physician workload and hospital reimbursement: …faculty.london.edu/nsavva/pss_final.pdf ·...

MANUFACTURING & SERVICEOPERATIONS MANAGEMENT

Vol. 14, No. 4, Fall 2012, pp. 512–528ISSN 1523-4614 (print) � ISSN 1526-5498 (online) http://dx.doi.org/10.1287/msom.1120.0384

© 2012 INFORMS

Physician Workload and Hospital Reimbursement:Overworked Physicians Generate Less

Revenue per Patient

Adam Powell, Sergei SavinThe Wharton School, University of Pennsylvania, Philadelphia, Pennsylvania 19104,

{[email protected], [email protected]}

Nicos SavvaLondon Business School, Regent’s Park, London NW1 4SA, United Kingdom,

[email protected]

We study the impact of physician workload on hospital reimbursement utilizing a detailed data set fromthe trauma department of a major urban hospital. We find that the proportion of patients assigned a

“high-severity” status for reimbursement purposes, which maps, on average, to a 47.8% higher payment forthe hospital, is substantially reduced as the workload of the discharging physician increases. This effect per-sists after we control for a number of systematic differences in patient characteristics, condition, and time ofdischarge. Furthermore, we show that it is unlikely to be caused by selection bias or endogeneity in either dis-charge timing or allocation of discharges to physicians. We attribute this phenomenon to a workload-inducedreduction in diligence of paperwork execution. We estimate the associated monetary loss to be approximately101% (95% confidence interval, 004%–109%) of the department’s annual revenue.

Key words : empirical; hospital operations; healthcare reimbursement; workload managementHistory : Received: March 28, 2011; accepted: January 6, 2012. Published online in Articles in Advance

May 4, 2012.

1. IntroductionHospitals in the developed world are facing a grow-ing set of challenges. Not only are they treating anaging population that places a higher demand onhospital resources, but they are also under constantpressure by public as well as private payers to sub-stantially reduce their operating costs. The combi-nation of these effects has led to a substantial andsustained increase in the daily workload of med-ical practitioners. One suggested response to thisworkload increase has been for physicians to spendless time doing paperwork, such as postdischargerecord keeping, to free up time to devote to treatingpatients (Gottschalk and Flocke 2005).1 From an oper-ations management perspective, reducing the timethat highly skilled and expensive servers spend per-forming secondary tasks is desirable. However, reduc-ing the time physicians devote to ancillary activitiessuch as paperwork might have unanticipated yetsignificant clinical and financial consequences. Empir-ical studies have demonstrated that an increase in

1 McCormick et al. (2004) found that the level of paperwork-relatedstress today is so great that two-thirds of physicians would be will-ing to give up 10% of their income for a substantial reduction inpaperwork.

physician workload has an adverse impact on thequality of residents’ discharge summaries (Coit et al.2010). Workload-induced degradation in the qualityof medical notes and, in particular, of the dischargenote, can have implications for follow-up care. Fur-thermore, because the discharge note is the primaryinput into the billing process, it may also have impli-cations for hospital reimbursement, which is the focusof this paper.

Utilizing detailed reimbursement data from thetrauma department of a major urban hospital,we study the impact of patient-generated workloadon hospital reimbursement. Our main finding is thatthe proportion of patients who are assigned a high-severity (as opposed to a low-severity) diagnosisrelated group (DRG) code, which maps, on average,to a 4708% higher reimbursement payment for thedepartment, is significantly reduced when physiciansexperience higher than average workloads. As a pre-view of our results, Figure 1 shows the proportion ofpatients assigned a high-severity DRG code as a func-tion of the number of same-day discharges completedby the discharging physician. As evident in Figure 1,a physician who discharges a single patient on anygiven day, and therefore has only one discharge note

512

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.

Powell, Savin, and Savva: Physician Workload and Hospital ReimbursementManufacturing & Service Operations Management 14(4), pp. 512–528, © 2012 INFORMS 513

Figure 1 Proportion of Patients Assigned a High-Severity DRG Code(and 90% Confidence Interval) vs. Number of Same-DayDischarges Performed by Discharging Physician for theTrauma Department of a Major Nonprofit Urban Hospital,January 1, 2006, to Septemper 28, 2010

1 2 3 4+

Number of same-day discharges by physician

25

30

35

40

45

Pro

port

ion

of p

atie

nts

assi

gned

50

55

high

-sev

erity

DR

G (

%)

Note. A high-severity DRG code assignment corresponds, on average, toa 47.8% higher reimbursement payment than a low-severity DRG codeassignment.

to write, has a 15% higher chance of getting a high-severity assignment for his patient compared witha physician who has to do three or more same-daydischarges.

The significant reduction in high-severity assign-ment as workload increases persists even after wecontrol for differences in patient characteristics (gen-der and age), treatment characteristics (length of stay,condition, and physician fixed effects), and differ-ences in discharge day (day of the week and calendar-month fixed effects). We find that this result is robustto alternative model specifications, and we performseveral robustness checks to rule out endogeneity indischarge timing, endogeneity in the allocations ofdischarges to physicians, and selection bias based onobservable factors as the root causes of this observa-tion. We also find that besides the number of same-day discharges a physician performs, the numberof inpatients under a physician’s care also has animpact on the probability that a discharged patientis assigned a high-severity DRG code. Interestingly,we find that department-level workload does notseem to have an impact on severity assignment,which provides support for the hypothesis that physi-cians do not share their paperwork load.

We attribute this significant reduction in high-severity assignments to workload-related degradationin the quality of discharge notes. Notes compiledunder high-workload conditions are more likelyto miss information on patient complications andcomorbidities (CCs), leading to these patients beingassigned a lower-severity DRG code than the onemedically warranted, a phenomenon one could call

“undercoding.”2 This explanation is consistent with acommonly held opinion regarding the value of dili-gent paperwork execution during the discharge pro-cess: Esposito et al. (2006) report that 63% of generalsurgeons surveyed do not believe that paying closerattention to discharge notes will increase reimburse-ment. Nonetheless, Reed et al. (2003) demonstrate thatincreased verification and collaboration in the traumabilling process can increase payments.

Because a high-severity DRG code assignment isassociated with an average extra payment of 4708%,we estimate the relative loss of revenue to the depart-ment due to workload-related undercoding to beapproximately 101% (95% confidence interval (CI),0.4%–1.9%). This amount, even at the lower end ofthe confidence interval, is substantial, and it wouldmore than cover the cost of reasonable changes inthe hospital’s discharge process designed to preventsuch prevalent undercoding. For example, even at thelower end it would be enough to cover the salaries oftwo full-time nurses.

Furthermore, we find that the impact of workloadon the probability that a patient is undercoded ismoderated by the frequency with which the traumadepartment is handling the patient’s condition. Themore frequent a condition is to the department, thesmaller the impact an extra same-day discharge willhave on the probability that a patient is assigned ahigh-severity DRG code. This reflects the likely under-lying causes of DRG undercoding. Physicians, in gen-eral, do their best to produce high-fidelity dischargenotes that include all relevant information regardingthe patient, the treatment, and any complications orcomorbidities. Such complete documentation wouldlead to high-severity patients being correctly assignedthe high-severity DRG code. However, when otheractivities compete for their time and attention, physi-cians may not put as much time and/or effort intowriting the discharge note. This is less problematic forfrequent cases for which the department has devel-oped organizational routines and therefore physiciansknow what needs to be written in the discharge note.It is, however, a problem for less common cases inwhich physicians would have to do more extensiveresearch on what information should be recorded.

Beyond hospital reimbursement, our paper haswider implications for operations management.It points to a behavioral impact on system perfor-mance that, to the best of our knowledge, has notreceived substantial attention in the operations man-agement literature. The increase in workload not onlyaffects the speed and the quality of the primary ser-vice (see Kc and Terwiesch 2009), but also compro-mises the ancillary activities that are secondary to

2 We would like to thank the anonymous reviewer for suggestingthis term.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.

Powell, Savin, and Savva: Physician Workload and Hospital Reimbursement514 Manufacturing & Service Operations Management 14(4), pp. 512–528, © 2012 INFORMS

the quality of outcomes but essential to generatingincome. Moreover, this has implications for the opti-mal design of prospective reimbursement systems inhealthcare and elsewhere because it shows that theability to accurately observe the system, a critical fac-tor in the successful implementation of prospectivereimbursements, may depend on the workload levelthe system is subjected to.

The rest of this paper is organized as follows. In §2we provide a detailed account of hospital reimburse-ment procedures and of the operations of the traumadepartment we studied. We then present a literaturereview in §3, followed by a discussion of our mainhypotheses in §4. We present data and results in §§5and 6. We conclude with a discussion of our mainfindings and of implications for hospital operationsalong with suggestions on mitigating measures in §7.

2. Hospital Reimbursement andthe Trauma Department

2.1. Hospital Reimbursement and DRG CodesThe Centers for Medicare and Medicaid Services(CMS) operates a Prospective Payment System (PPS)to reimburse hospitals for the care of Medicarepatients (Mayes 2007).3 Under this system, CMS reim-burses hospitals for the treatment of eligible patientsby predetermined amounts based on patients’ diag-noses (summarized by single three-digit DRG codes),rather than on actual procedures performed. The dif-ference between the costs a hospital incurs and theMedicare reimbursement amount is the responsibil-ity of the hospital (Goldsmith 1984). Thus, althoughthe treatment provided to a patient determines thecost of rendered services, it does not directly influ-ence the revenues that the hospital receives underthe PPS, which are determined by diagnosis andpatient characteristics, rather than by physician treat-ment decisions. CMS deliberately uses prospectivepayments rather than cost-based payments to elimi-nate the incentive for physicians to overtreat patientsto increase revenue. This system has long beenrecognized as promoting “yardstick” competition(Shleifer 1985).

