Organizational Structure and Moral Hazard among Emergency Department Physicians

Transcription

1 Organizational Structure and Moral Hazard among Emergency Department Physicians David C. Chan JOB MARKET PAPER (October 1, 2012) Abstract Does organizational structure affect physician behavior? I investigate this question by studying emergency department (ED) physicians who are observed to work in two organizational systems that differ in the extent of autonomy granted to physicians: a traditional system in which physicians are assigned patients by a triage nurse manager, and a selfmanaged system in which physicians decide among themselves which patients they will see. Taking advantage of several sources of quasi-random variation, I estimate that the self-managed system induces a 10-13% increase in productivity. Essentially all of this productivity increase can be accounted for by a decline in what I call foot-dragging: because of asymmetric information between physicians and the triage nurse, physicians appear to engage in a form of moral hazard in which they prolong the length of stay of patients in order to appear busier and avoid new patients. I show that foot-dragging is sensitive to the presence of and relationship between peers, which suggests that peers observe information about each other. Finally, I show that the assignment of new patients is more efficient in the self-managed system, which suggests that peers in the self-managed system use their information to assign work. Keywords: Physician behavior, organizational structure, social incentives, moral hazard PRELIMINARY AND INCOMPLETE. I am very grateful to David Cutler, Joe Doyle, Bob Gibbons, and Jon Gruber for their guidance and support. I also thank Alberto Abadie, Josh Angrist, David Bates, Amy Finkelstein, Brigham Frandsen, Nathaniel Hendren, Erin Johnson, Bruce Landon, Danielle Li, David Molitor, Michael Powell, Stephen Ryan, Heidi Williams, and an extensive list of MIT graduate students for helpful comments and suggestions. I gratefully acknowledge support from the NBER Health and Aging Fellowship, under the National Institute of Aging Grant Number T32-AG000186; the Charles A. King Trust Postdoctoral Fellowship, the Medical Foundation; the MIT George and Obie Schultz Fund; and the Agency for Healthcare Research and Quality Ruth L. Kirschstein Individual Postdoctoral Fellowship 1-F32-HS

2 1 Introduction Improving health care delivery hinges on influencing physicians to behave more efficiently. However, past emphasis on financial incentives alone has yielded disappointing results both in reducing costs and in improving quality (Brook, 2010). Rather, there is a growing recognition that organizational structure is what sets high-performing health care institutions or more precisely, the physicians practicing within them apart from the rest (McCarthy and Blumenthal, 2006; Oliver, 2007; Institute of Medicine, 2012). More broadly, despite well-known attention on financial incentives to address moral hazard (Holmstrom, 1982; Lazear and Rosen, 1981), recent research has shown that management and organization can matter greatly in productivity (e.g., Bloom, Eifert, Mahajan, McKenzie, and Roberts, 2011; Hamilton, Nickerson, and Owan, 2003). 1 However, disentangling the effect of organizations on productivity still faces several challenges. For example, organizational differences across firms are often difficult to separate from worker selection or other firm-level exposures. Even in natural or randomized experiments, organizational differences usually involve multiple features, which makes isolating behavioral mechanisms difficult. I study a natural experiment in which emergency department (ED) physicians are observed to work in two different organizational systems that differ in only one respect. In one system, which I call traditional, two physicians who work in the same location ( pod ) are assigned work by a triage nurse manager. In the second system, which I call self-managed, the manager first assigns work to a pod that is shared by two physicians, and then these physicians decide between themselves who will care for each patient that arrives. This comparison illustrates an important dimension of organization over how much autonomy workers are given to manage themselves. Outcomes in the two systems may differ via several mechanisms. In the traditional system, under asymmetric information, workers may want to avoid more work by appearing to be busier than they are, keeping patients longer than necessary ( foot-dragging ). In the selfmanaged system, if physicians have better information on each other s true workload and use 1 Other landmark studies include Bloom and Van Reenen (2007); Ichniowski, Shaw, and Prennushi (1997); Bertrand, Duflo, and Mullainathan (2004). 1

3 it in assignment, they can reduce foot-dragging relative to the traditional system. Workers in the self-managed system may also seek to avoid work by waiting for their peer to pick patients first ( free-riding ), although this need not be significant under sufficient joint physician information, commitment, or costs of having patients remain unattended. Finally, outcomes may differ through advantageous selection in the self-managed system, either according to skill or availability, that is unrelated to moral hazard. Several features of this empirical design allow for identification both of the overall effect of self-managed teams and of foot-dragging and free-riding as specific mechanisms. First, one of the two pods changed from a traditional system to a self-managed system during the sample period, allowing me to separate time-invariant and unobservable differences between the two pods from the effect of the self-managed system. Second, I observe the same health care providers physicians, nurses, and residents working in both pods, which allows me to control for worker identities. Third, physician schedules are arranged far in advance and do not allow physicians to choose shifts with specificity. I construct a measure based on the exogenous flow of work to the ED to isolate foot-dragging, and I use exogenous variation in the assignment of peers to evaluate peer effects on foot-dragging. Fourth, the detailed nature of physician orders allows me to infer when physicians start working and isolate free-riding. I focus on the time a physicians spends on a patient, length of stay, as my primary outcome measure. This resembles measures of throughput productivity used in other studies of worker productivity (Mas and Moretti, 2009; Bandiera, Barankay, and Rasul, 2009, 2010). In the ED, throughput is especially relevant because it impacts waiting times, a key determinant to patient satisfaction and outcomes (Bernstein, Aronsky, Duseja, Epstein, Handel, Hwang, McCarthy, John McConnell, Pines, and Rathlev, 2008; Thompson, Yarnold, Williams, and Adams, 1996). Waiting times are a public good that depend on the aggregate throughput of individual physicians. Of course, given that health care productivity also involves quality, I examine secondary outputs but do not expect them to differ much at the individual level if the foot-dragging moral hazard is the primary difference between organizational systems. I find that physicians perform 10-13% faster in the self-managed system than in the traditional system. There is no difference to the time that physicians write their first order, which 2

4 suggests that free-riding is not a significant mechanism. I further examine foot-dragging by testing a prediction that foot-dragging should respond to increases in expected future work, while other mechanisms should not. Physicians increase lengths of stay in the traditional system with expected future work, while no such increase occurs in the self-managed system. The overall effect of the self-managed system can be accounted for by this mitigation of foot-dragging. Consistent with foot-dragging that only prolongs length of stay, other quality outcomes (e.g., hospital admissions) and process measures (e.g., the number of medications ordered) are not different between the two systems as expected future work increases. Second, social incentives play a crucial role in modifying foot-dragging. The presence of a peer in the same pod especially a more senior peer substantially reduces foot-dragging in the traditional system, relative to when there is another physician in the ED but not in the same pod. The location of the other physician does not affect total work but does affect the ability to monitor work. Finally, foot-dragging in the traditional system is affected by the number of patients (the census ) of both the index physician and of the peer, which could reflect strategic behavior or social incentives. Third, I directly study patient assignment in the two systems. In particular, I test a prediction reminiscent of Milgrom and Roberts (1988): If a manager is aware of foot-dragging but cannot observe it, she can be strictly better off by making assignment based less on censuses that are informative but subject to distortion, i.e., by reducing ex post assignment efficiency. Consistent with this, I find that patient assignment is less correlated with censuses in the traditional system. I also study patient assignment around the transition of the pod switching to a self-managed system and find evidence of enforcement against foot-dragging in the short term, prior to improving ex post efficiency. The remainder of the paper proceeds as follows. The next Section describes the ED institutional setting and data. Section 3 outlines a simple model of asymmetric information that highlights the link between foot-dragging and patient assignment and the related but independent pathways of organizational structure and social incentives that reduce foot-dragging. Section 4 reports the overall effect of self-managed teams. Sections 5 and 6 discusses the main evidence for foot-dragging and its mitigation by organization structure and the presence of peers. 3

5 Section 7 explores patient assignment in the two systems over time. Section 8 concludes. 2 Institutional Setting and Data I study a large, tertiary-care ED with a high frequency of patients visits greater than 60,000 visits per year (or 165 visits per day) with a total of 380,699 visits over 6 years. For each visit, I observe detailed information for each time in the process of care patient arrival to the ED, arrival to the pod floor, entry of discharge order, and discharge with discharge destination as well as all physician orders written during the visit (approximately 14 orders on average per visit). Because all actions taken by physicians must be documented and time stamped as orders, these data provide uniquely detailed process measures of physician effort and patient care. Patient care in the ED is delivered by an attending physician ( physician ), a nurse, and sometimes a resident physician ( resident ). The physician is responsible for directing patient care, while the resident, who is still training, may assist the physician with varying levels of autonomy. Nurses are responsible for executing physician orders and reporting to physicians any concerns. I observe 92 unique physicians, 364 unique nurses, and 986 unique residents in the data. 75 unique physicians, 334 unique nurses, and 882 unique residents are observed to work in both organizational systems, with 11,865 unique physician-nurse-resident trios working in both pods Outcomes I observe patient-level outcome measures of length of stay, 30-day mortality, hospital admission, bounce-backs, defined as patients returning to the ED within 14 days, and costs. These represent common metrics for ED quality and performance both at the institutional level and at the broader national policy level (Schuur and Venkatesh, 2012; Forster, Stiell, Wells, Lee, and Van Walraven, 2003; Lerman and Kobernick, 1987). Length of stay is my primary outcome measure because it directly relates to ED throughput 2 Essentially all providers who do not work in both systems either are occasional moonlighters or represent errors in recording the correct provider. For example, the number of visits corresponding to median resident is 1,525, while this number is 17 for residents who are observed to work in only one system. 4

