatient department withthe objective of meeting a certain delay performance criterion (Green et al. 2006, Vassilocopoulos 1985a).Vassilocopoulos (1985b) and Gerchak et al. (1996) study the problem of allocating inpatient resources (bedsand operating room capacity respectively) when emergency patients cannot be turned away and need to beadmitted immediately.

More recently, several papers have developed detailed simulation models of ED operations (see Marmoret al. 2009, Cochran and Roche 2009) with the objective of providing decision support tools for specifichospitals. However, with a few exceptions, most of these papers do not explicitly model ambulance diver-sion. Kolker (2008) employs a simulation model to investigate the relationship between percent of timeon diversion and length of stay in the ED. Ramirez et al. (2009) also use simulation to examine the im-pact of different diversion policies on the tradeoff between time on diversion and average waiting time.However, these papers focus on individual hospitals our paper differs from these in the use of queueingapproximations to derive hypotheses about extent of diversion and the use of empirical methods to test thesehypotheses using data on a cross-section of hospitals.

2.2. Empirical estimation of queueing modelsThere is limited literature on empirical estimation of queueing models. Joskow (1980) models a hospital asan M/M/s queue and using normal approximation for the delay probability, shows that the average reservemargin of the hospital (number of beds less the average occupancy) varies proportionally to the square-root of the average hospital occupancy. He then estimates this reserve margin as a function of the averageoccupancy and various measures of market competition and regulation. Ramdas and Williams (2008) ex-amine the tradeoff between aircraft utilization and on-time performance for US airlines using a queueingparadigm. In conformance with the predictions from queueing models, they find that utilization of aircraftsis negatively associated with on-time performance and that this effect is more negative for aircrafts withhigher variability in their routes. More recently, Broyles and Cochran (2007) present evidence that renegingin an ED queue can be well estimated by nonlinear regression on the blocking probability in a finite-bufferqueue. Morton and Bevan (2008) study the impact of the number of physicians on the service time, anddemonstrate that service times may increase with the size of the staff.

Our empirical approach is different from these papers in many significant ways. First, formulation of theempirical model (specifically the definition of independent variables) is guided by the theoretical analysis ofthe underlying queueing network in different regimes. Indeed, our empirical results show that raw measuresof inpatient utilization and ED size do not show meaningful association with the extent of diversion. Second,we allow for the possibility that hospital managers may not use queueing rationale in deciding whetherto divert incoming ambulances or not. We use a selection model to endogenize EDs’ decision to go ondiversion and use the queueing framework to estimate the extent of diversion conditional on this decision.

Allon, Deo, Lin: Ambulance Diversion4 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

2.3. Empirical investigation of ED overcrowding and ambulance diversionGiven their prevalence, ED overcrowding and ambulance diversion are among the most actively researchedissues in emergency medicine. A number of indicators have been used to assess its extent: the time patientswait to receive service (waiting time), the total time spent by patients in the ED (length of stay), the per-centage of patients who leave without being seen, the number of patients who remain in the ED after thedecision is made to admit or transfer them (boarding patients), and the number of hours for which incom-ing ambulances are diverted to other hospitals (ambulance diversion) (Derlet et al. 2001, Burt and McCaig2006, GAO 2003). In this paper, we focus on ambulance diversion as the key indicator of ED overcrowdingbecause of its serious impact on various aspects of the public health system and its widespread prevalence.

Ambulance diversion increases transit time for patients (Schull et al. 2003b) and consequently the risk ofpoorer patient outcomes for certain conditions (Schull et al. 2004). It reduces the responsiveness of EMSagencies as ambulances need to find an open ED and/or wait for the ED staff to accept patients (Ecksteinand Chan 2004, Kennedy et al. 2004). It also results in substantial loss of revenue for hospitals since around40% of all hospital admissions come through the ED (Melnick et al. 2004, Merrill and Elixhauser 2005).

In trying to understand the causes of ambulance diversion, Asplin et al. (2003) present a conceptualframework that partitions the ED system into three interdependent components: input (patient arrival),throughput (ED operations) and output (patient disposal including inpatient admission) and Soldberg et al.(2003) report an accompanying comprehensive list of performance measures across the three components.Our queueing model attempts to present a stylized formal representation of this framework and provide therequisite theory to link important measures in each component and derive empirically testable hypotheses.

Recent empirical studies in this literature have focused on either the analysis of time-series data fromsingle hospitals or surveys of multiple hospitals. Using the data from one ED in Toronto, Schull et al.(2003a) found that the number of boarded patients in the ED was associated with duration of ambulancediversion episodes. Han et al. (2007) found that adding ED beds did not reduce the ambulance diversionhours in an urban, academic trauma center. In contrast, McConnell et al. (2005) found that increasing thenumber of ICU beds at an academic acute care facility decreased the time spent on ambulance diversion andreduced the ED length of stay for ICU patients. Similarly Forster et al. (2003) found that higher hospitaloccupancy was associated with shorter length of stay in the ED at an acute care teaching hospital. Whilethese single location studies contribute to our understanding of the causes of ambulance diversion, thegeneralizability of their results to a larger population of EDs is unclear. Moreover, by design, these studiesare not capable of testing the impact of structural characteristics such as the size, type and location ofthe hospital with ambulance diversion. We partly overcome these limitations by conducting our empiricalanalysis using a cross-sectional sample.

Burt and McCaig (2006) also use a cross-sectional sample and their data suggest that the time spent byhospitals on diversion is higher for larger hospitals. This runs counter to the well-known principle of statisti-cal economies of scale in systems susceptible to congestion. According to this principle, larger systems tendto provide better delay performance to their customers than smaller systems for similar levels of utilization.This implies that larger hospitals, everything else being the same, should spend less time on ambulancediversion. We reconcile these two seemingly disparate observations by employing diffusion approximationsof the queueing model, which allow us to appropriately redefine the measures of ED size and inpatientoccupancy by explicitly accounting for the size of the hospital.

3. Data Preview and Plan of AnalysisThe objective of our analysis is to understand the impact of inpatient occupancy and ED size on the extentof ambulance diversion. We study ambulance diversion in hospitals licensed in the state of California. Weuse two separate datasets containing self-reported data on capacity and utilization available from the Officeof Statewide Health Planning and Development (OSHPD). The first dataset called State Utilization DataFile for Hospitals contains basic licensing information (ownership, location and type of the facility) andannual utilization information (licensed bed capacity and patient census for different bed types, number of

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 5

visits to the ED and total number of diversion hours) for 491 hospitals. A second dataset called HospitalAnnual Financial Data File contains financial data (net patient revenue and costs by payer type, balancesheet, staff productivity) in addition to some licensing and utilization data for 451 hospitals2. After mergingthe two datasets and retaining only the records common to both datasets, we end up with 431 hospitals.Since our objective is to study the extent of ambulance diversion in emergency rooms, we drop the hospitalswhich either did not have an ED (zero ED beds reported) or did not operate the ED for the period underconsideration (zero ED visits reported). This results in 318 hospitals being retained for analysis.

Based on earlier studies, we use the average occupancy of ICU beds as a measure of the inpatient uti-lization. Hence, we drop hospitals that do not have an ICU (zero ICU licensed beds reported) retaining 295hospitals for our analysis dropping 23 hospital without an ICU. Comparison of the two groups using t-testshow that hospitals with ICUs have significantly higher number of ED beds and higher number of diversionhours (p < 0.01). Since our analysis focuses only on the hospitals with ICUs, the results of our empiricalanalysis cannot be directly generalized to hospitals without ICU.

We first conducted an empirical analysis using raw measures of inpatient occupancy and ED size butdid not find any statistically significant impact on the extent of diversion. Hence, we decided to develop ananalytical model of patient flow through the ED and the inpatient department to derive more meaningfulmeasures of occupancy and size, which would then be included in the empirical analysis. Given the aggre-gate nature of our data and foreseeable analytical complexity, we formulated a simplified queueing modelto capture the essential elements of this patient flow, which is described next.

4. Model FormulationOur primary objective, here, is to develop a parsimonious representation of patient flow to derive analyticalcharacterizations of appropriate measures of ED size and inpatient capacity and to form empirically testablehypotheses between these measures and the extent of ambulance diversion. Next, we describe two maincomponents of our model: the arrival process and the service process.

4.1. Arrival patternIn most EDs, the arrival rate changes depending on hour of the day (see Green et al. 2006, Figure 1)and also across days for the same hour (see McCarthy et al. 2008, Figure 3). Figure 1 shows a schematicrepresentation of the hourly arrivals to the ED in a typical hospital. Panel (a) shows the average hourlyarrival rate across several days (dashed line) and the actual realization of the hourly arrival rate on oneparticular day (solid line with markers). This arrival process can be thought of as consisting of two types ofvariability. First, predictable variability, refers to the fact that the average number of hourly arrivals variesover the course of a day and follows a certain pattern (Panel b). Second, unpredictable variability, refersto the fact that the actual realization of hourly arrivals on any given day fluctuates around this averagedue to the stochastic variability in interarrival times (Panel c). A natural model to capture both types ofvariability is a nonhomogeneous Poisson process (McCarthy et al. 2008), which, makes the exact analysisof our system (described below) intractable.

4.2. A two-station service modelFigure 2 depicts the model of patient flow through the ED and the inpatient department of the hospital. Forthe summary of all notation, see Table 1. We model the ED as a multi-server station where N1 denotes thenumber of beds in the ED. Patients arrive at the ED either by ambulance (ambulance patients) or by self-transportation (walk-in patients) at rates of λa(t) and λw(t), respectively, where the argument t denotes thatarrival rates can change over time. We assume that the ambulance patients receive non-preemptive priorityover the walk-in patients: An empty ED bed is never allocated to a walk-in patient if there is an ambulance

2 Both utilization and diversion data are reported on an annual and hence aggregate level. While data on these measures might beavailable at a finer level of detail for individual hospitals, it is difficult to obtain such data for a wide cross-section of hospitals. Weconduct a simulation study to demonstrate the impact of using aggregate data in Appendix F.

Allon, Deo, Lin: Ambulance Diversion6 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

Figure 1 Schematic Representation of Arrival Pattern for ED Patients

Predictable variability

Actual Arrival Pattern

Time

(b)Actual Arrival Pattern (b)

TimeUnpredictable variability(a)

Time

( )(c)

Figure 2 A Two-Station Queueing Model

PatientsDischarged

by Hospital

queue 1

queue 2

Inpatient DepartmentED

N1N

λw

λa

1− pa

pa

pw

1− pw

λn

patient waiting. However, a patient cannot be removed from her bed during treatment. After being treated inthe ED, a fraction pa of the ambulance patients and pw of the walk-in patients are admitted to the inpatientdepartment for further treatment and the remaining patients are discharged from the hospital.

We model the inpatient department as a multi-server station where N2 denotes the number of inpatientbeds. The inpatient department receives two streams of arrivals: emergency patients who are admitted fromthe ED and non-emergency patients who arrive directly. If all inpatient beds are occupied, we assume that thenon-emergency patients join a queue whereas the emergency patients continue to occupy ED beds (calledboarding patients) since there is typically no waiting room between the ED and the inpatient department.Anecdotal evidence suggests that hospitals use a threshold on the number of boarding patients to formulatetheir ambulance diversion policy. For instance, Columbia Presbyterian Hospital in New York City goes on

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 7

diversion when 15 or more patients are boarding (Green 2002) whereas Northwestern Memorial Hospital inChicago goes on diversion when 14 or more patients are boarding (Adams 2008). Based on this evidence,we assume that the hospital goes on ambulance diversion if the number of boarding patients in the EDexceeds a predetermined threshold denoted by K.3

However, even this simplified queueing model does not immediately lead to analytical measures of in-patient capacity and ED size that could be used directly in our empirical analysis. Hence, we develop twoapproximations that focus separately on each of the two underlying sources of variability in arrivals. Thediffusion approximation focuses on the unpredictable variability by assuming the average arrival rate doesnot vary over time. The fluid approximation focuses on the predictable variability, i.e., does not model thestochasticity of interarrival times. Each approximation results in different measures of inpatient capacityand ED size and their relationship with the extent of ambulance diversion, which are empirically testable.

5. Unpredictable VariabilityIn this section, we focus on the unpredictable variability arising from the randomness of the interarrivaltimes and service times. We assume that the average arrival rates of ambulances and walk-ins are constantover time and denoted by λa and λw, respectively and that arrivals follow Poisson processes. The servicetime of each patient (irrespective of the mode of arrival) in the ED is exponentially distributed with meanm1 (see Green and Nguyen (2001) for a discussion of this assumption) and that the length of stay of eachpatient (emergency as well as non-emergency) in the inpatient department is exponentially distributed withmean m2. Recent evidence suggests that length of stay might be dependent on the level of congestion inthe inpatient department (Kc and Terwiesch 2009). We do not incorporate this effect to maintain analyticaltractability.

