UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

The effects of British Columbia hospital closures on delivery of health care services and the population's… Panagiotoglou, Dimitra 2016

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2017_february_panagiotoglou_dimitra.pdf [ 2.45MB ]
Metadata
JSON: 24-1.0320896.json
JSON-LD: 24-1.0320896-ld.json
RDF/XML (Pretty): 24-1.0320896-rdf.xml
RDF/JSON: 24-1.0320896-rdf.json
Turtle: 24-1.0320896-turtle.txt
N-Triples: 24-1.0320896-rdf-ntriples.txt
Original Record: 24-1.0320896-source.json
Full Text
24-1.0320896-fulltext.txt
Citation
24-1.0320896.ris

Full Text

  THE EFFECTS OF BRITISH COLUMBIA HOSPITAL CLOSURES ON DELIVERY OF HEALTH CARE SERVICES AND THE POPULATION’S HEALTH by  Dimitra Panagiotoglou  B.A.Sc., University of Toronto, 2009 M.P.H., Columbia University, 2012  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF  DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Population and Public Health)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)  November 2016  © Dimitra Panagiotoglou, 2016      ii  Abstract Background: In 2002, British Columbia (BC) began redistributing its hospital services. Existing facilities closed, were downsized or became specialized. Reasons for these changes included the centralization of the health authorities and subsequently the redistribution of services, along with safety concerns regarding small units, difficulties recruiting and retaining staff, and policy changes in the province’s Standards of Accessibility. At the same time, innovations in service delivery, such as inter-hospital transfer practices and telehealth initiatives (e.g. telestroke), modified how health care is provided. Effective health service delivery is a complex matter. Over a decade since redistribution began, there has been no evaluation of the changes in service distribution and their impact on patient health. Methods: This is a retrospective cohort study of all adult (18 years and over) acute myocardial infarction (AMI), stroke, and severe trauma events that occurred within the province between April 1 1999 and March 31 2013. Using administrative data, segmented regression and hierarchical hazards modelling techniques, this study examines the effect of service redistribution on patients’ mortality outcomes. Results: The interrupted time series models found service redistribution was not associated with changes in 30-day mortality outcomes, and was likely a response to facility underutilization. Although there was extensive variation in patient access to care (travel burden) across health authorities, the hierarchical Cox proportional hazards models showed that long travel time (>30 mins) was not associated with patient short term mortality after controlling for appropriateness and quality of care along with compensating mechanisms such as inter-hospital transfers, and telehealth services. Conclusion: This work demonstrates that efficiencies in health system delivery can be gained by eliminating underutilized acute care services but also identifies challenges in ensuring equitable access to care. iii  Preface This dissertation is my original, unpublished work. The analysis reported is covered by UBC Behavioural Research Ethics Board Certificate number H14-00641. I was the lead investigator for all analysis, and responsible for the design of the research program, data analysis, and manuscript composition. Kim McGrail (KM) was the supervisory author on all research and involved throughout study design and manuscript edits. Committee members Michael Law (ML) and Stirling Bryan (SB) provided input on analysis and contributed to manuscript revisions. The involvement of my committee members is reflected in the transition between active singular and active plural voice within chapters. While I led all of the research, the use of active voice reflects the contributions of my committee members to the analyses described. All inferences, opinions, and conclusions drawn in this publication are those of the author, and do not reflect the opinions or policies of the Data Stewards. A version of Chapter 5 is currently under peer-review for publication: Panagiotoglou D, McGrail K. “Factors associated with STEMI guideline adherence, and its effects on thirty-day mortality in British Columbia: 1999-2013”. I led the design, analysis and drafting of the manuscript. KM provided input to the study design and draft, and revised the document. A version of Chapter 7 has recently been accepted for publication and is reused here with permission: Panagiotoglou D, Law M, McGrail K. “Effect of hospital closures on acute care outcomes in British Columbia, Canada: an interrupted time series study”. Medical Care, 2016; DP, ML and KM designed the study. I took the primary role of designing the study, analyzing the data and preparing the manuscript. ML and KM provided input to the draft and contributed to revisions. A version of Chapter 8 is currently undergoing peer-review for publication: Panagiotoglou D, Law M, McGrail K. “How appropriateness and quality can compensate for access to acute care for time-sensitive medical emergencies: a cohort study”. DP and ML designed the study. I led the analysis and prepared the manuscript. KM provided input and revised the manuscript. Funding for this research was provided through two CIHR Strategic Training Programs: (1) the Western Regional Training Centre (WRTC), and (2) Public Health and the Agricultural Rural Ecosystem (PHARE); and partner institutes including the Institute of Health Services and Policy iv  Research, Institute of Circulatory and Respiratory Health, Institute of Infection and Immunity, and the Institute of Population and Public Health. v  Table of Contents  Abstract .......................................................................................................................................... ii Preface ........................................................................................................................................... iii Table of Contents ...........................................................................................................................v List of Tables ................................................................................................................................ xi List of Figures ............................................................................................................................. xiv List of Abbreviations ................................................................................................................. xvi Acknowledgements .................................................................................................................... xix Chapter 1: Introduction ................................................................................................................1 1.1 Rationale ......................................................................................................................... 1 1.2 Thesis overview .............................................................................................................. 2 1.2.1 Research question and hypotheses .............................................................................. 2 1.2.2 Organization of the thesis ........................................................................................... 3 Chapter 2: Literature review and theoretical framework .........................................................4 2.1 Methods........................................................................................................................... 4 2.2 Results ............................................................................................................................. 5 2.2.1 Existing frameworks ................................................................................................... 5 2.2.2 Incorporating access, appropriateness, and quality into a single framework ............. 6 2.3 Access ............................................................................................................................. 7 2.4 Appropriateness .............................................................................................................. 8 2.5 Quality............................................................................................................................. 8 2.6 Time sensitive medical conditions of interest ............................................................... 10 vi  2.6.1 Acute myocardial infarction ..................................................................................... 10 2.6.2 Stroke ........................................................................................................................ 11 2.6.3 Trauma ...................................................................................................................... 13 2.7 Summary ....................................................................................................................... 14 Chapter 3: Data sources and measures ......................................................................................16 3.1 Data preparation ............................................................................................................ 16 3.1.1 Study sample ............................................................................................................. 17 3.2 Inclusion and exclusion criteria .................................................................................... 18 3.3 Variable development ................................................................................................... 21 3.3.1 Patient level characteristics ....................................................................................... 21 3.3.1.1 Comorbidity ...................................................................................................... 22 3.3.2 Hospital characteristics ............................................................................................. 24 3.3.2.1 Condition-specific volume ................................................................................ 24 3.3.2.2 Hospital peer groups ......................................................................................... 24 3.3.3 Episode of hospitalization ......................................................................................... 25 3.3.3.1 Length of stay per hospitalization and entire episode of care ........................... 26 3.3.3.2 In-hospital ......................................................................................................... 26 3.3.4 Access ....................................................................................................................... 28 3.3.4.1 Inter-hospital transfers ...................................................................................... 28 3.3.4.2 Total travel burden ............................................................................................ 29 3.3.5 Appropriateness ........................................................................................................ 29 3.3.6 Quality....................................................................................................................... 30 3.3.7 Outcomes .................................................................................................................. 31 vii  3.3.8 Summary ................................................................................................................... 32 Chapter 4: Access to care ............................................................................................................34 4.1 Methods......................................................................................................................... 34 4.1.1 Travel matrix ............................................................................................................. 34 4.2 Identifying a list of relevant facilities and their geographic coordinates ...................... 35 4.2.1 Identifying all residential postal codes and centroid coordinates ............................. 36 4.2.1.1 ArcGIS versus Google Maps ............................................................................ 36 4.2.2 Linking the travel matrix with patient records .......................................................... 39 4.2.3 Troubleshooting ........................................................................................................ 39 4.2.4 Emergency health services Ambulance data............................................................. 41 4.3 Results ........................................................................................................................... 41 4.4 Summary ....................................................................................................................... 45 Chapter 5: Appropriateness of care ...........................................................................................47 5.1 Methods......................................................................................................................... 47 5.1.1 AMI standards of care and protocols ........................................................................ 48 5.1.2 Stroke standards of care and protocols ..................................................................... 53 5.1.2.1 Telestroke services ............................................................................................ 54 5.1.3 Trauma standards of care and protocols ................................................................... 56 5.1.4 Emergency services offered outside a hospital ......................................................... 56 5.2 Results ........................................................................................................................... 56 5.2.1 AMI ........................................................................................................................... 56 5.2.2 Stroke ........................................................................................................................ 60 5.2.3 Trauma ...................................................................................................................... 61 viii  5.3 Summary ....................................................................................................................... 63 Chapter 6: Quality of care ..........................................................................................................64 6.1 Methods......................................................................................................................... 64 6.1.1 Data preparation ........................................................................................................ 65 6.1.2 Method variations and sensitivity analysis ............................................................... 66 6.2 Results ........................................................................................................................... 66 6.3 Summary ....................................................................................................................... 71 Chapter 7: Investigating the impact of redistribution .............................................................73 7.1 Methods......................................................................................................................... 74 7.1.1 Data preparation ........................................................................................................ 74 7.1.1.1 Intervention group membership ........................................................................ 74 7.1.1.2 Matched controls ............................................................................................... 75 7.1.1.3 Pre- vs. post-intervention observation time ...................................................... 76 7.1.1.4 Outcomes .......................................................................................................... 76 7.1.2 ITS models ................................................................................................................ 76 7.2 Results ........................................................................................................................... 78 7.2.1 Eligible facilities ....................................................................................................... 78 7.2.2 Data restrictions ........................................................................................................ 85 7.2.3 Patient demographics ................................................................................................ 85 7.2.4 30-day mortality rates over time ............................................................................... 88 7.2.5 Proportion of patients not taken to their closest facility ........................................... 93 7.2.5.1 Methods............................................................................................................. 94 7.2.5.2 Proportion mismatch results ............................................................................. 95 ix  7.3 Summary ..................................................................................................................... 104 Chapter 8: Determining the impacts of access, appropriateness, and quality on patient outcomes......................................................................................................................................105 8.1 Methods....................................................................................................................... 105 8.1.1 Variable selection.................................................................................................... 105 8.1.2 Model building and selection .................................................................................. 106 8.1.3 Sensitivity analysis.................................................................................................. 108 8.2 Results ......................................................................................................................... 108 8.2.1 Complete study period (1999-2013) ....................................................................... 108 8.2.2 Restricted study period (2009-2013) ...................................................................... 124 8.3 Summary ..................................................................................................................... 138 Chapter 9: Discussion ................................................................................................................139 9.1 Summary of results ..................................................................................................... 139 9.1.1 Hospitals which ceased treatment ........................................................................... 140 9.1.2 Appropriateness of care .......................................................................................... 142 9.1.3 Quality of care......................................................................................................... 145 9.1.4 Access to care ......................................................................................................... 146 9.1.5 Compensating mechanisms ..................................................................................... 150 9.1.5.1 Inter-hospital transfers .................................................................................... 150 9.1.5.2 Telehealth ........................................................................................................ 151 9.1.6 Other explanatory variables that influence patient outcomes ................................. 153 9.1.6.1 Patient characteristics...................................................................................... 153 9.2 Limitations .................................................................................................................. 154 x  9.2.1 Access to care ......................................................................................................... 154 9.2.2 Appropriateness of care .......................................................................................... 156 9.2.3 Condition severity ................................................................................................... 156 9.3 Generalizability ........................................................................................................... 157 9.4 Recommendations ....................................................................................................... 157 9.4.1 Policy recommendations ......................................................................................... 158 9.4.2 Research recommendations .................................................................................... 159 Chapter 10: Conclusion .............................................................................................................161 References ...................................................................................................................................163 Appendix A: Condition codes used to calculate Charlson Comorbidity Index ...................... 194 Appendix B: Propensity score matched interrupted time series ............................................. 195 Appendix C: Results of parsimonious proportional hazards models ...................................... 196 C.1 Method .................................................................................................................... 196 C.2 Results ..................................................................................................................... 196 xi  List of Tables Table 3-1. National trauma registry comprehensive data classification of external cause of trauma codes defined as urgent ..................................................................................................... 20 Table 3-2. ICD-9/10 external cause of trauma codes used as exclusion criteria in capturing trauma cohort ................................................................................................................................ 20 Table 3-3. Charlson Comorbidity Index weights assigned to each condition .............................. 22 Table 3-4. Revised Charlson Comorbidity Index with modern listing and weight assignments .. 23 Table 3-5. Summary of variables developed for subsequent analysis, source of data, method used, and range of values .............................................................................................................. 33 Table 4-1. Median time to closest facility in minutes by health authority, sex, and year of event, per condition ................................................................................................................................. 43 Table 4-2. Median time to first facility of admission in minutes by health authority, sex and year, per condition ................................................................................................................................. 44 Table 4-3. Correlation between time to care estimate and true time ............................................ 45 Table 5-1. Contraindications for patients with ST segment elevated myocardial infarction by patient and facility/staff level characteristics and year of event ................................................... 51 Table 5-2. Contraindications (errors of commission) for patients with ischemic stroke by year . 54 Table 5-3. Telestroke fee items, MSP codes, and descriptions .................................................... 55 Table 5-4. AMI patient demographics by appropriateness of care ............................................... 58 Table 5-5. Treatment received by AMI patient admitted to hospital, organized by appropriateness....................................................................................................................................................... 59 Table 5-6. Stroke patient demographics by appropriateness of care ............................................ 61 Table 5-7. Trauma patient demographics by appropriateness of care .......................................... 62 xii  Table 6-1. Number of hospitals that satisfy inclusion and exclusion criteria, and flagged as performance outliers using funnel plots ........................................................................................ 66 Table 6-2. Hospitals flagged as poor performers by condition and interval, using full models ... 70 Table 6-3. Hospitals flagged as poor performers by condition and interval, using parsimonious models ........................................................................................................................................... 71 Table 7-1. Facilities accepting AMI patients by last year of admission (16 hospitals and 26 non-hospital facilities) .......................................................................................................................... 80 Table 7-2. Facilities accepting stroke patients by last year of admission (14 hospitals and 27 non-hospital facilities) .......................................................................................................................... 81 Table 7-3. Facilities accepting trauma patients by last year of admission (14 hospitals and 28 non-hospital facilities) .................................................................................................................. 82 Table 7-4. Demographics by condition and interrupted time series intervention group (cases and controls) ........................................................................................................................................ 87 Table 7-5. Interrupted time series output for 30-day mortality rates, per condition ..................... 90 Table 7-6. Interrupted time series output for 90-day mortality rates, per condition ..................... 92 Table 7-7. Interrupted time series output for bypass rate, per condition ...................................... 98 Table 7-8. Interrupted time series output for proportion of patients transferred, per condition . 100 Table 8-1. Extent of missing estimated distance and reasons for it, per condition..................... 110 Table 8-2. Demographic differences between patients with and without time to care estimates, AMI ............................................................................................................................................. 112 Table 8-3. Demographic differences between patients with and without time to care estimates, stroke ........................................................................................................................................... 113 Table 8-4. Demographic differences between patients with and without time to care, trauma .. 114 xiii  Table 8-5. Results of univariate Cox Proportional Hazards models, AMI ................................. 116 Table 8-6. Results of univariate Cox Proportional Hazards models, stroke ............................... 117 Table 8-7. Results of univariate Cox Proportional Hazards models, trauma.............................. 118 Table 8-8. Results of multivariate Cox Proportional Hazards models, AMI.............................. 120 Table 8-9. Results of multivariate Cox Proportional Hazards models, stroke............................ 122 Table 8-10. Results of multivariate Cox Proportional Hazards models, trauma ........................ 123 Table 8-11. Provision of true time to care by reason for missing estimated time to care, per condition ..................................................................................................................................... 126 Table 8-12. Patient demographics by travel method 2009-2013, AMI ...................................... 127 Table 8-13. Patient demographics by travel method 2009-2013, stroke .................................... 128 Table 8-14. Patient demographics by travel method 2009-2013, trauma ................................... 129 Table 8-15. Results of multivariate Cox Proportional Hazards models using ambulance data (5) and estimated time to care (5a), AMI ......................................................................................... 133 Table 8-16. Results of multivariate Cox Proportional Hazards models using ambulance data (5) and estimated time to care (5a), stroke ....................................................................................... 135 Table 8-17. Results of multivariate Cox Proportional Hazards models using ambulance data (5) and estimated time to care (5a), trauma ...................................................................................... 137 Table A-1.Coding algorithms for Charlson Comorbidity Index………………………………. 194 Table C-1. Results of parsimonious multivariate Cox Proportional Hazards models using estimated time to care (1999-2013), trauma……………………………………………………197 Table C-2. Results of multivariate Cox Proportional Hazards models using ambulance data (5), estimated time to care for patients with ambulance data (5a), and all admissions between 2009 and 2013 (5b), trauma…………………………………………………………………………..198 xiv  List of Figures Figure 2-1. The three pillars of health service distribution influencing patient outcomes ............. 7 Figure 3-1. Proportion of AMI events that died within thirty days, by location of death (in or out of hospital) by fiscal year .............................................................................................................. 26 Figure 3-2. Proportion of stroke events that died within thirty days, by location of death (in or out of hospital) by fiscal year........................................................................................................ 27 Figure 3-3. Proportion of trauma events that died within thirty days, by location of death (in or out of hospital) by fiscal year........................................................................................................ 27 Figure 4-1. Matching process and results ..................................................................................... 42 Figure 5-1. Appropriate care for patient presenting with ST segment elevated myocardial infarction ....................................................................................................................................... 50 Figure 6-1. Trauma RSMR for 30-day mortality rates ................................................................. 68 Figure 7-1. Interrupted time series with control schematic .......................................................... 78 Figure 7-2. Hospitals by type, 1999 .............................................................................................. 84 Figure 7-3. Hospitals by type, 2013 .............................................................................................. 84 Figure 7-4. AMI, 30-day mortality rate ........................................................................................ 88 Figure 7-5. Stroke, 30-day mortality rate...................................................................................... 89 Figure 7-6. Trauma, 30-day mortality rate.................................................................................... 89 Figure 7-7. Mismatch between true facility visited and closest for AMI ..................................... 96 Figure 7-8. Mismatch between true facility visited and closest for Stroke .................................. 96 Figure 7-9. Mismatch between true facility visited and closest for Trauma ................................ 97 Figure 7-10. Number of patients admitted by hospital type ......................................................... 99 xv  Figure 7-11. Proportions of patients by HSDA traveling to hospital other than closest for AMI, 1999............................................................................................................................................. 101 Figure 7-12. Proportions of patients by HSDA traveling to hospital other than closest for AMI, 2013............................................................................................................................................. 101 Figure 7-13. Proportions of patients by health service delivery area traveling to hospital other than closest for trauma, 1999 ...................................................................................................... 102 Figure 7-14. Proportions of patients by health service delivery area traveling to hospital other than closest for trauma, 2013 ...................................................................................................... 102 Figure 7-15. Proportions of patients by health service delivery area traveling to hospital other than closest for stroke, 1999 ....................................................................................................... 103 Figure 7-16. Proportions of patients by health service delivery area traveling to hospital other than closest for stroke, 2013 ....................................................................................................... 103 Figure 8-1. Probability of survival by time to care, AMI ........................................................... 109 Figure 8-2. Probability of survival by time to facility, AMI ...................................................... 115 Figure 8-3. Probability of survival by time to facility, stroke .................................................... 115 Figure 8-4. Probability of survival by time to facility, trauma ................................................... 116 Figure B-1. Propensity score matched interrupted time series, AMI, 30-day mortality rate….. 195  xvi  List of Abbreviations aHR Adjusted hazards ratio AIC Akaike Information Criterion AIDS Acquired immune deficiency syndrome AMI Acute Myocardial Infarction API Application programming interface BC British Columbia BCEHS British Columbia Emergency Health Services bpm Beats per minute CA California CABG Coronary Artery Bypass Graft CAEP Canadian Association of Emergency Physicians CCI Charlson comorbidity index CCU Cardiac Care Unit CI Confidence interval CIHI Canadian Institute for Health Information CT Computed tomography CTAS Canadian Emergency Department Triage & Acuity Scale CVD Cerebrovascular disease DAD Discharge abstract database ECG/EKG Electrocardiography ED Emergency department EOH Episode of hospitalization EMS Emergency medical services Est Estimate FSA Forward sortation area GCS Glasgow coma scale GIS Geographic information systems HA Health Authority HIV Human immunodeficiency virus xvii  HR Heart rate Hazard ratio HRMR Hospital risk-adjusted mortality ratio Hrs Hours HSDA Health Services Delivery Area HSMR Hospital standardized mortality ratio ICD International classification of diseases ICU Intensive care unit ID Identifier ISS Injury Severity Score ITS Interrupted time series km Kilometers LOS Length of stay MHP Ministry of Health Planning mins Minutes MIZ Metropolitan influenced zone mmHg Millimeters of mercury MRI Magnetic resonance imaging MSP Medical services plan NC North Carolina OD Origin-Destination OLS Ordinary least squares PC Postal code PCI Percutaneous coronary intervention PHN Patient Health Number pPCI Primary percutaneous coronary intervention PTCA Percutaneous transluminal coronary angioplasty RR Respiratory rate Relative risk rtPA Recombinant tissue plasminogen activator xviii  RSMR Risk Standardized Mortality Ratio RTS Revised trauma score SBP Systolic blood pressure SD Standard deviation SES Socioeconomic status SLI Single link indicator STEMI ST Segment Elevation Myocardial Infarction tPA Tissue plasminogen activator UCOD Underlying cause of death UK United Kingdom USA United States of America   xix  Acknowledgements I am sincerely grateful to my supervisor Dr. Kimberlyn McGrail and my committee members Drs. Michael Law and Stirling Bryan, for their support throughout my doctoral training. In particular, I offer my extended gratitude to Dr. Kimberlyn McGrail who took me on as a student after my first year of doctoral studies and has provided academic and professional guidance, and opportunities to collaborate with others in the health services research arena. I cannot thank her enough for the encouragement and expertise she has offered over the years. I would like to thank Dr. Michael Law for his invaluable methodological advice, and Dr. Stirling Bryan for his thoughtful feedback. I would also like to thank my colleagues at CHSPR and in particular wish to acknowledge the contribution that Dawn Mooney has provided in making my results palatable to a larger audience. I also wish to recognize the Canadian Institutes for Health Research, who provided financial support for my doctoral studies through two Strategic Training Initiatives in Health Research awards. Special thanks are owed to my parents Christos and Maria Panagiotoglou, and brother George Panagiotoglou for their love and encouragement. Finally, thank you to my partner Philip Wolfsberger for his unwavering support and companionship. 1  Chapter 1: Introduction On December 12th 2001, the government of British Columbia (BC) announced a re-organization of its regional health authorities. The 52 health authorities in existence at the time were collapsed into six: five geographic divisions (subdivided into sixteen health service delivery areas, HSDAs) and a sixth to oversee the organization and delivery of provincial programs and highly specialized services such as the BC Centre for Disease Control and BC Children’s Hospital.1 Following this announcement, BC began redistributing its hospital services.2 Existing facilities closed, were downsized or specialized, while new centres opened elsewhere.2,3 In some cases, although facilities were not closed, the hours of operation were shortened from 24/7 care to business hours—effectively reducing emergency rooms to acute care clinics.4 Reasons for these changes included the centralization of the health authorities and subsequently the redistribution of services, along with safety concerns regarding small units, difficulties recruiting and retaining staff, and policy changes in the province’s Standards of Accessibility.3,5 At the same time, innovations in service delivery, such as inter-hospital transfer practices and telehealth initiatives (e.g. telestroke), modified how health care is provided.6–8 1.1 Rationale This is not the first time a province in Canada has gone through the process of centralizing its services in an effort to cull small, underutilized centres and improve system efficiency.34 Nor is it the first time that British Columbia has restructured the services it offers.35 Given the evidence that regional variations in disease burden and outcomes have been noted both locally and abroad,18–24 and studies show rural citizens have poorer health,17 there is the potential for said changes in the distribution of services to have exacerbated regional health inequalities. It is important to understand the effects of the recent redistribution of hospital services, and any influence these changes have on equitable access to care as they relate to the Canada Health Act.33  Effective health service delivery for acute patients is a complex matter. Over a decade since redistribution began, there has been no evaluation of the changes in service distribution and their impact on patient health. Similarly, the safety and quality of care delivered in small and/or rural facilities across the province continues to be contested.3,36 This is due, in part, to the limited literature exploring the consequences of closures within Canada34 and internationally37–40, and the 2  paucity of work studying the effects of geographic variability in access to care2 unexplained by the socioeconomic factors of communities. It is unclear what influence changes in access to appropriate care have had on medical emergencies where patients’ health deteriorates within minutes to hours. Following from this is the lack of consensus on the window of time available for treatment to optimize patient recovery. Lastly, it is unclear how innovations in service delivery via triaging, stabilization and transferring have influenced patient outcomes.41 Assessing the effects of the redistribution on patient outcomes, and understanding how the BC health system has evolved over time to care for patients experiencing acute time-sensitive conditions, is the focus of this thesis. 1.2 Thesis overview Recent changes in service distribution present an excellent natural experiment42 to explore these issues. By capitalizing on the longitudinal element of service redistribution, insight into the roles that access, quality and appropriateness play in patient health will be gained. This is imperative in effective and appropriate service distribution planning14. 1.2.1 Research question and hypotheses My research question is “What effect did British Columbia’s redistribution of hospital services have on residents’ outcomes, and did quality and appropriateness of care compensate for changes in access to acute care?” To answer this question, I will test the following hypotheses: 1. Redistribution has had no measurable effect on patient outcomes. 2. Redistribution has had no measurable effect on patients’ access to services. Testing these initial hypotheses then allows me to investigate the final hypothesis which joins three major themes of quality, appropriateness and access, and evaluate how they interact to describe patient outcomes, specifically: 3. Access to care has no measureable effect on patient outcomes after controlling for quality and appropriateness of acute care services, and compensating mechanisms such as inter-hospital transfers and telehealth services. While changes in hospital service distribution may have wide-ranging effects, the most pointed and immediate effects would likely be seen for health events or conditions that are time-sensitive. 3  For this reason, the analyses will focus on three distinct clinical events: acute myocardial infarctions (AMI, heart attacks), stroke, and trauma. 1.2.2 Organization of the thesis Before developing the study, I conducted a substantial review of the literature to identify the key factors related to hospital service delivery and patient outcomes, and existing frameworks that help define relationships among these factors. Based on this review, I chose to develop a theoretical framework that synthesizes core ideas in the literature and guides my subsequent analyses (Chapter 2). The literature review in Chapter 2 also helped focus my empirical analyses on three specific clinical events: acute myocardial infarction, stroke and trauma events.  I describe how I select my cohorts (Chapter 3) and the data I worked with, before dedicating three chapters (Chapters 4-6) to the development and descriptive analysis of variables as defined within the theoretical framework. In Chapter 7, I use interrupted time series models with segmented regression to test my first two hypotheses (see 1.2.1 Research question and hypotheses). My third hypothesis is tested using hierarchical survival analysis techniques (Chapter 8). After the thorough analysis of each question, I discuss the results of my entire dissertation and make recommendations for future research and policy in Chapter 9. The final chapter of this thesis (Chapter 10) is a formal conclusion revisiting my initial objectives. 4  Chapter 2: Literature review and theoretical framework The interest of this thesis is how the health care system responds to very specific events that are considered to be time-sensitive in nature and require specialized services – the very events most likely to be affected by changes in the location and availability of care. To examine the impact of service redistribution in BC, I drew from several schools of thought related to acute care and favourable outcomes. A systematic review was not amenable to my goal as it would likely have led to the exclusion of important work not directly associated with access but nonetheless relevant to both acute care and patient outcomes following a medical emergency. As such, I opted to conduct a traditional literature review44 to identify the key concepts underpinning health services research specific to acute care service redistribution, and focus my overall research question. 2.1 Methods I used Pubmed and Google Scholar to identify relevant literature. In the first instance, my search terms included “access”, “accessibility”, “distance”, “time to care”, “acute care”, “hospital”, “time sensitive event”, “medical emergency”, and “closure”. I used combinations of these terms to restrict the number of articles I retrieved, and from this initial review I expanded my search to include literature on “quality” and “appropriateness” – two concepts regularly tied to access and/or patient outcomes following time sensitive medical emergencies. Additionally, I used snowball retrieval techniques to identify literature not published in review journals or initially captured using my search terms.45 As part of the literature review, I also examined health services conceptual frameworks to understand the theoretical underpinnings necessary to evaluate the effects of BC’s redistribution of hospital services on patient outcomes. The objective was to choose or refine a framework that could help guide the empirical approach to my research. Based on the search terms above, Pubmed retrieved 1574 articles published as recently as June 30, 2015. After removing publications identified as duplicates; non-English; medical events other than cardiac, stroke, trauma or obstetrics events; prevention, elective procedures or post-acute care (e.g. rehabilitation, long-term care); and burden for caregiver—I was left with 228 articles. Articles related to obstetrics were retained because much of the hospital closure literature in British 5  Columbia focused on the effects of service loss on obstetric events. I supplemented this selection with 41 additional references retrieved through reviews of publications’ reference lists. In total, 269 articles were reviewed for the development of the background chapters (Chapters 2 – 6) although not all articles were directly referenced. 2.2 Results The iterative and expanding nature of this literature search and review resulted in the formulation of my research question as: “What effect did British Columbia’s redistribution of hospital services have on residents’ outcomes, and did quality and appropriateness of care compensate for changes in access to acute care?” The review revealed that in medical emergencies, the first thing facing patients is access to services, or how quickly they can get to what they need. Once care is accessed, the expectation is that patients will receive appropriate care or care that is best suited to their need, as determined by the health event and patient characteristics such as age and other existing conditions, by experienced medical personnel. Once appropriateness of care is determined, there is an expectation that the care provided is of similar quality to that generally provided (if not better). The concepts of access, appropriateness and quality are of course not as clearly distinguished as this overview implies but summarize how patients interact with the health care system following time sensitive events. 2.2.1 Existing frameworks The review identified three existing and potentially relevant frameworks: Donabedian’s Quality of Care framework first presented in 196646 and revisited in 2005;47 an updated version of Andersen’s Behavioral Model of Health Services Use;48 and Penchansky’s 1981 Access framework.49 All three frameworks have some relevance to the question at hand, but all are also limited in their ability to integrate and explain the mounting empirical evidence on the associations between access, appropriateness and quality of care on patient outcomes. Although Donabedian’s framework is highly flexible and thus applicable to many research endeavours, the compartments of structure, process, and outcome ignore the effects of patient, economic and social characteristics on the care received. As such, the framework is limited in its ability to incorporate how these external but important factors may explain some of the variation in service use, care provided, and outcomes observed.50 In particular, the framework fails to 6  describe how patient characteristics and choices should be incorporated into health systems research, and to distinguish access from quality in a satisfying manner. Conversely, Andersen’s Behavioral Model of Health Services Use focuses on the patient, and explains the use of health services as the combination of three concepts: the patient’s predisposition to use care, the resources that enable or impede use, and the need for care. The more recent adaptation of the original model elaborates on these concepts and allows for discrete compartment analysis of any concept or its subcomponents (e.g. experience, perceived vs. evaluated health, and patient characteristics). The revised framework is particularly useful because it recognizes that utilization patterns can be shaped by previous experience (e.g. consumer satisfaction). However, it minimizes the role that health care providers play on how services are used, and it considers predisposing characteristics (i.e. climate and culture, staff expertise, and patient attitudes) immutable. The model also fails to distinguish between utilization and availability of services; and the appropriateness and quality of care independent from consumers’ satisfaction.51 Pechansky’s framework focuses on the fit between patients’ needs and the system’s ability to meet (supply) them. He describes five dimensions: availability, accessibility, accommodation, affordability, and acceptability. The framework allows for the direct evaluation of how services are provided within a geo-spatial context. However, the vagueness of the dimensions means that accessibility may be defined in terms of the other four, and it is unclear whether potential or real utilization defines accessibility.52 2.2.2 Incorporating access, appropriateness, and quality into a single framework Given the strengths and limitations of these existing frameworks, I drew from but expanded these, developing the theoretical framework in Figure 2-1. I built the framework using the variables identified within the literature and organized it into three pillars related to health services distribution: access, appropriateness, and quality of care.   7    Access Appropriateness Quality   Supply Need Treatment Competence Expected Observed Factor level Patient  Urgency Condition   Standardized outcome (age, sex, comorbidities) Discharge state  Proximity   Length of stay   Contraindication  Readmission Facility Hours  Geographic remoteness Physician Size       Services  Technology Supporting staff Rurality    Proximity to care Standards of care Care relative to average facility Figure 2-1. The three pillars of health service distribution influencing patient outcomes I organized key elements tied to each pillar (i.e. urgency, proximity, hours, and services in the “Access” pillar) into patient and facility levels. These levels describe the source of the information for the element, and subsequently guide the modeling exercises. For example, urgency is a patient level factor because it describes patients’ states. Each pillar is also described in two dimensions according to the literature: access is about the supply of medical services meeting patient need; appropriateness is the comparison between the treatment provided and experience of the provider with that recommended according to relevant standards of best practices; and quality compares the observed outcome of treatment with the expected (did the treatment do what it was supposed to do?). The framework is not meant to imply causal pathways, though there is an inherent time element in the sense that access to a service precedes the quality of service delivery. Rather than fixing specific causal pathways, this framework summarizes how the literature describes each pillar within the health care system. Each pillar is explained in detail below using seminal works. 