6 productivity, consistent with other studies of worker productivity (e.g., Mas and Moretti, 2009). Given ED bed capacity constraints and a waiting room almost always with patients, length of stay is directly related to waiting times, another prominent ED-wide measure believed to be important for patient satisfaction and outcomes for conditions such as myocardial infarction and pneumonia (Thompson, Yarnold, Williams, and Adams, 1996; Bernstein, Aronsky, Duseja, Epstein, Handel, Hwang, McCarthy, John McConnell, Pines, and Rathlev, 2008). Measured from pod arrival to discharge order, length of stay is unaffected by inpatient bed availability, patient home transportation, or clinical care or patient adherence after the ED visit. The other outcomes represent patient-level quality measures that are important in health care. Thirty-day mortality occurs in about 2% of the sample. Hospital admissions for closer monitoring, diagnosis, and treatment represents a resource-intensive option for ED discharge that is believed to sometimes substitute for appropriate care in the ED and occurs in 25% of the sample. Bounce-backs, occurring in about 7% of the sample, represent the complementary quality issue of patients who were discharged home but needed to return to the ED. I observe total direct costs for each patient encounter, including any costs incurred from a resulting hospital admission. Finally, because I have data on all orders, I consider detailed process measures that capture all aspects of patient care, including nursing, medication, laboratory, and radiology orders. I do not observe the time that a physician officially signs up to see a patient, but to proxy for this, I use the time that the first physician order is written. 2.2 Systems, Pods, and Shifts Physicians are observed to work in two different organizational systems. All patients first must enter through a waiting room, or triage, where a triage nurse that decides where and when to send them. In the traditional system, patients are assigned directly to physicians by a triage nurse, even though two physician peers work in the same pod. The triage nurse serves as a manager in the sense that she allocates work at a given time according to a physician she thinks is available or able to do the work. 3 In the self-managed system, two physician peers in 3 The assignment of patients by nurses or non-medical staff is the predominant system of work assignment in hospital and ED settings. In many of these settings, however, these staff have no discretion but merely follow rules. Also, of course, these managers do not hire or fire nor set financial incentives. 5

7 the pod are jointly responsible for dividing work sent to the pod by the triage nurse. See Figure 1 for a schematic of patient workflow. The assignment of patients to nurses and residents does not differ between the two systems; in both systems, nurses are assigned patients, and residents choose patients. Basic information about patients cared for by each physician is publicly available to physicians at the triage nurse from the computer interface (see Figures 2 and 3). The most important public measure of workload that could be used in patient assignment is the number of patients already cared for by each physician (or his census ). However, censuses can be directly distorted by prolonging the time to discharge (i.e., foot-dragging ). Physicians may have superior information relative to the triage nurse about the true workload of their peers not only from differences in medical knowledge but also because they are in the same room and can directly observe peer behavior and patient status updates. In the self-managed system, physicians may also use this information to assign patients. Physicians in the ED work in 8-9 hour prescheduled shifts in two geographic locations or pods, which I call Alpha and Bravo. Alpha and Bravo pods are similar in the number of beds and staffing. The two pods are largely identical in layout and have remained so over time. The triage nurse can decide to send any patient to either pod, based on bed availability. However, one important difference between Alpha and Bravo is that Alpha pod was always open 24 hours, while Bravo pod always closed at night. As a result, patients who needed to stay longer, either because they were sicker or had conditions that might make discharge difficult (e.g., psychiatric patients), tended to be sent to Alpha pod. See Appendix Table A-1 for summary statistics comparing the pods. Alpha pod always had a self-managed system. In contrast, in March 2010, Bravo pod switched from a traditional system to a self-managed one. The regime change in Bravo pod resulted from a simple intervention in which beds that physicians previously owned were allowed to be shared so that the physicians were then allowed to choose among patients entering that pod. The reason for this switch, according to the ED administration, was to allow greater flexibility in patient assignment within pod. The switch was not considered a significant change in organizational structure by either administration or the physicians working in it, and overall 6

8 implications for efficiency were not apparent. 4 Importantly, physician schedules, staffing, and algorithms for patient assignment to beds, nurses, and residents for both Alpha and Bravo pods remained stable with respect to the regime change. The actual assignment of patients between the pods was relatively stable and continuous over time; if anything, Bravo pod received patients who were more likely to stay longer over time, as the ED became busier over time (see Appendix Figure A-1). Financial incentives for physicians are fixed; they are paid a salary with a 10% productivity bonus based on both clinical volume (i.e., number of Relative Value Units) and modified by research, teaching, and administrative metrics. 2.3 Physician Schedules and Exogenous Variation In addition to the natural experiment in the organizational system between pods and across time, this study exploits other identifying variation due to physician shift assignment. Physician schedules are determined one year in advance, and physicians are only able to request rare specific shifts off, such as holidays or vacation days. General preferences, such as whether they would like to work at night, may be voluntarily stated but not honored fully, and all physicians are expected to be open for shifts at all times of the day and days of the week. Once working on a shift, physicians cannot control the volume and types of patients arriving in the ED nor the types of patients assigned by the triage nurse to the pod. 5 Conditional on the month-year, day of the week, and hour of the day, I find that physicians are exposed to identical ED conditions and patients types arriving to their pod. Tables 1 and 2 show that faster physicians are exposed on average to patient with similar characteristics and numbers of patients arriving to the ED as slower physicians. Appendix A shows further support for exogenous physician assignment: Physicians with different preferences to choose certain patients (defined as the likelihood to choose that patient when in the self-managed system) are 4 In fact, in May 2011, the ED attempted a redesign that moved both pods towards the traditional system, only to later discover that it significantly reduced efficiency. They reversed this organizational change in January Physicians may rarely (e.g., <1-2% of operating times) put the ED on divert for up to an hour when the flow of patients is unusually high and the entire ED lacks capacity to see more patients. Even when this happens, this only affects some ambulances (which as a whole constitute 15% of visits) carrying serious emergencies, as opposed to the majority of patients, some of whom walk in. ED flow is largely unaffected. 7

9 exposed to ex ante identical patient types and ED volume conditions; physicians of different productivities are equally likely to work with high- and low-productivity peers; and physician identities do not have any joint significance in predicting available patient types, ED conditions, or peers types. The observed variation is not only exogenous but also rich for several reasons: First, overall ED volume and the types of patients are notoriously wide-ranging, even conditioning on the time of the day. 6 Second, physicians work very few shifts per week, usually 1-2 with the maximum being 4, and are expected to work in all types of shifts. As a result, I observe all physicians working in both locations, during all time categories, and with essentially all physician peers. Third, there is substantial variation in the tenure of physicians. While some physicians are observed to remain on staff for the entire 6-year period, other physicians are newly hired or leave the hospital during the observation period. I observe employment details, such as the place and date of medical school and residency. I use these data to construct descriptors of peer relationships, including hierarchy in tenure, the number of shifts previously shared, demographic similarity, and shared training. 3 Theoretical Framework In this Section, I outline a simple model of asymmetric information between physician workers and the triage nurse manager in order to consider organizational structure and moral hazard. Physicians foot-drag, prolonging patient lengths of stay, in order to distort signals of true workload and avoid new work that the manager in the traditional system would like to assign by physician workload. At the same time, a manager who takes this into account can be better off by committing to an ex post inefficiency policy of not always sending new patients to physicians she knows are less busy. I then show how self-management reduces foot-dragging. In the process, I will separately consider social incentives to be clear about pathways. I assume that physicians observe better information relative to the triage nurse about each other s true workload, which is necessary for 6 At any given hour, the number of patients arriving may range from close to none to the mid-twenties. Patients may require a simple prescription or pregnancy test, or they may have a gunshot wound. 8

10 both pathways. Social incentives reduce foot-dragging because physicians do not want to be seen engaging in it. Self-management reduces foot-dragging because physicians can use information on each other s true workload to assign new work. The two pathways are independent of each other. Finally, I discuss that free-riding can manifest in the self-managed system from the same moral hazard to avoid work but can be limited with sufficient joint physician information or commitment. 3.1 Model Setup Consider the following simple game: Two physicians j {1, 2} work in a single pod at the same time. They both have one patient each, which can either be difficult or easy. In addition to the time that they take on their current patients, physicians also care about new work a third patient that might be assigned to them. Physician utility is given by u P j = (t j θ j ) 2 K P (θ j ) I {j (3) = j}, (1) where t j is the time that physician j keeps his patient, θ j { θ, θ } is the workload entailed by his patient and unobservable by the triage nurse, and K P (θ j ) > 0 is the cost of getting a potential third patient if the physician started with workload θ j, multiplied by an indicator function for whether j is assigned the third patient, j(3). Because each physician starts with exactly one patient and otherwise has identical preferences, I will also refer to θ j as physician j s type. Type θ occurs with probability p, and type θ occur with probability 1 p. Physician types are never observed by the triage nurse, but with probability ψ, peers can observe each other s type. In contrast, the census of each physician, c j, is public information at any time. The action that each physician takes is t j. Absent any strategic behavior, each physician would like to discharge his patient at t j = θ j, which I also assume is the socially optimal time for treatment that generically captures all concerns of care (e.g., patient health and satisfaction, malpractice concerns, physician effort and boredom). 7 Because patient assignment depends on the organizational system, I discuss the triage nurse and 7 I thus abstract away from any principal-agent problem between physician and patient or between physician and ED in this term, other than the cost incurred by the new patient, discussed below. 9

