Repositorio Institucional de Documentos

2013 80 Lijie Song Improve primary care performance through operations management: an application to emergency care and preventive care Departamento Director/es Zaragoza Logistics Center Sáenz Gil de Gómez, María Jesús Horatius, Nicole Director/es Tesis Doctoral Autor Repositorio de la Universidad de Zaragoza – Zaguan http://zaguan.unizar.es UNIVERSIDAD DE ZARAGOZA Departamento Director/es Lijie Song IMPROVE PRIMARY CARE PERFORMANCE THROUGH OPERATIONS MANAGEMENT: AN APPLICATION TO EMERGENCY CARE AND PREVENTIVE CARE Director/es Zaragoza Logistics Center Sáenz Gil de Gómez, María Jesús Horatius, Nicole Tesis Doctoral Autor 2013 Repositorio de la Universidad de Zaragoza – Zaguan http://zaguan.unizar.es UNIVERSIDAD DE ZARAGOZA Departamento Director/es Director/es Tesis Doctoral Autor Repositorio de la Universidad de Zaragoza – Zaguan http://zaguan.unizar.es UNIVERSIDAD DE ZARAGOZA UNIVERSIDAD DE ZARAGOZA TESIS DOCTORAL Mejora del rendimiento en atención primaria a través de la gestión de operaciones: aplicación a atención de urgencia y al cuidado preventivo SONG, Lijie Máster en administración de la gestión logística, Universidad RMIT (Australia) Máster en ingeniería de la logística y de la gestión de la cadena de suministros, Universidad de Zaragoza (España) 22 de marzo de 2013 ©Song, Lijie. Reservados todos los derechos Autor: Dña. Song, Lijie, Doctorando Director de tesis: Dra. Nicole de Horatius Profesor de Gestión de Operaciones, Universidad de Chicago Profesor Adjunto del Programa Internacional de Logística MIT-Zaragoza Co-Director de tesis: Dra. María Jesús Sáenz Profesor de Gestión de la Cadena de Suministro, Programa Internacional de Logística MIT-Zaragoza Profesora Titular de la Universidad de Zaragoza Director del Programa de Doctorado, Zaragoza Logistics Center (ZLC) Director del Centro de Logística de Zaragoza (ZLC): Dr. David Gonsalvez Contenidos 1. Introducción...................................................................................................................................... 9 2. Reducción del tiempo de servicio y del tiempo de espera del departamento de urgencias a través del rediseño del proceso ………………………………………………………………………………… 15 2.1 Introducción................................................................................................................................. 15 2.2 Literatura relacionada .................................................................................................................... 17 2.3 Formulación de hipótesis................................................................................................................. 19 2.3.1 Tiempo de servicio........................................................................................................................ 20 2.3.2 Calidad del servicio................................................................................................................ 21 2.3.3 Tiempo de espera.......................................................................................................................... 21 2.4 Contexto de investigación........................................................................................................... 22 2.4.1 Configuración de la investigación............................................................................................... 23 2.4.2 Recopilación de datos y mediciones............................................................................................. 23 2.4.3 Variables de control...................................................................................................................... 25 2.4.4 Estadísticas descriptivas............................................................................................................... 21 2.5 Cálculo y resultados ........................................................................................................................ 30 2.5.1 Correspondencia del grado de propensión y generación de grupo de cuasi control..................... 30 2.5.2 Tiempo de servicio........................................................................................................................ 32 2.5.3 Calidad del servicio.................................................................................................................... 33 2.5.4 Tiempo de espera...........................................................................................................................34 2.6 Debate e investigación futura................................................................................................. 38 3.Comprensión de las preferencias del cliente en cuidado preventivo.................................................. 41 3.1 Introducción................................................................................................................................. 41 3.2 Revisión de la literatura................................................................................................................. 43 3.3 Metodología..................................................................................................................................... 46 3.4 Métodos ...................................................................................................................................... 47 3.4.1 Grupo destinatario conociendo las instalaciones actuales............................................................ 47 3.4.2 Diseño de la evaluación................................................................................................................ 49 3.4.3 Tamaño de la muestra................................................................................................................... 51 3.5 Datos................................................................................................................................................ 51 3.5.1 Antecedentes................................................................................................................................. 51 3.5.2 Estadísticas descriptivas............................................................................................................... 52 3.6 Resultados.................................................................................................................................... 53 3.6.1 Resultado numérico................................................................................................................ 53 3.6.2 Ejemplo de aplicación................................................................................................................... 55 3.7 Análisis de clase latente.............................................................................................................. 58 3.7.1 Número de posibles grupos latentes...................................................................................... 59 3.7.2 Comparación grupal de preferencias de atributos.................................................................. 59 3.7.3 Comparación grupal de características demográficas................................................................... 60 3.8 Análisis de simulación................................................................................................................. 62 3.8.1 Modelo básico............................................................................................................................... 62 3.8.2 Modelo de ajuste 1........................................................................................................................ 65 3.8.3 Modelo de ajuste 2 y de ajuste 3................................................................................................... 66 3.9 Conclusión....................................................................................................................................... 67 4. Conclusión......................................................................................................................................... 69 UNIVERSIDAD DE ZARAGOZA TESIS DOCTORAL Improve primary care performance through operations managementan application to emergency care and preventive care SONG, Lijie Master of Business in Logistics Management, RMIT University, Australia Master of Engineering in Logistics and Supply Chain Management, Universidad de Zaragoza, España March 22, 2013 ©Song, Lijie. All rights reserved Author: Song, Lijie Thesis Advisor: Dr. Nicole de Horatius Professor of Operations Management, University of Chicago Adjunct Professor of Supply Chain Management at the MIT-Zaragoza International Logistics Program Thesis Co-Advisor: Dr. María Jesús Sáenz Professor of Supply Chain Management, MIT-Zaragoza International Logistics Program Professor, University of Zaragoza PhD Program Director, Zaragoza Logistics Center (ZLC) Director, Zaragoza Logistics Center (ZLC): Dr. David Gonsalvez 1 Table of Contents 1. Summary ......................................................................................................................................................... 9 2. Point-of-Care Testing: Improving Emergency Department Performance through Process Redesign ........ 15 2.1. Introduction .......................................................................................................................................... 15 2.2. Related Literature ................................................................................................................................. 17 2.3. Hypothesis Formulation ....................................................................................................................... 19 2.3.1. Service Time ............................................................................................................................... 20 2.3.2. Service Quality ............................................................................................................................ 21 2.3.3. Waiting Time .............................................................................................................................. 21 2.4. Research Context ................................................................................................................................. 22 2.4.1. Research Setting .......................................................................................................................... 22 2.4.2. Data Collection and Measures .................................................................................................... 23 2.4.3. Control Variables ........................................................................................................................ 25 2.5. Estimation and Results ......................................................................................................................... 30 2.5.1. Propensity Score Analysis and Generation of Quasi-control Group .......................................... 30 2.5.2. Service Time ............................................................................................................................... 32 2.5.3. Service Quality ............................................................................................................................ 33 2.5.4. Waiting Time .............................................................................................................................. 34 2.6. Discussion and Future Research .......................................................................................................... 38 3. Understand Client Preferences for Preventive Care ..................................................................................... 41 3.1. Introduction .......................................................................................................................................... 41 3.2. Literature review .................................................................................................................................. 43 3.3. Methodology ........................................................................................................................................ 46 3.4. Methods ................................................................................................................................................ 47 3.4.1. Focus group meeting current facilities ........................................................................................ 47 3.4.2. Survey design .............................................................................................................................. 49 3.4.3. Sample size ................................................................................................................................. 51 3.5. Data ...................................................................................................................................................... 51 3.5.1. Background ................................................................................................................................. 51 3.5.2. Descriptive statistics .................................................................................................................... 52 3.6. Results .................................................................................................................................................. 53 3.6.1. Numerical Result ......................................................................................................................... 53 3.6.2. Application example ................................................................................................................... 55 3.7. Latent Class Analysis ........................................................................................................................... 58 3.7.1. Number of possible latent groups ............................................................................................... 59 3.7.2. Group Comparison of Attribute Preferences ............................................................................... 59 3.7.3. Group Comparison of Demographic Characteristics .................................................................. 60 2 3.8. Simulation Analysis ............................................................................................................................. 62 3.8.1. Basic model ................................................................................................................................. 62 3.8.2. Adjust1 model ............................................................................................................................. 65 3.8.3. Adjust2 and Adjust3 model ......................................................................................................... 66 3.9. Conclusion ........................................................................................................................................... 67 4. Conclusion .................................................................................................................................................... 69 3 List of Figures Figure 2-1: Stylized Emergency Department Patient Flow Process .................................................................... 20! Figure 2-2: Illustration of Test Status .................................................................................................................... 23! Figure 2-3: Matching Process ................................................................................................................................ 