To more accurately reflect the costs of providingservices, as well as to reduce the potential incen-tive for hospitals to selectively treat more prof-itable patients, CMS enables differential paymentsto be made, according to the severity of a patient’scondition. Thus, several diagnoses have more thanone DRG codes associated with them (Cleverley et al.2010, Chap. 3). For instance, there are three different

3 Many private insurers have chosen to follow suit and use the diag-nosis categories provided by CMS as the basis of their paymentsystems (Reinhardt 2006).

DRG codes associated with the diagnosis “Traumaticstupor and coma of greater than one hour” (Medi-care Severity–Diagnosis Related Group, version 27).4

Patients without CCs are classified as DRG 084,whereas patients with CCs are classified as DRG083, and patients with major CCs are classified asDRG 082.

The actual DRG code assigned to a patient isdetermined by a three-step process (Cleverley et al.2010, Chap. 2). First, during the patient’s care pro-cess but mostly during the patient’s discharge, physi-cians record notes. After the patient is dischargedfrom the hospital, a professional coder translates thephysicians’ notes into diagnostic and treatment codes.Once coding is complete, the diagnosis and procedurecodes, along with patient characteristics (age, sex, dis-charge status, complications, and comorbidities) arepassed to a “DRG grouper” software application,which determines the ultimate DRG code assignment(Centers for Medicare and Medicaid Services 2007).Note that although a patient might be assigned multi-ple diagnosis and procedure codes, the patient is ulti-mately assigned a single DRG code for the durationof their stay at the hospital. The professional coders incharge of transcribing the discharge notes come froma central pool shared by the hospital. Because thebilling processing does not need to be done urgently,a variable time lag may occur between the time apatient is discharged and the time a patient’s dis-charge note is processed by a coder. Furthermore,coders are assigned to transcribe notes from differ-ent parts of the hospital in no systematic fashion.Although coding is considered to be accurate enoughto be used as the basis for most U.S. hospital pay-ments, Fisher et al. (1992) found that there exists avariation in coding accuracy across medical condi-tions. This implies that although the accuracy of cod-ing may be related to the nature of the content beingcoded, for any given condition there is no substantialvariation in accuracy. It is worth noting that becauseof the time lag between discharge and coding, andbecause of coders being a pooled resource shared bya number of departments across the hospital, it isunlikely that coding errors (if any) made by profes-sional coders would be correlated with physician dis-charge workload.

There is an ongoing debate in economic literatureas to whether hospitals engage in “upcoding” behav-ior, where patients are switched from low-paying tohigh-paying DRG codes on grounds other than med-ical. Early studies (Carter et al. 1990) found littleevidence of upcoding, whereas more recent studies

4 See http://www.cms.hhs.gov/AcuteInpatientPPS/downloads/FY_2010_FR_Table_5.zip (accessed June 16, 2011) for a complete list.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


(Silverman and Skinner 2004, Dafny 2005) demon-strated that hospitals did engage in upcoding forselective DRGs where such behavior is profitable.They showed that for-profit hospitals engage in theseactivities more actively than nonprofit or governmenthospitals. These studies are based on data from thelate 1980s and early 1990s, when coding practiceswhere not audited as rigorously as they are today(in particular, under Section 302 of the Tax Reliefand Healthcare Act of 2006).5 Because the hospital westudy is a not-for-profit institution, and because of therigorous auditing regime that hospitals are subject to,it is unlikely that upcoding is a significant problem atthe hospital we investigate. Therefore, any reductionin DRG severity assignment associated with physiciandischarge workload is unlikely to be due to workloadpreventing physicians from attempting to “upcode”patients.

2.2. The Operations of the Trauma DepartmentWe have decided to focus on the trauma departmentsince we believe that econometrically it offers thecleanest test case for our enquiry, because a num-ber of confounding factors are less pronounced in thetrauma treatment environment.

The first advantage of studying the trauma depart-ment has to do with the nature of the admissionprocess. Trauma department patients, because theyare admitted as a result of accidents or violence,6

have a restricted ability to exercise choice, as do thephysicians working for the trauma department. Morespecifically, neither trauma patients nor their physi-cians are able to select the location or timing of theiradmission. Patients are allocated to the admittingphysician who happens to be on duty at the time theyarrive at the hospital and have neither the time northe ability to research or select their care providersor even the hospital they are admitted to. For thesame reason, trauma physicians are also restricted inboth the number and the case mix of patients theytreat. They treat the patients that happen to be admit-ted while they are on admission duty. This inabilityto load balance admissions creates greater workloadvariability than is present in departments in whichpatient admissions may be scheduled in advance.

The second reason we chose the trauma departmenthas to do with the fact that, during the patient’s entirehospital stay, his or her care is coordinated mainly byone physician, known as the “practitioner of record.”The practitioner of record in the vast majority of cases

5 See http://www.cms.gov/recovery-audit-program/ (accessedJune 16, 2011).6 Patients arrive at a typical trauma center both through groundtransportation and by helicopter, traveling distances of up to40 miles (Cocanour et al. 1997).

is the admitting physician. While this physician maybe supported by a team of residents and nurses, theprovided treatment is ultimately his or her responsi-bility. Similarly, although the team contributes to writ-ing the discharge notes, it is the practitioner of recordwho is responsible for their contents.

Third, trauma department physicians have less flex-ibility than their colleagues in other departments indetermining when to discharge patients under theircare. As the patient recovers, the trauma departmentmonitors the patient’s health through the case man-agement team, which is responsible along with thepractitioner of record for planning the patient’s even-tual discharge. Because of the complexity of mosttrauma cases, the trauma department remains incharge of patients even if the patient is transferredto another department within the hospital. The casemanagement team meets every weekday to discusswhen patients will be ready to leave the hospital andwhat services they will need once they depart (Curtis2007). When the case management team decides thata patient is ready to leave, the discharge process isinitiated by the practitioner of record.

The fourth advantage of the trauma departmentfor the purposes of studying the impact of work-load on severity assignment, and thus on income,has to do with the proportion of patients allocatedto a high-severity DRG code in the trauma depart-ment. As is shown in Figure 1, this proportion isnear 50%, which is relatively high compared to theother departments within the hospital that we stud-ied, which makes identification of any workload-induced systematic changes in severity assignmenteasier to observe.

3. Literature ReviewOur research relates to the growing body of empiri-cal research on healthcare operations in general andto the literature on the impact of workload in particu-lar. On one hand, there is evidence that higher physi-cian volume, i.e., higher number of patients treated,leads to higher-quality care for a number of con-ditions (e.g., cancer treatment (Hillner et al. 2000),coronary angioplasty (Jollis et al. 1997), and pneu-monia treatment (Lin et al. 2008)). Similarly, thereis also evidence that higher volume at the hospitallevel leads to higher-quality care as well (Luft et al.1979, Jollis et al. 1997, Hillner et al. 2000, Birkmeyeret al. 2002, Macias et al. 2009). Both of these effectscan be explained by individual and organizationallearning effects as well as the benefits of specializa-tion. On the other hand, such an increase in vol-ume, when coupled with an increase in providerworkload, i.e., an increase in the number of patientsseen by individual nurses, physicians, or the depart-ment per unit of time, has unintended consequences.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Kc and Terwiesch (2009) find that workload affectsboth service times and mortality rates. They showthat as the workload increases, procedures take lesstime to complete, but that this excess speed comesat the expense of patient safety as mortality ratesincrease during high-workload periods. Furthermore,they show that the decrease in completion time is notsustainable, because medical practitioners eventuallytire and slow down. Kuntz et al. (2012) find that orga-nizational workload has a nonlinear impact on qualityof care. They show that outcomes deteriorate whenworkload increases from already high workload lev-els, whereas they improve if workload is increasedfrom low workload levels. Green et al. (2010) studya different effect of workload on the healthcare sys-tem. They show that workload is linked to nurseabsenteeism and find that absenteeism is exacerbatedon days when the workload is anticipated to behigher because of insufficient staffing levels. Work-load can also inhibit organizations from learning frompast mistakes. For example, Tucker and Edmondson(2003) note, based on in-depth field research, thatheavy workload is in part responsible for nurses’focus on providing “locally optimal” solutions forindividual process failures at the expense of spendingtime identifying and rectifying the root cases of suchfailures. Such behavior does not create a conduciveenvironment for long-term and systemwide changesthat can improve hospital efficiency. Our work com-plements this line of research by demonstrating anunintended consequence of increased workload onhospital reimbursement rates. Thus, our paper alsocontributes to the debate on how human factors affectthe productivity of a system (Schultz et al. 1998, 1999;Oliva and Sterman 2001; Boudreau et al. 2003; Masand Moretti 2009).

When organizations perform tasks repeatedly, theylearn and develop routines adapted to their needsand environments (Nelson and Winter 1982). Rou-tines reduce the need for organizations to discoversolutions every time they face a problem, and theseroutines evolve over time through experiential learn-ing. Although repetition allows an organization todevelop routines, Zollo and Winter (2002) argue thatthe frequency with which an organization executesspecific tasks has a dramatic impact on the devel-opment of such routines. They argue that tasks per-formed at low frequency suffer significant losses intheir capability-building power. Furthermore, theyargue that for such tasks, explicit learning mech-anisms, such as knowledge codification, will bemore effective in building organizational routines asopposed to relying on tacit experience accumulation.Although our research does not explicitly test theimpact of frequency on learning to develop organiza-tional routines, by investigating the moderating effect

of task frequency on workload-related undercoding,we indirectly examine one of its consequences. Ourfindings are consistent with the hypothesis that devel-oping organizational routines is more effective forhigh-frequency tasks.

