You may notice some images loading slow across the Open Collections website. Thank you for your patience as we rebuild the cache to make images load faster.

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

The utility of atrial fibrillation clinical registry data in the modelling of change in patient-reported… Kwon, Jae-Yung 2020

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

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
24-ubc_2020_november_kwon_jae-yung.pdf [ 3.23MB ]
Metadata
JSON: 24-1.0390462.json
JSON-LD: 24-1.0390462-ld.json
RDF/XML (Pretty): 24-1.0390462-rdf.xml
RDF/JSON: 24-1.0390462-rdf.json
Turtle: 24-1.0390462-turtle.txt
N-Triples: 24-1.0390462-rdf-ntriples.txt
Original Record: 24-1.0390462-source.json
Full Text
24-1.0390462-fulltext.txt
Citation
24-1.0390462.ris

Full Text

THE UTILITY OF ATRIAL FIBRILLATION CLINICAL REGISTRY DATA IN THE MODELLING OF CHANGE IN PATIENT-REPORTED OUTCOMES by Jae-Yung Kwon  M.S.N., The University of British Columbia, 2014 B.S.N., The University of British Columbia, 2012 B.H.K., The University of British Columbia, 2010  A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF  DOCTOR OF PHILOSOPHY in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Nursing)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)   April 2020  © Jae-Yung Kwon, 2020 ii  The following individuals certify that they have read, and recommend to the Faculty of Graduate and Postdoctoral Studies for acceptance, the dissertation entitled:  The Utility of Atrial Fibrillation Clinical Registry Data in the Modelling of Change in Patient-Reported Outcomes  submitted by Jae-Yung Kwon in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Nursing  Examining Committee: Dr. Jennifer Baumbusch, Associate Professor, School of Nursing, UBC Co-supervisor Dr. Richard Sawatzky, Professor, School of Nursing, Trinity Western University Co-supervisor  Dr. Jason Sutherland, Professor, School of Population and Public Health, UBC University Examiner Dr. Bernie Garrett, Associate Professor, School of Nursing, UBC  University Examiner  Additional Supervisory Committee Members: Dr. Sandra Lauck, Clinical Associate Professor, School of Nursing, UBC Supervisory Committee Member Dr. Pamela Ratner, Vice-Provost and Associate Vice-President, UBC Supervisory Committee Member     iii  Abstract Patient-reported data collected in clinical registries are increasingly playing a role in improving patients’ outcomes by generating real-world information about the impact of treatment from patients’ perspectives. However, there is lack of methodological guidance on how to use these data to inform practice. Thus, the purpose of this study was to explore the analytical potential of patient-reported data stored in clinical registries with the aim of explicating the process of conducting such analyses and evaluating their utility. We examined patient-reported outcome trajectories of patients with atrial fibrillation in specialized multidisciplinary clinics and factors that predicted their different trajectories. The theoretical perspective that informed this study was the revised Wilson and Cleary framework using the Atrial Fibrillation Effect on QualiTy-of-Life (AFEQT) Questionnaire as the outcome. A retrospective analysis of provincial health registry data, collected by Cardiac Services BC, between 2008 and 2016, linked with administrative health data from Population Data BC was conducted using growth mixture models. We found that rather than a single health outcome trajectory, there were three different trajectories, with different patient characteristics (i.e., differences in age, gender, stroke risk score, and when ablation or anticoagulant therapy was received) that could inform tailored interventions. However, the findings highlighted the need for substantial transparency in the methods applied by the data stewards and researchers, including the need for well-defined data preparation and modelling specifications when conducting registry-based studies.  To maximize the use of registry data, clinicians must share the responsibility in understanding informed consent and confidentiality, confirming data accuracy and structure, and selecting variables and cohorts of interest. Key questions in their use include: (1) What is the iv  purpose of the registry and the rationale for its development? (2) What is the process and the context in which the registry data have been collected? and (3) What procedures have been applied to correct for sources of data error? Answering these questions will bring greater alignment between the quality of the patient-reported data, the questions they can help answer, and their intended purpose to integrate patients’ perspectives in registries and to effectively support patient-centred care.    v  Lay Summary Patient-reported data stored in registries maintained by clinicians play an important role in improving patients’ outcomes. These data are defined as patients’ own reports of their health conditions. However, there is lack of practical guidance on how clinicians and researchers can analyze the data kept in these registries. This study examined how these data can be best used to inform patient care. The data collected from patients with atrial fibrillation were studied as an example.  Using advanced growth mixture models, we found that not all health trajectories are the same and using this model can help to tailor interventions. However, model specifications and data quality must be carefully considered when clinicians interpret findings. Analysis of registry data requires a range of skills. To optimize their value, data quality, the research questions posed, and the purpose of the registry should align to generate meaningful information that best supports patient-centred care.      vi  Preface This doctoral dissertation is the original, independent work of the author, J.-Y. Kwon. However, the pronoun “we” and “us” was used to recognize that scientific pursuit is very much a social enterprise influenced by a culmination of discussion from people with different expertise and background. All research activities were approved by the University of British Columbia Behavioural Research Ethics Board (Certificate H16-03439). The members of the supervisory committee, Dr. J. L. Baumbusch, Dr. R. Sawatzky, Dr. S. Lauck, and Dr. P. A. Ratner provided guidance with the development of the research project, including the purpose, data analysis and writing of the dissertation.   All inferences, opinions, and conclusions drawn in this dissertation are those of the author, and do not reflect the opinions or policies of the Data Stewards.   vii  Table of Contents Abstract ......................................................................................................................................... iii Lay Summary .................................................................................................................................v Preface ........................................................................................................................................... vi Table of Contents ........................................................................................................................ vii List of Tables ..................................................................................................................................x List of Figures ............................................................................................................................... xi List of Abbreviations ................................................................................................................. xiii Glossary of Terms ...................................................................................................................... xvi Acknowledgements ................................................................................................................... xvii Dedication ................................................................................................................................. xviii Chapter 1: Background .................................................................................................................1 1.1 Patient-Reported Outcomes Measurement ..................................................................... 3 1.2 Use of Registries in the Management of Atrial Fibrillation ........................................... 6 1.3 Purpose and Significance of the Study ........................................................................... 9 1.4 Research Question ........................................................................................................ 10 1.5 Summary ....................................................................................................................... 11 Chapter 2: Literature Review .....................................................................................................13 2.1 Registries and Barriers to their Use .............................................................................. 13 2.2 Patient-Reported Outcomes .......................................................................................... 19 2.3 Atrial Fibrillation .......................................................................................................... 22 2.4 Trajectories of Change in Atrial Fibrillation ................................................................ 27 2.5 Selection of a Patient-Reported Outcome Measure ...................................................... 29 2.6 The Conceptual Frameworks ........................................................................................ 33 2.7 Summary ....................................................................................................................... 39 Chapter 3: Methods .....................................................................................................................40 3.1 Study Design and Setting .............................................................................................. 41 viii  3.2 AF Clinic Processes ...................................................................................................... 43 3.3 Data Sources and Variables .......................................................................................... 49 3.3.1 Atrial fibrillation clinic registry database. ................................................................ 49 3.3.2 Selection of variables and instruments. .................................................................... 53 3.3.3 Ethical considerations. .............................................................................................. 62 3.4 Information that Can be Extracted from Linked Data Sources ..................................... 64 3.5 Accommodation for Missing Data ................................................................................ 77 3.6 Data Analysis Strategy .................................................................................................. 89 3.6.1 How to best analyze longitudinal PROMs data. ....................................................... 90 3.6.2 How to best represent variability in the frequency and timing of measurement occurrences............................................................................................................................ 99 3.6.3 How to best represent the shape of the trajectories. ................................................ 103 3.6.4 How to best represent individual differences in trajectories. .................................. 107 3.6.5 How to identify factors that explain variability in individual trajectories. ............. 113 3.7 Summary ..................................................................................................................... 118 Chapter 4: Results......................................................................................................................121 4.1 Study Sample .............................................................................................................. 121 4.2 Model Selection .......................................................................................................... 127 4.3 Predictors of Latent Class Membership ...................................................................... 131 4.4 Summary ..................................................................................................................... 135 Chapter 5: Discussion ................................................................................................................137 5.1 Descriptive Results with Some Practical Implications ............................................... 137 5.2 Results Pertaining to Methods Applied to Analyze PROMs with Registry Data ....... 139 5.3 Results Pertaining to the Conceptual Framework ....................................................... 145 5.4 Strengths and Limitations ........................................................................................... 149 5.4.1 Strengths of the study.............................................................................................. 149 5.4.2 Limitations of the study. ......................................................................................... 152 5.5 Recommendations for Nursing ................................................................................... 155 ix  5.5.1 Understanding informed consent and confidentiality. ............................................ 156 5.5.2 Checking data accuracy. ......................................................................................... 160 5.5.3 Selecting variables and cohorts of interest.............................................................. 163 5.5.4 Transforming data structures. ................................................................................. 166 5.5.5 Key guiding questions............................................................................................. 170 Chapter 6: Conclusion ...............................................................................................................177 6.1 Overview ..................................................................................................................... 177 6.2 Future Research .......................................................................................................... 179 6.3 Conclusion .................................................................................................................. 181 References ...................................................................................................................................183 Appendices ..................................................................................................................................232  AFEQT Questionnaire ...................................................................................... 232  List of Comorbidities and Treatment Algorithms ............................................. 234  SAS code for calculating the Charlson Comorbidity Index .............................. 239  Characteristics of Variables with Missing Data ................................................ 244  Characteristics of Respondents and Non-Respondents during Study Period  (2008-2016) ............................................................................................................................ 245  Auxiliary Variables ............................................................................................ 247  Characteristics of Respondents (resp) and Non-Respondents (non-resp) at Baseline (N = 7,439) and during Follow-Up (N = 4,412) ...................................................... 248  Model Syntax Mplus ......................................................................................... 252  x  List of Tables Table 2-1. Atrial Fibrillation Symptoms and their Effects on Health and Quality of Life ........... 24 Table 2-2. Summary of Patient-Reported Outcome Measures in AF ........................................... 26 Table 3-1. Post-Ablation Pathway ................................................................................................ 48 Table 3-2. Time Intervals between Repeated AFEQT Questionnaire Completions ..................... 76 Table 3-3. Comparisons between Longitudinal Data Analysis Methods ...................................... 97 Table 3-4. Likelihood Statistics for Models of Change with Four Time Points ......................... 108 Table 3-5. Descriptive Statistics of AFEQT Questionnaire Scores) by Follow-Up Visit  (T0-T10) ...................................................................................................................................... 109  Table 3-6. Unadjusted Odds Ratios for Predictors of Latent Class Membership ...................... 115  Table 4-1. Characteristics of Patients (N = 7,439) .................................................................... 122 Table 4-2. Number of AFEQT Questionnaires Completed by Patients’ Characteristics ........... 125 Table 4-3. Mean Likelihood and Information Criteria for the Growth Mixture Models ............ 127 Table 4-4. Mean Parameter Estimates of the Three-Class Restricted Standard Model ............. 129 Table 4-5. Relative Frequencies by Class with Odds Ratios for Predictors of Latent Class Membership (N=7,439) .............................................................................................................. 132    Table 4-6. Analysis of Predictive Values of the Growth Mixture Model with Predictors .......... 135    xi  List of Figures Figure 2-1. Wilson and Cleary Model. ......................................................................................... 33 Figure 2-2. Ferrans et al.'s (2005) Revised Wilson and Cleary Model. ....................................... 35 Figure 2-3. Variables and Instruments Associated with the Conceptual Framework .................. 36 Figure 3-1. Location of Atrial Fibrillation Clinics in BC ............................................................. 42 Figure 3-2. Patient Journey through the Atrial Fibrillation Clinic ............................................... 45 Figure 3-3. General Treatment Pathway ...................................................................................... 46 Figure 3-4. Medication Pathway .................................................................................................. 47 Figure 3-5. Canadian Cardiovascular Society Severity in Atrial Fibrillation Scale ..................... 60 Figure 3-6. Description of Analytic Variables ............................................................................. 62 Figure 3-7. Flowchart of Eligible Study Cohort ........................................................................... 65 Figure 3-8. Combining Multiple Datasets into One Dataset ........................................................ 67 Figure 3-9. Identifying Comorbidities in the MSP and DAD Datasets ........................................ 68 Figure 3-10. Matching PharmaNet Datasets ................................................................................ 70 Figure 3-11. Assigning the Eligible Study Cohort to the Anticoagulant Variable ....................... 71 Figure 3-12. Flowchart of the AFEQT Questionnaire Dataset ..................................................... 72 Figure 3-13. Number of First AFEQT Questionnaires Completed by Time after  Initial Consultation (N=4,412) ...................................................................................................... 73 Figure 3-14. Number of First AFEQT Questionnaires Completed by Quarter  and by Clinic (N=4,412) ............................................................................................................... 74 Figure 3-15. Number of AFEQT Questionnaires Completed by 40 Randomly  Selected Patients by Time (in years) ............................................................................................. 75 Figure 3-16. Number of Ablations Performed Following Initial Consultation  by Time (in Years) ........................................................................................................................ 77 Figure 3-17. Frequency of Missing Data Patterns ........................................................................ 80 Figure 3-18. Distribution of the AFEQT Questionnaire Items..................................................... 82 xii  Figure 3-19. Density Plots of Covariates with Imputed Values (N = 7,439) ............................... 87 Figure 3-20. Density Plots of AFEQT Questionnaire Items with Imputed Values ...................... 88 Figure 3-21. Latent Growth Model with Three Measurement Occasions  of a Continous Outcome ............................................................................................................... 95 Figure 3-22. Smooth Fitting Curve of the AFEQT Questionnaire Scores  by Number of Follow-up Visits .................................................................................................. 100 Figure 3-23. Smooth Fitting Curve of the AFEQT Questionnaire Scores  with the Number of AFEQT Questionnaire Administrations by Time....................................... 101 Figure 3-24. AFEQT Questionnaire Scores by Follow-Up ........................................................ 104 Figure 3-25. Growth Mixture Model with Three Continuous Outcomes ................................... 105 Figure 3-26. A Summary of the Key Challenges in Analyzing PROMs Data  Stored in a Clinical Registry ....................................................................................................... 119 Figure 4-1. Three-Class Trajectory Model (AFEQT Questionnaire Scores). ............................ 130 Figure 5-1. Suggested Revisions to the Conceptual Framework. .............................................. 145 Figure 5-2. Matching Two Different Datasets ........................................................................... 161 Figure 5-3. ICD-10 Coding Structure ......................................................................................... 164 Figure 5-4. Transforming Data from a Long to Wide Format ................................................... 167 Figure 5-5. An Example of Inconsistent Imputation with the Long Format Data Structure ...... 168    xiii  List of Abbreviations AF    Atrial fibrillation  AFEQT  Atrial Fibrillation Effect on QualiTy-of-life  AFSS   Atrial Fibrillation Severity Scale AIC   Akaike’s information criterion ANOVA  Analysis of variance AR1   Autoregressive structure BIC   Bayesian information criterion CCI   Charlson Comorbidity Index CCS-SAF  Canadian Cardiovascular Society Severity in Atrial Fibrillation Scale  CHADS2  Stroke risk score CIHR   Canadian Institutes of Health Research ClinRO  Clinician-reported outcome DAD   Discharge Abstracts Database DIN   Drug identification number EP   Electrophysiologist EQ-5D   EuroQol FIML   Full information maximum likelihood GAD-7  Generalized Anxiety Disorder GCM   Growth curve model GMM   Growth mixture model HLM   Hierarchical linear model ICD   International Classification of Diseases xiv  LCGA   Latent class growth analysis LGM   Latent growth model MAR   Missing at random MCAR  Missing completely at random MICE   Multivariate imputation by chained equations MID   Minimal important difference MLM   Multilevel modelling  MNAR  Missing not at random MSP   Medical Services Plan NP   Nurse practitioner ObsRO  Observer-reported outcome PerfO   Performance outcome PHN   Personal Health Number PRO   Patient-reported outcome PROMs  Patient-reported outcome measures QOL   Quality of life SABIC  Sample-size adjusted BIC SCL   Symptom checklist SeiQOL  Schedule for the Evaluation of Quality of Life SEM   Structural equation modelling SF-36   Short-Form Health Survey SRE   Secure Research Environment T0   Time 0 – Initial consultation xv  T1   Time 1 – 1st follow-up visit (individually-varying) T2   Time 2 – 2nd follow-up visit (individually-varying) T3   Time 3 – 3rd follow-up visit (individually-varying) TCPS2   Tri-Council Policy Statement xvi  Glossary of Terms Clinical registry   A computer database that collects patient information as part of everyday care. Entropy Indicator of accuracy of latent class assignment when individuals are assigned to subgroups of trajectories.  Full information maximum likelihood (FIML) Missing data technique that uses all information from variables included in the statistical model to compute parameter estimates. Growth curve model (GCM)  A wide array of models that estimate the between-person differences in within-person change. Growth mixture model (GMM) An extension of the growth curve model that assigns individuals who share similar patterns of scores into unobserved subgroups called latent classes and accounts for individually-varying times of observations.  Latent class growth analysis (LCGA) The most parsimonious growth mixture model with the intercept and slope variances constrained to be equal within each latent class.  Latent growth model (LGM)  A growth curve model within the structural equation modelling framework with the intercept and slope represented as latent or unobserved variables. Multilevel modelling (MLM) A growth curve model within the multilevel framework with the intercept and slope represented as regression coefficients that correspond to more than one level. For example, repeated measurement occurrences (level 1 units) nested within subjects (level 2 units) in longitudinal studies. Multilevel multiple imputation  Missing data technique that creates multiple copies of a dataset with missing values replaced by imputed values while accounting for multilevel structure in a dataset (level 1 and 2 units). Patient-reported outcome measures (PROMs) Multi-item questionnaires administered to patients to assess their perspectives of their health and quality of life.   xvii  Acknowledgements This project was a culmination of support from many people. First, I would like to thank my doctoral co-supervisors, Dr. Jennifer Baumbusch and Dr. Richard Sawatzky. Jennifer, thank you for your support and encouragement through all the ups and downs and teaching me to be persistent in the creative process of scholarship. Rick, this project would not have been possible without your methodological expertise. Your mentorship at each step has been an inspiration and has influenced me in all areas of scholarship. Thank you for opening so many doors of knowledge and opportunity. I would also like to thank Dr. Sandra Lauck and Dr. Pamela Ratner, my committee members. Sandra, thank you for helping to navigate the clinical context and sharing your insights. Pam, thank you for sharing your wisdom. You have been the foundation that brought such dynamic and leading scholars to the same table to make this whole project “greater than the sum of its parts”.   I would like to acknowledge the funding support from the University of British Columbia, the Lyle Creelman Endowment Fund, and the Canadian Nurses Foundation. I would also like to acknowledge Cardiac Services BC for providing the registry data and Population Data BC for coordinating the project and for their student cost waiver program.   Lastly, I would like to thank my family and colleagues as well as my fellow PhD cohort (Lillian, Chantelle, Melissa, and Bubli) who shared this learning journey with me.  xviii  Dedication To my mother who taught me the true meaning of unconditional love.   Providing resources for capable and committed individuals to follow their instincts leads to the best of all discoveries, the totally surprising, the most unexpected, and the most useful ones. - Michael Smith, Nobel Laureate                  1  Chapter 1: Background Patient-reported data in clinical registries are increasingly being discussed in health care as having tremendous potential to improve health outcomes by generating real-world data in everyday practice shaped by patients’ perspectives (Daugherty et al., 2018; McDonald, Malcolm, Ramagopalan, & Syrad, 2019; Rivera et al., 2019). However, there is a gap in knowledge about how best to analyze these data to inform practice. The purpose of this study was to address this gap by exploring the analytical potential of patient-reported data stored in clinical registries, with the aim to explicate the process of conducting such an analysis of data from patients with atrial fibrillation in specialized clinics by answering the research question: Can the trajectories of change in patient-reported outcome measures for outpatients with atrial fibrillation, after initial consultation, be explained by different biological functioning or individual and environmental characteristics? Without understanding the underlying processes of analyzing these data, important insights from practice will not be fully utilized to truly support patient-centred care. To provide an understanding of the current shift towards patient-centredness in health care, we contrast the underlying philosophical context of evidence-based medicine to patient-centred care before describing patient-reported outcomes and the use of registries in the forthcoming sections.   Since its introduction in the early 1990s, the concept and practice of evidence-based medicine has emerged as a dominant paradigm in health care (Smith & Rennie, 2014). In general, evidence-based medicine emphasizes the use of evidence, primarily from clinical research, in making decisions about patient care (Sackett, Rosenberg, Gray, Haynes, & Richardson, 1996). Over several decades, evidence-based medicine has underpinned thousands of clinical guidelines or protocols designed to provide recommendations to clinicians that can be applied across different situations (Weaver, 2015). This approach, however, has been criticized 2  for being based on a “positivistic” orientation that overlooks important factors from patients’ points of view (Greenhalgh, Howick, & Maskrey, 2014; Thorne & Sawatzky, 2014). The unintended consequence of adopting a positivistic approach is the complexity in applying objectively-derived knowledge from clinical trials (considered as the “gold standard”) to particular situations (Thorne & Sawatzky, 2014). For example, nurses may be inclined to attend only to domains that already have a blueprint for providing care (e.g., treatment of a disease) and may simultaneously ignore less apparent domains that could contribute more significantly to a patient’s health and quality of life (e.g., a psychosocial aspect). The adoption of a positivistic orientation may inadvertently devalue the type of knowledge that should be taken as evidence (e.g., knowledge derived from a patient’s perspective) and also neglects the uniqueness and the complexities of the individual – an idea that is at the heart of nursing’s disciplinary epistemology (Thorne & Sawatzky, 2014).   Increasingly, there has been a shift towards patient-centred care as a response to the noted limitations of evidence-based medicine to incorporate individual circumstances and preferences into healthcare interactions (Greenhalgh et al., 2014). This shift has been perpetuated by other contextual factors, which include economic constraints on health care, reduced length of hospital stays, and the shifting roles of patients and families as they have become more active, informed, and influential (Bauman, Fardy, & Harris, 2003; Carman et al., 2013). As a result, the Institute of Medicine (2001) endorsed patient-centred care as one of six aims for health system improvement in the landmark report, “Crossing the Quality Chasm.” More recently and locally, the British Columbia (BC) Ministry of Health (2014) articulated the need to shift the culture of health care from being disease-centred and provider-focused to being patient-centred as the first of eight priorities for the healthcare system. Although there is a range of definitions available, patient-3  centred care is typically defined as a “partnership among practitioners, patients, and their families (when appropriate) to ensure that decisions respect patients’ wants, needs, and preferences and that patients have the education and support they need to make decisions and participate in their own care” (Institute of Medicine, 2001, p. 7). In contrast to the positivistic approach, patient-centred care is said to be grounded in a constructionist orientation that recognizes the personal, subjective experiences of individuals (Lincoln, Lynham, & Guba, 2011).  The implicit assumption of patient-centred care is that the patient is the best source of knowledge (Epstein & Street, 2011). The unintended consequence of adopting this approach is that nurses may inadvertently privilege subjectively derived knowledge and devalue other types of knowledge (e.g., empirical knowledge) (Epstein & Street, 2011; Romana, 2006; Thorne & Sawatzky, 2014). Thus, there is growing recognition that the best care should attempt to improve health outcomes for individual patients by relying on multiple sources of knowledge that take into account patients’ preferences in the healthcare decision-making process (Miller, Steele Gray, Kuluski, & Cott, 2015; Sacristán, 2013). Although there are different types of patient-reported data (e.g., patient-reported experiences), we focused on patient-reported outcomes as they are primarily used as clinical endpoints for healthcare services and interventions (Kingsley & Patel, 2017).  1.1 Patient-Reported Outcomes Measurement  The history of patient-reported outcomes (PROs), or measurement of health and quality of life in health research, can be traced back to a movement that originated in the 1970s, designed to give the perspective of patients a much needed voice in the face of the traditional dominance of biomedical approaches to care (Bowling, 2001). This movement helped to bring to the fore the importance of understanding the impact of illness and healthcare services on 4  people’s daily lives, including taking into account perspectives regarding individuals’ symptoms, functional status, and physical, social, and emotional well-being (Weldring & Smith, 2013). With increasing emphasis on patient-centred care by healthcare organizations and funding bodies, PROs have gained renewed attention as a critical component to better respond to the needs of patients and to improve quality of care (Black, 2013; Epstein, Fiscella, Lesser, & Stange, 2010; Weldring & Smith, 2013). Over the past decade, provinces and territories in Canada have further emphasized this importance by including PROs in their strategic priorities and directions (Canadian Institute for Health Information, 2017).  Although there are different ways to collect data on PROs, one way is through the use of patient-reported outcome measures (PROMs), which are multi-item questionnaires administered to patients to assess their perspectives of their health and quality of life (Deshpande, Rajan, Sudeepthi, & Abdul Nazir, 2011). There are three types of PROMs: generic, condition-specific, and individualized measures. Generic measures are used for the general population regardless of the disease or condition, which allows comparisons of outcomes with other groups. However, because these measures are designed to be broadly applicable rather than specific, they tend not to provide a sufficient level of detail or response to condition-specific treatment effects. Condition-specific measures focus on specific or unique manifestations (e.g., particular symptoms) of a given condition and its treatment. These measures are designed to be more sensitive and responsive to changes in patients with a particular condition. In contrast to the previous two types, individualized measures allow each patient to individually define the domains and weights to be assessed, which may create a unique measure but make it difficult to compare across individuals or groups. While PROMs of all types are increasingly being integrated into clinical practice, there has been tremendous growth in the development of 5  condition-specific PROMs to systematically assess symptom-based conditions (Weldring & Smith, 2013). Since there often has been a mismatch between objectively and subjectively derived outcome measures (e.g., clinical indicators versus self-reported health) (Chassany, Le-Jeunne, Duracinsky, Schwalm, & Mathieu, 2006; Flores et al., 2012; Janssen et al., 2016), the focus on PROMs can be useful in supporting patient-centred care in nursing. For example, nurses can use these measures to ask patients whether the treatment that they received helped to improve outcomes that were important to them and to identify unmet needs. Despite the potential to enhance patient-centred care, PROMs have mostly been used in clinical trials and research settings, which have been often criticized as not reflecting the patient population encountered in clinical practice (Blonde, Khunti, Harris, Meizinger, & Skolnik, 2018; Heneghan, Goldacre, & Mahtani, 2017).  There is, however, a source of information from a much broader and more representative patient population that realistically represents what occurs in the actual clinical environment that could complement information derived from clinical trials (Gliklich, Dreyer, & Leavy, 2014). Registries — “an organized system that uses observational study methods to collect uniform data (clinical and other) to evaluate specified outcomes for a population defined by a particular disease, condition, or exposure, and that serves one or more predetermined scientific, clinical, or policy purposes” — are seen as powerful sources of “real-world” data because they do not have restrictive exclusion criteria or pre-specified protocols (Gliklich et al., 2014, para. 8). The inclusion of PROMs in registries has the potential to provide additional insights about outcomes during actual practice from the patient’s perspective which may be of relevance to various stakeholders including decision makers, researchers, and clinicians in supporting patient-centred 6  care (Calkins et al., 2018; Norekvål, Fålun, & Fridlund, 2016; Rumsfeld et al., 2013). At the individual level, PROMs can be used to monitor the patient’s progress over time, guide shared decision-making, and facilitate communication between clinicians and the patient or family (Benzimra et al., 2018; Van Der Wees et al., 2014). At the aggregate level, PROMs can be used to compare groups or treatments to guide quality improvement, health system management, and payment policies (Forsberg et al., 2015; Segal, Holve, & Sabharwal, 2013; Van Der Wees et al., 2014). PROMs are also essential for patient-centred outcomes research that can serve as a bridge between facilitating healthcare interactions at the individual level and quality healthcare systems that align with the perspectives, interests, and values of patients (Largent et al., 2018; Patient-Centered Outcomes Research Institute, 2019).  Since registries have many potential uses in addressing a range of questions, nurses have a critical opportunity to ensure that registry data integrate the patient “voice” and provide relevant and meaningful information to better respond to the needs of patients. For example, using PROMs as tools, nurses can better support patient-centred care by raising awareness of problems or issues that are of concern to the patient at a single point in time or over time. As the measurement and reporting of health outcomes in Canadian health services shift towards the integration of PROMs in routine practice, nurses will need to be equipped with knowledge and skills necessary to make best use of these measures to not only maximize their benefits in advancing patient-centred care but also to begin participating in and influencing broader contextual issues (e.g., clinical, policy, and funding decisions) that shape healthcare delivery.  1.2 Use of Registries in the Management of Atrial Fibrillation  While there are registries for many other populations, we focused on patients with atrial fibrillation (AF) because essential questions regarding the clinical course, risks, and management 7  of AF in everyday practice remain unanswered with clinical trials data (Lip, Al-Khatib, et al., 2014). In the context of AF, many clinical registries exist with considerable variety in their design and methodology (Mazurek, Huisman, & Lip, 2017). There are national registries (e.g., Swedish and Danish National Patient Registries) where every patient who is diagnosed with AF is enrolled, and international registries that were initiated by the European Society of Cardiology (e.g., EURObservational Research Programme Atrial Fibrillation General Pilot Registry) (Lip, Laroche, et al., 2014; Ludvigsson et al., 2011; Schmidt et al., 2015). Some registries are industry-sponsored or -funded registries, such as the Global Registry on Long-Term Oral Antithrombotic Treatment in Patients with Atrial Fibrillation (GLORIA-AF) and the Global Anticoagulant Registry in the FIELD – Atrial Fibrillation (GARFIELD-AF) (Huisman et al., 2014; Kakkar et al., 2012). A few registries enrol only outpatients, such as the Outcomes Registry for Better Informed Treatment of Atrial Fibrillation (ORBIT-AF), whereas others include only inpatients such as the Get With the Guidelines-AFIB Registry (Lewis et al., 2014; Piccini et al., 2011). Many of these registries have recently been initiated because of improvements in AF diagnosis and management (Mazurek et al., 2017). However, the predominant focus of these registries is on monitoring quality using select indicators that are reported within and between sites, which do not necessarily reflect important outcomes from patients’ perspectives.  The routine use of PROMs is particularly important in the context of AF (Calkins et al., 2019). For example, AF is known to cause a range of symptoms that can change over time (Rienstra et al., 2012). Some patients may not be aware that their hearts are fibrillating until some time has passed, whereas others may be immediately aware of changes in their heart’s rhythm. Symptoms associated with AF can range from mild fatigue to shortness of breath and 8  chest pain. Although living with AF does not necessarily equate with having symptoms (e.g., some arrhythmias have no associated symptoms), it is important to monitor changes in patients’ experiences to ensure appropriate treatment. Among the many types of arrythmias (e.g., bradycardia and atrial flutter), AF is the most frequently encountered in clinical practice, affecting 2% of the general adult population in North America (Zoni-Berisso, Lercari, Carazza, & Domenicucci, 2014), and up to 18% of people over 80 years of age (Heeringa et al., 2006). In the past decade, the management of AF has become increasingly complex, compounded by advances in diagnostic and therapeutic procedures, multifaceted clinical presentations due in part to the aging population, polypharmacy, and variable side-effects of treatment (Calkins et al., 2019; Chugh, Roth, Gillum, & Mensah, 2014; Fumagalli et al., 2016). Since AF is a chronic condition that needs to be managed in this ever-changing context (McCabe, 2011), the routine collection of PROMs may help clinicians evaluate patients’ responses to treatment and tailor interventions (Black, 2013). For example, clinicians may have questions about the type of treatment strategies that should be developed for different groups of patients, and their likely impact on various expected changes with treatment. These questions can be answered by routinely collecting PROMs and describing “individual” trajectories based on changes to PROMs scores over time. Understanding trajectories of change over time allows for the identification of those at greatest risk for adverse trajectories and events, identification of factors that influence change in health status, and examination of possible time points in which interventions can be targeted (Henly, Wyman, & Findorff, 2011). In addition, such an approach can also allow the subsequent identification of low-risk patients who might need less attention. Some patients who are diagnosed with AF, for example, may report relatively low PROMs scores initially and gradually improve, indicating that they may need more focused education and 9  support at the beginning of their trajectory. Other patients may experience a gradual decline over time, which may indicate that they may need regular follow-up throughout their trajectory. Examining routinely collected PROMs data in registries provides an opportunity to analyze different trajectories as people engage with healthcare services over time. Identification of individual differences in trajectories aligns with the patient-centred emphasis in nursing practice because they help to explain variations in illness responses and can support the evaluation of the effects of health interventions at the individual and group levels (Henly et al., 2011). In real world settings, individual trajectories can provide details regarding the diversity of outcomes in a way that allows clinicians to prepare for next steps and to make more informed decisions about care and as a basis for quality improvement. Thus, it is critical that nurses, who have a direct influence on patients’ experiences and the outcomes of care, have a better understanding of the possible trajectories and underlying causes of change in health status over time. More specifically, the knowledge of individual trajectories may facilitate the development of a longitudinal design perspective to better respond to the needs of patients and tailor interventions for individual patients with AF in light of highly variable and complex health trajectories. 1.3 Purpose and Significance of the Study   The purpose of this study was to explore the analytical potential of patient-reported data stored in clinical registries, with the aim to explicate the process of conducting such an analysis of data from patients with AF in specialized clinics. To begin this process, we focused on exploring the trajectories of these patients and factors that predicted their different trajectories, which could inform tailored interventions for patients who may be at higher risk of experiencing poor outcomes. 10  This study stemmed from discussions with healthcare administrators and clinicians at Cardiac Services BC, an agency of the Provincial Health Services Authority, which oversees cardiac services in British Columbia. It was thought that by examining a large, registry-based PROMs dataset, novel insights could be gained about practices, processes, and outcomes to help inform patient-centred healthcare delivery (Nelson et al., 2016). In relation to nursing, the use of routinely collected patient-reported data in registries could be useful to better address outcomes that are important to patients as they adapt to the changing circumstances of their condition and treatment.  Traditionally, clinical and evaluation studies have focused on examining group-level effects of treatment while heterogeneity in individuals’ responses to treatment and change processes have largely been ignored (Schneider et al., 2012). As a result, there is limited understanding of how to improve treatment for individuals whose trajectories do not correspond to the average. Examining heterogeneity in individual trajectories is not only central to facilitating efforts to tailor interventions for individual patients (i.e., by identifying and changing interventions for individuals who have poorer health) but also to learn about the factors that are beneficial for individual patients who have better health outcomes. This dynamic individual trajectory perspective aligns with patient-centred care in nursing practice by focusing on the individual person in a way that is consistent with the human experience of health and illness over time (Henly et al., 2011).  1.4 Research Question Routine collection of PROMs in registries is increasingly used to enhance patient-centred care. However, most longitudinal studies have focused on average trajectories with little consideration of individual differences with distinct patterns of change – differences that could 11  provide insights into the diversity of outcomes and tailoring of health care for different patients. This study addressed this gap by generating new insights to the analysis of routinely collected PROMs data in registries for AF, as an example, which may be applicable in other clinical contexts to improve patient outcomes. Thus, the purpose of this study was to explore the analytical potential of patient-reported data stored in clinical registries, with the aim to explicate the process of conducting such an analysis by answering the following research question:  Can the trajectories of change in patient-reported outcome measures for outpatients with atrial fibrillation, after initial consultation, be explained by different biological functioning or individual and environmental characteristics?  Answering this main research question allowed us to address three key methodological questions explained in the methods section:  1) What information can be extracted from linked data sources? (3.4) 2) How best to accommodate missing data? (3.5) 3) What is an appropriate data analysis strategy? (3.6) 1.5 Summary As health care shifts toward patient-centred care, registry-based PROMs data are becoming an important clinical tool to not only guide clinical decision making, but also to inform healthcare programs and delivery systems; however, their use has received limited attention in the AF literature. By exploring the research question stated above, this study was positioned to provide methodological insights into analyzing routinely collected PROMs for AF and guidance in generating information from complex real-world registry data that could facilitate patient-centred care, and which may be applicable in improving healthcare processes and outcomes more generally. For clinicians, administrators, and healthcare decision makers to shift their focus to 12  include the use of PROMs in routine practice, more research is needed to identify the trajectories of change and factors that influence change in patients’ PROMs over time. For nurses to exercise more autonomy and to make decisions that facilitate patient-centred care, knowledge from PROMs data will increasingly be relied upon to capture outcomes that are important and meaningful to patients – information that cannot be collected through standard means, such as mortality and morbidity measures. Thus, nurses will need to ensure that they are equipped with relevant knowledge and skills necessary to participate in wider cross-disciplinary conversations to improve the use of PROMs, to encourage their inclusion in clinical registries, and to address barriers to their use.  13  Chapter 2: Literature Review The literature review presented here provides an overview of registries and barriers to their use, and the relevance of patient-reported outcomes in the context of AF. This section also explains how examining trajectories of change in AF may help to inform patient-centred care. Finally, the conceptual framework that underlies the study is presented. 2.1 Registries and Barriers to their Use A registry is a large organized system that relies on observational methods to collect uniform data to evaluate specified outcomes of a population of patients, and that serves a predefined scientific, clinical, or policy purpose (Gliklich et al., 2014; Hoque et al., 2017). The reliance on observational rather than interventional methods means that the care provided and recorded in the registry is determined by clinical judgement rather than by a study protocol. Historically, data have been entered into registries manually via medical record abstraction, although recent registries (e.g., The National Cardiovascular Data Registry) are able to extract electronic data directly from electronic health records (Bhatt et al., 2015).  Registries are often referred to as real-world data because large volumes of data are collected for every clinical encounter for nearly all patients (compared with clinical trials that are conducted under strict inclusion/exclusion criteria) (Gliklich et al., 2014). Existing registries have served critical functions to observe the course of a condition and to identify the likely causative factors; to understand variations in treatment and outcomes; to describe appropriateness and disparities in the delivery of care; to assess effectiveness; and to measure the quality of care delivered (Gliklich et al., 2014; Ludvigsson et al., 2011; Rolfson et al., 2016). Registries can either be prospective or retrospective in design and are developed and operated by different entities, including academia, professional societies, non-profit organizations, 14  government agencies, and industry (Gliklich et al., 2014). Registries typically include personal, exposure, and outcome information (Gliklich et al., 2014). Personal information includes patient demographics, clinical histories, and identifiers. Exposure information includes patients’ experiences with a product, condition, or procedure. Outcome information includes patients’ outcomes that are of interest to the registry stakeholders, such as adverse events and, at times, patient-reported outcomes. In addition, registries may be linked to other data sources such as administrative health data (e.g., data collected by government for record keeping) and pharmacy records to answer additional questions (Gliklich et al., 2014). Different stakeholders may recognize the value of registries and benefit from them in various ways. For clinicians, a registry can be used to track and monitor a condition of a large number of patients over a long period of time to provide a real-world picture of the progress of a condition, treatment patterns, and patient outcomes (Gliklich et al., 2014). For group practices, a registry might provide data that can be used to compare the quality of the care provided for practice improvement. For patients, a registry can help to monitor individual progress of a condition through examination of individual trajectories of change over time and facilitate the development of best practices or research on treatment. For payers or commercial sponsors, registries can provide information about how procedures, products or pharmaceuticals are actually used and about their performance in routine clinical environments. The benefits of using registries for all stakeholders in the healthcare system have led to the development of hundreds of registries throughout the world covering conditions that range from cancer to dementia (Garcia-Ptacek et al., 2014; Hayat, Howlader, Reichman, & Edwards, 2007). For example, the United Kingdom has over 50 clinical registries, while the United States and Sweden have over 15  100 registries (Emilsson, Lindahl, Köster, Lambe, & Ludvigsson, 2015; Healthcare Quality Improvement Partnership, 2018; Lyu, Cooper, Patel, Daniel, & Makary, 2016).  Despite their value to various stakeholders, one of the major barriers to the use of registries is uncertainty regarding the quality and the completeness of the data (Olmo, McGettigan, & Kurz, 2019). Since registries do not typically dictate patient visit schedules or a specific treatment protocol, patients may miss a scheduled visit or decline to undergo a particular procedure, and clinicians may choose to forego expected tests for treatment for certain subsets of their patients (Mack, Su, & Westreich, 2018). In addition, patients may refuse providing certain information and there may be errors in documentation (e.g., inconsistent or out-of-range values, which could lead to inaccuracy and missing data in the registry database (Mack et al., 2018)). More importantly, registries operate within a variety of sociological, technical, and policy contexts that make access to the data difficult for both patients and researchers to guide clinical practice or policy. There is much evidence to suggest that the sharing and re-use of data increases data quality, and thus their utility (Banzi et al., 2019; Naudet et al., 2018; Ohmann et al., 2017; Vickers, 2016). A case study of clinical registries in five countries showed that providing transparency by sharing best practices and providing regular performance reports can improve data quality and care (Levay, 2016). The important role of sharing and re-using data has mobilized a large multistakeholder task force to develop principles and recommendations related to the provision of access to individual patient-level data from clinical trials (Ohmann et al., 2017). Similar efforts for more consistent procedures and processes related to registries, including consent for data sharing, data standards, rights, and data management would help to reduce barriers to access, while recognizing that further work may need to be done to address wider national or regional privacy regulations. 16  Most registries still collect data via manual clinical chart abstractions (due to a lack of consistent structure and standardization of electronic health records data), which can lead to errors in the data entry process (Mack et al., 2018). In registries, data quality can be divided into two components: quality assurance and quality control (Arts, de Keizer, & Scheffer, 2002; Whitney, Lind, & Wahl, 1998). Quality assurance takes place before data collection to ensure that the data are of the highest possible quality at the time of data collection. Ideally, registries would have a data dictionary and definitions that describe both a common set of core data elements and how those derived quality indicators are interpreted (Gliklich et al., 2014). A typical data dictionary may include field length with start and end position of the variable and a detailed description of each variable, along with the coding information. For example, the term “health service delivery area” would be defined as health services near patients’ residences with coding information that range from 11 (e.g., East Kootenay) to 53 (e.g., Northeast) of the specific area. These core data elements and formats contained within a data dictionary would be harmonized across all participating centres to facilitate the standardization of data collection (European Medicines Agency, 2018). In contrast, quality control takes place during and after data collection to identify and correct for sources of error. Some examples of quality control procedures include using software algorithms to flag any errors, chart audits, and frequent use of the database to check for outliers, inconsistencies, and omissions (Arts et al., 2002; Williams, 2010).  