31! Figure 3-1: Coverage of the 9 clinics .................................................................................................................... 57! Figure 3-2: Marginal change of probability .......................................................................................................... 58! Figure 3-3: Simplified Arena model ...................................................................................................................... 64! 4 List of Tables Table 2-1: Summary of Observed Patient Visits .................................................................................................. 24! Table 2-2: List of Variables .................................................................................................................................. 28! Table 2-3: Correlation Table ................................................................................................................................ 28! Table 2-4: Descriptive Statistics for Complete Data and by Test Status (2007 and 2008) ................................... 28! Table 2-5: Descriptive Statistics for Unmatched Data (TEST Patients) .............................................................. 29! Table 2-6: Descriptive Statistics for Unmatched Data for (NO-TEST Patients) .................................................. 29! Table 2-7: Datasets Used for Hypothesis Testing ................................................................................................. 31! Table 2-8: Regression Results for H1a, H1b, and H2 ........................................................................................... 33! Table 2-9: Regression Results for H3a at Peak Hours .......................................................................................... 35! Table 2-10: Regression Results for H3a during Off-peak Hours .......................................................................... 35! Table 2-11: Regression Results for H3b ............................................................................................................... 37! Table 2-12: P-value for F-test between POCT2, POCT3 and POCT4 .................................................................. 37! Table 2-13: Predicted percentage change in waiting time for patients in each priority class ............................... 37! Table 3-1: Final attributes and levels .................................................................................................................... 48! Table 3-2: Respondents' characteristics to be collected ........................................................................................ 48! Table 3-3: Example of choice task ........................................................................................................................ 50! Table 3-4: Summary of respondents' characteristics ............................................................................................. 52! Table 3-5: Summary of attribute level for chosen clinics ..................................................................................... 53! Table 3-6: Hierarchical regression result .............................................................................................................. 54! Table 3-7: A realistic choice set faced by a woman lives in Westmount .............................................................. 57! Table 3-8: Predicted probability of choosing each clinic ...................................................................................... 57! Table 3-9: Summary of CAIC and Relative Chi Square of best replications ........................................................ 59! Table 3-10: Attribute importance of two latent groups ......................................................................................... 60! Table 3-11: Demographic characteristics’ comparison between two groups ........................................................ 61! Table 3-12: Logit prediction model result ............................................................................................................. 61! Table 3-13: Montreal population zone and corresponding participation rate ....................................................... 62! Table 3-14: Arena result from Basic model and Adjust1 model ........................................................................... 64! Table 3-15: Arena result from Basic model and Adjust1 model ........................................................................... 65! Table 3-16: Arena result from Adjust2 model and Adjust3 model ....................................................................... 66! 5 6 Abstract Improve primary care performance through operations managementan application to emergency care and preventive care The main purpose of this thesis is to apply operations management method to improve health care providers’ performance in two major component of primary care: emergency care and primary care. Operations Management (OM) and Operations Research (OR) has been applied to health care for many years to improve health care delivery efficiency. The center of medical care system is primary care, whose key functions include providing an entry point, delivering core medical and preventive care, and helping patients coordinate and integrate care, all of which are critical in improving not only health outcome of patients, but also cost performance of the whole health care system. In a study on primary care and health system performance, US reported much higher Emergency Department (ED) use rates than the other three countries, accompanied by lowest percentage of adults that have regular doctor, places or clinics to go when they are sick. Chapter 2 of this dissertation therefore target improving emergency room department through process redesign. Another key finding of the survey is that Canada has the lowest rate of screening for Pap test and mammogram. Given the significance of preventive care in saving lives and reducing cost, chapter 3 of this dissertation explores how to improve government funded preventive care program through network design. Chapter 2 set the context in a tertiary hospital emergency department (ED) that have an annual census of 55,000 patients, and look at how does redesigning process for a specific blood test impact the congestion of the ED. More specifically, we look at the change in three performance metrics after the analysis of patients blood sample for troponin level was moved from the hospital’s central lab to inside the ED. Use priority queueing theory, we generate hypotheses on the following performance measures: waiting time (defined as the time difference between patient intake registration and bed assignment), service time (defined as the time difference between bed assignment and disposition), and service quality (defined as patients’ 72-hour revisit rate). Using difference-in-difference model, we find the process redesign to be associated with statistically significant improvements in nearly all measures of operational performance. Specifically, we find the adoption of POCT to be associated with a 21.6% reduction in service time among test patients during peak hours and a 5.9% to 35.5% reduction in waiting time depending on the patient’s priority class during peak hours. Moreover, we find the adoption of POCT to be associated with improved service quality as patients’ predicted probability of bounce back decreased by 0.64% during its usage. We also find system wide spillover effects for patients 13 such programs (Verter and Lapierre 2002). In chapter 3 of this dissertation, we turn our focus from one aspect of primary care at unit level, improving emergency care performance, to another important piece of primary care under a broader context, improving preventive care performance at regional level. We chose breast cancer screening (mammogram) program in Canada as our studied care program due to the importance of early detection in reducing mortality of breast cancer, and the lower screening rate in the chosen country. Breast cancer is one of the most commom cancers in the world, accounts for 18% of women’s cancers worldwide (Hamilton and Barlow 2003). In US alone in 2004, breast cancer is responsible for the death of 40954 women, among 186,772 diagnosed. It is also the second leading cancer cause of death among women in Canada (2011). Despite the fact that early detection of breast cancer through high quality mammographic screening has been proved to be able to reduce breast cancer mortality significantly (Aro et al. 1999), uptakes of government funded breast cancer screening are not meeting targets in many countries, especially for rate of rescreening participation (Lechner et al. 1997). Studies suggest that factors such as social demographic characteristics and health behaviors of participants play an important role in the decision to participate in preventive care programs, other factors, such as type of facilities, accessibility by public transport may also choices of facilities to visit (Gerard et al. 2003; Hamilton and Barlow 2003; Maheswaran et al. 2006). Efficiency of cancer screening depends heavily on the frequency of screening (Cohen et al. 2008), therefore improving rescreening participate rate of these programs can have great impact on the health outcome of target population. There is very few studies aim at improving rescreening rate specifically for breast cancer screening. Existing research on improving breast cancer screening participation usually focus on three types of factors: (1) social-demographic factors, (2) psychological factors, (3) interventions (Aro et al. 1999; Kee et al. 1992; Munn 1993; Rimer et al. 1998; Sutton et al. 1994). Social-demographic factors are personal characteristics of population, such as their education level, income level, race group, age, living condition, etc. These studies use such information to predict the probability of participation for a known population. Psychological factors include self-perceived level of risk of cancer and concern about pain. Aro et al. (1999) conducted interview on women who were invited to a first round mammography in Finland, and found that high risk group may be related to more frequent earlier mammogram and weekly breast self examination, although their conclusion is contrary to that of Sutton et al. (1994), in which the authors found that women with self perceived high and low risk of breast cancer have lower attendance rate than women with moderate amount of perceived risk. Kee et al. (1992) found that attitudes rather than access play the most important role in influencing uptake. In their interview with women invited for mammography and declined attendance, the most cited 14 reasons are feelings of indifference or ignorance of screening issues and fear of pain or embarrassment, only less than 4% interviewed women expressed preference for more accessible screening unit. Munn (1993) and Rimer et al. (1989) found similar results in their interviews of women who declined participation of mammography invitations. However, these results cannot be applied to women that have participated in the first round screening, but decided not to attend rescreening, where non-attitude related factors are more likely to be the main barrier. In our study, using stated preference discrete choice modeling (SPDCM), we identify factors that affect current women’s screening decision, and their choice of facilities when choosing among a set of screening clinics for their regular biannual breast cancer screening. We conducted survey on the current participants of the Québec Breast Cancer Screening Program (PQDCS) in Montreal, Canada, a population based breast cancer screening. This is, to the best of our knowledge, the first paper that combines focus group meeting and SPDCM on population based cancer screening program to generate managerial implications for configuration of preventive care facilities, from the perspective of service and accessibility of facilities. We also combine survey result with aggregated data at population level in a simulation model to predict change in certain performance metrics for the overall system based on different configuration of service attributes. Our result shows that nursing staff’s manner and knowledge regarding breast cancer and screening, as well as waiting time for appointment are the most influential factors in choice of clinics. Using latent class analysis, we are also able to identify the homogeneity of program participants, judging by their preference over studied attributes. Our simulation models shows that when taking into account of clients preferences, great improvement to appointment time, utilization rate, and screening volume can be achieved in the current system. Such change can lead to significant increase in rescreening rate at very little cost, without an overhaul of the current facility network. 15 2. Point-of-Care Testing: Improving Emergency Department Performance through Process Redesign 2.1. Introduction For more than a decade, U.S. hospital emergency departments (EDs) have been concerned about the adverse health consequences of overcrowding (Melissa et al. 2009), which has been linked to an increase in patient mortality (Bernstein et al. 2009, Richardson 2006), longer patient wait times (Melissa et al. 2009), lower patient satisfaction (Booth et al. 1992), higher ambulance diversion rates (Allon et al. 2009), and a growing proportion of patients leaving without being seen (Bernstein et al. 2009, Weiss et al. 2005). Among the most frequently cited reasons for overcrowding are closure of existing hospitals (Institute of Medicine 2006), use of EDs for routine medical care (Murray and Berwick 2003), and lack of ED physician and nursing staff (Green 2008). Given the complexity of the overcrowding problem, hospital administrators looking to accommodate more patients within existing infrastructure have debated, among other process changes, point-of-care testing (POCT). POCT refers to the testing of specimens at or near the location of patient care (Jahn and Aken 2003) instead of transporting specimens drawn at bedside to a central lab for testing. Potential benefits of POCT include reductions in time to medical decision-making, length of patient stay, and ED overcrowding, and a concomitant improvement in patient satisfaction (Murray et al. 1999). Previous studies of POCT adoption and benefits are inconclusive. Our study differs from these in that we control for such alternative explanations for observed reductions in length of stay as patient severity, ED congestion, and seasonality. Nor does our study, as do studies that employ randomized trials, depend on the compliance of hospital staff during the intervention period. Lack of compliance is often advanced as a reason for the lack of significance in POCT studies (Rust et al. 2008). Moreover, whereas previous studies have tended to focus exclusively on patients who receive POCT, we examine what happens to all patients, test patients and otherwise, who arrive at the ED at the same time upon adoption of POCT. 16 Our study, conducted in a large, urban, tertiary, academic hospital ED with an annual census of 55,000 patients, involves evaluating key ED metrics like service time, waiting time, and service quality pre and post POCT adoption. We use queuing theory to generate, and the data we collect to test, a number of hypotheses regarding ED performance. We examine not only how POCT adoption affects patients whose samples are analyzed at bedside, but also spillover