2.3 Access Access, also described as “accessibility” in the literature, is the ability to receive care when it is needed53,54 and the result of the “relationship between the location of supply and the clients’ [need], taking account of client transportation resources and travel time, distance and cost.”49 Access is typically measured in the ability to obtain urgent and emergent care on evenings and weekends, total hours of coverage, proximity to home, and the ability to begin services whenever the patient wants.21,28 By extension, “the proof of access is use of services, not simply the presence of a facility.”55 8  Many studies use proximity to patients’ homes, measured in time to care, as the overall proxy for access to care. The term golden hour has been used since at least 1969 to define the critical period of time to acute care following a time-sensitive event (typically trauma).56 However, there is a dearth of research examining the costs and benefits of attempting to deliver patients to definitive care within the hour, particularly in light of improved at-scene stabilization and inter-facility transfer methods.56 In response to the limited research supporting the golden hour, a small number of studies have shown that patients stabilized en route or at an outpost before inter-hospital transfer have comparable outcomes to patients within major urban facilities’ catchments, and superior outcomes to patients travelling longer distances to access definitive care directly.57–59 At the same time, other studies caution against stabilization at inappropriate facilities for urgent trauma incidents.60–63 Given that a growing proportion of BC patients experience inter-hospital transfers41 it is important to understand how this modification in service delivery, complicated by shifts in facility availability, influences patient outcomes. 2.4 Appropriateness The appropriateness of care is “the extent that the expected health benefits of a procedure exceed its expected negative consequences by a sufficiently wide margin [such] that the procedure is worth doing.”64 This is determined by assessing the efficacy, risks and costs of a procedure or treatment, and is influenced by the level of care offered by the health facility; the supply and training level of staff, general practitioners, and specialists; availability of key health technologies; geographic remoteness; and community level characteristics as they relate to resource needs.65,66 Despite variability in the context of care provision, appropriateness of care is often standardized with protocols that evolve regularly over time. These standards reflect the two key parameters of appropriate care (see Figure 2-1): the treatment or procedure provided, and the experience of the providers. For the purposes of this dissertation, I rely on provincial or where necessary, national standards and protocols to assess whether or not appropriate care was provided. 2.5 Quality Quality of care is defined by “the degree to which health services for individuals and populations increase the likelihood of desired health outcomes and are consistent with current 9  professional knowledge”.67 Thus, given this dependence on current professional knowledge, there is overlap between the concepts of quality and appropriateness. However, beyond receiving the appropriate care given a patient’s condition, the quality of care can be interpreted as whether the care results in the expected outcome. Thus, quality of care for the purposes of this dissertation is an outcome measure rather than the often-used process measure which overlaps extensively with appropriateness of care. Put differently, given standard protocols and patient profiles, quality describes the probability of favourable patient outcomes. If this is so, then the patient outcome as observed can be compared to the expected outcome. Aggregated at the facility level and risk adjusted for patient profiles, this can describe whether or not hospitals meet a minimum threshold for quality. Several studies have found patients admitted to large facilities have better outcomes than their counterparts at smaller centres and surmise that quality is tied to patient volume.68,69 The researchers suggest that differences in staff experience acquired over repeated exposure to particular patient profiles, and access to cutting edge technology influence the quality of care received.70 An exhaustive systematic review examining the relationship between volume and outcome found that the rigour of the studies involved was heterogeneous, with modest effects overall.69 Many of the studies (44%) failed to include risk adjustment as part of the analysis, and the systematic review did not explore the types of interventions used to treat patients. However, despite these methodological shortcomings, the belief that volume explains gains in quality persists. The reasoning holds that medical teams of larger facilities are able to specialize their skills where staff of smaller sites must be generalists. Thus, two patients receiving the same care may have different outcomes owing to how well it was provided characterized by the competence (or experience), comfort and skill of the provider. Similarly, some literature describes the capacity of facilities as predictors of quality. As such, larger hospitals which have a wider variety of specialists available to address the diverse needs of patients are thought to be more likely to provide better quality care.9 Lastly, publications as recent as the mid 2000s71 hold that patients treated in rural centres tend to fare poorly compared with similar patients treated in urban facilities. Reasons for this include failure to adhere to appropriateness of care standards, physician mix and skill, and lack of resources. However, a 10  growing body of literature is contesting the belief that urban hospitals outperform rural facilities and explains discrepancies in past findings as a result of unmeasured confounding effects.72–74 Size of the facility notwithstanding, these authors acknowledge that the location of the facility (urban, peri-urban, rural) may influence patient outcomes owing to differences in patient access patterns (e.g. rural patients may put off travelling to a hospital until the severity of their incident crosses a threshold higher than their urban counterparts due to the distance they must travel), resource availability (e.g. staff and equipment availability) and geographic isolation which may influence inter-hospital transfer patterns. However, the influence of these factors on the quality of care seems to be much smaller than once thought and is not always obvious in the literature.73 2.6 Time sensitive medical conditions of interest Service changes have directly affected the time to care for patients. Given that, the most pointed test of effects on patient outcomes will come from examining services that are time-sensitive. Three time-sensitive medical events and their outcomes will be used to measure how the redistribution of hospital services has influenced BC patients’ health. These are acute myocardial infarctions (AMI), ischemic stroke, and trauma.18–20,25,75–78 The conditions were selected because their outcomes are heavily tied to the access, appropriateness, and quality of care received 79,80 and are relevant (most significantly trauma) across the entire population. Each of the next three sections briefly highlights where appropriate treatment can be provided within the province, but Chapter 5 is dedicated to describing in detail the recommended treatments within the context of service delivery across BC. 2.6.1 Acute myocardial infarction Aside from AMIs being time sensitive medical emergencies, several factors lend AMIs well to inclusion in this study. Cardiovascular disease (CVD) is the leading cause of death for Canadians over the age of 65 and the second most common cause of death overall.81–83 Within CVD, AMIs make up a significant proportion of the burden of disease with an age-standardized rate of 207 events per 100,000 Canadians in 201284 and account for approximately 23% of all CVD deaths85. Although the number of deaths from CVD are equal between males and females, males experience more AMIs and at a younger age than females but females have a higher in-hospital mortality rate.83,86–90 This may be a result of differences in AMI symptoms between the sexes leading to poorer early recognition and first response times for females; or subsequent treatment upon 11  hospital admittance.88,89,91 Additionally, regional differences in both the occurrence and survival rates of AMIs have been observed, with rural and remote communities experiencing higher incidence and poorer outcomes.92 This may be due to differences in demographics (e.g. age, sex, and socioeconomic status of populations), underlying health risks, access to care, or quality of services; all factors that will be investigated in this dissertation. Despite AMIs being a difficult condition to diagnose accurately in the field without electrocardiographs and blood work, the sensitivity and specificity of troponin tests within hospital settings identify ST Segment Elevation Myocardial Infarctions (STEMIs, the focus of this dissertation) with great reliability.93 Lastly, if identified and treated appropriately within an hour and a half of symptom initiation, prognosis is significantly improved.94–97 Ideal treatment of an AMI starts with reperfusion (return of blood to the heart by removing the thrombus) and may be followed with a procedure to ensure the artery remains unblocked. Reperfusion can be achieved in the short term for most patients through pharmacotherapy (use of a thrombolytic, clot busting drug). Longer term restoration of blood flow is achieved through the more invasive mechanical procedures of balloon catheterization and stent insertion, or coronary artery bypass graft. Approximately 60-80% of patients with a STEMI need catheterization to diagnose or definitively identify the extent of damage following thrombolysis, and to determine the need for subsequent treatment.98 British Columbia has five hospitals that offer tertiary level cardiac care (catheterization options): Kelowna General Hospital, Royal Columbian Hospital, Royal Jubilee Hospital, St. Paul’s Hospital and Vancouver General Hospital. The facilities are located within major urban centres, all in the south of the province.99 As such, care for residents of the Northern Health Authority and those not within a reasonable drive to this small set of hospitals may need to incorporate inter-hospital ground transfer and/or air transportation to receive definitive care.8 The combination of high incidence, outcomes contingent on time to definitive care, uneven distribution of services, clear diagnostic criteria, and high quality data make AMI outcomes an ideal litmus test for the impacts of redistribution of care on patient outcomes. 2.6.2 Stroke Cerebrovascular disease is the third leading cause of death within Canada, a leading cause of morbidity, and disproportionately affects older Canadians.81 Acute ischemia underlies the majority 12  of strokes worldwide (72-87%).100 Similar to AMIs, the primary objective of treatment is to restore blood flow to the area of the brain affected by an arterial occlusion.101 The sooner care is received the more favourable the outcome, including reduced length of stay, shorter rehabilitation, greater return of neurological functions, and reduced mortality.102–104 Barriers to timely treatment include late presentation, contraindications with recombinant tissue plasminogen activator (rtPA) or endovascular therapy,105 and access to ground and air transport.78 Over the duration of the study five comprehensive stroke centres and five primary stroke centres have been established in British Columbia. Vancouver General, St. Paul’s, Kelowna General, Royal Inland, and Victoria General Hospitals all serve as comprehensive stroke centres, providing patients with advanced thrombolytic treatment, neurosurgical options, stroke units, advanced neurovascular imaging capabilities, and interdisciplinary stroke teams.106,107 These centres are also responsible for coordinating stroke services across the province. The primary stroke centres provide acute thrombolysis, have stroke units and interdisciplinary stroke teams (although not as complete or available as in comprehensive centres), and computed tomography (CT).106 Primary stroke centres are Cariboo Memorial, East Kootenay, Kootenay Boundary Regional, Penticton Regional, and Vernon Jubilee Hospitals.108 Given that treatment with tPA is predicated on CT or magnetic resonance imaging (MRI) scans, and thrombolysis is the preferred treatment for anyone under the age of 80, the vast majority of patients needing treatment must present to either a comprehensive or primary stroke centre for definitive care. Of the three conditions studied here, stroke care is the only one to benefit from the introduction of remote telehealth services during the study interval. Telestroke services first began on July 7th, 2009 and have expanded to become one of the key elements of the BC Stroke Strategy.109 Telestroke relies on telecommunication methods to link referring and consulting sites for real-time assessments of stroke patients. This improves triaging of suspected stroke patients and is a mechanism for increased access to stroke treatment at facilities which lack specially trained on-site personnel. It is supported by a formally organized network of continuously available staff for consultation.110 Its primary focus is to increase the use of tPA at rural and remote sites. Some northern facilities first reliant on the program have since become self-sufficient with tPA administration thanks to the established 24/7 neurology on-call service.111 13  While redistribution contributed to the closure of many facilities, it also resulted in more concentrated and specialized centres. Stroke care has been one area that has potentially benefitted from the increased level of dedicated resource allocation. This makes it an ideal candidate to explore shifts from generalist to specialist approaches within the resource distribution question. Stroke also presents the opportunity to investigate how telehealth services have influenced patient health outcomes in recent years. 2.6.3 Trauma Trauma is the leading cause of death in Canadians younger than 45 years of age and is an important cause of long-term disability.112 In 2008 there were 14,065 major injury cases reported across Canada.113 Of these, there were 1,605 deaths (11.4%) either in emergency departments or in-hospital.113 The large majority of trauma (79%) in Canada is attributed to motor vehicle collisions (41%) and falls (38%).113 Furthermore, injury rates due to motor vehicle collisions,114–116 suicide rates, and occupational injury116 and injury death rates from all causes117 are higher in rural and remote communities. This may be attributed to a combination of factors including prolonged discovery times, delays accessing emergency medical services, the training of available emergency medical services (EMS), and access to hospital care and definitive or tertiary trauma services.25,115,118 The Canadian Association of Emergency Physicians (CAEP) developed the Canadian Emergency Department Triage & Acuity Scale (CTAS) to describe patients and align them with appropriate care. The most severe patients are triaged with a Level I condition requiring immediate intervention to manage a threat to life or limb. Level II patients are those with emergent needs to a potential threat to life or limb and require rapid medical intervention, while Levels III-V are progressively weaker in their urgency of medical care needs. Across Canada, hospital trauma capacity is defined by the maximum degree of care that facilities can provide based on the Trauma Association of Canada’s five-point scale which lines up with the patient’s CTAS score. Major hospitals that can treat the most severe patients are Level I centres.119  In comparison, smaller urban or rural centres that can manage minimally complex patients are Level III and outposts capable of stabilizing patients are Level V. Hameed et al. recently showed that approximately 22.5% of Canadians live more than one hour’s drive from a Level I or II centre. In BC, a similar estimate holds for the five Level I centres 14  managing the more than four million residents.120 To remedy this, the province has developed an integrated network of local hospitals (Level III and IV) to provide interim care and triage to Level I centres for definitive care.120 British Columbia has six hospitals which serve as a Level I or II trauma centres that can provide definitive care. These are Vancouver General Hospital, Royal Columbian Hospital, Victoria General Hospital, Royal Jubilee Hospital, Royal Inland, and Kelowna General.120 Trauma events and their outcomes were selected for inclusion in this study because they are sensitive to access/timeliness of treatment and can highlight access issues for younger Canadians. 2.7 Summary While there are many influences on a population’s health, the focus of this work is on the health care system and the effects of changes in service organization and delivery. The health care system responds to health care needs through a sequence of decisions and events. Although these decisions typically happen quickly and seamlessly in a single location, at times the process can take longer and involve multiple facilities and providers. Changes in health care delivery can disrupt or modify this process impacting patient outcome for time sensitive events. Where access to care is diminished, patient outcomes can be negatively affected. However, the system may adopt new treatment methodologies to compensate for losses in access (e.g. an increase in stabilization at small hospital followed by inter-hospital transfer, or telehealth services such as telestroke). There remains uncertainty regarding the impact of service redistribution on patient outcomes, in part because of methodological limitations. At least one study claims that redistribution decreases access to acute care services for already vulnerable populations, and implies that this excess travel burden  contributes to poor patient outcomes.38 Others find the centralization of services has had little effect on access,37 improves procedural outcomes34,121 and does not affect  mortality.122 To determine what, if any, changes to access, appropriateness, and quality of care have resulted due to health service redistribution (see section 1.2.1) is thus complex and requires disentangling multiple factors and their relationships with one another. This thesis uses the framework outlined in Figure 2.1 to help anchor analyses that attempt to understand the effects of changes in hospital service delivery. While the framework is broad and provides definition, it is important to acknowledge that the concepts represented (appropriateness, access, quality) cannot be measured 15  fully. In part, this is because of the limitations of existing data available for this research. It is also because the concepts are large and thus it is natural that only aspects of them can be captured in any operational definition.   Nevertheless, the complexity of the question, as illustrated here, is no reason to avoid attempting to answer it. To do so, I will focus on the time-sensitive events of AMI, trauma and stroke and their outcomes, and have organized my thesis to first operationalize and examine each pillar independently (Chapters 4 to 6) before bringing them together for a final examination. It is important to understand how the redistribution of hospital services has affected BC patients given the comparability of challenges brought by geography in other parts of Canada, and the ongoing debates of what constitutes equitable care delivery. 16  Chapter 3: Data sources and measures This thesis addresses the question “What effect did British Columbia’s redistribution of hospital services have on residents’ outcomes, and did quality and appropriateness of care compensate for changes in access to acute care?”  This study uses a quantitative analysis of existing administrative data to answer this question. This requires the construction of relevant cohorts for analysis and the development of variables that can represent, or at least approximate, access, quality and appropriateness. Each of these variables is developed in more detail in subsequent chapters. The present chapter provides information on the data used and definitions for the analytic cohorts. 3.1 Data preparation I used five datasets to assemble pertinent details of acute care episodes within British Columbia just prior to and following the centralization of hospital services within the province. Four of these data sets are maintained by the BC Ministry of Health, are provided through Population Data BC (PopData), and span the fiscal years April 1 1999 to March 31 2013. I requested data beginning in fiscal year 1999 because I wanted to capture baseline measures of how the health system was providing acute care before hospital closures began in 2002. The fifth data set includes records of ambulance transports provided by British Columbia Emergency Health Services, which is responsible for pre-hospital emergency services for the entire province. Owing to data quality concerns, these data were only provided for events that occurred between April 1 2009 and March 31 2013. 1. Population Data BC’s Discharge Abstract Database (DAD, Hospital Separations Files) captures pertinent administrative and clinical information on all in-patient and day surgery discharges for BC residents. Each record contains information on the primary (medical event) and subsequent (comorbidities) diagnoses, medical procedures, level of care, hospital(s) of treatment, discharge state, admission and separation dates, and primary physician of care.123 2. The accompanying Consolidation file (to the DAD) supplements each patient’s record with the individual’s sex, age, and neighbourhood socioeconomic status.123 17  3. The Medical Services Plan (MSP) Payment Information File captures services billed by physicians at outpost facilities. Of particular interest are telehealth transactions tied to the time-sensitive medical events such as telestroke’s early diagnosis and treatment with tPA.124 4. Vital Statistics Mortality, provided by the Ministry of Health through Population Data BC, is used to capture mortality that occurred outside hospitalizations (e.g. post hospital discharge or for patients never admitted to hospital). 5. BC Emergency Health Services’ (BCEHS) Ambulance data provides details on diagnosis at both dispatch and arrival, time to patient, time to facility (duration of trip), en route procedures and stabilization measures (such as airway management and intravenous treatment), and transfers made. As en route care grows as a strategy to deliver timely care for conditions such as AMI, there is interest in equipping ambulances with ECGs and training paramedics to administer thrombolytics.125 Additionally, since the timing of first receiving thrombolytics, along with the travel time to the closest hospital providing percutaneous coronary interventions (at time of pick-up) dictates where the patient is taken, this information is crucial for understanding the travel decisions made by providers. Population Data BC is responsible for linkage of data sets. Researchers receive record-level data, including a study ID that is consistently applied across datasets. In the research process, datasets were merged using that study-specific identification number. In other words, records between independent files (DAD, MSP, Mortality, and Ambulance) were joined only where study IDs were exact matches. All administrative data sets except for the Ambulance data have been previously tested for quality control (submitted data matches internal hospital records of event).123,126 In particular, nonclinical data (e.g. age, sex, date of admission) is reported with high reliability (>99%)126 and major difficulties in setting linkages were not experienced. 3.1.1 Study sample As this study is a population level analysis, all patient records within the study period were included for analysis. In this way, potential selection bias was virtually eliminated, the heterogeneity of patients was maintained, and the precision of reported differences was strengthened. In all cases, the number of parameters in the models used does not exceed 10% of the number of cases observed, which helps to ensure the stability of the models.127,128 All final 18  models were first developed and tested on random subsets of the data before they were re-run using the entire sub-cohort (where appropriate) to report measures of effects. This was done to ensure convergence criteria of models were met and to test complicated relationships between covariates in the most efficient way possible. More details are provided in Chapter 8. 3.2 Inclusion and exclusion criteria The study population includes all adult (eighteen years of age or older) residents of British Columbia who were treated within the province for an AMI, stroke or trauma event between April 1 1999 and March 31 2013. All variable development and analyses are condition-specific. The cohorts stem from two unique data sources: the Discharge Abstract Database (DAD) and the Vital Statistics Mortality Database. All events were captured using international classification of disease (ICD) 9 or 10 codes from the diagnosis fields in the DAD, or from the underlying cause of death (UCOD)i or record axisii codes in the Mortality Database. For both AMI and stroke, inclusion criteria were simple. The AMI cohort was comprised of all events with a primary diagnosis, UCOD or record axis code of ICD-9: 410.x or ICD-10: I21-I22.x. Based on validation studies, these codes best captured heart attack events.129 Similarly, the stroke cohort captured all events with a primary diagnosis, UCOD or record axis code of ICD-9: 430.x, 431.x, 433.x1, 434.x, 436, 362.3, or ICD-10: I60-I64.x, H34.1, or G45.x.130,131 Between April 1 1999 and March 31 2013 there were 96,672 unique AMI events of which 1,985 (1.84%) died without a hospital admission (even if post-mortem). Similarly, there were 94,863 ischemic stroke events of which 694 (0.73%) died without admission. For the trauma cohort, inclusion was based on a two-step selection process to ensure that a cohort comparable in urgency to that of AMI and stroke events was used in subsequent analyses. The reason for this was that while AMI and stroke events are considered medical emergencies where time to care is critical to survival, trauma is a much larger classification and can include incidents that do not pose immediate risk to life or limb but still result in short term hospitalization.                                                  i The disease or injury which initiated the series of events leading to death. An algorithm determines which ICD-9 or 10 code is the underlying cause that led to the cascade of events. ii Record axis codes summarize the overall medical certification of the death certificate. Multiple conditions are listed and then organized by standardized disease classification procedures. The first code in this field is usually the same as that listed in the UCOD. Redundant conditions are deleted and certain conditions are combined. 19  Since the focus of this study is effect of changes in service redistribution on time-sensitive emergencies, it was important to restrict analysis to comparable trauma events, particularly since this cohort was the only one of the three that was representative of the entire adult BC population (i.e. included younger adults). Many studies that look at trauma outcomes and access to care use the Injury Severity Score (ISS) provided by participating hospital or trauma networks to create cutoffs of urgency (e.g. restricting events to those with an ISS≥15).25,132–134 However, both the ISS and the variables used to create it were not available in the administrative data sets. As such, the Canadian Institute for Health Information’s recommendations for inclusion and exclusion criteria were used. Severe trauma events were any record with a primary diagnosis of ICD-9: 800-959.x or 900-999.x, or ICD-10: S00-T35.x or T66-T79.x and with any of the following: an external cause listed in Table 3-1, admission lasting at least two days, or admission to the intensive care unit (ICU).132,135,136 All mortalities with an UCOD or record axis code of ICD-9: 800-959.x or 990-999.x, or ICD-10: S00-T35.x or T66-T79.x were also included. Medical events as a result of poison, deprivation, medical interventions or drugs were excluded (see Table 3-2) irrespective of the length of hospitalization or admission to ICU. Although these events are captured within the range of ICD-9 and ICD-10 trauma codes listed above, they are not considered to be trauma incidents and are thus regularly excluded from studies looking at trauma events and access to care, as is the case here. Between April 1 1999 and March 31 2013, 220,151 trauma events occurred, and only 751 (0.34%) were never admitted to hospital (even if post-mortem). Condition ICD 9 Code ICD 10 Code Transport incident E800-E849 V01-99 Land transport incident E800-E829 V01-06, V09-V90 Water transport incident E830 V91-V94 Air and space transport incident E840-E845 V95-V97 Other unspecified incident E846-E849 V98-V99 Unintentional fall E880-E888 W00-W19 Exposure to inanimate mechanical force E916-E920 W20-W46, W49 Exposure to animate mechanical force E922 W50-W60, W64 Unintentional drowning or submersion E910 W65-W70, W73-W77, W81 Other unintentional threats to breathing  W83, W84 Exposure to smoke, fire, flames E890-E899 X00-X06, X08, X09 Contact with heat/hot substance E921, E923, E924 X10-X19 Exposure to forces of nature E926 X30-X39  E925  20  Condition ICD 9 Code ICD 10 Code Overexertion and strenuous or repetitive movements E927 X50 Prolonged stay in weightless environment E928 X52 Unintentional exposure to other unspecified factors  X58-X59 Intentional self-harm, excluding poisoning E950-E959 X70-X84 Assault excluding poisoning E960-E969 X86, X91-X99, X00-X05, X07-X09  E979  Event of undetermined intent, excluding poisonings E980-E989 Y20-Y34 Legal intervention and operations E970-E978 Y35-Y36 Table 3-1. National trauma registry comprehensive data classification of external cause of trauma codes defined as urgent  Condition ICD 9 Code ICD 10 Code Inhalation of gastric contents E911-E915 W78 Inhalation and ingestion of food causing obstruction  W79 Inhalation and ingestion of another object causing obstruction  W80 Contact with venomous animals or plants E900-E909 X20-X29 Unintentional poisonings and exposure to noxious substances E860-E869 X40-X49 Poisoning of undetermined intent E850-E858 Y10-Y19 Assault by poisoning  X85, X87-X90 Travel and motion  X51 Lack of food  X53 Lack of water  X54 Unspecified privation  X57 Neglect and abandonment  Y06 Drugs, medicaments and biological substances causing adverse effects in therapeutic use E930-E949 Y40-Y59 Misadventures to patients during surgical/medical care E870-E876 Y60-Y69 Medical devices associated with adverse incidents in diagnostic/therapeutic use  Y70-Y82 Surgical and other medical procedures as cause of abnormal reaction and complications E878-E879 Y83-Y84 Sequelae of external cause of morbidity/mortality  Y85-Y89 Table 3-2. ICD-9/10 external cause of trauma codes used as exclusion criteria in capturing trauma cohort The use of both datasets (DAD and VS Mortality) to build the cohort, while largely overlapping, allowed the inclusion of persons who died from a time-sensitive condition before being admitted to hospital. Thus, while the majority of patients were found in the DAD, including Vital Stats Mortality ensured that the few patients who were declared dead by paramedics or physicians prior to admission (e.g. dead on arrival at scene, died en route to care, or died in the emergency department of a hospital) were included in the analyses. This alleviated concerns of selection bias 21  that systematically excludes persons unable to reach appropriate medical care. Since the primary objective of this dissertation is to understand the role that redistribution of services has played on health outcomes, and this manifests itself in changes to access, it is critical that all residents of BC who experienced a time sensitive event are included. The data were organized in long form to aid in variable development and allow for simple aggregation of records for analyses. This means that each individual line was a specific record of an event defined by the admission date and hospital of the first admission, or the record of death in cases where the patient died without admission to hospital. The implication is that a single individual may be represented in the data set more than once, e.g. if a patient experienced both a trauma and an AMI during the study period. The complete patient cohort consisted of 411,686 records of hospital admits and deaths outside the hospital related to incidences of AMI, stroke and trauma. 3.3 Variable development 3.3.1 Patient level characteristics Both patient age at time of medical event and sex are directly available in the data provided by PopData (DAD and VS Mortality datasets). However, the socioeconomic status (SES) of the individual patient is not. Instead, an ordinal variable describing the patient’s neighborhood income as a proxy of SES is included at both the quintile and decile levels of the national range. There is strong evidence demonstrating that a person’s SES is similar to the community within which he or she lives. Thus, it is reasonable to assume that the neighborhood SES level provided for each patient is a reasonable proxy for the true SES of that person.137–139 Similarly, the extent to which the patient’s neighborhood is urbanized is provided by the Statistical Area Classification groupings categorizing the community as a metropolitan (urban), census agglomeration or census metropolitan influenced zone (MIZ). Census metropolitan areas and census agglomerations are considered urban. Meanwhile MIZ are peri-urban to non-urban depending on their proximity to metropolitan areas and the extent to which their population works within these areas. Communities classified as “no MIZ” mean that very few to none of the residents are employed in urban areas and thus the community is recognized to be remote. Finally, territories are also non-urban and described as remote.140,141 22  3.3.1.1 Comorbidity Patient comorbidity was calculated using a revised Charlson comorbidity index. The original Charlson comorbidity index served to quantify burden of disease using a series of comorbid classifications and weights to predict one-year mortality.142 First developed in 1987, it predicted one-year mortality by accounting for both the number and seriousness of patients’ comorbidities based on chart reviews. The adjusted relative risks determined from 10-year survival data were used to examine and develop the index for diseases whose relative risk of mortality within the year exceeded 1.2 after adjusting for age.142 The final weighted index is show in Table 3-3. Assigned weight Conditions 1 (1.2≤ RR < 1.5) Myocardial infarct Congestive heart failure Peripheral vascular disease Cerebrovascular disease Dementia Chronic pulmonary disease Rheumatologic disease Ulcer disease Mild liver disease Diabetes without chronic complications 2 (1.5 ≤ RR < 2.5) Hemiplegia or paraplegia Renal disease Diabetes with complications Any malignancy, including leukemia and lymphoma 3 (2.5 ≤ RR < 3.5) Moderate or severe liver disease 6 (RR ≥ 6.0) Metastatic solid tumor AIDS/HIV Maximum score: 29  Table 3-3. Charlson Comorbidity Index weights assigned to each condition  RR=relative risk  An individual’s Charlson comorbidity index was thus the sum of the weights based on his or her comorbidities and an additional point for each decade over the age of forty up to a maximum score of 29.142 The relative risk of mortality for each additional point on the comorbidity index was 2.3 (95% CI: 1.9-2.8) and for each decade of age 2.4 (95% CI: 2.0-2.9).142 The Charlson comorbidity index has been widely used across populations, repeatedly validated143–145 and at times modified for specific purposes146. However, it was not updated to accommodate advances in the effectiveness of treatments until 2010 when a team of Canadian researchers at the University of Calgary reassigned weights using discharge data from Australia’s 23  Victoria State, and national data from Canada, France, Japan, New Zealand, and Switzerland.147 This re-evaluation found that myocardial infarction, diabetes without chronic complications, peripheral vascular disease, peptic ulcer disease, and cerebrovascular disease were not associated with one-year mortality following hospitalization.147 The revised Charlson comorbidity index has since proven to align better with other comorbidity measures developed using administrative data.148 Thus, I used Quan et al.’s updated Charlson comorbidity index (see Table 3-4) to tabulate each patient’s comorbidity level. Assigned weight Conditions 1 Chronic pulmonary disease Diabetes with chronic complications Renal disease Rheumatologic disease 2 Congestive heart failure Dementia Hemiplegia or paraplegia Mild liver disease Any malignancy, including leukemia and lymphoma 4  Moderate or severe liver disease AIDS/HIV 6 Metastatic solid tumor Maximum score: 24  Table 3-4. Revised Charlson Comorbidity Index with modern listing and weight assignments   This revised score uses 12 conditions instead of 17, has adjusted the assigned weights, and has a maximum score of 24 compared with the original index’s maximum of 29 points. Additionally, no points are assigned for age. Instead, when determining patients’ overall health profiles for use in case mix adjustment, age is now an independent variable.147 While some users of the Charlson comorbidity index have found it useful to list each comorbid condition independently in a fitted model as this should enhance the ability to risk adjust given very specific populations, Quan et al. recognized that studies with small sample sizes will find that “rare comorbidities (such as AIDS/HIV) may cause instability in model performance when included in the model as dummy variables”.147 As such, since I am looking at hospitals of varying sizes, I too opted to use comorbidity scores to adjust for patients pre-existing conditions as part of case mix instead of listing the 12 individual comorbidities independently in my models. Thus, each patient’s comorbidity score ranging from 0 (no comorbidity) to 24 (maximum comorbidity) was added to the overall case mix model for the risk-standardized mortality ratio calculations 24  performed in Chapter 5. Importantly, the Charlson comorbidity index has only been applied to adult patients (≥18 years of age).142 Thus, for the purposes of this dissertation, and owing to the complexity of patient management of children, I too restricted my patients to adults.149,150 3.3.2 Hospital characteristics For the purposes of this study, a hospital is any facility that offers some level of emergency department services. At a minimum, this means any facility that provides 24/7 access to a physician and needed diagnostic equipment where the patient's condition can be stabilized before being transferred to definitive care. This definition is an extension of CIHI's, which identifies acute care hospitals as those that must submit data to the province's DAD. My more generous definition of a hospital expands the realm of eligible facilities to include those with appropriate diagnostic equipment such as electrocardiographs (ECG/EKG) and access to an experienced physician. At the same time, extended care facilities and residences are excluded. Over the study period, 101 hospitals were identified as being operational and capable of providing minimum thresholds of care according to standard protocols for my time sensitive events. This includes 100 hospitals identified at the beginning of the study and the Abbotsford Regional Hospital and Cancer Centre which opened in 2008. 3.3.2.1 Condition-specific volume While the DAD includes hospital size as one of the variables offered, it was found to be incomplete the majority of the time. To remedy this, I created a condition specific volume variable based on the tabulated number of patients admitted per condition per year.149,150 3.3.2.2 Hospital peer groups A hospital peer group variable was created based on CIHI’s designations.151 Teaching hospitals are facilities with confirmed teaching status from the provincial Ministry of Health or identified as such in the Canadian MIS Database. All other hospitals are split into the following three categories: large, medium or small community hospital based on total patient volume (not condition-specific) and complexity. Specialty hospitals are excluded (e.g. cancer centres, children’s hospitals). However, unlike the CIHI peer group designation, I included an additional category for facilities visited by patients that are not acute care hospitals according to CIHI but are equipped to diagnose and/or stabilize patients. 25  3.3.3 Episode of hospitalization As mentioned, the dataset was set up to have each line represent a hospitalization record (or non-hospitalized death) based on the date of admittance and hospital identifier. This means that the data provided for analysis was not organized per person or even per medical event but rather per admission. The data needed to be restructured to link records by episode of hospitalization so that patients transferred between hospitals for same-day procedures (bidirectional transfers), and permanent transfers (unidirectional transfers) would not be analyzed as unique events but rather as a string of medical services provided for a specific medical emergency. Failing to do so would inflate the number of medical events in the province, and underestimate the resources needed to manage individual emergencies.152 Thus, two circumstances arose in the data described as episode of hospitalization and episode of care: 1. Episode of care: describes all subsequent care and outcomes tied to a unique medical emergency. This includes any care received prior to hospital admittance, the first hospital admitted to, subsequent inter-hospital transfers for same day and definitive care, and any readmissions that occur within 30 days of the last discharge date. 2. Episode of hospitalization: describes all the medical care received within an episode of care before being discharged for at least a day. Thus, readmissions are not tabulated here and do not contribute time to the calculated episode of hospitalization length of stay but inter-hospital transfers are included. The implication is that an episode of care may be made up of multiple episodes of hospitalization if the patient was discharged for more than 24 hours (but less than 30 days) before being readmitted to treat the same medical issue. To calculate these variables, all the observations in the dataset were sorted by studyid (uniquely identifying individual patients), diag (primary diagnosis), hosp (hospital) and addate (date of admittance). Patients with multiple hospital records for the same medical event were grouped together and sorted by date of admittance and facility such that the dates increased from earliest hospitalization to latest. Using the LAG function in SAS, the patient’s addate was compared with the previous line’s sepdate (date of discharge) using a temporary gap variable (gap = addate – previous sepdate). The value of the gap variable dictated how the subsequent record should be categorized. Contiguous hospitalizations with changes in diagnosis code are not included as they were never captured during cohort selection phase. 26  3.3.3.1 Length of stay per hospitalization and entire episode of care The length of stay (LOS) for an episode of hospitalization was calculated using the first addate of a series of hospitalizations within an episode of care and the final sepdate for the series (excluding readmissions). This ensured that LOS did not double count nested hospitalizations as would be the case if a strict summation function were applied within SAS based on records and lengths of stay per record. A complete length of stay (C_LOS) was similarly calculated using the first date of admission and the last date of discharge including readmissions for the episode of care. 3.3.3.2 In-hospital  The in hospital designation describes whether an event outcome (mortality) occurred while the patient was admitted in hospital, or after the patient was discharged to the community. It was determined using hospital discharge state, and confirmed using VS Mortality data to confirm deathdate > sepdate where deaths were flagged for occurring outside the hospital. Figures 3-1 to 3-3 show the proportion of events that resulted in a death within thirty days, and the location of that event. Over time, the proportion of deaths declined. However, the proportion of deaths that occurred after discharge flat-lined or slowly increased.   Figure 3-1. Proportion of AMI events that died within thirty days, by location of death (in or out of hospital) by fiscal year 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012Death during admission 11.5 11.94 12.6 10.74 10.36 9.85 10.18 9.57 7.97 7.17 6.25 5.78 5.87 5.48Death after discharge 1.98 2.39 2.3 2.72 2.49 7.03 5.63 6.11 4.53 4.34 3.95 3.94 3.3 3.56Cumulative 13.48 14.33 14.9 13.46 12.85 16.88 15.81 15.68 12.5 11.51 10.2 9.72 9.17 9.0402468101214161820Proportion of AMI events  that died within 30 days27   Figure 3-2. Proportion of stroke events that died within thirty days, by location of death (in or out of hospital) by fiscal year   Figure 3-3. Proportion of trauma events that died within thirty days, by location of death (in or out of hospital) by fiscal year  1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012Death during admission 14.9 15.04 14.3 13.48 13.23 13.57 12.89 13.59 15.21 13.69 11.75 12.36 10.74 11.29Death after discharge 2.51 1.94 2.09 1.76 1.98 2.79 2.89 3.15 3.14 3.48 3.48 3.11 3.19 3.47Cumulative 17.41 16.98 16.39 15.24 15.21 16.36 15.78 16.74 18.35 17.17 15.23 15.47 13.93 14.7602468101214161820Proportion of stroke events that died within 30 days1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012Death during admission 3.47 3.32 3.95 3.01 3.22 3.17 3.15 3.05 2.8 3 3.05 2.72 2.94 2.79Death after discharge 0.75 0.72 0.77 0.91 0.91 1.3 1.42 1.34 1.53 1.45 1.52 1.47 1.52 1.29Cumulative 4.22 4.04 4.72 3.92 4.13 4.47 4.57 4.39 4.33 4.45 4.57 4.19 4.46 4.0802468101214161820Proportion of trauma events that died within 30 days28  3.3.4 Access Referring to the theoretical framework illustrated in Figure 2-1, access is best understood as the relationship between supply and demand. This manifests itself as the patient’s need for care and the provision of said care. When all other variables are accounted for (patient need as defined by urgency and complexity; and provision in terms of hours and services available), proximity to care dictates the extent to which care can be accessed. Proximity is thus the focus in this dissertation, and is estimated using postal code centroid to hospital coordinate ground travel time measures described in detail in Chapter 4.  3.3.4.1 Inter-hospital transfers Patients who were first seen in an emergency department without an admission and then transferred to a hospital for definitive care were identified using the hospfrom variable in the DAD and no observed record of admission. These patients, as well as patients that experienced nested hospitalizations (e.g. day surgery) and permanent transfers were categorized in the dataset as inter-hospital transfers. However, their distinction persists and was necessary in the tabulation of the outcome variable length of stay as discussed below. 1. Nested hospitalizations were identified where gap<0. These occurred when the patient was transferred from one facility to another for a procedure (typically same day) and was returned to the original hospital. 2. Inter-hospital transfers were identified where 0≤gap≤1 or no record of admission for the first hosp was recorded. In such circumstances the patient was discharged from one facility, transferred and admitted to another as part of the continuum of care or they were seen in the emergency department before being redirected for subsequent admission. The gap in previous sepdate to current addate could be as great as one day to allow for transfers that occurred over midnight. 3. Single admissions occurred when the specific medical event did not span multiple hospitalizations and represented a specific episode of hospitalization. For subsequent analysis purposes, each observation record for the same episode of care included details of transfers and nested events. In this way, if a single observation was needed to represent an episode of care, any of the affiliated records for that episode would contain the 29  pertinent information. As will be explained later, both the first line per episode of care, and the last line were used for the creation and testing of quality of care variable. Of the 411,686 independent episodes of care: 3,430 (0.83%) were deaths without admission to hospital; 86,883 (21.10%) experienced at least one transfer between facilities; and 10,141 (2.5%) had readmissions distinguishing episodes of care from episodes of hospitalization. 3.3.4.2 Total travel burden Where length of stays excluded nested transfers in their calculation, total travel burden was tabulated by summing the distance traveled during each transfer for that hospitalization. This means that patients transferred for a nested procedure had both the distance to and from the hospital offering the procedure included in the total travel burden. Distances traveled for readmissions were not included in the total travel burden calculations nor were distances traveled to return home from hospitalization. 3.3.5 Appropriateness Small facilities and outposts are not equipped to offer the extent of services needed to meet urgent patient needs. For the purposes of this dissertation, this concept will be unpacked under the “appropriateness of care” umbrella of variables representing the care employed and its compatibility with patients’ needs. Using provincial and national protocols consistent with international guidelines and evidence based practices describing the equipment, staff and procedures necessary for each condition’s complete care, I developed thresholds of minimally suitable care per condition (AMI, stroke, trauma). For many patients, the first centre of care visited did not provide definitive care. Instead, such centres served to stabilize patients before transferring them to facilities better equipped to manage complex patient needs. That said, the appropriateness of care variable considered the full travel trajectory and extent of services provided en route to definitive care per episode of hospitalization. As standards of care evolved during the study period, I evaluated whether or not patients received appropriate care with respect to when the event occurred. In other words, I compared the services received with the existing standards of care for that condition at the time of the event. The details of this exercise are discussed in Chapter 5. 