Due to the high cost, resources, and efforts needed to maintain a registry, the extent of the work to ensure data quality is highly dependent on the individual registry. A review of the registries in the United States indicated that a majority of healthcare organizations (e.g., 95%; 36/38) had conducted some auditing of their data (e.g., automated audit, remote or third-party 17  audits, comparisons with source data or on-site audits), while a few reported that their registries were new and that their audit methodologies had not been developed yet (Blumenthal, 2017). The underlying issue in considering the quality of the data in any information system such as registries is their static nature; in the real world of practice, data are changing at any point along the continuum of data collection and processing (Richesson & Andrews, 2012). For example, clinicians’ ratings or interpretation of results may change over time (Nahm, 2012). In addition, some errors may not even be open for detection, such as patients who may deliberately provide inaccurate answers on a questionnaire (Nahm, 2012). For these reasons, data quality is deemed a relative concept, defined as the “totality of features and characteristics of a data set that bear on its ability to satisfy the needs that result from the intended use of the data” (Arts et al., 2002, p. 602). Thus, some researchers maintain that the data are of sufficient quality if they satisfy the requirements and needs from the data user’s perspective (Arts et al., 2002; Keller, Korkmaz, Orr, Schroeder, & Shipp, 2017). Despite the significant efforts made in maintaining registries and generating information that is robust and reliable, data must be analyzed carefully and conscientiously (Urschel, 2015). For research purposes, registries could potentially be used to generate hypotheses for testing in prospective trials following rigorous protocols, verification of the observations, and facilitation of descriptive studies (Hoque et al., 2017). Another barrier to the use of registries is the complex statistical methods and techniques required to perform certain kinds of analysis (European Medicines Agency, 2018). For example, depending on the specific research question, a comparison group may be needed to assess for differences in treatment outcomes or the strength of associations between groups (Gliklich et al., 2014). Registries without comparison groups can be used for descriptive purposes, such as 18  characterizing the natural history of a disease or condition (Gliklich et al., 2014). But the addition of a comparison group can add significant complexity and challenges to the interpretation of the results. For example, patients who receive a new treatment may have different risk factors for adverse events compared with those who receive a usual treatment. One frequently used strategy for comparative analyses is individual matching of exposed patients and comparators with regard to key socio-demographic factors such as age and gender (Gliklich et al., 2014). Other techniques include matching study subjects on the basis of a large number of risk factors whilst using propensity scores (i.e., the predicted probability of use of one treatment over another based on clinical characteristics) to create strata of patients with similar risk profiles (Gliklich et al., 2014).  Despite these limitations, clinical registries are becoming an essential source of real-world data to answer questions that may be of relevance to various stakeholders (Nason & Husereau, 2014). As international organizations (such as the Clinical Data Interchange Standards Consortium) actively work to improve the standardization and interoperability of healthcare data (Clinical Data Interchange Standards Consortium, 2019), there will be increasing need for registries to include patients’ perspectives in their design, governance, and operation. Most registries have focussed on biological measures such as radiologic data or laboratory tests, which may not be directly meaningful to patients and therefore nurses (Kelley, Lipson, Daly, & Douglas, 2013). A greater understanding of the needs of patients in using registries is increasingly acknowledged, which may also help to facilitate improved access to data and to address issues of privacy (Nason & Husereau, 2014). Thus, the integration of patients’ perspectives of their outcomes within registries is crucial in making the data more meaningful to 19  patients as well as their nurses, in supporting shared decision making, and in shifting health services to the delivery of patient-centred care (Kelley et al., 2013; Nelson et al., 2016).  2.2 Patient-Reported Outcomes   There is tremendous interest in the use of PROs to inform healthcare decision makers, researchers, and clinicians about patients’ experiences, especially with respect to chronic illnesses. PROs provide information obtained directly from patients about the effects of a health condition and its management, and include measures of quality of life, the impact of a disease state on daily living and social functioning, symptom information, satisfaction with interventions, and other dimensions of self-reported health status (Deshpande et al., 2011). Types of PRO data collection range from oral medical histories and discussions with healthcare providers, to cognitive interviews and surveys that have been validated with patients (Deshpande et al., 2011). The focus on PROs is driven by an increasing interest of patients to not only frame their disease experiences within the greater viewpoint of their lives, but also to actively understand, participate in, and influence their healthcare decisions (Ocloo & Matthews, 2016). Although inconsistently operationalized, measured, and utilized, PROs have been studied in various settings, including the clinical, research, and policy arenas (Weldring & Smith, 2013). The measurement of PROs in relation to quality of life can be traced back to the transition period in the early 20th century from a predominant focus on curative interventions in health care (e.g., quantity of life) to an emphasis on the side-effects of treatments and the impact of disease and illness on physical, social, and emotional wellbeing (Pennacchini, Bertolaso, Elvira, & De Marinis, 2011). This transition was occurring as healthcare innovations helped to prolong life, which raised issues regarding the quality of life of individual patients. For example, advances in renal dialysis, transplantation, and new forms of cancer therapy succeeded in 20  improving life expectancy but often failed to make those extra years “satisfying” (Armstrong & Caldwell, 2004). Thus, there was growing recognition that new measures of success were needed to not only move away from the biomedical approach to care but also to ensure that healthcare innovations were patient centred.  To help shift the focus from the biomedical model and towards a biopsychosocial model, the World Health Organization (1946) broadened its definition of health to “a state of physical, mental, and social well-being and not merely the absence of disease” (p. 100). This laid the groundwork for conceptualizing PROs as a multidimensional construct that considers patients’ perspectives. Since then, various tools or instruments to measure PROs have spread to inform healthcare decision making at various micro-, meso-, and macro-levels (Krawczyk et al., 2018; Sutherland & Till, 1993). For example, the micro-level focuses on the use of PROs in clinical practice where individual patient data are tracked over time to monitor a patient’s health status and to inform clinical decisions about that patient’s goal of care. At the meso- and macro-levels, individual level data are aggregated to inform decisions about quality improvement, healthcare programs, and delivery systems (Greenhalgh et al., 2018). The terminology used to describe the conceptualization of PROs has also continued to evolve with the terms “quality of life” and “health-related quality of life” being used interchangeably in the PRO literature; however, the term “PRO” (and by extension “PROM”) has been increasingly adopted by researchers, clinicians, regulatory agencies, and policy makers. Since we are interested in instruments used to measure PROs, the term “PROM” will be used herein to describe these outcomes. The conceptualization of PROMs can be best understood within a larger context of four categories of clinical outcome assessment as used by the US Food and Drug Administration:   21  1) Patient-reported outcome (PRO) measures;  2) Clinician-reported outcome (ClinRO) measures; 3) Observer-reported outcome (ObsRO) measures; and 4) Performance outcome (PerfO) measures.    Clinical outcome assessments depend on the rater, whether a patient (e.g., self-observation), a clinician, or a non-clinician observer (Walton et al., 2015). Because people with different backgrounds, experiences, and training are likely to have different perspectives and skills, their ratings with regards to patients’ health status may vary. For example, PRO measures directly assess patients’ perceptions about their own health conditions and their treatment without interpretation by anyone else. Symptoms or other unobservable concepts (e.g., pain severity or nausea) are best measured on the basis of information provided by the patient experiencing them. Clinician-reported outcome measures (ClinRO) are based on reports that come from trained healthcare professionals after they have observed patients’ health conditions. These measures involve clinical judgement of observable signs, behaviour, or other physical manifestations thought to be related to a disease or condition, yet feelings or functions that are known only to the patient (e.g., pain intensity) cannot be directly assessed. Observer-reported outcome measures (ObsRO) are those for which observations can be made, appraised, and recorded by a person other than a patient and do not require specialized professional training. This may be a parent, spouse, or other non-clinical or informal caregiver who is in a position to observe and describe a specific aspect of a patient’s health status, yet these too are unable to “gain access” to the subjective experience of the patient. And last are performance outcome measures (PerfO), a type of measurement based on a task performed by a patient, but no rater’s judgement affects the 22  measurement. This includes instruments such as six-minute walking tests, the scoring of which are independent of the patient’s subjective experience.  Since it is not possible for clinicians or other observers to know patients’ health and quality of life without directly asking them, the importance of PROMs has been repeatedly highlighted over the years: by the Patient-Reported Outcomes Measurement Information System initially funded by the US National Institutes of Health in 2004 (Cella et al., 2007), by the US Food and Drug Administration’s release of the “Guidance for Industry Patient-Reported Outcome Measures: Use in Medical Product Development to Support Labeling Claims” in 2009 (U.S. Department of Health and Human Services Food and Drug Administration, 2009), and in this past decade by the US Government’s establishment of the Patient-Centered Outcomes Research Institute in 2010 (Clancy & Collins, 2010). In the context of cardiovascular health research, PROMs are gaining momentum and are strengthening the awareness of a broader model of cardiac health services that includes physical, mental, and social health status (Anker et al., 2014; Rumsfeld et al., 2013). This awareness led to the release of an American Heart Association Scientific Statement that advocated for greater use of PROMs as key outcomes of cardiovascular health in clinical research, practice, and disease surveillance (Rumsfeld et al., 2013). However, despite an increasing focus on PROMs, they have received limited attention in the AF literature. To date, very few studies have evaluated routinely collected PROMs in registries of patients with AF.  2.3 Atrial Fibrillation  Atrial fibrillation (AF) is an irregular and often rapid heartbeat, or a quivering of the upper chambers of the heart called the atria. AF occurs as a result of a malfunction in the heart’s electrical system and is the most common cardiac arrhythmia. Although different patients may 23  experience different symptoms, some patients describe AF as feeling like a skipped heartbeat, followed by a racing of the heart to “catch up” (McCabe, Schumacher, & Barnason, 2011). Others may describe it as a hard pounding or even a fish flopping around in the chest (McCabe, Schumacher, et al., 2011). Patients may also feel pressures in their chests that mimic a heart attack (McCabe, Schumacher, et al., 2011).  During and after AF, many patients may experience persistent tiredness, loss of energy, and shortness of breath (McCabe, Schumacher, et al., 2011). For many patients, AF can lead to high levels of anxiety, which may cause patients to be afraid to leave their homes and to prevent them from performing activities that they enjoy for fear of recurrence. Over time, AF may progressively increase in frequency and duration, lasting for a few days (paroxysmal), weeks (persistent), or even years (permanent) (National Heart, Lung, and Blood Institute, 2014). While AF is not usually life-threatening by itself, it can cause serious debilitating conditions. Since cardiac output (i.e., blood flow from the atria into the ventricles and then to the rest of the body) is restricted, blood clots can form in the atria, which can travel through the body, block a brain artery, and cause a stroke (National Heart, Lung, and Blood Institute, 2014). AF can also damage or overwork the heart muscles, and over a long period of time can cause heart failure and other heart-related complications such as vascular dementia and respiratory failure (National Heart, Lung, and Blood Institute, 2014). The inability to perform activities of daily living and the anxiety associated with some of these debilitating symptoms may contribute to impaired health and quality of life (see Table 2-1).    24  Table 2-1.  Atrial Fibrillation Symptoms and their Effects on Health and Quality of Life Symptoms of Atrial Fibrillation Effects on Health and Quality of Life Chronic fatigue, dyspnea, and  impaired exercise tolerance   Decreased exercise, depression, cognitive impairment, impaired social life, and reduced ability to provide self-care  Syncope or presyncope   Inability to drive, injury secondary to an associated fall, decreased independence Palpitations Anxiety, insomnia, and fear of leaving home   Chest Pain Anxiety, fear of heart attack, and/or stroke, anticoagulation-related lifestyle limitations, frequent visits to the emergency department  Since symptoms of AF can vary widely (e.g., some individuals may be asymptomatic whereas others can experience debilitating conditions), current guidelines in the management of AF (apart from anticoagulants to prevent strokes) are focused predominantly on reducing symptoms and improving quality of life (January et al., 2019; Kirchhof et al., 2016). The two main treatments for AF include rate control to slow the heart rate and rhythm control to restore the heart’s normal sinus rhythm through anti-arrhythmic medications and catheter ablation (a minimally invasive procedure in which a catheter is guided into the heart to selectively destroy the tissue that is responsible for the irregular heart rhythm) (January et al., 2014). Yet even well-established treatment of AF, such as ablation, remains an “imperfect” therapy in that some patients may require more than one procedure or may not find the therapy to be effective (Liang & Santangeli, 2016), which can evoke further anxiety, fear, and frustration. For example, many patients have reported feeling worried about having palpitations at any moment, being paralyzed and being a burden to others, and not being able to enjoy any social life, such as visiting friends, shopping, or travelling (Altiok, Yilmaz, & Rencüsoğullari, 2015). PROMs are regarded as being 25  particularly relevant in this population because AF and its treatment affect many aspects of patients’ health and quality of life. Despite increased emphasis on assessing patients’ health and quality of life in AF, only a small minority of clinical studies reported collecting PROMs. For example, results from the Clinicaltrials.gov analysis found that among 1,709 AF studies posted from 1999 to 2018, only 238 studies (14%) included PROMs, and among these studies most (n = 198; 83%) used generic PROMs with only a few (n = 40; 17%) reported using AF-specific PROM (Steinberg et al., 2019). This finding not only suggests that there is an under-emphasis of outcome measures that are most relevant to patients, but also suboptimal implementation of AF-specific measures that have been shown to be more sensitive and responsive than generic measures (Björkenheim et al., 2018). The few studies that have used AF-specific measures are highlighted in Table 2-2.   26  Table 2-2.   Summary of Patient-Reported Outcome Measures in AF Patient-reported  outcome measure  Domain Clinical Study Implementation AF effect on quality of life (AFEQT) 4 domains: 1) symptoms, 2) daily activities, 3) treatment concerns, 4) treatment satisfaction ORBIT-AF registrya CABANA trialb University of Toronto AF severity scale (AFSS) 4 domains: 1) AF burden, 2) global well-being, 3) AF symptom score, 4) health care utilization RACE II trialc CTAF triald AF quality of life scales (AF-QOL) 3 domain: 1) psychological, 2) physical, 3) sexual SARA triale Mayo AF-specific symptom inventory (MAFSI) Single inventory of AF symptoms CABANA trialb  Symptom checklist (SCL) Single inventory of AF symptom frequency and severity AFFIRM trialf Arrhythmia-specific questionnaire in tachycardia and arrhythmia (ASTA) Single inventory of arrythmia symptoms SMURF studyg Note. a(Freeman et al., 2015). b(Packer et al., 2018). c(Groenveld et al., 2011). d(Dorian et al., 2002). e(Mont et al., 2014). f(Jenkins et al., 2005). g(Charitakis, Barmano, Walfridsson, & Walfridsson, 2017).  Adapted from “Tackling patient-reported outcomes in atrial fibrillation and heart failure: Identifying disease-specific symptoms,” by B. A. Steinberg and J. P. Piccini, 2019, Cardiology Clinics, 37(2), 140. Copyright by Elsevier Inc.  Some of these studies have found AF-specific PROMs to be useful in evaluating the effects of treatment. For example, the CABANA trial found more favourable outcomes for the catheter ablation group compared with the drug therapy group using the AFEQT and the MAFSI questionnaires (Mark et al., 2019). These findings suggest that AF-specific measures can help to better guide decisions in the management of AF.  Unfortunately, PROMs in general have yet to enter routine clinical practice, partly due to questions about their validity, relevance, and usefulness as well as challenges of efficient data collection methods (de Groot et al., 2017; Foster, Croot, Brazier, Harris, & O’Cathain, 2018). 27  Thus, more research is needed to understand how the routinely collected PROMs data in registries can be best analyzed to generate information that can be useful in facilitating patient-centred care.  2.4 Trajectories of Change in Atrial Fibrillation Most longitudinal PROMs studies to date (that have utilized conventional growth modelling techniques) have focused on “average trajectories”, which assume that all individuals in a given population follow the same pattern of change. Such an assumption may be too simplistic in heterogeneous clinical populations (Henly et al., 2011). For example, there is not a single average trajectory for patients with AF because the onset of AF varies among individuals. Nor does the trajectory of AF become standardized over time once it occurs. For some patients, AF can begin in short bursts that resolve within a week. For others, it can lead to more persistent episodes over time that require more aggressive treatment such as ablation (or re-ablation if there is recurrent AF) (Nattel et al., 2014). In addition, the treatment of AF may affect patients differently, and depending on the type of AF, certain treatment strategies may not be appropriate (e.g., patients who are diagnosed with permanent AF) (Frankel, Kamrul, Kosar, & Jensen, 2013). Another consideration with regards to the treatment of AF is that patients’ beliefs about their illness may play an important role in influencing their health trajectory at a given point in time (McCabe, Barnason, & Houfek, 2011). For example, patients who present in the emergency department with abrupt symptoms of chest pain and palpitations may have high levels of anxiety because of a false belief that they are having a heart attack. Their initial level of anxiety, and therefore their initial trajectory, may stabilize once these patients are provided education about AF and how to control its symptoms (McCabe, 2011). In contrast, some patients may have less dramatic symptoms, such as fatigue or dyspnea, but may experience them for many months. In 28  these cases, their health trajectories may gradually decline over time unless they seek treatment and recognize the serious nature of AF and its consequences (McCabe, 2011). It may also be difficult to discern whether the impact on PROMs is solely related to AF or concomitant comorbidities because AF is often associated with other underlying cardiac conditions (e.g., valvular disease, heart failure, and hypertension) (Coyne, Margolis, Grandy, & Zimetbaum, 2005) and even high endurance exercise (Li, Cui, Xuan, Xuan, & Xu, 2018). Instead of considering individual variation around a single trajectory, the focus on models used to analyze individual differences in trajectories (methods that allow different groups of individuals to vary around different mean growth curves) can have several benefits in nursing practice and research (Henly et al., 2011). First, modelling individual difference can inform theories about how patients change. If certain groups of patients’ health trajectories get worse before they get better, then theoretical explanations for this observed response may be needed. For example, time can be used to explain that some patients may have more difficulty than others in managing their symptoms at the initiation of treatment (Brant, Beck, & Miaskowski, 2010). Second, trajectories of change over time can inform expectations regarding the magnitude and timing of change over the duration of treatment. An example of this can be found in Flint et al. (2017). After finding that the identified health trajectories for outpatients with heart failure showed little change, Flint et al. (2017) proposed that this information can help to set realistic expectations and guide shared decision making. Third, identifying subgroups of trajectories can help tailor interventions at the point of care (Henly et al., 2011). For example, patients with positive health trajectories may need minimal intervention (e.g., usual care), whereas patients with worsening health trajectories may need more intensive intervention or referral to specialized services. In addition to these particular examples, individual trajectories can describe the 29  dynamic (changing) course of health and illness, which naturally aligns with nursing’s patient-centred focus in understanding the uniqueness and the complexities of the individual. Thus, individual trajectories can help to provide relevant insights into the diversity of outcomes, improve health care, and open new areas in advancing nursing science and practice (Henly et al., 2011).  Since AF can affect PROMs in multiple ways depending on the type of AF, presence of comorbidities, and array of potential interventions—all of which create both methodological challenges and opportunities—the intent of this study was not to draw substantive conclusions to direct clinical care. Rather, we sought to identify important insights to aid the analysis of routinely collected PROMs data in AF—which may also be applicable in other clinical contexts as registries become increasingly used to improve healthcare processes and outcomes.  2.5 Selection of a Patient-Reported Outcome Measure A challenge for clinicians is to select an appropriate PROM that will best capture patients’ personal, subjective perceptions. Although there have been different methods in AF research to measure health and quality of life, there is no uniformly accepted PROM specifically for AF (Kotecha et al., 2016). In everyday clinical practice, the selection of an appropriate PROM may be limited and based more on pragmatic reasons or even because of its previous use in clinical trials (particularly in the early stages of its integration). At present, the most widely used PROM for patients with AF is the Atrial Fibrillation Effect on QualiTy-of-Life (AFEQT) questionnaire (Spertus et al., 2011). The AFEQT was developed in accordance with the recommendations of the Food and Drug Administration PRO document (2009) to address the limitations of other measures (in particular their lack of responsiveness to patients’ clinical changes). It contains 20 questions/items with a seven-point Likert-type response scale that assess 30  four domains: symptoms (4 items), daily activities (8 items), treatment concerns (6 items), and treatment satisfaction (2 items), and a summary score that incorporates data from the first 3 domains (see Appendix A). The total score for items in the domain is transformed into a 0 to 100 scale, where a higher score indicates better health status. These domains are based on a theoretical foundation and empirical studies related to health and quality of life, interviews with patients and clinicians, and a factor analysis of patients’ responses to an initial set of indicators (Spertus et al., 2011). To be clinically useful and well accepted in practice, the AFEQT questionnaire was designed to satisfy the basic properties of validity, reliability, and responsiveness. According to modern validity theory, an instrument must demonstrate construct validity, which refers to the justification of inferences or interpretations, actions, and decisions based on instrument scores (Messick, 1989; Sawatzky et al., 2017). This is a unified view that makes construct validity the whole of validity; other validity types (e.g., content and criterion-related validity) are additions to the evidentiary basis for (construct) validity (Messick, 1995). Reliability refers to whether the instrument yields the same consistent results on repeated measures each time they are assessed. Responsiveness refers to the instrument’s ability to detect change such as improvement or deterioration over time.  Although these psychometric properties are interrelated, each is independently important and complex to determine. In particular, the process of validation is a matter of degree, suggesting that the validity of our inferences may change over time. As Messick (1989) asserted, “Because evidence is always incomplete, validation is essentially a matter of making the most reasonable case to guide both current use of the test [instrument] and current research to advance understanding of what test [the instrument’s] scores mean.” Over time, “the existing evidence 31  becomes enhanced (or contravened) by new findings” (p. 13). Thus, in PROMs research, instruments can never be conclusively “proven” to be valid. Instead, the process of validation involves presenting cumulative evidence to support the inference and to show that alternative or competing inferences are not more convincing. To this end, the accumulated evidence of validity, reliability, and responsiveness of the selected instrument are described in detail.  In the original study of Spertus et al. (2011), the psychometric properties of the AFEQT questionnaire were evaluated for evidence of measurement validity, reliability, and responsiveness to clinical change using various criterion standards. The AFEQT questionnaire showed good internal consistency with a high Cronbach’s alpha coefficient of > .88 for all of its subdomains (Spertus et al., 2011). A factor analysis of the items identified three domains, which included symptoms (4 items), daily activities (8 items), and treatment concerns (6 items) (Spertus et al., 2011). Test-retest reliability was indicated when the ratings of 43 patients in the “no treatment change” group, at baseline and one-month follow-up, provided intraclass correlation coefficients for Overall (.80), Daily Activities (.80), Treatment Concerns (.70),  Treatment Satisfaction (.70), and Symptoms domain (.50) (Spertus et al., 2011). Evidence of construct validity include acceptable metrics of convergent and divergent validity (r ≥ .40 and < .40, respectively) based on the correlations of the AFEQT domains with other commonly used questionnaires, including generic PROMs (e.g., the Short-Form Health Survey (SF-36) (Ware & Sherbourne, 1992) and the EuroQol (EQ-5D) (EuroQol Group, 1990)), the symptom checklist (SCL) (Bubien, Kay, & Jenkins, 1993), the Generalized Anxiety Disorder (GAD-7) (Spitzer, Kroenke, Williams, & Löwe, 2006), and the Atrial Fibrillation Severity Scale (AFSS) (Dorian et al., 2000). In addition, the AFEQT questionnaire showed a greater magnitude of change for patients treated with pharmacological therapy or with AF ablation compared with the generic 32  health status measure, including the SF-36 and the EQ-5D. Raine et al. (2015) also found that the AFEQT questionnaire was more responsive and correlated more closely with arrhythmia outcomes compared with the SF-36.  Based on the reliability and validity evidence, the AFEQT questionnaire should be interpreted in the context of the following potential limitations. First, the samples in the validation studies were relatively homogenous group of symptomatic patients, and there may be subtle manifestations of the disease in asymptomatic individuals that were not captured. Second, there were discrepancies in the strength of the test-retest reliability of the AFEQT questionnaire domains: the intraclass correlation coefficient was .70 in the treatment concerns domain compared with .50 in the symptom domain (which might be associated with the variable nature of AF symptoms). While Bland and Altman (1997) recommended a minimum level of reliability of .90 for clinical practice, the AFEQT questionnaire has been commonly used because it has more information about its psychometric properties in comparison to other measures based on the COSMIN checklist (Kotecha et al., 2016). Third, the original validation study did not provide guidance on what may be considered a meaningful change in scores. Based on a later study by Dorian et al. (2013), however, (in which 210 patients were assessed at baseline and again at three months to estimate a meaningful improvement using the global ratings of change score), a change score between 6 to 19 points has been suggested as a minimal important improvement.  Thus, the AFEQT questionnaire currently offers the best available option for an AF-specific PROM (Steinberg et al., 2019), which has been widely used in clinical research to compare differences in treatment outcomes (Bai et al., 2015; Ikemura et al., 2019; Mark et al., 2019) and to evaluate models of care delivery (Smigorowsky, Norris, McMurtry, & Tsuyuki, 2017). 33  2.6 The Conceptual Frameworks   The revised Wilson and Cleary (1995) framework has been widely used in PROMs research, which was adapted by Spertus et al. (2002) for study with the cardiac patient population. Each of the revised models within the Wilson and Cleary (1995) framework are discussed below to begin explicating the process of analyzing PROMs data.  Wilson and Cleary (1995) developed a health-related quality of life (QOL) model to specify how different types of patient outcome measures interrelate (see Figure 2-1).  Figure 2-1. Wilson and Cleary Model. Reprinted from “Linking Clinical Variables with Health-Related Quality of Life: A Conceptual Model of Patient Outcomes,” by I. B. Wilson and P.D. Cleary, 1995, JAMA, 271(1), 60. Copyright by the American Medical Association.  Their aim was to integrate both biomedical and social science paradigms: the former focuses on the understanding of causal relationships and classifying patients into therapeutic groups, and the latter aims to understand how the social context influences patients’ illness experiences. Their overarching model describes a series of dominant relationships of five core concepts that link 34  biological and physiological factors, symptom status, functional status, and general health perceptions to overall QOL. The first concept, biological and physiological factors, focuses on the functioning of cells, organs, and organ systems including diagnoses of disease, physical examination of findings, and laboratory values. Although this concept does not represent a QOL domain per se, it delineates the basis for the rest of the concepts in the model that can be measured in terms of PROs. The second concept, symptom status, encompasses abnormal physical, emotional or cognitive states perceived by the patient. The third concept, functional status, includes physical, social, role, and psychological functioning. The fourth concept, general health perceptions, refers to patients’ evaluations and integration of all the preceding concepts. Lastly, overall QOL refers to patients’ well-being, which stems from satisfaction or dissatisfaction with the areas of life that are important to them. The model also links individual and environmental characteristics with four concepts (with the exception of the physiological and biological factors). Although the Wilson and Cleary (1995) model has been widely applied to different populations and clinical settings (Ojelabi, Graham, Haighton, & Ling, 2017), one of the major criticisms of the model is that the definitions of the individual and environmental characteristics are not made explicit (Bakas et al., 2012). This is important to address because individual characteristics of cognitive and behavioural status, as well as aspects of environmental characteristics such as social support are recognized as major factors that influence QOL (Tilburgs, Nijkamp, Bakker, & van der Hoeven, 2015; Zubritsky et al., 2013). Therefore, it has been argued that Ferrans et al.’s (2005) model is a better model that clearly indicates the importance of individual and environmental factors to QOL. 35  Ferrans et al. (2005) revised the Wilson and Cleary (1995) model based on a review of the literature and a thorough examination of the theoretical underpinnings of each of the concepts (see Figure 2-2).   Figure 2-2. Ferrans et al.'s (2005) Revised Wilson and Cleary Model. Reprinted from “Conceptual Model of Health-related Quality of Life,” by C. E. Ferrans et al., 2005, Journal of Nursing Scholarship, 37(4), 338. Copyright by John Wiley and Sons.  While Ferrans et al. (2005) kept the five core concepts (e.g., biological, symptoms, functional status, general health perceptions, and overall QOL), the revised model removed the labels on the arrows that tended to restrict the characterization of the relationships and made explicit the definition of individual and environmental characteristics. According to Bakas (2012), these revisions have better facilitated the use and interpretation of the concept of QOL in nursing and health care. In addition, Ferrans et al. (2005) argued that the characteristics of the individual and the environment are theoretically related to all five concepts of the model, including biological function, because individual genetic characteristics and exposure to environmental pathogens can 36  predispose people to various diseases. Based on the ecological model of McLeory et al. (1988), and later revised by Eyler et al. (2002), Ferrans et al.’s (2005) model theorizes that intrapersonal (characteristics of the individual), interpersonal (formal and informal social support), institutional (healthcare facilities), community factors (informal social networks), and public policy (local and national laws and policies) interact at the level of the individual and thus influence PROMs. Therefore, the revised model has a much broader conceptualization with the characteristics of the individual categorized as demographic, developmental, psychological, biological, and characteristics of the environment encompassing both social and physical factors that influence health outcomes (Ferrans et al., 2005).  Spertus et al. (2002) adapted the Wilson and Cleary model for cardiac patients. Their revised model includes the physiological determinants of a disease, symptoms specific to the disease, self-reported health (referring to the ways in which the underlying disease manifests itself for patients), and general quality of life indicators (see Figure 2-3).   Figure 2-3. Variables and Instruments Associated with the Conceptual Framework 37  One of the major limitations of the original model developed by Spertus et al. (2002) is the exclusion of individual (e.g., age, gender, education level) and environmental characteristics (e.g., social support), which may be associated with poor patient outcomes in patients with AF (Gleason et al., 2019; Kang, 2011). Thus, there is an opportunity to combine the adaptations made by Spertus et al. (2002) and Ferrans et al. (2005) to inform future research in patients with AF. In addition (as mentioned previously), time is an essential component in understanding how individuals experience health and illness, which is often missing in nursing theory (Henly et al., 2011). Although the concept of time is implied by the flow from biological function, to clinician-reported symptoms and ultimately self-reported health, we explicitly added an arrow underneath the model to recognize that individual health is experienced as changes over time. For example, the start of treatment may vary among individuals, as well as their trajectories of health over time and after treatment.  In Figure 2-3, a conceptual understanding of the way AF influences QOL is provided and combined with common measurement instruments that have been previously developed and used in studies on this population. In this framework, biological function (comorbidities and stroke risk score) influences clinician-reported symptoms (CCS-SAF; grading classification score from 0 (asymptomatic) to 4 (severe impact of symptoms)). Although the original conceptualization of the symptoms domain is based on patients’ perspectives, the CCS-SAF is based on a rationale that clinicians would integrate results from biological function (e.g., comorbidities and electrocardiograms) to assign patients’ symptom category (which would be situated between biological function and self-reported health status). This in turn shapes the self-reported health score (AFEQT questionnaire), which is further modified by individual and environmental 38  characteristics (age, gender, and distance to clinic). The conceptual framework is supplemented by treatment (ablation and medication) as a component of biological function.  While the ultimate focus on QOL is important, the concept of “Quality of Life” could not be fully operationalized by existing instruments such as the AFEQT because it does not capture a full range of QOL domains that are unique to individuals (even though it purports to measure QOL by its very title). For example, the AFEQT questionnaire imposes standardized models of quality of life and preselected domains such as symptoms, daily activities, and treatment concerns that are more in line with general health status rather than quality of life per se. Several authors have pointed out that a simple listing of QOL domains is not a satisfactory way of measuring QOL because it is unknown whether all important domains have been included (Karimi & Brazier, 2016; Post, 2014). In other words, QOL is dependent on individuals’ preferences and priorities in life; therefore, PROMs that purport to measure QOL should allow patients to specify areas of life that are most important to them, such as assessed by the Schedule for the Evaluation of Quality of Life (SeiQOL) (Hickey et al., 1996). In addition, standardized measures have been confounded by the “disability paradox” because patients’ QOL is not necessarily commensurate with their health status. For example, patients with significant health and functional issues or symptoms may report having high QOL despite difficulties performing daily activities and being socially isolated (Albrecht & Devlieger, 1999). Thus, the AFEQT questionnaire was deemed to more closely align with capturing self-reported health status, rather than QOL, because it does not address all the factors that are important to patients.  This study used part of the conceptual framework as a basis for exploring the analytical potential of PROMs data in registries of patients with AF. To this end, we sought to characterize changes in the self-reported health (AFEQT questionnaire) of patients with AF, over time, and to 39  investigate the factors determining its course. Factors including age, gender, distance to clinics, stroke risk score, ablation, and anticoagulation therapy were used to predict the changes to AFEQT scores from initial consultation to follow-up. 2.7 Summary This chapter provided an overview of registries and the relevance of routinely collected PROMs data in the context of AF. Despite growing interest in the evaluation of health outcomes, the study of PROMs in registries is still in its infancy. Within the AF context, in particular, few researchers have used data stored in registries in evaluating PROMs in part because AF may affect PROMs in various ways that create significant methodological challenges (as well as opportunities to address them). To better inform daily clinical practice, future studies must address how routinely collected PROMs data in registries can be best used longitudinally, such that the PROMs scores become part of clinical reviews. This requires exploration of the trajectories of change over time and factors that predict different trajectories, and at the same time consideration of ways to disentangle the complex relationships between concomitant comorbidities, heterogeneous treatment effects, and the various complexities of analyzing “real-world” clinical data.  40  Chapter 3: Methods Methods describe the specific tools and techniques used by researchers to answer their research questions (McGregor & Murnane, 2010). Since there is a lack of methodological guidance on how to use patient-reported data in registries to inform practice, we aimed to explicate the process of conducting such analyses. This chapter is organized by briefly outlining the research design (3.1) and the AF clinic processes (3.2) to contextualize the data and its sources (3.3), and that allowed us to address three key questions when analyzing PROMs data collected for and stored in clinical registries:  1) What information can be extracted from linked data sources? (3.4) 2) How best to accommodate missing data? (3.5) 3) What is an appropriate data analysis strategy? (3.6) With regard to the last question, several sub-questions needed to be addressed: a) How best to analyze longitudinal PROMs data? (3.6.1) b) How best to represent the variability in frequency and timing of measurement occurrences? (3.6.2) c) How best to represent the shape of different trajectories? (3.6.3) d) How best to represent individual differences in trajectories? (3.6.4) e) How best to identify factors that explain the variability in individual trajectories? (3.6.5) Our analysis revealed that substantial data preparation and cleaning were required before we could extract relevant patient information from linked data sources. We used the recommended multiple imputation and full information maximum likelihood procedures to accommodate missing data. To analyze longitudinal PROMs data, we compared several methods including 41  ANOVA, multilevel, and latent growth models. However, we found that an emerging statistical approach known as growth mixture modelling allowed for the identification of multiple subgroups of trajectories and nesting of time to account for individually-varying times of observation. To focus on demonstrating this relatively new approach, we limited the range of possible models to three parameterizations: (1) the unrestricted random effects model that freely estimates all parameters but may lead to convergence issues; (2) the restricted random effects model that aids in convergence by constraining the slope variance to zero and reflects the underlying data structure; and (3) the restricted random effects plus autoregressive structure (AR1) model that addresses autocorrelation issues, which recognizes that closely timed repeated measures are more correlated than measures that are timed further apart. After modelling, we described the relationships between theoretically-derived variables (time-invariant and time-varying) and different identified trajectories.1 3.1 Study Design and Setting This is a retrospective cohort study of 2008 to 2016 AF registry data of registered patients from British Columbia (BC). Retrospective cohort studies are longitudinal observational studies that involve a case-defined population at the time follow-up has been completed (Mann, 2003). This design was suitable for this study because the data were already collected from patients who had been referred to an AF clinic and followed between 2008 and 2016. The aim of the retrospective cohort study was to look back in time to examine the trajectories of change over  1 It is important to note that attempts to answer these methodological questions were primarily used to demonstrate the analytical potential of patient-reported data in registries, and not necessarily to reach definitive clinical conclusions regarding typologies based on individuals’ responses to the management of AF.  42  time of this patient cohort and to determine whether differences in the patterns of change could be predicted by theoretically-derived variables.  The registry database contains patients’ health information (e.g., demographic, clinical and medication information), interventions that the patients received, and their outcomes including those measured with PROMs. This information is routinely collected by the staff and healthcare providers in the AF clinics. The AF registry was envisioned as a clinical-point-of-care data capture system so that all data entered would be part of a patient’s medical and health record. The data were specifically derived from five AF clinics in BC (see Figure 3-1). Four clinics began operation between November 2009 and October 2010, with a fifth opening in late 2011.  Figure 3-1. Location of Atrial Fibrillation Clinics in BC  43  Cardiac Services BC, as an agency of the Provincial Health Services Authority, is a provincially-funded organization with a mandate to plan, coordinate, monitor, evaluate, and fund cardiovascular disease-related treatment services. Key priority areas include quality, access, and sustainability throughout the system. As part of this mandate, Cardiac Services BC created the cardiac registry (20-year publicly funded) for purposes of supporting research, quality improvement and knowledge translation projects designed to improve patient care and the BC health system. The cardiac registry data collects data from all five AF clinics as part of a broader provincial plan to streamline care for patients who have had suboptimal responses to standard therapies, and therefore require more complex medication management or catheter ablation (Shalansky, Basi, & Yu, 2011). The AF clinics offer multi-disciplinary patient care, harnessing the knowledge and expertise of electrophysiologists, cardiologists, nurse practitioners, nurses, and pharmacists. The clinics provide several services (e.g., triaging, patient education sessions, telephone intake, post-ablation follow-up, anticoagulation services, and electrophysiology procedures) and augment the care of AF in the community. The AF clinics fit within a growing trend towards hospital-based ambulatory outpatient clinics (i.e., patients not admitted for an overnight stay) to provide timely access to the most appropriate care in the most appropriate place (Paul et al., 2006). Thus, registry data from these clinics provide a unique opportunity to assess patients with AF who are managed within this specialized multidisciplinary model (in which patients in collaboration with clinicians participate in their own care) in order to use that data to guide the potential expansion of the AF clinics. 3.2 AF Clinic Processes The initial creation of the AF clinics did not entail adherence to a prescribed model of care; instead, each clinic was encouraged to establish its own model of multi-disciplinary care 44  and its own operating and data entry processes. While each of the AF clinics developed independently, shared personnel were employed to “shadow” one another at the clinics, and the establishment of inter-team meetings have helped create similar yet distinct models of care. Similarities include requiring that patients must have documented AF with an electrocardiogram in order to access the AF clinics. Patients who have a new onset of AF management or require anticoagulation therapy are also generally referred to a cardiologist (from general practitioners or emergency departments), and patients who require more complex AF management are typically referred to an electrophysiologist (by an internist or cardiologist), with a request for ablation consideration. Yet discharge criteria and follow-up visit protocols (e.g., post-ablation and medication titration) at the AF clinics are site-specific. The timing, mode (in-person or telephone-based), and staffing (physician, nurse practitioner, nurse, and pharmacist) during the processes of care also vary by clinic. While each AF clinic has variations in their models of care delivery and operations, common features of the clinics include their focus on patient education, enhancing patient participation in treatment selection, and close collaboration among the electrophysiologists (EP), cardiologists, registered nurses (RN),2 nurse practitioners (NP),3 and pharmacists.4 The operating models of key AF clinic functions are summarized in Figure 3-2.   2 The role of an RN is referral and telephone triage and assessment, patient education, obtaining an initial history and assessment during the initial consultation, case management and telephone follow-up.  3 The role of an NP includes patient triage and case reviews, patient assessment and diagnosis of health status, medical management of patients and follow-up visits to assess response to treatment, ordering diagnostics, prescribing and titrating medications, and providing psychosocial support for patients and professional support for clinical staff. 4 The role of a pharmacist is medication reconciliation, patient education and counselling, and monitoring of anticoagulation protocols and anti-arrhythmic medications. 45   Figure 3-2. Patient Journey through the Atrial Fibrillation Clinic. Adapted from “Improving Access and Optimizing Care: Development of an Atrial Fibrillation Clinic to Implement Canadian Cardiovascular Society Guidelines,” by S. Lauck et al., 2011, Canadian Journal of Cardiology, 27(5), S199–S200. Copyright by Elsevier.  Patients are referred (or re-referred if they have been discharged) directly to the AF clinic by general practitioners or internists. Patients can also self-refer to a group education session. If patients do not have a primary care provider at the time of referral, AF clinic staff facilitate finding a family physician to provide ongoing care. Patients are generally triaged by either the RN or the NP to review medical histories, current and past treatments for AF, and underlying comorbidities. CCS symptomatic AF scale (Dorian et al., 2006) and stroke risk (Gage et al., 2001) scores are determined to assess the severity of the symptoms and urgency of the need for consultation. Patients are invited to attend a group education session about AF and treatment options before their initial consultation. During the initial on-site consultation visit, the patient’s care plan is co-developed by the multidisciplinary team.  46  For all patients, the treatment of AF begins with prevention of thromboembolism based on the stroke risk score and an attempt at rate control. Depending on their symptoms, patients can continue or modify their rate control management or be considered for rhythm control (e.g., cardioversion, antiarrhythmics, and catheter ablation) if they have high symptom burden (see Figure 3-3).   Figure 3-3. General Treatment Pathway. Adapted from Pocket Guides, by Canadian Cardiovascular Society, 2019.  Since patients on medication, such as antiarrhythmics, require ongoing monitoring due to side effects, the clinics have developed a common general pathway based on how patients respond to their treatment. For example, if a patient is tolerating the medication but is having suboptimal rhythm control, then the dosage may be increased, and follow-up to a clinic arranged. On the other hand, if a patient is not tolerating the medication and is continuing to have symptoms, immediate follow-up is arranged to change or discontinue the medication (see Figure 3-4).  47   Figure 3-4. Medication Pathway. Adapted from Atrial Fibrillation Clinic, by Cardiac Services BC, 2019.  For patients undergoing catheter ablation, the clinics have developed a general follow-up pathway from pre-ablation up to one-year post ablation to assess for complications (see Table 3-1).    48  Table 3-1.  Post-Ablation Pathway   Time Frame5  Assessment Diagnostic Tests Pre-ablation Consultation • If decision made, book ablation and provide pre-ablation instructions • Echocardiogram • ECG • Holter monitor • Heart CT scan • Blood work Ablation Procedure 1-2 weeks Post-ablation Teleconsultation by RN/Pharmacist • Assessment for complications • Review medications • Advise to continue oral anticoagulants until reassessed • ECG • Holter monitor 8-10 weeks Post-ablation Clinic visit with NP • Clinical assessment  • Review Holter monitor results • Assess need for rate/rhythm control  • Re-assess oral anticoagulants  • ECG • Holter monitor 14-18 weeks Post-ablation Clinic visit with EP  • Clinical assessment • Re-assess rate/rhythm medication • Formulate strategy for ongoing AF management • ECG • Holter monitor 1-year Post-ablation Clinic visit with EP or NP • Clinical assessment • Re-assess rate/rhythm medication • Formulate a discharge plan • ECG • Holter monitor • Echocardiogram Note. ECG = Electrocardiogram; CT = Computerized Tomography. Adapted from Atrial Fibrillation Clinic, by Cardiac Services BC, 2019.  5 The specific follow-up time frame may vary depending on the clinic and the patient’s condition.  49  Once the patients are on pharmacotherapy, an ablation management pathway, or both, they are followed until their care objectives are met and then they are discharged. 3.3 Data Sources and Variables As noted previously, the revised Wilson and Cleary (1995) framework that underpinned this study linked biological function (supplemented with treatment), clinician-reported symptom status, as well as individual and environmental characteristics, to PROMs (conceptualized as self-reported health status). Each of the variables were selected to represent these concepts. However, since registries themselves may not contain all the necessary variables, they are often linked to other data sources. The selection of variables and instruments was largely determined by their availability in these linked data sources. The following provides descriptions of each of the data sources used—held by Population Data BC (www.popdata.bc.ca)—and the selection and definition of the variables of interest in the context of the revised framework (see Figure 2-3). All inferences, opinions, and conclusions drawn from the various data sources are those of the author, and they do not reflect the opinions or policies of the data steward(s).  3.3.1 Atrial fibrillation clinic registry database. Cardiac Services BC maintains the AF clinic registry database as part of the HEARTis system, which houses multi-year cardiac data related to clinical processes and patient outcomes across the province of BC (Cardiac Services BC, 2018). This system was implemented as a clinical point-of-care data capture system, with all data entered being part of a patient’s medical and health record. A working Minimum Data Set was developed based on clinical goals during the AF clinic planning process at Cardiac Services BC. Although determination of the accuracy and completeness of the AF registry database is not possible without a comprehensive chart 50  audit, a data quality assessment was undertaken in the fall of 2011, which led to all of the clinics retrospectively entering mandatory data elements since inception.  From this system, we requested three data files: (1) the AFEQT data file, (2) the AF clinic data file, and (3) the ablation data file. The AFEQT data file contains information of patients who completed the AFEQT questionnaire. The AFEQT data are usually collected when patients visit the AF clinic during their initial consultation. After the plan of care is determined (e.g., medication or ablation therapy), subsequent AFEQT questionnaires are collected each time that the patient visits the clinic until they are discharged (one clinic mails the AFEQT questionnaire to their patients during each follow-up).6 Once the AFEQT questionnaires are completed onsite or returned by mail, the clerical staff enter the information into the registry.  