Table 1 Summary of Notation

Notation Descriptionλa arrival rate of ambulance patients

λw arrival rate of walk-in patients

δ diversion probability

N1 number of beds in the ED

N2 number of inpatient beds

m1 mean service time in the emergency department

m2 mean service time in the inpatient department

pa fraction of ambulance patients that are admitted to the inpatient department

pw fraction of walk-in patients that are admitted to the inpatient department

K diversion threshold

B the average number of blocked beds in the ED

β2 (1− ρ2)√N2, normalized excess capacity

κ K/√N2, normalized diversion threshold

ρ2 (λapa +λwpw)m2/N2, traffic intensity

The patient flow model described in Section 4 is a tandem queueing system with finite buffer and blockingafter service. Under the above homogeneity and exponential assumptions, it has been studied extensively

3 In some cases, diversion might be driven by other measures of crowding such as the total number of patients in the ED rather thanthe number of boarding patients. We do not consider such policies here for analytical tractability and ease of exposition.

Allon, Deo, Lin: Ambulance Diversion8 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

in the literature using two main approaches. The first approach involves exact analysis under special casessuch as ‘infinite capacity’ or ‘saturation’ at the first stage, which then requires only analyzing the second(downstream) stage in greater detail. (See Perros (1990), Buzacott and Shanthikumar (1993) and referencestherein.) However, these special cases are too restrictive and do not reflect the situation in our model (neitherdoes the ED have infinite capacity nor is it saturated). The second approach involves approximate analysis ofa multi-stage (single server at each stage) system by decomposing it into one-stage or two-stage subsystemsto improve tractability. These approximations require numerical solution of a system of equations in orderto obtain steady state queue length distribution. While these methods have been shown to be effective ingeneral, they are not adequate for our purpose since they yield numerical solution for the average systemperformance but do not allow for analytical characterization which is needed to develop measures for ourempirical analysis.

5.1. A Simplified Queueing ModelThe exact characterization of dynamics of the two-station queueing network involves a multi-dimensionalMarkov process, where the multiple dimensions correspond to number of patients in the ED, number ofpatients in the inpatient department and number of boarding patients. Hence, like most research in theliterature, we introduce the following approximations to decouple the queueing network and simplify thequeueing dynamics:• We approximate the overall arrival process to the ED by a Poisson process with rate λw + (1− δ)λa,

where δ is the fraction of time on diversion. While the arrival process to the ED is not Poisson due to am-bulance diversion, such approximation can be justified when δλa << λw. This condition is easily satisfiedfor U.S. hospitals since around 15% of the arrivals are ambulance patients and the percentage of time ondiversion is less than 10% (Burt and McCaig 2006, GAO 2003).• We approximate the number of inpatient beds dedicated to emergency patients byN2 = λE

λE+λnN where

λE = λa(1− δ)pa +λwpw is the rate at which emergency patients are admitted to the inpatient departmentand λn is the arrival rate of the non-emergency patients. Thus, this approximation practically divides theinpatient bed capacity into two separate pools for emergency and non-emergency patients in the proportionof the arrival rate of the respective streams.• We approximate the arrival of the emergency patients to the inpatient department by a Poisson pro-

cess with rate λapa + λwpw. This approximation will work well when λa(1− δ)pa >> λwpw, i.e., whenthe majority of emergency patients admitted to the inpatient department are ambulance arrivals. This con-dition seems reasonable for U.S. hospitals since around 40% of the ambulance arrivals and around 10%of the walk-in patients are admitted on average (Burt et al. 2006). We also validate the accuracy of thisapproximation when this condition is violated using numerical simulation (see Appendix D)• We approximate the number of available beds in the ED by N1 −B, where B is the average number

of blocked beds in the ED.Using these approximations, the two-station queueing network in Figure 2 is reduced to two separate

single-station queueing systems, as shown in Figure 3, where the ED is approximated by an M/M/(N1−B)system (station 1) and the inpatient department is approximated by an M/M/N2/K system (station 2).Next, we apply the diffusion approximation to each of these single-station queueing systems4. Diffusionapproximations present tractable approaches to model unpredictable variability through the use of first twomoments – mean and standard deviation – rather than the entire distribution in the traditional queueingtheory. Diffusion approximation has been shown to be a good approximation for the performance measuresof the original stochastic system, provided its traffic intensity is sufficiently close to 1. We refer the readerto Appendix A for background on heavy traffic theory that justifies diffusion approximation, which drawsheavily on Halfin and Whitt (1981) and Whitt (2004).

4 Conventional (non-heavy-traffic) analysis of these sub-systems is cumbersome and does not result in analytical characterizationof the measures of occupancy and size that are required for our empirical analysis.

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 9

Figure 3 A Simplified Queueing Model

Station 2Station 1

λa(1− δ(K)) +λw

N1−B(K)N2

λapa +λwpw

buffer size ≥Kdiverted when

5.2. The diffusion approximationUsing the results from Halfin and Whitt (1981) and Whitt (2004), we approximate the queue length pro-cesses at station 1 and station 2 (Q1 andQ2) by two diffusion processes (Q̃1 and Q̃2) and use them to obtainan approximation for the delay probability (probability that an incoming patient has to wait for a bed) Pdand the percentage of diversion hours δ for the original system. See Halfin and Whitt (1981) and Whitt(2004) for the precise description of Q̃1 and Q̃2.

We first consider station 2 (M/M/N2/K system), which is the approximation of the inpatient department.Following the development in Appendix A and using (7.5) in Whitt (2004), we can approximate the fractionof time on diversion δ by

δ̃(N2, β2, κ) = P(Q̃2(∞)≥N2 +K) =β2e−κβ2

(√N2−β2)

(1− e−κβ2 +β2

Φ(β2)

φ(β2)

) , (1)

where β2 = (1− ρ2)√N2, κ=K/

√N2, where ρ2 = (λapa +λwpw)m2/N2 is the traffic intensity.

We can interpret β2 as a measure of normalized spare capacity in the inpatient department, where√N2

denotes the economies of scale effect - the buffer capacity required to mitigate stochastic (unpredictable)variability grows less than linearly with the size of the system. Similarly, κ can be interpreted as the nor-malized threshold in the ED, again accounting for the economies of scale effect. Below, we summarizekey properties regarding the impact of these underlying measures of occupancy and size on the extent ofdiversion captured in the function δ̃, which we will use for our empirical analysis:

PROPOSITION 1. The function δ̃ satisfies:(i) If Kρ2 ≥ 2, δ̃(N2, β2, κ) is decreasing in β2.

(ii) δ̃(N2, β2, κ) is decreasing in κ.

Note that the condition in result (i) is easily satisfied for ρ2 ≥ 0.8 andK ≥ 3, which is reasonable for mosthospitals. Thus, result (i) states that the fraction of time the ED is on diversion is decreasing in the excesscapacity of the inpatient department, normalized for its size. Similarly, result (ii) implies that increasingthe diversion threshold K (hence κ) will reduce the percentage of time on diversion δ. However, increasingκ might also result in more congestion in the ED and increase the likelihood that an incoming patient hasto wait for a bed (delay probability, Pd). Thus hospitals face a trade-off between two critical performancemeasures, the fraction of time on diversion and the delay probability.

In our model, we assume that the ED chooses the appropriate level of diversion thresholdK∗ to minimizethe fraction of time on diversion while maintaining a certain level of delay probability P̄d since its primarymission is to provide timely service to its arriving patients (Green 2002, Green and Nguyen 2001). Hence,in order to characterizeK∗, we first need to characterize how the delay probability depends on the diversionthreshold K (or equivalently κ).

Allon, Deo, Lin: Ambulance Diversion10 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

For this, consider station 1 (M/M/N1−B system), which is the approximation of the ED. Using Propo-sition 1 of Halfin and Whitt (1981) we approximate the delay probability as

P̃d = P(Q̃1(∞)≥N1−B) =

[1 +

β̃1Φ(β̃1)

φ(β̃1)

]−1

, (2)

where β̃1 depends on N1,N2, β2, and κ. Since the expression is quite complicated, it has been relegatedto Appendix A for the ease of exposition. In what follows, we use the notation P̃d(N1,N2, β2, κ) so as toemphasize P̃d as a function of N1,N2, β2 and κ. The next Proposition formalizes the intuition that higher κyields higher delay probability Pd.

PROPOSITION 2. The function P̃d(N1,N2, β2, κ) is increasing in κ.

Proposition 2 implies that for any N1,N2, β2 there is a unique κ that solves P̃d(N1,N2, β2, κ) = P̄ddenoted by κ∗(N1,N2, β2, P̄d). We can then approximate the appropriate level of diversion threshold byK̃∗(N1,N2, β2, P̄d) =

√N2κ

∗(N1,N2, β2, P̄d).

PROPOSITION 3. The function K̃∗(N1,N2, β2, P̄d) is increasing in N1.

Proposition 3 is quite intuitive and implies that, everything else being equal, a hospital with larger EDwill set a higher diversion threshold in terms of the number of boarding patients5. Using Proposition 3 andresult(ii) of Proposition 1, it is easy to see that the fraction of time on diversion is decreasing in N1√

N2.

To summarize, the analysis of our queueing network model using the diffusion approximation allows usto formulate the following hypotheses regarding how the fraction of time on diversion depends on the keyoperational factors of the system:

1. The fraction of time spent on diversion is decreasing in the spare capacity of the inpatient department,appropriately normalized for its size (β2), as observed from result (i) of Proposition 1.

2. The fraction of time spent on diversion is decreasing in the size of the ED, appropriately normalizedfor the size of the inpatient department (N1/

√N2), as observed from Proposition 3 and result(ii) of

Proposition 1.3. There is an interaction between the two explanatory variables mentioned above, as observed from (1).

6. Predictable VariabilityIn this section, we develop a fluid approximation of the two-station queueing model described in Section4 as an alternative to the diffusion approximation developed in Section 5. Specifically, we approximatethe stochastic arrival (ambulance and walk-in) and service (ED and inpatient department) processes bydeterministic and continuous flows thus focusing on the impact of predictable variability on the extent ofdiversion. Simplifications arising from this approximation allow us to focus on an alternative explanationof diversion, i.e., predictable variations in arrival pattern characterized by a peak period and an off-peakperiod, which results in ED experiencing intermittent periods of overloading and underloading 6. See Figure4 (a) and (b).

We consider a period of length T (say 24 hours) that repeats itself. The peak arrival period is from timet = 0 to t = τ and the off-peak period is from t = τ to t = T . The arrival rates for ambulances and thewalk-ins during the peak period are given by λha and λhw respectively. Similarly, the arrival rates during theoff-peak period are given by λla and λlw. In order to ensure tractability and to eliminate cases which clearlydo not correspond to the data, we make the following approximations and assumptions about the serviceand arrival rates:

5 This result will continue to hold in the presence of patient reneging, i.e., when patients leave without being seen, as long as thereneging rate is independent of the size of the ED. Moreover, one can also see that reneging does not impact the analysis of thesecond stage.6 The fluid model with constant arrival rate will result in a diversion probability of either 0 or 1, which is clearly not supported byour data. Hence, we do not analyze such a stationary fluid model.

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 11

Figure 4 Arrival rate pattern and inventory buildup for fluid model

������������������������������

������������������������������

������������������������������������������������������������

������������������������������������������������������������

(d)

T

Q2(t)

N2 +K

N2

K0

ττ ′

−r2r1λlw

λha

λla

λhw

T

Tτ

τT

Q1(t)

ττ ′

N1

λhw− (N1−K)µ1

λha +λh

w−N1µ1

N1µ1−λla−λl

w

(b)

(a)(c)

• We approximate the service rate of the ED by N1/m1 when not on diversion and by (N1 −K)/m1

when on diversion, where K denotes the diversion threshold based on the blocked beds as before and weapproximate the service rate in the inpatient department by N2/m2.• We assume that capacity in the inpatient department is sufficient to serve the arrivals during the off-

peak period but not enough to serve all the arrivals during the peak period, i.e., λlapa + λlwpw <N2/m2 <λhapa +λhwpw. Note that if this condition is not true, the system would be always either underloaded (henceno diversion) or overloaded (hence always diversion), which are not observed in our data.• We assume that capacity in the inpatient department is enough to serve all walk-in patients, i.e.,

N2/m2 >λhwpw. Clearly, diversion is not useful as a mechanism to mitigate overcrowding in the absence of

this condition.• We assume that capacity in the ED is large enough to serve all those patients who would get admitted

during the peak period, i.e., N1/m1 >λhapa +λhwpw >N2/m2.

Similar to our approach for the diffusion approximation, in order to derive an expression for the diversionprobability, we first consider the dynamics in the second stage for a given diversion threshold K. Then, weconsider the dynamics in the ED and derive the optimal threshold K such that the ED meets an exogenouslyspecified delay probability. Interested reader is referred to Appendix E for further details. Following thefluid approximation allows us to formulate the following hypotheses regarding how the fraction of timespent by the ED on diversion depends on the key structural characteristics of the system:

1. The fraction of the time spent on diversion is decreasing in the square root of the utilization of theinpatient department (

√1− ρ2)

2. The fraction of the time spent on diversion is decreasing in the size of the ED, appropriately normalizedfor the size of the inpatient department (N1/N2)

Allon, Deo, Lin: Ambulance Diversion12 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

Before turning to empirical analysis, it is instructive to compare and contrast the two approximationapproaches. Under both approximations, the extent of diversion is found to be decreasing in the relative sizeof the ED and decreasing in the normalized measure of the utilization of the inpatient department. However,the definition of these measures is quite different under both approaches. In the diffusion approximation,both of these measures are normalized using the square root of the number of inpatient beds. In the fluidapproximation the utilization of the inpatient department is not normalized, whereas the size of the ED isnormalized using the absolute number of inpatient beds.