11 assignment further in the following subsections. Physicians incur a cost if they treat a new patient. I parameterize this cost as K P, θ = θ K P (θ) = K P, θ = θ, where K P > K P > 0. This reflects that neither physician would like to get the new patient, but that it is more costly for the busy physician, 8 either because he must exert more additional effort or because he will have worse outcomes for this new patient. The timing of the game is as follows: At time t = 0 both physicians receive one patient each, and they discover the workload entailed by that patient, θ j { θ, θ } with respective probabilities p and 1 p. Physicians commit to how long they will keep their patients (t j ). With probability ψ > 0, physicians observe each other s θ j at this time. It is common knowledge that exactly one patient will arrive with uniform probability distributed across the time interval t [ θ, θ ]. 9 When the patient arrives, this new patient is assigned to a physician by the triage nurse (in the traditional system) or the physicians themselves (in the self-managed system). Finally, physicians complete their work on the 1 or 2 patients under their care and end their shifts. This model is highlights the tension between using signals (censuses c j ) of private information (types θ j ) for patient assignment and the fact that these signals can be distorted (through t j ). Physicians have moral hazard to avoid new work regardless of their type (through K P (θ) > 0), but otherwise I assume that physicians have no incentive to keep patients longer than socially optimal. Although the triage nurse never observes the physicians types, physicians observe each other s type with some positive probability ψ. This superior information allows greater efficiency through two separate pathways social incentives and self-management. 8 To be precise, the busy physician is a physician who was given the more difficult patient (i.e., the one with the higher type). Because I only want t j to be a costly signal of θ j, by this use of the word I do not allow that physicians are made busier by keeping their patients longer. 9 The assumption of a single new patient simplifies the model because I do not need to model discharge decisions for new patients in the face of potentially more new patients. The additional assumption that physicians must commit to a discharge time for their initial patients prior to t = θ is made for convenience because of the certainty of one patient arriving implies an increasing hazard over time. 10

12 3.2 Traditional System In the traditional system, the triage nurse assigns the new patient to a physician. In my baseline model, I assume that physicians cannot report their types or anything else to the triage nurse, but that the triage nurse can credibly commit to an assignment policy. 10 I believe that this scenario is most realistic. Physicians will find it difficult to describe the workload entailed by their patients in practice, but the assignment policy based on censuses is easily observable, relatively simple, and can be enforced in a repeated game. 11 The nurse s utility is u N = D (t j θ j ) 2 ( ) K N θj(3), (2) j {1,2} which is similar but potentially different from that of the physicians. D is an indicator that allows the triage nurse to care about the treatment times of the first two patients as outcomes (if D = 1). Remember that the socially optimal discharge times for patients is t j = θ j, which is universally agreed upon. The second term, K N (θ), is the cost of assigning the new patient to a physician of type θ, specified as 0, θ = θ K N (θ) = K N, θ = θ, where K N > 0. As before, this represents that there is some cost (now to the nurse) in assigning the new patient to a physician with greater workload, again either in terms of time or lower quality. I do not restrict the the value of K N relative to K P K P. 12 The nurse s action is defined by an assignment policy function π (c 1, c 2 ), where recall that 10 I also assume that there is no credible way for physicians to report each other s workload to the triage nurse. Moore and Repullo (1988) have formalized such a subgame perfect mechanism. However, with either limited financial incentives or social incentives such as reciprocity, such a mechanism may not be implementable. In addition, as I will discuss below, I believe physicians have limited scope for even reporting their own workloads to the triage nurse. 11 In Appendix B, I consider two alternative scenarios the pure signaling game in which physicians can neither report their types nor can the nurse commit to an assignment policy, and the full mechanism design game in which physicians can both report types and the nurse can commit to a policy. Results are similar, with reporting and commitment both improving efficiency. 12 Note also that if D = 0, then it does not matter what value K N takes, as long as it is some positive number. 11

13 censuses c j {0, 1}. To simplify the analysis, I impose that π (0, 0) = π (1, 1) = 1 2 and π π (0, 1) = 1 π (1, 0). That is, when both physicians have equal censuses, the nurse should have no preference to send the new patient to one physician or the other, since she has no other information about who is less busy at that time. Also, probabilities must sum to 1. Note that π = 1 represents what I mean by ex post efficiency, since the triage nurse can infer that if c j = 0 and c j = 1, then j certainly had the lower workload. I use a Perfect Bayesian Equilibrium as the equilibrium concept, comprised of actions t j by the physicians and π by the triage nurse that are best responses of each other under consistent beliefs. To analyze the model, first note that the triage nurse will never want to send the new patient with greater probability to a physician with c j > c j. So high-type physicians will never want to mimic low-type physicians, but low-type physicians have some reason to mimic high-type ones. For a given π π (0, 1) previously chosen by the nurse, the low-type physician s [ optimization problem of max tj E u P ( j tj ; π, θ )], where the utility is given in (1), yields the firstorder condition t = θ + K ( P 2 ( θ θ ) π 1 ). (3) 2 There is a first-order gain in temporary mimicry relative to a second-order loss, as long as π > 1 2. In other words, as long as the nurse is more likely to send patients to physicians she believes are less-busy, low-type physicians will foot-drag in order to mimic high-type physicians. 13 The triage nurse will then commit to π such that her expected utility is maximized. For simplicity, I will only present the case with D = 0, in which lengths of stay for the first two patients are only signals and not direct arguments in the nurse s utility function; I leave the more general case for Appendix B. Substituting (3) into her expected utility and solving the first-order condition yields the optimal assignment rule under commitment π = ( θ θ ) 2 K P. (4) 13 A nice feature of this simple two-type model is that the first-order condition does not depend on what the peer s type or strategy is, because there is only one patient each and thus one policy parameter π. Keeping this initial patient longer by dt decreases the likelihood of getting the new patient by (π 1 /2) dt/ ( θ θ ) regardless of the peer s census. For this reason, parameters like p do not matter. This is shown in detail in Appendix B. 12

14 That is, the nurse s choice of π only depends on the low-type physician s cost of getting the new patient, because she balances correct assignment against distortion caused by the low-type physician s foot-dragging. As the pressure for foot-dragging by low-type physicians increases, either as as K P increases or as θ θ decreases, the ex post efficiency of assignment (π ) decreases. 14 The important general point is that the triage nurse may commit to an assignment policy function π < 1. Even if she only cares about the assignment of the third patient, an ex post inefficient assignment policy may improve her expected utility, which is similar to Milgrom and Roberts (1988) finding that managers can be better off if they commit to not listen to subordinates who could undertake costly influence activities. This commitment also decreases foot-dragging, shown in Appendix B. This simple model assumes a single patient will arrive in the interval t [ θ, θ ], which is convenient for capturing the temptation for moral hazard by the low-type physician. However, there are of course in practice many new patients, and I identify foot-dragging as the response of length of stay to the flow of expected future work, defined in terms of numbers of patients that arrive to the ED triage. To capture this intuition, I can extend the model by changing the interval over which the single patient is expected to arrive, which results in replacing the interval θ θ in the denominator of Equation (3) with some t θ θ, as long as t remains an interior solution. I show details in Appendix B, but the intuition is straightforward. With an infinite flow of patients to the ED (as t 0), physicians should expect to get a new patient the minute they discharge one. With no expected future patients (as t ), there is no incentive to foot-drag. 3.3 Physician Self-monitoring and Social Incentives As asymmetric information between the triage nurse and physicians allows for moral hazard, better information between physician peers reduces this moral hazard. Before I discuss the effect of self-management, I will first extend the model of the traditional system to include social 14 In Appendix B, I show that the optimal assignment policy π is even lower when she also cares about lengths of stay for the first two patients as outcomes (i.e., when D = 1). In the simple case here with D = 0, the policy function does not depend on the relative likelihood of having a low-type, p, because her tradeoff is only relevant conditional on there being a low-type physician. It also does not matter what K N is, because this cost is the only thing she cares about. 13