The AF clinic data file contains demographic and clinical information about the patients who have visited the clinic. The information in the AF clinic data file is first collected when patients are referred to the AF clinic via a referral form. From this referral form, basic demographic information, such as age, gender, and address (including postal code) is entered by the clerical staff. During the initial consultation, the stroke risk score (CHADS2) is calculated and the CCS-SAF is assigned by the clinician, which is then entered into the file by the clerical staff, as well.  The ablation data file includes information about patients who have undergone catheter ablation and other diagnostic procedures. The date of the procedure and the type of procedure are entered by the clerical staff each time the patient undergoes a procedure.     6 Visits are determined by many factors. Although patients who are in the medication pathway do not have a set follow-up schedule (unless the patient is having symptoms) those on the ablation pathway follow a more structured schedule for up to one year (see Table 3-1). 51  Consolidation File  The BC Ministry of Health Consolidation data file is a central demographics database maintained by Population Data BC containing information about all individuals who are registered as being eligible to receive or who actually do receive medical services in BC (British Columbia Ministry of Health, 2018a). It is a comprehensive data source that includes basic socio-demographic information of almost the entire population of the province in any given year.7 The content of the file is regularly cleaned and maintained to ensure consistency in demographic information (Population Data BC, 2018a). This data file was requested to provide the socio-demographic profile of patients with AF, including their age and gender. Discharge Abstracts Database (Hospital Separations) The Discharge Abstracts Database (DAD) file contains information about all acute, chronic or rehabilitation inpatient or same-day surgical admissions to acute care hospitals (using the Canadian Classification of Health Interventions), and includes the codes for the patients’ diagnoses (ICD-10 codes) (Canadian Institute for Health Information, 2018). Comprehensive and regular quality checks and audits are conducted to ensure the accuracy of the data (Canadian Institute for Health Information, 2019). This data file was requested to identify co-morbidities and interventions patients with AF received during their hospital stays.     7 In Canada, public health insurance is available to eligible residents. Canadian citizens and permanent residents can apply for provincial health insurance. Under the Medicare Protection Act, enrolment is mandatory for all eligible residents and their dependents. Residents must live in BC, be a citizen of Canada or lawfully admitted to Canada for permanent residence, and must be physically present in BC at least six months in a calendar year (these rules changed in January 2020). Dependents must also be BC residents and may include a spouse, children, and dependent post-secondary students. The First Nations Health Authority enrols all “Status Indians” who are residents of BC (excluding people who receive health benefits by way of a First Nations organization pursuant to self-government agreements with Canada). Certain other people, such as some holders of study or work permits, issued under the federal Immigration and Refugee Protection Act, may be deemed residents. 52  Medical Services Plan Payment Information File The Medical Services Plan (MSP) payment data file covers all fee-for-service claims in BC, including the date of service, the fee item indicating the service provided, and the diagnostic codes for each patient visit using the three-digit ICD-9 codes (British Columbia Ministry of Health, 2018b). These codes were developed by the World Health Organization and include additional non-ICD-9 diagnostic codes to represent particular services within the MSP (e.g., X-ray examinations and lab tests) (British Columbia Ministry of Health, 2018). The majority of the records are submitted electronically by practitioners’ offices to MSP, which are then regularly audited and checked for quality assurance (Population Data BC, 2018c). This data file was requested to identify the co-morbidities of the patients with AF.  PharmaNet The PharmaNet data file contains all prescriptions for drugs and medical supplies dispensed from community pharmacies in BC as well as prescriptions dispensed from hospital outpatient pharmacies (British Columbia Ministry of Health, 2018c). This data file was requested to determine the type of cardiac medications that the patients with AF were receiving (although the drugs prescribed when the patients were hospital inpatients are not included in this database). Vital Statistics - Deaths  The BC vital statistics data file contains information about all deaths registered in the province of BC (British Columbia Vital Statistics Agency, 2018). The specific year, month and day of death were requested to determine whether failure of follow-up with AF clinic patients was the result of their having died (which had implications for the treatment of missing data).   53  3.3.2 Selection of variables and instruments.  To understand the data in the context of the AF clinic processes, site visits were conducted to discuss with the clinical staff how the data were collected, and to shadow the clinicians during their consultations with patients. This step provided a more informed approach to the selection of the variables that could be used and the inclusion/exclusion of patients for the analyses (for example, the exclusion of patients without an initial consultation date was informed by clinicians’ reports that they may not have followed the same clinic processes, enabling stronger alignment of the data to the practice setting). The following explains the relevant study variables and instruments, based on the revised framework, categorized as: a) individual and environmental characteristics, b) biological function, c) clinician-reported symptoms, and d) self-reported health status.  a) Individual and environmental characteristics: The revised Wilson and Cleary framework incorporates individual characteristics, including demographic, developmental, psychological, and/or biological factors, and environmental characteristics categorized as either social or physical (Ferrans et al., 2005). Variables that are indicators of these characteristics in this domain and available in the linked registry database included age, gender, and distance to the AF clinic. Age: Although research shows that increasing age amplifies the risk of AF (Wasmer, Eckardt, & Breithardt, 2017), there is limited understanding of the relationship between aging and PROMs in the AF population. Because there are many indications for the type of treatment offered, the ages of the patients in the AF clinics can vary widely. To derive age, the patient’s year of birth (from the Consolidation file) was subtracted from the year of the initial consultation 54  (from the AF clinic registry). The distribution of age in years was divided into quartiles (less than 60 years, 60-67, 68-75, and 76 or older).  Gender: There is a significant gap in the understanding of how men and women with AF respond to PROMs. Because sex is biological, the term gender was adopted for this study; men’s and women’s responses to PROMs may be influenced by their social contexts and cultural relationships. For example, previous research suggests that women are more likely to endorse items that describe emotions of vulnerability (Lix et al., 2016). Items with gender-specific endorsements may inflate negative scores at the scale level for women compared with men, reflecting measurement bias rather than true differences in health status. Therefore, gender is more appropriate to recognize the complexity of how men and women are situated and contextualized in their experiences of AF, their adaptation following treatment, and their responses to PROMs. Although the term gender itself is an evolving concept, the Canadian Institutes of Health Research (CIHR) Institute of Gender and Health (2012) conceptualized gender as referring to “socially constructed roles, relationships, behaviours, relative power, and other traits that societies ascribe to women and men” (para. 3). In the AF clinic registry, gender is entered when the patient is first added in the database with ‘M’ for men or ‘F’ for women. While some could argue that these indicators (M/F) pertain to biological sex, they are characterized as indicators of gender to highlight that social differences between men and women may contribute to differences in their health.   Distance to AF clinic: British Columbians who require specialized treatment for AF (e.g., catheter ablation) must travel to the larger metropolitan areas of the province to access specialized services. Access and continued follow-up for AF management may have implications for patients, especially those who reside in rural or remote communities. For example, the 55  recommended minimum follow-up times post-ablation are 3, 6, and 12 months (Calkins et al., 2018). In addition, there are ongoing education sessions for patients in these clinics to manage their AF and to reduce their anxiety. It is not known whether the need to travel long distances to access specialized care and to maintain appropriate follow-up influences patients’ anxiety levels and their capacity for self-care management, and in turn their health status. The distance in kilometres between the AF clinic and the patients’ places of residence, based on their postal codes, was provided by Cardiac Services BC. To obtain an indicator of accessibility, the distance was dichotomized to patients living ≥ 100 or < 100 kilometres away, as a proxy measure of the ease of transportation to the AF clinic.  b) Biological function: In the revised Wilson and Clearly (1995) model, the concept of biological function focuses on the function of cells, organs, and systems, and is assessed with indicators including medical diagnoses, physical examination findings, laboratory values, and treatment. To fully capture the clinical characteristics pertinent to patients with AF, common comorbidities and different types of treatment were described, including cardiac procedures and AF medications. Comorbidities: For descriptive purposes, comorbidity information was obtained because AF often develops subsequent to and in parallel with other conditions (Chamberlain et al., 2017). The comorbid conditions were based on information the clinicians collected during their intake assessment. Researchers have identified validated algorithms using administrative data (Tonelli et al., 2015), which were applied to categorize the information according to the following systems: (1) cardiac (heart failure, myocardial infarction, hypertension, peripheral vascular disease, stroke or transient ischemia attack, (2) respiratory (chronic obstructive pulmonary 56  disease and sleep disorders), (3) gastrointestinal (gastrointestinal bleed and peptic ulcer disease), (4) gastro-urinary/endocrine (chronic kidney disease, hypothyroidism and diabetes, and (5) mental health (depression). Based on the algorithm, patients were identified as having a given condition if they had a hospitalization (using the inpatient DAD file) or two or more claims (using the outpatient claims MSP file) within one year of the initial consultation date. These algorithms were identified as having good predictive value and sensitivity compared with an acceptable gold standard, such as a medical chart review (Tonelli et al., 2015) (see table in Appendix B). Charlson Comorbidity Index (CCI): For descriptive purposes, the CCI was estimated as a proxy measure of the patients’ overall disease burden (Quan et al., 2005). The CCI is a method of assessing the comorbid status of patients based on their number and severity of conditions. This index is based on the International Classification of Diseases (ICD) codes for 17 conditions (myocardial infarction, congestive heart failure, peripheral vascular disease, cerebrovascular disease or transient ischemic disease, hemiplegia, asthma/chronic obstructive pulmonary disease, diabetes, renal disease, liver disease, gastric or peptic ulcers, cancer, Alzheimer’s disease or dementia, rheumatic or connective tissue disease, HIV or AIDS). Each of these conditions are assigned a weight from 1 to 6; a summary score is then computed with higher scores indicating greater comorbid burden (range from 0 to 24). The CCI has been shown to be a valid and reliable indicator of disease burden for diverse clinical cohorts in a variety of healthcare settings including patients with AF (LaMori et al., 2013; Quan et al., 2011).  In this study, the longitudinal CCI was calculated because patients had multiple visits over time, which meant that the patients could be identified as having a condition during any episode of care. The longitudinal index was calculated from both the MSP and the DAD file 57  based on Quan et al.’s (2005) coding algorithm (see Appendix C SAS code). This code takes into account all of the 17 conditions for all episodes of care by identifying a comorbidity category only once and not increasing the overall index score when the same category occurs more than once. The CCI was treated as a continuous variable in the descriptive analysis.  CHADS2: The CHADS2 was recorded at the time of the initial consultation. The CHADS2 is a clinical tool that estimates the risk of stroke in the AF population (Gage et al., 2001). The score is derived by assigning one point for each stroke risk factor that a patient has, including having congestive heart failure, hypertension, or diabetes, and being greater than 75 years of age. Two points are assigned for a history of stroke or transient ischemic attack; the maximum score is six points (Gage et al., 2001). A score of 0 is low risk, one point is considered moderate risk, and a score of more than 1 represents a higher risk for stroke (Gage et al., 2001). Based on clinical guidelines, an individual with AF is recommended to initiate treatment with an oral anticoagulant if their CHADS2 score is ≥ 1 (Camm et al., 2010). Although there has not been rigorous reliability and validity testing of the tool, several studies have shown that increasing CHADS2 scores are associated with physiological changes in the heart and blood vessels (Li et al., 2013; Winkle, Mead, Engel, Kong, & Patrawala, 2014) and greater risk of adverse cardiovascular events (Chen et al., 2013; Welles et al., 2011). Although an updated tool is available (CHA2DS2-VASc), the CHADS2 was used because the AF clinics were using the score at the time of data collection.   Treatment: Prior history of cardiac surgery was obtained because researchers have reported that AF is the most frequent postoperative complication that occurs in about 35% of cardiac surgery cases, including coronary artery bypass surgery and valve-replacement surgery (Greenberg, Lancaster, Schuessler, & Melby, 2017). Thus, a comprehensive clinical profile of patients who 58  have undergone cardiac procedures are described including cardioversion, ablation (e.g., atrioventricular node ablation, pulmonary vein isolation ablation, and maze procedure), coronary artery bypass surgery, percutaneous coronary intervention, valve surgery, pacemaker implantation, and cardioverter defibrillator implantation. Patients were identified as having received a given intervention if they met the algorithm criterion of a hospitalization (using the inpatient DAD file) before the initial consultation date, based on the Canadian Classification of Health Interventions coding assignment (Canadian Institute for Health Information, 2015). In BC, about 96% of the interventions reported in the DAD were confirmed in the chart review while about 90% of the interventions recorded in the chart review were present in the DAD (Canadian Institute for Health Information, 2009; Canadian Institute for Health Information, 2010; Canadian Institute for Health Information, 2012).  For AF medications, the specific data in each category were based on the medications that were reviewed in the clinics: (1) anticoagulants (dabigatran, rivaroxaban, apixaban, warfarin), (2) antiplatelets (acetylsalicylic acid), (3) beta-blockers (atenolol, bisoprolol, metoprolol, carvedilol), (4) calcium channel blockers (diltiazem, verapamil), (5) antiarrhythmics (amiodarone, flecainide, propafenone, sotalol, dronedarone), and (6) digoxin. A one-year exposure ascertainment window ending on the date of the initial consultation was constructed (consistent with the algorithm to identify comorbidities within the one-year period) to classify patients with AF as having been exposed to a given medication recorded in the PharmaNet data file.    Among the treatment variables, ablation or anticoagulation therapy were specified as time-varying variables that could affect the AFEQT questionnaire item 16 (“Worrying about complications or side effects from procedures like catheter ablation…”) and item 17 (“Worrying 59  about side effects of blood thinners such as nosebleeds…”). To capture ablation as a time-varying covariate during the patients’ trajectory of care, the follow-up periods were divided based on the median time between the multiple AFEQT questionnaire completions (see Table 3-2). The same follow-up periods were used to capture anticoagulation therapy as a time-varying covariate.   c) Clinician-Reported Symptoms: The Canadian Cardiovascular Society Severity in Atrial Fibrillation Scale (CCS-SAF) was developed as a condition-specific tool for clinicians to classify AF patients’ symptoms. The scale is constructed with a four-level grading of symptom severity: (S) symptoms attributable to AF; (A) associations between symptoms (palpitations, dyspnea, dizziness/syncope, chest pain, weakness/fatigue) and documentation of AF or therapies for AF (i.e., therapy associated symptoms); and (F) functional consequences of these symptoms on the patient’s daily function and quality of life (Dorian et al., 2006). The AF class is then rated on a scale from 0 (asymptomatic) to 4 (severe impact of symptoms on QOL and activities of daily living) (see Figure 3-5). In the AF clinic database, the clinician enters the CCS-SAF scale during the on-site consultation as a measure of their patients’ symptom burden.    60  CCS-SAF Scale Definition 0 Asymptomatic with respect to AF 1 Symptoms attributable to AF with a minimal effect on a patient’s general quality of life 2 Symptoms attributable to AF with a minor effect on a patient’s general quality of life 3 Symptoms attributable to AF with a moderate effect on a patient’s general quality of life 4 Symptoms attributable to AF with a severe effect on a patient’s general quality of life  Figure 3-5. Canadian Cardiovascular Society Severity in Atrial Fibrillation Scale. Adopted from “A Novel, Simple Scale for Assessing the Symptom Severity of Atrial Fibrillation at the Bedside: The CCS-SAF Scale,” by P. Dorian et al., 2006, Canadian Journal of Cardiology, 22(5), 385. Copyright by Elsevier Inc.  d) Self-Reported Health: According to the established framework, self- reported health status includes aspects of symptoms, daily activities, and treatment concerns. To capture these domains, the Atrial Fibrillation Effect on Quality-of-Life questionnaire was used. AFEQT: The Atrial Fibrillation Effect on Quality-of-Life (AFEQT) questionnaire is a condition-specific PROM for AF. The AFEQT questionnaire contains 20 questions with a 4-week recall period. Questions 1-18 assess 3 domains: symptoms (4 items), daily activities (8 items) and treatment concerns (6 items), with the final two questions inquiring about treatment satisfaction (2 items). Each domain contributes insight into different aspects of patients’ self-reported health status, with higher scores representing better health (each question is answered on a seven-point Likert scale, with the first three domains scored between (1) for “Not at all…” to (7) for “Extremely…”).  61  For this study, the first three domains (i.e., the subscales) and a summary score were the outcomes of interest; we excluded treatment satisfaction because it did not capture self-reported health status, the focus of the study. The psychometric properties and the applicability of the questionnaire to the AF population have been previously discussed (a copy of the instrument, its scoring instructions, implementation manual, and a licence for its use can be obtained at www.AFEQT.org). At the AF clinic, patients are provided opportunity to complete the AFEQT questionnaire when they visit the clinic or have it mailed to their home. It is important to note that the AFEQT questionnaires were collected with varying intervals throughout the follow-up period rather than at key time points. For example, the AFEQT questionnaire could have been collected days or even years after the initial consultation, depending on the complexity of the management of the patient’s AF, the clinic wait times, or whether the patient decided to complete the questionnaire at all or at every visit. Based in part by what was available in the linked registry database, we categorized the final analytic variables according to the domains of the conceptual framework (see Figure 3-6).    62  Category Variables Values Primary or Derived Individual and Environmental Characteristics Age group (years) Categorical variable dummy coded as No (0) or Yes (1) Both – based on Consolidation file and AF clinic registry             Less than 60             60 – 67             68 – 75             76 or older  Gender Men (0) or Women (1) Primary – based on AF clinic registry   Distance to AF clinic < 100 km (0)  ≥ 100 km (1) Primary – based on AF clinic registry   Biological Function CHADS2 Continuous variable (range 0 to 6) Primary – based on AF clinic registry   Ablation  0 to 6 months 6 months to 1 year 1 year to 1.5 years 1.5 to 2 years more than 2 years   No (0) Yes (1) No (0) Yes (1) No (0) Yes (1) No (0) Yes (1) No (0) Yes (1) Derived – based on DAD. Anticoagulant 0 to 6 months 6 months to 1 year 1 year to 1.5 years 1.5 to 2 years more than 2 years   No (0) Yes (1) No (0) Yes (1) No (0) Yes (1) No (0) Yes (1) No (0) Yes (1) Derived – based on PharmaNet. Self-Reported Health AFEQT Continuous variable  (range 0 to 100) Primary – based on AF clinic registry    Figure 3-6. Description of Analytic Variables8  3.3.3 Ethical considerations. This study involved secondary analyses of existing data, which meant that the study did not involve collection of data directly from patients. Most concerns about secondary data  8 The AFEQT score was used as the dependent variable while age group, gender, distance to AF clinic, CHADS2, and ablation and anticoagulant at specified follow-up times were used as independent variables.    63  analysis pertain to the potential harm to individuals and issues of consent (Tripathy, 2013). We took several steps to ensure that patient confidentiality was maintained and protected. First, this study was reviewed by the doctoral supervisory committee before undergoing formal research application to the Data Stewards and the ethical review board. Second, the author completed the Tri-Council Policy Statement (TCPS2) ethics course to adhere to the principles outlined related to the secondary use of information for research purposes (Government of Canada, 2019). Since it was not possible to seek consent from every individual in the study, we took additional measures to protect the confidentiality of their data and to safeguard potential identifiable information. For example, once the study was approved by the University of British Columbia Behavioural Research Ethics Board (Certificate H16-03439) and the Data Stewards (through the data access request process via Population Data BC), we entered into an information-sharing agreement with Cardiac Services BC and Population Data BC that described the conditions of collection, use, dissemination, retention, and disposal of the data. We followed the guidelines in accordance with this document to minimize privacy risks and threats to the security of the patients’ information. For example, the de-identified data were accessed and analyzed through the Secure Research Environment (SRE), which is a central server accessible only via an encrypted Virtual Private Network through a firewall and use of a Yubikey® token for authentication. All data files generated from the study were stored in the SRE to safeguard the information. Any file transfers out of the SRE (e.g., figures and aggregate results of analyses) were conducted through a transfer program that automatically screens for identifiable information (e.g., StudyID) and creates an audit trail that is routinely reviewed to minimize accidental releases of data. Data access through the SRE will be terminated on October 15, 2021 and Population Data BC will follow the destruction procedures outlined in the agreement. 64  3.4 Information that Can be Extracted from Linked Data Sources To begin extracting information from linked data sources, substantial data preparation and cleaning were required. Data cleaning refers to the correction of data problems, including missing values, incorrect or out-of-range values, entries that are logically inconsistent with other responses in the database, and duplicate patient records (Gliklich et al., 2014). This process required complex computer programming and knowledge of the following analytical software: SPSS (IBM Corp, 2017) to unzip the *.gz project files and insert break numbers for each variable in the Data Dictionary provided by Population Data BC; SAS (SAS Institute Inc, 2014) to calculate the longitudinal Charlson Comorbidity Index; R statistical software (R Core Team, 2019) to prepare, clean, and analyze the data; and Mplus (Muthén & Muthén, 2017) to model the trajectory of change over time.  The following provides steps that were taken to arrive at the eligible study cohort of 13,113 outpatients with AF who had an initial consultation date occurring between 2008 and 2016. The initial linkage to obtain the study population occurred between Cardiac Services BC (to obtain the AF registry cohort) and Population Data BC (to link the AF registry cohort to the Population Directory that included all the individuals about whom Population Data BC had information) (Population Data BC, 2018b). Cardiac Services BC created the cohort of the approved study population of patients with AF who were referred, between 2008 to 2016, to the AF clinics, and securely sent the cohort with the approved Personal Health Number (PHN), which is an individual-specific health number assigned to all permanent residents of BC (see flowchart in Figure 3-7). 65   Figure 3-7. Flowchart of Eligible Study Cohort                   Population Data BC then used the PHN of the 16,524 patients to facilitate the linkage to their Population Directory, resulting in a total of 16,392 patients (after 52 unlinked9 and 80 missing PHNs were excluded). The PHN of this linked cohort was replaced with a project-specific anonymized study identification number (StudyID), which was used as the basis for analysis. Since this StudyID cohort may include patients who had moved out of the province, we excluded those not registered in the provincial Medical Services Plan (n = 30) during the study period to serve as a proxy measure for residence in BC. Among the 16,362 patients, we excluded those without initial consultation dates (n = 2,902) because the clinicians in the AF clinics noted  9 The Population Directory covers most (but not all) of the BC population so a small number of unlinked records are common.  16 524 patients Before 2008 (n = 0) After 2016 (n = 347) 16 392 patients (Linked by PHNs)  Unlinked PHNs (n = 52) Missing PHNs (n = 80)  16 362 patients  13 460 patients  13 113 eligible patients  Not registered in MSP (n = 30)  No initial consult date (n = 2,903) 66  that those who did not have initial dates may have been referred to the education-only sessions or may have been inappropriate referrals, which meant that they would not have had a consultation for AF management and would not have followed the clinic processes, as shown in Figure 3-2. We further applied the eligibility criteria to those who had an initial consultation date between 2008 and 2016 (by excluding 347 patients who had initial dates after 2016), resulting in 13,113 patients who met the final eligibility criteria of outpatients with AF who had initial consultations between 2008 and 2016. The next step was to combine both the MSP and the DAD datasets with the eligible study cohort data to obtain descriptive information about their comorbidities. We first selected the variables of interest in each of the datasets. For the MSP dataset, we selected “ICD9” (International Classification of Disease, version 9) and the “Service Date” (date in which the service was rendered by a healthcare professional). For the DAD dataset, we selected “DIAGX1” up to “DIAGX25”, which contained the ICD-10-CA (International Classification of Diseases, 10th revision, Clinical Modification) codes, and the year prior to hospitalization “VER_YEAR”. We then combined multiple years of the MSP datasets (i.e., 10 datasets that ranged in years between 2008 and 2016) into one dataset as well as the multiple years of the DAD datasets into one dataset for ease of analysis (see Figure 3-8). Note that the actual column values in the figures are hypothetical and are provided for illustrative purposes.  67   Figure 3-8. Combining Multiple Datasets into One Dataset Note. *DIAGX variable has up to 25 variables.  Second, we identified the presence of chronic conditions using validated algorithms known to have moderate to high predictive validity and sensitivity (Tonelli et al., 2015). For the combined MSP dataset, we used the ICD-9 codes to identify each condition per patient. For example, if the ICD-9 code was ‘428’ (heart failure), we created a column variable “Heart Failure” and assigned a value of ‘1’ if the patient had at least two claims records of ‘428’ that occurred within one year prior to the initial consultation date (e.g., Service Date <= Initial Consult Date & Service Date > One Year Before Initial Consultation Date); otherwise, the variable was assigned a value of ‘0’. Similarly, for the combined DAD dataset, we used the ICD-10-CA codes to identify each condition on any of the “DIAGX1” up to “DIAGX25” columns per patient. For example, if the ICD-10-CA code was ‘I43’ (heart failure), we created a column variable “Heart Failure” and assigned a value of ‘1’ if the patient had at least one hospitalization 68  record of ‘I43’ one year prior to the initial consultation date (e.g., VER_YR (one year prior to hospitalization) = Initial Consultation Year). Then, the MSP cohort (N = 13,113) and the DAD cohort (N = 11,966) were combined to identity heart failure for the eligible study cohort (see Figure 3-9).  Figure 3-9. Identifying Comorbidities in the MSP and DAD Datasets   As illustrated in Figure 3-9, patients in the eligible study cohort were assigned a value of ‘1’ to the “Heart Failure” column if heart failure was identified in the MSP or DAD dataset separately or in both (e.g., Eligible StudyID 1, 2, 3 & 5). In contrast, patients in the eligible study cohort were assigned a value of ‘0’ if heart failure was not identified in both the MSP and the DAD datasets (e.g., Eligible StudyID 4 & 6). If only the DAD dataset (N = 11,966) was used, the remaining 1,147 patients that were not linked to the eligible study cohort were treated as missing; however, the MSP dataset (N = 13,113) included all 13,113 patients and augmented the missing comorbidity information for the final eligible study cohort.  To identify patients who had interventions, we used the same combined DAD dataset and selected the variables containing “ICODE1” to “ICODE20”, which included comprehensive 69  codes for diagnostic, therapeutic, and other associated healthcare interventions known as the Canadian Classification of Health Interventions, and “PDATE1” to “PDATE20”, which were the procedure dates of the corresponding “ICODE” variables. According to the DAD data dictionary, the date of the procedure may be coded only once (“PDATE1”) for single surgical episodes with multiple procedures. Thus, “PDATE1” was used for patients who had multiple procedures; otherwise, the corresponding “ICODE” with its respective “PDATE” was used (e.g., ICODE3 was matched with PDATE3). Based on the algorithm, we counted an intervention (e.g., ablation) if the patient had at least one hospitalization record of ‘1HH59’ on any of the “ICODE1” to “ICODE20” variables with its respective “PDATE1” to “PDATE20” occurring on or before the initial consultation date (e.g., PDATE1 <= Initial Consultation Date). For the intervention category, there was only one procedure date missing for ablation, which was replaced with the date from the AF registry ablation dataset.  To identify patients who had been prescribed medications, we combined two separate datasets within the PharmaNet data source: Dispensing-Event and Health Products. Dispensing-Event contained records of medication dispensing events and included the date of service (SRV_DATE) and the drug identification number (DIN), which is assigned by Health Canada to uniquely identify particular drugs by chemical, dosage, form, and manufacturer. Health Products contain information about the drug identification number (DIN) related to the name (generic drug), dosage, and measurement units (e.g., mls, grams). Both the Dispensing-Event and the Health Products datasets were matched one-to-one based on the drug identification number (see Figure 3-10). 70               Figure 3-10. Matching PharmaNet Datasets  This combined dataset was used to identify patients who had medication exposures at baseline by grouping them into one of six medication categories (i.e., anticoagulants, antiplatelets, beta-blockers, calcium channel blockers, antiarrhythmics, and digoxin) and assessing whether the dispensed date was within one year of the initial consult date (e.g., SRV_DATE <= Initial Consultation Date & SRV_DATE > One Year Before Initial Consultation Date). An example of this process using patients who were grouped into the anticoagulant category is provided in Figure 3-11.  71   Figure 3-11. Assigning the Eligible Study Cohort to the Anticoagulant Variable       In the PharmaNet dataset, patients were assigned a value of ‘1’ in the “Anticoagulant” column if they had a record of either dabigatran, rivaroxaban, apixaban, or warfarin within one year of the initial consultation date (resulting in 8,170 patients in the anticoagulant group). The eligible study cohort (N = 13,113) was linked to this anticoagulant group by one-to-one matching based on the StudyID. There were 15 non-matching StudyIDs with no PharmaNet records during the linkage process, and they were treated as missing. The next step was to prepare the AFEQT questionnaire dataset (see flowchart in Figure 3-12). 72   Figure 3-12. Flowchart of the AFEQT Questionnaire Dataset  Note. n refers to number of questionnaires (not StudyIDs)  The AFEQT questionnaire dataset (n = 16,005) was linked to the eligible cohort by one-to-one matching based on StudyID, resulting in 14,408 questionnaires (after excluding 1,597 unlinked questionnaires). There were 2,121 AFEQT questionnaires completed before the initial consultation date, which were excluded because they did not fit the inclusion criteria. Because the population was limited to those who had an initial consultation date between 2008 and 2016, we excluded 1,803 AFEQT questionnaires that were completed in 2017 or later. Then, 31 duplicates with identical rows of data were identified. There were 27 AFEQT questionnaires with the same survey date but different score items (possibly due to data entry errors although the exact reason for the inconsistency could not be identified). Questionnaires that were 16 005 AFEQT  Duplicates  (n=31) Unattributable  (n=27) 14 408 AFEQT  (Linked based on StudyID)  Unlinked  (n=1,597) 12 287 AFEQT   10 484 AFEQT  10 426 AFEQT  Before initial consult  (n=2,121)  Before 2008   (n=0) After 2017     (n=1,803) 73  duplicates or present with unattributable errors were removed, resulting in a total of 10,426 AFEQT questionnaires available for further analyses.  To gain a general sense of the distribution of the AFEQT questionnaire dataset, we plotted a histogram of the first AFEQT questionnaire completed by time completed, in years, after the initial consultation date (see Figure 3-13).   Figure 3-13. Number of First AFEQT Questionnaires Completed by Time after  Initial Consultation (N=4,412)  The majority of the first AFEQT questionnaires were completed within one year of the initial consultation. We also graphed the distribution of the first AFEQT completed by quarter years broken down by clinic site (see Figure 3-14).   74    Figure 3-14. Number of First AFEQT Questionnaires Completed by Quarter  and by Clinic (N=4,412)  The pattern of the first AFEQT questionnaire collection between Q1 2011 until the end of Q4 2016 shows that the AFEQT questionnaire data collection began tentatively, in line with the opening of the AF clinics, with the amount of data gradually increasing during the first two years, and a period of varying amounts of data collected after Q3 2012. Clinic 5 appeared to have collected the most AFEQT questionnaire data compared with the other clinics while clinic 1 collected the least, possibly due to the volume of patients seen and compliance with the data collection protocol given the demands of the clinics (the rate of AFEQT questionnaire completion in clinic 5 was 39% (1843/4721), clinic 2: 54% (1360/2516), clinic 4: 60% (948/1592), clinic 1: 61% (1160/1893), and clinic 3: 89% (2128/2391)).  75  To get an overview of the time intervals between the AFEQT questionnaire completion, we graphed a time series plot of 40 randomly selected patients from the date of their initial consultation (see Figure 3-15).    Figure 3-15. Number of AFEQT Questionnaires Completed by 40 Randomly  Selected Patients by Time (in years)  The topmost lines in the figure show patients who had two or more questionnaires completed. The patient at the very top (i.e., StudyID 36) had five AFEQT questionnaires completed within 2.5 years of their initial consultation.  The time intervals (in months) of all the AFEQT questionnaires completed during follow-up are shown in Table 3-2.  76  Table 3-2.  Time Intervals between Repeated AFEQT Questionnaire Completions Follow-up Median interval in months between questionnaire completions (IQR) T0 (n = 4,040)  4.5 (1.1-16.5)  T1 (n = 4,412)  6.6 (3.2-13.8) T2 (n = 1,288)  5.6 (2.8-8.8) T3 (n = 403)  4.9 (2.3-7.2) T4 (n = 154)  4.1 (2.3-6.9) T5 (n = 70)  3.4 (1.6-6.2) T6 (n = 30)  6.6 (3.8-9.6) T7 (n = 17)  5.6 (3.6-11.9) T8 (n = 7)  4.2 (3.2-15.5) T9 (n = 4)  3.3 (3.3-3.3) T10 (n = 1)  Note. T0 = Initial consultation; T1:T10 = Number of follow-up visits; IQR = Interquartile range  While there is variability between the follow-up time points, the median interval was about 5-6 months between repeated administrations of the AFEQT questionnaire.  The next step was to construct the time intervals for the time-varying variables (dates of ablation and anticoagulation therapy) by examining the DAD and the PharmaNet dataset. To capture the time-varying nature of when ablation was performed, a similar process of identifying patients who had one or both of the interventions was applied, apart from identifying ablation after the initial consultation (e.g., PDATE > Initial Consultation Date). In the combined DAD dataset, 4,896 ablations were performed after the initial consultation date. We plotted a histogram of the number of ablation procedures performed by time in years after the initial consultation (see Figure 3-16).  77   Figure 3-16. Number of Ablations Performed Following Initial Consultation by Time (in Years)  The procedure was most often performed within one to two years after the initial consultation date. We categorized each of the ablation procedures based on when they were performed relative to the specified follow-up period (e.g., within 6 months (> 0 to ≤ 0.5year), 1 year (> 0.5 to ≤ 1), 1.5 years (> 1 to ≤ 1.5), 2 years (> 1.5 to ≤ 2), and > 2 years)). To capture the time-varying nature of anticoagulation therapy, the time intervals were divided into similar categories. The number of patients in each of the follow-up periods are described in the findings section.  3.5 Accommodation for Missing Data  One of the major challenges with analyzing registry-based PROMs data is the bias that may be introduced through the presence of missing data because the characteristics of patients who do or do not complete PROMs at various times (e.g., when they die, miss a scheduled clinic visit, or fail to respond to a questionnaire) may be different from those who complete all their follow-up assessments. While issues of missing data are well-documented, the most common 78  practice of accommodating missing data is to simply delete all cases with missing value (i.e., listwise deletion) (Pampaka, Hutcheson, & Williams, 2016). Part of the reason for this practice is that many of the standard statistical analysis software packages default to listwise deletion for most modelling procedures (Kang, 2013). The disadvantage of listwise deletion is that it could lead to a loss of large amounts of potentially useful information and to selection bias; thus, its use is generally not recommended unless those with missing data are less than 5% of the cases (Fayers & Machin, 2007). Pairwise deletion, which attempts to minimize the loss that occurs in listwise deletion by deleting a case when it is missing a variable for a particular analysis but is otherwise retained, can also create problems because different sets of patients contribute information at different instances and thus make comparisons unclear (Fayers & Machin, 2007).  To appropriately accommodate missing data, it is important to first understand the three patterns of missingness: missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR) (Rubin, 1976). In MCAR, the pattern of missingness is equally likely for any individual, meaning that missing values would be randomly scattered throughout the data for reasons unrelated to the study. MCAR is an ideal pattern because these missing cases can often be excluded without biasing the results; however, attributing most missingness to MCAR is often difficult to support. An example of MCAR would be patients who did not complete the PROM because they moved to another geographic location or staff who forgot to administer the PROM.  In MAR, the pattern of missingness depends on some observed variable in the dataset. MAR is a more likely scenario in most PROM-based studies because patients who are in good health (e.g., lower CCS-SAF with fewer comorbidities) are expected to have less missing information than those who are not doing as well (e.g., higher CCS-SAF with multiple 79  comorbidities) (Biering, Hjollund, & Frydenberg, 2015). Finally, the data are missing not at random (MNAR) if they are neither MCAR nor MAR.  In MNAR, the pattern of missing data depends on the missing variable itself, even when observed data are taken into account. For example, patients may have had high initial PROMs and low symptom burden of AF, but during follow-up the same patients may suddenly worsen and drop out. If these patients were lost to follow-up and there was no observation in the middle to reflect the worsened condition, the missingness is MNAR. Missing data that are MNAR are the most problematic because the data needed to test the assumption are not directly available and standard statistical analysis models do not take into account this pattern of missingness (Dong & Peng, 2013; Du, Hahn, & Cella, 2011).  It is important to note that all three types of missing data may be present in a given data set and even for a given variable. Thus, we provide the patterns of missingness for each variable and the appropriate methods for handling missing data based on the patterns observed. Missing data ranged from 0% to 15% depending on the variable. A descriptive table of all of the missing data for each variable in the dataset is provided in Appendix D.  The frequency of missing data patterns for the study variables is shown in Figure 3-17.  80   Figure 3-17. Frequency of Missing Data Patterns Note. Blue is observed, red is missing values. Measurement occasion = data collected on an individual at a particular moment in time.  The left vertical axis provides the number of unique measurement occasions for each missing data pattern while the right vertical axis provides the number of variables with missing data in that pattern. For example, the first row shows 10,150 observations with 0 occurrences of missing data on any of the variables. The second row shows 1,643 observations with missing data on only one variable (CHADS2; identified by the red cell). The x-axis provides the total number of missing values for each variable. We further tested whether the data were MCAR by conducting Little’s MCAR test, which indicated that the pattern was not MCAR (χ2 = 163.68, p < .001). 81  We then examined the characteristics of the patients who responded to the AFEQT questionnaire (respondents) compared with those who did not complete a questionnaire (non-respondents) during the study period (see Appendix E). Patients who were non-respondents (43.3%) differed from the respondents (56.7%) by living further away from the clinic, being at higher risk for stroke (higher CHADS2 scores) and more symptomatic (higher CCS-SAF scores). Although there may be various reasons why a large percentage of patients did not complete the AFEQT questionnaire, such as being too ill to complete it, the most cited reason by the clinicians at the clinics was that there was an administrative failing (e.g., the staff forgot to administer the questionnaire or were too busy to ensure entry of the results into the registry). These characteristics and other differences show that complete case analyses under MCAR assumptions would not have been appropriate because the missingness was related to variables observed in the dataset (e.g., distance to clinic, CHADS2, and CCS-SAF).  We also examined the missingness of the AFEQT questionnaires at the item level. Of the 10,426 AFEQT questionnaires, there were 7,033 questionnaires with all items complete while the remaining 3,393 questionnaires had one or more missing items. The missing items included Q12 (“Doing vigorous activities such as lifting or moving heavy furniture…”), which had the greatest number of missing responses, at 10.8%, and Q5 (“Ability to have recreational pastimes…”) having the least at 2.9%. A stacked bar graph of the distribution of each of the AFEQT questionnaire items is shown in Figure 3-18.  82   Figure 3-18. Distribution of the AFEQT Questionnaire Items  Note. Q1:Q4 = Symptom; Q5:Q14 = Daily activities; Q15:Q18 = Treatment concerns.  A sizeable percentage of the sample reported “Not at all” bothered or “Hardly any difficulty” to the symptoms, daily activities, and treatment concerns domains (ranging between 37% and 67%). Among the items, questionnaire item 3 (Q3 “A pause in heart activity”) showed that most responses were in the “Not at all” or “Hardly bothered” response categories while questionnaire item 12 (Q12 “Doing vigorous activities…”) showed the fewest responses in these categories. All of the items were positively skewed in the direction of better self-reported health (skewness ranging from 0.2 to 1.7). Kurtosis ranged from -1.5 to 1.8. For the AFEQT questionnaire items, we tested the MCAR assumption by conducting Little’s MCAR test, which indicated that the pattern was not MCAR (χ2 = 11,736.09, p < .001). This further supported the 83  conclusion that the missing items would not be appropriately handled with listwise or pairwise deletion, but that imputed data (using multiple imputation) were needed (Pedersen et al., 2017). We noted at the outset that while the methods of this study were exploratory in nature, the results of these approaches allowed us to have imputed datasets that provided a more accurate representation of the patients’ characteristics and their responses to the AFEQT questionnaires, an advantage that would not have otherwise been possible.  To better handle the missing data, we included the information of patients who were lost to follow-up because they had died. This was done by one-to-one matching of the Vital Statistics Deaths dataset (“date of death” variable) with the eligible cohort based on StudyID. In the literature, researchers have excluded patients who died during follow-up, assigned a score of 0 or the minimum possible score on a scale, or provided an indicator of death to be included in their imputation model (Bell & Fairclough, 2014; Biering et al., 2015). If we excluded from the analysis patients who died during follow-up, the results may have been too optimistic; on the other hand, assigning 0 to the AFEQT questionnaire items would have meant that patients who died were recorded to have experienced severe symptoms and extreme difficulty with daily activities, which did not make sense unless the scales were explicitly anchored to death (as is the case with utility measures such as HUI3). Thus, we opted to include a “death” indicator in the imputation model for patients who could not complete the AFEQT questionnaire because they had died. A total of 932 patients died during follow-up and after their initial consultation (452 respondents and 480 non-respondents).  The current recommended approach to the handling of missing data is to use multiple imputation with auxiliary variables (Ayilara et al., 2019; Little, Jorgensen, Lang, & Moore, 2014). Multiple imputation works by creating multiple copies of a dataset with missing values 84  replaced by imputed values. Auxiliary variables (variables that are potential correlates of missingness) are often included during the imputation to increase precision and reduce the bias of the imputed values (Ayilara et al., 2019). For the imputation of missing data, there are generally two widely available methods: multivariate normal imputation, which imputes data based on probabilities constructed from continuous values (Schafer, 1997) and imputation by chained equations, which imputes data on a variable-by-variable basis (also known as fully conditional specification) (Raghunathan, Lepkowski, Ryn, & Solenberger, 2001; van Buuren, Brand, Groothuis-Oudshoorn, & Rubin, 2006). Although the two methods are based on different theoretical assumptions and different computational methods, both approaches lead to comparable results when distributional assumptions are adequately met (Lee & Carlin, 2010; Mistler & Enders, 2017). Lee and Carlin (2010), however, noted that imputation by chained equations may provide a more flexible method because it does not rely on the assumption of a multivariate normal distribution, which is often not plausible with categorical variables. Imputation by chained equations runs a series of regression models whereby each variable with missing data is modelled conditional upon the other variables (Azur, Stuart, Frangakis, & Leaf, 2011). This means that each variable can be modelled according to its distributions with categorical variables (e.g., binary, ordinal, and nominal) modelled using logistic models and continuous variables modelled using linear models. For this study, the imputation by chained equations was implemented using the ‘MICE’ package (an acronym for multivariate imputation by chained equations) in the R software (Buuren & Groothuis-Oudshoorn, 2011).  Since our longitudinal dataset was organized in a long format (having multiple rows per individual) with unequal time intervals (i.e.,  there are multiple (level 1) measurements for each individual), we used the multilevel variant of multiple imputation using the ‘miceadds’ package 85  that provides additional features (Robitzsch, Grund, & Henke, 2017). Accounting for missingness with a multilevel structure is a relatively recent development in which repeated measurements in PROMs (known as level 1) are nested within individuals (known as level 2) (Enders, Mistler, & Keller, 2016). Among the available imputation methods, we chose the recommended predictive mean matching algorithm (De Silva, Moreno-Betancur, De Livera, Lee, & Simpson, 2019) and accounted for the nested structure by specifying individuals into clusters and applied the matching algorithm to level 1 variables (e.g., Q1-Q12) and level 2 variables (e.g., CHADS2 and ablation therapy status) (Van Buuren, 2018). The predictive mean matching is a standard multiple imputation technique that has been found to be robust in a variety of scenarios and with variables of different types (Kleinke, 2017). The predictive mean matching works not by replacing missing values with predicted values from the imputation model, but by selecting a set of observed values with the closest predicted value as the missing one and imputing the missing data by a random draw from that set (Van Buuren, 2018). Since the matching is based on observed values, the method can be applied to variables with skewed distributions and non-continuous variables (e.g., binary variables) (Van Buuren, 2018).  To ensure that the imputation model has the most information possible, we used all available variables in the dataset as auxiliary variables (see Appendix F for the auxiliary variables and applied algorithm). For the number of imputed datasets, we followed the recommendations by Graham et al. (2007) of using 20 imputed datasets for 10% to 30% missing information. The 20 imputed datasets with five iterations for each imputation (the default method) were successfully run without any error messages on the “respondent” sample (those who completed a questionnaire).  86  After obtaining 20 imputed datasets organized in a long format without any missing values, the next step was to prepare these datasets for modelling in wide format (having only one row per each individual). However, significant issues to address were the follow-up times, which were unequally spaced, and the number and timing of the follow-up measurements, which differed. As a result, when these datasets were transformed into wide format, there were disproportionate amounts of additional “missing” values that were never intended to be collected. In this case, instead of conducting multiple imputation again, another advanced data technique called full information maximum likelihood (FIML) is recommended (Little et al., 2014). FIML does not impute missing values but, similar to multiple imputation, it uses all information from variables included in the statistical model to compute parameter estimates. This method involves sequentially estimating different parameter values until the fit to the data is optimal. Since power is maximized using all available data from the variables in the model, FIML has been shown to produce unbiased parameter estimates and standard errors under MAR and MCAR assumptions (Schafer & Graham, 2002). This step was conducted using the default feature in Mplus during the modelling process (Muthén & Muthén, 2017). To visualize the plausibility of values, we checked the density estimates for the marginal distributions (total probability densities of the 20 imputed datasets) of the observed data (in blue) in comparison to the imputed data (in red) known as density plots (see Figure 3-19). Although density plots are used for continuous variables (e.g., CHADS2 and CCS-SAF), we plotted these for the categorical variables to get a general sense of the imputed values.  87   Figure 3-19. Density Plots of Covariates with Imputed Values (N = 7,439)                                     Note. Blue lines are observed values. Red lines are imputed values.   Since some variables had only a few values imputed (e.g., CCS-SAF variable with 12 missing values per dataset and the antiplatelet variable with 15 missing values per dataset), the peaks in the categorical observed values (in blue) may show some discrepancies with the imputed ones (in red). While we do not expect the imputed values to be identical to the observed ones (due to the MAR assumption, the use of auxiliary variables, and the multivariate normality assumption that leads to imputed data being naturally more continuous), one would expect the shape of the distributions to be relatively similar (e.g., with similar spread and the modes occurring at the same location).  The next set of density plots include each of the AFEQT questionnaire items; they show a similar trend with the distributions of the observed and the imputed values being more or less 88  similar (with the possible exception of Q3 due to the skewness in the responses) (see Figure 3-20).  Figure 3-20. Density Plots of AFEQT Questionnaire Items with Imputed Values  Note. Blue lines are observed values. Red lines are imputed values.   After imputation, the subscale scores for symptoms, daily activities, and treatment concerns as well as the total score (i.e., the AFEQT score) were constructed with the imputed values using the post-processing (also known as passive imputation) feature in the ‘MICE’ package. The post-processing is a convenient feature that enables the calculation of the total score and its subscale scores after imputing each item in one step without needing to calculate them for each of the imputed datasets separately. The 20 imputed datasets were saved in a new file to begin the modelling process.  