30  3.3.6 Quality Quality, in the context of acute care, attempts to describe the link between clinical practice and patient outcomes;153 in other words, whether the care received resulted in the intended effect.54 While the complexity of the undertaking is apparent (particularly because the outcome of a procedure is tied to the procedure’s appropriateness), the pursuit of a simple indicator that can distinguish between “good” and “bad” hospitals has led to the development of two popular risk adjustment techniques. The first, and more commonly used is the hospital standardized mortality ratio (HSMR).154–158 If quality influences patient outcomes, this measure attempts to adjust for the factors other than hospital processes that can affect patient outcomes. Thus, it standardizes the case-mix differences between hospitals before comparing the expected versus observed outcomes of patients within the hospital.74,159 Standardization normally calls for adjustment of both patient mix (age, sex, socio-economic status (SES), and comorbidity), and hospital traits (e.g. size9 and geographic remoteness71,74) using hierarchical regression methods159. Once completed, the observed outcomes of patients are compared with the expected outcomes. These expected outcomes are estimated using the HSMR factor calculated from the regression analysis based on the patient mix of that hospital compared with an average hospital in the sample. A hospital whose observed versus expected mortality ratio is lower than 1.0 is considered to offer “good” quality, whereas facilities whose ratio falls significantly higher than 1.0 are said to perform poorly and thus offer “poor” quality.153 Importantly, HSMRs are calculated using a large list of conditions that contribute to 80% of deaths in the hospital. Thus, when standardized, the mortality ratio is a generic evaluation of performance at the hospital that is not condition-specific.150 Additionally, HSMRs are calculated using in-hospital mortalities and thus are sensitive to discharge bias (the systematic early discharge of patients to improve the proportion of patients discharged alive).149,155,156 An alternative to the HSMR is the risk-standardized mortality ratio (RSMR). This measure uses the patient as the unit of analysis, and typically 30-day all-cause mortality as the outcome of interest.149,157,160,161 Similar to the HSMR, the RSMR compares the observed number of deaths at a hospital to the expected, once the outcomes have been standardized for patient and hospital-specific characteristics. However, unlike the HSMR, it uses episodes of hospitalization rather than individual hospitalizations for the same medical event. Lastly, it is condition-specific.150,160 For 31  these reasons, most importantly because the cohort used in this dissertation is restricted to patients who experienced an AMI, stroke and trauma and thus make HSMR calculations impossible, the RSMR is used for subsequent analyses. 3.3.7 Outcomes The primary outcome of interest throughout this dissertation is mortality. Research in the field162 suggests that a pre-defined mortality period be used to evaluate risk as it relates to hospital performance in part to avoid concerns of discharge bias. However, a fine balance must be struck in selecting the duration of the outcome variable. Too short, or restricted to in-hospital mortality, and discharge bias may be at play, too long and community effects of care may be responsible for variations in mortality. As such, 30-day mortality and time to mortality censored at 30 days were used for Chapters 7 and 8 respectively because they are the preferred measure used in the literature (making subsequent results comparable to other studies) and because mean length of stay for the three conditions used here can range from a few days to weeks with wide variation between facilities and over time.163–166 Thirty-day mortality per episode of care was calculated using the LAG function in SAS. I compared the first addate for the episode with when the patient died. If the patient died within 30 days (mortality date – addate ≤ 30), then the 30-day mortality flag=1, and time to mortality was calculated as the difference between the admission and mortality date (mortality date – addate). Additionally, thirty-day readmissions were calculated using the LAG function in SAS by comparing the current observation’s addate with the previous observation’s sepdate. Thirty-day readmissions were identified where the gap between sepdate and addate was 1 < gap ≤ 30 such that a patient would be discharged and not readmitted for at least a day but admitted within 30 days. As noted above, readmissions were tied to a specific medical event and previous hospitalizations, they were not analyzed as part of the episode of hospitalization. Instead, readmission rates were the aggregated result of individual episode of care readmissions for a given condition (AMI, trauma or stroke) within a specific hospital and/or during an interval of time (typically a year). 32  3.3.8 Summary Table 3-5 summarizes the variables developed for inclusion in subsequent models. Chapters 4 to 6 describe how the variables for access, appropriateness, and quality were conceptualized from the literature and operationalized for the purposes of this dissertation. 33  Variable Data Sources Method/Formula Values/units Comorbidity DAD Quan et al.’s revised Charlson comorbidity index 0 (no comorbidity) to 24 (maximum comorbidity) Condition specific volume DAD Total unique episodes of care for the condition, per hospital and year ≥0 Hospital peer groups DAD CIHI designations 1=Non-acute acute care hospitals (outside CIHI’s designation) 2=Small community hospitals 3=Medium community hospitals 4=Large community hospitals 5=Teaching hospitals Episode of care: C_LOS DAD Mortality Using LAG function: Last discharge/mortality date (excluding transfers) – first admit date Number of days Episode of hospitalization: LOS DAD Mortality Discharge/mortality date (excluding transfers) – admit date Number of days Inter-hospital transfers DAD Using LAG function: addate – previous sepdate 0 – 1 days Total travel burden DAD Travel matrix Sum of resident to hospital and all inter-hospital distances travelled Total km by road Total mins by road Thirty-day mortality DAD and VS Mortality Using LAG function: If (addate – mortality date)≤30 0 or 1  In hospital DAD and VS Mortality If (deathdate>sepdate) or discharge status≠dead, in_hospital=0 0 or 1 Time to mortality DAD and VS Mortality Using LAG function: addate – mortality date Number of days (can be blank) Table 3-5. Summary of variables developed for subsequent analysis, source of data, method used, and range of values 34  Chapter 4: Access to care Access is the ability to receive care when it is needed.53,54 It reflects the ability to obtain urgent and emergent care on evenings and weekends, total hours of coverage, proximity to home, and the ability to begin services whenever the patient wants.21,28 However, access is often reduced to a simple analysis of proximity to patients’ homes, measured in time to care and discussed as travel burden. This dimension of access as time to care is relevant in that it captures “access” that is most directly affected by redistribution of hospital services. I developed time to care estimates using proximity from patients’ residence to facilities equipped to diagnose and provide initial condition specific treatment. These facilities had to operate 24/7 and include at least one physician available to attend to emergent events at all hours. 4.1 Methods For the purposes of this dissertation, access was represented with three proxies of “burden of travel” necessary for spatial accessibility (also defined elsewhere in the literature as travel cost or burden)167: ground travel distance in kilometers, aerial distance (as the crow flies, Euclidean) in kilometers, and drive time in minutes.168,169 Euclidean and ground travel distance measures are computationally simple, and improvements in spatial data processing have made drive times easy enough to generate14 provided that the computational power and time to do so is available. Euclidean linear arc estimates were calculated using SAS 9.2 software (SAS Institute Cary, NC)170, while road travel times and distance estimates were determined using ArcGIS’s Network Analyst version 10.2 (ESRI Redlands, CA). 4.1.1 Travel matrix The objective was to calculate distances and ground time to care for each event. To ensure the privacy of all healthcare users included in the study, I generated a generic matrix of every possible BC postal code to hospital distance and time combination and provided this matrix to Population Data BC. PopData staff replaced the six-digit postal code with a study specific postal code ID also included in my data extract of patient records. I used deterministic matching to link each patient record to the appropriate series of travel burden calculations within the matrix via patient’s de-identified postal code and hospital fields in the DAD. This means that each patient who was admitted to a hospital was matched with a time to care, and two distance to care proxies – the 35  distance to the closest treating hospital and the distance to the actual treating hospital, which may or may not be the closest facility. Additionally, all patients including those never admitted to hospital, were matched to the closest hospital they could have visited based on the hospitals that were in operation at the time of their event as explained in detail in Chapter 7. Populating the travel matrix involved the following steps which will be described in turn: 1) identifying a list of relevant facilities and their geographic coordinates; 2) identifying all residential postal codes and centroid coordinates; 3) choosing a methodology for calculating distances: ArcGIS vs. Google Maps; and 4) trouble shooting where needed. 4.2 Identifying a list of relevant facilities and their geographic coordinates To generate the travel matrix, I relied on publicly accessible CanMap Postal Code Suite v2013.3 (formerly known as the Platinum Postal Suite) DMTI spatial files171 available through the University of British Columbia’s Abacus portal. These files included a set of latitude and longitude coordinates to represent the centroid of each six-digit postal code. The CanMap RouteLogistics 2005172 file provided the necessary road network files with speed limits and road restrictions. The dataset also included an exhaustive list of medical facilities with location details within the province. This list was compared with the Ministry of Health’s address list first effective in September 2002 and updated in 2014 to ensure no hospitals were missing for the travel burden exercise.173 The address list was especially useful because it organized facilities by original designation: hospital, extended care, diagnostic and treatment, and outpost. The CanMap list duplicated and included additional health centres. Combining the two lists and restricting to facilities that offered some level of emergency services to the public reduced the list of operational hospitals available in 2001 to 100 facilities. Excluded facilities included the Canadian Red Cross Society Outpost in Vancouver, several children’s centres, group homes, community care clinics, and mental health centres. We excluded these facilities because of their irrelevance to the study question. To capture any facilities that had become operational after 2001, an online search of BC hospitals by health authority was conducted. The Abbotsford Regional Hospital and Cancer Centre, which opened to patients in 2008, was added bringing the final list of hospitals for inclusion to 101. The location of each hospital was cross-verified and latitude and longitude coordinates were added using Google Maps (Google Inc. Mountain View, CA USA) where needed.174 36  4.2.1 Identifying all residential postal codes and centroid coordinates Of the 185,000 postal codes available in the files, 109,887 unique postal codes with geographic coordinates to six decimal points of precision (within 2 meters) were included in the distance calculation matrix: 101,329 currently active postal codes, and 8558 postal codes retired no earlier than ten years prior to the commencement of the study. The remaining postal codes had either retired before 1989 or were duplicates. For postal codes with multiple sets of coordinates available, the single link indicator (SLI) was used to identify the main postal code record.175 Postal codes retired within 10 years of the study start were retained in the matrix because it is unclear how soon after a postal code is decommissioned by the Canada Post Corporation residents stop using it. 4.2.1.1 ArcGIS versus Google Maps The matrix that needed to be generated was very large (109,887 postal codes x 101 hospital locations = 11,098,587 unique residential to hospital calculations, and 10,100 hospital to hospital combinations). There were two possible strategies for generating postal code to hospital distance calculations and populating the matrix. These were to use Geographic Information Systems (GIS) software such as ArcGIS to approximate time to care and distance given road conditions, barriers and speed impositions, or to calculate the real time distance and travel time available through Google Maps’ Transport Time Estimate. GIS software is regularly used to generate burden of travel estimates for studies with geospatial components. As part of ArcGIS’s suite of tools, there are two methods available for modeling spatial networks: network analysis and raster analysis. Network analysis is a vector based modelling method which uses existing road network files with speed limits and other geographic features to find the quickest or shortest path as prioritized by the user.14 Conversely, raster analysis uses a series of regularly sized cells to form a travel grid and requires more intervention on the part of the user to specify the cost of traveling between cells (via speed impositions and travel restrictions.14 Network analysis is best suited for path calculations and approximates where the true travel environment is reliably mapped out. Alternatively, raster based analysis is superior in poorly defined service areas (i.e. regions lacking detailed transportation networks).14 Given that I had detailed road network files for the province, network analysis better suited my needs. Google Maps is quickly growing as a popular alternative in the geospatial community to the extent that it now offers a plug-in feature compatible with various software (including SAS) to 37  allow researchers to quickly generate travel burden calculations and analyze them with their preferred statistical software. In my case, I did not have access to the plug-in. Instead, I had access to a SAS macro developed by Mike Zdeb’s team for the SAS Community that pulls coordinate pairs from a table and runs them through Google Maps. This workaround allows the semi-automated extraction of needed coordinates to calculate travel and runs in real time. The major limitation of using the free Google Application Programming Interface (API) is Google’s public usage limit of 2500 unique searches per day.176 Thus, even with eight computers available for use at the School of Population and Public Health, given the number of postal code to hospital combinations, this amounted to 555 days of non-stop work (109,887 postal codes x 101 hospital locations / 2500 hits per day/8 computers). To manage this computational workload without compromising the representativeness of data, I considered the option of altering the precision of the postal code latitude and longitude coordinates with two geographers after coming across the suggestion in related literature.177–179 At the six decimal level of precision at my disposal, distance calculations were typically within two meters. By reducing the degree of precision for the coordinates, which in itself is an imprecise point estimate, the number of unique coordinate combinations was made more manageable. I examined the possibility of reducing all coordinates to two decimal points of precision. This way, the overall results of the study would not be significantly altered but the number of distance calculations needed to populate the matrix would be drastically fewer. In doing so, the unique postal code coordinates were reduced to 3,521. This did not mean that the number of codes included in the study was reduced. Rather, PCs whose centroids overlapped were represented by the same set of coordinates. Overlap of coordinates occurs predominately in high-density urban settings where, for example, individual apartment complexes are given a unique postal code owing to resident density. Since street blocks can support multiple complexes, the opportunity for overlap between centroids when precision is broadened is quite high and explains the significant drop in unique coordinates.178,180 Conversely, in rural communities where individual postal codes span much larger geographic areas, the opportunity for there to be overlap between postal code centroids is minimal. At the two digit precision level, Google Maps would introduce approximately one kilometer or roughly four minutes of drive time error177 which was deemed to be within a 38  reasonable estimate of true travel to care times. However, even at 3521 coordinate sets, it would take eighteen days (3521 x 101 / 2500 / 8) to populate the matrix using the Google Maps approach. Lastly, Google Maps calculates travel burden based on the weather and traffic conditions at the moment of calculation. This means that depending on the time of day and year the user runs the software, the burden in travel time can be dramatically different from another combination of date and time for the same route. Conversely, the network analysis method of ArcGIS assumes that every travel instance experiences the same conditions (as time, date, traffic patterns and weather are held constant) unless otherwise specified by the user.14 While at face value Google Maps’ approach may seem like an improvement, it reduces the internal comparability of travel burden because the estimates would be generated over eighteen days (as mentioned above) and be contingent on the weather and other extenuating details across the province during the data generation period (January 2015). Thus, ArcGIS Network Analyst was used instead of Google Maps Transport Time Estimate despite Google Maps’ improved precision in road time calculations.79 Reasons for doing so included: speed, internal validity, and non-sensitivity to time and date of calculation. Within ArcGIS’ network analysis suite, I used the Origin-Destination (OD) cost matrix operation to generate travel burden calculations for the complete travel matrix (instead of the reduced centroid solution explored when assessing Google Maps’ feasibility) because of the operation’s ability to incorporate road speed limits and other geographically pertinent information provided in road network files uploaded as part of the CanMap Suite and for its ability to generate many-to-many calculations (i.e. a single postal code to each hospital distance was calculated and repeated for every postal code). To better simulate emergency service driving conditions, one-way and turn restrictions were ignored and the quickest road travel time was prioritized. Once each batch of OD calculations successfully ran, the datasets were exported to comma-separated files and merged in SAS before Euclidean arclengths were calculated using the geodist function. I worked with a single personal computer to run the OD cost matrix in the Network Analyst tools package of ArcGIS 10.2.iii To mitigate calculation storage issues within ArcGIS, I ran batches                                                  iii For anyone wishing to replicate this work, it is important to know that within ArcGIS there are storage issues for large batch projects and no error comment explains when you don’t successfully produce a file. I used a Lenovo 39  of 10,000 postal codes to 101 hospital distances at a time. In total, it took 42 hours to populate the entire postal-code to hospital matrix. I also created a second matrix of hospital-to-hospital calculations to capture travel time incurred by patients transferred between facilities during the study. This second matrix of 100 hospitals x 100 hospitals took two minutes to complete using OD. By using the final matrix of six-digit de-identified postal code to hospital distance and travel time calculations and hospital to hospital travel time/distance calculations, very fine-grained measures per patient were created, and then converted as described above to protect patient confidentiality. 4.2.2 Linking the travel matrix with patient records The completed patient cohort was linked with the distance matrices before exclusion criteria were applied. Observations that were hospital transfers (as noted in the DAD) had the burden between the origin and destination hospitals appended through a linkage between the patient dataset and the hospital-to-hospital travel matrix, while patients arriving to their first point of care (initial hospital) had their travel burden appended based on a match between the patient’s 10-digit dummy code (represented without revealing the residential postal code) and the hospital within the postal code-to-hospital travel matrix. However, when the dataset was linked with the initial travel matrix, only 319,057 (77.5%) patient records were successfully matched with travel burden estimates. This left 92,629 (22.5%) records unmatched – far higher than would be expected based on missing postal codes within administrative data alone. 4.2.3 Troubleshooting While only 60,930 (14.8%) of all patient records were for rural residents as defined using Canada Post’s second digit (=0) in the three-digit forward sortation area (FSA) identifier, 45,573 (49.2%) of the patient records without distance matches were rural. In the original distance files generated using CanMap records there were 330 postal codes that had the second digit of the FSA set as “0”. Of the records that were left unmatched, using the FSA value explicitly provided, there                                                  ThinkPad equipped with Intel® Core™i5-3427U CPU at 1.80 GHz and 4.00GB of RAM (3.70GB operational) with success, after dividing the data into batches as described. 40  were 360 unique 6-digit rural postal codes that were included in patient records for which no match existed. Additionally, there were 3,598 non-rural PCs that were not included in the travel matrix. The CanMap file contains 185,000 BC current and retired postal codes while the Postal Code Conversion File offered by Canada Post only holds 115,000. I selected CanMap’s file to tabulate the postal code to hospital distance calculations with the assumption that it was the more complete record for BC resident postal codes and owing to its more recent road files – necessary for the time estimations using GIS. Based on the description provided by DMTI, the completeness of the file was not tested against other records. It should be noted, that to my knowledge no gold standard for a complete dataset of BC postal codes has been confirmed across studies. However, given the significant number of patients left unmatched, and the potential bias in data completion for rural residents, I supplemented the distance file calculations with postal codes present in the PCCF+ file that were absent from CanMap’s records. The completed distance/time to hospital calculations rose to 11,098,587 when the PCCF+ was used to supplement postal code to hospital calculations (15,488 additional unique postal codes). When the updated file was added into the secure research environment, the number of unmatched patient records went from 92,629 to 27,325 (-70.5% change). Thus, of the 411,686 patient records, 384,515 (93.4%) were matched with distance calculations (see Figure 4-1). The discrepancies in the completeness of the CanMap and PCCF+ files are likely due to how the postal codes are routinely collected. With respect to the PCCF+ file, which generally has fewer records in total, GIS topology rules are applied to the geocodes that are available. Where multiple geographic polygons overlap, particularly in high-density settings, stacked geocodes are not all successfully mapped. Given that the majority of the population of BC is urban, or peri-urban, this helps explain why overall there are fewer postal codes in the PCCF+ file. Conversely, the DMTI inherited Canada Post’s records of current and retired postal codes in 2005, and has since been updating the list every quarter. As it relies on a different mapping strategy to calculate relevant population centroids, it does not have the same stacking problem as the PCCF+. However, due to the reduced cross-sectional frequency with which it screens for new postal codes compared with the frequency employed by the PCCF+ (every quarter vs. every month respectively), there is the opportunity for some postal codes to never be added given how infrequently they are used. This 41  may explain why CanMap’s records are more complete for urban and peri-urban residents, but needed to be supplemented by PCCF+ records to close the gap for missing rural PCs. 4.2.4 Emergency health services Ambulance data Aside from tabulated travel burden estimates, I also had access to true time to care measures for patients transported by paramedics between April 1 2009 and March 31 2013. Matching ambulance data to patient records allowed me to conduct sensitivity analysis where model outcomes comparing travel estimates vs. true time were compared. However, since data was only available for patients who used paramedic services, these sub-analyses (described at length in Chapter 8) must exclude patients who self-transport. Despite all conditions studied being time sensitive and paramedic assistance strongly recommended, many patients did not arrive at the hospital via EMS. To be clear, although the Ambulance dataset included chute, dispatch, and en route times, I restricted my analysis to transport time as this was the direct comparator to the estimates developed. 4.3 Results After completing the second round of postal code matching, 27,534 of the 411,686 (6.7%) patient records were left unmatched (see Figure 4-1). Of these, 26,460 had a postal code that was unfamiliar (in other words could not be matched to my matrix) with 6,615 of those (25.0%) identified as rural or remote, and 19,845 (75.0%) as urban or peri-urban. As mentioned, geography here was based on the second digit of each patient’s FSA field available within the DAD. Rural/remote entries were characterized with a 0, urban and peri-urban by a digit from 1 to 9 and not-defined were entries with a missing value in the FSA field. At the same time, I matched 18,104 AMI, 16,150 stroke, and 47,039 trauma patients whose events occurred between fiscal years 2009-2013 with true transport time measures using BCEHS’s Ambulance data.  42   Figure 4-1. Matching process and results  Table 4-1 and Table 4-2 show the results of the travel matrix exercise by median and interquartile range of number of minutes to care. These measures were used to better represent the data in light of patients with some extreme distance tabulations skewing the results when reporting means and ranges. As can be seen, time to closest facility and time to first facility admitted vary significantly by Health Authority (HA) and sex, and there is a slight upward trend in travel burden to closest facility over the study period which can be expected given that facilities closed over time. According to the tables, residents in Vancouver Coastal have the smallest and most uniform travel burden across the three conditions. Conversely, residents in Northern travel the longest and have the widest interquartile range, indicating larger variance in time traveled by residents. However, while Northern residents actively travel the longest (see Table 4-2), residents in Interior Patient records411,686Patient records matched with CanMap319,057 (77.5%)Patient records unmatched 92,629 (22.5%)Patient records matched with PCCF+65,095 (70.3%)Patient records left unmatched27,534 (29.7%)Patient record missing postal code1,074 (3.9%)Patient record with unfamiliar postal code26,460 (96.1%)Rural or remote postal code6,615 (25.0%)Urban or peri-urban postal code19,845 (75.0%)43  have to travel the longest to access their closest facility (see Table 4-1). Across all conditions there is a difference in travel burden to closest compared with the facility first admitted to, suggesting some degree of hospital selection behavior (i.e. paramedics or patients traveling longer to access care for reasons not captured in the data) and/or the assumption that events occur at home is misaligned with reality for a proportion of the population. On a related note, women consistently travel less time to care across all three conditions. This discrepancy may also be a result of differences in location of the medical emergency and/or hospital selection behavior.   Condition Health Authority AMI Median (IQR), mins Stroke Median (IQR), mins Trauma Median (IQR), mins Fraser 18.9 (11.1 – 28.4) 18.0 (10.4 – 27.0) 18.4 (10.9 – 27.5) Interior 23.1 (7.8 – 63.5) 16.1 (7.2 – 60.5) 23.0 (7.6 – 63.7) Northern 14.0 (6.6 – 40.3) 12.0 (5.7 – 38.7) 12.5 (5.7 – 38.7) Vancouver Island 23.6 (14.0 – 46.2) 22.4 (13.2 – 39.2) 22.4 (13.4 – 39.7) Vancouver Coastal 14.9 (7.9 – 22.5) 14.3 (7.8 – 21.3) 13.7 (7.2 – 21.0) Sex    Male 19.6 (10.1 – 35.4) 17.8 (9.4 – 31.7) 18.3 (9.5 – 33.9) Female 17.4 (9.0 – 31.4) 16.7 (8.7 – 28.4) 17.2 (8.9 – 30.8) Year of event    1999-2000 18.0 (8.9 – 30.4) 16.9 (8.6 – 28.6) 17.2 (8.4 – 31.0) 2000-2001 17.5 (8.5 – 31.3) 16.1 (8.0 – 28.1) 17.1 (8.3 – 31.7) 2001-2002 18.0 (9.2 – 31.8) 16.4 (8.4 – 29.8) 16.9 (8.1 – 31.1) 2002-2003 18.4 (9.7 – 32.7) 16.7 (8.7 – 29.0) 17.4 (8.7 – 31.9) 2003-2004 17.6 (9.0 – 32.2) 16.9 (8.4 – 30.0) 17.5 (8.8 – 32.3) 2004-2005 18.8 (9.5 – 34.8) 16.8 (8.8 – 29.0) 17.7 (9.0 – 32.5) 2005-2006 18.3 (9.6 – 33.6) 17.1 (9.0 – 30.1) 17.3 (9.0 – 31.8) 2006-2007 18.4 (9.3 – 34.4) 17.4 (9.2 – 29.2) 17.8 (9.2 – 32.6) 2007-2008 18.3 (10.1 – 35.4) 17.5 (9.3 – 30.1) 17.9 (9.4 – 32.1) 2008-2009 19.6 (10.2 – 35.4) 17.2 (9.1 – 30.5) 18.0 (9.4 – 32.5) 2009-2010 19.4 (10.1 – 35.7) 17.4 (9.2 – 30.4) 18.4 (9.7 – 32.9) 2010-2011 19.8 (10.3 – 35.6) 18.1 (9.8 – 31.2) 18.6 (9.9 – 33.0) 2011-2012 19.2 (9.9 – 35.6) 18.5 (9.9 – 32.8) 18.3 (9.8 – 33.3) 2012-2013 19.9 (10.6 – 35.5) 18.0 (10.0 – 30.9) 18.4 (9.8 – 32.8) Table 4-1. Median time to closest facility in minutes by health authority, sex, and year of event, per condition IQR=Interquartile range; mins=minutes   44   Condition Health Authority AMI Median (IQR), mins Stroke Median (IQR), mins Trauma Median (IQR), mins Fraser 42.0 (18.8 – 105.2) 29.2 (14.8 – 66.4) 42.0 (18.8 – 105.2) Interior 88.4 (18.3 – 376.9) 42.4 (11.2 – 212.8) 88.4 (18.4 – 377.2) Northern 255.4 (12.8 – 1428.5) 34.0 (8.8 – 643.2) 255.2 (12.8 – 1428.4) Vancouver Island 45.1 (21.8 – 219.7) 39.6 (20.0 – 169.2) 45.2 (22.0 – 219.6) Vancouver Coastal 25.6 (14.0 – 58.4) 23.6 (13.6 – 38.4) 25.6 (14.0 – 58.4) Sex    Male 50.2 (19.6 – 239.4) 31.6 (14.8 – 98.8) 50.4 (19.6 – 239.2) Female 34.1 (15.2 – 127.2) 28.8 (13.2 – 38.4) 34.0 (15.2 – 127.2) Year of event    1999-2000 43.0 (16.9 – 228.0) 31.2 (14.0 – 129.6) 42.8 (16.8 – 228.0) 2000-2001 44.0 (17.0 – 220.1) 31.2 (13.6 – 113.2) 44.0 (17.2 – 220.0) 2001-2002 45.8 (17.0 – 208.2) 31.2 (13.2 – 105.2) 41.2 (17.2 – 208.4) 2002-2003 45.7 (17.6 – 220.5) 30.8 (13.6 – 95.2) 46.0 (17.6 – 220.5) 2003-2004 44.4 (18.1 – 197.0) 32.4 (13.6 – 105.2) 45.7 (18.0 – 197.2) 2004-2005 42.7 (17.8 – 198.0) 32.8 (14.0 – 101.2) 44.4 (18.0 -194.0) 2005-2006 40.0 (17.4 – 203.2) 32.4 (14.8 – 94.0) 42.8 (17.6 – 203.2) 2006-2007 41.8 (17.6 – 181.7) 30.8 (14.4 – 87.2) 40.0 (17.6 – 203.2) 2007-2008 40.3 (17.8 – 176.2) 29.6 (14.0 – 82.0) 41.6 (17.6 – 181.6) 2008-2009 40.2 (17.2 – 194.8) 29.2 (14.0 – 82.4) 41.6 (17.6 – 200.8) 2009-2010 49.2 (19.2 – 192.9) 28.8 (14.0 – 83.2) 40.4 (17.6 – 176.4) 2010-2011 38.1 (17.0 – 159.5) 29.2 (14.0 – 78.0) 49.2 (19.2 – 192.8) 2011-2012 29.1 (14.4 – 81.9) 25.6 (13.2 – 53.2) 29.2 (14.4 – 82.0) 2012-2013 29.7 (14.0 – 94.3) 31.2 (14.8 – 86.0) 38.0 (16.8 – 159.6) Table 4-2. Median time to first facility of admission in minutes by health authority, sex and year, per condition IQR=Interquartile range; mins=minutes  Although time to closest facility increased marginally and consistently over the years (see Table 4-1), time to the first facility visited varied dramatically between years and health authorities, and dropped dramatically in the last two years (2011-2012, 2012-2013) of the study period (see Table 4-2). The general variation may be due to hospital selection behavior and the precipitous drop may be a consequence of the restructuring and absorption of paramedic services by health authorities that began in 2010-11.181 As part of their strategy to improve access to care, and better triage patients, the frequency of patients stabilized locally before transfer for definitive care was prioritized. Thus, beginning one year after the restructuring exercise, patients that appeared stable or less urgent received care at a closer facility before being transferred.182 45  For all three conditions, although time to care estimates extended beyond 240 mins (four hours), according to ambulance data provided by BCEHS, no patient travelled longer than four hours to access care. In fact, the vast majority of patients travelled within one hour to receive care (99.93% for AMI, 99.67% for trauma, and 99.86% for stroke). The same cannot be said for patients that did not utilize emergency health services to access care. Table 4-3 shows the extent of linear correlation between estimated and true time to care measures for all three conditions. As can be seen, while there is some linear association, it is weak – likely owing to the underlying assumptions required to create the time to care estimate. This finding along with patient demographics related to ambulance use are further explained in Chapter 8. Condition Number of patients Time to care est., mean (SD) True time, mean (SD) Pearson correlation coefficient, r P-value AMI 18,104 29.09 (41.60) 10.23 (16.20) 0.23272 <0.0001 Stroke 16,150 19.44 (34.57) 8.15 (11.29) 0.20418 <0.0001 Trauma 47,039 32.23 (47.68) 10.03 (12.37) 0.24448 <0.0001 Table 4-3. Correlation between time to care estimate and true time Est=estimated; SD=standard deviation 4.4 Summary Access to care is a key explanatory variable for analysis of the redistribution of hospital services. In this thesis, “access” is measured as the travel distance and time between residential postal code centroid and hospital coordinates. Although as the crow flies (aerial) distances have been used in the past as a proxy for access, this dissertation prioritizes drive time to care (in minutes) calculated using ArcGIS. However, both aerial and land distances were also tabulated for sensitivity analyses. Of the 411,686 medical emergencies included in this study, 384,152 (93.3%) were successfully matched with time to care estimates. Time to closest facility varied extensively by health authority with more urbanized populations travelling less time and experiencing less variation in travel burden than patients residing in more sparsely populated health authorities. Additionally, time to closest facility increased over the years, as was expected given facilities downgraded or closed during the study period. The estimated time to the facility first admitted for care was consistently longer than the time to the closest facility based on travel matrix calculations and the assumption that all events occurred close to home. This may be explained by preferential selection of hospitals other than the 46  closest, or by a mismatch between the true location of the incident and that assumed (at home). However, in the last two years of the study period (fiscal years 2011 and 2012), patients traveled less time to first point of admission. Changes in paramedic triaging activity may explain this variation, although it is not studied at length here. Additionally, comparisons between estimated and true time to care show that estimated time to care consistently overestimated the travel burden by a significant degree, such that there was limited correlation between true and estimated travel times. This finding suggests that patients did not travel for as long as suggested using the travel matrix method likely in part because during medical emergencies drivers exceed speed limits – not easily replicated in origin-distance tabulations, and that the assumption that all patients travelled from home is likely false. However, there is no reason to believe this overestimation systematically biases any particular group of patients. To assuage these concerns, models using both estimated and true travel burden will be compared when testing the third hypothesis: Access to care has no measureable effect on patient outcomes after controlling for quality and appropriateness of acute care services, and compensating mechanisms such as inter-hospital transfers and telehealth services. 47  Chapter 5: Appropriateness of care Appropriateness of care is the comparison between the treatment provided and that recommended given the event’s circumstances (e.g. condition severity; patient comorbidities and age; facility equipment; and staff procedural experience). The expectation of any treatment is that the benefits outweigh the risks, and as such standardized protocols have evolved to reflect clinical findings of effective care. Assessing the appropriateness of care using administrative data is challenging for time sensitive medical emergencies because documentation on best practices strongly emphasizes the timeliness of care. Iezzoni’s 1997 paper “Assessing quality using administrative data”, remarks on the limitations of using administrative data, in particular with regards to assessing appropriateness in reference to Donabedian’s quality of care framework. She posits that appropriateness of care can be described in two ways: “errors of omission (failing to do necessary things) and errors of commission (doing unnecessary things). Both errors can be related to another important dimension of quality, access to health care. In errors of omission, access may be impeded [while in] errors of commission, access may be too easy or inducements to perform procedures too great.”183 Although the lack of timestamps in hospitalization data means it is not possible to assess the timeliness of procedures received once admitted, administrative data can still be used to evaluate both errors of omission (incomplete care) and commission (not recommended or contraindicated). This is accomplished by comparing the treatment received with that recommended in national guidelines describing the equipment, staff and procedures necessary for complete, appropriate care. 5.1 Methods Using national and provincial guidelines on care, I compared the treatment received with that recommended. Since protocols evolved during my study period, I created time specific thresholds that reflected the period for which they were relevant, described in detail below. I then evaluated the appropriateness of care each patient received by the year of event occurrence. I considered a one-year lag from when the protocol was published to when it would be fully implemented in the 48  field. Thus, patients whose event occurred the same year as the protocol change had their appropriateness of care evaluated against the previous protocols which were more relaxed. In this way, I avoided penalizing facilities transitioning to new guidelines. For patients who were transferred between facilities to receive definitive care, I evaluated the care received during the entire episode of hospitalization. These standards and the thresholds developed are summarized below. The relevant procedure codes for each condition’s treatment were selected using multiple documents and complied with the internal classification of clinical modification rules and the accompanying Canadian specific coding documents (i.e. Canadian Coding of Health Interventions, and Canadian Classification of Diagnostic, Therapeutic, and Surgical Procedures). I used independent sample t-tests to compare continuous variables between appropriateness of care groups where distribution was normal, Wilcoxon-Mann-Whitney tests when distribution was not normal, and chi-square tests for categorical variables. 5.1.1 AMI standards of care and protocols Of the three medical events of interest, treatment for AMIs (specifically STEMIs which are the focus of this dissertation) underwent the most dramatic and consistent evolution during the study period owing to ongoing technological advancements and regular changes to evidence based practices. During the study period, pharmaceutical thrombolysis became the preferred initial treatment for all AMI events – offering immediate, non-invasive treatment of peripheral arterial or venous occlusions.188 While tissue plasminogen activator (tPA) is the most common thrombolytic agent, it is not the only form of pharmaceutical thrombolysis. Aside from tPA, other similarly performing drugs include lanoteplase, reteplase, staphylokinase, streptokinase (the first thrombolytic ever produced),188 tenecteplase, and urikinase.189 Owing to tPA’s ubiquitous use in treating blood occlusions, the terms thrombolysis and tPA are used interchangeably in American and Canadian standards of care for STEMI patients.190–195 Thus, although contraindications for use of tPA are normally cited, they are considered contraindications for the use of all thrombolytic agents due to the risk of internal bleeding.188–195 Successful reperfusion (initial restoration of blood flow) can be achieved in 50-60% of STEMIs using pharmacological thrombolytics.188,196 Although thrombolytics do not eliminate the risk of 49  re-occlusion, they dissolve clots causing immediate threat and extend the window of time to definitive care. Definitive care is performed either by widening the narrowed arterial or venous passages supporting heart tissue or redirecting blood flow where needed. The techniques to do so include the minimally invasive use of a balloon catheterization to compress the blockage against the arterial wall (percutaneous transluminal coronary angioplasty, PTCA)197, a similar procedure where once the balloon has been inflated a stent is put in place to keep the blockage held back (percutaneous coronary intervention (PCI) or angioplasty with stent)198 or redirection of blood by bypassing the blocked portion of artery188 through the use of alternate blood vessels or grafts (coronary artery bypass graft surgery, CABG)199. Although PCI evolved from PTCA, for ease of subsequent discussion PCI is used as the generic term for all balloon catheterization procedures (percutaneous coronary intervention, percutaneous transluminal coronary angioplasty, and angioplasty with stent). All three methods are known as surgical or mechanical revascularization. Regardless of the intervention used, approximately 60-80% of all events require catheterization (for diagnosis e.g. angiograms and/or treatment e.g. PCI) during the episode of hospitalization for the ischemia.98,200 As technologies to treat AMIs evolved, procedures fell out of fashion and acceptable timelines for treatment changed to reflect both the improvements in care and the evidence supporting said procedures. Figure 5-1 is a streamlined version of the minimum benchmarks for appropriate care for STEMI patients in British Columbia. Notice that for each circumstance (e.g. paramedics are the first point of medical contact), there are changes to the care trajectory as best practices progressed but definitive care is characterized by some form of mechanical revascularization. While a patient with a STEMI in 2000 who was given thrombolytics within 30 minutes and arrived at a PCI-enabled facility just after 1.5 hours of first contact with paramedics would be described as having received appropriate care, that same trajectory of care was considered inappropriate in 2006.190–195 Additionally, whether appropriate care was provided is also contingent on patient and facility level circumstances. 50   Figure 5-1. Appropriate care for patient presenting with ST segment elevated myocardial infarction 51  Table 5-1 summarizes when thrombolytics, PCI and CABG are contraindicated given the patient’s condition and when the contraindication was first recorded in treatment guidelines.  Contraindications (errors by commission) Year Treatment Patient characteristics Facility/staff characteristics 1996      Thrombolysis >74 years Shock Pulmonary congestion Tachycardia (HR>100bpm) Hypotension (SBP ≤ 100 mmHg) Intercranial hemorrhage       1999 Congestive heart failure  2004 Pregnancy  1996  Primary PCI   Physician performs < 75 PCIs/year Hospital oversees < 200 PCIs/year 1999    Septal defect Angioplasty fail (PTCA) Mitral valve insufficiency Persistent AMI     2004  Ischemic stroke 3hrs ≤ time < 3 mths  Hospital does not offer cardiac surgery Hospital oversees < 36 pPCIs/year 2004 CABG Stable angina  Table 5-1. Contraindications for patients with ST segment elevated myocardial infarction by patient and facility/staff level characteristics and year of event HR=heart rate; SBP=systolic blood pressure; mmHg=millimeters of mercury; PTCA=percutaneous transluminal coronary angioplasty; AMI=acute myocardial infarction; PCI=percutaneous coronary intervention; pPCI=primary percutaneous coronary intervention (no thrombolysis before procedure); hrs=hours; mths=months  Using the information summarized in Figure 5-1 and Table 5-1, I created an appropriateness of care variable to assess the treatment received against that recommended for AMI events. Some details could not be captured owing to no availability of Emergency Department (ED) data.  For instance, from the data available in my datasets, it is not obvious if patients who arrived at facilities unable to provide PCI services were given thrombolytics before transfer. What is clear is when patients are transferred from a non-PCI centre to a PCI-enabled centre. Similarly, without access to ED data I cannot comment on whether or not patients received thrombolytics within the window of time recommended. Nevertheless, I was able to determine if patients were routed to the correct facilities given their health status, the distance from said facilities (under the assumption that the 52  event occurred at home), and the facility’s capacity to perform the procedure (given physician and support staff’s experience with the treatment method as evaluated annually). The appropriateness of care was assessed per episode of hospitalization based on recorded procedures. For AMI events, appropriateness of care was divided into three categories: patients who were mechanically revascularized without contraindications, appropriateness=1 (received appropriate care); patients who were revascularized against medical guidelines, appropriateness=0 (care contraindicated, treatment by commission); and patients who did not receive mechanical revascularization during their episode of hospitalization were defined as appropriateness=99 (appropriateness of care unknown, plausible treatment by omission). Care was flagged as fitting one of the above three categories based on the patient’s co-existing medical conditions, the procedures performed, and the expertise of the primary physician and the experience of the hospital. Co-existing medical conditions were identified through the diagnostic codes present in the hospitalization data for that event. Both physician experience and hospital experience attempt to describe the experience of the physician and staff treating the patient. For centres equipped with the technology to perform a procedure, the treatment may still be considered contraindicated according to standards if overall annual volumes are low. For simplicity, a primary physician was determined to be experienced if he or she performed the minimum number of procedures stated in the standards in the preceding 365 days (e.g. >75 PCIs). Similarly, a hospital’s staff were considered experienced if that hospital had overseen a minimum number of procedures (e.g. >200 PCIs in that year) as described in Table 5-1. To interpret Table 5-1, each year adds additional contraindications to the preceding list; there were no factors identified as contraindications for a treatment in one year that were later removed. 53  5.1.2 Stroke standards of care and protocols Unlike AMI-STEMI care, stroke standards of care and respective protocols did not change considerably between 1999 and 2013.190,201–209 In fact, despite new releases of the standards of care produced in 2003, 2005, 2007 and 2013; there were no changes to patient care until 2013, aside from the expanding list of relevant contraindications.204,208 Like AMIs, thrombolytic medication is the primary treatment for blocked ischemic stroke, and similarly tPA is the most common thrombolytic drug in use. Thus between 1999 and 2013, appropriate ischemic stroke care was defined as receiving tPA within three hours of initial symptom onset, and importantly within one hour of arrival at hospital. However, before tPA was administered, patients must have met with their stroke team, and received the results of a diagnostic computer tomography scan. Finally, within three hours of arrival at hospital, patients with an ischemic stroke should have been admitted, where possible, to the stroke unit for further care and analysis. In 2013, the primary window of interest, namely time from symptom onset to thrombolysis was extended from three hours to 4.5 hours. However, patients were still expected to be diagnosed and treated with thrombolytics within one hour of arrival at hospital and admitted to the stroke unit within three hours. As with AMI, Table 5-2 outlines the contraindications for thrombolysis treatment, and when they were added to the list (first noted in guidelines). Since, as mentioned, I could not determine if some services (namely those within the ED) were received within the one-hour window upon arrival at hospital, I used this table to flag patients who received inappropriate care (treatment by commission).   