effects on other patients in the ED at the same time. We further examine whether POCT adoption has a uniform impact across all patients or a differential impact based on patient severity, and whether the benefits of adoption vary across different periods throughout the day (i.e., peak vs. nonpeak). A summary of key findings follows. We find converting a single test from central lab processing to POCT to be associated, among test patients who arrive during peak hours, with a statistically significant decrease in service time of, on average, 21.6% for test patients and 4.7% for patients who do not receive the test (hereafter referred to as no-test patients).1 For test patients, we attribute the service time reduction largely to the difference in the time it takes to execute the test in the central lab compared to at bedside, for no-test patients, to the reduced demand for central lab resources, which may improve the lab’s ability to process other tests required for existing patients. We further show that the observed decrease in service time does not come at the expense of service quality. Using the 72-hour re-admission rate, a common measure of ED service quality (Mayer et al. 1998, Guttmann et al. 2006), we find that service quality improves upon adoption of POCT, the predicted probability of bounceback decreasing, in absolute terms, by 0.6% for a typical patient. We find the impact of POCT adoption on waiting time to be similar across test and no-test patients during peak hours, and the magnitude of the effect to differ depending on patient severity. Patients who have the lowest priority, identified as those who are the least sick, experience the greatest improvement in waiting time, 35.5% during peak hours compared to patients in higher priority classes. This is because lower priority patients benefit from service time reductions in higher priority patients, whereas high priority patients are typically unaffected by the service times of lower priority patients. The rest of the paper is organized as follows. We review in Section 2 the relevant literature on ED operational performance and POCT. In Section 3, we present our analytical queuing model and formulate our hypotheses. Our research setting and the data we collected are described in Section 4, the empirical methods 1 Although we analyze these data for both peak and non-peak periods, for brevity we describe only the peak hour results. In an overview of this analysis in Section 6, we reveal the impact of POCT adoption to be greater during high congestion, peak hour periods compared to other periods (non-peak hours). 17 used to test our hypotheses delineated, and our results summarized, in Section 5. Implication of our results and limitations of our research are discussed in Section 6. 2.2. Related Literature We draw from and contribute to both the operations management and emergency care management research streams. The operations management literature includes numerous studies that address issues pertaining to the management of EDs, and calls for operations researchers to aid hospital administrators in developing a more efficient service delivery system are increasing (see, for example, Guarisco and Samuelson 2011 and Green 2008). Such studies in the medical literature as pertain to understanding and improving ED operations tend to focus on therapeutic rather than operational matters. Relevant to our work are the medical studies that examine specifically POCT, the findings of which we detail below. Operations researchers have sought to improve ED performance by optimizing ED staff scheduling for observed seasonality in patient arrivals over the course of the day and throughout the week (Green et al. 2006, Vassilocopoulous 1985a), proposing novel triage systems (Saghafian et al. 2011), identifying drivers of ambulance diversions (Allon et al. 2009), and improving inpatient bed allocation (Vassilocopoulous 1985b, Green and Nguyen 2001, Balaji and Brownlee 2009). Wiler et al. (2011) review the operational literature pertaining to emergency department patient flow. Hospital administrators have also explored process redesign as a way to improve the efficiency of EDs (Meislin et al. 1988, Considine et al. 2008). Cooke et al. (2004) and Gorelick et al. (2005) propose deferring part of the registration process until a patient is assigned a room and a physician has completed an initial evaluation. A “fast-track” system, by employing nurse practitioners to serve low priority patients (i.e., patients with minor, specific ambulatory or acute illness or injury), prevents long wait times and overcrowding among nonemergency care patients (Meislin et al. 1988, Yoon 2003). Kilgore et al. (1998) and Jahn and Aken (2003) posit POCT as one way to alleviate the bottleneck that results when a centralized laboratory serves the ED, inpatient units, and outpatient clinics. Converting a test ordinarily processed in the central lab to POCT, because the patient specimen no longer has to travel to, or await processing by, the central laboratory, may reduce the time required for diagnosis and treatment (i.e., service time). Deo and Sohoni (2011), approaching POCT adoption from a public policy perspective, examine, in a context in which resources are limited, which facilities in a network ought to adopt. 18 Our work is closest to studies in emergency medicine that evaluate POCT as an alternative to the traditional diagnostic process that relies on a central laboratory. Fermann and Suyama (2002), reviewing the emergency medicine literature on POCT, find results on the impact of adoption on the reduction of patient length of stay (LOS), being the total time a patient spends in the ED, to be inconclusive. Whereas Murray et al. (1999) find, among discharged patients who received POCT, a reduction in LOS of nearly an hour, but no decrease for similarly treated patients who were admitted, Kendall et al. (1998) and Renaud et al. (2008) find no reduction in LOS associated with the adoption of POCT. Lee-Lewandrowski et al. (2003), having observed no change in LOS from shifting any single test, argue that any meaningful reduction in LOS resulting from POCT adoption relies on shifting multiple tests to bedside. Our study differs from this previous work in the following ways. One, we do not conduct a prospective, randomized control trial to measure the differential impact of POCT versus central lab testing on patients who experience the former. Such studies rely on the notion of appropriate sampling to control for plausible alternative explanations like ED congestion and patient complexity. We instead perform a retrospective analysis of the impact of POCT among patients who are tested, observing as well patients who are not tested and explicitly control for other factors through the use of propensity score analysis (as detailed in Section 5.1). Moreover, whereas previous studies examine primarily changes in patient length of stay, we divide LOS into two parts, waiting time and service time, and use queuing theory to generate separate hypotheses on the expected impact of POCT adoption on each of these elements. In the emergency medicine literature, our study most closely resembles that of Parvin et al. (1996), who evaluate the adoption of POCT for a specific test and compare a period of adoption to a control period during which patient samples are processed by a central lab. They describe the impact of POCT on different patient populations (e.g., high complexity versus low complexity) by dividing the observed sample into subgroups according to such factors as presenting condition or disposition. We control for factors like patient complexity through regression and propensity score analysis. Although Parvin et al. (1996) evaluate both test and no-test patients, they document for neither patient population during their short, five-week window of observation any substantial difference in LOS resulting from POCT. We compare ED performance metrics during a four-month window post-introduction of POCT to three time periods during which only central lab processing was used, (1) the four-month window immediately preceding adoption of POCT, (2) the four months of the previous year that correspond to the POCT post-introduction period, and (3) the prior year fourth-month window that corresponds to the POCT pre-adoption period. We observe, after controlling for factors like ED congestion and patient 19 complexity, differences in patient service time, waiting time, and service quality, and find that POCT adoption not only benefits the patients who are tested, but also has important spillover effects for other patients concurrently in the ED. In sum, our research makes unique contributions at the intersection of the ED and OM literatures. Being the first study of POCT rooted in a formal model that explains the source of potential gains or penalties in operational performance, ours is among the few papers in emergency medicine that employs the structure of a queuing model to make predictions within the ED, and among the few papers within OM to empirically test such predictions. Methodologically, ours is, to the best of our knowledge, the first study in the emergency medicine literature to apply propensity score analysis and use difference-in-differences to proxy for a randomized control trial. This methodological approach enables us to empirically test system-level impacts on the entire patient population (i.e., patients that receive POCT as well as those that do not), and thereby measure as well as the main effect of POCT adoption on test patients any spillover effects of adoption on other patients. 2.3. Hypothesis Formulation In this section, we formalize our research questions by analyzing POCT from the perspective of a stylized, but data-independent, queuing model of the ED patient flow process (see Figure 1). Arriving patients receive a bed assignment Priority based on an initial assessment of severity, then Wait a randomly distributed amount of time until space becomes available and they receive a Bed Assignment. Patients assigned a bed spend a randomly distributed amount of time in Service, which time includes all activities related to diagnosis, treatment, and, if required, boarding. The Disposition of treated patients is either discharge from the ED or admission to the hospital. We incorporate POCT by assuming a single diagnostic test to be a candidate for conversion from central lab to bedside processing. This gives two types of patients: Test patients, whose candidate tests are processed by the central lab (control period) or at the point of care (study period); No Test patients given any number of non-candidate tests that are processed by the central lab. Note that, independent of POCT, Test patients may also have non-candidate tests ordered and thus experience central lab delays that will form part of their service times. Service Wait Bed Assignment Priority Disposition Waiting Time Service Time Length of Stay 20 Figure 2-1: Stylized Emergency Department Patient Flow Process Formally, for a fixed number of beds (n) we can model patient flow by a G/G/n:NPP (Non-Preemptive Priority) queuing system. Patient type t, t=1, 2 denotes Test and No Test, respectively. Priority class i, i=1, 2, … K denotes a patient’s priority, where i = 1 is the highest, and K the lowest, priority, l it is the arrival rate for patients of type t in priority class i, and the mean and second moment of service time for patient type t in priority class i are denoted by mit and m it (2), respectively. Utilization for patients of type t in priority class i is thus defined by rit = litmit/n. Note that, because such a model assumes a stationary system, we study subperiods of the data in which stationarity may be reasonably assumed. We use this model to generate hypotheses about the impact of POCT testing on service time, service quality, and waiting time for both test and no-test patients. 2.3.1. Service Time Two countervailing forces are at play with respect to the impact of adoption of POCT on service time. Shifting a single test from central lab processing to POCT can decrease service time by eliminating transport time and central lab congestion effects. But POCT could also increase service time by increasing the point of care workload. Assuming, as is standard for most queuing models, independence of service times, the average impact on Test patients of a shift from central lab to point-of-care processing is primarily captured by the change in mi1 for any class i, and suggests the following hypothesis. H1a. In a stationary system, patients who receive a candidate test that is processed at the point of care experience shorter expected service times than those who receive the same test processed by a central lab. Under the G/G/n:NPP approximation, we assume stationarity, and that staffing is sufficient such that the number of beds (n) constitutes the primary bottleneck.2 We thus would not expect conversion of one test from central lab processing to POCT to affect the service time of No Test patients (t = 2) for patients of any priority class. Hence, we hypothesize as follows. H1b. In a stationary system, patients who do not receive a candidate test experience the same service times before and after the introduction of POCT. 2 If staffing is a binding constraint, the added burden of bedside test processing may result in shifting resources away from No Test patients. 21 2.3.2. Service Quality With POCT, as with any process redesign, hospital administrators expect the quality of care to be at least maintained, if not improved. In fact, improving service quality is often a key motivation for adopting POCT, as patients are expected to receive lab results more quickly and providers begin their therapeutic courses sooner, with a concomitant improvement in patient outcomes (Jahn and Aken 2003). Kc and Terwiesch (2009), however, empirically demonstrate that in some hospital settings service time reduction may be negatively associated with service quality. Specifically, they show, in a study of cardiothoracic surgery patients, that the decreased service time associated with early discharge results in higher rates of mortality. In a different context, Oliva and Sterman (2001) demonstrate a negative relationship between speed and quality. For operational improvements that accrue to the introduction of POCT to satisfy the paramount health objective, we frame our null hypothesis as follows. H2. Converting a candidate test from central lab processing to POCT does not reduce the quality of care delivered by the ED. 2.3.3. Waiting Time Intuitively, service time reduction to a patient in priority class i should at least affect the waiting time of all patients in the same class, regardless of type, because a server (bed) becomes available more quickly. Formally, we assume that within a priority class i all patients are treated on a first-come, first-served (FCFS) basis independent of type (test or no-test). We define the nominal utilization for priority class i as 𝜌!=𝜌!"∀!!. Let r(i) = r1 +r2 +…+ri be the utilization for priority class i across all types of the same and higher priority, and W(i) be the waiting time (delay) for a patient with priority i. We use the approximation for waits in such systems given on page 88 of Buzacott and Shanthikumar (1993), as follows: E𝑾(𝒊) 𝑮/𝑮/𝒏:𝑵𝑷𝑷 ≈ E𝑾(𝒊) 𝑴/𝑮/𝟏:𝑵𝑷𝑷 E𝑾(𝒊) 𝑴/𝑮/𝟏:𝑭𝑪𝑭𝑺 E𝑾(𝒊) 𝑮/𝑮/𝒏:𝑭𝑪𝑭𝑺 ≈E𝑾(𝒊) 𝑴/𝑮/𝟏:𝑵𝑷𝑷 𝒄𝒂 𝟐!𝒄𝒔 𝟐 𝟏!