89  In summary, we used the latest recommended approaches to handle the missing data in our analyses. Specifically, we used multiple imputation to account for the uncertainty about the missing data by creating a total of 20 different plausible imputed datasets. In particular, we used the multilevel variant of multiple imputation to correct for biases associated with the multilevel structure by specifying level 1 and level 2 variables. To increase the accuracy of the imputed values in the imputation model, we included all available auxiliary variables, including the variable that noted if a patient had died during the study period. We also used the FIML method to account for variation in the frequency and timing of follow-up data. Unlike multiple imputation, FIML does not impute missing data but rather estimates parameters from variables already included in the analysis. In this way, we were able to make full use of all available data to minimize the biased estimates and standard errors typically caused by missing data.  3.6 Data Analysis Strategy We have provided a detailed overview of the setting, data sources, and data preparation as preliminary steps toward addressing the research question. What follows is a description of our longitudinal data analysis strategy and additional challenges (e.g., how to best represent variability in frequency and timing of measurement occurrences, shape and individual differences in trajectories, and how to identify factors that explain variability in individual trajectories) to answer our research question. Our analytical objective was to explore how PROM scores in patients with AF change over time and factors associated with this change by first examining the variation in individual trajectories of change and subsequently determining which identified patterns were associated with age, gender, distance to clinic, CHADS2 score, and treatment (ablation or anticoagulation therapy) at follow-up periods.  90  3.6.1 How to best analyze longitudinal PROMs data. Various statistical methods are available to analyze longitudinal data when PROMs are collected at multiple follow-up times. In clinical studies, longitudinal analysis is often used to describe patterns of change over time among groups (e.g., different treatment groups), and to explain how and why group changes may or may not occur (Ghisletta et al., 2015; Schober & Vetter, 2018). The most common method in longitudinal analysis has been to use repeated analysis of variance (ANOVA), in which the mean outcomes measured at different time points in the same patients are compared under different treatment conditions (Gibbons, Hedeker, & DuToit, 2010). Despite its appeal, this method has several limitations that may be of limited use with registry data. For example, it is quite common in the clinical setting for each patient to complete a different number of PROMs at different times. This may be due to patients being discharged at different times, when their care objectives have been met, or simply having different schedules of availability. The disadvantage of ANOVA is that patients who have unevenly spaced and an unequal number of measurement occurrences are simply removed from the analysis (Gibbons et al., 2010). Yet if patients who are followed for a longer period of time with more measurement occurrences are different from those who are followed for a shorter period of time with fewer measurement occurrences, then there is a risk of bias because the results are not representative of the original target population. Another related issue is that ANOVA does not allow time-varying covariates (i.e., characteristics that change across time such as changes in treatment) into the model, which are often essential to modelling dynamic relationships between predictors and outcomes (Gibbons et al., 2010). In addition, ANOVA focuses on the comparison of group means and provides no information regarding how individual patients change across time (Gibbons et al., 2010). Given the limitations of ANOVA 91  in analyzing complex longitudinal data (e.g., irregularly spaced measurement occasions and missing time points), a host of new, powerful analytic methods have been developed in recent years (Curran, Obeidat, & Losardo, 2010).  A Growth Curve Model (CGM) is a newer, more flexible approach that refers to a wide array of models that estimate the between-person (inter-individual) differences in within-person (intra-individual) change, the shape and direction of change, and the determinants that are associated with the change (Grimm, Ram, & Estabrook, 2017). Variants of GCM are known by different names including general linear mixed models, random effects or random coefficients models, or hierarchical linear models. These models have become an important analytical approach in diverse fields, including education, psychology, and health sciences (Heck, Thomas, & Tabata, 2013). Several characteristics are common to these models. First, GCM allows for the estimation of fixed and random effects. Fixed effects refer to the mean of the trajectory of all individuals within the sample. Random effects refer to the variance of the individual trajectories around the group means. Thus, GCM is able to describe each individual patient’s trend across time and the variation of this individual trend from the average. In the clinical context, this information is of great interest to understand how patients vary in their treatment responses in their baseline and their rates of change (Gerig, Fishbaugh, & Sadeghi, 2016). Second, GCM does not require balanced data across different times, which allows it to handle unequal sample sizes, inconsistent time intervals, and missing data (Curran et al., 2010). Third, GCM estimates the change parameters with greater precision as the number of time points in measurement increase (Shek & Ma, 2011). This improves the reliability of the growth parameters by reducing the standard errors of the intra-individual change (Shek & Ma, 2011). 92  The most basic growth model has both fixed and random effects to best capture the individual trajectories of change over time (Curran et al., 2010). For a linear trajectory, the fixed effects are estimates of the mean intercept (i.e., baseline) and mean slope (i.e., rate of change) of the entire sample; the random effects capture the between-person variability around the estimated mean intercept and slope, which allows the growth trajectory to differ randomly for each individual. Since there are two general approaches used to fit a basic growth model, we provide a brief background to the two approaches, and consider their strengths and limitations.  The first approach is to fit a growth model within the multilevel modelling (MLM) framework, which treats the repeated measures as univariate (also known as the “long” data format) (Bryk & Raudenbush, 1987; Raudenbush & Bryk, 2002). The multilevel model was originally developed to allow for nesting of multiple individuals within a group and can be applied to multiple repeated measures nested within individuals. In MLM, the fixed and random effects correspond to a hierarchy of levels with the repeated, correlated measurement occurring among all of the lower level units for each particular upper level unit (Heck et al., 2013). For example, the level 1 model refers to an intra-individual change model (e.g., repeated measurements over time) and describes variations within individuals over time. The level 2 model captures whether the rate of change varies across individuals (Heck et al., 2013). Taken together, these levels capture the general characteristics of change, with repeated measurement occurrences (level 1 units), nested within subjects (level 2 units). The predictor variables can be included at all existing levels with time-varying covariates at level 1 and time-invariant covariates at level 2.  93  Since an understanding of the linear regression equation is essential to understanding multilevel model building approaches, a brief explanation of the two-level linear model is provided. The below equation describes a standard linear regression analyses model:   Yti = b0i + b1Xti + eti   (1) In the above equation, Yti is the predicted outcome score for individual i at time t (e.g., a PROMs score). Xti is the observed explanatory variable for individual i at time t. b0 is the intercept, which is a constant when the value of the Xti variable is 0. b1 is the regression coefficient (slope) given to multiply the Xti to predict the outcome variable, which is interpreted as the amount of change in Yi per unit change in Xti. eti is a residual error. The Level 1 component of a multilevel analysis “fits” separate regression lines for each individual denoted by i, producing estimates of the intercept (b0i) and slope (b1i) for all Level 2 units (e.g., patients) as shown below:   b0i = G00 + U0i   (2) b1i = G10 + U1i  The equation above shows two estimates: (a) G00 and G10 are fixed effects, which depict the mean of the Level 1 coefficients across Level 2 units, and (b) U0i and U1i are random effects, which reflect how much dispersion there is in the Level 1 coefficients across Level 2 units. For the equation predicting initial status (b0i), G00 represents initial status across all individuals, and U0i is the Level 2 regression residual indicating the difference between that particular individual’s initial baseline status. Likewise, for the equation predicting slope (b1i), G10 represents the strength and direction of this association averaged across all individuals and 94  provides an estimate of the bivariate association between the predictor and outcome variables. U1i is the Level 2 regression residual indicating the difference between that particular individual’s slope and the average slope of the sample. Depending on the study, this model can be extended to higher levels (e.g., level 3 units across clinics). After the basic unconditional model has been specified, variables of theoretical or clinical interest can be included as predictors of either intercepts, slopes, and/or shape parameters. The aim of this step is to examine whether various characteristics (e.g., clinical or socio-demographic variables) can help to explain the variability in the temporal change in the PROMs scores. The inclusion of predictors in the model is often called a conditional growth or mixed effects model because the fixed and random effects are now “conditioned on” the predictors (Curran et al., 2010).  The second approach is to fit a growth model within the structural equation modelling (SEM) framework (Bollen & Curran, 2006; McArdle, 1988; Meredith & Tisak, 1990). Within the SEM approach, the b0i (intercept) and b1i (slope) from equation 2 are represented as latent variables rather than as regression coefficients – hence the name latent growth model (LGM). Thus, the only difference between MLM and LGM is the way time is represented in the model (Hox & Stoel, 2005). In MLM, time is introduced as an individually-varying explanatory variable where an additional variable is added to the model, whereas in the LGM, it is introduced via the factor loadings to be constrained to represent time. The consequence of this is that MLM treats repeated measures as observations of the same variable or univariate (“long” data format), whereas LGM treats each repeated measure as a separate variable or multivariate (“wide” data format) (Stoel & Wittenboer, 2003).  The general LGM with three repeated measures, which is considered to be a minimum for the study of change (Duncan & Duncan, 2009), is shown in Figure 3-21. 95   Figure 3-21. Latent Growth Model with Three Measurement Occasions of a Continuous Outcome  The Yi0 – Yi2 represents the outcomes at the varying time points for individual i. These variables are used as indicators of latent variables that represent different aspects of individuals’ change known as latent growth factors. There are two latent variables (sometimes called random coefficients). The first is the latent intercept, which represents the level of the outcome when time is zero (baseline), and thus the intercept factor loadings are all fixed to one as a constant. The second is the latent slope, which represents the change in the outcome over time. The slope loadings are fixed to reflect the time since initial baseline status.  In a basic LGM, each individual has an estimated intercept and slope, which are allowed to vary across individuals. This variability across individuals is estimated as the variance of the latent intercept and slope and is depicted as a double headed arrow that points to the same variable. The intercept and slope are shown to covary and modelled in the figure to show how individuals’ start values relate to their rate of change. The latent variables also have means to reflect the average of all individuals’ intercepts and slopes. In addition, individuals have their 96  own deviations from these means at each time point known as residual/error variance, which are depicted as εi0-εi2. Although the two approaches are similar in many ways, there are key differences. Table 3-3 compares differences between the MLM and the LGM approaches, along with the traditional ANOVA method.     97  Table 3-3.  Comparisons between Longitudinal Data Analysis Methods  ANOVA Multi-level model Latent growth model Method classes Uses intercept only approach Uses mixed effect approach, which treats repeated measures as univariate (“long” data format) Uses latent-curve model approach, which treats repeated measures as multivariate (“wide” data format) Analysis Strategies Assesses group means over time  Each post-baseline time period can be compared to all adjacent time periods Allows higher levels of nesting (e.g., repeated measures nested within each individual, and individual nested within clinics)  More limited in the estimation of comprehensive measurement models  Allows for specification of latent variables representing growth curve parameters (intercept and slope)  Very flexible modelling that allows for factor loadings for the growth parameters to be freely estimated to correspond to the data structure Nonlinear growth Cannot model nonlinear growth For specific forms of nonlinear growth (cubic and quadratic), the MLM is preferred because it can estimate the model directly For complex nonlinear growth, the SEM approach is preferred because the values of the slopes can be estimated from the data (“free curve”) rather than fixed to specific values Small samples Can be used with few people Accommodates growth models with few people Generally requires a large sample Time-unstructured data Cannot handle time-unstructured data (complete cases only) Minimally distinguishes between time-structured and time unstructured data Handles time-unstructured data in most cases, provided that additional steps are taken Time-varying covariates Not allowed Allowed  Allowed Model fit t-tests or post-hoc test Comparative model fit indices (e.g., AIC, BIC) Global fit indices (e.g., RMSEA, CFI, and TLI except when fitting individually varying time metric as factor loading matrix varies across individuals)  Residual structures Assumes equal residual structures Applies preprogrammed residual structures but is less flexible Highly flexible and requires all relations to be manually specified Note. ANOVA = Analysis of variance; AIC = Akaike’s Information Criterion; BIC = Bayesian Information Criterion; RMSEA= Root Mean Square Error of Approximation; CFI = Comparative Fit Index; TLI = Tucker-Lewis Index.   98  As reflected in Table 3-3, the limitations of repeated measures ANOVA in handling time-unstructured data with missing data make it unsuitable for the analysis of real-world longitudinal data compared with the growth curve model within either the MLM or the LGM approaches. The MLM may be suitable for straightforward models (e.g., a simple growth trajectory with smaller samples), time-unstructured data, or data with multiple levels of nested structure (McNeish & Matta, 2018). In contrast, LGM may be more appropriate for complex growth models (e.g., complex nonlinear growth) that require more flexibility and options to compare groups and to constrain parameters within and across groups (McNeish & Matta, 2018). In practice, however, the choice between the MLM and LGM approaches may not be straightforward because the decision may come down to situations whereby one approach is easier to implement compared with another (Hox & Stoel, 2005). In addition, recent developments in statistical models and software have blurred the boundaries between the two approaches. For example, advanced structural equation modelling software such as Mplus (Muthén & Muthén, 2017) incorporates some multilevel features whereas some multilevel modelling software such as HLM (Raudenbush, Bryk, Cheong, & Congdon, 2019) offers linear relations between the growth parameters. While we considered the above longitudinal data methods including ANOVA, MLM, and LGM, there were several additional challenges (discussed in the following sections) that led us to apply a new modelling approach known as growth mixture modelling, which has the ability to identify multiple subgroups of trajectories and account for individually-varying times of observations. 99  3.6.2 How to best represent variability in the frequency and timing of measurement occurrences. A key step in fitting an appropriate growth curve model is to first identify the metric of time that would be most useful to understand the within-person change processes (Grimm et al., 2017). Since we had the exact date of when the AFEQT questionnaire was collected for each patient, we had the option of either modelling time as the number of follow-up visits (tied to the frequency of AFEQT questionnaires completed), as the difference between the initial consultation and survey dates (tied to the exact time of when AFEQT questionnaires were completed), or possibly both. If the time metric was indexed by the number of follow-up visits, time would have represented a set of clinic-related processes that described the extent of exposure to the clinic. For example, patients receiving more resource-intensive treatment such as ablation would have more frequent exposure to the clinic because of the subsequent follow-up pathway. To examine this time metric in our dataset, we plotted the overall trend of the AFEQT scores, by the number of follow-up visits (see Figure 3-22).  100      Figure 3-22. Smooth Fitting Curve of the AFEQT Questionnaire Scores  by Number of Follow-up Visits  Although the curve appears curvilinear, there is a general upward trend in the AFEQT scores over the number of follow-up visits. The shaded area is the 95% confidence interval of the curve, which gradually expands as the data points become increasingly sparse.  In contrast, if the time metric was indexed by continuous time (e.g., years), time would have represented a set of processes prompted by the initial meeting with the clinical staff. For example, after initial consultation, the patients may have been asked to either return to the clinic within a certain time frame or to visit the clinic at shorter intervals because their symptoms may not have been adequately controlled. For comparison of this time metric, we plotted the overall trend of the AFEQT questionnaire scores by time (in years) with the number of AFEQT administrations (see Figure 3-23). 101   Figure 3-23. Smooth Fitting Curve of the AFEQT Questionnaire Scores  with the Number of AFEQT Questionnaire Administrations by Time   The top plot shows a rapid increase in the AFEQT scores until about one-year of follow-up then a gradual decrease over time. The bottom plot shows that most of the AFEQT questionnaires were administered within 6 months. The gradual reduction in AFEQT scores over time makes sense from a clinical perspective because the patients who were seen for longer periods, by the clinics, may have had poorer health, and they also experienced the effects of aging.  Both Figures 3-22 and 3-23 suggest that there may have been multiple processes involved with different trajectories based on the specific time metric used. For example, if we modelled time based on the number of follow-up visits, we see that the patients’ health trajectories were 102  slightly improving over time. However, if we modelled time based on time elapsed since the initial consultation, we see that the patients had a rapid increase, then a gradual deterioration in health over time. Nevertheless, in practical applications, it may be difficult to completely separate the number of follow-up visits from the time passed since the initial consultation. For example, any changes in the AFEQT questionnaire scores may have been influenced by both clinic-related processes, when the patients followed a certain treatment protocol, and time-related processes, when the sicker patients may have been followed by the clinic for a longer period of time. Thus, the metric of time had to account for the individually-varying times of observations, which refers to the situation in which each individual was followed-up at different intervals (e.g., 3 week→10 weeks→4 months) rather than even intervals (e.g., every 6 months). In other words, the model had to represent the variability in frequency and timing of the measurement occurrences.   With the LGM approach, the major issue became the complexity introduced in trying to fit each patient who was measured at varying frequency and intervals. Since LGM treats repeated measures as a multivariate (i.e., each time point requires its own unique column in the data), it was unclear which value parameters needed to be constrained and how to organize the columns within the data (McNeish & Matta, 2018). While one option was to equivalently categorize the time periods for all patients, studies have shown that coarsening continuous time periods when they are not can bias the parameter estimates (Aydin, Leite, & Algina, 2014; Singer & Willett, 2003), which in turn can magnify the bias depending on how variably spaced the measurement occurrences are (Coulombe, Selig, & Delaney, 2016; Singer & Willett, 2003). In contrast, the MLM approach was far more straightforward with respect to implementing the unique 103  measurement occurrences because it processes the data in the “long” univariate format with time explicitly featured as a variable in the model (McNeish & Matta, 2018).  While MLM was also considered, another issue was its underlying variable-centred approach. The variable centred-approach is the traditional and dominant method in the social sciences used to explain relationships between variables of interest in a population (Howard & Hoffman, 2018). Variable-centred approaches (e.g., regression-based analyses) underlie a view that people are a medium through which predictor variables affect outcomes and assume that trajectories are similarly experienced by all individuals (an assumption that underlies most clinical research) (Laursen & Hoff, 2006). While these types of analyses provide a single set of parameters that may reveal general health improvements over time (e.g., whether a treatment, on average, works for most people), they can mask subgroups of trajectories inherent to each clinical population, and therefore may not be representative of all patients (Owen et al., 2015).  In the context of AF, it seemed unlikely that all patients with AF come from a single population that follow a single health trajectory, and as such, important subgroups of patients may demonstrate distinct patterns of change. Thus, it was clear that a more flexible modelling method was needed that could address both the subgroups of trajectories in patients with AF as well as the variability in the frequency and timing of the measurement occurrences.  3.6.3 How to best represent the shape of the trajectories. To develop an appropriate model that represented the shape of the trajectories from multiple subgroups, we plotted a random subsample (n = 1,143) of individual growth trajectories for three sequential follow-up visits (see Figure 3-24). 104   Figure 3-24. AFEQT Questionnaire Scores by Follow-Up  The line graph comparing the AFEQT scores by follow-up shows wider variability in the intercepts compared with the slopes. For example, some patients had initial scores as low as about 20 points while others had initial scores as high as about 100 points, and those who had low scores appeared to have similar linear increases over time while those who had high scores remained relatively unchanged. Based on Figure 3-24, it was difficult to consider the variable-centred approach to modelling (e.g., MLM), which assumes a single health trajectory. Rather, we conjectured that at least two different groups would emerge. For one group, we anticipated that the AFEQT questionnaire scores would be low at the initial assessment and then gradually improve over time. For another group, the initial level of the AFEQT questionnaire scores would be high and they would remain relatively unchanged or slightly decrease.  An emerging statistical approach to empirically classify individuals based on their growth trajectory, first introduced by Muthén and Shedden (1999), is known as Growth Mixture 105  Modelling (GMM). Unlike the variable-centred approaches that provide a single set of population parameters, GMM provides many sets of parameters. GMM is considered to be a patient-centred modelling approach because it underlies the view that people are the agents that affect the outcome with predictor variables deemed to be properties of the individual, and assumes that trajectories are different across individuals (Laursen & Hoff, 2006). GMM works by assigning individuals who share similar patterns of scores into unobserved subgroups called latent classes. These latent classes are based on probabilities in which each individual receives fractional membership in all classes to reflect varying degrees of precision in their classification (Berlin, Parra, & Williams, 2014). In addition, GMM allows the loading of individually-varying times of observation at each measurement occasion, allowing each individual’s slope loadings to have unique time points. For example, the latent slope can be specified to represent individual specific times at which their AFEQT questionnaire was collected. The general GMM with three repeated measures is shown in Figure 3-25.  Figure 3-25. Growth Mixture Model with Three Continuous Outcomes 106  GMM extends both the LGM and MLM approaches to describe differences among individuals. GMM extends the LGM approach because it incorporates a categorical latent variable, which represents mixtures of subgroups where membership is not known a priori but must be inferred from the data. In this way, latent classes are operationalized as collections of individuals who follow approximately the same trajectory. For example, patients with AF classified into different latent subgroups may have different trajectories of change with their own initial status (intercept) and change rates (slope). GMM also extends the MLM approach because the slope loadings are constrained to allow nesting of time observations within individuals labelled as ti0 – ti2, which allows each individual’s slope to represent the unique times at which their AFEQT questionnaires were collected. For example, the individually-varying times of observation can be included in Mplus by using the TSCORES command with an AT option (see Example 6.12 in the Mplus User’s Guide; Muthén & Muthén, 2017). However, since the slope loadings vary across individuals, Mplus does not provide traditional SEM goodness-of-fit statistics or the mixture model statistical comparison tests such as the Lo-Mendell-Rubin Test (LMRT) (Lo, Mendell, & Rubin, 2001) and the Bootstrap Likelihood Ratio Test (BLRT) (McLachlan & Peel, 2000). Instead, the comparative fit of the models is primarily assessed using the Akaike’s information criterion (AIC), Bayesian information criterion (BIC), and the sample-size adjusted BIC (SABIC). The AIC and BIC are both based on maximum likelihood estimates (-2 log likelihood) plus a penalty function for the number of estimated parameters, but the BIC penalizes for model complexity more heavily. The sample size-adjusted BIC is similar to BIC except that it further adjusts for sample size.  Among the fit indices, the general principle of selecting the preferable GMM is to choose the lowest BIC and SABIC values (lower values indicate better model-data fit) because they 107  provide more accuracy in identifying the latent classes compared with the AIC (Chen, Luo, Palardy, Glaman, & McEnturff, 2017). Other evaluation criteria include entropy, which is a summary index of the accuracy of latent class assignments for all individuals (Celeux & Soromenho, 1996). While there is no conventional threshold value for entropy, values closer to 1 indicate greater accuracy of latent class assignment (Celeux & Soromenho, 1996). Overall, GMM is extremely flexible and allows the estimation of multiple models by changing the slope factor loadings, adding additional latent variables (e.g., quadratic functions) or imposing various constraints. For example, one approach is to constrain the intercept and slope variance to zero (referred to as latent class growth analysis) or just the slope (referred to as restricted random effects). It is also common to specify homoscedastic or heteroscedastic models by constraining residual variances across time points or classes. However, this flexibility comes at a price because the framework is built on assumptions that should be realistic for the data, and to estimate the model, all relations between observed and latent variables have to be specified (Lubke & Luningham, 2017). While the exploratory characteristics need to be taken into account when interpreting these models, such an analysis can provide a great deal of insight into the processes and interrelations among different factors when the traditional approach of summarizing an entire population together with a single set of parameters (i.e., assuming population homogeneity) is unrealistic.  3.6.4 How to best represent individual differences in trajectories. To begin an initial examination of possible differences in the patients’ trajectories, a series of GMMs were specified and subsequently estimated. According to Gilthorpe et al. (2014), the selection of a suitable GMM with the “correct” number of latent classes is heavily influenced by the method used to parameterize the random effects within the model. For example, one 108  approach is to freely estimate growth factor variances and covariances for each latent class (referred to as unrestricted GMM). In contrast, an extreme form of model parsimony is to constrain all growth factor variances to be equal, referred to as latent class growth analysis (LCGA) (Jung & Wickrama, 2008). To identify the best baseline model prior to GMM specification, we conducted several single-group LCGAs using four time points on the first dataset. These included intercept only, linear, and quadratic models (see Table 3-4). The intercept only model represented the initial (baseline) levels of the AFEQT questionnaire scores, whereas the linear model included the linear change in the AFEQT questionnaire scores (i.e., a slope). For the quadratic model, an additional latent variable was added to the linear model to estimate a nonlinear pattern.  Table 3-4.   Likelihood Statistics for Models of Change with Four Time Points  -2LL df AIC BIC SABIC LCGA       Intercept -46,156.95 6 92,325.90 92,367.36 92,348.32   Linear -46,065.15 9 92,148.29 92,210.52 92,181.92   Quadratic -45,910.73 13 91,847.46 91,937.35 91,896.04 GMM - quadratic       1-class -45,910.73 13 91,846.46 91,937.35 91,896.04   2-class -45,910.73 27 91,875.47 92,062.16 91,976.36 Note. LCGA = Latent growth curve analysis; GMM = Growth mixture model; LL = Log-likelihood; df = degrees of freedom; AIC = Akaike’s Information Criterion; BIC = Bayesian information criterion; SABIC = Sample size-adjusted BIC.  The AICs, BICs, and SABICs for the LCGA models suggested that the quadratic model was a better fit than the intercept only or linear model. This baseline quadratic model was used to run the GMM by freely estimating all the parameters (i.e., the latent means, variances/covariances, and residuals). However, the 2-class quadratic GMM resulted in poorer 109  model fit than the baseline 1-class model, which indicated that the model was not appropriate for the given data.  Part of the issue with the GMM may have been insufficient sample size, especially in the amount of data available for the latter of the four time points because studies have shown that small samples can lead to convergence issues, improper solutions, and the inability to identify meaningful subgroups (Berlin et al., 2014; Vargha, Bergman, & Takács, 2016). While what constitutes “adequate” sample size is difficult to determine, because it depends on the specification of the model, the distribution of the variables, the amount of missing data, and the strength of the relations among the variables are relevant (Muthén & Muthén, 2002); in general, large sample sizes (> 500) are often deemed most appropriate for complex mixture models (Meyer & Morin, 2016). In our dataset, the second follow-up visit (T2) met this criterion with a sample size of 1,285; there was a large reduction for the third visit (T3) with a sample size of 401 (see Table 3-5).   Table 3-5.   Descriptive Statistics of AFEQT Questionnaire Scores by Follow-Up Visit (T0-T10) Variable N Mean SD Min Max Skewness Kurtosis AFEQT T0 4040 63.70 23.31 0.93 100 -0.39 -0.75 AFEQT T1 4412 71.25 23.15 2.78 100 -0.68 -0.43 AFEQT T2 1285 75.19 22.82 2.78 100 -0.90 -0.10 AFEQT T3 401 77.98 20.88 5.56 100 -1.16 0.70 AFEQT T4 158 73.81 22.47 6.48 100 -0.73 -0.31 AFEQT T5 70 74.88 21.12 19.44 100 -0.56 -0.70 AFEQT T6 30 75.31 21.12 29.63 100 -0.63 -0.70 AFEQT T7 18 80.04 19.59 37.04 100 -1.01 -0.20 AFEQT T8 7 82.54 19.37 47.22 100 -0.73 -1.03  AFEQT T9 <5 84.72 13.30 67.59 98.15 -0.14 -1.81 AFEQT T10 <5 99.07 0.0 99.07 99.07 0.00 0.00 Note. T0 = initial consultation. T1-T10 = corresponding number of follow-up visits. Cell size < 5 not reported as per policy. 110   Although it is not unusual to have declining samples in longitudinal studies (Bell, Kenward, Fairclough, & Horton, 2013), due in part to early discharge from a clinic or attrition, we narrowed our analysis to three time points (T0-T2) to ensure appropriate identification of the trajectories and their meaningful subgroups (Berlin et al., 2014). Although we were limited to the linear GMM (due to having only three time points), many combinations of constraints were still possible (e.g., fixing the intercepts, slopes, residual variances or a combination of all three), which could affect the selection and interpretation of the GMMs. While acknowledging these other possible model specifications, this study was not necessarily a full exploration of GMM but a guided and constrained exploration of the random structure recommended by Gilthorpe et al. (2014) and the underlying data generation processes using three GMM parameterizations: (a) unrestricted random effects, (b) restricted random effects (random intercepts only and no covariances), and (c) restricted random effects plus AR1 (an autoregressive structure).   The unrestricted random effects model is specified to freely estimate all parameters, including the latent means of the intercept and the slope, variances, and covariances. However, such free estimation can lead to convergence issues and thus some constraints are typically applied. The restricted random effects model may aid convergence by constraining the slope variance to zero. Based on Figure 3-24, the restricted random effects model also reflects the underlying structure of the data because there is less variability in the slope than the intercept. The restricted random effects plus AR1 was modelled because specifying constraints can lead to autocorrelation issues. In other words, rather than assuming that each observation of the AFEQT scores was independent of the others, the autoregressive structure recognizes that closely timed 111  repeated measures are more strongly correlated than measures that are timed further apart. Simulation studies have shown that explicitly modelling the autoregressive structure in longitudinal data when the variance-covariance structure is constrained can lead to model improvement and can more accurately capture the underlying data generation process (Gilthorpe et al., 2014). For all models, the measurement occasion-specific variances were constrained to be identical across the time points within each class, which is consistent with the recommended approach and other mixture modelling approaches (e.g., hierarchical linear models). Details of each model specification are provided in Appendix H. Growth Mixture Model Specification As noted previously, all of the models were estimated using Mplus (Muthén & Muthén, 2017) and the expectation-maximization algorithm was used to obtain estimates (with robust standard errors) of all the growth curve parameters and posterior probabilities (i.e., the probability of assigning individuals to groups given the estimates). The expectation-maximization algorithm is an iterative procedure that attempts to converge on the global maximum solution (e.g., parameters associated with the largest loglikelihood). However, the algorithm may have difficulty identifying whether the parameters are based on the largest value for the entire curve (global maximum) or the largest value for only a given area on the curve (local maximum) (Jung & Wickrama, 2008). The convergence to a local maximum, instead of a global maximum, becomes more common in complex models with a greater number of latent classes and indicators (Uebersax, 2009). To address the issue of local solutions, Mplus (by default) incorporates the use of random starting values to ensure that a global maximum value is identified. For example, the STARTS syntax specifies the number of random sets of starting values (default = 10) followed by the number of final optimization (default = 2), which optimizes 112  the two best sets identified by the highest loglikelihood values. During our initial model estimation for two or more latent classes, we increased the number of random starts (100 random starts with 10 final optimizations) to establish a global maxima and to avoid local solutions (Jung & Wickrama, 2008). Before fitting the set of models described above, we had to account for multiple imputed datasets because our estimates and standard errors were averaged over the set of analyses (Rubin, 1987; Schafer, 1997). A common issue that arises during the identification of latent classes across multiple imputed datasets is label switching (Tueller, Drotar, & Lubke, 2011); that is, the latent class identifier is arbitrarily assigned to individuals, resulting in misclassification from one dataset to another. For example, an individual may have Class 1 probability of 60% and Class 2 probability of 40%, and in the second dataset, the same individual may have Class 1 probability of 41% and Class 2 probability of 59%. Here, the class labels between the two datasets are clearly switched. To avoid label switching during the identification of the latent classes, we ran a 2-class GMM on one of the imputed datasets and obtained the starting values (using the SVALUES command in Mplus) for both the overall and the class-specific model. Once we obtained the starting values for each of the parameters, we used these values for the analysis across the imputed datasets (Muthen, 2007). This same procedure was used to run each additional class for the models.   The confidence in the final solution was evaluated based on several statistical fit indices including the AIC, BIC, and SABIC. As mentioned previously, traditional SEM goodness-of-fit statistics or the mixture model statistical comparison tests such as the LMRT and the BLRT were not available because the slope values varied across individuals. Lower AIC, BIC, and SABIC values indicate a better fitting model. Other evaluation criteria included entropy to assess the 113  accuracy of class assignment for each individual, examination of parameter estimates, and visualization of the selected model. While some models had difficulty converging (discussed in the results), we were able to generate reasonably well-fitting solutions with the restricted random effects model up to 3 classes. 3.6.5 How to identify factors that explain variability in individual trajectories. Once the final GMM was determined, we examined the predictors that explained the variability in the subgroups of the identified trajectories (latent classes). There are two general approaches regarding how to include the predictors or covariates and the outcomes of the latent classes in GMM: a one-step (joint model estimation) approach and a three-step (stepwise estimation) approach. The one-step approach uses a joint model that combines the latent class model with a latent class regression model such that the latent classes are conditioned on the covariates (Asparouhov & Muthén, 2014a). While the one-step approach may result in improved accuracy if the correct covariates are included (e.g., smaller standard errors), the most prominent disadvantage is that the inclusion of covariates may affect the type of classes found as well as class membership (Vermunt, 2010). For example, both the latent class model and the latent class regression model needs to be re-estimated each time a covariate is added. This may not only be impractical in most exploratory studies with many covariates but may cause the latent class variable to lose its meaning because it is no longer based on the original indicator variables (Vermunt, 2010).  To address this issue, the three-step approach has been developed to independently evaluate the relationships between the latent classes and the predictor variables, such that the addition of predictor variables into the model does not change class membership itself (Asparouhov & Muthén, 2014a; Vermunt, 2010). The three-step approach involves first 114  estimating the GMM using only latent class indicator variables (e.g., the AFEQT questionnaire scores) without covariates. In the second step, the most likely latent classes are created based on the posterior probabilities obtained in the prior step. In the third step, the latent classes are regressed on the predictor variables with multinomial logistic regression while adjusting for classification uncertainty in the second step (Asparouhov & Muthén, 2014a). To apply the three-step approach, we used the R3STEP method in Mplus to conduct both bivariable and multivariable logistic regression analyses. The advantages of using the R3STEP is that the three-step procedure is implemented automatically rather than having to perform each step manually. The limitation of this approach is that the R3STEP does not allow for hierarchical (i.e., stepwise) regression models because model fit occurs only at the level of the GMM, which means that the fit indices will be the same regardless of the predictor variables entered. This is necessary to hold the class proportions fixed at the values identified when each predictors are entered in the model. In the case with GMM, the predictor variables explain complex relationships between both within-class variation as well as the probability of class membership. Thus, we followed the variable selection approach recommended by Hosmer and Lemeshow (2000) with modifications based on the identified health trajectories. We conducted a series of univariate analyses for each of the predictor variables to the final GMM that compared the “poor and improving health” group as a reference category to both the “good and stable heath” and the “excellent and stable health” groups (derived in section 4.2) to identify predictors of patients who were at higher risk of experiencing poor outcomes (see Table 3-6).  For the age category with 4 levels (less than 60 years, 60-67, 68-75, and 76 or older), we dummy coded each category with age 76 or older as reference because of increasing prevalence of AF with age. For ablation and anticoagulant therapy, each of the follow-up time period was 115  treated as a separate independent binary variable to capture the time-varying nature of ablation and anticoagulation therapy reflected in practice that are not necessarily mutually exclusive. For example, patients who have ablation within 6 months could have another ablation between 1 year to 1.5 years after initial consultation. Similarly, patients who have been prescribed anticoagulant within 6 months could have their prescription renewed again between 1 year to 1.5 years after initial consultation. Variables were retained for the multivariate model if the alpha level in both group comparison was ≤ .25.  Table 3-6.   Unadjusted Odds Ratios for Predictors of Latent Class Membership  Poor and improving healtha vs. good and stable health p value Poor and improving healtha vs. excellent and stable health  p value  Odds ratio (95% CI)  Odds ratio (95% CI)  Age group (years)     Less than 60 0.98 (0.81–1.18) .82 0.97 (0.75–1.23) .77 60 – 67  0.96 (0.79–1.16) .64 1.05 (0.82–1.34) .73 68 – 75 0.93 (0.77–1.13) .46 1.16 (0.92–1.48) .25 Age 76 or oldera - - - -      Women 0.53 (0.44–0.63) .00 0.40 (0.31–0.52) .00      Distance ≥100km 0.54 (0.41–0.70) .00 0.79 (0.59–1.05) .07      CHADS2 0.86 (0.79–0.92) .00 0.82 (0.74–0.91) .00      Ablation therapy      0 to 6 months (Y/N)b 0.23 (0.15–0.34) .00 0.72 (0.54–0.96) .01 6 months to 1 year (Y/N)b 0.25 (0.14–0.44) .00 0.54 (0.34–0.85) .00   1 year to 1.5 years (Y/N)b 0.37 (0.20–0.69) .00 0.42 (0.20–0.90) .00   1.5 to 2 years (Y/N)b 0.41 (0.21–0.80) .00 0.57 (0.28–1.17) .04   More than 2 years (Y/N)b 0.21 (0.11–0.40) .00 0.29 (0.15–0.56) .00 Anticoagulation therapy       0 to 6 months (Y/N)b 0.51 (0.43–0.61) .00 0.48 (0.38–0.60) .00 6 months to 1 year (Y/N)b  0.60 (0.51–0.71) .00 0.46 (0.37–0.57) .00 1 year to 1.5 years (Y/N)b 0.64 (0.54–0.76) .00 0.46 (0.36–0.57) .00 1.5 to 2 years (Y/N)b 0.70 (0.59–0.82) .00 0.45 (0.35–0.57) .00 More than 2 years (Y/N)b 0.65 (0.55–0.77) .00 0.50 (0.39–0.62) .00 Note. a The reference category; CI = confidence interval. b Each variable treated as an independent binary variable.  116  The univariate analysis reported in Table 3-6 identified gender, distance from clinic, CHADS2, time of ablation therapy and time of anticoagulation therapy as potential predictors of class membership. During the iterative multivariate fitting, ablation therapy during specified intervals (6 months to 1 year, 1 year to 1.5 years, 1.5 to 2 years, and more than 2 years) were eliminated one at a time because they did not reach the 0.1 alpha level, and when taken out, had minimal change to the remaining parameter estimates (less than 20%). The maximum p value of the remaining variables was less than 0.1 for at least one of the two group comparisons, at which point the age category variable originally set aside was reconsidered. The age group made it back to the model because it was significant at the 0.1 alpha level. Thus, the final multivariate model with age, gender, distance, CHADS2, ablation at follow-up (0 to 6 months, 6 month to 1 year, 1 year to 1.5 years, 1.5 to 2 years, more than 2 years) and anticoagulation at follow-up (0 to 6 months) were reported in the results.  To evaluate the effects of the predictor variables for the multivariate model, we reported the odds ratios and the confidence intervals, and refrained from reporting the p-values. In a recent article, Wassertein, Schirm, and Lazar (2019) made a strong argument for caution in basing conclusions on an arbitrary threshold of p < .05, not using the phrase “statistically significant”, and being thoughtful about the use and interpretation of p-values if they are used at all (e.g., reporting p-values as a continuous, descriptive statistic). Considering a movement towards discouraging the reporting of p-values and some journals banning the use of p-values altogether (e.g., Basic and Applied Social Psychology), we did not report p-values and instead provide confidence intervals to assess the effects of the predictors on the latent classes. While confidence intervals could be used as an alternative for statistical significance testing (e.g., if the 117  value does not cross 1), they provide more context by providing the effect size and accounting for the uncertainty in estimation (Szumilas, 2010).  To better interpret the odds ratios, we estimated the distributions of each of the predictors within each of the classes. We followed the recommendations of Asparouhov and Muthén (2014b) and used the BCH procedure, which has been shown to be more robust with varying sample sizes and entropy levels compared with other methods; however, since the BCH procedure was not available for multiple imputed datasets (i.e., TYPE = IMPUTATION in Mplus), the means and the standard errors were averaged across the 20 imputed datasets using Excel.  Because Mplus does not provide a way to evaluate the fit of GMM with predictors, we conducted a separate multinomial logistic regression analysis, based on the predicted class membership, in SPSS (IBM Corp, 2017). The limitation of this approach is that the classes are treated as being discrete (i.e., analysis does not take class entropy into account) and should be cautiously interpreted in light of this limitation. The data were checked for linearity of the logit and multicollinearity; both assumptions were met with a deviation of linearity test of the continuous independent variable (CHADS2) (F(6,7432) = 1.13 p = .34) and a low variance inflation factor (ranged from 1.03–1.64). The overall model was evaluated using Nagelkerke’s R2 statistic and predictive values. Among the pseudo R2 values, Nagelkerke’s R2 statistic was reported for ease of interpretation with the explained variation ranging from 0 to 1, and values closer to zero indicating that the model had no predictive power. We also evaluated predictive values such as the sensitivity, specificity, positive predictive value, and negative predictive value. For multi-class classification with class 1 as an example, sensitivity tested the ability of the model to correctly identify all class 1 instances that were classified as class 1, whereas 118  specificity tested the ability of the model to correctly identify all non-class 1 instances that were not classified as class 1. Positive predictive value is the probability that those with the predicted class 1 classification are truly classified as class 1. Negative predictive value is the probability that those with predicted class 1 classification are truly not classified as class 1 (for further explanation see Beleites, Salzer, & Sergo, 2013).  3.7 Summary We undertook extensive data preparation steps to arrive at the eligible study cohort of 13,113 outpatients with AF. A large part of the study was focused on how to represent time, and then we established a defensible process to represent the individual variation in the patients’ trajectories. The following summarizes the key challenges in analyzing the PROMs data stored within the registry (see Figure 3-26).  1. What information can be extracted from linked data sources? • Substantial data preparation and cleaning were needed, requiring complex computer programming and knowledge of various analytical software packages. • The eligible study cohort was created based on the eligibility criteria of the study and information from the clinicians about the clinic processes.   • The StudyID (project-specific anonymized study identification number from the PHN) was used to link the different data sources (MSP, DAD, and PharmaNet).  • Relevant information including patients’ comorbidities and interventions received were extracted with validated algorithms known to have moderate to high predictive validity and sensitivity.  • Various graphical techniques were employed to reveal relevant information, such as stacked bar plots and time series plots, and to construct time intervals for the time-varying covariates. 2. How to Best Accommodate Missing Data • Accommodating missing data required understanding the missing data mechanisms: missing completely at random, missing at random, or missing not at random.  • The missing data mechanism was assumed to be missing at random (most common in PROM-based studies) because the patients who were in good health were expected to have less missing information than those who are not as well.  • To accommodate the missing at random assumptions, two recommended techniques (multiple imputation and full information maximum likelihood) were conducted.  • Multilevel variant of multiple imputation was used to correct for the multilevel structure by specifying level 1 and level 2 variables.  • To increase the accuracy of the imputed values in the imputation model, all available auxiliary variables were added, including an indicator of death occurring during the observation period.  119  • Full information maximum likelihood was used after multiple imputation to account for the follow-up times that were unequally spaced.  3. How to Best Represent the Variability in Frequency and Timing of the Measurement Occurrences? • Several options for modelling time were discussed based on the exact date of when the AFEQT questionnaires was administered to each patient.  • The smooth fitting curves in Figure 3-22 and 3-23 showed that multiple processes were involved requiring the metric of time to account for both the frequency and timing of the measurement occurrences (individually-varying times of observations).  • Both the LGM and MLM approaches were compared with acknowledgement that MLM was easier to implement using time as a univariate factor (“long” format). • While MLM was considered, there was an issue with the underlying variable-centred approach that provides only a single set of parameters, which would not have been representative of all the patients with AF. Thus, a need for a more flexible modelling method was highlighted.  4. How to Best Represent the Shape of the Trajectories? • The line graph in Figure 3-24 showed that most of the patients’ trajectories were linear. However, there was wide variability in the intercepts compared with the slopes, which suggested at least two different trajectory groups.  • To identity multiple subgroups of trajectories, an emerging statistical approach, known as GMM, was introduced. GMM is a highly flexible method that incorporated latent classes to assign individuals who shared similar patterns of scores into unobserved subgroups and allowed for the nesting of time to account for the individually-varying times of observation. • The model was primarily assessed using AIC, BIC, and SABIC statistics because the traditional SEM goodness-of-fit statistics or mixture model statistical comparison tests such as LMR or BLRT were not available when the slope loadings varied across individuals.  5. How to Best Represent Individual Differences in Trajectories? • Several single-group LCGA models with four time points were estimated to identify the best baseline model prior to the GMM specification, which suggested that the quadratic model was a better fit. However, the two-class quadratic GMM resulted in convergence issues due to insufficient sample size, and subsequent modelling was limited to three time points.  • While acknowledging other GMM specifications, a structured approach was followed based on the recommendations of Gilthorpe et al. (2014) and the underlying data generation processes using three parameterizations: (a) an unrestricted random effects model, (b) a restricted random effects model, and (c) a restricted random effects plus AR1 model. 6. How to identify factors that explain variability in individual trajectories? • Two general approaches regarding how to include the predictors or covariates and outcomes of the latent classes in GMM were explored: a one-step (joint model estimation) approach and a three-step (stepwise estimation) approach. • The three-step approach was selected because it addresses the limitations of the one-step approach such that the addition of predictor variables does not change the class membership itself. • Based on variable selection approach recommended by Hosmer and Lemeshow (2000) with modifications based on the GMM, age, gender, distance from clinic, CHADS2, time at ablation therapy and time of anticoagulation therapy were explored as predictors of class membership.  Figure 3-26. A Summary of the Key Challenges in Analyzing PROMs Data  Stored in a Clinical Registry  120  Overall, our analytical objective was to examine individual trajectories of change and to determine which identified patterns were associated with the theoretically-derived predictor variables. The PROMs data in the AF registry showed that the longitudinal analysis method had to account for both the different subgroups of patients with AF and variability in the frequency and timing of the measurement occurrences. To address these issues, an alternative modelling approach known as GMM was introduced that closely aligned with the patient-centred approach. While acknowledging other possible model specifications, we followed a relatively guided and constrained exploration of GMM based on the recommendations of Gilthorpe et al. (2014) and the underlying data generation processes using three GMM parameterizations: (a) an unrestricted random effects model that freely estimated all the parameters but may have led to convergence issues, (b) a restricted random effects model that aided convergence by constraining the slope variance to zero, and reflected the underlying data structure, and (c) a restricted random effects plus AR1 model that addressed autocorrelation issues, which recognizes that closely timed repeated measures are more strongly correlated than measures that are timed further apart. In addition, several time invariant and time-varying covariates were examined to further understand the trajectories of the patients’ PROMs scores.  121  Chapter 4: Results In this chapter, we provide the results in three key sections. In the first section, descriptive statistics of the study sample at their initial consultation and at their follow-up visits are reported. Then, details of the number of AFEQT questionnaires completed by the patients are provided. In the second section, the fit of the three GMM parameterizations (unrestricted random effects, restricted random effects, and restricted random effects plus AR1) are evaluated to support the final model selection, and then the corresponding model findings are examined and plotted. In the third section, predictors of latent class membership (subgroups of trajectories) based on the odds of class membership (with odds ratios reported) are examined.   