54   Contraindications Year Treatment Patient characteristics 1996          Thrombolysis (tPA) < 3 hrs from symptom onset Heparin administered within previous 48 hrs Thrombocytopenia (platelet count<100000/mm3) Hypoglycemia (blood glucose concentration < 50mg/dl) Hypertension (SBP>185 or DBP>110mmHg) History of intracranial hemorrhage Stroke <3months ago Subarachnoid hemorrhage Active internal bleeding (internal hemorrhage unspecified) Head trauma <3 months Gastrointestinal or urinary tract hemorrhage <21 days 2003    Acute bleeding diathesis Major surgery or trauma <14 days AMI <3 months Arterial puncture at non compressible site <7 days 2004   Pregnancy Intracranial or intraspinal surgery Intracranial neoplasm, arteriovenous malformation, aneurysm Table 5-2. Contraindications (errors of commission) for patients with ischemic stroke by year mm3=cubic millimeters; mg/dl=milligrams per deciliter; mmHg=millimeters of mercury  5.1.2.1 Telestroke services Beginning in 2006, the Heart and Stroke Foundation with the Ministry of Health began piloting telestroke services in British Columbia. Vancouver Island first offered the services in July 2009 while the Lower Mainland began in February 2010. By 2011, the provincial government had invested five million dollars in the BC Stroke Strategy to develop telestroke initiatives (pilots and subsequent expansion).210 In simplest terms, telestroke services are the live interactions between a referring site (hospital providing care) with a patient, and an approved consulting site (location of the neurologist) for the assessment and management of stroke patients using a variety of telecommunications technology.210,211 This hub and spoke model allows neurologist services to extend well beyond their physical location. Thus, the ideal sites to participate as spokes are currently restricted to primary stroke centres – facilities that can provide laboratory, diagnostic imaging in the form of CT scans, ultrasounds and CT angiography services but who do not have a resident neurologist available 24/7 to confirm diagnosis. Comprehensive stroke centres which offer these services plus 55  ensure the availability of said neurologist are ideal for hubs. By using real-time telecommunications, primary stroke centres can link remotely to a neurologist at a comprehensive centre who can accurately diagnose and recommend next steps for appropriate stroke care.212  This appropriate care is typically in the form of tPA thrombolysis within 4.5 hours of initial symptom onset given the absence of contraindications. Telestroke is particularly relevant for this dissertation because it is specifically developed to offset some issues related to the concentration of resources owing to the centralization of services. By creating secure networks of telecommunication, telehealth increases the capacity of facilities that would otherwise not be able to provide appropriate care to extend their services to include the diagnosis and primary treatment of ischemic strokes. Thus, patients who received telestroke services need to be identified in the data and this compensating mechanism flagged. Relying on the fee schedule published by the Medical Services Commission, patients who received telestroke services were any patients whose MSP fee items included the codes “40441” – “40444” (see Table 5-3).212 Fee item Fee name Fee description 40441 Telestroke Consultation Video conference examination and review of diagnostics 40442 Follow-up Telestroke neurological clinical monitoring and treatment of symptoms without administration of tPA used to monitor patient every ½ hour 40443 Follow-up Telestroke neurological clinical monitoring and treatment of symptoms with administration of tPA used to monitor patient every ½ hour 40444 Follow-up Telestroke relapse intervention used for monitoring treated patient over 72 hours Table 5-3. Telestroke fee items, MSP codes, and descriptions As observations in the dataset are organized in long form, telestroke services were linked to the first observation for an episode of hospitalization where the servdate and addate were at most one day apart (to account for overnight and delayed hospital admissions), and only where the cause of hospitalization was stroke. Alternatively, in the case of a mortality where a patient was never admitted to hospital with stroke as a cause of death, the same rule applied. Since this analysis is only interested in telestroke as a compensating mechanism for acute care services, billings for 40441-diagnostic consultations that did not lead to a stroke diagnosis were excluded from subsequent analysis. 56  5.1.3 Trauma standards of care and protocols The Canadian Emergency Department Triage & Acuity Scale lists patient severity on a five-level scale corresponding with the relevant minimum level of treatment necessary as described by the Trauma Association of Canada. The CTAS level of severity mirrors CAEP’s hospital designations and provides a guide to time to care from entry into hospital emergency departments as fractile response rates.213 Ninety-eight percent of Level I patients should be seen by a physician at a Level I hospital immediately. Similarly, 95% of patients with a trauma severity of Level II should be seen within 15 minutes of arrival at the emergency department of a Level II trauma centre or higher. Patients with CTAS scores III-IV can be seen anywhere between 30 minutes and two hours, reflecting the reduced urgency of their needs.213 The cohort I captured using CIHI restrictions described in Chapter 3 only includes patients experiencing a Level I or II trauma event. Here, patients who received appropriate care were first admitted to a Level II trauma centre or higher, while those flagged as receiving inappropriate care were first admitted to a lower grade facility (treatment by omission). 5.1.4 Emergency services offered outside a hospital For both AMI and trauma, an MSP billing code exists for emergency acute care services provided by a physician outside of a hospital. Fee item “00081” is appropriate for the evaluation, diagnosis and treatment of patients with life-threatening injuries or illnesses including cardiac arrest, multiple trauma, acute respiratory failure, coma, shock, cardiac arrhythmia with haemodynamic compromise, and hypothermia.214 As part of the treatment offered, fee item “00081” can be billed for the use and monitoring of pharmacologic agents such as thrombolytic drugs. In the data, this fee item was used only 18 times, and thus despite the potentially promising addition to the analysis, this was not explored further in subsequent analyses. 5.2 Results 5.2.1 AMI Table 5-4 shows the demographic differences between those who received appropriate AMI care and those who did not. Approximately 28% (n=27,473) of the 96,672 patients who underwent STEMI treatment between 1999 and 2013 received some form of mechanical revascularization (PCI/PTCA or CABG) during the hospitalization following thrombolytic treatment. Of these 57  27,473 patients, 12,653 (46%) were re-vascularized but did not meet treatment guidelines (i.e. error by commission because treated at low-volume facility, treated by inexperienced lead physician, or treated despite contraindicated co-morbidities). In examining Table 5-4, we see that women were less likely to be revascularized (error by omission) for an AMI event than men, as were frailer and older patients. Patients who were revascularized in accordance with the most current guidelines at the time were more likely to be transferred at some point (proportion transferred=55.72% vs. 32.27%, p-value<0.0001) and traveled further for their first admission (median=29.23 mins vs. 13.60 mins). Appropriate revascularization was associated with lower readmission and short-term mortality rates. These are strictly descriptive results. They have not been adjusted for potential confounding effects such as age, which has been found to explain sex differences in treatment in the past. Adjusted associations will be explored in Chapter 8. As such, readers should avoid drawing conclusions here, but can begin to consider the potential implications of these findings.   58  Table 5-4. AMI patient demographics by appropriateness of care IQR=interquartile range; mins=minutes; km=kilometers Patient demographics Mechanical revascularization No revascularization Total P-value Appropriate care Inappropriate care (error by commission) Inappropriate care (error by omission) Died without admission N (%) 14820 (15.33) 12653 (13.09) 67214 (62.30) 1985 (1.84) 96672  Sex       Female 3717 (25.08) 3328 (26.30) 24327 (36.19) 664 (33.45) 32036 (29.69) <0.0001 Male 11103 (74.92) 9325 (73.70) 42887 (63.81) 1321 (66.55) 64636 (59.91)  Income quintile       Lowest income quintile 3170 (21.39) 2703 (21.36) 15739 (23.42) 484 (24.38) 22096 (20.48) <0.0001 2nd income quintile 3093 (20.87) 2544 (20.11) 14208 (21.14) 453 (22.82) 20298 (18.81)  3rd income quintile 2920 (19.70) 2454 (19.39) 12837 (19.10) 392 (19.75) 18603 (17.24)  4th income quintile 2786 (18.80) 2424 (19.16) 11751 (17.48) 365 (18.39) 17326 (16.06)  Highest income quintile 2597 (17.52) 2208 (17.45) 10760 (16.01) 271 (13.65) 15836 (14.68)  Missing 214 (1.44) 274 (2.17) 1811 (2.69) 20 (1.01) 2319 (2.15)  Charlson comorbidity index     N=94687  CCI=0 11291 (76.19) 9442 (74.62) 41832 (62.24)  62565 (59.07) <0.0001 CCI=1 2887 (19.48) 1603 (12.67) 9479 (14.10)  13969 (13.19)  CCI=2 446 (3.01) 1064 (8.41) 9654 (14.36)  11164 (10.54)  CCI=3 136 (0.92) 422 (3.34) 4085 (6.08)  4643 (4.38)  CCI=4+ 60 (0.40) 122 (0.96) 2164 (3.22)  2346 (2.22)  Self-transport 2347 (15.84) 1483 (11.72) 16004 (23.81)  19834 (18.73) <0.0001 Transfers 8257 (55.72) 4083 (32.27) 17587 (26.17)  29927 (28.26) <0.0001 Readmissions 550 (3.71) 346 (2.73) 2165 (3.22)  3061 (2.89) <0.0001 30-day mortality 382 (2.58) 823 (6.50) 9503 (14.14)  12693 (11.98) <0.0001 Age (year), Median (IQR) 64 (55 – 74) 64 (55 – 75) 72 (60 – 81) 75 (66 – 83) 70 (59 – 80) <0.0001 Length of stay N=14799 N=12653 N=58619  N=94687  Median days (IQR) 4 (3 – 6) 5 (3 – 7) 6 (4 – 9)  5 (4 – 9) 0.0070 Euclidean distance N=12922 N=10228 N=54994  N=91722  Median distance, km (IQR) 23.01 (7.61 – 95.98) 8.98 (4.32 – 41.57) 5.36 (2.45 – 19.92)  6.82 (2.91 – 25.12)  Closest facility N=14700 N=12485 N=59333 N=1971 N=88489  Median time, mins (IQR) 5.71 (3.18 – 9.42) 5.31 (3.04 – 8.85) 5.26 (2.74 – 9.33) 4.72 (2.62 – 8.39) 5.32 (2.83 – 9.21) <0.0001 Facility visited N=12922 N=10228 N=54997  N=78147  Median time, mins (IQR) 29.23 (11.43 – 134.68) 13.60 (6.43 – 45.03) 8.16 (3.89 – 22.83)  10.05 (4.52 – 30.28) <0.0001 59   Mechanical revascularization No revascularization  Treatment characteristics Appropriate (%) Inappropriate (%) (error by commission) Inappropriate (%) (error by omission) Total (%) Treatment 14820 (15.33) 12653 (13.09) 67214 (69.53) 96672 Died within 30 days 382 (2.58) 823 (6.50) 9503 (14.14) 12693 (13.13) Treatment method     CABG 1031 (6.96) 96 (0.76)  1127 (1.17) Angiogram 13673 (92.26) 11035 (87.21) 11469 (17.06) 36177 (37.42) PCI/PTCA 13816 (93.23) 12636 (99.87)  26452 (27.36) Catheterization 14512 (97.92) 12653 (100.00) 11469 (17.09) 38634 (39.96)      Inappropriate PCI by reason  12636 (99.87)   Under-experienced physician  237 (1.87)   Under-experienced facility  4678 (36.97)   Both  7721 (61.02)   Contraindication (age, comorbidities)  17 (0.13)   Table 5-5. Treatment received by AMI patient admitted to hospital, organized by appropriateness CABG=coronary artery bypass grafting; PCI=percutaneous coronary intervention; PTCA=percutaneous transluminal coronary angioplasty Meanwhile, Table 5-5 shows the unadjusted mortality count and rate for each type of treatment.  The mortality rate for patients who received inappropriate mechanical revascularization (irrespective of reason) is more than double that of patients whose treatment complied with standards of care guidelines. However, receiving any form of mechanical revascularization appears to be associated with lower mortality than receiving just thrombolytic therapy, where over 14% of patients died.  Of the reasons surgical intervention was deemed inappropriate, the experience of both the lead physician and the facility dominated, dispelling concerns that the method of evaluating appropriateness of care penalized new physicians in their first year. Instead, Table 5-6 suggests that facilities are not meeting patient thresholds associated with developing facility expertise. This may be attributed to the general low patient density at any particular facility, or preference to not perform mechanical revascularization. 60  5.2.2 Stroke Table 5-6 shows patient demographics as organized by appropriateness of care received. Of all patients that experienced a stroke and were taken to hospital, very few (6.5%) were treated against standard practice and only 694 were never admitted to hospital. Similar to AMI results, female patients were less likely to receive appropriate care and a larger proportion of patients who received appropriate care were transfers. Appropriate care recipients had fewer readmissions but higher 30-day mortality rates. Patient demographics Appropriate care Inappropriate care (error by commission) Died without admission Total P-value N (%) 88010 (92.78) 6159 (6.49) 694 (0.73) 94863  Sex      Female 42559 (48.36) 4200 (68.19) 301 (43.47) 47060 (49.61) <0.0001 Male 45295 (51.47) 1940 (31.50) 393 (56.63) 47628 (50.21)  Unknown 156 (0.18) 19 (0.31)  175 (0.18)  Income quintile      Lowest income quintile 20648 (23.50) 1592 (25.87) 194 (27.95) 22343 (23.68) <0.0001 2nd income quintile 18329 (20.86) 1266 (20.57) 148 (21.33) 19743 (20.84)  3rd income quintile 16924 (19.26) 1079 (17.53) 151 (21.76) 18154 (19.17)  4th income quintile 14967 (17.03) 948 (15.40) 94 (13.54) 16009 (16.90)  Highest income quintile 14792 (16.83) 1051 (17.08) 101 (14.55) 15944 (16.83)  Missing (141 no data) 2214 (2.52) 218 (3.54) 6 (0.86) 2438 (2.57)  Charlson comorbidity index      CCI=0 57891 (65.78) 4253 (69.05)  62838 (66.24) <0.0001 CCI=1 12467 (14.17) 516 (8.38)  12983 (13.69)  CCI=2 12388 (14.08) 1078 (17.50)  13466 (14.20)  CCI=3 2557 (2.91) 172 (2.79)  2729 (2.88)  CCI=4+ 2707 (3.08) 140 (2.27)  2847 (3.01)  Self-transport 14502 (16.48) 436 (7.08)  14938 (15.75) <0.0001 Transfers 10262 (11.66) 142 (2.31)  10404 (10.97) <0.0001 Readmissions 3032 (3.44) 361 (6.83)  3393 (3.58) <0.0001 30-day mortality 14111 (16.03) 831 (13.59)  14942 (15.75) <0.0001 Age (year), Median (IQR) 76 (65 – 83) 84 (78 – 89) 75 (66 – 82) 76 (66 – 84) <0.0001 Length of stay N=87890 N=6151  N=94041  61  Patient demographics Appropriate care Inappropriate care (error by commission) Died without admission Total P-value Median days (IQR) 7 (3 – 19) 11 (6 – 24)  8 (3 – 19) <0.0001 Euclidean distance N=72819 N=3600  N=76419  Median distance, km (IQR) 5.76 (2.59 – 18.12) 5.66 (2.42 – 21.51)  5.76 (2.59 – 18.36) 0.0150 Closest facility N=87302 N=6108  N=93410  Median time, mins (IQR) 4.90 (2.67 – 8.22) 4.58 (2.34 – 7.68) 4.95 (2.89 – 8.76) 4.88 (2.65 – 8.20) <0.0001 Facility visited N=72822 N=3600  N=76422  Median time, mins (IQR) 8.31 (4.07 – 21.15) 8.33 (3.89 – 25.73)  8.32 (4.05 – 21.36) 0.0700 Table 5-6. Stroke patient demographics by appropriateness of care IQR=interquartile range; mins=minutes; km=kilometers  5.2.3 Trauma Table 5-7 shows the demographic differences between trauma patients who received appropriate care and those who did not. Of all trauma incidents included in the analysis, 34.8% were directly transported to a Level I or II trauma centre (received appropriate care). Patients who received appropriate care were more likely to be male, young, with fewer comorbidities and less likely to have arrived at hospital by self-transport. They also traveled longer to access care, and lived farther away from the nearest facility able to stabilize trauma incidents. Without a severity measure it is unclear if the number of patients included in the analysis is restricted to the most severe trauma in need of urgent care. However, comparing the two groups, patients treated at Level I and II trauma centres had longer stays (14.25 vs. 12.18 days, Table 5-7 only shows median times) and spent more time in the ICU (1.85 vs. 1.14 days). 62  Patient demographics Appropriate care Inappropriate care (error by omission) Died without admission Total P-value N (%) 76632 (34.81) 143519 (65.19) 751 (0.34) 220151  Sex      Female 32927 (43.04) 76609 (53.61) 246 (32.76) 109782 (49.87) <0.0001 Male 43515 (56.88) 66092 (46.25) 505 (67.24) 110112 (50.02)  Unknown 60 (0.08) 197 (0.14)  257 (0.12)  Income quintile      Lowest income quintile 17824 (23.40) 34516 (24.23) 190 (25.30) 52530 (23.95) <0.0001 2nd income quintile 15566 (20.44) 28210 (19.80) 154 (20.51) 43930 (20.03)  3rd income quintile 14004 (18.39) 26195 (18.39) 119 (15.85) 40318 (18.38)  4th income quintile 13366 (17.55) 24606 (17.27) 149 (19.84) 38121 (17.38)  Highest income quintile 12715 (16.69) 23922 (16.79) 127 (16.91) 36764 (16.76)  Missing (793 no data) 2688 (3.53) 4995 (3.50) 12 (1.60) 7682 (3.50)  Charlson comorbidity index      CCI=0 66872 (87.41) 117095 (81.94)  184718 (83.91) <0.0001 CCI=1 4300 (5.62) 9018 (6.31)  13318 (6.05)  CCI=2 3726 (4.87) 12065 (8.44)  15791 (7.17)  CCI=3 820 (1.07) 2359 (1.65)  3179 (1.44)  CCI=4+ 914 (1.19) 2982 (2.08)  3145 (1.43)  Self-transport 9830 (12.83) 20455 (14.25)  30285 (13.76) <0.0001 Transfers 3483 (4.55) 6428 (4.48)  9911 (4.50) 0.4777 Readmissions 1438 (1.88) 2283 (1.59)  3721 (1.69) 0.0190 30-day mortality 3484 (4.55) 5759 (4.03)  9243 (4.20) 0.0450 Age (year), Median (IQR) 52 (31 – 75) 67 (42 – 83) 59 (33 – 77) 61 (38 – 81) <0.0001 Length of stay      Median days (IQR) 7 (4 – 13) 7 (4 – 13)  7 (4 – 13) <0.0001 Euclidean distance      Median distance, km (IQR) 8.84 (4.31 – 45.48) 6.87 (2.61 – 33.22)  8.01 (3.21 – 36.56) <0.0001 Closest facility      Median time, mins (IQR) 5.33 (3.02 – 8.55) 4.87 (2.47 – 8.81) 5.34 (2.89 – 10.28) 5.04 (2.68 – 8.70) <0.0001 Facility visited      Median time, mins (IQR) 12.79 (6.52 – 51.13) 9.71 (4.13 – 35.99)  11.22 (4.92 – 42.64) <0.0001 Table 5-7. Trauma patient demographics by appropriateness of care IQR=interquartile range; mins=minutes; km=kilometers63  5.3 Summary Appropriateness is matching the service delivered to the needs of the patient and the skills and experience of the people and facility providing care. Appropriateness can be defined both by doing the right thing and by avoiding things that are contraindicated.183 Administrative data are not able to address all aspects of appropriateness care. In the context of this thesis, the important missing information has to do with whether service was provided within specified time frames according to treatment guidelines. However, the data can support assessment of patient-level co-morbidities that should influence care choices, and the experience of physicians and facilities as measured by yearly volume of service. For this dissertation, appropriateness identified errors by omission (failing to provide a service) and by commission (providing a service contrary to recommendations owing to physician or facility expertise, or patient comorbidities). Many patients were flagged for receiving inappropriate care according to the most current guidelines at the time of their event. For AMI, 45.9% of patients who received mechanical revascularization were classified as receiving inappropriate care, mainly owing to the lack of personnel experience (99.87%), while only 0.13% did so because of co-morbidities contraindicating treatment (see Table 5-5). Similarly for trauma patients, 65.2% were not initially admitted to a facility capable of providing definitive care for urgent trauma emergencies (Level I or II trauma centre). However, the majority of stroke patients (92.8%) were not flagged for inappropriate care likely due to the simplicity of the way appropriate care can be assessed using administrative data (simply examining whether tPA was provided, and not examining whether this was done within the recommended time frame). Care appropriateness differed by sex, age, and comorbidity for all three conditions. This may indicate treatment preferences favouring certain demographics, which will be examined in multivariate models in subsequent chapters, or selection bias (healthier patients more likely to be given more invasive procedure/treatment).  Lastly, patients who received appropriate care typically traveled farther for their first admission than patients who did not. 64  Chapter 6: Quality of care Quality of care describes the probability of a favourable event outcome given the patient’s characteristics and the facility’s capacity and experience. At its heart, quality of care compares the observed event outcome with the expected following treatment. Aggregated at the facility level and risk adjusted for patient profiles, the difference between expected and observed outcomes can highlight facilities that do not perform as well as other hospitals given the same circumstances. To operationalize quality of care, I used risk-standardized mortality ratios (RSMRs) and funnel plots to compare hospitals’ performance after controlling for variations in patient populations seen across hospitals.150,215 The funnel plots were instrumental in identifying hospitals with potential performance issues and creating a quality variable in the dataset for subsequent analysis purposes. While some researchers and health systems evaluations use process measures to operationalize ideas of quality,154 studies have shown that these techniques can neglect factors such as the level of attention received from nursing staff which may contribute to survival.216 These differences in human factors may compensate for oversights in minor process measures. By using an outcome measure that compares the outcomes observed with those expected, quality in this context captures the patient experience beyond access and appropriateness of care, and circumvents methods that bias against small facilities. 6.1 Methods I used random intercept logistic regression models within the GLIMMIX procedure in SAS217 to adjust for patient age, sex, comorbidity, and condition severity; and hospital condition specific volume, CIHI hospital type designation (i.e. teaching, large to small community, or non-acute centre), and geographic location (rural or not). I also clustered by hospital to correctly adjust standard error calculations. By exponentiating the random intercept output per hospital, I had an RSMR for each facility relative to the average hospital in BC (set as RSMR=1). From there, I calculated the expected number of deaths by dividing the observed number of deaths by the RSMR. I plotted the expected number of deaths and the RSMR on funnel plots. Using the confidence intervals of each hospital, I developed confidence bands to visually identify performance outliers. Lastly, I drew 90% and 95% confidence interval bands on the funnel plots using the random effects’ effect measures and standard errors, and set the y-axis (RSMR) to the logarithmic scale. 65  Any hospital outside the upper confidence interval band (95%) was flagged as underperforming. The data in this study covers three very distinct medical events where quality of care may differ between hospitals. As such and given the 14-year duration of the study period, RSMRs were calculated per condition and over discrete time intervals within the study period. 6.1.1 Data preparation Only patients admitted to hospitals were included in the analyses. The patient level characteristics age, sex, SES, time from care, and comorbidity were modeled as fixed effects. Hospital characteristics were modeled as random effects and included condition specific patient volume and hospital peer group. Covariates for inclusion in the models were selected based on previously identified factors in the literature149,150 and in the case of time to care an attempt to add a new variable that has not, to the best of my knowledge, been used elsewhere for these purposes but which has demonstrated considerable variability across the population (see Chapter 4). Patients who received care at multiple facilities during their episode of hospitalization were assigned to the facility to which they were first admitted. Studies have shown that assigning patients to hospitals in this way increases the number of visits attributed to small facilities while providing no impact on the observed outcomes in larger centres.161,218 Since this dissertation is interested, in part, in understanding how quality is tied to facility size, it is important to incorporate methods that favor small facility inclusion in analysis. The only caveat is that some non-hospital acute care centres may be included in the comparison activity if enough patients are first taken there for care before transferring to a hospital. Several reports limit comparing institutional performance to hospitals by specifying patient volume minimums. Such minimums can range from five patients admitted per year159 to fifty admitted over three years150 or alternatively, to hospitals with at least one expected mortality event218. Here, I restricted my comparisons to facilities that admitted at least five patients in a given year. This threshold was intentionally low enough to include as many small hospitals as possible, without incorporating the effects of non-hospitals in the models. Additionally, given that mortality is a comparatively rare event, this allowed me to include small hospitals that would otherwise have been excluded. 66  6.1.2 Method variations and sensitivity analysis As part of this dissertation, several variations of the risk-adjusted mortality ratio were tested for robustness, and validity. The first such variation was over the duration of time intervals. I tested one-year, two-year and three-year time intervals. Second, I repeated the exercise using 90-day mortality rates as a sensitivity analysis of the plots. Third, I used parsimonious GLIMMIX models to calculate RSMRs. 6.2 Results Of the three time intervals tested (one-year, two-year and three-year durations), the three-year interval was found to be superior for two reasons. Like the two-year interval it was much more stable than the one-year interval, which was prone to dramatic differences in hospital performance from year to year (particularly for small facilities). This was noted in the movement of hospitals within funnel plots across time. However, the three-year interval also allowed for the greatest number of hospitals to meet the minimum of one expected mortality event over the study duration. The final decision was to use three-year intervals as this is still short enough to notice differences in quality for a given hospital across time but long enough that the models are stable and outliers are clearly identified. Thus, four models and funnel plots were run for each condition (2001-2003, 2004-2006, 2007-2009, and 2010-2012).  Interval 2001-2003 2004-2006 2007-2009 2010-2012 AMI Hospitals ≥ 5 patients 83 78 72 72 Hospitals ≥ 1 expected death 71 68 66 67 Hospitals with RSMR > 95% CI 2 3 5 1 Stroke Hospitals ≥ 5 patients 99 81 80 75 Hospitals ≥ 1 expected death 79 63 67 67 Hospitals with RSMR > 95% CI 4 3 2 3 Trauma Hospitals ≥ 5 patients 110 90 86 79 Hospitals ≥ 1 expected death 79 63 62 57 Hospitals with RSMR > 95% CI 4 2 4 3 Table 6-1. Number of hospitals that satisfy inclusion and exclusion criteria, and flagged as performance outliers using funnel plots CI=confidence interval; RSMR=risk standardized mortality ratio  67  Table 6-1 shows that the number of hospitals accepting five or more patients per year with at least one expected death per interval decreases over time. This can be explained by the reduction in hospitals providing these services over time owing to the centralization of services. Figure 6-1 is a panel of four funnel plots created for trauma. The jump in expected mortalities seen for the period 2010-2012 coincides with the timing of the restructuring of paramedic services in the province and the change in how patients are triaged, first mentioned in Chapter 4. Therefore, the combined effect of the centralization of services through the closure of smaller facilities along with shifts in triage behavior may explain the larger jump in admissions and expected mortalities. This is further corroborated by the observation that no patients were first admitted to a non-acute care facility during 2010-2012. 68       Figure 6-1. Trauma RSMR for 30-day mortality rates0.220 20 40 60 80 100 120Risk Standardized Mortality RatioExpected Number of Deaths 2001-20030.220 20 40 60 80 100 120Risk Standardized Mortality RatioExpected Number of Deaths 2004-20060.220 20 40 60 80 100 120Risk Standardized Mortality RatioExpected Number of Deaths 2007-20090.220 20 40 60 80 100 120Risk Standardized Mortality RatioExpected Number of Deaths 2010-2012Teaching HospitalLarge Community HospitalMedium Community HospitalSmall Community HospitalNot Acute Hospital69  These funnel plots along with those created for both AMI and stroke (not shown here) were used to identify hospitals whose RSMR exceeded the expected 30-day mortality rate after adjusting for patient, provider, and facility differences. Any hospital that fell outside the upper 95% CI band (upper thick dashed line) was flagged as offering sub-optimal quality (variable name: qualityflag95). A second variable (qualitfyflag90) flagged an additional group of hospitals whose quality was within the 95% CI but exceeded the 90% CI for subsequent sensitivity analysis. Across the literature, 30-day mortality rates were used to identify facilities with higher than expected mortality rates.149,162,218 A sensitivity analysis was run to determine what, if any, effect there was in changing the outcome of interest from 30-day mortality rates per hospital to 90-day mortality rates. Although the expected number of deaths increased (marginally) there was no major change in the position of the hospitals relative to the confidence bands. Research in the field162 suggests that longer mortality periods be used to fully capture the risk of mortality as it relates to hospital performance. However, separating the effects of patient quality in acute care versus the community is beyond the scope of this project. Instead, the 90-day mortality analysis was used to ensure that notable differences in hospital performance were not missed. Although the figures shown for comparison are for trauma, the same results were drawn across all conditions and discrete time intervals. In other words, no additional hospitals were flagged as poor performers when I extended the mortality rate interval from 30 to 90 days. For this reason, and because it is more difficult to tie 90-day mortality rate to hospital quality alone, the 30-day mortality rate was used as the quality of care outcome. Table 6-2 lists the outcomes of the independent funnel plots by discrete interval and flag. These hospitals were identified using models controlling on all covariates within the data. 70    Flag AMI Stroke Trauma 2001-03 95% Sechelt Cowichan District  East Kootenay Regional Mount Saint Joseph UBC Campbell River and District General Surrey Memorial Burnaby General 90% East Kootenay Regional UBC  Ridge Meadows Kootenay Boundary Regional Cowichan District 2004-06 95% South Okanagan General East Kootenay Regional Eagle Ridge UBC Mount Saint Joseph Eagle Ridge St. Paul’s 90% Nanaimo General  Saanich Peninsula St. Joseph’s General 2007-09 95% Abbotsford Regional East Kootenay Regional UBC Hospital of Northern BC Royal Jubilee Kelowna General UBC Delta Richmond Burnaby General Vernon Jubilee 90%    2010-12 95% St. Paul’s Delta Penticton Regional Lion’s Gate Cowichan District Richmond Burnaby General 90%  Fort St. John General Richmond  Table 6-2. Hospitals flagged as poor performers by condition and interval, using full models *=closed during study period 71  However, if I compare the outcomes of this model type with those using parsimonious models (see Table 6-3) there is considerable difference in the facilities that are flagged as poor performers. In using the parsimonious model, only large community and teaching hospitals are flagged as outliers, and when compared to other health systems’ use of the RSMR method, it appears this method may systematically bias against large facilities with complex patients.150,219 Consequently, I chose to work with the full models rather than the parsimonious ones with the understanding that this method better allowed me to control for all relevant covariates. This meant including chosen covariates even if they were not found to be associated with mortality during univariate analysis.  Flag AMI Stroke Trauma 2001-03 95% UBC Royal Columbia Victoria General Royal Inland Prince George Kitimat Vancouver General Royal Columbian Royal Jubilee Victoria General Royal Inland Nanaimo General 90% St. Vincent’s Nanaimo General Lions Gate Kelowna General 2004-06 95% St. Paul’s Victoria General Ridge Meadows Victoria General Royal Columbian Victoria General 90% Nanaimo General MSA General*   2007-09 95% Victoria General  Victoria General  90% MSA General  Kelowna General Royal Columbian 2010-12 95% Royal Jubilee Victoria General Royal Inland  90%  Royal Columbian Kelowna General  Table 6-3. Hospitals flagged as poor performers by condition and interval, using parsimonious models MSA=Matsui-Sumas Abbotsford; *=closed during study period  6.3 Summary My interest in quality is narrowly focused on observed vs. expected mortality, after controlling for access and appropriateness. The measure of quality chosen here is based on RSMRs rather than HSMRs. HSMRs use hospitals as the unit of analysis, consider mortality among a select set of conditions that account for about 80% of deaths, and consequently provide a generic measure of 72  overall hospital performance. RSMRs in contrast use the patient as the unit of analysis and are condition-specific. Given the focus of this thesis, and the centrality of condition-specific care and outcome, RSMRs were chose as the appropriate measure. Using the RSMR and funnel plot method, I was able to identify hospitals whose observed mortalities were higher than expected for a facility of their size and location, given patients’ conditions. Contrary to some research associating poor performance with small size, the hospitals that I identified as performance outliers using full models were not predominately of any specific hospital type or size. However, when I used parsimonious models, large teaching or community hospitals were systematically flagged as poor performers. Since these larger facilities were likely experiencing patient loads that were unique and more complex that what can readily be captured in administrative data, it is possible that the parsimonious models were not adjusting on factors that would remove bias against the performance of these larger facilities. As such, I worked with the output from the full models moving forward. What does remain consistent across both modeling exercises is that small facilities are not routinely flagged as outliers, as some literature would suggest. Thus, for the purposes of the rest of this thesis, episodes of hospitalization that occurred within these flagged hospitals during the interval of underperformance were coded with a flag (e.g. qualityflag95) set to “1”, while episodes of hospitalization that did not include care at such hospitals had that same flag set to “0”. Patients who died without receiving care, or were treated at a facility that could not be evaluated given the criteria I used (≥5 events admitted, and ≥ 1 mortality during evaluation interval) were coded as qualityflag95=99 for future analysis purposes.  73  Chapter 7: Investigating the impact of redistribution People in rural and remote communities tend to have elevated disease burden, higher hospitalization rates, and poorer health outcomes compared with their urban counterparts.17,31,32 Differences in age, socio-economic status and health behaviours account for some of this discrepancy.28,29,32 However, regional variation is generally attributed to access to health services. Although rural residents are more likely to need acute medical services, the distribution of hospitals favours high-density urban centres.2,31,32,78 This, combined with the unique travel burdens rural residents encounter (e.g. geographic obstacles, poor road networks, and inclement weather) reduces access to services and may increase the risk of mortality.25,76,220 In 2002 the province of British Columbia began redistributing its acute care services during a general restructuring of the health authorities. This is not the first time a province has undergone a process of centralization, nor is it unique to the Canadian context. Service redistribution has gained momentum over the past three decades across mature health systems (e.g. UK, USA, and Australia) as providers (public and private) work to improve system safety and efficiency and achieve economies of scale by concentrating services in well-equipped large urban centres.5,34,39,221–223 In BC, this period of service redistribution led to the closures of many small, rural facilities and potentially exacerbated regional variations in disease burden and outcomes. Research examining the effects of service redistribution lacks consensus. Some studies show that redistribution decreases access to acute care services for already vulnerable populations and implicate this excess travel burden with poorer patient outcomes.38 Others find the centralization of services has had little effect on access,37 improves procedural outcomes34,121 and does not contribute to added mortality.122 Overall, the consequences of service redistribution remain unclear owing to a lack of generalizability and methodological limitations. The purpose of this chapter is to understand the impact redistribution has had directly on BC patient outcomes for three time sensitive medical events (AMI, stroke, and trauma). This chapter focused on testing the first two hypotheses: 1. Redistribution has had no measurable effect on patient outcomes. 2. Redistribution has had no measurable effect on patients’ access to services.  74  7.1 Methods I used interrupted time series (ITS) models to examine how the recent centralization of hospital services in British Columbia affected patient outcomes for AMI, stroke and trauma events. The redistribution of acute care services presents an excellent natural experiment to test how access to care influences patient outcomes immediately and over time.224 To strengthen the design, I included a control group comprised of patients who were not affected by local hospital closures. In doing so, issues of internal validity such as history, instrumentation, and selection, were further mitigated in the ITS models.42,224 7.1.1 Data preparation Each episode of hospitalization during the study period (April 1 1999 to March 31 2013) counted as a single observation. Patients with multiple discrete medical events appeared in the data per episode. Patients whose underlying cause of death was AMI, stroke or trauma and who expired without being admitted to hospital were also included as unique observations. 7.1.1.1 Intervention group membership The first task was to flag patients affected by redistribution as members of the intervention group. These were individuals whose local facility either closed during the study period or ceased providing services for patients with the diagnoses of interest. In other words, “stopped” is a condition-specific variable describing patients in communities whose closest facility changed (access changed). An alternative would be to differentiate full closure from loss of service. This method of structuring the data was not pursued because it was not clear within the data which facilities were full closures. In any case, the questions here are condition-specific, so it is most important to define access of appropriate care more than to define facilities that are “open”. To start, I used my postal code to hospital travel matrix and sorted the list to find the closest relevant (i.e. capable of providing appropriate care) hospital per postal code. This created a list of nearest facilities as of 1999. I appended the name of the nearest hospital to each event, irrespective of whether or not the patient visited this hospital or received treatment elsewhere. To determine whether the nearest hospital provided care at the time of the patient’s event, I assigned each event to the first facility visited and created a second matrix of number of admissions per condition by month and year. This inflated the number of patients who received care at small 75  hospitals without threatening large hospitals’ role in providing care. Hospitals that admitted zero patients for three consecutive quarters were flagged as no longer providing care (stopped=1) for the condition on the first quarter their patient admission was zero. Again, this was irrespective of whether or not the facility closed entirely. Thus, given that my data set runs to March 31 2013, the cutoff date for facilities flagged as no longer admitting patients was June 30, 2012. I chose the three quarter period because upon initial inspection of the data I noticed that some facilities accepted no patients in one quarter but several in the next. This typically occurred in small facilities where local events could easily fluctuate and provide spurious stop dates. I also observed that facilities that admitted no patients for at least two quarters stopped admitting indefinitely. Thus, to avoid misclassifying facilities as stopped, I used empirical evidence to set the three quarter threshold as the appropriate indicator for permanent termination of services. Patients whose closest facility was flagged as stopped (at some point during the study period) were categorized as members of the intervention group (intervention=1), while all other patients were non-members (intervention=0) and potential control matches. 7.1.1.2 Matched controls Unlike other study designs that rely on carefully selected matched controls to eliminate bias in analysis, ITS controls can be more loosely matched to “cases”. In this study, cases were all patients identified as members of the intervention group – ones whose local facility stopped admitting patients with their condition at some point during the study period. Meanwhile, controls were patients who were first admitted to a hospital where the condition-specific volume was ±10 patients of the hospital where cases received care (not necessarily the closest facility). Simulations loosening the treating facilities’ volume restriction from (±10 to as much as ±75 patients) only added an additional 5-7% of controls to each condition but sacrificed the general comparability in treating facilities between groups since quality may be tied to volume225 and volume differences that large describe very different hospital settings. Controls were also matched to cases by the year of the event occurrence. I used the matching macro provided by Jennifer L. Waller from the Medical College of Georgia with minor corrections for the match selecting exercise in SAS.226 I also re-ran the interrupted time series models matching controls to cases using propensity score matching and controlling for all possible covariates related with the probability of receiving the intervention (living in a community where the local facility stopped treating patients). These 76  were: age, sex, comorbidity, extent of urbanization using metropolitan mix (MIZ), neighborhood income quintile, method of transport to hospital, and event year. I used the propensity score matching macro provided to the SAS community by Lori S. Parsons from the Ovation Research Group in Washington and conducted nearest neighbor 1:1 matching.227 However, I found that my analyses were underpowered owing to difficulties matching cases with controls (see Appendix B). Preferring to include all cases and compare pre- versus post-intervention differences in this group with a control group that ruled out the history threat to validity, I opted to work with the stable, loosely matched models first described. 7.1.1.3 Pre- vs. post-intervention observation time Facilities stopped offering care for given conditions at different times within the study period. Thus, instead of structuring my ITS models using calendar time, I worked with study time. For members in the intervention group, the event’s occurrence was tabulated based on whether the quarter it occurred in was before or after the quarter the facility ceased admitting patients. Patients who experienced a medical event after their closest hospital stopped treating the condition were part of the post-intervention group (positive time since intervention), while patients whose event occurred before the cessation of service were part of the pre-intervention group (negative time since intervention). Controls were assigned time based on their matched cases’ study time. 7.1.1.4 Outcomes The primary outcome of interest was 30-day mortality rate. I also conducted a sensitivity analysis using 90-day mortality rate. Both 30-day and 90-day mortality rates were tabulated per intervention group by quarter of interest. Mortality rates included patients who expired without admission to a hospital. Similarly, changes in proportions of patients transferred were explored across all three conditions. 7.1.2 ITS models While data were prepared using SAS, they were analyzed using R. I plotted the intervention group and control group data to visually inspect it for outliers, linear trends, and quality issues. I applied a standard ordinary least squares (OLS) regression with time series specification and used three methods to check for autocorrelation (patterns in the data over time). The first was the two-sided Durbin-Watson test that provides a statistical test to check for autocorrelation and its 77  direction (results < 2 suggest positive correlation, results > 2 suggest negative correlation). Since I organized my data by annual quarters, I tested for a maximum lag of four to ensure that no annual seasonality was responsible for underlying patterns in the data. Although I tested for autocorrelation, I was not expecting to find any given I was working with study rather than calendar time. I followed this with a plot of residuals from the OLS to visually inspect for any relationship between consecutive points. Lastly, I used autocorrelation plots to visually confirm the absence of both autocorrelation and moving averages – or where needed, to correct the OLS to adjust for their presence. Once concerns of autocorrelation were ruled out, I ran the final ITS models using the basic time series with control group as follows: 𝑜𝑢𝑡𝑐𝑜𝑚𝑒𝑗𝑡 = 𝛽0 +  𝛽1 ∙ 𝑡𝑖𝑚𝑒𝑡 +  𝛽2 ∙ 𝑔𝑟𝑜𝑢𝑝𝑘 + 𝛽3 ∙ 𝑔𝑟𝑜𝑢𝑝𝑘 ∙  𝑡𝑖𝑚𝑒𝑡 +  𝛽4 ∙ 𝑙𝑒𝑣𝑒𝑙𝑡 +  𝛽5∙ 𝑡𝑟𝑒𝑛𝑑𝑡 +  𝛽6 ∙ 𝑙𝑒𝑣𝑒𝑙𝑡 ∙ 𝑔𝑟𝑜𝑢𝑝𝑘 +  𝛽7 ∙ 𝑡𝑟𝑒𝑛𝑑𝑡 ∙ 𝑔𝑟𝑜𝑢𝑝𝑘 +  𝜀𝑗𝑘𝑡 The intervention j was stopped, and t was study time in three-month intervals pre- or post-intervention. In these models, 𝛽0 estimated the baseline (pre-intervention) thirty-day mortality rate using controls; 𝛽1 estimated the pre-intervention change in trend for controls; 𝛽2 estimated the pre-intervention difference in mortality between cases and controls; 𝛽3 estimated the pre-intervention difference in trend between cases and controls; 𝛽4 and 𝛽5 estimated the changes in level and trend for controls post-intervention, respectively; 𝛽6 estimated the difference in level between cases and controls following the intervention; 𝛽7 estimated the difference in rate between cases and controls post-intervention; and 𝜀𝑗𝑘𝑡 was the error term. Figure 7-1 is a visual representation of what a standard interrupted time series with control looks like. The solid blue line represents the pre- and post-intervention control group while the green represents the observed changes pre- and post-intervention for the case group. The two main covariates of interest in the output of the models are 𝛽6 and 𝛽7 as they highlight how the intervention group differs from the control group and from its own level and trend prior to the intervention.  