𝒄𝒔 𝟐𝒏 (1) where the second inequality follows by using the simple heavy-traffic approximation for delays in 𝐺/𝐺/ 𝑛:𝐹𝐶𝐹𝑆 queues (e.g., Whitt 1993, eqn (2.13)), ca is the coefficient of variation for the arrival distribution across all types and classes, and cs is the coefficient of variation for the service distribution across all types and classes. Our assumption that the overall variability ratio 𝑐! !+𝑐! !/1+𝑐! ! in (1) is not significantly affected by the 22 implementation of POCT will be the case if arrivals are approximately Poisson, which would imply that ca is approximately equal to 1 and the ratio is therefore also approximately equal to 1. Now E𝑾(𝒊) 𝑴/𝑮/𝟏:𝑵𝑷𝑷 = 𝝀𝒋𝟏𝒎𝒋𝟏 (𝟐) 𝑲 𝒋!𝟏!𝝀𝒋𝟐𝒎𝒋𝟐 (𝟐) 𝟐𝟏!𝝆(𝒊)𝟏!𝝆(𝒊!𝟏) (2) (e.g., Kleinrock 1976, eqn (3.31)). Note that under this equation, delays in queue for some priority class i are proportional to 1/((1-r1-r2-…-ri)(1-r1-r2-…-ri-1)). The direct effect of introducing a POCT test being a change in mean service time for test patients, for any class i where Test patients are present, the resulting change in ri and expected waiting time E[W(i)] is the same for both Test and No Test patients. This suggests the following hypothesis. H3a. Within the same patient priority class, Test and No Test patients experience the same decrease in waiting time following the introduction of POCT for Test patients. More significant, note that a change in waiting time is also experienced by patients of any priority class j >= i. Intuitively, waiting time changes from Test patients in priority class i trickle down to all lower priority patients because the higher priority queues clear faster. Where test patients appear in multiple classes, we would expect waiting time impact to accumulate across classes. The greatest change should thus accrue to patients in the lowest priority class. This suggests the following hypothesis. H3b. Assuming that test patients appear in multiple classes, we expect a greater reduction in waiting time to be experienced by lower priority (class j >= i) than by higher priority (class i) patients. 2.4. Research Context 2.4.1. Research Setting Our research site is an academic, urban, tertiary US hospital ED with an annual census of 55,000 patients and capacity of 47 beds. In 2008, the hospital ED decided to convert all cardiac troponin test processing from the central lab to point-of-care (i.e., bedside). Cardiac troponins in a patient’s blood reliably indicate heart muscle damage and the need for critical cardiac care within the ED. The standard of care in our research setting stipulates that the troponin test be ordered for every patient who presents to the ED with chest pain. According to the 2007 National Health Statistics Report (Niska et al. 2010), chest pain is the second most frequent complaint among patients between the ages of 15 and 64 who present to the ED. Among the ED patients observed, 16.3% required a troponin test. 29 Max 47,286 365,704 1 39,417 321,125 1 47,286 365,704 1 Tables 2-5 and 2-6, in presenting our descriptive statistics for test and no-test patients, respectively, across the four time periods of interest to us most closely resemble the empirical strategy employed in Section 5.0. We divide the data into four groups: Month1 to Month4 2007, Month1 to Month4 2008, Month5 to Month8 2007, and Month5 to Month8 2008 (the period during which POCT was adopted). We observe that, among test patients, average waiting time is nearly 9.7% lower during the period of bedside testing compared to the average of all other periods. Similarly, among test patients, service time is nearly 22.3% lower during the period of bedside testing compared to the average of all other periods. We observe the bounceback rate to be fairly stable among test patients across all periods. Table 2-5: Descriptive Statistics for Unmatched Data (TEST Patients) Month1-4, 07 (lab processed) Month1-4, 08 (lab processed) Month5-8, 07 (lab processed) Month5-8, 08 (bedside) 2,594 obs. 2,696 obs. 2,378 obs. 2,012 obs. Variable Mean σ Mean σ Mean σ Mean σ Waiting (Sec) 3,367.7 4,630.3 3,600.1 5,138.9 3,264.6 4,064.0 3,088.0 4,094.4 Service (Sec) 26,848.9 24,528.8 29,187.1 26,294.8 26,511.6 23,421.8 21,418.2 13,828.1 Bounceback 1.2% 1.1% 1.3% 1.1% 1.2% 1.1% 1.3% 1.1% For no-test patients, average waiting time after conversion to POCT is 8.6% lower than the average of the other three periods. Average service time among no-test patients, however, is 2% higher after conversion to POCT, without controlling for differences in patient population or system characteristics. The proportion of patients who bounce back during the period following adoption of POCT is 0.16% lower than during the other periods combined. Table 2-6: Descriptive Statistics for Unmatched Data for (NO-TEST Patients) Month1-4, 07 Month1-4, 08 Month5-8, 07 Month5-8, 08 12,150 obs. 12,411 obs. 12,144 obs. 13,032 obs. Variable Mean σ Mean σ Mean σ Mean σ Waiting (Sec) 4,846.1 5,266.5 5,425.9 6,164. 5,243.285 5,308.6 4,729.0 5,362.6 Service (Sec) 15,888.8 16,188.1 17,332.2 18,300.6 17,132.4 17,804.6 17,120.9 16,961.6 Bounceback 2.3% 1.5% 2.4% 1.5% 2.9% 1.7% 2.4% 1.5% Patient gender, age, and disposition and number of orders are, as noted above, among the control variables included in our analysis. Our patients are mostly female (59%), average age, 42 years, most often discharged (73%), and associated with 7.85 tests during their stay. We interpret our results for a discharged female in Section 5.0. On average, the system is hosting 47 patients (waiting or in service) when a patient presents to the ED. 30 2.5. Estimation and Results We test our hypotheses using an application of the difference-in-differences approach to compare SERVICE, WAITING, and BOUNCEBACK among different patient populations (test and no-test) before and after adoption of POCT. We employ propensity score analysis, a widely used technique, to construct a control group of test and no-test patients similar on key dimensions (i.e., patient and system characteristics) to the group of test and no-test patients who arrive during the POCT period. There are for both patient populations essentially four clusters of data: pre period 2007, post period 2007, pre period 2008, and post period 2008. The ED adopted POCT only in the post period 2008 quarter. To establish whether POCT had an impact on SERVICE, WAITING, or BOUNCEBACK, we could simply compare the months before and after adoption, but given the well-documented trend in ED usage and overcrowding, we used data from the same months in the prior year as a reference. This enables us to compare SERVICE, WAITING, and BOUNCEBACK preand post-POCT adoption, while controlling for any observed year-over-year changes. Our expectation, given that we have controlled for observed hospital-specific effects common to test and no-test patients, is that any incremental differences in SERVICE, WAITING, and BOUNCEBACK beyond those found when comparing pre-period 2007 to pre-period 2008 can be attributed to the adoption of POCT. 2.5.1. Propensity Score Analysis and Generation of Quasicontrol Group We identify quasi-control groups of test patients and no-test patients by matching test patients arriving in a given month in 2007 to test patients arriving in the same month in 2008, and no-test patients arriving in a given month in 2007 to no-test patients arriving in the same month in 2008. We match based on propensity score, which is the probability of ED arrival conditional on covariates (i.e., patient and system characteristics) for each patient population (test and no-test), and generate the probabilities for each patient within the population of test (no-test) patients using a logit model that predicts the probability of ED arrival controlling for patient AGE, GENDER, NUM_ORDERS, and DISCHARGE and such system characteristics as PEAK, MON-SUN, WAITING, SERVICE, and TOTAL (Becker and Ichino 2002). WAITING was not included as a covariate to generate our matched paired samples for the testing of any hypotheses in which WAITING was the outcome 31 variable, nor SERVICE included as a covariate to generate our matched paired samples for the testing of any hypotheses in which SERVICE was the outcome variable. After estimating from our logit models a propensity score for each patient, we apply one-to-one nearestneighbor matching (i.e., identify two patients with the closest propensity score) without replacement in order to pair each 2008 Test (No Test) patient with a 2007 Test (No Test) patient whose ED visit occurs in the same month (Leuven and Sianesi 2003). One-to-one nearest neighbor matching is preferred here because it matches each treated unit with only one unique control unit. Becker and Ichino (2002) observe that no one matching method is necessarily better than any other. Figure 3 depicts our matching process. Month/ Year Mar Apr May Jun Jul Aug Sept Oct 2007 T N T N T N T N T N T N T N T N 2008 T N T N T N T N POCT N POCT N POCT N POCT N Figure 2-3: Matching Process Note that descriptive statistics for our matched samples and unmatched data are similar, only a small proportion of data (3.3%) having been dropped during the matching process. We use our selected matched pair samples to test our hypotheses on service time, waiting time, and service quality. We estimate our regression model on our matched pair sample with the inclusion of controls, and present our results. For robustness, we implemented the alternative approach of estimating our model using the full sample of observations (results are available upon request). Our results differ slightly with the approach employed, but our overall conclusion regarding the impact of POCT on SERVICE, WAITING, and BOUNCEBACK remains unchanged. Table 2-7 describes the quasi-control groups generated for each hypothesis tested. Table 2-7: Datasets Used for Hypothesis Testing Hypothesis H1a H1b H2 H3a, H3b Patient included in the sample Dataset with just the matched test patients Dataset with just the matched no-test patients Union of the matched test patient dataset with the matched no-test patient dataset Union of the matched test patient dataset with the matched no-test patient dataset 32 2.5.2. Service Time We hypothesize service time to differ among test patients upon adoption of the test (H1a) and to remain the same among no-test patients (H1b). To test this hypothesis, we estimate on our matched pair sample of test patients for H1a and no-test patients for H1b the following model: 𝑆𝐸𝑅𝑉𝐼𝐶𝐸 =𝛽!+𝛽!𝑌+𝛽!𝑃+𝛽!𝑃𝑂𝐶𝑇 +𝜷𝟒𝑿𝒊+𝜷𝟓𝒁𝒊 (3) where 𝑋! is the set of system level control variables that includes MON-SUN, MONTH1-8, TOTAL, and WAITING and 𝑍! is the set of patient characteristics that includes AGE, GENDER, NUM_ORDERS, and DISCHARGE. The coefficient of interest is 𝛽!. Because the queuing model assumes stationarity, we estimate the model twice for H1a and twice for H1b. Specifically, we estimate it for test patients (H1a) and no-test patients (H1b) who arrive during peak (i.e., PEAK=1) and nonpeak hours. This enables us to assess whether the impact of POCT adoption differs during different periods of ED staffing and crowding.3 Our estimation results are reported in Table 2-8, columns 1-4 (we omit in all tables coefficients of control variables Month1-Month8, Mon-Sun, Age, and Gender). In support of H1a, we find the POCT to be negative and significant in both peak (𝛽!= -5,251.94, t= -3.97) and off-peak periods (𝛽!= -4,204.75, t= -3.92). POCT adoption is associated with a statistically significant reduction in service time during peak hours of, on average, approximately 87 minutes (=5251.94/60) or 1.45 hours. Peak hour service time for a typical test patient4 in 2007 March-July is 6.65 hours and in July-Nov 7.19 hours. Service time for a typical patient in 2008 March-July is 6.82 hours and in July-Nov 5.90 hours. This equates to a 21.6% drop in service time during the peak period.5 The percentage change during off-peak periods is, by a similar calculation, 16.5%, using predicted service time for off-peak periods (not presented here). 3 We purposefully created subsamples consisting of peak and non-peak periods in order to run our models. Although we could have presented a single model that included a dichotomous variable defining peak and nonpeak that we interacted with POCT, we found interpreting this interaction to be more complicated than analyzing each period separately. The two approaches nevertheless yield equivalent results. 4 We define a typical patient as female, discharged, of average age (42.56), and associated with the average number of orders (7.41) who arrives at the ED on an average day of the week in an average month during which there are, on average, 44.80 other patients in the system. We do not use the overall means in the sample presented in Section 4, but rather the mean prior to the adoption of POCT (i.e., the variable means from the March-July 2007 period). 5 (5.9-6.82)/6.82-(7.19-6.65)/6.65= -21.6%. 33 Our results fail to support H1b for peak or off-peak periods. Among no-test patients, we observe a statistically significant decline in service time (𝛽!= -877.32, t= -2.05, peak; 𝛽!= -630.62, t= -1.91, off-peak). On average, POCT adoption is associated with a 10-15 minute decrease in service time among no-test patients (15 minutes for peak, 10 minutes for off-peak). This suggests unexpected spillover effects into other patient populations. Such spillover effects from the adoption of POCT have heretofore not been discussed in the literature. We observe for a typical no-test patient a reduction in service time during peak hours of 4.7% and during off-peak hours of 3.1%. We discuss this unexpected result in Section 6. Table 2-8: Regression Results for H1a, H1b, and H2 H1a-peak H1a-off-peak H1b-peak H1b-off-peak H2 Variables Service Service Service Service Bounceback Test patients only No-test patients only All patients Y 606.09 -706.80 -122.62 -805.06*** 0.17** (953.54) (834.27) (313.66) (229.94) (0.08) P 2,192.35* 888.72 1,025.81** 102.02 0.24* (1,295.31) (1,189.64) (467.16) (363.91) (0.13) POCT(Y*P) -5,251.94*** -4,204.75*** -877.32** -630.62* -0.32*** (1,324.01) (1,072.77) (428.06) (330.67) (0.11) Waiting 0.30*** 0.36*** 0.07*** -0.01 0.04 (0.08) (0.08) (0.02) (0.02) (0.00) Service 0.13*** (0.00) Total 50.83 -40.00 50.41*** 7.66 -0.01*** (44.42) (25.06) (13.25) (7.78) (0.00) Num_orders 1,361.51*** 1,642.55*** 1,732.78*** 1,804.74*** -0.04*** (78.97) (73.89) (49.17) (39.44) (0.01) Peak 0.11* (0.07) Discharge 12,554.48*** 13,693.82*** 4,537.38*** 5,565.01*** -0.19** (1,049.85) (818.45) (522.60) (420.62) (0.09) Constant -6,587.40** -2,289.05 -3,063.48*** -2,311.78*** -2.93*** (2,951.94) (2,019.15) (1,022.97) (700.43) (0.20) Obs 3,465 5,661 18,160 30,168 57,428 R-squared 0.22 0.25 0.30 0.31 -2ll -6137.758*** Robust standard errors in parentheses (*** p<0.01, ** p<0.05, * p<0.1) 2.5.3. Service Quality We hypothesize that service quality does not decline upon adoption of POCT (H2). Being concerned with service quality for all patients, test and no-test, we estimate on the combined matched pair samples (both test and no-test patients) the following logistic model: 𝐿𝑜𝑔𝑖𝑡(𝐵𝑂𝑈𝑁𝐶𝐸𝐵𝐴𝐶𝐾)=𝛽!+𝛽!𝑌+𝛽!𝑃+𝛽!𝑃𝑂𝐶𝑇 +𝜷𝟒𝑿𝒊+𝜷𝟓𝒁𝒊 (4) where 𝑋! and 𝑍! include all control variables in equation (3) as well as SERVICE. Again, 𝛽! is the coefficient of interest. 34 We find support for our hypothesis (Table 2-8, column 5, 𝛽!