15 incentives. A small but growing empirical literature has shown evidence of social incentives that can reverse temptations for moral hazard, specifically when peers are observed (Bandiera, Barankay, and Rasul, 2005, 2009; Mas and Moretti, 2009; Jackson and Schneider, 2010). With probability ψ > 0 physicians can observe each other s true workload, and following Kandel and Lazear (1992), I allow for social costs conditional on this observation as a function S (t j θ j ) increasing for t j > θ j. Social costs are incurred conditional on being observed footdragging (which occurs with probability ψ). The expected utility for the low-type physician is then E [ u P j (t j ; π, θ) ] = (t j θ) 2 K P Pr { j (3) = j tj, π, p } ψs (t j θ). It is straightforward to show that low-type physicians will foot-drag less because they anticipate social costs to foot-dragging. Since they foot-drag less, assignment will be more efficient (ex ante and ex post). Empirically, I can test for the joint effect of ψs ( ) on foot-dragging, which requires both observation and social costs conditional on being observed. One way to identify ψs ( ) is to measure foot-dragging when another physician is in the same pod and can observe foot-dragging (I call this physician a peer ) versus whether the other physician is in another pod. In the traditional system, there is no reason why foot-dragging should differ between these two scenarios if there are no social incentives. Another method would be to measure differences in ψs ( ) depending on the type of peer present in the pod. 3.4 Self-managed System Finally, foot-dragging can be modified by organizational structure. This is a natural extension of an influential and striking experimental literature that shows that cooperation and punishment depends not only on social preferences but on the rules of the game (Ostrom, Walker, and Gardner, 1992; Fehr and Gachter, 2000). Social incentives alone are often insufficient to control moral hazard, and changes in organizational structure, under the same social incentives and information structure, may lead to greater efficiency. For the self-managed system, I assume the same physician utilities (including social incen- 14

16 tives) and information structure (i.e., that they observe each other s θ j with probability ψ > 0 and nothing otherwise) as before. The only difference is that physicians themselves, not a triage nurse, are responsible for deciding who gets the new patient. Physician actions include both t j and an action that determines the assignment of the new patient. In Appendix B, I detail two microfoundations of this assignment in the self-managed system, both continuing the baseline assumption that physicians cannot report their types. First, physicians may act completely non-cooperatively and only choose patients when it is in their best interest too do so. In this case, I assume that there is a cost imposed on both physicians by a patient remaining unattended on the pod and that this cost is higher for the high-type physician. If physicians observe each other s type, with probability ψ and if types are different, then the low-type physician will choose the new patient immediately upon arrival by subgame perfect reasoning similar to Rubinstein (1982). If they do not observe each other s type, however, then they engage in free-riding through a war of attrition (e.g., Bliss and Nalebuff, 1984), inefficiently leaving patients unattended with some probability. There is no foot-dragging in this case. Second, physicians may commit to some policy function that takes the form of π (c 1, c 2 ; o 1, o 2 ), where o j is the physician j s type observed by his peer, equal to θ j with probability ψ and Ø with probability 1 ψ. This case is very similar to the traditional system: I find the Perfect Bayesian Equilibrium in which physicians first commit to π (c 1, c 2 ; o 1, o 2 ) and then choose t j. Physicians naturally should commit to π ( c 1, c 2 ; θ, θ ) = 1, for all c 1 and c 2. So with probability ψ, foot-dragging will have no effect on assignment; in particular, if there is a low-type and a high-type physician, the low-type physician will receive the new patient with probability 1. With probability 1 ψ, physicians do not observe each other s type, but they follow their policy function, summarized by a single parameter π π (0, 1; Ø, Ø). Prior to knowing their types, physicians solve for the optimal π that maximizes their ex ante expected utility with the knowledge that it is more costly for a busy physician to accept the new patient but that they will attempt to foot-drag regardless after they discover they have lower workloads. Commitment to a policy function again implies some ex post assignment inefficiency, but less than in the 15

17 traditional system, again primarily because foot-dragging is pointless with probability ψ > I consider both cases because the truth likely lies in between two extremes. Physicians are not forced to see patients in the self-managed system, but they very likely have a strong cultural norm that prevents patients from waiting on the pod unattended. There is little free-riding with sufficient physician information (high ψ), commitment, or costs of leaving patients unattended. 4 Overall Effect of the Self-managed System In this Section, I aim to estimate the overall effect of the self-managed system on a given team of providers and on a given patient. That is, if the same patient and providers were assigned to each other in a different organizational system, what would their outcomes be? I can control for pod-specific time-invariant unobservable differences by the fact that I observe one of the two pods (Bravo) switching from a traditional system to a self-managed one. I also can directly control for providers because I observe essentially all providers physicians, residents, and nurses working in both pods over time. Of course, I do not observe the same patient visit in both pods, but I do observe a rich set of characteristics to control on. As my baseline analysis, I thus estimate the following equation: Y ijkt = αself it + βx it + γp od it + ηi t + ν jk + ε ijkt, (5) where outcome Y ijkt is indexed for patient i, physician j, resident-nurse k, and time t. The variable of interest in Equation (5) is Self it, which is an indicator for whether patient i was sent to a pod with a self-managed system at time t. It also controls for patient characteristics X it, pod identities P od it, a vector of time dummies I t (for month-year, day of the week, and hour of the day), and physician-resident-nurse trio identities ν jk. My approach of controlling for observable patient characteristics and provider identities is similar to a matching approach that has been described as semiparametric differences-in-differences (Abadie, 2005; Heckman, 15 In Appendix B, I also show that another reason for improved ex post assignment inefficiency is that physicians likely care more about inefficient assignment, since the cost of inefficient assignment is scaled relative to treating their own patients, while the triage nurse scales this relative to treating all patients. 16

18 Ichimura, and Todd, 1997). 16 The coefficient of interest is α, the effect of the self-managed system. Because I cannot control for patient unobservables, for this estimate to be unbiased, I must assume that conditional on patient observables, the patients sent to Alpha versus Bravo did not change over time. In Table 3, I estimate several versions of (5), where I include progressively more controls for patient characteristics. The estimate for effect of self-managed teams on log length of stay remains stable (and slightly increases) from -10% to -13% upon adding a progressively rich set of controls. The fact that the estimate increases is consistent with the fact that sicker patients were sent to Bravo pod over time (see Appendix Figure A-1) and is suggestive that other characteristics that I cannot observe may follow the same pattern. This overall effect represents a significant difference in length of stay due to a simple organizational change in which physicians assign work among themselves, while the physicians themselves and financial incentives were held fixed. As a comparison, this effect is roughly equivalent to 1 standard deviation in physician productivity fixed effects: Physicians who are 1 standard deviation faster have lengths of stay of about 11% shorter. While I find a significant effect of self-managed teams on length of stay, I find no effect for other outcomes of 30-day mortality, admissions, 14-day bounce-backs, and total direct costs, reported in Table Physicians are not decreasing length of stay by reducing the quality of care, at least on the measures that I observe. I begin discussing direct evidence of foot-dragging in the next Section. However, I note here that alternative mechanisms of free-riding and advantageous selection could affect the quality of care, because they mean either that specific patients are being made to wait for care or are seen by physicians who are better (or more available) to see them. In contrast, under pure foot-dragging, only the discharge of patients is delayed in order to prevent more work, then patients who enter the self-managed vs. traditional system should not necessarily have different quality outcomes. Any negative impact in terms of waiting times would be shared by all patients in the ED. That is, the lack of effect on quality measures 16 As discussed in Section 2, there were secular trends between the two pods. Specifically, more and sicker patients were sent to Bravo pod over time, and new nurses generally spent more time in Bravo pod to fill the need of higher volume. I show unconditional results and discuss this in Appendix C. 17 Given the low percentage of mortality, unfortunately the estimate for the effect on mortality is relatively imprecise. 17

19 between the two systems is more consistent with foot-dragging than the other mechanisms. Table 4 also reports the effect on the time to the first physician order. The effect of the selfmanaged system on this time is insignificant from 0 and slightly negative. I use this outcome measure as a proxy for the time that the physician chooses to see the patient. Significant freeriding would imply a significant positive coefficient on the self-managed system with respect to this proxy. The effect on length of stay due to Bravo s regime change to a self-managed system can also be seen graphically. Figure 4 shows month-year-pod fixed effects over time for the two pods estimated by this equation: M Y 1 Y ijkt = α myp I m I y Self p it (1 Self it) 1 p + βx it + ζĩt + ν jk + ε ijkt, (6) m=1 y=1 p=0 where the coefficients of interest α myp are plotted for each month, year, and pod interaction; I m and I y are indicator functions for the month and year, respectively; and Ĩt is a revised vector of time dummies that only includes day of the week and hour of the day. Figure 4 shows a persistent discontinuity at the regime change that is consistent with my baseline estimates in Table 3 that self-managed teams in Bravo decreased patient lengths of stay. In addition, Figure 4 both shows that the (conditional) parallel trends assumption is valid and hints at inference based on the long time series of observations. An issue that arises in difference-in-differences estimation is with constructing appropriate standard errors for inference (Bertrand, Duflo, and Mullainathan, 2004). 18 The baseline specification clusters standard errors by physician, which is equivalent to thinking of an experiment in which sampling occurs at the level of physicians assigned to shifts that imply pods and organizational systems, before and after the regime change in Bravo. This is the thought experiment I wish to consider, as physicians are in both pods before and after the regime change, and as physicians shifts are randomly assigned conditional on rough time categories. However, there is the additional statistical issue of unobserved and potentially correlated podlevel shocks over time. Therefore, I consider two alternative thought experiments for inference, 18 This issue is largely only relevant for estimating the overall effect. Specific mechanisms use additional variation. In particular, foot-dragging relies on the effect of exogenous expected future work. 18