78    Figure 7-1. Interrupted time series with control schematic Blue represents the control group, while green represents the treatment or case group 7.2 Results 7.2.1 Eligible facilities Table 7-1 to Table 7-3 list all the facilities patients were first taken to and are organized by the year of the last date they admitted patients, per condition. As a reminder, for the purposes of this study, a hospital was any facility identified as an acute care hospital according to CIHI or one that provided 24/7 access to a physician capable of diagnosing and stabilizing a patient before transfer to definitive care. At the same time, extended care facilities and residences were excluded. This generous definition of a hospital expanded the realm of facilities that provide care to include those with appropriate diagnostic equipment such as ECGs and access to an experienced physician as recommended in condition specific guidelines. As can be seen from the list, several facilities in 79  BC have “hospital” written in their name but do not meet the inclusion criteria while others that describe themselves as diagnostic centres and clinics meet the minimum definition of hospital for this study. Since the main objective of this dissertation is to understand the effects of centralization or redistribution of care, it made sense to include facilities that patients were regularly taken to for immediate care of AMIs, strokes or trauma beyond CIHI designated hospitals. By extending the definition to include these facilities and by capturing how the consolidation of services has influenced their continued service provision, my study can shed light on the true impact of centralization. As can be seen from the tables, there were many facilities that were the initial point of care for patients that do not fall within my definition of hospital. The third column of Table 7-1 to Table 7-3 identifies which facilities fail (Hospital designation=N) to meet the criteria. Final year Facility Hospital designation 2002 Enderby & District Memorial Hospital Kimberley & District Hospital Fraser Lake Diagnostic & Treatment Centre Port Alice Hospital Nisgaa Valey Diagnostic & Treatment Centre Riverview Hospital Y N Y Y N N 2003 St. Vincent’s Hospital (Heather) Saint Mary’s Hospital (New Westminster) Chase Diagnostic & Treatment Centre N Y Y 2004 Overlander Extended Care Hospital Menno Extended Care Hospital Mount St. Francis Extended Care Hospital Elkford & District Diagnostic & Treatment Centre Stewart General Hospital N N N Y Y 2005 Holy Family Hospital BC Rehab Society (George Pearson Centre) Fellburn Hospital Trillium Extended Care Hospital N N N N 2006 Mount St. Mary Extended Care Hospital Juan de Fuca Extended Care Hospital Keremeos Diagnositc & Treatment Centre Pouce Coupe Community Extended Care Hospital N N N N 2007 St. Vincent’s Langara Hospital Brock Fahmi Extended Care Unit Gorge Road Hospital St. Bartholomew’s Hospital N N N Y 2008 Louis Brier Extended Care Hospital N 80  Final year Facility Hospital designation St. Michael’s Centre Extended Care Hospital MSA General Hospital N Y 2009 Summerland Health Centre N 2010 Chemainus General Hospital Slocan Community Health Centre Victorian Community Health Centre of Kaslo Sparwood General Hospital N N N Y 2011 Ashcroft & District General Hospital Stuart Lake Hospital Y Y 2012 Pacific Health Care Society Extended Care Hospital Princeton General Hospital Dr. Helmcken Memorial Hospital Ladysmith & District General Hospital MacKenzie & District Hospital Tumbler Ridge Health Care Centre N Y Y Y Y N Table 7-1. Facilities accepting AMI patients by last year of admission (16 hospitals and 26 non-hospital facilities)  Final year Facility Hospital designation 2001 Fraser Lake & District Diagnostic & Treatment Centre Tahsis Health Centre Y Y 2002 Fellburn Hospital Enderby & District Memorial Hospital Keremeos Diagnostic & Treatment Centre Bastion Place Kimberley & District Hospital Sparwood General Hospital Master Misericordiae Health Care Facility Nisga’a Valley Diagnostic & Treatment Centre Terraceview Lodge N Y N N N Y N N N 2003 Castlegar & District Community Health Centre Y 2004 St. Vincent’s Hospital (Heather) St. Mary’s Hospital (New Westminster) Pleasant Valley Health Centre Mount St. Francis Extended Care Hospital Peace Lutheran Extended Care Riverview Hospital N Y N N N N 2005 Trillium Extended Care Hospital Pouce Coupe Community Extended Care Hospital N N 2006 Stewart General Hospital Y 2007 St. Michael’s Centre Extended Care Hospital N 81  Final year Facility Hospital designation Gorge Road Hospital Mount St. Mary Extended Care Hospital Ladysmith & District General Hospital Slocan Community Health Centre of Kaslo N N Y N 2008 Louis Brier Extended Care Hospital Brock Fahmi Extended Care Unit St. Bartholomew’s Hospital MSA General Hospital Menno Extended Care Hospital N N Y Y N 2009 St. Vincent’s Langara Hospital Juan de Fuca Extended Care Hospital Overlander Extended Care Hospital Houston Diagnostic & Treatment Centre N N N Y 2010 Chemainus General Hospital Cormorant Island Community Health Centre (formerly St. George’s Hospital) N Y 2011 Princeton General Hospital Summerland Health Centre Stuart Lake Hospital Y N Y 2012 Holy Family Hospital N Table 7-2. Facilities accepting stroke patients by last year of admission (14 hospitals and 27 non-hospital facilities)   Final year Facility Hospital designation 2001 Cumberland Diagnostic & Treatment Centre Master Misericordiae Health Care Facility N N 2002 Enderby & District Memorial Hospital Bastion Place Kimberley & District Hospital Sparwood General Hospital Hudson’s Hope Gething Diagnostic & Treatment Centre Peace Lutheran Extended Care Ocean Falls Diagnostic & Treatment Centre Nisga’a Valley Diagnostic & Treatment Centre Terraceview Lodge Riverview Hospital Y N N Y Y N N N N N 2003 Saint Mary’s Hospital (New Westminster) Pleasant Valley Health Centre Pemberton Diagnostic & Treatment Centre Y N Y 2004 St. Vincent’s Hospital (Heather) Fellburn Hospital N N 82  Final year Facility Hospital designation Mount St. Francis Extended Care Hospital Castlegar & District Community Health Centre Stewart General Hospital N Y Y 2005 Trillium Extended Care Hospital N 2007 Gorge Road Hospital Ladysmith & District General Hospital Fraser Lake Diagnostic & Treatment Centre N Y Y 2008 Brock Fahmi Extended Care Unit Mount St. Mary Extended Care Hospital MSA General Hospital Menno Extended Care Hospital Pouce Coupe Community Extended Care Hospital Houston Diagnostic & Treatment Centre N N Y N N Y 2009 Louis Brier Extended Care Hospital St. Michael’s Centre Extended Care Hospital St. Vincent’s Langara Hospital Juan de Fuca Extended Care Hospital Overlander Extended Care Hospital N N N N N 2010 Summerland Health Centre N 2011 Holy Family Hospital Slocan Community Health Centre of Kaslo Stuart Lake Hospital N N Y 2012 Port McNeill & District Hospital Victorian Community Health Centre Kaslo R.W. Large Memorial Hospital Y N Y Table 7-3. Facilities accepting trauma patients by last year of admission (14 hospitals and 28 non-hospital facilities)  To interpret Table 7-1, there were 42 facilities that ceased to admit AMI patients sometime before June 30th, 2012. Of these facilities, 16 qualified under my extended definition of “hospital”. These 16 facilities were either hospitals according to CIHI or were able to provide minimum standards of care in treating AMIs (i.e. access to appropriate diagnostic equipment and trained physician 24/7 that can provide a thrombolytic intravenously). This does not mean that each hospital was equipped to provide definitive care such as percutaneous coronary intervention, but these facilities were able to confirm the event was an AMI and provide recommended pharmacotherapy. Additionally, some patients were taken to one of 26 centres that lacked the diagnostic equipment and necessary staff to provide initial diagnosis and/or stabilization. Similar results were found for stroke patients and shown in Table 7-2. Of the 41 facilities that stopped 83  accepting patients, 14 were hospitals and thus indicate a true decrease in the number of accessible facilities capable of providing initial treatment, while 25 were centres that should not have admitted stroke patients. Lastly, as shown in Table 7-3, 14 hospitals able to initially treat major trauma events, stopped doing so. In other words, over the duration of the study period, some patients were inappropriately taken to any of 28 centres not capable of stabilizing major trauma. Put differently, in 1999 patients with a heart attack were admitted to any of 121 facilities across the province. By 2013, this number declined to 79. However, there were 16 hospitals which stopped admitting patients. The additional “loss” of services actually accounted for corrections of transport and admission processes that in the past were inappropriate. This could be attributed to improvements in triaging, and turning away self-transports. The same can be said for ischemic stroke and major trauma patients. In 1999, ischemic stroke events were admitted to any of 120 facilities in 1999 but by 2013, they were only admitted to 79 hospitals. Patients with urgent trauma needs were first treated at 122 facilities in 1999 but by 2013 there were only 80 facilities treating them. Thus, there is some merit to concerns of hospital closures beginning in 2001 under BC’s centralization of services. However, my results suggest two things. First, a small proportion of patients were inappropriately admitted to non-hospital facilities for serious medical events. Second, redistribution led to the consolidation of care at hospital centres. This means patients stopped receiving care at non-hospital sites and were instead taken directly to hospitals. At the same time, some small hospitals ceased admitting patients. 84   Figure 7-2. Hospitals by type, 1999 Figure 7-3. Hospitals by type, 2013 For simplicity’s sake, Figure 7-2 and Figure 7-3 map the location of hospitals and show the extent of changes to admission over time. Figure 7-2 plots every designated hospital that accepted AMI, trauma and stroke patients in 1999 (N=92) while Figure 7-3 has every hospital that still accepted said patients in 2013 (N=79). Thirteen non-acute hospital centres (by CIHI’s definition, but still appropriate for initial intervention) and small hospitals stopped admitting all three types of patients by 2013. 85  7.2.2 Data restrictions There were 7469 controls matched to 8630 cases for AMI events, 6372 controls matched to 7843 cases for stroke, and 14,318 controls matched to 16,056 cases of trauma events. However, restricting the study populations to patients who visited hospitals as defined in this study and thus had relevant distance to care calculations left me with 6881 (-7.8%) controls and 8000 (-7.3%) cases for AMI analysis, 6362 (-0.2%) controls and 7827 (-0.2%) cases for stroke analysis, and 11,680 (-18.4%) controls and 13,430 (-16.4%) cases for trauma analysis. Although the study duration was extensive enough to provide 54 periods (13.5 years) before and 52 periods (12 years) after the intervention we restricted the analysis to just three years before the intervention and three years following the intervention (12 quarters pre- and post-intervention). This was because as we looked at periods further from the intervention time (time zero), there were fewer admissions that occurred making the analysis sensitive to unstable event rates. At the same time, we agreed that the important impacts of redistribution would be assessed comparing patient outcomes immediately before and after local closures. Expanding the ITS to include as many as 54 periods before and 52 after weakened the effects as communities and health authorities likely adjusted to the changes in access to services. Thus, further restricting the samples to those whose event occurred within three years of the intervention or matched controls left 1996 (-59.2%) controls and 3267 (-52.5%) cases for AMI analysis, 1604 (-74.8%) controls and 2852 (-63.6%) cases for stroke analysis, and 3640 (-68.7%) controls and 6318 (-53.0%) cases for trauma analysis. This meant that for AMI there were a minimum of 71 patient events to a maximum of 164 per period, 70 to 142 respectively for stroke, and 189 to 299 respectively for trauma incidents. For comparison, when using propensity score matched samples, I had 25-60 events per treatment group per condition for each time period. 7.2.3 Patient demographics Interrupted time series models were restricted to patients who visited a hospital where cases were members of the intervention group (pre or post), and controls were patients whose local facility never stopped treating patients but were treated at a hospital with similar patient volume to the cases (±10 patients). Control events were matched to case events by patient volume for two reasons: a) ITS does not require strict matching criteria as most concerns about validity are resolved by the study design; 86  b) if the outcome of interest is 30-day mortality, and hospital size/patient volume/facility experience do matter, then it makes sense to restrict controls to patients who were treated at similar facilities.  Additionally, control events were matched to cases without replacement; meaning once a control event was matched with a case event it could not be used as a match for another case as well. Table 7-4 shows the demographic results of the matched cases and controls per condition. Differences in proportions were evaluated using Chi-square tests and t-tests were used for means and standard deviations. Cases were generally poorer (lower income quintiles), lived farther away from the treating hospital (time to care), but were healthier (lower Charlson comorbidity index). AMI cases were transferred less frequently (17.5% versus 30.1%, p-value<0.001); trauma cases were younger (55.2 (SD=24.8) vs. 61.0 (SD=24.3) years, p-value<0.001); stroke cases were more likely female (50.2% vs. 46.0%, p-value=0.007); but trauma cases were less likely female (46.8% vs. 52.7%, p-value<0.001) than controls. A high proportion (17.8% to 22.1%) of patients self-transported (did not use paramedic services) to hospital across conditions without significant differences between cases and controls. These differences in case and control subjects are accounted for in the existing levels and trends of my models. 87    AMI Stroke Trauma Controls N=1996 (%) Cases N=3267 (%) P-value Controls N=1604 (%) Cases N=2852 (%) P-value Controls N=3640 (%) Cases N=6318 (%) P-value Female 656 (32.9) 1117 (34.2) 0.32 738 (46.0) 1432 (50.2) 0.007 1918 (52.7) 2957 (46.8) <0.001 Income quintile Lowest Second Third Fourth Highest Missing 429 (21.5) 435 (21.8) 405 (20.3) 367 (18.4) 342 (17.1) 18 (0.9) 921 (28.2) 694 (21.2) 552 (16.9) 485 (14.9) 473 (14.5) 142 (4.4) <0.001 321 (20.0) 361 (22.5) 361 (22.5) 276 (17.2) 271 (16.9) 14 (0.9) 761 (26.7) 642 (22.5) 501 (17.6) 394 (13.8) 437 (15.4) 117 (4.1) <0.001 887 (24.4) 725 (19.9) 718 (19.7) 694 (19.1) 573 (15.7) 43 (1.2) 1790 (28.3) 1194 (18.9) 959 (15.2) 937 (14.8) 1083 (17.1) 355 (5.6) <0.001 Charlson comorbidity index CCI=0 CCI=1 CCI=2 CCI=3 CCI=4+ 1361 (68.2) 388 (19.4) 125 (6.3) 78 (3.9) 44 (2.2) 2301 (70.4) 431 (13.2) 346 (10.6) 128 (3.9) 61 (1.8) 0.09 976 (60.9) 391 (24.4) 134 (8.4) 64 (4.0) 39 (2.4) 1929 (67.6) 452 (15.9) 335 (11.8) 62 (2.2) 74 (2.6) <0.001 3038 (83.5) 266 (7.3) 226 (6.2) 60 (1.7) 50 (1.4) 5521 (87.4) 337 (5.3) 350 (5.5) 55 (0.9) 55 (0.9) <0.001 Transferred 601 (30.1) 572 (17.5) <0.001 82 (5.1) 194 (6.8) 0.02 91 (2.5) 159 (2.5) 0.96 Self-transport 420 (21.0) 722 (22.1) 0.37 323 (20.1) 580 (20.3) 0.87 648 (17.8) 1160 (18.4) 0.45 Age Mean (SD) Range 68.00 (15.3) 18 - 104 68.24 (15.3) 18 – 105 0.58 73.00 (14.8) 18 – 101 72.77 (14.4) 18 – 101 0.69 61.05 (24.3) 18 - 106 55.17 (24.8) 18 – 104 <0.001 Time to care, mins Mean (SD) Range N=1642 45.77 (123.5) 0.25 – 1413 N=2657 78.43 (171.9) 0.09 – 1503 <0.001 N=1356 32.11 (105.9) 0.21 – 1507 N=2330 55.22 (159.8) 159.81 – 2343 <0.001 N=2997 57.57 (157.3) 0.11 – 1753 N=5291 92.10 (203.0) 0.01 – 1691 <0.001 Table 7-4. Demographics by condition and interrupted time series intervention group (cases and controls) Controls were matched on condition, year of event, and hospital condition specific volume ±10 CCI=Charlson comorbidity index; SD=standard deviation; mins=minutes88  7.2.4 30-day mortality rates over time Figure 7-4 to Figure 7-6 show the visual output of the ITS with segmented regression lines plotted for each model. From visual inspection, the 30-day mortality rate appears consistently higher in the intervention group (lost access) than in controls (no change in access) irrespective of the condition. However, given underlying differences in the cases and control group, differences between the groups may be biased. As such, the control group is only used to confirm that no other policy change or event explains pre- versus post-treatment differences in the case group. For purposes of illustration, the dashed lines in the figures represent the counterfactual line segments. The counterfactual of the intervention group takes into consideration the observed level and trend changes in the control group. Hence it rarely will appear as a continuation of the observed outcome preceding the intervention as is the case for the control counterfactual.   Figure 7-4. AMI, 30-day mortality rate 89   Figure 7-5. Stroke, 30-day mortality rate  Figure 7-6. Trauma, 30-day mortality rate 90  Table 7-5 provides the statistical output of my segmented regressions comparing the intervention group with the control (refer to Figure 7-1 for a reminder of each covariate), and patient outcomes pre- and post-intervention. Patients in the intervention group had a significantly different mortality rate than those in the control group (see β2) but there was no statistical difference in change in level or trend in the pre- and post-intervention periods of the respective groups. Model 1 AMI N=5,263 Stroke N=4,456 Trauma N=9,958 β (95% CI) P-value β (95% CI) P-value β (95% CI) P-value Controls, pre-stop (β0) 7.04% (4.04, 10.05) <0.001 5.25% (2.37, 8.13) <0.001 1.20% (0.31, 2.10) 0.012 Controls’ trend, pre-stop (β1) -0.19% (-0.61, 0.24) 0.397 0.00% (-0.41, 0.40) 0.987  -0.03% (-0.16, 0.10) 0.647 Difference between cases and controls, pre-stop (β2) 7.31% (3.06, 11.56) 0.002 6.78% (2.70, 10.84) 0.002 1.86% (0.59, 3.12) 0.006 Difference in trend between cases and controls, pre-stop (β3) 0.11% (-0.49, 0.71) 0.714 -0.21% (-0.79, 0.36) 0.464 0.08% (-0.10, 0.26) 0.390 Controls, post-stop (β4) 3.34% (-1.29, 7.97) 0.165 4.42% (-0.20, 8.86) 0.058 0.70% (-0.68, 2.08) 0.327 Controls’ trend, post-stop (β5) 0.15% (-0.49, 0.79) 0.649 -0.07% (-0.68, 0.54) 0.817 -0.04% (-0.15, 0.23) 0.713 Difference between cases and controls, post-stop (β6) -3.91% (-10.46, 2.64) 0.249 0.09% (-5.36, 7.21) 0.774 -1.54% (-3.49, 0.41) 0.129 Difference in trend between cases and controls, pre-stop (β7) -0.12% (-0.79, 1.02) 0.789 -0.77% (-0.94, 0.79) 0.863 0.03% (-0.24, 0.30) 0.838 Table 7-5. Interrupted time series output for 30-day mortality rates, per condition Controls were matched on condition, year of event, and hospital condition specific volume ±10 N’s are total of cases and controls per condition P-values < 0.05 are significant 91  Table 7-6 provides the output of the segmented regressions where the exercise was repeated but the outcome of interest changed from 30-day mortality rates to 90-day mortality rates (model 1a). As with the quality variable in Chapter 6, this sensitivity analysis was conducted to see if the patterns held across longer periods of time. Recall that the number of patients per period was low for some periods (as little as 70 for stroke). By extending the mortality period to 90 days, the mortality rates per period of time stabilized making the segmented regressions less sensitive to mortality rate outliers. As can be seen in Table 7-6, the patterns seen in Table 7-5 persist. In these cases, patients in the intervention group had a significantly higher mortality rate at 90 days than their control counterparts for both AMI and stroke. In the case of trauma, the difference in mortality rate became non-significant. This may be because the overall mortality rate between groups was low and by extending the mortality period from 30-days to 90-days the effects of outliers were muted. 92  Model 1a AMI N=5,263 Stroke N=4,456 Trauma N=9,958 β (95% CI) P-value β (95% CI) P-value β (95% CI) P-value Controls, pre-stop (β0) 8.27% (5.00, 11.55) <0.001 7.19% (3.56, 10.82) <0.001 3.17% (1.58, 4.77) <0.001 Controls’ trend, pre-stop (β1) 0.04% (-0.43, 0.50) 0.872 -0.15% (-0.66, 0.37) 0.574 0.03% (-0.20, 0.25) 0.826 Difference between cases and controls, pre-stop (β2) 4.95% (0.32, 9.58) 0.042 8.97% (3.84, 14.11) 0.001 2.55% (0.30, 4.81) 0.032 Difference in trend between cases and controls, pre-stop (β3) -0.30% (-0.96, 0.36) 0.374 -0.46% (-1.19, 0.27) 0.220 -0.01% (-0.34, 0.30) 0.915 Controls, post-stop (β4) -1.52% (-6.57, 3.54) 0.559 2.50% (-3.10, 8.10) 0.387 0.51% (-1.94, 2.97) 0.685 Controls’ trend, post-stop (β5) 0.04% (-0.66, 0.73) 0.922 0.71% (-0.70, 0.84) 0.858 0.03% (-0.31, 0.37) 0.849 Difference between cases and controls, post-stop (β6) -0.87% (-7.23, 7.06) 0.981 -1.05% (-8.96, 6.86) 0.796 -1.09% (-4.57, 2.39) 0.544 Difference in trend between cases and controls, pre-stop (β7) 4.74% (-0.51, 1.46) 0.352 0.98% (-0.11, 2.08) 0.085 1.48% (-0.33, 0.63) 0.550 Table 7-6. Interrupted time series output for 90-day mortality rates, per condition Controls were matched on condition, year of event, and hospital condition specific volume ±10 N’s are total sum of cases and controls P-values < 0.05 are significant93  The patterns seen in both sets of ITS models fail to reject the null hypothesis that there was no impact on patient health following changes in local access to acute care; put more simply, hospital redistribution does not appear to alter patient outcomes, as measured by 30-day mortality. However, they show that patients in the control group fared better than cases irrespective of the timing of their event relative to local closures. There are two possible interpretations of this. The first is if patients’ outcomes are indeed strongly dictated by time to care, then time to care has not changed as a result of local facility closures. If this is so, then the difference in mortality rate between the cases and controls may be an artifact of the underlying practice of transporting the majority of cases to hospitals other than their closest facility even before the local centre ceased to treat the patient type entirely. The second possibility is there may be a difference in patients’ health between cases and controls that contributes to consistently poorer health outcomes for cases. I was able to test the first possibility using ITS with segmented regressions and an additional analysis described in the section below. This is further investigated in Chapter 8 when I use survival analysis to examine how time to care affects patient outcomes after controlling for patient demographic differences. Table 7-4 shows the demographic differences between cases and controls, and although cases are typically older and poorer, they do not have poorer health when described using the comorbidity measure. 7.2.5 Proportion of patients not taken to their closest facility Section 7.2.4 examined changes in access to the closest facility, described as a loss of access when facilities stopped admitting patients for a given condition. The ITS analysis showed that after a change in access, cases did not experience statistically significant different mortality rates than before the intervention. This analysis thus tested, and helped answer the first hypothesis stated in Chapter 1: “Redistribution has had no measureable effect on patient outcomes”. In testing the hypothesis, the assumption was that patients had to travel longer to access care once the closest facility stopped treating their condition. Thus, a change in time to care (and therefore access) was implicit in the constructing of the case group. However, as has been described in the literature and summarized in Chapter 2, time (or distance) to care is one proxy for access. Constructed as such, it fails to capture a second element of the theme of access, namely that “the proof of access is use of services, not simply the presence of a facility.”55 Thus, hypothesis 2: “Redistribution has had no measureable effect on patients’ access 94  of services” was tested by examining whether cases and controls were using the closest facility prior to the intervention differently. This is in effect a test of whether or to what extent redistribution of hospital services affected existing patterns of health care services use for the three conditions of interest. 7.2.5.1 Methods For this analysis I repeated the methods explained above with a few revisions. In the initial ITS examining whether changes in access affected patient outcomes, I matched each patient with the closest appropriate hospital for the event available at the start of the intervention period (still available by the fourth quarter of 2000). I then organized patient events by intervention group if, as mentioned, the closest facility did not admit patients for three consecutive quarters by June 30th 2012. Structuring my analysis like this assumes that patients indeed traveled to the closest facility. In reality, this might not have been the case. This analysis corrects for that assumption by exploring whether patients went to the closest facility or opted for another and whether this preference to travel beyond the closest differed between communities that lost their local hospital and those that did not. Here, I retained the link to the closest facility but added an additional link matching each patient (both cases and controls) to the closest facility that continued to accept patients at the end of the study. I then proceeded to create two mismatch variables. Mismatch00 compared the hospital the patient was first admitted to for the episode of hospitalization with the closest facility as of 2000. If there was a mismatch between the closest facility and the true facility visited (i.e. the patient bypassed the closest facility for care) mismatch00=1, and if the patient went to the closest facility, mismatch00=0. I repeated the exercise comparing the true facility visited at the time of the event with the closest facility available at the end of the study period. Thus, mismatch13=0 if the patient went to the closest facility that admitted patients in 2013 and 1 if not. The reasoning here was that if patients were bypassing local care but time to care matters, then they would be going to a local facility but perhaps not the closest for a variety of reasons (such as assumptions of quality, capacity, and competence). For controls, mismatch00=mismatch13 since the closest facility never changed. For this group, the extent that they used facilities other than the closest would likely be because the event did not occur within the patients’ residential postal code, and there is no reason to believe this changed 95  dramatically over time. Since the administrative data used in this project only provided de-identified patient postal codes but no information on the location of the medical emergency, this mismatch within controls helped capture events that occurred away from home (e.g. at the patient’s place of work, on the road, leisure activity or elsewhere all of which are likely to be outside the home postal code geography but beyond the abilities of this project to confirm). Alternatively, for cases, events that occurred before the local hospital closure (pre-intervention) looked at mismatch00 while post-intervention events compared the facility visited with that closest in 2013 (mismatch13). Although some of the differences were likely due to the same reasons as the mismatches in the controls, differences in mismatches pre- vs. post-intervention, and compared with the controls would signal differences in use patterns. I finalized the preparation for the ITS by developing the prop_mismatch outcome variable. This was the proportion of mismatches based on whether a patient was a case or control and the event occurred pre- or post-intervention by quarter. 7.2.5.2 Proportion mismatch results As Figure 7-7, 7-8 and 7-9 along with Table 7-7 show, unlike the results of Figure 7-4 to Figure 7-6, there were noticeable differences between cases and controls again but also in the pre- vs. post-intervention periods. Anywhere between 60-80% of cases in the pre-intervention period were first admitted to a facility other than the closest. However, following the intervention, this proportion consistently dropped across conditions to about 50% and approached that observed in controls. The range in the pre-intervention group fits my earlier hypothesis that cases bypassed their closest facility to receive care even before that facility stopped offering care entirely as has been demonstrated in other studies.40 This finding suggests facilities stopped treating the condition as a reflection rather than modification of existing patterns of service use. 96   Figure 7-7. Mismatch between true facility visited and closest for AMI   Figure 7-8. Mismatch between true facility visited and closest for Stroke 97   Figure 7-9. Mismatch between true facility visited and closest for Trauma 98   Model 2 AMI N=5,263 Stroke N=4,456 Trauma N=9,958 β (95% CI) P-value β (95% CI) P-value β (95% CI) P-value Controls, pre-closures (β0) 32.97% (27.47, 38.49) <0.001 29.11% (22.90, 35.31) <0.001 44.69% (41.50, 47.88) <0.001 Controls’ trend, pre-closures (β1) -0.05% (-0.83, 0.73) 0.903 -0.79% (-1.66, 0.92) 0.087 0.43% (-0.02, 0.88) 0.071 Difference between cases and controls, pre-closures (β2) 50.06% (42.27, 57.85) <0.001 36.16% (27.38, 44.93) <0.001 32.24% (27.73, 36.76) <0.001 Difference in trend between cases and controls, pre-closures (β3) 1.10% (-0.00, 2.20) 0.057 0.96% (-0.28, 2.20) 0.136 -0.60% (-1.24, 0.03) 0.070 Controls, post-closures (β4) -4.51% (-13.00, 3.99) 0.305 7.90% (-1.67, 17.47) 0.113 -4.18% (-9.10, 0.74) 0.103 Controls’ trend, post-closures (β5) 0.08% (-1.09, 1.26) 0.890 0.47% (-0.86, 1.79) 0.494 -0.04% (-0.73, 0.63) 0.897 Difference between cases and controls, post-closures (β6) -15.50% (-27.52, -3.49) 0.015 -25.27% (-38.80, -11.73) <0.001 -22.69% (-29.64, -15.72) <0.001 Difference in trend between cases and controls, pre-closures (β7) -2.62% (-4.29, -0.96) 0.004 -1.01% (-2.88, 0.86) 0.297 0.46% (-0.50, 1.43) 0.355 Table 7-7. Interrupted time series output for bypass rate, per condition Controls were matched on condition, year of event, and hospital condition specific volume ±10 N’s are sum of cases and controls P-values < 0.05 are significant99  Across all facilities that ceased treating patients by condition, the general trend in aggregated data showed a gradual decline over time in the number of patients admitted while larger hospitals accepted more (see Figure 7-10).  Figure 7-10. Number of patients admitted by hospital type Figure 7-11 to 7-14 are maps of the proportion of patients who went to facilities other than the closest (proportion mismatched) in 2000 and again in 2013. As suggested by the results of the ITS model shown in Table 7-7, as many as 75% of patients in an HSDA traveled beyond the closest facility to be admitted to care. This may not be as surprising as it initially seems. While time to care is acknowledged as a key predictor in patient outcomes for the conditions to which analysis is restricted, several studies show that patients triaged at inappropriate facilities fare worse than those taken directly to definitive care.60–63 At the very least, this analysis reveals a bit more on how service is utilized in BC for time-sensitive events. This is corroborated by results of another ITS that looked at proportions of patients transferred by condition. For trauma, where publications have recommended against stabilizing at a small facility before transferring to definitive care, there was no difference between patient groups or periods of time in terms of proportions transferred (see Table 7-8).60,61,63 0500100015002000250030003500400045001999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012Number of patients admittedYearTeaching Hospital Large Community Hospital Medium Community HospitalSmall Community Hospital Not Acute Hospital100  Model 3 AMI N=5,263 Stroke N=4,456 Trauma N=9,958 β (95% CI) P-value β (95% CI) P-value β (95% CI) P-value Controls, pre-closures (β0) 11.58% (7.59, 15.58) <0.001 2.94% (0.61, 5.27) 0.017 1.05% (-0.06, 2.17) 0.072 Controls’ trend, pre-closures (β1) 0.41% (-0.15, 0.98) 0.160 0.13% (-0.20, 0.45) 0.459 -0.02% (-0.18, 0.14) 0.802 Difference between cases and controls, pre-closures (β2) 14.31% (8.67, 19.97) <0.001 5.27% (1.98, 8.56) 0.003 1.60% (0.03, 3.18) 0.053 Difference in trend between cases and controls, pre-closures (β3) 0.60% (-0.20, 1.40) 0.148 0.17% (-0.30, 0.63) 0.481 0.06% (-0.17, 0.28) 0.616 Controls, post-closures (β4) 1.08% (-5.08, 7.23) 0.733 0.13% (-3.46, 3.71) 0.945 0.70% (-1.02, 2.42) 0.433 Controls’ trend, post-closures (β5) -0.72% (-1.18, 0.13) 0.105 0.17% (-0.33, 0.67) 0.506 -0.01% (-0.25, 0.23) 0.940 Difference between cases and controls, post-closures (β6) -3.4% (-12.13, 5.29) 0.445 3.51% (-1.57, 8.58) 0.183 -0.41% (-2.84, 2.02) 0.741 Difference in trend between cases and controls, pre-closures (β7) -0.95% (-2.15, 0.25) 0.129 -1.10% (-1.18, -0.40) 0.004 -0.18% (-0.35, 0.32) 0.919 Table 7-8. Interrupted time series output for proportion of patients transferred, per condition Controls were matched on condition, year of event, and hospital condition specific volume ±10 N’s are sum of cases and controls P-values < 0.05 are significant101  This aligns with Figure 7-13 and 7-14that show the greatest mismatch in the proportions of patients travelling beyond their closest facility was for trauma patients, and that  difference persists to a greater degree across the two time periods (pre- and post-intervention). Meanwhile, since both AMI and stroke can be initially treated with thrombolytics before reaching definitive care, it makes sense that cases have a higher proportion of transfers than controls.  Figure 7-11. Proportions of patients by HSDA traveling to hospital other than closest for AMI, 1999 Figure 7-12. Proportions of patients by HSDA traveling to hospital other than closest for AMI, 2013 102   Figure 7-13. Proportions of patients by health service delivery area traveling to hospital other than closest for trauma, 1999 Figure 7-14. Proportions of patients by health service delivery area traveling to hospital other than closest for trauma, 2013 103   Figure 7-15. Proportions of patients by health service delivery area traveling to hospital other than closest for stroke, 1999 Figure 7-16. Proportions of patients by health service delivery area traveling to hospital other than closest for stroke, 2013 104  7.3 Summary This chapter addressed the first two hypothesis of this dissertation, testing whether redistribution had a measurable effect on patient outcomes and on patients’ access to services. During the study period, 16 hospitals stopped treating AMIs, and 14 each for both stroke and trauma events. Additionally, a number of small facilities that were not set up to provide appropriate care as defined by this thesis, ceased admitting patients. Using ITS models to compare outcomes for patients whose closest capable hospital stopped admitting patients with those of patients treated in similarly sized hospitals that never stopped (controls) revealed that both 30- and 90-day mortality rates did not differ pre- and post-intervention period for cases. Additionally, patients in communities whose closest facility stopped admitting were transferred more frequently than control counterparts. However, the bypass behavior prior to local facility closures was very high for cases, but following the local stoppage, became similar to that seen in the control groups. This suggests that cessation of treatment may have been in response to underutilization and a reflection of use patterns rather than a driver of change in how services were accessed. 105  Chapter 8: Determining the impacts of access, appropriateness, and quality on patient outcomes Previous chapters developed and provided descriptive analytics of measures of care that have been shown in the literature to affect patients’ outcomes following medical emergencies. A significant methodological limitation of existing work is that these measures (access, appropriateness, and quality of care) are analyzed independent of each other. In this chapter, I test the hypothesis that access to care has no measureable effect on patient outcomes after controlling for quality, and appropriateness of acute care services and compensating mechanisms such as inter-hospital transfers and telehealth services.  By combining the results of Chapter 7, which showed redistribution did not affect patients’ outcomes, with the findings of this chapter, it is my intention to answer the question “What effect did British Columbia’s redistribution of hospital services have on residents’ outcomes, and did quality and appropriateness of care compensate for variations in access of acute care?” 8.1 Methods The goal of this chapter was to build models that estimated the effect that differences in time to care have on patient mortality when controlling for appropriateness and quality of care. 8.1.1 Variable selection These analyses rely on covariates developed earlier and use all observations of patient admissions (that is, excluding patients never admitted to hospital) rather than restricting analyses to patients affected by redistribution and their matched controls as was the case in Chapter 7 (which included patients never admitted to hospital). For access to care, I had three variants to describe travel burden. I chose ground travel time as my access to care proxy but retained Euclidean and ground distance measures for sensitivity analyses (especially to relate to older publications which relied on as the crow flies measures to describe the effects of access to care on patient outcomes). Although aerial distance can serve as an acceptable proxy for access,168 ground travel estimates are superior for describing barriers to access and impacts on health.14 Additionally, using time to care allowed me to assess whether or not the contested “golden hour” mentioned in the literature was relevant in contemporary access 106  to care discussions. Although Newgard et al., demonstrated no significant association between EMS intervals within one hour with those beyond,23 the term continues to be used in North America228 and abroad229,230, and to shape service distribution. It remains a common term in the literature and for comparative reasons, I use time to care to examine it here as well. Lastly, time to care better approximates burden of travel in rural and remote areas, given that road conditions were incorporated in the origin-destination network analysis algorithm. There are two ways of investigating access to care. The first uses an “intention to treat” approach, looking not at where the patient was taken but rather the closest facility the patient could be taken relative to his or her postal code. The second approach, best described as “non-intention to treat”, looks at treatment as received rather than assigned or assumed. Instead of assigning patients to the closest facility, this method looks at the facility to which the patient was admitted and matches a distance from the travel matrix using both the hospital identifier and patient postal code fields. Given how often patients are not taken to their closest facility, as demonstrated in Chapter 7, and that measures of appropriateness and quality reflect the facility to which patients were admitted, this chapter focuses on the actual treating facility. The underlying assumption on which measures of time to care rest is that events occur close to home. This will clearly be untrue in some cases, but the bias it introduces is expected to be non-differential (i.e. no reason to suspect differences in cases and controls in Chapter 7 or by time interval here). Additionally, this chapter uses true time to care provided by ambulance data for events that occurred between April 1 2009 and March 31 2013 as a sensitivity analysis to investigate the extent to which estimated measures of time to care impact the models’ validity. 8.1.2 Model building and selection I used hierarchical Cox Proportional Hazards models clustered by hospital using sandwich covariate estimates to test the relationship between access to care and mortality, censoring at thirty days. Cox models are preferred to logistic models because they do not ignore the effects of censoring and time to event on the relationship between covariates and the outcome.231 Time to event matters in so much as the effects of the covariates can change over time. Therefore a static measure of effect captured at thirty days may not appropriately adjust for this variation over time.232 Although logistic regression can, to some extent, incorporate the time element if events are grouped into intervals of occurrence, it remains inferior to the survival analysis technique 107  because it does not require that modelers correct for changes in effects of covariates across time.233 When corrections have been made, as is the requirement for proportional hazards models, the hazard ratio which evaluates risk instantaneously can be assumed to be the same across the time interval. The validity of hazards models (and thereby interpretation of constant risk over time) is predicated on two conditions: a) proportionality – the hazard ratio between two observations remains independent of time, and b) linearity – the model is only valid for time-independent covariates.231 I used both Schoenfeld and Martingale Residual plots within R to visually inspect the proportionality and linearity assumptions of my models, respectively. Where covariates did not meet the proportionality assumption, I created interactions with time to correct for the violation and where this failed, stratified by the covariate. In the case where continuous variables failed the linearity assumption, as demonstrated using Martingale Residual plots, I relied on the assess feature of the PROC PHREG function within SAS to determine what transformations would best approach linearity and in some cases converted continuous covariates into categorical ones with cut-offs driven by the data. Once the process was determined to be a success, data were clustered by hospital in PROC PHREG and a time variant covariate, in hospital, was added to adjust for any discharge bias. This variable accounts for the fact that patients were not necessarily in hospital until censoring and that the hazard of mortality should be tied to hospitalization. If this is not the case, then there is indication of systematic premature discharge to the detriment of patients’ health. Once the proportionality and linearity conditions were satisfied, I produced results for univariate effects of access (time) to care, appropriateness of care, and quality of care on patient mortality. I worked with hazards ratios developed from maximum likelihood estimates and ensured my sandwich covariate estimates for the clustering effect by hospital were always appropriate (p-value<0.05). Additionally, I tested for interactions between covariates and relied on AIC output to validate their inclusion in the final models. In interpreting the results, the reference group for time to care is patients who traveled 0 to 30 mins to their first admission. For appropriate care, the reference group is comprised of patients who did not receive appropriate care and for poor quality of care, the reference group is patients who received care at a facility where the observed mortalities were within the 95% confidence interval (CI) of expected mortalities based on the risk standardized mortality ratio and funnel plot exercise from Chapter 6. 108  8.1.3 Sensitivity analysis For sensitivity analysis, I replaced my estimated time to care covariate with a true time to care measure provided within the BCEHS’s Ambulance data. I restricted my analyses to patients that were treated between April 1 2009 and March 31 2013 as this was the period the Ambulance data covered. All the variables captured in the initial model-building exercise were included in two subsequent analyses. The first used true time to care as determined from BCEHS ambulance data, and the second included the same sub-cohort (patients transported by paramedics captured in Ambulance data) but used the estimated time to care measure. This allowed a direct comparison of outcomes using a time estimate rather than true travel burden. To be clear, although the Ambulance dataset included chute, dispatch, and en route times, I restricted my analysis to transport time as this was the only time I could estimate for my overall study using ArcGIS as described in Chapter 4. 8.2 Results 8.2.1 Complete study period (1999-2013) Figure 8-1 is a set of Kaplan Meier curves of probability of survival for the first 30 days following an AMI event (top panel probability extends to 0%). While differences in time to care do not seem to predict variations in patient outcome, patients without a distance to care estimate (no distance data) appear to do a lot worse than their complete data counterparts. Similar patterns of survival probability were found for the other two conditions (stroke and trauma – figures not shown). 109    Figure 8-1. Probability of survival by time to care, AMI    110  Table 8-1 reports the extent to which observations have missing distance values and the reasons for the missing data. The majority of observations with missing time to care estimates are events captured in the DAD that were either first admitted to a facility but the hospital code was missing or the first admitted facility was not a hospital (hospital identifier dependent). At the same time, missing values owing to a missing or unfamiliar postal code as outlined impact my analysis irrespective of the method of assignment used (intention or non-intention to treat).    AMI Stroke Trauma Total number of observations 96,672 94,863 220,151  Observations with time to care estimates, n (%) 79,704 (82.4) 83,872 (88.4) 179,702 (81.6)   Mean time to hospital of admission, mins (SD) 48.88 (139.27) 41.25 (128.29) 61.55 (161.08)   Mean time to closest hospital, mins (SD) 8.18 (11.22) 7.60 (11.34) 8.18 (12.44)   Mortality rate, deaths/100 events 14.3 15.3 2.4  Observations without time to care estimates, n (%) 16,968 (17.6) 10,991 (11.6) 40,449 (18.4)   1. Not admitted, n (%) 1985 (11.7) 694 (6.3) 751 (1.9)   Mean time to closest hospital, mins (SD) 7.98 (10.74) 8.03 (11.65) 8.13 (12.01)   Mortality rate, deaths/100 events 100 100 100   2. Admitted to unknown hospital, n (%) 6,320 (37.2) 1668 (15.2) 15,365 (38.0)   Mean time to closest hospital, mins (SD) 8.03 (11.46) 7.32 (12.53) 7.22 (10.80)   Mortality rate, deaths/100 events 5.5 6.5 4.4   3. Admitted to facility other than hospital, n (%) 3,042 (17.9) 5,371 (48.9) 6,167 (15.2)   Mean time to closest hospital, mins (SD) 7.95 (10.31) 7.67 (11.03) 7.32 (10.47)   Mortality rate, deaths/100 events 2.9 4.3 1.4   4. Observation missing postal code, n (%) 195 (1.1) 116 (1.1) 767 (1.9)   Mortality rate, deaths/100 events 8.4 9.7 2.5   5. Observation has unfamiliar postal code, n (%) 5,426 (32.0) 3,142 (28.6) 17,399 (43.0)   Mortality rate, deaths/100 events 23.3 29.0 11.1 Table 8-1. Extent of missing estimated distance and reasons for it, per condition n=number of observations, SD=standard deviation, mins=minutes 111  Interpreting the table, there are 16,968 records that lack a time to care estimate in the AMI cohort. Of these, 1985 (11.7%) are missing because the patient was never admitted to hospital. These records are excluded from further analyses because values for access, appropriateness and quality of care cannot be calculated. Conversely, patient observations without a time to care estimate but successfully admitted to hospital (n=14,983) were retained in the proportional hazards exercise. Access was designated as “Unknown time to care”. I included the 30-day mortality rate, and means and standard deviations of time to closest facility in Table 8-1. Examining the results for differences in time to closest minimally capable hospital, what minor differences exist are clinically negligible. Without knowing anything further about why patients were not admitted to hospital, it does not appear that access to care was the primary driver of a lack of admission. Examining the patients with missing or unfamiliar postal codes suggests there may be something inherently different in these observations. Table 8-2 to Table 8-4 show how patients with time to care estimates differ from those without. Patients without time to care estimates are more likely to be female, poor, old, frail, transferred less frequently, from rural communities, and have longer hospital stays. By including observations that fall under missing data reasons 2-5, I hoped to reduce the potential selection bias introduced if I were to exclude all patients lacking time to care estimates. Additionally, since my ambulance data provides true time to care irrespective of whether patients’ postal code or hospital of admission were properly recorded, this should help correct missing data effects in my sensitivity analysis.  112  Patient Demographics With time to care estimates Without time to care estimates P-value Total 79,704 16,968  Female, n (%) 25,186 (31.6) 7,289 (43.0) <0.0001 Income quintile, n (%) Lowest income quintile 2nd income quintile 3rd income quintile 4th income quintile Highest income quintile  18,252 (22.9) 16,818 (21.1) 15,463 (19.4) 14,187 (17.8) 13,390 (16.8)  4,178 (24.6) 3,560 (21.0) 2,985 (17.6) 2,834 (16.7) 2,666 (15.7)  <0.0001 Charlson comorbidity index, n (%) None (CCI=0) Low (CCI=1-2) Moderate (CCI=3-4) High (CCI=5+)  53,880 (67.6) 20,245 (25.4) 4,941 (6.1) 638 (0.8)  9,741 (57.4) 5,263 (31.0) 1,671 (9.9) 292 (1.7)  <0.0001 Transferred, n (%) 22,237 (27.9) 3,612 (21.3) <0.0001 Rural, n (%) 9,405 (11.8) 2,730 (16.1) <0.0001 Age, years Mean (SD) Range  68.70 (13.58) 0-105  75.40 (13.27) 0-107  <0.0001 Length of stay, days Mean (SD) Range  8.27 (50.64) 1 - 4032  10.14 (37.89) 1 - 2475  <0.0001 Table 8-2. Demographic differences between patients with and without time to care estimates, AMI n=number of observations; CCI=Charlson comorbidity index; SD=standard deviation    113  Patient Demographics With time to care estimates Without time to care estimates P-value Total 83,872 10,991  Female, n (%) 40,007 (47.7) 6,476 (58.9) <0.0001 Income quintile, n (%) Lowest income quintile 2nd income quintile 3rd income quintile 4th income quintile Highest income quintile  19,542 (23.3) 17,362 (20.7) 16,187 (19.3) 14,426 (17.2) 14,426 (17.2)  2,804 (25.5) 2,362 (21.5) 1,950 (17.7) 1,692 (15.4) 1,703 (15.6)  <0.0001 Charlson comorbidity index, n (%) None (CCI=0) Low (CCI=1-2) Moderate (CCI=3-4) High (CCI=5+)  64,355 (67.8) 25,499 (26.9) 3,946 (4.2) 1,062 (1.1)  6,437 (58.6) 3,581 (32.6) 773 (7.0) 200 (1.8)  <0.0001 Transferred, n (%) 11,260 (11.9) 1,438 (13.1) <0.0001 Rural, n (%) 15,301 (16.1) 2,117 (19.3) <0.0001 Age, years Mean (SD) Range  72.59 (14.17) 0 – 105  79.79 (12.32) 0 - 107  <0.0001 Length of stay, days Mean (SD) Range  15.98 (41.43) 1 – 4124  23.18 (33.59) 1 - 980  <0.0001 Table 8-3. Demographic differences between patients with and without time to care estimates, stroke n=number of observations; CCI=Charlson comorbidity index; SD=standard deviation     114  Patient Demographics With time to care estimates Without time to care estimates P-value Total 179,702 40,449  Female, n (%) 84,100 (46.8) 25,811 (63.8) <0.0001 Income quintile, n (%) Lowest income quintile 2nd income quintile 3rd income quintile 4th income quintile Highest income quintile  42,410 (23.6) 36,120 (20.1) 33,245 (18.5) 31,628 (17.6) 30,549 (17.0)  10,294 (25.5) 7,952 (19.7) 7,115 (17.6) 6,549 (16.2) 6,415 (15.9)  <0.0001 Charlson comorbidity index, n (%) None (CCI=0) Low (CCI=1-2) Moderate (CCI=3-4) High (CCI=5+)  15,7599 (87.7) 18,330 (10.2) 2,696 (1.5) 899 (0.5)  27,493 (68.0) 10,399 (25.7) 1,958 (4.8) 599 (1.5)  <0.0001 Transferred, n (%) 31,987 (17.8) 6,435 (15.9) <0.0001 Rural, n (%) 28,393 (15.8) 7,366 (18.2) <0.0001 Age, years Mean (SD) Range  54.37 (24.71) 0 – 107  74.73 (21.21) 0 – 112  <0.0001 Length of stay, days Mean (SD) Range  12.05 (44.00) 1 – 4834  17.87 (32.22) 1 – 2665  <0.0001 Table 8-4. Demographic differences between patients with and without time to care, trauma n=number of observations; CCI=Charlson comorbidity index; SD=standard deviation  Figure 8-2 to Figure 8-4 show the survival curves for each condition when plotting is restricted to patients with time to care values. Aside from the trauma cohort, it is difficult to establish any relationship between time to care and patient survival. For all three cohorts, it is premature to assume a relationship between time to care and survival outcome exists based on crude estimates alone. 115   Figure 8-2. Probability of survival by time to facility, AMI   Figure 8-3. Probability of survival by time to facility, stroke 116   Figure 8-4. Probability of survival by time to facility, trauma   Model  Parameter N=94,687 Hazard Ratio 95% CI P-value Wald statistic 1 Ref: Time to care < 30 mins 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240 Time to care (mins) ≥ 240 Unknown time to care  1.004 0.958 0.864 0.865 1.640  (0.588, 1.712) (0.861, 1.065) (0.678, 1.100) (0.780, 0.959) (1.277, 2.107)  0.9897 0.4251 0.2345 0.0059 <0.0001 251.80 2 Ref: Inappropriate care (error by commission) Appropriate care Inappropriate care (error by omission)  0.396 2.174  (0.350, 0.448) (2.019, 2.340)  <0.0001 <0.0001 291.14 3 Ref: Quality of care not flagged as poor Poor quality care  1.156  (0.559, 2.388)  0.6364 51.23 Table 8-5. Results of univariate Cox Proportional Hazards models, AMI Ref=reference category; mins=minutes  Table 8-5 shows the crude hazards ratios for the AMI cohort do not suggest a relationship between time to care and mortality within 30 days except for patients admitted to hospital with unknown time to care whose hazard of mortality is clearly higher (HR=1.640, 95% CI: 1.277, 117  2.107) and for patients who live over four hours from the facility where they received treatment (HR=0.865, 95% CI: 0.780, 0.959). The protective estimate for patients who live more than four hours from care may be because patients were much closer to the facility at the time of the event (i.e. patients did not travel from home) or they did not use ground transport (i.e. received air transport with advanced life support). Patients treated appropriately were those who received mechanical revascularization and whose treatment was not contraindicated or against guideline recommendations. For comparison, the reference group is patients who received mechanical revascularization but whose treatment was contraindicated or against guidelines (error by commission). The last group, also defined as inappropriate care, is unique to the AMI group. It describes patients who received basic care, but not definitive as recommended by the guidelines (error by omission). The results show that receiving appropriate care is protective (crude HR=0.396, 95% CI: 0.350, 0.448) and inappropriate care by omission is deleterious (crude HR=2.174, 95% CI: 2.019, 2.340) relative to inappropriate care by commission. In other words, receiving revascularization treatment, even against guideline recommendations is preferable to not being revascularized at all. Lastly, receiving care at a facility with questionable quality is not associated with patient mortality risk (crude HR=1.156, 95% CI: 0.559, 2.388).  