4.1 Study Sample The analysis cohort (N = 7,439) consisted of 2,897 women (38.9%) and 4,542 men (61.1%) mostly in the 60 and older age category (72.9%). The majority of the patients lived less than 100 kilometres from their AF clinic (83.5%). For CHADS2 score, most of the patients were at moderate stroke risk with CHADS2 score greater than 1 (65.3%). Of the comorbidities, hypertension was the most frequent (n = 2,620; 35.2%) followed by heart failure (n = 1,164; 15.6%). Of the prior interventions performed, the most frequent was cardioversion (n = 2,001; 15.5%) followed by ablation (n = 453; 6.1%). Most of the patients were on anticoagulant therapy (n = 4,614; 62.0%).  The patients’ characteristics at the time of the initial consultation and at the follow-up visits were compared to examine whether they were similar. For example, in the AF clinic, many patients had an initial consultation without any follow-up visit or they may have had initial consultation but did not complete the AFEQT questionnaire. As Table 4-1 shows, there were some differences in the patients’ characteristics at the time of the initial consultation and at 122  follow-up. The patients who were seen at a follow-up appointment were less likely to be in the less than 60 age group (OR = 0.75, 95% CI: [0.68–0.83]) and more likely to be 76 or older (OR = 1.19, 95% CI: [1.07–1.33]), live further from their respective clinics (OR = 2.70, 95% CI: [2.38–3.07]), were less likely to be enrolled at clinic sites 1 (OR = 0.17, 95% CI: [0.15–0.21]), 2 (OR = 0.24, 95% CI: [0.21–0.27]), or 3 (OR = 0.33, 95% CI: [0.30–0.37]), and were more likely to be enrolled at clinic sites 4 (OR = 4.01, 95% CI: [3.45–4.68]), or 5 (OR = 20.38, 95% CI: [17.35–24.06]), were less likely to report symptoms of AF (CCS-SAF ≥ 1; OR = 0.71, 95% CI: [0.63–0.80]), were twice as likely to have had ablation therapy (OR = 2.36, 95% CI: [1.94–2.88]), and almost three times more likely to have had a defibrillator implantation (OR = 2.84, 95% CI: [1.39–6.27]) compared with the larger group of patients observed at an initial consultation seen only once.  Table 4-1.   Characteristics of Patients (N = 7,439)    Characteristic All Patients at initial consultation  (limited to patients seen only once)  Patients at first follow-up (after having had an initial consultation) Patients at initial consultation vs. patients at first follow-upa N = 7,439 (100%) n = 4,040 (54.3%) n =  3,399 (45.7%) Odds ratio (95% CI) Age group (years)     Less than 60 2,021 (27.2) 1,202 (29.8) 819 (24.1) 0.75 (0.68–0.83) 60 – 67  1,829 (24.6) 967 (23.9) 862 (25.4) 1.08 (0.97–1.20) 68 – 75 1,842 (24.8) 981 (34.3) 861 (25.3) 1.06 (0.95–1.18) 76 or older 1,747 (23.5) 890 (22.0) 857 (25.2) 1.19 (1.07–1.33) Women  2,897 (38.9) 1,552 (38.4) 1,345 (39.6) 1.05 (0.96–1.15) Distance to clinic       ≥ 100 km 1,231 (16.5) 420 (10.4) 811 (23.9) 2.70 (2.38–3.07) Clinic site     1 1,160 (15.6) 980 (24.3) 180 (5.3) 0.17 (0.15–0.21) 2 1,360 (18.3) 1,086 (26.9) 274 (8.1) 0.24 (0.21–0.27) 3 2,128 (28.6) 1,547 (38.3) 581 (17.1) 0.33 (0.30–0.37) 4 948 (12.7) 246 (6.1) 702 (20.7) 4.01 (3.45–4.68) 5 1,843 (24.8) 181 (4.5) 1,662 (48.9) 20.38 (17.35–24.06) CHADS2     123  Table 4-1.   Characteristics of Patients (N = 7,439)    Characteristic All Patients at initial consultation  (limited to patients seen only once)  Patients at first follow-up (after having had an initial consultation) Patients at initial consultation vs. patients at first follow-upa N = 7,439 (100%) n = 4,040 (54.3%) n =  3,399 (45.7%) Odds ratio (95% CI) 0 2,583 (34.7) 1,439 (35.6) 1,144 (33.7) 0.92 (0.83–1.01) 1 2,349 (31.6) 1,261 (31.2) 1,088 (32.0) 1.04 (0.94–1.14) ≥2 2,507 (33.7) 1,340 (33.2) 1,167 (34.3) 1.05 (0.96–1.16) CCS-SAF ≥ 1 6,152 (82.7) 3,431 (84.9) 2,721 (80.1) 0.71 (0.63–0.80) Charlson score      0 1,308 (17.6) 700 (17.3) 608 (17.9) 1.04 (0.92–1.17) 1 1,517 (20.4) 825 (20.4) 692 (20.4) 1.00 (0.89–1.12) ≥2 4,614 (62.0) 2,515 (62.3) 2,099 (61.8) 0.98 (0.89–1.08) Cardiac condition       Hypertension 2,620 (35.2) 1,428 (35.3) 1,192 (35.1) 0.99 (0.90–1.09)   Heart failure 1,164 (15.6) 689 (17.1) 475 (14.0) 0.79 (0.70–0.90)   Stroke/TIA  555 (7.5) 312 (7.7) 243 (7.1) 0.92 (0.77–1.09)   Myocardial infarction 169 (2.3) 87 (2.2) 82 (2.4) 1.12 (0.83–1.53) Peripheral vascular disease 26 (0.3) 19 (0.5) 7 (0.2) 0.44 (0.17–1.02) Prior interventions     Cardioversion 1,151 (15.5) 601 (14.9) 550 (16.2) 1.10 (0.97–1.25) Ablation 453 (6.1) 157 (3.9) 296 (8.7) 2.36 (1.94–2.88) Pacemaker 232 (3.1) 123 (3.0) 109 (3.2) 1.06 (0.81–1.37) PCI 215 (2.9) 111 (2.7) 104 (3.1) 1.12 (0.85–1.47) Coronary artery bypass graft 153 (2.1) 98 (2.4) 55 (1.6) 0.66 (0.47–0.92) Valve surgery 128 (1.7) 73 (1.8) 55 (1.6) 0.89 (0.63–1.27) Dialysis 31 (0.4) 21 (0.5) 10 (0.3) 0.57 (0.25–1.19) Defibrillator implantation 34 (0.5) 10 (0.2) 24 (0.7) 2.84 (1.39–6.27) Comorbidities       Diabetes 998 (13.4) 575 (14.2) 423 (12.4) 0.86 (0.75–0.98) COPD 420 (5.6) 214 (5.3) 206 (6.1) 1.15 (0.95–1.40)   Depression 330 (4.4) 184 (4.6) 146 (4.3) 0.94 (0.75–1.17)   Chronic kidney disease 266 (3.6) 164 (4.1) 102 (3.0) 0.73 (0.57–0.94)   Sleep disorder 210 (2.8) 106 (2.6) 104 (3.1) 1.17 (0.89–1.54) Hypothyroidism 164 (2.2) 91 (2.3) 73 (2.1) 0.95 (0.70–1.30)   Gastrointestinal bleed 150 (2.0) 83 (2.1) 67 (2.0) 0.96 (0.69–1.33)   Peptic ulcer 15 (0.2) 7 (0.2) 8 (0.2) 1.36 (0.48–3.94) Prior medications     Anticoagulants 4,614 (62.0) 2,456 (60.8) 2,158 (63.5) 1.12 (1.02–1.23)   Beta-blockers 4,362 (58.6) 2,391 (59.2) 1,971 (58.0) 0.95 (0.87–1.04)   Antiarrhythmics 2,689 (36.1) 1,450 (35.9) 1,239 (36.5) 1.02 (0.93–1.13)   Calcium channel blockers 1,735 (23.3) 882 (21.8) 853 (25.1) 1.20 (1.08–1.34)   Digoxin 876 (11.8) 469 (11.6) 407 (12.0) 1.034 (0.90–1.20)   Antiplatelets 353 (4.7) 210 (5.2) 143 (4.2) 0.80 (0.64–0.99) Note. Results based on an imputed dataset 1. CCS-SAF = Canadian Cardiovascular Society Severity in Atrial Fibrillation Scale; TIA = Transient Ischemic Attack; PCI = Percutaneous Coronary Intervention; COPD = Chronic Obstructive Pulmonary Disease; a Reference category. 124  Most patients did not complete an AFEQT questionnaire at each clinic visit, with 56.0% providing data at six months after their initial consultation, and 22.6% at one year, 14.1% at 18 months, 10.3% at 2 years, and 21.8% at more than 2 years (see Appendix G). Further description of the number of AFEQT questionnaires completed by the patients is provided in Table 4-2. Patients who completed two or more AFEQT questionnaires had higher rates of having received ablation or anticoagulation therapy compared with those who completed only one AFEQT questionnaire: patients who completed two or three AFEQT questionnaires had higher rates of ablation therapy reported by the time of a six-month clinic visit compared with those who completed only one questionnaire (OR = 1.21, 95% CI: [1.06–1.39]).     125  Table 4-2.   Number of AFEQT Questionnaires Completed by Patients’ Characteristics     Number of AFEQT Questionnaires Completed Two and three vs. one questionnairea  Odds ratio (95% CI) Characteristics One (n = 5,412) Two (n = 1,756) Three (n = 271) Age group (years)     Less than 60 1,503 (27.) 439 (25.0) 79 (29.2) 0.89 (0.79–1.00) 60 – 67  1,292 (23.9) 459 (26.1) 78 (28.8) 1.15 (1.02–1.29) 68 – 75 1,346 (24.9) 435 (24.8) 61 (22.5) 0.98 (0.87–1.10) 76 or older 1,271 (23.5) 423 (24.1) 53 (19.6) 1.00 (0.89–1.13) Women 2,089 (38.6) 708 (40.3) 100 (36.9) 1.05 (0.95–1.17) Distance to clinic        ≥ 100 km 928 (17.1) 286 (16.3) 17 (6.3) 0.85 (0.74–0.98) Clinic site     1 1,053 (19.5) 98 (5.6) 9 (3.3) 0.23 (0.19–0.28) 2 1,227 (22.7) 120 (6.8) 13 (4.8) 0.24 (0.20–0.29) 3 1,360 (25.1) 609 (34.7) 159 (58.7) 1.82 (1.63–2.03) 4 385 (7.1) 479 (27.3) 84 (31.0) 5.02 (4.36–5.79) 5 1,387 (25.6) 450 (25.6) 6 (2.2) 0.84 (0.75–0.95) CHADS2 ≥ 1 3,560 (65.8) 1,125 (64.1) 171 (63.1) 0.92 (0.83–1.03) 0 1,852 (34.2) 631 (35.9) 100 (36.9) 1.08 (0.97–1.21) 1 1,696 (31.3) 567 (32.3) 86 (31.7) 1.04 (0.93–1.16) ≥2 1,864 (34.4) 558 (31.8) 85 (31.4) 0.88 (0.79–0.99) CCS-SAF ≥ 1 4,543 (83.9) 1,385 (78.9) 224 (82.7) 0.74 (0.65–0.84) Charlson score     0 968 (17.9) 296 (16.9) 44 (16.2) 0.93 (0.81–1.06) 1 1,060 (19.6) 403 (22.9) 54 (19.9) 1.20 (1.05–1.35) ≥2 3,384 (62.5) 1,057 (60.2) 173 (63.8) 0.92 (0.83–1.03) Cardiac condition       Hypertension 1,895 (35.0) 635 (36.2) 90 (33.2) 1.03 (0.93–1.15)   Heart failure 847 (15.7) 264 (15.0) 53 (19.6) 1.00 (0.87–1.15)   Stroke/TIA 402 (7.4) 141 (8.0) 12 (4.4) 1.02 (0.84–1.23)   Myocardial infarction 131 (2.4) 37 (2.1) <5 0.77 (0.53–1.10) Peripheral vascular disease 23 (0.4) <5 0  0.36 (0.08–1.05) Prior interventions     Cardioversion 836 (15.4) 265 (15.1) 50 (18.5) 1.01 (0.87–1.16) Pacemaker 184 (3.4) 44 (2.5) <5 0.69 (0.49–0.95) PCI 165 (3.0) 45 (2.6) 5 (1.8) 0.81 (0.58–1.10) Coronary artery bypass graft 127 (2.3) 23 (1.3) <5 0.54 (0.35–0.82) Valve surgery 103 (1.9) 22 (1.3) <5 0.65 (0.41–0.99) Dialysis 28 (0.5) <5 0  0.30 (0.07–0.85) Defibrillator implantation 28 (0.5) 6 (0.3) 0  0.58 (0.22–1.32) Comorbidities       Diabetes 751 (13.9) 206 (11.7) 41 (15.1) 0.86 (0.74–1.00)   COPD 289 (5.3) 114 (6.5) 17 (6.3) 1.23 (0.99–1.51)   Depression 235 (4.3) 82 (4.7) 13 (4.8) 1.08 (0.85–1.38)   Chronic kidney disease 219 (4.0) 42 (2.4) 5 (1.8) 0.56 (0.41–0.77)   Sleep disorder 152 (2.8) 50 (2.8) 8 (3.0) 1.02 (0.75–1.38) 126  Table 4-2.   Number of AFEQT Questionnaires Completed by Patients’ Characteristics     Number of AFEQT Questionnaires Completed Two and three vs. one questionnairea  Odds ratio (95% CI) Characteristics One (n = 5,412) Two (n = 1,756) Three (n = 271)   Hypothyroidism 116 (2.1) 42 (2.4) 6 (2.2) 1.11 (0.78–1.55)   Gastrointestinal bleed 113 (2.1) 33 (1.9) <5 0.87 (0.59–1.26)   Peptic ulcer 13 (0.2) <5 <5 0.44 (0.06–1.60) Prior medications       Beta-blockers 3,122 (57.7) 1,073 (61.1) 167 (61.6) 1.16 (1.04–1.28)   Antiarrhythmics 1,996 (36.9) 583 (33.2) 110 (40.6) 0.89 (0.80–0.99)   Calcium channel blockers 1,302 (24.1) 366 (20.8) 67 (24.7) 0.86 (0.76–0.97)   Digoxin 631 (11.7) 211 (12.0) 34 (12.5) 1.04 (0.89–1.22)   Antiplatelets 265 (4.9) 71 (4.0) 17 (6.3) 0.88 (0.69–1.12) Ablation       0 to <6 months 824 (15.2) 291 (16.6) 72 (26.6) 1.21 (1.06–1.39)   6 months to <1 year 390 (7.2) 143 (8.1) 47 (17.3) 1.33 (1.11–1.60) 1 year to <1.5 years 193 (3.6) 82 (4.7) 25 (9.2) 1.51 (1.18–1.92) 1.5 to <2 years 143 (2.6) 80 (4.6) 18 (6.6) 1.87 (1.44–2.43) 2+ years 325 (6.0) 184 (10.5) 35 (12.9) 1.90 (1.58–2.27) Anticoagulation        0 to <6 months 3,769 (69.6) 1,297 (73.9) 216 (79.7) 1.28 (1.14–1.44)   6 months to <1 year 3,238 (66.0) 1,173 (66.8) 211 (77.9) 1.44 (1.30–1.61) 1 year to <1.5 years 2,834 (52.4) 1,063 (60.5) 195 (72.0) 1.49 (1.34–1.65) 1.5 to <2 years 2,471 (45.7) 974 (55.5) 177 (65.3) 1.56 (1.41–1.73) 2+ years 2,445 (45.2) 1,024 (58.3) 184 (67.9) 1.79 (1.61–1.99) Note. Data shown as N (% within respondent category) unless stated otherwise. Cell size < 5 not reported as per policy. Results based on imputed dataset 1. CCS-SAF = Canadian Cardiovascular Society Severity in Atrial Fibrillation Scale; TIA = Transient Ischemic Attack; PCI = Percutaneous Coronary Intervention; COPD = Chronic Obstructive Pulmonary Disease; a Reference category.   In summary, the AFEQT questionnaires were completed at varying frequencies and intervals (e.g., some patients may have completed only one AFEQT questionnaire while others may have completed three). There was considerable variability in the frequency and timing of the measurement occurrences with completion being associated to some extent with age, comorbidities, and interventions provided including ablation and anticoagulation therapy.  127  4.2 Model Selection The results of the three GMM models are provided with the mean fit indices and entropy values, which were averaged across the 20 imputed datasets (see Table 4-3).   Table 4-3.   Mean Likelihood and Information Criteria for the Growth Mixture Models  Mean -2LL (SD)  df Mean  AIC Mean  BIC Mean SABIC Mean Entropy Unrestricted model        1-class -44,280.12 (13.11) 6 88,572.23 (26.22) 88,613.72 (26.22) 88,594.65 (26.22) -   2-class -43,469.74 (9.53) 13 86,965.49 (19.06) 87,055.37 (19.06) 87,014.06 (19.06) .63   3-class - - - - - - Restricted standard model        1-class -44,307.32 (12.65) 4 88,622.63 (25.29) 88,650.29 (25.29) 88,637.58 (25.29) -   2-class -43,492.79 (11.12) 9 87,003.57 (22.24) 87,065.80 (20) 87,037.20 (22.24) .63   3-class -43,229.48 (13.71) 14 86,486.96 (27.41) 86,583.76 (27.41) 86,539.27 (27.41) .66   4-class - - - - - - Restricted AR1 model        1-class -44,301.81 (12.57) 5 88,613.63 (25.14) 88,648.20 (25.14) 88,632.31 (25.14) -   2-class - - - - - - Note. LL = log-likelihood; df = degrees of freedom; AIC = Akaike’s information criterion; BIC = Bayesian information criterion; SABIC = Sample size-adjusted BIC.  The rows left blank indicate that the models did not converge. The TSCORES in Mplus do not provide Lo-Mendell-Rubin Tests or the Bootstrap Likelihood Ratio Test.   Each of the models specified with one latent class successfully converged without any error messages. However, we encountered several issues when estimating two or more latent classes. For the restricted AR1 two-class model, while the best loglikelihood value has been replicated to obtain the starting vales, there was an error message below this result that one or more parameters were fixed to avoid singularity of the information matrix; the singularity is most likely because the model is not identified, or because of empty cells in the joint distribution of 128  the categorical variables in the model. Because there was no variability in the correlations, the additional parameter for the autoregressive structure was fixed to 0. Thus, we did not further estimate this model because autocorrelation was not an issue in our data structure.  For the unrestricted model, the three-class model specification, on one of the imputed datasets, was estimated; however, when the starting values were applied for all of the 20 imputed datasets, the model did not converge on any and generated error messages indicating a non-positive definite covariance matrix. Despite considerable effort (increasing the number of random starts to 1,000, increasing the number of iterations, and decreasing the STSCALE (the random start scale, which controls the dispersion of the random perturbations) to achieve convergence as suggested in the Mplus discussion board), we were not able to find a proper fitting solution. To identify whether the source of the issue was identification or problems with local maxima, a four-class model was estimated, which resulted in the same error message suggesting the problem was likely due to the model not being identified. Thus, we were able to generate a proper fitting solution with up to two latent classes. For the restricted standard model, the four-class model generated the same error message indicating a non-positive definite covariance matrix despite similar efforts (increasing the number of random starts to 1,000, increasing the number of iterations and decreasing the STSCALE value) to find a proper fitting solution. This model converged without any error messages with up to three latent classes.  Based on the above identified models, increasing the number of latent classes resulted in better (i.e., lower) mean AICs, BICs, and SABICs. For example, there was improvement in fit in the two-class model compared with the one-class model for both the unrestricted and the restricted standard models. However, lower fit indices supported the three-class restricted 129  standard model. In addition, the three-class model resulted in a slightly better classification of patients (entropy of .66) compared with the two-class model (entropy of .63). These results indicate that the observed data were best represented by three latent classes or subgroups of trajectories by the restricted standard model. As recommended by Ram and Grimm (2009), we examined the mean parameter estimates of the three-class restricted model to assess for out-of-bound parameters (e.g., negative variances) or problems in estimation (see Table 4-4).   Of the three classes, Class 2 included the most patients (n = 4,7301.0, 63.6%; implied sample size based on probabilistic membership), followed by Class 3 (n = 2,076.6, 27.9%), and Class 1 (n = 631.4, 8.5%). Distinct mean intercepts were observed across the classes with Class 1 having the highest average initial AFEQT questionnaire score (97.08), followed by Class 3 Table 4-4.   Mean Parameter Estimates of the Three-Class Restricted Standard Model  Class 1 (n = 631.4; 8.5%) Est. (se) Class 2 (n = 4,7301.0; 63.6%) Est. (se) Class 3 (n = 2,076.6; 27.9%) Est. (se) Restricted standard model   Mean    Intercept  97.08 (1.48) 55.77 (1.04) 83.65 (2.21) Slope*  0.37 (0.29) 2.59 (0.31) 2.17 (0.41) Variance    Intercept  0.87 (0.79) 49.07 (14.19) 0.00 (0.01) Slope* 0 0 0 Residual variance of AFEQT scores at:    T0, T1, T2 5.89 (4.61) 428.91 (16.91) 81.72 (15.75) Note. Sample size is not in whole numbers because they reflect probabilistic membership. Restricted standard model fixes linear slope to zero and constrains residual variances to be equal within class.  T0 = Initial consultation, T1 = First follow-up, T2 = Second follow-up. *Since individually-varying time periods were used, the slope parameter here refers to individual slope loadings. 130  (83.65), and Class 2 (55.77). For the mean slope, Class 2 had a steeper increasing slope (2.59) compared with Class 3 (2.17) and Class 1 (0.37). With respect to the intercept variances, Class 2 had more variability (49.07) compared with Class 1 (0.87) and Class 3 (0.00). The variances of the slopes were zero for all three classes because of how the restricted standard model was specified. Reviewing the residual variances, Class 2 showed the greatest variability (428.91), in part because it had the larger sample size, followed by Class 3 (81.72) and Class 1 (5.89).  Finally, the latent class trajectories were plotted to visualize the results of the models (see Figure 4-1).    Figure 4-1. Three-Class Trajectory Model (AFEQT Questionnaire Scores).  Note. The mean AFEQT questionnaire scores, at each follow-up, for each latent group, are indicated in the table below the figure. The duration of the interval between observation occurrences, in mean years, is presented below the line plot.  The most common trajectory (Class 2 with 63.6% of the sample) was for patients who had AFEQT scores that averaged 52.9 points at the initial consultation and gradually improved at each follow-up (according to Freeman et al. (2015), scores of ≤ 65.7 are considered to be 131  indicative of poor health status). This class was labelled “poor and improving health.” The second most common trajectory (Class 3 with 27.9% of the sample) was for patients that started with higher baseline AFEQT questionnaire scores of about 82.3 points (Freeman et al. (2015) considered scores of 81.9 to 93.1 points to be indicative of good health status) but showed less improvement over time. This class was labelled “good and stable health.” The least common trajectory (Class 1 with 8.5%) showed that some patients’ initial baseline AFEQT questionnaire scores were very high at about 96.8 points (Freeman et al. (2015) observed that ≥ 93.5 points is equivalent to having excellent health status) with little change through the rest of the follow-up period. This class was labelled “excellent and stable health.” The mean time (years) at each follow-up varied for the three trajectory classes ranging from 0.77 to 0.92 years for the first follow-up and from 1.19 to 1.67 years for the second follow-up.  4.3 Predictors of Latent Class Membership The relative frequencies of the predictors of latent class membership, by class, with odds ratios and 95% confidence intervals are shown in Table 4-5.   132  Table 4-5.   Relative Frequencies by Class with Adjusted Odds Ratios for Predictors of Latent Class Membership (N=7,439)  Poor and  improving health Good and stable health Excellent  and stable health Poor and improving healtha  vs. good and stable health Poor and improving healtha vs. excellent and stable health   Relative frequencies Relative frequencies Relative frequencies Odds ratio (95% CI) Odds ratio (95% CI) Individual/Environmental Factors Age category       Less than 60 27.3% 26.9% 26.6% 0.59 (0.44–0.80) 0.61 (0.40–0.91) 60 - 67 24.7% 23.9% 25.6% 0.71 (0.53–0.94) 0.83 (0.57–1.23) 68 - 75 24.9% 23.5% 27.8% 0.73 (0.56–0.95) 1.03 (0.72–1.47) 76 or oldera 23.0% 25.6% 19.9% - -       Women (Men)  44.8% 30.0% 24.5% 0.48 (0.40–0.58) 0.36 (0.27–0.47)       Distance ≥100km (<100km)    19.0% 11.1% 15.6% 0.63 (0.48–0.83) 0.79 (0.59–1.08)  Biological Function/Treatment CHADS2, means (SD) 1.25 (.02) 1.04 (.04) 1.00 (.06) 0.78 (0.71–0.86) 0.82 (0.72–0.94)       Ablation therapy         0 to 6 months (Y/N)b 20.6% 5.4% 15.8% 0.28 (0.19–0.41) 0.83 (0.62–1.12) 6 months to 1 year (Y/N)b 10.3% 2.7% 5.8% 0.27 (0.15–0.46) 0.54 (0.34–0.88) 1 year to 1.5 years (Y/N)b 5.2% 2.0% 2.3% 0.46 (0.25–0.86) 0.53 (0.27–1.07) 1.5 to 2 years (Y/N)b 4.0% 1.7% 2.4% 0.59 (0.30–1.15) 0.75 (0.36–1.56)   More than 2 years (Y/N)b 10.1% 2.3% 3.2% 0.25 (0.14–0.44) 0.29 (0.15–0.56) Anticoagulation therapy        0 to 6 months (Y/N)b 76.3% 62.1% 60.6% 0.70 (0.56–0.86) 0.53 (0.41–0.69) Note. a The reference category; SE = standard errors; CI = confidence interval. b Each variable treated as an independent binary variable.   There were differences in the ages of the patients within the latent classes or trajectory groups. For example, for the less than 60 years of age category (with reference to age 76 or older), both the “good and stable health” (26.9%) and “excellent and stable health” (26.6%) groups had lower percentages compared to the “poor and improving health” group (27.3%) (OR = 0.59, 95% CI: [0.44–0.80] and OR = 0.61, 95% CI: [0.40–0.91], respectively). For the age group between 60 to 67 years of age (with reference to age 76 or older), the “good and stable health” group had a lower percentage (23.9%) compared with “poor and improving health” 133  group (24.7%) (OR = 0.71, 95% CI: [0.53–0.94]). For the age group between 68 to 75 years of age (with reference to age 76 or older), the “good and stable health” group also had a lower percentage (23.5%) compared with “poor and improving health” group (24.9%) (OR = 0.73, 95% CI: [0.56–0.95]). For gender, both the “good and stable health” (30.0%) and “excellent and stable health” (24.5%) groups had lower percentages of women compared to the “poor and improving health” group (44.8%) (OR = 0.48 95% CI: [0.40–0.58] and OR = 0.36, 95% CI: [0.27–0.47], respectively). For distance to clinic, the “good and stable health” group (11.1%) had a smaller proportion of patients living a great distance from their clinics compared with the “poor and improving health” group (19.0%) (OR = 0.63, 95% CI: [0.48–0.83]). For the CHADS2 scores, both the “good and stable health” (mean = 1.04) and “excellent and stable health” (mean = 1.00) groups had lower average stroke risk score compared with the “poor and improving health” group (OR = 0.78 95% CI: [0.71–0.86] and OR = 0.82, 95% CI: [0.72–0.94], respectively).  For ablation therapy, there were differences in the time intervals within the groups. For example, within 0 to 6 months after initial consultation, the “good and stable health” group (5.4%) had a smaller percentage of patients who received ablation therapy compared to the “poor and improving health” group (20.6%) (OR = 0.28, 95% CI: [0.19–0.41]). Between 6 months to 1 year after initial consultation, both the “good and stable health” (2.7%) and “excellent and stable health” (5.8%) groups had lower percentages of patients who received ablation therapy compared to the “poor and improving health” group (10.3%) (OR = 0.27 95% CI: [0.15–0.46] and OR = 0.54, 95% CI: [0.34–0.88], respectively). Between 1 year to 1.5 years after initial consultation, the “good and stable health” group (2.0%) had a lower percentage of patients who received ablation therapy compared to the “poor and improving health” group (5.2%) (OR = 134  0.46, 95% CI: [0.25–0.86]). For more than 2 years after initial consultation, both the “good and stable health” (2.3%) and “excellent and stable health” (3.2%) groups had lower percentages of patients who received ablation therapy compared to the “poor and improving health” group (10.1%) (OR = 0.25 95% CI: [0.14–0.44] and OR = 0.29, 95% CI: [0.15–0.56], respectively). For anticoagulation therapy, there were only differences in one specified time interval. For example, within 0 to 6 months after initial consultation, both the “good and stable health” (62.1%) and “excellent and stable health” (60.6%) groups had lower percentages of patients who received ablation therapy compared to the “poor and improving health” group (76.3%) (OR = 0.70 95% CI: [0.56–0.86] and OR = 0.53, 95% CI: [0.41–0.69], respectively). Stated the other way, the “poor and improving health” group compared to both the “good and stable health” and “excellent and stable health” groups was more likely to have greater proportion of patients less than 60 years of age with 76 years or older as reference (OR = 1.69 and OR = 1.64, respectively), women (OR = 2.08 and OR = 2.78, respectively), with higher risk of stroke (OR = 1.28 and OR = 1.22, respectively), and those more likely to receive ablation therapy within 6 months to 1 year (OR = 3.70 and OR = 1.85, respectively) and more than 2 years (OR = 4.00 and OR = 3.45, respectively), and have been prescribed anticoagulation therapy within 6 months (OR = 1.43 and OR = 1.89, respectively) from initial consultation.  To evaluate the overall model, a multinomial logistic regression analysis was performed by regressing the identified three-class trajectory groups on the predictor variables (see Table 4-6).    135  Table 4-6.   Analysis of Predictive Values of the Growth Mixture Model with Predictors  Sensitivity Specificity Positive predictive value Negative predictive value Excellent and stable health 0% 100% N/A 83.9% Good and stable health 4.7% 96.3% 37.9% 67.3% Poor and improving health 97.4% 3.4% 59.9% 46.8% Note. N/A = Not applicable because the model did not identify anyone in this group.  The model explained 7.6% (Nagelkerke R2) of the variance in the 3-class trajectory group and correctly classified 97.4% of the cases in the “poor and improving health” group, 4.7% of the cases in the “good and stable health” group, and 0% of the cases in the “excellent and stable health” group (the model did not predict anyone in this group). 4.4 Summary Using the restricted standard model, we found three groups of patients with distinct trajectories. The “poor and improving health” group consisted of 63.6% of the sample, who started at a relatively low baseline scores and gradually improved at each follow-up visit. The “good and stable health” group consisted of 27.9% of the sample, and were patients with higher baseline scores that remained relatively stable over time. The “excellent and stable health” group consisted of 8.5% of the sample and were patients with very high baseline scores that remained little changed throughout the follow-up period. We also found characteristics that predicted the likelihood of membership in these three groups. When comparing the “poor and improving health group to both the “good and stable health” and “excellent and stable health” groups, the “poor and improving health” group was more likely to have less than 60 years of age category (with reference to age 76 or older), more likely to have been women, had higher CHADS2 scores, received ablation therapy within 6 months to 1 year and more than 2 years after the consultation, and been prescribed anticoagulation therapy within 6 months of the initial 136  consultation. Although multinomial logistic regression showed that the overall model had low explanatory power (Nagelkerke R2 = 7.6%), the model showed high classification rate for the “poor and improving health” group.   137  Chapter 5: Discussion The purpose of this study was to explore the analytical potential of patient-reported data stored in clinical registries, with the aim to explicate the process of conducting relevant and meaningful analyses with such data. The case example was patients with atrial fibrillation who were seen in specialized clinics. In this chapter, the significance of the results is interpreted and explained in light of this purpose. This chapter consists of six major sections. The first describes the results while noting some practical implications. The second discusses the results pertaining to model building, with a focus on the methodological implications of analyzing PROMs data collected for and stored in clinical registries. The third provides the results pertaining to the relationships among the variables, with a discussion of the implications for the established conceptual framework. The fourth section notes the strengths and limitations of the study. The fifth highlights some recommendations for nursing with key guiding questions. Lastly, future research directions are discussed.  5.1 Descriptive Results with Some Practical Implications This is believed to be the first study that examined PROMs in patients with AF using registry data to identify their patterns of change over time and the factors that contributed to experiencing a particular trajectory. We considered several growth curve models and found that GMM provided a relatively flexible method compared with traditional growth models (e.g., MLM and LGM). It closely aligned with the patient-centred approach by identifying multiple unobserved subgroups of patients with AF and accounted for individually-varying times of observation. The restricted standard model indicated that the optimal fitting model was a three-class trajectory model. The three groups that emerged were characterized as (a) poor and improving, (b) good and stable health, and (c) excellent and stable health. 138  A notable finding in our study was that there were multiple subgroups identified with distinct patterns of change. By using GMM, we were able to describe the longitudinal change within each subgroup and examine the differences in change among these subgroups. One of the key advantages of this method was that group differences in longitudinal change can be identified and described even when the grouping variable is latent or unobserved. In our study, we found that the three trajectory patterns were primarily distinguished by their baseline values with distinct patterns of change over the duration of the follow-up period. For example, the “poor and improving health” group began with the lowest average baseline value of 52.9 points (T0) which increased by about 12-points at the second follow-up visit (T2 = 65.1). This 12-point change is associated with a “minimal important improvement” (ranging between 6 to 19 units), though it did not exceed the 19-point change that is considered to be moderately large (Dorian et al., 2013). The “good and stable health” group began with an average baseline value of 82.3 points (T0) and increased by about 6-points at the second follow-up (T2 = 88.1), which barely meets the minimal important improvement from a patient’s perspective. The “excellent and stable health” group began with an average baseline value of 96.8 points (T0) and reached somewhat of a plateau with about a 1-point change at the second follow-up (T2 = 97.7) because of a “ceiling effect” where further potential improvement could not be observed. The change in health trajectories found in our study was consistent with the results of a closely related study by Flint et al. (2017), which suggested that the relative improvement in health status in outpatient settings may be modest at best. Although registry-based studies are generally hypothesis generating (i.e., developing hypotheses after the data are collected) and more rigorous studies would need to be conducted, there is great potential in GMM and the identification of different health trajectories, which 139  clinicians can potentially use to assess for differences in patients’ health trajectories and provide tailored interventions based on those differing trajectories.    In addition, we examined how potential covariates (age, gender, distance to clinic, CHADS2 score, ablation therapy and anticoagulation therapy) predicted trajectory membership. The findings indicated that compared to “good and stable health” and “excellent and stable health” group, the “poor and improving health” group included a higher proportion of younger age category (less than 60 years of age with reference to age 76 or older), women, those at increased stroke risk (CHADS2), and who had ablation and anticoagulation therapies during follow-up. The practical implication of this finding is that different mechanisms can affect patients with different health trajectories. For example, a covariate found to affect patients in one health trajectory may not have the same effect on patients in another health trajectory. Clinicians can also potentially use this and other information about covariates to inform patient teaching and provide tailored interventions for different subgroups of patients who may be at higher risk of poor outcomes. Nevertheless, GMM is a constrained exploratory technique that is limited by the specific bounds imposed during modelling specification. Thus, we focus the following discussion on the methods, including the processes underlying the model building decisions since such decisions can affect the results, potentially leading to different clinical conclusions. From this perspective, our contribution to the field is in bringing more transparency to the method of analysis in highlighting issues related to model building with regards to PROMs data in registries. 5.2 Results Pertaining to Methods Applied to Analyze PROMs with Registry Data One of the major issues in identifying and estimating our model was selecting an appropriate time metric because the interpretation of the parameters (i.e., the intercepts and 140  slopes) and subsequently how the trajectories actually look depended on this very choice. In longitudinal studies, and in registries that routinely collect patient-reported data in particular, individuals are often assessed at different time points and the number of assessments for each individual may vary, resulting in an unbalanced design (van de Schoot, Sijbrandij, Winter, Depaoli, & Vermunt, 2017). Simulation studies have shown that ignoring individual differences in time points can lead to biased estimates in the baselines (intercepts) and rates of change (slopes) in the trajectories (Aydin et al., 2014; Coulombe et al., 2016). In our study, we found that how we coded time could affect the results. For example, if we coded the time metric as the number of follow-up visits, the health trajectories might have shown subgroups of patients with more rapid improvements over time. In contrast, if we coded the time metric as a continuous variable in years, the health trajectories might have shown a subgroup of patients whose health was apparently decreasing over time. However, it was obvious that ignoring the spacing of the observations would not have accurately represented the underlying processes of daily practice where patients have different numbers of follow-up visits at varying time intervals. Studies have shown that the length of the follow-up period and the spacing in between time points affect the number of trajectories that can be found (Eggleston, Laub, & Sampson, 2004; Piquero, 2008). Therefore, the usefulness of the model depends not only on transparently reporting the underlying time-related processes but also on how the time metric in the model is specified.  According to van de Schoot et al. (2017), the fit of the model should not be used to determine the specification of the time metric but rather decided on prior to running the analyses. In registry-based studies, it is quite likely that there will be more than one possibility for coding of time. Thus, in our study, we plotted several graphs of the different time metrics, including the AFEQT questionnaire scores by the number of follow-up visits and by time in years, and 141  provided a rationale that the time metric had to account for the individually-varying times of observation to reflect what actually occurs in the daily processes of the clinics. These graphs represented a more complete analysis of the trajectories of change over time and provided opportunities to gain new insights that were not readily apparent from the output of the final model.   Following the determination of the time metric, another modelling issue was how to specify the growth parameters (i.e., intercepts and slopes), and in particular their random effects (i.e., the variances of the intercepts and slopes) “because the latent classes and random effects compete for the same variability in the trajectories” (Palardy & Vermunt, 2010, p. 555). Previous studies suggest that in some situations, constraining the random effects could lead to over extraction (Bauer & Curran, 2004; Lubke & Neale, 2006). Our results were consistent with these observations. For example, compared with the unrestricted model with no constraints, the restricted model with constraints on the slopes identified an additional class (i.e., three classes with the restricted model and two classes with the unrestricted model). In our study, a simple strategy to protect against over extraction would have been to choose a model with no constraints or the unrestricted model. However, in the context of GMM, the intercepts and slopes varied far more within the groups than between the groups, and the variance in the group slope is usually much smaller than the variance in the group intercept (Palardy & Vermunt, 2010). Thus, the variance of the groups’ slopes from the model with no constraints or the unrestricted model tend to be very small, which can result in a drastic slowing of parameter estimation and sometimes even non-convergence (Palardy & Vermunt, 2010). In these cases, it may be better to constrain the variance in the group slopes to zero and conduct a sensitivity analysis to examine whether 142  placing constraints impacts the number of classes identified (as was done in our case running both the restricted and the unrestricted random effects model).  Because of these issues, the decision in selecting the final model is not always straightforward. In our study, the statistical fit indices (i.e., AIC, BIC, and SABIC) supported the restricted standard 3-class model compared with the unrestricted 2-class model; however, some researchers may choose the unrestricted 2-class model based on whether the additional class adds to the substantive interpretation of the findings. For example, some could argue that the additional class identified (e.g., excellent and stable health group) paralleled another class that was larger in size (i.e., good and stable health group), and that the unrestricted 2-class model would have been a preferred model. While the selection of the model is often based on statistical fit, there is some level of subjectivity in arriving at what constitutes “fit” and therefore in selecting a final solution.  Once the number of classes with an appropriate GMM were specified, the next step was to include the covariates/predictors to predict class membership. One particular concern was that the addition of covariates in the model could alter the distribution of the random effects and thus affect the number and composition of the classes identified (Palardy & Vermunt, 2010). This implies that the assignment of individual patients to classes or subgroups based on their posterior membership probabilities may also be altered.  While there is general agreement that incorporating covariates is beneficial to provide more accurate parameter estimates and recover the correct number of classes (Huang, Brecht, Hara, & Hser, 2010; Li & Hser, 2011), there has been little consensus on when these covariates should be included. Some researchers argue that covariates of latent group membership should be included when deciding on the number of latent classes (Lubke & Muthén, 2007; Muthén, 143  2003), while others advocate their inclusion only after the number of classes have been identified (Bauer, 2007; Nylund-Gibson & Masyn, 2016). This distinction has played an important role in the development of many new analytic techniques in handling covariates, including variations of the 1-step (joint model estimation) approach that supports the former argument and the 3-step (stepwise estimation) approach that supports the latter.  Although the 1-step approach provides more accurate parameter estimates when appropriate covariates are included, most researchers may prefer using the 3-step approach for several reasons. The first is that the construction of a growth trajectory and examining how covariates affect these trajectories are often seen as two different steps in the analysis (Bakk, Tekle, & Vermunt, 2013). For example, the latent classes that we identified were based on the different trajectories, and the covariates were based on different biological functioning and individual/environmental characteristics (e.g., age, gender, stroke risk score, and distance from clinic). It would be difficult to argue that these covariates should be included at the same time as the data used to identify the different trajectory groups as a means of examining the predictive validity of the latent classification. Another problem with the 1-step approach is that simultaneously including a large number of covariates in a single step may be too cumbersome due to the sparseness of the frequencies and the increased computation time (Clark & Muthen, 2009). However, the 3-step approach is not without its own challenges, as this method does not account for the classification errors that may systematically underestimate the association between the predictors and class membership (Bakk et al., 2013; Bolck, Croon, & Hagenaars, 2004; Vermunt, 2010).  In our study, we used the modified 3-step BCH method that uses a weighting procedure to account for this possible classification error, which has been shown to provide less biased 144  estimates (Asparouhov & Muthén, 2014a). Since many methods have been proposed in predicting latent classes that could invalidate the results and it is difficult to verify outside of the artificial context of simulation studies, articulating the methods and the rationale for each decision should be given more standing because they enable researchers to examine model specifications that could potentially change the results (and therefore their interpretation).  In addition, we also included time-varying predictors (ablation and anticoagulation therapy) to recognize that the start of treatment may vary among patients, which in turn may affect patients’ health trajectories. For example, patients who received ablation earlier in their trajectory may have been different from those who received ablation at a later stage when they may have exhausted all other options for treatment. In GMM, time-varying predictors are often directly included in the model as predictors of the repeated measures (Bollen & Curran, 2006; Muthén & Curran, 1997). Although the direct inclusion of time-varying predictors makes sense if the patients had received their treatment at the exact same time that the PROMs data had been collected, this is often not the case in practice. For example, if a patient had a total of three clinic visits including an initial consultation at time 0, a second visit at 1 month, and a third visit 6 months later, the patient could have had ablation at 3 months or possibly re-ablation 1 year later. If ablation was directly included in the model as a predictor of the repeated measures, it would have led to an inconsistency from what actually occurred in practice (i.e., the ablation was not directly tied to the clinic visit when the data were collected). To be consistent with the clinical setting, ablation therapy required being specified as individually-varying for each patient. According to Therneau and Crowson (2019), a simple solution to this issue is to use time intervals. In our study, ablation and anticoagulation therapy were specified within 6-month intervals based on their temporal distribution following the initial consultation. In this way, the 145  inclusion of time intervals for predictors brought greater alignment with patient-centred approaches that can better inform when certain interventions may be appropriate for different patients.  5.3 Results Pertaining to the Conceptual Framework The time component in our revised Wilson and Cleary (1995) framework was helpful in building longitudinal models that reflected the trajectories of change over time for each patient. In particular, the use of GMM allowed us to account for the exact time of each patient’s follow-up visits (i.e., individually-varying times of observation) and to identify “individual” trajectories. However, consistent with the results of our study, we can make further modifications to the conceptual framework (see Figure 5-1).   Figure 5-1. Suggested Revisions to the Conceptual Framework.  For example, we modified the time component to  “Heterogeneity in Individual Trajectories Over Time” to highlight the different subgroups and positioned this new component as ranging “over” the other variables. We made this change because GMM assumes that the population is 146  heterogenous with multiple individual trajectories before the association between variables and different trajectories are explored. This suggested revision to the framework may bring more clarity to modelling that considers different patterns of trajectories at the outset and to distinguish the models from a variable-centred approach (an assumption that underlies most clinical research), which would treat all individuals as a single group before any variables are examined. Another change we made was the shape of the treatment component with an arrow to highlight the continuum of treatment in the care delivery. In our study, we found that the “poor and improving health” group was more likely to have received ablation within 6 months to 1 year and more than 2 years of the initial consultation, and been prescribed anticoagulation therapy within 6 months of the initial consultation, compared with the other two groups (“good and stable health” and “excellent and stable health”). This type of information is pivotal for clinicians to guide the timing and evaluation of interventions. For example, during the first six months, clinicians can focus their attention on helping patients in the “poor and improving health” group accept their diagnosis and understand the benefits of anticoagulation therapy to help improve their perceptions of their health.  To further strengthen the conceptual framework, we included other predisposing factors in the individual and environmental characteristics domain. In our study, the low explanatory value of the model (Nagelkerke R2 = 7.6%) indicated that the current variables did not have strong predictive power. Based on an integrative literature review of associated factors and self-reported health in patients with AF (Son, Baek, Lee, & Seo, 2019), we included the following predisposing factors: for the individual characteristic domain, we included exercise intervention (Lakkireddy et al., 2013), alcohol use (Dixit et al., 2017), sleep (Maryniak et al., 2006), employment (Fransson et al., 2015), and psychological profile (includes anxiety) (Polikandrioti 147  et al., 2018); for characteristics of the environment, we included financial burden/income (Guhl et al., 2019). Exploring the extent to which these predisposing factors affect GMM and improve its overall predictive value is an important area of future research.  While adding additional predisposing factors may help to further explain the variation found in the model, we cannot completely remove from consideration whether the AFEQT questionnaire itself was an appropriate tool to assess self-reported health. In our study, we found that two trajectories reached close to a ceiling effect (“good and stable health” and “excellent and stable health” groups), suggesting that subtle changes in patients’ health cannot be evaluated using the AFEQT questionnaire. The importance of selecting an appropriate PROM is highlighted based on previous studies showing that PROMs scores may improve over time regardless of treatment efficacy (Fichtner et al., 2012; Wokhlu et al., 2010), which may relate to the low sensitivity of certain PROMs to detect changes associated with treatment. In this case, a more focused AF-specific symptom assessment such as the Mayo AF-Specific Symptom Inventory (MAFSI) (which includes a checklist of 12 symptoms with a Likert-type scale that ranges from 0 (never) to 4 (always)) may be a better endpoint for treatment that can more directly reflect changes in rhythm status.  The revised conceptual framework may also be an opportunity to address new types of research questions for nursing, such as “What emergent subgroups of trajectories can be identified following ablation?” and “Can the differences in the subgroups be explained by different individual and environmental characteristics?” Since these questions suggest the existence of identifiable subgroups, the revised framework can be used to generate new evidence that can better inform clinical care by highlighting differences in trajectories and tailoring interventions based on those differing trajectories. For example, nurses can look for worsening 148  subgroups of patients following ablation, and then examine various individual and environmental characteristics (e.g., exercise, sleep, and financial burden) to assess the type of interventions that may be necessary for these group of patients. The use of the revised framework is especially pertinent in advancing nursing science and practice because it accounts for the uniqueness of “individual” health trajectories and also considers the complexity of broader individual and environmental characteristics associated with patient outcomes.   In our study, we identified several distinct health trajectories, which were also distinguished by the patients’ status on predictors of biological functioning (supplemented with treatments undertaken), and individual and environmental characteristics. For example, one identified trajectory (“poor and improving health”) was more likely to consist of patients: (a) younger age category (less than 60 years of age with 76 or older as reference), (b) who were female, (c) with higher stroke risk scores, (d) who had received ablation therapy within 6 months to 1 year and more than 2 years after the consultation, and (e) who had been prescribed anticoagulation therapy within 6 months of the initial consultation compared with the other trajectories (“good and stable health” and “excellent and stable health”). This type of information provides a more nuanced understanding about patients who may be more or less at risk for adverse trajectories, rather than assuming that all variables similarly apply to all patients. The explicit time component, at the outset, was also helpful in considering how variables may vary among patients over time. This allowed us to consider treatments such as ablation and anticoagulation as time-varying variables by including time intervals to indicate that the commencement of treatment varied, which could provide information about possible times in which interventions could be targeted for better patient outcomes or when patients required particular types of support. Although we included clinician-reported symptoms and quality of life as part of the revised 149  conceptual framework, we did not fully test these relationships because of the nature of our research question and the limitations of the data collected with the existing instrument for the registry. We recommend further testing of the revised conceptual framework that includes clinician-reported symptoms with the ultimate focus on quality of life, which would capture various domains unique to individual patients. As previously mentioned in Chapter 2, the quality of life concept can be assessed with the Schedule for the Evaluation of Quality of Life, which allows patients to specify areas of life most important to them. The results of the fully developed and tested conceptual framework would provide unique information about patients that could identify more specific targets of intervention, which may not be fully captured using standardized outcome measures, and that would enhance the understanding of patients’ unique perspectives.  5.4 Strengths and Limitations  5.4.1 Strengths of the study. The main strength of this study is the novel evaluation of the registry-based population and a longitudinal design perspective in the identification of health trajectories. The study population was from a real-world clinical setting who required management of AF from five clinics in BC. We had comprehensive descriptive information about the patients who completed AFEQT questionnaires as well as those who did not. In addition, the longitudinal design perspective allowed us to examine trajectories of change in patient-reported outcome measures using complex statistical models.  We considered various conventional growth models (e.g., multi-level and latent growth models) and used GMM to identify unobserved subgroups of patients with different health trajectories. The characteristics of these subgroups produced additional insights that enhanced 150  our understanding of patients who may be at higher risk of experiencing poor outcomes, and from a clinical perspective could provide useful information regarding the identification of patients likely to benefit the most from treatment. For example, in the context of AF, this may pertain to important clinical questions such as:  a) What are the health trajectories of patients with AF?  b) How do these patients change over time after treatment or clinic visits?  c) What factors are associated with these different trajectories?  Answering these questions can potentially inform the development of health programs for continuing care and tailoring of such programs for different subgroups of patients.  Another strength of the study was that the analysis was informed by the modified Wilson and Cleary (1995) framework, which guided the characterization of the change model and the selection of the variables. As part of this modification, we included the component “trajectories of change over time” to explicitly indicate that individual health is experienced as change over time. The specification of our GMM reflected this component by coding the time metric as individually-varying as well as the interventions received (ablation and anticoagulation), which were individually-varying. Although we did not fully test this modified framework, we made a beginning attempt to understand how biological function and its treatment as well as individual and environmental characteristics affect self-reported health status. Moreover, the data preparation and cleaning steps were partly informed by the author’s visits to the clinics to ask questions about their processes and workflow. These site visits, where the author had the opportunity to interact with clinicians (i.e., nurses, nurse practitioners, a cardiologist, an electrophysiologist, and a pharmacist) and observe interactions with patients, provided more 151  contextual information about the underlying variables and the selection of the eligible cohort that closely followed the AF clinic processes, which may not have been possible otherwise.  This study was also strengthened by using the latest recommended missing data techniques to address biases that may have been introduced at various stages. For example, instead of performing single-level multiple imputation, we used its multilevel variant in our longitudinal dataset. If we used the single-level multiple imputation, it would have led to inconsistent results from the underlying multilevel structure in which repeated measurements in PROMs (level 1) were nested within individuals (level 2). Moreover, we included auxiliary variables using all available variables including death status to ensure that the imputation model had the most information possible for a more accurate imputation. To account for patients who had unequal measurement occurrences when our dataset was transformed to the wide format for modelling, we used FIML to obtain unbiased parameter estimates and standard errors. The use of these missing data techniques reflects the most recent best practices to address the challenges of missing data.  But perhaps this study’s greatest strength is how it brings transparency to a method of analysis in the conduct of research with PROMs registry data. This approach detailed above allowed us to examine the underlying assumptions of different models (e.g., variable- and patient-centred approaches) and explore modelling issues (e.g., selection of an appropriate time metric, specifying growth parameters, and when to include covariates to predict class membership) that could have potentially led to different conclusions. Thus, this study brings important new insights about the analysis of PROMs data stored in registries that can be used to generate information to provide greater alignment with patient-centred approaches. 152  5.4.2 Limitations of the study. The principal concern with registries, and a major limitation of this study, is the quality of the data, since quality standards have not previously been well established or consistently reported (Gliklich et al., 2014). In practice, incorrect patients can be registered or data items can be inaccurately recorded or not recorded at all (Arts et al., 2002). The resulting degree of data quality affects how data can be used and, ultimately, the level of confidence in research findings or other decisions based on the data. Registries have several reasons for missing data.  