The main reason for the difference in these relationships is the fact that the diffusion approximationaccounts for the economies of scale effect, i.e., the impact of unpredictable variation reduces as the systemsize increases. This is not the case for the fluid approximation since it does not include unpredictablevariation.

7. Empirical AnalysisOur analysis of the queueing model using approximations from §5 and §6 provides us with empiricallytestable hypotheses regarding the relationships between the fraction of time spent on diversion and variousmeasures of inpatient capacity and ED size. In this section, we attempt to understand which of these hy-potheses are better supported empirically. For comparison, we also estimate a basic empirical model usingraw measures of inpatient capacity and ED size which is not explicitly based on a specific queueing approxi-mation. We first describe a common modeling approach shared by these three empirical specifications. Thenwe describe the specifications themselves followed by construction of the relevant measures. Finally, wedescribe the estimation results for these specifications and conclude with sensitivity analysis and discussionof the limitations of our analysis.

7.1. Modeling ApproachAs discussed earlier, ambulance diversion is one of several options available to EDs to mitigate overcrowd-ing. Other options include creating temporary surge capacity by placing beds and stretchers in hallways andearly discharge for inpatients in stable condition. Some hospitals might adopt a strategic decision to acceptall incoming ambulances irrespective of the extent of overcrowding (GAO 2003) due either to their loca-tion, e.g., rural or their mission, e.g., community hospitals. In other words, it is possible that EDs self-selectthemselves into the sub-sample that has positive diversion hours. Hence estimating the extent of diversiononly for this sub-sample would lead to biased coefficient estimates for the entire population. We mitigatethis problem by endogenizing the EDs’ decision to use ambulance diversion and by estimating it jointlywith the extent of diversion using a selection model (Amemiya 1984, Greene 2008) shown below:

y1i = α′Zi + ε1i (3a)

y2i =

{γ′Xi + ε2i if y1i > 00 otherwise, (3b)

where (3a) is referred to as the choice equation and (3b) is referred to as the level equation. The choiceequation governs whether the EDs choose to employ ambulance diversion (y1 > 0) or not (y1 ≤ 0). Giventhat an ED chooses diversion, the level equation determines the extent of diversion. The set of independentvariables in the choice equation (Z) typically contains variables that are not included in the set of inde-pendent variables in the level equation (X). This is commonly referred to as exclusion restriction in theliterature. We later discuss the basis for exclusion in our model and its implication for our results. Followingconvention about parametric specifications, we assume that the error terms ε1 and ε2 are assumed to be dis-tributed according to a bivariate normal distribution. We assume that only the sign of the selection variabley1 can be inferred but not its magnitude. Hence the selection equation (3a) is reformulated as follows byintroducing another variable w1i where w1i = 1 if y1i > 0 and w1i = 0 otherwise:

P (w1i = 1|Zi) = Φ(α′Zi) , (4)

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 13

where Φ(·) is the cumulative normal distribution.We use a two-step procedure known as Limited Information Maximum Likelihood (LIML) to estimate

our model. This method explicitly recognizes the problem of estimating equation (3b) based on the sub-sample as that of an omitted variable bias (Heckman 1979). The first step involves estimating (4) using a

Probit model and calculating λ̂i =φ(α′Zi)Φ(α′Zi)

known as the Inverse Mills Ratio (IMR). In the second step, IMRis introduced in the level equation (3b), which is then estimated using OLS on the sub-sample. While beingapproximate, it is easier to compute and also more robust to misspecifications (Leung and Yu 2000). Asecond method, called Full Information Maximum Likelihood (FIML) involves the likelihood function overthe bivariate normal distribution of error terms. While this method is more exact, the likelihood functionis not guaranteed to be concave and maximizing it might also present some computational difficulties. Itis also known to be sensitive to deviation from normality of error terms and measurement errors in thedependent variable (Stapleton and Young 1984). For completeness, we report results from this method forall specifications and model families in Appendix G.

7.2. Specifications and measures: Level equationSince the observation period for all hospitals is the same (calendar year 2004), we use total hours on diver-sion as a proxy for the fraction of time on diversion (δ) and take its log transform to reduce the skewness.We denote this measure by (DIV HOURS)7.

For the diffusion approximation, we show in §5 that δ is decreasing in the (normalized) spare capacityof the inpatient department β2 = (1 − ρ2)

√N2 and decreasing in the (normalized) size of the ED N1√

N2.

Similarly, for the fluid approximation, we show that δ is decreasing in the square root of the spare capacityof the inpatient department

√1− ρ2 and increasing in the relative size of the ED N1

N2. Thus, we first need

to construct measures for the underlying variables ρ2,N2 and N1. Here N1 represents the number of bedsin the ED, N2 represents the number of inpatient beds that are relevant to the flow of patient admissionsfrom the ED and ρ2 represents the utilization or occupancy of these beds and then use these to derive themeasures for independent variables of interest.

Our dataset contains the number of licensed inpatient beds for the following categories: general acutecare (GAC), acute psychiatric, skilled nursing and intermediate care beds. GAC is further subclassified intomedical/surgical, pediatric, perinatal, intensive care, coronary care, acute respiratory, burn, rehabilitation,and neonatal intensive care beds. A number of empirical studies have highlighted the impact of ICU bedson the extent of ambulance diversion in the ED (McConnell et al. 2005). Similarly, past surveys have alsohighlighted the lack of ICU beds as one of the most important reasons for going on ambulance diversion(GAO 2003, McManus 2001). Based on these studies, we choose the number of ICU beds to construct ameasure for N2. However, licensed beds typically do not reflect the true bed capacity since hospitals mightstaff only a fraction of all the licensed beds depending on the patient load (Green 2002). Data on staffedbeds is available only in aggregate whereas data on licensed beds is available for all the subcategoriesmentioned above. Hence, we calculate the number of staffed ICU beds (NI)

8 and use it as a measure ofN2. Similarly, we calculate the theoretical utilization of staffed ICU beds ρI using the actual census of ICUpatients reported and the staffed ICU beds NI calculated above.

We then use these measures to construct variables of our interest as follows. For the dif-fusion approximation, we define the relevant independent variables (X) as REL ED SIZE HT( N1√

NI) and NORM ICU CAP ((1− ρI)

√NI). For the fluid approximation, we define (X) as

REL ED SIZE FL (N1N2

) and SQRT ICU CAP (√

1− ρI). Due to the presence of the interaction term(REL ED SIZE HT)*(NORM ICU CAP) for the diffusion specification, we center the variables around

7 This includes both “trauma diversion” and “medical diversion”, where the former is used to divert only trauma patients. We do nothave information regarding these separate categories in our data and hence we use trauma center designation as a control variablein our model.8 staffed icu beds = licensed ICU beds

total licensed beds∗total staffed beds

Allon, Deo, Lin: Ambulance Diversion14 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

their mean to create (REL ED SIZE HT CTR) and (NORM ICU CAP CTR) to reduce nonessential mul-ticollinearity (Cohen et al. 2003).

7.3. Specifications and measures: Choice equationWe use a binary variable (AMB DIV=1) to denote whether the hospital had positive diversion hours. Thedecision of whether to employ ambulance diversion or not might itself depend on other characteristics ofthe hospital such as location and ownership structure. For example, many hospitals in rural areas mightdecide not to go on ambulance diversion since there are no alternate hospitals in the vicinity. Similarly,it is plausible that nonprofit or academic hospitals are less likely to divert ambulances because of theirmission. Also, trauma centers might adopt different ambulance diversion policies than EDs which are notdesignated as trauma centers. Similar to the level equation, it is likely that the operational variables impactthe choice of employing diversion. For instance, a hospital that on average has lower bed utilization andlarger EDs might decide not to employ diversion for occasional overcrowding but employ other mitigationmechanisms, which presumes a significant fixed cost of employing and managing the diversion mechanism.

In our data, hospitals are designated as rural hospitals based on Section 124840 of the California Healthand Safety Code which includes the following criteria: acute care hospital with less than 76 beds and locatedin a census dwelling place with less than 15000 residents as per the 1980 census. We create a dummyvariable to indicate the location of the hospital (RURAL=1 for rural and RURAL=0 for urban). We alsoclassify ownership control of the hospitals into the following categories: GOVERNMENT (City and/orCounty, District), NONPROFIT (Not-for-profit including university hospitals), and INVESTOR (for-profitconcerns including partnerships, and public and private companies). The California EMS Authority assignsone of four designations to each ED with a trauma center depending on its capabilities to treat and stabilizetrauma patients with level 1 being the most capable and level 4 being the least capable. We create a dummyvariable to indicate if an ED was designated as a trauma center in one of these categories (TRAUMA=1) ornot (TRAUMA=0). This combined with the choice equation above will be referred to as Specification A.

7.4. Descriptive StatisticsTable 2 presents the descriptive statistics for our variables. We find a huge variation in DIV HOURS (fromzero to 7170 hours) with approximately 35% of the EDs having zero diversion hours, again suggestingthat some hospitals make a strategic decision of not going on diversion as discussed above. Moreover, thelong tail and the fact that standard deviation is much larger than the mean indicates that the data is notnormally distributed. Hence we transformed our dependent variable by taking its natural logarithm. Notethat the utilization of the ICU beds exceeds 100% in some cases. This is possible because ICU patients areoccasionally placed in other wards if ICU beds are full.

7.5. Estimation and Results7.5.1. Diffusion approximation.Table 3 contains the results the specification based on the diffusion ap-proximation, which we refer to as Specification A. Before turning our attention to the level equation, whichis our primary focus since our queueing models are used to formulate hypotheses on the extent of diversion,we briefly discuss the results for the choice equation.

We find that rural hospitals are less likely to use ambulance diversion whereas trauma centers and privatehospitals are more likely to use it. Many rural hospitals might be forced to accept all ambulances due tothe lack of other hospitals in their catchment area. On the other hand, diverting ambulances with traumapatients might be more effective in ensuring prompt care than accepting them in an overcrowded ED. Wealso found that the hospitals whose relative size of the ED compared to the ICU is larger are more likely tochoose ambulance diversion.

Returning to the level equation, the coefficient of the Inverse Mills Ratio (IMR) is statistically significantin the LIML estimation. This rejects the null hypothesis that the sub-sample of hospitals with positivediversion hours is randomly selected from the larger sample (Leung and Yu 1996, Melino 1982). Thus, wefind evidence that some hospitals make explicit decisions to accept all ambulances irrespective of the extent

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 15

Table 2 Descriptive Statistics.

Variable Name Mean Std. dev. Min. Max.DIV HOURS 790.17 1261.62 0.00 7170.00

N1 17.95 11.08 1.00 64.00

NI 17.09 17.42 2.00 189.00

ρI 0.66 0.31 0.00 2.05

REL ED SIZE HT 4.65 1.94 0.29 11.63

NORM ICU CAP 1.30 1.31 -2.99 6.65

RURAL 0.13

INVESTOR 0.60

GOVERNMENT 0.16

TRAUMA 0.18

Table 3 Estimates for selection models based on diffusion approximation. Standard Errors are shown in parentheses.∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

LEVEL EQUATIONINTERCEPT 6.54∗∗∗ (0.29)

REL ED SIZE HT CTR -0.07 (0.06)

NORM ICU CAP CTR -0.06 (0.09)

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) -0.06∗ (0.04)

RURAL

INVESTOR

GOVERNMENT

INVERSE MILLS RATIO -0.98∗ (0.53)

R2 0.03

F-value 1.64∗

CHOICE EQUATIONINTERCEPT 0.46∗∗∗ (0.11)

REL ED SIZE HT CTR 0.13∗∗∗ (0.04)

NORM ICU CAP CTR -0.04 (0.06)

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) 0.02 (0.03)

RURAL -1.28∗∗∗ (0.27)

INVESTOR 0.68∗∗∗ (0.22)

GOVERNMENT -0.46 (0.3)

TRAUMA 0.38∗ (0.23)Log Likelihood -154.55

Allon, Deo, Lin: Ambulance Diversion16 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

of overcrowding. The interaction term REL ED SIZE HT*NORM ICU CAP is significant at 10% level inSpecification A indicating that REL ED SIZE HT and NORM ICU CAP moderate each other’s effect onthe extent of diversion, conditional on the hospital’s decision to use ambulance diversion. This supports ourHypothesis 3 from Section 5.2 is in conformance with the analytical expression for the fraction of time ondiversion (1) which includes a product term κβ2. This implies that the magnitude of the marginal impactof the spare capacity of inpatient beds on the extent of diversion is decreasing in the size of the ED. Toour knowledge, this interaction between the capacity of the inpatient department and the size of the ED hasnot been studied before in the literature. It is not possible to empirically observe this interaction in singlehospital studies due to lack of meaningful variation in ED size, which is not the case in our cross-sectionalanalysis.