Finally, our work is related to the literature onincentives in hospital reimbursement (see a reviewby Newhouse 1996). For example, Shleifer (1985) dis-cusses how the Medicare DRG-related reimbursementpolicy provides incentives for monopolist providersto reduce costs. Fuloria and Zenios (2001) present aprincipal–agent payment system and show that socialwelfare is maximized when reimbursements are out-come adjusted. In general, this literature assumes thateven if the treatment is private information for theprovider, the payer is able to verify the diagnosisand health outcome of a patient. As our researchshows, this assumption might not be valid, becausethe documentation generated by physicians is sensi-tive to operational conditions such as system work-load. Therefore any reimbursement plan that aims toalign incentives between payers and providers needsto take this endogeneity into account.

4. Development of HypothesesBroadly speaking, there are three main activities thatcompete for the time and attention of healthcareproviders. These are the admission of new patients,the monitoring and continuation of care for alreadyadmitted patients, and the discharging of sufficientlyrecovered patients. The admissions process at thetrauma department we study is decoupled from careand discharge which are treated more as business asusual. Admitting physicians serve at the admissionsbay in shifts, and while they are working on admis-sions, they are not treating or discharging any of theirpreviously admitted patients. For the development ofour hypotheses, we focus on the last two activities.

Each individual discharge generates a substantialamount of physician-related workload, and, as a con-sequence, physicians have a limited amount of timeto devote to each discharge note. Similarly, the timea physician spends attending to inpatients competeswith the time he or she devotes to writing dischargenotes. Consistent with previous research in this area(Coit et al. 2010), we conjecture that the quality andthe level of detail of each discharge note will decreaseas the workload, defined as the number of same-daydischarges performed by, or the number of in-patientsunder the direct care of, a physician increases. Morespecifically, we conjecture that a discharge note com-piled under high-workload conditions will have aless detailed record of the patient’s comorbidities andcomplications. This conjecture reflects the fact thatcompiling such evidence takes time, effort, and atten-tion to detail, all of which are likely to be in short

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


supply when the discharging physician is under highworkload pressure. Although the adverse impact ofworkload on the quality of the discharge note mightbe symmetric across patients of different severities,we expect it to have an asymmetric impact on sever-ity assignment, because for a patient to be awardedthe higher-severity status, the discharge note needsto provide detailed evidence for comorbidities andcomplications. If such evidence is missing or is notcomplete, then a patient whose condition should havebeen classified as high severity may be given a low-severity assignment. Conversely, if such evidence ismissing from the discharge note of a patient whosecondition should be classified as low severity, thenthe patient is not more likely to be assigned thehigh-severity status. Therefore, as the discharge work-load increases we would expect to see fewer patientsassigned a high-severity DRG code.

Hypothesis 1. As the number of same-day dischargesperformed by a physician increases, the probability thateach of these discharged patients is awarded a high-severityDRG code decreases.

Hypothesis 2. As the number of inpatients under thecare of the discharging physician increases, the probabilityof a discharged patient being awarded a high-severity DRGcode decreases.

The benefits of learning from experience havelong been a part of the management literature,from Smith’s (1776) arguments for the benefits ofspecialization in pin production, to Weber’s (1978)description of bureaucracies undergoing experien-tial learning, to Cyert and March’s (1963) A Behav-ioral Theory of the Firm. When organizations performtasks repeatedly, they learn and develop routinesadapted to their needs and environments (Nelson andWinter 1982). Routines reduce the need for organi-zations to discover solutions every time they face aproblem, and these routines evolve over time throughexperiential learning. It is likely that trauma depart-ments have more developed routines for handlingpatients with common conditions than those withrare conditions, because common conditions providemore opportunities for experiential learning. Such fre-quent interactions are well suited for the develop-ment of a “capability-building mechanism based ontacit accumulation of experiences in the minds of‘expert’ personnel” (Zollo and Winter 2002, p. 347).When physicians at the trauma department are underhigh workload pressure, they can fall back on opti-mized organizational routines when handling patientswith familiar conditions, but they must expend moreeffort on discovering the care and documentation pro-cess when handling patients with rarer conditionsfor which such routines are not as well established.

Because these routines are established at the level ofthe department as opposed to the individual physi-cian, our third hypothesis is that the impact of dis-charge and inpatient workload on the probability thata patient is undercoded is moderated by the fre-quency with which the trauma department treats aspecific condition.

Hypothesis 3. The negative relationship betweenphysician workload on discharge day and probability ofhigh-severity DRG code assignment is moderated by thefrequency with which the patient’s condition is treated bythe trauma department.

5. Data DescriptionTo empirically test the hypotheses articulated inthe previous section, we utilize a detailed data setfrom the trauma department of a large urban hos-pital. Our data span a period of 45 months, start-ing on January 1, 2006, and ending on September28, 2010, and represent a complete description oftreatment activities and patient characteristics for all7,100 patients admitted to the trauma department forthe reported period. For each admission we knowthe admission date, discharge date, patient charac-teristics (gender, age, race, insurance plan), treatmentcharacteristics (assigned DRG, days stayed in hos-pital, physician identifier), and billing data. Duringthe period for which we collect data, CMS movedfrom version 24 of the DRG code through to ver-sion 27. The differences between versions 25 to 27were minor, whereas the differences between ver-sion 24 and version 25, which was implemented onOctober 1, 2007, were more drastic. In version 24,many DRG classifications were “paired” to reflect thepresence of CCs, whereas in version 25 the pairingwas, in many instances, replaced with a three-tieredsystem: absence of CCs, presence of CCs, and a higherlevel of presence of major CCs (see Centers for Medi-care and Medicaid Services 2007). To allow a like-to-like comparison the hospital reassigned all the claimsin our data set using the official software for the DRGversion 25 system.

To be able to examine how severity assignments aremade, we have lumped related DRGs that are onlydifferentiated by severity into the same “base DRG.”For those diagnoses with two or three severity levels,we created a separate severity variable, which was setto 0 for the lowest severity and 1 for the other sever-ity level(s). For the example of “Traumatic stupor andcoma of greater than one hour,” which has three dif-ferent DRGs associated with it—084, 083, and 082 forpatients without CCs, with CCs, and with major CCs,respectively—all of these DRGs would fall under thesame base-DRG code, whereas patients with DRG 084would be mapped to a severity level of 0, and patients

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Table 1 Descriptive Statistics and Correlation Table

Descriptive statistics Correlation table

Variables Obs. Mean SD Min Max (1) (2) (3) (4) (5) (6)

(1) High-Severity Assignment 61165 0047 0050 0 1 1000(2) Physician’s Discharge Workload 61165 1070 0091 1 6 −0010∗∗∗ 1000(3) Physician’s Admission Workload 61165 0026 0083 0 11 0000 0002 1000(4) Physician’s Inpatient Workload 61165 3069 2050 0 15 −0011∗∗∗ 0019∗∗∗ −0000 1000(5) Total Patient Discharges 61165 5048 2048 1 15 0001 0039∗∗∗ 0002∗ 0006∗∗∗ 1000(6) Total Inpatients 61165 25041 7048 0 50 −0001 0006∗∗∗ 0003∗∗ 0038∗∗∗ 0016∗∗∗ 1000(7) Total Patient Admissions 61165 4034 2030 0 13 −0000 0007∗∗∗ 0011∗∗∗ 0019∗∗∗ 0020∗∗∗ 0044∗∗∗

∗∗∗, ∗∗, ∗ denote statistical significance of the correlation coefficient at 1%, 5%, and 10% confidence levels, respectively.

with DRG 082 and 083 would be mapped to a severitylevel of 1.

To construct workload measures, such as the num-ber of inpatients treated at the trauma department,we rely on the admission and discharge data. For thefirst few days of the data set, we do not have an accu-rate picture of how many patients the department istreating because these patients were admitted beforeour data set starts. To avoid this left-censoring prob-lem, we exclude from the workload regressions the 59discharges that took place in the first 20 days of thedata set. Because the length of stay is less than 20 daysfor about 96% of patients, the choice of 20 days asa cutoff is a conservative one. Similarly, we cannotconstruct workload measures for patients dischargedafter September 28, 2010, when we no longer haveadmissions data. To avoid this right-censoring issue,we exclude the 11 discharges that occurred after thelast admission date.

Although we use all remaining patients to con-struct workload measures, for the severity assign-ment regressions we also exclude any base-DRG codethat does not have a high-severity assignment (thisexcludes 657 observations) and any base-DRG codesthat have only high- or only low-severity assignmentsin our data sample (188 data points). For example, onesuch excluded DRG is that for chest pain, DRG 313,which has no high-severity counterpart. We also dropan additional 11 observations where the practitionerof record was either not available or was not a traumadepartment regular, as well as 9 patients whose agewas recorded to be between 124 and 128 years, whichwe consider a typo. This leaves 6,165 observations.

Our billing data include the actual payment the hos-pital received and not just the amount charged. Theactual payment can be very different from the amountcharged, because patients might not pay because theyare uninsured, or because insurers negotiate prefer-ential treatment or volume discounts, etc. Havingthe actual payment allows us to examine the impactof workload-induced undercoding on actual financialperformance as opposed to estimated financial per-formance based on charges or prospective payments.