20 both of which exploit the long time series and can be understood by Figure 4. First, I address sampling variation at the pod level across time but with a more parametric form, assuming a month-year-pod shock that is correlated by an AR1 process across months within pod. Second, in the spirit of systematic placebo tests (Abadie, Diamond, and Hainmueller, 2011; Abadie, 2010) and randomization inference (Rosenbaum, 2002), I consider the thought experiment that, under the sharp null of no effect of self-managed teams, there should be no significant difference between my obtained estimates and those I would obtain if I consider placebo regime changes over pod and month. Rather than sampling variation, this approach considers randomization at the level of the treatment with the placebo regime changes. As detailed in Appendix C, both approaches yield a high degree of statistical significance, with p-values less than 0.01 and not appreciably larger than in my baseline approach. 5 Main Evidence of Foot-dragging In order to identify foot-dragging separately from other mechanisms that could contribute to the overall effect, it is important to recognize that the expected gains to physicians by footdragging depends on expectations of future work. If no further patients arrive to the ED, then foot-dragging is not needed to prevent new work. But if there is an endless supply of patients waiting to be seen, then discharging a patient directly leads to having to see another one, and the incentive to foot-drag is extremely strong. Expectations of future work therefore depend on the number of patients arriving to the ED. Perhaps more concretely, physicians can click on their computer interface to view the patients waiting to be seen. I identify and quantify foot-dragging by increases in length of stay as expected future work increases. Conceptually, other than through the moral hazard of foot-dragging, there is no other reason why lengths of stay should increase with expected future work, holding actual work constant. 19 I therefore interpret any increase in lengths of stay with expected future work as evidence of foot-dragging. This is independent of the organizational system in which I observe foot-dragging, 19 When both pods are open, patients that arrive to the ED while the index patient has just arrived on the pod may be sent to either pod. This allows me to separate expectations of future work (the number of patients arriving to the ED) from actual future work (the number of patients who will arrive to the pod). I discuss this further below. 19

21 but I am also interested in comparing foot-dragging between different organizational systems. Also, as outlined in Section 3 above, it is important to note that the costs of foot-dragging derive from both direct moral hazard and inefficient assignment that results from this moral hazard. Both of these are directly related to expectations of future work, and both are jointly determined in equilibrium. I do not separately identify these effects here, although I separately address assignment in Section 7. For my baseline estimation of foot-dragging, I estimate equations of the following form for log length of stay Y ijkt for patient i, physician j, resident and nurse k, and time t: Y ijkt = α 2 EDWork t + α 2 Self it EDWork t + α 3 Self it + βx it + γp od it + ηi t + ν jk + ε ijkt, (7) where as before, Self it represents whether patient i was seen on a self-managed team, X it controls for patient characteristics, P od it controls for time-invariant pod unobservables, and other controls include time fixed effects ηi t and provider-trio fixed effects ν jk. EDWork t represents expected future work based on ED patient volume, defined in two ways. First, I consider the number of patients arriving to triage in the hour prior to the patient i s arrival to the pod. The arrival of these patients is not controlled by physicians. They are also seen by physicians via the computer interface, but their ultimate destination is unknown. 20 Second, I also consider the number of patients (the census) in the waiting room at the time of patient i s arrival to the pod; although physicians presumably can affect the number of patients waiting, this is a more salient measure of expected future work. The coefficients of interest in (7) are α 1, α 2, and α 3. A positive α 1 indicates that physicians increase length of stay as expected future work increases (i.e., they foot-drag) in the traditional system, while a negative α 2 indicates that the self-managed system mitigates foot-dragging. Similar to the coefficient of interest in Equation (5), α 3 represents the effect of the self-managed system after controlling for foot-dragging, or the effect related to expected future work. 20 Given average waiting times and the average length of stay, most patients will not even be assigned until after patient i is discharged. I also try alternative time windows for this measure and find that my results are robust to them. 20

22 Table 5 reports estimates for (7) for both measures of expected future work. With an additional patient arriving hourly to triage or waiting in triage, length of stay increases by 0.6 percentage points in the traditional system. The estimate of foot-dragging in the traditional system is equivalent to a length-of-stay elasticity of 0.10 with respect to expected future work. 21 The coefficient on the interaction between expected future work and the self-managed system suggests that this effect is essentially reversed in the self-managed system. an additional patient in each measure of expected future work does not affect lengths of stay in the self-managed system. Finally, the coefficient on the self-managed system is insignificant and even slightly positive after accounting for the effect of expected future work. In addition to the baseline specification in (7), I also include a number of controls for physician workload with pod-level volume at the time of patient arrival. Results are robust to including these controls. These results suggest large and robust evidence of foot-dragging in traditional teams and equally large mitigation of it in self-managed teams. Estimates are robust under both measures of expected future work: ED patient volume and waiting room census. Also, controlling for actual pod-level work, either current or future does not change results, suggesting that increased patient length of stay is due to expectations of future work and insensitive to actual current or future workload. Again, this distinguishes foot-dragging from any other mechanism. Without moral hazard that seeks to prevent future work, physicians should not increase length of stay as expected future work increases. In contrast, other mechanisms should only depend on actual work that reaches the pod. 22 Finally, foot-dragging as a function of expected future work is quantitatively large enough to explain the difference in performance between self-managed and traditional systems. The remaining effect in α 3 ranges from -1% to 3%, suggesting in some specifications that physicians may actually perform 3% slower in the self-managed system at the lowest levels of expected future work, although this is not statistically significant. 21 I estimate this by using log ED volume as EDWork t. This is not my preferred specification because ED volume is roughly normally distributed. However, results are qualitatively the same in this specification, with large foot-dragging in the traditional system and essentially none in the self-managed system. 22 Actual work is highly pod-specific. Two potential exceptions of spillovers between pods are waiting for a radiology test or a hospital bed. However, I find no difference in foot-dragging between patients likely and unlikely to receive radiology tests. I also show in Table 6 that radiology testing is not affected by expected future work. The time waiting for a hospital bed is excluded from my measure of length of stay, since I record the time of the discharge order. In contrast, Table 6 also shows that outcomes like admission are affected by the volume of work. 21

23 Figures 5 and 6 plot log length of stay coefficients of each decile of ED patient volume interacted with organizational system, estimated by Y ijkt = 10 d=2 10 α d 01 (Self it = 0, D (EDWork t ) = d) + α11 d (Self it = 1, D (EDWork t ) = d) + d=1 βx it + γp od it + δp odw ork it + ηi t + ν jk + ε ijkt, (8) where D (EDWork t ) denotes deciles of expected future work, measured as the patient hourly arrival rate and waiting room census, respectively. The coefficients { α0 2,..., α10 0 ; α1 1,..., } α10 1 can be interpreted as the relative expected length of stay for patients in different organizational systems and under different ED volume conditions, where the expected length of stay of patients in traditional teams under the first decile of patient volume is normalized to 0. As shown in Figures 5 and 6, the traditional team has progressively longer expected lengths of stay as ED volume increases. The increase in log length of stay is roughly linear by deciles, which suggests that physicians engage in foot-dragging continuously with expected future work. In contrast, the self-managed team has roughly the same expected length of stay as the traditional team at low patient volumes, and its expected length of stay does not change with patient volume. My measures of ED patient volume are likely to be noisy representations of physicians expectations of future work (e.g., they may expect more patients even when there is no one in the waiting room). Therefore this estimate is a lower bound on true foot-dragging: It is biased downwards to the extent that I do not capture true expectations of future work. More generally, I consider any increase in length of stay with respect to expected future work as foot-dragging. But it may reasonable to think that physicians in the absence of moral hazard should actually work faster, for example if they care about not having patients wait too long in the waiting room. In this sense, my estimate is therefore also a conservative benchmark, which considers foot-dragging relative to physicians not paying attention to future work, for example. Note that since length of stay does not increase with expected future work in the self-managed system, foot-dragging relative to 0 and foot-dragging relative to self-managed teams are roughly the 22

24 same in magnitude. 23 I also estimate Equation 7 for other outcomes and process measures: 30-day mortality, admissions, 14-day bounce-backs, total direct costs, and a host of detailed process measures including medication, laboratory, and radiology orders. I show these outcomes and a subset of the process measures in Table 6. I find no differential effect of expected future work between the two systems for any of these outcomes or process measures. Some outcomes do reflect a slight effect of ED volume on hospital capacity for both systems, such as hospital admissions and total costs that can include those incurred in admissions. Estimates for process measures are tightly estimated and show that the care provided while foot-dragging is not substantively different. In addition to the result previously shown in Table 4, this confirms that foot-dragging solely acts through the channel of delaying the time of patient discharge, rather than through increasing the quality or content of medical care Peer Effects on Foot-dragging Foot-dragging could be reduced if physicians have better information than the triage nurse and if they also care about being seen foot-dragging. In this Section, I test for the existence of social incentives through peer effects on foot-dragging with three different types of analyses. 25 First, I examine physician foot-dragging when there is a peer in the pod compared to when there is no peer present but someone else working in the ED. In the traditional system, there is no reason for foot-dragging to be affected by the location of other physicians, except through social incentives. I also use two additional settings physicians working without a peer in a pod that is officially self-managed (but with another physician in the ED), and physicians working alone 23 This interpretation is also supported by the fact that length of stay does not respond to expected future work when there is only one physician in the ED, which I discuss further in Section 6.1. With only one physician in the ED, there is no other physician present to foot-drag upon. However, the physician may still foot-drag on future physicians, and I technically cannot separate expected future work from actual future work, given that there is only one pod open. 24 If anything, some of the process measures show a slight decrease in number of orders for patients, although this is attenuated when accounting for actual pod-level work. 25 Note that peer effects on foot-dragging is a more narrowly defined peer effect, which is the interaction of peers and the effect of expected future work. That is, I specifically deal with the effect of observation on moral hazard, as opposed to general peer effects on productivity, such as productivity spillovers. I discuss more general peer effects in Appendix D, where I show results similar to Mas and Moretti (2009) in which productive peers increase the productivity of physicians. 23