Due to centering, one needs to be careful in interpreting the coefficients of REL ED SIZE HT CTR andNORM ICU CAP CTR. The estimates reported in Table 3 are the conditional effects of each variable at themean of the other. Hence, in order to assess statistical significance more thoroughly, we calculate the esti-mates of the simple slopes and standard errors of each variable at three values (mean, two standard deviationbelow the mean and two standard deviation above the mean) of the other variable involved in the interaction(Cohen et al. 2003). These results are shown in Table 4. We find that coefficient of NORM ICU CAP issignificant at 10% level for very high values of REL ED SIZE HT and coefficient of REL ED SIZE HT issignificant at 10% for high values of NORM ICU CAP thus providing support for our Hypotheses 1 and 2from Section 5.2.

Table 4 Simple slope for Relative ED size and normalized ICU spare capacity for different values of the other interactingvariable . σN is the standard deviation of the normalized ICU spare capacity in the sample. Standard Errors are shown in

parentheses. ∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

Simple slope ofREL ED SIZE

Simple slope ofNORM ICU CAP

NORM ICU CAP CTR = -2*σN 0.12 (0.12)

NORM ICU CAP CTR = 0 -0.07 (0.06)

NORM ICU CAP CTR = 2*σN -0.26∗(0.15)REL ED SIZE HT CTR = -2*σR 0.22 (0.19)

REL ED SIZE HT CTR = 0 -0.06 (0.09)

REL ED SIZE HT CTR = 2*σR -0.34∗ (0.20)

This provides an interesting interpretation of the bottleneck in the patient flow process. The ICU bedsare the bottleneck when the size of the ED is relatively large compared to the ICU (higher values ofREL ED SIZE HT): the congestion in the ED and the consequent ambulance diversion is increasing in theutilization of the ICU beds. On the other hand, the ED is the bottleneck when the utilization of the ICUnormalized for its size is very low (high values of NORM ICU CAP): the extent of ambulance diversionis decreasing in the number of ICU beds. This is partly in conformance with the findings from previoussingle-hospital studies that find ICU and other inpatient departments to be the primary driver for ambulancediversion. However, if the size of the ED relative to the ICU is sufficiently small, the fraction of time spenton diversion is not significantly associated with the utilization of the ICU. These findings suggest that oneneeds to qualify the findings from single hospital studies to a larger population and that a “one size fits all”policy aimed at improving inpatient capacity might not reduce the extent of ambulance diversion across theboard.

7.5.2. Fluid approximation.The results for the specifications based on the fluid approximation of thequeueing system are shown in Table 5 below.

Similar to the diffusion approximation, we focus mainly on the estimates of the level equation. First,note that the Inverse Mills Ratio is significant at 10% for indicating evidence of sample selection. Next,the coefficient of REL ED SIZE FL is not significant. Also, as in the diffusion model, we find that rural

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 17

Table 5 Estimates for selection model based on fluid approximation. Standard Errors are shown in parentheses.∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

LEVEL EQUATIONINTERCEPT 6.83∗∗∗ (0.33)

SQRT ICU CAP 0.08 (0.51)

REL ED SIZE FL -2.33 (2.86)

RURAL

INVESTOR

GOVERNMENT

INVERSE MILLS RATIO -1.1∗ (0.64)R2 0.04

F-value 2.23∗∗

CHOICE EQUATIONINTERCEPT 1.45∗∗∗ (0.24)

SQRT ICU CAP -1.23∗∗∗ (0.29)

REL ED SIZE FL -4.22∗∗∗ (1.37)

RURAL -0.98∗∗∗ (0.28)

INVESTOR 0.51∗∗ (0.21)

GOVERNMENT -0.6∗∗ (0.3)

TRAUMA 0.2 (0.22)Log Likelihood -151.74

hospitals and private hospitals are less likely to employ diversion compared to urban and non-profit hospi-tals, respectively. Similarly, for-profit hospitals are more likely to employ diversion compared to non-profithospitals.

7.5.3. Basic model.In addition to the diffusion approximation and fluid approximation of the queueingmodel described in Section 4, we now consider a basic model, where the operational variables are rawmeasures of ED SIZE (N1), ICU CAP (1 − ρI) and ICU SIZE (NI). For fair comparison between thedifferent models, we do not treat this model as a naive model, and thus we do not change the non-operationalvariables such as ownership, location and trauma center designation. Unlike the previous two, this basicmodel does not rely on an analytical model to automatically generate its hypotheses. However, analogousto them, we hypothesize that the extent of diversion is decreasing in ICU CAP and decreasing in ED SIZE.However, we do not have a direct analog for the hypothesis about ICU SIZE and we refrain from forming anew one.

Results in Table 6 show that coefficients of ICU CAP and ED SIZE are not significant in both specifi-cations. The coefficient of ICU SIZE is positive and significant indicating certain diseconomies of scale.However, we could not find any previous result that highlights this effect and could not articulate a strongjustification for it. Morton and Bevan (2008) do observe a certain diseconomies of scale in health care set-ting but their set up is markedly different from ours. They study the variation in service times rather than adelay measure and size is operationalized as the number of doctors rather than number of beds. In the nextsection, we analyze the comparative statistical fit of the three models.

One drawback of the basic model is that it does not incorporate any nonlinearity unlike the diffusionmodel, which might lead statistically insignificant results. To explore this further, we developed two alter-

Allon, Deo, Lin: Ambulance Diversion18 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

Table 6 Estimates for selection model based on basic model. Standard Errors are shown in parentheses.∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

LEVEL EQUATIONINTERCEPT 5.99∗∗∗ (0.59)

ICU CAP 0.21 (0.31)

ICU SIZE 0.01∗ (0.01)

ED SIZE 0.01 (0.01)

RURAL

INVESTOR

GOVERNMENT

INVERSE MILLS RATIO -0.69 (0.57)R2 0.06

F-value 3.08∗∗

CHOICE EQUATIONINTERCEPT -0.67∗∗ (0.28)

ICU CAP 0.49∗∗ (0.22)

ICU SIZE 0.01 (0.01)

ED SIZE 0.03∗∗ (0.01)

RURAL -0.87∗∗∗ (0.29)

INVESTOR 0.75∗∗∗ (0.22)

GOVERNMENT -0.42 (0.29)

TRAUMA -0.09 (0.24)Log Likelihood -149.97

nate nonlinear specifications of the basic model. In the first specification we included an interaction betweenICU SIZE and ICU CAP in addition to the individual variables. In the second specification, we used squareroot of ICU SIZE instead of the ICU SIZE to capture economies of scale. The results of these two speci-fications is shown in Table 7. Again, we do not find any statistically significant results for these nonlinearspecifications.

7.5.4. Comparison of alternative models.Note that one cannot use a simple likelihood ratio test to com-pare alternative models since they are not nested within each other. Hence, we conducted Vuong’s test(Greene 2008) for their comparative evaluation. This test is based on the fact that the difference in the like-lihood ratios of two alternative specifications is normally distributed and allows for both specifications tobe false under the null hypothesis. In other words, it only tests if one non-nested model is a better fit tothe data than the other, even if both are incorrect. Using this test, we found that the diffusion model dom-inated the fluid model (p < 0.01) and the basic model (p < 0.01). Hence, we conclude that the diffusionapproximation provides a better explanation of the data and retain it for further analysis. This suggests thata model based on the unpredictable variability of arrival pattern better describes the differences in extent ofdiversion across hospitals in our sample than the one that is based on predictable variability9.

9 Note that we do not claim that unpredictable variability is a more dominant driver of diversions than predictable variability.

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 19

Table 7 Two non-linear specifications of the basic model

Level equationINTERCEPT 5.48∗∗∗ (0.72) 6.02∗∗∗ (0.64)

ICU CAP 0.42 (0.35) 0.15 (0.47)

ED SIZE 0.0003 (0.01) 0.0006 (0.01)

ICU SIZE -0.0005 (0.02)

ICU SIZE*ICU CAP 0.02 (0.03)

SQRT ICU SIZE 0.16 (0.1)

LAMBDA -0.68 (0.57) -0.64 (0.54)Choice equationIntercept -0.99∗∗∗ (0.36) -0.44 (0.3)

ICU CAP 0.61∗∗∗ (0.23) 0.08 (0.3)

ED SIZE 0.03∗∗ (0.01) 0.03∗∗ (0.01)

ICU SIZE -0.02 (0.01)

ICU SIZE*ICU CAP 0.06∗∗ (0.02)

SQRT ICU SIZE 0.11 (0.08)

TRAUMA 0.06 (0.25) -0.01 (0.26)

INVESTOR 0.82∗∗∗ (0.22) 0.87∗∗∗ (0.23)

GOVERNMENT -0.43 (0.3) -0.36 (0.3)

RURAL -0.84∗∗ (0.29) -0.87∗∗∗ (0.29)

7.6. Instrumental variables approachIn the above analysis, we assumed that the explanatory variables of inpatient capacity and ED size are ex-ogenous and uncorrelated with the error term. However, it is plausible that these variables are endogenousin that the hospitals decide on these values based on the attractiveness and size of the market for hospitalservices, which are not included in the econometric specification, and are unobservable to the econometri-cian. This could lead to an omitted variables bias in our estimates. In this section, we employ appropriateinstrumental variables to account for this possible endogeneity of our main independent variables. Since,we have seen that the diffusion model performs better than the fluid model, we focus our attention on thediffusion model in this section.

As explained above, we hypothesize that the hospitals’ capacity decisions will depend on the size andattractiveness of the market for hospital services. Since these market characteristics were not available inour dataset, we decided to utilize the US Census data. We obtained (i) the population density, (ii) the medianhousehold income and (iii) the median family income by zip code for the state of California and linked thesedata to our original dataset by the zip code of the hospital. We combined these with the following variablesfrom our dataset: (i) the total number of patients arriving to the ED, (ii) the fraction of total visits to the EDpaid by Medicare, and (iii) the total number of inpatient surgeries in the hospital. While the first of these isclearly a measure of demand for hospital services (albeit censored), the latter two capture the profitabilityof the hospital’s patients since Medicare patients are typically less profitable compared to privately insuredpatients and surgeries/procedures are more profitable than treating chronic conditions.

The standard Heckman’s two-step estimator could not be directly used to estimate the model using in-strumental variables. As a result, we used a GMM estimator based approach (Meijer and Wansbeek 2007).

Allon, Deo, Lin: Ambulance Diversion20 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

We first conducted an instrumental variable Probit estimation in the first stage (using ivprobit statement inSTATA) to obtain the Inverse Mills Ratio. We then estimated the level equation using GMM estimator, withthe above instruments and our original independent variables as potential endogenous regressors. We alsoused (Meijer and Wansbeek 2007) and the GMM framework to estimate robust standard errors as againstour previous approach of estimating the regular standard errors. We used this method to estimate all threespecifications (A, B, C) of the heavy traffic model from the previous version of the paper but report theresults for specification A here for brevity (Table 8).

Table 8 Estimates for Specification A using instrumental variables. Standard errors are shown in parantheses and clusterederrors are shown in square brackets. ∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

LEVEL EQUATIONINTERCEPT 5.93∗∗∗ (0.42)

[0.40]

REL ED SIZE HT CTR -0.03 (0.12)[0.12]

NORM ICU CAP CTR 0.37∗∗∗ (0.13)[0.14]

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) -0.45∗∗∗ (0.08)[0.07]

RURAL

INVESTOR

GOVERNMENT

INVERSE MILLS RATIO -0.57 (0.71)[0.40]

CHOICE EQUATIONINTERCEPT 0.31∗∗ (0.12)

[0.12]REL ED SIZE HT CTR 0.34∗∗∗ (0.06)

[0.04]

NORM ICU CAP CTR -0.04 (0.07)[0.06]

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) 0.03 (0.03)[0.03]

RURAL -0.98∗∗∗ (0.28)[0.26]

INVESTOR 0.83∗∗∗ (0.23)[0.22]

GOVERNMENT -0.4 (0.32)[0.31]

TRAUMA 0.24 (0.24)[0.23]

Note that the interaction term in the level equation is of a greater magnitude and more statistically sig-nificant compared to the original model. We also repeated the analysis for calculating the marginal effects

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 21

(simple slopes) of one variable involved in the interaction term for different values of the other variable.These results are shown below in Table 9.

Table 9 Simple slopes estimation using instrumental variables

REL ED SIZE NORM ICU CAPNORM ICU CAP CTR = -2*σN 1.36∗∗∗ (0.41)

NORM ICU CAP CTR = 0 -0.03 (0.12)

NORM ICU CAP CTR = 2*σN -1.41∗∗∗ (0.42)REL ED SIZE HT CTR = -2*σR 2.47∗∗∗ (0.61)

REL ED SIZE HT CTR = 0 0.37∗∗∗ (0.13)

REL ED SIZE HT CTR = 2*σR -1.74∗∗∗ (0.63)

The main conclusion is the robustness of our qualitative finding from the previous analysis: ICU is thebottleneck when ED is relatively small and ED is the bottleneck when the ICU is relatively under utilized.In addition, we note that all the effects are larger in magnitude and have higher statistical significance.The intuitive explanation for these is as follows: If capacity and size are indeed chosen optimally, theircorrelation with the instrumental variables (market potential) will have similar sign as that with the out-come variable, the level of diversion. Hence, our earlier model specification that did not account for thisendogeneity underestimated the impact of the capacity and size variables on the extent of diversion.