To fully reflect the true severity-related differences inreimbursement, we use reimbursement data startingon October 1, 2007, when DRG version 25 went intoeffect. We also exclude any observations in which thepayment amount is not reported, mainly because thebill has not yet been settled. This happens mostly inthe last three months of data. This leads to a final countof 3,793 patient observations for the reimbursementregression. Of the 3,793 patients we observe, 257 (or6078%) did not pay for the delivered care.

Table 1 shows descriptive statistics for the depen-dent and independent workload-related variables andtheir correlations. The unit of analysis is individ-ual patient discharges. The dependent variable High-Severity Assignment takes the value 1 if a patientreceives a high-severity DRG code and 0 otherwise.The Physician’s Discharge Workload variable stands forthe total number of same-day discharges performedby the discharging physician, whereas the Total PatientDischarges variable designates the respective numberof patients discharged by the entire trauma depart-ment on each discharge date. Note that the averagenumber for inpatient admissions in Table 1 is calcu-lated using only the days on which there was at least asingle discharge. The rest of the variables are similarlydefined. Note that some of the pairwise correlationcoefficients are significant; however, we discoveredno significant multicollinearity problems. Descriptivestatistics for hospital payments and for patient ageand length of stay appear in Table 2. All analysesand econometric tests presented were implemented inSTATA/IC 10.0 for Windows.

6. Econometric AnalysisOur first model focuses on testing our hypotheses onthe influence of the discharge and inpatient work-loads on the probability of high-severity DRG codeassignments. In particular, we estimate the followinglogistic regression:

ln(Prob4SevAssigni1 s1 t = 15

Prob4SevAssigni1 s1 t = 05

)

= �0 +�cControlsi1 s1 t +�1PhysDisLoads1 t

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Table 2 Descriptive Statistics of Patient Age and Length of StayVariables

Variables Obs. Mean SD Min Max

Patient Age (years) 61165 4303 1908 12 106Length of Stay (days) 61165 5049 5049 0 140

+�2PhysAdmLoads1 t +�3PhysInpLoads1 t+�4DepDist +�5DepAdmt +�6DepInpt1 (1)

where the unit of analysis is patient i, dischargedby physician s on date t. The variable SevAssigni1 s1 t

is the indicator for patient severity assignment.It takes the value 1 when the patient receives a high-severity assignment and 0 otherwise. The variablesPhysDisLoads1 t , PhysAdmLoads1 t , and PhysInpLoads1 tdenote the number of discharges, admissions, andinpatients looked after by the discharging physicians on the discharge date t, respectively. The variablesDepDist , DepAdmt , and DepInpt denote the number ofdischarges, admissions, and inpatients of the traumadepartment on the discharge day t, respectively.

Models I and II, presented in Table 3, have differ-ent controls Controlsi1 s1 t . Model I controls for patientcharacteristics (age, gender, and length of stay), andphysician and time fixed effects (calendar month andday of the week), whereas Model II adds base-DRGfixed effects. This last control allows us to modelheterogeneity in the propensity of specific condi-tions to generate high-severity assignments. However,this is computationally and data intensive, because itrequires an additional 111 controls. To address poten-tial correlation of error terms, we cluster standarderrors in both models on physicians.

As evident from Table 3, we find strong support forboth Hypotheses 1 and 2. Both models show that thedischarging physician’s discharge workload as wellas the number of inpatients assigned to the discharg-ing physician have a significant and negative impacton the probability that a patient is assigned highseverity. According to Model I (Model II), the coeffi-cient of the physician discharge workload is −000938,p-value < 0000 (−000853, p-value = 0002), whereas thecoefficient of the physicians inpatient workload is−000289, p-value = 0002 (−000261, p-value = 0003). Fur-thermore, Model I (Model II) suggests that the aver-age marginal effect of an extra same-day dischargeby the discharging physician on the probability ofhigh-severity assignment is −203%, p-value < 00000,(−201%, p-value = 0002), whereas the marginal effectof an extra inpatient under the care of the dischargingphysician is −007%, p-value = 00015 (−006%, p-value =

0003). The physician’s admission workload and allthe department workload measures are not signif-icantly different from zero, confirming that admis-sions are decoupled from the rest of the discharge

Table 3 Results of Models I and II

Dependent variable: Was the patientassigned high severity?

Model I Model IIVariables Logit coefficient Logit coefficient

Physician’s Discharge Workload −000938∗∗∗ −000853∗∗

40002085 40003515Physician’s Admission Workload 000696 000415

40004705 40006075Physician’s Inpatient Workload −000289∗∗ −000261∗∗

40001185 40001205Total Patient Discharges 000097 000193

40001255 40001265Total Inpatients −000031 0000111

400008855 400009485Total Patient Admissions 000099 −0000209

40001145 40001425Constant −1089∗∗∗ −2011

4004395 410825Fixed Effects

Calendar Month Yes YesDay of the Week Yes YesPatient Gender Yes YesPatient Age (5 categorical variables) Yes YesLength of Stay Yes Yes

(5 categorical variables)Physician Yes YesBase DRG No Yes

Observations 6,165 6,165Log likelihood −31273 −21506Pseudo-R2 00233 00412

Note. Errors shown in parentheses are clustered on physicians.∗∗∗, ∗∗, ∗ denote statistical significance at 1%, 5%, and 10% confidence

levels, respectively.

workload and that physicians do not share workloadwith one another. Turning to goodness-of-fit mea-sures, the Hosmer–Lemeshow statistic (with 10 and20 groups) rejects the hypothesis that the logit modelis not the correct model for both Models I and II.As a robustness check we also ran two additionalmodels. First, we estimated a regression model withall the controls of Model I and with the physician-level workload variables, but without the unit-levelworkload variables. Second, we estimated a regres-sion model with all the controls of Model I and withthe unit-level workload variables but without thephysician-level workload variables. The former modelproduced results very similar to the ones presentedin Table 3, whereas the latter produced no signifi-cant results. This confirms that physician-level dis-charge and inpatient workload matter when it comesto severity assignment, whereas unit-level workloadshave no explanatory power in predicting severityassignment.

The linear specification used in Models I and II(Table 3) is, to a large extent, ad hoc. After all,

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


it seems plausible that the impact of an increase inworkload only “kicks in” after a certain thresholdhas been exceeded. To this end, we present a non-linear model specification for the three physician-related explanatory variables in Table 4. The nonlinear

Table 4 Results of Model III

Dependent variable: Was thepatient assigned high severity?

Model IIIVariables Logit coefficient

DV = 1 if Physician’s Discharge −000576Workload= 2 40008715

DV = 1 if Physician’s Discharge −00217∗∗∗

Workload= 3 40007225DV = 1 if Physician’s Admission 000875

Workload> 0 4001205DV = 1 if Physician’s Inpatient 000643

Workload in 2nd Quartile 4001055DV = 1 if Physician’s Inpatient −00129∗

Workload in 3rd Quartile 40007385DV = 1 if Physician’s Inpatient −00139

Workload in 4th Quartile 4001065DV = 1 if Total Patient 00224∗∗∗

Discharges in 2nd Quartile 40007815DV = 1 if Total Patient 000736

Discharges in 3rd Quartile 4001105DV = 1 if Total Patient 001534

Discharges in 4th Quartile 4001095DV = 1 if Total Inpatients in 2nd Quartile 00153

4001125DV = 1 if Total Inpatients in 3rd Quartile 00140

4001445DV = 1 if Total Inpatients in 4th Quartile 000614

4001605DV = 1 if Total Patient 000537

Admissions in 2nd Quartile 4001395DV = 1 if Total Patient 000271

Admissions in 3rd Quartile 40008515DV = 1 if Total Patient 00121

Admissions in 4th Quartile 4001185Constant −2021

410055Fixed Effects

Calendar Month YesDay of the Week YesPatient Gender YesPatient Age (5 categorical variables) YesLength of Stay Yes

(5 categorical variables)Physician YesBase DRG Yes

Observations 6,165Log likelihood −21502Pseudo-R2 00413

Notes. DV, dummy variable. Errors shown in parentheses are clustered onphysicians.

∗∗∗, ∗∗, ∗ denote statistical significance at 1%, 5%, and 10% confidencelevels, respectively.

specification breaks the physician’s discharge work-load into three categories (with one, two, and threeor more discharges, containing 5305%, 3000%, and1505% of the observations, respectively), the physi-cian’s admission workload into two categories (zeroand one or more, containing 8702% and 1208% ofobservations, respectively), and the rest of the vari-ables in quartiles. For each category, we define adummy variable that takes the value 1 when theexplanatory variable is in that category and 0 oth-erwise. For each explanatory variable, we omit thefirst category so that the coefficients should be inter-preted as the difference from the lightest workloadstate. As evident from Table 4, the main finding is thatphysician discharge workload has a substantial andsignificant effect once the discharge workload is equalto or greater than three (marginal effect = −503%,coefficient = −00217, p-value = 00003). Similarly, whenthe number of patients assigned to the same physicianis in the third or fourth quartile, then the probabilitythat a discharged patient is assigned the high-severity DRG code is reduced significantly in thecase of the third quartile (marginal effect = −302%,coefficient = −00129, p-value = 0008), but not signif-icantly in the case of the fourth (marginal effect =

−304%, coefficient = −00139, p-value = 0019). The lackof significance in the case of the fourth quartile,which is observed despite the fact that the coef-ficient is larger in absolute terms than the coeffi-cient of the third quartile, might be indicative ofthe variability in the workload generated by inpa-tients because the standard error of the coefficientis quite high. An individual physicians admissionworkload does not play a significant role in sever-ity assignment. Looking at the department-level vari-ables, there seems to be a positive effect when thedepartment processes a smaller number of dischargesthan the median (marginal effect = 506%, coefficient =

00228, p-value = 0001). The rest of the variables are notsignificant.