The first reason is questionnaire nonresponse, which occurs when patients do not complete a questionnaire during one or more follow-up periods. In our study, among the 13,113 eligible patients, 7,439 patients completed at least one questionnaire during follow-up while the remaining 5,674 patients did not complete a questionnaire at all. Upon closer examination, the patients who did not respond differed from those who responded by living further away from their respective clinics, being at higher risk for stroke, and being more symptomatic. The most concerning issue with questionnaire nonresponse is selection bias because it can lead to misleading inferences that patients in the clinics are generally improving when their health may actually be declining over time. This may also in part explain why we did not observe any subgroup of patients with decreasing health trajectories. While there are statistical methods to adjust for selection bias (e.g., propensity score matching), the results may differ depending on the type of method used, and the inherent challenges to making appropriate inferences would still remain (Stukel et al., 2007). In addition, questionnaire nonresponse also leads to a reduction in sample size, which not only affects the precision of parameter estimates but also the type of growth model that can be estimated. For example, we initially attempted to run a quadratic GMM with four time points but were unable to obtain good fit due to limited sample size (n = 153  401). Thus, we were only able to run a linear GMM with three time points (due to sample size being larger; n = 1,285), which was another limitation because we were not able to examine non-linear trajectories of change to identify subset of patients who may have decreased and then improved. The second reason for missing data is item nonresponse, which occurs when patients complete a questionnaire without providing a response for one or more of the items. In our study, of the 10,426 AFEQT questionnaires, 3,393 questionnaires had more than one missing item. While there may be system checks in place to ensure complete data entry, item nonresponse may still occur because these data elements are optional and are often not mandatory fields, which has implications in supporting patient-centred objectives within the clinic.  Another limitation is that the author was not directly involved in the data collection process, which is often inherent to the nature of secondary analysis of registry data (Cheng & Phillips, 2014). Therefore, we may have missed contextual features or glitches in the data collection that are important to the interpretation of the analyses. Although we examined the data dictionary for variable definitions and changes in data collection procedures, as well as reviewing descriptive statistics and frequency distribution graphs, to scan for coding inaccuracies and incorrect values, it is possible that many important nuances could have been missed that could have affected the analysis and interpretation of our results. For example, anecdotal evidence from the staff suggests that the CCS-SAF may not be interpreted in the same way across the clinics, which may raise questions about the suitability of this covariate for inclusion. However, since we were focused on exploring the analytical potential of patient-reported data in registries and not necessarily on making substantive clinical conclusions, the possible drawbacks of specific variables were considered to be of lesser importance. Lastly, the same limitations of retrospective observational studies apply as correlation does not imply causation. Thus, caution 154  is warranted in deriving causal interpretations and findings would need to be replicated in a well-designed prospective study.  Nevertheless, these findings show that while registries can be of major value to the research community, despite the effort invested in generating and maintaining them, there is still much work needed to improve the integration of PROMs in evaluation frameworks and clinical practice. Van Der Wees et al. (2014) identified three main factors associated with the establishment of routine data collection: (1) the availability of electronic data capture, (2) the need to avoid disrupting the workflow, and (3) the need to obtain high response rates from patients. Electronic data capture for patient-reported data can certainly help to provide immediate feedback to clinicians and avoid some of the issues of manually entering data, which can also potentially reduce the disruption of workflow (by automatically capturing data) and improve response rates from patients (by allowing patients to complete questionnaires via electronic devices to be directly uploaded to the database). To obtain high response rates from patients, patients must believe that the results are actually being used by their clinicians to inform their care planning. However, all of these strategies involve substantial upfront and ongoing investment in people and information technology capabilities that require engagement of multiple stakeholders, including patients, clinicians, and administrators who see a clear value in collecting PROMs. Without greater clarity and a shared vision for the use of PROMs, the ability of the healthcare system to provide patient-centred care will not be fully realized. Thus, the use of PROMs requires full engagement of multiple stakeholders, including the public, and ongoing efforts to establish trust to maximize their impact to improve patient outcomes through the entire continuum of care.  155  5.5 Recommendations for Nursing  Increasingly, nurses of all stripes, including practising nurses, nurse educators, and nurse scientists are expected to participate in evidence-based practice to ensure the use of the best available evidence in making clinical decisions (Curtis, Fry, Shaban, & Considine, 2017; Majid et al., 2011). Registry data, which are a collection of real-world data, can be of great value to this endeavour (Blonde et al., 2018). For example, PROMs data in registries, collected in daily practice, can provide insights about patients’ perspectives about their health and quality of life that can be used to inform individualized patient care. While registries may not be ideal vehicles for the collection of events that “occur” outside practice or when patients need to recall all their clinical history (Christian et al., 2018), they can be enriched by linkages to additional data sources that can help obtain higher levels of complete data and to answer a wider range of clinical questions that were not possible before.  However, there is a lack of practical guidance in the literature about the specific skills and competencies necessary for meaningful analyses for those interested in using registries in their work. For example, a critical skillset to have before one uses registry data is to be familiar with a range of statistical programing software such as R, SAS and Mplus. While nurse researchers may be most familiar with using the graphical user interface of SPSS to manage their data, it may not be suitable for complex analyses and modelling tasks. Although there is a steep learning curve to become familiar with different programming languages, there are many freely available online resources and platforms to ask questions (e.g., Stack Overflow website). If nurse researchers do not become familiar with some of these programming languages, they may overlook many modelling considerations and not be able to assess studies using complex statistical analyses that could potentially lead to misleading results, which in turn could influence the decisions made in 156  patient care. Being versed in a range of statistical programing languages will also enable participation in wider conversations with researchers in other disciplines to ensure that registry data is shaped by nursing knowledge to improve relevance in practice and help shape healthcare service delivery to be more patient-centred. Another issue is that there is a major gap in knowledge translation and education resources to support the analysis of registry data, which leaves massive amounts of data untapped that can potentially provide useful insights to inform practice. Studies have shown that despite the potential of registries to improve patient-centred outcomes, registry data are fraught with methodological challenges that requires in-depth knowledge and expertise in working with complex statistical analyses (Cox, Kartsonaki, & Keogh, 2018; Gliklich et al., 2014). By providing practical examples of key skills and competencies required in preparing for the analysis of registry data, we are taking an important step in knowledge mobilization. Thus, the following section describes these skills and competencies (with helpful coding examples where applicable) to help facilitate the analysis of registry data. These skillsets include understanding informed consent and confidentiality, checking data accuracy, selecting variables and cohorts of interest, and transforming data structures. This section concludes by providing key guiding questions for nurse researchers interested in pursuing the use of registry data for their own research.  5.5.1 Understanding informed consent and confidentiality. In preparing for the use of registries, nurse researchers should be familiar with issues of informed consent (patient permission for data use) and confidentiality (non-disclosure of identifiable patient data and security measures). One of the most important discussions in the context of clinical registries is whether informed consent should be obtained from each enrolled patient. For large scale registry data, it is particularly challenging if consent is not universal 157  because the requirement of written consent (which is particularly important in intervention studies) introduces additional time, costs, and most importantly selection bias, such that the participating patients are no longer representative of patients in the registry and subsequently compromise the external validity of the research (Huang, Shih, Chang, & Chou, 2007; Tu et al., 2004). For these reasons, registry-based studies do not generally need to obtain informed consent as long as there is evidence that the study produces minimal risk, as assessed by Research Ethics Boards.  According to the Canadian Government’s Tri-council Policy Statement (2019), informed consent is waived when five main conditions are met: (1) the research involves no more than minimal risk to participants, (2) the lack of consent is unlikely to adversely affect the welfare of the participants, (3) the research could not practicably be carried out without the waiver or alteration, (4) when appropriate, the participants will be provided with additional pertinent information after participation, and (5) the waived consent does not involve a therapeutic intervention. Among these conditions, the main risk to the participant relates to condition 2 (i.e., the lack of consent is unlikely to adversely affect the welfare of the participant) based on the extent to which the information proposed for use in research is identifiable. The following categories provide guidance on assessing the extent to which information could be used to identify the participant (from most sensitive to the least) (Government of Canada, 2019): (1) directly identifying information (includes direct identifiers such as name, social insurance number or personal health number); (2) indirectly identifying information (includes combination of indirect identifiers such as date of birth or place of residence); (3) coded information (direct identifiers are replaced with a code so that re-identification is only possible by those with access to the code); (4) anonymized information (direct identifiers are irrevocably removed so that re-158  identification is not possible); (5) anonymous information (direct identifiers were never collected such as anonymous surveys). Based on these categories, while the least risk of identification of individuals is through the collection and use of anonymized or anonymous data (i.e., category 4 and 5), the usefulness of registry data becomes more limited because it cannot be linked to other data sources.  The next best alternative that is often used is de-identified data (i.e., category 3), which are provided to the researcher in de-identified form and the coded information is only accessible to a trusted third party who is independent of the researcher and the data stewards who provide the data. For example, certain multi-university organizations (e.g., Population Data BC) provide de-identified data while the existing key code is held in their storage facility to link new information (Pencarrick Hertzman, Meagher, & McGrail, 2013). Some trusted third party may provide a central server for data storage and analysis as well as other security measures including encrypted private networks through a firewall, token-based authentication, and an audit trail that is routinely reviewed to minimize accidental releases of data to provide further protection (e.g., The Secure Research Environment provided by Population Data BC fills this role). As an additional precautionary measure to reduce identification/self-identification, certain public authorities may set minimum cell size policy for the display of their data. For example, the policy may stipulate that the number of cases or events in a cell (e.g., admissions, discharges, patients, or services) containing a value of 1 to 5 cannot be reported directly. In these cases, the nurse researcher can further aggregate the data (e.g., combining multiple years or collapsing across categories) or coarsen the value by reporting < 5 in a given cell. Although all these measures will not necessarily guarantee confidentiality, these efforts in concert make the disclosure substantially less likely.  159   Once the Research Ethics Board approves the research project and waives the informed consent requirement, the nurse researcher will then enter into an information-sharing agreement with the data stewards (organizations providing the data) on the specifics of creating data linkages and de-identifying data, and the security protocols in managing the data to ensure confidentiality. The data sharing agreement is a legally binding contract between the researcher and the data stewards that includes governance in the collection, transmission, storage, security, analysis, archiving, and destruction of data. Depending on the project, this process may involve meetings with the data stewards to discuss the minimum data set (e.g., set of variables that are used to collect and report data in the registry) that can be linked, the rationale for the use of each variable, and how the data will be managed. In some cases, the nurse researcher may need to contact their home institutions’ legal counsel for advice. It is helpful if the nurse researcher already has some familiarity with the people and the clinical context where the data are collected (e.g., through previous work experience) in navigating some of the administrative hurdles with the data delivery and unforeseen delays. Some data stewards may even request the completion of a data management plan tool (which may be required for some U.S. funding agencies) that asks questions about how the data will be managed through the entire life of a project (including how the data will be stored and backed up during research, what the anticipated storage requirements for the project will be, and the length of time the data will be stored) (www.dmptool.org).  It should be noted that some of the legal requirements are in place to protect patients and to adequately protect the data provided, not to prevent legitimate research. With careful planning, registry data can be of great benefit to participants by maximizing the value of the data already collected and reducing the burden by not asking them to complete additional questionnaires. However, these sometimes-lengthy processes may influence decisions regarding the selection of 160  certain variables and, in some cases, the feasibility of using registry data for secondary research purposes. For example, the wait time for data access requests and data releases can vary widely from several months to over a year, which should be taken into account when planning a registry-based study (Population Data BC, 2018d).  5.5.2 Checking data accuracy. While registry data present many opportunities to address a wide range of questions, with regard to the quality of the data there are inherent challenges to making appropriate inferences. While registries are not required to meet any set standard (e.g., the International Conference on Harmonization Good Clinical Practice Guideline developed for clinical trials) to ensure scientific validity and credibility of the data, data stewards may have some form of quality checks in place (e.g., chart audits). However, since many healthcare organizations that hold registry data may not have the resources to maintain quality checks, the registry data that the nurse researcher receives may be prone to inaccurate data values. Thus, whenever possible, it is important to request another dataset from clinical or administrative sources to check the accuracy (e.g., data entry errors), data completeness, or out-of-range values. For example, if a patient’s age has an impossible value or is missing in the registry data, an appropriate age can be substituted from an administrative dataset. This is possible because the patient from the registry can be linked to information on the same patient from another data source.  To link individual-level patient data from different sources, there are generally two possible ways: deterministic and probabilistic record linkage (Sayers, Ben-Shlomo, Blom, & Steele, 2016). Deterministic (exact) record linkage is the process of linking information that perfectly matches a unique set of identifiers. For example, in Canada, a unique permanent health identification number is assigned to all eligible residents, which enables deterministic record 161  linkages. In many jurisdictions, an individual-specific health number is replaced with a project-specific anonymized study identification number before the data are provided to researchers (with a “trusted third party” only having access to the identifier) as a best practice to protect confidentiality (Harron et al., 2017).  If an individual-specific health number is not available to provide an exact match, probabilistic record linkages can be used. Probabilistic record linkages attempt to link information based on non-unique identifiers including date of birth, gender, and place of residence. However, the use of the probabilistic approach introduces considerable complexity in terms of errors in the linkage keys (e.g., differences in how the data are captured and maintained in different datasets) and addressing the lack of unique keys for linkage (Sayers et al., 2016). While the use of either linkage approach may depend on several factors (including resources, the research question, and the availability and quality of the data), nurse researchers should at least be familiar with how to perform deterministic linkage. An example of applying this linkage and issues that may arise are shown with two separate datasets: the registry dataset containing age and the PharmaNet dataset containing both age and drug information (see Figure 5-2).    Figure 5-2. Matching Two Different Datasets  162    The left dataset is a sample dataset from the registry with a project-specific anonymized study identification number (e.g., StudyID) while the right dataset is a sample dataset from PharmaNet with the same number. The deterministic record linkage can be used to match the two datasets using the unique StudyID to match one-to-one (e.g., using the merge function in R statistical software). After merging, the Registry StudyID 1 with age as 400 can be replaced with a more appropriate age of 64 from the PharmaNet StudyID 1. For StudyID 3, there seems to be discrepancy in age with the registry dataset reporting 59 while the PharmaNet dataset reporting 54. Since both ages seem to be plausible, it may be necessary to check every row for additional discrepancies, and a decision would need to be made on which data are more accurately coded. In addition, the sample size shows that there are more patients in the registry dataset (N = 910) compared with the PharmaNet dataset (N = 900), which suggests that there were some missing medication records. Thus, it is important to also report variable completeness (e.g., % of missing values), descriptive statistics (e.g., mean, SD, min, max, skewness, and kurtosis), frequency tables, and distribution graphs of the main variables in the analysis. This provides information about the use of the coding pattern for each variable and about the profile of missing data for each variable.  Other data checking considerations include data formatting (e.g., dates that are stored in different formats such as 11/27/2019 versus 27/11/2019), chronological sequences of dates (e.g., the date of initial consultation should occur before the date of discharge), and consistent information (e.g., patients reported as having had an intervention should continue to have that intervention reported throughout the subsequent years). Because of these complexities, analysis of registry data can easily contain hundreds of steps and lack of reproducibility in this analysis process can be a severe issue. Nurse researchers who continue to use point-and-click statistical 163  software will be ill-prepared when they encounter this type of data. Thus, the use of a programming language provides a more transparent and reproducible approach that enables documentation of each step of the analysis. For example, a line of code would change a data value, and a comment code could be included that explains what the code is doing and why. These codes can also be used to go back to the analysis steps or to verify with other researchers when necessary. This way of “reproducible research” in the method of analyses should arguably be given more standing as the path to the research findings may be equally, if not, more important than the findings themselves.  5.5.3 Selecting variables and cohorts of interest. Because the use of registry data may not involve the same control over what data have been collected or are available, it is important for nurse researchers to prepare a minimum set of variables that would be required to answer their research question. The minimum dataset contains a set of mandatory data elements that are collected for every patient, which can be contrasted from optional data elements. The selection of variables should primarily be from the mandatory data elements; otherwise, the nurse researcher may risk incomplete data entry by using optional variables, which may lead to issues during subsequent data analyses. However, even when mandatory data elements have been identified, nurse researchers should be aware of underlying biases in the data. For example, the nurse researcher may find biases in the cardiac procedure data if only a small non-representative sample of patients undergo this procedure, which could make this variable unreliable for inclusion. In these cases, nurse researchers can use their clinical knowledge to identify whether variables should be included for further analyses. This clinical expertise is also useful in understanding the processes and workflow of care to 164  interpret the results of the data. If there are unresolved issues with data accuracy, the nurse researcher can directly contact the data stewards or a trusted third party for more information.  After obtaining the minimum required dataset, a special coding skillset that nurse researchers will need to know is selecting the cohort of interest or obtaining comorbidity information by using the International Classification of Diseases (ICD). Developed by the World Health Organization, the ICD is the international standard to identify health conditions, and used in health care to report diagnoses, track epidemiological trends and to assist in medical reimbursements (World Health Organization, 2019).10  Although the ICD codes undergo periodic updates, the current diagnosis coding system used in most healthcare systems is the ICD-10th version, which is organized into 21 chapters and has more than 68,000 codes (Cartwright, 2013). The codes in the ICD-10 are structured to have upwards of seven characters, depending on the level of specificity required for the given diagnosis (see Figure 5-3).   Figure 5-3. ICD-10 Coding Structure   10 Several countries have developed clinical modifications of the ICD codes to be used in their own hospitals and medical facilities, including Australia (AM), Canada (CA), Germany (GM), Korea (KM), Thailand (TM), and the United States (CM) (Jetté et al., 2010). 165  The first three characters designate the general category of the diagnosis, which starts with a letter, followed by two numeric digits. In this instance, the letter “I” designates that the diagnosis relates to the cardiovascular system. The numerals “7” and “0” indicate that the diagnosis is related to atherosclerosis. This category can stand on its own or can be further categorized, which further explains the etiology, location, severity, and other details. In this case, the numerals “2”, “0” and “1” indicate a diagnosis of “unspecified atherosclerosis of native arteries of extremities – right leg.” The last character, extension, describes the type of encounter from the initial encounter (A), subsequent encounter (D), and encounters related to a previous diagnosis are listed with the term sequela (S). In practice, however, not all ICD-10 codes are classified to this level of detail and it may often be the case that four characters are enough to describe the diagnosis and extension. In some instances, datasets may also contain ICD-9 codes, which differ from the ICD-10, by using three-digit numbers (i.e., 001 to 999) for the category, followed by a decimal and up to two digits for etiology (e.g., 440.20 for the same atherosclerosis diagnosis). In Canada, the discharge abstract database (which contains information related to hospitalized patients) uses the ICD-10 code, and the physician claims data (which contains patient information managed at outpatient settings) uses the ICD-9 code. It should be noted that some administrative datasets may use different coding systems for procedures. For example, the Canadian Classification of Health Interventions is the companion classification system to the ICD-10-CA for coding procedures and interventions in Canada.   Identifying chronic conditions using administrative data can be simple (e.g., using a single dataset in a given year) or complex (e.g., using both inpatient and outpatient encounters across multiple years). However, the use of administrative data for such purposes assumes that the data provide valid information about diagnoses, comorbidities, and clinical services (Walker 166  et al., 2012). Therefore, it is important to use previously validated algorithms because prevalence and incidence of chronic conditions can vary widely.  For example, Tonelli et al. (2015) provided administrative algorithms for 30 chronic conditions, which can be reliably used to facilitate comparisons of data about comorbidities across settings. In addition, several studies have shown that the different administrative data sources used and the duration of the “look-back period” or the time prior to hospitalization that represents the index event can influence the prevalence of chronic conditions (Chen et al., 2016; Rassen, Bartels, Schneeweiss, Patrick, & Murk, 2018). According to Chen et al. (2016), a minimum of one year look-back period is recommended for recovering information about patients’ comorbidities that influence the probability of survival.  These codes in combination can be used to identify a range of patient diagnoses for inclusion in a study or to calculate a comorbidity index for the patient population. For example, researchers have developed programming codes from the ICD codes to calculate different types of comorbidity scores (e.g., R package ‘comorbidity’ (Gasparini, Hojjat, & Williman, 2019) and SAS macro (Quan et al., 2005)).  5.5.4 Transforming data structures.  Another key skillset is to be able to transform data structures as needed. Typically, longitudinal data come in two different formats: long and wide. In the long format, there are multiple rows for each individual depending on the number of repeated measures in time. In contrast, in the wide format, there is only one row per individual with the repeated measures in time contained in multiple columns. Registry data come in different formats, and certain analyses or visualizations are designed for one specific form. For example, to obtain descriptive characteristics of the sample such as the average score or when fitting certain types of models 167  such as structural equation models, the data structure needs to be in wide format. For plotting and fitting models such as multilevel models, the data structure needs to be in long format. However, issues of data structure often arise when applying missing data techniques, particularly in registries that have collected data continuously with no clear cut-offs in time. An example of this issue is shown when the long format has been transformed to the wide format for individuals with a PROM score at each follow-up in Figure 5-4.   Figure 5-4. Transforming Data from a Long to Wide Format  In the long format data structure, there is only one missing value for a patient with StudyID 1 at time 0. However, if this data structure is transformed to the wide format, there is a total of five missing values (i.e., the patient with StudyID 1 has three missing values and the patient with StudyID 2 has two missing values). The number of missing values has increased because the patients did not complete the PROM at every follow-up from time 0 to time 3. If a single-level multiple imputation (e.g., standard multiple imputation method) is conducted with the wide data structure, the imputation would now have to account for the disproportionate number of missing values that have been introduced. On the other hand, if a single-level multiple imputation is conducted in the long format, it ignores the information from the other rows for the 168  same patient because it treats each row as an independently sampled case. This can lead to bias (e.g., underestimation of standard errors) and inconsistent imputations (Enders et al., 2016).  A hypothetical example is shown below to illustrate what can happen if a single-level multiple imputation is conducted with the long format data structure (see Figure 5-5).   Figure 5-5. An Example of Inconsistent Imputation with the Long Format Data Structure  The baseline age of a patient with StudyID 1 is imputed with a different value in the third row even though the first two rows indicated that the patient’s baseline age was 65 years of age, creating an inconsistent value for the patient that could not have been observed. In this case, multilevel multiple imputation is recommended when the data structure is unbalanced (e.g., when the data are collected continuously) and when transformation to the wide data structure leads to variables with high proportions of missing values (Grund, Lüdtke, & Robitzsch, 2018).  Multilevel multiple imputation works by considering both the repeated measurement (e.g., time) at level 1 and the individual (e.g., StudyID) at level 2 to impute missing values and avoids the problem of having inconsistent imputations. Unfortunately, most statistical software programs are not intuitively designed to handle missing data in long format data structure. For example, in R, the ‘MICE’ package in conjunction with the ‘miceadds’ needs to be installed to specify the level 2 variables during imputation (Grund et al., 2018). In Mplus, additional lines of 169  code need to be included by specifying CLUSTER = StudyID in the VARIABLE command, and TYPE = TWOLEVEL in the ANALYSIS command (see Example 11.7 in the Mplus user guide).  The performance of the multilevel multiple imputation also depends on other factors. For example, auxiliary variables (variables that are not part of the main analyses but may be associated with missingness) should be included in the imputation model (Collins, Schafer, & Kam, 2001; Graham, 2009; Schafer & Graham, 2002). When more information is included from the auxiliary variables, the imputation model can increase its ability to infer from the observed data with greater accuracy (for a step-by-step example of how to include auxiliary variables in multilevel multiple imputation, see Grund, Lüdtke, and Robitzsch (2018)). In addition, the inferences obtained from the multilevel multiple imputation can be improved by generating a larger number of imputations. Although more imputations can certainly improve the uncertainty in parameter estimation, it can also increase computational time. According to Graham et al. (2007), at least 20 imputations are recommended for 10% to 30% missing information in the dataset.  Once multiple long form datasets have been created (after multilevel multiple imputation), the dataset may need to be transformed to a wide format to conduct certain modelling techniques such as structural equation modelling. As shown in Figure 5-4, the transformation from long to wide format will naturally increase the number of missing values. However, instead of conducting multiple imputation again, another advanced missing data technique called full information maximum likelihood is recommended. In this method, missing values are not imputed but rather handled within the analysis model by assuming missing at random using variables included in the analysis information (in contrast to multiple imputation that can accommodate MAR based on additional variables that are available in the dataset but 170  not in the analysis) to obtain parameter estimates. This method is the default feature in Mplus when the model is run; in R it can be specified with an additional line of code (i.e., missing = ‘fiml’ in the ‘lavaan’ package).  After all the missing data have been accounted for, the final step is to pool multiple sets of results from each of the imputed datasets into a single set of results. The estimates are pooled by using Rubin’s rules (Rubin, 1987), which are calculated by accounting for the uncertainty of the imputation within a dataset (e.g., differences in predicted values from the dataset regarding each observation) and the imputation variance (e.g., differences between multiple datasets). This can be done by specifying additional lines of code (e.g., TYPE = IMPUTATION in Mplus or the pool function in the ‘MICE’ package in R) within the analysis model.  5.5.5 Key guiding questions.  Due to many challenges and unforeseen issues that may arise during analyses, nurse researchers may find that registry data are not suitable for immediate statistical analysis, and considerable preparation and cleaning work are needed. In addition, the scope of the analysis and the ability to answer a research question may often be limited by the quality and completeness of the data. For example, the nurse researcher may have a general research question in mind before undertaking research using registry data. However, the nurse researcher may find that the registry data may not contain all the necessary variables or may not be of sufficient quality to address the research question. This would often result in either modification of the research question or the analysis plan based on the best available data. In an effort to guide nurse researchers in the use of registry data, three key guiding questions are provided:   171  1) What is the purpose of the registry and the rationale for its development?  The purposes for registries can vary widely. Registries are used for many purposes including the recruitment of patients for clinical trials, development of treatment or monitoring of outcomes over time, generating research hypotheses, or for improving and monitoring the quality of health care (Blumenthal, 2019; Gliklich et al., 2014; Hoque et al., 2017). They may also pursue a specific, focused research agenda, collecting data for a limited time, or may collect data on an indefinite basis to answer a variety of existing and emerging research questions (Gliklich et al., 2014).  Analysis of registries involve either primary or secondary data sources, based on the relationship of data to the registry purpose (Gliklich et al., 2014). Primary data sources are collected for direct purposes of the registry (i.e., primarily for the registry) and are typically used when the data of interest are not available elsewhere or, if available, are deemed not to be sufficient for use. In contrast, secondary data sources have already been collected for purposes other than the registry (e.g., standard care or billing records) and are typically used to examine a novel research question. Although there are many advantages and disadvantages to using either data source with regard to time and resources, the main distinction is that researchers can control the collection of study variables using primary data sources (Kodra et al., 2018). This is an important distinction because the analysis of registry data often involves researchers who are not directly involved in the design and collection of the data. While registries may be developed to serve more than one purpose, the nurse researcher who is interested in the analysis of registry data must work within the limitations of secondary data sources (which have already been collected) to answer their research questions. For example, depending on the purpose of the 172  registry, the data may have been collected at different geographic regions or participating sites, in different years, or even in a different target population altogether.  In addition, depending on the intended purpose, the quality of the registry data will vary, which has implications for the type of questions that can be asked. For example, a registry intended to serve as primary evidence for decision making (e.g., product safety evaluations or performance-based payment) will require higher levels of quality assurance than a registry that describes the natural course of a condition (Gliklich et al., 2014). This is where the joint, iterative process of the “researcher question-driven approach” and the “data-driven approach” begins because the research question or the data requirements may change depending on the availability and the quality of the data (Cheng & Phillips, 2014). If the registry data do not contain all the necessary variables to address the research question, it may be possible to link to other data sources (e.g., administrative data) to find suitable variables. In some cases, an entirely different registry data may need to be considered. Thus at the outset, nurse researchers will need to have a basic analytic plan (with an understanding that the research question may need to be modified based on the available data) and knowledge of the underlying motivation for the registry development to place the data in a context that will further illuminate other characteristics of the data (e.g., target population) and the choices that will be made in the analyses and, in turn, the interpretation of the results.  2) What is the process by which the registry data have been collected? Most registries are structured to have a central coordinating centre that sets up and maintains the database and several participating sites (e.g., clinics and hospitals) where the data are collected (Arts et al., 2002). The quality of the registry data depends on those sites, which may or may not follow the same data collection procedures (Arts et al., 2002). For example, in a 173  condition-specific registry, some sites may enrol and record only confirmed cases of patients with a specific condition while other sites may enrol and record all patients with particular symptoms or signs. In addition, the collection of certain variables may involve varying degrees of judgement and interpretation. For example, while physical examination findings (e.g., height and weight) or laboratory findings (e.g., white blood cell counts) are relatively straightforward, manual abstraction of unformatted medical records may be more difficult, particularly when there is free text involved (Gliklich et al., 2014). The personnel, who may either be trained or untrained in analyzing medical records, may need to decipher illegible handwriting, translate complex medical abbreviations and acronyms, and understand enough about the clinical context to adequately extract the desired information (Gliklich et al., 2014). If some data were collected through questionnaires (e.g., PROMs), how the instrument is presented (e.g., font size and language) and delivered (e.g., paper-and-pencil, computer, telephone or voice inputs) may also affect the results of the study (Gliklich et al., 2014).  In addition, other factors that may affect registry data are selection bias from voluntary participation of selected clinicians, sites, and patients, as well as information bias from systematic distortions during data collection (e.g., underreporting of adverse events) (Trenner, Eckstein, Kallmayer, Reutersberg, & Kühnl, 2019). To optimize the quality and usability of the data, each registry has a detailed manual that describes the protocols, policies, and procedures, any survey instruments, and a listing of variables and their definitions. The manual may also include mandatory fields that need to be completed on every patient and updates to new procedures that have been introduced. Some registries may have documentation of their data collection procedures in the agency website, in published articles, other third-party entities 174  responsible for making the data available or may even be available from personal communication with relevant individuals.11  In practice, however, the details of the manuals may vary widely, and some data stewards may not grant access to these internal documents (Kodra et al., 2018). Thus, to the extent possible, a guidebook or relevant codebook should be obtained to assess the type of problems encountered in the data collection process and ensure that the meaning of information captured at the different sites are the same. It may also be helpful to obtain feedback from experts in the measurement of health outcomes or a biostatistician familiar with registry-based data, as they may notice potential analysis issues that need to be considered during variable selection. 3) What procedures have been applied to correct for sources of data error? Broadly, there are two ways to correct for data error: automatic and manual data validation. Many registries will have some form of automatic data entry checks via pre-set rules, which can detect discrepancies in the data and when items are left blank (Arts et al., 2002; Hoeijmakers, Beck, Wouters, Prins, & Steup, 2018; Tan, Armstrong, Close, & Harris, 2019). However, not all data errors can be detected automatically. For example, while certain out-of-range lab values can be “flagged,” values that are within the predefined range and erroneous will not be detected. Therefore, the use of automated data validation rules may not be fully adequate to ensure accuracy of the data and some manual data checking and cleaning will be necessary. The coordinating centre of a registry will often visit the participating sites and perform data audits (e.g., checking the entered data with paper-based patient records) (Horton, Krijnen, Molenaar, & Schipper, 2017; Pellen, Guéganton, Pougheon Bertrand, & Rault, 2018). For paper- 11 For more information the National Institutes of Health (2019) provides a listing of registry web sites. 175  based entries, automated data checks may not be available at the time of filling out paper forms but can be incorporated when the data are later entered into the database.  According to Gliklich et al. (2014), procedures to address potential sources of data error include:  1) Training of data collectors/abstractors in a structured manner. Specific training processes and an infrastructure for training of data collectors should be provided on an ongoing basis to account for unanticipated changes and turn over of personnel who regularly enter data into registries. In multi-site registries, a central training session would ensure that all personnel are trained in a standardized manner.  2) Providing participating sites with timely and ongoing feedback or reports on data inconsistencies or comparisons across sites or over time. This may be useful for reviewing procedures and improving data quality.  3) Onsite audits for participating sites or “for-cause” audits (only when there is indication of a problem) to ensure that standard procedures are implemented and followed and to identify data entry errors (e.g., discrepancies between enrolment and screening logs or comparing sample registry data with original source data). The audit plan will vary in scope, frequency, and location, depending on the purpose of a registry and funding constraints.  It is important to note that it is unrealistic to expect registry data to be completely free of errors. Some errors may remain undetected and uncorrected regardless of the various procedures conducted to improve data quality. The frequency and extent of monitoring procedures may also need to achieve a balance between reliable data integrity and the ease of workflow of the sites 176  with regards to enrolment and follow-up. Thus, an important question that the nurse researcher would want to answer is, “How good does the data quality need to be to support the analysis?” This is not an easy question to answer and may, in part, depend on the research question because the intended purpose of the registry determines the necessary properties of its data (Arts et al., 2002). For example, in a registry that is used to determine incidence rates of a particular condition, all existing patient cases need to be included. In contrast, if a registry is used for case-control studies, the diagnoses and characteristics of the registered patients need to be recorded correctly while the exact number of included patients may be of lesser importance (Arts et al., 2002). As such, and within the context of a given research question, pursuing data quality to a greater extent than needed to support the conclusion may not be completely necessary. In other words, the assessment of data quality should correspond with its intended use.  In summary, nurse researchers are advised to assess the level of data quality that is appropriate for their particular research question, which will inevitably increase as stakes become higher from a purely descriptive study aimed at understanding the characteristics of people who develop a condition and how it progresses over time to a highly focused question intended to support clinical decision making. In this way, there is greater alignment between the quality of the data, the research question, and the intended purpose of the registry to generate meaningful information that can truly support patient-centred care.  177  Chapter 6: Conclusion 6.1 Overview We addressed a substantial gap in the literature on how to analyze PROMs data in registries to inform practice. The purpose of this study was to explore the analytical potential of patient-reported data stored in clinical registries, with the aim to explicate the process of conducting such an analysis of data from patients with AF in specialized clinics by answering the following research question:  Can the trajectories of change in patient-reported outcome measures for outpatients with atrial fibrillation, after initial consultation, be explained by different biological functioning or individual and environmental characteristics?  Answering this main research question allowed us to address three key methodological questions: (1) What information can be extracted from linked data sources? (2) How best to accommodate missing data? and (3) What is an appropriate data analysis strategy? In Chapter 1 (Background), we explained how  patient-reported outcomes can be used to support patient-centred care. AF was introduced as an example context as the condition can cause a range of symptoms that can change over time. Although PROMs have received limited attention in the AF, increased interest and opportunity for using routinely collected PROMs data in registries to better understand trajectories of change was emphasized.  In Chapter 2 (Literature Review), we provided an overview of registries and the relevance of routinely collected PROMs data in the context of AF. Despite increasing interest in the use of PROMs in registries, limited studies in the AF literature has been noted because of significant methodological challenges including data quality issues. The revised Wilson and Cleary (1995) framework was also presented that informed the research question.  178  In Chapter 3 (Methods), we outlined the research design and the AF clinic processes to provide more context to the data, and addressed three key methodological question. To extract information from linked data sources, relevant information including patients’ comorbidities and interventions received were extracted with validated algorithms and graphical techniques were employed to visualize relevant information (e.g., stacked bar plots and times series plots). To accommodate missing data, multilevel multiple imputation was used with all available auxiliary variables. Full information maximum likelihood was used to account for unequal number and duration of follow-ups. To model patient trajectories, several methods including multilevel and latent growth models were compared. However, an emerging statistical approach known as growth mixture modelling was selected because it more closely aligned with the patient-centred approach by being able to identify unobserved subgroups of patients with AF and account for individually-varying times of observations. In Chapter 4 (Results), we provided the results of the growth mixture model that  identified three groups of patients with distinct trajectories: poor and improving health, good and stable health, and excellent and stable health. We also found that differences in age, gender, stroke risk score, and ablation and anticoagulation therapy at specified follow-up predicted membership in the lowest health trajectory.  In Chapter 5 (Discussion), we discussed practical and methodological implications as well as further modifications to the conceptual framework. Strengths and limitations of the study and recommendations were also provided. For practical implications, we highlighted the potential opportunity for clinicians to assess for differences in trajectories and tailor interventions based on those differing trajectories. Regarding the methods, we discussed the importance of exploring modelling issues (e.g., selection of an appropriate time metric, specifying growth 179  parameters, and when to include covariates to predict class membership). Regarding the conceptual framework, the different subgroups of trajectories, the varying times of treatment and additional predisposing factors were added to be consistent with our results. For strengths and limitations, one of the key strengths highlighted was bringing more transparency to the methods of analysis; however, data quality issues presented major challenges as quality standards were not consistently reported. Thus, we argued that clinician researchers must share responsibility by checking data accuracy, selecting relevant variables and cohorts of interest to effectively support patient-centred care.  6.2 Future Research Registries operate within a complex system of governance and data sources, which come with both challenges in their use and opportunities for future research. One of the major directions for future research is the inclusion of patients as partners in the governance structure of registries and to collect additional data that supports patient-driven research questions. As explained in Chapter 1, registries do not often provide information about quality of care or reflect important outcomes from patients’ perspectives. Thus, patients can provide their perspectives on what is important for a registry to be able to meet their needs and expectations, how researchers can best engage with patients, how to recognize gaps in existing data and best ways to collect that information, and how the registry could provide additional value to patients beyond data collection. An overarching goal of these patient-centred research designs is to conduct more meaningful patient-centred research by building in education, support, and other community resources for patients. For example, patients may find support and benefit in sharing their experiences with others who are experiencing similar challenges to their health and economic, social, and emotional well-being. Facilitating these relationships through support groups, online 180  forums for engagement (to discuss relevant literature and patients’ experiences), and sharing information about the study through the enrolment, implementation, and dissemination phases may help build a foundation for long-term registry engagement. Another avenue for future research is in developing standardized methodologies and guidelines in assessing and reporting data quality in registries. As explained in Chapter 2, data quality is highly dependent on the individual registry. Although there are many guidance tools to assess the quality of evidence and grading the strength of recommendations for clinical trials (e.g., Grading of Recommendations Assessment Development, and Evaluation (GRADE) system (Atkins et al., 2004)), similar criteria for registry-based studies are still lacking. For example, there are no clear guidelines on the type of criteria that need to be applied to ensure enough confidence in the data quality to inform clinical practice. Thus, research studies on the development and evaluation of standardized tools that assess the quality of evidence for registry-based studies are important next steps to ensure that registry-based studies maintain the rigor needed to provide clinically meaningful results, to aid in clinical and policy decisions, and that align with the best interest of patients.  There are also opportunities for further theoretical development regarding “individual” trajectories. As explained in Chapter 5, we suggested a modification to the revised Wilson and Cleary (1995) framework to reflect multiple subgroups of patients’ trajectories that align with patient-centred approaches to modelling. However, with increasing use of digital technologies such as smartphones and monitoring devices that allow opportunities to collect “intensive longitudinal data” (i.e., many measurement occurrences), there is growing interest in patient-specific approaches to studying effects that may be idiosyncratic to one or more individual patient (Howard & Hoffman, 2018). Using patient-specific approaches to modelling would entail 181  adapting new perspectives about conducting research because inferences would not necessarily be made about populations or subgroups, but about the individual from data that has been collected across multiple time points. Some common patient-specific techniques currently used include state-space modelling, dynamic factor analysis, and p-technique factor analysis (Ram, Brose, & Molenaar, 2013). By using these patient-specific approaches, researchers can better understand the individualized and contextualized nature of a patient’s trajectory, which could in turn serve as a basis for theoretical development or testing of theories. For example, patient-specific analyses may inform future case-study research, which is often used in initial theory development (Crowe et al., 2011), or to understand when certain theories may no longer be accurate or informative. With the ever-increasing use of digital technologies that allow continuous, real-time data collection, it is likely that patient-specific research and the development of theories that build on the individual as the basic unit of analysis will continue to expand.  6.3 Conclusion Patient-reported data in registries have the potential to provide valuable information on treatment practices and patient characteristics in the real-world of daily practice. To this end, we examined trajectories of change in the PROM scores of patients with AF and factors associated with their change. We considered several growth models and ultimately settled on using GMM because it more closely aligned with the patient-centred approach by being able to identify unobserved subgroups of patients with AF and account for individually-varying times of observations. We found that instead of a single health trajectory, there were several different health trajectories, with different patient characteristics associated with these trajectories. Based on these findings, we argue that this model of looking at trajectories shows us that not all 182  trajectories are the same and using this method can help us to tailor interventions for different groups of patients. However, several challenges in analyzing PROMs data stored in registries were encountered, which offered us the opportunity to examine the underlying assumptions of different models (e.g., variable and patient-centred approaches) and explore modelling issues (e.g., the selection of appropriate time metrics, specifying growth parameters, and when to include covariates to predict class membership) that potentially could have led to different conclusions. Nurses, including practising nurses, nurse educators, and nurse scientists, have an opportunity to ensure that registry data are used to inform the tailoring of interventions for different patients. To maximize these opportunities, nurses should be aware of the critical skills and competencies necessary for their meaningful use. These skillsets include understanding informed consent and confidentiality, checking data accuracy, selecting variables and cohorts of interest, and transforming data structures. When using registry data, we recommend that nurse researchers be guided by three key questions: (1) What is the purpose of the registry and the rationale for its development? (2) What is the process and the context in which the registry data have been collected? and (3) What procedures have been applied to correct for sources of data error? The consequences of asking these questions is to bring greater alignment between the quality of the data, the research question, and the intended purpose of the registry to truly support better patient-centred care. 183  References Albrecht, G. L., & Devlieger, P. J. (1999). The disability paradox: High quality of life against all odds. Social Science & Medicine (1982), 48(8), 977–988. Altiok, M., Yilmaz, M., & Rencüsoğullari, I. (2015). Living with atrial fibrillation: An analysis of patients’ perspectives. Asian Nursing Research, 9(4), 305–311. https://doi.org/10.1016/j.anr.2015.10.001 Anker, S. D., Agewall, S., Borggrefe, M., Calvert, M., Jaime Caro, J., Cowie, M. R., … Kirchhof, P. (2014). The importance of patient-reported outcomes: A call for their comprehensive integration in cardiovascular clinical trials. European Heart Journal, 35(30), 2001–2009. https://doi.org/10.1093/eurheartj/ehu205 Armstrong, D., & Caldwell, D. (2004). Origins of the concept of quality of life in health care: A rhetorical solution to a political problem. Social Theory & Health, 2(4), 361–371. https://doi.org/10.1057/palgrave.sth.8700038 Arnason, T., Wells, P. S., van Walraven, C., & Forster, A. J. (2006). Accuracy of coding for possible warfarin complications in hospital discharge abstracts. Thrombosis Research, 118(2), 253–262. https://doi.org/10.1016/j.thromres.2005.06.015 Arts, D. G. T., de Keizer, N. F., & Scheffer, G.