25 in the entire ED as falsification tests for my identification of foot-dragging. Second, I explore whether the type of peer present matters for foot-dragging in traditional and self-managed teams and find that senior peers reduce foot-dragging among their juniors, suggesting greater social incentives from senior to junior peers. Third, I test whether foot-dragging differs depending on own- and peer-physician censuses. This constitutes a joint test of strategic behavior and social incentives. 6.1 Presence of a Peer During certain times in the work schedule, physicians who will be or have been part of a team suddenly find themselves without a peer. This is because shifts in each pod are staggered and because staffing adjusts upward in the morning when patient volume incresases and downward in the night when it decreases. I use this fact to identify how physicians respond to the presence or absence of a peer in the same pod, during times in which there is still someone else working in the ED, so that foot-dragging still entails a negative externality. If foot-dragging is purely self-interested behavior unaffected by social incentives, then physicians working in the traditional system should foot-drag equally conditional on ED volume per physician. It should not matter whether the other physician in the ED is in the same pod or in the other pod. In addition, two similar analyses can serve as falsification tests for the identification of foot-dragging as increases in length of stay with expected future work. First, when there is no peer in the self-managed pod, physicians will suddenly find themselves in an effective traditional system: Every patient that arrives is in fact assigned to them by the triage nurse. They should then exhibit foot-dragging behavior as if working in the traditional system (and more than when working with a peer in the traditional system, if there are social incentives). Second, when there is only one physician in the entire ED, there is essentially no assignment problem. That physician owns all patients who arrive to the ED. With no one else to foot-drag against, physicians have no incentive for foot-dragging Physicians can still foot-drag against future physicians, but this theoretically is no different at any other time. Also, I cannot control for actual future work in this scenario, because all work that comes to the ED eventually goes to the same pod and physician. However, this would only bias measured foot-dragging upwards if the omitted volume of actual work is positively correlated with expected future work and increases length of stay. 24

26 In Table 7, I present results for regressions of the form Y ijkt = α 1 EDW ork t + α 2 NoP eer jt EDWork t + α 3 NoP eer jt + βx it + γp od it + ηi t + ν jk + ε ijkt, (9) for traditional-team and self-managed-team samples separately. EDWork t represents the patient volume arriving to ED triage in the hour prior to the index patient s arrival to the pod, and NoP eer jt is a dummy for whether physician j has no peer in the same pod. These regressions test whether physicians increase length of stay as expected future work increases within system and are also restricted to times when there are at least two physicians in the ED. I also perform the pooled regression Y = 4 1 (P eerstate jt = s) (α s EDWork t + δ s ) + (10) s=1 βx it + γp od it + ηi t + ν jk + ε ijkt, which estimates the degree of foot-dragging with increasing expected future work for each of the 4 following states: Working with peer in the self-managed system, working with peer in the traditional system, working alone in pod but not alone in ED, and working alone in ED. Results in Table 7 are consistent with previously estimated coefficients for the increase in length of stay with expected future work, in both systems when a peer is present. When a peer is not present, however, length of stay increases much more quickly with expected future work. Estimates suggest that, without a peer present, the response to expected future work quintuples in the traditional pod when no peer is present and increases in the self-managed pod (but effective traditional system) to almost triple the magnitude in the traditional team with a peer. Results from the pooled regression in Equation (10), shown in the third column of Table 7, confirm this and also show that physicians do not increase lengths of stay with expected future work when they are alone in the ED. These results suggest that physicians reduce their foot-dragging moral hazard when a peer is present, which is consistent with social incentives and the fact that peers can observe each other s 25

27 true workload better than physicians in different locations. My interpretation of increases in length of stay with expected future work as a foot-dragging moral hazard, that can be controlled by either social incentives or self-management, is also consistent with the results from the two falsification tests. First, physicians in the pod that is officially self-managed if a peer were present revert to foot-dragging, at levels even higher than in the traditional system with a peer, when there is no peer present because they are then practicing in an effective traditional system without a peer. Second, when a single physician is responsible for all patients entering the ED, I find no evidence of foot-dragging. Both regressions (9) and (10) include time dummies for hour of the day. More generally, these regressions (and all other regressions identifying foot-dragging) are interested in the response of length of stay with respect to expected future work, rather than levels of length of stay. In addition, essentially all observations with only one physician in the pod occur during transition times of 2-3 hours and correspond to shifts in which I see the same physician working with a peer. This implies that in practice I compare the behavior of the same physician and shift but under different peer conditions. ED conditions are generally unchanged in this short window of time relative to nearby times. One obvious difference when a peer is suddenly not present is that the ED workload is effectively distributed among one fewer physician. However, results are qualitatively unchanged when normalizing the ED patient volume measure of expected future work for the number of physicians in the ED. 6.2 Peer Type Given that foot-dragging is reduced by the presence of a peer, I next consider different types of social relationships between physicians and their peers. First, social connectedness driven by demographics or shared history has been shown to be important, presumably because people care more about what these peers think, in previous literature (Bandiera, Barankay, and Rasul, 2010, 2005; Jackson and Schneider, 2010). I consider peers of the same sex, similar age, or same place of residency training as potentially more connected to each other. Second, I consider peers who are more familiar with each other s behavior in the workplace by their history of time 26

28 working with each other, who could have better information or established reputations. 27 Third, I consider peers who are faster (or more productive) than median. The effect of this peer type may be both social and strategic. To see the strategic considerations, note that slower peers will cause more work to be redirected to physicians unless they slow down as well. Finally, I consider peers who have at least two years greater tenure than the index physician. Social hierarchy is a common feature in many workplaces and not least in medicine. Unlike other settings among entry-level workers, important social influences in workplaces with professionals, long tenures, or strong work cultures may be unilateral and driven by hierarchy. For each of these peer types I estimate separate regressions of this form on samples from traditional and self-managed teams: Y ijkt = α 1 EDWork t + α 2 P eert ype m jt EDWork t + α 3 P eert ype m jt + βx it + γp od it + ηi t + ν jk + ε ijkt, (11) where P eert ype m jt is an indicator that the peer for physician j at time t is of type m. I am interested in the coefficient α 2 in Equation (11) as the effect of working with a peer type on foot-dragging, again identified by increases in length of stay with respect to expected future work. In Table 8, I report results for three of the peer types. 28 As before, I interpret increases in length of stay with expected future work as foot-dragging. Foot-dragging is largely unchanged with each peer type, except for physicians working with senior peers. In the traditional system, working with a senior peer decreases foot-dragging by half, from an increase of 0.8% for each patient arriving to the ED to an increase of 0.4%. In the pooled regression, it also appears that senior peers further reduce foot-dragging in the self-managed system. For other peer types, including productive peers and familiar peers shown in Table 8, there is no significant change in foot-dragging. These results suggest that the most important social relationships with peers may be unilateral ones based on hierarchy, as opposed to ones that are based on connectedness, 27 I use a threshold of at least 60 hours, which is at the 75th percentile, to describe peers that are familiar with each other. Given physician turnover and a large number of shift times and locations, the same physician pair having a longer history working together in the same pod is relatively uncommon. 28 For brevity, I omit other peer types from Table 8, which also show no effect on foot-dragging. 27

29 productivity, or familiarity. The effect of peer relationships on efficiency may have implications on the efficient construction of teams. Of course, this depends on what relationships are most important. If hierarchical relationships are most important in mitigating foot-dragging, then one possible implementation will be to have teams composed of physicians with different tenures. However, this effect appears small relative to the effect of having any peer present or of the self-managed system. 6.3 Foot-dragging as a Census Policy As physicians in the traditional system both foot-drag and pay attention to the presence and identity of their peers, it follows that foot-dragging may also depend on what the peer is doing. As discussed above, an important summary statistic for state of work in a pod is the distribution of patients between a physician and his peer. Foot-dragging as a policy that depends on own and peer censuses could be influenced both by strategic behavior and social incentives. Strategically, in the traditional system, censuses influence patient assignment by the triage nurse. For example, in the theoretical framework, physicians with high censuses should have less of an incentive to foot-drag because they are already unlikely to receive patients. On the other hand, foot-dragging may be socially less acceptable at certain joint censuses. In particular, the experimental literature has convincingly shown reciprocity in games with public goods: Agents cooperatively refrain from moral hazard as long as their partners also refrain from moral hazard, and they also punish partners who deviate even if punishment is costly (Fehr and Gachter, 2000). Such cooperation and punishment are inconsistent with purely rational behavior but could be explained by social preferences. In a theory of reciprocity and punishment, Fehr and Schmidt (1999) show that these experimental findings can be simply explained by an aversion toward inequity, which again is a concept best summarized by censuses. For now, I aim to jointly test for strategic behavior and social incentives. I therefore evaluate foot-dragging as a policy function that depends on the state space of physician and peer censuses. I consider the census Census jt for physician j at time t, defined as the number of patients that physician j cared for that were on the ED floor at time t, and the corresponding 28