We next examine the impact of the frequencywith which the trauma department treats specificconditions on severity assignment. We measure thefrequency by counting the number of instances abase-DRG code appears in the 45 months for whichwe have data. Models IV and V, presented inTable 5, include the interaction of DRG frequencywith the continuous workload variables of Model IIand the dummy workload variables of Model IIIthat were statistically significant. We find that fre-quency has a significant moderating impact on theeffect of discharge workload on severity assign-ment (coefficient = 00000659, p-value = 0002). Theaverage marginal effect of an extra discharge on theprobability a patient is assigned the high-severity

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Table 5 Results of Models IV and V


Model IV Model VVariables Logit coefficient Logit coefficient

DV = 1 if Physician’s Discharge −00152Workload= 2 4001675

DV = 1 if Physician’s Discharge −00474∗∗∗

Workload= 3 4001535Base-DRG Freq× DV = 1 if 00000629

Physician’s Discharge Workload= 2 4000008575Base-DRG Freq× DV = 1 if 0000171∗∗

Physician’s Discharge Workload= 3 4000007615DV = 1 if Physician’s Admission 000912

Workload> 0 4001195DV = 1 if Physician’s Inpatient 000709

Workload in 2nd Quartile 4001045DV = 1 if Physician’s Inpatient −00118

Workload in 3rd Quartile 40007775DV = 1 if Physician’s Inpatient −00131

Workload in 4th Quartile 4001135Base-DRG Freq× DV = 1 if Physician’s 2.73e−05

Inpatient Workload in 2nd Quartile 4000007615Base-DRG Freq× DV = 1 if Physician’s −00000142

Inpatient Workload in 3rd Quartile 4000008775Base-DRG Freq× DV = 1 if Physician’s −00000221

Inpatient Workload in 4th Quartile 4000009425Total Patient Discharges 000202 000202∗

40001255 40001195Total Inpatients 0000123 00000792

400009555 400009415Total Patient Admissions −0000118 −0000110

40001425 40001465Physician’s Discharge Workload −00183∗∗∗

40005615Base-DRG Freq×Physician’s 00000657∗∗

Discharges 4000002795Physician’s Admission Workload 000418

40005975Physician’s Inpatient Workload 000102

40001855Base-DRG Freq×Physician’s −00000241∗∗

Inpatients 4000001055Constant −2001 −10439



(5 categorical variables)Physician Yes YesBase DRG Yes Yes


Notes. DV, dummy variable. Errors shown in parentheses are clustered onphysicians.

∗∗∗, ∗∗, ∗ denote statistical significance at 1%, 5%, and 10% confidencelevels, respectively.

DRG code7 is increased (i.e., becomes less negative)and is significant for most patients as the numberof times the trauma department sees a conditionincreases. Averaged over all patients, an increase inthe frequency with which the trauma department seesa condition from just under 1 per month (43 instancesin our data set) to 308 per month (i.e., 170 instances inour data set) halves the marginal effect of dischargeworkload (from −202% to −101%). The frequency withwhich the trauma department treats specific condi-tions does not seem to play a role in severity assign-ment when the number of patients a physician isin charge of is higher than the median, althoughthis could be because of substantial variability in theworkload generated by inpatients. Therefore, we findpartial support for our third hypothesis.

6.1. Robustness ChecksOne possible concern with respect to our results isthat patients may be less likely to be awarded thehigh-severity assignment on a busy day not becauseof undercoding, but rather because physicians chooseto discharge patients of lower severity when they arebusy, perhaps to free up capacity (Chan et al. 2011).Although one of the reasons for choosing the traumadepartment is that each case is monitored by thecase management team, and, therefore, the individ-ual physician has less flexibility as opposed to otherhospital departments, the problem remains that thephysicians to some extent can choose when to dis-charge a patient. It is plausible that when a physicianis busy, he or she might decide to reduce the thresh-old of recovery at which they discharge a patient (seeKc and Terwiesch 2009). If they do so selectively, thatis, if they choose to reduce the discharge thresholdfor patients of low severity more than they reduce thethreshold for patients of high severity, that would alsogenerate results similar to ours. However, if physi-cians do selectively discharge easier cases duringbusy times, one would expect that patients remainingin the hospital after a busy day would be, on average,of higher severity. Therefore, if the day(s) before thecurrent discharge day were busy, one would expectto see an increase in the proportion of high-severityassignments for the patients released on the current

7 To calculate the marginal effect of an extra discharge as a functionof the frequency, we follow Ai and Norton (2003). Despite the factthat the interaction coefficient is positive and significant, as Ai andNorton (2003) point out, the nonlinear nature of the models weuse implies that for some patients the impact of a change in thefrequency on the marginal effect of an extra discharge could benegative. Although this change of sign is due to a “mechanicalsaturation effect” that occurs as the probability of high-DRG codeassignment approaches either 0 or 1 (Kolasinski and Siegel 2010,p. 1), one needs to take into account that the interaction effect isdifferent for different patients.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


day. An alternative but related mechanism that couldproduce results similar to ours is the following. Whena physician is busy, he or she might choose not todischarge fully recovered high-severity patients, butstill go ahead and discharge any fully recovered low-severity patients who would generate little adminis-trative burden. However, if this were the case, thenone would expect that these high-severity fully recov-ered patients would be discharged on the very nextday or in the next few days. Therefore, if the day(s)before the current discharge day were busier than nor-mal, we would observe an increase in the proportionof high-severity assignments.

We test this possibility by explicitly includinglagged workload in our model as an explanatoryvariable. In Model VI, presented in Table 6, we

Table 6 Results of Models VI and VII


Model VI Model VIIVariables Logit coefficient Logit coefficient

St. Dev. Discharges×Physician’s 000223Discharge Workload 40007095

St. Dev. Discharges 001844001725

Physician’s Discharge Workload −00108∗∗ −00157∗

40005505 40009415Physician’s Discharge Workload −000411

Previous Day 40005545Physician’s Admission Workload −00000340 000426

40007365 40006115Physician’s Inpatient Workload −000305 −000254∗∗

40002915 40001245Physician’s Inpatient Workload 000256

Previous Day 40003535Total Patient Discharges 000175 0000219

40001935 40001865Total Inpatients −0000720 00000613

40001185 400009695Total Patient Admissions 000290∗ −0000104

40001605 40001465Constant −20222∗∗∗ −20795∗∗



(5 categorical variables)Physician Yes YesBase DRG No Yes


Note. Errors shown in parentheses are clustered on physicians.∗∗∗, ∗∗, ∗ denote statistical significance at 1%, 5%, and 10% confidence


include previous-day (lag 1) physician workload vari-ables (i.e., the number of discharges as well as thenumber of inpatients assigned to a physician onthe previous day). Because for approximately two-thirds of our sample, any given physician doesnot discharge patients on two consecutive days,this reduces the data points on which we can fitthe model to 21611. Nevertheless, the impact ofcontemporaneous discharge workload remains sig-nificant (coefficient = −00108, p-value = 0005), andthe impact of the lagged discharge workload vari-able is nonpositive and statistically indistinguishablefrom 0 (coefficient = −000412, p-value = 0046). Theimpact of inpatient workload is no longer significant(coefficient = −000305, p-value = 0029), but neither isthe impact of the previous-day inpatient workload(coefficient = 000256, p-value = 0047). The second testwe perform is to include lags from the previous sevendays. To avoid the problem of further reducing thesample size with the introduction of lags, we set thelagged workload equal to 0 when it is missing, andwe include seven dummy variables in the model (onefor each lagged variable) that take the value 1 whenwe have a missing data point and 0 otherwise. Todeal with left-censoring of the lagged variables, weadd another 7 days to the first 20 days we excludefrom the regression. The estimated coefficients of thelags for the discharge and inpatient workload vari-ables are reported in Figure 2, along with their 95%confidence intervals. The first observation is that thereis no systematic pattern in the lags. The second isthat none of the lagged discharge workloads has asignificant positive effect (even at the 90% confidencelevel), whereas the contemporaneous discharge work-load remains negative and significant. Similarly, allthe inpatient lagged workload variables are statisti-cally indistinguishable from 0. This is consistent withthe hypothesis that when physicians choose to dis-charge patients “early” because of increased work-load, it seems that they do not discriminate, at leastnot in a statistically significant way, between “nearly”fully recovered patients that were of high severityor “nearly” fully recovered patients that were of lowseverity. This provides evidence that the measuredreduction in high-severity DRG code assignments isnot due to endogeneity but to undercoding.