Table 8 also shows the clustered standard errors, where hospitals were clustered at the level of a county.The clustered errors might account for potential correlation across hospitals due to common diversion guide-lines or ambulance routing policies within the county or physical spill over of arrivals during crowdedperiods. Our results show that clustering does not alter the standard errors and statistical significance ofthe coefficients substantially. We repeated our analysis for different definitions of clusters such as city andHealth Services Area (HSA) with little change in the results.

7.7. Network effectResults from several recent studies suggest that ambulance diversion is a network phenomenon, i.e., diver-sion at one hospital affects the congestion and hence consequently the diversion at neighboring hospitals(see Sun et al. 2006). There is also some anecdotal evidence of “defensive diversion”, wherein a hospitalgoes on diversion immediately after its neighboring hospital goes on diversion in anticipation of the in-creased load and congestion (Vilkes et al. 2004). Thus, it can be hypothesized that in the presence of sucha network effect, the extent of diversion at a hospital does not depend solely on its own characteristics butalso on the configuration of the network of hospitals around it. Hence, we developed a measure of net-work density for each ED as follows. We calculated distance between each pair of hospitals in our sampleand then calculated the number of hospitals in a five mile radius around each hospital. We then created aset of categorical variables depending on this number: no hospitals (NET NON), less than five hospitals(NET LOW), between five and ten hospitals (NET MED) and more than ten hospitals (NET HIGH). Weincluded this measure of network density in our level equation.

The results for this specification are shown in Table 10. We find that most coefficients have similar valuescompared to the earlier specification, though some lose statistical significance. Moreover, the sign of thecoefficients for the network variables are as expected: hospitals in less dense networks have lower extent ofdiversion after controlling for operational variables such as inpatient capacity and ED size. Especially, theextent of diversion is significantly lower in hospitals with less than five hospitals in its neighborhood. Whilethere has been extensive discussion about the network effect in the literature, there have been very fewstudies that have empirically documented the network effect (Sun et al. 2006, Vilkes et al. 2004) and none ofthe studies to our knowledge considers the joint effect of hospital characteristics and network characteristicson the extent of diversion.

Allon, Deo, Lin: Ambulance Diversion22 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

Table 10 Estimates for selection models based on diffusion approximation with network effect. Standard Errors are shown inparentheses. ∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

LEVEL EQUATIONINTERCEPT 7.06∗∗∗ (0.44)

NORM ICU CAP CTR -0.03 (0.08)

REL ED SIZE HT CTR -0.02 (0.06)

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) -0.04 (0.04)

NET NON -1.44∗∗∗ (0.49)

NET LOW -0.92∗∗ (0.4)

NET MED -0.23 (0.42)

INVERSE MILLS RATIO -0.58 (0.56)R2 0.10

F-value 2.80∗∗∗

CHOICE EQUATIONINTERCEPT 0.46∗∗∗ (0.11)

NORM ICU CAP CTR -0.06 (0.06)

REL ED SIZE HT CTR 0.12∗∗∗ (0.04)

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) 0.01 (0.03)

RURAL -1.27∗∗∗ (0.27)

INVESTOR 0.63∗∗∗ (0.21)

GOVERNMENT -0.46 (0.3)

TRAUMA 0.26 (0.22)Log Likelihood -154.56

7.8. Sensitivity analysisWe estimated two alternate specifications for each model family. For Specification B, the choice equation in-cluded all operational and non-operational variables whereas the level equation included all operational andnon-operational variables other than Trauma center. As discussed above, the extent of diversion is impactedby the delay probability, which itself might depend on non-operational factors. This provides a natural ba-sis for exclusion restriction in our model. For example, rural hospitals might be willing to tolerate longerdelays because of lack of appropriate alternatives in their geographical vicinity. Similarly, hospitals withdifferent ownership structures such as private, government and non-profits might differ significantly in theirtolerance for delays. For instance, privately owned hospitals might tolerate longer delays than non-profitsas a way to reduce foregone revenues via diversion. On the other hand, non-profits might tolerate longerdelays due to adverse social consequences of diversion, which might not be measurable in pecuniary terms.Nonetheless, it is reasonable to expect that the delay probability and hence the level of diversion woulddiffer significantly based on ownership. We hypothesize that trauma centers should not have different targetdelay probability than other EDs without such designation since they only differ in the arrival character-istics for high priority patients, who typically do not face any delay. Given that Specification B has fewerexclusion restrictions compared to Specification A, we expect that the standard errors will be inflated andestimates will be inefficient in Specification B.

For Specification C, the level equation included only operational variables and the choice equation in-cluded only non-operational variables. The results for these specifications based on diffusion approximation,

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 23

fluid approximation and the basic model are shown in Tables 11, 14 and Table 15 below. As can be seen,these estimates are very similar to Specifications A emphasizing the robustness of our results. Specifically,for the diffusion approximation, we see that the estimates of simple slopes for ED size and normalized ICUspare capacity are also very similar across all three specifications (Tables 12 and 13 compared to Table 4).

Table 11 Estimates for alternate specifications based on diffusion approximation. Standard Errors are shown in parentheses.∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10.

Specification B Specification CLEVEL EQUATIONINTERCEPT 6.34∗∗∗ (0.83) 6.69∗∗∗ (0.3)

REL ED SIZE HT CTR -0.07 (0.09) -0.01 (0.06)

NORM ICU CAP CTR -0.05 (0.11) -0.08 (0.08)

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) -0.06 (0.04) -0.05 (0.04)

RURAL -1.86 (1.55)

INVESTOR 0.17 (0.76)

GOVERNMENT -0.29 (0.53)

INVERSE MILLS RATIO -0.24 (1.63) -1.28∗∗ (0.54)

R2 0.07 0.05

F-value 1.91∗ 2.22∗

CHOICE EQUATIONINTERCEPT 0.46∗∗∗ (0.11) 0.47∗∗∗ (0.11)

REL ED SIZE HT CTR 0.13∗∗∗ (0.04)

NORM ICU CAP CTR -0.04 (0.06)

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) 0.02 (0.03)

RURAL -1.28∗∗∗ (0.27) -1.37∗∗∗ (0.26)

INVESTOR 0.68∗∗∗ (0.22) 0.5∗∗ (0.21)

GOVERNMENT -0.46 (0.3) -0.47∗ (0.29)

TRAUMA 0.38∗ (0.23) 0.36∗ (0.22)Log Likelihood -154.55 -161.92

Table 12 Simple slopes for Relative ED size for different values of normalized ICU spare capacity. σN is the standarddeviation of the normalized ICU spare capacity in the sample. Standard Errors are shown in parentheses.

∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

Specification B Specification CNORM ICU CAP CTR = -2*σN 0.13 (0.13) 0.15 (0.12)

NORM ICU CAP CTR = 0 -0.05 (0.11) -0.01 (0.06)

NORM ICU CAP CTR = 2*σN -0.24 (0.19) -0.17 (0.14)

7.9. Discussion and LimitationsPreceding analysis highlights the complementary value of analytical and empirical models. Specifically, ouranalysis of the two station queueing model provided the theoretical foundation for the measures of inpatientcapacity and ED size that were used in the empirical models. On the other hand, empirical analysis provided

Allon, Deo, Lin: Ambulance Diversion24 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

Table 13 Simple slope for Normalized ICU spare capacity for different values of relative ED size. σR is the standard deviationof the relative ED size in the sample. Standard Errors are shown in parentheses. ∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

Specification B Specification CREL ED SIZE HT CTR = -2*σR 0.21 (0.21) 0.16 (0.19)

REL ED SIZE HT CTR = 0 -0.07 (0.09) -0.08 (0.08)

REL ED SIZE HT CTR = 2*σR -0.35∗ (0.20) -0.32∗ (0.2)

Table 14 Estimates for alternate specifications based on fluid approximation. Standard Errors are shown in parentheses.∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

Specification B Specification CLEVEL EQUATIONINTERCEPT 6.80∗∗∗ (0.37) 7.32∗∗∗ (0.4)

SQRT ICU CAP 0.07 (1.24) -0.48 (0.41)

REL ED SIZE FL -2.51 (5.05) -4.24∗ (2.42)

RURAL -1.45 (1.64)

INVESTOR -0.31 (0.52)

GOVERNMENT 0.38 (0.98)

INVERSE MILLS RATIO -0.7 (2.23) -1.22∗∗ (0.55)R2 0.07 0.05

F-value 2.23∗∗ 3.24∗∗

CHOICE EQUATIONINTERCEPT 1.45∗∗∗ (0.24) 0.48∗∗∗ (0.11)

SQRT ICU CAP -1.23∗∗∗ (0.29)

REL ED SIZE FL -4.22∗∗∗ (1.37)

RURAL -0.98∗∗∗ (0.28) -1.36∗∗∗ (0.26)

INVESTOR 0.51∗∗ (0.21) 0.45∗∗ (0.2)

GOVERNMENT -0.6∗∗ (0.3) -0.48∗ (0.29)

TRAUMA 0.20 (0.22) 0.26 (0.21)Log Likelihood -151.74 -164.18

a method of comparing the validity of the different analytical models. Based on Vuong’s test for non-nestedmodels, we conclude that the diffusion model provides a better explanation of our data than the fluid andbasic models.

Our empirical analysis has several limitations stemming partly from the limitations of our data and partlyfrom the complexities of the queueing model. Our study sample is not drawn randomly from the populationof hospitals in the U.S. and hence our results cannot be directly generalized beyond California. While it isreasonable to expect that similar operational factors impact ambulance diversion in other states as well, Cal-ifornia differs from other states in certain important respects such as not requiring express approval (calledCertificate of Need) from local health planning agencies prior to expanding their capacity (Cimasi 2005),existence of its “Medicaid Disproportionate Share Hospital” program and specific regulations governing thevalidity and appropriateness of ambulance diversions.

Our data is aggregate and represents annual averages for inpatient occupancy and staffed beds and yearend number for licensed beds. It has been widely observed that demand for emergency services follows a

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 25

cyclic pattern with typically more visits on the weekdays than the weekends and more arrivals in morningsthan afternoons (Green et al. 2006, Burt and McCaig 2006). Inpatient admission and discharge processesalso follow similar cyclic patterns (McManus et al. 2003). To the extent that the number of staffed bedscannot be continuously adjusted to match these patterns, our aggregate data would have underestimatedthe magnitude of undercapacity and overcapacity thus reducing the variation in the underlying data. Thispotentially creates a bias against finding significant effect on inpatient occupancy on ED overcrowding andambulance diversion. We confirm this by conducting a simulation study in Appendix F.

Our measure of staffed ICU beds was calculated based on the licensed ICU beds and the overall ratio ofstaffed and licensed beds in the hospital. We tested the robustness of our results using alternative measuresof inpatient capacity. Using licensed ICU beds instead of staffed ICU beds did not significantly changeour results. However, using all the beds designated to general acute care instead of only intensive careresulted in loss of significance for most of our variables indicating that ICU occupancy is the main driverfor crowding in the ED. Moreover, since our measure of inpatient capacity is based on the utilization andthe number of beds in ICU, our results cannot be generalized to hospitals without ICUs. However, giventhat very few hospitals are without ICUs and congestion in ICU has been cited as a driver for ED crowdingin the literature, we believe that this is not a serious limitation.

Table 15 Estimates for alternate specifications based on basic model. Standard Errors are shown in parentheses.∗∗∗ p< 0.01; ∗∗ p< 0.05; ∗ p< 0.10

Specification B Specification CLEVEL EQUATIONINTERCEPT 2.82 (2.00) 5.82∗∗∗ (0.45)

ICU CAP 0.79∗ (0.47) -0.51 (0.31)

ICU SIZE 0.03∗∗ (0.01) 0.01∗ (0.01)

ED SIZE 0.04 (0.03) 0.01 (0.01)

RURAL -3.75∗∗ (1.5)

INVESTOR 0.93 (0.71)

GOVERNMENT -0.47 (0.79)

INVERSE MILLS RATIO 2.93 (2.05) -1.11∗∗ (0.53)R2 0.09 0.07

F-value 2.76∗∗∗ 3.71∗∗∗

CHOICE EQUATIONINTERCEPT -0.67∗∗ (0.28) -0.72∗∗∗ (0.27)

ICU CAP 0.49∗∗ (0.22)

ICU SIZE 0.01 (0.01)

ED SIZE 0.03∗∗ (0.01)

RURAL -0.87∗∗∗ (0.29) -1.37∗∗∗ (0.26)

INVESTOR 0.75∗∗∗ (0.22) 0.81∗∗∗ (0.22)

GOVERNMENT -0.42 (0.29) -0.43 (0.30)

TRAUMA -0.09 (0.24) 0.36∗ (0.22)Log Likelihood -149.96 -161.92

Even the stylized queueing network model of patient flow from the ED to the inpatient department isquite complicated and requires a series of approximations to derive analytical results. Moreover, these ana-

Allon, Deo, Lin: Ambulance Diversion26 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

lytical expressions are highly nonlinear in the parameters that need to be estimated. This prevents us fromundertaking a structural estimation of the parameters of our queueing model.