= -0.32, z= -2.80, dividing both waiting and service time by 10,000) in that we observe a statistically significant decline in bounceback rate post POCT adoption. It thus appears that the decrease in service time does not come at the expense of quality, and that the decline in service time among all patients results in higher quality care overall. The predicted probability for ED 72-hour revisit for a typical patient in 2007 March-July is 1.83%, in JulyNov 2.32%, in 2008 March-July 2.17%, and in July-Nov 2.01%. This equates to a 0.6%6 drop that can be associated with the adoption of POCT. 2.5.4. Waiting Time We hypothesize that test and no-test patients experience the same decrease in waiting time upon adoption of POCT, controlling for patient severity (H3a). We regress waiting time POCT for each priority class separately, controlling for patient and system characteristics. Utilizing our combined sampled of matched test and no-test patients, we estimate 𝑾𝐴𝐼𝑇𝐼𝑁𝐺 =𝛽!+𝛽!𝑌+𝛽!𝑃+𝛽!𝑇+𝛽!𝑌∙𝑃+𝛽!𝑃∙𝑇+𝛽!𝑌∙𝑇+𝛽!𝑃𝑂𝐶𝑇 +𝜷𝟖𝑿𝒊+𝜷𝟗𝒁𝒊 (5) where 𝑋! is the set of system level control variables that includes MON-SUN, MONTH1-8, and TOTAL and 𝑍! the set of patient characteristic control variables that includes AGE, GENDER, NUM_ORDERS, and DISCHARGE. Our coefficient of interest is 𝛽!. Tables 2-9 and 2-10 present the results of our test of hypothesis H3a for peak and off-peak hours, respectively. Supporting H3a, we find that the impact of POCT adoption on waiting time does not differ with test status (the coefficient of POCT is insignificant) during peak or off-peak hours. This is true across each of the four priority classes. We further observe that the main effect of POCT (Y*P) is insignificant in all but one of the priority classes, PRIORITY4, during peak hours. In other words, during peak hours POCT has a statistically significant impact on waiting time only among low priority patients. 6 Essentially, our approach compares the mean of each outcome variable between preand post-treatment for patient visits during 2008 and subtracts the difference between preand post-mean outcome variables for patient visits during 2007 (Card and Krueger 1994). We derive the percentage difference by calculating the difference in the predicted probability of a typical patient from July-Nov 2007 and March-July 2007 (=0.48%) and subtracting this value from the difference in the predicted probability of a typical patient from July-Nov 2008 and March-July 2008 (=-.16%). We find the percentage change associated with the adoption of POCT to be 0.48%-(-0.16%)=0.64%. 35 Table 2-9: Regression Results for H3a at Peak Hours Priority1 Priority2 Priority3 Priority4 Variables Waiting Waiting Waiting Waiting Y -804.15*** -227.59 -921.49*** -71.11 (269.66) (231.68) (246.72) (234.64) P 471.48 508.88* 979.14*** 1,003.67*** (351.26) (308.01) (333.07) (327.17) T -1,395.83*** -227.86 -1,560.44* 2,343.86* (279.83) (297.33) (827.68) (1,282.17) Y*P -149.83 -335.92 -295.53 -1,738.75*** (369.75) (319.31) (338.24) (323.95) P*T -312.09 -441.95 2,716.27* -1,817.51 (381.04) (415.32) (1,589.19) (1,326.20) Y*T 52.63 -575.00 661.73 -282.77 (382.78) (472.33) (1,816.51) (2,899.21) POCT (Y*P*T) 425.96 17.46 -613.05 781.76 (538.15) (647.12) (2,621.41) (3,394.67) Total 173.20*** 218.75*** 222.42*** 191.48*** (8.84) (8.52) (10.14) (9.60) Num_orders -55.50*** 1.67 78.18 373.04** (11.86) (43.80) (104.63) (158.61) Discharge 1,677.94*** 1,603.33*** 1,768.75*** 3,115.69*** (176.58) (169.01) (363.09) (436.07) Constant -3,330.54*** -6,792.15*** -6,784.73*** -7,678.29*** (537.56) (633.30) (819.69) (729.91) Obs 5,388 6,201 5,196 4,923 R-squared 0.20 0.21 0.16 0.13 Robust standard errors in parentheses (*** p<0.01, ** p<0.05, * p<0.1) Table 2-10: Regression Results for H3a during Off-peak Hours Priority1 Priority2 Priority3 Priority4 Variables Waiting Waiting Waiting Waiting Y -818.96*** -219.99 -341.19** -3.62 (180.54) (160.79) (143.20) (126.53) P 116.91 924.14*** 533.81** 995.16*** (225.29) (200.33) (209.06) (179.94) T -878.66*** -317.93 -375.09 -323.47 (179.12) (217.27) (634.03) (1,426.96) Y*P -152.79 -777.26*** -382.58** -903.38*** (238.37) (207.95) (194.46) (171.82) P*T 61.56 -169.33 -1,041.10 386.34 (245.80) (279.80) (743.76) (2,175.94) Y*T 354.71 -165.18 -424.39 -233.80 (248.91) (319.79) (1,033.47) (2,110.99) POCT (Y*P*T) 249.26 516.29 1,742.73 922.89 (338.40) (416.86) (1,240.41) (3,332.05) Total 128.17*** 170.88*** 190.28*** 168.44*** (4.26) (4.57) (4.77) (4.21) Num_orders -23.62*** -43.73 97.63 165.96* (7.90) (28.93) (60.87) (85.66) Discharge 982.62*** 968.00*** 1,732.80*** 2,057.80*** (105.86) (107.52) (211.36) (274.02) Constant -1,982.90*** -4,549.81*** -6,550.27*** -5,887.82*** (297.04) (387.37) (448.11) (408.31) 36 Obs 8,632 9,853 8,503 8,738 R-squared 0.17 0.22 0.23 0.22 These results differ slightly for off-peak hours, during which POCT adoption is associated with a statistically significant reduction in waiting time across all priority classes, excepting the most severe patients. The impact of POCT on waiting time seeming to differ depending on priority class, we formally test in H3b an interaction between POCT adoption and patient severity. Our second waiting time hypothesis (H3b) expects a greater reduction in waiting time to be experienced by lower than by higher priority patients, independent of test status.7 To test this hypothesis, we estimate, using our combined sample of test and no-test patients, the following regression model: 𝑊𝐴𝐼𝑇𝐼𝑁𝐺 =𝛽!+𝛽!𝑌+𝛽!𝑃+𝛽!𝑌𝑃 +𝜷𝟑𝑷𝑹𝑰𝑶𝑹𝑰𝑻𝒀𝒊+𝜷𝟓𝑷𝑹𝑰𝑶𝑹𝑰𝑻𝒀𝒊∙𝑷+𝜷𝟔𝑷𝑹𝑰𝑶𝑹𝑰𝑻𝒀𝒊∙𝒀+ 𝜷𝟕𝑷𝑶𝑪𝑻 +𝜷𝟖𝑿𝒊+𝜷𝟗𝒁𝒊 (6) where 𝑋! is same as in equation (4) and 𝑍! includes AGE, GENDER, and DISCHARGE. A significant coefficient on 𝛽! indicates that POCT adoption has a differential impact on waiting time that depends on patient priority. We use PROIRITY1 patients, the most severe, as our base case for comparison. Table 2-11 presents our results from the regression. We find support for H3b, as, collectively, the interaction terms between the time period (P), year (Y), and PRIORITY (2-4) variables are statistically significant during both peak and off-peak hours (joint F-test p-value of 0.0001 during peak hours and 0.0027 during off-peak hours). This means that the impact of POCT on waiting time varies with patient priority class. During peak hours, we find the impact of POCT on waiting times to not be statistically different from one another among priority classes 1-3 (Table 2-12), but to differ substantially for the least severe patients (PRIORITY4). As highlighted in Table 2-13, POCT is associated, for the typical patient, with a nearly 35% reduction in waiting time (~30 minutes) for the least severe patients. Post POCT adoption, we observe an approximate 6% reduction in waiting time among PRIORITY3, and approximate 7% reduction in waiting time among PRIORITY2, patients, both of which equate, in absolute terms, to approximately six minutes. The typical PRIORITY1 patient, however, experiences a slight, and in absolute terms relatively small in magnitude, increase in waiting time of 2% (~ 2 minutes). We do not have a strong explanation for this increase, but, given its small magnitude, question its importance. 7 This hypothesis assumes that we have test patients in multiple priority classes. As part of our initial analysis, we verified that we do, indeed, have test patients across all four priority classes. 37 Table 2-11: Regression Results for H3b H3b-peak H3b-off-peak Variables Waiting Waiting Y -977.35*** -842.65*** (195.58) (126.51) P 141.04 453.48*** (233.32) (147.27) Priority2 117.04 197.84 (192.44) (127.69) Priority3 423.06* 106.84 (215.88) (131.82) Priority4 -312.22 -74.56 (212.88) (124.45) Y*P 140.29 -6.59 (276.05) (173.66) Priority2*P 290.66 89.33 (269.18) (173.38) Priority3*P 767.84*** 252.39 (291.19) (175.47) Priority4*P 1,361.31*** 570.17*** (285.37) (164.86) Priority2*Y 737.87*** 623.82*** (279.25) (188.58) Priority3*Y 210.30 670.50*** (306.92) (190.82) Priority4*Y 892.70*** 850.58*** (300.62) (180.21) POCT2 (priority2*P*Y) -491.33 -688.45*** (392.57) (251.31) POCT3 (priority3*P*Y) -516.52 -427.14* (433.69) (259.62) POCT4(priority4*P*Y) -1,949.94*** -875.23*** (424.45) (245.35) Total 201.35*** 163.85*** (4.63) (2.24) Discharge 1,944.94*** 1,258.89*** (108.19) (66.20) Constant -6,009.26*** -4,583.96*** (292.47) (160.74) Obs 21,708 35,726 R-squared 0.18 0.21 Robust standard errors in parentheses (*** p<0.01, ** p<0.05, * p<0.1) Table 2-12: P-value for F-test between POCT2, POCT3 and POCT4 POCT2= POCT3 POCT 3= POCT4 POCT 2= POCT4 Peak 0.9538 0.0020 0.0006 Off-Peak 0.3236 0.0836 0.4564 Table 2-13: Predicted percentage change in waiting time for patients in each priority class Peak (%) Off-peak (%) Peak (sec) Off-peak (sec) Priority Class 1 2.42% -0.56% 140.29 -6.59 Priority Class 2 -6.52% -14.11% -351.04 -695.04 Priority Class 3 -5.88% -8.77% -376.24 -433.73 38 Priority Class 4 -35.54% -18.07% -1809.66 -881.82 2.6. Discussion and Future Research We provide evidence that ED process redesign, specifically the adoption of POCT, has a substantial and valuable impact on ED operations. POCT adoption reduces the service time not only of patients who receive the troponin test at the point of care, but also of other patients within the ED who do not require a troponin test. That we further observe the service time impacts of POCT adoption to be greater during peak than during offpeak hours among both test and no-test patients suggests that POCT can have a measurable impact on ED performance during crucial operating periods. POCT adoption is also associated with improved service quality. All patients who presented to the ED during the POCT pilot period experienced a lower bounceback rate than in the comparison periods. Some of this quality improvement may follow from the priority queue model predictions, whereby high and low priority patients experience differential impacts. Empirically, lower priority patients have a greater propensity for bounceback (Pham et al. 2011, White et al. 2011). We speculate that less severe patients may, a priori, receive less attention. As service time savings accrue, as with priority queuing, less severe patients stand to reap the greatest benefit from any surplus physician attention. Finally, we find adoption of POCT to have a positive effect on waiting time for both test and no-test patients. We observe the lowest priority patients to experience the greatest, and highest priority patients the smallest, decrease in waiting time upon adoption of POCT, supporting our prediction of ED behavior based on queuing theory. It further suggests that POCT leads to operational improvement not only through a direct impact on the service time of test patients, but also through an indirect impact on waiting time for all who present to the ED when POCT is in use. Our study admits several limitations. First, as our data are drawn from a single, urban, level 1 trauma center, generalizing our results to different institutions with different patient populations is not possible, even when converting the same test. Second, as with all research that combines analytical modeling with empirical testing, one can identify settings in which the modeling assumptions may not hold. For example, we follow the priority queue model and assume that beds are the primary bottleneck in this ED. If the nursing staff performing POCT is a bottleneck, then adopting POCT will exceed the bedside staffs’ workload capacity, compromising 45 The main contribution of our work is our extension to existing studies of preventive cancer screening in two areas, methodology and research focus. In the area of breast cancer screening, extensive review suggests that past studies mainly focus on the relationship between participation and three types of factors: 1) socialdemographic factors, 2) psychological factors, and 3) interventions. The first stream of study focuses on the personal characteristics of population, such as their education level, income level, race group, age, living condition, and use these information to predict the probability of participation. Psychological factors that have been studied include self-perceived level of risk of cancer and concern about pain. For example, Sutton et al. (1994) found that main difference between attenders and non-attenders in their prospective design survey are health related behavior and attitude, belief and intention. Aro et al. (1999) conducted interview on women who were invited to a first round mammography in Finland, and found that high risk group may be related to more frequent earlier mammogram and weekly breast self examination, although their conclusion is contrary to that of Sutton et al. (1994), in which the authors found that women with self perceived high and low risk of breast cancer have lower attendance rate than women with moderate amount of perceived risk. Contrary to normal belief that accessibility is the biggest barrier to breast screening participation, Kee et al. (1992) found that attitudes rather than access play the most important role in influencing uptake. In their interview with women invited for mammography and declined attendance, the most cited reasons are feelings of indifference or ignorance of screening issues and fear of pain or embarrassment, only less than 4% interviewed women expressed preference for more accessible screening unit. Munn (1993) and Rimer et al. (1989) found similar results in their interviews of women who declined participation of mammography invitations. However, these results can not be applied to women that have participated in the first round screening, but decided not to attend rescreening, where non-attitude related factors are more likely to be the main barrier. Research on breast screening interventions mainly uses random control trials to explore the impact on participation rate of certain type of interventions, such as physician recommendation, mobile mammography, media campaign, etc. This stream of literatures usually focuses on interventions directed at participants, few studies can be found that focus on intervention conducted at facility level. Our research method enables us to present a close-to-reality choice setting for current participants of breast cancer screening and force them to make trade-offs based on the priority they assign to each studied attribute of facility service. This is an area that has been largely overlooked in the past, but can have significant impact on the rescreening decision of existing patients. Efficiency of cancer screening depends heavily on the frequency of screening (Cohen et al. 2008), 46 therefore improving rescreening participate rate of these programs can have great impact on the health outcome of target population. 3.3. Methodology We use Stated Preference Discrete Choice Modeling (SPDCM) as the basic framework of our survey design. Preference data come in different forms (Louviere et al. 2000). Market data, or Revealed Preference (RP) data, are data obtained through market observations (Samuelson 1947). Stated Preference (SP) data, on the other hand, are choice responses from the same economic agents (eg. choosers), but draw in hypothetical markets. While RP data contain information about behavior of interest of current market, SP data are rich in attribute tradeoff information and therefore more useful for forecasting changes in behavior (Louviere et al. 2000). Discrete Choice Modeling (DCM) is the generation and analysis of choice data, using hypothetical markets that are constructed to suit relevant research question. In DCM survey, respondents are given a set (or several sets) of alternatives, each alternative is described by a set of attributes, and each attribute can take on one of several levels. Levels are ranges over which attributes vary across alternatives. Respondents make a discrete choice of yes or no for each alternative, e.g. whether to visit a hypothetical store or not, based on their evaluation of overall utility of each alternative. The resulting choice can then be used to estimate the contribution of each attribute/level to the overall utility (Lancsar and Louviere 2008). The SPDCM method is derived from McFadden’s (1974) random utility theory (RUT) and Lancaster’s economic theory of value (Lancaster 1966). Let Uij be the utility of jth alternative to ith individual, RUT posits that this utility can be decomposed into two parts, an systematics component Vij that is explainable, and εij, a random part component: Uij=Vij+ εij (1) Consumer will choose alternative j if the utility gained from j is greater than utility gained from other alternatives. Let Yij=1 if alternative j is chosen by respondent i, then: Pr(Yij=1)=Pr(Uij>Uik)=Pr(Vij+ εij > Vik+ εik )=Pr(Vij - Vik > εik - εij), for all k≠j (2) Lancaster’s economic theory of value suggests that a commodity can be decomposed into separable attributes, therefore allow for examination of preference of different choices by their attribute level, including those that are achievable yet not available (Gerard et al. 2003). Let Xij denote a vector of attribute of alternative j presented to respondent i, we have: Vij=βXij (3) 47 When β are the utility parameters, and εij can be assumed to be identically distributed (iid) with the Weibull distribution, the probability of choosing alternative i can be calculated as (McFadden 1974): 𝐏𝐫(𝐘𝐢𝐣 =𝟏|𝟏,𝟐,…𝑱)=𝐞𝐱𝐩!