Although patient care at the trauma departmentis coordinated by the “practitioner of record” andhandovers are rare, another possible, even thoughunlikely, explanation that would generate results con-sistent with ours is that physicians coordinate indeciding who gets to discharge which patients toequate their workload. More specifically, physicianscould “allocate” patient discharges so that one physi-cian does fewer but more time consuming (i.e., moresevere) cases and another physician does a larger

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Figure 2 Coefficients of Lagged Discharge and Inpatient Workload Variables (and 95% Confidence Intervals)

76

5 4 3 21

7

–0.30

–0.25

–0.20

–0.15

–0.10

–0.05

0

0.05

0.10

0.15

0.20

0.25

–0.08

–0.06

–0.04

–0.02

0

0

0.04

0.02

0.06

0.08

0.10

0.12

6 5 4 3 2 1

0

Discharge workload lag Inpatients workload lag

Logi

stic

reg

ress

ion

coef

ficie

nt

Logi

stic

reg

ress

ion

coef

ficie

nt

Notes. The coefficients were estimated by including the lags in the logistic regression presented in Model I along with a matrix of dummy variables that takethe value 1 when a lagged variable is missing and 0 otherwise. The dummy variables for the missing values were not significant (at the 5% level) except forlag 5, which was significant and negative.

number of less time consuming (i.e., less severe) cases.To test whether this is the case, we measure the stan-dard deviation of the number of discharges done byphysicians on any given day and use this as a mea-sure of heterogeneity in the number of dischargedpatients. If the hypothesis that physicians coordi-nate their discharge allocations holds true, we wouldexpect the standard deviation to have an amplify-ing effect on the impact of workload on severityassignments; that is, on days when there is a greateramount of heterogeneity in the number of dischargesperformed (i.e., some physicians discharge very fewbut more severe cases and others discharge a greaternumber of less severe cases), we would expect tofind the probability of a high-severity assignmentfor those physicians with a high workload to beeven lower compared to days with less heterogene-ity. As evident from Model VII, presented in Table 6,this is not the case. The coefficient of the interactionterm between standard deviation and number of dis-charges is positive instead of negative, and is statisti-cally indistinguishable from zero (coefficient = 000223,p-value = 0076).

The third robustness test we performed is designedto address a potential selection bias problem. Clearly,it is impossible to observe the severity assignment forthe same patient under both high- and low-dischargeworkload conditions. Therefore, one potential con-cern is that patients discharged under high-workloadconditions are, for one reason or another, systemat-ically different from patients discharged under low-workload conditions. This could be problematic forour conclusions, because any difference in severityallocation might not be due to a causal link betweenseverity assignment and workload, but instead tothis selection bias. To partially address this problem,we included several control variables in all logisticregressions (see, for example, Equation (1)). However,the question remains as to whether the selection bias

is appropriately and fully addressed by the con-trol variables and by the chosen linear specification.An alternative approach is to match patients dis-charged under high-workload conditions to patientsdischarged under low-workload conditions, so thatthese individuals are almost identical in all observablecharacteristics except the workload conditions on dis-charge. Although this does not alleviate the concernthat the patients are heterogenous based on a variablenot unobservable to us, the researchers, but observ-able to physicians, it does provide some extra supportbecause it relaxes the linearity restriction imposed bythe logistic regression.

To do the matching, we define a categorical vari-able that takes the value 1 if the patient is dis-charged under high-workload conditions (i.e., thephysician has discharged three or more patients onthat day or is looking after more patients than themedian) and 0 otherwise. We use this as the treat-ment variable. Because we have a number of observ-able factors (patient age and gender, length of stay,physician, day of the week, month, and conditionfixed effects), several of which are continuous vari-ables, finding an exact match for every patient isnot advisable because of the curse of dimensionality.Instead, we rely on propensity score matching meth-ods, which provide a natural weighting scheme thatcan yield unbiased estimates of the impact of work-load on severity allocation (Dehejia and Wahba 2002).We estimate the propensity score by running a logis-tic regression of the treatment variable on the above-mentioned observable characteristics. We exclude 12observations that fall outside the common support(Caliendo and Kopeinig 2008). We use the propen-sity score to generate a sample of low-workloaddischarges that matches as closely as possible thehigh-workload discharge sample. To do so we use thenearest-neighbor (with replacement), the two-nearest-neighbors (with replacement), and a nonlinear kernel

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Table 7 Propensity Score Matching Results

Impact of high workload onseverity assignment

Sample Average effect Standard error T -stat.

Unmatched −00122 000127 −9060Unmatched with controls −000616 000191 −3022Matched nearest neighbor −000683 000228 −3000Matched nearest 2 neighbors −000603 000201 −3000Matched Epanechnikov kernel −000683 000228 −3000

Notes. Patients are divided into two treatments, those discharged underhigh-workload conditions (i.e., the physician has discharged three or morepatients on that day or is looking after more patients than the median) and therest. The column “Average effect” presents the difference in the probability apatient is assigned a high-severity code between the two treatments for fivedifferent models. The standard error and the T -statistic of the average effectare presented in the next two columns. The row “Unmatched” presents thedifference in probability of high-severity assignment without controlling forany difference between the two treatments. The row “Unmatched with con-trols” presents the difference in probability of high-severity assignment aftercontrolling for systematic differences based on patient age, gender, lengthof stay, physician, day of the week, month, and condition fixed effects usinga logistic regression. The next three rows present the difference in proba-bility of high-severity assignment using the nearest-neighbor, two-nearest-neighbors, and Epanechnikov kernel matching algorithms.

(Epanechnikov kernel) matching algorithms. Usingthe significance test on pseudo-R2 before and aftermatching (Sianesi 2004), we find that the two-nearest-neighbors approach provides a good match quality(the observable factors are not jointly significant evenat the 10% level in predicting the treatment category),whereas the nearest-neighbor or the kernel approachis not entirely successful (the observable factors werejointly significant at the 5% level in predicting thetreatment category). However, for completeness, wereport results for all three methods in Table 7.

In Table 7, the first row reports the difference inaverage severity assignment between patients dis-charged under high-workload and low-workload con-ditions using all of the observations in our data set.The second row reports the marginal effect (aver-aged over all observations) of workload treatmentafter estimating a logit model where all observablefactors (patient age and gender, length of stay, physi-cian of record, day of the week, month, and condi-tion fixed effects) are used as controls. The last threerows present the difference in average severity assign-ment between high- and low-workload patients afterrunning a propensity score matching algorithm (near-est neighbor, two nearest neighbors, and Epanech-nikov kernel). For the matched samples, the standarderrors reported are based on the approximation byLechner (2001).8 As the average effect of being dis-charged on a high-workload day in the unmatched

8 Errors based on bootstrapping were also estimated, and they wereessentially identical to the ones reported in Table 7.

sample (coefficient = −00122, standard error = 000127)is significantly lower compared to all other modelspresented in Table 7, we can conclude that patientsdischarged under high workloads are indeed system-atically different from patients discharged under lowworkloads. However, controlling by using observ-able factors with propensity score matching or witha logistic regression specification yields very simi-lar results: the probability a patient discharged underhigh-workload conditions receives the high-severityassignment is approximately 6% (standard error 2%)lower than a matched patient discharged under low-workload conditions. More importantly, this obser-vation provides additional evidence that our findingregarding the impact of workload on the probabilityof a patient being assigned a high-severity DRG codeis unlikely to be due to selection bias on observablefactors.

As a last robustness test, we turn to the findingthat the frequency with which the trauma depart-ment treats specific conditions has a moderatingimpact on workload. One alternative explantation forwhat we find is that less frequent conditions requireinherently longer and more complex discharge notes,which are more time consuming for the discharg-ing physicians than more frequent conditions. There-fore, the moderating impact of frequency is not dueto frequency fostering the development of workload-mitigating organizational routines, but rather due tofrequency being negatively correlated with the inher-ent complexity of the discharge note. To investigateif this is a plausible correlated omitted variable prob-lem, we check directly whether the frequency of thecondition is related to the complexity of the dischargenote. To do so we use the total number of diagnosisand procedure codes the professional coder assignedto each discharged patient based on the dischargenote as a proxy for complexity (i.e., higher paperworkcomplexity would generate, on average, more diag-nosis and procedure codes). These codes, along withpatient characteristics, are the inputs that determinethe DRG code assigned to a patient. For all conditions(DRGs codes) treated by the trauma department, wecalculate the average number of codes recorded forpatients discharged during nonbusy times (i.e., thephysician has discharged two or fewer patients onthat day or is looking after fewer patients than themedian). We restrict our attention to nonbusy times,because the physicians might systematically misstreatment and diagnostic codes during busy times.We find that, on average, a low-severity conditiongenerates 6 codes (95% CI, 5022–6079), whereas a high-severity condition generates an additional 604 codes(95% CI, 5001–7085). The frequency with which thedepartment treats a condition has no significant rela-tionship with the number of codes assigned irrespec-tive of the condition’s severity. This suggests that

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


conditions treated more frequently are not inherentlymore complex in terms of paperwork (i.e., they arenot associated with more diagnosis and procedurecodes) than less-frequently-treated conditions.

We conclude this section by noting that the mod-erating impact of the frequency of a condition onthe probability a patient is assigned the high-severityDRG code provides an additional piece of evidencethat supports the hypothesis that workload leads toundercoding rather than the alternative hypothesis ofthis being a side effect of endogeneity in dischargetiming or coordination or selection bias. If it was oneof the latter, then it is not clear why frequency wouldhave such a moderating effect on the probability of ahigh-severity DRG code assignment.

6.2. The Impact of Workload on HospitalReimbursement: A Counterfactual Study

Having established that discharge-day workload con-ditions reduce the probability a patient is classified ashigh severity, we now assess the impact of this effecton hospital reimbursement. To do this we would ide-ally like to know which patients were undercodedand how much revenue was lost per undercodedpatient. However, we only have a probabilistic assess-ment on who was undercoded. Furthermore, we donot always know the counterfactual payment becausethe high-severity assignment commands different pre-miums for different diagnoses and for different pay-ers. This is further complicated by the fact that 608%(standard error = 0041%) of the patients did not paymainly because they were uninsured.9 To overcomethese difficulties, we proceed in two different ways.