30 census of his peer Census jt. 29 I summarize these censuses into quintiles, where I denote an indicator for physician j s census at time t being in the m th quintile as QU m jt. I define the state space as the physician-peer quintile-quintile interaction, and I estimate the foot-dragging coefficient in each of these states in this regression: Y ijkt = 5 5 m=1 n=1 5 5 m=1 n=1 α m,n 1 QU m jt QU n jt EDWork t + α m,n 2 QU m jt QU n jt + βx it + γp od it + ηi t + ν vk + ε ijkt. (12) EDWork t is again the number of patients arriving to the ED in the hour before patient assignment to the pod and represents expected future work. I estimate this model separately for traditional and self-managed teams; I am interested in testing for heterogeneity in foot-dragging in the traditional system, while I am mostly using the self-managed system as a falsification test for heterogeneity, as I do not see any foot-dragging in the self-managed system anyway. The coefficients of interest are α m,n 1, which represent the degree of foot-dragging by physician j when his census is in the m th quintile and his peer s census is in the n th quintile. As before, I consider foot-dragging as any increase in length of stay with increasing expected future work. Table 9 presents estimated foot-dragging coefficients α m,n 1 for (12), with the first panel corresponding to traditional teams and the second panel corresponding to self-managed teams. These estimates specific to each quintile-quintile interaction reveal several features of policy functions in the two organizational systems. First, using 0 as the benchmark for no foot-dragging, there is virtually no foot-dragging in the self-managed system, regardless of censuses. This is consistent with earlier findings of little to no foot-dragging in the self-managed system. Second, the traditional system policy function shows remarkable dependency and predictability according to the exact quintile-quintile joint-census state, also shown in Figure 7. As predicted by strategic behavior, physicians foot-drag more when they have lower censuses. Although I do not separate strategic behavior from social incentives in this reduced-form analysis, physicians notably refrain 29 I assume that patients appear on physician j s census once they arrive on the pod. This is certainly true in the traditional system. In the self-managed system, it abstracts away from the fact that physician j has to choose patients on his census when working in self-managed teams. 29

31 from foot-draging when both censuses are normal in the third quintile, which is suggestive of cooperation. Of course, however, physicians usually fail to refrain from foot-dragging in the traditional system. 7 Patient Assignment and Equilibrium Building According to the management literature, self-managed teams improve efficiency by monitoring and managing work process and progress (Pallak and Perloff, 1986). More specifically, in the model in Section 3, if physicians sometimes use information about true relative workloads to choose patients in the self-managed system, foot-dragging should not only decrease, but the ex post efficiency of assignment should also increase. Patients should be assigned more often according to observable measures of workload such as censuses. In addition to studying patient assignment in steady state, a related but distinct issue is how physicians build the new equilibrium after Bravo pod switched from a traditional to a self-managed system. In Figure 8, I show that foot-dragging does not immediately disappear in Bravo by estimating the effect of expected future work, interacted with four-month interval dummies, on length of stay. Rather, it takes at least 5 months to disappear. Therefore, a second question relates to the assignment of patients during this transition period and in particular whether foot-dragging physicians are more likely to be assigned new patients. 30 To address these questions, I study the relationship between censuses and patient assignment. Censuses are a good measure of workload and are publicly observable, but they are distortable by foot-dragging. According to theory, patient assignment by the triage nurse will make limited use of census information because of the threat of foot-dragging, but as this threat is reduced in the self-managed system, assignment may become more correlated with censuses. In addition to equilibrium assignment of both systems, assignment during the transition of Bravo pod will shed light on how the no foot-dragging equilibrium is eventually enforced by assignment. In particular, assigning patients to foot-dragging physicians should result in a positive correlation 30 For a more in depth discussion about how patients are assigned in the self-managed system, see Section 3.4 and Appendix B. Regardless of whether physicians can commit to an assignment policy and of the amount of information they observe, patients should be assigned to physicians who foot-drag with greater probability. 30

32 with censuses. In particular, I study the correlation between censuses as a signal of workload and new patient assignment for both pods over time. For my baseline specification, I estimate the linear probability model I ijt = αcensus jt + βshift jt + η j + ν it + ε ijt, (13) where the outcome I ijt is an indicator variable for whether patient i who arrives on the pod at time t is assigned to physician j. Census jt denotes the number of patients under the care of physician j at time t and is the variable of interest. Shift jt includes time indicators of physician j s shift at time t, since physicians are less likely to be assigned new patients as they near the end of their shift, regardless of their censuses. I also control for physician identities by the fixed effect η j, which allows that some physicians are more likely to take new patients regardless of census or observed behavior by their peers. The term ν it ensures that two physicians could be assigned each patient i and that only one physician is assigned the patient while the other is not; this is equivalent to restricting the sample to patients that meet these criteria. This linear probability model is also equivalent to estimating a differenced model in which the variable of interest is Census jt Census jt with a coefficient algebraically equal to α because at most two physicians are available to be assigned a patient conditional on the pod. Figure 9 shows a plot of the set of coefficients α in (13) over time and in both pods. 31 I estimate α over each month by using triangular kernels with 45 days on each side of the first of the month; for months immediately before and after the regime change, I only use 45 days on the side away from the regime change. α represents the incremental likelihood, averaged over different shift times, for a physician to receive a new patient entering the pod if he has one extra patient on his census. Prior to the Bravo regime change, both Figures show relatively stable assignment in both pods with the traditional system in Bravo and the self-managed system in Alpha. In both systems, physicians with lower censuses are more likely to be assigned patients, but this likelihood is much greater in the self-managed system. At the time of the regime change in March 2010, there is a striking jump in which physicians with higher censuses are actually more likely to 31 See Appendix E for a set of plots representing the same estimates but with confidence intervals. 31

33 receive new patients. This spike rapidly disappears, and after 3 months, patients are again more likely to be assigned to physicians with lower censuses, even more so than prior to the regime change. In addition, Bravo s assignment function in equilibrium (after the spike) is the same as Alpha s. The qualitative features of these assignment policy functions are robust to a number of different specifications, including logit estimation, omission of physician fixed effects η j, alternative kernel bandwiths, and controlling for other workload observables such as the number of patients under different acuity levels. Answering the first question, these results show that, in equilibrium, the self-managed system improves the ex post efficiency of assignment according to publicly observable signals of workload. I also utilize the difference-in-differences framework to show that these assignment functions are not specific to pods, but rather to the organizational systems. This is consistent with the theory in Section 3: Because foot-dragging is reduced, new patients can be more readily assigned to physicians with lower censuses. It also provides empirical evidence that mechanism for reducing foot-dragging in the self-managed system is not trivial, for example through randomly assigning patients regardless of workload, but substantive in that it involves superior information among physicians. The answer to the second question of the assignment after the regime change in Bravo pod is admittedly more speculative. Most standard models, including mine, are silent on building new equilibria, since behavior is considered already in equilibrium. However, there is reason to believe that the transition will not be immediate. A well-known regularity in experimental economics is that players do not usually settle immediately on equilibria, even when they are unique. Experiments have also shown that players are influenced by past experiences that should be irrelevant to the current game (Bohnet and Huck, 2004) or by referring to the same game by different names (Liberman, Samuels, and Ross, 2004). Specifically in this empirical setting, although physicians have worked in self-managed teams in Alpha, the regime change in Bravo was not announced as a move to replicate the organizational system in Alpha, but rather was a simple merger of bed ownership with nothing else changed (e.g., Bravo was still closed at night). As mentioned in Section 2, most physicians work sporadically and could be new to the regime change in Bravo even after a week or more. Finally, the self-managed system requires 32

34 physicians to work more as a team, and therefore outcomes are more dependent on the beliefs and strategies of both peers. Given that full cooperation is not immediate, the question is how full cooperation is eventually established. Again, some insight can be gained from the experimental literature. Enforcement in public goods games has been studied in seminal research by Ostrom, Walker, and Gardner (1992) and Fehr and Gachter (2000). They have found that when allowed to do so, by the rules of the game akin to organizational structure in my setting, players can enforce full cooperation through punishment. Unlike these tightly designed experiments, I cannot definitively say that punishment occurs at the transition. I do however see that physicians who have higher censuses, during the same time when there is residual foot-dragging in the self-managed system, are more likely to be assigned new patients. This is consistent with foot-dragging physicians being assigned patients. Whether this is punishment, or simply an increase in the likelihood of assignment to physicians who have lower true workloads as my model predicts, this assignment behavior supports the new equilibrium with no foot-dragging. 8 Conclusion This paper presents evidence for organizational structure as an important means to improve the performance of physicians working together. In particular, this setting draws a contrast between two systems in which physicians have different managerial authority. In a traditional system, physicians are assigned work by a triage-nurse manager, while in the other selfmanaged system they are responsible for assigning work among themselves. The self-managed system produces significantly shorter lengths of stay than the traditional system, and this is essentially due to reducing foot-dragging, a moral hazard in which physicians keep patients longer to prevent new work from the triage nurse. Foot-dragging is reduced by social incentives, most notably by the presence of a peer and by hierarchical relationship with peers, but physicians in the traditional system are generally unable to eliminate foot-dragging. The self-managed system further reduces foot-dragging by both monitoring and managing patient assignment. Better information between physician peers both prevents moral hazard and allows for more 33