-J. (2002). Defining and improving data quality in medical registries: A literature review, case study, and generic framework. Journal of the American Medical Informatics Association, 9(6), 600–611. https://doi.org/10.1197/jamia.M1087 Asparouhov, T., & Muthén, B. (2014a). Auxiliary variables in mixture modeling: Three-step approaches using Mplus. Structural Equation Modeling, 21(3), 329–341. https://doi.org/10.1080/10705511.2014.915181 184  Asparouhov, T., & Muthén, B. (2014b). Auxiliary variables in mixture modeling: Using the BCH method in Mplus to estimate a distal outcome model and an arbitrary secondary model. Retrieved from https://www.statmodel.com/examples/webnotes/ webnote21.pdf Atkins, D., Best, D., Briss, P. A., Eccles, M., Falck-Ytter, Y., Flottorp, S., … GRADE Working Group. (2004). Grading quality of evidence and strength of recommendations. BMJ (Clinical Research Ed.), 328(7454), 1490. https://doi.org/10.1136/bmj.328.7454.1490 Aydin, B., Leite, W. L., & Algina, J. (2014). The consequences of ignoring variability in measurement occasions within data collection waves in latent growth models. Multivariate Behavioral Research, 49(2), 149–160. https://doi.org/10.1080/00273171.2014.887901 Ayilara, O. F., Zhang, L., Sajobi, T. T., Sawatzky, R., Bohm, E., & Lix, L. M. (2019). Impact of missing data on bias and precision when estimating change in patient-reported outcomes from a clinical registry. Health and Quality of Life Outcomes, 17(1), 106. https://doi.org/10.1186/s12955-019-1181-2 Azur, M. J., Stuart, E. A., Frangakis, C., & Leaf, P. J. (2011). Multiple imputation by chained equations: What is it and how does it work? International Journal of Methods in Psychiatric Research, 20(1), 40–49. https://doi.org/10.1002/mpr.329 Bai, Y., Bai, R., Wu, J.-H., Zhang, T., Liu, N., Shi, X.-B., … Ma, C.-S. (2015). Differences in quality of life between atrial fibrillation patients with low stroke risk treated with and without catheter ablation. Journal of the American Heart Association, 4(9), e002130. https://doi.org/10.1161/JAHA.115.002130 185  Bakas, T., McLennon, S. M., Carpenter, J. S., Buelow, J. M., Otte, J. L., Hanna, K. M., … Welch, J. L. (2012). Systematic review of health-related quality of life models. Health and Quality of Life Outcomes, 10(1), 134. https://doi.org/10.1186/1477-7525-10-134 Bakk, Z., Tekle, F. B., & Vermunt, J. K. (2013). Estimating the association between latent class membership and external variables using bias-adjusted three-step approaches. Sociological Methodology, 43(1), 272–311. https://doi.org/10.1177/0081175012470644 Banzi, R., Canham, S., Kuchinke, W., Krleza-Jeric, K., Demotes-Mainard, J., & Ohmann, C. (2019). Evaluation of repositories for sharing individual-participant data from clinical studies. Trials, 20(1), 169. https://doi.org/10.1186/s13063-019-3253-3 Bauer, D. J. (2007). Observations on the use of growth mixture models in psychological research. Multivariate Behavioral Research, 42(4), 757–786. https://doi.org/10.1080/00273170701710338 Bauer, D. J., & Curran, P. J. (2004). The integration of continuous and discrete latent variable models: Potential problems and promising opportunities. Psychological Methods, 9(1), 3–29. https://doi.org/10.1037/1082-989X.9.1.3 Bauman, A. E., Fardy, H. J., & Harris, P. G. (2003). Getting it right: Why bother with patient-centred care? Medical Journal of Australia, 179(5), 253–256. Beleites, C., Salzer, R., & Sergo, V. (2013). Validation of soft classification models using partial class memberships: An extended concept of sensitivity & co. applied to grading of astrocytoma tissues. Chemometrics and Intelligent Laboratory Systems, 122, 12–22. https://doi.org/10.1016/j.chemolab.2012.12.003 186  Bell, M. L., & Fairclough, D. L. (2014). Practical and statistical issues in missing data for longitudinal patient-reported outcomes. Statistical Methods in Medical Research, 23(5), 440–459. https://doi.org/10.1177/0962280213476378 Bell, M. L., Kenward, M. G., Fairclough, D. L., & Horton, N. J. (2013). Differential dropout and bias in randomised controlled trials: When it matters and when it may not. The BMJ, 346, e8668. https://doi.org/10.1136/bmj.e8668 Benzimra, M., Bonnamour, B., Duracinsky, M., Lalanne, C., Aubert, J.-P., Chassany, O., … Mahé, I. (2018). Real-life experience of quality of life, treatment satisfaction, and adherence in patients receiving oral anticoagulants for atrial fibrillation. Patient Preference and Adherence, 12, 79–87. https://doi.org/10.2147/PPA.S131158 Berlin, K. S., Parra, G. R., & Williams, N. A. (2014). An introduction to latent variable mixture modeling (part 2): Longitudinal latent class growth analysis and growth mixture models. Journal of Pediatric Psychology, 39(2), 188–203. https://doi.org/10.1093/jpepsy/jst085 Bhatt, D. L., Drozda, J. P., Shahian, D. M., Chan, P. S., Fonarow, G. C., Heidenreich, P. A., … Welke, K. F. (2015). ACC/AHA/STS Statement on the future of registries and the performance measurement enterprise: A report of the American College of Cardiology/American Heart Association Task Force on Performance Measures and The Society of Thoracic Surgeons. Journal of the American College of Cardiology, 66(20), 2230–2245. https://doi.org/10.1016/j.jacc.2015.07.010 Biering, K., Hjollund, N. H., & Frydenberg, M. (2015). Using multiple imputation to deal with missing data and attrition in longitudinal studies with repeated measures of patient-reported outcomes. Clinical Epidemiology, 7, 91–106. https://doi.org/10.2147/CLEP.S72247 187  Björkenheim, A., Brandes, A., Magnuson, A., Chemnitz, A., Edvardsson, N., & Poçi, D. (2018). Patient-reported outcomes in relation to continuously monitored rhythm before and during 2 years after atrial fibrillation ablation using a disease-specific and a generic instrument. Journal of the American Heart Association, 7(5), e008362. https://doi.org/10.1161/JAHA.117.008362 Black, N. (2013). Patient reported outcome measures could help transform healthcare. BMJ, 346, f167. https://doi.org/10.1136/bmj.f167 Bland, J. M., & Altman, D. G. (1997). Cronbach’s alpha. BMJ, 314(7080), 572. Blonde, L., Khunti, K., Harris, S. B., Meizinger, C., & Skolnik, N. S. (2018). Interpretation and impact of real-world clinical data for the practicing clinician. Advances in Therapy, 35(11), 1763–1774. https://doi.org/10.1007/s12325-018-0805-y Blumenthal, S. (2017). The use of clinical registries in the United States: A landscape survey. EGEMs, 5(1), 26. https://doi.org/10.5334/egems.248 Blumenthal, S. (2019). The NQRN registry maturational framework: Evaluating the capability and use of clinical registries. EGEMs, 7(1), 29. https://doi.org/10.5334/egems.278 Bolck, A., Croon, M., & Hagenaars, J. (2004). Estimating latent structure models with categorical variables: One-step versus three-step estimators. Political Analysis, 12(1), 3–27. https://doi.org/www.jstor.org/stable/25791751 Bollen, K. A., & Curran, P. J. (2006). Latent Curve Models: A Structural Equation Perspective. Hoboken, NJ: Wiley. Bowling, A. (2001). Measuring disease: A review of disease specific quality of life measurement scales (2 edition). Philadelphia: Open University Press. 188  Brant, J. M., Beck, S., & Miaskowski, C. (2010). Building dynamic models and theories to advance the science of symptom management research. Journal of Advanced Nursing, 66(1), 228–240. https://doi.org/10.1111/j.1365-2648.2009.05179.x British Columbia Ministry of Health. (2014). Setting priorities for B.C. health—Province of British Columbia. Retrieved January 30, 2017, from http://www2.gov.bc.ca/gov/content/health/about-bc-s-health-care-system/health-priorities/setting-priorities-for-bc-health British Columbia Ministry of Health. (2018). Diagnostic Code Descriptions (ICD-9)—Province of British Columbia. Retrieved December 22, 2018, from https://www2.gov.bc.ca/gov/content/health/practitioner-professional-resources/msp/physicians/diagnostic-code-descriptions-icd-9 British Columbia Ministry of Health [creator]. (2018a). Consolidation File (MSP Registration & Premium Billing). Population Data BC [publisher]. Data Extract. MOH (2018). http://www.popdata.bc.ca/data British Columbia Ministry of Health [creator]. (2018b). Medical Services Plan (MSP) Payment Information File. Population Data BC [publisher]. Data Extract. MOH (2018). http://www.popdata.bc.ca/data British Columbia Ministry of Health [creator]. (2018c). PharmaNet. BC Ministry of Health [publisher]. Data Extract. Data Stewardship Committee (2018). http://www.popdata.bc.ca/data British Columbia Vital Statistics Agency [creator]. (2018). Vital Statistics Deaths. Population Data BC [publisher]. Data Extract BC Vital Statistics Agency (2018). http://www.popdata.bc.ca/data 189  Bryk, A. S., & Raudenbush, S. W. (1987). Application of hierarchical linear models to assessing change. Psychological Bulletin, 101(1), 147–158. https://doi.org/10.1037/0033-2909.101.1.147 Bubien, R., Kay, G., & Jenkins, L. (1993). Test specifications for symptoms checklist: Frequency and severity. Milwaukee, WI: University of Wisconsin. Buuren, S. van, & Groothuis-Oudshoorn, K. (2011). MICE: Multivariate imputation by chained equations in R. Journal of Statistical Software, 45(1), 1–67. https://doi.org/10.18637/jss.v045.i03 Calkins, H., Gliklich, R. E., Leavy, M. B., Piccini, J. P., Hsu, J. C., Mohanty, S., … Turakhia, M. P. (2019). Harmonized outcome measures for use in atrial fibrillation patient registries and clinical practice: Endorsed by the Heart Rhythm Society Board of Trustees. Heart Rhythm, 16(1), e3–e16. https://doi.org/10.1016/j.hrthm.2018.09.021 Calkins, H., Hindricks, G., Cappato, R., Kim, Y.-H., Saad, E. B., Aguinaga, L., … Document Reviewers: (2018). 2017 HRS/EHRA/ECAS/APHRS/SOLAECE expert consensus statement on catheter and surgical ablation of atrial fibrillation. EP Europace, 20(1), e1–e160. https://doi.org/10.1093/europace/eux274 Camm, A. J., Kirchhof, P., Lip, G. Y. H., Schotten, U., Savelieva, I., Ernst, S., … Rutten, F. H. (2010). Guidelines for the management of atrial fibrillation: The Task Force for the Management of Atrial Fibrillation of the European Society of Cardiology (ESC). European Heart Journal, 31(19), 2369–2429. https://doi.org/10.1093/eurheartj/ehq278 Canadian Cardiovascular Society. (2019). Pocket Guides. Retrieved December 13, 2019, from https://www.ccs.ca/index.php/en/resources/pocket-guides 190  Canadian Institute for Health Information. (2009). CIHI Data Quality Study of the 2007-2008 Discharge Abstract Database. Ottawa. Canadian Institute for Health Information. (2010). CIHI Data Quality Study of the 2008-2009 Discharge Abstract Database. Ottawa. Canadian Institute for Health Information. (2012). CIHI Data Quality Study of the 2009-2010 Discharge Abstract Database. Ottawa. Canadian Institute for Health Information. (2015). Canadian classification of health interventions. Retrieved from https://www.cihi.ca/en/cci_volume_four_2015_en.pdf Canadian Institute for Health Information. (2017). Patient-centred measurement and reporting in Canada: Launching the discussion toward a future state. Retrieved from https://www.cihi.ca/sites/default/files/document/visioning-day-paper-en-web.pdf Canadian Institute for Health Information. (2019). Discharge Abstract Database Metadata (DAD). Retrieved April 24, 2019, from https://www.cihi.ca/en/discharge-abstract-database-metadata Canadian Institute for Health Information [creator]. (2018). Discharge Abstract Database (Hospital Separations). Population Data BC [publisher]. Data Extract. MOH (2018). http://www.popdata.bc.ca/data Canadian Institutes of Health Research, Institute of Gender and Health. (2012). What a difference sex and gender make: A gender, sex and health research casebook. Retrieved from http://www.cihr-irsc.gc.ca/e/44734.html Cardiac Services BC. (2019). Atrial fibrillation clinic. Internal report: unpublished. 191  Cardiac Services BC [creator]. (2018). Cardiac Services Registry (HEARTis). Population Data BC [publisher]. Data Extract. CSBC (2018). http://www.cardiacbc.ca/our-services/evaluation-surveillance Carman, K. L., Dardess, P., Maurer, M., Sofaer, S., Adams, K., Bechtel, C., & Sweeney, J. (2013). Patient and family engagement: A framework for understanding the elements and developing interventions and policies. Health Affairs, 32(2), 223–231. https://doi.org/10.1377/hlthaff.2012.1133 Cartwright, D. J. (2013). ICD-9-CM to ICD-10-CM Codes: What? Why? How? Advances in Wound Care, 2(10), 588–592. https://doi.org/10.1089/wound.2013.0478 Celeux, G., & Soromenho, G. (1996). An entropy criterion for assessing the number of clusters in a mixture model. Journal of Classification, 13(2), 195–212. https://doi.org/10.1007/BF01246098 Cella, D., Yount, S., Rothrock, N., Gershon, R., Cook, K., Reeve, B., … PROMIS Cooperative Group. (2007). The Patient-Reported Outcomes Measurement Information System (PROMIS): Progress of an NIH Roadmap cooperative group during its first two years. Medical Care, 45(5 Suppl 1), S3–S11. https://doi.org/10.1097/01.mlr.0000258615.42478.55 Chamberlain, A., Alonso, A., Gersh, B., Manemann, S., Killian, J., Weston, S., … Roger, V. (2017). Multimorbidity and the risk of hospitalization and death in atrial fibrillation: A population-based study. American Heart Journal, 185, 74–84. https://doi.org/10.1016/j.ahj.2016.11.008 Charitakis, E., Barmano, N., Walfridsson, U., & Walfridsson, H. (2017). Factors predicting arrhythmia-related symptoms and health-related quality of life in patients referred for 192  radiofrequency ablation of atrial fibrillation: An observational study (the SMURF study). JACC. Clinical Electrophysiology, 3(5), 494–502. https://doi.org/10.1016/j.jacep.2016.12.004 Chassany, O., Le-Jeunne, P., Duracinsky, M., Schwalm, M.-S., & Mathieu, M. (2006). Discrepancies between patient-reported outcomes and clinician-reported outcomes in chronic venous disease, irritable bowel syndrome, and peripheral arterial occlusive disease. Value in Health, 9(1), 39–46. https://doi.org/10.1111/j.1524-4733.2006.00079.x Chen, G., Lix, L., Tu, K., Hemmelgarn, B. R., Campbell, N. R. C., McAlister, F. A., & Quan, H. (2016). Influence of using different databases and ‘look back’ intervals to define comorbidity profiles for patients with newly diagnosed hypertension: Implications for health services researchers. PLoS ONE, 11(9). https://doi.org/10.1371/journal.pone.0162074 Chen, J.-Y., Zhang, A.-D., Lu, H.-Y., Guo, J., Wang, F.-F., & Li, Z.-C. (2013). CHADS2 versus CHA2DS2-VASc score in assessing the stroke and thromboembolism risk stratification in patients with atrial fibrillation: A systematic review and meta-analysis. Journal of Geriatric Cardiology, 10(3), 258–266. https://doi.org/10.3969/j.issn.1671-5411.2013.03.004 Chen, Q., Luo, W., Palardy, G. J., Glaman, R., & McEnturff, A. (2017). The efficacy of common fit indices for enumerating classes in growth mixture models when nested data structure is ignored: A Monte Carlo study. SAGE Open, 7(1), 2158244017700459. https://doi.org/10.1177/2158244017700459 193  Cheng, H. G., & Phillips, M. R. (2014). Secondary analysis of existing data: Opportunities and  implementation. Shanghai Archives of Psychiatry, 26(6), 371–375. https://doi.org/10.11919/j.issn.1002-0829.214171 Christian, J., Dasgupta, N., Jordan, M., Juneja, M., Nilsen, W., & Reites, J. (2018). Digital health and patient registries: Today, tomorrow, and the future. In R. E. Gliklich, N. A. Dreyer, & M. B. Leavy (Eds.), Registries for evaluating patient outcomes: A user’s guide, addendum [Internet] (3rd edition). Retrieved from https://www.ncbi.nlm.nih.gov/books/NBK493822/ Chugh, S. S., Roth, G. A., Gillum, R. F., & Mensah, G. A. (2014). Global burden of atrial fibrillation in developed and developing nations. Global Heart, 9(1), 113–119. https://doi.org/10.1016/j.gheart.2014.01.004 Clancy, C., & Collins, F. S. (2010). Patient-Centered Outcomes Research Institute: The intersection of science and health care. Science Translational Medicine, 2(37), 37cm18. https://doi.org/10.1126/scitranslmed.3001235 Clark, S. L., & Muthen, B. (2009). Relating latent class analysis results to variables not included in the analysis. Retrieved from https://www.statmodel.com/download/relatinglca.pdf Clinical Data Interchange Standards Consortium. (2019). CDISC [Text]. Retrieved October 31, 2019, from CDISC website: https://www.cdisc.org/ Collins, L. M., Schafer, J. L., & Kam, C. M. (2001). A comparison of inclusive and restrictive strategies in modern missing data procedures. Psychological Methods, 6(4), 330–351. Coulombe, P., Selig, J. P., & Delaney, H. D. (2016). Ignoring individual differences in times of assessment in growth curve modeling. International Journal of Behavioral Development, 40(1), 76–86. https://doi.org/10.1177/0165025415577684 194  Cox, D. R., Kartsonaki, C., & Keogh, R. H. (2018). Big data: Some statistical issues. Statistics & Probability Letters, 136, 111–115. https://doi.org/10.1016/j.spl.2018.02.015 Coyne, K., Margolis, M. K., Grandy, S., & Zimetbaum, P. (2005). The state of patient-reported outcomes in atrial fibrillation. PharmacoEconomics, 23(7), 687–708. https://doi.org/10.2165/00019053-200523070-00004 Crowe, S., Cresswell, K., Robertson, A., Huby, G., Avery, A., & Sheikh, A. (2011). The case study approach. BMC Medical Research Methodology, 11(1), 100. https://doi.org/10.1186/1471-2288-11-100 Curran, P. J., Obeidat, K., & Losardo, D. (2010). Twelve frequently asked questions about growth curve modeling. Journal of Cognition and Development, 11(2), 121–136. https://doi.org/10.1080/15248371003699969 Curtis, K., Fry, M., Shaban, R. Z., & Considine, J. (2017). Translating research findings to clinical nursing practice. Journal of Clinical Nursing, 26(5–6), 862–872. https://doi.org/10.1111/jocn.13586 Daugherty, S. E., Lee, S. B., Nowell, B., Peay, H., Solomon, D., Valbrun, T. G., … Whicher, D. (2018). The increasing focus on the patient in patient registries. In R. E. Gliklich, N. A. Dreyer, & M. B. Leavy (Eds.), Registries for evaluating patient outcomes: A user’s guide, addendum [Internet] (3rd ed.). Retrieved from https://www.ncbi.nlm.nih.gov/books/NBK493820/ de Groot, S., van der Linden, N., Franken, M. G., Blommestein, H. M., Leeneman, B., van Rooijen, E., … Uyl-de Groot, C. A. (2017). Balancing the optimal and the feasible: A practical guide for setting up patient registries for the collection of real-world data for 195  health care decision making based on dutch experiences. Value in Health, 20(4), 627–636. https://doi.org/10.1016/j.jval.2016.02.007 De Silva, A. P., Moreno-Betancur, M., De Livera, A. M., Lee, K. J., & Simpson, J. A. (2019). Multiple imputation methods for handling missing values in a longitudinal categorical variable with restrictions on transitions over time: A simulation study. BMC Medical Research Methodology, 19, 1–14. https://doi.org/10.1186/s12874-018-0653-0 Deshpande, P. R., Rajan, S., Sudeepthi, B. L., & Abdul Nazir, C. P. (2011). Patient-reported outcomes: A new era in clinical research. Perspectives in Clinical Research, 2(4), 137–144. https://doi.org/10.4103/2229-3485.86879 Dixit, S., Alonso, A., Vittinghoff, E., Soliman, E., Chen, L. Y., & Marcus, G. M. (2017). Past alcohol consumption and incident atrial fibrillation: The Atherosclerosis Risk in Communities (ARIC) Study. PLoS ONE, 12(10), e0185228. https://doi.org/10.1371/journal.pone.0185228 Dong, Y., & Peng, C.-Y. J. (2013). Principled missing data methods for researchers. SpringerPlus, 2(1), 222. https://doi.org/10.1186/2193-1801-2-222 Dorian, P., Burk, C., Mullin, C. M., Bubien, R., Godejohn, D., Reynolds, M. R., … Spertus, J. (2013). Interpreting changes in quality of life in atrial fibrillation: How much change is meaningful? American Heart Journal, 166(2), 381-387.e8. https://doi.org/10.1016/j.ahj.2013.04.015 Dorian, P., Cvitkovic, S. S., Kerr, C. R., Crystal, E., Gillis, A. M., Guerra, P. G., … Wyse, D. G. (2006). A novel, simple scale for assessing the symptom severity of atrial fibrillation at the bedside: The CCS-SAF scale. The Canadian Journal of Cardiology, 22(5), 383–386. 196  Dorian, P., Jung, W., Newman, D., Paquette, M., Wood, K., Ayers, G. M., … Luderitz, B. (2000). The impairment of health-related quality of life in patients with intermittent atrial fibrillation: Implications for the assessment of investigational therapy. Journal of the American College of Cardiology, 36(4), 1303–1309. https://doi.org/10.1016/S0735-1097(00)00886-X Dorian, P., Paquette, M., Newman, D., Green, M., Connolly, S. J., Talajic, M., & Roy, D. (2002). Quality of life improves with treatment in the Canadian Trial of Atrial Fibrillation. American Heart Journal, 143(6), 984–990. https://doi.org/10.1067/mhj.2002.122518 Du, H., Hahn, E. A., & Cella, D. (2011). The impact of missing data on estimation of health-related quality of life outcomes: An analysis of a randomized longitudinal clinical trial. Health Services and Outcomes Research Methodology, 11(3), 134–144. https://doi.org/10.1007/s10742-011-0074-6 Duncan, T. E., & Duncan, S. C. (2009). The abc’s of LGM: An introductory guide to latent variable growth curve modeling. Social and Personality Psychology Compass, 3(6), 979–991. https://doi.org/10.1111/j.1751-9004.2009.00224.x Eggleston, E. P., Laub, J. H., & Sampson, R. J. (2004). Methodological sensitivities to latent class analysis of long-term criminal trajectories. Journal of Quantitative Criminology, 20(1), 1–26. https://doi.org/www.jstor.org/stable/23366817 Emilsson, L., Lindahl, B., Köster, M., Lambe, M., & Ludvigsson, J. F. (2015). Review of 103 Swedish healthcare quality registries. Journal of Internal Medicine, 277(1), 94–136. https://doi.org/10.1111/joim.12303 197  Enders, C. K., Mistler, S. A., & Keller, B. T. (2016). Multilevel multiple imputation: A review and evaluation of joint modeling and chained equations imputation. Psychological Methods, 21(2), 222–240. https://doi.org/10.1037/met0000063 Epstein, R. M., Fiscella, K., Lesser, C. S., & Stange, K. C. (2010). Why the nation needs a policy push on patient-centered health care. Health Affairs, 29(8), 1489–1495. https://doi.org/10.1377/hlthaff.2009.0888 Epstein, R. M., & Street, R. L. (2011). The values and value of patient-centered care. The Annals of Family Medicine, 9(2), 100–103. https://doi.org/10.1370/afm.1239 European Medicines Agency. (2018). Use of patient disease registries for regulatory purposes—Methodological and operational considerations. Retrieved from https://www.ema.europa.eu/documents/other/discussion-paper-use-patient-disease-registries-regulatory-purposes-methodological-operational_en.docx EuroQol Group. (1990). EuroQol—A new facility for the measurement of health-related quality of life. Health Policy, 16(3), 199–208. https://doi.org/10.1016/0168-8510(90)90421-9 Eyler, A. E., Wilcox, S., Matson-Koffman, D., Evenson, K. R., Sanderson, B., Thompson, J., … Rohm-Young, D. (2002). Correlates of physical activity among women from diverse racial/ethnic groups. Journal of Women’s Health & Gender-Based Medicine, 11(3), 239–253. https://doi.org/10.1089/152460902753668448 Fayers, P., & Machin, D. (2007). Missing data. In Quality of life: The assessment, analysis and interpretation of patient-reported outcomes (2nd ed., pp. 355–384). New York, NY: Wiley. 198  Ferrans, C. E., Zerwic, J. J., Wilbur, J. E., & Larson, J. L. (2005). Conceptual model of health-related quality of life. Journal of Nursing Scholarship, 37(4), 336–342. https://doi.org/10.1111/j.1547-5069.2005.00058.x Fichtner, S., Deisenhofer, I., Kindsmüller, S., Dzijan-Horn, M., Tzeis, S., Reents, T., … Ladwig, K.-H. (2012). Prospective assessment of short- and long-term quality of life after ablation for atrial fibrillation. Journal of Cardiovascular Electrophysiology, 23(2), 121–127. https://doi.org/10.1111/j.1540-8167.2011.02165.x Flint, K. M., Schmiege, S. J., Allen, L. A., Fendler, T. J., Rumsfeld, J., & Bekelman, D. (2017). Health status trajectories among outpatients with heart failure. Journal of Pain and Symptom Management, 53(2), 224–231. https://doi.org/10.1016/j.jpainsymman.2016.08.018 Flores, L. T., Bennett, A. V., Law, E. B., Hajj, C., Griffith, M. P., & Goodman, K. A. (2012). Patient-reported outcomes vs. Clinician symptom reporting during chemoradiation for rectal cancer. Gastrointestinal Cancer Research, 5(4), 119–124. Forsberg, H. H., Nelson, E. C., Reid, R., Grossman, D., Mastanduno, M. P., Weiss, L. T., … Weinstein, J. N. (2015). Using patient-reported outcomes in routine practice: Three novel use cases and implications. The Journal of Ambulatory Care Management, 38(2), 188–195. https://doi.org/10.1097/JAC.0000000000000052 Foster, A., Croot, L., Brazier, J., Harris, J., & O’Cathain, A. (2018). The facilitators and barriers to implementing patient reported outcome measures in organisations delivering health related services: A systematic review of reviews. Journal of Patient-Reported Outcomes, 2, 46. https://doi.org/10.1186/s41687-018-0072-3 199  Frankel, G., Kamrul, R., Kosar, L., & Jensen, B. (2013). Rate versus rhythm control in atrial fibrillation. Canadian Family Physician, 59(2), 161–168. Fransson, E. I., Stadin, M., Nordin, M., Malm, D., Knutsson, A., Alfredsson, L., & Westerholm, P. J. M. (2015). The association between job strain and atrial fibrillation: Results from the Swedish WOLF study. BioMed Research International, 2015. https://doi.org/10.1155/2015/371905 Freeman, J. V., Simon, D. N., Go, A. S., Spertus, J., Fonarow, G. C., Gersh, B. J., … Piccini, J. P. (2015). Association between atrial fibrillation symptoms, quality of life, and patient outcomes results from the outcomes registry for better informed treatment of atrial fibrillation (ORBIT-AF). Circulation: Cardiovascular Quality and Outcomes, 8(4), 393–402. https://doi.org/10.1161/CIRCOUTCOMES.114.001303 Fumagalli, S., Chen, J., Dobreanu, D., Madrid, A. H., Tilz, R., & Dagres, N. (2016). The role of the Arrhythmia Team, an integrated, multidisciplinary approach to treatment of patients with cardiac arrhythmias: Results of the European Heart Rhythm Association survey. Europace, 18(4), 623–627. https://doi.org/10.1093/europace/euw090 Gage, B. F., Waterman, A. D., Shannon, W., Boechler, M., Rich, M. W., & Radford, M. J. (2001). Validation of clinical classification schemes for predicting stroke: Results from the national registry of atrial fibrillation. JAMA, 285(22), 2864–2870. Garcia-Ptacek, S., Farahmand, B., Kåreholt, I., Religa, D., Cuadrado, M. L., & Eriksdotter, M. (2014). Mortality risk after dementia diagnosis by dementia type and underlying factors: A cohort of 15,209 patients based on the Swedish dementia registry. Journal of Alzheimer’s Disease, 41(2), 467–477. https://doi.org/10.3233/JAD-131856 200  Gasparini, A., Hojjat, S., & Williman, J. (2019). Comorbidity: Computing comorbidity scores (Version 0.5.2). Retrieved from https://CRAN.R-project.org/package=comorbidity Gerig, G., Fishbaugh, J., & Sadeghi, N. (2016). Longitudinal modeling of appearance and shape and its potential for clinical use. Medical Image Analysis, 33, 114–121. https://doi.org/10.1016/j.media.2016.06.014 Ghisletta, P., Renaud, O., Jacot, N., & Courvoisier, D. (2015). Linear mixed-effects and latent  curve models for longitudinal life course analyses. In C. Burton-Jeangros, S. Cullati, A. Sacker, & D. Blane (Eds.), A Life Course Perspective on Health Trajectories and Transitions. Springer. http://www.ncbi.nlm.nih.gov/books/NBK385371/ Gibbons, R. D., Hedeker, D., & DuToit, S. (2010). Advances in analysis of longitudinal data. Annual Review of Clinical Psychology, 6, 79–107. https://doi.org/10.1146/annurev.clinpsy.032408.153550 Gilthorpe, M. S., Dahly, D. L., Tu, Y.-K., Kubzansky, L. D., & Goodman, E. (2014). Challenges in modelling the random structure correctly in growth mixture models and the impact this has on model mixtures. Journal of Developmental Origins of Health and Disease, 5(3), 197–205. https://doi.org/10.1017/S2040174414000130 Gleason, K. T., Dennison Himmelfarb, C. R., Ford, D. E., Lehmann, H., Samuel, L., Han, H. R., … Nazarian, S. (2019). Association of sex, age and education level with patient reported outcomes in atrial fibrillation. BMC Cardiovascular Disorders, 19, 85. https://doi.org/10.1186/s12872-019-1059-6 Gliklich, R. E., Dreyer, N. A., & Leavy, M. B. (2014). Registries for evaluating patient outcomes: A user’s guide. In Registries for evaluating patient outcomes: A user’s guide (3rd edition). Retrieved from https://www.ncbi.nlm.nih.gov/books/NBK208628/ 201  Government of Canada, I. A. P. on R. E. (2019, April 1). Chapter 5: Privacy and confidentiality. Retrieved September 24, 2019, from http://pre.ethics.gc.ca/eng/tcps2-eptc2_2018_chapter5-chapitre5.html Graham, J. W. (2009). Missing data analysis: Making it work in the real world. Annual Review of Psychology, 60, 549–576. https://doi.org/10.1146/annurev.psych.58.110405.085530 Graham, J. W., Olchowski, A. E., & Gilreath, T. D. (2007). How many imputations are really needed? Some practical clarifications of multiple imputation theory. Prevention Science, 8(3), 206–213. https://doi.org/10.1007/s11121-007-0070-9 Greenberg, J. W., Lancaster, T. S., Schuessler, R. B., & Melby, S. J. (2017). Postoperative atrial fibrillation following cardiac surgery: A persistent complication. European Journal of Cardio-Thoracic Surgery, 52(4), 665–672. https://doi.org/10.1093/ejcts/ezx039 Greenhalgh, J., Dalkin, S., Gibbons, E., Wright, J., Valderas, J. M., Meads, D., & Black, N. (2018). How do aggregated patient-reported outcome measures data stimulate health care improvement? A realist synthesis. Journal of Health Services Research & Policy, 23(1), 57–65. https://doi.org/10.1177/1355819617740925 Greenhalgh, T., Howick, J., & Maskrey, N. (2014). Evidence based medicine: A movement in crisis? BMJ, 348, g3725. https://doi.org/10.1136/bmj.g3725 Grimm, K. J., Ram, N., & Estabrook, R. (2017). Growth modeling: Structural equation and multilevel modeling approaches. New York, NY: Guilford Publications. Groenveld, H. F., Crijns, H. J. G. M., Van den Berg, M. P., Van Sonderen, E., Alings, A. M., Tijssen, J. G. P., … RACE II Investigators. (2011). The effect of rate control on quality of life in patients with permanent atrial fibrillation: Data from the RACE II (Rate Control 202  Efficacy in Permanent Atrial Fibrillation II) study. Journal of the American College of Cardiology, 58(17), 1795–1803. https://doi.org/10.1016/j.jacc.2011.06.055 Grund, S., Lüdtke, O., & Robitzsch, A. (2018). Multiple imputation of missing data for multilevel models: Simulations and recommendations. Organizational Research Methods, 21(1), 111–149. https://doi.org/10.1177/1094428117703686 Guhl, E., Althouse, A., Sharbaugh, M., Pusateri, A. M., Paasche-Orlow, M., & Magnani, J. W. (2019). Association of income and health-related quality of life in atrial fibrillation. Open Heart, 6(1). https://doi.org/10.1136/openhrt-2018-000974 Harron, K., Dibben, C., Boyd, J., Hjern, A., Azimaee, M., Barreto, M. L., & Goldstein, H. (2017). Challenges in administrative data linkage for research. Big Data & Society, 4(2), 1–12. https://doi.org/10.1177/2053951717745678 Hayat, M. J., Howlader, N., Reichman, M. E., & Edwards, B. K. (2007). Cancer statistics, trends, and multiple primary cancer analyses from the surveillance, epidemiology, and end results (SEER) program. The Oncologist, 12(1), 20–37. https://doi.org/10.1634/theoncologist.12-1-20 Healthcare Quality Improvement Partnership. (2018). NHS England quality accounts list. 2018-19. Retrieved from https://www.hqip.org.uk/national-programmes/quality-accounts/#.W_Wv1ehKiUk Heck, R. H., Thomas, S. L., & Tabata, L. N. (2013). Multilevel and longitudinal modeling with IBM SPSS. New York, NY: Routledge. Heeringa, J., van der Kuip, D. A. M., Hofman, A., Kors, J. A., van Herpen, G., Stricker, B. H. C., … Witteman, J. C. M. (2006). Prevalence, incidence and lifetime risk of atrial 203  fibrillation: The Rotterdam study. European Heart Journal, 27(8), 949–953. https://doi.org/10.1093/eurheartj/ehi825 Heneghan, C., Goldacre, B., & Mahtani, K. R. (2017). Why clinical trial outcomes fail to translate into benefits for patients. Trials, 18(1), 122. https://doi.org/10.1186/s13063-017-1870-2 Henly, S. J., Wyman, J. F., & Findorff, M. J. (2011). Health and illness over time: The trajectory perspective in nursing science. Nursing Research, 60(3 Suppl), S5-14. https://doi.org/10.1097/NNR.0b013e318216dfd3 Hickey, A. M., Bury, G., O’Boyle, C. A., Bradley, F., O’Kelly, F. D., & Shannon, W. (1996). A new short form individual quality of life measure (SEIQoL-DW): Application in a cohort of individuals with HIV/AIDS. BMJ, 313(7048), 29–33. https://doi.org/10.1136/bmj.313.7048.29 Hoeijmakers, F., Beck, N., Wouters, M. W. J. M., Prins, H. A., & Steup, W. H. (2018). National quality registries: How to improve the quality of data? Journal of Thoracic Disease, 10, S3490–S3499. Hoque, D. M. E., Kumari, V., Hoque, M., Ruseckaite, R., Romero, L., & Evans, S. M. (2017). Impact of clinical registries on quality of patient care and clinical outcomes: A systematic review. PloS One, 12(9), e0183667. https://doi.org/10.1371/journal.pone.0183667 Horton, E. E., Krijnen, P., Molenaar, H. M., & Schipper, I. B. (2017). Are the registry data reliable? An audit of a regional trauma registry in the Netherlands. International Journal for Quality in Health Care, 29(1), 98–103. https://doi.org/10.1093/intqhc/mzw142 Hosmer, D. W., & Lemeshow, S. (2000). Applied Logistic Regression. New York, NY: Wiley. 204  Howard, M. C., & Hoffman, M. E. (2018). Variable-centered, person-centered, and person-specific approaches: Where theory meets the method. Organizational Research Methods, 21(4), 846–876. https://doi.org/10.1177/1094428117744021 Hox, J., & Stoel, R. D. (2005). Multilevel and SEM approaches to growth curve modeling. In B. S. Everitt & D. C. Howell (Eds.), Encyclopedia of Statistics in Behavioral Science (pp. 1296–1305). Chichester, England: Wiley. Huang, A. R., Redpath, C. J., & van Walraven, C. (2015). The influence of cholinesterase inhibitor therapy for dementia on risk of cardiac pacemaker insertion: A retrospective, population-based, health administrative databases study in Ontario, Canada. BMC Neurology, 15(1), 66. https://doi.org/10.1186/s12883-015-0325-1 Huang, D., Brecht, M.-L., Hara, M., & Hser, Y.-I. (2010). Influences of a covariate on growth mixture modeling. Journal of Drug Issues, 40(1), 173–194. Huang, N., Shih, S.-F., Chang, H.-Y., & Chou, Y.-J. (2007). Record linkage research and informed consent: Who consents? BMC Health Services Research, 7(1), 18. https://doi.org/10.1186/1472-6963-7-18 Huisman, M. V., Lip, G. Y. H., Diener, H. C., Dubner, S. J., Halperin, J. L., Ma, C. S., … Bartels, D. B. (2014). Design and rationale of Global Registry on Long-Term Oral Antithrombotic Treatment in Patients with Atrial Fibrillation: A global registry program on long-term oral antithrombotic treatment in patients with atrial fibrillation. American Heart Journal, 167(3), 329–334. https://doi.org/10.1016/j.ahj.2013.12.006 IBM Corp. (2017). IBM SPSS Statistics for Windows (Version 25.0). Armonk, NY: IBM Corp. Ikemura, N., Kohsaka, S., Kimura, T., Ueda, I., Katsumata, Y., Nishiyama, T., … Takatsuki, S. (2019). Assessment of sex differences in the initial symptom burden, applied treatment 205  strategy, and quality of life in japanese patients with atrial fibrillation. JAMA Network Open, 2(3), e191145. https://doi.org/10.1001/jamanetworkopen.2019.1145 Institute of Medicine. (2001). Envisioning the national health care quality report. Retrieved from https://www.ncbi.nlm.nih.gov/books/NBK223308/ Janssen, S. J., van Rein, E. A. J., Paulino Pereira, N. R., Raskin, K. A., Ferrone, M. L., Hornicek, F. J., … Schwab, J. H. (2016). The discrepancy between patient and clinician reported function in extremity bone metastases. Sarcoma, 2016, 1014248. https://doi.org/10.1155/2016/1014248 January, C. T., Wann, L. S., Alpert, J. S., Calkins, H., Cigarroa, J. E., Cleveland, J. C., … Yancy, C. W. (2014). 2014 AHA/ACC/HRS guideline for the management of patients with atrial fibrillation: A report of the American College of Cardiology/American Heart Association Task Force on practice guidelines and the Heart Rhythm Society. Journal of the American College of Cardiology, 64(21), e1–e76. https://doi.org/10.1016/j.jacc.2014.03.022 January, C. T., Wann, L. S., Calkins, H., Chen, L. Y., Cigarroa, J. E., Cleveland, J. C., … Yancy, C. W. (2019). 2019 AHA/ACC/HRS focused update of the 2014 AHA/ACC/HRS guideline for the management of patients with atrial fibrillation: A report of the American College of Cardiology/American Heart Association Task Force on clinical practice guidelines and the Heart Rhythm Society. Journal of the American College of Cardiology, 140(2), e125–e151. https://doi.org/10.1016/j.jacc.2019.01.011 Jenkins, L. S., Brodsky, M., Schron, E., Chung, M., Rocco, T., Lader, E., … Shemanski, L. (2005). Quality of life in atrial fibrillation: The Atrial Fibrillation Follow-up Investigation 206  of Rhythm Management (AFFIRM) study. American Heart Journal, 149(1), 112–120. https://doi.org/10.1016/j.ahj.2004.03.065 Jetté, N., Quan, H., Hemmelgarn, B., Drosler, S., Maass, C., Oec, D.-G., … Investigators,  the I. (2010). The development, evolution, and modifications of ICD-10: Challenges to the international comparability of morbidity data. Medical Care, 48(12), 1105–1110. https://doi.org/www.jstor.org/stable/25767019 Jolley, R. J., Liang, Z., Peng, M., Pendharkar, S. R., Tsai, W., Chen, G., … Ronksley, P. E. (2018). Identifying cases of sleep disorders through International Classification of Diseases (ICD) codes in administrative data. International Journal of Population Data Science, 3(1), 3–13. https://doi.org/10.23889/ijpds.v3i1.448 Jung, T., & Wickrama, K. a. S. (2008). An introduction to latent class growth analysis and growth mixture modeling. Social and Personality Psychology Compass, 2(1), 302–317. https://doi.org/10.1111/j.1751-9004.2007.00054.x Kakkar, A. K., Mueller, I., Bassand, J.-P., Fitzmaurice, D. A., Goldhaber, S. Z., Goto, S., … Turpie, A. G. G. (2012). International longitudinal registry of patients with atrial fibrillation at risk of stroke: Global Anticoagulant Registry in the FIELD (GARFIELD). American Heart Journal, 163(1), 13-19.e1. https://doi.org/10.1016/j.ahj.2011.09.011 Kang, H. (2013). The prevention and handling of the missing data. Korean Journal of Anesthesiology, 64(5), 402–406. https://doi.org/10.4097/kjae.2013.64.5.402 Kang, Y. (2011). The relationships between uncertainty and its antecedents in Korean patients with atrial fibrillation. Journal of Clinical Nursing, 20(13–14), 1880–1886. https://doi.org/10.1111/j.1365-2702.2010.03416.x 207  Karimi, M., & Brazier, J. (2016). Health, health-related quality of life, and quality of life: What is the difference? PharmacoEconomics, 34(7), 645–649. https://doi.org/10.1007/s40273-016-0389-9 Keller, S., Korkmaz, G., Orr, M., Schroeder, A., & Shipp, S. (2017). The evolution of data quality: Understanding the transdisciplinary origins of data quality concepts and approaches. Annual Review of Statistics and Its Application, 4(1), 85–108. https://doi.org/10.1146/annurev-statistics-060116-054114 Kelley, C. G., Lipson, A. R., Daly, B., & Douglas, S. L. (2013). Using a psychosocial registry as a data source for nurses. Clinical Journal of Oncology Nursing, 17(4), 444–446. https://doi.org/10.1188/13.CJON.444-446 Kingsley, C., & Patel, S. (2017). Patient-reported outcome measures and patient-reported experience measures. BJA Education, 17(4), 137–144. https://doi.org/10.1093/bjaed/mkw060 Kirchhof, P., Benussi, S., Kotecha, D., Ahlsson, A., Atar, D., Casadei, B., … ESC Scientific Document Group. (2016). 2016 ESC Guidelines for the management of atrial fibrillation developed in collaboration with EACTS. European Heart Journal, 37(38), 2893–2962. https://doi.org/10.1093/eurheartj/ehw210 Kleinke, K. (2017). Multiple imputation under violated distributional assumptions: A systematic evaluation of the assumed robustness of predictive mean matching. Journal of Educational and Behavioral Statistics, 42(4), 371–404. https://doi.org/10.3102/1076998616687084 Kodra, Y., Weinbach, J., Posada-de-la-Paz, M., Coi, A., Lemonnier, S. L., van Enckevort, D., … Taruscio, D. (2018). Recommendations for improving the quality of rare disease 208  registries. International Journal of Environmental Research and Public Health, 15(8), e1644. https://doi.org/10.3390/ijerph15081644 Kotecha, D., Ahmed, A., Calvert, M., Lencioni, M., Terwee, C. B., & Lane, D. A. (2016). Patient-reported outcomes for quality of life assessment in atrial fibrillation: A systematic review of measurement properties. PloS One, 11(11), e0165790. https://doi.org/10.1371/journal.pone.0165790 Krahn, A. D., Wagar, B., Jackson, H., Tang, A., & McFarlane, A. (2014). Pacemaker implant rates: A national perspective. Canadian Journal of Cardiology, 30(10), S294–S295. https://doi.org/10.1016/j.cjca.2014.07.526 Krawczyk, M., Sawatzky, R., Schick-Makaroff, K., Stajduhar, K., Öhlen, J., Reimer-Kirkham, S., … Cohen, R. (2018). Micro-meso-macro practice tensions in using patient-reported outcome and experience measures in hospital palliative care. Qualitative Health Research, 29, 510–521. https://doi.org/10.1177/1049732318761366 Lakkireddy, D., Atkins, D., Pillarisetti, J., Ryschon, K., Bommana, S., Drisko, J., … Dawn, B. (2013). Effect of yoga on arrhythmia burden, anxiety, depression, and quality of life in paroxysmal atrial fibrillation: The YOGA My Heart Study. Journal of the American College of Cardiology, 61(11), 1177–1182. https://doi.org/10.1016/j.jacc.2012.11.060 Lam, J., Gomes, T., Juurlink, D. N., Mamdani, M. M., Pullenayegum, E. M., Kearon, C., … Holbrook, A. M. (2013). Hospitalization for hemorrhage among warfarin recipients prescribed amiodarone. The American Journal of Cardiology, 112(3), 420–423. https://doi.org/10.1016/j.amjcard.2013.03.051 LaMori, J. C., Mody, S. H., Gross, H. J., daCosta DiBonaventura, M., Patel, A. A., Schein, J. R., & Nelson, W. W. (2013). Burden of comorbidities among patients with atrial fibrillation. 209  Therapeutic Advances in Cardiovascular Disease, 7(2), 53–62. https://doi.org/10.1177/1753944712464101 Largent, E. A., Weissman, J. S., Gupta, A., Abraham, M., Rozenblum, R., Lynch, H. F., & Cohen, I. G. (2018). Patient-centered outcomes research: Stakeholder perspectives and ethical and regulatory oversight issues. IRB, 40(1), 7–17. Lauck, S., Chakrabarti, S., Kutcher, S., Barr, S., Basi, S., Furlan, B., … Kerr, C. R. (2011). Improving access and optimizing care: Development of an atrial fibrillation clinic to implement Canadian cardiovascular society guidelines. Canadian Journal of Cardiology, 27(5), S199–S200. https://doi.org/10.1016/j.cjca.2011.07.322 Laursen, B. P., & Hoff, E. (2006). Person-centered and variable-centered approaches to longitudinal data. Merrill-Palmer Quarterly, 52(3), 377–389. https://doi.org/10.1353/mpq.2006.0029 Lee, D. S., Stitt, A., Wang, X., Yu, J. S., Gurevich, Y., Kingsbury, K. J., … Tu, J. V. (2013). Administrative hospitalization database validation of cardiac procedure codes. Medical Care, 51(4), e22-26. https://doi.org/10.1097/MLR.0b013e3182329778 Lee, K. J., & Carlin, J. B. (2010). Multiple imputation for missing data: Fully conditional specification versus multivariate normal imputation. American Journal of Epidemiology, 171(5), 624–632. https://doi.org/10.1093/aje/kwp425 Levay, C. (2016). Policies to foster quality improvement registries: Lessons from the Swedish case. Journal of Internal Medicine, 279(2), 160–172. https://doi.org/10.1111/joim.12438 Lewis, W. R., Piccini, J. P., Turakhia, M. P., Curtis, A. B., Fang, M., Suter, R. E., … Fonarow, G. C. (2014). Get With The Guidelines AFIB: Novel quality improvement registry for 210  hospitalized patients with atrial fibrillation. Circulation: Cardiovascular Quality and Outcomes, 7(5), 770–777. https://doi.org/10.1161/CIRCOUTCOMES.114.001263 Li, L., & Hser, Y.-I. (2011). On inclusion of covariates for class enumeration of growth mixture models. Multivariate Behavioral Research, 46(2), 266–302. https://doi.org/10.1080/00273171.2011.556549 Li, X., Cui, S., Xuan, D., Xuan, C., & Xu, D. (2018). Atrial fibrillation in athletes and general population: A systematic review and meta-analysis. Medicine, 97(49), e13405. https://doi.org/10.1097/MD.0000000000013405 Li, Y., Ding, W., Wang, H., Song, N., Lin, L., Wang, Z., … Zhang, W. (2013). Relationship of CHA2DS2-VASc and CHADS2 score to left atrial remodeling detected by velocity vector imaging in patients with atrial fibrillation. PloS One, 8(10), e77653. https://doi.org/10.1371/journal.pone.0077653 Liang, J. J., & Santangeli, P. (2016). Contact force sensing during atrial fibrillation ablation: Clinical experience and effects on outcomes. Expert Review of Cardiovascular Therapy, 14(6), 749–759. https://doi.org/10.1586/14779072.2016.1168695 Lincoln, Y. S., Lynham, S., & Guba, E. G. (2011). Paradigmatic controversies, contradictions, and emerging confluences, revisited. In N. K. Denzin & Y. S. Lincoln (Eds.), The Sage Handbook of Qualitative Research (4th ed., pp. 97–128). Thousand Oaks, CA: Sage. Lip, G. Y. H., Al-Khatib, S. M., Cosio, F. G., Banerjee, A., Savelieva, I., Ruskin, J., … Camm, A. J. (2014). Contemporary management of atrial fibrillation: What can clinical registries tell us about stroke prevention and current therapeutic approaches? Journal of the American Heart Association, 3(4), e001179. https://doi.org/10.1161/JAHA.114.001179 211  Lip, G. Y. H., Laroche, C., Dan, G.-A., Santini, M., Kalarus, Z., Rasmussen, L. H., … Maggioni, A. P. (2014). A prospective survey in European Society of Cardiology member countries of atrial fibrillation management: Baseline results of EURObservational Research Programme Atrial Fibrillation (EORP-AF) Pilot General Registry. Europace, 16(3), 308–319. https://doi.org/10.1093/europace/eut373 Little, T. D., Jorgensen, T. D., Lang, K. M., & Moore, E. W. G. (2014). On the joys of missing data. Journal of Pediatric Psychology, 39(2), 151–162. https://doi.org/10.1093/jpepsy/jst048 Lix, L. M., Wu, X., Hopman, W., Mayo, N., Sajobi, T. T., Liu, J., … Sawatzky, R. (2016). Differential item functioning in the SF-36 physical functioning and mental health sub-scales: A population-based investigation in the Canadian Multicentre Osteoporosis study. PloS One, 11(3), e0151519. https://doi.org/10.1371/journal.pone.0151519 Lo, Y., Mendell, N. R., & Rubin, D. B. (2001). Testing the number of components in a normal mixture. Biometrika, 88(3), 767–778. https://doi.org/10.1093/biomet/88.3.767 Lubke, G. H., & Luningham, J. (2017). Fitting latent variable mixture models. Behaviour Research and Therapy, 98, 91–102. https://doi.org/10.1016/j.brat.2017.04.003 Lubke, G., & Muthén, B. O. (2007). Performance of factor mixture models as a function of model size, covariate effects, and class-specific parameters. Structural Equation Modeling, 14(1), 26–47. https://doi.org/10.1080/10705510709336735 Lubke, G., & Neale, M. C. (2006). Distinguishing between latent classes and continuous factors: Resolution by maximum likelihood? Multivariate Behavioral Research, 41(4), 499–532. https://doi.org/10.1207/s15327906mbr4104_4 212  Ludvigsson, J. F., Andersson, E., Ekbom, A., Feychting, M., Kim, J.-L., Reuterwall, C., … Olausson, P. O. (2011). External review and validation of the Swedish national inpatient register. BMC Public Health, 11, 450. https://doi.org/10.1186/1471-2458-11-450 Lyu, H., Cooper, M., Patel, K., Daniel, M., & Makary, M. A. (2016). Prevalence and data transparency of national clinical registries in the United States. Journal for Healthcare Quality, 38(4), 223–234. https://doi.org/10.1097/JHQ.0000000000000001 Mack, C., Su, Z., & Westreich, D. (2018). Managing missing data in patient registries: Addendum [Internet]. In AHRQ registries for evaluating patient outcomes: A user’s guide (4th ed.). Retrieved from https://www.ncbi.nlm.nih.gov/books/NBK493608/ Majid, S., Foo, S., Luyt, B., Zhang, X., Theng, Y.-L., Chang, Y.-K., & Mokhtar, I. A. (2011). Adopting evidence-based practice in clinical decision making: Nurses’ perceptions, knowledge, and barriers. Journal of the Medical Library Association, 99(3), 229–236. https://doi.org/10.3163/1536-5050.99.3.010 Mann, C. J. (2003). Observational research methods. Research design II: Cohort, cross sectional, and case-control studies. Emergency Medicine Journal, 20(1), 54–60. https://doi.org/10.1136/emj.20.1.54 Mark, D. B., Anstrom, K. J., Sheng, S., Piccini, J. P., Baloch, K. N., Monahan, K. H., … CABANA Investigators. (2019). Effect of catheter ablation vs medical therapy on quality of life among patients with atrial fibrillation: The CABANA randomized clinical trial. JAMA, 321(13), 1275–1285. https://doi.org/10.1001/jama.2019.0692 Maryniak, A., Walczak, F., Bodalski, R., Szumowski, L., Derejko, P., Urbanek, P., … Szufladowicz, E. (2006). Atrial fibrillation onset circumstances and their relation to patients’ quality of life. Kardiologia Polska, 64(10), 1102–1108; discussion 1109. 213  Mazurek, M., Huisman, M. V., & Lip, G. Y. H. (2017). Registries in atrial fibrillation: From trials to real-life clinical practice. The American Journal of Medicine, 130(2), 135–145. https://doi.org/10.1016/j.amjmed.2016.09.012 McArdle, J. J. (1988). Dynamic but structural equation modeling of repeated measures data. In J. R. Nesselroade & R. B. Cattell (Eds.), Handbook of Multivariate Experimental Psychology (pp. 561–614). https://doi.org/10.1007/978-1-4613-0893-5_17 McCabe, P. J. (2011). What patients want and need to know about atrial fibrillation. Journal of Multidisciplinary Healthcare, 4, 413–419. https://doi.org/10.2147/JMDH.S19315 McCabe, P. J., Barnason, S. A., & Houfek, J. (2011). Illness beliefs in patients with recurrent symptomatic atrial fibrillation. Pacing and Clinical Electrophysiology: PACE, 34(7), 810–820. https://doi.org/10.1111/j.1540-8159.2011.03105.x McCabe, P. J., Schumacher, K., & Barnason, S. A. (2011). Living with atrial fibrillation: A qualitative study. The Journal of Cardiovascular Nursing, 26(4), 336–344. https://doi.org/10.1097/JCN.0b013e31820019b9 McDonald, L., Malcolm, B., Ramagopalan, S., & Syrad, H. (2019). Real-world data and the patient perspective: The PROmise of social media? BMC Medicine, 17(1), 11. https://doi.org/10.1186/s12916-018-1247-8 McGregor, S. L. T., & Murnane, J. A. (2010). Paradigm, methodology and method: Intellectual integrity in consumer scholarship. International Journal of Consumer Studies, 34(4), 419–427. https://doi.org/10.1111/j.1470-6431.2010.00883.x McLachlan, G., & Peel, D. (2000). Finite mixture models. New York: Wiley. McLeroy, K. R., Bibeau, D., Steckler, A., & Glanz, K. (1988). An ecological perspective on health promotion programs. Health Education Quarterly, 15(4), 351–377. 214  McNeish, D., & Matta, T. (2018). Differentiating between mixed-effects and latent-curve approaches to growth modeling. Behavior Research Methods, 50(4), 1398–1414. https://doi.org/10.3758/s13428-017-0976-5 Meredith, W., & Tisak, J. (1990). Latent curve analysis. Psychometrika, 55(1), 107–122. https://doi.org/10.1007/BF02294746 Messick, S. (1989). Validity. In S. A. (Ed.), Educational measurement (3rd ed., pp. 13–103). Washington (DC): American Council on Education. Messick, S. (1995). Validity of psychological assessment: Validation of inferences from persons’ responses and performances as scientific inquiry into score meaning. American Psychologist, 50(9), 741–749. https://doi.org/10.1037/0003-066X.50.9.741 Meyer, J. P., & Morin, A. J. S. (2016). A person‐centered approach to commitment research: Theory, research, and methodology. Journal of Organizational Behavior, 4(37), 584–612. https://doi.org/10.1002/job.2085 Miller, D., Steele Gray, C., Kuluski, K., & Cott, C. (2015). Patient-centered care and patient-reported measures: Let’s look before we leap. The Patient, 8(4), 293–299. https://doi.org/10.1007/s40271-014-0095-7 Mistler, S. A., & Enders, C. K. (2017). A comparison of joint model and fully conditional specification imputation for multilevel missing data. Journal of Educational and Behavioral Statistics, 42(4), 432–466. https://doi.org/10.3102/1076998617690869 Mont, L., Bisbal, F., Hernández-Madrid, A., Pérez-Castellano, N., Viñolas, X., Arenal, A., … SARA investigators. (2014). Catheter ablation vs. antiarrhythmic drug treatment of persistent atrial fibrillation: A multicentre, randomized, controlled trial (SARA study). European Heart Journal, 35(8), 501–507. https://doi.org/10.1093/eurheartj/eht457 215  Muthén, B. (2003). Statistical and substantive checking in growth mixture modeling: Comment on Bauer and Curran (2003). Psychological Methods, 8(3), 369–377. https://doi.org/10.1037/1082-989X.8.3.369 Muthen, B. (2007). Re: Missing values, MI, and growth mixture [Mplus discussion]. Retrieved from http://www.statmodel.com/discussion/messages/22/2082.html?1492907612 Muthén, B. O., & Curran, P. J. (1997). General longitudinal modeling of individual differences in experimental designs: A latent variable framework for analysis and power estimation. Psychological Methods, 2(4), 371–402. https://doi.org/10.1037/1082-989X.2.4.371 Muthén, B., & Shedden, K. (1999). Finite mixture modeling with mixture outcomes using the EM algorithm. Biometrics, 55(2), 463–469. https://doi.org/10.1111/j.0006-341x.1999.00463.x Muthén, L. K., & Muthén, B. O. (2002). How to use a Monte Carlo study to decide on sample size and determine power. Structural Equation Modeling, 9(4), 599–620. https://doi.org/10.1207/S15328007SEM0904_8 Muthén, L. K., & Muthén, B. O. (2017). Mplus user’s guide. Eighth edition. Los Angeles, CA: Muthén & Muthén. Nahm, M. (2012). Data quality in clinical research. In R. L. Richesson & J. E. Andrews (Eds.), Clinical Research Informatics (pp. 175–201). London: Springer. Nason, E., & Husereau, D. (2014). Roundtable on real world evidence: System readiness—Are we ready to use routinely collected data to improve health system performance? Retrieved from https://www.ihe.ca/events/past/conferences/real-world-evidence/about-rwe 216  National Heart, Lung, and Blood Institute. (2014). What is atrial fibrillation? Retrieved September 19, 2016, from https://www.nhlbi.nih.gov/health/health-topics/topics/af National Institutes of Health. (2019). List of Registries. Retrieved December 13, 2019, from National Institutes of Health (NIH) website: https://www.nih.gov/health-information/nih-clinical-research-trials-you/list-registries Nattel, S., Guasch, E., Savelieva, I., Cosio, F. G., Valverde, I., Halperin, J. L., … Camm, A. J. (2014). Early management of atrial fibrillation to prevent cardiovascular complications. European Heart Journal, 35(22), 1448–1456. https://doi.org/10.1093/eurheartj/ehu028 Naudet, F., Sakarovitch, C., Janiaud, P., Cristea, I., Fanelli, D., Moher, D., & Ioannidis, J. P. A. (2018). Data sharing and reanalysis of randomized controlled trials in leading biomedical journals with a full data sharing policy: Survey of studies published in The BMJ and PLOS Medicine. BMJ (Clinical Research Ed.), 360, k400. https://doi.org/10.1136/bmj.k400 Nelson, E. C., Dixon-Woods, M., Batalden, P. B., Homa, K., Citters, A. D. V., Morgan, T. S., … Lindblad, S. (2016). Patient focused registries can improve health, care, and science. BMJ, 354, i3319. https://doi.org/10.1136/bmj.i3319 Norekvål, T. M., Fålun, N., & Fridlund, B. (2016). Patient-reported outcomes on the agenda in cardiovascular clinical practice. European Journal of Cardiovascular Nursing, 15(2), 108–111. https://doi.org/10.1177/1474515115614133 Nylund-Gibson, K., & Masyn, K. E. (2016). Covariates and mixture modeling: Results of a simulation study exploring the impact of misspecified effects on class enumeration. Structural Equation Modeling, 23(6), 782–797. https://doi.org/10.1080/10705511.2016.1221313 217  Ocloo, J., & Matthews, R. (2016). From tokenism to empowerment: Progressing patient and public involvement in healthcare improvement. BMJ Quality & Safety, 25(8), 626–632. https://doi.org/10.1136/bmjqs-2015-004839 Ohmann, C., Banzi, R., Canham, S., Battaglia, S., Matei, M., Ariyo, C., … Demotes-Mainard, J. (2017). Sharing and reuse of individual participant data from clinical trials: Principles and recommendations. BMJ Open, 7(12). https://doi.org/10.1136/bmjopen-2017-018647 Ojelabi, A. O., Graham, Y., Haighton, C., & Ling, J. (2017). A systematic review of the application of Wilson and Cleary health-related quality of life model in chronic diseases. Health and Quality of Life Outcomes, 15(1), 241. https://doi.org/10.1186/s12955-017-0818-2 Olmo, C. A., McGettigan, P., & Kurz, X. (2019). Barriers and opportunities for use of patient registries in medicines regulation. Clinical Pharmacology & Therapeutics, 106(1), 39–42. https://doi.org/10.1002/cpt.1414 Owen, J., Adelson, J., Budge, S., Wampold, B., Kopta, M., Minami, T., & Miller, S. (2015). Trajectories of change in psychotherapy. Journal of Clinical Psychology, 71(9), 817–827. https://doi.org/10.1002/jclp.22191 Packer, D. L., Mark, D. B., Robb, R. A., Monahan, K. H., Bahnson, T. D., Moretz, K., … CABANA Investigators. (2018). Catheter ablation versus antiarrhythmic drug therapy for atrial fibrillation (CABANA) trial: Study rationale and design. American Heart Journal, 199, 192–199. https://doi.org/10.1016/j.ahj.2018.02.015 Palardy, G. J., & Vermunt, J. K. (2010). Multilevel growth mixture models for classifying groups. Journal of Educational and Behavioral Statistics, 35(5), 532–565. https://doi.org/10.3102/1076998610376895 218  Pampaka, M., Hutcheson, G., & Williams, J. (2016). Handling missing data: Analysis of a challenging data set using multiple imputation. International Journal of Research & Method in Education, 39(1), 19–37. https://doi.org/10.1080/1743727X.2014.979146 Patient-Centered Outcomes Research Institute. (2019). PCORI. Retrieved October 24, 2019, from https://www.pcori.org/ Paul, J., Seeman, N., Gagliardi, A., Mahindra, S., Blackstien-Hirsch, P., & Brown, A. (2006). New measures of ambulatory care performance in Ontario: Preliminary system snapshot, 2006. Retrieved from hsprn.ca/uploads/files/Preliminary%20system%20snapshot.pdf Pedersen, A. B., Mikkelsen, E. M., Cronin-Fenton, D., Kristensen, N. R., Pham, T. M., Pedersen, L., & Petersen, I. (2017). Missing data and multiple imputation in clinical epidemiological research. Clinical Epidemiology, 9, 157–166. https://doi.org/10.2147/CLEP.S129785 Pellen, N., Guéganton, L., Pougheon Bertrand, D., & Rault, G. (2018). Lessons from the on-site quality audit of data transmitted to the French cystic fibrosis registry. Orphanet Journal of Rare Diseases, 13(1), 9. https://doi.org/10.1186/s13023-017-0750-x Pencarrick Hertzman, C., Meagher, N., & McGrail, K. M. (2013). Privacy by design at Population Data BC: A case study describing the technical, administrative, and physical controls for privacy-sensitive secondary use of personal information for research in the public interest. Journal of the American Medical Informatics Association, 20(1), 25–28. https://doi.org/10.1136/amiajnl-2012-001011 Pennacchini, M., Bertolaso, M., Elvira, M. M., & De Marinis, M. G. (2011). A brief history of the quality of life: Its use in medicine and in philosophy. La Clinica Terapeutica, 162(3), e99–e103. 