(𝜷𝑿𝒊𝒋) 𝐞𝐱𝐩!(𝜷𝑿𝒊𝒌) 𝑱 𝒌!𝟏 (4) where J is the total number of alternatives presented to respondent i. This is conditional logit model that are used in our data analysis. 3.4. Methods 3.4.1. Focus group meeting current facilities To ensure we can better understand the decision process of choosing screening unit, we conducted a focus group meeting that consists of five women. Among them four have breast screening experiences, either from government program or on their own, the other one shared experience from primary care facility selection. The meeting was voice-recorded and transcribed later. Two of the co-authors acted as moderators to facilitate the flow of discussion. We first gave an introduction about the research background, such as motivation and research setting, specifically, we stressed that this is for government-funded free screening. We then gave them few examples of type of factor that we are interested, such as traveling time, clinic opening hour, etc. The onehour meeting was very helpful in shaping out the scope of attributes to be considered in the selection process, and revealed some interesting aspects that fail to be captured during literature review. Topics covered including accessibility such as traveling time, mode and parking availability, staffs’ communication skill, and clinic’s physical surroundings and so on. Besides providing information about facility attributes, focus group members were also asked about what type of personal factors will affect their choice process, such as income, education level. Overall, focus group provided an interactive dynamic for developing, challenging, and refining ideas (Hamilton and Barlow 2003). One of the potential problems with SP data is it can be affected by the degree of `contextual realism' established for respondents (Louviere et al. 2000). When hypothetical settings of the alternatives are not achievable or realistic enough, respondents might not take the choice task seriously and therefore compromise the validity of the survey, as well as managerial implication generated from survey result. We therefore conducted phone survey with all current government designated screening centers in the studied area, and used 48 the data to finalize the attribute levels used in the survey. Data collected from current facilities include invitation process, accessibility by public transport, parking availability/price, opening hour, waiting time for appointment, waiting time inside clinic, and waiting time for screening result. Table 3-1 below lists our final attributes and attribute levels shaped by focus groups and clinic survey, name of the corresponding variables are in parentheses. Table 3-2 lists personal characteristics to be collected about survey respondents. Household income used 2009 government-reported median family income of the studied area as the dividing point. Table 3-1: Final attributes and levels Attributes Level Waiting time for appointment (WAITAPP) Less than 2 weeks [1] Between 2-6 weeks [2] Longer than 6 weeks [3] Travel time (TRAVEL) Less than 20 minutes [1] 21-40 minutes [2] Longer than 40 minutes [3] Parking availability (PARKING) Free parking on-street or clinic parking [1] On-street parking at $1.5 per hour [2] Off-street parking at $4.5 per hour [3] Clinic opening hours (HOUR) 8am-4pm weekday only [1] 8am-8pm weekday only [2] 9am-5pm weekday and 9am-1pm weekend [3] Waiting time inside clinic (WAITCLINIC) Less than 15 minutes [1] 16 – 30 minutes [2] 31-60 minutes [3] Nursing stuff (NURSE) Knowledgeable, capable of answering breast cancer/screening related questions [1] Provide information sheet, but does not answer questions [2] Provide no information sheet, does not answer questions [3] Screening process (SCREENING) Technician explain the process while doing the screening [1] Minimal amount of communication during the screening [2] Waiting time for result (WAITRESULT) Less than 4 days [1] 4-15 days [2] 16-30 days [3] Table 3-2: Respondents' characteristics to be collected Respondent Characteristics Level Age 50-55 [1] 56-60 [2] 60-69 [3] Education Level High school diploma or lower [1] Undergraduate degree [2] Graduate degree or higher [3] Household annual income Below $66,000 [1] 49 Above $66,000 [0] Past screening experiences First time screener [1] Regular screener [2] Irregular screener, please specify the time since your last screening [3] First language English [1] French [2] Others [3] Do you have a referral from a family physician for current visit Yes [1] No [0] Are you aware of the recommended rescreening frequency of once every two years for women after age 50 Yes [1] No [0] Are you knowledgeable of the risks of breast cancer for women of age 50 and above? Yes [1] No [0] Do you have family members, friends, or acquaintances diagnosed with breast cancer? Yes [1] No [0] 3.4.2. Survey design 3.4.2.1 QuestionnaireOrthogonal main effect Let J denote number of alternatives in a choice set, and Lj be the number of levels of attribute J, then the total number of alternatives (combination of alternative/level) is the full factorial: ∏𝒋!𝟏 𝑱𝑳𝒋 (5) Although a full factorial design allow the estimation of both main effects of each attribute and interaction effects between two or more attributes, it’s hard to implement in reality due to the fact that the resulting number grows exponentially as the number of attributes or attribute levels grows (Lancsar and Louviere 2008). Therefore when only main effects are of interest, many design use what’s called fractional factorial, where only a subset of the alternatives are used in the choice set. This subset, however, usually are not selected randomly, instead there exist a large range of sampling methods that lead to practical designs with the purpose of estimating the effects of interest as efficiently as possible (Louviere et al. 2000). We used orthogonal main effect design for designing our choice set to ensure zero correlation among attributes/levels, so that main effects of each attribute/level can be estimated independently and unbiased (Louviere et al. 2000), we are not interested in any interaction between our chosen attributes. We obtained our orthogonal array from an online orthogonal library where number of attributes and levels suits our data (2011). 50 3.4.2.2 Size of choice set Most empirical work in discrete choice model have used 1-16 choice tasks per person, although there’re reports of 32 and 64 choice (Louviere et al. 2000). The number is usually context specific and depends on many factors, such as minimal number of choice set needed to test orthogonal main effect design. There are evidence that as the number of attributes and levels increase, task complexity increase, and might lead to increased unobserved variability and reduce choice consistency (Lancsar and Louviere 2008; Louviere et al. 2008). Our initial choice set consists of 18 alternatives. Respondents are asked to make a yes or no decision for each alternative. Feedback from a pilot study of 5 women revealed that the number of alternative was too big. When this happens, after a few initial choice tasks, respondent will tend to start targeting on only one attribute that she considers the most important, and making Yes/No decisions solely on this one criteria. We subsequently divided the alternative into two groups of nine alternatives, questionnaire version one (V1) and two (V2). We randomly chose half of our respondents to answer version one, the other half version two to ensure balanced number of all 18 alternatives. Although larger number of choice tasks per respondent can help to “blow up” the total number of choice tasks, it has the risk of violating the IID assumption for conditional logit model we prefer, where choice decision is assumed to be independent of each choice. In reality, very often a respondent’s decision of one choice task might be affect by a previous choice task. In this view, the fewer choice tasks per respondent, the more confident we can be of our estimation of model parameters (Louviere et al. 2000). Therefore splitting alternatives into two groups is actually preferred from model estimation point of view. Table 3-3 below is an example of choice task. Respondents are free to choose any number of clinics to patronize from the nine alternatives. Table 3-3: Example of choice task Attributes Clinic 1 Waiting time for appointment Less than 2 weeks Travel time Less than 20 minutes Parking availability Free on-street parking or clinic parking Opening hours 8am - 4pm weekdays only Waiting time inside clinic Less than 15 minutes Nursing staff Knowledgeable, capable of answering breast cancer/screening related questions Screening process Technician explains the process while doing the screening Waiting time for result Less than 4 days Please indicate whether you will choose the clinic Yes No 51 3.4.3. Sample size The purpose of the DCM survey is to measure choice probability (or proportion) with a desired level of accuracy (Louviere et al. 2000). Let n is our target sample size, p be the actual (current) participate rate (choice probability) of studied program. If we want the estimated probability to be within a percent of the current value p with probability 𝛾 or greater, we can calculate minimal sample size as (Louviere et al. 2000): 𝒏≥𝒒 𝒑(𝒂 𝟏𝟎𝟎)𝟐𝜱!𝟏(𝟏!𝜸 𝟐) 𝟐 (6) where q equals to 1-p, and Φ-1(·) is the inverse cumulative standard normal distribution function (CDF). In our design, we set to achieve within 10% of the true participation rate of studied area with probability of 95% or greater, this gave us a minimal sample size of 460. This sample size, however, is the number of choice task. The real sample size in terms of number of returned questionnaires needed, give that each respondents will be performing 9 choice tasks, is 52. 3.5. Data 3.5.1. Background We conducted our survey in Montreal, Canada. The breast screening program in Montreal is part of The Québec Breast Cancer Screening Program (PQDCS). PQDCS was launched in 1998 to offer free breast cancer screening for women between 50 to 69 years old on a biannual basis. Every two year, an eligible woman receives an invitation letter sent by PQDCS. The letter also acts as prescription, hence referral from a personal physician is not necessary to participate in the program. There are a total of 15 designated screening center (CDD) in Montreal that are chosen by the program based on certain quality criteria, as well as to meet the requirement of geographical coverage. Participants are free to choose any of the 15 CDDs. Reported participate rate of Montreal has been very low at 45.5 as of 2010, below the province’s overall participation rate of 57.6% (2011), and even far below the targeted screening rate of 70% set by the Quebec government (Zhang et al. 2011). Our target population are all women in Montreal that are participating in PQDCS program. Given the fact that differences in average income, education and other demographic characteristics exist among residents of different regions, we try to cover as broadly geographically as possible. In an effort to do so, we approached all 15 CDDs and had 4 clinics agree to participate in the survey. By a pleasant coincidence, the final 4 clinics 52 happen to locate evenly across the city Montreal. Although we didn’t compare our final sample to the target population, this still gave us reasonable confidence that final sample can be moderately representative of the true demographics characteristics of target population. Our survey was conducted in all four clinics simultaneously for around two months. At each clinic, all current participants of PQDCS program that visit the clinic for breast screening were approached. Questionnaire V1 and V2 were given out interchangeably. Participants complete the survey at the clinics and return them before leaving. The complete questionnaire usually takes no more than 10 minutes. This includes reading the instruction, complete the choice task, and fill in demographic information. 3.5.2. Descriptive statistics Our final number of returned survey is 287. Among them, eight are unusable due to all Yes or all No answers to all 9 alternatives. This gave us a total of 278 usable questionnaires and 2502 choice responses, greatly exceeding our target sample size. Estimated overall response rate is …. There are 140 V1 and 138 V2 questionnaires. Median of total number of chosen clinics is 3. A summary of demographic characteristics of respondents is shown in table 3-4. Table 3-4: Summary of respondents' characteristics Respondent Characteristics Level Frequency (278 respondents) Age (AGE) 50-55 [1] 100 56-60 [2] 76 60-69 [3] 79 Education Level (EDU) High school diploma or lower [1] 107 Undergraduate degree [2] 74 Graduate degree or higher [3] 48 Household annual income (INC) Below $66,000 [0] 139 Above $66,000 [1] 105 Past screening experiences (EXP) First time screener [1] 20 Regular screener [2] 200 Irregular screener, please specify the time since your last screening [3] 31 First language (LAN) English [1] 44 French [2] 194 Others [3] 17 Do you have a referral from a family physician for current visit (REF) Yes [1] No [2] 148 106 Are you aware of the recommended rescreening frequency of once every two years for women after age 50 (FRE) Yes [1] No [2] 245 6 53 Are you knowledgeable of the risks of breast cancer for women of age 50 and above? (RISK) Yes [1] No [2] 245 12 Do you have family members, friends, or acquaintances diagnosed with breast cancer? (FAM) Yes [1] No [2] 156 94 Table 3-5 summarizes the level of attributes of chosen clinics. A clear and intuitive preference can be seen in all attributes except opening hour. Two sample binary t-test indicate differences in chosen levels are significant for all attributes at 95% confidence level, except for between TRAVEL1 and TRAVEL2, HOUR1 and HOUR3. Table 3-5: Summary of attribute level for chosen clinics Attributes Level Frequency WAITAPP Less than 2 weeks [1] 375 Between 2-6 weeks [2] 295 Longer than 6 weeks [3] 134 TRAVEL Less than 20 minutes [1] 324 21-40 minutes [2] 299 Longer than 40 minutes [3] 181 PARKING Free parking on-street or clinic parking [1] 326 On-street parking at $1.5 per hour [2] 293 Off-street parking at $4.5 per hour [3] 185 HOUR 8am-4pm weekday only [1] 250 8am-8pm weekday only [2] 302 9am-5pm weekday and 9am-1pm weekend [3] 252 WAITCLINIC Less than 15 minutes [1] 286 16 – 30 minutes [2] 332 31-60 minutes [3] 186 NURSE Knowledgeable, capable of answering breast cancer/screening related questions [1] 389 Provide information sheet, but does not answer questions [2] 248 Provide no information sheet, does not answer questions [3] 167 SCREENING Technician explain the process while doing the screening [1] 534 Minimal amount of communication during the screening [2] 270 WAITRESULT Less than 4 days [1] 327 4-15 days [2] 266 16-30 days [3] 211 3.6. Results 3.6.1. Numerical Result One requirement for using the conditional logit model, which is based on McFadden’s random utility theory, is the iid restriction on error term. The requirement of iid is to guarantee the independence of irrelevant alternatives (IIA) property on respondent’s choice. The IIA property states that the odds ratio of selecting one 54 alternative over another is independent of the number or presence of other alternatives. This is a very strict requirement which leads to the very few application of conditional logit in DCM surveys . We test this assumption on our data using the popular Hausman Specification Test (Louviere et al. 2000) and the STATA command “suest” . We first run a conditional logit model on the complete data set, call it the full model, then a partial model that excluded one of the alternatives. If the IIA property holds, the coefficients on attribute should not differ significantly. We randomly selected alternative 5 from V2 to be excluded in the partial model, test result suggests IIA property indeed holds (chi2=-1.32 for Hausman and p=0.1383 for suest). We divided our attributes into four groups and used hierarchical regression . The four groups are: Group 1: WAITAPP Group 2: TRAVEL, PARKING, HOUR Group 3: WAITCLINIC, NURSE, SCREENING Group 4: WAITRESULT WAITAPP indicates the level of crowding of the clinic. TRAVEL, PARKING, and HOUR are all related to the ease of accessibility on the day of screening. Group 3 variables are concerned with service quality. While some might expect waiting time for result depends on the congestion level like WAITAPP, our survey of CDDs implies no relationship between WAITAPP and WAITREULT. We therefore put WAITRESULT in a separate group. Table 3-6 summarizes the regression result for hierarchical regression. Both coefficient and odds ratio are shown in the table. Table 3-6: Hierarchical regression result ! Step!1! Step!2! Step!3! Step!4! Attributes! Coef.! O.R.! Coef.! O.R.! Coef.! O.R.! Coef.! O.R.! waitapp1' 1.54***' 4.69***' 1.57***' 4.81***' 1.97***' 7.18***' 1.8***' 6.06***' waitapp2' 1.12***' 3.07***' 1.12***' 3.06***' 1.45***' 4.25***' 1.36***' 3.91***' travel1' ' ' 0.86***' 2.35***' 1.30***' 3.68***' 1.24***' 3.44***' travel2' ' ' 0.74***' 2.10***' 1.09***' 2.98***' 1.00***' 2.73***' parking1' ' ' 1.54***' 4.68***' 1.30***' 3.67***' 1.61***' 5.01***' parking2' ' ' 1.04***' 2.82***' 0.51**' 1.67**' 0.81***' 2.26***' hour1' ' ' (0.4)***' 0.67***' (0.48)***' 0.62***' (0.29)' 0.75' hour2' ' ' (0.39)**' 0.68**' (0.09)' 0.92' (0.14)' 0.87' waitclinic1' ' ' ' ' 0.62***' 1.86***' 0.65***' 1.91***' waitclinic2' ' ' ' ' 0.40' 1.48' 0.42*' 1.53*' nurse1' ' ' ' ' 1.95***' 7.04***' 1.92***' 6.83***' nurse2' ' ' ' ' 0.94***' 2.56***' 0.76***' 2.13***' screening1' ' ' ' ' 1.39***' 4.03***' 1.3***' 3.67***' waitresult1' ' ' ' ' ' ' 0.86***' 2.36***' waitresult2' ' ' ' ' ' ' 0.56***' 1.75***' 61 sample binary t-test. P-value is for alternative hypothesis that there’s difference between the two groups in these social-demographic classes. We see that none of the p-values are small enough to reject the null hypothesis (same for p-values for other alternative hypothesis, e.g. group 1> group 2, and group 1< group 2), indicating there’s no evidence of significant differences between the two groups. We also run a logit model using group 1 membership as dependent variable, and all listed social-demographic variables as predictor, again found no predictors to be significant. Table 3-12 summarizes the result from logit prediction model. Table 3-11: Demographic characteristics’ comparison between two groups group 1 group 2 z-value p-value AGE 50-55 38,6 39,5 -0.1375 0.8906 56-60 29,5 29,9 -0.0655 0.9478 61-70 31,8 30,5 0.2100 0.8337 EDU Highs school diploma or lower 44,4 48,0 -0.5117 0.6089 Undergraduate degree 33,3 31,8 0.2439 0.8073 Graduate degree or higher 22,2 20,3 0.3470 0.7286 Below average income (INC) 60,9 54,8 -0.9282 0.3533 EXP First time screener 7,0 8,5 -0.4187 0.6754 Regular screener 82,6 78,2 0.8178 0.4135 Irregular screener 10,5 13,3 -0.6555 0.5122 LAN English 15,9 18,0 -0.4129 0.6797 French 76,1 76,0 0.0157 0.9874 Others 8,0 6,0 0.5985 0.5495 Has family doctor referral (REF) 60,7 57,0 0.5712 0.5679 Aware of recommended frequency of 2 years (FRE) 97,7 97,6 0.0486 0.9613 Aware of risk of breast cancer for women age 50 and above (RISK) 94,4 100,0 -0.4944 0.6210 Has family or friends diagnosed with breast cancer (FAM) 64,7 61,2 0.5402 0.5890 Table 3-12: Logit prediction model result Group 1 Coef. Std. Err. z P>z age1 -.2216667 .3781772 -0.59 0.558 age2 -.2289872 .3906771 -0.59 0.558 edu1 -.4385267 .4209581 -1.04 0.298 edu2 -.203273 .4257177 -0.48 0.633 inc -.2829239 .3269181 -0.87 0.387 exp1 .1246325 .7204419 0.17 0.863 exp2 .6044719 .4844607 1.25 0.212 lan1 -.5103114 .6820413 -0.75 0.454 lan2 -.1838424 .5904224 -0.31 0.756 ref .1294552 .3171549 0.41 0.683 fre .0096676 .9751531 0.01 0.992 62 risk -.7372943 .6704594 -1.10 0.271 fam .0720999 .3130271 0.23 0.818 _cons .236171 1341235,00 0.18 0.860 3.8. Simulation Analysis Our final step is to show how our analysis can be used in the close-to -reality scenario of Montreal breast cancer screening at a macro level, we use Arena simulation that incorporates real population data and econometrics analysis to illustrate how changing of attribute levels will affect the screening rate of current clinics. 3.8.1. Basic model Our basic model aims to simulate current screening situation in Montreal. To achieve that, we used data provided from PQDCS program office to separate Montreal into 12 population zones. Table 3-13 lists total eligible population (women 50-69 years old), current participation rate and number of participants for each population zone (as of December 2010). Table 3-13: Montreal population zone and corresponding participation rate Population!Zone! 1! 2! 3! 4! 5! 6! 7! 8! 9! 10! 11! 12! Participation!Rate! 50%' 38,3%' 47,3%' 50%' 45,6%' 48,1%' 47,9%' 38%' 47,1%' 41,3%' 48,1%' 43,4%' Eligible!Populatoin! 27117' 14556' 21072' 19923' 10441' 12446' 15265' 23495' 17287' 14727' 18055' 28543' #!of!Participants! 13559' 5575' 9967' 9962' 4761' 5987' 7312' 8928' 8142' 6082' 8684' 12388' The process of the basic model is as follows: Step 1: Participants arrival There’re a total of 12 entity modules, each correspondents to an arrival from a population zone. The total number of opening hour of current 15 CDDs is 2100 hours per year at the time of interview, not consider holidays. We therefore assume 250 working days per year and 8 working hours per day (2000 hours per year). Current participation rate is reported on a biannual basis, i.e. number of individual screening in the past 24 month. For each population zone, we assume possion arrival, and calculate the arrival interval as 4000 hours divided by eligible population. Step 2: Screening decision We have one decision module for each arrival, where a decision of whether to screen or not is made. This is a two-way by chance decision module, where participation rate for each zone is used as probability to screen. 63 Step 3: Clinic choice A second decision module follows each screening decision, where participants need to decide which clinic to visit. This is a 15-way by chance module, where participants from each zone are assigned 15 probabilities, represents probability of visiting each of the clinics, calculated using results from conditional logit model (column of step 4, table 3-6). We believe this is more realistic than sending participants to the one clinic with the highest probability, as in reality, people don’t always go to the choice with highest utility level due to different limitations. We try to mimic the true clinic configurations as closely as possible. We used interview data on WAITAPP, PARKING, HOUR, WAITCLINIC, and WAITRESULT for most clinics where available, and randomly assigned value on NURSE and SCREENING. For attribute TRAVEL, we generate a 12 by 15 matrix, represents travel time from each of the 12 population zone to each of the 15 CDDs, roughly calculated based on geographic distance. This means for each clinic, while 7 of the attributes are invariant across all population zones, TRAVEL varies for each clinic-population zone pair. Step 4: Waiting for appointment We add a HOLD module before each SCREENING module to hold participants until the queue in its corresponding SCREENING module reaches zero. This is a close proxy for the waiting time for appointment, although by doing so, our HOLD module also captures some waiting time inside the clinic. Currently all 15 CDDs in Montreal accept participants by appointment only. Screening time for mammography is fairly constant at 20-30 minutes, waiting time inside clinics are usually not very long, hardly goes over one hour. Therefore incorporating it into waiting time for appointment wouldn’t affect accuracy since the latter is usually measured by days, sometimes even months. Step 5: Screening We have 15 process module, each represent one screening clinic. We set service time for all clinics to be constant at 30 minutes. True number of mammography machines in each CDD is used in assigning resources. Clinic 5 has 3 machines, clinic 3, 8 and 12 each has 2 machines, the rest of clinics each has one machine only. Step 6: Disposition Participants leave clinic after screening. Figure 3 below presents the basic process of the Arena model. There is another dispose module after each screening decision module, represent non-participants, that we didn’t show here since it’s not of our interest. 64 Note there are 12 modules for each of Arrival, Screening decision, and Clinic choice, represent 12 population zones. Waiting for appointment, Screening, and Disposition each has 15 modules, represents 15 current CDDs. Figure 3-3: Simplified Arena model We set up Arena to fun for 2000 hours (one year) at 8 hours a day for 10 replications, and report the average of the 10 replications for the rest of the section. Our key performance metrics are number of screenings, waiting time for appointment, utilization rate, and total number of participants waiting for screening (participants with an appointment). Columns labeled Basic in Table 3-14 below presents result from our basic model. With current configuration, we see that participants are far from evenly distributed among 15 clinics. While 11 out of 15 clinics has zero days of waiting time of appointment, mostly with an utilization rate of below 25%, clinic 4 and 7 both have waiting time of almost 3 months, together holding almost 10000 people waiting in line for screening. Clinic 10 also has an average of 40 days waiting time for appointment. Clinic 3, 5, 8 and 12 all have multiple machines, therefore ideally their total number of screening should also be greater than the rest of the clinics, which is not the case in basic model. CDD2 and 12 have the lowest screening, with 115 screenings performed in CDD2 and as few as 22 in CDD12. At the time of our interview, 6 of current CDDs have waiting time for appointment of less than 1 week, another one can make next day appointment. 3 CDDs have waiting time of 2 to 3 weeks. Two CDDs have waiting time of 1 to 2 month, and the other 3 CDDs all have waiting time longer than 3 months. Although the attribute level of current clinics that we used in calculating probability distributions are not 100% accurate, our simulation result is nonetheless reflective of reality of unbalanced screening load among current CDDs. Our first step then is to adjust current system by changing some of the attribute levels. Table 3-14: Arena result from Basic model and Adjust1 model Total number of screening Waiting time for appointment Utilization rate Number of participants waiting for screening Clinics Basic Adjust1 Basic Adjust1 Basic Adjust1 Basic Adjust1 1 995 1092 0 0 24.88% 27.31% 0 0 2 115 401 0 0 2.87% 10.03% 0 0 3 1137 6293 0 0 14.21% 78.67% 0 1 4 4000 2597 90 0 100.00% 64.94% 5214 0 Arrival! Screening! decision! Clinic!choice! Waiting!for! appointment! Screening! Disposition! 65 5 1938 10538 0 0 16.15% 87.83% 0 2 6 743 817 0 0 18.56% 20.44% 0 0 7 4000 3740 85 0 100.00% 93.51% 4286 6 8 1397 4761 0 0 17.47% 59.53% 0 0 9 2655 2833 0 0 66.38% 70.83% 0 0 10 3999 3906 40 1 99.99% 97.67% 959 18 11 3986 3998 5 21 99.67% 99.95% 76 399 12 22 335 0 0 0.56% 8.37% 0 0 13 3197 3486 0 0 79.94% 87.16% 1 3 14 789 867 0 0 19.73% 21.67% 0 0 15 581 4161 0 0 7.27% 52.02% 0 0 Total 29554 49825 10537 429 3.8.2. Adjust1 model Our adjustment mainly focuses on three attributes, waiting time for appointment (WAITAPP), parking availability (PARKING), and nurses’ knowledge (NURSE). Table 3-15 below summarizes the change of attributes from Basic model to Adjust1 model. Note to make the screening load more balanced, on one hand we improved attributes level for some of the clinics, at the same time we try to make clinic 4, 7 less attractive by changing their waiting time for appointment longer than before. In reality, this can easily be accomplished by arbitrarily offering a much later appointment than necessary. Clinic 2, 12 and 15 all have a long waiting time for appointment. Although appointment time has a big impact on clinic screening rate, we avoid changing it to acknowledge the fact that this can’t be done in a short time, unless extra resources (mammography machines) are available. Table 3-15: Arena result from Basic model and Adjust1 model Attribute level WAITAPP PARKING NURSE Clinics Basic Adjust1 Basic Adjust1 Basic Adjust1 2 - - - - 1 2 3 - - 3 1 - - 4 1 3 - - - - 5 - - 3 1 - - 7 2 3 - - - - 8 - - - - 1 2 12 - - 2 1 1 3 15 - - - - 1 3 Columns labeled Adjust1 in table 3-14 show the key performance metrics for our adjusted model. The most notable improvement to the system is total number of screening improved by almost 70% from 29554 to close to 50000. Note the total eligible population is 222927, at the current 45% biannual screening rate, we 66 expect the system to performance around 50000 screenings (222927*45%*0.5). Only one out of all 15 clinics now have a queue for appointment (clinic 11). Utilization rate of most clinics have improved significantly for most previously under-used clinics, and number of participants waiting for screening is now only a small fraction of the number from basic model. Although our basic model didn’t reflect the current screening volume due to lack of information on true attribute level, our adjusted model demonstrates that current system is far from ideal, as the same screen rate can be achieved with much better performance, for example, measured by appointment time. 3.8.3. Adjust2 and Adjust3 model The targeted screening rate set by the Quebec government is 70%. The current 15 CDDs together can handle a screening rate of 79%8 at full capacity. In our Adjust2 model summarized in table 3-17, we keep all attributes level same as Adjust1 model, but increase screening rate of all population zones by 25% to reach 70% screening rate. We see from table 3-16 that more than half of the clinics now have a queue for appointment, most of them with waiting time between 3 weeks to 2 months. This, however, hasn’t even taken into account that all these CDDs also perform non-PQDCS screening. In our final model Adjust2, we add one extra clinic (clinic 16) in the same population zone where clinic 1 located, after considering the total screening load of the area. We then assigned TRAVEL level to be same as clinic 1, WAITAPP=2, PARKING=2, HOUR=2, WAITCLINIC=2, NURSE=1, SCREENING=1, and WAITRESULT=1 for clinic 16, and present the result for this scenario in table 3-16 labeled Adjust2. Our result shows that clinic 16 is a bit overloaded. However performance of other previously congested clinics all improved dramatically in terms of waiting time for appointment and number of participants waiting for screening. Total number of screening increased by 3108, total number of waiting in queue reduced by more than 1500. Utilization rate of most clinics decreased a little due to added capacity from clinic 16. Overall the system is much less crowded as before, although there’s still room for improvement by further adjusting the attribute levels of some clinics. Table 3-16: Arena result from Adjust2 model and Adjust3 model Total number of screening Waiting time for appointment Utilization rate Number of participants waiting for screening Clinics Adjust2 Adjust3 Adjust2 Adjust3 Adjust2 Adjust3 Adjust2 Adjust3 1 1710 1390 0 0 42.75% 34.75% 0 0 8 This can easily be computed by dividing the total number of machine minutes of all 15 CDDs per 2 years by 30 minutes per screening. 