The first and more direct approach to estimating thefinancial impact of workload-induced undercodinginvolves estimating, for each patient, the extra rev-enue generated by high-severity assignment as wellas the probability of undercoding and using thesefigures to estimate the lost revenue. To estimate theextra revenue generated by high-severity assignment,we run a least-squares regression of the logarithmof payments (for those patients that do pay) on theseverity assignment with base DRG, insurance plan,and month fixed effects as controls. We use monthfixed effects to control for any systematic changes toreimbursements over time. The model estimated isgiven by

ln4Paymenti5 = a0 + a1Severityi+ acControlsi +Errori1 (2)

and the results are summarized in the first col-umn of Table 8. We find that, conditional on pay-ing, a high-severity patient generates, on average,

9 After controlling for patient age, neither severity nor workloadconditions predict whether a patient will pay or not.

4708% (= exp4003915 − 1) more revenue per patientthan a low-severity patient. The 95% confidence inter-val suggests that the average revenue lost for eachundercoded patient is between 3700% and 5806%.10

We next calculate the probability an individual patientis undercoded. To do this, we first use Model II toestimate the probability of high-severity assignmentgiven the workload conditions at discharge and allother patient-specific information. Then, we calculatethe probability of high-severity assignment assumingthe patient was discharged under low-workload con-ditions (i.e., the physician did no more than one dis-charge, and the number of inpatients is no greaterthan the one). The probability a patient is undercodedis the difference between the latter and the former.We also estimate the probability that each patient willpay their bill. Combining these estimated parameters(and their errors), we can estimate that the hospital islosing, on average, what could be approximately anextra 101% of revenue (with a 95% CI of 0.4%–1.9%).

For the second, and less direct, approach we run aleast-squares regression on the logarithm of paymentsagainst a dummy variable called Busy that takes thevalue 1 if a patient is discharged under high-workloadconditions (where physicians either look after morepatients than the median or perform three or moredischarges or both) with base DRG, insurance plan,and month fixed effects as controls. The model esti-mated is given by

ln4Paymenti5=d0 +d1Busyi+dcControlsi+Errori1 (3)

and the results are summarized in the secondcolumn of Table 8. We find the coefficient of theworkload dummy variable to be negative and sig-nificant (coefficient = −000748, p-value = 00005), indi-cating that, conditional on getting paid, the hospital,on average (averaging over all base DRGs and insur-ance plans), receives a 702% = 41 − exp4−0007485)lower payment (95% CI, 2.4%–12.0%) for patients dis-charged under high-workload conditions comparedto patients discharged under low-workload condi-tions. Considering that 5306% of the patients are dis-charged under high-workload conditions and takinginto account the probability of nonpayment, we canestimate the overall lost revenue to the hospital tobe 306% (95% CI, 1.2%–6.0%). It is not surprising thatthe second method produces a higher and more noisyestimate of the impact of workload on revenues thanthe first method. This is most likely because, unlike

10 A similar analysis using the official CMS (version 25) prospec-tive rates (also known as DRG weights) suggests that high-severitypatients generate 78% more revenue per patient for the hospital(95% CI, 75%–81%) than low-severity patients. This suggests thatCMS is more aggressive in adjusting payments for patient severitythan other payers.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Table 8 Linear Regression of the Logarithm of Hospital Payments

Linear regressionVariables Dependent variable: ln(Payment)

High-Severity Assignment 00391∗∗∗

40002935Busy dummy variable −000748∗∗∗

40002605Fixed EffectsCalendar Month Yes YesBase DRG Yes YesInsurance Plan Yes YesObservations 3,536 3,536R2 00643 00625

Note. Errors shown in parentheses are clustered on base DRG.∗∗∗, ∗∗, ∗ denote statistical significance at 1%, 5%, and 10% confidence


the first method, the second method does not con-trol for any heterogeneity in the patients dischargedunder high-workload conditions, which was shownto overestimate the impact of workload on the prob-ability a patient is undercoded (see, for example, thepropensity score matching analysis of §6). Note thatin the regression of Equation (3) we did not control forpatient severity. If we do include the severity dummyvariable as well as the interaction between severityand workload conditions, the workload conditions atdischarge have no further impact on hospital pay-ments, i.e., given that a discharged patient is assignedthe high-severity (or low-severity) DRG code, howbusy the discharging physician is does not make anydifference to revenues. This provides further confir-mation that discharge workload conditions affect hos-pital revenue only through severity assignment.

Although the two methods result in slightlydifferent estimates, they both suggest that the finan-cial impact of workload-induced undercoding issubstantial.

7. Discussion and ConclusionsBy examining detailed reimbursement data at a majortrauma department, we are able to show that theproportion of patients assigned a high-severity DRGcode is linked to workload. In particular, we estab-lish that the probability of a patient receiving a high-severity assignment decreases with the discharge andthe inpatient workload performed by the patient’sphysician of record. Consistent with the literature,we attribute this probability reduction to workload-induced degradation in the quality of the dischargenotes, and it would be interesting for further researchto identify the exact mechanism of how workloadaffects the quality of the discharge note. Our mainfinding suggests that the quality of the discharge note,besides having an impact on continuity of care, also

has adverse financial implications. We also find thatthis reduction is moderated by the frequency withwhich the trauma department treats specific condi-tions. This is an example of human behavior havingan impact on performance in a systematic mannerthat, to the best of our knowledge, has not been stud-ied before. More importantly, we show that the impactof this understandable human deficiency is financiallyimportant. We are able to estimate that the hospitalloses approximately 101% (95% confidence interval,0.4%–1.9%) of its annual revenue due to this effect.

To the extent that hospitals are profit maximiz-ers, our finding suggests that an increase in physi-cian utilization is only guaranteed to increase profitswhen complemented by operational changes that mit-igate the adverse impact of such an increase on theability of the system to keep reliable patient and ser-vice records. Just like in the case of (the more seri-ous) medical errors (see Insitute of Medicine 1999),we believe coding errors are not due to “bad peo-ple,” but to “system failures” (Leape and Berwick2005). System interventions such as automation of cer-tain clinical functions, complemented by training ofclinical and nonclinical staff in quality management,have been shown to reduce medical errors (Aronet al. 2011), and we believe that similar interventionscan help with workload-induced coding errors. Morespecifically, information systems already deployed inhospitals that automatically sense and record theactions of clinical agents (Ball 2003) can providethe discharging physician with relevant content thatcould make writing the discharge summaries easierand faster, and serve as props for presenting a morecomplete account of the diagnosis and the treatment.Moreover, managerial review systems that generatereports and summary statistics of deviations from pre-scribed or expected procedures (Aron et al. 2011) canbe programmed to provide feedback to clinicians andhospital management about individual coding perfor-mance, particularly when the discharging physicianis under a high workload. Complementing automa-tion with training that reinforces to clinical staff theimportance of complete discharge summaries both forensuring quality of patient follow-up care (Kripalaniet al. 2007) as well as for correct reimbursement, alongwith providing specific templates and guidelines (Raoet al. 2005) for generating high-quality discharge sum-maries, might alleviate the under-reimbursement doc-umented by our study. Such deliberate investmentin human capital is shown to be effective in creat-ing organizational routines even for low-frequencyevents (Zollo and Winter 2002). Furthermore, a morefundamental reorganization, where a fully trainednurse contributes to the discharge process, particu-larly in nonroutine cases during busy times, would be

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


another possibility. Whether and which of these inter-ventions are effective would be an interesting follow-up research question.

Our work has implications for the design of reim-bursement systems, such as the prospective system inuse by healthcare payers, or any other service systemwhere payments are contingent on system-generatedreports. The prospective system is designed to reim-burse service providers based on the average costof providing a specific service. When tending to theneeds of undercoded patients, the service providerincurs a cost that is likely to be higher than the reim-bursement received. Therefore, the prospective pay-ment system, which is designed to allow the averagehospital to break even, may not always be adequate:service providers may incur losses that have noth-ing to do with operational efficiency, but instead aredue to workload-induced undercoding. Such systemsneed to take into account that the quality of system-generated reports and, therefore, system verifiabilityare endogenous to the workload the service systemis subjected to and should specify operational proce-dures that explicitly decouple, as far as possible, sys-tem workload from information collection.

Further research needs to investigate whether othersettings, in healthcare and beyond, exhibit a similarresponse to workload. Any setting subject to stochas-tic variability with reports generated during (or at theend of) normal operations could in principle displaybehavior similar to that of the system we study. Forexample, relying on managerial reports to compilenear miss statistics (Dillon and Tinsley 2008) might beproblematic if near misses are more likely when thesystem is operating at a higher-than-average work-load and the quality of the data collection process iscompromised at higher workloads. Similarly, sophisti-cated knowledge management systems, an importantdriver of performance for knowledge-based organi-zations such as consulting or financial services firms(see Ofek and Sarvary 2001), might be less effective inhelping organizations deal with high-workload condi-tions if the quality (and quantity) of content that goesinto the system is compromised under high-workloadconditions. Our empirical findings open up oppor-tunities for further empirical and analytical work inthese areas.

AcknowledgmentsThe authors thank the traumatologists and administratorsof the hospital examined in this paper and Chris P. Lee foruseful discussions and for assistance in obtaining the dataused in this paper. They are also grateful to Gérard Cachon,Serguei Netessine, Chris Parker, Kamalini Ramdas, StefanScholtes, Stefanos Zenios, the associate editor, and threeanonymous reviewers for comments that greatly improvedthis paper.