8. ConclusionIn this paper we present theoretical as well as empirical analysis of the phenomenon of ED overcrowdingand ambulance diversion in hospitals. While earlier investigations have shown that higher occupancy inthe inpatient departments, such as the ICU, is associated with greater extent of ambulance diversion inthe ED, these studies have lacked the theoretical framework to compare the effect of hospital size on thestrength of this association. Moreover, queueing theory predicts that the effect of hospital size can be verydifferent depending on whether one focuses on unpredictable variability or predictable variability of thearrival pattern.

We bridge this gap by employing a queueing network model for the flow of patients between the ED andthe inpatient department of the hospital and analyze it using two approximations for large and busy systemsthat focus on unpredictable and predictable variability. Our analysis shows that the measures based on themodel of unpredictable variability (diffusion approximation) better explain the differences in the extent ofdiversion across hospitals than those based on the model of predictable variability (fluid approximation) anda basic model containing raw measures of inpatient capacity and ED size.

For the diffusion approximation, we find that the capacity of the inpatient department and the ED interactwith each other in determining their impact on ambulance diversion, an effect that has not yet been identifiedin the literature. As a result of this interaction, tje inpatient department tends to be the bottleneck in hospitalswhere the ED is large relative to the hospital. Conversely, ED tends to be the bottleneck in hospitals wherethe inpatient department is has relatively higher spare capacity.

8.1. Policy ImplicationsThese results have important implications for the policy debate around drivers of ambulance diversion.

First, our result on unpredictable variability suggests that if one ED has higher unpredictable variabilitycompared another ED, it is likely to have greater number of hours on diversion, everything else being equal.This implies that ED administrators who want to reduce their diversion hours should focus on managingthis unpredictable variability by adding appropriate level of safety capacity in terms of staffing and beds.

Second, our result on interaction between ED and inpatient department capacities help characterize twotypes of hospitals based on the bottleneck in the patient flow process and the driver of ambulance diversion.In the first type of hospital, where the ED is large relative to the hospital, the inpatient department tendsto be the bottleneck. In the second type, where the inpatient department has relatively higher spare capac-ity, the ED tends to be the bottleneck. This finding suggests that different measures might be required tomitigate ambulance diversion in these two types of hospitals. This qualifies the current policy recommenda-tion that ambulance diversion and ED crowding can be reduced by creating more capacity in the inpatientdepartment, typically ICU.

Third, our result on interaction further suggests that hospitals could improve their responsiveness, i.e.,reduce delay probability, by appropriately choosing the number of beds in the ED depending on the extentof spare capacity in the inpatient department. In our experience, such coordinated capacity decisions arerarely made in the U.S. hospitals. Future work could provide qualitative guidance regarding the relationshipbetween relative size of the ED and the spare capacity in the inpatient department for an optimized patientflow.

ReferencesJ. Adams. Personal communication, 2008.

T. Amemiya. Tobit models: A survey. Journal of Econometrics, 24(1-2):3–61, 1984.

B. R. Asplin, D. J. Magid, K. V. Rhodes, L. I. Soldberg, N. Lurie, and C. A. Carmago. A conceptual model ofemergency department overcrowding. Annals of Emergency Medicine, 42(2):173–180, 2003.

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 27

J. Broyles and J. Cochran. Estimating business loss to a hospital emergency department from patient reneging byqueuing-based regression. In Proceedings of the 2007 Industial Engineering Research Conference, pages 613–618, 2007.

C. W. Burt and L. F. McCaig. Staffing,capacity,and ambulance diversion in emergency departments: Unitedstates,2003–04. CDC Advance data from vital and health statistics, 376, 2006.

C. W. Burt, L. F. McCaig, and R. Valverde. Analysis of ambulance transpors and diversions among us emergencydepartments. AEM, 47(4):317–326, 2006.

J. Buzacott and J. Shanthikumar. Stochastic models of manufacturing systems. Prentice Hall, 1993.

R. J. Cimasi. U.S. healthcare certificate of need sourcebook. Frederick, MD: Beard Books, 2005.

J. Cochran and K. Roche. A multi-class queuing network analysis methodology for improving hospital emergencydepartment performance. Computers and Operations Research, 36(5):1497–1512, 2009.

J. Cohen, P. Cohen, S. G. West, and L. S. Aiken. Applied multiple regression / correlation analysis for the behavioralsciences (3rd ed.). Mahwah, NJ: Lawrence Erlbaum Associates, Inc., 2003.

R. W. Derlet and J. R. Richards. Overcrowding in the nation’s emergency departments: Complex causes and disturbingeffects. Annals of Emergency Medicine, 35(1):63–68, 2000.

R. W. Derlet, J. R. Richards, and R. L. Kravitz. Frequent overcrowding in u.s. emergency departments. AcademicEmergency Medicine, 8(2):151–155, 2001.

M. Eckstein and L. S. Chan. The effect of emergency department crowding on paramedic ambulance availability.Annals of Emergency Medicine, 43(1):100–105, 2004.

J. A. Fitzsimmons. A methodology for emergency ambulance deployment. Management Science, 19(6):627–636,1973.

A. J. Forster, I. Stiell, G. Wells, A. J. Lee, and C. Van Walraven. The effect of hospital occupancy on emergencydepartment length of stay and patient disposition. Academic Emergency Medicine, 10(2):127–133, 2003.

GAO. Hospital emergency departments: Crowded conditions vary among hospitals and communities. WashingtonDC: U.S. General Accounting Office, 2003.

Y. Gerchak, D. Gupta, and M. Henig. Reservation planning for elective surgery under uncertain demand for emergencysurgery. Management Science, 42(3):321–334, 1996.

L. V. Green. How many hospital beds? Inquiry, 39:400–412, 2002.

L. V. Green and P. J. Kolesar. Improving emergency reponsiveness with management science. Management Science,50(8):1001–10014, 2004.

L. V. Green and V. Nguyen. Strategies for cutting hospital beds: The impact on patient service. Health ServicesResearch, 36(2):421–442, 2001.

L. V. Green, J. Soares, J. F. Giglio, and R. A. Green. Using queueing theory to increase the effectiveness of emergencydepartment provider staffing. Academic Emergency Medicine, 13(1):61–68, 2006.

W. H. Greene. Econometric analysis (6th ed.). Upper Saddle River, NJ: Pearson Education, Inc., 2008.

S. Halfin and W. Whitt. Heavy-traffic limits for queues with many exponential servers. Operations Research, 29(3):567–588, 1981.

J. H. Han, C. Zhou, D. J. France, S. Zhong, I. Jones, A. B. Storrow, and D. Aronsky. The effect of emergencydepartment expansion on emergency department overcrowding. Academic Emergency Medicine, 14(4):338–343,2007.

J. J. Heckman. Sample selection bias as a specification error. Econometrica, 47(1):153–161, 1979.

P. L. Joskow. The effects of competition and regulation on hospital bed supply and the reservation quality of thehospital. The Bell Journal of Economics, 11(2):421–447, 1980.

D. S. Kc and C. Terwiesch. An Econometric Analysis of Patient Flows in the Cardiac ICU. Working Paper, 2009.

J. Kennedy, K. Rhodes, C. A. Walls, and B. Asplin. Access to emergency care: Restricted by long waiting times andcost and coverage concerns. Annals of Emergency Medicine, 43(5):567–573, 2004.

A. Kolker. Process modeling of emergency department patient flow: Effect of patient length of stay on ED diversion.Journal of Medical Systems, 32(5):389–401, 2008.

Allon, Deo, Lin: Ambulance Diversion28 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

S. F. Leung and S. Yu. On the choice between sample selection and two-part models. Journal of Econometrics, 72(1-2):197–229, 1996.

S. F. Leung and S. Yu. Collinearity and two-stage estimation of sample selection models: Problems, origins andremedies. Computational Economics, 15(3):173–199, 2000.

A. Mandelbaum and S. Zeltyn. Service Engineering: Data-Based Course Development and Teaching. Full Version. Afull version of the present document, under review for INFORMS Transactions on Education, Special Issue onTeaching Service and Retail Operations Management, 1(1.3):2, 2009.

Y. Marmor, A. Shtub, A. Mandelbaum, S. Vasserkrug, S. Zeltyn, B. Carmeli, O. Greenshpan, and Y. Masika. To-ward simulation-based real-time decision-support-systems for emergency departments. In Winter SimulationConference, 2009.

M. McCarthy, S. Zeger, R. Ding, D. Aronsky, N. Hoot, and G. Kelen. The challenge of predicting demand for emer-gency department services. Academic Emergency Medicine, 15(4):337–346, 2008.

K. J. McConnell, C. F. Richards, M. Daya, S. H. Bernell, C. C. Weathers, and R. A. Lowe. Effect of increased icucapacity on emergency department length of stay and ambulance diversion. Annals of Emergency Medicine, 45(5):471–478, 2005.

M. McManus. Emergency department overcrowding in Massachusetts: Making room in our hospitals. Waltham, MA:The Massachusetts Health Policy Forum, 2001.

M. L. McManus, M. C. Long, A. Cooper, J. Mandell, D. M. Berwick, M. Pagano, and E. Litvak. Variability in surgicalcaseload and access to intensive care services. Anesthesiology, 98(6):1491–1496, 2003.

E. Meijer and T. Wansbeek. The sample selection model from a method of moments perspective. EconometricReviews, 26(1):25–51, 2007.

A. Melino. Testing for sample selection bias. The Review of Economic Studies, 49(1):151–153, 1982.

G. A. Melnick, A. C. Nawathe, A. Bamezai, and L. Green. Emergency department capacity and access in california,1990–2001: An economic analysis. Health Affairs., 23(3), 2004.

C. T. Merrill and A. Elixhauser. Hospitalization in the United States, 2002: HCUP Fact Book No. 6. Rockville, MD:Agency for Healthcare Research and Quality, 2005.

A. Morton and G. Bevan. What’s in a wait? Contrasting management science and economic perspectives on waitingfor emergency care. Health policy, 85(2):207–217, 2008.

H. Perros. Approximation algorithms for open queueing networks with blocking. Stochastic Analysis of Computerand Communication Systems, pages 451–98, 1990.

K. Ramdas and J. Williams. An empirical investigation into the tradeoffs that impact on-time performance in theairline industry. Working Paper, 2008.

A. Ramirez, J. Fowler, and T. Wu. Analysis of ambulance diversion policies for a large-size hospital. In WinterSimulation Conference (WSC), Proceedings of the 2009, pages 1875–1886, 2009.

E. S. Savas. Simulation and cost-effectiveness analysis of new york’s emergency ambulance service. ManagementScience, 15(12):B608–B627, 1969.

M. J. Schull, K. Lazier, M. Vermeulen, S. Mawhinney, and L. J. Morrison. Emergency department contributors toambulance diversion: A quantitative analysis. Annals of Emergency Medicine, 41(4):467–476, 2003a.

M. J. Schull, L. J. Morrison, M. Vermeulen, and D. A. Redelmeier. Emergency department gridlock and out-of-hospitaldelays for cardiac patients. Academic Emergency Medicine, 10(7):709–716, 2003b.

M. J. Schull, M. Vermuelen, G. Slaughter, L. Morrison, and P. Daly. Emergency department crowding and thrombolysisdelays in acute myocardial infarction. Annals of Emergency Medicine, 44(6):577–585, 2004.

L. T. Soldberg, B. R. Asplin, R. M. Weinick, and D. J. Magid. Emergency department crowding: Consensus develop-ment of potential measures. Annals of Emergency Medicine, 42(6):824–834, 2003.

D. Stapleton and D. Young. Censored normal regression with measurement error on the dependent variable. ECON,52(3):737–760, 1984.

B. C. Sun, S. A. Mohanty, R. Weiss, R. Tadeo, M. Hasbrouck, W. Koenig, C. Meyer, and S. Asch. Effects of hospitalclosures and hospital characteristics on emergency department ambulance diversion, los angeles county, 1998 to2004. Annals of Emergency Medicine, 47(4):309–316, 2006.

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 29

C. Swoveland, D. Uyeno, I. Vertinsky, and R. Vickson. Ambulance locations: A probabilistic enumeration approach.Management Science, 20(4):686–698, 1973.

The Lewin Group. Emergency department overload: A growing crisis - The results of the AHA survey of emergencydepartment (ED) and hospital capacity. Falls Church, VA: American Hospital Association, 2002.

G. Vassilocopoulos. A simulation model for bed allocation to hospital inpatient departments. Simulation, 45(5):233–241, 1985a.

G. Vassilocopoulos. Allocating doctors to shifts in an accident and emergency department. Journal of the OperationalResearch Society, 36(6):517–523, 1985b.

G. M. Vilkes, E. M. Castillo, M. A. Metz, L. U. Ray, P. A. Murrin, R. Lev, and T. C. Chan. Community trial to decreaseambulance diversion hours: The san diego county patient destination trial. Annals of Emergency Medicine, 44(4):295–303, 2004.