219  Piccini, J. P., Fraulo, E. S., Ansell, J. E., Fonarow, G. C., Gersh, B. J., Go, A. S., … Peterson, E. D. (2011). Outcomes registry for better informed treatment of atrial fibrillation: Rationale and design of ORBIT-AF. American Heart Journal, 162(4), 606-612.e1. https://doi.org/10.1016/j.ahj.2011.07.001 Piquero, A. R. (2008). Taking stock of developmental trajectories of criminal activity over the life course. In A. M. Liberman (Ed.), The long view of crime: A synthesis of longitudinal research (pp. 23–78). New York, NY: Springer. Polikandrioti, M., Koutelekos, I., Vasilopoulos, G., Gerogianni, G., Gourni, M., Zyga, S., & Panoutsopoulos, G. (2018). Anxiety and depression in patients with permanent atrial fibrillation: Prevalence and associated factors. Cardiology Research and Practice, 2018. https://doi.org/10.1155/2018/7408129 Population Data BC. (2018a, November). Consolidation file. Retrieved April 24, 2019, from https://www.popdata.bc.ca/data/population/consolidationfile Population Data BC. (2018b, November). Data linkage. Retrieved April 24, 2019, from https://www.popdata.bc.ca/datalinkage Population Data BC. (2018c, November). Medical Services Plan (MSP) Payment Information. Retrieved April 24, 2019, from https://www.popdata.bc.ca/data/health/msp Population Data BC. (2018d, December 4). Application processing time: Recent examples. Retrieved October 11, 2019, from https://www.popdata.bc.ca/dataaccess/app_processing/graphs Post, M. W. M. (2014). Definitions of quality of life: What has happened and how to move on. Topics in Spinal Cord Injury Rehabilitation, 20(3), 167–180. https://doi.org/10.1310/sci2003-167 220  Quan, H., Li, B., Couris, C. M., Fushimi, K., Graham, P., Hider, P., … Sundararajan, V. (2011). Updating and validating the Charlson comorbidity index and score for risk adjustment in hospital discharge abstracts using data from 6 countries. American Journal of Epidemiology, 173(6), 676–682. https://doi.org/10.1093/aje/kwq433 Quan, H., Sundararajan, V., Halfon, P., Fong, A., Burnand, B., Luthi, J.-C., … Ghali, W. A. (2005). Coding algorithms for defining comorbidities in ICD-9-CM and ICD-10 administrative data. Medical Care, 43(11), 1130–1139. R Core Team. (2019). R: A language and environment for statistical computing (Version 3.6.1). Retrieved from http://www.R-project.org Raghunathan, T. E., Lepkowski, J. M., Ryn, J., & Solenberger, P. (2001). A multivariate technique for multiply imputing missing values using a sequence of regression models. Survey Methodology, 27(1), 85–95. Raine, D., Langley, P., Shepherd, E., Lord, S., Murray, S., Murray, A., & Bourke, J. P. (2015). Effect of catheter ablation on quality of life in patients with atrial fibrillation and its correlation with arrhythmia outcome. Open Heart, 2(1), e000302. https://doi.org/10.1136/openhrt-2015-000302 Ram, N., Brose, A., & Molenaar, P. C. M. (2013). Dynamic factor analysis: Modeling person-specific process. In Oxford Library of Psychology. The Oxford handbook of quantitative methods: Statistical analysis, Vol. 2 (pp. 441–457). New York, NY, US: Oxford University Press. Ram, N., & Grimm, K. J. (2009). Growth mixture modeling: A method for identifying differences in longitudinal change among unobserved groups. International Journal of Behavioral Development, 33(6), 565–576. https://doi.org/10.1177/0165025409343765 221  Rassen, J. A., Bartels, D. B., Schneeweiss, S., Patrick, A. R., & Murk, W. (2018). Measuring prevalence and incidence of chronic conditions in claims and electronic health record databases. Clinical Epidemiology, 11, 1–15. https://doi.org/10.2147/CLEP.S181242 Raudenbush, S. W., & Bryk, A. S. (2002). Hierarchical linear models: Applications and data analysis methods (2nd ed.). Thousand Oaks: Sage. Raudenbush, S. W., Bryk, A. S., Cheong, Y. F., & Congdon, R. (2019). HLM 8 for Windows. Skokie, IL: Scientific Software International, Inc. Richesson, R., & Andrews, J. (2012). Clinical research informatics. Cham, Switzerland: Springer. Rienstra, M., Lubitz, S. A., Mahida, S., Magnani, J. W., Fontes, J. D., Sinner, M. F., … Benjamin, E. J. (2012). Symptoms and functional status of patients with atrial fibrillation: State-of-the-art and future research opportunities. Circulation, 125(23), 2933–2943. https://doi.org/10.1161/CIRCULATIONAHA.111.069450 Rivera, S. C., Kyte, D. G., Aiyegbusi, O. L., Slade, A. L., McMullan, C., & Calvert, M. J. (2019). The impact of patient-reported outcome (PRO) data from clinical trials: A systematic review and critical analysis. Health and Quality of Life Outcomes, 17(1), 156. https://doi.org/10.1186/s12955-019-1220-z Robitzsch, A., Grund, S., & Henke, T. (2017). Package “miceadds.” Retrieved from https://cran.r-project.org/web/packages/miceadds/miceadds.pdf Rolfson, O., Bohm, E., Franklin, P., Lyman, S., Denissen, G., Dawson, J., … Lübbeke, A. (2016). Patient-reported outcome measures in arthroplasty registries. Acta Orthopaedica, 87(Suppl 1), 9–23. https://doi.org/10.1080/17453674.2016.1181816 222  Romana, H.-W. (2006). Is evidence-based medicine patient-centered and is patient-centered care evidence-based? Health Services Research, 41(1), 1–8. https://doi.org/10.1111/j.1475-6773.2006.00504.x Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581–592. https://doi.org/10.2307/2335739 Rubin, D. B. (1987). Multiple imputation for nonresponse in surveys. New York, NY: Wiley. Rumsfeld, J. S., Alexander, K. P., Goff, D. C., Graham, M. M., Ho, P. M., Masoudi, F. A., … Zerwic, J. J. (2013). Cardiovascular health: The importance of measuring patient-reported health status: A scientific statement from the American Heart Association. Circulation, 127(22), 2233–2249. https://doi.org/10.1161/CIR.0b013e3182949a2e Sackett, D. L., Rosenberg, W. M., Gray, J. A., Haynes, R. B., & Richardson, W. S. (1996). Evidence based medicine: What it is and what it isn’t. BMJ (Clinical Research Ed.), 312(7023), 71–72. Sacristán, J. A. (2013). Patient-centered medicine and patient-oriented research: Improving health outcomes for individual patients. BMC Medical Informatics and Decision Making, 13, 6. https://doi.org/10.1186/1472-6947-13-6 SAS Institute Inc. (2014). SAS/ACCESS® 9.4 Interface to ADABAS (Version 9.4). Cary, NC: SAS Institute. Sawatzky, R., Chan, E. K. H., Zumbo, B. D., Ahmed, S., Bartlett, S. J., Bingham, C. O., … Lix, L. M. (2017). Montreal Accord on Patient-Reported Outcomes (PROs) use series–Paper 7: Modern perspectives of measurement validation emphasize justification of inferences based on patient reported outcome scores. Journal of Clinical Epidemiology, 89, 154–159. https://doi.org/10.1016/j.jclinepi.2016.12.002 223  Sayers, A., Ben-Shlomo, Y., Blom, A. W., & Steele, F. (2016). Probabilistic record linkage. International Journal of Epidemiology, 45(3), 954–964. https://doi.org/10.1093/ije/dyv322 Schafer, J. L. (1997). Analysis of incomplete multivariate data. London: Chapman & Hall. Schafer, J. L., & Graham, J. W. (2002). Missing data: Our view of the state of the art. Psychological Methods, 7(2), 147–177. Schmidt, M., Schmidt, S. A. J., Sandegaard, J. L., Ehrenstein, V., Pedersen, L., & Sørensen, H. T. (2015). The Danish National Patient Registry: A review of content, data quality, and research potential. Clinical Epidemiology, 7, 449–490. https://doi.org/10.2147/CLEP.S91125 Schneider, S., Junghaenel, D. U., Keefe, F. J., Schwartz, J. E., Stone, A. A., & Broderick, J. E. (2012). Individual differences in the day-to-day variability of pain, fatigue, and well-being in patients with rheumatic disease: Associations with psychological variables. Pain, 153(4), 813–822. https://doi.org/10.1016/j.pain.2012.01.001 Schober, P., & Vetter, T. R. (2018). Repeated measures designs and analysis of longitudinal data: If at first you do not succeed—try, try again. Anesthesia and Analgesia, 127(2), 569–575. https://doi.org/10.1213/ANE.0000000000003511 Segal, C., Holve, E., & Sabharwal, R. (2013). Collecting and using patient-reported outcomes (PRO) for comparative effectiveness research (CER) and patient-centered outcomes research (PCOR): Challenges and opportunities. Retrieved from http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.831.9388&rep=rep1&type=pdf Shalansky, S., Basi, S., & Yu, C. (2011). Practice spotlight: Pharmacists in a multidisciplinary atrial fibrillation clinic. The Canadian Journal of Hospital Pharmacy, 64(5), 370–371. 224  Shek, D. T. L., & Ma, C. M. S. (2011). Longitudinal data analyses using linear mixed models in SPSS: Concepts, procedures and illustrations. Scientific World Journal, 11, 42–76. https://doi.org/10.1100/tsw.2011.2 Singer, J. D., & Willett, J. B. (2003). Applied longitudinal data analysis: Modeling change and event occurrence. New York: Oxford University Press. Smigorowsky, M. J., Norris, C. M., McMurtry, M. S., & Tsuyuki, R. T. (2017). Measuring the effect of nurse practitioner (NP)-led care on health-related quality of life in adult patients with atrial fibrillation: Study protocol for a randomized controlled trial. Trials, 18(1), 364. https://doi.org/10.1186/s13063-017-2111-4 Smith, R., & Rennie, D. (2014). Evidence based medicine—An oral history. BMJ, 348, g371. https://doi.org/10.1136/bmj.g371 Son, Y.-J., Baek, K.-H., Lee, S. J., & Seo, E. J. (2019). Health-related quality of life and associated factors in patients with atrial fibrillation: An integrative literature review. International Journal of Environmental Research and Public Health, 16(17). https://doi.org/10.3390/ijerph16173042 Spertus, J., Dorian, P., Bubien, R., Lewis, S., Godejohn, D., Reynolds, M. R., … Burk, C. (2011). Development and validation of the Atrial Fibrillation Effect on QualiTy-of-Life (AFEQT) Questionnaire in patients with atrial fibrillation. Circulation. Arrhythmia and Electrophysiology, 4(1), 15–25. https://doi.org/10.1161/CIRCEP.110.958033 Spertus, J., Tooley, J., Jones, P., Poston, C., Mahoney, E., Deedwania, P., … Weintraub, W. S. (2002). Expanding the outcomes in clinical trials of heart failure: The quality of life and economic components of EPHESUS (EPlerenone’s neuroHormonal efficacy and survival 225  study). American Heart Journal, 143(4), 636–642. https://doi.org/10.1067/mhj.2002.120775 Spitzer, R. L., Kroenke, K., Williams, J. B. W., & Löwe, B. (2006). A brief measure for assessing generalized anxiety disorder: The GAD-7. Archives of Internal Medicine, 166(10), 1092–1097. https://doi.org/10.1001/archinte.166.10.1092 Steinberg, B. A., Dorian, P., Anstrom, K. J., Hess, R., Mark, D. B., Noseworthy, P. A., … Piccini, J. P. (2019). Patient-reported outcomes in atrial fibrillation research: Results of a clinicaltrials.gov analysis. JACC: Clinical Electrophysiology, 5(5), 599–605. https://doi.org/10.1016/j.jacep.2019.03.008 Steinberg, B. A., & Piccini, J. P. (2019). Tackling patient-reported outcomes in atrial fibrillation and heart failure: Identifying disease-specific symptoms? Cardiology Clinics, 37(2), 139–146. https://doi.org/10.1016/j.ccl.2019.01.013 Stoel, R. D., & Wittenboer, G. V. D. (2003). Analyzing longitudinal data using multilevel regression and latent growth curve analysis. Metodología de las ciencias del comportamiento, 5(1), 21–42. Stukel, T. A., Fisher, E. S., Wennberg, D. E., Alter, D. A., Gottlieb, D. J., & Vermeulen, M. J. (2007). Analysis of observational studies in the presence of treatment selection bias: Effects of invasive cardiac management on AMI survival using propensity score and instrumental variable methods. JAMA, 297(3), 278–285. https://doi.org/10.1001/jama.297.3.278 Sutherland, H. J., & Till, J. E. (1993). Quality of life assessments and levels of decision making: Differentiating objectives. Quality of Life Research, 2(4), 297–303. https://doi.org/10.1007/BF00434801 226  Szumilas, M. (2010). Explaining odds ratios. Journal of the Canadian Academy of Child and Adolescent Psychiatry, 19(3), 227–229. Tan, A. C., Armstrong, E., Close, J., & Harris, I. A. (2019). Data quality audit of a clinical quality registry: A generic framework and case study of the Australian and New Zealand Hip Fracture Registry. BMJ Open Quality, 8(3), e000490. https://doi.org/10.1136/bmjoq-2018-000490 Therneau, T., & Crowson, C. (2019). Using time dependent covariates and time dependent coefficients in the cox model. Retrieved from http://cran.r-project.org/web/packages/survival/vignettes/timedep.pdf Thorne, S., & Sawatzky, R. (2014). Particularizing the general: Sustaining theoretical integrity in the context of an evidence-based practice agenda. Advances in Nursing Science, 37(1), 5–18. https://doi.org/10.1097/ANS.0000000000000011 Tilburgs, B., Nijkamp, M. D., Bakker, E. C., & van der Hoeven, H. (2015). The influence of social support on patients’ quality of life after an intensive care unit discharge: A cross-sectional survey. Intensive and Critical Care Nursing, 31(6), 336–342. https://doi.org/10.1016/j.iccn.2015.07.002 Tonelli, M., Wiebe, N., Fortin, M., Guthrie, B., Hemmelgarn, B. R., James, M. T., … Quan, H. (2015). Methods for identifying 30 chronic conditions: Application to administrative data. BMC Medical Informatics and Decision Making, 15–31. https://doi.org/10.1186/s12911-015-0155-5 Trenner, M., Eckstein, H.-H., Kallmayer, M. A., Reutersberg, B., & Kühnl, A. (2019). Secondary analysis of statutorily collected routine data. Gefässchirurgie, 24(3), 220–227. https://doi.org/10.1007/s00772-019-0524-y 227  Tripathy, J. P. (2013). Secondary data analysis: Ethical issues and challenges. Iranian Journal of Public Health, 42(12), 1478–1479. Tu, J. V., Willison, D. J., Silver, F. L., Fang, J., Richards, J. A., Laupacis, A., & Kapral, M. K. (2004). Impracticability of informed consent in the registry of the Canadian stroke network. New England Journal of Medicine, 350(14), 1414–1421. https://doi.org/10.1056/NEJMsa031697 Tueller, S. J., Drotar, S., & Lubke, G. H. (2011). Addressing the problem of switched class labels in latent variable mixture model simulation studies. Structural Equation Modeling, 18(1), 110–131. https://doi.org/10.1080/10705511.2011.534695 Uebersax, J. (2009). LCA frequently asked questions (FAQ). Retrieved from http://www.john-uebersax.com/stat/faq.htm Urschel, S. (2015). Apples, oranges, and statistical magic: Limitations of registry studies and need for collaborative studies. The Journal of Heart and Lung Transplantation, 34(9), 1136–1138. https://doi.org/10.1016/j.healun.2015.05.013 U.S. Department of Health and Human Services Food and Drug Administration. (2009). Guidance for industry: Patient-reported outcome measures: use in medical product development to support labeling claims: Draft guidance. Retrieved from www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM193282.pdf Van Buuren, S. (2018). Flexible Imputation of Missing Data, Second Edition. Boca Raton, FL: CRC Press. 228  van Buuren, S., Brand, J. P. L., Groothuis-Oudshoorn, C. G. M., & Rubin, D. B. (2006). Fully conditional specification in multivariate imputation. Journal of Statistical Computation and Simulation, 76(12), 1049–1064. https://doi.org/10.1080/10629360600810434 van de Schoot, R., Sijbrandij, M., Winter, S. D., Depaoli, S., & Vermunt, J. K. (2017). The GRoLTS-Checklist: Guidelines for reporting on latent trajectory studies. Structural Equation Modeling, 24(3), 451–467. https://doi.org/10.1080/10705511.2016.1247646 Van Der Wees, P. J., Nijhuis-Van Der Sanden, M. W. G., Ayanian, J. Z., Black, N., Westert, G. P., & Schneider, E. C. (2014). Integrating the use of patient-reported outcomes for both clinical practice and performance measurement: Views of experts from 3 countries. The Milbank Quarterly, 92(4), 754–775. https://doi.org/10.1111/1468-0009.12091 Vargha, A., Bergman, L., & Takács, S. (2016). Performing cluster analysis within a person-oriented context: Some methods for evaluating the quality of cluster solutions. Journal for Person-Oriented Research, 2, 78–86. https://doi.org/10.17505/jpor.2016.08 Vermunt, J. K. (2010). Latent class modeling with covariates: Two improved three-step approaches. Political Analysis, 18(4), 450–469. http://www.jstor.org/stable/25792024 Vickers, A. J. (2016). Sharing raw data from clinical trials: What progress since we first asked “Whose data set is it anyway?” Trials, 17. https://doi.org/10.1186/s13063-016-1369-2 Walker, R. L., Hennessy, D. A., Johansen, H., Sambell, C., Lix, L., & Quan, H. (2012). Implementation of ICD-10 in Canada: How has it impacted coded hospital discharge data? BMC Health Services Research, 12(1), 149. https://doi.org/10.1186/1472-6963-12-149 Walton, M. K., Powers, J. H., Hobart, J., Patrick, D., Marquis, P., Vamvakas, S., … International Society for Pharmacoeconomics and Outcomes Research Task Force for Clinical 229  Outcomes Assessment. (2015). Clinical outcome assessments: Conceptual foundation-report of the ISPOR clinical outcomes assessment - emerging good practices for outcomes research task force. Value in Health, 18(6), 741–752. https://doi.org/10.1016/j.jval.2015.08.006 Ware, J. E., & Sherbourne, C. D. (1992). The MOS 36-item short-form health survey (SF-36). I. Conceptual framework and item selection. Medical Care, 30(6), 473–483. Wasmer, K., Eckardt, L., & Breithardt, G. (2017). Predisposing factors for atrial fibrillation in the elderly. Journal of Geriatric Cardiology, 14(3), 179–184. https://doi.org/10.11909/j.issn.1671-5411.2017.03.010 Wasserstein, R. L., Schirm, A. L., & Lazar, N. A. (2019). Moving to a world beyond “p < 0.05.” The American Statistician, 73(sup1), 1–19. https://doi.org/10.1080/00031305.2019.1583913 Weaver, R. R. (2015). Reconciling evidence-based medicine and patient-centred care: Defining evidence-based inputs to patient-centred decisions. Journal of Evaluation in Clinical Practice, 21(6), 1076–1080. https://doi.org/10.1111/jep.12465 Weldring, T., & Smith, S. M. S. (2013). Patient-reported outcomes (PROs) and patient-reported outcome measures (PROMs). Health Services Insights, 6, 61–68. https://doi.org/10.4137/HSI.S11093 Welles, C. C., Whooley, M. A., Na, B., Ganz, P., Schiller, N. B., & Turakhia, M. P. (2011). The CHADS2 score predicts ischemic stroke in the absence of atrial fibrillation among patients with coronary heart disease: Data from the heart and soul study. American Heart Journal, 162(3), 555–561. https://doi.org/10.1016/j.ahj.2011.05.023 230  Whitney, C. W., Lind, B. K., & Wahl, P. W. (1998). Quality assurance and quality control in longitudinal studies. Epidemiologic Reviews, 20(1), 71–80. https://doi.org/10.1093/oxfordjournals.epirev.a017973 Williams, W. G. (2010). Uses and limitations of registry and academic databases. Seminars in Thoracic and Cardiovascular Surgery: Pediatric Cardiac Surgery Annual, 13(1), 66–70. https://doi.org/10.1053/j.pcsu.2010.02.007 Wilson, I. B., & Cleary, P. D. (1995). Linking clinical variables with health-related quality of life: A conceptual model of patient outcomes. JAMA, 273(1), 59–65. https://doi.org/10.1001/jama.1995.03520250075037 Wilson IB, & Cleary PD. (1995). Linking clinical variables with health-related quality of life: A conceptual model of patient outcomes. JAMA, 273(1), 59–65. https://doi.org/10.1001/jama.1995.03520250075037 Winkle, R. A., Mead, R. H., Engel, G., Kong, M. H., & Patrawala, R. A. (2014). Comparison of CHADS2 and CHA2DS2-VASC anticoagulation recommendations: Evaluation in a cohort of atrial fibrillation ablation patients. Europace, 16(2), 195–201. https://doi.org/10.1093/europace/eut244 Wokhlu, A., Monahan, K. H., Hodge, D. O., Asirvatham, S. J., Friedman, P. A., Munger, T. M., … Packer, D. L. (2010). Long-term quality of life after ablation of atrial fibrillation the impact of recurrence, symptom relief, and placebo effect. Journal of the American College of Cardiology, 55(21), 2308–2316. https://doi.org/10.1016/j.jacc.2010.01.040 World Health Organization. (1946). Preamble to the Constitution of WHO as adopted by the International Health Conference, New York, 19 June—22 July 1946; signed on 22 July 1946 by the representatives of 61 States (Official Records of WHO, no. 2, p. 100) and 231  entered into force on 7 April 1948. Retrieved from http://www.who.int/suggestions/faq/en/ World Health Organization. (2019). International Classification of Diseases (ICD) Information Sheet. Retrieved October 14, 2019, from WHO website: http://www.who.int/classifications/icd/factsheet/en/ Zoni-Berisso, M., Lercari, F., Carazza, T., & Domenicucci, S. (2014). Epidemiology of atrial fibrillation: European perspective. Clinical Epidemiology, 6, 213–220. https://doi.org/10.2147/CLEP.S47385 Zubritsky, C., Abbott, K. M., Hirschman, K. B., Bowles, K. H., Foust, J. B., & Naylor, M. D. (2013). Health-related quality of life: Expanding a conceptual framework to include older adults who receive long-term services and supports. The Gerontologist, 53(2), 205–210. https://doi.org/10.1093/geront/gns093 232  Appendices AFEQT Questionnaire  233       234  List of Comorbidities and Treatment Algorithms System Category Algorithm ICD-9 CM ICD-10/CCI Description Positive predictive value/ sensitivity Cardiac Heart failure 1 hospitalization or 2 claims in 1 year 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 (A428, I428, N428, R428)* I09.9 Rheumatic heart disease High  (PPV 72%  Sn 91%) ICD9  Low  (PPV 69% Sn 90% ICD10 (Tonelli et al., 2015) I25.5 Ischemic cardiomyopathy I42.0 Dilated cardiomyopathy I42.5-I42.9  Other cardiomyopathy (alcohol, drug, unspecified)  I43 Cardiomyopathy in diseases classified elsewhere I50 Heart failure (left ventricular, systolic, diastolic, combined, unspecified) Myocardial infarction 1 hospitalization in 1 year 410 I21 Includes ST elevation at different sites (anterior wall, inferior wall, other sites), and other types of myocardial infarction  High  (PPV 89% Sn 89% ICD9 (Tonelli et al., 2015) I22 Subsequent ST elevation and non-ST elevation myocardial infarction Hypertension 1 hospitalization or 2 claims in 1 year 401-405 I10 Essential (primary) hypertension High  (PPV 95% Sn 79% ICD9  Moderate  (PPV 93%  Sn 68%) ICD10 (Tonelli et al., 2015) I11 Hypertensive heart disease I12 Hypertensive renal disease I13 Hypertensive heart and renal disease I15 Secondary hypertension  Peripheral vascular disease 1 hospitalization or 1 claim in 1 year 440.2 I70.2 Atherosclerosis of native arteries of the extremities High  (PPV 94% Sn 77%) ICD9 (Tonelli et al., 2015) Stroke/ Transient ischemic attack 1 hospitalization or 1 claim in 1 year 362.3, 430 (A430, D430, H430, I430, N430, R430)*, 431, 433, 434, 435, 436 G45.0-G45.3, G45.8-G45.9 Transient cerebral ischemic attacks and related syndromes  Moderate  (PPV 90%  Sn na) ICD9  Moderate  (PPV 92%  Sn na) ICD10 (Tonelli et al., 2015) H34.1 Central retinal artery occlusion I60 Nontraumatic subarachnoid hemorrhage I61 Nontraumatic intracerebral hemorrhage I63 Cerebral infarction  I64 Stroke, not specified as hemorrhage or infarction 235  System Category Algorithm ICD-9 CM ICD-10/CCI Description Positive predictive value/ sensitivity Intervention Cardioversion 1 hospitalization  1HZ09 In CCI, cardioversion is classified to the generic intervention of “stimulation.” na Ablation  1 hospitalization   1HH59 In CCI, ablation is classified to “destruction, cardiac conduction system” (e.g., using cryoprobe and radiofrequency).   na Coronary artery bypass surgery (CABG) 1 hospitalization   1IJ76 In CCI, CABG is classified to bypass, coronary arteries (e.g., using saphenous vein, internal thoracic artery, and synthetic tissue).  PPV 98% (Lee et al., 2013) Percutaneous coronary intervention (PCI) 1 hospitalization   1IJ50  Dilation, coronary arteries (e.g., angioplasty). PPV 96% (Lee et al., 2013) 1IJ57 Extraction, coronary arteries (e.g., endarterectomy). Valve surgery 1 hospitalization   1HU90, 1HU80 Mitral valve replacement/repair PPV 97% (Lee et al., 2013) 1HV90, 1HV80 Aortic valve replacement/repair  1HT90, 1HT80  Pulmonary valve replacement/repair 1HS90, 1HS80 Tricuspid valve replacement/repair  1HW Valve annulus surgery not elsewhere classified Pacemaker implant  1 hospitalization   1HZ53GRFR, 1HZ53LAFR, 1HZ53SYFR Cardiac resynchronization therapy pacemaker  na (Huang, Redpath, & van Walraven, 2015; Krahn, Wagar, Jackson, Tang, & McFarlane, 2014) 1HZ53GRNM,  1HZ53LANM, 1HZ53QANM Single chamber pacemaker 1HZ53GRNK, 1HZ53LANK, 1HZ53QANK,  Dual chamber pacemaker  1HZ53GRNL, 1HZ53LANL, 1HZ53QANL, Fixed rate pacemaker  1HZ53GRNN, 1HZ53LANN Temporary pacemakers Defibrillator implantation 1 hospitalization   1HZ53GRFS, Includes implantable cardioverter/defibrillator na 236  System Category Algorithm ICD-9 CM ICD-10/CCI Description Positive predictive value/ sensitivity 1HZ53HAFS, 1HZ53LAFS, 1HZ53SYFS (transluminal, open, or combined approach).  (Huang et al., 2015; Krahn et al., 2014) Respiratory Chronic obstructive pulmonary disease 1 hospitalization or 2 claims in 1 year 416.8, 416.9, 490, 491 (A491,D491, H491, I491, R491)*, 492, 494-505, 506.4, 508.1, 508.8 I27.8, I27.9,  Other specified/ unspecified pulmonary heart disease Moderate(PPV 92% Sn 55%) ICD9 Moderate (PPV 91% Sn 53%) ICD10 (Tonelli et al., 2015) J40.x- J47.x  Chronic lower respiratory diseases (e.g., bronchitis, emphysema, asthma) J60.x-J67.x Lung disease due to external agents (e.g., Pneumoconiosis, airway disease due to organic dust) J68.4, Chronic respiratory conditions due to chemicals, gases, fumes and vapors J70.1, J70.3 Chronic and other pulmonary manifestations due to radiation/drug-induced interstitial lung disorders Sleep disorder 1 hospitalization or 1 claim in 1 year 291.82,292.8, 307.4x,327.1, 327.12,327.2, 327.3x,327.4x, 327.5x,327.8, 333.94,347.x, 780.5x,8917, 8918,V69.4  F51.x Nonorganic sleep disorders PPV 88.8% Sn 34.0% (Jolley et al., 2018)  G25.8 Other specified extrapyramidal and movement disorders G47.x Other sleep disorders R06.81 Apnea, not elsewhere classified Gastro-intestinal Gastro-intestinal bleed  1 hospitalization or 1 claim in 1 year 531.0, 531.2, 531.4, 531.6, 532.0, 532.2, 532.4, 532.6, 533.0, 533.2, 533.4, 533.6, 534.0, 534.2, 534.4, 534.6, 578.0, I85.0 Esophageal varices PPV 93% Sn 91% (Arnason, Wells, van Walraven, & Forster, 2006; Lam et al., 2013) I98.20, I98.3 Esophageal varices with bleeding classified elsewhere K22.10, K22.12,  Ulcer of esophagus, acute with bleeding/perforation K22.14, K22.16 Ulcer of esophagus, chronic with bleeding/perforation K25.0, K25.2 Gastric ulcer, acute with bleeding/perforation K25.4, K25.6  Gastric ulcer, chronic with bleeding/perforation K26.0, K26.2  Duodenal ulcer, acute with bleeding/perforation K26.4, K26.6 Duodenal ulcer, chronic with bleeding/perforation 237  System Category Algorithm ICD-9 CM ICD-10/CCI Description Positive predictive value/ sensitivity 578.1, 578.9, 569.3  K27.0, K27.2,  Peptic ulcer, acute with bleeding/perforation K27.4, K27.6,  Peptic ulcer, chronic with bleeding/perforation K28.0, K28.2,  Gastrojejunal ulcer, acute with bleeding/perforation K28.4, K28.6,  Gastrojejunal ulcer, chronic with bleeding/perforation K29.0 Acute bleeding gastritis K31.80 Angiodysplasia of stomach and duodenum with bleeding K63.80 Angiodysplasia of small intestine, except duodenum with bleeding K55.20 Angiodysplasia of colon with bleeding K62.5 Bleeding of anus and rectum K92.2 Gastrointestinal bleeding, unspecified Peptic ulcer disease 1 hospitalization or 2 claims in 1 year  531.7, 531.9, 532.7, 532.9, 533.7, 533.9, 534.7, 534.9 K25.7, K25.9 Gastric ulcer (acute or chronic) without hemorrhage or perforation Moderate (PPV 84% Sn: 37%) ICD9 Moderate (PPV 77%  Sn 40%) ICD10 (Tonelli et al., 2015) K26.7, K26.9 Duodenal ulcer (acute or chronic) without hemorrhage or perforation K27.7, K27.9  Peptic ulcer (acute or chronic) without hemorrhage or perforation K28.7, K28.9 Gastrojejunal ulcer (acute or chronic) without hemorrhage or perforation Gastro-urinary/ Endocrine Chronic kidney disease 1 hospitalization or 3 claims in 1 year 583, 584, 585 (A585, C585, D585, H585, I585, N585, R585)*, 586, 592, 593.9 N00-N08 Glomerular diseases Low (PPV 64% Sn: 12%) ICD9/ICD10 (Tonelli et al., 2015) N10-N16 Renal tubule-interstitial diseases N17-N19 Acute kidney failure and chronic kidney diseases N20-N23 Urolithiasis Hypo-thyroidism 1 hospitalization or 2 claims in 1 year 240.9, 243, 244, 246.1, 246.8 E01 Iodine-deficiency-related thyroid disorders and allied conditions Moderate (PPV 93% Sn 65) ICD9 Moderate (PPV 93  Sn 39) ICD10  (Tonelli et al., 2015)  E02 Subclinical iodine-deficiency hypothyroidism  E03 Other hypothyroidism E89.0 Postprocedural hypothyroidism 238  System Category Algorithm ICD-9 CM ICD-10/CCI Description Positive predictive value/ sensitivity Diabetes  1 hospitalization or 2 claims in 1 year 250 (H250, I250, N250, R250)* E10 Type 1 diabetes mellitus High (PPV 80% Sn 86%) ICD9 (Tonelli et al., 2015) E11 Type 2 diabetes mellitus E12 Malnutrition-related diabetes mellitus  E13 Other specified diabetes mellitus  E14 Unspecified diabetes mellitus  Mental health Depression 1 hospitalization or 2 claims in 1 year 296.2, 296.3, 296.5, 300.4, 309, 311 F20.4 Post-schizophrenic depression Moderate (PPV: 80% Sn: 57%) ICD9 Moderate (PPV 92%  Sn 45%) ICD10 (Tonelli et al., 2015)  F31.3-F31.5,  Bipolar affective disorder, current episode depression (mild/moderate/severe) with or without psychotic symptoms F32 Depressive episode F33 Recurrent depressive disorder F34.1 Dysthymia F41.2  Mixed anxiety and depressive disorder F43.2 Adjustment disorders Note. ICD-9 CM International Classification of Diseases 9th Revision Clinical Modification; ICD-10 International Classification of Diseases 10th Revision; CCI Canadian Classification of Health Interventions; n/a Not available; *MSP-only diagnostic codes include A, D, H, I, N, R. Variable is considered to be of high validity if both positive predictive value and sensitivity ≥ 70% compared to an acceptable gold standard such as chart review. Negative predictive value or specificity is not considered because they are generally > 90% (Tonelli et al., 2015).   239  SAS code for calculating the Charlson Comorbidity Index /* This SAS code calculates Charlson Comorbidity Index scores & generates an overall (longitudinal)  score for individuals based on ALL episodes of care (DADs or MSP over specified period of time), by collapsing the calculated individual scores into one record per person. The original SAS code was based on Hude Quan's group at the University of Calgary, in Calgary Alberta, and has been modified to work with DADs or MSP data housed within Population Data BC*/  *DATE: January 4, 2019 AUTHOR: Jae-Yung Kwon***;  proc import out=work.hospital_msp datafile='R:\working\MSP & DADs comorbidity\msp_hosp_dat.csv' dbms = csv replace; run; proc print data=work.hospital_msp(obs=10);  run;  data charlson_hosp; set work.hospital_msp;  /* set up array for DIAGX1-25 */ array DX(25) DIAGX1-DIAGX25;    /*Myocardial Infarction*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('I21','I22','I252') then do; CC_GRP_1 =1; end; end;  /*Congestive Heart Failure*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('I41','I50','I099','I110','I130','I132','I255','I420','I425','I426',       'I427','I428','I429','P290')then do;  CC_GRP_2 =1; end; end;  /*Peripheral Vascular Disease*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('I70','I71','I731','I738','I739','I771','I790','I792','K551','K558',       'K559','Z958','Z959')then do; CC_GRP_3 =1; end; end;  /*Cerebrovascular Disease*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('G45','G46','I60','I61','I62','I63','I64','I65','I66','I67','I68',       'I69','H340')then do; CC_GRP_4 =1; end; end;  /*Dementia*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('F00','F01','F02','F03','G30','F051','G311')then do;  240  CC_GRP_5 =1;  end; end;  /*Chronic Pulmonary Disease*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('J40','J41','J42','J43','J44','J45','J46','J47','J60','J61','J62','J63',       'J64','J65','J66','J67','I278','I279','J684','J701','J703')then do; CC_GRP_6 =1; end; end;  /*Connective Tissue Disease-Rheumatic Disease*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('M05','M32','M33','M34','M06','M315','M351','M353','M360')then do; CC_GRP_7 =1; end; end;  /*Peptic Ulcer Disease*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('K25','K26','K27','K28')then do;  CC_GRP_8 =1; end; end;  /*Mild Liver Disease*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('B18','K73','K74','K700','K701','K702','K703','K709','K717','K713',       'K714','K715','K760','K762','K763','K764','K768','K769','Z944')then do;  CC_GRP_9 =1; end; end;  /*Diabetes w/o complications */  do i=1 to 25 until (DX{i}=''); if DX(i) in:('E100','E101','E106','E108','E109','E110','E111','E116','E118','E119',       'E120','E121','E126','E128','E129','E130','E131','E136','E138','E139',       'E140','E141','E146','E148','E149')then do;  CC_GRP_10 =1; end; end;  /*Diabetes w complications */  do i=1 to 25 until (DX{i}=''); if DX(i) in:('E102','E103','E104','E105','E107','E112','E113','E114','E115','E117',       'E122','E123','E124','E125','E127','E132','E133','E134','E135','E137',       'E142','E143','E144','E145','E147')then do; CC_GRP_11 =1;  end; end;  /*Paraplegia and Hemiplegia*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('G81','G82','G041','G114','G801','G802','G830','G831','G832','G833',       'G834','G839')then do; CC_GRP_12 =1; end; end;  /*Renal Disease*/  do i=1 to 25 until (DX{i}=''); 241  if DX(i) in:('N18','N19','N052','N053','N054','N055','N056','N057','N250','I120',       'I131','N032','N033','N034','N035','N036','N037','Z490','Z491','Z492',       'Z940','Z992')then do; CC_GRP_13 =1; end; end;  /*Cancer*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('C00','C01','C02','C03','C04','C05','C06','C07','C08','C09','C10','C11',       'C12','C13','C14','C15','C16','C17','C18','C19','C20','C21','C22','C23',       'C24','C25','C26','C30','C31','C32','C33','C34','C37','C38','C39','C40',       'C41','C43','C45','C46','C47','C48','C49','C50','C51','C52','C53','C54',       'C55','C56','C57','C58','C60','C61','C62','C63','C64','C65','C66','C67',       'C68','C69','C70','C71','C72','C73','C74','C75','C76','C81','C82','C83',       'C84','C85','C88','C90','C91','C92','C93','C94','C95','C96','C97')then do;  CC_GRP_14 =1; end; end;  /*Moderate or Severe Liver Disease*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('K704','K711','K721','K729','K765','K766','K767','I850','I859','I864','I982')then do; CC_GRP_15 =1; end; end;  /*Metastatic Carcinoma*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('C77','C78','C79','C80')then do; CC_GRP_16 =1; end; end;  /*AIDS/HIV*/  do i=1 to 25 until (DX{i}=''); if DX(i) in:('B20','B21','B22','B24')then do; CC_GRP_17 =1; end;  end;  /*Count total number of groups for each record*/  * TOT_GRP = CC_GRP_1 + CC_GRP_2 + CC_GRP_3 + CC_GRP_4 + CC_GRP_5 + CC_GRP_6 + CC_GRP_7 + CC_GRP_8 +CC_GRP_9 + CC_GRP_10 + CC_GRP_11 + CC_GRP_12 + CC_GRP_13 + CC_GRP_14 + CC_GRP_15 +   CC_GRP_16 +CC_GRP_17;  * LABEL TOT_GRP = 'Total CCI Groups per record'; * run;  proc print data=charlson_hosp (obs=10);  run;  data charlson_msp; set charlson_hosp;  *****************CHARLSON MSP******************; grp_1 = (icd9 in ('410','412')); 242  label grp_1 = 'Myocardial Infarction';   grp_2 = (icd9 in ('398','402','425','428')); label grp_2 = 'Congestive Heart Failure';  grp_3 = (icd9 in ('440','441','443','447','557')); label grp_3 = 'Peripheral Vascular Disease';  grp_4 = (icd9 in ('430','431','432','433','434','435','436','437','438')); label grp_4 = 'Cerebrovascular Disease';  grp_5 = (icd9 in ('290','294','331')); label grp_5 = 'Dementia';  grp_6 = (icd9 in ('416','490','491','492','493','494','495','496','500','501','502','503','504','505')); label grp_6 = 'Chronic Pulmonary Disease';  grp_7 = (icd9 in ('446','710','714','725')); label grp_7 = 'Connective Tissue Disease-Rheumatic Disease';  grp_8 = (icd9 in ('531','532','533','534')); label grp_8 = 'Peptic Ulcer Disease';  grp_9 = (icd9 in ('070','570','571','573')); label grp_9 = 'Mild Liver Disease';  grp_10 = (icd9 in ('250')); label grp_10 = 'Diabetes without complications';  grp_11 = 0; *** overlap with #10; label grp_11 = 'Diabetes with complications';  grp_12 = (icd9 in ('334','342','343','344')); label grp_12 = 'Paraplegia and Hemiplegia';  grp_13 = (icd9 in ('403','582','583','585','586','588','V56')); label grp_13 = 'Renal Disease';  grp_14 = (icd9 in ('140','141','142','143','144','145','146','147','148','149',        '150','151','152','153','154','155','156','157','158','159',        '160','161','162','163','164','165','170','171','172','174',        '175','176','179','180','181','182','183','184','185','186',        '187','188','189','190','191','192','193','194','195','200',        '201','202','203','204','205','206','207','208','238')); label grp_14 = 'Cancer';  grp_15 = (icd9 in ('456','572')); label grp_15 = 'Moderate or Severe Liver Disease';  grp_16 = (icd9 in ('196','197','198','199')); label grp_16 = 'Metastatic Carcinoma';  grp_17 = (icd9 in ('042','043','044')); label grp_17 = 'AIDS/HIV'; 243   run;   data charlson_all; set charlson_msp;  proc sort data= charlson_all;  by STUDYID;  run; data charlson_longitudinal_all;  set charlson_all (keep=STUDYID CC_GRP_1-CC_GRP_17 grp_1-grp_17);  by STUDYID; /* retain statement retains the value it had in the previous iteration */ retain ccgrp1 ccgrp2 ccgrp3 ccgrp4 ccgrp5 ccgrp6 ccgrp7 ccgrp8 ccgrp9      ccgrp10 ccgrp11 ccgrp12 ccgrp13 ccgrp14 ccgrp15 ccgrp16 ccgrp17;  if first.STUDYID then do;           ccgrp1=0; ccgrp2=0; ccgrp3=0; ccgrp4=0; ccgrp5=0; ccgrp6=0; ccgrp7=0; ccgrp8=0; ccgrp9=0;     ccgrp10=0; ccgrp11=0; ccgrp12=0; ccgrp13=0; ccgrp14=0; ccgrp15=0; ccgrp16=0; ccgrp17=0; end;  ***comorbidity variables generated from hospital; array hsp{17} cc_grp_1 cc_grp_2 cc_grp_3 cc_grp_4 cc_grp_5 cc_grp_6 cc_grp_7 cc_grp_8 cc_grp_9      cc_grp_10 cc_grp_11 cc_grp_12 cc_grp_13 cc_grp_14 cc_grp_15 cc_grp_16 cc_grp_17;  ***comorbidity variables generated from msp; array msp{17} grp_1 grp_2 grp_3 grp_4 grp_5 grp_6 grp_7 grp_8 grp_9      grp_10 grp_11 grp_12 grp_13 grp_14 grp_15 grp_16 grp_17;  ***summary comorbidity; array tot{17}ccgrp1 ccgrp2 ccgrp3 ccgrp4 ccgrp5 ccgrp6 ccgrp7 ccgrp8 ccgrp9      ccgrp10 ccgrp11 ccgrp12 ccgrp13 ccgrp14 ccgrp15 ccgrp16 ccgrp17;  do i =1 to 17; if hsp{i}=1 or msp{i} then tot{i}=1; end;  if last.STUDYID then do;   totalcc=sum(of ccgrp1-ccgrp17);    ***Use weighted score;   wgtcc=sum(of ccgrp1-ccgrp10)+ccgrp11*2+ccgrp12*2+ccgrp13*2+ccgrp14*2+ccgrp15*3+ccgrp16*6+ccgrp17*6;  output; end;   keep STUDYID totalcc wgtcc ccgrp1 ccgrp2 ccgrp3 ccgrp4 ccgrp5 ccgrp6 ccgrp7 ccgrp8 ccgrp9      ccgrp10 ccgrp11 ccgrp12 ccgrp13 ccgrp14 ccgrp15 ccgrp16 ccgrp17; run;    244  Characteristics of Variables with Missing Data Variables Values Missing, n (%) Age Continuous in years 0 (0) Gender Men (0) or Women (1) 0 (0) Distance to clinic < 100 km (0) ≥ 100 km (1) 185 (1.4) Clinic site Clinic coded from 1-5 0 (0) CHADS2 Continuous 1,830 (15.0) CCS-SAF Continuous 12 (.1) CCI Continuous 0 (0) Hypertension No (0) or Yes (1) 0 (0) Heart failure No (0) or Yes (1) 0 (0) Stroke/TIA No (0) or Yes (1) 0 (0) Myocardial infarction No (0) or Yes (1) 0 (0) Peripheral vascular disease No (0) or Yes (1) 0 (0) Cardioversion No (0) or Yes (1) 1,147 (8.7) Ablation (time-varying) No (0) or Yes (1) 1,147 (8.7) Pacemaker No (0) or Yes (1) 1,147 (8.7) PCI No (0) or Yes (1) 1,147 (8.7) CABG No (0) or Yes (1) 1,147 (8.7) Valve surgery No (0) or Yes (1) 1,147 (8.7) Dialysis No (0) or Yes (1) 1,147 (8.7) Defibrillator implantation No (0) or Yes (1) 1,147 (8.7) Diabetes No (0) or Yes (1) 0 (0) COPD No (0) or Yes (1) 0 (0) Depression No (0) or Yes (1) 0 (0) Chronic kidney disease No (0) or Yes (1) 0 (0) Sleep disorder No (0) or Yes (1) 0 (0) Hypothyroidism No (0) or Yes (1) 0 (0) Gastrointestinal bleed No (0) or Yes (1) 0 (0) Peptic ulcer No (0) or Yes (1) 0 (0) Anticoagulants (time-varying) No (0) or Yes (1) 15 (.1) Beta-blockers No (0) or Yes (1) 15 (.1) Antiarrhythmics No (0) or Yes (1) 15 (.1) Calcium channel blockers No (0) or Yes (1) 15 (.1) Digoxin No (0) or Yes (1) 15 (.1) Antiplatelets No (0) or Yes (1) 15 (.1) Note. CCS-SAF = Canadian Cardiovascular Society Severity in Atrial Fibrillation Scale; CCI = Charlson Comorbidity Index; TIA = Transient Ischemic Attack; PCI = Percutaneous coronary intervention; CABG = Coronary artery bypass graft; COPD = Chronic Obstructive Pulmonary Disease.      245  Characteristics of Respondents and Non-Respondents during Study Period  (2008-2016) Characteristics Eligible Cohort (based on criteria)  Respondents (completed a questionnaire) Non-respondents (did not complete a questionnaire) p value N=13,113 (100%) n=7,439 (56.7%) n=5,674 (43.3%) Age, median (IQR) 67 (16) 67 (16) 67 (17) .14 Women 5,069 (38.7) 2,897 (38.9) 2,172 (38.3) .45 Distance to clinic    <.001   ≥ 100 km 2,594 (20.1) 1,223 (16.7) 1,371 (24.5)    <100 km 10,334 (79.9) 6,111 (83.3) 4,223 (75.5)  Missing  105 (.01) 80 (.01)  Clinic site    <.001 1 1,893 (14.4) 1,160 (15.6) 733 (12.9)  2 2,516 (19.2) 1,360 (18.3) 1,156 (20.4)  3 2,391 (18.2) 2,128 (28.6) 263 (4.6)  4 1,592 (12.1) 948 (12.7) 644 (11.4)  5 4,721 (36.0) 1,843 (24.8) 2,878 (50.7)  CHADS2, mean (SD) 1.19 (1.17) 1.15 (1.14) 1.24 (1.17) <.001        Missing 1,830 (14.0) 1,230 (.17) 600 (.11)  CCS-SAF,  mean (SD) 1.73 (1.10) 1.66 (1.08) 1.82 (1.13) <.001 Missing 12 (.1) 5 (.0) 7 (.0)  Charlson score, mean (SD)  3.24 (3.27) 3.16 (3.16) 3.34 (3.4) .10 Cardiac condition       Hypertension 4,604 (35.1) 2,620 (35.2) 1,984 (35.0) .78   Heart failure 1,962 (15.0) 1,164 (15.6) 798 (14.1) <.05   Stroke/TIA 979 (7.5) 555 (7.5) 424 (7.5) .99   MI 276 (2.1) 169 (2.3) 107 (1.9) .14 PVD 53 (.4) 26 (.3) 27 (.5) .32 Prior interventions     Cardioversion 1,897 (14.5)  1,091 (14.7) 806 (14.2) .57 Ablation 842 (7.0) 425 (6.2) 417 (8.1) <.001 Pacemaker 393 (3.3) 218 (3.2) 175 (3.4) .60 PCI 372 (3.1) 210 (3.1) 162 (3.1) .90 CABG 248 (2.1) 147 (2.2) 101 (2.0) .48 Valve surgery 230 (1.9) 126 (1.9) 104 (2.0) .56 Dialysis 66 (.6) 31 (.5) 35 (.7) .13 Defibrillator implantation 52 (.4) 30 (.4) 22 (.4) .99 Missing 1,147 (8.7) 631 (8.5) 516 (9.1)  Comorbidities       Diabetes 1,800 (13.7) 998 (13.4) 802 (14.1) .25 COPD 779 (5.9) 420 (5.6) 359 (6.3) .11   Depression 599 (4.6) 330 (4.4) 269 (4.7) .43 246  Characteristics Eligible Cohort (based on criteria)  Respondents (completed a questionnaire) Non-respondents (did not complete a questionnaire) p value N=13,113 (100%) n=7,439 (56.7%) n=5,674 (43.3%)   CKD 496 (3.8) 266 (3.6) 230 (4.1) .17   Sleep disorder 385 (2.9) 210 (2.8) 175 (3.1) .41 Hypothyroidism 288 (2.2) 164 (2.2) 124 (2.2) .99   GI bleed 242 (1.8) 150 (2.0) 92 (1.6) .11   Peptic ulcer 28 (.2) 15 (.2) 13 (.2) .88 Prior medications     Anticoagulants 8,170 (62.4) 4,611 (62.1) 3,559 (62.8) .40   Beta-blockers 7,654 (58.4) 4,360 (58.7) 3,294 (58.1) .53   Antiarrhythmics 4,624 (35.3) 2,685 (35.9) 1,939 (34.2) <.05   Calcium channel blockers 3,152 (24.1) 1,733 (23.3) 1,419 (25.0) <.05   Digoxin 1,613 (12.3) 873 (11.7) 740 (13.1) <.05   Antiplatelets 646 (4.9) 352 (4.7) 294 (5.2) .26 Missing 15 (.1) 9 (.0) 6 (.0)  Note. Continuous variables compared using the Mann-Whitney U test. Categorical variables compared using Χ2 test. Results based on imputed dataset 1. CCS-SAF = The Canadian Cardiovascular Society Severity in Atrial Fibrillation Scale; TIA = Transient ischemic attack; MI = Myocardial infarction; PVD = Peripheral vascular disease; PCI = Percutaneous coronary intervention; CABG = Coronary artery bypass graft; COPD = Chronic obstructive pulmonary disease; CKD = Chronic kidney disease; GI bleed = Gastrointestinal bleed.          247  Auxiliary Variables  Auxiliary Variables Algorithm Age NA Gender NA CHADS2 2lonly.pmm CCS-SAF 2lonly.pmm Distance to clinic 2lonly.pmm Clinic site NA Heart failure NA Hypertension NA Peripheral vascular disease NA Stroke NA Cardioversion 2lonly.pmm Ablation (time-varying) 2lonly.pmm Coronary artery bypass graft 2lonly.pmm Percutaneous coronary intervention 2lonly.pmm Valve surgery 2lonly.pmm Pacemaker 2lonly.pmm Defibrillator implantation 2lonly.pmm Dialysis 2lonly.pmm Chronic obstructive pulmonary disease NA Sleep disorder NA Gastro-intestinal bleed NA Peptic ulcer NA Chronic kidney disease NA Hypothyroid NA Diabetes NA Depression NA Charlson score NA Anticoagulants (time-varying) 2lonly.pmm Antiplatelet 2lonly.pmm Betablockers 2lonly.pmm Antiarrhythmics 2lonly.pmm Calcium channel blockers 2lonly.pmm Digoxin 2lonly.pmm Death indicator NA Q1:Q12  2l.pmm Note. NA=Not Applicable or no missing data; 2lonly.pmm=imputation at level-2 by predictive mean matching; 2l.pmm=imputation at level-1 by predictive mean matching.   248  Characteristics of Respondents (resp) and Non-Respondents (non-resp) at Baseline (N = 7,439) and during Follow-Up (N = 4,412)    Characteristics Total  (N = 7,439) Follow-up Assessments (N = 4,412) Initial consult Resp n=4,040 Non-resp n=3,398 6 months  (>0 to ≤0.5year) Resp n=2,472 (56.0%) Non-resp n=1,940 (44.0%) 1 year (>0.5 to ≤1) Resp n=996  (22.6%) Non-resp n=3,416 (77.4%) 1.5 years (>1 to ≤1.5) Resp n=624 (14.1%) Non-resp n=3,788 (85.9%) 2 years (>1.5 to ≤2) Resp n=454 (10.3%) Non-resp n=3,958 (89.7%) >2 years Resp n=962 (21.8%) Non-resp n=3,450 (78.2%) Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Time-invariant             Age, median (IQR) 66  (16) 68  (16) 69  (16) 65  (16) 68  (17) 67  (16) 66  (14) 68  (17) 66  (15) 67  (16) 64  (14) 68  (17) Women 1,552 (38.4) 1,345 (39.6) 1,013 (41.0) 738 (38.0) 380        (38.2) 1,371 (40.1) 237 (38.0) 1,514 (40.0) 176 (38.8) 1,575 (39.8) 365 (37.9) 1,386 (40.2) Distance to clinic               ≥ 100 km 420 (10.4) 811 (23.9) 478 (19.3) 418 (21.5) 145 (14.6) 751 (22.0) 108 (17.3) 788 (20.8) 120 (26.4) 776 (19.6) 229 (23.8) 667 (19.3) Clinic site             1 980 (24.3) 180 (5.3) 137  (5.5) 132  (6.8) 44  (4.4) 225 (6.6) 29  (4.6) 240 (6.3) 20  (4.4) 249 (6.3) 60  (6.2) 209 (6.1) 2 1,086 (26.9) 274 (8.1) 125  (5.1) 244 (12.6) 48  (4.8) 321 (9.4) 47  (7.5) 322 (8.5) 40  (8.8) 329 (8.3) 139 (14.4) 230 (6.7) 3 1,547 (38.3) 581 (17.1) 558 (22.6) 646 (33.3) 351 (35.2) 853 (25.0) 231 (37.0) 973 (25.7) 135 (29.7) 1,069 (27.0) 211 (21.9) 993 (28.8) 4 246 (6.1) 702 (20.7) 726 (29.4) 129  (6.6) 316 (31.7) 539 (15.8) 120 (19.2) 735 (19.4) 60 (13.2) 795 (20.1) 83  (8.6) 772 (22.4) 5 181 (4.5) 1,662 (48.9) 926 (37.5) 789 (40.7) 237 (23.8) 1,478 (43.3) 197 (31.6) 1,518 (40.1) 199 (43.8) 1,516 (38.3) 469 (48.8) 1,246 (36.1) CHADS2,  mean (SD) 1.16 (1.16) 1.18 (1.14) 1.27 (1.14) 1.03  (1.12) 1.23 (1.15) 1.15 (1.13) 1.04 (1.08) 1.19 (1.15) 0.98 (1.11) 1.19 (1.14) 0.93 (1.07) 1.23 (1.15) CCS-SAF,  mean (SD) 1.67 (1.03) 1.66 (1.13) 1.53 (1.15) 1.78 (1.03) 1.55 (1.12) 1.67 (1.10) 1.71 (1.09) 1.63 (1.11) 1.81 (1.06) 1.62 (1.11) 1.90 (1.01) 1.57 (1.12) Charlson score mean (SD)  3.18 (3.18) 3.14 (3.14) 3.18 (3.26) 3.11 (2.97) 3.19 (3.11) 3.13 (3.14) 3.07 (3.06) 3.16 (3.15) 3.05 (3.04) 3.16 (3.14) 2.90 (2.79) 3.22 (3.22) 249     Characteristics Total  (N = 7,439) Follow-up Assessments (N = 4,412) Initial consult Resp n=4,040 Non-resp n=3,398 6 months  (>0 to ≤0.5year) Resp n=2,472 (56.0%) Non-resp n=1,940 (44.0%) 1 year (>0.5 to ≤1) Resp n=996  (22.6%) Non-resp n=3,416 (77.4%) 1.5 years (>1 to ≤1.5) Resp n=624 (14.1%) Non-resp n=3,788 (85.9%) 2 years (>1.5 to ≤2) Resp n=454 (10.3%) Non-resp n=3,958 (89.7%) >2 years Resp n=962 (21.8%) Non-resp n=3,450 (78.2%) Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Cardiac condition               Hypertension 1,428 (35.3) 1,192 (35.1) 949 (38.4) 596 (30.7) 355 (35.6) 1,190 (34.8) 207 (33.2) 1,338 (35.3) 151 (33.3) 1,394 (35.2) 291 (30.2) 1,254 (36.3)   Heart failure 689 (17.1) 475 (14.0) 396 (16.0) 280 (14.4) 156 (15.7) 520 (15.2) 99 (15.9) 577 (15.2) 61 (13.4) 615 (15.5) 106 (11.0) 570 (16.5)   Stroke/TIA 312 (7.7) 243 (7.1) 189  (7.6) 136  (7.0) 82  (8.2) 243 (7.1) 41  (6.6) 284 (7.5) 26  (5.7) 299 (7.6) 60  (6.2) 265 (7.7)   MI 87  (2.2) 82  (2.4) 63  (2.5) 35  (1.8) 27  (2.7) 71  (2.1) 7  (1.1) 91  (2.4) 8  (1.8) 90  (2.3) 17  (1.8) 81  (2.3) PVD 19  (.5) 7  (.2) 7  (.3) <5   <5   6  (.2) <5   7  (.2) 0   8  (.2) 0   8  (.2) Prior interventions            Cardioversion 601 (14.9) 550 (16.2) 364 (14.7) 330  (17.0) 149 (15.0) 545 (16.0) 131 (21.0) 563 (14.9) 87 (19.2) 607 (15.3) 176 (18.3) 518 (15.0) Pacemaker 123 (3.0) 109 (3.2) 91  (3.7) 34 (1.8) 22  (2.2) 103 (3.0) 14  (2.2) 111 (2.9) 9  (2.0) 116 (2.9) 13  (1.4) 112 (3.2) PCI 111 (2.7) 104 (3.1) 82  (3.3) 48  (2.5) 28  (2.8) 102 (3.0) 9  (1.4) 121 (3.2) 14  (3.1) 116 (2.9) 20  (2.1) 110 (3.2) CABG 98  (2.4) 55  (1.6) 50  (2.0) 22  (1.1) 19 (1.9) 53  (1.6) 7  (1.1) 65  (1.7) <5   68  (1.7) 7  (.7) 65  (1.9) Valve surgery 73  (1.8) 55  (1.6) 46  (1.9) 20  (1.0) 7  (.7) 59  (1.7) 12  (1.9) 54  (1.4) 6  (1.3) 60  (1.5) 11  (1.1) 55  (1.6) Dialysis 21  (.5) 10  (.3) 11  (.4) 0   <5  10  (.3) 0   11  (.3) 0   11  (.3) 0   11  (.3) Defibrillator implantation 10  (.2) 24  (.7) 17  (.7) 8  (.4) <5   21  (.6) <5   21 (.6) <5  24  (.6) <5   21  (.6) Comorbidities             250     Characteristics Total  (N = 7,439) Follow-up Assessments (N = 4,412) Initial consult Resp n=4,040 Non-resp n=3,398 6 months  (>0 to ≤0.5year) Resp n=2,472 (56.0%) Non-resp n=1,940 (44.0%) 1 year (>0.5 to ≤1) Resp n=996  (22.6%) Non-resp n=3,416 (77.4%) 1.5 years (>1 to ≤1.5) Resp n=624 (14.1%) Non-resp n=3,788 (85.9%) 2 years (>1.5 to ≤2) Resp n=454 (10.3%) Non-resp n=3,958 (89.7%) >2 years Resp n=962 (21.8%) Non-resp n=3,450 (78.2%) Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp   Diabetes 575 (14.2) 423 (12.4) 329 (13.3) 240  (12.4) 144 (14.5) 425 (12.4) 79 (12.7) 490 (12.9) 46 (10.1) 523 (13.2) 89  (9.3) 480 (13.9) COPD 214 (5.3) 206 (6.1) 173  (7.0) 94  (4.8) 59  (5.9) 208 (6.1) 36  (5.8) 231 (6.1) 33  (7.3) 234 (5.9) 40  (4.2) 227 (6.6)   Depression 184 (4.6) 146 (4.3) 118  (4.8) 77  (4.0) 43  (4.3) 152 (4.4) 25  (4.0) 170 (4.5) 15  (3.3) 180 (4.5) 38  (4.0) 157 (4.6)   CKD 164 (4.1) 102 (3.0) 78  (3.2) 50  (2.6) 24  (2.4) 104 (3.0) 13  (2.1) 115 (3.0) 9  (2.0) 119 (3.0) 19  (2.0) 109 (3.2)   Sleep disorder 106 (2.6) 104 (3.1) 84  (3.4) 48  (2.5) 21  (2.1) 111 (3.2) 22  (3.5) 110 (2.9) 14  (3.1) 118 (3.0) 22  (2.3) 110 (3.2) Hypothyroidism 91  (2.3) 73  (2.1) 57  (2.3) 38  (2.0) 26  (2.6) 69  (2.0) 12  (1.9) 83  (2.2) 7  (1.5) 88  (2.2) 17  (1.8) 78  (2.3)   GI bleed 83  (2.1) 67  (2.0) 56  (2.3) 33  (1.7) 15  (1.5) 74  (2.2) 12  (1.9) 77  (2.0) 6  (1.3) 83  (2.1) 10  (1.0) 79  (2.3)   Peptic ulcer 7  (.2) 8  (.2) 8  (.3) <5   0   10  (.3) <5 9  (.2) 0   10  (.3) <5  9  (.3) Prior medications               Beta-blockers 2,391 (59.2) 1,971 (58.0) 1,487 (60.2) 1,111 (57.3) 612 (61.4) 1,986 (58.1) 398 (63.8) 2,200 (58.1) 269 (59.3) 2,329 (58.8) 532 (55.3) 2,066 (59.9)   Antiarrhythmics 1,450 (35.9) 1,239 (36.5) 761 (30.8) 822  (42.4) 307 (30.8) 1,276 (37.4) 252 (40.4) 1,331 (35.1) 218 (48.0) 1,365 (34.5) 456 (47.4) 1,127 (32.7)   Calcium channel blockers 882 (21.8) 853 (25.1) 573 (23.2) 492  (25.4) 189 (19.0) 876 (25.6) 146 (23.4) 919 (24.3) 118 (26.0) 947 (23.9) 267 (27.8) 798 (23.1)   Digoxin 469 (11.6) 407 (12.0) 286 (11.6) 241  (12.4) 101 (10.1) 426 (12.5) 73 (11.7) 454 (12.0) 69 (15.2) 458 (11.6) 123 (12.8) 404 (11.7)   Antiplatelets 210 (5.2) 143 (4.2) 123 (5.0) 69  (3.6) 46  (4.6) 146 (4.3) 25  (4.0) 167 (4.4) 15  (3.3) 177 (4.5) 28  (2.9) 164 (4.8) Time-varying             251     Characteristics Total  (N = 7,439) Follow-up Assessments (N = 4,412) Initial consult Resp n=4,040 Non-resp n=3,398 6 months  (>0 to ≤0.5year) Resp n=2,472 (56.0%) Non-resp n=1,940 (44.0%) 1 year (>0.5 to ≤1) Resp n=996  (22.6%) Non-resp n=3,416 (77.4%) 1.5 years (>1 to ≤1.5) Resp n=624 (14.1%) Non-resp n=3,788 (85.9%) 2 years (>1.5 to ≤2) Resp n=454 (10.3%) Non-resp n=3,958 (89.7%) >2 years Resp n=962 (21.8%) Non-resp n=3,450 (78.2%) Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Resp Non-resp Ablation  157 (3.9) 296 (8.7) 369 (14.9) 398  (20.5) 73 (7.3) 303 (8.9) 44  (7.1) 154 (4.1) 40  (8.8) 132 (3.3) 261 (27.1) 148 (4.3)   Anticoagulant 2,456 (60.8) 2,158 (63.5) 1,870 (75.6) 1,380 (71.1) 729 (73.2) 2,185 (64.0) 445 (71.3) 2,189 (57.8) 302 (66.5) 2,068 (52.2) 796 (82.7) 1,691 (49.0) Note. Data shown as N (% within responder category) unless stated otherwise. Cell size < 5 not reported as per policy. Results based on imputed dataset 1. CCS-SAF = The Canadian Cardiovascular Society Severity in Atrial Fibrillation Scale; TIA = Transient ischemic attack; MI = Myocardial infarction; PVD = Peripheral vascular disease; PCI = Percutaneous coronary intervention; CABG = Coronary artery bypass graft; COPD = Chronic obstructive pulmonary disease; CKD = Chronic kidney disease; GI bleed = Gastrointestinal bleed. 252  Model Syntax Mplus DATA: FILE IS imp1.dat; !TYPE=imputation;  !Used after starting values are obtained for imputed datasets  DATA LONGTOWIDE:     !Dataset transformed from long to wide format LONG=afeqt|symp|act|conc|diff; WIDE=afeqt_0-afeqt_10|symp_0-symp_10|act_0-act_10|conc_0-conc_10|diff_0-diff_10; IDVARIABLE=id; REPETITION=follow;  VARIABLE: NAMES ARE id consultD age sex CHADS2 CCS_SAF distance clinic hf MI hyperten PVC stroke cardio_v ablation CABG PCI valve_sx pacemake defib dialysis COPD sleep bleed ulcer CKD hypothyr diabetes depress VASC CCI anticoag antiplat betabloc antiarry cal_bloc digoxin six_m_ab one_y_ab one5y_ab two_y_ab twopl_ab six_m_co one_y_co one5y_co two_y_co twopl_co consultI six_m one_y one5_y two_y moretwoy deathD deathI SurveyD Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Q12 Q13 Q14 Q15 Q16 Q17 Q18 follow diff afeqt symp act conc; MISSING=.;  IDVARIABLE=id; USEVARIABLES=afeqt_0 afeqt_1 afeqt_2 diff_0 diff_1 diff_2; TSCORES=diff_0 diff_1 diff_2;    !Individually-varying times (years) variable  CLASSES=c(1);       !Used for GMM to request # of requested classes  ANALYSIS: TYPE = mixture random; !Used for GMM with individually varying times of observation                      !Starts=100 10;     !Increases random starts (2 or more classes)                        Processors=4(starts);    !Reduces computation time  MODEL: !UNRESTRICTED MODEL  %OVERALL%       !Specifies the overall model  i s | afeqt_0 afeqt_1 afeqt_2 AT diff_0 diff_1 diff_2;    %C#1%       !Specific estimates for class 1, change to %C#2% for                                                                                          class 2  [i s];                           !Freely estimated latent means (intercept and slope)  i s with i s;       !Freely estimated covariances  i;s;       !Freely estimated variances  afeqt_0-afeqt_2(resvar1); !Residual variances fixed within each class (resvar2 for second class)  !RESTRICTED STANDARD MODEL %OVERALL%       !Specifies the overall model  i s | afeqt_0 afeqt_1 afeqt_2 AT diff_0 diff_1 diff_2;  s@0;       !Fixes slope variances to zero within class  i;       !Freely estimated intercept   %C#1%       !Specific estimates for class 1, change to %C#2% for         class 2  [i s];                           !Freely estimated latent means (intercept and slope)  i;       !Freely estimated intercept variance 253   afeqt_0-afeqt_2(resvar1); !Residual variances fixed within each class (resvar2 for class 2)  !RESTRICTED AR1 MODEL %OVERALL% !specifies the overall model   !Specifies the overall model  i s | afeqt_0 afeqt_1 afeqt_2 AT diff_0 diff_1 diff_2;  s@0;                            ! Fixes slope variances to zero within class  %C#1%  !Specific estimates for class 1, change to %C#2% for                                                                                 class 2  [i s];                           !Freely estimated latent means (intercept and slope)  afeqt_0-afeqt_2 (resvar1);   !Residual variances fixed within each class (resvar2 for class 2)  afeqt_0-afeqt_1 PWITH afeqt_1-afeqt_2(p11);  !Paired residual correlations (p12 for class 2)  afeqt_0 WITH afeqt_2(p31);    !Correlated residuals (p32 for class 2)  MODEL CONSTRAINT: NEW (corr);       p11=resvar1*corr;p31=resvar1*corr**2; !Specifies autoregressive structure AR1 for class 1 residual OUTPUT: svalues;  !Used to obtain starting values to run multiple imputed datasets      

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}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            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:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0390462/manifest

Comment

Related Items