ReferencesAi C, Norton EC (2003) Interaction terms in logit and probit

models. Econom. Lett. 80(1):123–129.Aron R, Dutta S, Janakiraman R, Pathak PA (2011) The impact of

automation of systems on medical errors: Evidence from fieldresearch. Inform. Systems Res. 22(3):429–446.

Ball MJ (2003) Hospital information systems: Perspectives on prob-lems and prospects, 1979 and 2002. Internat. J. Medical Informat-ics 69(2–3):83–89.

Birkmeyer JD, Siewers AE, Finlayson EVA, Stukel TA, Lucas FL,Batista I, Welch HG, Wennberg DE (2002) Hospital volume andsurgical mortality in the united states. New England J. Medicine346(1):1128–1137.

Boudreau J, Hopp W, McClain JO, Thomas LJ (2003) Com-missioned paper: On the interface between operations andhuman resources management. Manufacturing Service Oper.Management 5(3):179–202.

Caliendo M, Kopeinig S (2008) Some practical guidance forthe implementation of propensity score matching. J. Econom.Surveys 22(1):31–72.

Carter GM, Newhouse JP, Relles DA (1990) How much change inthe case mix index is DRG creep? J. Health Econom. 9:441–428.

Centers for Medicare and Medicaid Services (2007) Medicare pro-gram; changes to the hospital inpatient prospective paymentsystems and fiscal year 2008 rates. Federal Register 72(162):47130–48175.

Chan C, Farias V, Bambos N, Escobar G (2011) Optimizing ICUdischarge decisions with patient readmissions. Working paper,Graduate School of Business, Columbia University, New York.

Cleverley WO, Cleverley JO, Song PH (2010) Essentials of HealthCare Finance (Jones & Bartlett Publishers, Sudbury, MA).

Cocanour CS, Fischer RP, Ursic CM (1997) Are scene flights forpenetrating trauma justified? J. Trauma 43(1):83–88.

Coit M, Katz J, McMahon G (2010) The effect of workload reduc-tion on the quality of residents discharge summaries. J. GeneralInternal Medicine 26(1):28–32.

Curtis K (2007) Trauma nurse case management. Trauma.org(June 2), http://www.trauma.org/index.php/main/article/532.

Cyert R, March J (1963) A Behavioral Theory of the Firm (Prentice-Hall, Englewood Cliffs, NJ).

Dafny LS (2005) How do hospitals respond to price changes? Amer.Econom. Rev. 95(5):1525–1547.

Dehejia RH, Wahba S (2002) Propensity score-matching meth-ods for nonexperimental causal studies. Rev. Econom. Statist.84(1):151–161.

Dillon RL, Tinsley CH (2008) How near-misses influence deci-sion making under risk: A missed opportunity for learning.Management Sci. 54(8):1425–1440.

Esposito TJ, Rotondo M, Barie PS, Reilly P, Pasquale MD (2006)Making the case for a paradigm shift in trauma surgery.J. Amer. College Surgeons 202(4):655–666.

Fisher ES, Whaley FS, Krushat M, Malenka DJ, Fleming C, BaronJA, Hsia DC (1992) The accuracy of Medicare’s hospital claimsdata: Progress has been made, but problems remain. Amer.J. Public Health 82(2):243–248.

Fuloria PC, Zenios SA (2001) Outcomes-adjusted reimbursement ina health-care delivery system. Management Sci. 47(6):735–751.

Goldsmith J (1984) Death of a paradigm: The challenge ofcompetition. Health Affairs 3:5–19.

Gottschalk A, Flocke SA (2005) Time spent in face-to-face patientcare and work outside the examination room. Ann. FamilyMedicine 3(6):488–493.

Green LV, Savin S, Savva N (2010) Nursevendor problem: Personnelstaffing in the presence of endogenous absenteeism. Workingpaper, Columbia University, New York.

Hillner BE, Smith TJ, Desch CE (2000) Hospital and physicianvolume or specialization and outcomes in cancer treatment:Importance in quality of cancer care. J. Clinical Oncology18(11):2327–2340.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.


Insitute of Medicine (1999) To err is human: Building a safer healthsystem. Committee on Quality Health Care in America, Insti-tute of Medicine, Washington, DC.

Jollis JG, Peterson ED, Nelson CL, Stafford JA, DeLong ER,Muhlbaier LH, Mark DB (1997) Relationship between physi-cian and hospital coronary angioplasty volume and outcomein elderly patients. Circulation 95(11):2467–70.

Kc DS, Terwiesch C (2009) Impact of workload on service time andpatient safety: An econometric analysis of hospital operations.Management Sci. 55(9):1486–1498.

Kolasinski A, Siegel A (2010) On the economic meaning of interac-tion term coefficients in non-linear binary response regressionmodels. Working paper, University of Washington, Seattle.

Kripalani S, Jackson AT, Schnipper JL, Coleman EA (2007) Pro-moting effective transitions of care at hospital discharge:A review of key issues for hospitalists. J. Hospital Medicine2(5):314–323.

Kuntz L, Mennicken R, Scholtes S (2012) Stress on the ward: Empir-ical evidence for safety tipping points in hospitals. Workingpaper, Judge Business School, University of Cambridge, Cam-bridge, UK.

Leape LL, Berwick DM (2005) Five years after to err is human.JAMA: J. Amer. Medical Assoc. 293(19):2384.

Lechner M (2001) Identification and estimation of causal effectsof multiple treatments under the conditional independenceassumption. Lechner M, Pfeiffer F, eds. Econometric Evalua-tion of Labour Market Policies (Physica-Verlag, Heidelberg, Ger-many), 1–18.

Lin H-C, Xirasagar S, Chen C-H, Hwang Y-T (2008) Physician’scase volume of intensive care unit pneumonia admissions andin-hospital mortality. Amer. J. Respiratory Critical Care Medicine177:989–994.

Luft HS, Bunker JP, Enthoven AC (1979) Should operations beregionalized? The empirical relation between surgical volumeand mortality. New England J. Medicine 301:1364–1269.

Macias CA, Rosengart MR, Puyana J-C, Linde-Zwirble WT,Smith W, Peitzman A, Angus DC (2009) The effects of traumacenter care, admission volume, and surgical volume on paraly-sis after traumatic spinal cord injury. Ann. Surgery 249(1):10–17.

Mas A, Moretti E (2009) Peers at work. Amer. Econom. Rev.99(34):112–145.

Mayes R (2007) The origins, development, and passage of medi-care’s revolutionary prospective payment system. J. HistoryMedicine Allied Sci. 62(1):21–55.

McCormick D, Himmelstein DU, Woolhandler S, Bor DH (2004)Single-payer national health insurance: Physicians’ views.Arch. Intern. Med. 164:300–304.

Nelson R, Winter S (1982) An Evolutionary Theory of Economic Change(Belknap Press, Cambridge, MA).

Newhouse JP (1996) Reimbursing health plans and healthproviders: Efficiency in production versus selection. J. Econom.Literature 34(3):1236–1263.

Ofek E, Sarvary M (2001) Leveraging the customer base: Creat-ing competitive advantage through knowledge management.Management Sci. 47(11):1441–1456.

Oliva R, Sterman JD (2001) Cutting corners and working over-time: Quality erosion in the service industry. Management Sci.47(7):894–914.

Rao P, Andrei A, Fried A, Gonzalez D, Shine D (2005) Assessingquality and efficiency of discharge summaries. Amer. J. MedicalQuality 20(6):337–343.

Reed RL II, Davis KA, Silver GM, Esposito TJ, Tsitlik V, O’HernT, Gamelli RL (2003) Reducing trauma payment denials withcomputerized collaborative billing. J. Trauma 55:762–770.

Reinhardt UE (2006) The pricing of U.S. hospital services: Chaosbehind a veil of secrecy. Health Affairs 25(1):57–69.

Schultz KL, Juran DC, Boudreau JW (1999) The effects of low inven-tory on the development of productivity norms. ManagementSci. 45(12):1664–1678.

Schultz KL, Juran DC, Boudreau JW, McClain JO, Thomas LJ (1998)Modeling and worker motivation in JIT production systems.Management Sci. 44(12):1595–1607.

Shleifer A (1985) A theory of yardstick competition. RANDJ. Econom. 16(3):319–327.

Sianesi B (2004) An evaluation of the Swedish system of activelabor market programs in the 1990s. Rev. Econom. Statist.86(1):133–155.

Silverman E, Skinner J (2004) Medicare upcoding and hospitalownership. J. Health Econom. 23(2):369–389.

Smith A (1776) An Inquiry Into the Nature and Causes of the Wealthof Nations (Methuen & Co., London).

Tucker AL, Edmondson AC (2003) Why hospitals don’t learnfrom failures: Organizational and psychological dynamics thatinhibit system change. California Management Rev. 45(2):55–72.

Weber M (1978) Economy and Society: An Outline of Interpretive Soci-ology (University of California Press, Berkeley).

Zollo M, Winter SG (2002) Deliberate learning and the evolution ofdynamic capabilities. Organ. Sci. 13(3):339–351.

INFORMS

holds

copyrightto

this

article

and

distrib

uted

this

copy

asa

courtesy

tothe

author(s).

Add

ition

alinform

ation,

includ

ingrig

htsan

dpe

rmission

policies,

isav

ailableat

http://journa

ls.in

form

s.org/.

physician workload and hospital reimbursement: …faculty.london.edu/nsavva/pss_final.pdf ·...

Documents