35 efficient assignment, as physicians learn that foot-dragging does not pay. Although this paper shows that a significant improvement in productivity occurs with a relatively simple rearrangement in organizational structure, it does not inform a horserace between organizational and financial instruments of controlling physician behavior. In addition, self-managed teams and indeed health care organizations in practice differ a wide variety of ways, including for example the number of health care workers in the same team and the nature of work relationships among team members. This study shows only a small, although controlled and well-described, experiment in the arrangement of teams. Other arrangements may have significantly different effects on physician behavior and should be further studied. Also, although I find that alternative mechanisms, such as matching patients and physicians, do not explain a significant amount of the overall effect of self-managed systems, I do find preliminary evidence that physicians choose patients whom they are better at seeing. This is beyond the scope of this paper and is an area of further research. Finally, for practical reasons, I focus on observable productivity, but there may be a host of other considerations, such conflict among physicians, that could arise from self-management but may be more difficult to measure in this study. However, many of these issues have been addressed extensively in a literature outside of economics and point quite favorably toward self-management (Yeatts and Hyten, 1997; Pallak and Perloff, 1986). References Abadie, A. (2005): Semiparametric difference-in-differences estimators, Review of Economic Studies, 72(1), (2010): Synthetic control methods for comparative case studies: Estimating the effect of California s tobacco control program, Journal of the American Statistical Association, 105(490), Abadie, A., A. Diamond, and J. Hainmueller (2011): Comparative politics and the synthetic control method, MIT Political Science Department Research Paper, No Bandiera, O., I. Barankay, and I. Rasul (2005): Social preferences and the response to incentives: Evidence from personnel data, The Quarterly Journal of Economics, 120(3), (2009): Social connections and incentives in the workplace: Evidence from personnel data, Econometrica, 77(4),

36 (2010): Social incentives in the workplace, The Review of Economic Studies, 77(2), Bernstein, S. L., D. Aronsky, R. Duseja, S. Epstein, D. Handel, U. Hwang, M. Mc- Carthy, K. John McConnell, J. M. Pines, and N. Rathlev (2008): The effect of emergency department crowding on clinically oriented outcomes, Academic Emergency Medicine, 16(1), Bertrand, M., E. Duflo, and S. Mullainathan (2004): How much should we trust differences-in-differences estimates?, The Quarterly Journal of Economics, 119(1), Bliss, C., and B. Nalebuff (1984): Dragon-slaying and ballroom dancing: The private supply of a public good, Journal of Public Economics, 25(1), Bloom, N., B. Eifert, A. Mahajan, D. McKenzie, and J. Roberts (2011): Does management matter? Evidence from India, Discussion paper, National Bureau of Economic Research. Bloom, N., and J. Van Reenen (2007): Measuring and Explaining Management Practices Across Firms and Countries, Quarterly Journal of Economics, 122(4), Bohnet, I., and S. Huck (2004): Repetition and reputation: Implications for trust and trustworthiness when institutions change, The American Economic Review, 94(2), Brook, R. H. (2010): Physician compensation, cost, and quality, The Journal of the American Medical Association, 304(7), Fehr, E., and S. Gachter (2000): Cooperation and punishment in public goods experiments, The American Economic Review, 90(4), Fehr, E., and K. M. Schmidt (1999): A theory of fairness, competition, and cooperation, The Quarterly Journal of Economics, 114(3), Forster, A. J., I. Stiell, G. Wells, A. J. Lee, and C. Van Walraven (2003): The effect of hospital occupancy on emergency department length of stay and patient disposition, Academic Emergency Medicine, 10(2), Hamilton, B. H., J. A. Nickerson, and H. Owan (2003): Team incentives and worker heterogeneity: An empirical analysis of the impact of teams on productivity and participation, The Journal of Political Economy, 111(3), Heckman, J. J., H. Ichimura, and P. E. Todd (1997): Matching as an econometric evaluation estimator: Evidence from evaluating a job training programme, The Review of Economic Studies, pp Holmstrom, B. (1982): Moral hazard in teams, The Bell Journal of Economics, 13(2), Ichniowski, C., K. Shaw, and G. Prennushi (1997): The effects of human resource management practices on productivity: A study of steel finishing lines, The American Economic Review, 87(3),

37 Institute of Medicine (2012): Best Care at Lower Cost: The Path to Continuously Learning Health Care in America, Discussion paper, National Academies Press, Washington, DC. Jackson, C. K., and H. S. Schneider (2010): Do social connections reduce moral hazard? Evidence from the New York City taxi industry, Discussion paper, National Bureau of Economic Research. Kandel, E., and E. P. Lazear (1992): Peer pressure and partnerships, The Journal of Political Economy, pp Lazear, E. P., and S. Rosen (1981): Rank-order tournaments as optimum labor contracts, The Journal of Political Economy, 89(5), Lerman, B., and M. S. Kobernick (1987): Return visits to the emergency department, The Journal of Emergency Medicine, 5(5), Liberman, V., S. M. Samuels, and L. Ross (2004): The name of the game: Predictive power of reputations versus situational labels in determining prisoner s dilemma game moves, Personality and Social Psychology Bulletin, 30(9), Manski, C. F. (1993): Identification of endogenous social effects: The reflection problem, The Review of Economic Studies, 60(3), Mas, A., and E. Moretti (2009): Peers at work, The American Economic Review, 99(1), McCarthy, D., and D. Blumenthal (2006): Stories from the sharp end: case studies in safety improvement, Milbank Quarterly, 84(1), Milgrom, P., and J. Roberts (1988): An economic approach to influence activities in organizations, The American Journal of Sociology, 94(Supplement), Moore, J., and R. Repullo (1988): Subgame perfect implementation, Econometrica, 56(5), Oliver, A. (2007): The Veterans Health Administration: an American success story?, Milbank Quarterly, 85(1), Ostrom, E., J. Walker, and R. Gardner (1992): Covenants with and without a sword: Self-governance is possible, The American Political Science Review, pp Pallak, M. S., and R. O. Perloff (1986): Psychology and Work: Productivity, Change, and Employment. American Psychological Association, Washington DC. Rosenbaum, P. (2002): Covariance adjustment in randomized experiments and observational studies, Statistical Science, 17(3), Rubinstein, A. (1982): Perfect equilibrium in a bargaining model, Econometrica, 50(1), Schuur, J. D., and A. K. Venkatesh (2012): The growing role of emergency departments in hospital admissions, The New England Journal of Medicine, 367,

38 Thompson, D. A., P. R. Yarnold, D. R. Williams, and S. L. Adams (1996): Effects of actual waiting time, perceived waiting time, information delivery, and expressive quality on patient satisfaction in the emergency department., Annals of emergency medicine, 28(6), Yeatts, D. E., and C. Hyten (1997): High-performing self-managed work teams: A comparison of theory to practice. Sage Publications, Incorporated. 37

39 Figure 1: Patient-to-physician Assignment Algorithm Patient arrives in ED triage Self-managed Triage nurse decides on pod Traditional Triage nurse decides on bed Triage nurse decides on bed Patient arrives in pod bed Patient arrives in pod bed Physicians decide on assignment Physicians own beds Physician assigned Physician assigned Note: This figure shows the patient assignment algorithm starting with patient arrival to ED triage and ending with assignment to a physician. In ED triage, the triage nurse decides which pod to send the patient. If she decides to send the patient to a pod with a self-managed system, then she does not assign the physician. The physicians currently working in the self-managed pod will decide among themselves on that assignment. If the triage nurse decides to send the patient to a pod with a traditional system (if one exists), then she also makes the decision on which physician will be assigned the patient. Although not shown in the figure, the triage nurse always assigns the bed and the nurse; she never assigns the resident, since residents in either pod choose their own patients or are told by physicians to see patients. 38

40 Figure 2: Computer Schematic of Alpha Pod Note: This figure shows a patient-deidentified computer screen layout of Alpha pod, which is both a geographic representation of the physical pod and the interface for physicians to select patients, examine the electronic medical record, and enter orders. Slots represent beds, with two beds per room, and filled slots are represented by slots with information. Colors represent various patient states, for example, whether an order needs to be taken off or whether the patient has been ordered for discharge. Identifying patient information has been removed, but information when used includes patient last name, chief complaint, age, sex, physician, resident, and nurse, emergency severity index, and minutes in ED (including triage) and in pod. 39

41 Figure 3: Computer Schematic of Bravo Pod Note: This figure shows a patient-deidentified computer screen layout of Bravo pod, which is both a geographic representation of the physical pod and the interface for physicians to select patients (when applicable under the self-managed system), examine the electronic medical record, and enter orders. Information represented is the same as for Alpha pod, shown and explained in Figure 2. When Bravo pod was under the traditional system, the geographic layout and the screen was simply divided in half, with generally one physician occupying each half. The image above was taken with Bravo under the self-managed system. 40