67 2 625 506 0 0 15.63% 12.65% 0 0 3 7996 7903 23 1 99.97% 98.79% 891 32 4 3973 3276 3 0 99.33% 81.91% 46 1 5 11998 11997 33 12 100.00% 99.99% 2126 643 6 1255 1009 0 0 31.36% 25.23% 0 0 7 3999 3997 39 18 99.98% 99.94% 923 342 8 7366 5955 0 0 92.09% 74.44% 4 1 9 3997 3581 12 0 99.93% 89.54% 208 3 10 4000 3998 41 25 99.99% 99.96% 996 489 11 4000 3999 58 42 99.99% 99.98% 1725 1004 12 530 416 0 0 13.24% 10.40% 0 0 13 3998 3996 31 12 99.96% 99.91% 683 211 14 1321 1070 0 0 33.02% 26.76% 0 0 15 6521 5303 0 0 81.53% 66.30% 1 0 16 7999 57 99.99% 3330 Total 63287 66395 7602 6057 3.9. Conclusion Breast cancer is one of the most common cancers in the world, and the second leading cancer cause of death among women in Canada (Hamilton and Barlow 2003). Early detection of breast cancer through mammographic screening can lead to significant reduction of breast cancer mortality (Aro et al. 1999). Optimal design of government-funded breast cancer screening program therefore plays an important role in the prevention and early diagnosis of breast cancer. In this paper, use stated preference discrete choice modeling, we conduct a survey in Montreal, Canada, under the setting the Quebec breast cancer screening programPQDCS program, to explore client preference regarding the service configuration of screening clinics. Of the eight studied service attributes, we found nursing staff’s manner and knowledge regarding screening and breast caner, as well as waiting time for appointment are the most influential factors in choice of clinics, followed by parking availability. We found clinics’ opening hours to be the only attributes that are not significant in the decision-making. Use latent class analysis, we are able to identify the homogeneity among PQDCS current participants, i.e., there’s no evidence of difference among participants’ preference of studied attributes. Use the collected data on 12 population zones in metropolitan Montreal, and 15 current designated screening clinics, we build an Arena simulation model to explore the change in the overall system if preferences are taking into account. Our results show significant improvement in all three metrics, measured by number of screening per year, utilization rate of mammography machines, as well as number of women waiting for appointment. 68 The main contribution of our research is to provide a tool that can be used to improve service quality of preventive care program significantly, without incurring vast capital investment. All the studied attributes can be adjusted at low or even no cost. And although not presented here, our survey result can also be used in a probit model to generate re-screening participation function. Our survey respondents are a convenient sample of current participants of the PQDCS program. Future research can use the same approach, but include nonparticipant in the survey as well, to generate a general participant function to gain more managerial insights. 69 4. Conclusion In chapter 2, we provide evidence that ED process redesign, specifically the adoption of POCT, has a substantial and valuable impact on ED operations. POCT adoption reduces the service time not only of patients who receive the troponin test at the point of care, but also of other patients within the ED who do not require a troponin test. That we further observe the service time impacts of POCT adoption to be greater during peak than during offpeak hours among both test and no-test patients suggests that POCT can have a measurable impact on ED performance during crucial operating periods. POCT adoption is also associated with improved service quality. All patients who presented to the ED during the POCT pilot period experienced a lower bounceback rate than in the comparison periods. Some of this quality improvement may follow from the priority queue model predictions, whereby high and low priority patients experience differential impacts. Empirically, lower priority patients have a greater propensity for bounceback. We speculate that less severe patients may, a priori, receive less attention. As service time savings accrue, as with priority queuing, less severe patients stand to reap the greatest benefit from any surplus physician attention. Finally, we find adoption of POCT to have a positive effect on waiting time for both test and no-test patients. We observe the lowest priority patients to experience the greatest, and highest priority patients the smallest, decrease in waiting time upon adoption of POCT, supporting our prediction of ED behavior based on queuing theory. It further suggests that POCT leads to operational improvement not only through a direct impact on the service time of test patients, but also through an indirect impact on waiting time for all who present to the ED when POCT is in use. There are several possible extensions of this research. First, what model should be used to help an administrator determine the optimal number of tests to be converted to POCT? This study analyzes a setting in which a single test is converted to POCT. That nursing workloads increase with each additional test converted suggests diminishing marginal returns. What distribution of tests between bedside and central lab minimizes service times without compromising service quality subject to staffing constraints? Second, what model would help an administrator select the optimal test(s) to be converted to POCT? Prior studies reveal the rationale for test selection to date to be largely ad hoc. What are the critical selection criteria and potential for scale economies, in both cost and quality, from test analysis? Finally, how do administrators model the financial impact of POCT conversion? In our setting and that of many prior studies, the conversion to POCT is justified solely on the basis 70 of service time and service quality. In the context of rising healthcare costs, however, POCT adoption incurs both a capital equipment expense and higher marginal per-test costs. Whether POCT can decrease ED and central lab congestion, and increase patient volume, sufficiently to offset these costs warrants further study. In chapter 3, study on breast cancer screening, we found nursing staff’s manner and knowledge regarding screening and breast caner, as well as waiting time for appointment are the most influential factors in choice of clinics, followed by parking availability. We found clinics’ opening hours to be the only attributes that are not significant in the decision-making. Use latent class analysis, we are able to identify the homogeneity among PQDCS current participants, i.e., there’s no evidence of difference among participants’ preference of studied attributes. Use the collected data on 12 population zones in metropolitan Montreal, and 15 current designated screening clinics, we build an Arena simulation model to explore the change in the overall system if preferences are taking into account. Our results show significant improvement in all three metrics, measured by number of screening per year, utilization rate of mammography machines, as well as number of women waiting for appointment. The main contribution of our research is to provide a tool that can be used to improve service quality of preventive care program significantly, without incurring vast capital investment. All the studied attributes can be adjusted at low or even no cost. And although not presented here, our survey result can also be used in a probit model to generate re-screening participation function. Our survey respondents are a convenient sample of current participants of the PQDCS program. Future research can use the same approach, but include nonparticipant in the survey as well, to generate a general participant function to gain more managerial insights. 77 Olivares, M., C. Terwiesch ,L. Cassorla. 2008. Structural estimation of the newsvendor model: An application to reserving operating room time. Management Science 54(1) 41-55. Parker, B. R. ,V. Srinivasan. 1976. A consumer preference approach to the planning of rural primary health-care facilities. Operations Research 24(5) 991-1025. Partovi, S.N., B.K. Nelson, E.D. Bryan, M.J. Walsh. 2001. Faculty triage shortens emergency department length of stay. Academic Emergency Medicine 8(10) 990-995. Parvin, C.A., S.F. Lo, S.M. Deuser, L.G. Weaver, L.M. Lewis, M.G. Scott. 1996. Impact of point-of-care on patients’ length of stay in a large emergency department. Clinical Chemistry 42(5) 711-717. Pascoe, G. C. 1983. Patient satisfaction in primary health care: A literature review and analysis. Evaluation and Program Planning 6(3-4) 185-210. Pham, J.C., T.D. Kirsch, P.M. Hill, K. DeRuggerio, B. Hoffmann. 2011. Seventy-two-hour returns may not be a good indicator of safety in the emergency department: A national study. Academic Emergency Medicine 18(4) 390-397. Pitts, S.R., R.W. Niska, J. Xu, C.W. Burt. 2008. National hospital ambulatory medical care survey: 2006 emergency department summary. National Center for Health Statistics, US Centers for Disease Control and Prevention. National Health Statistics Reports 4 1-32. http://www.cdc.gov/nchs/data/nhsr/nhsr004.pdf. Renaud, B., P. Maison, A. Ngako, P. Cunin, A. Santin, J. Hervé, M. Salloum, M.-J. Calmettes, C. Boraud, V. Lemiale, J.C. Grégo, M. Debacker, F. Hemery, E. Roupie. 2008. Impact of point-of-care testing in the emergency department evaluation and treatment of patients with suspected acute coronary syndromes. Academic Emergency Medicine 15(3) 216-224. Richardson, D. 2006. Increase in patient mortality at 10 days associated with emergency department overcrowding. Medical Journal of Australia 184(5) 213-216. Rimer, B. K., M. K. Keintz, H. B. Kessler, P. F. Engstrom ,J. R. Rosan. 1989. Why women resist screening mammography: Patient-related barriers. Radiology 172(1) 243-246. Rivera, A.J., B.-T. Karsh. 2010. Interruptions and distractions in healthcare: Review and reappraisal. Quality and Safety in Health Care 19(4): 304-312. Rust, G., M. Gailor, E. Daniels, B. Mcmillan-Persaud, H. Strothers, R. Mayberry. 2008. Point of care testing to improve glycemic control. International Journal of Health Care Quality Assurance 21(3) 325-335. 78 Saghafian, S., W.J. Hopp, M.P. Van Oyen, J.S. Desmond, S.L. Kronick. 2010. Patient streaming as a mechanism for improving responsiveness in emergency departments. Working Paper Series #1162, Ross School of Business, University of Michigan, Ann Arbor, MI. http://hdl.handle.net/2027.42/85792. Saghafian, S., W.J. Hopp, M.P. Van Oyen, J.S. Desmond, S.L. Kronick. 2011. Complexity-based triage: A tool for improving patient safety and operational efficiency. Working Paper Series #1161, Ross School of Business, University of Michigan, Ann Arbor, MI. http://hdl.handle.net/2027.42/85791. Samuelson, P. A. 1947. Foundations of economic analysis, Harvard University Press. Sawtooth Software, I. 2004. The cbc latent class technical paper (version 3) [Online]. www.sawtoothsoftware.com. 2012]. Sawtooth Software, I. 2007. Latent class v4 manual [Online]. www.sawtoothsoftware.com. 2012]. Schoen, C., R. Osborn, P. T. Huynh, M. Doty, K. Davis, K. Zapert ,J. Peugh. 2004. Primary care and health system performance: Adults¬í experiences in five countries. Health Affairs. Schull, M.J., A. Kiss, J.-P. Szalai. 2007. The effect of low-complexity patients on emergency department waiting times. Annals of Emergency Medicine 49(3) 257-264. Starfield, B. 1998. Primary care, New York, Oxford University Press. Stinnett, A. A. ,A. D. Paltiel. 1996. Mathematical programming for the efficient allocation of health care resources. Journal of Health Economics 15(5) 641-653. Stockbridge, H., R. Hardy ,C. Glueck. 1989. Public cholesterol screening: Motivation for participation, followup outcome, self-knowledge, and coronary heart disease risk factor intervention. The Journal of Laboratory and Clinical Medicine 114(2) 142-151. Sutton, S., G. Bickler, J. Sancho-Aldridge ,G. Saidi. 1994. Prospective study of predictors of attendance for breast screening in inner london. Journal of Epidemiology and Community Health 48(1) 65-73. Vance, J., P. Sprivulis. 2005. Triage nurses validly and reliably estimate emergency department patient complexity. Emergency Medicine Australasia 17(4) 382-386. Vassilacopoulos, G. 1985a. A simulation model for bed allocation to hospital inpatient departments. Simulation 45(5) 233-241. Vassilocopoulos, G. 1985b. Allocating doctors to shifts in an accident and emergency department. Journal of the Operational Research Society 36(6) 517-523. Vernon, S. W., E. A. Laville ,G. L. Jackson. 1990. Participation in breast screening programs: A review. Social Science & Medicine 30(10) 1107-1118. 79 Verter, V. ,S. D. Lapierre. 2002. Location of preventive health care facilities. Annals of Operations Research 110 123-132. Welch, S.J., S.S. Jones, T. Allen. 2007. Mapping the 24-hour emergency department cycle to improve patient flow. Joint Commission Journal on Quality and Patient Safety 33(5) 247-255. Weiss, S.J., A.A. Ernst, R. Derlet, R. King, A. Bair, T.G. Nick. 2005. Relationship between the national ED overcrowding scale and the number of patients who leave without being seen in an academic ED. The American Journal of Emergency Medicine 23(3) 288-294. White, D., L. Kaplan, E. Linda. 2011. Characteristics of patients who return to the emergency department within 72 hours in one community hospital. Advanced Emergency Nursing Journal 33(3) 344-353. Whitt, W. 1993. Approximations for the gi/g/m queue. Production and Operations Management 2(2) 114-161. Wiler, J., R. Griffey, T. Olsen. 2011. Review of modeling approaches for emergency department patient flow and crowding research. Academic Emergency Medicine 18(12) 1371-1379. Yoon, P. 2003. Emergency department fast-track system. Health Technology Assessment Unit Initiative #10, Alberta Heritage Foundation for Medical Research, Edmonton, Alberta, Canada. http://www.ihe.ca/documents/HTA-FR10.pdf. Zeithaml, V. A. 1988. Consumer perceptions of price, quality, and value: A means-end model and synthesis of evidence. The Journal of Marketing 52(3) 2-22. Zhang, Y., O. Berman ,V. Verter. 2011. The impact of client choice on preventive healthcare facility network design. OR Spectrum 1-22.