W. Whitt. A diffusion approximation for the G/GI/n/m queue. Operations Research, 52(6):922–941, 2004.

Allon, Deo, Lin: Ambulance Diversion30 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

Appendix

A. Background of heavy traffic theoryTo illustrate the approach, consider an M/M/s system with arrival rate λ and service rate µ for each server.To derive the heavy traffic limit, we consider a sequence of M/M/n systems, each with service rate µ andindexed by n= 1,2, .., that satisfies the following conditions:(A1) For the nth system, the number of servers is n and the arrival rate is λn, where the superscript ndenotes the nth system(A2) λs = λ, i.e., the sth system in the sequence is the original system, and(A3)

√n(1− ρn)→ β for some constant β as n→∞, where ρn = λn/(nµ) is the traffic intensity of the

nth system.Condition (A3) is critical and needs more explanation. It states that for large systems, the excess capacity

(1− ρ) should be approximately inversely proportional to the square-root of the number of servers. It is notonly consistent with the common wisdom that the appropriate level of server utilization in service systemsshould increase with the size of the system but also further specifies that this increases at a rate proportionalto the square root of the system size. This condition is also the theoretical underpinning for the famous“square-root” staffing rule that is used in call center management.

Halfin and Whitt (1981) provide theoretical justification for (A3) by showing that the probability of delayis approximately equal for each of the systems in the sequence if and only if this condition is satisfied. Ifthe fraction of excess capacity goes down at a rate faster than 1/

√n, then an arriving customer has to wait

almost surely. On the other hand, if the fraction of excess capacity goes down at a rate slower than 1/√n,

an arriving customer almost never waits.Halfin and Whitt (1981) further show that under condition (A3), in steady state, the normalized queue

length Q̃n(∞) =(Qn(∞)− n

)/√n converges to a diffusion process Q̃(∞) in distribution. Therefore, we

can approximate Qn(∞) by√nQ̃(∞) + n, a property that will be used extensively to characterize the

performance measures for the system of our interest.Similar approach can be adopted for an M/M/s/K system (Whitt 2004) by considering the limit of a

sequence of systems, each with service rate µ and indexed by n such that:(B1) The nth system is an M/M/n/Kn system with arrival rate λn

(B2) λs = λ and Ks =K, i.e., the sth system in the sequence is the original system, and(B3)

√n(1− ρn)→ β and Kn/

√n→ κ for some constants β and κ as n→∞.

B. Estimation on Delay Probability PdIn this section, we derive the function P̃d(N1,N2, β2, κ) that we use to approximate the delay probabilityPd.

Let ρ1 denote the traffic intensity of station 1 (the ED). It is clear that

ρ1 = (λw + (1− δ)λa)m1/(N1−B). (5)

Using Proposition 1 of Halfin and Whitt (1981), we can approximate the delay probability in the ED by[1 + β1Φ(β1)

φ(β1)

]−1

, where

β1 =√N1−B(1− ρ1). (6)

Therefore, to obtain an estimate of Pd, we need estimates of B, the average number of boarding patients,and ρ1, the traffic intensity of the ED.

We first approximate B using the steady state distribution of Q̃2 given in Whitt (2004),

B̃(N2, β2, κ) =√N2

1− e−κβ2(1 +κ)

β2(1− e−κβ2 +β φ(β2)

Φ(β2))

(7)

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 31

Next, from (5), we can approximate ρ1 as

ρ̃1(N1,N2, β2, κ) =λ1(1− δ̃(N2, β2, κ)) +λ2

N1− B̃(N2, β2, κ)m1, (8)

where δ̃ and B̃ are given in (1) and (7) respectively. We can then substitute (8) and (7) in (6) to obtain

β̃1(N1,N2, β2, κ). And P̃d(N1,N2, β2, κ) =

[1 + β̃1(N1,N2,β2,κ)Φ(β̃1(N1,N2,β2,κ))

φ(β̃1(N1,N2,β2,κ))

]−1

.

C. ProofsProof of Proposition 1

To prove result (i), let g1(β2) = (√N2 − β2)eκβ2/2 and g2(β2) = eκβ2/2−e−κβ2/2

β2+eκβ2/2Φ(β2)φ(β2)

. We know that

δ̃(N2, β2, κ) = [g1(β2)g2(β2)]−1. Hence, it is sufficient to show that g1 is nondecreasing and g2 is increasingin β2. To show g1 is nondecreasing, we look at the first order derivative:

g′1(β2) =eκβ2/2[κ(√N2−β2)/2− 1] (9)

=eκβ2/2(Kρ2/2− 1)≥ 0. (10)

For g2, it is easy to see that eκβ2/2Φ(β2)/φ(β2) is increasing in β2. Therefore, it remains to show thatg3(β2) = (eκβ2/2− e−κβ2/2)/β2 is nondecreasing in β2. Again, we take the first order derivative:

g′3(β2) = [(κβ2/2)(eκβ2/2 + e−κβ2/2)− eκβ2/2 + e−κβ2/2]/β22 = f(κβ2/2)/β2

2 ,

where f(x) = x(ex + e−x)− ex + e−x. Obviously f(0) = 0. Moreover f(x) is nondecreasing in x becausef ′(x) = x(ex− e−x)≥ 0. Therefore, f(x)≥ 0 for x≥ 0, proving g′3(β2)≥ 0.

To prove result (ii), we can rewrite (1) as

δ̃(N2, β2, κ) =β2

(√N2−β2)

(eκβ2 − 1 + eκβ2β2

Φ(β2)

φ(β2)

) .

The result then follows because the function(eκβ2 − 1 + eκβ2β2

Φ(β2)

φ(β2)

)is increasing in κ. �

Proof of Proposition 2 We can rewrite equation (7) as

B̃(N2, β2, κ) =√N2

(1

β2

− κe−κβ2

(1− e−κβ2)

)/

(1 +

β2Φ(β2)

φ(β2)(1− e−κβ2)

).

It is easy to see that B̃ is increasing in κ because κe−κβ2

1−e−κβ2= κ/(eκβ2 − 1) is decreasing in κ. Since δ̃ is

decreasing in κ, β̃1 is decreasing in κ. Therefore, P̃d is increasing in κ. �Proof of Proposition 3 It is equivalent to show that κ∗ is increasing inN1. It is easy to see that P̃d is decreas-ing in N1 because, clearly, β̃1 is increasing in N1. Now for any N1,1 <N1,2, let κ1 = κ∗(N1,1,N2, β2, P̄d)and κ2 = κ∗(N1,2,N2, β2, P̄d). It suffices to show κ1 <κ2.

Suppose κ1 ≥ κ2. Because N1,1 < N1,2, we have P̃d(N1,1,N2, β2, P̄d) > P̃d(N1,2,N2, β2, P̄d), whichcontradicts with the fact that P̃d(N1,1,N2, β2, P̄d) = P̃d(N1,2,N2, β2, P̄d) = P̄d. �

D. SimulationIn this section, we simulate the diffusion model described in section 4 to assess the accuracy of the simpli-fications and approximations made during our analysis in section 5.1. The performance measure of primaryinterest is the fraction of time on diversion but we also assess the accuracy of the delay probability esti-mation. We compute the fraction of time on diversion using the two approximations and a discrete-event

Allon, Deo, Lin: Ambulance Diversion32 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

simulation tool where we vary the size of the hospital and fix the other parameters. We choose our parame-ter values to roughly match with the median values in our data set. In the validated setting, we assume thatK = 7, and N1 = 15, the arrival rates are λa = 6, λw = 12, and λn = 5, the service times are m1 = 0.75,andm2 = 2. We also assume that the likelihood that patients arriving to the emergency department continueto an inpatient department are pa = 0.6 and pw = 0.005. Table 16 shows the estimates of fraction of timeon diversion using both diffusion approximation and simulation, and the difference between along with theestimated traffic intensity level of the inpatient department. First, observe that the accuracy of the approxi-

Table 16 Estimation of fraction of time on diversion: approximation vs simulation

N ρ2 Approximation Simulation Difference18 0.962 0.0805 0.0829 3.00%

20 0.866 0.0468 0.0486 3.76%

22 0.787 0.0269 0.0281 4.16%

24 0.722 0.0156 0.0163 4.39%

26 0.666 0.0091 0.0095 4.62%

28 0.619 0.0054 0.0058 7.42%

30 0.577 0.0032 0.0035 7.52%

mation is extremely high as long as the traffic intensity of the hospital is high enough. Second, as we fix thearrival rates and the size of the ED, the fraction of time on diversion decreases with increasing number ofinpatient beds, as estimated by the simulation and computed by the approximation.

Next, we validate the accuracy of the delay probability estimates via the diffusion approximation relativeto a simulation-based estimation. We fix the parameter as described above. In addition, we fix the number ofinpatient beds to N = 18. Table 17 reports, for different sizes of the ED, the delay probability as computedby the diffusion approximation and a discrete-event simulation, as well as the difference between the twoestimates. The table also reports the traffic intensity of the emergency department. Again, observe that theaccuracy of the delay probability estimate increases as the traffic intensity of the emergency departmentincreases. These results suggest that the series of approximations introduced in Section 4 and the diffusionapproximation provide a very good description of the actual dynamics of the original network.

Table 17 Estimation delay probability: approximation vs simulation

N1 ρ1 Approximation Simulation Difference15 0.7995 0.3584 0.3695 3.01%

16 0.7421 0.2387 0.2531 5.66%

17 0.6924 0.1581 0.1723 8.26%

18 0.6489 0.1040 0.1166 10.78%

19 0.6106 0.0680 0.0783 13.11%

20 0.5765 0.0442 0.0522 15.27%

21 0.5460 0.0286 0.0346 17.41%

We also conducted a numerical study to validate the accuracy of the approximation when varying thethreshold level at which the hospital begin diverting ambulances. We fix the parameters described above aswell as the ICU size at 18 beds and the ED size at 15 beds. We vary the threshold and compute the delayprobability and the diversion probability. Table 18 reports these probabilities as well as the accuracy of thediversion probability approximation. Consistent with our explanation above, as the hospital increases thethreshold, the diversion probability decreases, yet the probability of being delayed increases.

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 33

Table 18 Estimation diversion probability, approximation vs simulation, as a function of the boarding-beds-threshold

K Delay probability Diversion Approximation Diversion Simulation Difference5 0.2520 0.1034 0.1093 5.36%

6 0.3011 0.0907 0.0951 4.63%

7 0.3584 0.0805 0.0829 3.00%

8 0.4250 0.0720 0.0743 3.14%

9 0.5023 0.0649 0.0665 2.36%

10 0.5916 0.0589 0.0599 1.63%

11 0.6944 0.0537 0.0542 0.87%

Table 19 Estimation diversion probability, approximation vs simulation, as a function of the probability of admitting awalk-in patient

pw ρ2 Approximation Simulation Difference0.05 0.613 0.0041 0.0042 2.01%

0.1 0.653 0.0057 0.0059 3.41%

0.2 0.733 0.0112 0.0117 3.96%

0.3 0.813 0.0220 0.0231 4.95%

0.4 0.893 0.0404 0.0433 6.72%

0.5 0.973 0.0681 0.0744 8.49%

Lastly, we conducted a numerical study that aimed at testing the sensitivity of the approximation toviolation of the condition λa(1− δ)pa >> λwpw. We fixed the parameters discussed above as well as thesize of the hospital to 30 beds and computed the diversion probability for different values of pw. As weincrease the probability above 0.2 the condition is violated. Table 19 reports the comparison between theapproximation and the simulation and shows that even when the condition is violated the approximationis still fairly accurate. The main conclusion from this table is that while the assumption is required for thepurpose of analytical tractability, even if the condition is not satisfied, the accuracy of the model is quitehigh.

E. Details of the fluid approximationE.1. Diversion ProbabilityLet Q2(t) denote the number of patients that have completed service in the ED but are yet to completeservice in the inpatient department including the boarding patients (Figure 4(c)). K0 (< N2) denotes thenumber of patients in the inpatient department at time 0 and let τ ′ represent the first time that diversionstarts, i.e., τ ′ = min{t :Q2(t) =N2 +K}.

From t = 0 to t = τ ′, i.e., when not on diversion, the nominal arrival rate to the inpatient departmentcan be approximated by λhapa + λhwpw since ambulance arrivals receive priority over walk-in arrivals andwe assume that λhapa >> λhwpw and N1/m1 > λhapa + λhwpw > N2/m2. Thus, from t = 0 to t = τ ′, thenumber of patients in the inpatient department including the boarding patients increase at a rate of r1 =λhapa +λhwpw−N2/m2. Thus,

N2 +K =K0 + r1τ′. (11)

Next, consider the diversion period, i.e., from t = τ ′ to t = τ . Since the fluid model does not captureindividual patient level dynamics, we model diversion as a fraction of ambulances that are not acceptedby the ED, using a parameter 0< δ′ < 1. Thus, the nominal arrival rate to the inpatient department during

Allon, Deo, Lin: Ambulance Diversion34 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

diversion is λhapa(1− δ′) + λhwpw. Since Q2(t) is equal to N2 +K at all times during the diversion, theflows in and out of the ED must be balanced. Hence,

λhapa(1− δ′) +λhwpw =N2/m2. (12)

At time τ , when the peak period ends and arrival rates jump from high to low, the number of patientsQ2(t) depletes at a rate of r2 =N2/m2−λlapa−λlwpw and drops below N2 +K thereby ending diversion.Since time T is also time 0 of next decision cycle, Q2(T ) = Q2(0) = K0. Note that balancing the flowbetween times t= τ and t= T , we obtain

N2 +K −K0 = r2(T − τ). (13)

Now, we can approximate the diversion probability as δ = δ′(τ − τ ′)/T since the ED diverts a fraction δ′

of the ambulances during the time between τ ′ and τ . Thus, combining (12), (11) and (13) we have

δ=λhapa +λhwpw−N2/m2

λhapa

(1− (N2 +K −K0)(r1 + r2)

r1r2T

).