Model Parameter N=94,169 Hazard Ratio 95% CI P-value Wald statistic 1 Ref: Time to care <30 mins 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240 Time to care (mins) ≥ 240 Unknown time to care  0.873 0.906 0.952 0.800 0.973  (0.707, 1.079) (0.788, 1.043) (0.719, 1.258) (0.721, 0.887) (0.519, 1.823)  0.2084 0.1719 0.7284 <0.0001 0.9318 42.1909 2 Ref: Inappropriate care Appropriate care  0.227  (0.163, 0.318)  <0.0001 47.6805 3 Ref: Quality of care not flagged as poor Poor quality care  1.495  (1.276, 1.752)  <0.0001 114.4167 Table 8-6. Results of univariate Cox Proportional Hazards models, stroke Ref=reference category; mins=minutes  The results of univariate Cox Proportional Hazards models (Table 8-6) for stroke appear to lack a pattern for access to care. As with AMI, patients living over four hours from the hospital of 118  admission had lower hazards of mortality (crude HR=0.800, 95% CI: 0.721, 0.887). The crude estimates for a protective effect of appropriate care (crude HR=0.227, 95% CI: 0.163, 0.318), and negative effect of poor quality care (crude HR=1.495, 95% CI: 1.276, 1.752) are in the hypothesized directions. Model  Parameter N=219,400 Hazard Ratio 95% CI P-value Wald  statistic  1 Ref: Time to care <30 mins 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240 Time to care (mins) ≥ 240 Unknown time to care  0.849 0.755 0.646 0.639 2.770  (0.684, 1.055) (0.608, 0.939) (0.508, 0.821) (0.554, 0.738) (1.882, 4.079)  0.1401 0.0115 0.0004 <0.0001 <0.0001 1387.3582  2 Ref: Inappropriate care Appropriate care  0.975  (0.821, 1.157)  0.4753 1.2581  3 Ref: Quality of care not flagged as poor Poor quality care  1.084  (1.027, 1.113)  0.0126 16.5462  Table 8-7. Results of univariate Cox Proportional Hazards models, trauma Ref=reference category; mins=minutes  For trauma patients (see Table 8-7), the crude estimates suggest a protective effect for patients who are over an hour from the hospital of admission and an increased hazard of mortality for those with unknown time to care (HR=2.770, 95% CI: 1.882, 4.079). Both appropriateness and quality of care are not related to patient outcomes. Table 8-8 provides the results of the multivariate models finalized for the AMI cohort. Model 4 is a full model including both the results of system-specific and patient-level covariates. It includes time to care, appropriateness care, and quality of care, as well as the method of transportation (reference group: patients transported without paramedic involvement, also known as self-transport), and whether the patient was taken directly to definitive care (transferred), patient age, sex, and neighborhood income quintile, as well as comorbidity and length of stay. The model suggests that after controlling for the set of patient and system characteristics developed here using administrative data, time to care is significantly associated with mortality for AMI patients for patients who live 60-119 minutes away, since these patients have higher mortality risks (adjusted hazard ratio (aHR)=1.124, 95% CI: 1.002, 1.262). The parsimonious model shown in Appendix C also found that time to care impacts patients’ hazards of mortality, but found a stronger stepwise dose-response relationship between time to care and increased hazards of mortality. 119  Model 4 N=94,687 Hazard Ratio 95% CI P-value Wald statistic Ref: Time to care <30 mins 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240 Time to care (mins) ≥ 240 Unknown time to care  1.051 1.124 1.045 1.171 1.059  (0.945, 1.169) (1.002, 1.262) (0.894, 1.220) (1.000, 1.371) (0.947, 1.185)  0.3236 0.0470 0.5814 0.0507 0.3160 37.6316 Ref: Inappropriate care (error by commission) Appropriate care Inappropriate care (error by omission)  0.893 1.200  (0.776, 1.027) (1.074, 1.341)  0.1132 0.0013 26.8877 Ref: Quality of care not flagged as poor Poor quality  1.070  (0.945, 1.212)  0.2839 8.7942 Ref: Death occurred outside hospital In-hospital  762.597  (372.608, 1560.768)  <0.0001 134.6215 Ref: EMS ground transport EMS air transport EMS combination (air, ground, and/or water) Self-transport  0.993 0.784 0.957  (0.777, 1.270) (0.585, 1.051) (0.915, 1.002)  0.9565 0.1034 0.0595 165.4412 Ref: No inter-hospital transfer during episode Inter-hospital transfer  1.208  (1.142, 1.278)  <0.0001 39.9811 Ref: Age ≤ 45 45 < age (years) ≤ 75 Age > 77  1.670 2.380  (1.457, 1.913) (2.053, 2.759)  <0.0001 <0.0001 530.3594 Ref: Male Female  1.009  (0.982, 1.037)  0.5096 1.6294 Ref: No comorbidity defined (CCI: 0) Low comorbidity (CCI: 1-2) Moderate comorbidity (CCI: 3-4) High comorbidity (CCI: 5+)  1.141 1.204 1.289  (1.103, 1.180) (1.141, 1.270) (1.137, 1.461)  <0.0001 <0.0001 <0.0001 22.0403 Ref: Length of stay < 1 day Length of stay (1 day) Length of stay (2-7 days) Length of stay (8-14 days) Length of stay (15 days +)  0.343 0.079 0.021 0.001  (0.299, 0.392) (0.068, 0.093) (0.015, 0.030) (0.001, 0.001)  <0.0001 <0.0001 <0.0001 <0.0001 1270.4515 Ref: Lowest income quintile Income (2nd quintile) Income (3rd quintile) Income (4th quintile) Income (5th quintile) Income (unknown)  0.966 0.985 0.915 0.906 0.964  (0.930, 1.004) (0.940, 1.034) (0.872, 0.960) (0.868, 0.947) (0.866, 1.074)  0.0784 0.5473 0.0003 <0.0001 0.5092 8.6826 Ref: Year:1999 Year of event  1.021  (1.006, 1.035)  0.0049 336.1523 120  Table 8-8. Results of multivariate Cox Proportional Hazards models, AMI Models were stratified for: hospital peer group (teaching, large, medium or small community, or non-acute facility); hospital condition specific volume (low=1-120, moderate=121-600, high=601+ admissions per year); and location of hospital (rural or non-rural) Ref=reference category; mins=minutes; CCI=Charlson comorbidity index; CI=confidence interval  Among patients who received mechanical revascularization following thrombolytic therapy, the appropriateness of care (physician or facility expertise) did not impact patient outcomes (aHR=0.893, 95% CI: 0.776, 1.027). However, patients who were not revascularized had significantly higher hazards of mortality across 30 days (aHR=1.200, 95% CI: 1.074, 1.341). This aligns with the crude estimate that found appropriate care was associated with lower mortality. . Meanwhile, and again in line with the crude estimates, the quality of care does not impact patient mortality (aHR=1.070, 95% CI: 0.945, 1.212). Furthermore, self-transport is not associated with mortality (aHR=0.957, 95% CI: 0.915, 1.02) despite being protective in the parsimonious one shown in Appendix C (aHR=0.731, 95% CI: 0.684, 0.782) when compared with ground ambulance transport. This is likely the result of including hospitalization length of stay, which may be acting as a proxy for severity, as is suggested by other literature.234 Without a measure for event severity, it is unclear if patients who are self-transported are experiencing a milder heart attack, or that waiting for paramedics is associated with higher mortality. Inter-hospital transfer is found to be hazardous to survival (aHR=1.208, 95% CI: 1.142, 1.278) despite studies elsewhere showing its utility, and the parsimonious model also demonstrating favourable outcomes. Patient characteristics such as age and comorbidity affect survival, with advanced age and comorbidity increasing the hazards of mortality in the expected direction; but female sex has no effect on hazards (aHR=1.009, 95%CI: 0.982, 1.037). These findings are also replicated in the parsimonious model. Despite univariate analysis suggesting income is not associated with hazard of mortality (not shown here), the findings from the model show that patients in the 4th or higher income quintile are more likely to survive than patients in the lowest income quintile. Lastly, patient survival worsens with year of event. This implies patient outcomes have worsened with time even after controlling for quality and appropriateness of care as well as patient demographics, and may be related to redistribution efforts. 121  Importantly, three factors included in the analysis but stratified to ensure proportionality do not have their own hazards ratios in each table. These are: hospital peer group (teaching, large/medium/small community hospital, and non-acute facility), hospital volume by condition (0<low≤120, 120<medium≤600, 600<high), and hospital rurality. Stratification by these variables showed consistent results across all three conditions.  Model 4 N=94,169 Hazard Ratio 95% CI P-value Wald statistic Ref: Time to care <30 mins 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240 Time to care (mins) ≥ 240 Unknown time to care  0.993 1.038 1.075 1.068 1.118  (0.918, 1.074) (0.943, 1.143) (0.932, 1.240) (1.002, 1.147) (1.067, 1.173)  0.8572 0.4425 0.3209 0.0421 <0.0001 77.7005 Ref: Inappropriate care (error by commission) Appropriate care  0.966  (0.906, 1.085)  0.2978 16.6218 Ref: Quality of care not flagged as poor Poor quality  1.145  (0.993, 1.320)  0.0620 27.1244 Ref: Death occurred outside hospital In-hospital  2833.817  (505.130, 15897.940)  <0.0001 425.9573 Ref: EMS ground transport EMS air transport EMS combination (air, ground, and/or water) Self-transport  1.077 0.896 0.682  (0.932, 1.244) (0.794, 1.011) (0.635, 0.734)  0.3240 0.0752 <0.0001 400.2305 Ref: Did not receive telehealth Received telehealth  1.547  (1.087, 2.200)  0.0152 2.5892 Ref: No inter-hospital transfer during episode Inter-hospital transfer  1.187  (1.110, 1.278)  <0.0001 13.5624 Ref: Age ≤ 45 45 < age (years) ≤ 75 Age > 77  1.187 1.451  (1.094, 1.288) (1.338, 1.575)  <0.0001 <0.0001 390.7886 Ref: Male Female  1.019  (0.997, 1.040)  0.0894 2.0664 Ref: No comorbidity defined (CCI: 0) Low comorbidity (CCI: 1-2) Moderate comorbidity  (CCI: 3-4) High comorbidity (CCI: 5+)  0.983 1.045 1.239  (0.943, 1.024) (0.987, 1.107) (1.141, 1.345)  0.4015 0.1297 <0.0001 414.3075 Ref: Length of stay < 1 day Length of stay (1 day) Length of stay (2-7 days) Length of stay (8-14 days)  0.353 0.071 0.011  (0.306, 0.407) (0.043, 0.118) (0.004, 0.029)  <0.0001 <0.0001 <0.0001 768.1541 122  Model 4 N=94,169 Hazard Ratio 95% CI P-value Wald statistic Ref: Lowest income quintile Income (2nd quintile) Income (3rd quintile) Income (4th quintile) Income (5th quintile) Income (unknown)  0.992 0.975 0.994 0.970 1.003  (0.945, 1.041) (0.937, 1.014) (0.956, 1.034) (0.934, 1.007) (0.923, 1.089)  0.7453 0.2083 0.7716 0.1067 0.9512 6.2268 Ref: Year:1999 Year of event  1.547  (1.087, 2.200)  <0.0001 80.5885 Table 8-9. Results of multivariate Cox Proportional Hazards models, stroke Models were stratified for: hospital peer group (teaching, large, medium or small community, or non-acute facility); hospital condition specific volume (low=1-120, moderate=121-600, high=601+ admissions per year); and location of hospital (rural or non-rural) Ref=reference category; mins=minutes; CCI=Charlson comorbidity index; CI=confidence interval  Table 8-9 shows that for stroke events, time to care is not associated with increasing risk of mortality; and appropriate and quality care are also not significantly associated with adjusted hazards of mortality. Increasing age and comorbidity are deleterious to survival. Telehealth services are found to exacerbate hazards of mortality (aHR=1.547, 95% CI: 1.087, 2.200), as is year of event. Furthermore, self-transport remains protective for stroke patients (aHR=0.682, 95% CI: 0.635, 0.734). Compared with patients that did not receive telehealth services, patients who did were more likely to be male (53.7 vs. 50.0%), of middle income, with no comorbidities (74.7 vs. 66.0%), receive a transfer (45.1 vs. 12.0%), and be rural residents (16.7 vs. 14.8%). While it is not clear why telehealth recipients in general differed so much from patients that were not recipients, since telehealth is related to directing complex patients to comprehensive stroke centres, it logically follows that these patients are more likely to have been transferred at some point during their episode of hospitalization.  Model 4 N=219,400 Hazard Ratio 95% CI P-value Wald statistic Ref: Time to care <30 mins 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240 Time to care (mins) ≥ 240 Unknown time to care  0.974 0.997 0.975 1.091 1.357  (0.902, 1.052) (0.877, 1.133) (0.876, 1.087) (0.975, 1.221) (1.226, 1.501)  0.5012 0.9640 0.6524 0.1279 <0.0001 130.8522 Ref: Inappropriate care Appropriate care  1.030  (0.948, 1.120)  0.4832 1.2358 123  Model 4 N=219,400 Hazard Ratio 95% CI P-value Wald statistic Ref: Quality of care not flagged as poor Poor quality  1.134  (0.994, 1.293)  0.0614 0.5637 Ref: Death occurred outside hospital In-hospital  808.047  (263.153, 2481.217)  <0.0001 486.3286 Ref: EMS ground transport EMS air transport EMS combination (air, ground, and/or water) Self-transport  1.174 0.838 0.757  (0.777, 1.774) (0.671, 1.048) (0.715, 0.801)  0.4449 0.1219 <0.0001 486.3283 Ref: No inter-hospital transfer during episode Inter-hospital transfer  1.517  (1.345, 1.712)  <0.0001 5.5951 Ref: Age ≤ 45 45 < age (years) ≤ 75 Age > 77  1.660 3.267  (1.442, 1.910) (2.595, 4.112)  <0.0001 <0.0001 560.6703 Ref: Male Female  0.867  (0.830, 0.906)  <0.0001 150.6603 Ref: No comorbidity defined (CCI: 0) Low comorbidity (CCI: 1-2) Moderate comorbidity (CCI: 3-4) High comorbidity (CCI: 5+)  1.382 1.473 2.000  (1.269, 1.506) (1.298, 1.670) (1.691, 2.366)  <0.0001 <0.0001 <0.0001 621.8254 Ref: Length of stay < 1 day Length of stay (1 day) Length of stay (2-7 days) Length of stay (8-14 days)  0.418 0.057 0.016  (0.369, 0.474) (0.048, 0.068) (0.009, 0.028)  <0.0001 <0.0001 <0.0001 1761.0959 Ref: Lowest income quintile Income (2nd quintile) Income (3rd quintile) Income (4th quintile) Income (5th quintile) Income (unknown)  0.962 0.961 0.985 0.921 0.956  (0.914, 1.013) (0.906, 1.020) (0.932, 1.041) (0.875, 0.969) (0.875, 1.044)  0.1398 0.1885 0.5903 0.0016 0.3142 10.0388 Ref: Year:1999 Year of event  1.038  (1.018, 1.058)  <0.0001 57.5189 Table 8-10. Results of multivariate Cox Proportional Hazards models, trauma Models were stratified for: hospital peer group (teaching, large, medium or small community, or non-acute facility); hospital condition specific volume (low=1-120, moderate=121-600, high=601+ admissions per year); and location of hospital (rural or non-rural) Ref=reference category; mins=minutes; CCI=Charlson comorbidity index; CI=confidence interval  Table 8-10 shows that time to care is not statistically significantly associated with hazards of early mortality for trauma events after controlling for all other factors, except in cases for which time to care is unknown (aHR=1.357, 95% CI: 1.226, 1.501). Both appropriateness and quality of 124  care also show no association, although quality is only marginally insignificant (aHR=1.134, 95% CI: 0.994, 1.293). Method of transport remains significant in both the full and parsimonious models, with self-transport protective (aHR: 0.757, 95% CI: 0.715, 0.801) which shows a dose relationship with improved hazards of mortality the longer a patient survives. Again, both advanced age and comorbidity increase the risk of mortality while female sex is protective (adjusted HR: 0.867, 95% CI: 0.830, 0.906) and patients from the highest neighborhood income quintile have significantly reduced hazards of mortality (aHR=0.921, 95% CI: 0.875, 0.969). As with the AMI models, year of event is significant with each year increasing the hazard of mortality by 3.8% compared with events that occurred in 1999. Across the three conditions, the in-hospital covariate shows that discharge bias is not a likely problem within BC as hazard of mortality is significantly and likely clinically higher while the patient is in hospital. In other words, hazards of mortality are much higher while in-hospital than not, simply because patients are not discharged until it is safe to do so. Although the argument could be made that quality and in-hospital mortality are related, when I included interaction terms in all versions of the models, they were not significant and indeed raised the AIC scores. Although the models can be run without the inclusion of this covariate and still be significant for all other covariates, the inclusion of in-hospital lowers the AIC score enough that it is best to include the covariate as it helps the model better fit the data. 8.2.2 Restricted study period (2009-2013) For this section, I restricted my analysis to patients that were treated between April 1 2009 and March 31 2013. All the variables mentioned earlier were included with the exception that the time to care estimate was replaced with true time to care as determined from BCEHS ambulance data for comparison purposes (Model 5). True time to care uses travel time from point of pick up by paramedics (i.e. it does not include dispatch, en route or chute times). Despite all conditions studied being time sensitive and paramedic assistance strongly recommended, many patients did not arrive at the hospital via EMS. I created Table 8-11 in Table 8-1’s likeness to organize patients with ambulance data by reason for missing time to care estimates. However, Table 8-11 also provides true time to care and allows a comparison of those patients with complete data, with those with missing estimates. Table 8-11 shows that time to care using postal code to hospital origin-destination measures are grossly 125  overestimated when compared with true time traveled. The travel burden estimate was exaggerated by 51.2% for stroke patients and up to 79.6% for trauma patients (almost twice as long as true time to care) for patients who received paramedic assistance. This is most likely caused by the assumption that patients were transported from home when in reality events may have occurred at work, during leisure or while driving. Furthermore, true time to care was based on travel time recorded by paramedics. Ambulance drivers have the added benefit of exceeding speed limits and ignoring traffic rules (one way streets, traffic signals), further aiding in reduced time to care, as seen here. When compared with the time estimates for the whole study period (1999-2013), the time to care estimated for patients who experienced a medical emergency between April 1 2009 and March 31 2013 is considerably longer: AMI1999-2013=48.88 mins (SD=139.27) compared with AMI2009-13=68.91 mins (SD=169.34), p-value<0.0001; trauma1999-2013=61.55 mins (SD=161.08) versus trauma2009-13=66.30 mins (SD=170.67), p-value<0.0001; stroke1999-2013=41.25 mins (SD=128.29) and stroke2009-13=42.19 mins (SD=131.41), p-value<0.0001. This is likely related to shifts in transport patterns over time as services became centralized. 126     AMI Stroke Trauma Total number of observations of events that occurred between 2009 and 2013 40,285 27,831 73,032  Observations with true time to care, n (%) 31,382 (77.9) 22,404 (80.5) 57,184 (78.3)   Mean true time to hospital of admission, mins (SD) 42.07 (68.41) 28.00 (45.98) 36.91 (54.83)   Mean est. time to hospital of admission, mins (SD) 68.91 (169.34) 42.19 (131.41) 66.30 (170.67)   Mean est. time to closest hospital, mins (SD) 8.79 (12.20) 7.98 (11.63) 9.34 (15.54)   Mortality rate, deaths/100 events 11.2 13.0 2.4  Observations without time to care estimates, n (%) 8,432 (22.1) 5,423 (19.5) 15,848 (21.7)   Not admitted, n (%) 854 (10.1) 155 (2.9) 0 (0.0)   Mean time to closest hospital, mins (SD) 7.98 (10.74) 8.03 (11.65) 8.13 (12.01)   Mortality rate, deaths/100 events 100 100 100   Admitted to unknown hospital, n (%) 3,045 (36.1) 2,211 (40.8) 6,970 (44.0)   Mean true time to hospital of admission, mins (SD) 60.43 (83.96) 34.85(57.43) 47.55 (68.19)   Mortality rate, deaths/100 events 11.5 21.3 6.7   Admitted to facility other than hospital, n (%) 0 (0.0) 0 (0.0) 0 (0.0)   Observation missing postal code, n (%) 42 (0.5) 28 (0.5) 41 (0.2)   Mean true time to hospital of admission, mins (SD) 54.92 (91.81) 38.79 (60.35) 37.17 (51.19)   Mortality rate, deaths/100 events 2.4 7.1 2.4   Observation has unfamiliar postal code, n (%) 5,426 (32.0) 3,142 (28.6) 17,399 (43.0)   Mean true time to hospital of admission, mins (SD) 52.82 (78.03) 31.99 (52.97) 44.79 (65.01)   Mortality rate, deaths/100 events 23.3 29.0 11.1 Table 8-11. Provision of true time to care by reason for missing estimated time to care, per condition n=number of observations; SD=standard deviation; mins=minutes 127  Table 8-12 to Table 8-14 show the demographics of patients taken to hospital by paramedics versus those not. In the case of AMI patients (see Table 8-12), although patients are similar for most demographic features, patients transported by paramedics are older (mean age=70.08 years (SD=15.54) versus 67.47 years (SD=14.38)), frailer, have longer episodes of hospitalization (mean episode of hospitalization=10.58 days (SD=20.20) versus 2.91 days (SD=60.25)), are transferred less frequently for definitive care (probability of transfer=26.12 vs. 36.57%) and experience a lower 30-day mortality rate (11.21 versus 18.09%). These results help validate the assumptions of the patients who self-transport as described above in my analyses using complete study period data. Patient Demographics Transported by paramedics Self-transport P-value Total 18,104 14,076  Female, n (%) 6,513 (36.0) 4,435 (31.5) <0.0001 Income quintile, n (%) Lowest income quintile 2nd income quintile 3rd income quintile 4th income quintile Highest income quintile Missing  4,233 (23.5) 3,963 (22.0) 3,485 (19.3) 3,173 (17.6) 2,937 (16.3) 246 (1.4)  3,083 (22.0) 3,054 (21.8) 2,755 (19.6) 2,603 (18.6) 2,361 (16.8) 170 (1.2)  <0.0001 Charlson comorbidity index, n (%) None (CCI=0) Low (CCI=1-2) Moderate (CCI=3-4) High (CCI=5+)  11,059 (61.1) 5,446 (30.1) 1,415 (7.8) 183 (1.0)  9,926 (70.5) 3,382 (24.0) 680 (4.8) 88 (0.6)  <0.0001 Transferred, n (%) 4728 (26.1) 5148 (36.6) <0.0001 Rural, n (%) 3086 (17.1) 2470 (17.6) 0.2391 30-day mortality rate 2,029 (11.2) 2,546 (18.1) <0.0001 Age, years Mean (SD) Range  70.08 (15.54) 0-105  67.47 (14.38) 0-103  <0.0001 Length of stay, days Mean (SD) Range  10.58 (20.20) 0 – 274  2.91 (60.25) 0 – 888  0.0056 Driving time in mins, est (SD) Mean Range  65.94 (157.51) 0.09 – 2730.26  37.41 (120.45) 0.01 – 2924.10  <0.0001 Table 8-12. Patient demographics by travel method 2009-2013, AMI n=number of observations; CCI=Charlson comorbidity index; SD=standard deviation 128  Table 8-13 demonstrates older (mean age=75.82 years (SD=14.31) versus 70.73 years SD=14.73)), frailer, and female patients are more likely to travel by ambulance for stroke events. Additionally, these patients have longer lengths of stay (19.20 (SD=26.45) days versus 17.87 (SD=26.48) days) and are more likely to die within 30 days (30-day mortality rate=18.34% versus 5.67%). Patient Demographics Transported by paramedics Self-transport P-value Total 16,150 8,360  Female, n (%) 8,329 (51.57) 3,670 (43.89) <0.0001 Income quintile, n (%) Lowest income quintile 2nd income quintile 3rd income quintile 4th income quintile Highest income quintile Missing  3835 (23.80) 3466 (21.51) 3114 (19.32) 2797 (17.36) 2722 (16.89) 182 (1.13)  1779 (21.36) 1782 (21.39) 1616 (19.40) 1517 (18.21) 1547 (18.57) 89 (1.07)  <0.0001 Charlson comorbidity index, n (%) None (CCI=0) Low (CCI=1-2) Moderate (CCI=3-4) High (CCI=5+)  9079 (56.22) 5921 (36.66) 927 (5.74) 223 (1.38)  5408 (64.67) 2621 (31.34) 254 (3.04) 79 (0.94)  <0.0001 Transferred, n (%) 910 (5.63) 306 (3.66) <0.0001 Rural, n (%) 2187 (13.54) 1303 (15.58) <0.0001 Received telehealth, n (%) 91 (0.56) 64 (0.77) 0.0491 30-day mortality rate 1,976 (4.2) 172 (1.0) <0.0001 Age, years Mean (SD) Range  75.82 (14.31) 0 – 104  70.73 (14.73) 0 – 105  <0.0001 Length of stay, days Mean (SD) Range  19.20 (26.45) 0 – 278  17.87 (26.48) 0 – 287  0.0002 Driving time in mins, est (SD) Mean Range  36.05 (111.35) 0.15 – 2342.65  34.03 (114.31) 0.12 – 2291.73  0.1822 Table 8-13. Patient demographics by travel method 2009-2013, stroke n=number of observations; CCI=Charlson comorbidity index; SD=standard deviation  129  Patient Demographics Transported by paramedics Self-transport P-value Total 47,039 16,958  Female, n (%) 25,741 (54.7) 8,061 (47.5) <0.0001 Income quintile, n (%) Lowest income quintile 2nd income quintile 3rd income quintile 4th income quintile Highest income quintile Missing  11,376 (24.4) 9,757 (20.9) 8,797 (18.8) 8,256 (17.7) 7,802 (16.7) 720 (1.6)  3,757 (22.4) 3,410 (20.3) 3,179 (18.9) 3,130 (18.6) 3,101 (18.5) 223 (1.3)  <0.0001 Charlson comorbidity index, n (%) None (CCI=0) Low (CCI=1-2) Moderate (CCI=3-4) High (CCI=5+)  37,455 (79.6) 7,939 (16.9) 1,250 (2.7) 85 (0.8)  15,070 (88.9) 1,597 (9.4) 206 (1.2) 85 (0.5)  <0.0001 Transferred, n (%) 1,163 (2.5) 313 (1.9) 0.0001 Rural, n (%) 6,981 (14.8) 2,439 (14.4) 0.1472 30-day mortality rate 1,976 (4.2) 172 (1.0) <0.0001 Age, years Mean (SD) Range  65.55 (23.70) 0 – 107  54.20 (24.20) 0 – 106  <0.0001 Length of stay, days Mean (SD) Range  19.44 (33.82) 0 – 174  17.64 (26.47) 0 – 204  0.0012 Driving time in mins, est (SD) Mean Range  0.09 (1.01) 0 – 104  0.06 (0.66) 0 – 34  <0.0001 Table 8-14. Patient demographics by travel method 2009-2013, trauma n=number of observations; CCI=Charlson comorbidity index; SD=standard deviation  The pattern is similar for trauma patients (see Table 8-14), although not as pronounced in some respects. Again, patients transported to hospital by paramedics are older (mean age=65.55 years (SD=23.71) versus 54.20 years (SD=24.20)), frailer, and have longer episodes of hospitalization (mean episode of hospitalization=19.44 days (SD=33.81) versus 17.64 days (SD=26.47)). Furthermore, patients transported by paramedics have higher 30-day mortality rate (4.20 versus 1.01%). I re-ran my Cox proportional hazards models as before comparing the results of the true time (using ambulance data) versus time to care (GIS estimate) covariates across all three conditions. For this sensitivity analysis, models were restricted to patients who had both true time and time to 130  care estimates (April 1 2009 to March 31 2013). When comparing the fit statistics between models in which analysis is restricted to patients with true time to care, the models using true travel time consistently fit the data better than those using estimated times (e.g. AIC for AMI models: 29687.81 vs. 29718.31). The results show that both models are informative, albeit using estimates produces slightly poorer fitting models. Given that there is currently no alternative for capturing time to care for patients that self-transport, estimated time to care will likely remain the only option when using administrative data. At the same time, this variable includes the assumption that all events occur at home, whereas true time to care does not have this problem. Thus, in reviewing the results of Table 8-15 to Table 8-17, the effect of time to care is inconsistent across models for AMI patients. In Model 5, which uses BCEHS travel time, travel beyond the first sixty minutes was found to be detrimental (aHR=1.520, 95% CI: 1.047, 2.206). This suggests that time to care really does matter, but minimally within the first hour of travel. Looking at model 5a, the time to care estimate does not approximate the access effect as well (although model fits are very similar). This is corroborated by looking at the other two conditions. Lastly, the models for AMI show that other covariates are more critical determinants of hazard of mortality than time to care (e.g. appropriateness, quality, and patient characteristics). Model 5 shows patients who were appropriately revascularized had better hazards of survival, although not significant at α=0.05 level (aHR=0.884, 95% CI: 0.771, 1.013) and appropriateness is protective although again not statistically significant (aHR=0.823, 95% CI: 0.635, 1.068). These results are the same as those found in model 4 (see Table 8-8). For stroke patients, appropriateness of care is again significant (using model 5 in Table 8-16) while poor quality of care is again marginally not significant, and this is consistent with the results seen across the overall cohort (using model 4 in Table 8-9). For trauma, it appears quality is more important than appropriateness of care (which here means whether or not a patient was directly admitted to a level I or II trauma centre). Patients treated at facilities flagged as performance outliers had higher hazards of mortality (aHR=1.243, 95% CI: 1.046, 1.476). Meanwhile, inter-hospital transfers repeated the findings observed earlier. Although literature suggests inter-hospital transfers are protective for AMI and stroke patients, these models and their sensitivity analyses show that when controlling for all other factors, inter-hospital transfers increase hazards of mortality across all three conditions. 131  As before, stroke patients that received telehealth services had higher hazards of mortality. This may have to do with delays in receiving overall care. Lastly, age and comorbidity also add to the hazard of early mortality, but female sex does not for AMI and stroke patients, and is protective once more for trauma patients. For trauma patients, access and appropriateness of care were not significant. However, as before, poor quality of care was associated with higher hazards of mortality (aHR=1.243, 95% CI: 1.046, 1.476 for model 5 or aHR=1.255, 95% CI: 1.066, 1.478 for model 5a).  132   Model 5 N=18,104 Model 5a N=18,104 Parameter Hazard Ratio 95% CI P-value Wald statistic Hazard Ratio 95% CI P-value Wald statistic Ref: Time to care <30 mins 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120  1.236 1.520  (0.824, 1.854) (1.047, 2.206)  0.3061 0.0277 0.5267  1.038 1.211  (0.859, 1.254) (0.991, 1.480)  0.7009 0.0618 0.4306 Ref: Inappropriate care (error by commission) Appropriate care Inappropriate care (error by omission)  0.884 1.314  (0.771, 1.013) (1.145, 1.507)  0.0759 <0.0001 0.6284  0.880 1.270  (0.771, 1.004) (1.108, 1.455)  0.0583 <0.0001 0.7023 Ref: Quality of care not flagged as poor Poor quality  1.116  (0.985, 1.264)  0.0838 0.0032  1.117  (0.976, 1.278)  0.1089 9.7367 Ref: Death occurred outside hospital In-hospital  504.214  (183.175, 1387.912)  <0.0001 756.4039  451.692  (341.393, 597.627)  <0.0001 749.2661 Ref: EMS ground transport EMS air transport EMS combination (air, ground, and/or water)  1.136 1.075  (0.851, 1.517) (0.802, 1.440)  0.3855 0.6303 6.6613  1.051 0.946  (0.826, 1.338) (0.713, 1.255)  0.6854 0.6993 6.5277 Ref: No inter-hospital transfer during episode Inter-hospital transfer  1.685  (1.456, 1.949)  <0.0001 0.5939  1.620  (1.404, 1.869)  <0.0001 0.5921 Ref: Age ≤ 45 45 < age (years) ≤ 75 Age > 77  1.554 2.188  (1.147, 2.107) (1.652, 2.899)  0.0045 <0.0001 152.1256  1.569 2.213  (1.163, 2.117) (1.671, 2.931)  0.0032 <0.0001 170.0677 Ref: Male Female  1.025  (0.977, 1.074)  0.3118 89.9997  1.021  (0.973, 1.072)  0.3887 81.5094 Ref: No comorbidity defined (CCI: 0) Low comorbidity (CCI: 1-2) Moderate comorbidity (CCI: 3-4) High comorbidity (CCI: 5+)  1.050 1.141 1.280  (0.980, 1.125) (1.028, 1.265) (1.017, 1.612)  0.1618 0.0127 0.0354 18.8028  1.044 1.128 1.292  (0.972, 1.121) (1.015, 1.253) (1.023, 1.632)  0.2361 0.0256 0.0314 19.5491 Ref: Length of stay < 1 day Length of stay (1 day) Length of stay (2-7 days) Length of stay (8-14 days)  0.330 0.084 0.026  (0.263, 0.414) (0.066, 0.106) (0.015, 0.044)  <0.0001 <0.0001 <0.0001 1802.8174  0.326 0.083 0.026  (0.258, 0.411) (0.065, 0.105) (0.015, 0.044)  <0.0001 <0.0001 <0.0001 1799.8457 133  Table 8-15. Results of multivariate Cox Proportional Hazards models using ambulance data (5) and estimated time to care (5a), AMI Models were stratified for: hospital peer group (teaching, large, medium or small community, or non-acute facility); hospital condition specific volume (low=1-120, moderate=121-600, high=601+ admissions per year); and location of hospital (rural or non-rural). Ref=reference category; mins=minutes; CCI=Charlson comorbidity index; CI=confidence interval  Ref: Lowest income quintile Income (2nd quintile) Income (3rd quintile) Income (4th quintile) Income (5th quintile) Income (unknown)  0.999 1.020 0.926 0.920 1.003  (0.926, 1.077) (0.919, 1.130) (0.843, 1.018) (0.824, 1.026) (0.789, 1.276)  0.9694 0.7137 0.1106 0.1337 0.9807 1.1913  0.994 1.025 0.924 0.914 0.950  (0.924, 1.069) (0.924, 1.137) (0.843, 1.013) (0.822, 1.016) (0.750, 1.203)  0.8624 0.6396 0.0933 0.0949 0.6694 1.0535 Ref: Year:2009 Year of event  0.991  (0.956, 1.027)  0.6163 0.0213  0.995  (0.961, 1.030)  0.7639 0.0333 134   Model 5 N=16,150 Model 5a N=16,150 Parameter Hazard Ratio 95% CI P-value Wald statistic Hazard Ratio 95% CI P-value Wald statistic Ref: Time to care <30 mins 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120  0.860 0.951  (0.649, 1.140) (0.768, 1.178)  0.2940 0.6545 3.7702  1.105 1.107  (0.983, 1.242) (0.981, 1.249)  0.0932 0.0988 15.2516 Ref: Inappropriate care (error by commission) Appropriate care  0.872  (0.775, 0.982)  0.0232 0.4230  0.904  (0.800, 1.021)  0.1047 0.6483 Ref: Quality of care not flagged as poor Poor quality  1.139  (0.982, 1.321)  0.0845 12.0449  1.131  (0.967, 1.323)  0.1227 10.3394 Ref: Death occurred outside hospital In-hospital  836.946  (143.657, 4876.066)  <0.0001 42.7741  857.610  (148.211, 4962.91)  <0.0001 42.2282 Ref: EMS ground transport EMS air transport EMS combination (air, ground, and/or water)  1.363 0.898  (1.142, 1.627) (0.763, 1.055)  0.0006 0.1906 1.9107  1.175 0.808  (0.991, 1.393) (0.671, 0.974)  0.0639 0.0253 2.2598 Ref: No inter-hospital transfer during episode Inter-hospital transfer  1.522  (1.338, 1.731)  <0.0001 0.4421  1.488  (1.289, 1.719)  0.0039 0.2263 Ref: Did not receive telehealth Received telehealth  1.398  (1.094, 1.787)  0.0074 0.1117  1.414  (1.122, 1.780)  0.0033 0.1027 Ref: Age ≤ 45 45 < age (years) ≤ 75 Age > 77  1.403 1.734  (1.207, 1.632) (1.429, 2.104)  <0.0001 <0.0001 4.5302  1.414 1.722  (1.233, 1.622) (1.456, 2.037)  <0.0001 <0.0001 4.5004 Ref: Male Female  1.065  (1.017, 1.115)  0.0074 129.8792  1.051  (1.005, 1.098)  0.0291 132.7806 Ref: No comorbidity defined (CCI: 0) Low comorbidity (CCI: 1-2) Moderate comorbidity (CCI: 3-4) High comorbidity (CCI: 5+)  0.937 1.091 1.540  (0.882, 0.996) (0.944, 1.261) (1.237, 1.916)  0.0372 0.2372 0.0001 40.5421  0.921 1.061 1.476  (0.871, 0.974) (0.920, 1.223) (1.187, 1.834)  0.0039 0.4153 0.0005 41.0149 Ref: Length of stay < 1 day Length of stay (1 day) Length of stay (2-7 days) Length of stay (8-14 days)  0.281 0.075 0.015  (0.240, 0.328) (0.045, 0.127) (0.006, 0.042)  <0.0001 <0.0001 <0.0001 955.0392  0.282 0.075 0.015  (0.240, 0.332) (0.045, 0.127) (0.006, 0.043)  <0.0001 <0.0001 <0.0001 936.5391 135  Table 8-16. Results of multivariate Cox Proportional Hazards models using ambulance data (5) and estimated time to care (5a), stroke Models were stratified for: hospital peer group (teaching, large, medium or small community, or non-acute facility); hospital condition specific volume (low=1-120, moderate=121-600, high=601+ admissions per year); and location of hospital (rural or non-rural. Ref=reference category; mins=minutes; CCI=Charlson comorbidity index; CI=confidence interval  Ref: Lowest income quintile Income (2nd quintile) Income (3rd quintile) Income (4th quintile) Income (5th quintile) Income (unknown)  0.969 0.953 0.987 0.957 1.267  (0.870, 1.080) (0.870 1.044) (0.896, 1.088) (0.864, 1.059) (0.939, 1.709)  0.5746 0.2986 0.7978 0.3950 0.1215 0.0321  0.966 0.977 1.003 0.977 1.226  (0.866, 1.076) (0.898, 1.064) (0.910, 1.105) (0.887, 1.077) (0.909, 1.653)  0.5287 0.5986 0.6442 0.1814 0.1924 2.1872 Ref: Year: 2009 Year of event  0.998  (0.964, 1.033)  0.9121 1.0369  0.998  (0.963, 1.034)  0.8990 0.9270 136   Model 5 N=47,039 Model 5a N=47,039 Parameter Hazard Ratio 95% CI P-value Wald statistic Hazard Ratio 95% CI P-value Wald statistic Ref: Time to care <30 mins 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240  1.020 1.180 1.016  (0.703, 1.478) (0.887, 1.570) (0.795, 1.299)  0.9181 0.2558 0.8967 0.2098  1.097 1.130 1.159  (0.937, 1.284) (0.884, 1.444) (0.960, 1.400)  0.2490 0.3301 0.1240 0.0311 Ref: Inappropriate care (error by commission) Appropriate care  0.957  (0.867, 1.057)  0.3871 0.0076  1.063  (0.957, 1.181)  0.2514 0.0091 Ref: Quality of care not flagged as poor Poor quality  1.243  (1.046, 1.476)  0.0134 1.2864  1.255  (1.066, 1.478)  0.0064 4.3687 Ref: Death occurred outside hospital In-hospital  509.561  (133.268, 1948.351)  <0.0001 13.2803  516.105  (136.989, 1944.423)  <0.0001 13.2691 Ref: EMS ground transport EMS air transport EMS combination (air, ground, and/or water)  1.026 0.989  (0.599, 1.758) (0.701, 1.394)  0.9247 0.9482 0.0747  0.960 0.897  (0.555, 1.660) (0.637, 1.264)  0.8841 0.5341 0.0756 Ref: No inter-hospital transfer during episode Inter-hospital transfer  1.941  (1.669, 2.257)  <0.0001 1.5716  1.971  (1.707, 2.276)  <0.0001 1.3353 Ref: Age ≤ 45 45 < age (years) ≤ 75 Age > 77  1.432 2.713  (1.169, 1.755) (2.036, 3.615)  <0.0001 <0.0001 5.2661  2.590 1.971  (1.294, 1.616) (1.977, 3.394)  0.0007 <0.0001 5.2017 Ref: Male Female  0.922  (0.870, 0.977)  0.0061 4.2759  1.446  (0.850, 0.964)  0.0019 4.3942 Ref: No comorbidity defined (CCI: 0) Low comorbidity (CCI: 1-2) Moderate comorbidity (CCI: 3-4) High comorbidity (CCI: 5+)  1.305 1.402 2.348  (1.125, 1.514) (1.120, 1.756) (1.830, 3.011)  0.0004 0.0032 0.0001 20.7551  1.285 1.338 2.350  (1.111, 1.487) (1.068, 1.676) (1.873, 2.948)  0.0008 0.0113 <0.0001 19.9905 Ref: Length of stay < 1 day Length of stay (1 day) Length of stay (2-7 days) Length of stay (8-14 days)  0.096 0.006 0.002  (0.059, 0.159) (0.002, 0.015) (0.000, 0.008)  <0.0001 <0.0001 <0.0001 314.9969  0.099 0.006 0.002  (0.060, 0.165) (0.002, 0.016) (0.000, 0.008)  <0.0001 <0.0001 <0.0001 326.1280 137   Model 5 N=47,039 Model 5a N=47,039 Parameter Hazard Ratio 95% CI P-value Wald statistic Hazard Ratio 95% CI P-value Wald statistic Ref: Lowest income quintile Income (2nd quintile) Income (3rd quintile) Income (4th quintile) Income (5th quintile) Income (unknown)  0.906 0.897 1.009 0.874 1.009  (0.815, 1.009) (0.814, 0.987) (0.926, 1.100) (0.779, 0.980) (0.772, 1.318)  0.0714 0.0266 0.8356 0.0213 0.9489 0.1649  0.894 0.890 0.999 0.883 1.000  (0.807, 0.990) (0.811, 0.978) (0.915, 1.090) (0.787, 0.991) (0.778, 1.285)  0.0306 0.0150 0.9779 0.0347 0.9998 0.1607 Ref: Year: 2009 Year of event  1.000  (0.970, 1.031)  0.9964 0.0811  0.998  (0.966, 1.032)  0.9214 0.0791 Table 8-17. Results of multivariate Cox Proportional Hazards models using ambulance data (5) and estimated time to care (5a), trauma Models were stratified for: hospital peer group (teaching, large, medium or small community, or non-acute facility); hospital condition specific volume (low=1-120, moderate=121-600, high=601+ admissions per year); and location of hospital (rural or non-rural. Ref=reference category; mins=minutes; CCI=Charlson comorbidity index; CI=confidence interval  138  8.3 Summary The Cox proportional hazards models used in this exercise incorporated the three pillars of health service delivery discussed throughout this dissertation. By controlling for patient characteristics, I attempted to create a unifying explanation of how differences in access, quality or appropriateness can compensate for one another and effectively treat patients. When the entire cohort and estimated travel burden were used, time to care, appropriateness and quality of care had different effects across the three conditions but with generally consistent findings. The models showed that travel longer than 30 minutes to one hour significantly reduced patients’ likelihood of survival, even after adjusting for quality and appropriateness. Some models demonstrated a dose-response relationship which further strengthened the general findings of the models. Furthermore, the mode of transportation relative to ground paramedic transport made a difference. According to all models, self-transport was found to be protective while air transport was found to increase the hazards of mortality. Without a true measure for severity in the data it is difficult to tell if self-transport was associated with less urgent emergencies or healthier patients and thus overall survival was expected to be more favourable than patients transported by paramedics. However, Table 8-12 to Table 8-14 demonstrate that patients who used EMS for transport were generally older, frailer and stayed in hospital longer following the procedure, so this suggests that at least general health may be at play and the effects seen in these models is imperfectly controlled confounding by indication. Careful interpretation of the general findings of models 4, 5, and 5a suggest paramedics play an integral role in triaging and providing early care that distinguishes the models and explains why appropriateness, quality, travel burden and inter-hospital transfers do not always match the findings in the models which include patients who self-transport. These findings may indicate that EMS involvement in the care continuum is critical in compensating for differences in access and by taking patients where care can be best delivered (considering access, appropriateness and quality all at once). The findings of both Chapters 7 and 8 will be discussed in greater detail in the following chapter and set in context both locally (within Canada) and internationally. 139  Chapter 9: Discussion 9.1 Summary of results In 2002, BC’s Ministry of Health began redistributing hospital services, effectively closing or converting some hospitals into specialty clinics. Throughout the process, local communities and clinicians have been concerned that the changes in access to emergent services will negatively impact patients’ health. Meanwhile, proponents of the closures have supported the changes with evidence that small and/or rural hospitals provide suboptimal care. Despite these competing perspectives, there has yet to be a thorough examination of the effects of the hospital closures in the province and evidence of the effects of the centralization efforts remain contested. This dissertation aimed to understand the impact the redistribution of hospital services has had on BC residents’ health outcomes following medical emergencies. Using a retrospective cohort study design, I selected all BC patients admitted to hospitals or who died from acute myocardial infarction, trauma or ischemic stroke events between fiscal years 1999 and 2013. I developed a theoretical framework that incorporated themes of access, appropriateness and quality of care and constructed my covariates from data available within the Ministry of Health’s Discharge Abstract Database and consolidation file, the Medical Services Plan Payment Information File, Vital Statistics’ Mortality Data, the British Columbia Emergency Health Services’ Ambulance file, and publicly accessible data such as the Postal Code Conversion File. I created three variations for access of care: ground travel distance in kilometers, aerial distance (as the crow flies, Euclidean) in kilometers, and ground travel time in minutes. Ground travel distance and time estimates were origin-destination cost matrix calculations from the network analysis toolkit in ArcGIS 10.2 while aerial distance was calculated using Euclidean linear arc estimates in SAS 9.2. Based on the recommendations in current literature and the results of my own modeling exercises, I ultimately worked with ground travel time as my proxy for travel burden. For appropriateness of care, I used British Columbian or where unavailable, Canadian standards of care, published for each condition. Since my administrative data lacked time stamps for each procedure but most of the standards of care are developed around the timing of treatments (e.g. recommended door-to-needle time) appropriateness in this dissertation was whether or not the patient was taken to a facility capable of treating him or her and whether or not the treatment 140  received was contraindicated given the patient’s characteristics (e.g. comorbidities and age). The data lack the detail needed to assess whether patients received timely care, most importantly any information on the time of initial presentation at the facility. Time of admission can sometimes be captured, but admission and initial presentation are not necessarily close to the same, and the difference between the two will vary across individuals.  Lastly, for quality of care I used funnel plots to graph the risk standardized mortality ratio of each hospital relative to an average BC hospital given patient and hospital characteristics per condition. If the observed mortalities for patients admitted to the facility exceeded the 95% CI of the expected number of risk standardized mortalities for that hospital for a discrete time interval, that hospital was flagged as offering poor quality (also described as a poor performer) for the period of time. Once developed, I used these covariates along with others readily determined (e.g. comorbidity using the revised Charlson comorbidity index) from the administrative data to test the following hypotheses: 1. Redistribution has had no measurable effect on patient outcomes a. Redistribution has had no measurable effect on patients’ access to services. 2. Access to care has no measureable effect on patient outcomes after controlling for quality, and appropriateness of acute care services and compensating mechanisms such as inter-hospital transfers and telehealth services. For the first two hypotheses, I used controlled quasi-experimental study designs and interrupted time series analyses to investigate the impact local hospital condition-specific treatment cessation had on access to care. I tested the last hypothesis by conducting a series of univariate and multivariate mixed effects survival analyses. 9.1.1 Hospitals which ceased treatment To prepare for the analyses I identified which hospitals stopped treating patients per given condition and when (calendar quarter) this occurred. I classified facilities within BC as hospitals if they offered some level of emergency services (at minimum 24/7 access to a physician and necessary diagnostic equipment e.g. ECG). In doing so, 101 facilities, including the Abbotsford Regional Hospital and Cancer Centre which opened in 2008, were identified as hospitals available at some point over the duration of the study period. The criteria captured centres not classified as 141  hospitals by CIHI which identifies hospitals (acute care centres) based on whether or not the inpatient institution is mandated by the provincial ministry of health to submit to the DAD. However, this does not guarantee that the centre will oblige as some acute care centres do not always provide data. At the same time other facilities not recognized by CIHI consistently do, and appear within the DAD all the same.234 Between 1999 and 2012, 26 non-hospital facilities and 16 hospitals ceased admitting patients. In other words, at the beginning of the study period, patients were admitted to one of 121 facilities (some of which were inappropriate) within the discharge abstract database. By 2013, patients were only treated at one of 79 facilities, all of which provided a minimum threshold of care needed for initial treatment of AMIs. Stroke patients in 1999 were taken to one of 120 facilities but by 2013 were only admitted to 79 hospitals. Between 1999 and 2013, 16 hospitals and 25 non-hospital facilities stopped admitting them as patients. Similarly for trauma patients, in 1999 there were 122 facilities that treated patients: 80 hospitals which remained open, 14 hospitals which closed during the study period, and 28 non-hospital facilities that stopped accepting patients by 2013. Thus, over the duration of the study period 804 (0.86%) AMI, 1230 (1.59%) stroke, and 4989 (2.59%) trauma events were taken to non-acute care facilities. These results can be interpreted in two ways. On the one hand, redistribution appears to have led to the consolidation of services, as expected. By the end of the study period, all patients were taken to facilities capable of diagnosing and providing preliminary care. This was achieved by the closure or down-grading of some facilities that should not have been treating patients according to accepted clinical guidelines. On the other hand, since these facilities had treated patients, their ceasing to do so could be interpreted as a loss of services. Without a clear and ubiquitous definition of what constitutes a hospital, it is easy to understand why the public perceives redistribution as a much larger loss of access than it may in fact be. However, from this analysis, it appears that, especially before redistribution, a considerable number of facilities were inappropriately visited for emergency care needs. This misuse of local facilities declined with time either as a byproduct of shifts in patient transport patterns and the eventual closure of the facilities or because the centres refused to admit certain patients, which forced shifts in patient transport. 142  9.1.2 Appropriateness of care All three conditions used to investigate the impact of the hospital closures in BC are time sensitive. This means that the appropriateness of care variable ideally incorporates the timing of treatment as well as where the treatment was provided (e.g. a facility capable of managing subsequent complications or able to transfer the patient for definitive care) and the type of treatment used (e.g. primary PCI or thrombolytics) as outlined within the standards of care currently in place (see Chapter 5). However, given the limitations of using administrative data mentioned earlier, my appropriateness of care variable captured whether or not a treatment provided was contraindicated based on the patient’s health, age, and condition; as well as whether the facility was appropriately equipped and if the attending physician and staff had the necessary experience to perform invasive procedures. Despite the variable’s limited capacity to capture appropriateness of care in its entirety, the results of Chapter 5 shed important light on patterns of care for time-sensitive conditions. For example, just 34.5% of trauma patients were directly transported to a Level I or II trauma centre while an additional 1.6% were transferred there for definitive care (2.6% of patients initially taken to an inappropriate facility). Given that the trauma cohort was restricted to patients with the most severe conditions, there are several possible explanations for the finding: (a) the majority of patients inappropriately transferred to inadequately equipped hospitals are inappropriately triaged, (b) transportation methods heavily influence where a patient is first admitted or (c) the selection criteria outlined in Chapter 2 do not restrict observations to the most severe patients. Chapter 8’s Table 8-14 showed that 73.5% of all trauma patients between 2009 and 2013 (i.e. those for whom there are ambulance data) were transported to hospital by paramedics (ground, air, or combination). Hence, explanation (b) is not a strong explanation for why the majority of patients were not treated at Level I or II centres. Meanwhile patients treated at a trauma centre spent 1.85 days (SD=63.05) in the intensive care unit compared with patients not taken to a Level I or II centre who spent a mean 1.14 days (SD=28.78), p-value<0.0001. This suggests that patients taken to Level I or II centres may indeed have more urgent and severe injuries and helps dispel concerns of inappropriate triage or explanation (a). To investigate explanation (c)’s likelihood, I consulted Dr. David Evans, the Provincial Medical Director of Trauma Services BC. He agreed that the inclusion criteria were fitting for the purposes 143  of this thesis but cautioned that the exclusion criteria may not be specific enough to prevent less urgent trauma incidents from diluting the analysis. Hence the criteria potentially lacked selection specificity. He acknowledged that the absence of severity indications in the Ambulance and Administrative data meant the true severity of the trauma incident beyond what was suggested by CIHI is unknown.236 Thus, without additional details on patients’ severity, it seems more likely that the patients in true need of urgent care were adequately triaged to Level I or II centres. Additionally, a higher proportion of injured males (56.9%) compared with females (43.0%) were triaged to a Level I or II trauma facility but were less likely to use EMS (47.57% versus 54.77%). This suggests one of two things: (a) males experienced more severe trauma thereby explaining the differences in Level I and II trauma centre admissions or b) the presence of triage selection bias. While studies have shown that men are more likely to engage in risky activities and experience severe trauma237 given the extent of self-transport, this alone does not explain triage patterns. Despite gender neutral triage guidelines, research shows that paramedics and physicians respond to patients’ needs differently based on patients’ gender. For example, a Canadian retrospective cohort study of 98,871 adult trauma patients with an injury severity score >15 or who died within 24 hours of hospitalization found that of severely injured women, 49.6% received care in a trauma centre compared with 63.2% of males.238 Similarly, in a retrospective study looking at EMS response to cardiac events in Dallas County, Texas, women were more likely to be seriously delayed (adjusted odds ratio 1.52, 95% CI 1.32 – 1.74 for >15 mins beyond median arrival time) than men.239 Suggested reasons for these gendered responses include perceived differences in injury severity, the assumed benefit from trauma centre care, or a potential underlying gender bias.240 Thus, the overall proportion of patients treated appropriately for trauma incidents may not be as inaccurate as initially assumed and indeed, some patients may not be getting the care they are supposed to according to trauma guidelines. On the other end of the condition spectrum, of all AMI events, only 13.7% received mechanical revascularization after tPA, 1.8% died without any admission, 84.5% were only treated with tPA, and only 56.3% were transported by paramedics over the entire study duration. Across all three conditions, EMS use was below ideal. A study using the National Registry of Myocardial Infarction 2 found that over the four-year period 1994-98 a similar proportion (53.4%) used EMS services across the United States241 while a more recent one found it closer to 49.8%242. 