While the above is a closed-form expression for the diversion probability δ, it contains several parametersthat are not observable in the data. Hence, we use the fact that the utilization of the inpatient department

is given by ρ2 = 1−

((N2−K0)2

2r1+

(N2−K0)2

2r2

)TN2

. In addition, we restrict ourselves to analyzing the diversionprobability δ for a sequence of hospitals where the arrival rate scales linearly with the size of the inpatientdepartment, i.e., λji = βjiN2, i∈ {a,w}, j ∈ {h, l}.

Making these two substitutions, we obtain:

δ=βhapa +βhwpw− 1/m2

βhapa

(1−

√2(1− ρ2)

T− K

N2T

)(14)

It is clear from the above expression that δ is decreasing in θ2 =√

1− ρ2 and decreasing in κ′ = KN2

wherewe use κ′ to distinguish it from κ used in the previous section. Note that if the arrival rate increases lessthan linearly with the size of the hospital, the diversion probability will tend to zero for very large hospitals.On the other hand, if the arrival rate increases more than linearly with the size of the hospital, the diversionprobability will converge to one for very large hospitals. Since we do not observe both of these scenarios inour data, we believe that choosing a linear scaling provides the best explanatory power to the model basedon fluid approximation.E.2. Delay ProbabilityFigure 4(d) depicts the build-up and draw-down of the number of patients that have not completed theirservice in the ED, Q1(t). We assume that the day begins with an empty ED, i.e., Q1(0) = 0. Before thediversion begins, i.e., from t = 0 to t = τ ′, the number of patients in the ED builds up since λha + λhw >N1/m1. During diversion, i.e., from t = τ ′ to t = τ , rate of build up (or draw down) changes to λhw −(N1 −K)/m1 due to simultaneous changes in arrival rate (due to ambulance diversion) and service rate(due to blocking). After diversion, i.e., from t= τ to t= T , the number of patients draws down at the rateof N1/m1 − λal − λwl until all patients are discharged. Note that in this model, the ED operates under twoextreme regimes: arrivals to the ED either surely wait for a bed when Q1(t)<N1 or surely do not wait forthe bed when Q1(t)>N1. We define the fraction of time spent in the latter regime as the delay probabilityfor this model.

In order to show that an optimal threshold K is increasing in N1, we make two observations. First, forfixed N1, an increase in K leads to an increase in the slope of Q1(t) during periods of diversion, i.e., fromt = τ ′ t = τ thereby increasing Q1(t) ∀ t > τ ′ and hence increasing the probability of delay. Second, forfixed K, an increase in N1 leads to a reduction in slope of Q1(t) from t= 0 to t= τ and an increase in theslope for t > τ thereby reducing Q1(t) ∀ t and hence decreasing the probability of delay. Combining theabove two arguments, we can see that for a fixed delay probability, the threshold K is increasing in N1.

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 35

F. Simulation study on the effect of aggregate dataOne of the main limitations of our data is that it is aggregate at the annual level, which introduces certaindifficulties in our analysis and in interpreting our results. For example, we may be underestimating the im-portance of temporal undercapacity or overcapacity. In order to understand the impact of data aggregation,we conducted a simulation study. In this study we chose a sample of 30 random hospitals from our dataset. For each of these hospitals, we had the arrival rate, the number of beds, and the diversion probability,which allowed us to find, the service rate that resulted in the reported diversion probability (using a simpleline search). Based on this data, we simulated 12 months of data for this random cohort of 30 EDs. We thenestimated an OLS model on the cross-sectional data that represented the monthly average of each hospi-tal. Then, we also estimated a fixed effects model on the panel data (12 observations of 30 hospitals) thataccounted for hospital and time fixed effects. The results of this analysis are shown in Table 20.

Note that it is not possible to estimate the effect of ED size (REL ED SIZE) in the panel data becausethe number of ED beds does not vary over time in our simulation and hence is subsumed within the hospitalfixed-effect. We find that the interaction effect is significant in panel estimation even though it is not so inthe cross-sectional estimation. The coefficient of normalized ICU capacity is higher in absolute magnitudein the cross-sectional estimation but note that it is only at the average value of the relative ED size. Inconclusion, based on this simulation study, we believe that our empirical findings will be stronger if paneldata is used instead of the cross-sectional data.

G. FIML estimates for all models and specifications

Allon, Deo, Lin: Ambulance Diversion36 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

Table 20 Results of estimating fixed-effect and OLS models on panel and cross-sectional simulated data respectivelyPanel estimation Cross-sectional estimation

Intercept -1.69∗∗∗ (0.09) -1.66∗∗∗ (0.12)REL ED SIZE CTR 0.12∗∗ (0.04)NORM ICU CAP CTR -0.07∗ (0.04) -0.38∗∗∗ (0.09)REL ED SIZE CTR*NORM ICU CAP CTR 0.11∗∗∗ (0.03) -0.02 (0.02)CS1 0.21∗∗ (0.1)CS2 0.09 (0.11)CS3 1.57∗∗∗ (0.22)CS4 0.09 (0.11)CS5 1.01∗∗∗ (0.11)CS6 0.1 (0.11)CS7 0.55∗∗∗ (0.11)CS8 1.55∗∗∗ (0.12)CS9 1.92∗∗∗ (0.13)CS10 0.15 (0.11)CS11 0.64∗∗∗ (0.11)CS12 0.36∗∗∗ (0.11)CS13 1.22∗∗∗ (0.12)CS14 1.62∗∗∗ (0.13)CS15 0 (0.16)CS16 1.05∗∗∗ (0.14)CS17 0.23∗∗ (0.1)CS18 0.49∗∗∗ (0.11)CS19 -0.27∗ (0.14)CS20 1.15∗∗∗ (0.12)CS21 -0.2 (0.18)CS22 0.28∗∗∗ (0.1)CS23 0.05 (0.28)CS24 -0.2 (0.16)CS25 -0.08 (0.16)CS26 1.43∗∗∗ (0.12)CS27 -0.1 (0.23)CS28 -0.14 (0.14)CS29 1.27∗∗∗ (0.12)TS1 -1.01∗∗∗ (0.07)TS2 -0.91∗∗∗ (0.07)TS3 -0.83∗∗∗ (0.07)TS4 -0.68∗∗∗ (0.07)TS5 -0.58∗∗∗ (0.07)TS6 -0.5∗∗∗ (0.07)TS7 -0.4∗∗∗ (0.07)TS8 -0.29∗∗∗ (0.07)TS9 -0.22∗∗∗ (0.07)TS10 -0.11∗ (0.07)TS11 -0.09 (0.07)

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 37

Table 21 Estimation results for selection models based on diffusion approximation.

Spec A Spec B Spec CLEVEL EQUATIONINTERCEPT 7.33∗∗∗

(0.16)7.40∗∗∗

(0.19)7.35∗∗∗

(0.15)

REL ED SIZE HT CTR -0.16∗∗

(0.07)-0.19∗∗

(0.07)-0.04(0.05)

NORM ICU CAP CTR 0.01(0.10)

-0.01(0.10)

-0.08∗∗∗

(0.03)

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) -0.06(0.04)

-0.08(0.05)

-0.04∗

(0.02)

RURAL 1.11∗∗

(0.56)INVESTOR -1.06∗∗∗

(0.35)GOVERNMENT 0.69

(0.59)CHOICE EQUATIONINTERCEPT 0.55∗∗∗

(0.08)0.51∗∗∗

(0.10)0.47∗∗∗

(0.09)

REL ED SIZE HT CTR 0.08∗∗∗

(0.13)0.08∗∗

(0.04)NORM ICU CAP CTR -0.01

(0.04)-0.02(0.05)

(REL ED SIZE HT CTR)*(NORM ICU CAP CTR) 0.03(0.02)

0.03(0.03)

RURAL -0.48∗∗∗

(0.04)-0.95∗∗∗

(0.24)-0.39∗∗∗

(0.04)

INVESTOR -0.24∗∗∗

(0.04)0.23(0.16)

-0.15∗∗∗

(0.03)

GOVERNMENT -0.28∗∗∗

(0.03)-0.59∗∗

(0.26)-0.27∗∗∗

(0.02)

TRAUMA 0.11(0.08)

0.08(0.15)

0.15(0.18)

Log Likelihood -517.19 -507.33 -520.62

Notes: Standard Errors are shown in parentheses. Log Likelihood for the LIML estimation method is correspondingto the Probit model for the selection equation. ∗∗∗p < 0.01; ∗∗p < 0.05; ∗p < 0.10

Allon, Deo, Lin: Ambulance Diversion38 Operations Research 00(0), pp. 000–000, c© 0000 INFORMS

Table 22 Estimation results for selection models based on fluid approximation.

Specification A Specification B Specification CLEVEL EQUATION

INTERCEPT 6.3∗∗∗ (0.37) 6.5∗∗∗ (0.38) 7.73∗∗∗ (0.31)

SQRT ICU CAP 1.55∗∗∗ (0.47) 1.37∗∗∗ (0.47) 0.04 (0.13)

REL ED SIZE FL 4.04 (2.61) 2.85 (2.54) -0.06 (0.06)

RURAL 1.08∗ (0.57)

INVESTOR -0.82∗∗ (0.33)

GOVERNMENT 0.69 (0.58)

CHOICE EQUATIONINTERCEPT 1.08∗∗∗ (0.17) 1.07∗∗∗ (0.18) 0.44∗∗∗ (0.09)

SQRT ICU CAP -0.62∗∗∗ (0.2) -0.64∗∗∗ (0.22)

REL ED SIZE FL -3.64∗∗∗ (1.3) -3.22∗∗∗ (1.15)

RURAL -0.2 (0.17) -0.67∗∗∗ (0.25) -0.40∗∗∗ (0.06)

INVESTOR -0.16∗∗∗ (0.05) 0.21 (0.15) -0.16∗∗∗ (0.06)

GOVERNMENT -0.37 (0.06) -0.69∗∗∗ (0.26) 0.31∗ (0.16)

TRAUMA 0.23∗ (0.12) 0.2∗∗∗ (0.07) -0.27∗∗∗ (0.03)Log Likelihood -512.71 -507.11

Notes: Standard Errors are shown in parentheses. Log Likelihood for the LIML estimation method is correspondingto the Probit model for the selection equation. ∗∗∗p < 0.01; ∗∗p < 0.05; ∗p < 0.10

Allon, Deo, Lin: Ambulance DiversionOperations Research 00(0), pp. 000–000, c© 0000 INFORMS 39

Table 23 Estimation results for selection models based on the basic model.

Specification A Specification B Specification CLEVEL EQUATION

INTERCEPT 7.44∗∗∗ (0.5) 5.21∗∗∗ (0.64) 7.19∗∗∗ (0.17)

ICU CAP -0.1 (0.37) 0.46 (0.31) -0.10 (0.08)

ICU SIZE 0.01 (0.01) 0.02∗∗ (0.01) 0.00 (0.01)

ED SIZE -0.02 (0.02) 0.01 (0.02) 0.00 (0.01)

RURAL -1.97∗∗∗ (0.76) -0.31 (0.24)

INVESTOR 0.13 (0.34)

GOVERNMENT 0.2 (0.61)

CHOICE EQUATION

INTERCEPT -0.27 (0.24) -0.68∗∗ (0.28) -0.31 (0.24)

ICU CAP 0.32∗ (0.19) -0.89∗∗∗ (0.29)

ICU SIZE 0 (0.01) 0.05 (0.25)

ED SIZE 0.02∗∗ (0.01) 0.79∗∗∗ (0.22)

RURAL -0.6∗∗ (0.23) -0.42 (0.29) -0.45∗∗∗ (0.04)

INVESTOR 0.17 (0.19) 0.45∗∗ (0.22) 0.18 (0.19)

GOVERNMENT -0.31∗ (0.19) 0.01 (0.01) -0.33∗ (0.19)

TRAUMA -0.03 (0.17) 0.03∗∗∗ (0.01) 0.01 (0.06)Log Likelihood -519.31 -521.96

Notes: Standard Errors are shown in parentheses. Log Likelihood for the LIML estimation method is correspondingto the Probit model for the selection equation. ∗∗∗p < 0.01; ∗∗p < 0.05; ∗p < 0.10