144  In Canada, a similar study (2010-2013) found it to be 60.1% overall (81.8% of patients taken to PCI centres compared with 49.2% treated elsewhere).243 Given the extent that time to care is stressed in the literature and the documented role of paramedics in the timely diagnosis and transportation of patients, my results indicate much needed behavioural changes in how the public recognizes and responds to AMI symptoms. In this study, females accounted for 34.9% of all events, 35.7% of patients treated appropriately and 36.0% of those transported via EMS. These results are consistent with other studies that find more women, older patients, and those with comorbidities are more likely to use EMS services for an AMI episode.244,245 Retrospective cohort studies conducted in Sweden and the USA found that women have better age-adjusted prognosis following a heart attack but are less likely to be prescribed appropriate care such as early medical treatments, acute reperfusion therapies, timely reperfusion (pharmacological or mechanical) and invasive procedures.246,247 This is consistent with findings in this thesis, both in appropriateness of care by demographics and in the overall results of the Cox proportional hazards models which found that despite differences in appropriateness of care women and men have similar survival outcomes. Explanations for differences in appropriateness of care include ongoing difficulties in accurately diagnosing women with STEMI which consequently impacts door-to-balloon time, and the delayed onset of coronary disease in women, in part due to the protective effect of estrogen, propagating the misperception that heart disease and AMIs are a male medical concern.248 As such, alternative etiologic reasons behind women’s symptoms are first explored before AMI is diagnosed, consequently delaying appropriate care.249 Unfortunately, since my appropriateness of care indicator does not reflect whether care was provided in a timely manner, I am unable to elaborate further on whether women in this study indeed received appropriate care or simply arrived at capable centres and were eventually treated without contraindication owing to their higher EMS usage. For stroke, the findings were similar. Of all events admitted to hospital, 34.6% were not transported using EMS services. This high proportion of self-transport is typical of other research studies, and the motivation behind many stroke response campaigns geared towards the public.250 Although sex was associated with EMS use in my study, this was not the case in others where age, 145  comorbidities and location of stroke were associated with EMS use but sex, race, and prior stroke were not.251 The results of the univariate Cox Proportional Hazards models in Chapter 8 did not show a consistent association between appropriateness of care and 30-day mortality. However, a strong protective effect was found for stroke patients, event after controlling for all other system and patient relative characteristics. The crude HR=0.227 (95% CI: 0.163, 0.318) and an adjusted HRstroke=0.436 (95% CI: 0.298, 0.637). 9.1.3 Quality of care In developing my theoretical framework, I came across studies that demonstrated hospitals that treated a higher volume of condition specific events enjoyed better patient outcomes—an  argument routinely used to justify small and rural hospital closures.252 I also read studies that showed quality and adherence to process measures contributed to improved outcomes. Hence, some of the analyses here seek to determine whether quality contributed to patient outcomes after adjusting for condition specific volume, patient demographics and access and appropriateness of care. The quality variable used here was derived from risk standardized mortality ratios and funnel plots to single out hospitals that had higher than expected mortality rates for admitted patients after accounting for patient and hospital characteristics. Thus, instead of using process or structure indicators these analyses relied on what mattered most, patient outcomes, to flag performance outliers. Unlike many studies, I found various hospital types (teaching to small community hospitals) to be performance outliers (see Table 6-2) when using a full model and only large community and teaching hospitals when using a parsimonious model (Table 6-3). These findings conflict with studies that demonstrate smaller facilities offer poorer quality healthcare.213 Such studies also typically laud teaching hospitals as the pinnacle of excellence in patient care and outcomes.254 However, there is at least one study showing that small facilities can provide similar care to large teaching centres, and that staff of these smaller centres know when to transfer patients to optimize outcomes, helping support the present findings.72 Without an indication for patient severity, it remains unclear whether large teaching hospitals experience higher than expected mortalities owing to poor quality of care, or simply as an artifact of unmeasured patient characteristics. 146  Additionally, Chapter 6 showed that the extent of patients treated at flagged hospitals declined over time. In other words, fewer patients were treated at poor performing facilities as time progressed, independent of local redistribution effects. The Cox Proportional Hazards models of Chapter 8 were inconclusive on the effect of poor quality on hazards of 30-day mortality. Univariate analysis of poor quality of care and 30-day mortality outcomes found a crude HRAMI=1.156 (95% CI=0.559, 2.388), HRtrauma=1.084 (95% CI: 1.027, 1.113) and crude HRstroke=1.495 (95% CI: 1.276, 1.752). Meanwhile the adjusted hazards were significant for trauma (4-5a) and showed poor quality increased hazards of mortality. In other words, only trauma outcomes are associated with quality of care after controlling for confounders according to my models. Although the lack of association between quality of care and hazards of 30-day mortality may at first appear surprising, a systematic review exploring the relationship between risk standardized or adjusted mortality rates to adherence to existing standards of clinical care did not find a strong association.255 The authors of the literature review provided three possible explanations for the lack of a consistent association between RSMR and adherence to clinical standards: (a) differences in patient prognosis not captured by comorbidity, (b) differences in definitions of conditions and treatment, and (c) genuine differences in care not captured through the process measures (e.g. staff vigilance in coding of care delivered). While explanation (b) does not seem relevant to this study, as all conditions were identified using standardized definitions, without detailed time stamps it is not possible to confirm whether or not process measures were properly adhered to, and thus I cannot definitively discuss what sensitivity or specificity shortcomings the RSMR may have in identifying poor performance. Similarly, my analysis strategy did not capture other process measure details (c). However, these results do indicate that regionalization did not contribute to changes in quality of care directly. Given that treatment received at a poor performing hospital did not contribute to elevated hazards ratios it is reasonable to understand why quality was not a criteria for hospital closures. Instead, the closures were likely the result of high operating costs and low treatment margins as is typical in other occurrences of hospital closures.256 9.1.4 Access to care Chapter 7 used segmented regression to determine whether communities that lost a facility experienced a higher 30-day mortality rate than communities that did not. The underlying 147  assumption was of course that given a local facility closure, the travel burden for patients in these communities increased. Notable differences in the trend or level of 30-day mortality following the intervention (loss of service) would indicate that changes in access did indeed influence patient outcomes. Across all three conditions there was no significant change in level or trend of 30-day mortality rates following the intervention (see Table 7-5) and the findings persisted when the outcome was extended to 90-day mortality. However, communities that lost access to care had a higher baseline 30-day mortality rate that remained throughout the analysis interval compared with those that experienced no change in access: AMI 7.3% (95% CI: 3.1, 11.6%), stroke 6.8% (95% CI: 2.7, 10.8%), trauma 1.86% (95% CI: 0.6, 3.1%). An overall higher mortality rate and no change following the intervention might be the result of systematically bypassing the closest facility for care at a centre better suited to treat the emergency. Bypassing the closest facility means an extended travel burden that could contribute to the higher mortality rate. Analyses of bypass rates showed higher rates in communities whose closest facility stopped admitting for the condition even prior to redistribution of services (see Table 7-7). Following redistribution, large drops in bypass rates were seen across all three conditions: AMI dropped 15.5% (95% CI: 3.5, 27.5%), stroke dropped 25.3% (95% CI: 11.7, 38.8), and trauma dropped 22.7% (95% CI: 15.7, 29.6%). Furthermore, the bypass rate following the intervention in communities that lost care approached that of communities that did not. Together these results help explain why local facility closures did not perturb 30-day mortality rates. They also mirror other reports of closures in Canada which attribute considerable underutilization of local services122 to their loss/closure. Specific to the results of the bypass rates in my analyses, a retrospective cohort study investigating rural bypass activity across seven American states between 1991 and 1996 found an overall bypass rate of 30%.257 The authors found that patients with managed care or commercial insurance were more likely to bypass. While these explanations are not likely drivers of bypass activity within BC, both severity and diagnosis-related groups were also found to be associated with bypass activity in the American study and may explain the varying degrees of bypass activity found here. Other factors associated with bypass activity in the literature and potentially relevant to my findings included age, income, satisfaction with local hospital, and travel burden.258 148  Thus, given the results of Chapter 7, it does not appear that the redistribution of hospital services negatively affected patients’ health for time sensitive events because many patients were not using the facilities that closed but had been bypassing them for some time (at least three years prior to the closure). This goes back to the perceived versus true loss of care as understood from the public’s perspective. As governments look to better manage health services costs while ensuring positive patient outcomes, these findings along with others259 suggest that the centralization of services does not have to hinder patient outcomes if done strategically. At the same time, it once more highlights the need for the public to understand what is being lost and whether or not it is of consequence to their health outcomes. As described in the conceptual framework for this thesis (Chapter 2), access is about the presence and absence of care, but also about the proximity to care. The interest here was in whether access as measured by time to care influenced likelihood of mortality, and whether there is truth to the still-often-used but contested notion of the ‘golden hour’ in relation to patient outcomes. Travel burden was defined as ground travel time broken into time increments: 0-30 mins, 31-60 mins, 60-120 mins, 121-240 mins, 240+ mins. With the 0-30 minute interval as the reference group, multivariate analyses showed no statistically significant relationship between time to care and hazard of 30-day mortality the longer the time to care estimate across AMI, stroke and trauma. This may have been because the time to care estimate used in this study did not reflect actual time to care (ie. It assumed all events occurred at home, and that patients transporting themselves did so at posted speed limits). To test and correct for this, further analyses restricted observations to events that occurred between 2009 and 2013, one using ground time estimates as before and a second replacing the estimated burden of travel (Models 5a) with the true burden as captured using Ambulance data (Models 5). This true time to care was restricted to travel time and did not include call, dispatch, en route and scene times. Using true time, no patient that accessed EMS travelled longer than four hours (240 minutes). Furthermore, adjusted hazards ratios did not demonstrate an expected dose response across all conditions. In the case of AMI, true time to care did not increase the hazard ratio unless the patient travelled longer than an hour to access care (adjusted HRAMI=1.520, 95% CI: 1.047, 1.854). This pattern mirrored the results of the complete cohort analysis (1999-2013) where the adjusted hazards of mortality for AMI patients travelling longer than 60 minutes was 149  aHR=1.124 (95% CI: 1.002, 1.262). Meanwhile for stroke and trauma the adjusted HRs for travel burden found no statistically significant effects. When comparing driving time estimates between patients that used EMS and those who were self-transported, AMI patients that relied on EMS for transport were considerably farther from care than patients that self-transported (mean=65.94 SD=157.51 mins compared with mean=37.41 SD=120.45 mins, p-value<0.0001). Similarly, trauma patients transported by paramedics lived a mean=57.35 mins (SD=149.72) from the hospital compared with a mean=48.23 mins (SD=146.07) for the self-transported group (p-value=0.0009). However, there was no significant difference in the estimated travel time between stroke patients who used EMS and those who self-transported (mean=36.05 SD=111.35 mins vs. 34.03 SD=114.31 mins, p-value=0.1822). Thus, the models suggest that access to care alone does not determine patient outcomes but rather is one covariate that influences patient results. In models that looked at both system and patient level characteristics, it appears that appropriateness of care along with other measures compensate for access. Despite some limitation to the appropriateness variable used here, these results show that if care is provided by capable institutions and is not contraindicated, it can compensate for time delays – something not regularly reported elsewhere. This is not to diminish the role that time to care plays on patient outcome, but rather to suggest that there are multiple opportunities to support favourable outcomes, and that proximity to care is not the sole determinant of patient survival. For example, a large observational study found that the sooner patients receive intravenous tPA, the better their outcomes.260 There are two ways of achieving this. The patient can be proximal to definitive care, or as suggested in other studies, paramedics can provide tPA after accurate diagnosis of a stroke while en route. To date, BC paramedics do not offer fibrinolytic therapy, but randomized controlled trials and subsequent pilot projects on the utility and success of paramedic administered tPA have proven successful across the country.261–264 Given the positive results found in other jurisdictions, it seems reasonable to suggest that well trained paramedics should have their scope of practice expanded to include the drug as part of their treatment regimen. 150  9.1.5 Compensating mechanisms 9.1.5.1 Inter-hospital transfers This thesis tested whether compensating mechanisms attenuate the influence that access plays on patient outcomes following time sensitive events. I incorporated both inter-hospital transfers and telehealth services as part of my system specific covariates included in my Cox Proportional Hazards models described in Chapter 8. Inter-hospital transfers occurred when a patient first visited one facility (admitted or not) and then transferred to another hospital for care. In cases where admitted patients were transferred, they were typically brought to the second facility for a day procedure before returning to the original hospital of admission for recovery. In total 12.2% of all patients included in the study experienced some form of transfer. AMI events had the highest frequency of transfers (29,927 events or 28.3%) followed by stroke events (10,404 events or 11.0%) and trauma (9,911 events of 4.5%). Inter-hospital transfers were found to be increase hazards of mortality across all three conditions in every model (4-5a) when using full models, but were protective for AMI and stroke patients (adjusted HRAMI=0.744, 95% CI: 0.689, 0.803; adjusted HRstroke=0.785, 95% CI: 0.664, 0.927) and deleterious for trauma (adjusted HRtrauma=1.477, 95% CI: 1.233, 1.770) when relying on parsimonious models (Appendix B). My AMI model findings conflict with others’ work that demonstrates that inter-hospital transfers are beneficial for AMI patients. Specifically, after controlling for patient characteristics such as age and comorbidity, patients stabilized locally and then transferred to definitive care fare better than those who are not.265–267 However, such studies did not include other hospital-specific characteristics (e.g. access, appropriateness, and quality), method of transport, or patients who were brought from another hospital without an initial admission (i.e. in the case where someone was self-transported to a facility ill-equipped to treat the condition (A) and was immediately transported to another hospital (B) by paramedics without first being admitted to facility A. This distinction between hospital coordinated transfers, and transfers due to lack of initial triage may explain why I found inter-hospital transfers hazardous when using my complete models to accommodate for all patient and system characteristics within the data. The story for trauma patients is very different. According to my results, local stabilization followed by transfer to definitive care is detrimental to patients’ health and incurs a higher hazard 151  of mortality. Similar results were found in a study examining the outcomes of patients taken directly to trauma centres versus those first stabilized locally before subsequent transfer, especially when patients were first taken to rural hospitals.268 A more recent systematic review found however, that there is no clear harm for trauma patients transferred.269 The authors of the systematic review cautioned that the heterogeneity of the papers included in the review may explain the lack of an association, and found that the majority of papers excluded mortalities incurred at the outlying facility, likely inflating the risk of mortality in transfers. Lastly, the authors recommend the prospective use of trauma registries to establish whether the process of transfer to trauma centres compared with direct admission negatively impact patient outcomes. While I did not conduct a prospective study, I did use administrative data and controlled for direct admission to a Level I or II trauma centre (appropriateness of care). My analysis helps further the discussion on the merits of accessible trauma centres since transfers do not appear to be a compensating mechanism. Finally, a large study investigating the merits of inter-hospital transfer for stroke patients suggests it is difficult to determine a priori who will best respond to transfers, and thus do not find an association between transfer and patient outcome.270 While inter-hospital transfers were significant during the entire study period and for the restricted cohort, they were hazardous. Again this may be because patients transferred in my study include those initially transported to an inappropriate facility via self-transport. 9.1.5.2 Telehealth The second compensating mechanism I was able to incorporate in my analysis was telestroke, a telehealth subcategory specific to stroke patients. In 2006, the province began to pilot telehealth projects and by June 2009 had established fee items for physician remuneration specific to stroke care. Telestroke works on the hub and spoke model of care. A patient admitted to a primary stroke centre (spoke) has the diagnostics run, but the tests are reviewed via video conferencing by a neurologist at a comprehensive stroke centre (hub). Based on the resident neurologists’ virtual consultation results, decisions are subsequently made about whether or not the patient should be transferred to a comprehensive stroke centre capable of managing the stroke and its sequelae. Telehealth initiatives are currently very popular because they offer a solution to the lack of specialists in remote facilities. However, there were short-comings to the Telestroke model during 152  the period I evaluated. Patients needed to be transported to primary stroke centres in order to benefit from the program. As of 2013 these were all in relatively well-populated areas: Cariboo Memorial (Thompson Cariboo Shuswap, Interior), East Kootenay (East Kootenay, Interior), Kootenay Boundary Regional (Kootenay Boundary, Interior), Penticton Regional (Okanagan, Interior), and Vernon Jubilee Hospitals (Okanagan, Interior). Given comprehensive stroke centres were at Vancouver General (Vancouver, Vancouver Coastal), St. Paul’s (Vancouver, Vancouver Coastal), Kelowna General (Okanagan, Interior), Royal Inland (Kamloops, Interior), and Victoria General (Greater Victoria, Vancouver Island), this left patients in the Fraser and Northern Health Authorities out of receiving the purported benefits in telestroke organization. This is not to say that expansion was not forthcoming, but simply to highlight vulnerabilities in the model during my study period. Of the 24,512 stroke events that occurred between April 2009 and March 2013, 155 (0.63%) received telestroke services. Ninety-one (58.7%) were transported by paramedics and 64 (41.2%) were self-transported. In the survival regression models run in Chapter 8, telestroke services were associated with poorer outcomes (models 4-5b). This conflicts with the most recent literature which finds telestroke services beneficial. A review of telestroke found it aids in the administration of tPA to patients in community hospitals with results similar to those in comprehensive stroke centres, and in particular finds it useful in shortening the time to needle for patients in rural communities.271 However, because of the technological needs (e.g. high enough bandwidth to transmit images/video, and diagnostic equipment), telestroke is currently limited to larger centres equipped with CT scans and for the time being does not really serve the truly rural communities in BC as well as it might. Taking this one step further, if telestroke is currently limited to primary stroke centres in larger hospitals, then perhaps the elevated hazards ratios in these analysis are due to general severity of stroke. These primary stroke centres are able to treat standard strokes through drip and keep: the administration of tPA following general diagnosis.  For the time being then, perhaps telestroke consultations in BC are being used to determine whether or not a patient will need subsequent invasive surgery (e.g. cases where there is large middle cerebral artery ischaemia, basilar artery occlusions or intracerebral haemorrhage).271 Therefore, the elevated hazards ratio may be due to confounding by indication owing to imbalances in the underlying risk not properly captured in the data, since the data lack indicators for stroke severity.272 Furthermore, since 2009 153  only 155 stroke admissions received telehealth services making any estimate susceptible to rare events (e.g. mortality).  9.1.6 Other explanatory variables that influence patient outcomes 9.1.6.1 Patient characteristics Of all the patient demographics included in initial analyses (e.g. neighborhood income, rurality, age, sex, and comorbidity), age, sex, and comorbidity as measured using the Charlson Comorbidity Index proved to be significant cofactors of patient outcomes, although not uniformly across all three conditions. Age was transformed from a continuous variable to three categories following the interpretation of functional form output in SAS whilst developing my survival models: young patients (0-45 years of age), middle age patients (46-75 years), and old (76+). Across all three conditions, age demonstrated a predictable dose response with hazards of mortality increasing with age. This is to be expected as studies consistently report a penalty for age and in the case of trauma, male sex.273–276 With regards to patient sex, female sex was found to be protective in trauma events (adjusted HRtrauma=0.635, 95%CI: 0.594, 0.678) but was found to be insignificant for both stroke and AMI outcomes. This may be because despite delays in diagnosing and thereby treating women for AMI events, they have similar survival outcomes although poorer quality of life measures.249,277 Patients’ Charlson Comorbidity Index values were also found to be significant in a dose response relationship where added comorbidity increased the hazard for mortality even after controlling for age. My findings overall agree with other studies that have found characteristics such as age and comorbidities almost always predict patient outcomes, but sex and income do not. Studies specific to stroke and trauma have also found severity of the event to predict mortality, but owing to the limitations of the data, I cannot report on its impact in patient survival in this study group.219 Lastly, length of stay showed that hazards of mortality improved after the initial day of admission. This is consistent with other studies when considering the mean length of stay,234 and that in the case of trauma patients, inclusion required at least a two day stay, admission to ICU or mortality within the first two days. 154  9.2 Limitations Throughout the discussion of my results I have noted some limitations of my covariates and the methodology. Here I summarize the general limitations of the study, describe their potential impact and recommend strategies for future research. 9.2.1 Access to care With redistribution efforts underway, the primary concern to communities has been whether the loss of services and thereby the added travel burden to patients has affected their health outcomes. As such, the methods for developing the access to care covariate can significantly impact the results describing regionalization’s influence on mortality outcomes for communities that lost access to care. A critical examination of how access was estimated (time to care) reveals both data and analysis limitations. Firstly, given that the data systems used did not include information on the true location of the event, time to care estimates assume that all events occurred within the home, and that self-transport to hospital occurred at the posted speed limits of roads used. Furthermore, without patient street addresses, home to hospital distance had to be estimated using postal code centroids. Of the two assumptions, the location of the event is likely to have contributed to a much larger travel burden estimate than actually experienced by patients. Table 8-18 in Chapter 8 compares patients with complete data (both travel burden estimate and true travel time) and shows true time to care is much shorter than that suggested by the estimates. This is likely because many patients did not experience their events at home and were taken to the most appropriate facility relative to the location of the event.  While it is not obvious in my data where patients experienced the event (the location field in the ambulance data is not consistently populated for BC and there is no equivalent in the DAD), some studies and reports demonstrate a significant proportion of patients experience their event away from home. For example, Ontario’s 2009 Trauma Registry Report indicates that 41% of major trauma events occur on a street or highway, and only 27% of events occur at home.278 Meanwhile, the vast majority of stroke events are reported to occur in the home.279 This can explain the additional discrepancy in time to care estimates vs. true time across the conditions, as the overestimation was much less (51%) in stroke patients compared with trauma patients (79%). Given this assumption, the sensitivity analysis in section 8.2.2 allows me to ensure that my travel 155  burden assumption does not impact my proportional hazards models to the extent that they are uninformative. Comparing models in section 8.2.1 with 8.2.2 we can see that travel burden matters but can be compensated for by other mechanisms. The repeated results and similar effects of my covariates across conditions and cohorts allow me to be confident in my general findings. With regards to the second assumption that postal code centroid to hospital distance will suffice in lieu of precise measures, this method to accommodate for data limitations has been extensively used and examined by other researchers. Compared with other population weighted models, the centroid model is the best approach for estimating time to care when the exact location of the event is unknown.280,281 Since it is very important that patient privacy is retained, it does not seem likely that the DAD will include patient information for more precise estimates of time to care (i.e. location of event) anytime in the near future. However, the BCEHS data are a source of more precise information, and future studies should concentrate on sources of data that provide the precision in time to care needed without compromising patient confidentiality. In the case of Ambulance data this was achieved by providing times for each segment of dispatch and delivery without including patient specific information. Unfortunately, even with the use of Ambulance data, patients who self-transport will still need to have travel burden estimated much as was done in this dissertation for lack of other sources of time to care. Another series of assumptions tied to the travel burden estimates was that the road conditions were consistent across all patient trips (i.e. weather, traffic) and that the same road networks existed in 1999 as they did in 2013, given I worked with 2013 road network files. With regards to correcting the first limitation, although GIS methods have become quite sophisticated and are the preferred method of estimating travel burden (Google’s products are still too slow, and date/time specific for the public’s use), again using Ambulance data moving forward would eliminate workarounds such as these since the travel time captured is a direct reflection of road and traffic conditions actually experienced. Furthermore, although several road improvement and expansion projects were underway in BC during the study period, it is unlikely that they impacted travel time for a majority of the population. However, they may have contributed to reduced time to care estimates for rural populations as they were the most affected by the projects (e.g. first portion of Cariboo connector started in 2007 and completed in 2011).282 Again this was corrected for using Ambulance data. 156  9.2.2 Appropriateness of care The appropriateness of care covariate was limited in its ability to assess whether or not appropriate care was provided. This was because appropriate care for AMI and stroke events includes recommendations on the timing of treatments for optimal patient recovery and this information is not captured in the administrative datasets available for this thesis, nor could it be requested owing to the current access issues for researchers with regards to Emergency Department records. As such, I was unable to determine whether needed diagnostic and fibrinolytic therapy was delivered within the timelines recommended in our standards of care. This limitation affected the sensitivity but not the specificity of the variable constructed, as it was possible to flag patients who received inappropriate treatment (errors of commission or omission) within the dataset. It is a fair assumption that increased time to hospital likely leads to delays in treatment which may result in poorer patient outcomes. Despite an inability to comment directly on the timeliness of care achieved beyond travel burden, the quality and appropriateness of care covariates still help capture process concerns related to the timeliness of care. The quality covariate is outcome instead of process driven, which means although it does not confirm process issues (such as whether or not patients were prescribed beta blockers following AMI), it does help indicate where a facility may not be meeting process measures by controlling for patient demographics and facility size. Thus, observed outliers may be explained by timeliness of care or other unmeasured markers of quality such as staff vigilance – either way raising flags where needed. At the same time, although the appropriateness covariate does not indicate whether or not timelines were achieved, it does flag treatment provided by inexperienced physicians or staff, and against treatment protocols. Since patient survival is the ultimate outcome, these two variables combined still capture necessary red flags even if there is as yet no definitive explanation why hospital outliers underperform or confirmation whether timeliness was achieved. 9.2.3 Condition severity Another data limitation was the lack of a severity score, particularly for trauma patients. Without it, it was not possible to confirm definitively that trauma events were restricted to severe incidents, which may have diluted results and possibly led to the underestimation of effect measures. It is unclear to what extent these analyses incorporated non-life threatening trauma events. However, given that my measures of association were still significant and matched findings 157  across the other conditions (namely that time to care matters but can be compensated with appropriateness of care), it appears the dilution in severity was not critical. Similarly, severity measures for AMI and stroke would help determine which patients should have been transported for definitive care as opposed to just observing which ones were. The measure would also have helped determine whether telestroke is indeed futile as is it currently being applied or if the association seen here was due to confounding by indication. Moving forward, Ambulance data should consistently capture the severity of the incident and data integrity needs to be ensured.  9.3 Generalizability This thesis shows that although redistribution contributed to the closure of sixteen facilities, it did not negatively impact patient health outcomes following acute myocardial infarctions, stroke and major trauma events, as examined here. Although it captured system level effects on patient mortality, the work presented did not assess the other effects of service redistribution on families of patients, and the out-of-pocket costs incurred in accessing care. Additionally, it is unclear how generalizable the findings are to other hospital services such as cancer treatment, elective procedures, obstetrics, and mental health emergencies. As such, the findings of this study are not meant to imply that hospital closures were inconsequential. It does demonstrate that the system was able to absorb changes in redistribution in so much as patient mortality did not increase following the centralization of acute care. 9.4 Recommendations While the closure of hospitals was apparent, for the most part, patients ceased using the facilities for the services studied here several years before the official cessation of services. The results in Chapter 7 demonstrated a significantly higher bypass rate in communities that eventually lost their local facility than in communities that retained theirs. These results may also help explain the higher mortality rate in communities that eventually lost services compared with their controls. Furthermore, the extensive modeling of Chapter 8 showed that although time to care can play an important role in patient outcomes, the appropriateness of care plays an important mediating role. Likewise, patient sex, age and mode of transport to hospital were all critical determinants of patient outcomes. Thus, these results demonstrate that the provincial exercise of redistributing hospital services was not, in and of itself, problematic for patient survival, provided that compensating 158  mechanisms were in place to accommodate the changes. Based on these findings, the following are recommendations for both future policy and research initiatives within British Columbia. 9.4.1 Policy recommendations Time to care has been associated with patient outcomes in many studies and reports. Rural patients in particular travel farther to access care. For AMI and stroke, there are opportunities to begin to provide treatment to patients sooner than is currently the case in British Columbia. Several international studies have reported paramedics’ ability to accurately identify STEMI and ischemic stroke patients, the effectiveness of fibrinolytic intervention en route, the benefits of this early administration, and paramedics’ willingness to extend their scope of practice.263,283,284 As Canadian pilot projects continue to prove the feasibility of paramedic initiated pharmacologic intervention, BC should implement it as part of advanced paramedic training. This is particularly important since an estimated 30-50% of patients with AMI fail to receive treatment within 90 minutes of the event, thereby needing more invasive and costly care.125 Regarding strokes, in such a model, where access to reperfusion is provided early en route, evidence supports the direct transport of stroke patients with suspected complications to comprehensive stroke centres rather than being transferred after the condition worsens.285 Closely related to this recommendation is the need to increase the public’s awareness of both the urgency of suspected AMI, trauma and stroke events, and the potential benefit of EMS assistance in transporting patients. As it stands, a large proportion of British Columbians continue to self-transport to hospital despite campaigns urging otherwise. The benefits of paramedic-initiated reperfusion cannot be enjoyed if patients do not utilize the services. Additionally, paramedics are trained to triage patients properly.286 Conversely, self-transport leads to unnecessary time delays by barring early access en route, the speed with which the patient can be transported to hospital, and the lack of triaging to designated facilities prepared to treat the incoming patient and informed of his or her condition. Another finding from my study and related to self-transport, was that early in the study period many facilities ill-equipped to treat these time sensitive events still admitted patients. As time progressed, these facilities ceased to admit these patients. However, with over 30% of patients self-transporting, it is critical that the public understands the difference between a hospital and 159  alternate facilities which may be perceived as hospitals (i.e. include hospital in their name) as this confusion adds unnecessary delays to accessing care. 9.4.2 Research recommendations There were several limitations to these analyses that prevent definitive conclusions regarding how access, appropriateness and quality of care impact patient outcomes. The subsequent recommendations are in part to correct for these limitations, as well as push for ongoing research in health services delivery related to acute care. The first recommendation is for a severity metric to be included in the hospital data (DAD) and to be routinely tabulated by paramedics with high reliability in the Ambulance (BCEHS) data. The absence of this variable made it difficult to determine whether patients who self-transported were different from those who sought paramedic services (aside from age, sex and general comorbidity) and whether hospitals were performance outliers because they did indeed provide poorer quality of care, or because patient severity indicated direct transport to these facilities. Tied to this recommendation would be the inclusion of emergency department data. As it stands, I was unable to use the administrative data at my disposal to confirm whether patients received care in the guideline specified time frames. As such, my appropriateness of care metric lacked a critical assessment of the timeliness of care. Future research may find that while the guidelines are not being met, timeliness rather than the expertise of the overseeing physician or staff may account for poor patient outcomes, or that there is an interaction between expertise and timeliness that explains why experience is critical. Additionally, research examining the distribution of paramedic services and wait times to access EMS should be conducted. Such a study would help illuminate the costs versus benefits of waiting for EMS services relative to self-transport, particularly when accounting for geographic variations in access. Furthermore, such research, especially if augmented by a better severity score, would help explain why air transport is repeatedly associated with increased hazards of mortality. Ideally, such a study would have true time to care for self-transport as opposed to estimated based on home location. Lastly, more research is needed to understand whether a patient’s sex biases paramedics’ and physicians’ treatment of their trauma event, and whether women continue to experience delays in 160  diagnoses and treatment of AMI events. Both age and sex were significant determinants of patient survival and appropriate treatment. As the burgeoning use of administrative data in health services research suggests, improvements in the breadth and quality of data available will enable policy makers to better allocate services and improve patient outcomes.   161  Chapter 10: Conclusion This dissertation sought to answer the question “What effect did British Columbia’s redistribution of hospital services have on residents’ outcomes, and did quality and appropriateness of care compensate for changes in access to acute care?”  In preparation of the analyses, I developed a theoretical framework that integrated and improved key ideas discussed in Donabedian’s Quality of Care, Anderson’s Behavioural Model, and Penchansky’s Access to Care frameworks to better reflect salient factors associated with acute outcomes identified in health services research (i.e. access, appropriateness, and quality of care). The framework was not meant to imply causal pathways but rather to summarize seminal work and guide the analysis process. Although it proved helpful, it is important to acknowledge that the concepts represented in the framework (i.e. appropriateness, access, quality) could not be measured fully. This was a consequence of the limitations of existing data available for this research, but also the complex concepts the framework described. Nevertheless, the framework helped simplify a complex question and guided the analysis strategy.  To answer the question, the analyses used administrative data for three time sensitive medical events (AMI, trauma and stroke) and the measurable effects of redistribution efforts on patient outcomes; whether redistribution affected patients’ access to services; and ultimately, what effect access to care had on patient outcomes after adjusting for relevant patient and system level characteristics.  A main finding is that recent redistribution efforts have not increased mortality rates for AMI, trauma or stroke events. This is likely because redistribution was related to facility underutilization before the change, whether pre-planned or not. Perhaps more importantly, these results show that the effects of longer travel burden on patient outcomes can be compensated for with appropriate and quality care. However, timely access to effective treatment should continue to be sought. This work demonstrates that efficiencies in health system delivery can be gained by redistributing acute care services without harming patient outcomes. This can be done by paying close attention to processes that can mediate the effects of proximity to care. For example, service delivery methods which allow treatment to begin sooner (i.e. en route) should be explored, along with the distribution of paramedic services, and efforts to increase utilization of EMS by the public 162  during emergencies. At the same time, it does not suggest that the redistribution of health services did not impact patient outcomes for conditions or factors (e.g. out of pocket costs) not studied here.  163  References 1. Ministry of Health Planning. A New Era for Patient-Centred Health Care. (Ministry of Health Planning, 2001). 2. Fleet, R., Archambault, P., Plant, J. & Poitras, J. Access to emergency care in rural Canada: should we be concerned? Can. J. Emerg. Med. 15, 191–193 (2013). 3. Grzybowski, S., Kornelsen, J. & Schuurman, N. Planning the optimal level of local maternity service for small rural communities: A systems study in British Columbia. Health Policy (2009). doi:10.1016/j.healthpol.2009.03.007 4. Emergency room in Kaslo, BC, set to cut hours. CBC News (2013). 5. Barer, M. L., Morgan, S. G. & Evans, R. G. Strangulation or Regionalization? Costs and Access in Canadian Hospitals. Longwoods Rev. 7, (2003). 6. BC Patient Transfer Network (PTN) (formerly BC Bedline). (2013). Available at: http://www.health.gov.bc.ca/ehsc/bcbedline/. 7. British Columbia Ministry of Health. Telehealth: Current State. Available at: http://www.health.gov.bc.ca/ehealth/telehealth_project.html. (Accessed: 29th December 2013) 8. Chan, A. W. et al. In-hospital outcomes of a regional ST-Segment Elevation Myocardial Infarction Acute Transfer and Repatriation Program. Can. J. Cardiol. 27, 664.e1-664.e8 (2011). 9. BC Ministries of Health Services and Health Planning. Standards of Accessibility and Guidelines for Provision of Sustainable Acute Care Services by Health Authorities. (Government of British Columbia, 2002). 10. Ministry of Health. Evergreen Document. (Ministry of Health, 1997). 164  11. Parker, E. B. & Campbell, J. L. Measuring access to primary medical care: some examples of the use of geographical information systems. Health Place 4, 183–193 (1998). 12. Jones, S. G., Ashby, A. J., Momin, S. R. & Naidoo, A. Spatial implications associated with using Euclidean distance measurements and geographic centroid imputation in health care research. Health Serv. Res. 45, 316–327 (2010). 13. Apparicio, P., Abdelmajid, M., Riva, M. & Shearmur, R. Comparing alternative approaches to measuring the geographical accessibility of urban health services: Distance types and aggregation-error issues. Int. J. Health Geogr. 7, 7 (2008). 14. Delamater, P. L., Messina, J. P., Shortridge, A. M. & Grady, S. C. Measuring geographic access to health care: raster and network-based methods. Int. J. Health Geogr. 11, 15 (2012). 15. Haynes, R., Jones, A. P., Sauerzapf, V. & Zhao, H. Validation of travel times to hospital estimated by GIS. Int. J. Health Geogr. 5, 40 (2006). 16. Phibbs, C. S. & Luft, H. S. Correlation of travel time on roads versus straight line distance. Med. Care Res. Rev. MCRR 52, 532–542 (1995). 17. DesMeules, M. et al. How healthy are rural Canadians? An assessment of their health status and health determinants. (Canadian Institute for Health Information, 2006). 18. Feero, S., Hedges, J. R., Simmons, E. & Irwin, L. Does out-of-hospital EMS time affect trauma survival? Am. J. Emerg. Med. 13, 133–135 (1995). 19. Gonzalez, R. P., Cummings, G. R., Phelan, H. A., Mulekar, M. S. & Rodning, C. B. Does increased emergency medical services prehospital time affect patient mortality in rural motor vehicle crashes? A statewide analysis. Am. J. Surg. 197, 30–34 (2009). 20. Grossman, D. C. et al. Urban-rural differences in prehospital care of major trauma. J. Trauma 42, 723–729 (1997). 165  21. Eisenberg, M. S., Bergner, L. & Hallstrom, A. Cardiac resuscitation in the community. Importance of rapid provision and implications for program planning. JAMA J. Am. Med. Assoc. 241, 1905–1907 (1979). 22. Roth, R., Stewart, R. D., Rogers, K. & Cannon, G. M. Out-of-hospital cardiac arrest: factors associated with survival. Ann. Emerg. Med. 13, 237–243 (1984). 23. Newgard, C. D. et al. Emergency medical services intervals and survival in trauma: assessment of the ‘golden hour’ in a North American prospective cohort. Ann. Emerg. Med. 55, 235–246.e4 (2010). 24. Casey, Q. National trauma divide must be narrowed. Can. Med. Assoc. J. 182, (2010). 25. Simons, R. et al. A population-based analysis of injury-related deaths and access to trauma care in rural-remote Northwest British Columbia. J. Trauma 69, 11–19 (2010). 26. Alter, D. A., Austin, P. C. & Tu, J. V. Community factors, hospital characteristics and inter-regional outcome variations following acute myocardial infarction in Canada. Can. J. Cardiol. 21, 247–255 (2005). 27. Gamble, J.-M. et al. Patterns of Care and Outcomes Differ for Urban Versus Rural Patients With Newly Diagnosed Heart Failure, Even in a Universal Healthcare System. Circ. Heart Fail. 4, 317–323 (2011). 28. Iii, A. G. M. & Kohrs, F. P. A comparison of health status between rural and urban adults. J. Community Health 20, 423–431 (1995). 29. Casey, M. M., Thiede Call, K. & Klingner, J. M. Are rural residents less likely to obtain recommended preventive healthcare services? Am. J. Prev. Med. 21, 182–188 (2001). 30. Lavergne, M. R. Understanding Geographic Variation in Health Care Costs in British Columbia. (University of British Columbia, 2015). 166  31. Grzybowski, S. & Kornelsen, J. Rural Health Services: Finding the Light at the End of the Tunnel. Healthc. Policy 8, 10–16 (2013). 32. Smith, K. B., Humphreys, J. S. & Wilson, M. G. A. Addressing the health disadvantage of rural populations: how does epidemiological evidence inform rural health policies and research? Aust. J. Rural Health 16, 56–66 (2008). 33. Minister of Health. Canada Health Act - Annual Report 2009-2010. (Ministry of Health, 2010). 34. Hemmelgarn, B. R., Ghali, W. A. & Quan, H. A case study of hospital closure and centralization of coronary revascularization procedures. CMAJ Can. Med. Assoc. J. 164, 1431–1435 (2001). 35. Sheps, S. B. et al. Hospital downsizing and trends in health care use among elderly people in British Columbia. CMAJ 163, 397–401 (2000). 36. Canadian Hospital Reporting Project Technical Notes. (Canadian Institute for Health Information, 2013). 37. McKay, N. L. & Coventry, J. A. Access implications of rural hospital closures and conversions. Hosp. Health Serv. Adm. 40, 227–246 (1995). 38. Hsia, R. Y.-J. & Shen, Y.-C. Rising Closures Of Hospital Trauma Centers Disproportionately Burden Vulnerable Populations. Health Aff. (Millwood) 30, 1912–1920 (2011). 39. Kaufman, B. G. et al. The Rising Rate of Rural Hospital Closures. J. Rural Health Off. J. Am. Rural Health Assoc. Natl. Rural Health Care Assoc. (2015). doi:10.1111/jrh.12128 40. Rosenbach, M. L. & Dayhoff, D. A. Access to care in rural America: impact of hospital closures. Health Care Financ. Rev. 17, 15–37 (1995). 167  41. Schull, M. J. et al. Underuse of prehospital strategies to reduce time to reperfusion for ST-elevation myocardial infarction patients in 5 Canadian provinces. Can. J. Emerg. Med. 11, 473–90 (2008). 42. Shadish, W. R., Cook, T. D. & Campbell, D. T. Experimental and Quasi-Experimental Designs for Generalized Causal Inference. (Wadsworth). 43. McGuirk, M. A. & Porell, F. W. Spatial Patterns of Hospital Utilization: The Impact of Distance and Time. Inquiry 21, 84–95 (1984). 44. Cronin, P., Ryan, F. & Coughlan, M. Undertaking a literature review: a step-by-step approach. Br. J. Nurs. 17, (2008). 45. Levesque, J.-F., Harris, M. F. & Russell, G. Patient-centred access to health care: conceptualising access at the interface of health systems and populations. Int. J. Equity Health 12, 18 (2013). 46. Donabedian, A. Evaluating the Quality of Medical Care. Milbank Mem. Fund Q. 44, 166–206 (1966). 47. Donabedian, A. Evaluating the Quality of Medical Care. Milbank Q. 83, 691–729 (2005). 48. Andersen, R. M. Revisiting the behavioral model and access to medical care: does it matter? J. Health Soc. Behav. 36, 1–10 (1995). 49. Penchansky, R. & Thomas, J. W. The Concept of Access: Definition and Relationship to Consumer Satisfaction. Med. Care 19, 127–140 (1981). 50. McDonald, K. M. et al. Conceptual Frameworks and Their Application to Evaluating Care Coordination Interventions. (Agency for Healthcare Research and Quality (US), 2007). 168  51. Fortney, J. C., Burgess, J. F., Bosworth, H. B., Booth, B. M. & Kaboli, P. J. A Re-conceptualization of Access for 21st Century Healthcare. J. Gen. Intern. Med. 26, 639–647 (2011). 52. Ngui, A. N. & Apparicio, P. Optimizing the two-step floating catchment area method for measuring spatial accessibility to medical clinics in Montreal. BMC Health Serv. Res. 11, 166 (2011). 53. Phillippi, J. Women’s perceptions of access to prenatal care in the United States: a literature review. J. Midwifery Womens Health 54, 219–25 (2009). 54. Campbell, S. M., Roland, M. O. & Buetow, S. A. Defining quality of care. Soc. Sci. Med. 51, 1611–1625 (2000). 55. Donabedian, A. Models for Organizing the Delivery of Personal Health Services and Criteria for Evaluating Them. Milbank Mem. Fund Q. 50, 103–154 (1972). 56. Lerner, E. B. & Moscati, R. M. The Golden Hour: Scientific Fact or Medical ‘Urban Legend’? Acad. Emerg. Med. 8, 758–760 (2001). 57. Grines, C. L. et al. A randomized trial of transfer for primary angioplasty versus on-site thrombolysis in patients with high-risk myocardial infarctionThe air primary angioplasty in myocardial infarction study. J. Am. Coll. Cardiol. 39, 1713–1719 (2002). 58. Veenema, K. R. & Rodewald, L. E. Stabilization of Rural Multiple-Trauma Patients at Level III Emergency Departments Before Transfer to a Level I Regional Trauma Center. Ann. Emerg. Med. 25, 175–181 (1995). 59. Fried, M. J., Bruce, J., Colquhoun, R. & Smith, G. Inter-hospital transfers of acutely ill adults in Scotland. Anaesthesia 65, 136–144 (2010). 169  60. Smith, R. M. & Conn, A. K. Prehospital care − Scoop and run or stay and play? Injury 40, Supplement 4, S23–S26 (2009). 61. Nirula, R., Maier, R., Moore, E., Sperry, J. & Gentilello, L. Scoop and Run to the Trauma Center or Stay and Play at the Local Hospital: Hospital Transferʼs Effect on Mortality: J. Trauma Inj. Infect. Crit. Care 69, 595–601 (2010). 62. Haas, B. & Nathens, A. B. Pro/con debate: Is the scoop and run approach the best approach to trauma services organization? Crit. Care 12, 224 (2008). 63. Deakin, C. D. in Anaesthesia, Pain, Intensive Care and Emergency Medicine - A.P.I.C.E. (ed. Gullo, A.) (Springer-Verlag, 1996). 64. Chassin, M. R. et al. Does inappropriate use explain geographic variations in the use of health care services? A study of three procedures. (The RAND Corporation, 1987). 65. Hanley, G., Janssen, P. A. & Greyson, D. Regional Variation in the Cesarean Delivery and Assisted Vaginal Delivery Rates. Obstet. Gynecol. 115, (2010). 66. Leduc, E. Defining rurality: A General Practice Rurality Index for Canada. Can. J. Rural Med. 2, 125 (1997). 67. Institute of Medicine of the National Academies. Crossing the Quality Chasm: The IOM Health Care Quality Initiative. (national Academy of Sciences, 2013). 68. Thiemann, D. R., Coresh, J., Oetgen, W. J. & Powe, N. R. The Association between Hospital Volume and Survival after Acute Myocardial Infarction in Elderly Patients. N. Engl. J. Med. 340, 1640–1648 (1999). 69. Halm, E. A., Lee, C. & Chassin, M. R. Is volume related to outcome in health care? A systematic review and methodologic critique of the literature. Ann. Intern. Med. 137, 511–520 (2002). 170  70. Goldman, B. Beware of Small, Rural Hospitals. White Coat, Black Art (2011). 71. Baldwin, L.-M. et al. Quality of Care for Acute Myocardial Infarction in Rural and Urban US Hospitals. J. Rural Health 20, 99–108 (2004). 72. Glenn, L. L. & Jijon, C. R. Risk-adjusted In-hospital Death Rates for Peer Hospitals in Rural and Urban Regions. J. Rural Health 15, 94–107 (1999). 73. Hartz, A. & James, P. A. A Systematic Review of Studies Comparing Myocardial Infarction Mortality for Generalists and Specialists: Lessons for Research and Health Policy. J. Am. Board Fam. Med. 19, 291–302 (2006). 74. James, P. A., Li, P. & Ward, M. M. Myocardial Infarction Mortality in Rural and Urban Hospitals: Rethinking Measures of Quality of Care. Ann. Fam. Med. 5, 105–111 (2007). 75. Grzybowski, S., Stoll, K. & Kornelsen, J. Distance matters: a population based study examining access to maternity services for rural women. BMC Health Serv. Res. 11, (2011). 76. Schuurman, N., Fiedler, R. S., Grzybowski, S. C. & Grund, D. Defining rational hospital catchments for non-urban areas based on travel-time. Int. J. Health Geogr. 5, 43 (2006). 77. Grzybowski, S., Kornelsen, J. & Cooper, E. Rural maternity care services under stress: the experience of providers. Can. J. Rural Med. 12, (2007). 78. Albright, K. C. et al. ACCESS: acute cerebrovascular care in emergency stroke systems. Arch. Neurol. 67, 1210–1218 (2010). 79. Wallace, D. D. et al. Accuracy of Prehospital Transport Time Estimation. Acad. Emerg. Med. 21, 10–16 (2014). 80. British Columbia Ambulance Service. BC Ambulance Services Treatment Guidelines: Introduction. (2007). 81. Statistics Canada. The 10 leading causes of death, 2011. (Statistics Canada, 2014). 171  82. Public Health Agency of Canada. Leading Causes of Death and Hospilization in Canada. (2006). 83. Statistics Canada. Causes of Death, 2004. (Statistics Canada, 2007). 84. Canadian Institute for Health Information. Health Indicators Interactive Tool. (CIHI, 2014). 85. Statistics Canada. Mortality, summary list of causes 2008. (Statistics Canada, 2011). 86. Heart and Stroke Foundation of Canada. 2007 Annual Report on Canadians’ Health. (Heart and Stroke Foundation of Canada, 2007). 87. Alter, D. A., Naylor, C. D., Austin, P. C. & Tu, J. V. Biology or Bias: Practice Patterns and Long-Term Outcomes for Men and Women with Acute Myocardial Infarction. J. Am. Coll. Cardiol. 39, 1909–16 (2002). 88. Bertoni, A. G., Bonds, D. E., Lovato, J., Goff, D. C. & Brancati, F. L. Sex disparities in procedure use for acute myocardial infarction in the United States, 1995 to 2001. Am. Heart J. 147, 1054–60 (2004). 89. Fang, J. & Alderman, M. H. Gender differences of revascularization in patients with acute myocardial infarction. Am. J. Cardiol. 97, 1722–26 (2006). 90. Tu, J. V. et al. Outcomes of acute myocardial infarction in Canada. Can. J. Cardiol. 19, 893–901 (2003). 91. Randall R. Fransoo et al. Age difference explains gender difference in cardiac intervention rates after acute myocardial infarction. Health Policy 6, 88–103 (2010). 92. Hall, R. E. & Tu, J. V. Chapter 7: Hospitalization rates and length of stay for cardiovascular conditions in Canada, 1994 to 1999. Can. J. Cardiol. 19, 1123–31 (2003). 93. Thygesen, K., Alpert, J. S. & White, H. D. Universal definition of myocardial infarction. J. Am. Coll. Cardiol. 50, 2173–95 (2007). 172  94. Danchin, N. et al. Comparison of thrombolysis followed by broad use of percutaneous coronary intervention with primary percutaneous coronory intervention for ST-Segment-Elevation acute myocardial infarction. Circulation 118, 268–276 (2008). 95. Chan, A. W. et al. Improved Survival Associated with Pre-Hospital Triage Strategy in a Large Regional ST-Segment Elevation Myocardial Infarction Program. JACC Cardiovasc. Interv. 5, 1239–46 (2012). 96. Welsh, R. C., Travers, A., Senaratne, M., Williams, R. & Armstrong, P. W. Feasibility and applicability of paramedic-based prehospital fibrinolysis in a large North American center. Am. Heart J. 152, 1007–14 97. Eirc Boersma, Arthur C P Maas, Jaap W Deckers & Maarten L Simons. Early thrombolytic treatment in acute myocardial infarction: reappraisal of the golden hour. Lancet 348, 771–75 (1996). 98. Grines, C. L. Should every patient undergo cardiac catheterization after myocardial infarction? J. Nucl. Cardiol. Off. Publ. Am. Soc. Nucl. Cardiol. 1, S131-133 (1994). 99. Cardiac Services BC. Cardiac Services BC Annual Report 2011. (2011). 100. Robinson, T., Zaheer, Z. & Mistri, A. K. Thrombolysis in Acute Ischaemic Stroke: An Update. Ther. Adv. Chronic Dis. 2, 119–131 (2011). 101. Joseph Kwan, Peter Hand & Peter Sandercock. A systematic review of barriers to delivery of thrombolysis for acute stroke. Age Ageing 33, 116–121 (2004). 102. Association of outcome with early stroke treatment: pooled analysis of ATLANTIS, ECASS, and NINDS rt-PA stroke trials. The Lancet 363, 768–774 (2004). 173  103. The National Institute of Neurological Disorders and Stroke rt-PA Stroke Study Group. Tissue plasminogen activator for acute ischemic stroke. N. Engl. J. Med. 333, 1581–1587 (1995). 104. Marler, J. R. et al. Early stroke treatment associated with better outcome The NINDS rt-PA Stroke Study. Neurology 55, 1649–1655 (2000). 105. Nogueira, R. G. & Smith, W. S. Emergency treatment of acute ischemic stroke: expanding the time window. Curr. Treat. Options Neurol. 11, 433–443 (2009). 106. Ganesh, A. et al. The quality of treatment of hyperacute ischemic stroke in Canada: a retrospective chart audit. Can. Med. Assoc. J. 2, 233–239 (2014). 107. Stroke Collaborative – Success Stories | BC Patient Safety & Quality Council. (2012). 108. Collier, T. Systematic Stroke Care. (2014). 109. Ministry of Health Services & Heart and Stroke Foundation of Canada. Telestroke Update. (Ministry of Health Services, 2009). 110. Heart and Stroke Foundation of Canada & Canadian Stroke Network. Telestroke Implementation Toolkit. (Canadian Stroke Network, 2013). 111. Canadian Stroke Network. Increasing access to quality stroke care by implementing telestroke in Canada. (Canadian Stroke Network, 2014). 112. Peden, M., McGee, K. & (Eds), E. K. Injury: A leading Cause of the Global Burden of Disease, 2000. (World Health Organization, 2002). 113. Haas, B., Poon, V., Waller, B., Sidhom, P. & Fortin, C. M. National Trauma Registry 2011 Report: Hospitalizations for Major Injury in Canada, 2008-2009 Data. (Canadian Institute for Health Information, 2011). 174  114. Chen, B., Maio, R. F., Green, P. E. & Burney, R. E. Geographical variation in preventable deaths from motor vehicle crashes. J. Trauma 38, 228–32 (1995). 115. Gonzalez, R. P., Cummings, G. R., Mulekar, M. S. & Rodning, C. B. Increased mortality in rural vehicular trauma: identifying contributing factors through data linkage. J. Trauma 61, 404–9 (2006). 116. Peek-Asa, C., Zwerling, C. & Stallones, L. Acute traumatic injuries in rural populations. Am. J. Public Health 94, 1689–93 (2004). 117. Rogers, F. B., Shackford, S. R., Osler, T., Vane, D. W. & Davis, J. H. Rural Trauma: the challenge for the next decade. J. Trauma 47, 802–27 (1999). 118. Shafi, S., Nathens, A. B., Elliot, A. C. & Gentiello, L. Effect of trauma system on motor vehicle occupant mortality: a comparison between states with and without a formal system. J. Trauma 61, 1374–78 (2006). 119. Trauma Association of Canada & Association Canadienne de Traumatologie. Trauma System Accreditation Guidelines. (Trauma Association of Canada, 2011). 120. Hameed, S. M. et al. Access to trauma in Canada. J. Trauma 69, 1350–1361 (2010). 121. Grumbach K, Anderson GM, Luft HS, Roos LL & Brook R. Regionalization of cardiac surgery in the United States and Canada: Geographic access, choice, and outcomes. JAMA 274, 1282–1288 (1995). 122. Liu, L., Hader, J., Brossart, B., White, R. & Lewis, S. Impact of rural hospital closures in Saskatchewan, Canada. Soc. Sci. Med. 52, 1793–1804 (2001). 123. Canadian Institute for Health Information [creator] (2013). Discharge Abstract Database (Hospital Separations File). V2. Population Data BC [publisher]. Data Extract. MOH. (2015). Available at: http://www.popdata.bc.ca/data. 175  124. British Columbia Ministry of Health [creator] (2013). Medical Services Plan (MSP) Payment Information File. V2. Population Data BC [publisher]. Data Extract. MOH. (2015). Available at: https://www.popdata.bc.ca/data. 125. Johnston, S., Brightwell, R. & Ziman, M. Paramedics and pre‐hospital management of acute myocardial infarction: diagnosis and reperfusion. Emerg. Med. J. EMJ 23, 331–334 (2006). 126. Canadian Institute for Health Information. CIHI Data Quality Study of the 2005-2006 Discharge Abstract Database. (CIHI, 2009). 127. Rothman, K. J., Greenland, S. & Lash, T. L. Modern Epidemiology. (Lippincott Williams and Wilkins, 2008). 128. Kutner, M. H., Nachtsheim, C. & Neter, J. Applied linear regression models. (McGraw-Hill/Irwin, 2004). 129. Metcalfe, A. et al. Case definitions for acute myocardial infarction in administrative databases and their impact on in-hospital mortality rates. Health Serv. Res. 48, 290–318 (2013). 130. Kokotailo, R. A. & Hill, M. D. Coding of Stroke and Stroke Risk Factors Using International Classification of Diseases, Revisions 9 and 10. Stroke 36, 1776–1781 (2005). 131. Tirschwell, D. L. & Jr., W. T. L. Validating Administrative Data in Stroke Research. Stroke 33, 2465–2470 (2002). 132. Tabitha Garwe, Linda D. Cowan, Barbara R. Neas, John C. Sacra & Roxie M. Albrecht. Directness of Transport of Major Trauma Patients to a Level I Trauma Center: A Propensity-Adjusted Survival Analysis of the Impact on Short-Term Mortality. J. Trauma Inj. Infect. Crit. Care 70, 1118–1127 (2011). 176  133. Bulger, E. M. et al. Impact of prehospital mode of transport after severe injury: A multicenter evaluation from the Resuscitation Outcomes Consortium. J. Trauma 72, 567–575 (2012). 134. Palmer, C. Major Trauma and the Injury Severity Score - Where Should We Set the Bar? Annu. Proc. Assoc. Adv. Automot. Med. 51, 13–29 (2007). 135. Langley, J. & Brenner, R. What is an injury? Inj. Prev. 10, 69–71 (2004). 136. National Trauma Registry Comprehensive Data Set - Data Dictionary. (Canadian Institute for Health Information, 2012). 137. Kawachi, I. & Subramanian, S. V. Neighbourhood influences on health. J. Epidemiol. Community Health 61, 3–4 (2007). 138. Pilote, L. et al. Socioeconomic Status, Access to Health Care, and Outcomes after Acute Myocardial Infarction in Canada’s Universal Health Care System. Med. Care 45, 638–646 (2007). 139. Dailey, A. B., Kasl, S. V., Holford, T. R., Calvocoressi, L. & Jones, B. A. Neighborhood-Level Socioeconomic Predictors of Nonadherence to Mammography Screening Guidelines. Cancer Epidemiol. Biomarkers Prev. 16, 2293–2303 (2007). 140. Figure 3 Statistical Area Classification (SAC) hierarchy - Census Dictionary. Available at: https://www12.statcan.gc.ca/census-recensement/2011/ref/dict/figures/figure3-dict-eng.cfm. (Accessed: 26th October 2015) 141. More information on Statistical Area Classification (SAC). Available at: http://www12.statcan.ca/census-recensement/2006/ref/dict/geo045a-eng.cfm. (Accessed: 26th October 2015) 177  142. Charlson, M. E., Pompei, P., Ales, K. L. & MacKenzie, C. R. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J. Chronic Dis. 40, 373–383 (1987). 143. Huntley, A. L., Johnson, R., Purdy, S., Valderas, J. M. & Salisbury, C. Measures of Multimorbidity and Morbidity Burden for Use in Primary Care and Community Settings: A Systematic Review and Guide. Ann. Fam. Med. 10, 134–141 (2012). 144. de Groot, V., Beckerman, H., Lankhorst, G. J. & Bouter, L. M. How to measure comorbidity: a critical review of available methods. J. Clin. Epidemiol. 56, 221–229 (2003). 145. Sundararajan, V. et al. New ICD-10 version of the Charlson comorbidity index predicted in-hospital mortality. J. Clin. Epidemiol. 57, 1288–1294 (2004). 146. Deyo, R. A., Cherkin, D. C. & Ciol, M. A. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J. Clin. Epidemiol. 45, 613–619 (1992). 147. Quan, H. et al. Updating and validating the Charlson comorbidity index and score for risk adjustment in hospital discharge abstracts using data from 6 countries. Am. J. Epidemiol. 173, 676–682 (2011). 148. Elixhauser, A., Steiner, C., Harris, D. R. & Coffey, R. M. Comorbidity measures for use with administrative data. Med. Care 36, 8–27 (1998). 149. Clench-Aas, J. et al. Methodological development and evaluation of 30-day mortality as quality indicator for Norwegian hospitals. 200 (The Norwegian Knowledge Centre for the Health Services (NOKC), 2005). 150. Bureau of Health Information. Spotlight on measurement: 30-day mortality following hospitalization, five clinical conditions, NSW, July 2009-June 2012. 1–28 (Bureau of Health Information, 2013). 178  151. Canadian Institute for Health Information. Hospital Standardized Mortality Ratio: Technical Notes. (Canadian Institute for Health Information, 2015). 152. Osman, M., Quail, J., Hudema, N. & Hu, N. Using SAS to Create Episodes-of-Hospitalization for Health Services Research. in 3281–2015, (Saskatchewan Health Quality Council, 2015). 153. Wright, J. & Shojania, K. G. Measuring the Quality of Hospital Care. BMJ 338, 783–784 (2009). 154. Werner, R. M., Bradlow, E. T. & Asch, D. A. Does Hospital Performance on Process Measures Directly Measure High Quality Care or Is It a Marker of Unmeasured Care? Health Serv. Res. 43, 1464–1484 (2008). 155. Pouw, M. E., Peelen, L. M., Moons, K. G. M., Kalkman, C. J. & Lingsma, H. F. Including post-discharge mortality in calculation of hospital standardised mortality ratios: retrospective analysis of hospital episode statistics. BMJ 347, f5913 (2013). 156. Vasilevskis, E. E. et al. Relationship between discharge practices and intensive care unit in-hospital mortality performance: evidence of a discharge bias. Med. Care 47, 803–812 (2009). 157. Ben-Tovim, D. I., Pointer, S. C., Woodman, R., Hakendorf, P. H. & Harrison, J. E. Routine use of administrative data for safety and quality purposes--hospital mortality. Med. J. Aust. 193, S100-103 (2010). 158. Bottle, A., Jarman, B. & Aylin, P. Hospital standardized mortality ratios: sensitivity analyses on the impact of coding. Health Serv. Res. 46, 1741–1761 (2011). 179  159. Lee, K. C. L., Sethuraman, K. & Yong, J. On the Hospital Volume and Outcome Relationship: Does Specialization Matter More Than Volume? Health Serv. Res. (2015). doi:10.1111/1475-6773.12302 160. Australian Commission on Safety and Quality in Health Care. National core, hospital-based outcome indicator specification (2012). (2012). 161. Normand, S.-L. T., Glickman, M. E. & Ryan, T. J. in Case Studies in Bayesian Statistics (eds. Gatsonis, C. et al.) 155–236 (Springer New York, 1997). 162. Byrne, B. E., Mamidanna, R., Vincent, C. A. & Faiz, O. Population-based cohort study comparing 30- and 90-day institutional mortality rates after colorectal surgery. Br. J. Surg. 100, 1810–1817 (2013). 163. Palmer, C. Major Trauma and the Injury Severity Score - Where Should We Set the Bar? Annu. Proc. Assoc. Adv. Automot. Med. 51, 13–29 (2007). 164. Chen, E. & Naylor, C. D. Variation in hospital length of stay for acute myocardial infarction in Ontario, Canada. Med. Care 32, 420–435 (1994). 165. Li, Q. et al. National trends in hospital length of stay for acute myocardial infarction in China. BMC Cardiovasc. Disord. 15, 9 (2015). 166. Russo, C. A., Ho, K. & Elixhauser, A. in Healthcare Cost and Utilization Project (HCUP) Statistical Briefs (Agency for Health Care Policy and Research (US), 2006). 167. Guagliardo, M. F. Spatial accessibility of primary care: concepts, methods and challenges. Int. J. Health Geogr. 3, 3 (2004). 168. Bliss, R. L., Katz, J. N., Wright, E. A. & Losina, E. Estimating Proximity to Care: Are straight line and zipcode centroid distances acceptable proxy measures? Med. Care 50, 99–106 180  169. Boscoe, F. P., Henry, K. A. & Zdeb, M. S. A nationwide comparison of driving distance versus straight-line distance to hospitals. Prof. Geogr. 64, 1–12 (2012). 170. SAS Institute Inc. The SAS system for Windows. (SAS Inst., 2011). 171. DMTI Spatial, Inc. [creator]. CanMap Postal Code Suite. (DMTI Spatial, 2013). 172. DMTI Spatial, Inc. CanMap RouteLogistics. (DMTI Spatial, 2009). 173. BC Ministry of Health. Hospital Address List. (2014). 174. Raknes, G. & Hunskaar, S. Method Paper – Distance and Travel Time to Casualty Clinics in Norway Based on Crowdsourced Postcode Coordinates: A Comparison with Other Methods. PLOS ONE 9, e89287 (2014). 175. DMTI Spatial, Inc. CanMap Postal Code Suite, v2013.3. (2013). 176. The Google geocoding API. (2014). 177. Benjamin F. Miller et al. Primary Care, Behavioural Health, Provider Colocation, and Rurality. J. Am. Board Fam. Pract. 27, 367–374 (2014). 178. Henry, K. A. & Boscoe, F. P. Estimating the accuracy of geographical imputation. Int. J. Health Geogr. 7, (2008). 179. Statistics Canada. Representative point. (Statistics Canada, 2013). 180. Statistics Canada. How postal codes map to geographic areas. (Statistics Canada, 2007). 181. Heart and Stroke Foundation of Canada. Annual Report 2010/11. (Provincial Health Services Authority and British Columbia Ambulance Service, 2011). 182. Vancouver Coastal Health. Patient Transportation Options. (2010). 183. Iezzoni, L. I. Assessing quality using administrative data. Ann. Intern. Med. 127, 666–674 (1997). 181  184. Canadian Institute for Health Information. Canadian Classification of Health Interventions: Volume Three - Tabular List. (CIHI, 2012). 185. Canadian Institute for Health Information. Coding Standards for Version 2009 ICD-10-CA and CCI, Revised September 2009. (CIHI, 2009). 186. ICD-9/CCP and ICD-9-CM | CIHI. (2015). Available at: https://www.cihi.ca/en/data-and-standards/standards/classification-and-coding/icd-9ccp-and-icd-9-cm. (Accessed: 3rd November 2015) 187. ICD-10-CA | CIHI. (2015). Available at: https://www.cihi.ca/en/data-and-standards/standards/classification-and-coding/icd-10-ca. (Accessed: 3rd November 2015) 188. Ouriel, K. A history of thrombolytic therapy. J. Endovasc. Ther. Off. J. Int. Soc. Endovasc. Spec. 11 Suppl 2, II128-133 (2004). 189. Thrombolytic therapy: MedlinePlus Medical Encyclopedia. Available at: http://www.nlm.nih.gov/medlineplus/ency/article/007089.htm. (Accessed: 22nd May 2015) 190. Ryan, T. J. et al. ACC/AHA guidelines for the management of patients with acute myocardial infarctionA report of the American College of cardiology/American Heart Association Task Force on Practice Guidelines (Committee on Management of Acute Myocardial Infarction). J. Am. Coll. Cardiol. 28, 1328–1419 (1996). 191. Ryan, T. J. et al. 1999 Update: ACC/AHA Guidelines for the Management of Patients With Acute Myocardial Infarction: Executive Summary and Recommendations A Report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines (Committee on Management of Acute Myocardial Infarction). Circulation 100, 1016–1030 (1999). 182  192. Antman, E. M. et al. ACC/AHA Guidelines for the Management of Patients With ST-Elevation Myocardial Infarction—Executive Summary A Report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines (Writing Committee to Revise the 1999 Guidelines for the Management of Patients With Acute Myocardial Infarction). Circulation 110, 588–636 (2004). 193. Armstrong, P. W. et al. The 2004 ACC/AHA Guidelines: a perspective and adaptation for Canada by the Canadian Cardiovascular Society Working Group. Can. J. Cardiol. 20, 1075–1079 (2004). 194. Welsh, R. C., Travers, A., Huynh, T. & Cantor, W. J. Canadian Cardiovascular Society Working Group: Providing a perspective on the 2007 focused update of the American College of Cardiology and American Heart Association 2004 guidelines for the management of ST elevation myocardial infarction. Can. J. Cardiol. 25, 25–32 (2009). 195. Antman, E. M. et al. 2007 Focused Update of the ACC/AHA 2004 Guidelines for the Management of Patients With ST-Elevation Myocardial Infarction. J. Am. Coll. Cardiol. 51, 210–247 (2008). 196. Bolooki, H. M. & Askari, A. Acute Myocardial Infarction. Disease Management (2010). Available at: http://www.clevelandclinicmeded.com/medicalpubs/diseasemanagement/cardiology/acute-myocardial-infarction/. (Accessed: 23rd March 2016) 197. Percutaneous transluminal coronary angioplasty (PTCA) - Anatomy Video: MedlinePlus Medical Encyclopedia. Available at: http://www.nlm.nih.gov/medlineplus/ency/anatomyvideos/000096.htm. (Accessed: 22nd May 2015) 183  198. Heart disease - Percutaneous coronary intervention (PCI or angioplasty with stent) - Heart and Stroke Foundation of Canada. heartandstroke.ca Available at: http://www.heartandstroke.com/site/c.ikIQLcMWJtE/b.3831925/k.4F32/Heart_disease__Percutaneous_coronary_intervention_PCI_or_angioplasty_with_stent.htm. (Accessed: 22nd May 2015) 199. The Johns Hopkins University, The Johns Hopkins Hospital & Johns Hopkins Health System. Coronary Artery Bypass Graft Surgery (CABG). Health Library (2012). 200. What Is Cardiac Catheterization? - NHLBI, NIH. Available at: http://www.nhlbi.nih.gov/health/health-topics/topics/cath. (Accessed: 22nd March 2016) 201. Adams, H. P. et al. Guidelines for the Early Management of Patients With Ischemic Stroke A Scientific Statement From the Stroke Council of the American Stroke Association. Stroke 34, 1056–1083 (2003). 202. Adams, H., Adams, R., Zoppo, G. D. & Goldstein, L. B. Guidelines for the Early Management of Patients With Ischemic Stroke 2005 Guidelines Update A Scientific Statement From the Stroke Council of the American Heart Association/American Stroke Association. Stroke 36, 916–923 (2005). 203. Adams, H. P. et al. Guidelines for the Early Management of Adults With Ischemic Stroke A Guideline From the American Heart Association/American Stroke Association Stroke Council, Clinical Cardiology Council, Cardiovascular Radiology and Intervention Council, and the Atherosclerotic Peripheral Vascular Disease and Quality of Care Outcomes in Research Interdisciplinary Working Groups: The American Academy of Neurology affirms the value of this guideline as an educational tool for neurologists. Circulation 115, e478–e534 (2007). 184  204. Jauch, E. C. et al. Guidelines for the Early Management of Patients With Acute Ischemic Stroke A Guideline for Healthcare Professionals From the American Heart Association/American Stroke Association. Stroke 44, 870–947 (2013). 205. Canadian Stroke Network & Heart and Stroke Foundation of Canada. Canadian Best Practice Recommendations for Stroke Care: 2006. (2006). 206. Lindsay, P. et al. Canadian best practice recommendations for stroke care (updated 2008). Can. Med. Assoc. J. 179, S1–S25 (2008). 207. Lindsay, P. et al. Canadian best practice recommendations for stroke care (Update 2010). (On behalf of the Canadian Stroke Strategy Best Practices and Standards Writing Group, 2010). 208. The Canadian Stroke Best Practices and Standards Working Group. Canadian Best Practice Recommendations for Stroke Care (4th edition). (2013). 209. Health, M. of. Stroke & TIA - BC Guidelines. Available at: http://www2.gov.bc.ca/gov/topic.page?id=421363BE5B684D349E6304B9BCE9E4EE. (Accessed: 22nd May 2015) 210. Canadian Stroke Network. The Quality of Stroke Care in Canada. (2011). 211. Telestroke the next best thing to being there, study finds - Heart and Stroke Foundation of B.C. and Yukon. heartandstroke.ca Available at: http://www.heartandstroke.bc.ca/site/apps/nlnet/content2.aspx?c=kpIPKXOyFmG&b=7759923&ct=11251381. (Accessed: 20th August 2015) 212. Medical Services Commission. MSC Payment Schedule: Neurology. (British Columbia Ministry of Health, 2014). 185  213. Murray, M. J. The Canadian Triage and Acuity Scale: A Canadian perspective on emergency department triage. Emerg. Med. 15, 6–10 (2003). 214. Medical Services Commission. MSP tutor: Fee item 00081 - Emergency Care. (British Columbia Ministry of Health, 2014). 215. Spiegelhalter, D. J. Funnel plots for comparing institutional performance. Stat. Med. 24, 1185–1202 (2005). 216. Mant, J. Process versus outcome indicators in the assessment of quality of health care. Int. J. Qual. Health Care 13, 475–480 (2001). 217. Dai, J., Li, Z. & Rocke, D. Hierarchical Logistic Regression Modelling with SAS GLIMMIX. (2006). 218. Bureau of Health Information. Spotlight on measurement: 30-day mortality following hospitalisation: Considering approaches for ongoing reporting in NSW. 1–62 (Bureau of Health Information, 2015). 219. Katzan, I. L. et al. Risk adjustment of ischemic stroke outcomes for comparing hospital performance: a statement for healthcare professionals from the American Heart Association/American Stroke Association. Stroke J. Cereb. Circ. 45, 918–944 (2014). 220. Nicholl, J., West, J., Goodacre, S. & Turner, J. The relationship between distance to hospital and patient mortality in emergencies: an observational study. Emerg. Med. J. EMJ 24, 665–668 (2007). 221. Succi, M. J., Lee, S. Y. & Alexander, J. A. Effects of market position and competition on rural hospital closures. Health Serv. Res. 31, 679–699 (1997). 222. Mullner, R. M. & McNeil, D. Rural and urban hospital closures: a comparison. Health Aff. (Millwood) 5, 131–141 (1986). 186  223. Prior, M., Farmer, J., Godden, D. J. & Taylor, J. More than health: The added value of health services in remote Scotland and Australia. Health Place 16, 1136–1144 (2010). 224. Wagner, A. K., Soumerai, S. B., Zhang, F. & Ross-Degnan, D. Segmented regression analysis of interrupted time series studies in medication use research. J. Clin. Pharm. Ther. 27, 299–309 (2002). 225. Chowdhury, M. M., Dagash, H. & Pierro, A. A systematic review of the impact of volume of surgery and specialization on patient outcome. Br. J. Surg. 94, 145–161 (2007). 226. Waller, J. L., Brantley, V. C. & Podolsky, R. H. Where’s the Match? Matching Cases and Controls after Data Collection. in (WUSS, 2010). 227. Parsons, L. & Ovation Research Group. Performing a 1:N Case-Control Match on Propensity Score. Stat. Data Anal. 165, 228. Newgard, C. D. et al. Revisiting the ‘Golden Hour’: An Evaluation of Out-of-Hospital Time in Shock and Traumatic Brain Injury. Ann. Emerg. Med. 66, 30–41, 41–3 (2015). 229. Vanderschuren, M. & McKune, D. Emergency care facility access in rural areas within the golden hour?: Western Cape case study. Int. J. Health Geogr. 14, (2015). 230. Kotwal RS, Howard JT, Orman JA & et al. The effect of a golden hour policy on the morbidity and mortality of combat casualties. JAMA Surg. 151, 15–24 (2016). 231. Singh, R. & Mukhopadhyay, K. Survival analysis in clinical trials: Basics and must know areas. Perspect. Clin. Res. 2, 145–148 (2011). 232. in Cochrane Handbook for Systematic Reviews of Interventions (eds. Julian PT Higgins & Sally Green) (The Cochrane Collaboration, 2011). 233. Abbott, R. D. Logistic regression in survival analysis. Am. J. Epidemiol. 121, 465–471 (1985). 187  234. Appelros, P. Prediction of length of stay for stroke patients. Acta Neurol. Scand. 116, 15–19 (2007). 235. Canadian Institute for Health Information. Data Quality Documentation for External Users: Discharge Abstract Database, 2010-2011. (CIHI, 2011). 236. Evans, D. Trauma data question. (2015). 237. Tolin, D. F. & Foa, E. B. Sex differences in trauma and posttraumatic stress disorder: A quantitative review of 25 years of research. Psychol. Bull. 132, 959–992 (2006). 238. Hill, A., Pinto, R., Nathens, A. & Fowler, R. in C94. Understanding and overcoming the effects of health disparities on respiratory disease management A4953–A4953 (American Thoracic Society, 2013). 239. Concannon, T. W. et al. Elapsed Time in Emergency Medical Services for Patients with Cardiac Complaints: Are Some Patients at Greater Risk for Delay? Circ. Cardiovasc. Qual. Outcomes 2, 9–15 (2009). 240. Gomez, D. et al. Gender-associated differences in access to trauma center care: A population-based analysis. Surgery 152, 179–185 (2012). 241. Canto, J. G. et al. Use of emergency medical services in acute myocardial infarction and subsequent quality of care: observations from the National Registry of Myocardial Infarction 2. Circulation 106, 3018–3023 (2002). 242. Tataris, K., Kivlehan, S. & Govindarajan, P. National Trends in the Utilization of Emergency Medical Services for Acute Myocardial Infarction and Stroke. West. J. Emerg. Med. 15, 744–748 (2014). 188  243. Mercuri, M. et al. Providing optimal regional care for ST-segment elevation myocardial infarction: a prospective cohort study of patients in the Hamilton Niagara Haldimand Brant Local Health Integration Network. CMAJ Open 3, E1-7 (2015). 244. Goldberg, R. J., Kramer, D. G., Yarzebski, J., Lessard, D. & Gore, J. M. Prehospital Transport of Patients with Acute Myocardial Infarction: A Community-Wide Perspective. Heart Lung J. Crit. Care 37, 266–274 (2008). 245. Goldberg, R. J. et al. Community Trends in the Utilization and Characteristics of Persons With Acute Myocardial Infarction Transported by Emergency Medical Services. Heart Lung J. Crit. Care doi:10.1016/j.hrtlng.2012.02.007 246. Redfors, B. et al. Trends in Gender Differences in Cardiac Care and Outcome After Acute Myocardial Infarction in Western Sweden: A Report From the Swedish Web System for Enhancement of Evidence‐Based Care in Heart Disease Evaluated According to Recommended Therapies (SWEDEHEART). J. Am. Heart Assoc. 4, e001995 (2015). 247. Jneid, H. et al. Sex differences in medical care and early death after acute myocardial infarction. Circulation 118, 2803–2810 (2008). 248. Dreyer, R. P. et al. Evaluation of Gender Differences in Door-to-Balloon Time in ST-Elevation Myocardial Infarction. Heart Lung Circ. 22, 861–869 (2013). 249. Gopalakrishnan, P., Ragland, M. M. & Tak, T. Gender Differences in Coronary Artery Disease: Review of Diagnostic Challenges and Current Treatment. Postgrad. Med. 121, 60–68 (2009). 250. Ekundayo, O. J. et al. Patterns of Emergency Medical Services Use and Its Association With Timely Stroke Treatment Findings From Get With the Guidelines-Stroke. Circ. 189  Cardiovasc. Qual. Outcomes CIRCOUTCOMES.113.000089 (2013). doi:10.1161/CIRCOUTCOMES.113.000089 251. Adeoye, O. et al. Emergency Medical Services Utilization by Stroke Patients: A Population-Based Study. Am. J. Emerg. Med. 27, 141–145 (2009). 252. Joynt, K. E., Harris, Y., Orav, E. J. & Jha, A. K. Quality of Care and Patient Outcomes in Critical Access Hospitals. Jama 306, 45–52 (2011). 253. Votruba, M. E. & Cebul, R. D. Redirecting patients to improve stroke outcomes: implications of a volume-based approach in one urban market. Med. Care 44, 1129–1136 (2006). 254. Ayanian, J. Z. & Weissman, J. S. Teaching Hospitals and Quality of Care: A Review of the Literature. Milbank Q. 80, 569–593 (2002). 255. Pitches, D. W., Mohammed, M. A. & Lilford, R. J. What is the empirical evidence that hospitals with higher-risk adjusted mortality rates provide poorer quality care? A systematic review of the literature. BMC Health Serv. Res. 7, 91 (2007). 256. Ly, D. P., Jha, A. K. & Epstein, A. M. The association between hospital margins, quality of care, and closure or other change in operating status. J. Gen. Intern. Med. 26, 1291–1296 (2011). 257. Radcliff, T. A., Brasure, M., Moscovice, I. S. & Stensland, J. T. Understanding Rural Hospital Bypass Behavior. J. Rural Health 19, 252–259 (2003). 258. Liu, J. J., Bellamy, G. R. & McCormick, M. Patient bypass behavior and critical access hospitals: implications for patient retention. J. Rural Health Off. J. Am. Rural Health Assoc. Natl. Rural Health Care Assoc. 23, 17–24 (2007). 190  259. Morris, S. et al. Impact of centralising acute stroke services in English metropolitan areas on mortality and length of hospital stay: difference-in-differences analysis. BMJ 349, g4757 (2014). 260. Saver JL, Fonarow GC, Smith EE & et al. Time to treatment with intravenous tissue plasminogen activator and outcome from acute ischemic stroke. JAMA 309, 2480–2488 (2013). 261. Solhpour, A. et al. Comparison of Outcomes for Patients ≥75 Years of Age Treated With Pre-Hospital Reduced-Dose Fibrinolysis Followed by Percutaneous Coronary Intervention Versus Percutaneous Coronary Intervention Alone for Treatment of ST-Elevation Myocardial Infarction. Am. J. Cardiol. 113, 60–63 (2014). 262. Morrow, D. A. et al. Evaluation of the time saved byprehospital initiation of reteplase forST-elevation myocardial infarction: Results of the early retavase-thrombolysisin myocardial infarction (ER-TIMI) 19 trial. J. Am. Coll. Cardiol. 40, 71–77 (2002). 263. Morrison, L. J., Verbeek, P. R., McDonald, A. C., Sawadsky, B. V. & Cook, D. J. Mortality and prehospital thrombolysis for acute myocardial infarction: A meta-analysis. JAMA 283, 2686–2692 (2000). 264. Kapasi, H., Kelly, L. & Morgan, J. Thrombolysis in the air. Air-ambulance paramedics flying to remote communities treat patients before hospitalization. Can. Fam. Physician Médecin Fam. Can. 46, 1313–1319 (2000). 265. Ranasinghe, I., Barzi, F., Brieger, D. & Gallagher, M. Long-term mortality following interhospital transfer for acute myocardial infarction. Heart heartjnl-2014-306966 (2015). doi:10.1136/heartjnl-2014-306966 191  266. Steg, P. G. et al. ESC Guidelines for the management of acute myocardial infarction in patients presenting with ST-segment elevation. Eur. Heart J. 33, 2569–2619 (2012). 267. Gale, C. P., Laar, M. van & Hamm, C. Acute myocardial infarction and inter-hospital transfer. Heart heartjnl-2015-307508 (2015). doi:10.1136/heartjnl-2015-307508 268. Young, J. S. et al. Interhospital versus direct scene transfer of major trauma patients in a rural trauma system. Am. Surg. 64, 88-91-92 (1998). 269. Hill, A. D., Fowler, R. A. & Nathens, A. B. Impact of interhospital transfer on outcomes for trauma patients: a systematic review. J. Trauma 71, 1885–1900; discussion 1901 (2011). 270. Fuentes, B. et al. Futile Interhospital Transfer for Endovascular Treatment in Acute Ischemic Stroke: The Madrid Stroke Network Experience. Stroke J. Cereb. Circ. 46, 2156–2161 (2015). 271. Hess, D. C. & Audebert, H. J. The history and future of telestroke. Nat. Rev. Neurol. 9, 340–350 (2013). 272. Signorello, L. B. et al. Confounding by indication in epidemiologic studies of commonly used analgesics. Am. J. Ther. 9, 199–205 (2002). 273. Jones, J. M., Skaga, N. O., Sovik, S., Lossius, H. M. & Eken, T. Norwegian survival prediction model in trauma: modelling effects of anatomic injury, acute physiology, age, and co-morbidity. Acta Anaesthesiol. Scand. 58, 303–315 (2014). 274. Prusmack, C., Rochman, A. & Levi, A. The Effect of Age on Survival Following Traumatic Spinal Cord Injury. Top. Spinal Cord Inj. Rehabil. 12, 49–57 (2006). 275. Mosenthal, A. C. et al. The effect of age on functional outcome in mild traumatic brain injury: 6-month report of a prospective multicenter trial. J. Trauma 56, 1042–1048 (2004). 192  276. Morris, J. A., Mackenzie, E. J., Damiano, A. M. & Bass, S. M. Mortality in Trauma Patients: The Interaction between Host F... : Journal of Trauma and Acute Care Surgery. J Trauma 30, 1476–82 277. Dreyer, R. P. et al. Gender Differences in the Trajectory of Recovery in Health Status Among Young Patients With Acute Myocardial Infarction: Results From the VIRGO Study. Circulation CIRCULATIONAHA.114.014503 (2015). doi:10.1161/CIRCULATIONAHA.114.014503 278. Ontario Trauma Registry 2009 Report: Major Injury in Ontario (Includes 2008-2009 Data). (Canadian Institute for Health Information, 2010). 279. Kelly-Hayes, M. et al. Temporal Patterns of Stroke Onset The Framingham Study. Stroke 26, 1343–1347 (1995). 280. Rosenthal, M. B., Zaslavsky, A. & Newhouse, J. P. The Geographic Distribution of Physicians Revisited. Health Serv. Res. 40, 1931–1952 (2005). 281. Berke, E. M. & Shi, X. Computing travel time when the exact address is unknown: a comparison of point and polygon ZIP code approximation methods. Int. J. Health Geogr. 8, 23 (2009). 282. Infrastructure, M. of T. and. Highway Infrastructure Projects - Province of British Columbia. Available at: http://www2.gov.bc.ca/gov/content/transportation/transportation-infrastructure/projects. (Accessed: 8th January 2016) 283. Keeling, P., Hughes, D., Price, L., Shaw, S. & Barton, A. Safety and feasibility of prehospital thrombolysis carried out by paramedics. BMJ 327, 27–28 (2003). 193  284. Björklund, E. et al. Pre-hospital thrombolysis delivered by paramedics is associated with reduced time delay and mortality in ambulance-transported real-life patients with ST-elevation myocardial infarction. Eur. Heart J. 27, 1146–1152 (2006). 285. Gladstone, D. J. et al. A citywide prehospital protocol increases access to stroke thrombolysis in Toronto. Stroke J. Cereb. Circ. 40, 3841–3844 (2009). 286. Concannon, T. W. et al. Comparative Effectiveness of STEMI Regionalization Strategies. Circ. Cardiovasc. Qual. Outcomes 3, 506–513 (2010).     194  Appendix A: Condition codes used to calculate Charlson Comorbidity Index Comorbidities ICD-9-CM ICD-10 Myocardial infarct 410.x, 412.x I21.x, I22.x, I25.2 Congestive heart failure 398.91, 402.01, 402.11, 402.91, 404.01, 404.03, 404.11, 404.13, 404.91, 404.93, 425.4-425.9, 428.x I09.9, I11.0, I13.2, I25.5, I42.0, I42.5-I42.9, I43.x, I50.x, P29.0 Peripheral vascular disease 93.0, 437.3, 440.x, 441.x, 443.1-443.9, 447.1, 557.1, 557.9, V43.4 I70.x, I71.x, I73.1, I73.8, I73.9, I77.1, I79.0, I79.2, K55.1, K55.1, K55.8, K55.9, Z95.8, Z95.9 Cerebrovascular disease 362.34, 430.x-438.x G45.x, G46.x, H34.0, I60.x-I69.x Dementia 290.x, 294.1, 331.2 F00.x-F03.x, F05.1, G30.x, G31.1 Chronic pulmonary disease 416.8, 416.9, 490.x-505.x, 506.4, 508.1, 508.8 I27.8, I27.9, J40.x-J47.x, J60.x-J67.x, J68.4, J70.1, J70.3 Rheumatic disease 446.5, 710.0-710.4, 714.0-714.2, 714.8, 725.x M05.x, M06.x, M31.5, M32.x, M34.x, M35.1, M35.3, M36.0 Peptic ulcer disease 531.x-534.x K25.x-K28.x Mild liver disease 70.22, 70.23, 70.32, 70.33, 70.44, 70.54, 70.6, 70.9, 570.x, 571.x, 573.3, 573.4, 573.8, 573.9, V42.7 B18.x, K70.0-K70.3, K70.9, K71.3-K71.5, K71.7, K73.x, K74.x, K76.0, K76.2-K76.4, k76.8, K76.9, Z94.4 Diabetes without chronic complication 250.0-250.3, 250.8, 250.9 E10.0, E10.1, E10.6, E10.8, E10.9, E11.0, E11.1, E11.6, E11.8, E11.9, E12.0, E12.1, E12.6, E12.8, E12.9, E13.0, E13.1, E13.6, E13.8, E13.9 Diabetes with chronic complication 250.4-250.7 E10.2-E10.5, E10.7, E11.2, E11.5, E11.7, E12.2-E12.5, E12.7, E13.2-E13.5, E13.7, E14.2-E14.5, E14.7 Hemiplegia or paraplegia 334.1, 342.x, 344.0-344.6, 344.9 G04.1, G11.4, G80.1, G80.2, G81.x, G82.x, G83.0-G83.4, G83.9 Renal disease 403.01, 403.11, 403.91, 404.02, 404.03, 404.12, 404.13, 404.92, 404.93, 582.x, 583.0-583.7, 585.x, 586.x, 588.0, V42.0, V45.1, V56.x I12.0, I13.1, N03.2-N03.7, N05.2-N05.7, N18.x, N19.x, N25.0, Z49.0-Z49.2, Z94.0, Z99.2 Any malignancy (incl. lymphoma, leukemia, except malignant neoplasm of skin) 140.x-172.x, 174.x-195.8, 200.x-208.x, 238.6 C00.x-C26.x, C30.x-C34.x, C37.x-C41.x, C43.x, C45.x-C58.x, C60.x-C76.x, C81.x-C85.x, C88.x, C90.x-C97.x Moderate or severe liver disease 456.0-456.2, 572.2-572.8 I85.0, I85.9, I86.4, I98.2, K70.4, K71.1, K72.1, K72.9, K76.5, K76.6, K76.7 Metastatic solid tumor 196.x-199.x C77.x-C80.x AIDS/HIV 42.x-44.x B20.x-B22.x, B24.x Table A-1.Coding algorithms for Charlson Comorbidity Index140,276   195  Appendix B: Propensity score matched interrupted time series Chapter 7 describes the case control matching techniques used to create my interrupted time series models. As mentioned, I attempted to use propensity score matching to match cases with controls with similar probabilities of having received the intervention but did not. Unfortunately, there was very little overlap in probabilities of receiving the intervention between cases and controls. This led to very few matched cases and controls per quarter (25-60 events per group compared with 70-200 using the loosely matched method), and power problems. As the figure below shows, when dealing with a rare event outcome, such as 30-day mortality, and a small sample, the results of the segmented regression are prone to wild outliers and poor fit. For these reasons, the propensity score matching technique was abandoned in favour of the loose matching method which allowed me to compare cases pre- and post-intervention and mitigated concerns of historical events other than cessation of treatment at the local facility to account for any differences observed pre- vs. post-intervention.  Figure B-1. Propensity score matched interrupted time series, AMI, 30-day mortality rate   196  Appendix C: Results of parsimonious proportional hazards models Chapter 8 describes in detail the outcome of my survival analysis using full models. Here, I describe how I developed parsimonious models and their output to compare the differences between modeling techniques. C.1 Method I started by running a series of univariate survival tests in SAS to determine which covariates I would include in my multivariate regressions. Covariates that were statistically significant using a generous p-value (cut off p-value<0.20) were the first to be added to subsequent multivariate models. Following the results of the univariate analyses, I used stepwise modeling to build my models and both R and SAS to test the validity of my models. As in Chapter 8, I worked with Cox Proportional Hazards models censoring at 30 days after the event and used Schoenfeld and Martingale Residual plots within R to identify linearity and proportionality issues. Lastly, as part of the model building exercise, I used the AIC to distinguish between models.  C.2 Results Tables C-1 and C-2 show the results of my parsimonious models for trauma. Unlike the full models, time to care exhibits a dose response relationship with hazards of 30-day mortality in model 4 but that relationship disappears when working with true time to care instead of estimated, and restricting to patients who used paramedic services.   197  Model Parameter Hazard Ratio 95% CI P-value 4 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240 Time to care (mins) ≥ 240 Unknown time to care Appropriate care Poor quality In-hospital Air transport Combination Self-transport Inter-hospital transfer 45 < age (years) ≤ 75 Age > 75 Female Low comorbidity (CCI: 1-2) Moderate comorbidity (CCI: 3-4) High comorbidity (CCI: 5+) 1.189 1.247 1.158 1.102 1.532 0.865 1.151 16.902 0.990 0.522 0.653 1.477 1.421 2.947 0.635 1.152 1.539 2.237 (1.046, 1.351) (1.033, 1.505) (0.894, 1.501) (0.936, 1.298) (1.328, 1.767) (0.699, 1.071) (0.939, 1.410) (8.410, 33.969) (0.610, 1.609) (0.376, 0.725) (0.613, 0.696) (1.233, 1.770) (1.256, 1.606) (2.424, 3.588) (0.594, 0.678) (1.019, 1.302) (1.290, 1.836) (1.793, 2.791) 0.0079 0.0218 0.2671 0.2448 <0.0001 0.1834 0.1769 <0.0001 0.9686 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0237 Table C-1. Results of parsimonious multivariate Cox Proportional Hazards models using estimated time to care (1999-2013), trauma   198  Model Parameter Hazard Ratio 95% CI P-value 5 30 ≤ True time (mins) < 60 60 ≤ True time (mins) < 120 120 ≤ True time (mins) < 240 Appropriate care Inter-hospital transfer Air transport Combination In hospital 45 < age (years) ≤ 75 Age > 75 Female Low comorbidity (CCI: 1-2) Moderate comorbidity (CCI: 3-4) High comorbidity (CCI: 5+) 0.714 0.734 0.791 1.139 1.085 2.201 0.847 23.514 1.510 4.753 0.482 1.841 2.796 6.102 (0.397, 1.283) (0.487, 1.104) (0.567, 1.103) (0.946, 1.370) (0.951, 1.237) (1.309, 3.702) (0.445, 1.611) (9.617, 57.491) (1.220, 1.870) (3.871, 5.836) (0.435, 0.534) (1.605, 2.111) (2.200, 3.553) (4.502, 8.272) 0.2596 0.1373 0.1666 0.1711 0.2233 0.0029 0.6120 <0.0001 0.0002 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 5a 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240 Appropriate care Inter-hospital transfer Air transport Combination In hospital 45 < age (years) ≤ 75 Age > 75 Female Low comorbidity (CCI: 1-2) Moderate comorbidity (CCI: 3-4) High comorbidity (CCI: 5+) 1.118 1.023 1.188 1.131 1.030 1.991 0.715 8.671 1.517 4.823 0.484 1.839 2.799 6.095 (0.970, 1.454) (0.770, 1.360) (0.835, 1.691) (0.935, 1.368) (0.898, 1.182) (1.072, 3.698) (0.348, 1.469) (4.314, 17.427) (1.216, 1.893) (3.877, 6.000) (0.437, 0.535) (1.599, 2.114) (2.195, 3.570) (4.472, 8.307) 0.0934 0.8731 0.3379 0.2041 0.6747 0.0293 0.3611 <0.0001 0.0002 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 199  Model Parameter Hazard Ratio 95% CI P-value 5b 30 ≤ Time to care (mins) < 60 60 ≤ Time to care (mins) < 120 120 ≤ Time to care (mins) < 240 Time to care (mins) ≥ 240 Appropriate care Inter-hospital transfer Air transport Combination Self-transport In hospital 45 < age (years) ≤ 75 Age > 75 Female Low comorbidity (CCI: 1-2) Moderate comorbidity (CCI: 3-4) High comorbidity (CCI: 5+) 1.413 1.166 1.173 1.143 0.851 0.783 4.574 1.253 0.531 15.314 2.421 5.916 0.664 1.782 2.345 6.102 (1.273, 1.569) (0.916, 1.484) (0.984, 1.398) (1.007, 1.298) (0.761, 0.950) (0.680, 0.902) (3.112, 6.722) (0.965, 1.627) (0.482, 0.585) (8.179, 28.673) (1.814, 3.231) (3.903, 8.967) (0.643, 0.686) (1.578, 1.986) (2.199, 2.691) (4.502, 7.679) <0.0001 0.1873 0.3334 0.0998 <0.0001 <0.0001 <0.0001 0.4013 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 0.0043 <0.0001 <0.0001 Table C-2. Results of multivariate Cox Proportional Hazards models using ambulance data (5), estimated time to care for patients with ambulance data (5a), and all admissions between 2009 and 2013 (5b), trauma   

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0320896/manifest

Comment

Related Items